llvm/test/CodeGen/AMDGPU/call-waw-waitcnt.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass=si-insert-waitcnts %s -o - | FileCheck -check-prefix=GCN %s

# $sgpr30_sgpr31 will hold the return address. We need a waitcnt before SI_CALL so
# that the return address is not clobbered in the callee by the outstanding load.

--- |
  define amdgpu_kernel void @call_waw_waitcnt() #0 {
    %1 = call i32 @func()
    ret void
  }

  declare hidden i32 @func() #0

  attributes #0 = { nounwind }
...

---
name:            call_waw_waitcnt
tracksRegLiveness: true
body:             |
  bb.0:
    liveins: $sgpr4_sgpr5, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3

    ; GCN-LABEL: name: call_waw_waitcnt
    ; GCN: liveins: $sgpr4_sgpr5, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3
    ; GCN-NEXT: {{  $}}
    ; GCN-NEXT: S_WAITCNT 0
    ; GCN-NEXT: $sgpr30_sgpr31 = S_LOAD_DWORDX2_IMM $sgpr4_sgpr5, 0, 0
    ; GCN-NEXT: $sgpr33 = S_MOV_B32 killed $sgpr7
    ; GCN-NEXT: $flat_scr_lo = S_ADD_U32 killed $sgpr4, $sgpr33, implicit-def $scc
    ; GCN-NEXT: $flat_scr_hi = S_ADDC_U32 killed $sgpr5, 0, implicit-def $scc, implicit killed $scc
    ; GCN-NEXT: BUNDLE implicit-def $sgpr4_sgpr5, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $scc {
    ; GCN-NEXT:   $sgpr4_sgpr5 = S_GETPC_B64
    ; GCN-NEXT:   $sgpr4 = S_ADD_U32 internal $sgpr4, target-flags(amdgpu-rel32-lo) @func + 4, implicit-def $scc
    ; GCN-NEXT:   $sgpr5 = S_ADDC_U32 internal $sgpr5, target-flags(amdgpu-rel32-hi) @func + 4, implicit-def $scc, implicit internal $scc
    ; GCN-NEXT: }
    ; GCN-NEXT: $sgpr32 = S_MOV_B32 $sgpr33
    ; GCN-NEXT: S_WAITCNT 49279
    ; GCN-NEXT: dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr4_sgpr5, @func, csr_amdgpu, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit-def dead $vgpr0
    ; GCN-NEXT: S_ENDPGM 0
    $sgpr30_sgpr31 = S_LOAD_DWORDX2_IMM $sgpr4_sgpr5, 0, 0
    $sgpr33 = S_MOV_B32 killed $sgpr7
    $flat_scr_lo = S_ADD_U32 killed $sgpr4, $sgpr33, implicit-def $scc
    $flat_scr_hi = S_ADDC_U32 killed $sgpr5, 0, implicit-def $scc, implicit killed $scc
    BUNDLE implicit-def $sgpr4_sgpr5, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $scc {
      $sgpr4_sgpr5 = S_GETPC_B64
      $sgpr4 = S_ADD_U32 internal $sgpr4, target-flags(amdgpu-rel32-lo) @func + 4, implicit-def $scc
      $sgpr5 = S_ADDC_U32 internal $sgpr5, target-flags(amdgpu-rel32-hi) @func + 4, implicit-def $scc, implicit internal $scc
    }
    $sgpr32 = S_MOV_B32 $sgpr33
    dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr4_sgpr5, @func, csr_amdgpu, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit-def dead $vgpr0
    S_ENDPGM 0
...