llvm/test/CodeGen/AMDGPU/back-off-barrier-subtarget-feature.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9-NO-BACKOFF %s
   3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9-BACKOFF %s
   4 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9-BACKOFF %s
   5 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -mattr=-back-off-barrier -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9-NO-BACKOFF %s
   6 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-BACKOFF %s
   7 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11-BACKOFF %s
   8
   9 ; Subtargets must wait for outstanding memory instructions before a barrier if
  10 ; they cannot back off of the barrier.
  11
  12 define void @back_off_barrier_no_fence(ptr %in, ptr %out) #0 {
  13 ; GFX9-NO-BACKOFF-LABEL: back_off_barrier_no_fence:
  14 ; GFX9-NO-BACKOFF:       ; %bb.0:
  15 ; GFX9-NO-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  16 ; GFX9-NO-BACKOFF-NEXT:    flat_load_dword v0, v[0:1]
  17 ; GFX9-NO-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  18 ; GFX9-NO-BACKOFF-NEXT:    s_barrier
  19 ; GFX9-NO-BACKOFF-NEXT:    flat_store_dword v[2:3], v0
  20 ; GFX9-NO-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  21 ; GFX9-NO-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
  22 ;
  23 ; GFX9-BACKOFF-LABEL: back_off_barrier_no_fence:
  24 ; GFX9-BACKOFF:       ; %bb.0:
  25 ; GFX9-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  26 ; GFX9-BACKOFF-NEXT:    flat_load_dword v0, v[0:1]
  27 ; GFX9-BACKOFF-NEXT:    s_barrier
  28 ; GFX9-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  29 ; GFX9-BACKOFF-NEXT:    flat_store_dword v[2:3], v0
  30 ; GFX9-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  31 ; GFX9-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
  32 ;
  33 ; GFX10-BACKOFF-LABEL: back_off_barrier_no_fence:
  34 ; GFX10-BACKOFF:       ; %bb.0:
  35 ; GFX10-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  36 ; GFX10-BACKOFF-NEXT:    flat_load_dword v0, v[0:1]
  37 ; GFX10-BACKOFF-NEXT:    s_barrier
  38 ; GFX10-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  39 ; GFX10-BACKOFF-NEXT:    flat_store_dword v[2:3], v0
  40 ; GFX10-BACKOFF-NEXT:    s_waitcnt lgkmcnt(0)
  41 ; GFX10-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
  42 ;
  43 ; GFX11-BACKOFF-LABEL: back_off_barrier_no_fence:
  44 ; GFX11-BACKOFF:       ; %bb.0:
  45 ; GFX11-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  46 ; GFX11-BACKOFF-NEXT:    flat_load_b32 v0, v[0:1]
  47 ; GFX11-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  48 ; GFX11-BACKOFF-NEXT:    s_barrier
  49 ; GFX11-BACKOFF-NEXT:    flat_store_b32 v[2:3], v0
  50 ; GFX11-BACKOFF-NEXT:    s_waitcnt lgkmcnt(0)
  51 ; GFX11-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
  52   %load = load i32, ptr %in
  53   call void @llvm.amdgcn.s.barrier()
  54   store i32 %load, ptr %out
  55   ret void
  56 }
  57
  58 define void @back_off_barrier_with_fence(ptr %in, ptr %out) #0 {
  59 ; GFX9-NO-BACKOFF-LABEL: back_off_barrier_with_fence:
  60 ; GFX9-NO-BACKOFF:       ; %bb.0:
  61 ; GFX9-NO-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  62 ; GFX9-NO-BACKOFF-NEXT:    flat_load_dword v0, v[0:1]
  63 ; GFX9-NO-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  64 ; GFX9-NO-BACKOFF-NEXT:    s_barrier
  65 ; GFX9-NO-BACKOFF-NEXT:    s_waitcnt lgkmcnt(0)
  66 ; GFX9-NO-BACKOFF-NEXT:    flat_store_dword v[2:3], v0
  67 ; GFX9-NO-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  68 ; GFX9-NO-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
  69 ;
  70 ; GFX9-BACKOFF-LABEL: back_off_barrier_with_fence:
  71 ; GFX9-BACKOFF:       ; %bb.0:
  72 ; GFX9-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  73 ; GFX9-BACKOFF-NEXT:    flat_load_dword v0, v[0:1]
  74 ; GFX9-BACKOFF-NEXT:    s_waitcnt lgkmcnt(0)
  75 ; GFX9-BACKOFF-NEXT:    s_barrier
  76 ; GFX9-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  77 ; GFX9-BACKOFF-NEXT:    flat_store_dword v[2:3], v0
  78 ; GFX9-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  79 ; GFX9-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
  80 ;
  81 ; GFX10-BACKOFF-LABEL: back_off_barrier_with_fence:
  82 ; GFX10-BACKOFF:       ; %bb.0:
  83 ; GFX10-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  84 ; GFX10-BACKOFF-NEXT:    flat_load_dword v0, v[0:1]
  85 ; GFX10-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  86 ; GFX10-BACKOFF-NEXT:    s_waitcnt_vscnt null, 0x0
  87 ; GFX10-BACKOFF-NEXT:    s_barrier
  88 ; GFX10-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
  89 ; GFX10-BACKOFF-NEXT:    s_waitcnt_vscnt null, 0x0
  90 ; GFX10-BACKOFF-NEXT:    buffer_gl0_inv
  91 ; GFX10-BACKOFF-NEXT:    flat_store_dword v[2:3], v0
  92 ; GFX10-BACKOFF-NEXT:    s_waitcnt lgkmcnt(0)
  93 ; GFX10-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
  94 ;
  95 ; GFX11-BACKOFF-LABEL: back_off_barrier_with_fence:
  96 ; GFX11-BACKOFF:       ; %bb.0:
  97 ; GFX11-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
  98 ; GFX11-BACKOFF-NEXT:    flat_load_b32 v0, v[0:1]
  99 ; GFX11-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 100 ; GFX11-BACKOFF-NEXT:    s_waitcnt_vscnt null, 0x0
 101 ; GFX11-BACKOFF-NEXT:    s_barrier
 102 ; GFX11-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 103 ; GFX11-BACKOFF-NEXT:    s_waitcnt_vscnt null, 0x0
 104 ; GFX11-BACKOFF-NEXT:    buffer_gl0_inv
 105 ; GFX11-BACKOFF-NEXT:    flat_store_b32 v[2:3], v0
 106 ; GFX11-BACKOFF-NEXT:    s_waitcnt lgkmcnt(0)
 107 ; GFX11-BACKOFF-NEXT:    s_setpc_b64 s[30:31]
 108   %load = load i32, ptr %in
 109   fence syncscope("workgroup") release
 110   call void @llvm.amdgcn.s.barrier()
 111   fence syncscope("workgroup") acquire
 112   store i32 %load, ptr %out
 113   ret void
 114 }
 115
 116 declare void @llvm.amdgcn.s.barrier()
 117
 118 attributes #0 = { nounwind }