llvm/test/CodeGen/AMDGPU/GlobalISel/load-divergent.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -mattr=+unaligned-access-mode,-real-true16 < %s | FileCheck --check-prefixes=GFX12,GFX12-True16 %s
; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -mattr=+unaligned-access-mode,+real-true16 < %s | FileCheck --check-prefixes=GFX12,GFX12-NoTrue16 %s

define amdgpu_ps void @load_divergent_P0_i8_any_extending(ptr addrspace(0) %ptra, ptr addrspace(0) %out) {
; GFX12-LABEL: load_divergent_P0_i8_any_extending:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    flat_load_u8 v0, v[0:1]
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    flat_store_b8 v[2:3], v0
; GFX12-NEXT:    s_endpgm
  %a = load i8, ptr addrspace(0) %ptra
  store i8 %a, ptr addrspace(0) %out
  ret void
}

; with true16, S16 16-bit load
; without true16, S32 16-bit any-extending load
define amdgpu_ps void @load_divergent_P0_i16(ptr addrspace(0) %ptra, ptr addrspace(0) %out) {
; GFX12-True16-LABEL: load_divergent_P0_i16:
; GFX12-True16:       ; %bb.0:
; GFX12-True16-NEXT:    flat_load_u16 v0, v[0:1]
; GFX12-True16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-True16-NEXT:    flat_store_b16 v[2:3], v0
; GFX12-True16-NEXT:    s_endpgm
;
; GFX12-NoTrue16-LABEL: load_divergent_P0_i16:
; GFX12-NoTrue16:       ; %bb.0:
; GFX12-NoTrue16-NEXT:    flat_load_d16_b16 v0, v[0:1]
; GFX12-NoTrue16-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NoTrue16-NEXT:    flat_store_b16 v[2:3], v0
; GFX12-NoTrue16-NEXT:    s_endpgm
  %a = load i16, ptr addrspace(0) %ptra
  store i16 %a, ptr addrspace(0) %out
  ret void
}

define amdgpu_ps void @load_divergent_P0_i32(ptr addrspace(0) %ptra, ptr addrspace(0) %out) {
; GFX12-LABEL: load_divergent_P0_i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    flat_load_b32 v0, v[0:1]
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    flat_store_b32 v[2:3], v0
; GFX12-NEXT:    s_endpgm
  %a = load i32, ptr addrspace(0) %ptra
  store i32 %a, ptr addrspace(0) %out
  ret void
}

define amdgpu_ps void @load_divergent_P0_v2i32(ptr addrspace(0) %ptra, ptr addrspace(0) %out) {
; GFX12-LABEL: load_divergent_P0_v2i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    flat_load_b64 v[0:1], v[0:1]
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    flat_store_b64 v[2:3], v[0:1]
; GFX12-NEXT:    s_endpgm
  %a = load <2 x i32>, ptr addrspace(0) %ptra
  store <2 x i32> %a, ptr addrspace(0) %out
  ret void
}

define amdgpu_ps void @load_divergent_P0_v3i32(ptr addrspace(0) %ptra, ptr addrspace(0) %out) {
; GFX12-LABEL: load_divergent_P0_v3i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    flat_load_b96 v[4:6], v[0:1]
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    flat_store_b96 v[2:3], v[4:6]
; GFX12-NEXT:    s_endpgm
  %a = load <3 x i32>, ptr addrspace(0) %ptra
  store <3 x i32> %a, ptr addrspace(0) %out
  ret void
}

define amdgpu_ps void @load_divergent_P0_v4i32(ptr addrspace(0) %ptra, ptr addrspace(0) %out) {
; GFX12-LABEL: load_divergent_P0_v4i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    flat_load_b128 v[4:7], v[0:1]
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    flat_store_b128 v[2:3], v[4:7]
; GFX12-NEXT:    s_endpgm
  %a = load <4 x i32>, ptr addrspace(0) %ptra
  store <4 x i32> %a, ptr addrspace(0) %out
  ret void
}



define amdgpu_ps void @load_divergent_P1_i8_any_extending(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P1_i8_any_extending:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_u8 v0, v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b8 v[2:3], v0, off
; GFX12-NEXT:    s_endpgm
  %a = load i8, ptr addrspace(1) %ptra
  store i8 %a, ptr addrspace(1) %out
  ret void
}

; with true16, S16 16-bit load
; without true16, S32 16-bit any-extending load
define amdgpu_ps void @load_divergent_P1_i16(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-True16-LABEL: load_divergent_P1_i16:
; GFX12-True16:       ; %bb.0:
; GFX12-True16-NEXT:    global_load_u16 v0, v[0:1], off
; GFX12-True16-NEXT:    s_wait_loadcnt 0x0
; GFX12-True16-NEXT:    global_store_b16 v[2:3], v0, off
; GFX12-True16-NEXT:    s_endpgm
;
; GFX12-NoTrue16-LABEL: load_divergent_P1_i16:
; GFX12-NoTrue16:       ; %bb.0:
; GFX12-NoTrue16-NEXT:    global_load_d16_b16 v0, v[0:1], off
; GFX12-NoTrue16-NEXT:    s_wait_loadcnt 0x0
; GFX12-NoTrue16-NEXT:    global_store_b16 v[2:3], v0, off
; GFX12-NoTrue16-NEXT:    s_endpgm
  %a = load i16, ptr addrspace(1) %ptra
  store i16 %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P1_i32(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P1_i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b32 v0, v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b32 v[2:3], v0, off
; GFX12-NEXT:    s_endpgm
  %a = load i32, ptr addrspace(1) %ptra
  store i32 %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P1_v2i32(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P1_v2i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b64 v[0:1], v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b64 v[2:3], v[0:1], off
; GFX12-NEXT:    s_endpgm
  %a = load <2 x i32>, ptr addrspace(1) %ptra
  store <2 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P1_v3i32(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P1_v3i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b96 v[4:6], v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b96 v[2:3], v[4:6], off
; GFX12-NEXT:    s_endpgm
  %a = load <3 x i32>, ptr addrspace(1) %ptra
  store <3 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P1_v4i32(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P1_v4i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b128 v[4:7], v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b128 v[2:3], v[4:7], off
; GFX12-NEXT:    s_endpgm
  %a = load <4 x i32>, ptr addrspace(1) %ptra
  store <4 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P1_v8i32(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P1_v8i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_clause 0x1
; GFX12-NEXT:    global_load_b128 v[4:7], v[0:1], off
; GFX12-NEXT:    global_load_b128 v[8:11], v[0:1], off offset:16
; GFX12-NEXT:    s_wait_loadcnt 0x1
; GFX12-NEXT:    global_store_b128 v[2:3], v[4:7], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b128 v[2:3], v[8:11], off offset:16
; GFX12-NEXT:    s_endpgm
  %a = load <8 x i32>, ptr addrspace(1) %ptra
  store <8 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P1_v16i32(ptr addrspace(1) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P1_v16i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_clause 0x3
; GFX12-NEXT:    global_load_b128 v[4:7], v[0:1], off
; GFX12-NEXT:    global_load_b128 v[8:11], v[0:1], off offset:16
; GFX12-NEXT:    global_load_b128 v[12:15], v[0:1], off offset:32
; GFX12-NEXT:    global_load_b128 v[16:19], v[0:1], off offset:48
; GFX12-NEXT:    s_wait_loadcnt 0x3
; GFX12-NEXT:    global_store_b128 v[2:3], v[4:7], off
; GFX12-NEXT:    s_wait_loadcnt 0x2
; GFX12-NEXT:    global_store_b128 v[2:3], v[8:11], off offset:16
; GFX12-NEXT:    s_wait_loadcnt 0x1
; GFX12-NEXT:    global_store_b128 v[2:3], v[12:15], off offset:32
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b128 v[2:3], v[16:19], off offset:48
; GFX12-NEXT:    s_endpgm
  %a = load <16 x i32>, ptr addrspace(1) %ptra
  store <16 x i32> %a, ptr addrspace(1) %out
  ret void
}



define amdgpu_ps void @load_divergent_P3_i8_any_extending(ptr addrspace(3) %ptra, ptr addrspace(3) %out) {
; GFX12-LABEL: load_divergent_P3_i8_any_extending:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    ds_load_u8 v0, v0
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    ds_store_b8 v1, v0
; GFX12-NEXT:    s_endpgm
  %a = load i8, ptr addrspace(3) %ptra
  store i8 %a, ptr addrspace(3) %out
  ret void
}

; with true16, S16 16-bit load
; without true16, S32 16-bit any-extending load
define amdgpu_ps void @load_divergent_P3_i16(ptr addrspace(3) %ptra, ptr addrspace(3) %out) {
; GFX12-True16-LABEL: load_divergent_P3_i16:
; GFX12-True16:       ; %bb.0:
; GFX12-True16-NEXT:    ds_load_u16 v0, v0
; GFX12-True16-NEXT:    s_wait_dscnt 0x0
; GFX12-True16-NEXT:    ds_store_b16 v1, v0
; GFX12-True16-NEXT:    s_endpgm
;
; GFX12-NoTrue16-LABEL: load_divergent_P3_i16:
; GFX12-NoTrue16:       ; %bb.0:
; GFX12-NoTrue16-NEXT:    ds_load_u16_d16 v0, v0
; GFX12-NoTrue16-NEXT:    s_wait_dscnt 0x0
; GFX12-NoTrue16-NEXT:    ds_store_b16 v1, v0
; GFX12-NoTrue16-NEXT:    s_endpgm
  %a = load i16, ptr addrspace(3) %ptra
  store i16 %a, ptr addrspace(3) %out
  ret void
}

define amdgpu_ps void @load_divergent_P3_i32(ptr addrspace(3) %ptra, ptr addrspace(3) %out) {
; GFX12-LABEL: load_divergent_P3_i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    ds_load_b32 v0, v0
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    ds_store_b32 v1, v0
; GFX12-NEXT:    s_endpgm
  %a = load i32, ptr addrspace(3) %ptra
  store i32 %a, ptr addrspace(3) %out
  ret void
}

define amdgpu_ps void @load_divergent_P3_v2i32(ptr addrspace(3) %ptra, ptr addrspace(3) %out) {
; GFX12-LABEL: load_divergent_P3_v2i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    ds_load_b64 v[2:3], v0
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    ds_store_b64 v1, v[2:3]
; GFX12-NEXT:    s_endpgm
  %a = load <2 x i32>, ptr addrspace(3) %ptra
  store <2 x i32> %a, ptr addrspace(3) %out
  ret void
}

define amdgpu_ps void @load_divergent_P3_v3i32(ptr addrspace(3) %ptra, ptr addrspace(3) %out) {
; GFX12-LABEL: load_divergent_P3_v3i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    ds_load_b96 v[2:4], v0
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    ds_store_b96 v1, v[2:4]
; GFX12-NEXT:    s_endpgm
  %a = load <3 x i32>, ptr addrspace(3) %ptra
  store <3 x i32> %a, ptr addrspace(3) %out
  ret void
}

define amdgpu_ps void @load_divergent_P3_v4i32(ptr addrspace(3) %ptra, ptr addrspace(3) %out) {
; GFX12-LABEL: load_divergent_P3_v4i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    ds_load_b128 v[2:5], v0
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    ds_store_b128 v1, v[2:5]
; GFX12-NEXT:    s_endpgm
  %a = load <4 x i32>, ptr addrspace(3) %ptra
  store <4 x i32> %a, ptr addrspace(3) %out
  ret void
}



define amdgpu_ps void @load_divergent_P4_i8_any_extending(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P4_i8_any_extending:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_u8 v0, v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b8 v[2:3], v0, off
; GFX12-NEXT:    s_endpgm
  %a = load i8, ptr addrspace(4) %ptra
  store i8 %a, ptr addrspace(1) %out
  ret void
}

; with true16, S16 16-bit load
; without true16, S32 16-bit any-extending load
define amdgpu_ps void @load_divergent_P4_i16(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-True16-LABEL: load_divergent_P4_i16:
; GFX12-True16:       ; %bb.0:
; GFX12-True16-NEXT:    global_load_u16 v0, v[0:1], off
; GFX12-True16-NEXT:    s_wait_loadcnt 0x0
; GFX12-True16-NEXT:    global_store_b16 v[2:3], v0, off
; GFX12-True16-NEXT:    s_endpgm
;
; GFX12-NoTrue16-LABEL: load_divergent_P4_i16:
; GFX12-NoTrue16:       ; %bb.0:
; GFX12-NoTrue16-NEXT:    global_load_d16_b16 v0, v[0:1], off
; GFX12-NoTrue16-NEXT:    s_wait_loadcnt 0x0
; GFX12-NoTrue16-NEXT:    global_store_b16 v[2:3], v0, off
; GFX12-NoTrue16-NEXT:    s_endpgm
  %a = load i16, ptr addrspace(4) %ptra
  store i16 %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P4_i32(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P4_i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b32 v0, v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b32 v[2:3], v0, off
; GFX12-NEXT:    s_endpgm
  %a = load i32, ptr addrspace(4) %ptra
  store i32 %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P4_v2i32(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P4_v2i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b64 v[0:1], v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b64 v[2:3], v[0:1], off
; GFX12-NEXT:    s_endpgm
  %a = load <2 x i32>, ptr addrspace(4) %ptra
  store <2 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P4_v3i32(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P4_v3i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b96 v[4:6], v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b96 v[2:3], v[4:6], off
; GFX12-NEXT:    s_endpgm
  %a = load <3 x i32>, ptr addrspace(4) %ptra
  store <3 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P4_v4i32(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P4_v4i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_load_b128 v[4:7], v[0:1], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b128 v[2:3], v[4:7], off
; GFX12-NEXT:    s_endpgm
  %a = load <4 x i32>, ptr addrspace(4) %ptra
  store <4 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P4_v8i32(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P4_v8i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_clause 0x1
; GFX12-NEXT:    global_load_b128 v[4:7], v[0:1], off
; GFX12-NEXT:    global_load_b128 v[8:11], v[0:1], off offset:16
; GFX12-NEXT:    s_wait_loadcnt 0x1
; GFX12-NEXT:    global_store_b128 v[2:3], v[4:7], off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b128 v[2:3], v[8:11], off offset:16
; GFX12-NEXT:    s_endpgm
  %a = load <8 x i32>, ptr addrspace(4) %ptra
  store <8 x i32> %a, ptr addrspace(1) %out
  ret void
}

define amdgpu_ps void @load_divergent_P4_v16i32(ptr addrspace(4) %ptra, ptr addrspace(1) %out) {
; GFX12-LABEL: load_divergent_P4_v16i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_clause 0x3
; GFX12-NEXT:    global_load_b128 v[4:7], v[0:1], off
; GFX12-NEXT:    global_load_b128 v[8:11], v[0:1], off offset:16
; GFX12-NEXT:    global_load_b128 v[12:15], v[0:1], off offset:32
; GFX12-NEXT:    global_load_b128 v[16:19], v[0:1], off offset:48
; GFX12-NEXT:    s_wait_loadcnt 0x3
; GFX12-NEXT:    global_store_b128 v[2:3], v[4:7], off
; GFX12-NEXT:    s_wait_loadcnt 0x2
; GFX12-NEXT:    global_store_b128 v[2:3], v[8:11], off offset:16
; GFX12-NEXT:    s_wait_loadcnt 0x1
; GFX12-NEXT:    global_store_b128 v[2:3], v[12:15], off offset:32
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_store_b128 v[2:3], v[16:19], off offset:48
; GFX12-NEXT:    s_endpgm
  %a = load <16 x i32>, ptr addrspace(4) %ptra
  store <16 x i32> %a, ptr addrspace(1) %out
  ret void
}



define amdgpu_ps void @load_divergent_P5_i8_any_extending(ptr addrspace(5) %ptra, ptr addrspace(5) %out) {
; GFX12-LABEL: load_divergent_P5_i8_any_extending:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    scratch_load_u8 v0, v0, off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    scratch_store_b8 v1, v0, off
; GFX12-NEXT:    s_endpgm
  %a = load i8, ptr addrspace(5) %ptra
  store i8 %a, ptr addrspace(5) %out
  ret void
}

; with true16, S16 16-bit load
; without true16, S32 16-bit any-extending load
define amdgpu_ps void @load_divergent_P5_i16(ptr addrspace(5) %ptra, ptr addrspace(5) %out) {
; GFX12-True16-LABEL: load_divergent_P5_i16:
; GFX12-True16:       ; %bb.0:
; GFX12-True16-NEXT:    scratch_load_u16 v0, v0, off
; GFX12-True16-NEXT:    s_wait_loadcnt 0x0
; GFX12-True16-NEXT:    scratch_store_b16 v1, v0, off
; GFX12-True16-NEXT:    s_endpgm
;
; GFX12-NoTrue16-LABEL: load_divergent_P5_i16:
; GFX12-NoTrue16:       ; %bb.0:
; GFX12-NoTrue16-NEXT:    scratch_load_d16_b16 v0, v0, off
; GFX12-NoTrue16-NEXT:    s_wait_loadcnt 0x0
; GFX12-NoTrue16-NEXT:    scratch_store_b16 v1, v0, off
; GFX12-NoTrue16-NEXT:    s_endpgm
  %a = load i16, ptr addrspace(5) %ptra
  store i16 %a, ptr addrspace(5) %out
  ret void
}

define amdgpu_ps void @load_divergent_P5_i32(ptr addrspace(5) %ptra, ptr addrspace(5) %out) {
; GFX12-LABEL: load_divergent_P5_i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    scratch_load_b32 v0, v0, off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    scratch_store_b32 v1, v0, off
; GFX12-NEXT:    s_endpgm
  %a = load i32, ptr addrspace(5) %ptra
  store i32 %a, ptr addrspace(5) %out
  ret void
}

define amdgpu_ps void @load_divergent_P5_v2i32(ptr addrspace(5) %ptra, ptr addrspace(5) %out) {
; GFX12-LABEL: load_divergent_P5_v2i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    scratch_load_b64 v[2:3], v0, off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    scratch_store_b64 v1, v[2:3], off
; GFX12-NEXT:    s_endpgm
  %a = load <2 x i32>, ptr addrspace(5) %ptra
  store <2 x i32> %a, ptr addrspace(5) %out
  ret void
}

define amdgpu_ps void @load_divergent_P5_v3i32(ptr addrspace(5) %ptra, ptr addrspace(5) %out) {
; GFX12-LABEL: load_divergent_P5_v3i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    scratch_load_b96 v[2:4], v0, off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    scratch_store_b96 v1, v[2:4], off
; GFX12-NEXT:    s_endpgm
  %a = load <3 x i32>, ptr addrspace(5) %ptra
  store <3 x i32> %a, ptr addrspace(5) %out
  ret void
}

define amdgpu_ps void @load_divergent_P5_v4i32(ptr addrspace(5) %ptra, ptr addrspace(5) %out) {
; GFX12-LABEL: load_divergent_P5_v4i32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    scratch_load_b128 v[2:5], v0, off
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    scratch_store_b128 v1, v[2:5], off
; GFX12-NEXT:    s_endpgm
  %a = load <4 x i32>, ptr addrspace(5) %ptra
  store <4 x i32> %a, ptr addrspace(5) %out
  ret void
}