llvm/test/CodeGen/AMDGPU/s_or_saveexec_xor_combine.mir

   1 # RUN: llc -march=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32 -run-pass=si-optimize-exec-masking -verify-machineinstrs %s -o - | FileCheck --check-prefixes=GCN,WAVE32 %s
   2 # RUN: llc -march=amdgcn -mcpu=gfx1010 -mattr=-wavefrontsize32,+wavefrontsize64 -run-pass=si-optimize-exec-masking -verify-machineinstrs %s -o - | FileCheck --check-prefixes=GCN,WAVE64 %s
   3
   4 ---
   5
   6 # After the Optimize exec masking (post-RA) pass, codegen can end up with the following sequence:
   7 # s_or_saveexec_b32 s0, s0
   8 # s_xor_b32 exec_lo, exec_lo, s0
   9 #
  10 # This can be combined into one instruction:
  11 # s_andn2_saveexec_b32 s0, s0
  12
  13 # Ensure the transformation gets applied in the b32 case.
  14 # GCN-LABEL: name: s_or_saveexec_xor_combine_b32
  15 # WAVE32: S_ANDN2_SAVEEXEC_B32
  16 name: s_or_saveexec_xor_combine_b32
  17 tracksRegLiveness: true
  18 body: |
  19   bb.0:
  20     liveins: $sgpr0
  21     renamable $sgpr0 = S_OR_SAVEEXEC_B32 killed renamable $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec
  22     $exec_lo = S_XOR_B32 $exec_lo, renamable $sgpr0, implicit-def $scc
  23 ...
  24
  25 ---
  26
  27 # Ensure the transformation gets applied in the b64 case.
  28 # GCN-LABEL: name: s_or_saveexec_xor_combine_b64
  29 # WAVE64: S_ANDN2_SAVEEXEC_B64
  30 name: s_or_saveexec_xor_combine_b64
  31 tracksRegLiveness: true
  32 body: |
  33   bb.0:
  34     liveins: $sgpr0_sgpr1
  35     renamable $sgpr0_sgpr1 = S_OR_SAVEEXEC_B64 killed renamable $sgpr0_sgpr1, implicit-def $exec, implicit-def $scc, implicit $exec
  36     $exec = S_XOR_B64 $exec, renamable $sgpr0_sgpr1, implicit-def $scc
  37 ...
  38
  39 ---
  40
  41 # Ensure the transformation does get applied even if the operands are swapped.
  42 # GCN-LABEL: name: s_or_saveexec_xor_combine_b32_swap
  43 # WAVE32: S_ANDN2_SAVEEXEC_B32
  44 name: s_or_saveexec_xor_combine_b32_swap
  45 tracksRegLiveness: true
  46 body: |
  47   bb.0:
  48     liveins: $sgpr0
  49     renamable $sgpr0 = S_OR_SAVEEXEC_B32 killed renamable $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec
  50     $exec_lo = S_XOR_B32 renamable $sgpr0, $exec_lo, implicit-def $scc
  51 ...
  52
  53 ---
  54
  55 # Ensure the transformation does get applied if source and dest operand for s_or_saveeexec are not equal.
  56 # GCN-LABEL: name: s_or_saveexec_xor_combine_b32_inequal_operands
  57 # WAVE32: S_ANDN2_SAVEEXEC
  58 name: s_or_saveexec_xor_combine_b32_inequal_operands
  59 tracksRegLiveness: true
  60 body: |
  61   bb.0:
  62     liveins: $sgpr0, $sgpr1
  63     renamable $sgpr0 = S_OR_SAVEEXEC_B32 killed renamable $sgpr1, implicit-def $exec, implicit-def $scc, implicit $exec
  64     $exec_lo = S_XOR_B32 $exec_lo, renamable $sgpr0, implicit-def $scc
  65 ...
  66
  67 ---
  68
  69 # Ensure the transformation does not get applied if s_xor does not use the dest as input operand.
  70 # GCN-LABEL: name: s_or_saveexec_xor_combine_b32_wrong_input
  71 # WAVE32: S_OR_SAVEEXEC
  72 # WAVE32: S_XOR_B32
  73 name: s_or_saveexec_xor_combine_b32_wrong_input
  74 tracksRegLiveness: true
  75 body: |
  76   bb.0:
  77     liveins: $sgpr0, $sgpr1
  78     renamable $sgpr0 = S_OR_SAVEEXEC_B32 killed renamable $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec
  79     $exec_lo = S_XOR_B32 $exec_lo, renamable $sgpr1, implicit-def $scc
  80 ...
  81
  82 ---
  83
  84
  85 # Ensure the transformation does not get applied if the instructions don't appear sequentially.
  86 # GCN-LABEL: name: s_or_saveexec_xor_combine_b32_non_sequence
  87 # WAVE32: S_OR_SAVEEXEC
  88 # WAVE32: S_MOV_B32
  89 # WAVE32: S_XOR_B32
  90 name: s_or_saveexec_xor_combine_b32_non_sequence
  91 tracksRegLiveness: true
  92 body: |
  93   bb.0:
  94     liveins: $sgpr0, $sgpr1
  95     renamable $sgpr0 = S_OR_SAVEEXEC_B32 killed renamable $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec
  96     renamable $sgpr1 = S_MOV_B32 renamable $sgpr0
  97     $exec_lo = S_XOR_B32 $exec_lo, renamable $sgpr1, implicit-def $scc
  98 ...
  99
 100 ---
 101
 102 # Don't apply the transformation if the basic block only has a single instruction.
 103
 104 # GCN-LABEL: name: s_or_saveexec_xor_combine_b32_last_inst
 105 # WAVE32: S_OR_SAVEEXEC
 106 name: s_or_saveexec_xor_combine_b32_last_inst
 107 tracksRegLiveness: true
 108 body: |
 109   bb.0:
 110     liveins: $sgpr0
 111     renamable $sgpr0 = S_OR_SAVEEXEC_B32 killed renamable $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec
 112 ...
 113
 114 ---
 115
 116 # Don't apply the transformation if the basic block ends with an S_OR_SAVEEXEC_B32 instruction.
 117
 118 # GCN-LABEL: name: s_or_saveexec_xor_combine_b32_or_saveexec_terminator
 119 # WAVE32: S_MOV_B32
 120 # WAVE32: S_OR_SAVEEXEC
 121 name: s_or_saveexec_xor_combine_b32_or_saveexec_terminator
 122 tracksRegLiveness: true
 123 body: |
 124   bb.0:
 125     liveins: $sgpr0, $sgpr1
 126     renamable $sgpr1 = S_MOV_B32 renamable $sgpr0
 127     renamable $sgpr0 = S_OR_SAVEEXEC_B32 killed renamable $sgpr0, implicit-def $exec, implicit-def $scc, implicit $exec