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

   1 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
   2 # RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass=si-insert-waitcnts %s -o - | FileCheck -check-prefix=GCN %s
   3
   4 # $sgpr30_sgpr31 will hold the return address. We need a waitcnt before SI_CALL so
   5 # that the return address is not clobbered in the callee by the outstanding load.
   6
   7 --- |
   8   define amdgpu_kernel void @call_waw_waitcnt() #0 {
   9     %1 = call i32 @func()
  10     ret void
  11   }
  12
  13   declare hidden i32 @func() #0
  14
  15   attributes #0 = { nounwind }
  16 ...
  17
  18 ---
  19 name:            call_waw_waitcnt
  20 tracksRegLiveness: true
  21 body:             |
  22   bb.0:
  23     liveins: $sgpr4_sgpr5, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3
  24
  25     ; GCN-LABEL: name: call_waw_waitcnt
  26     ; GCN: liveins: $sgpr4_sgpr5, $sgpr7, $sgpr0_sgpr1_sgpr2_sgpr3
  27     ; GCN: S_WAITCNT 0
  28     ; GCN: $sgpr30_sgpr31 = S_LOAD_DWORDX2_IMM $sgpr4_sgpr5, 0, 0
  29     ; GCN: $sgpr33 = S_MOV_B32 killed $sgpr7
  30     ; GCN: $flat_scr_lo = S_ADD_U32 killed $sgpr4, $sgpr33, implicit-def $scc
  31     ; GCN: $flat_scr_hi = S_ADDC_U32 killed $sgpr5, 0, implicit-def $scc, implicit killed $scc
  32     ; GCN: BUNDLE implicit-def $sgpr4_sgpr5, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $scc {
  33     ; GCN:   $sgpr4_sgpr5 = S_GETPC_B64
  34     ; GCN:   $sgpr4 = S_ADD_U32 internal $sgpr4, target-flags(amdgpu-rel32-lo) @func + 4, implicit-def $scc
  35     ; GCN:   $sgpr5 = S_ADDC_U32 internal $sgpr5, target-flags(amdgpu-rel32-hi) @func + 4, implicit-def $scc, implicit internal $scc
  36     ; GCN: }
  37     ; GCN: $sgpr32 = S_MOV_B32 $sgpr33
  38     ; GCN: S_WAITCNT 49279
  39     ; GCN: dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr4_sgpr5, @func, csr_amdgpu_highregs, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit-def dead $vgpr0
  40     ; GCN: S_ENDPGM 0
  41     $sgpr30_sgpr31 = S_LOAD_DWORDX2_IMM $sgpr4_sgpr5, 0, 0
  42     $sgpr33 = S_MOV_B32 killed $sgpr7
  43     $flat_scr_lo = S_ADD_U32 killed $sgpr4, $sgpr33, implicit-def $scc
  44     $flat_scr_hi = S_ADDC_U32 killed $sgpr5, 0, implicit-def $scc, implicit killed $scc
  45     BUNDLE implicit-def $sgpr4_sgpr5, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $scc {
  46       $sgpr4_sgpr5 = S_GETPC_B64
  47       $sgpr4 = S_ADD_U32 internal $sgpr4, target-flags(amdgpu-rel32-lo) @func + 4, implicit-def $scc
  48       $sgpr5 = S_ADDC_U32 internal $sgpr5, target-flags(amdgpu-rel32-hi) @func + 4, implicit-def $scc, implicit internal $scc
  49     }
  50     $sgpr32 = S_MOV_B32 $sgpr33
  51     dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr4_sgpr5, @func, csr_amdgpu_highregs, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit-def dead $vgpr0
  52     S_ENDPGM 0
  53 ...