llvm/test/CodeGen/AMDGPU/schedule-pressure-bundle-subreg-def.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6
# RUN: llc -mtriple=amdgpu11.00 -run-pass=machine-scheduler %s -o - | FileCheck %s

# Each bundle defines one lane of %0. PressureDiffs do not track lanes, so the
# scheduler must not use them for these bundles. The BUNDLE header only has
# implicit operands, so the instructions inside the bundle have to be checked.
# Previously the cached PressureDiff counted each partial def as a def of the
# whole register and tripped the EXPENSIVE_CHECKS pressure cross-check in
# GCNSchedStrategy.

---
name: bundle_subreg_defs
tracksRegLiveness: true
body: |
  bb.0:
    liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $sgpr0_sgpr1
    ; CHECK-LABEL: name: bundle_subreg_defs
    ; CHECK: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $sgpr0_sgpr1
    ; CHECK-NEXT: {{  $}}
    ; CHECK-NEXT: BUNDLE implicit-def undef %0.sub3, implicit $vgpr3, implicit $exec {
    ; CHECK-NEXT:   undef [[V_MOV_B32_e32_:%[0-9]+]].sub3:vreg_128 = V_MOV_B32_e32 $vgpr3, implicit $exec
    ; CHECK-NEXT: }
    ; CHECK-NEXT: BUNDLE implicit-def [[V_MOV_B32_e32_]].sub2, implicit $vgpr2, implicit $exec {
    ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]].sub2:vreg_128 = V_MOV_B32_e32 $vgpr2, implicit $exec
    ; CHECK-NEXT: }
    ; CHECK-NEXT: BUNDLE implicit-def [[V_MOV_B32_e32_]].sub1, implicit $vgpr1, implicit $exec {
    ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]].sub1:vreg_128 = V_MOV_B32_e32 $vgpr1, implicit $exec
    ; CHECK-NEXT: }
    ; CHECK-NEXT: BUNDLE implicit-def [[V_MOV_B32_e32_]].sub0, implicit $vgpr0, implicit $exec {
    ; CHECK-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]].sub0:vreg_128 = V_MOV_B32_e32 $vgpr0, implicit $exec
    ; CHECK-NEXT: }
    ; CHECK-NEXT: GLOBAL_STORE_DWORDX4_SADDR undef %1:vgpr_32, [[V_MOV_B32_e32_]], $sgpr0_sgpr1, 0, 0, implicit $exec
    ; CHECK-NEXT: S_ENDPGM 0
    BUNDLE implicit-def undef %0.sub3, implicit $vgpr3, implicit $exec {
      undef %0.sub3:vreg_128 = V_MOV_B32_e32 $vgpr3, implicit $exec
    }
    BUNDLE implicit-def %0.sub2, implicit $vgpr2, implicit $exec {
      %0.sub2:vreg_128 = V_MOV_B32_e32 $vgpr2, implicit $exec
    }
    BUNDLE implicit-def %0.sub1, implicit $vgpr1, implicit $exec {
      %0.sub1:vreg_128 = V_MOV_B32_e32 $vgpr1, implicit $exec
    }
    BUNDLE implicit-def %0.sub0, implicit $vgpr0, implicit $exec {
      %0.sub0:vreg_128 = V_MOV_B32_e32 $vgpr0, implicit $exec
    }
    GLOBAL_STORE_DWORDX4_SADDR undef %1:vgpr_32, %0, $sgpr0_sgpr1, 0, 0, implicit $exec
    S_ENDPGM 0
...