llvm.org GIT mirror llvm / stable test / CodeGen / AMDGPU / merge-load-store-physreg.mir
stable

Tree @stable (Download .tar.gz)

merge-load-store-physreg.mir @stableraw · history · blame

# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s

# Check that SILoadStoreOptimizer honors physregs defs/uses between moved
# instructions.
#
# The following IR snippet would usually be optimized by the peephole optimizer.
# However, an equivalent situation can occur with buffer instructions as well.

# CHECK-LABEL: name: scc_def_and_use_no_dependency
# CHECK: S_ADD_U32
# CHECK: S_ADDC_U32
# CHECK: DS_READ2_B32
--- |
  define amdgpu_kernel void @scc_def_and_use_no_dependency(i32 addrspace(3)* %ptr.0) nounwind {
    %ptr.4 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 1
    %ptr.64 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 16
    ret void
  }

  define amdgpu_kernel void @scc_def_and_use_dependency(i32 addrspace(3)* %ptr.0) nounwind {
    %ptr.4 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 1
    %ptr.64 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 16
    ret void
  }
...
---
name:            scc_def_and_use_no_dependency
alignment:       0
exposesReturnsTwice: false
legalized:       false
regBankSelected: false
selected:        false
tracksRegLiveness: false
liveins:
  - { reg: '$vgpr0' }
  - { reg: '$sgpr0' }
frameInfo:
  isFrameAddressTaken: false
  isReturnAddressTaken: false
  hasStackMap:     false
  hasPatchPoint:   false
  stackSize:       0
  offsetAdjustment: 0
  maxAlignment:    0
  adjustsStack:    false
  hasCalls:        false
  maxCallFrameSize: 0
  hasOpaqueSPAdjustment: false
  hasVAStart:      false
  hasMustTailInVarArgFunc: false
body:             |
  bb.0:
    liveins: $vgpr0, $sgpr0

    %1:vgpr_32 = COPY $vgpr0
    %10:sgpr_32 = COPY $sgpr0

    $m0 = S_MOV_B32 -1
    %2:vgpr_32 = DS_READ_B32 %1, 0, 0, implicit $m0, implicit $exec :: (load 4 from %ir.ptr.0)

    %11:sgpr_32 = S_ADD_U32 %10, 4, implicit-def $scc
    %12:sgpr_32 = S_ADDC_U32 %10, 0, implicit-def dead $scc, implicit $scc

    %3:vgpr_32 = DS_READ_B32 %1, 64, 0, implicit $m0, implicit $exec :: (load 4 from %ir.ptr.64)
    S_ENDPGM 0

...

# CHECK-LABEL: name: scc_def_and_use_dependency
# CHECK: DS_READ2_B32
# CHECK: S_ADD_U32
# CHECK: S_ADDC_U32
---
name:            scc_def_and_use_dependency
alignment:       0
exposesReturnsTwice: false
legalized:       false
regBankSelected: false
selected:        false
tracksRegLiveness: false
liveins:
  - { reg: '$vgpr0' }
  - { reg: '$sgpr0' }
frameInfo:
  isFrameAddressTaken: false
  isReturnAddressTaken: false
  hasStackMap:     false
  hasPatchPoint:   false
  stackSize:       0
  offsetAdjustment: 0
  maxAlignment:    0
  adjustsStack:    false
  hasCalls:        false
  maxCallFrameSize: 0
  hasOpaqueSPAdjustment: false
  hasVAStart:      false
  hasMustTailInVarArgFunc: false
body:             |
  bb.0:
    liveins: $vgpr0, $sgpr0

    %1:vgpr_32 = COPY $vgpr0
    %10:sgpr_32 = COPY $sgpr0

    $m0 = S_MOV_B32 -1
    %2:vgpr_32 = DS_READ_B32 %1, 0, 0, implicit $m0, implicit $exec :: (load 4 from %ir.ptr.0)
    %20:sgpr_32 = V_READFIRSTLANE_B32 %2, implicit $exec

    %21:sgpr_32 = S_ADD_U32 %20, 4, implicit-def $scc
    ; The S_ADDC_U32 depends on the first DS_READ_B32 only via SCC
    %11:sgpr_32 = S_ADDC_U32 %10, 0, implicit-def dead $scc, implicit $scc

    %3:vgpr_32 = DS_READ_B32 %1, 64, 0, implicit $m0, implicit $exec :: (load 4 from %ir.ptr.64)
    S_ENDPGM 0

...