llvm/test/CodeGen/AMDGPU/remat-csr.mir

# RUN: llc -mtriple=amdgpu9.00-amd-amdhsa -passes='greedy<sgpr>' %s -o - | FileCheck %s

--- |
  define void @one_use() { ret void }
  define void @implicit_def() { ret void }
  define void @two_uses() { ret void }
  define void @hot_loop_use() { ret void }
  declare void @callee()
...
# A value used once after the call is rematerialized instead of using a CSR.
---
name: one_use
tracksRegLiveness: true
frameInfo:
  adjustsStack: true
body: |
  ; CHECK-LABEL: name: one_use
  ; CHECK-NOT: assigned-phys: '$sgpr54'
  ; CHECK: %[[MASK:[0-9]+]]:sreg_32 = S_MOV_B32 2147483647
  ; CHECK-NEXT: %{{[0-9]+}}:vgpr_32 = V_BFI_B32_e64 %[[MASK]], %{{[0-9]+}}, %{{[0-9]+}}, implicit $exec
  bb.0:
    liveins: $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11, $sgpr12, $sgpr13, $sgpr14, $sgpr15, $sgpr32
    %0:sgpr_64 = COPY $sgpr4_sgpr5
    %1:sgpr_64 = COPY $sgpr6_sgpr7
    %2:sgpr_64 = COPY $sgpr8_sgpr9
    %3:sgpr_64 = COPY $sgpr10_sgpr11
    %4:sgpr_32 = COPY $sgpr12
    %5:sgpr_32 = COPY $sgpr13
    %6:sgpr_32 = COPY $sgpr14
    %7:sgpr_32 = COPY $sgpr15
    %8:sreg_32 = S_MOV_B32 2147483647
    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    %9:sreg_64_xexec = IMPLICIT_DEF
    dead $sgpr30_sgpr31 = SI_CALL %9, @callee, csr_amdgpu, implicit-def $vgpr0
    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    S_NOP 0, implicit %0, implicit %1, implicit %2, implicit %3, implicit %4, implicit %5, implicit %6, implicit %7
    %10:vgpr_32 = IMPLICIT_DEF
    %11:vgpr_32 = IMPLICIT_DEF
    %12:vgpr_32 = V_BFI_B32_e64 %8, %10, %11, implicit $exec
    S_NOP 0, implicit %12
    S_ENDPGM 0
...
# An implicit definition used after the call is also rematerialized.
---
name: implicit_def
tracksRegLiveness: true
frameInfo:
  adjustsStack: true
body: |
  ; CHECK-LABEL: name: implicit_def
  ; CHECK-NOT: assigned-phys: '$sgpr54'
  ; CHECK: %[[MASK:[0-9]+]]:sreg_32 = IMPLICIT_DEF
  ; CHECK-NEXT: %{{[0-9]+}}:vgpr_32 = V_BFI_B32_e64 %[[MASK]], %{{[0-9]+}}, %{{[0-9]+}}, implicit $exec
  bb.0:
    liveins: $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11, $sgpr12, $sgpr13, $sgpr14, $sgpr15, $sgpr32
    %0:sgpr_64 = COPY $sgpr4_sgpr5
    %1:sgpr_64 = COPY $sgpr6_sgpr7
    %2:sgpr_64 = COPY $sgpr8_sgpr9
    %3:sgpr_64 = COPY $sgpr10_sgpr11
    %4:sgpr_32 = COPY $sgpr12
    %5:sgpr_32 = COPY $sgpr13
    %6:sgpr_32 = COPY $sgpr14
    %7:sgpr_32 = COPY $sgpr15
    %8:sreg_32 = IMPLICIT_DEF
    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    %9:sreg_64_xexec = IMPLICIT_DEF
    dead $sgpr30_sgpr31 = SI_CALL %9, @callee, csr_amdgpu, implicit-def $vgpr0
    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    S_NOP 0, implicit %0, implicit %1, implicit %2, implicit %3, implicit %4, implicit %5, implicit %6, implicit %7
    %10:vgpr_32 = IMPLICIT_DEF
    %11:vgpr_32 = IMPLICIT_DEF
    %12:vgpr_32 = V_BFI_B32_e64 %8, %10, %11, implicit $exec
    S_NOP 0, implicit %12
    S_ENDPGM 0
...
# A value used on both sides of the call is rematerialized for the later use.
---
name: two_uses
tracksRegLiveness: true
frameInfo:
  adjustsStack: true
body: |
  ; CHECK-LABEL: name: two_uses
  ; CHECK-NOT: assigned-phys: '$sgpr54'
  ; CHECK: %[[MASK0:[0-9]+]]:sreg_32 = S_MOV_B32 2147483647
  ; CHECK: %{{[0-9]+}}:vgpr_32 = V_AND_B32_e64 %[[MASK0]], %{{[0-9]+}}, implicit $exec
  ; CHECK: %[[MASK1:[0-9]+]]:sreg_32 = S_MOV_B32 2147483647
  ; CHECK-NEXT: %{{[0-9]+}}:vgpr_32 = V_BFI_B32_e64 %[[MASK1]], %{{[0-9]+}}, %{{[0-9]+}}, implicit $exec
  bb.0:
    liveins: $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11, $sgpr12, $sgpr13, $sgpr14, $sgpr15, $sgpr32
    %0:sgpr_64 = COPY $sgpr4_sgpr5
    %1:sgpr_64 = COPY $sgpr6_sgpr7
    %2:sgpr_64 = COPY $sgpr8_sgpr9
    %3:sgpr_64 = COPY $sgpr10_sgpr11
    %4:sgpr_32 = COPY $sgpr12
    %5:sgpr_32 = COPY $sgpr13
    %6:sgpr_32 = COPY $sgpr14
    %7:sgpr_32 = COPY $sgpr15
    %8:sreg_32 = S_MOV_B32 2147483647
    %10:vgpr_32 = IMPLICIT_DEF
    %11:vgpr_32 = V_AND_B32_e64 %8, %10, implicit $exec
    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    %9:sreg_64_xexec = IMPLICIT_DEF
    dead $sgpr30_sgpr31 = SI_CALL %9, @callee, csr_amdgpu, implicit-def $vgpr0
    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    S_NOP 0, implicit %0, implicit %1, implicit %2, implicit %3, implicit %4, implicit %5, implicit %6, implicit %7
    %12:vgpr_32 = IMPLICIT_DEF
    %13:vgpr_32 = V_BFI_B32_e64 %8, %11, %12, implicit $exec
    S_NOP 0, implicit %13
    S_ENDPGM 0
...
# A use in a hot loop makes rematerialization costlier than assigning a CSR.
---
name: hot_loop_use
tracksRegLiveness: true
frameInfo:
  adjustsStack: true
body: |
  ; CHECK-LABEL: name: hot_loop_use
  ; The use in bb.1 is more expensive to rematerialize than using a CSR.
  ; CHECK: - { id: 8, class: sreg_32, preferred-register: '', flags: [  ], assigned-phys: '$sgpr53' }
  bb.0:
    successors: %bb.1
    liveins: $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11, $sgpr12, $sgpr13, $sgpr14, $sgpr15, $sgpr32
    %0:sgpr_64 = COPY $sgpr4_sgpr5
    %1:sgpr_64 = COPY $sgpr6_sgpr7
    %2:sgpr_64 = COPY $sgpr8_sgpr9
    %3:sgpr_64 = COPY $sgpr10_sgpr11
    %4:sgpr_32 = COPY $sgpr12
    %5:sgpr_32 = COPY $sgpr13
    %6:sgpr_32 = COPY $sgpr14
    %7:sgpr_32 = COPY $sgpr15
    %8:sreg_32 = S_MOV_B32 2147483647
    S_BRANCH %bb.1
  bb.1:
    successors: %bb.1(0x7f000000), %bb.2(0x01000000)
    liveins: $sgpr32
    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    %9:sreg_64_xexec = IMPLICIT_DEF
    dead $sgpr30_sgpr31 = SI_CALL %9, @callee, csr_amdgpu, implicit-def $vgpr0
    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
    %10:vgpr_32 = IMPLICIT_DEF
    %11:vgpr_32 = V_BFI_B32_e64 %8, %10, %10, implicit $exec
    S_CBRANCH_SCC1 %bb.1, implicit undef $scc
    S_BRANCH %bb.2
  bb.2:
    S_NOP 0, implicit %0, implicit %1, implicit %2, implicit %3, implicit %4, implicit %5, implicit %6, implicit %7, implicit %11
    S_ENDPGM 0
...