# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=SI %s

#  Unaligned access is assumed on for HSA, but not mesa
# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=bonaire -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=CI-HSA %s
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire -O0 -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck -check-prefix=CI-MESA %s

# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -O0 -run-pass=legalizer  -global-isel-abort=0 %s -o - | FileCheck -check-prefix=VI %s
# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -O0 -run-pass=legalizer  -global-isel-abort=0 %s -o - | FileCheck -check-prefixes=GFX9-HSA %s
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -O0 -run-pass=legalizer  -global-isel-abort=0 %s -o - | FileCheck -check-prefixes=GFX9-MESA %s

---
name: test_load_global_s1_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s1_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; SI: $vgpr0 = COPY [[AND]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s1_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; CI-HSA: $vgpr0 = COPY [[AND]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s1_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; CI-MESA: $vgpr0 = COPY [[AND]](s32)
    ; VI-LABEL: name: test_load_global_s1_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; VI: $vgpr0 = COPY [[AND]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s1_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; GFX9-HSA: $vgpr0 = COPY [[AND]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s1_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; GFX9-MESA: $vgpr0 = COPY [[AND]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s1) = G_LOAD %0 :: (load 1, align 1, addrspace 1)
    %2:_(s32) = G_ZEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s2_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s2_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; SI: $vgpr0 = COPY [[AND]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s2_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; CI-HSA: $vgpr0 = COPY [[AND]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s2_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; CI-MESA: $vgpr0 = COPY [[AND]](s32)
    ; VI-LABEL: name: test_load_global_s2_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; VI: $vgpr0 = COPY [[AND]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s2_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; GFX9-HSA: $vgpr0 = COPY [[AND]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s2_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; GFX9-MESA: $vgpr0 = COPY [[AND]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s2) = G_LOAD %0 :: (load 1, align 1, addrspace 1)
    %2:_(s32) = G_ZEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s8_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_s8_align4
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI: $vgpr0 = COPY [[COPY1]](s32)
    ; SI-LABEL: name: test_load_global_s8_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s8_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s8_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
    ; VI-LABEL: name: test_load_global_s8_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s8_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s8_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s8) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s8_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s8_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s8_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s8_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
    ; VI-LABEL: name: test_load_global_s8_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s8_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s8_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s8) = G_LOAD %0 :: (load 1, align 1, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s16_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s16_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s16_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s16_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
    ; VI-LABEL: name: test_load_global_s16_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s16_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s16_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s16) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s16_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s16_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s16_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s16_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
    ; VI-LABEL: name: test_load_global_s16_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s16_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s16_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s16) = G_LOAD %0 :: (load 2, align 2, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s16_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_s16_align1
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[C1]](s32)
    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[TRUNC]]
    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[AND1]](s32)
    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC2]]
    ; CI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; SI-LABEL: name: test_load_global_s16_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s16_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s16_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; VI-LABEL: name: test_load_global_s16_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s16_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s16_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s16) = G_LOAD %0 :: (load 2, align 1, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s32_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s32_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s32_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s32_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: $vgpr0 = COPY [[LOAD]](s32)
    ; VI-LABEL: name: test_load_global_s32_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s32_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s32_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s32) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_s32_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s32_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: $vgpr0 = COPY [[OR]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s32_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s32_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: $vgpr0 = COPY [[OR]](s32)
    ; VI-LABEL: name: test_load_global_s32_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: $vgpr0 = COPY [[OR]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s32_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s32_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: $vgpr0 = COPY [[OR]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s32) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_s32_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s32_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: $vgpr0 = COPY [[OR2]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s32_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s32_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: $vgpr0 = COPY [[OR2]](s32)
    ; VI-LABEL: name: test_load_global_s32_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: $vgpr0 = COPY [[OR2]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s32_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s32_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: $vgpr0 = COPY [[OR2]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s32) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_s24_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s24_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 8, addrspace 1)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s24_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 8, addrspace 1)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s24_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 8, addrspace 1)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
    ; VI-LABEL: name: test_load_global_s24_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 8, addrspace 1)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s24_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 8, addrspace 1)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s24_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 8, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s24) = G_LOAD %0 :: (load 3, align 8, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s24_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s24_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s24_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s24_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: $vgpr0 = COPY [[COPY1]](s32)
    ; VI-LABEL: name: test_load_global_s24_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s24_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s24_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s24) = G_LOAD %0 :: (load 3, align 4, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s24_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s24_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, align 2, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s24_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, align 2, addrspace 1)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL]], [[ZEXTLOAD]]
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[OR]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s24_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, align 2, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; VI-LABEL: name: test_load_global_s24_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, align 2, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s24_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, align 2, addrspace 1)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL]], [[ZEXTLOAD]]
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[OR]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY1]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s24_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, align 2, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s24) = G_LOAD %0 :: (load 3, align 2, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s24_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s24_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-HSA-LABEL: name: test_load_global_s24_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; CI-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[LOAD1]], [[C2]](s32)
    ; CI-HSA: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL]], [[AND]]
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[OR]](s32)
    ; CI-HSA: $vgpr0 = COPY [[COPY2]](s32)
    ; CI-MESA-LABEL: name: test_load_global_s24_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; VI-LABEL: name: test_load_global_s24_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_s24_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[LOAD1]], [[C2]](s32)
    ; GFX9-HSA: [[OR:%[0-9]+]]:_(s32) = G_OR [[SHL]], [[AND]]
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[OR]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[COPY2]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_s24_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s16) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(s24) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(s24) = G_INSERT [[DEF]], [[LOAD]](s16), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(s24) = G_INSERT [[INSERT]], [[LOAD1]](s8), 16
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[INSERT1]](s24)
    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s24) = G_LOAD %0 :: (load 3, align 1, addrspace 1)
    %2:_(s32) = G_ANYEXT %1
    $vgpr0 = COPY %2
...

---
name: test_load_global_s48_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_s48_align8
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 6, align 8, addrspace 1)
    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
    ; CI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
    ; CI: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
    ; CI: $vgpr0_vgpr1 = COPY [[AND]](s64)
    ; SI-LABEL: name: test_load_global_s48_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
    ; SI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
    ; SI: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
    ; SI: $vgpr0_vgpr1 = COPY [[AND]](s64)
    ; CI-HSA-LABEL: name: test_load_global_s48_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[AND]](s64)
    ; CI-MESA-LABEL: name: test_load_global_s48_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[AND]](s64)
    ; VI-LABEL: name: test_load_global_s48_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
    ; VI: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
    ; VI: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
    ; VI: $vgpr0_vgpr1 = COPY [[AND]](s64)
    ; GFX9-HSA-LABEL: name: test_load_global_s48_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[AND]](s64)
    ; GFX9-MESA-LABEL: name: test_load_global_s48_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 281474976710655
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s64) = COPY [[LOAD]](s64)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s64) = G_AND [[COPY1]], [[C]]
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[AND]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s48) = G_LOAD %0 :: (load 6, align 8, addrspace 1)
    %2:_(s64) = G_ZEXT %1
    $vgpr0_vgpr1 = COPY %2
...

---
name: test_load_global_s64_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s64_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; CI-HSA-LABEL: name: test_load_global_s64_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; CI-MESA-LABEL: name: test_load_global_s64_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; VI-LABEL: name: test_load_global_s64_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; GFX9-HSA-LABEL: name: test_load_global_s64_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; GFX9-MESA-LABEL: name: test_load_global_s64_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_s64_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s64_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; CI-HSA-LABEL: name: test_load_global_s64_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; CI-MESA-LABEL: name: test_load_global_s64_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; VI-LABEL: name: test_load_global_s64_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; GFX9-HSA-LABEL: name: test_load_global_s64_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; GFX9-MESA-LABEL: name: test_load_global_s64_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_s64_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s64_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[MV]](s64)
    ; CI-HSA-LABEL: name: test_load_global_s64_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; CI-MESA-LABEL: name: test_load_global_s64_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
    ; VI-LABEL: name: test_load_global_s64_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[MV]](s64)
    ; GFX9-HSA-LABEL: name: test_load_global_s64_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; GFX9-MESA-LABEL: name: test_load_global_s64_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 8, align 2, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_s64_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s64_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[MV]](s64)
    ; CI-HSA-LABEL: name: test_load_global_s64_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; CI-MESA-LABEL: name: test_load_global_s64_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
    ; VI-LABEL: name: test_load_global_s64_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[MV]](s64)
    ; GFX9-HSA-LABEL: name: test_load_global_s64_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; GFX9-MESA-LABEL: name: test_load_global_s64_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; GFX9-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; GFX9-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_s96_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s96_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[EXTRACT:%[0-9]+]]:_(<3 x s32>) = G_EXTRACT [[LOAD]](<4 x s32>), 0
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[EXTRACT]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-HSA-LABEL: name: test_load_global_s96_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-MESA-LABEL: name: test_load_global_s96_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; VI-LABEL: name: test_load_global_s96_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-HSA-LABEL: name: test_load_global_s96_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-MESA-LABEL: name: test_load_global_s96_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s96) = G_LOAD %0 :: (load 12, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_s96_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s96_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 8, align 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-HSA-LABEL: name: test_load_global_s96_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-MESA-LABEL: name: test_load_global_s96_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; VI-LABEL: name: test_load_global_s96_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-HSA-LABEL: name: test_load_global_s96_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-MESA-LABEL: name: test_load_global_s96_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s96) = G_LOAD %0 :: (load 12, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_s96_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s96_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-HSA-LABEL: name: test_load_global_s96_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-MESA-LABEL: name: test_load_global_s96_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; VI-LABEL: name: test_load_global_s96_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-HSA-LABEL: name: test_load_global_s96_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-MESA-LABEL: name: test_load_global_s96_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s96) = G_LOAD %0 :: (load 12, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_s96_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s96_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-HSA-LABEL: name: test_load_global_s96_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-MESA-LABEL: name: test_load_global_s96_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; VI-LABEL: name: test_load_global_s96_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-HSA-LABEL: name: test_load_global_s96_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-MESA-LABEL: name: test_load_global_s96_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s96) = G_LOAD %0 :: (load 12, align 2, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_s96_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s96_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-HSA-LABEL: name: test_load_global_s96_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; CI-MESA-LABEL: name: test_load_global_s96_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; VI-LABEL: name: test_load_global_s96_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; VI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; VI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; VI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; VI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; VI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; VI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-HSA-LABEL: name: test_load_global_s96_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    ; GFX9-MESA-LABEL: name: test_load_global_s96_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s96) = G_LOAD %0 :: (load 12, align 1, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_s160_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_s160_align4
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s256) = G_LOAD [[COPY]](p1) :: (load 20, align 4, addrspace 1)
    ; CI: [[TRUNC:%[0-9]+]]:_(s160) = G_TRUNC [[LOAD]](s256)
    ; CI: S_NOP 0, implicit [[TRUNC]](s160)
    ; SI-LABEL: name: test_load_global_s160_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 16, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<5 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 128
    ; SI: [[BITCAST:%[0-9]+]]:_(s160) = G_BITCAST [[INSERT1]](<5 x s32>)
    ; SI: S_NOP 0, implicit [[BITCAST]](s160)
    ; CI-HSA-LABEL: name: test_load_global_s160_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 16, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<5 x s32>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; CI-HSA: [[INSERT1:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 128
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s160) = G_BITCAST [[INSERT1]](<5 x s32>)
    ; CI-HSA: S_NOP 0, implicit [[BITCAST]](s160)
    ; CI-MESA-LABEL: name: test_load_global_s160_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 16, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<5 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 128
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s160) = G_BITCAST [[INSERT1]](<5 x s32>)
    ; CI-MESA: S_NOP 0, implicit [[BITCAST]](s160)
    ; VI-LABEL: name: test_load_global_s160_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 16, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<5 x s32>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 128
    ; VI: [[BITCAST:%[0-9]+]]:_(s160) = G_BITCAST [[INSERT1]](<5 x s32>)
    ; VI: S_NOP 0, implicit [[BITCAST]](s160)
    ; GFX9-HSA-LABEL: name: test_load_global_s160_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 16, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<5 x s32>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; GFX9-HSA: [[INSERT1:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 128
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s160) = G_BITCAST [[INSERT1]](<5 x s32>)
    ; GFX9-HSA: S_NOP 0, implicit [[BITCAST]](s160)
    ; GFX9-MESA-LABEL: name: test_load_global_s160_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 16, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<5 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<5 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 128
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s160) = G_BITCAST [[INSERT1]](<5 x s32>)
    ; GFX9-MESA: S_NOP 0, implicit [[BITCAST]](s160)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s160) = G_LOAD %0 :: (load 20, align 4, addrspace 1)
    S_NOP 0, implicit %1
...

---
name: test_load_global_s224_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s224_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, align 4, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 4 + 24, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD1]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD2]](s32), 64
    ; SI: [[DEF1:%[0-9]+]]:_(<7 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT2:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[DEF1]], [[LOAD]](<4 x s32>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[INSERT2]], [[INSERT1]](<3 x s32>), 128
    ; SI: [[BITCAST:%[0-9]+]]:_(s224) = G_BITCAST [[INSERT3]](<7 x s32>)
    ; SI: [[DEF2:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
    ; SI: [[INSERT4:%[0-9]+]]:_(s256) = G_INSERT [[DEF2]], [[BITCAST]](s224), 0
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT4]](s256)
    ; CI-HSA-LABEL: name: test_load_global_s224_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 16, align 4, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<7 x s32>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; CI-HSA: [[INSERT1:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](<3 x s32>), 128
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s224) = G_BITCAST [[INSERT1]](<7 x s32>)
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT2:%[0-9]+]]:_(s256) = G_INSERT [[DEF1]], [[BITCAST]](s224), 0
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](s256)
    ; CI-MESA-LABEL: name: test_load_global_s224_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 16, align 4, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<7 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](<3 x s32>), 128
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s224) = G_BITCAST [[INSERT1]](<7 x s32>)
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(s256) = G_INSERT [[DEF1]], [[BITCAST]](s224), 0
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](s256)
    ; VI-LABEL: name: test_load_global_s224_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 16, align 4, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<7 x s32>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](<3 x s32>), 128
    ; VI: [[BITCAST:%[0-9]+]]:_(s224) = G_BITCAST [[INSERT1]](<7 x s32>)
    ; VI: [[DEF1:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
    ; VI: [[INSERT2:%[0-9]+]]:_(s256) = G_INSERT [[DEF1]], [[BITCAST]](s224), 0
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](s256)
    ; GFX9-HSA-LABEL: name: test_load_global_s224_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 16, align 4, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<7 x s32>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; GFX9-HSA: [[INSERT1:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](<3 x s32>), 128
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s224) = G_BITCAST [[INSERT1]](<7 x s32>)
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT2:%[0-9]+]]:_(s256) = G_INSERT [[DEF1]], [[BITCAST]](s224), 0
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](s256)
    ; GFX9-MESA-LABEL: name: test_load_global_s224_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 16, align 4, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<7 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[DEF]], [[LOAD]](<4 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<7 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](<3 x s32>), 128
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s224) = G_BITCAST [[INSERT1]](<7 x s32>)
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(s256) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(s256) = G_INSERT [[DEF1]], [[BITCAST]](s224), 0
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](s256)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s224) = G_LOAD %0 :: (load 28, align 4, addrspace 1)
     %2:_(s256) = G_IMPLICIT_DEF
     %3:_(s256) = G_INSERT %2, %1, 0
     $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3

...

---
name: test_load_global_s128_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s128_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; CI-HSA-LABEL: name: test_load_global_s128_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; CI-MESA-LABEL: name: test_load_global_s128_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; VI-LABEL: name: test_load_global_s128_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; GFX9-HSA-LABEL: name: test_load_global_s128_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; GFX9-MESA-LABEL: name: test_load_global_s128_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s128) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_s128_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_s128_align4
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s128) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](s128)
    ; SI-LABEL: name: test_load_global_s128_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; CI-HSA-LABEL: name: test_load_global_s128_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; CI-MESA-LABEL: name: test_load_global_s128_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; VI-LABEL: name: test_load_global_s128_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; GFX9-HSA-LABEL: name: test_load_global_s128_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; GFX9-MESA-LABEL: name: test_load_global_s128_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s128) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_s128_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s128_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; SI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; SI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; SI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; SI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; SI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; SI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; SI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; SI: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; CI-HSA-LABEL: name: test_load_global_s128_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; CI-MESA-LABEL: name: test_load_global_s128_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; CI-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; CI-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; CI-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; CI-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; CI-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; VI-LABEL: name: test_load_global_s128_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; VI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; VI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; VI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; VI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; VI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; VI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; VI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; VI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; VI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; VI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; VI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; VI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; VI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; VI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; VI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; VI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; VI: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; GFX9-HSA-LABEL: name: test_load_global_s128_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    ; GFX9-MESA-LABEL: name: test_load_global_s128_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s128) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](s128)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s128) = G_LOAD %0 :: (load 16, align 1, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_s256_align32
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_s256_align32
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(s256) = G_BITCAST [[LOAD]](<8 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](s256)
    ; CI-HSA-LABEL: name: test_load_global_s256_align32
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s256) = G_BITCAST [[LOAD]](<8 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](s256)
    ; CI-MESA-LABEL: name: test_load_global_s256_align32
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s256) = G_BITCAST [[LOAD]](<8 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](s256)
    ; VI-LABEL: name: test_load_global_s256_align32
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s256) = G_BITCAST [[LOAD]](<8 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](s256)
    ; GFX9-HSA-LABEL: name: test_load_global_s256_align32
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s256) = G_BITCAST [[LOAD]](<8 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](s256)
    ; GFX9-MESA-LABEL: name: test_load_global_s256_align32
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s256) = G_BITCAST [[LOAD]](<8 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](s256)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s256) = G_LOAD %0 :: (load 32, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---
name: test_load_global_p1_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p1_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; CI-HSA-LABEL: name: test_load_global_p1_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; CI-MESA-LABEL: name: test_load_global_p1_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; VI-LABEL: name: test_load_global_p1_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; GFX9-HSA-LABEL: name: test_load_global_p1_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; GFX9-MESA-LABEL: name: test_load_global_p1_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p1) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_p1_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p1_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; CI-HSA-LABEL: name: test_load_global_p1_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; CI-MESA-LABEL: name: test_load_global_p1_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; VI-LABEL: name: test_load_global_p1_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; GFX9-HSA-LABEL: name: test_load_global_p1_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; GFX9-MESA-LABEL: name: test_load_global_p1_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p1) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_p1_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p1_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p1)
    ; CI-HSA-LABEL: name: test_load_global_p1_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; CI-MESA-LABEL: name: test_load_global_p1_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](p1)
    ; VI-LABEL: name: test_load_global_p1_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p1)
    ; GFX9-HSA-LABEL: name: test_load_global_p1_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
    ; GFX9-MESA-LABEL: name: test_load_global_p1_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; GFX9-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; GFX9-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p1) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](p1)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p1) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_p3_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p3_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: $vgpr0 = COPY [[LOAD]](p3)
    ; CI-HSA-LABEL: name: test_load_global_p3_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p3)
    ; CI-MESA-LABEL: name: test_load_global_p3_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: $vgpr0 = COPY [[LOAD]](p3)
    ; VI-LABEL: name: test_load_global_p3_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: $vgpr0 = COPY [[LOAD]](p3)
    ; GFX9-HSA-LABEL: name: test_load_global_p3_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p3)
    ; GFX9-MESA-LABEL: name: test_load_global_p3_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p3) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](p3)
     %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p3) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_p4_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_p4_align8
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; SI-LABEL: name: test_load_global_p4_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; CI-HSA-LABEL: name: test_load_global_p4_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; CI-MESA-LABEL: name: test_load_global_p4_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; VI-LABEL: name: test_load_global_p4_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; GFX9-HSA-LABEL: name: test_load_global_p4_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; GFX9-MESA-LABEL: name: test_load_global_p4_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p4) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_p4_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p4_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; CI-HSA-LABEL: name: test_load_global_p4_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; CI-MESA-LABEL: name: test_load_global_p4_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; VI-LABEL: name: test_load_global_p4_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; GFX9-HSA-LABEL: name: test_load_global_p4_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; GFX9-MESA-LABEL: name: test_load_global_p4_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p4) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_p4_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p4_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p4)
    ; CI-HSA-LABEL: name: test_load_global_p4_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; CI-MESA-LABEL: name: test_load_global_p4_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
    ; VI-LABEL: name: test_load_global_p4_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p4)
    ; GFX9-HSA-LABEL: name: test_load_global_p4_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; GFX9-MESA-LABEL: name: test_load_global_p4_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p4) = G_LOAD %0 :: (load 8, align 2, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_p4_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p4_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[MV]](p4)
    ; CI-HSA-LABEL: name: test_load_global_p4_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; CI-MESA-LABEL: name: test_load_global_p4_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
    ; VI-LABEL: name: test_load_global_p4_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[MV]](p4)
    ; GFX9-HSA-LABEL: name: test_load_global_p4_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p4) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](p4)
    ; GFX9-MESA-LABEL: name: test_load_global_p4_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; GFX9-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; GFX9-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(p4) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[MV]](p4)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p4) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_p5_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p5_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: $vgpr0 = COPY [[LOAD]](p5)
    ; CI-HSA-LABEL: name: test_load_global_p5_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p5)
    ; CI-MESA-LABEL: name: test_load_global_p5_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: $vgpr0 = COPY [[LOAD]](p5)
    ; VI-LABEL: name: test_load_global_p5_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: $vgpr0 = COPY [[LOAD]](p5)
    ; GFX9-HSA-LABEL: name: test_load_global_p5_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p5)
    ; GFX9-MESA-LABEL: name: test_load_global_p5_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](p5)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p5) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_p5_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p5_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR]](s32)
    ; SI: $vgpr0 = COPY [[INTTOPTR]](p5)
    ; CI-HSA-LABEL: name: test_load_global_p5_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p5)
    ; CI-MESA-LABEL: name: test_load_global_p5_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR]](s32)
    ; CI-MESA: $vgpr0 = COPY [[INTTOPTR]](p5)
    ; VI-LABEL: name: test_load_global_p5_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR]](s32)
    ; VI: $vgpr0 = COPY [[INTTOPTR]](p5)
    ; GFX9-HSA-LABEL: name: test_load_global_p5_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p5)
    ; GFX9-MESA-LABEL: name: test_load_global_p5_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[INTTOPTR]](p5)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p5) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_p5_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_p5_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR2]](s32)
    ; SI: $vgpr0 = COPY [[INTTOPTR]](p5)
    ; CI-HSA-LABEL: name: test_load_global_p5_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](p5)
    ; CI-MESA-LABEL: name: test_load_global_p5_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR2]](s32)
    ; CI-MESA: $vgpr0 = COPY [[INTTOPTR]](p5)
    ; VI-LABEL: name: test_load_global_p5_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR2]](s32)
    ; VI: $vgpr0 = COPY [[INTTOPTR]](p5)
    ; GFX9-HSA-LABEL: name: test_load_global_p5_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(p5) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](p5)
    ; GFX9-MESA-LABEL: name: test_load_global_p5_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[INTTOPTR:%[0-9]+]]:_(p5) = G_INTTOPTR [[OR2]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[INTTOPTR]](p5)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(p5) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v2s8_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s8_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; SI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-HSA-LABEL: name: test_load_global_v2s8_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
    ; CI-HSA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; CI-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-HSA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-MESA-LABEL: name: test_load_global_v2s8_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; VI-LABEL: name: test_load_global_v2s8_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; VI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s8_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
    ; GFX9-HSA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-HSA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
    ; GFX9-HSA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s8_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
    %2:_(s16) = G_BITCAST %1
    %3:_(s32) = G_ANYEXT %2
    $vgpr0 = COPY %3
...

---
name: test_load_global_v2s8_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s8_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; SI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-HSA-LABEL: name: test_load_global_v2s8_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
    ; CI-HSA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; CI-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-HSA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-MESA-LABEL: name: test_load_global_v2s8_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; VI-LABEL: name: test_load_global_v2s8_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; VI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s8_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
    ; GFX9-HSA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-HSA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
    ; GFX9-HSA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s8_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 2, addrspace 1)
    %2:_(s16) = G_BITCAST %1
    %3:_(s32) = G_ANYEXT %2
    $vgpr0 = COPY %3
...

---
name: test_load_global_v2s8_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s8_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-HSA-LABEL: name: test_load_global_v2s8_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C]](s32)
    ; CI-HSA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; CI-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-HSA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; CI-MESA-LABEL: name: test_load_global_v2s8_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; CI-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; VI-LABEL: name: test_load_global_v2s8_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; VI: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s8_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 1, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-HSA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C3]]
    ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
    ; GFX9-HSA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C3]]
    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-HSA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C4]](s16)
    ; GFX9-HSA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-HSA: $vgpr0 = COPY [[ANYEXT]](s32)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s8_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: $vgpr0 = COPY [[ANYEXT]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s8>) = G_LOAD %0 :: (load 2, align 1, addrspace 1)
    %2:_(s16) = G_BITCAST %1
    %3:_(s32) = G_ANYEXT %2
    $vgpr0 = COPY %3
...

---
name: test_load_global_v3s8_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s8_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; SI: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v3s8_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; CI-HSA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v3s8_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; CI-MESA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; VI-LABEL: name: test_load_global_v3s8_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; VI: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s8_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; GFX9-HSA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s8_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 4, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; GFX9-MESA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s8>) = G_LOAD %0 :: (load 3, align 4, addrspace 1)
    %2:_(<4 x s8>) = G_IMPLICIT_DEF
    %3:_(<4 x s8>) = G_INSERT %2, %1, 0
    $vgpr0 = COPY %3

...

---
name: test_load_global_v3s8_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s8_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; SI: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v3s8_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; CI-HSA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v3s8_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; CI-MESA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; VI-LABEL: name: test_load_global_v3s8_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; VI: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s8_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; GFX9-HSA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s8_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s8>) = G_LOAD [[COPY]](p1) :: (load 3, align 1, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s8>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s8>) = G_INSERT [[DEF]], [[LOAD]](<3 x s8>), 0
    ; GFX9-MESA: $vgpr0 = COPY [[INSERT]](<4 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s8>) = G_LOAD %0 :: (load 3, align 1, addrspace 1)
    %2:_(<4 x s8>) = G_IMPLICIT_DEF
    %3:_(<4 x s8>) = G_INSERT %2, %1, 0
    $vgpr0 = COPY %3
...

---
name: test_load_global_v4s8_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s8_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; SI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v4s8_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v4s8_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; VI-LABEL: name: test_load_global_v4s8_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; VI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s8_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s8_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s8>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v4s8_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_v4s8_align2
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 2, addrspace 1)
    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
    ; CI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI: [[GEP1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, align 2, addrspace 1)
    ; CI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI: [[GEP2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; SI-LABEL: name: test_load_global_v4s8_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32)
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32)
    ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; SI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v4s8_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v4s8_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32)
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32)
    ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; VI-LABEL: name: test_load_global_v4s8_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32)
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32)
    ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; VI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s8_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s8_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32)
    ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s8>) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v4s8_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s8_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; SI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v4s8_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v4s8_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; VI-LABEL: name: test_load_global_v4s8_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; VI: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s8_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s8_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<4 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s8>) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v8s8_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v8s8_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>)
    ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v8s8_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-HSA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v8s8_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
    ; VI-LABEL: name: test_load_global_v8s8_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>)
    ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v8s8_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v8s8_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>)
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v16s8_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v16s8_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; SI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; SI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; SI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; SI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; SI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; SI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; SI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v16s8_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-HSA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; CI-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; CI-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; CI-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; CI-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; CI-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; CI-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; CI-HSA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; CI-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; CI-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; CI-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; CI-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; CI-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; CI-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; CI-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; CI-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; CI-HSA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; CI-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v16s8_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
    ; VI-LABEL: name: test_load_global_v16s8_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v16s8_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>)
    ; GFX9-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; GFX9-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; GFX9-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; GFX9-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; GFX9-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>)
    ; GFX9-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; GFX9-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; GFX9-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; GFX9-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32)
    ; GFX9-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; GFX9-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v16s8_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>)
    ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>)
    ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32)
    ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>)
    ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v32s8_align32
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v32s8_align32
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; SI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; SI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; SI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; SI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; SI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; SI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; SI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; SI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32)
    ; SI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32)
    ; SI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32)
    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32)
    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; SI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32)
    ; SI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>)
    ; SI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32)
    ; SI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32)
    ; SI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32)
    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32)
    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; SI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; SI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>)
    ; SI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32)
    ; SI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32)
    ; SI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32)
    ; SI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32)
    ; SI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; SI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; SI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; SI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32)
    ; SI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>)
    ; SI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32)
    ; SI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32)
    ; SI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32)
    ; SI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32)
    ; SI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; SI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; SI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; SI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>)
    ; CI-HSA-LABEL: name: test_load_global_v32s8_align32
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-HSA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; CI-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; CI-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; CI-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; CI-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; CI-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; CI-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; CI-HSA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; CI-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; CI-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; CI-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; CI-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; CI-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; CI-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; CI-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; CI-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; CI-HSA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; CI-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; CI-HSA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32)
    ; CI-HSA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32)
    ; CI-HSA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32)
    ; CI-HSA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32)
    ; CI-HSA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; CI-HSA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; CI-HSA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; CI-HSA: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32)
    ; CI-HSA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>)
    ; CI-HSA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32)
    ; CI-HSA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32)
    ; CI-HSA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32)
    ; CI-HSA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32)
    ; CI-HSA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; CI-HSA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; CI-HSA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; CI-HSA: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32)
    ; CI-HSA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>)
    ; CI-HSA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32)
    ; CI-HSA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32)
    ; CI-HSA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32)
    ; CI-HSA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32)
    ; CI-HSA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; CI-HSA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; CI-HSA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; CI-HSA: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32)
    ; CI-HSA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>)
    ; CI-HSA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32)
    ; CI-HSA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32)
    ; CI-HSA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32)
    ; CI-HSA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32)
    ; CI-HSA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; CI-HSA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; CI-HSA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; CI-HSA: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; CI-HSA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>)
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>)
    ; CI-MESA-LABEL: name: test_load_global_v32s8_align32
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; CI-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32)
    ; CI-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32)
    ; CI-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32)
    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32)
    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; CI-MESA: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32)
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>)
    ; CI-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32)
    ; CI-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32)
    ; CI-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32)
    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32)
    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; CI-MESA: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>)
    ; CI-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32)
    ; CI-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32)
    ; CI-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32)
    ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32)
    ; CI-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; CI-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; CI-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; CI-MESA: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32)
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>)
    ; CI-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32)
    ; CI-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32)
    ; CI-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32)
    ; CI-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32)
    ; CI-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; CI-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; CI-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; CI-MESA: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>)
    ; VI-LABEL: name: test_load_global_v32s8_align32
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>)
    ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32)
    ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>)
    ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32)
    ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>)
    ; VI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32)
    ; VI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32)
    ; VI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32)
    ; VI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32)
    ; VI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; VI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; VI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; VI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32)
    ; VI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>)
    ; VI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32)
    ; VI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32)
    ; VI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32)
    ; VI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32)
    ; VI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; VI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; VI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; VI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32)
    ; VI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>)
    ; VI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32)
    ; VI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32)
    ; VI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32)
    ; VI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32)
    ; VI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; VI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; VI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; VI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32)
    ; VI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>)
    ; VI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32)
    ; VI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32)
    ; VI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32)
    ; VI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32)
    ; VI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; VI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; VI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; VI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; VI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>)
    ; GFX9-HSA-LABEL: name: test_load_global_v32s8_align32
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>)
    ; GFX9-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; GFX9-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; GFX9-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; GFX9-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; GFX9-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>)
    ; GFX9-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; GFX9-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; GFX9-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; GFX9-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32)
    ; GFX9-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; GFX9-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>)
    ; GFX9-HSA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32)
    ; GFX9-HSA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32)
    ; GFX9-HSA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32)
    ; GFX9-HSA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32)
    ; GFX9-HSA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; GFX9-HSA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>)
    ; GFX9-HSA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32)
    ; GFX9-HSA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32)
    ; GFX9-HSA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32)
    ; GFX9-HSA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32)
    ; GFX9-HSA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; GFX9-HSA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>)
    ; GFX9-HSA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32)
    ; GFX9-HSA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32)
    ; GFX9-HSA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32)
    ; GFX9-HSA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32)
    ; GFX9-HSA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; GFX9-HSA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>)
    ; GFX9-HSA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32)
    ; GFX9-HSA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32)
    ; GFX9-HSA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32)
    ; GFX9-HSA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32)
    ; GFX9-HSA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32)
    ; GFX9-HSA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; GFX9-HSA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>)
    ; GFX9-MESA-LABEL: name: test_load_global_v32s8_align32
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32)
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>)
    ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32)
    ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32)
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>)
    ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32)
    ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32)
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32)
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32)
    ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>)
    ; GFX9-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32)
    ; GFX9-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32)
    ; GFX9-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32)
    ; GFX9-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32)
    ; GFX9-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; GFX9-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>)
    ; GFX9-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32)
    ; GFX9-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32)
    ; GFX9-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32)
    ; GFX9-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32)
    ; GFX9-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; GFX9-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>)
    ; GFX9-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32)
    ; GFX9-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32)
    ; GFX9-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32)
    ; GFX9-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32)
    ; GFX9-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; GFX9-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>)
    ; GFX9-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32)
    ; GFX9-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32)
    ; GFX9-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32)
    ; GFX9-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32)
    ; GFX9-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32)
    ; GFX9-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; GFX9-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>)
    ; GFX9-MESA: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<32 x s8>) = G_LOAD %0 :: (load 32, align 32, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---

name: test_load_global_v2s16_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s16_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v2s16_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v2s16_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v2s16_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s16_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s16_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v2s16_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s16_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v2s16_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v2s16_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v2s16_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s16_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s16_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v2s16_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_v2s16_align1
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 1, addrspace 1)
    ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[C1]](s32)
    ; CI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[TRUNC]]
    ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
    ; CI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[AND1]](s32)
    ; CI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC2]]
    ; CI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI: [[GEP1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 1, addrspace 1)
    ; CI: [[GEP2:%[0-9]+]]:_(p1) = G_PTR_ADD [[GEP1]], [[C]](s64)
    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 1, addrspace 1)
    ; CI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[TRUNC]]
    ; CI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C6]](s32)
    ; CI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C4]]
    ; CI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[AND4]](s32)
    ; CI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND3]], [[TRUNC4]]
    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[OR]](s16), [[OR1]](s16)
    ; CI: $vgpr0 = COPY [[BUILD_VECTOR]](<2 x s16>)
    ; SI-LABEL: name: test_load_global_v2s16_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v2s16_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v2s16_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v2s16_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s16_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s16_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s16>) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v3s16_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s16_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[LOAD]](<4 x s16>), [[DEF]](<4 x s16>), [[DEF]](<4 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[UV]](<3 x s16>), 0
    ; SI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v3s16_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[LOAD]](<4 x s16>), [[DEF]](<4 x s16>), [[DEF]](<4 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[UV]](<3 x s16>), 0
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v3s16_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[LOAD]](<4 x s16>), [[DEF]](<4 x s16>), [[DEF]](<4 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[UV]](<3 x s16>), 0
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v3s16_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[LOAD]](<4 x s16>), [[DEF]](<4 x s16>), [[DEF]](<4 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[UV]](<3 x s16>), 0
    ; VI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[LOAD]](<4 x s16>), [[DEF]](<4 x s16>), [[DEF]](<4 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[UV]](<3 x s16>), 0
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[LOAD]](<4 x s16>), [[DEF]](<4 x s16>), [[DEF]](<4 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[UV]](<3 x s16>), 0
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 8, addrspace 1)
    %2:_(<4 x s16>) = G_IMPLICIT_DEF
    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
    $vgpr0_vgpr1 = COPY %3
...

---
name: test_load_global_v3s16_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_v3s16_align4
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; CI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; CI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; SI-LABEL: name: test_load_global_v3s16_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; SI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v3s16_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v3s16_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v3s16_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; VI: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 4, addrspace 1)
    %2:_(<4 x s16>) = G_IMPLICIT_DEF
    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
    $vgpr0_vgpr1 = COPY %3
...

---
name: test_load_global_v3s16_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s16_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST]](<2 x s16>), 0
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; SI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; SI: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; SI: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v3s16_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 2, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v3s16_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST]](<2 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; CI-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; CI-MESA: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v3s16_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST]](<2 x s16>), 0
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; VI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; VI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; VI: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; VI: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 2, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR_TRUNC]](<2 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; GFX9-MESA: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 2, addrspace 1)
    %2:_(<4 x s16>) = G_IMPLICIT_DEF
    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
    $vgpr0_vgpr1 = COPY %3
...

---
name: test_load_global_v3s16_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s16_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST]](<2 x s16>), 0
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR3]](s16), 32
    ; SI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; SI: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; SI: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v3s16_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 1, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v3s16_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST]](<2 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR3]](s16), 32
    ; CI-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; CI-MESA: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v3s16_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST]](<2 x s16>), 0
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; VI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR3]](s16), 32
    ; VI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; VI: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; VI: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s16_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s16>) = G_LOAD [[COPY]](p1) :: (load 6, align 1, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF]], [[LOAD]](<3 x s16>), 0
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[INSERT]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s16_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<12 x s16>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR_TRUNC]](<2 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR2]](s16), 32
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; GFX9-MESA: [[INSERT4:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV8]](<3 x s16>), 0
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[INSERT4]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s16>) = G_LOAD %0 :: (load 6, align 1, addrspace 1)
    %2:_(<4 x s16>) = G_IMPLICIT_DEF
    %3:_(<4 x s16>) = G_INSERT %2, %1, 0
    $vgpr0_vgpr1 = COPY %3
...

---
name: test_load_global_v4s16_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s16_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v4s16_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v4s16_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v4s16_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v4s16_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s16_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v4s16_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v4s16_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v4s16_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v4s16_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s16_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v4s16_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v4s16_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v4s16_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 2, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v4s16_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s16_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C4]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C5]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v4s16_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v4s16_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C4]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C5]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    ; VI-LABEL: name: test_load_global_v4s16_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C3]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C4]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s16_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<4 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s16_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[OR2]](s16)
    ; GFX9-MESA: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[OR3]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[ANYEXT3]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s16>) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v5s16_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v5s16_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; SI: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; SI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; SI: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<6 x s16>)
    ; SI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST5:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST5]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST6]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST7]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v5s16_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; CI-HSA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; CI-HSA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; CI-HSA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; CI-HSA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<6 x s16>)
    ; CI-HSA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; CI-HSA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-HSA: [[BITCAST5:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; CI-HSA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-HSA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; CI-HSA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-HSA: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-HSA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-HSA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; CI-HSA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; CI-HSA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-HSA: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-HSA: $vgpr0 = COPY [[BITCAST5]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[BITCAST7]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v5s16_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; CI-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; CI-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; CI-MESA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<6 x s16>)
    ; CI-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[BITCAST5:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST5]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[BITCAST7]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v5s16_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; VI: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; VI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; VI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; VI: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<6 x s16>)
    ; VI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[BITCAST5:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST5]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[BITCAST6]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[BITCAST7]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v5s16_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; GFX9-HSA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; GFX9-HSA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; GFX9-HSA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; GFX9-HSA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<6 x s16>)
    ; GFX9-HSA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC2]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v5s16_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; GFX9-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; GFX9-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; GFX9-MESA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<6 x s16>)
    ; GFX9-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC2]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<5 x s16>) = G_LOAD %0 :: (load 10, align 16, addrspace 1)
    %2:_(<5 x s16>) = G_IMPLICIT_DEF
    %3:_(<10 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6

...

---
name: test_load_global_v5s16_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v5s16_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 8, align 8, addrspace 1)
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; SI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[LOAD]](<4 x s16>), 0
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 64
    ; SI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; SI: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C1]](s32)
    ; SI: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C1]](s32)
    ; SI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C1]](s32)
    ; SI: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C2]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C2]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C1]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST4:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C2]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C2]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BITCAST5:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C2]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C2]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C1]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST4]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST5]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v5s16_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 8, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; CI-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v5s16_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 8, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; CI-MESA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; CI-MESA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v5s16_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 8, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; VI: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; VI: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v5s16_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 8, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; GFX9-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v5s16_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 8, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; GFX9-MESA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<5 x s16>) = G_LOAD %0 :: (load 10, align 8, addrspace 1)
    %2:_(<5 x s16>) = G_IMPLICIT_DEF
    %3:_(<10 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6

...

---
name: test_load_global_v5s16_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v5s16_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s16>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 8, align 4, addrspace 1)
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; SI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[LOAD]](<4 x s16>), 0
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 64
    ; SI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; SI: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C1]](s32)
    ; SI: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C1]](s32)
    ; SI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C1]](s32)
    ; SI: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C2]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C2]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C1]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST4:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C2]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C2]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BITCAST5:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C2]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C2]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C1]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST4]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST5]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v5s16_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 4, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; CI-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v5s16_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 4, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; CI-MESA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; CI-MESA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v5s16_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 4, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; VI: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; VI: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v5s16_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 4, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; GFX9-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v5s16_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 4, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; GFX9-MESA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<5 x s16>) = G_LOAD %0 :: (load 10, align 4, addrspace 1)
    %2:_(<5 x s16>) = G_IMPLICIT_DEF
    %3:_(<10 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6

...

---
name: test_load_global_v5s16_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v5s16_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; SI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 64
    ; SI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; SI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C4]](s32)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; SI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; SI: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; SI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; SI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; SI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR4]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v5s16_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 2, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; CI-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v5s16_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-MESA: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 64
    ; CI-MESA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; CI-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C4]](s32)
    ; CI-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; CI-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; CI-MESA: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; CI-MESA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; CI-MESA: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; CI-MESA: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR4]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v5s16_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; VI: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; VI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; VI: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; VI: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 64
    ; VI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; VI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C4]](s32)
    ; VI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; VI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; VI: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; VI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; VI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR4]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v5s16_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 2, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; GFX9-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v5s16_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-MESA: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 64
    ; GFX9-MESA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C4]](s32)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C4]](s32)
    ; GFX9-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C4]](s32)
    ; GFX9-MESA: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; GFX9-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC2]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC4]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<5 x s16>) = G_LOAD %0 :: (load 10, align 2, addrspace 1)
    %2:_(<5 x s16>) = G_IMPLICIT_DEF
    %3:_(<10 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6

...

---
name: test_load_global_v5s16_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v5s16_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C4]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C5]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; SI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; SI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; SI: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[OR6]](s16), 64
    ; SI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; SI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C5]](s32)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C5]](s32)
    ; SI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C5]](s32)
    ; SI: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; SI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C5]](s32)
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C8]]
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C8]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C5]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL7]]
    ; SI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR7]](s32)
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C8]]
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C8]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C5]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL8]]
    ; SI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR8]](s32)
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C8]]
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C8]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C5]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND14]], [[SHL9]]
    ; SI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR9]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v5s16_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 1, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; CI-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v5s16_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C4]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C5]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; CI-MESA: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[OR6]](s16), 64
    ; CI-MESA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; CI-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C5]](s32)
    ; CI-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C5]](s32)
    ; CI-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C5]](s32)
    ; CI-MESA: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; CI-MESA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C5]](s32)
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C8]]
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C8]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C5]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL7]]
    ; CI-MESA: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR7]](s32)
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C8]]
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C8]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C5]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL8]]
    ; CI-MESA: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR8]](s32)
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C8]]
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C8]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C5]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND14]], [[SHL9]]
    ; CI-MESA: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR9]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v5s16_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C3]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C4]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C1]]
    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C2]](s16)
    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; VI: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; VI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; VI: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; VI: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[OR6]](s16), 64
    ; VI: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; VI: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; VI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C4]](s32)
    ; VI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; VI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; VI: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; VI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C7]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C7]]
    ; VI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C4]](s32)
    ; VI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL7]]
    ; VI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR7]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C7]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C7]]
    ; VI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; VI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL8]]
    ; VI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR8]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C7]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C7]]
    ; VI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C4]](s32)
    ; VI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND14]], [[SHL9]]
    ; VI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR9]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v5s16_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<5 x s16>) = G_LOAD [[COPY]](p1) :: (load 10, align 1, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<30 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<10 x s16>) = G_CONCAT_VECTORS [[LOAD]](<5 x s16>), [[UV]](<5 x s16>)
    ; GFX9-HSA: [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>), [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<10 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV6]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV7]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV8]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v5s16_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[OR2]](s16)
    ; GFX9-MESA: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[OR3]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[ANYEXT3]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C1]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C2]](s16)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[DEF]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<5 x s16>), [[UV1:%[0-9]+]]:_(<5 x s16>), [[UV2:%[0-9]+]]:_(<5 x s16>), [[UV3:%[0-9]+]]:_(<5 x s16>), [[UV4:%[0-9]+]]:_(<5 x s16>), [[UV5:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<30 x s16>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV]](<5 x s16>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<30 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>), [[DEF1]](<6 x s16>)
    ; GFX9-MESA: [[UV6:%[0-9]+]]:_(<5 x s16>), [[UV7:%[0-9]+]]:_(<5 x s16>), [[UV8:%[0-9]+]]:_(<5 x s16>), [[UV9:%[0-9]+]]:_(<5 x s16>), [[UV10:%[0-9]+]]:_(<5 x s16>), [[UV11:%[0-9]+]]:_(<5 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<30 x s16>)
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[DEF1]], [[UV6]](<5 x s16>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<6 x s16>) = G_INSERT [[INSERT2]], [[OR4]](s16), 64
    ; GFX9-MESA: [[DEF2:%[0-9]+]]:_(<6 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT3]](<6 x s16>)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV12]](<2 x s16>)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C6]](s32)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV13]](<2 x s16>)
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C6]](s32)
    ; GFX9-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV14]](<2 x s16>)
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C6]](s32)
    ; GFX9-MESA: [[UV15:%[0-9]+]]:_(<2 x s16>), [[UV16:%[0-9]+]]:_(<2 x s16>), [[UV17:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF2]](<6 x s16>)
    ; GFX9-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV15]](<2 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C6]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC2]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC4]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<5 x s16>) = G_LOAD %0 :: (load 10, align 1, addrspace 1)
    %2:_(<5 x s16>) = G_IMPLICIT_DEF
    %3:_(<10 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6

...

---
name: test_load_global_v6s16_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v6s16_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[EXTRACT:%[0-9]+]]:_(<3 x s32>) = G_EXTRACT [[LOAD]](<4 x s32>), 0
    ; SI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[EXTRACT]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v6s16_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v6s16_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; VI-LABEL: name: test_load_global_v6s16_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v6s16_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v6s16_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<6 x s16>) = G_LOAD %0 :: (load 12, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_v6s16_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v6s16_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 8, align 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v6s16_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v6s16_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; VI-LABEL: name: test_load_global_v6s16_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v6s16_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v6s16_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 8, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<6 x s16>) = G_LOAD %0 :: (load 12, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_v6s16_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v6s16_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v6s16_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v6s16_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; VI-LABEL: name: test_load_global_v6s16_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v6s16_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v6s16_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<6 x s16>) = G_LOAD %0 :: (load 12, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_v6s16_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v6s16_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v6s16_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v6s16_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; VI-LABEL: name: test_load_global_v6s16_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; VI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v6s16_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v6s16_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<6 x s16>) = G_LOAD %0 :: (load 12, align 2, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_v6s16_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v6s16_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v6s16_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v6s16_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; VI-LABEL: name: test_load_global_v6s16_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; VI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; VI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; VI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; VI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; VI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; VI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; VI: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v6s16_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v6s16_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<6 x s16>) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[BITCAST]](<6 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<6 x s16>) = G_LOAD %0 :: (load 12, align 1, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_v7s16_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v7s16_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; SI: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; SI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; SI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; SI: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<8 x s16>)
    ; SI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C1]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; SI: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; SI: $vgpr3 = COPY [[BITCAST9]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v7s16_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; CI-HSA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; CI-HSA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; CI-HSA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; CI-HSA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
    ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; CI-HSA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<8 x s16>)
    ; CI-HSA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; CI-HSA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-HSA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; CI-HSA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-HSA: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; CI-HSA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-HSA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; CI-HSA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-HSA: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-HSA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; CI-HSA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; CI-HSA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-HSA: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-HSA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C1]]
    ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; CI-HSA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; CI-HSA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C]](s32)
    ; CI-HSA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; CI-HSA: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; CI-HSA: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; CI-HSA: $vgpr3 = COPY [[BITCAST9]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v7s16_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; CI-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; CI-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; CI-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; CI-MESA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<8 x s16>)
    ; CI-MESA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C1]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; CI-MESA: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; CI-MESA: $vgpr3 = COPY [[BITCAST9]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v7s16_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; VI: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; VI: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; VI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; VI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; VI: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<8 x s16>)
    ; VI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C1]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST6]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[BITCAST7]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[BITCAST8]](<2 x s16>)
    ; VI: $vgpr3 = COPY [[BITCAST9]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v7s16_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; GFX9-HSA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; GFX9-HSA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; GFX9-HSA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; GFX9-HSA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
    ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; GFX9-HSA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<8 x s16>)
    ; GFX9-HSA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-HSA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC2]](<2 x s16>)
    ; GFX9-HSA: $vgpr3 = COPY [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v7s16_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>), [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[BITCAST]](<8 x s16>)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
    ; GFX9-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
    ; GFX9-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
    ; GFX9-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
    ; GFX9-MESA: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>), [[UV6:%[0-9]+]]:_(<2 x s16>), [[UV7:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF1]](<8 x s16>)
    ; GFX9-MESA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC2]](<2 x s16>)
    ; GFX9-MESA: $vgpr3 = COPY [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<7 x s16>) = G_LOAD %0 :: (load 14, align 16, addrspace 1)
    %2:_(<7 x s16>) = G_IMPLICIT_DEF
    %3:_(<14 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>), %9:_(<2 x s16>), %10:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6
    $vgpr3 = COPY %7

...

---
name: test_load_global_v7s16_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v7s16_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; SI: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; SI: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; SI: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v7s16_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 8, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; CI-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; CI-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v7s16_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 8, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; CI-MESA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; CI-MESA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; CI-MESA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v7s16_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 8, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; VI: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; VI: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; VI: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v7s16_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 8, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; GFX9-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; GFX9-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v7s16_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 8, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; GFX9-MESA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; GFX9-MESA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<7 x s16>) = G_LOAD %0 :: (load 14, align 8, addrspace 1)
    %2:_(<7 x s16>) = G_IMPLICIT_DEF
    %3:_(<14 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>), %9:_(<2 x s16>), %10:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6
    $vgpr3 = COPY %7

...

---
name: test_load_global_v7s16_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v7s16_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 4, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; SI: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; SI: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; SI: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v7s16_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 4, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; CI-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; CI-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v7s16_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 4, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; CI-MESA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; CI-MESA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; CI-MESA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v7s16_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 4, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; VI: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; VI: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; VI: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v7s16_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 4, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; GFX9-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; GFX9-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v7s16_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 4, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; GFX9-MESA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; GFX9-MESA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<7 x s16>) = G_LOAD %0 :: (load 14, align 4, addrspace 1)
    %2:_(<7 x s16>) = G_IMPLICIT_DEF
    %3:_(<14 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>), %9:_(<2 x s16>), %10:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6
    $vgpr3 = COPY %7

...

---
name: test_load_global_v7s16_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v7s16_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST2]](<2 x s16>), 0
    ; SI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; SI: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; SI: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; SI: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; SI: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; SI: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; SI: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; SI: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; SI: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; SI: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; SI: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; SI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; SI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; SI: [[BITCAST6:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST6]], [[C4]](s32)
    ; SI: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; SI: [[BITCAST7:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST7]], [[C4]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; SI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; SI: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR4]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C4]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL5]]
    ; SI: [[BITCAST10:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[BITCAST6]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST7]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL6]]
    ; SI: [[BITCAST11:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR6]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST8]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST9]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST10]](<2 x s16>)
    ; SI: $vgpr3 = COPY [[BITCAST11]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v7s16_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 2, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; CI-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; CI-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v7s16_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST2]](<2 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; CI-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; CI-MESA: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; CI-MESA: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; CI-MESA: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; CI-MESA: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; CI-MESA: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; CI-MESA: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; CI-MESA: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; CI-MESA: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; CI-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; CI-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; CI-MESA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; CI-MESA: [[BITCAST6:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST6]], [[C4]](s32)
    ; CI-MESA: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; CI-MESA: [[BITCAST7:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST7]], [[C4]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; CI-MESA: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; CI-MESA: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR4]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C4]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL5]]
    ; CI-MESA: [[BITCAST10:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[BITCAST6]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST7]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL6]]
    ; CI-MESA: [[BITCAST11:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR6]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST8]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[BITCAST9]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[BITCAST10]](<2 x s16>)
    ; CI-MESA: $vgpr3 = COPY [[BITCAST11]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v7s16_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST2]](<2 x s16>), 0
    ; VI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; VI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; VI: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; VI: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; VI: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; VI: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; VI: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; VI: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; VI: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; VI: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; VI: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; VI: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; VI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; VI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; VI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; VI: [[BITCAST6:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST6]], [[C4]](s32)
    ; VI: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; VI: [[BITCAST7:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST7]], [[C4]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR3]](s32)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; VI: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR4]](s32)
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C4]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL5]]
    ; VI: [[BITCAST10:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[BITCAST6]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST7]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; VI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; VI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL6]]
    ; VI: [[BITCAST11:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR6]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST8]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[BITCAST9]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[BITCAST10]](<2 x s16>)
    ; VI: $vgpr3 = COPY [[BITCAST11]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v7s16_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 2, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; GFX9-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; GFX9-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v7s16_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR_TRUNC2]](<2 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[TRUNC]](s16), 32
    ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; GFX9-MESA: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; GFX9-MESA: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; GFX9-MESA: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; GFX9-MESA: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; GFX9-MESA: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; GFX9-MESA: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; GFX9-MESA: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; GFX9-MESA: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C4]](s32)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C4]](s32)
    ; GFX9-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C4]](s32)
    ; GFX9-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; GFX9-MESA: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; GFX9-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32)
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC4]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC5]](<2 x s16>)
    ; GFX9-MESA: $vgpr3 = COPY [[BUILD_VECTOR_TRUNC6]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<7 x s16>) = G_LOAD %0 :: (load 14, align 2, addrspace 1)
    %2:_(<7 x s16>) = G_IMPLICIT_DEF
    %3:_(<14 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>), %9:_(<2 x s16>), %10:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6
    $vgpr3 = COPY %7

...

---
name: test_load_global_v7s16_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v7s16_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C4]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C5]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C1]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; SI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; SI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C5]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL8]]
    ; SI: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR8]](s32)
    ; SI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; SI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C1]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST2]](<2 x s16>), 0
    ; SI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; SI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR9]](s16), 32
    ; SI: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; SI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; SI: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; SI: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; SI: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; SI: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; SI: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; SI: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; SI: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; SI: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; SI: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; SI: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; SI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C5]](s32)
    ; SI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C5]](s32)
    ; SI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C5]](s32)
    ; SI: [[BITCAST6:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST6]], [[C5]](s32)
    ; SI: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; SI: [[BITCAST7:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST7]], [[C5]](s32)
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C8]]
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C8]]
    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C5]](s32)
    ; SI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[AND14]], [[SHL10]]
    ; SI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR10]](s32)
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; SI: [[AND16:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C8]]
    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C8]]
    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[C5]](s32)
    ; SI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[AND16]], [[SHL11]]
    ; SI: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR11]](s32)
    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; SI: [[AND18:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C8]]
    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C8]]
    ; SI: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[C5]](s32)
    ; SI: [[OR12:%[0-9]+]]:_(s32) = G_OR [[AND18]], [[SHL12]]
    ; SI: [[BITCAST10:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR12]](s32)
    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[BITCAST6]](s32)
    ; SI: [[AND20:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C8]]
    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[BITCAST7]](s32)
    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C8]]
    ; SI: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[C5]](s32)
    ; SI: [[OR13:%[0-9]+]]:_(s32) = G_OR [[AND20]], [[SHL13]]
    ; SI: [[BITCAST11:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR13]](s32)
    ; SI: $vgpr0 = COPY [[BITCAST8]](<2 x s16>)
    ; SI: $vgpr1 = COPY [[BITCAST9]](<2 x s16>)
    ; SI: $vgpr2 = COPY [[BITCAST10]](<2 x s16>)
    ; SI: $vgpr3 = COPY [[BITCAST11]](<2 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v7s16_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 1, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; CI-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; CI-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; CI-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; CI-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; CI-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; CI-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; CI-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v7s16_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY2]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY4]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C4]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[COPY6]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL4]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C5]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C1]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; CI-MESA: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; CI-MESA: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C5]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL8]]
    ; CI-MESA: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR8]](s32)
    ; CI-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C1]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C2]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST2]](<2 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR9]](s16), 32
    ; CI-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; CI-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; CI-MESA: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; CI-MESA: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; CI-MESA: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; CI-MESA: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; CI-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; CI-MESA: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; CI-MESA: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; CI-MESA: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; CI-MESA: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; CI-MESA: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; CI-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C5]](s32)
    ; CI-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C5]](s32)
    ; CI-MESA: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C5]](s32)
    ; CI-MESA: [[BITCAST6:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST6]], [[C5]](s32)
    ; CI-MESA: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; CI-MESA: [[BITCAST7:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST7]], [[C5]](s32)
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C8]]
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C8]]
    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C5]](s32)
    ; CI-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[AND14]], [[SHL10]]
    ; CI-MESA: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR10]](s32)
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; CI-MESA: [[AND16:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C8]]
    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C8]]
    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[C5]](s32)
    ; CI-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[AND16]], [[SHL11]]
    ; CI-MESA: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR11]](s32)
    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; CI-MESA: [[AND18:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C8]]
    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C8]]
    ; CI-MESA: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[C5]](s32)
    ; CI-MESA: [[OR12:%[0-9]+]]:_(s32) = G_OR [[AND18]], [[SHL12]]
    ; CI-MESA: [[BITCAST10:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR12]](s32)
    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[BITCAST6]](s32)
    ; CI-MESA: [[AND20:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C8]]
    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[BITCAST7]](s32)
    ; CI-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C8]]
    ; CI-MESA: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[C5]](s32)
    ; CI-MESA: [[OR13:%[0-9]+]]:_(s32) = G_OR [[AND20]], [[SHL13]]
    ; CI-MESA: [[BITCAST11:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR13]](s32)
    ; CI-MESA: $vgpr0 = COPY [[BITCAST8]](<2 x s16>)
    ; CI-MESA: $vgpr1 = COPY [[BITCAST9]](<2 x s16>)
    ; CI-MESA: $vgpr2 = COPY [[BITCAST10]](<2 x s16>)
    ; CI-MESA: $vgpr3 = COPY [[BITCAST11]](<2 x s16>)
    ; VI-LABEL: name: test_load_global_v7s16_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C3]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
    ; VI: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
    ; VI: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR4]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C4]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR5]](s32)
    ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST]](<2 x s16>), [[BITCAST1]](<2 x s16>)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C1]]
    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C2]](s16)
    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C1]]
    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C1]]
    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C2]](s16)
    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL7]]
    ; VI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; VI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; VI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C4]](s32)
    ; VI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL8]]
    ; VI: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR8]](s32)
    ; VI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; VI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C1]]
    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C1]]
    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C2]](s16)
    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL9]]
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BITCAST2]](<2 x s16>), 0
    ; VI: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; VI: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR9]](s16), 32
    ; VI: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; VI: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; VI: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; VI: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; VI: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; VI: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; VI: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; VI: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; VI: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; VI: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; VI: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; VI: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; VI: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C4]](s32)
    ; VI: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C4]](s32)
    ; VI: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C4]](s32)
    ; VI: [[BITCAST6:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST6]], [[C4]](s32)
    ; VI: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; VI: [[BITCAST7:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST7]], [[C4]](s32)
    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C7]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C7]]
    ; VI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C4]](s32)
    ; VI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[AND14]], [[SHL10]]
    ; VI: [[BITCAST8:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR10]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; VI: [[AND16:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C7]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C7]]
    ; VI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[C4]](s32)
    ; VI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[AND16]], [[SHL11]]
    ; VI: [[BITCAST9:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR11]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST5]](s32)
    ; VI: [[AND18:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C7]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C7]]
    ; VI: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[C4]](s32)
    ; VI: [[OR12:%[0-9]+]]:_(s32) = G_OR [[AND18]], [[SHL12]]
    ; VI: [[BITCAST10:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR12]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST6]](s32)
    ; VI: [[AND20:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C7]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST7]](s32)
    ; VI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C7]]
    ; VI: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[C4]](s32)
    ; VI: [[OR13:%[0-9]+]]:_(s32) = G_OR [[AND20]], [[SHL13]]
    ; VI: [[BITCAST11:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR13]](s32)
    ; VI: $vgpr0 = COPY [[BITCAST8]](<2 x s16>)
    ; VI: $vgpr1 = COPY [[BITCAST9]](<2 x s16>)
    ; VI: $vgpr2 = COPY [[BITCAST10]](<2 x s16>)
    ; VI: $vgpr3 = COPY [[BITCAST11]](<2 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v7s16_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<7 x s16>) = G_LOAD [[COPY]](p1) :: (load 14, align 1, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>), [[DEF1]](<8 x s16>)
    ; GFX9-HSA: [[UV:%[0-9]+]]:_(<7 x s16>), [[UV1:%[0-9]+]]:_(<7 x s16>), [[UV2:%[0-9]+]]:_(<7 x s16>), [[UV3:%[0-9]+]]:_(<7 x s16>), [[UV4:%[0-9]+]]:_(<7 x s16>), [[UV5:%[0-9]+]]:_(<7 x s16>), [[UV6:%[0-9]+]]:_(<7 x s16>), [[UV7:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS]](<56 x s16>)
    ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<14 x s16>) = G_CONCAT_VECTORS [[LOAD]](<7 x s16>), [[UV]](<7 x s16>)
    ; GFX9-HSA: [[UV8:%[0-9]+]]:_(<2 x s16>), [[UV9:%[0-9]+]]:_(<2 x s16>), [[UV10:%[0-9]+]]:_(<2 x s16>), [[UV11:%[0-9]+]]:_(<2 x s16>), [[UV12:%[0-9]+]]:_(<2 x s16>), [[UV13:%[0-9]+]]:_(<2 x s16>), [[UV14:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<14 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[UV8]](<2 x s16>)
    ; GFX9-HSA: $vgpr1 = COPY [[UV9]](<2 x s16>)
    ; GFX9-HSA: $vgpr2 = COPY [[UV10]](<2 x s16>)
    ; GFX9-HSA: $vgpr3 = COPY [[UV11]](<2 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v7s16_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C1]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C2]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C1]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C2]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[OR]](s16)
    ; GFX9-MESA: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[OR1]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C1]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C1]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C2]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C1]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C1]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C2]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[OR2]](s16)
    ; GFX9-MESA: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[OR3]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[ANYEXT3]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C1]]
    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C1]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C2]](s16)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL4]]
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C1]]
    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C1]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C2]](s16)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL5]]
    ; GFX9-MESA: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[OR4]](s16)
    ; GFX9-MESA: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[OR5]](s16)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT4]](s32), [[ANYEXT5]](s32)
    ; GFX9-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C1]]
    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C1]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C2]](s16)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL6]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<4 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[DEF]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV:%[0-9]+]]:_(<3 x s16>), [[UV1:%[0-9]+]]:_(<3 x s16>), [[UV2:%[0-9]+]]:_(<3 x s16>), [[UV3:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS1]](<12 x s16>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT]], [[BUILD_VECTOR_TRUNC2]](<2 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT1]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV4:%[0-9]+]]:_(<3 x s16>), [[UV5:%[0-9]+]]:_(<3 x s16>), [[UV6:%[0-9]+]]:_(<3 x s16>), [[UV7:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS2]](<12 x s16>)
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[DEF1]], [[UV4]](<3 x s16>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<4 x s16>) = G_INSERT [[INSERT2]], [[OR6]](s16), 32
    ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<12 x s16>) = G_CONCAT_VECTORS [[INSERT3]](<4 x s16>), [[DEF1]](<4 x s16>), [[DEF1]](<4 x s16>)
    ; GFX9-MESA: [[UV8:%[0-9]+]]:_(<3 x s16>), [[UV9:%[0-9]+]]:_(<3 x s16>), [[UV10:%[0-9]+]]:_(<3 x s16>), [[UV11:%[0-9]+]]:_(<3 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS3]](<12 x s16>)
    ; GFX9-MESA: [[DEF2:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[DEF3:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[DEF2]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; GFX9-MESA: [[UV12:%[0-9]+]]:_(<7 x s16>), [[UV13:%[0-9]+]]:_(<7 x s16>), [[UV14:%[0-9]+]]:_(<7 x s16>), [[UV15:%[0-9]+]]:_(<7 x s16>), [[UV16:%[0-9]+]]:_(<7 x s16>), [[UV17:%[0-9]+]]:_(<7 x s16>), [[UV18:%[0-9]+]]:_(<7 x s16>), [[UV19:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS4]](<56 x s16>)
    ; GFX9-MESA: [[INSERT4:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV12]](<7 x s16>), 0
    ; GFX9-MESA: [[INSERT5:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT4]], [[CONCAT_VECTORS]](<4 x s16>), 0
    ; GFX9-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<56 x s16>) = G_CONCAT_VECTORS [[INSERT5]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>), [[DEF3]](<8 x s16>)
    ; GFX9-MESA: [[UV20:%[0-9]+]]:_(<7 x s16>), [[UV21:%[0-9]+]]:_(<7 x s16>), [[UV22:%[0-9]+]]:_(<7 x s16>), [[UV23:%[0-9]+]]:_(<7 x s16>), [[UV24:%[0-9]+]]:_(<7 x s16>), [[UV25:%[0-9]+]]:_(<7 x s16>), [[UV26:%[0-9]+]]:_(<7 x s16>), [[UV27:%[0-9]+]]:_(<7 x s16>) = G_UNMERGE_VALUES [[CONCAT_VECTORS5]](<56 x s16>)
    ; GFX9-MESA: [[INSERT6:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[DEF3]], [[UV20]](<7 x s16>), 0
    ; GFX9-MESA: [[INSERT7:%[0-9]+]]:_(<8 x s16>) = G_INSERT [[INSERT6]], [[UV8]](<3 x s16>), 64
    ; GFX9-MESA: [[DEF4:%[0-9]+]]:_(<8 x s16>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[UV28:%[0-9]+]]:_(<2 x s16>), [[UV29:%[0-9]+]]:_(<2 x s16>), [[UV30:%[0-9]+]]:_(<2 x s16>), [[UV31:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[INSERT7]](<8 x s16>)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV28]](<2 x s16>)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C6]](s32)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV29]](<2 x s16>)
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C6]](s32)
    ; GFX9-MESA: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV30]](<2 x s16>)
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C6]](s32)
    ; GFX9-MESA: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV31]](<2 x s16>)
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C6]](s32)
    ; GFX9-MESA: [[UV32:%[0-9]+]]:_(<2 x s16>), [[UV33:%[0-9]+]]:_(<2 x s16>), [[UV34:%[0-9]+]]:_(<2 x s16>), [[UV35:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[DEF4]](<8 x s16>)
    ; GFX9-MESA: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV32]](<2 x s16>)
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C6]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[BITCAST]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[BITCAST1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[BITCAST2]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[BITCAST3]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[BITCAST4]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: $vgpr1 = COPY [[BUILD_VECTOR_TRUNC4]](<2 x s16>)
    ; GFX9-MESA: $vgpr2 = COPY [[BUILD_VECTOR_TRUNC5]](<2 x s16>)
    ; GFX9-MESA: $vgpr3 = COPY [[BUILD_VECTOR_TRUNC6]](<2 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<7 x s16>) = G_LOAD %0 :: (load 14, align 1, addrspace 1)
    %2:_(<7 x s16>) = G_IMPLICIT_DEF
    %3:_(<14 x s16>) = G_CONCAT_VECTORS %1, %2
    %4:_(<2 x s16>), %5:_(<2 x s16>), %6:_(<2 x s16>), %7:_(<2 x s16>), %8:_(<2 x s16>), %9:_(<2 x s16>), %10:_(<2 x s16>) = G_UNMERGE_VALUES %3
    $vgpr0 = COPY %4
    $vgpr1 = COPY %5
    $vgpr2 = COPY %6
    $vgpr3 = COPY %7

...

---
name: test_load_global_v8s16_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v8s16_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v8s16_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v8s16_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; VI-LABEL: name: test_load_global_v8s16_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v8s16_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v8s16_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<8 x s16>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v8s16_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v8s16_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; CI-HSA-LABEL: name: test_load_global_v8s16_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; CI-MESA-LABEL: name: test_load_global_v8s16_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; VI-LABEL: name: test_load_global_v8s16_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; GFX9-HSA-LABEL: name: test_load_global_v8s16_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    ; GFX9-MESA-LABEL: name: test_load_global_v8s16_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s16>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<8 x s16>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<8 x s16>) = G_LOAD %0 :: (load 16, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2s32_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s32_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v2s32_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v2s32_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; VI-LABEL: name: test_load_global_v2s32_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s32_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s32_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v2s32_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s32_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v2s32_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v2s32_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; VI-LABEL: name: test_load_global_v2s32_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s32_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s32_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v2s32_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s32_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v2s32_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v2s32_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    ; VI-LABEL: name: test_load_global_v2s32_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s32_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s32_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 2, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v2s32_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s32_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v2s32_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v2s32_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    ; VI-LABEL: name: test_load_global_v2s32_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s32_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s32_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s32>) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v3s32_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s32_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[EXTRACT:%[0-9]+]]:_(<3 x s32>) = G_EXTRACT [[LOAD]](<4 x s32>), 0
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[EXTRACT]](<3 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v3s32_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v3s32_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; VI-LABEL: name: test_load_global_v3s32_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s32_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s32_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s32>) = G_LOAD %0 :: (load 12, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_v3s32_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s32_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[INSERT1]](<3 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v3s32_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v3s32_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; VI-LABEL: name: test_load_global_v3s32_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s32_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s32_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s32>) = G_LOAD %0 :: (load 12, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_load_global_v4s32_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s32_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v4s32_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v4s32_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; VI-LABEL: name: test_load_global_v4s32_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s32_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s32_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v4s32_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s32_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v4s32_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v4s32_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; VI-LABEL: name: test_load_global_v4s32_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s32_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s32_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v4s32_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s32_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v4s32_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v4s32_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; VI-LABEL: name: test_load_global_v4s32_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s32_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s32_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s32>) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v8s32_align32
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v8s32_align32
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<8 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v8s32_align32
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<8 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v8s32_align32
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<8 x s32>)
    ; VI-LABEL: name: test_load_global_v8s32_align32
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<8 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v8s32_align32
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<8 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v8s32_align32
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<8 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<8 x s32>) = G_LOAD %0 :: (load 32, align 32, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---
name: test_load_global_v16s32_align32
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_load_global_v16s32_align32
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, align 32, addrspace 1)
    ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI: [[GEP:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[GEP]](p1) :: (load 4, addrspace 1)
    ; CI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI: [[GEP1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[GEP1]](p1) :: (load 4, align 8, addrspace 1)
    ; CI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI: [[GEP2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[GEP2]](p1) :: (load 4, addrspace 1)
    ; CI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI: [[GEP3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[GEP3]](p1) :: (load 4, align 16, addrspace 1)
    ; CI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 20
    ; CI: [[GEP4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[GEP4]](p1) :: (load 4, addrspace 1)
    ; CI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
    ; CI: [[GEP5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[GEP5]](p1) :: (load 4, align 8, addrspace 1)
    ; CI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 28
    ; CI: [[GEP6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[GEP6]](p1) :: (load 4, addrspace 1)
    ; CI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
    ; CI: [[GEP7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[GEP7]](p1) :: (load 4, align 32, addrspace 1)
    ; CI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
    ; CI: [[GEP8:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[GEP8]](p1) :: (load 4, addrspace 1)
    ; CI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 40
    ; CI: [[GEP9:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[GEP9]](p1) :: (load 4, align 8, addrspace 1)
    ; CI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
    ; CI: [[GEP10:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C10]](s64)
    ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[GEP10]](p1) :: (load 4, addrspace 1)
    ; CI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
    ; CI: [[GEP11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[GEP11]](p1) :: (load 4, align 16, addrspace 1)
    ; CI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
    ; CI: [[GEP12:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C12]](s64)
    ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[GEP12]](p1) :: (load 4, addrspace 1)
    ; CI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 56
    ; CI: [[GEP13:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C13]](s64)
    ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[GEP13]](p1) :: (load 4, align 8, addrspace 1)
    ; CI: [[C14:%[0-9]+]]:_(s64) = G_CONSTANT i64 60
    ; CI: [[GEP14:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C14]](s64)
    ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[GEP14]](p1) :: (load 4, addrspace 1)
    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<16 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32), [[LOAD2]](s32), [[LOAD3]](s32), [[LOAD4]](s32), [[LOAD5]](s32), [[LOAD6]](s32), [[LOAD7]](s32), [[LOAD8]](s32), [[LOAD9]](s32), [[LOAD10]](s32), [[LOAD11]](s32), [[LOAD12]](s32), [[LOAD13]](s32), [[LOAD14]](s32), [[LOAD15]](s32)
    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[BUILD_VECTOR]](<16 x s32>)
    ; SI-LABEL: name: test_load_global_v16s32_align32
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[COPY]](p1) :: (load 64, align 32, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[LOAD]](<16 x s32>)
    ; CI-HSA-LABEL: name: test_load_global_v16s32_align32
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[COPY]](p1) :: (load 64, align 32, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[LOAD]](<16 x s32>)
    ; CI-MESA-LABEL: name: test_load_global_v16s32_align32
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[COPY]](p1) :: (load 64, align 32, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[LOAD]](<16 x s32>)
    ; VI-LABEL: name: test_load_global_v16s32_align32
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[COPY]](p1) :: (load 64, align 32, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[LOAD]](<16 x s32>)
    ; GFX9-HSA-LABEL: name: test_load_global_v16s32_align32
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[COPY]](p1) :: (load 64, align 32, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[LOAD]](<16 x s32>)
    ; GFX9-MESA-LABEL: name: test_load_global_v16s32_align32
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[COPY]](p1) :: (load 64, align 32, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY [[LOAD]](<16 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<16 x s32>) = G_LOAD %0 :: (load 64, align 32, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = COPY %1
...

---
name: test_load_global_v2s64_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s64_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v2s64_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v2s64_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; VI-LABEL: name: test_load_global_v2s64_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2s64_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s64_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v2s64_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v2s64_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; VI-LABEL: name: test_load_global_v2s64_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2s64_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s64_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v2s64_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v2s64_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; VI-LABEL: name: test_load_global_v2s64_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2s64_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s64_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR2]](s32), [[OR3]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v2s64_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 2, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v2s64_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR2]](s32), [[OR3]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    ; VI-LABEL: name: test_load_global_v2s64_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR2]](s32), [[OR3]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C4]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR]](s32), [[OR1]](s32)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR2]](s32), [[OR3]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 2, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2s64_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s64_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; SI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; SI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; SI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; SI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
    ; SI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; SI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C10]](s32)
    ; SI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; SI: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; SI: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C10]](s32)
    ; SI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v2s64_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v2s64_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; CI-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; CI-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; CI-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
    ; CI-MESA: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; CI-MESA: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C10]](s32)
    ; CI-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; CI-MESA: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; CI-MESA: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C10]](s32)
    ; CI-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    ; VI-LABEL: name: test_load_global_v2s64_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C10]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; VI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; VI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; VI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL7]]
    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
    ; VI: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
    ; VI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL8]]
    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL9]]
    ; VI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; VI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; VI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C9]](s32)
    ; VI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; VI: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; VI: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; VI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C9]](s32)
    ; VI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s64_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<2 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s64_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; GFX9-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; GFX9-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C10]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL7]]
    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL8]]
    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL9]]
    ; GFX9-MESA: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; GFX9-MESA: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C9]](s32)
    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; GFX9-MESA: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; GFX9-MESA: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C9]](s32)
    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s64>) = G_LOAD %0 :: (load 16, align 1, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2sp1_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2sp1_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-HSA-LABEL: name: test_load_global_v2sp1_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-MESA-LABEL: name: test_load_global_v2sp1_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; VI-LABEL: name: test_load_global_v2sp1_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2sp1_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2sp1_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v3s64_align32
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s64_align32
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; SI: [[EXTRACT:%[0-9]+]]:_(<3 x s64>) = G_EXTRACT [[LOAD]](<4 x s64>), 0
    ; SI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s64>), 0
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v3s64_align32
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-HSA: [[EXTRACT:%[0-9]+]]:_(<3 x s64>) = G_EXTRACT [[LOAD]](<4 x s64>), 0
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s64>), 0
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v3s64_align32
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-MESA: [[EXTRACT:%[0-9]+]]:_(<3 x s64>) = G_EXTRACT [[LOAD]](<4 x s64>), 0
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s64>), 0
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
    ; VI-LABEL: name: test_load_global_v3s64_align32
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; VI: [[EXTRACT:%[0-9]+]]:_(<3 x s64>) = G_EXTRACT [[LOAD]](<4 x s64>), 0
    ; VI: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s64>), 0
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s64_align32
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-HSA: [[EXTRACT:%[0-9]+]]:_(<3 x s64>) = G_EXTRACT [[LOAD]](<4 x s64>), 0
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s64>), 0
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s64_align32
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-MESA: [[EXTRACT:%[0-9]+]]:_(<3 x s64>) = G_EXTRACT [[LOAD]](<4 x s64>), 0
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF]], [[EXTRACT]](<3 x s64>), 0
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT]](<4 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s64>) = G_LOAD %0 :: (load 24, align 32, addrspace 1)
    %2:_(<4 x s64>) = G_IMPLICIT_DEF
    %3:_(<4 x s64>) = G_INSERT %2, %1, 0
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3
...

---
name: test_load_global_v3s64_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s64_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; SI: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; SI: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v3s64_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; CI-HSA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v3s64_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; VI-LABEL: name: test_load_global_v3s64_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; VI: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; VI: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s64_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; GFX9-HSA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s64_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s64>) = G_LOAD %0 :: (load 24, align 8, addrspace 1)
    %2:_(<4 x s64>) = G_IMPLICIT_DEF
    %3:_(<4 x s64>) = G_INSERT %2, %1, 0
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3
...

---
name: test_load_global_v3s64_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v3s64_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; SI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; SI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; SI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; SI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
    ; SI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; SI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C10]](s32)
    ; SI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; SI: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; SI: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C10]](s32)
    ; SI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; SI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; SI: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C12]](s64)
    ; SI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; SI: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; SI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; SI: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; SI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; SI: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; SI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; SI: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; SI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; SI: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; SI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; SI: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; SI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; SI: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; SI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; SI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; SI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
    ; SI: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
    ; SI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL12]](s32)
    ; SI: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
    ; SI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; SI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
    ; SI: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
    ; SI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL13]](s32)
    ; SI: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
    ; SI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; SI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
    ; SI: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
    ; SI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL14]](s32)
    ; SI: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
    ; SI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; SI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; SI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
    ; SI: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
    ; SI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL15]](s32)
    ; SI: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
    ; SI: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; SI: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; SI: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C10]](s32)
    ; SI: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; SI: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; SI: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; SI: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C10]](s32)
    ; SI: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s64>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[MV2]](s64), 128
    ; SI: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; SI: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v3s64_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, align 1, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; CI-HSA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; CI-HSA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v3s64_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; CI-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; CI-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; CI-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
    ; CI-MESA: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; CI-MESA: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C10]](s32)
    ; CI-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; CI-MESA: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; CI-MESA: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C10]](s32)
    ; CI-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; CI-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-MESA: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C12]](s64)
    ; CI-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; CI-MESA: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; CI-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; CI-MESA: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; CI-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; CI-MESA: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; CI-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; CI-MESA: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; CI-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; CI-MESA: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; CI-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; CI-MESA: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; CI-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; CI-MESA: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; CI-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; CI-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; CI-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; CI-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
    ; CI-MESA: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
    ; CI-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL12]](s32)
    ; CI-MESA: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
    ; CI-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; CI-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; CI-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
    ; CI-MESA: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
    ; CI-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL13]](s32)
    ; CI-MESA: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
    ; CI-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; CI-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; CI-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
    ; CI-MESA: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
    ; CI-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL14]](s32)
    ; CI-MESA: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
    ; CI-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; CI-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; CI-MESA: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
    ; CI-MESA: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
    ; CI-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL15]](s32)
    ; CI-MESA: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
    ; CI-MESA: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; CI-MESA: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; CI-MESA: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C10]](s32)
    ; CI-MESA: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; CI-MESA: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; CI-MESA: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; CI-MESA: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C10]](s32)
    ; CI-MESA: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; CI-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s64>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[MV2]](s64), 128
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; VI-LABEL: name: test_load_global_v3s64_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C10]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; VI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; VI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; VI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL7]]
    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
    ; VI: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
    ; VI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL8]]
    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL9]]
    ; VI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; VI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; VI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C9]](s32)
    ; VI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; VI: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; VI: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; VI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C9]](s32)
    ; VI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; VI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; VI: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; VI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; VI: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; VI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; VI: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; VI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; VI: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; VI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; VI: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; VI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; VI: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; VI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; VI: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; VI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; VI: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; VI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; VI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; VI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; VI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
    ; VI: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
    ; VI: [[SHL12:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
    ; VI: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL12]]
    ; VI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; VI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; VI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
    ; VI: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
    ; VI: [[SHL13:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
    ; VI: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL13]]
    ; VI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; VI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; VI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
    ; VI: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
    ; VI: [[SHL14:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
    ; VI: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL14]]
    ; VI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; VI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; VI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
    ; VI: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
    ; VI: [[SHL15:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
    ; VI: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL15]]
    ; VI: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; VI: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; VI: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C9]](s32)
    ; VI: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; VI: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; VI: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; VI: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C9]](s32)
    ; VI: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; VI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s64>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[MV2]](s64), 128
    ; VI: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; VI: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v3s64_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s64>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(s64) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 16, align 1, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[LOAD]](<2 x s64>), 0
    ; GFX9-HSA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[LOAD1]](s64), 128
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v3s64_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; GFX9-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; GFX9-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C10]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL7]]
    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL8]]
    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL9]]
    ; GFX9-MESA: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; GFX9-MESA: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C9]](s32)
    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; GFX9-MESA: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; GFX9-MESA: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C9]](s32)
    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64)
    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-MESA: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; GFX9-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; GFX9-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; GFX9-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; GFX9-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; GFX9-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
    ; GFX9-MESA: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
    ; GFX9-MESA: [[SHL12:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
    ; GFX9-MESA: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL12]]
    ; GFX9-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; GFX9-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; GFX9-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
    ; GFX9-MESA: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
    ; GFX9-MESA: [[SHL13:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
    ; GFX9-MESA: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL13]]
    ; GFX9-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; GFX9-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; GFX9-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
    ; GFX9-MESA: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
    ; GFX9-MESA: [[SHL14:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
    ; GFX9-MESA: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL14]]
    ; GFX9-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; GFX9-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; GFX9-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
    ; GFX9-MESA: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
    ; GFX9-MESA: [[SHL15:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
    ; GFX9-MESA: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL15]]
    ; GFX9-MESA: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; GFX9-MESA: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; GFX9-MESA: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C9]](s32)
    ; GFX9-MESA: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; GFX9-MESA: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; GFX9-MESA: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; GFX9-MESA: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C9]](s32)
    ; GFX9-MESA: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; GFX9-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s64>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[DEF]], [[BUILD_VECTOR]](<2 x s64>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s64>) = G_INSERT [[INSERT]], [[MV2]](s64), 128
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(<4 x s64>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<4 x s64>) = G_INSERT [[DEF1]], [[INSERT1]](<3 x s64>), 0
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[INSERT2]](<4 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s64>) = G_LOAD %0 :: (load 24, align 1, addrspace 1)
    %2:_(<4 x s64>) = G_IMPLICIT_DEF
    %3:_(<4 x s64>) = G_INSERT %2, %1, 0
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3
...

---
name: test_load_global_v4s64_align32
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s64_align32
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v4s64_align32
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v4s64_align32
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; VI-LABEL: name: test_load_global_v4s64_align32
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s64_align32
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s64_align32
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s64>) = G_LOAD %0 :: (load 32, align 32, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---
name: test_load_global_v4s64_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s64_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v4s64_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v4s64_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; VI-LABEL: name: test_load_global_v4s64_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s64_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s64_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s64>) = G_LOAD %0 :: (load 32, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---
name: test_load_global_v4s64_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4s64_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; SI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; SI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; SI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; SI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; SI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; SI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; SI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; SI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; SI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; SI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; SI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; SI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; SI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; SI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; SI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; SI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; SI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; SI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
    ; SI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
    ; SI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; SI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C10]](s32)
    ; SI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; SI: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; SI: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C10]](s32)
    ; SI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; SI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; SI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; SI: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C12]](s64)
    ; SI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; SI: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; SI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; SI: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; SI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; SI: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; SI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; SI: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; SI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; SI: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; SI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; SI: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; SI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; SI: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; SI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; SI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; SI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
    ; SI: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
    ; SI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL12]](s32)
    ; SI: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
    ; SI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; SI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
    ; SI: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
    ; SI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL13]](s32)
    ; SI: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
    ; SI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; SI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
    ; SI: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
    ; SI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL14]](s32)
    ; SI: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
    ; SI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; SI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; SI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
    ; SI: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
    ; SI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL15]](s32)
    ; SI: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
    ; SI: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; SI: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; SI: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C10]](s32)
    ; SI: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; SI: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; SI: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; SI: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C10]](s32)
    ; SI: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; SI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; SI: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
    ; SI: [[PTR_ADD23:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C13]](s64)
    ; SI: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD23]](p1) :: (load 1 + 24, addrspace 1)
    ; SI: [[PTR_ADD24:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C]](s64)
    ; SI: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD24]](p1) :: (load 1 + 25, addrspace 1)
    ; SI: [[PTR_ADD25:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C1]](s64)
    ; SI: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD25]](p1) :: (load 1 + 26, addrspace 1)
    ; SI: [[PTR_ADD26:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C2]](s64)
    ; SI: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD26]](p1) :: (load 1 + 27, addrspace 1)
    ; SI: [[PTR_ADD27:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C3]](s64)
    ; SI: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD27]](p1) :: (load 1 + 28, addrspace 1)
    ; SI: [[PTR_ADD28:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C4]](s64)
    ; SI: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD28]](p1) :: (load 1 + 29, addrspace 1)
    ; SI: [[PTR_ADD29:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C5]](s64)
    ; SI: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD29]](p1) :: (load 1 + 30, addrspace 1)
    ; SI: [[PTR_ADD30:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C6]](s64)
    ; SI: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD30]](p1) :: (load 1 + 31, addrspace 1)
    ; SI: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
    ; SI: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
    ; SI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD25]](s32)
    ; SI: [[AND25:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C9]]
    ; SI: [[SHL18:%[0-9]+]]:_(s32) = G_SHL [[AND25]], [[COPY24]](s32)
    ; SI: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[SHL18]](s32)
    ; SI: [[OR18:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[TRUNC25]]
    ; SI: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
    ; SI: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
    ; SI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LOAD27]](s32)
    ; SI: [[AND27:%[0-9]+]]:_(s32) = G_AND [[COPY27]], [[C9]]
    ; SI: [[SHL19:%[0-9]+]]:_(s32) = G_SHL [[AND27]], [[COPY26]](s32)
    ; SI: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[SHL19]](s32)
    ; SI: [[OR19:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[TRUNC27]]
    ; SI: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
    ; SI: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
    ; SI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LOAD29]](s32)
    ; SI: [[AND29:%[0-9]+]]:_(s32) = G_AND [[COPY29]], [[C9]]
    ; SI: [[SHL20:%[0-9]+]]:_(s32) = G_SHL [[AND29]], [[COPY28]](s32)
    ; SI: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[SHL20]](s32)
    ; SI: [[OR20:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[TRUNC29]]
    ; SI: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
    ; SI: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
    ; SI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; SI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LOAD31]](s32)
    ; SI: [[AND31:%[0-9]+]]:_(s32) = G_AND [[COPY31]], [[C9]]
    ; SI: [[SHL21:%[0-9]+]]:_(s32) = G_SHL [[AND31]], [[COPY30]](s32)
    ; SI: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[SHL21]](s32)
    ; SI: [[OR21:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[TRUNC31]]
    ; SI: [[ZEXT12:%[0-9]+]]:_(s32) = G_ZEXT [[OR18]](s16)
    ; SI: [[ZEXT13:%[0-9]+]]:_(s32) = G_ZEXT [[OR19]](s16)
    ; SI: [[SHL22:%[0-9]+]]:_(s32) = G_SHL [[ZEXT13]], [[C10]](s32)
    ; SI: [[OR22:%[0-9]+]]:_(s32) = G_OR [[ZEXT12]], [[SHL22]]
    ; SI: [[ZEXT14:%[0-9]+]]:_(s32) = G_ZEXT [[OR20]](s16)
    ; SI: [[ZEXT15:%[0-9]+]]:_(s32) = G_ZEXT [[OR21]](s16)
    ; SI: [[SHL23:%[0-9]+]]:_(s32) = G_SHL [[ZEXT15]], [[C10]](s32)
    ; SI: [[OR23:%[0-9]+]]:_(s32) = G_OR [[ZEXT14]], [[SHL23]]
    ; SI: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR22]](s32), [[OR23]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
    ; CI-HSA-LABEL: name: test_load_global_v4s64_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; CI-MESA-LABEL: name: test_load_global_v4s64_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C9]]
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C9]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
    ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C9]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[COPY5]](s32)
    ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[SHL2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[TRUNC5]]
    ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C9]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C8]](s32)
    ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[SHL3]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[TRUNC7]]
    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; CI-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C10]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; CI-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; CI-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C10]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; CI-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; CI-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; CI-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; CI-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C9]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[COPY8]](s32)
    ; CI-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[SHL6]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[TRUNC9]]
    ; CI-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C9]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[COPY10]](s32)
    ; CI-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[SHL7]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[TRUNC11]]
    ; CI-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C9]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[COPY12]](s32)
    ; CI-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[SHL8]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[TRUNC13]]
    ; CI-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C9]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[COPY14]](s32)
    ; CI-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[SHL9]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[TRUNC15]]
    ; CI-MESA: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; CI-MESA: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C10]](s32)
    ; CI-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; CI-MESA: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; CI-MESA: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C10]](s32)
    ; CI-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; CI-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; CI-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; CI-MESA: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C12]](s64)
    ; CI-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; CI-MESA: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; CI-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; CI-MESA: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; CI-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; CI-MESA: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; CI-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; CI-MESA: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; CI-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; CI-MESA: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; CI-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; CI-MESA: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; CI-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; CI-MESA: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; CI-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; CI-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; CI-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; CI-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C9]]
    ; CI-MESA: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[COPY16]](s32)
    ; CI-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[SHL12]](s32)
    ; CI-MESA: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[TRUNC17]]
    ; CI-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; CI-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; CI-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C9]]
    ; CI-MESA: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[COPY18]](s32)
    ; CI-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[SHL13]](s32)
    ; CI-MESA: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[TRUNC19]]
    ; CI-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; CI-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; CI-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C9]]
    ; CI-MESA: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[COPY20]](s32)
    ; CI-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[SHL14]](s32)
    ; CI-MESA: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[TRUNC21]]
    ; CI-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; CI-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; CI-MESA: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C9]]
    ; CI-MESA: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[COPY22]](s32)
    ; CI-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[SHL15]](s32)
    ; CI-MESA: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[TRUNC23]]
    ; CI-MESA: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; CI-MESA: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; CI-MESA: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C10]](s32)
    ; CI-MESA: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; CI-MESA: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; CI-MESA: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; CI-MESA: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C10]](s32)
    ; CI-MESA: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; CI-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; CI-MESA: [[C13:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
    ; CI-MESA: [[PTR_ADD23:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C13]](s64)
    ; CI-MESA: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD23]](p1) :: (load 1 + 24, addrspace 1)
    ; CI-MESA: [[PTR_ADD24:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C]](s64)
    ; CI-MESA: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD24]](p1) :: (load 1 + 25, addrspace 1)
    ; CI-MESA: [[PTR_ADD25:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C1]](s64)
    ; CI-MESA: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD25]](p1) :: (load 1 + 26, addrspace 1)
    ; CI-MESA: [[PTR_ADD26:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C2]](s64)
    ; CI-MESA: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD26]](p1) :: (load 1 + 27, addrspace 1)
    ; CI-MESA: [[PTR_ADD27:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C3]](s64)
    ; CI-MESA: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD27]](p1) :: (load 1 + 28, addrspace 1)
    ; CI-MESA: [[PTR_ADD28:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C4]](s64)
    ; CI-MESA: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD28]](p1) :: (load 1 + 29, addrspace 1)
    ; CI-MESA: [[PTR_ADD29:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C5]](s64)
    ; CI-MESA: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD29]](p1) :: (load 1 + 30, addrspace 1)
    ; CI-MESA: [[PTR_ADD30:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C6]](s64)
    ; CI-MESA: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD30]](p1) :: (load 1 + 31, addrspace 1)
    ; CI-MESA: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
    ; CI-MESA: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
    ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD25]](s32)
    ; CI-MESA: [[AND25:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C9]]
    ; CI-MESA: [[SHL18:%[0-9]+]]:_(s32) = G_SHL [[AND25]], [[COPY24]](s32)
    ; CI-MESA: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[SHL18]](s32)
    ; CI-MESA: [[OR18:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[TRUNC25]]
    ; CI-MESA: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
    ; CI-MESA: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
    ; CI-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LOAD27]](s32)
    ; CI-MESA: [[AND27:%[0-9]+]]:_(s32) = G_AND [[COPY27]], [[C9]]
    ; CI-MESA: [[SHL19:%[0-9]+]]:_(s32) = G_SHL [[AND27]], [[COPY26]](s32)
    ; CI-MESA: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[SHL19]](s32)
    ; CI-MESA: [[OR19:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[TRUNC27]]
    ; CI-MESA: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
    ; CI-MESA: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
    ; CI-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LOAD29]](s32)
    ; CI-MESA: [[AND29:%[0-9]+]]:_(s32) = G_AND [[COPY29]], [[C9]]
    ; CI-MESA: [[SHL20:%[0-9]+]]:_(s32) = G_SHL [[AND29]], [[COPY28]](s32)
    ; CI-MESA: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[SHL20]](s32)
    ; CI-MESA: [[OR20:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[TRUNC29]]
    ; CI-MESA: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
    ; CI-MESA: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
    ; CI-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[C8]](s32)
    ; CI-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LOAD31]](s32)
    ; CI-MESA: [[AND31:%[0-9]+]]:_(s32) = G_AND [[COPY31]], [[C9]]
    ; CI-MESA: [[SHL21:%[0-9]+]]:_(s32) = G_SHL [[AND31]], [[COPY30]](s32)
    ; CI-MESA: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[SHL21]](s32)
    ; CI-MESA: [[OR21:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[TRUNC31]]
    ; CI-MESA: [[ZEXT12:%[0-9]+]]:_(s32) = G_ZEXT [[OR18]](s16)
    ; CI-MESA: [[ZEXT13:%[0-9]+]]:_(s32) = G_ZEXT [[OR19]](s16)
    ; CI-MESA: [[SHL22:%[0-9]+]]:_(s32) = G_SHL [[ZEXT13]], [[C10]](s32)
    ; CI-MESA: [[OR22:%[0-9]+]]:_(s32) = G_OR [[ZEXT12]], [[SHL22]]
    ; CI-MESA: [[ZEXT14:%[0-9]+]]:_(s32) = G_ZEXT [[OR20]](s16)
    ; CI-MESA: [[ZEXT15:%[0-9]+]]:_(s32) = G_ZEXT [[OR21]](s16)
    ; CI-MESA: [[SHL23:%[0-9]+]]:_(s32) = G_SHL [[ZEXT15]], [[C10]](s32)
    ; CI-MESA: [[OR23:%[0-9]+]]:_(s32) = G_OR [[ZEXT14]], [[SHL23]]
    ; CI-MESA: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR22]](s32), [[OR23]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
    ; VI-LABEL: name: test_load_global_v4s64_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; VI: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; VI: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; VI: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; VI: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; VI: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; VI: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; VI: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; VI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; VI: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; VI: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; VI: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C10]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; VI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; VI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; VI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; VI: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; VI: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
    ; VI: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
    ; VI: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; VI: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
    ; VI: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
    ; VI: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL7]]
    ; VI: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; VI: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
    ; VI: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
    ; VI: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL8]]
    ; VI: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; VI: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
    ; VI: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
    ; VI: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL9]]
    ; VI: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; VI: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; VI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C9]](s32)
    ; VI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; VI: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; VI: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; VI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C9]](s32)
    ; VI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; VI: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; VI: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; VI: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; VI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; VI: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; VI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; VI: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; VI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; VI: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; VI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; VI: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; VI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; VI: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; VI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; VI: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; VI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; VI: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; VI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; VI: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; VI: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; VI: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
    ; VI: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
    ; VI: [[SHL12:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
    ; VI: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL12]]
    ; VI: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; VI: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; VI: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
    ; VI: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
    ; VI: [[SHL13:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
    ; VI: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL13]]
    ; VI: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; VI: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; VI: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
    ; VI: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
    ; VI: [[SHL14:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
    ; VI: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL14]]
    ; VI: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; VI: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; VI: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
    ; VI: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
    ; VI: [[SHL15:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
    ; VI: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL15]]
    ; VI: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; VI: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; VI: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C9]](s32)
    ; VI: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; VI: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; VI: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; VI: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C9]](s32)
    ; VI: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; VI: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; VI: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
    ; VI: [[PTR_ADD23:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C12]](s64)
    ; VI: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD23]](p1) :: (load 1 + 24, addrspace 1)
    ; VI: [[PTR_ADD24:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C]](s64)
    ; VI: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD24]](p1) :: (load 1 + 25, addrspace 1)
    ; VI: [[PTR_ADD25:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C1]](s64)
    ; VI: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD25]](p1) :: (load 1 + 26, addrspace 1)
    ; VI: [[PTR_ADD26:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C2]](s64)
    ; VI: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD26]](p1) :: (load 1 + 27, addrspace 1)
    ; VI: [[PTR_ADD27:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C3]](s64)
    ; VI: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD27]](p1) :: (load 1 + 28, addrspace 1)
    ; VI: [[PTR_ADD28:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C4]](s64)
    ; VI: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD28]](p1) :: (load 1 + 29, addrspace 1)
    ; VI: [[PTR_ADD29:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C5]](s64)
    ; VI: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD29]](p1) :: (load 1 + 30, addrspace 1)
    ; VI: [[PTR_ADD30:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C6]](s64)
    ; VI: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD30]](p1) :: (load 1 + 31, addrspace 1)
    ; VI: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
    ; VI: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
    ; VI: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD25]](s32)
    ; VI: [[AND25:%[0-9]+]]:_(s16) = G_AND [[TRUNC25]], [[C7]]
    ; VI: [[SHL18:%[0-9]+]]:_(s16) = G_SHL [[AND25]], [[C8]](s16)
    ; VI: [[OR18:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[SHL18]]
    ; VI: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
    ; VI: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
    ; VI: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD27]](s32)
    ; VI: [[AND27:%[0-9]+]]:_(s16) = G_AND [[TRUNC27]], [[C7]]
    ; VI: [[SHL19:%[0-9]+]]:_(s16) = G_SHL [[AND27]], [[C8]](s16)
    ; VI: [[OR19:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[SHL19]]
    ; VI: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
    ; VI: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
    ; VI: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD29]](s32)
    ; VI: [[AND29:%[0-9]+]]:_(s16) = G_AND [[TRUNC29]], [[C7]]
    ; VI: [[SHL20:%[0-9]+]]:_(s16) = G_SHL [[AND29]], [[C8]](s16)
    ; VI: [[OR20:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[SHL20]]
    ; VI: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
    ; VI: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
    ; VI: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD31]](s32)
    ; VI: [[AND31:%[0-9]+]]:_(s16) = G_AND [[TRUNC31]], [[C7]]
    ; VI: [[SHL21:%[0-9]+]]:_(s16) = G_SHL [[AND31]], [[C8]](s16)
    ; VI: [[OR21:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[SHL21]]
    ; VI: [[ZEXT12:%[0-9]+]]:_(s32) = G_ZEXT [[OR18]](s16)
    ; VI: [[ZEXT13:%[0-9]+]]:_(s32) = G_ZEXT [[OR19]](s16)
    ; VI: [[SHL22:%[0-9]+]]:_(s32) = G_SHL [[ZEXT13]], [[C9]](s32)
    ; VI: [[OR22:%[0-9]+]]:_(s32) = G_OR [[ZEXT12]], [[SHL22]]
    ; VI: [[ZEXT14:%[0-9]+]]:_(s32) = G_ZEXT [[OR20]](s16)
    ; VI: [[ZEXT15:%[0-9]+]]:_(s32) = G_ZEXT [[OR21]](s16)
    ; VI: [[SHL23:%[0-9]+]]:_(s32) = G_SHL [[ZEXT15]], [[C9]](s32)
    ; VI: [[OR23:%[0-9]+]]:_(s32) = G_OR [[ZEXT14]], [[SHL23]]
    ; VI: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR22]](s32), [[OR23]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4s64_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s64>) = G_LOAD [[COPY]](p1) :: (load 32, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[LOAD]](<4 x s64>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4s64_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 5
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 7
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C7]]
    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C7]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C8]](s16)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C7]]
    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C7]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C8]](s16)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s16) = G_AND [[TRUNC4]], [[C7]]
    ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s16) = G_AND [[TRUNC5]], [[C7]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C8]](s16)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s16) = G_AND [[TRUNC6]], [[C7]]
    ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s16) = G_AND [[TRUNC7]], [[C7]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C8]](s16)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C9]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL4]]
    ; GFX9-MESA: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16)
    ; GFX9-MESA: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[OR3]](s16)
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C9]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL5]]
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR4]](s32), [[OR5]](s32)
    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C10]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; GFX9-MESA: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s16) = G_AND [[TRUNC8]], [[C7]]
    ; GFX9-MESA: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s16) = G_AND [[TRUNC9]], [[C7]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s16) = G_SHL [[AND9]], [[C8]](s16)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s16) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s16) = G_AND [[TRUNC10]], [[C7]]
    ; GFX9-MESA: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s16) = G_AND [[TRUNC11]], [[C7]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s16) = G_SHL [[AND11]], [[C8]](s16)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s16) = G_OR [[AND10]], [[SHL7]]
    ; GFX9-MESA: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD12]](s32)
    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s16) = G_AND [[TRUNC12]], [[C7]]
    ; GFX9-MESA: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD13]](s32)
    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s16) = G_AND [[TRUNC13]], [[C7]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s16) = G_SHL [[AND13]], [[C8]](s16)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s16) = G_OR [[AND12]], [[SHL8]]
    ; GFX9-MESA: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD14]](s32)
    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s16) = G_AND [[TRUNC14]], [[C7]]
    ; GFX9-MESA: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD15]](s32)
    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s16) = G_AND [[TRUNC15]], [[C7]]
    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s16) = G_SHL [[AND15]], [[C8]](s16)
    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s16) = G_OR [[AND14]], [[SHL9]]
    ; GFX9-MESA: [[ZEXT4:%[0-9]+]]:_(s32) = G_ZEXT [[OR6]](s16)
    ; GFX9-MESA: [[ZEXT5:%[0-9]+]]:_(s32) = G_ZEXT [[OR7]](s16)
    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[ZEXT5]], [[C9]](s32)
    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[ZEXT4]], [[SHL10]]
    ; GFX9-MESA: [[ZEXT6:%[0-9]+]]:_(s32) = G_ZEXT [[OR8]](s16)
    ; GFX9-MESA: [[ZEXT7:%[0-9]+]]:_(s32) = G_ZEXT [[OR9]](s16)
    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[ZEXT7]], [[C9]](s32)
    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[ZEXT6]], [[SHL11]]
    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR10]](s32), [[OR11]](s32)
    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
    ; GFX9-MESA: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C11]](s64)
    ; GFX9-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; GFX9-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; GFX9-MESA: [[TRUNC16:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD16]](s32)
    ; GFX9-MESA: [[AND16:%[0-9]+]]:_(s16) = G_AND [[TRUNC16]], [[C7]]
    ; GFX9-MESA: [[TRUNC17:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD17]](s32)
    ; GFX9-MESA: [[AND17:%[0-9]+]]:_(s16) = G_AND [[TRUNC17]], [[C7]]
    ; GFX9-MESA: [[SHL12:%[0-9]+]]:_(s16) = G_SHL [[AND17]], [[C8]](s16)
    ; GFX9-MESA: [[OR12:%[0-9]+]]:_(s16) = G_OR [[AND16]], [[SHL12]]
    ; GFX9-MESA: [[TRUNC18:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD18]](s32)
    ; GFX9-MESA: [[AND18:%[0-9]+]]:_(s16) = G_AND [[TRUNC18]], [[C7]]
    ; GFX9-MESA: [[TRUNC19:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD19]](s32)
    ; GFX9-MESA: [[AND19:%[0-9]+]]:_(s16) = G_AND [[TRUNC19]], [[C7]]
    ; GFX9-MESA: [[SHL13:%[0-9]+]]:_(s16) = G_SHL [[AND19]], [[C8]](s16)
    ; GFX9-MESA: [[OR13:%[0-9]+]]:_(s16) = G_OR [[AND18]], [[SHL13]]
    ; GFX9-MESA: [[TRUNC20:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD20]](s32)
    ; GFX9-MESA: [[AND20:%[0-9]+]]:_(s16) = G_AND [[TRUNC20]], [[C7]]
    ; GFX9-MESA: [[TRUNC21:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD21]](s32)
    ; GFX9-MESA: [[AND21:%[0-9]+]]:_(s16) = G_AND [[TRUNC21]], [[C7]]
    ; GFX9-MESA: [[SHL14:%[0-9]+]]:_(s16) = G_SHL [[AND21]], [[C8]](s16)
    ; GFX9-MESA: [[OR14:%[0-9]+]]:_(s16) = G_OR [[AND20]], [[SHL14]]
    ; GFX9-MESA: [[TRUNC22:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD22]](s32)
    ; GFX9-MESA: [[AND22:%[0-9]+]]:_(s16) = G_AND [[TRUNC22]], [[C7]]
    ; GFX9-MESA: [[TRUNC23:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD23]](s32)
    ; GFX9-MESA: [[AND23:%[0-9]+]]:_(s16) = G_AND [[TRUNC23]], [[C7]]
    ; GFX9-MESA: [[SHL15:%[0-9]+]]:_(s16) = G_SHL [[AND23]], [[C8]](s16)
    ; GFX9-MESA: [[OR15:%[0-9]+]]:_(s16) = G_OR [[AND22]], [[SHL15]]
    ; GFX9-MESA: [[ZEXT8:%[0-9]+]]:_(s32) = G_ZEXT [[OR12]](s16)
    ; GFX9-MESA: [[ZEXT9:%[0-9]+]]:_(s32) = G_ZEXT [[OR13]](s16)
    ; GFX9-MESA: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[ZEXT9]], [[C9]](s32)
    ; GFX9-MESA: [[OR16:%[0-9]+]]:_(s32) = G_OR [[ZEXT8]], [[SHL16]]
    ; GFX9-MESA: [[ZEXT10:%[0-9]+]]:_(s32) = G_ZEXT [[OR14]](s16)
    ; GFX9-MESA: [[ZEXT11:%[0-9]+]]:_(s32) = G_ZEXT [[OR15]](s16)
    ; GFX9-MESA: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[ZEXT11]], [[C9]](s32)
    ; GFX9-MESA: [[OR17:%[0-9]+]]:_(s32) = G_OR [[ZEXT10]], [[SHL17]]
    ; GFX9-MESA: [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR16]](s32), [[OR17]](s32)
    ; GFX9-MESA: [[C12:%[0-9]+]]:_(s64) = G_CONSTANT i64 24
    ; GFX9-MESA: [[PTR_ADD23:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C12]](s64)
    ; GFX9-MESA: [[LOAD24:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD23]](p1) :: (load 1 + 24, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD24:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C]](s64)
    ; GFX9-MESA: [[LOAD25:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD24]](p1) :: (load 1 + 25, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD25:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD26:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD25]](p1) :: (load 1 + 26, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD26:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD27:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD26]](p1) :: (load 1 + 27, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD27:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD28:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD27]](p1) :: (load 1 + 28, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD28:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD29:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD28]](p1) :: (load 1 + 29, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD29:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD30:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD29]](p1) :: (load 1 + 30, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD30:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD23]], [[C6]](s64)
    ; GFX9-MESA: [[LOAD31:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD30]](p1) :: (load 1 + 31, addrspace 1)
    ; GFX9-MESA: [[TRUNC24:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD24]](s32)
    ; GFX9-MESA: [[AND24:%[0-9]+]]:_(s16) = G_AND [[TRUNC24]], [[C7]]
    ; GFX9-MESA: [[TRUNC25:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD25]](s32)
    ; GFX9-MESA: [[AND25:%[0-9]+]]:_(s16) = G_AND [[TRUNC25]], [[C7]]
    ; GFX9-MESA: [[SHL18:%[0-9]+]]:_(s16) = G_SHL [[AND25]], [[C8]](s16)
    ; GFX9-MESA: [[OR18:%[0-9]+]]:_(s16) = G_OR [[AND24]], [[SHL18]]
    ; GFX9-MESA: [[TRUNC26:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD26]](s32)
    ; GFX9-MESA: [[AND26:%[0-9]+]]:_(s16) = G_AND [[TRUNC26]], [[C7]]
    ; GFX9-MESA: [[TRUNC27:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD27]](s32)
    ; GFX9-MESA: [[AND27:%[0-9]+]]:_(s16) = G_AND [[TRUNC27]], [[C7]]
    ; GFX9-MESA: [[SHL19:%[0-9]+]]:_(s16) = G_SHL [[AND27]], [[C8]](s16)
    ; GFX9-MESA: [[OR19:%[0-9]+]]:_(s16) = G_OR [[AND26]], [[SHL19]]
    ; GFX9-MESA: [[TRUNC28:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD28]](s32)
    ; GFX9-MESA: [[AND28:%[0-9]+]]:_(s16) = G_AND [[TRUNC28]], [[C7]]
    ; GFX9-MESA: [[TRUNC29:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD29]](s32)
    ; GFX9-MESA: [[AND29:%[0-9]+]]:_(s16) = G_AND [[TRUNC29]], [[C7]]
    ; GFX9-MESA: [[SHL20:%[0-9]+]]:_(s16) = G_SHL [[AND29]], [[C8]](s16)
    ; GFX9-MESA: [[OR20:%[0-9]+]]:_(s16) = G_OR [[AND28]], [[SHL20]]
    ; GFX9-MESA: [[TRUNC30:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD30]](s32)
    ; GFX9-MESA: [[AND30:%[0-9]+]]:_(s16) = G_AND [[TRUNC30]], [[C7]]
    ; GFX9-MESA: [[TRUNC31:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD31]](s32)
    ; GFX9-MESA: [[AND31:%[0-9]+]]:_(s16) = G_AND [[TRUNC31]], [[C7]]
    ; GFX9-MESA: [[SHL21:%[0-9]+]]:_(s16) = G_SHL [[AND31]], [[C8]](s16)
    ; GFX9-MESA: [[OR21:%[0-9]+]]:_(s16) = G_OR [[AND30]], [[SHL21]]
    ; GFX9-MESA: [[ZEXT12:%[0-9]+]]:_(s32) = G_ZEXT [[OR18]](s16)
    ; GFX9-MESA: [[ZEXT13:%[0-9]+]]:_(s32) = G_ZEXT [[OR19]](s16)
    ; GFX9-MESA: [[SHL22:%[0-9]+]]:_(s32) = G_SHL [[ZEXT13]], [[C9]](s32)
    ; GFX9-MESA: [[OR22:%[0-9]+]]:_(s32) = G_OR [[ZEXT12]], [[SHL22]]
    ; GFX9-MESA: [[ZEXT14:%[0-9]+]]:_(s32) = G_ZEXT [[OR20]](s16)
    ; GFX9-MESA: [[ZEXT15:%[0-9]+]]:_(s32) = G_ZEXT [[OR21]](s16)
    ; GFX9-MESA: [[SHL23:%[0-9]+]]:_(s32) = G_SHL [[ZEXT15]], [[C9]](s32)
    ; GFX9-MESA: [[OR23:%[0-9]+]]:_(s32) = G_OR [[ZEXT14]], [[SHL23]]
    ; GFX9-MESA: [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[OR22]](s32), [[OR23]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s64>) = G_BUILD_VECTOR [[MV]](s64), [[MV1]](s64), [[MV2]](s64), [[MV3]](s64)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<4 x s64>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s64>) = G_LOAD %0 :: (load 32, align 1, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---
name: test_load_global_v2s128_align32
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2s128_align32
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x s128>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<2 x s128>)
    ; CI-HSA-LABEL: name: test_load_global_v2s128_align32
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<2 x s128>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<2 x s128>)
    ; CI-MESA-LABEL: name: test_load_global_v2s128_align32
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s128>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<2 x s128>)
    ; VI-LABEL: name: test_load_global_v2s128_align32
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x s128>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<2 x s128>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2s128_align32
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<2 x s128>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<2 x s128>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2s128_align32
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<2 x s128>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<2 x s128>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s128>) = G_LOAD %0 :: (load 32, align 32, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---
name: test_load_global_v2p1_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2p1_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-HSA-LABEL: name: test_load_global_v2p1_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-MESA-LABEL: name: test_load_global_v2p1_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; VI-LABEL: name: test_load_global_v2p1_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 16, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2p1_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2p1_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-HSA-LABEL: name: test_load_global_v2p1_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-MESA-LABEL: name: test_load_global_v2p1_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; VI-LABEL: name: test_load_global_v2p1_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 8, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2p1_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2p1_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-HSA-LABEL: name: test_load_global_v2p1_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-MESA-LABEL: name: test_load_global_v2p1_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; VI-LABEL: name: test_load_global_v2p1_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 4, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v2p1_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2p1_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; SI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; SI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; SI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; SI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; SI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; SI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; SI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; SI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-HSA-LABEL: name: test_load_global_v2p1_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; CI-MESA-LABEL: name: test_load_global_v2p1_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; CI-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; CI-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; CI-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; CI-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; CI-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; VI-LABEL: name: test_load_global_v2p1_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; VI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; VI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; VI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; VI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; VI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; VI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; VI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; VI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; VI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; VI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; VI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; VI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; VI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; VI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; VI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; VI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; VI: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2p1_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, align 1, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[LOAD]](<4 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2p1_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C3]]
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32), [[OR8]](s32), [[OR11]](s32)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<2 x p1>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<2 x p1>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p1>) = G_LOAD %0 :: (load 16, align 1, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_load_global_v4p1_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v4p1_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; SI: [[BITCAST:%[0-9]+]]:_(<4 x p1>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<4 x p1>)
    ; CI-HSA-LABEL: name: test_load_global_v4p1_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<4 x p1>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<4 x p1>)
    ; CI-MESA-LABEL: name: test_load_global_v4p1_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<4 x p1>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<4 x p1>)
    ; VI-LABEL: name: test_load_global_v4p1_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(<4 x p1>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<4 x p1>)
    ; GFX9-HSA-LABEL: name: test_load_global_v4p1_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<4 x p1>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<4 x p1>)
    ; GFX9-MESA-LABEL: name: test_load_global_v4p1_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, align 8, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<4 x p1>) = G_BITCAST [[LOAD]](<8 x s32>)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<4 x p1>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x p1>) = G_LOAD %0 :: (load 32, align 8, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1
...

---
name: test_load_global_v2p3_align8
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2p3_align8
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; CI-HSA-LABEL: name: test_load_global_v2p3_align8
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; CI-MESA-LABEL: name: test_load_global_v2p3_align8
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; VI-LABEL: name: test_load_global_v2p3_align8
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2p3_align8
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2p3_align8
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p3>) = G_LOAD %0 :: (load 8, align 8, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v2p3_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2p3_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; CI-HSA-LABEL: name: test_load_global_v2p3_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; CI-MESA-LABEL: name: test_load_global_v2p3_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; VI-LABEL: name: test_load_global_v2p3_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2p3_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2p3_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p3>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_load_global_v2p3_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v2p3_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[INTTOPTR:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR2]](s32)
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; SI: [[INTTOPTR1:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR5]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[INTTOPTR]](p3), [[INTTOPTR1]](p3)
    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
    ; CI-HSA-LABEL: name: test_load_global_v2p3_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; CI-MESA-LABEL: name: test_load_global_v2p3_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[INTTOPTR:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR2]](s32)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; CI-MESA: [[INTTOPTR1:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR5]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[INTTOPTR]](p3), [[INTTOPTR1]](p3)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
    ; VI-LABEL: name: test_load_global_v2p3_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[INTTOPTR:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR2]](s32)
    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; VI: [[INTTOPTR1:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR5]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[INTTOPTR]](p3), [[INTTOPTR1]](p3)
    ; VI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
    ; GFX9-HSA-LABEL: name: test_load_global_v2p3_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x p3>) = G_LOAD [[COPY]](p1) :: (load 8, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
    ; GFX9-MESA-LABEL: name: test_load_global_v2p3_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[INTTOPTR:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR2]](s32)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; GFX9-MESA: [[INTTOPTR1:%[0-9]+]]:_(p3) = G_INTTOPTR [[OR5]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x p3>) = G_BUILD_VECTOR [[INTTOPTR]](p3), [[INTTOPTR1]](p3)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x p3>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x p3>) = G_LOAD %0 :: (load 8, align 1, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_ext_load_global_s32_from_1_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_ext_load_global_s32_from_1_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; SI: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-HSA-LABEL: name: test_ext_load_global_s32_from_1_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-MESA-LABEL: name: test_ext_load_global_s32_from_1_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-MESA: $vgpr0 = COPY [[LOAD]](s32)
    ; VI-LABEL: name: test_ext_load_global_s32_from_1_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; VI: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s32_from_1_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s32_from_1_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s32) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_ext_load_global_s32_from_2_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_ext_load_global_s32_from_2_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; SI: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-HSA-LABEL: name: test_ext_load_global_s32_from_2_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; CI-MESA-LABEL: name: test_ext_load_global_s32_from_2_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-MESA: $vgpr0 = COPY [[LOAD]](s32)
    ; VI-LABEL: name: test_ext_load_global_s32_from_2_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; VI: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s32_from_2_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s32_from_2_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s32) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_ext_load_global_s64_from_1_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1


    ; SI-LABEL: name: test_ext_load_global_s64_from_1_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; SI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_1_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_1_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; VI-LABEL: name: test_ext_load_global_s64_from_1_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; VI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_1_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_1_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_ext_load_global_s64_from_2_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_ext_load_global_s64_from_2_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; SI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_2_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_2_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; VI-LABEL: name: test_ext_load_global_s64_from_2_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; VI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_2_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_2_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_ext_load_global_s64_from_4_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; CI-LABEL: name: test_ext_load_global_s64_from_4_align4
    ; CI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
    ; SI-LABEL: name: test_ext_load_global_s64_from_4_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_4_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_4_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; VI-LABEL: name: test_ext_load_global_s64_from_4_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_4_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_4_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_ext_load_global_s128_from_4_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_ext_load_global_s128_from_4_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
    ; SI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[DEF]](s32)
    ; SI: [[DEF1:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
    ; SI: [[MV1:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s64), [[DEF1]](s64)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV1]](s128)
    ; CI-HSA-LABEL: name: test_ext_load_global_s128_from_4_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
    ; CI-HSA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[DEF]](s32)
    ; CI-HSA: [[DEF1:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
    ; CI-HSA: [[MV1:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s64), [[DEF1]](s64)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV1]](s128)
    ; CI-MESA-LABEL: name: test_ext_load_global_s128_from_4_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
    ; CI-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[DEF]](s32)
    ; CI-MESA: [[DEF1:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
    ; CI-MESA: [[MV1:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s64), [[DEF1]](s64)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV1]](s128)
    ; VI-LABEL: name: test_ext_load_global_s128_from_4_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
    ; VI: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[DEF]](s32)
    ; VI: [[DEF1:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
    ; VI: [[MV1:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s64), [[DEF1]](s64)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV1]](s128)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s128_from_4_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[DEF]](s32)
    ; GFX9-HSA: [[DEF1:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
    ; GFX9-HSA: [[MV1:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s64), [[DEF1]](s64)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV1]](s128)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s128_from_4_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[LOAD]](s32), [[DEF]](s32)
    ; GFX9-MESA: [[DEF1:%[0-9]+]]:_(s64) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[MV1:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[MV]](s64), [[DEF1]](s64)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV1]](s128)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s128) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_ext_load_global_s64_from_2_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_ext_load_global_s64_from_2_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; SI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_2_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_2_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; VI-LABEL: name: test_ext_load_global_s64_from_2_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; VI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_2_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_2_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, align 4, addrspace 1)
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 2, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_ext_load_global_s64_from_1_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_ext_load_global_s64_from_1_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; SI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; SI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-HSA-LABEL: name: test_ext_load_global_s64_from_1_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; CI-MESA-LABEL: name: test_ext_load_global_s64_from_1_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; CI-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; VI-LABEL: name: test_ext_load_global_s64_from_1_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; VI: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; VI: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-HSA-LABEL: name: test_ext_load_global_s64_from_1_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    ; GFX9-MESA-LABEL: name: test_ext_load_global_s64_from_1_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, align 4, addrspace 1)
    ; GFX9-MESA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[ANYEXT]](s64)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(s64) = G_LOAD %0 :: (load 1, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_extload_global_v2s32_from_4_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v2s32_from_4_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-HSA-LABEL: name: test_extload_global_v2s32_from_4_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-MESA-LABEL: name: test_extload_global_v2s32_from_4_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; VI-LABEL: name: test_extload_global_v2s32_from_4_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-HSA-LABEL: name: test_extload_global_v2s32_from_4_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-MESA-LABEL: name: test_extload_global_v2s32_from_4_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 1, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 1, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_extload_global_v2s32_from_4_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v2s32_from_4_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-HSA-LABEL: name: test_extload_global_v2s32_from_4_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-MESA-LABEL: name: test_extload_global_v2s32_from_4_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; VI-LABEL: name: test_extload_global_v2s32_from_4_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-HSA-LABEL: name: test_extload_global_v2s32_from_4_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-MESA-LABEL: name: test_extload_global_v2s32_from_4_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, align 2, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 2, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_extload_global_v2s32_from_4_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v2s32_from_4_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-HSA-LABEL: name: test_extload_global_v2s32_from_4_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; CI-MESA-LABEL: name: test_extload_global_v2s32_from_4_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; VI-LABEL: name: test_extload_global_v2s32_from_4_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-HSA-LABEL: name: test_extload_global_v2s32_from_4_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    ; GFX9-MESA-LABEL: name: test_extload_global_v2s32_from_4_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s32>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0_vgpr1 = COPY %1
...

---
name: test_extload_global_v3s32_from_6_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v3s32_from_6_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; CI-HSA-LABEL: name: test_extload_global_v3s32_from_6_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; CI-MESA-LABEL: name: test_extload_global_v3s32_from_6_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; VI-LABEL: name: test_extload_global_v3s32_from_6_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; GFX9-HSA-LABEL: name: test_extload_global_v3s32_from_6_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    ; GFX9-MESA-LABEL: name: test_extload_global_v3s32_from_6_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 6, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[LOAD]](<3 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<3 x s32>) = G_LOAD %0 :: (load 6, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2 = COPY %1
...

---
name: test_extload_global_v4s32_from_8_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v4s32_from_8_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-HSA-LABEL: name: test_extload_global_v4s32_from_8_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; CI-MESA-LABEL: name: test_extload_global_v4s32_from_8_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; VI-LABEL: name: test_extload_global_v4s32_from_8_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-HSA-LABEL: name: test_extload_global_v4s32_from_8_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    ; GFX9-MESA-LABEL: name: test_extload_global_v4s32_from_8_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[LOAD]](<4 x s32>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<4 x s32>) = G_LOAD %0 :: (load 8, align 4, addrspace 1)
    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
...

---
name: test_extload_global_v2s96_from_24_align1
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align1
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; SI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; SI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; SI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; SI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; SI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; SI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; SI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; SI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; SI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[COPY13:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY13]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; SI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; SI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; SI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; SI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; SI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; SI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; SI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; SI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; SI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; SI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; SI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; SI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C3]]
    ; SI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; SI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; SI: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C7]](s64)
    ; SI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; SI: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; SI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; SI: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; SI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; SI: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; SI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LOAD16]](s32)
    ; SI: [[AND16:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C3]]
    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; SI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C3]]
    ; SI: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[C4]](s32)
    ; SI: [[OR12:%[0-9]+]]:_(s32) = G_OR [[AND16]], [[SHL12]]
    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LOAD18]](s32)
    ; SI: [[AND18:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C3]]
    ; SI: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND18]], [[C5]](s32)
    ; SI: [[OR13:%[0-9]+]]:_(s32) = G_OR [[OR12]], [[SHL13]]
    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; SI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C3]]
    ; SI: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[C6]](s32)
    ; SI: [[OR14:%[0-9]+]]:_(s32) = G_OR [[OR13]], [[SHL14]]
    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR11]](s32), [[OR14]](s32)
    ; SI: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C8]](s64)
    ; SI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; SI: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C]](s64)
    ; SI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; SI: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C1]](s64)
    ; SI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; SI: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C2]](s64)
    ; SI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LOAD20]](s32)
    ; SI: [[AND20:%[0-9]+]]:_(s32) = G_AND [[COPY22]], [[C3]]
    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; SI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C3]]
    ; SI: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[C4]](s32)
    ; SI: [[OR15:%[0-9]+]]:_(s32) = G_OR [[AND20]], [[SHL15]]
    ; SI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LOAD22]](s32)
    ; SI: [[AND22:%[0-9]+]]:_(s32) = G_AND [[COPY24]], [[C3]]
    ; SI: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[AND22]], [[C5]](s32)
    ; SI: [[OR16:%[0-9]+]]:_(s32) = G_OR [[OR15]], [[SHL16]]
    ; SI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; SI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C3]]
    ; SI: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[C6]](s32)
    ; SI: [[OR17:%[0-9]+]]:_(s32) = G_OR [[OR16]], [[SHL17]]
    ; SI: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR17]](s32), 64
    ; SI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; SI: [[COPY26:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; SI: [[COPY27:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY26]](s96)
    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY27]](s96)
    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align1
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 1, addrspace 1)
    ; CI-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align1
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; CI-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; CI-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; CI-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; CI-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; CI-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; CI-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; CI-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; CI-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY13]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; CI-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; CI-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; CI-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; CI-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; CI-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; CI-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; CI-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; CI-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; CI-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; CI-MESA: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; CI-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; CI-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; CI-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; CI-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C3]]
    ; CI-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; CI-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; CI-MESA: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C7]](s64)
    ; CI-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; CI-MESA: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; CI-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; CI-MESA: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; CI-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; CI-MESA: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; CI-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LOAD16]](s32)
    ; CI-MESA: [[AND16:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C3]]
    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; CI-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C3]]
    ; CI-MESA: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[C4]](s32)
    ; CI-MESA: [[OR12:%[0-9]+]]:_(s32) = G_OR [[AND16]], [[SHL12]]
    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LOAD18]](s32)
    ; CI-MESA: [[AND18:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C3]]
    ; CI-MESA: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND18]], [[C5]](s32)
    ; CI-MESA: [[OR13:%[0-9]+]]:_(s32) = G_OR [[OR12]], [[SHL13]]
    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; CI-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C3]]
    ; CI-MESA: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[C6]](s32)
    ; CI-MESA: [[OR14:%[0-9]+]]:_(s32) = G_OR [[OR13]], [[SHL14]]
    ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR11]](s32), [[OR14]](s32)
    ; CI-MESA: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C8]](s64)
    ; CI-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; CI-MESA: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C]](s64)
    ; CI-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; CI-MESA: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C1]](s64)
    ; CI-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; CI-MESA: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C2]](s64)
    ; CI-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LOAD20]](s32)
    ; CI-MESA: [[AND20:%[0-9]+]]:_(s32) = G_AND [[COPY22]], [[C3]]
    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; CI-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C3]]
    ; CI-MESA: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[C4]](s32)
    ; CI-MESA: [[OR15:%[0-9]+]]:_(s32) = G_OR [[AND20]], [[SHL15]]
    ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LOAD22]](s32)
    ; CI-MESA: [[AND22:%[0-9]+]]:_(s32) = G_AND [[COPY24]], [[C3]]
    ; CI-MESA: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[AND22]], [[C5]](s32)
    ; CI-MESA: [[OR16:%[0-9]+]]:_(s32) = G_OR [[OR15]], [[SHL16]]
    ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; CI-MESA: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C3]]
    ; CI-MESA: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[C6]](s32)
    ; CI-MESA: [[OR17:%[0-9]+]]:_(s32) = G_OR [[OR16]], [[SHL17]]
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR17]](s32), 64
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; CI-MESA: [[COPY26:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-MESA: [[COPY27:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY26]](s96)
    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY27]](s96)
    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align1
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; VI: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; VI: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; VI: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; VI: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; VI: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; VI: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; VI: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; VI: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; VI: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; VI: [[COPY13:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY13]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; VI: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; VI: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; VI: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; VI: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; VI: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; VI: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; VI: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; VI: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; VI: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; VI: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; VI: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; VI: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; VI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; VI: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C3]]
    ; VI: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; VI: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; VI: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C7]](s64)
    ; VI: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; VI: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; VI: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; VI: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; VI: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; VI: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; VI: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; VI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LOAD16]](s32)
    ; VI: [[AND16:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C3]]
    ; VI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; VI: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C3]]
    ; VI: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[C4]](s32)
    ; VI: [[OR12:%[0-9]+]]:_(s32) = G_OR [[AND16]], [[SHL12]]
    ; VI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LOAD18]](s32)
    ; VI: [[AND18:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C3]]
    ; VI: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND18]], [[C5]](s32)
    ; VI: [[OR13:%[0-9]+]]:_(s32) = G_OR [[OR12]], [[SHL13]]
    ; VI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; VI: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C3]]
    ; VI: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[C6]](s32)
    ; VI: [[OR14:%[0-9]+]]:_(s32) = G_OR [[OR13]], [[SHL14]]
    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR11]](s32), [[OR14]](s32)
    ; VI: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C8]](s64)
    ; VI: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; VI: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C]](s64)
    ; VI: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; VI: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C1]](s64)
    ; VI: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; VI: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C2]](s64)
    ; VI: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; VI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LOAD20]](s32)
    ; VI: [[AND20:%[0-9]+]]:_(s32) = G_AND [[COPY22]], [[C3]]
    ; VI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; VI: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C3]]
    ; VI: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[C4]](s32)
    ; VI: [[OR15:%[0-9]+]]:_(s32) = G_OR [[AND20]], [[SHL15]]
    ; VI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LOAD22]](s32)
    ; VI: [[AND22:%[0-9]+]]:_(s32) = G_AND [[COPY24]], [[C3]]
    ; VI: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[AND22]], [[C5]](s32)
    ; VI: [[OR16:%[0-9]+]]:_(s32) = G_OR [[OR15]], [[SHL16]]
    ; VI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; VI: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C3]]
    ; VI: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[C6]](s32)
    ; VI: [[OR17:%[0-9]+]]:_(s32) = G_OR [[OR16]], [[SHL17]]
    ; VI: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR17]](s32), 64
    ; VI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; VI: [[COPY26:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; VI: [[COPY27:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY26]](s96)
    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY27]](s96)
    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align1
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 1, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 1, addrspace 1)
    ; GFX9-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align1
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 1, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 1 + 1, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 1 + 2, addrspace 1)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 1 + 3, addrspace 1)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C3]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C3]]
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C4]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C3]]
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND2]], [[C5]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[OR]], [[SHL1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C3]]
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C6]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[OR1]], [[SHL2]]
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 1 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 1 + 5, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 1 + 6, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 1 + 7, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C3]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C3]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C4]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL3]]
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY7]], [[C3]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND6]], [[C5]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[OR3]], [[SHL4]]
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C3]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C6]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[OR4]], [[SHL5]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR2]](s32), [[OR5]](s32)
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C8]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 1 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 1 + 9, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 1 + 10, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 1 + 11, addrspace 1)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C3]]
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C3]]
    ; GFX9-MESA: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C4]](s32)
    ; GFX9-MESA: [[OR6:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL6]]
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C3]]
    ; GFX9-MESA: [[SHL7:%[0-9]+]]:_(s32) = G_SHL [[AND10]], [[C5]](s32)
    ; GFX9-MESA: [[OR7:%[0-9]+]]:_(s32) = G_OR [[OR6]], [[SHL7]]
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C3]]
    ; GFX9-MESA: [[SHL8:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C6]](s32)
    ; GFX9-MESA: [[OR8:%[0-9]+]]:_(s32) = G_OR [[OR7]], [[SHL8]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY13]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR8]](s32), 64
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-MESA: [[PTR_ADD11:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C9]](s64)
    ; GFX9-MESA: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p1) :: (load 1 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD12:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C]](s64)
    ; GFX9-MESA: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p1) :: (load 1 + 13, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD13:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p1) :: (load 1 + 14, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD14:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p1) :: (load 1 + 15, addrspace 1)
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32)
    ; GFX9-MESA: [[AND12:%[0-9]+]]:_(s32) = G_AND [[COPY14]], [[C3]]
    ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32)
    ; GFX9-MESA: [[AND13:%[0-9]+]]:_(s32) = G_AND [[COPY15]], [[C3]]
    ; GFX9-MESA: [[SHL9:%[0-9]+]]:_(s32) = G_SHL [[AND13]], [[C4]](s32)
    ; GFX9-MESA: [[OR9:%[0-9]+]]:_(s32) = G_OR [[AND12]], [[SHL9]]
    ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32)
    ; GFX9-MESA: [[AND14:%[0-9]+]]:_(s32) = G_AND [[COPY16]], [[C3]]
    ; GFX9-MESA: [[SHL10:%[0-9]+]]:_(s32) = G_SHL [[AND14]], [[C5]](s32)
    ; GFX9-MESA: [[OR10:%[0-9]+]]:_(s32) = G_OR [[OR9]], [[SHL10]]
    ; GFX9-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32)
    ; GFX9-MESA: [[AND15:%[0-9]+]]:_(s32) = G_AND [[COPY17]], [[C3]]
    ; GFX9-MESA: [[SHL11:%[0-9]+]]:_(s32) = G_SHL [[AND15]], [[C6]](s32)
    ; GFX9-MESA: [[OR11:%[0-9]+]]:_(s32) = G_OR [[OR10]], [[SHL11]]
    ; GFX9-MESA: [[PTR_ADD15:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C7]](s64)
    ; GFX9-MESA: [[LOAD16:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD15]](p1) :: (load 1 + 16, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD16:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C]](s64)
    ; GFX9-MESA: [[LOAD17:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD16]](p1) :: (load 1 + 17, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD17:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD18:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD17]](p1) :: (load 1 + 18, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD18:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD15]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD19:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD18]](p1) :: (load 1 + 19, addrspace 1)
    ; GFX9-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LOAD16]](s32)
    ; GFX9-MESA: [[AND16:%[0-9]+]]:_(s32) = G_AND [[COPY18]], [[C3]]
    ; GFX9-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LOAD17]](s32)
    ; GFX9-MESA: [[AND17:%[0-9]+]]:_(s32) = G_AND [[COPY19]], [[C3]]
    ; GFX9-MESA: [[SHL12:%[0-9]+]]:_(s32) = G_SHL [[AND17]], [[C4]](s32)
    ; GFX9-MESA: [[OR12:%[0-9]+]]:_(s32) = G_OR [[AND16]], [[SHL12]]
    ; GFX9-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LOAD18]](s32)
    ; GFX9-MESA: [[AND18:%[0-9]+]]:_(s32) = G_AND [[COPY20]], [[C3]]
    ; GFX9-MESA: [[SHL13:%[0-9]+]]:_(s32) = G_SHL [[AND18]], [[C5]](s32)
    ; GFX9-MESA: [[OR13:%[0-9]+]]:_(s32) = G_OR [[OR12]], [[SHL13]]
    ; GFX9-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LOAD19]](s32)
    ; GFX9-MESA: [[AND19:%[0-9]+]]:_(s32) = G_AND [[COPY21]], [[C3]]
    ; GFX9-MESA: [[SHL14:%[0-9]+]]:_(s32) = G_SHL [[AND19]], [[C6]](s32)
    ; GFX9-MESA: [[OR14:%[0-9]+]]:_(s32) = G_OR [[OR13]], [[SHL14]]
    ; GFX9-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR11]](s32), [[OR14]](s32)
    ; GFX9-MESA: [[PTR_ADD19:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD11]], [[C8]](s64)
    ; GFX9-MESA: [[LOAD20:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD19]](p1) :: (load 1 + 20, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD20:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C]](s64)
    ; GFX9-MESA: [[LOAD21:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD20]](p1) :: (load 1 + 21, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD21:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C1]](s64)
    ; GFX9-MESA: [[LOAD22:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD21]](p1) :: (load 1 + 22, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD22:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD19]], [[C2]](s64)
    ; GFX9-MESA: [[LOAD23:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD22]](p1) :: (load 1 + 23, addrspace 1)
    ; GFX9-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LOAD20]](s32)
    ; GFX9-MESA: [[AND20:%[0-9]+]]:_(s32) = G_AND [[COPY22]], [[C3]]
    ; GFX9-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LOAD21]](s32)
    ; GFX9-MESA: [[AND21:%[0-9]+]]:_(s32) = G_AND [[COPY23]], [[C3]]
    ; GFX9-MESA: [[SHL15:%[0-9]+]]:_(s32) = G_SHL [[AND21]], [[C4]](s32)
    ; GFX9-MESA: [[OR15:%[0-9]+]]:_(s32) = G_OR [[AND20]], [[SHL15]]
    ; GFX9-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LOAD22]](s32)
    ; GFX9-MESA: [[AND22:%[0-9]+]]:_(s32) = G_AND [[COPY24]], [[C3]]
    ; GFX9-MESA: [[SHL16:%[0-9]+]]:_(s32) = G_SHL [[AND22]], [[C5]](s32)
    ; GFX9-MESA: [[OR16:%[0-9]+]]:_(s32) = G_OR [[OR15]], [[SHL16]]
    ; GFX9-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LOAD23]](s32)
    ; GFX9-MESA: [[AND23:%[0-9]+]]:_(s32) = G_AND [[COPY25]], [[C3]]
    ; GFX9-MESA: [[SHL17:%[0-9]+]]:_(s32) = G_SHL [[AND23]], [[C6]](s32)
    ; GFX9-MESA: [[OR17:%[0-9]+]]:_(s32) = G_OR [[OR16]], [[SHL17]]
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR17]](s32), 64
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; GFX9-MESA: [[COPY26:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-MESA: [[COPY27:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY26]](s96)
    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY27]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 1, addrspace 1)
    %2:_(s96) = G_EXTRACT %1, 0
    %3:_(s96) = G_EXTRACT %1, 96
    $vgpr0_vgpr1_vgpr2 = COPY %2
    $vgpr3_vgpr4_vgpr5 = COPY %3
...

---
name: test_extload_global_v2s96_from_24_align2
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align2
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; SI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; SI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; SI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; SI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; SI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; SI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; SI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; SI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[COPY7:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY7]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; SI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; SI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; SI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; SI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C1]]
    ; SI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C2]](s32)
    ; SI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; SI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C3]](s64)
    ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 2 + 16, addrspace 1)
    ; SI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 2 + 18, addrspace 1)
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; SI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C1]]
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; SI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C1]]
    ; SI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C2]](s32)
    ; SI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR3]](s32), [[OR4]](s32)
    ; SI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C4]](s64)
    ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 2 + 20, addrspace 1)
    ; SI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 2 + 22, addrspace 1)
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; SI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C1]]
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; SI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C1]]
    ; SI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C2]](s32)
    ; SI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL5]]
    ; SI: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR5]](s32), 64
    ; SI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; SI: [[COPY14:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; SI: [[COPY15:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY14]](s96)
    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY15]](s96)
    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align2
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 2, addrspace 1)
    ; CI-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align2
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; CI-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; CI-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; CI-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; CI-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; CI-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; CI-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; CI-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; CI-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; CI-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; CI-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; CI-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; CI-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; CI-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; CI-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY7]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; CI-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; CI-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; CI-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; CI-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; CI-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; CI-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C1]]
    ; CI-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C2]](s32)
    ; CI-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; CI-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C3]](s64)
    ; CI-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 2 + 16, addrspace 1)
    ; CI-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; CI-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 2 + 18, addrspace 1)
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; CI-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C1]]
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; CI-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C1]]
    ; CI-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C2]](s32)
    ; CI-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR3]](s32), [[OR4]](s32)
    ; CI-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C4]](s64)
    ; CI-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 2 + 20, addrspace 1)
    ; CI-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; CI-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 2 + 22, addrspace 1)
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; CI-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C1]]
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; CI-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C1]]
    ; CI-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C2]](s32)
    ; CI-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL5]]
    ; CI-MESA: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; CI-MESA: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR5]](s32), 64
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY14]](s96)
    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY15]](s96)
    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align2
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; VI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; VI: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; VI: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; VI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; VI: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; VI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; VI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; VI: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; VI: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; VI: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; VI: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; VI: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; VI: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; VI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; VI: [[COPY7:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; VI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY7]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; VI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; VI: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; VI: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; VI: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; VI: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; VI: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C1]]
    ; VI: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C2]](s32)
    ; VI: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; VI: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C3]](s64)
    ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 2 + 16, addrspace 1)
    ; VI: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 2 + 18, addrspace 1)
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; VI: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C1]]
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; VI: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C1]]
    ; VI: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C2]](s32)
    ; VI: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR3]](s32), [[OR4]](s32)
    ; VI: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C4]](s64)
    ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 2 + 20, addrspace 1)
    ; VI: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 2 + 22, addrspace 1)
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; VI: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C1]]
    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; VI: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C1]]
    ; VI: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C2]](s32)
    ; VI: [[OR5:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL5]]
    ; VI: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; VI: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR5]](s32), 64
    ; VI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; VI: [[COPY14:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; VI: [[COPY15:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY14]](s96)
    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY15]](s96)
    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align2
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 2, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 2, addrspace 1)
    ; GFX9-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align2
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 2, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 2 + 2, addrspace 1)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C1]]
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C1]]
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[C2]](s32)
    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s32) = G_OR [[AND]], [[SHL]]
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 2 + 4, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; GFX9-MESA: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 2 + 6, addrspace 1)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32)
    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C1]]
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32)
    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C1]]
    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C2]](s32)
    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s32) = G_OR [[AND2]], [[SHL1]]
    ; GFX9-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR]](s32), [[OR1]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; GFX9-MESA: [[PTR_ADD3:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p1) :: (load 2 + 8, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD4:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD3]], [[C]](s64)
    ; GFX9-MESA: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p1) :: (load 2 + 10, addrspace 1)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32)
    ; GFX9-MESA: [[AND4:%[0-9]+]]:_(s32) = G_AND [[COPY5]], [[C1]]
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32)
    ; GFX9-MESA: [[AND5:%[0-9]+]]:_(s32) = G_AND [[COPY6]], [[C1]]
    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[AND5]], [[C2]](s32)
    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[AND4]], [[SHL2]]
    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; GFX9-MESA: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY7]], [[BUILD_VECTOR]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[OR2]](s32), 64
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-MESA: [[PTR_ADD5:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C5]](s64)
    ; GFX9-MESA: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p1) :: (load 2 + 12, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD6:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C]](s64)
    ; GFX9-MESA: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p1) :: (load 2 + 14, addrspace 1)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32)
    ; GFX9-MESA: [[AND6:%[0-9]+]]:_(s32) = G_AND [[COPY8]], [[C1]]
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32)
    ; GFX9-MESA: [[AND7:%[0-9]+]]:_(s32) = G_AND [[COPY9]], [[C1]]
    ; GFX9-MESA: [[SHL3:%[0-9]+]]:_(s32) = G_SHL [[AND7]], [[C2]](s32)
    ; GFX9-MESA: [[OR3:%[0-9]+]]:_(s32) = G_OR [[AND6]], [[SHL3]]
    ; GFX9-MESA: [[PTR_ADD7:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C3]](s64)
    ; GFX9-MESA: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p1) :: (load 2 + 16, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD8:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD7]], [[C]](s64)
    ; GFX9-MESA: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p1) :: (load 2 + 18, addrspace 1)
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32)
    ; GFX9-MESA: [[AND8:%[0-9]+]]:_(s32) = G_AND [[COPY10]], [[C1]]
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32)
    ; GFX9-MESA: [[AND9:%[0-9]+]]:_(s32) = G_AND [[COPY11]], [[C1]]
    ; GFX9-MESA: [[SHL4:%[0-9]+]]:_(s32) = G_SHL [[AND9]], [[C2]](s32)
    ; GFX9-MESA: [[OR4:%[0-9]+]]:_(s32) = G_OR [[AND8]], [[SHL4]]
    ; GFX9-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[OR3]](s32), [[OR4]](s32)
    ; GFX9-MESA: [[PTR_ADD9:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD5]], [[C4]](s64)
    ; GFX9-MESA: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p1) :: (load 2 + 20, addrspace 1)
    ; GFX9-MESA: [[PTR_ADD10:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD9]], [[C]](s64)
    ; GFX9-MESA: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p1) :: (load 2 + 22, addrspace 1)
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32)
    ; GFX9-MESA: [[AND10:%[0-9]+]]:_(s32) = G_AND [[COPY12]], [[C1]]
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32)
    ; GFX9-MESA: [[AND11:%[0-9]+]]:_(s32) = G_AND [[COPY13]], [[C1]]
    ; GFX9-MESA: [[SHL5:%[0-9]+]]:_(s32) = G_SHL [[AND11]], [[C2]](s32)
    ; GFX9-MESA: [[OR5:%[0-9]+]]:_(s32) = G_OR [[AND10]], [[SHL5]]
    ; GFX9-MESA: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[BUILD_VECTOR1]](<2 x s32>), 0
    ; GFX9-MESA: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[OR5]](s32), 64
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY14]](s96)
    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY15]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 2, addrspace 1)
    %2:_(s96) = G_EXTRACT %1, 0
    %3:_(s96) = G_EXTRACT %1, 96
    $vgpr0_vgpr1_vgpr2 = COPY %2
    $vgpr3_vgpr4_vgpr5 = COPY %3
...

---
name: test_extload_global_v2s96_from_24_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, align 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load 4 + 8, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[COPY1:%[0-9]+]]:_(<3 x s32>) = COPY [[DEF]](<3 x s32>)
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[COPY1]], [[LOAD]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD1]](s32), 64
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[PTR_ADD1]](p1) :: (load 8 + 12, align 4, addrspace 1)
    ; SI: [[PTR_ADD2:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD1]], [[C]](s64)
    ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p1) :: (load 4 + 20, addrspace 1)
    ; SI: [[INSERT2:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD2]](<2 x s32>), 0
    ; SI: [[INSERT3:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT2]], [[LOAD3]](s32), 64
    ; SI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT3]](<3 x s32>)
    ; SI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; SI: [[COPY3:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY2]](s96)
    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY3]](s96)
    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; CI-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; VI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; VI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; VI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; GFX9-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 4, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 4, addrspace 1)
    %2:_(s96) = G_EXTRACT %1, 0
    %3:_(s96) = G_EXTRACT %1, 96
    $vgpr0_vgpr1_vgpr2 = COPY %2
    $vgpr3_vgpr4_vgpr5 = COPY %3
...

---
name: test_extload_global_v2s96_from_24_align16
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_extload_global_v2s96_from_24_align16
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1)
    ; SI: [[EXTRACT:%[0-9]+]]:_(<3 x s32>) = G_EXTRACT [[LOAD]](<4 x s32>), 0
    ; SI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[EXTRACT]](<3 x s32>)
    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; SI: [[LOAD1:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 8 + 12, align 4, addrspace 1)
    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[PTR_ADD]], [[C1]](s64)
    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load 4 + 20, addrspace 1)
    ; SI: [[DEF:%[0-9]+]]:_(<3 x s32>) = G_IMPLICIT_DEF
    ; SI: [[INSERT:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[DEF]], [[LOAD1]](<2 x s32>), 0
    ; SI: [[INSERT1:%[0-9]+]]:_(<3 x s32>) = G_INSERT [[INSERT]], [[LOAD2]](s32), 64
    ; SI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[INSERT1]](<3 x s32>)
    ; SI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; SI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; SI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; SI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; CI-HSA-LABEL: name: test_extload_global_v2s96_from_24_align16
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; CI-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; CI-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; CI-MESA-LABEL: name: test_extload_global_v2s96_from_24_align16
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; CI-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; CI-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; CI-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; CI-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; VI-LABEL: name: test_extload_global_v2s96_from_24_align16
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; VI: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; VI: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; VI: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; VI: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; VI: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; VI: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; VI: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; GFX9-HSA-LABEL: name: test_extload_global_v2s96_from_24_align16
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-HSA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; GFX9-HSA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-HSA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; GFX9-HSA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    ; GFX9-MESA-LABEL: name: test_extload_global_v2s96_from_24_align16
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[COPY]](p1) :: (load 12, align 16, addrspace 1)
    ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD]](<3 x s32>)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(<3 x s32>) = G_LOAD [[PTR_ADD]](p1) :: (load 12 + 12, align 4, addrspace 1)
    ; GFX9-MESA: [[BITCAST1:%[0-9]+]]:_(s96) = G_BITCAST [[LOAD1]](<3 x s32>)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s96) = COPY [[BITCAST]](s96)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s96) = COPY [[BITCAST1]](s96)
    ; GFX9-MESA: $vgpr0_vgpr1_vgpr2 = COPY [[COPY1]](s96)
    ; GFX9-MESA: $vgpr3_vgpr4_vgpr5 = COPY [[COPY2]](s96)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<2 x s96>) = G_LOAD %0 :: (load 24, align 16, addrspace 1)
    %2:_(s96) = G_EXTRACT %1, 0
    %3:_(s96) = G_EXTRACT %1, 96
    $vgpr0_vgpr1_vgpr2 = COPY %2
    $vgpr3_vgpr4_vgpr5 = COPY %3
...

---
name: test_load_global_v32s1_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v32s1_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
    ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
    ; SI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; SI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C7]](s32)
    ; SI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
    ; SI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C8]](s32)
    ; SI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
    ; SI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C9]](s32)
    ; SI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
    ; SI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C10]](s32)
    ; SI: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; SI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C11]](s32)
    ; SI: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
    ; SI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C12]](s32)
    ; SI: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
    ; SI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C13]](s32)
    ; SI: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
    ; SI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C14]](s32)
    ; SI: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C15]](s32)
    ; SI: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 17
    ; SI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C16]](s32)
    ; SI: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 18
    ; SI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C17]](s32)
    ; SI: [[C18:%[0-9]+]]:_(s32) = G_CONSTANT i32 19
    ; SI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C18]](s32)
    ; SI: [[C19:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; SI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C19]](s32)
    ; SI: [[C20:%[0-9]+]]:_(s32) = G_CONSTANT i32 21
    ; SI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C20]](s32)
    ; SI: [[C21:%[0-9]+]]:_(s32) = G_CONSTANT i32 22
    ; SI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C21]](s32)
    ; SI: [[C22:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
    ; SI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C22]](s32)
    ; SI: [[C23:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C23]](s32)
    ; SI: [[C24:%[0-9]+]]:_(s32) = G_CONSTANT i32 25
    ; SI: [[LSHR24:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C24]](s32)
    ; SI: [[C25:%[0-9]+]]:_(s32) = G_CONSTANT i32 26
    ; SI: [[LSHR25:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C25]](s32)
    ; SI: [[C26:%[0-9]+]]:_(s32) = G_CONSTANT i32 27
    ; SI: [[LSHR26:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C26]](s32)
    ; SI: [[C27:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; SI: [[LSHR27:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C27]](s32)
    ; SI: [[C28:%[0-9]+]]:_(s32) = G_CONSTANT i32 29
    ; SI: [[LSHR28:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C28]](s32)
    ; SI: [[C29:%[0-9]+]]:_(s32) = G_CONSTANT i32 30
    ; SI: [[LSHR29:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C29]](s32)
    ; SI: [[C30:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
    ; SI: [[LSHR30:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C30]](s32)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; SI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; SI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; SI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR24]](s32)
    ; SI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR25]](s32)
    ; SI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR26]](s32)
    ; SI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LSHR27]](s32)
    ; SI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR28]](s32)
    ; SI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR29]](s32)
    ; SI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR30]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<32 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32), [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32), [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32), [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32), [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32), [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32), [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<32 x s1>) = G_TRUNC [[BUILD_VECTOR]](<32 x s32>)
    ; SI: $vgpr0 = COPY [[TRUNC]](<32 x s1>)
    ; CI-HSA-LABEL: name: test_load_global_v32s1_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; CI-HSA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
    ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; CI-HSA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
    ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; CI-HSA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
    ; CI-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; CI-HSA: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C7]](s32)
    ; CI-HSA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
    ; CI-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C8]](s32)
    ; CI-HSA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
    ; CI-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C9]](s32)
    ; CI-HSA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
    ; CI-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C10]](s32)
    ; CI-HSA: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; CI-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C11]](s32)
    ; CI-HSA: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
    ; CI-HSA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C12]](s32)
    ; CI-HSA: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
    ; CI-HSA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C13]](s32)
    ; CI-HSA: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
    ; CI-HSA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C14]](s32)
    ; CI-HSA: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C15]](s32)
    ; CI-HSA: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 17
    ; CI-HSA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C16]](s32)
    ; CI-HSA: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 18
    ; CI-HSA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C17]](s32)
    ; CI-HSA: [[C18:%[0-9]+]]:_(s32) = G_CONSTANT i32 19
    ; CI-HSA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C18]](s32)
    ; CI-HSA: [[C19:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; CI-HSA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C19]](s32)
    ; CI-HSA: [[C20:%[0-9]+]]:_(s32) = G_CONSTANT i32 21
    ; CI-HSA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C20]](s32)
    ; CI-HSA: [[C21:%[0-9]+]]:_(s32) = G_CONSTANT i32 22
    ; CI-HSA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C21]](s32)
    ; CI-HSA: [[C22:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
    ; CI-HSA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C22]](s32)
    ; CI-HSA: [[C23:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C23]](s32)
    ; CI-HSA: [[C24:%[0-9]+]]:_(s32) = G_CONSTANT i32 25
    ; CI-HSA: [[LSHR24:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C24]](s32)
    ; CI-HSA: [[C25:%[0-9]+]]:_(s32) = G_CONSTANT i32 26
    ; CI-HSA: [[LSHR25:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C25]](s32)
    ; CI-HSA: [[C26:%[0-9]+]]:_(s32) = G_CONSTANT i32 27
    ; CI-HSA: [[LSHR26:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C26]](s32)
    ; CI-HSA: [[C27:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; CI-HSA: [[LSHR27:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C27]](s32)
    ; CI-HSA: [[C28:%[0-9]+]]:_(s32) = G_CONSTANT i32 29
    ; CI-HSA: [[LSHR28:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C28]](s32)
    ; CI-HSA: [[C29:%[0-9]+]]:_(s32) = G_CONSTANT i32 30
    ; CI-HSA: [[LSHR29:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C29]](s32)
    ; CI-HSA: [[C30:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
    ; CI-HSA: [[LSHR30:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C30]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; CI-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; CI-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; CI-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; CI-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; CI-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; CI-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; CI-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; CI-HSA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; CI-HSA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; CI-HSA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; CI-HSA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; CI-HSA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; CI-HSA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; CI-HSA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; CI-HSA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; CI-HSA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; CI-HSA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR24]](s32)
    ; CI-HSA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR25]](s32)
    ; CI-HSA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR26]](s32)
    ; CI-HSA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LSHR27]](s32)
    ; CI-HSA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR28]](s32)
    ; CI-HSA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR29]](s32)
    ; CI-HSA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR30]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<32 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32), [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32), [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32), [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32), [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32), [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32), [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<32 x s1>) = G_TRUNC [[BUILD_VECTOR]](<32 x s32>)
    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](<32 x s1>)
    ; CI-MESA-LABEL: name: test_load_global_v32s1_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
    ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
    ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; CI-MESA: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C7]](s32)
    ; CI-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
    ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C8]](s32)
    ; CI-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
    ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C9]](s32)
    ; CI-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
    ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C10]](s32)
    ; CI-MESA: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C11]](s32)
    ; CI-MESA: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
    ; CI-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C12]](s32)
    ; CI-MESA: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
    ; CI-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C13]](s32)
    ; CI-MESA: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
    ; CI-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C14]](s32)
    ; CI-MESA: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C15]](s32)
    ; CI-MESA: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 17
    ; CI-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C16]](s32)
    ; CI-MESA: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 18
    ; CI-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C17]](s32)
    ; CI-MESA: [[C18:%[0-9]+]]:_(s32) = G_CONSTANT i32 19
    ; CI-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C18]](s32)
    ; CI-MESA: [[C19:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; CI-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C19]](s32)
    ; CI-MESA: [[C20:%[0-9]+]]:_(s32) = G_CONSTANT i32 21
    ; CI-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C20]](s32)
    ; CI-MESA: [[C21:%[0-9]+]]:_(s32) = G_CONSTANT i32 22
    ; CI-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C21]](s32)
    ; CI-MESA: [[C22:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
    ; CI-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C22]](s32)
    ; CI-MESA: [[C23:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C23]](s32)
    ; CI-MESA: [[C24:%[0-9]+]]:_(s32) = G_CONSTANT i32 25
    ; CI-MESA: [[LSHR24:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C24]](s32)
    ; CI-MESA: [[C25:%[0-9]+]]:_(s32) = G_CONSTANT i32 26
    ; CI-MESA: [[LSHR25:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C25]](s32)
    ; CI-MESA: [[C26:%[0-9]+]]:_(s32) = G_CONSTANT i32 27
    ; CI-MESA: [[LSHR26:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C26]](s32)
    ; CI-MESA: [[C27:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; CI-MESA: [[LSHR27:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C27]](s32)
    ; CI-MESA: [[C28:%[0-9]+]]:_(s32) = G_CONSTANT i32 29
    ; CI-MESA: [[LSHR28:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C28]](s32)
    ; CI-MESA: [[C29:%[0-9]+]]:_(s32) = G_CONSTANT i32 30
    ; CI-MESA: [[LSHR29:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C29]](s32)
    ; CI-MESA: [[C30:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
    ; CI-MESA: [[LSHR30:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C30]](s32)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; CI-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR24]](s32)
    ; CI-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR25]](s32)
    ; CI-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR26]](s32)
    ; CI-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LSHR27]](s32)
    ; CI-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR28]](s32)
    ; CI-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR29]](s32)
    ; CI-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR30]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<32 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32), [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32), [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32), [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32), [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32), [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32), [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<32 x s1>) = G_TRUNC [[BUILD_VECTOR]](<32 x s32>)
    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<32 x s1>)
    ; VI-LABEL: name: test_load_global_v32s1_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
    ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
    ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; VI: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C7]](s32)
    ; VI: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
    ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C8]](s32)
    ; VI: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
    ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C9]](s32)
    ; VI: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
    ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C10]](s32)
    ; VI: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C11]](s32)
    ; VI: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
    ; VI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C12]](s32)
    ; VI: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
    ; VI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C13]](s32)
    ; VI: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
    ; VI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C14]](s32)
    ; VI: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C15]](s32)
    ; VI: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 17
    ; VI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C16]](s32)
    ; VI: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 18
    ; VI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C17]](s32)
    ; VI: [[C18:%[0-9]+]]:_(s32) = G_CONSTANT i32 19
    ; VI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C18]](s32)
    ; VI: [[C19:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; VI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C19]](s32)
    ; VI: [[C20:%[0-9]+]]:_(s32) = G_CONSTANT i32 21
    ; VI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C20]](s32)
    ; VI: [[C21:%[0-9]+]]:_(s32) = G_CONSTANT i32 22
    ; VI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C21]](s32)
    ; VI: [[C22:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
    ; VI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C22]](s32)
    ; VI: [[C23:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C23]](s32)
    ; VI: [[C24:%[0-9]+]]:_(s32) = G_CONSTANT i32 25
    ; VI: [[LSHR24:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C24]](s32)
    ; VI: [[C25:%[0-9]+]]:_(s32) = G_CONSTANT i32 26
    ; VI: [[LSHR25:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C25]](s32)
    ; VI: [[C26:%[0-9]+]]:_(s32) = G_CONSTANT i32 27
    ; VI: [[LSHR26:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C26]](s32)
    ; VI: [[C27:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; VI: [[LSHR27:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C27]](s32)
    ; VI: [[C28:%[0-9]+]]:_(s32) = G_CONSTANT i32 29
    ; VI: [[LSHR28:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C28]](s32)
    ; VI: [[C29:%[0-9]+]]:_(s32) = G_CONSTANT i32 30
    ; VI: [[LSHR29:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C29]](s32)
    ; VI: [[C30:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
    ; VI: [[LSHR30:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C30]](s32)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; VI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; VI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; VI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; VI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; VI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; VI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; VI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; VI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; VI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; VI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR24]](s32)
    ; VI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR25]](s32)
    ; VI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR26]](s32)
    ; VI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LSHR27]](s32)
    ; VI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR28]](s32)
    ; VI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR29]](s32)
    ; VI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR30]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<32 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32), [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32), [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32), [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32), [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32), [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32), [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<32 x s1>) = G_TRUNC [[BUILD_VECTOR]](<32 x s32>)
    ; VI: $vgpr0 = COPY [[TRUNC]](<32 x s1>)
    ; GFX9-HSA-LABEL: name: test_load_global_v32s1_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
    ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; GFX9-HSA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
    ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; GFX9-HSA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
    ; GFX9-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; GFX9-HSA: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C7]](s32)
    ; GFX9-HSA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
    ; GFX9-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C8]](s32)
    ; GFX9-HSA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
    ; GFX9-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C9]](s32)
    ; GFX9-HSA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
    ; GFX9-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C10]](s32)
    ; GFX9-HSA: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; GFX9-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C11]](s32)
    ; GFX9-HSA: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
    ; GFX9-HSA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C12]](s32)
    ; GFX9-HSA: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
    ; GFX9-HSA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C13]](s32)
    ; GFX9-HSA: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
    ; GFX9-HSA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C14]](s32)
    ; GFX9-HSA: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C15]](s32)
    ; GFX9-HSA: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 17
    ; GFX9-HSA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C16]](s32)
    ; GFX9-HSA: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 18
    ; GFX9-HSA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C17]](s32)
    ; GFX9-HSA: [[C18:%[0-9]+]]:_(s32) = G_CONSTANT i32 19
    ; GFX9-HSA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C18]](s32)
    ; GFX9-HSA: [[C19:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; GFX9-HSA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C19]](s32)
    ; GFX9-HSA: [[C20:%[0-9]+]]:_(s32) = G_CONSTANT i32 21
    ; GFX9-HSA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C20]](s32)
    ; GFX9-HSA: [[C21:%[0-9]+]]:_(s32) = G_CONSTANT i32 22
    ; GFX9-HSA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C21]](s32)
    ; GFX9-HSA: [[C22:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
    ; GFX9-HSA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C22]](s32)
    ; GFX9-HSA: [[C23:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C23]](s32)
    ; GFX9-HSA: [[C24:%[0-9]+]]:_(s32) = G_CONSTANT i32 25
    ; GFX9-HSA: [[LSHR24:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C24]](s32)
    ; GFX9-HSA: [[C25:%[0-9]+]]:_(s32) = G_CONSTANT i32 26
    ; GFX9-HSA: [[LSHR25:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C25]](s32)
    ; GFX9-HSA: [[C26:%[0-9]+]]:_(s32) = G_CONSTANT i32 27
    ; GFX9-HSA: [[LSHR26:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C26]](s32)
    ; GFX9-HSA: [[C27:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; GFX9-HSA: [[LSHR27:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C27]](s32)
    ; GFX9-HSA: [[C28:%[0-9]+]]:_(s32) = G_CONSTANT i32 29
    ; GFX9-HSA: [[LSHR28:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C28]](s32)
    ; GFX9-HSA: [[C29:%[0-9]+]]:_(s32) = G_CONSTANT i32 30
    ; GFX9-HSA: [[LSHR29:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C29]](s32)
    ; GFX9-HSA: [[C30:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
    ; GFX9-HSA: [[LSHR30:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C30]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; GFX9-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; GFX9-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32)
    ; GFX9-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; GFX9-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32)
    ; GFX9-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; GFX9-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32)
    ; GFX9-HSA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; GFX9-HSA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32)
    ; GFX9-HSA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; GFX9-HSA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32)
    ; GFX9-HSA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; GFX9-HSA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32)
    ; GFX9-HSA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; GFX9-HSA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32)
    ; GFX9-HSA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; GFX9-HSA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR24]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32)
    ; GFX9-HSA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR25]](s32)
    ; GFX9-HSA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR26]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32)
    ; GFX9-HSA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LSHR27]](s32)
    ; GFX9-HSA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR28]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32)
    ; GFX9-HSA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR29]](s32)
    ; GFX9-HSA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR30]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>), [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>), [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>), [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>), [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>), [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>), [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>), [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<32 x s1>) = G_TRUNC [[CONCAT_VECTORS]](<32 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](<32 x s1>)
    ; GFX9-MESA-LABEL: name: test_load_global_v32s1_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 5
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 6
    ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 7
    ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; GFX9-MESA: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C7]](s32)
    ; GFX9-MESA: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 9
    ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C8]](s32)
    ; GFX9-MESA: [[C9:%[0-9]+]]:_(s32) = G_CONSTANT i32 10
    ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C9]](s32)
    ; GFX9-MESA: [[C10:%[0-9]+]]:_(s32) = G_CONSTANT i32 11
    ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C10]](s32)
    ; GFX9-MESA: [[C11:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C11]](s32)
    ; GFX9-MESA: [[C12:%[0-9]+]]:_(s32) = G_CONSTANT i32 13
    ; GFX9-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C12]](s32)
    ; GFX9-MESA: [[C13:%[0-9]+]]:_(s32) = G_CONSTANT i32 14
    ; GFX9-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C13]](s32)
    ; GFX9-MESA: [[C14:%[0-9]+]]:_(s32) = G_CONSTANT i32 15
    ; GFX9-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C14]](s32)
    ; GFX9-MESA: [[C15:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C15]](s32)
    ; GFX9-MESA: [[C16:%[0-9]+]]:_(s32) = G_CONSTANT i32 17
    ; GFX9-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C16]](s32)
    ; GFX9-MESA: [[C17:%[0-9]+]]:_(s32) = G_CONSTANT i32 18
    ; GFX9-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C17]](s32)
    ; GFX9-MESA: [[C18:%[0-9]+]]:_(s32) = G_CONSTANT i32 19
    ; GFX9-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C18]](s32)
    ; GFX9-MESA: [[C19:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; GFX9-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C19]](s32)
    ; GFX9-MESA: [[C20:%[0-9]+]]:_(s32) = G_CONSTANT i32 21
    ; GFX9-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C20]](s32)
    ; GFX9-MESA: [[C21:%[0-9]+]]:_(s32) = G_CONSTANT i32 22
    ; GFX9-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C21]](s32)
    ; GFX9-MESA: [[C22:%[0-9]+]]:_(s32) = G_CONSTANT i32 23
    ; GFX9-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C22]](s32)
    ; GFX9-MESA: [[C23:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C23]](s32)
    ; GFX9-MESA: [[C24:%[0-9]+]]:_(s32) = G_CONSTANT i32 25
    ; GFX9-MESA: [[LSHR24:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C24]](s32)
    ; GFX9-MESA: [[C25:%[0-9]+]]:_(s32) = G_CONSTANT i32 26
    ; GFX9-MESA: [[LSHR25:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C25]](s32)
    ; GFX9-MESA: [[C26:%[0-9]+]]:_(s32) = G_CONSTANT i32 27
    ; GFX9-MESA: [[LSHR26:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C26]](s32)
    ; GFX9-MESA: [[C27:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; GFX9-MESA: [[LSHR27:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C27]](s32)
    ; GFX9-MESA: [[C28:%[0-9]+]]:_(s32) = G_CONSTANT i32 29
    ; GFX9-MESA: [[LSHR28:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C28]](s32)
    ; GFX9-MESA: [[C29:%[0-9]+]]:_(s32) = G_CONSTANT i32 30
    ; GFX9-MESA: [[LSHR29:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C29]](s32)
    ; GFX9-MESA: [[C30:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
    ; GFX9-MESA: [[LSHR30:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C30]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32)
    ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32)
    ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32)
    ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32)
    ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32)
    ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32)
    ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32)
    ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32)
    ; GFX9-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32)
    ; GFX9-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32)
    ; GFX9-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32)
    ; GFX9-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32)
    ; GFX9-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32)
    ; GFX9-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32)
    ; GFX9-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32)
    ; GFX9-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32)
    ; GFX9-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32)
    ; GFX9-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR24]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32)
    ; GFX9-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR25]](s32)
    ; GFX9-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR26]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32)
    ; GFX9-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[LSHR27]](s32)
    ; GFX9-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR28]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32)
    ; GFX9-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR29]](s32)
    ; GFX9-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR30]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>), [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>), [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>), [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>), [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>), [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>), [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>), [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<32 x s1>) = G_TRUNC [[CONCAT_VECTORS]](<32 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<32 x s1>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<32 x s1>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0 = COPY %1
...

---
name: test_load_global_v8s4_align4
body: |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; SI-LABEL: name: test_load_global_v8s4_align4
    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; SI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; SI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; SI: [[TRUNC:%[0-9]+]]:_(<8 x s4>) = G_TRUNC [[BUILD_VECTOR]](<8 x s32>)
    ; SI: $vgpr0 = COPY [[TRUNC]](<8 x s4>)
    ; CI-HSA-LABEL: name: test_load_global_v8s4_align4
    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-HSA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; CI-HSA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; CI-HSA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; CI-HSA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; CI-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<8 x s4>) = G_TRUNC [[BUILD_VECTOR]](<8 x s32>)
    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](<8 x s4>)
    ; CI-MESA-LABEL: name: test_load_global_v8s4_align4
    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; CI-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<8 x s4>) = G_TRUNC [[BUILD_VECTOR]](<8 x s32>)
    ; CI-MESA: $vgpr0 = COPY [[TRUNC]](<8 x s4>)
    ; VI-LABEL: name: test_load_global_v8s4_align4
    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; VI: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32), [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32)
    ; VI: [[TRUNC:%[0-9]+]]:_(<8 x s4>) = G_TRUNC [[BUILD_VECTOR]](<8 x s32>)
    ; VI: $vgpr0 = COPY [[TRUNC]](<8 x s4>)
    ; GFX9-HSA-LABEL: name: test_load_global_v8s4_align4
    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-HSA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; GFX9-HSA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; GFX9-HSA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; GFX9-HSA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; GFX9-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>), [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<8 x s4>) = G_TRUNC [[CONCAT_VECTORS]](<8 x s16>)
    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](<8 x s4>)
    ; GFX9-MESA-LABEL: name: test_load_global_v8s4_align4
    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load 4, addrspace 1)
    ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4
    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C]](s32)
    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 12
    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
    ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 20
    ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C4]](s32)
    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
    ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C5]](s32)
    ; GFX9-MESA: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 28
    ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C6]](s32)
    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32)
    ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32)
    ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32)
    ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32)
    ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32)
    ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32)
    ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32)
    ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32)
    ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32)
    ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>), [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>)
    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<8 x s4>) = G_TRUNC [[CONCAT_VECTORS]](<8 x s16>)
    ; GFX9-MESA: $vgpr0 = COPY [[TRUNC]](<8 x s4>)
    %0:_(p1) = COPY $vgpr0_vgpr1
    %1:_(<8 x s4>) = G_LOAD %0 :: (load 4, align 4, addrspace 1)
    $vgpr0 = COPY %1
...