llvm/test/CodeGen/AMDGPU/waitcnt-vmem-waw.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 -o - %s | FileCheck -check-prefix=GFX9 %s
   3
   4 # Two buffer loads with overlapping outputs. No waitcnt required.
   5 ---
   6 name: buffer_buffer
   7 tracksRegLiveness: true
   8 body: |
   9   bb.0:
  10     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $sgpr5
  11     ; GFX9-LABEL: name: buffer_buffer
  12     ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $sgpr5
  13     ; GFX9-NEXT: {{  $}}
  14     ; GFX9-NEXT: S_WAITCNT 0
  15     ; GFX9-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = BUFFER_LOAD_DWORDX4_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  16     ; GFX9-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr5, 0, 0, 0, implicit $exec
  17     $vgpr0_vgpr1_vgpr2_vgpr3 = BUFFER_LOAD_DWORDX4_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  18     $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr5, 0, 0, 0, implicit $exec
  19 ...
  20
  21 # Two tbuffer loads with overlapping outputs. No waitcnt required.
  22 ---
  23 name: tbuffer_tbuffer
  24 tracksRegLiveness: true
  25 body: |
  26   bb.0:
  27     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $sgpr5
  28     ; GFX9-LABEL: name: tbuffer_tbuffer
  29     ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $sgpr5
  30     ; GFX9-NEXT: {{  $}}
  31     ; GFX9-NEXT: S_WAITCNT 0
  32     ; GFX9-NEXT: $vgpr0_vgpr1_vgpr2 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, 0, 4, 125, 0, 0, implicit $exec
  33     ; GFX9-NEXT: $vgpr0 = TBUFFER_LOAD_FORMAT_X_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, 0, 16, 116, 0, 0, implicit $exec
  34     $vgpr0_vgpr1_vgpr2 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, 0, 4, 125, 0, 0, implicit $exec
  35     $vgpr0 = TBUFFER_LOAD_FORMAT_X_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, 0, 16, 116, 0, 0, implicit $exec
  36 ...
  37
  38 # Two gathers with overlapping outputs. (Note gathers can't be trimmed because
  39 # dmask means something different.) No waitcnt required.
  40 ---
  41 name: gather_gather
  42 tracksRegLiveness: true
  43 body: |
  44   bb.0:
  45     liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0_vgpr1_vgpr2, $vgpr3_vgpr4_vgpr5
  46     ; GFX9-LABEL: name: gather_gather
  47     ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0_vgpr1_vgpr2, $vgpr3_vgpr4_vgpr5
  48     ; GFX9-NEXT: {{  $}}
  49     ; GFX9-NEXT: S_WAITCNT 0
  50     ; GFX9-NEXT: $vgpr10_vgpr11_vgpr12_vgpr13 = IMAGE_GATHER4_LZ_O_V4_V3 $vgpr0_vgpr1_vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load (s128))
  51     ; GFX9-NEXT: $vgpr13_vgpr14_vgpr15_vgpr16 = IMAGE_GATHER4_LZ_O_V4_V3 $vgpr3_vgpr4_vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load (s128))
  52     $vgpr10_vgpr11_vgpr12_vgpr13 = IMAGE_GATHER4_LZ_O_V4_V3 $vgpr0_vgpr1_vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load (s128))
  53     $vgpr13_vgpr14_vgpr15_vgpr16 = IMAGE_GATHER4_LZ_O_V4_V3 $vgpr3_vgpr4_vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load (s128))
  54 ...
  55
  56 # Image load vs image sample. Waitcnt required because they are not guaranteed
  57 # to write their results in order, despite both using the s_waitcnt vmcnt
  58 # counter.
  59 ---
  60 name: nosampler_sampler
  61 tracksRegLiveness: true
  62 body: |
  63   bb.0:
  64     liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0_vgpr1_vgpr2_vgpr3
  65     ; GFX9-LABEL: name: nosampler_sampler
  66     ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0_vgpr1_vgpr2_vgpr3
  67     ; GFX9-NEXT: {{  $}}
  68     ; GFX9-NEXT: S_WAITCNT 0
  69     ; GFX9-NEXT: $vgpr4 = IMAGE_LOAD_V1_V4 $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, 2, -1, 0, 0, 0, 0, 0, 0, implicit $exec :: (load (s128))
  70     ; GFX9-NEXT: S_WAITCNT 3952
  71     ; GFX9-NEXT: $vgpr4 = IMAGE_SAMPLE_L_V1_V4 $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 8, 0, 0, 0, 0, 0, -1, 0, implicit $exec :: (load (s128))
  72     $vgpr4 = IMAGE_LOAD_V1_V4 $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, 2, -1, 0, 0, 0, 0, 0, 0, implicit $exec :: (load (s128))
  73     $vgpr4 = IMAGE_SAMPLE_L_V1_V4 $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 8, 0, 0, 0, 0, 0, -1, 0, implicit $exec :: (load (s128))
  74 ...
  75 # (global_load + scratch_load + buffer_load)
  76 ---
  77 name: global_scratch_buffer
  78 tracksRegLiveness: true
  79 body: |
  80   bb.0:
  81     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0_vgpr1
  82     ; GFX9-LABEL: name: global_scratch_buffer
  83     ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0_vgpr1
  84     ; GFX9-NEXT: {{  $}}
  85     ; GFX9-NEXT: S_WAITCNT 0
  86     ; GFX9-NEXT: $vgpr2 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
  87     ; GFX9-NEXT: $vgpr2 = SCRATCH_LOAD_DWORD $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
  88     ; GFX9-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  89     $vgpr2 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
  90     $vgpr2 = SCRATCH_LOAD_DWORD $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
  91     $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
  92 ...
  93 # waw between flat and buffer should have a wait inserted between.
  94 # (flat + buffer)
  95 ---
  96 name: flat_buffer
  97 tracksRegLiveness: true
  98 body: |
  99   bb.0:
 100     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0_vgpr1
 101     ; GFX9-LABEL: name: flat_buffer
 102     ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0_vgpr1
 103     ; GFX9-NEXT: {{  $}}
 104     ; GFX9-NEXT: S_WAITCNT 0
 105     ; GFX9-NEXT: $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
 106     ; GFX9-NEXT: S_WAITCNT 49279
 107     ; GFX9-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 108     $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
 109     $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 110 ...
 111 # buffer + flat
 112 ---
 113 name: buffer_flat
 114 tracksRegLiveness: true
 115 body: |
 116   bb.0:
 117     liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0_vgpr1
 118     ; GFX9-LABEL: name: buffer_flat
 119     ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0_vgpr1
 120     ; GFX9-NEXT: {{  $}}
 121     ; GFX9-NEXT: S_WAITCNT 0
 122     ; GFX9-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 123     ; GFX9-NEXT: S_WAITCNT 3952
 124     ; GFX9-NEXT: $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
 125     $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
 126     $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
 127 ...