test/CodeGen/AMDGPU/pei-scavenge-sgpr-carry-out.mir

   1 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
   2 # RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -verify-machineinstrs -run-pass=prologepilog %s -o - | FileCheck %s
   3
   4 # Test what happens when an SGPR is unavailable for the unused add
   5 # carry out when materializing the frame index.
   6
   7
   8 # There are truly no free SGPRs, so the entire frame index expansion
   9 # needs to be inverted to restore the original frame register.
  10
  11 ---
  12 name: scavenge_sgpr_pei_no_sgprs
  13 tracksRegLiveness: true
  14
  15 stack:
  16   - { id: 0, type: default, offset: 0, size: 4, alignment: 8192 }
  17   - { id: 1, type: default, offset: 0, size: 4, alignment: 8192 }
  18
  19 machineFunctionInfo:
  20   isEntryFunction: false
  21   scratchRSrcReg:  $sgpr0_sgpr1_sgpr2_sgpr3
  22   scratchWaveOffsetReg: $sgpr34
  23   frameOffsetReg:  $sgpr33
  24   stackPtrOffsetReg:  $sgpr32
  25
  26 body:             |
  27   bb.0:
  28     liveins: $vgpr1
  29
  30     ; CHECK-LABEL: name: scavenge_sgpr_pei_no_sgprs
  31     ; CHECK: liveins: $vgpr1
  32     ; CHECK: $sgpr27 = frame-setup COPY $sgpr33
  33     ; CHECK: $sgpr4 = frame-setup S_ADD_U32 $sgpr32, 524224, implicit-def $scc
  34     ; CHECK: $sgpr33 = frame-setup S_AND_B32 killed $sgpr4, 4294443008, implicit-def $scc
  35     ; CHECK: $sgpr32 = frame-setup S_ADD_U32 $sgpr32, 1572864, implicit-def $scc
  36     ; CHECK: S_NOP 0, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $sgpr6, implicit-def $sgpr7, implicit-def $sgpr8, implicit-def $sgpr9, implicit-def $sgpr10, implicit-def $sgpr11, implicit-def $sgpr12, implicit-def $sgpr13, implicit-def $sgpr14, implicit-def $sgpr15, implicit-def $sgpr16, implicit-def $sgpr17, implicit-def $sgpr18, implicit-def $sgpr19, implicit-def $sgpr20, implicit-def $sgpr21, implicit-def $sgpr22, implicit-def $sgpr23, implicit-def $sgpr24, implicit-def $sgpr25, implicit-def $sgpr26, implicit-def $sgpr17, implicit-def $sgpr28, implicit-def $sgpr29, implicit-def $sgpr30, implicit-def $sgpr31, implicit-def $vcc
  37     ; CHECK: $sgpr33 = S_SUB_U32 $sgpr33, $sgpr34, implicit-def $scc
  38     ; CHECK: $sgpr33 = S_LSHR_B32 killed $sgpr33, 6, implicit-def $scc
  39     ; CHECK: $sgpr33 = S_ADD_U32 killed $sgpr33, 8192, implicit-def $scc
  40     ; CHECK: $vgpr2 = COPY killed $sgpr33
  41     ; CHECK: $sgpr33 = S_SUB_U32 killed $sgpr33, 8192, implicit-def $scc
  42     ; CHECK: $sgpr33 = S_LSHL_B32 killed $sgpr33, 6, implicit-def $scc
  43     ; CHECK: $sgpr33 = S_ADD_U32 $sgpr33, $sgpr34, implicit-def $scc
  44     ; CHECK: $vgpr0 = V_OR_B32_e32 killed $vgpr2, $vgpr1, implicit $exec, implicit $sgpr4, implicit $sgpr5, implicit $sgpr6, implicit $sgpr7, implicit $sgpr8, implicit $sgpr9, implicit $sgpr10, implicit $sgpr11, implicit $sgpr12, implicit $sgpr13, implicit $sgpr14, implicit $sgpr15, implicit $sgpr16, implicit $sgpr17, implicit $sgpr18, implicit $sgpr19, implicit $sgpr20, implicit $sgpr21, implicit $sgpr22, implicit $sgpr23, implicit $sgpr24, implicit $sgpr25, implicit $sgpr26, implicit $sgpr17, implicit $sgpr28, implicit $sgpr29, implicit $sgpr30, implicit $sgpr31
  45     ; CHECK: $sgpr32 = frame-destroy S_SUB_U32 $sgpr32, 1572864, implicit-def $scc
  46     ; CHECK: $sgpr33 = frame-setup COPY $sgpr27
  47     ; CHECK: S_ENDPGM 0, implicit $vcc
  48     S_NOP 0, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $sgpr6, implicit-def $sgpr7, implicit-def $sgpr8, implicit-def $sgpr9, implicit-def $sgpr10, implicit-def $sgpr11, implicit-def $sgpr12, implicit-def $sgpr13, implicit-def $sgpr14, implicit-def $sgpr15, implicit-def $sgpr16, implicit-def $sgpr17, implicit-def $sgpr18, implicit-def $sgpr19, implicit-def $sgpr20, implicit-def $sgpr21, implicit-def $sgpr22, implicit-def $sgpr23, implicit-def $sgpr24, implicit-def $sgpr25, implicit-def $sgpr26, implicit-def $sgpr17, implicit-def $sgpr28, implicit-def $sgpr29, implicit-def $sgpr30, implicit-def $sgpr31, implicit-def $vcc
  49     $vgpr0 = V_OR_B32_e32 %stack.1, $vgpr1, implicit $exec, implicit $sgpr4, implicit $sgpr5, implicit $sgpr6, implicit $sgpr7, implicit $sgpr8, implicit $sgpr9, implicit $sgpr10, implicit $sgpr11, implicit $sgpr12, implicit $sgpr13, implicit $sgpr14, implicit $sgpr15, implicit $sgpr16, implicit $sgpr17, implicit $sgpr18, implicit $sgpr19, implicit $sgpr20, implicit $sgpr21, implicit $sgpr22, implicit $sgpr23, implicit $sgpr24, implicit $sgpr25, implicit $sgpr26, implicit $sgpr17, implicit $sgpr28, implicit $sgpr29, implicit $sgpr30, implicit $sgpr31
  50     S_ENDPGM 0, implicit $vcc
  51 ...
  52
  53 # One 32-bit SGPR is available for the intermediate scale computation,
  54 # so only an extra copy to VALU is necessary.
  55
  56 ---
  57 name: scavenge_sgpr_pei_one_sgpr
  58 tracksRegLiveness: true
  59
  60 stack:
  61   - { id: 0, type: default, offset: 0, size: 4, alignment: 8192 }
  62   - { id: 1, type: default, offset: 0, size: 4, alignment: 8192 }
  63
  64 machineFunctionInfo:
  65   isEntryFunction: false
  66   scratchRSrcReg:  $sgpr0_sgpr1_sgpr2_sgpr3
  67   scratchWaveOffsetReg: $sgpr34
  68   frameOffsetReg:  $sgpr33
  69   stackPtrOffsetReg:  $sgpr32
  70
  71 body:             |
  72   bb.0:
  73     liveins: $vgpr1
  74
  75     ; CHECK-LABEL: name: scavenge_sgpr_pei_one_sgpr
  76     ; CHECK: liveins: $vgpr1
  77     ; CHECK: $sgpr27 = frame-setup COPY $sgpr33
  78     ; CHECK: $sgpr4 = frame-setup S_ADD_U32 $sgpr32, 524224, implicit-def $scc
  79     ; CHECK: $sgpr33 = frame-setup S_AND_B32 killed $sgpr4, 4294443008, implicit-def $scc
  80     ; CHECK: $sgpr32 = frame-setup S_ADD_U32 $sgpr32, 1572864, implicit-def $scc
  81     ; CHECK: S_NOP 0, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $sgpr6, implicit-def $sgpr7, implicit-def $sgpr8, implicit-def $sgpr9, implicit-def $sgpr10, implicit-def $sgpr11, implicit-def $sgpr12, implicit-def $sgpr13, implicit-def $sgpr14, implicit-def $sgpr15, implicit-def $sgpr16, implicit-def $sgpr17, implicit-def $sgpr18, implicit-def $sgpr19, implicit-def $sgpr20, implicit-def $sgpr21, implicit-def $sgpr22, implicit-def $sgpr23, implicit-def $sgpr24, implicit-def $sgpr25, implicit-def $sgpr26, implicit-def $sgpr17, implicit-def $sgpr28, implicit-def $sgpr30, implicit-def $sgpr31, implicit-def $vcc
  82     ; CHECK: $sgpr29 = S_SUB_U32 $sgpr33, $sgpr34, implicit-def $scc
  83     ; CHECK: $sgpr29 = S_LSHR_B32 killed $sgpr29, 6, implicit-def $scc
  84     ; CHECK: $sgpr29 = S_ADD_U32 killed $sgpr29, 8192, implicit-def $scc
  85     ; CHECK: $vgpr2 = COPY killed $sgpr29
  86     ; CHECK: $vgpr0 = V_OR_B32_e32 killed $vgpr2, $vgpr1, implicit $exec, implicit $sgpr4, implicit $sgpr5, implicit $sgpr6, implicit $sgpr7, implicit $sgpr8, implicit $sgpr9, implicit $sgpr10, implicit $sgpr11, implicit $sgpr12, implicit $sgpr13, implicit $sgpr14, implicit $sgpr15, implicit $sgpr16, implicit $sgpr17, implicit $sgpr18, implicit $sgpr19, implicit $sgpr20, implicit $sgpr21, implicit $sgpr22, implicit $sgpr23, implicit $sgpr24, implicit $sgpr25, implicit $sgpr26, implicit $sgpr17, implicit $sgpr28, implicit $sgpr31
  87     ; CHECK: $sgpr32 = frame-destroy S_SUB_U32 $sgpr32, 1572864, implicit-def $scc
  88     ; CHECK: $sgpr33 = frame-setup COPY $sgpr27
  89     ; CHECK: S_ENDPGM 0, implicit $vcc
  90     S_NOP 0, implicit-def $sgpr4, implicit-def $sgpr5, implicit-def $sgpr6, implicit-def $sgpr7, implicit-def $sgpr8, implicit-def $sgpr9, implicit-def $sgpr10, implicit-def $sgpr11, implicit-def $sgpr12, implicit-def $sgpr13, implicit-def $sgpr14, implicit-def $sgpr15, implicit-def $sgpr16, implicit-def $sgpr17, implicit-def $sgpr18, implicit-def $sgpr19, implicit-def $sgpr20, implicit-def $sgpr21, implicit-def $sgpr22, implicit-def $sgpr23, implicit-def $sgpr24, implicit-def $sgpr25, implicit-def $sgpr26, implicit-def $sgpr17, implicit-def $sgpr28, implicit-def $sgpr30, implicit-def $sgpr31, implicit-def $vcc
  91     $vgpr0 = V_OR_B32_e32 %stack.1, $vgpr1, implicit $exec, implicit $sgpr4, implicit $sgpr5, implicit $sgpr6, implicit $sgpr7, implicit $sgpr8, implicit $sgpr9, implicit $sgpr10, implicit $sgpr11, implicit $sgpr12, implicit $sgpr13, implicit $sgpr14, implicit $sgpr15, implicit $sgpr16, implicit $sgpr17, implicit $sgpr18, implicit $sgpr19, implicit $sgpr20, implicit $sgpr21, implicit $sgpr22, implicit $sgpr23, implicit $sgpr24, implicit $sgpr25, implicit $sgpr26, implicit $sgpr17, implicit $sgpr28, implicit $sgpr31
  92     S_ENDPGM 0, implicit $vcc
  93 ...