Home | History | Annotate | Download | only in AMDGPU
      1 # RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -run-pass si-optimize-exec-masking -verify-machineinstrs -o -  %s | FileCheck %s
      2 --- |
      3   define amdgpu_kernel void @undefined_physreg_sgpr_spill() #0 {
      4     unreachable
      5   }
      6 
      7   define amdgpu_kernel void @undefined_physreg_sgpr_spill_reorder() #0 {
      8     unreachable
      9   }
     10 
     11   attributes #0 = { nounwind "amdgpu-num-sgpr"="16" }
     12 
     13 ...
     14 ---
     15 
     16 # copy + s_and_b64 was turned into saveexec, deleting the copy,
     17 # leaving a spill of the undefined register.
     18 
     19 # CHECK-LABEL: name: undefined_physreg_sgpr_spill
     20 # CHECK: $sgpr0_sgpr1 = COPY $exec, implicit-def $exec
     21 # CHECK-NEXT: SI_SPILL_S64_SAVE $sgpr0_sgpr1,
     22 # CHECK-NEXT: $sgpr2_sgpr3 = S_AND_B64 killed $sgpr0_sgpr1, killed $vcc, implicit-def dead $scc
     23 # CHECK: $exec = COPY killed $sgpr2_sgpr3
     24 name:            undefined_physreg_sgpr_spill
     25 alignment:       0
     26 exposesReturnsTwice: false
     27 legalized:       false
     28 regBankSelected: false
     29 selected:        false
     30 tracksRegLiveness: true
     31 registers:
     32 liveins:
     33   - { reg: '$vgpr0', virtual-reg: '' }
     34   - { reg: '$sgpr4_sgpr5', virtual-reg: '' }
     35 stack:
     36   - { id: 0, name: '', type: spill-slot, offset: 0, size: 8, alignment: 4,
     37       stack-id: 1, callee-saved-register: '', callee-saved-restored: true,
     38       debug-info-variable: '', debug-info-expression: '',
     39       debug-info-location: '' }
     40 constants:
     41 body:             |
     42   bb.0:
     43     successors: %bb.1, %bb.2
     44     liveins: $vgpr0, $sgpr4_sgpr5
     45 
     46     $vgpr1_vgpr2 = COPY killed $sgpr4_sgpr5, implicit $exec
     47     $vgpr1 = GLOBAL_LOAD_UBYTE killed $vgpr1_vgpr2, 0, 0, 0, implicit $exec :: (non-temporal dereferenceable invariant load 1 from `i1 addrspace(2)* undef`)
     48     $vcc = V_CMP_NE_U32_e64 0, $vgpr0, implicit $exec
     49     $sgpr0_sgpr1 = V_CMP_EQ_U32_e64 1, killed $vgpr1, implicit $exec
     50     $vgpr1 = V_CNDMASK_B32_e64 0, -1, killed $sgpr0_sgpr1, implicit $exec
     51     $sgpr0_sgpr1 = COPY $exec, implicit-def $exec
     52     SI_SPILL_S64_SAVE $sgpr0_sgpr1, %stack.0, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11, implicit $sgpr13, implicit-def dead $m0 :: (store 8 into %stack.0, align 4, addrspace 5)
     53     $sgpr2_sgpr3 = S_AND_B64 killed $sgpr0_sgpr1, killed $vcc, implicit-def dead $scc
     54     $exec = S_MOV_B64_term killed $sgpr2_sgpr3
     55     SI_MASK_BRANCH %bb.2, implicit $exec
     56     S_BRANCH %bb.1
     57 
     58   bb.1:
     59     successors: %bb.3(0x80000000)
     60     liveins: $vgpr0, $vgpr1
     61 
     62     $sgpr2_sgpr3 = S_MOV_B64 0
     63     $vgpr2 = V_MOV_B32_e32 0, implicit $exec
     64     $sgpr4_sgpr5 = IMPLICIT_DEF
     65     S_BRANCH %bb.3
     66 
     67   bb.2:
     68     successors:
     69 
     70     $sgpr0_sgpr1 = SI_SPILL_S64_RESTORE %stack.0, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11, implicit $sgpr13, implicit-def dead $m0 :: (load 8 from %stack.0, align 4, addrspace 5)
     71     $exec = S_OR_B64 $exec, killed $sgpr0_sgpr1, implicit-def $scc
     72 
     73   bb.3:
     74     liveins: $vgpr0, $vgpr1, $vgpr2, $sgpr2_sgpr3, $sgpr4_sgpr5
     75 
     76     $vcc = COPY $vgpr1
     77     S_ENDPGM
     78 
     79 ...
     80 ---
     81 
     82 # Move spill to after future save instruction
     83 # CHECK-LABEL: {{^}}name: undefined_physreg_sgpr_spill_reorder
     84 # CHECK: $sgpr0_sgpr1 = COPY $exec, implicit-def $exec
     85 # CHECK: $sgpr2_sgpr3 = S_AND_B64 $sgpr0_sgpr1, killed $vcc, implicit-def dead $scc
     86 # CHECK: SI_SPILL_S64_SAVE killed $sgpr0_sgpr1, %stack.0, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11, implicit $sgpr13, implicit-def dead $m0 :: (store 8 into %stack.0, align 4, addrspace 5)
     87 # CHECK: $exec = COPY killed $sgpr2_sgpr3
     88 name:            undefined_physreg_sgpr_spill_reorder
     89 alignment:       0
     90 exposesReturnsTwice: false
     91 legalized:       false
     92 regBankSelected: false
     93 selected:        false
     94 tracksRegLiveness: true
     95 registers:
     96 liveins:
     97   - { reg: '$vgpr0', virtual-reg: '' }
     98   - { reg: '$sgpr4_sgpr5', virtual-reg: '' }
     99 stack:
    100   - { id: 0, name: '', type: spill-slot, offset: 0, size: 8, alignment: 4,
    101       stack-id: 1, callee-saved-register: '', callee-saved-restored: true,
    102       debug-info-variable: '', debug-info-expression: '',
    103       debug-info-location: '' }
    104 constants:
    105 body:             |
    106   bb.0:
    107     successors: %bb.1, %bb.2
    108     liveins: $vgpr0, $sgpr4_sgpr5
    109 
    110     $vgpr1_vgpr2 = COPY killed $sgpr4_sgpr5, implicit $exec
    111     $vgpr1 = GLOBAL_LOAD_UBYTE killed $vgpr1_vgpr2, 0, 0, 0, implicit $exec :: (non-temporal dereferenceable invariant load 1 from `i1 addrspace(2)* undef`)
    112     $vcc = V_CMP_NE_U32_e64 0, $vgpr0, implicit $exec
    113     $sgpr0_sgpr1 = V_CMP_EQ_U32_e64 1, killed $vgpr1, implicit $exec
    114     $vgpr1 = V_CNDMASK_B32_e64 0, -1, killed $sgpr0_sgpr1, implicit $exec
    115     $sgpr0_sgpr1 = COPY $exec, implicit-def $exec
    116     $sgpr2_sgpr3 = S_AND_B64 $sgpr0_sgpr1, killed $vcc, implicit-def dead $scc
    117     SI_SPILL_S64_SAVE killed $sgpr0_sgpr1, %stack.0, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11, implicit $sgpr13, implicit-def dead $m0 :: (store 8 into %stack.0, align 4, addrspace 5)
    118     $exec = S_MOV_B64_term killed $sgpr2_sgpr3
    119     SI_MASK_BRANCH %bb.2, implicit $exec
    120     S_BRANCH %bb.1
    121 
    122   bb.1:
    123     successors: %bb.3(0x80000000)
    124     liveins: $vgpr0, $vgpr1
    125 
    126     $sgpr2_sgpr3 = S_MOV_B64 0
    127     $vgpr2 = V_MOV_B32_e32 0, implicit $exec
    128     $sgpr4_sgpr5 = IMPLICIT_DEF
    129     S_BRANCH %bb.3
    130 
    131   bb.2:
    132     successors:
    133 
    134     $sgpr0_sgpr1 = SI_SPILL_S64_RESTORE %stack.0, implicit $exec, implicit $sgpr8_sgpr9_sgpr10_sgpr11, implicit $sgpr13, implicit-def dead $m0 :: (load 8 from %stack.0, align 4, addrspace 5)
    135     $exec = S_OR_B64 $exec, killed $sgpr0_sgpr1, implicit-def $scc
    136 
    137   bb.3:
    138     liveins: $vgpr0, $vgpr1, $vgpr2, $sgpr2_sgpr3, $sgpr4_sgpr5
    139 
    140     $vcc = COPY $vgpr1
    141     S_ENDPGM
    142 
    143 ...
    144