This is an archive of the discontinued LLVM Phabricator instance.

Differential D128636

AMDGPU: Make packed 32-bit instructions rematerializable
ClosedPublic

Authored by arsenm on Jun 27 2022, 5:30 AM.

Details

Reviewers
rampitec
Group Reviewers
Restricted Project

Diff Detail

Event Timeline

arsenm created this revision.Jun 27 2022, 5:30 AM
Herald added a project: Restricted Project. · View Herald TranscriptJun 27 2022, 5:30 AM
Herald added subscribers: kosarev, foad, kerbowa and 7 others. · View Herald Transcript
arsenm requested review of this revision.Jun 27 2022, 5:30 AM
Herald added a project: Restricted Project. · View Herald TranscriptJun 27 2022, 5:30 AM
Herald added a subscriber: wdng. · View Herald Transcript
Harbormaster completed remote builds in B172167: Diff 440181.Jun 27 2022, 6:32 AM
rampitec added inline comments.Jun 27 2022, 11:08 AM
llvm/test/CodeGen/AMDGPU/remat-vop-gfx940.mir
2 ↗(On Diff #440181)

These should work with gfx90a and the test can be merged with the existing remat-vop.mir.

arsenm updated this revision to Diff 440627.Jun 28 2022, 7:48 AM

Merge test

Harbormaster completed remote builds in B172486: Diff 440627.Jun 28 2022, 7:49 AM
rampitec accepted this revision.Jun 28 2022, 11:27 AM

LGTM

This revision is now accepted and ready to land.Jun 28 2022, 11:27 AM
arsenm closed this revision.Jun 29 2022, 8:58 AM

ffd6aaf5b6663836700663ae0f9a2d80f4056689

Revision Contents

PathSize
llvm/
lib/
Target/
AMDGPU/
2 lines
test/
CodeGen/
AMDGPU/
138 lines

Diff 440627

llvm/lib/Target/AMDGPU/VOP3PInstructions.td

Show First 20 Lines • Show All 661 Lines • ▼ Show 20 Linesdef MAIInstInfoTable : GenericTable {
let Fields = [ let Fields = [
"Opcode", "is_dgemm", "is_gfx940_xdl" "Opcode", "is_dgemm", "is_gfx940_xdl"
]; ];
let PrimaryKey = ["Opcode"]; let PrimaryKey = ["Opcode"];
let PrimaryKeyName = "getMAIInstInfoHelper"; let PrimaryKeyName = "getMAIInstInfoHelper";
} }
let SubtargetPredicate = HasPackedFP32Ops, isCommutable = 1 in { let SubtargetPredicate = HasPackedFP32Ops, isCommutable = 1, isReMaterializable = 1 in {
defm V_PK_FMA_F32 : VOP3PInst<"v_pk_fma_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fma>; defm V_PK_FMA_F32 : VOP3PInst<"v_pk_fma_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fma>;
defm V_PK_MUL_F32 : VOP3PInst<"v_pk_mul_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fmul>; defm V_PK_MUL_F32 : VOP3PInst<"v_pk_mul_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fmul>;
defm V_PK_ADD_F32 : VOP3PInst<"v_pk_add_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fadd>; defm V_PK_ADD_F32 : VOP3PInst<"v_pk_add_f32", VOP3P_Profile<VOP_V2F32_V2F32_V2F32, VOP3_PACKED>, any_fadd>;
defm V_PK_MOV_B32 : VOP3PInst<"v_pk_mov_b32", VOP3P_Profile<VOP_V2I32_V2I32_V2I32, VOP3_PACKED>>; defm V_PK_MOV_B32 : VOP3PInst<"v_pk_mov_b32", VOP3P_Profile<VOP_V2I32_V2I32_V2I32, VOP3_PACKED>>;
} // End SubtargetPredicate = HasPackedFP32Ops, isCommutable = 1 } // End SubtargetPredicate = HasPackedFP32Ops, isCommutable = 1
def : MnemonicAlias<"v_accvgpr_read", "v_accvgpr_read_b32">; def : MnemonicAlias<"v_accvgpr_read", "v_accvgpr_read_b32">;
def : MnemonicAlias<"v_accvgpr_write", "v_accvgpr_write_b32">; def : MnemonicAlias<"v_accvgpr_write", "v_accvgpr_write_b32">;
▲ Show 20 Lines • Show All 355 Lines • Show Last 20 Lines

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

Show First 20 Lines • Show All 5,241 Lines • ▼ Show 20 Lines bb.0:
%1:vgpr_32 = nofpexcept V_FMA_MIX_F32 8, %0, 8, %0, 8, %0, 0, 0, 0, implicit $mode, implicit $exec %1:vgpr_32 = nofpexcept V_FMA_MIX_F32 8, %0, 8, %0, 8, %0, 0, 0, 0, implicit $mode, implicit $exec
%2:vgpr_32 = nofpexcept V_FMA_MIX_F32 9, %0, 9, %0, 9, %0, 0, 0, 0, implicit $mode, implicit $exec %2:vgpr_32 = nofpexcept V_FMA_MIX_F32 9, %0, 9, %0, 9, %0, 0, 0, 0, implicit $mode, implicit $exec
%3:vgpr_32 = nofpexcept V_FMA_MIX_F32 10, %0, 10, %0, 10, %0, 0, 0, 0, implicit $mode, implicit $exec %3:vgpr_32 = nofpexcept V_FMA_MIX_F32 10, %0, 10, %0, 10, %0, 0, 0, 0, implicit $mode, implicit $exec
S_NOP 0, implicit %1 S_NOP 0, implicit %1
S_NOP 0, implicit %2 S_NOP 0, implicit %2
S_NOP 0, implicit %3 S_NOP 0, implicit %3
S_ENDPGM 0, implicit %0 S_ENDPGM 0, implicit %0
... ...
---
name: test_remat_v_pk_fma_f32
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GCN-LABEL: name: test_remat_v_pk_fma_f32
; GCN: liveins: $vgpr0_vgpr1
; GCN-NEXT: {{ $}}
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_FMA_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_FMA_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_FMA_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: S_ENDPGM 0, implicit killed renamable $vgpr0_vgpr1
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:vreg_64_align2 = nofpexcept V_PK_FMA_F32 8, %0, 8, %0, 11, %0, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
%2:vreg_64_align2 = nofpexcept V_PK_FMA_F32 8, %0, 8, %0, 11, %0, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
%3:vreg_64_align2 = nofpexcept V_PK_FMA_F32 8, %0, 8, %0, 11, %0, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
S_NOP 0, implicit %1
S_NOP 0, implicit %2
S_NOP 0, implicit %3
S_ENDPGM 0, implicit %0
...
---
name: test_no_remat_v_pk_fma_f32
tracksRegLiveness: true
machineFunctionInfo:
stackPtrOffsetReg: $sgpr32
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GCN-LABEL: name: test_no_remat_v_pk_fma_f32
; GCN: liveins: $vgpr0_vgpr1
; GCN-NEXT: {{ $}}
; GCN-NEXT: renamable $vgpr2_vgpr3 = V_PK_FMA_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: SI_SPILL_V64_SAVE killed $vgpr2_vgpr3, %stack.1, $sgpr32, 0, implicit $exec :: (store (s64) into %stack.1, align 4, addrspace 5)
; GCN-NEXT: renamable $vgpr2_vgpr3 = V_PK_FMA_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: SI_SPILL_V64_SAVE killed $vgpr2_vgpr3, %stack.0, $sgpr32, 0, implicit $exec :: (store (s64) into %stack.0, align 4, addrspace 5)
; GCN-NEXT: renamable $vgpr2_vgpr3 = V_PK_FMA_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: SI_SPILL_V64_SAVE killed $vgpr2_vgpr3, %stack.2, $sgpr32, 0, implicit $exec :: (store (s64) into %stack.2, align 4, addrspace 5)
; GCN-NEXT: renamable $vgpr2_vgpr3 = SI_SPILL_V64_RESTORE %stack.1, $sgpr32, 0, implicit $exec :: (load (s64) from %stack.1, align 4, addrspace 5)
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = SI_SPILL_V64_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s64) from %stack.0, align 4, addrspace 5)
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = SI_SPILL_V64_RESTORE %stack.2, $sgpr32, 0, implicit $exec :: (load (s64) from %stack.2, align 4, addrspace 5)
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: S_ENDPGM 0, implicit killed renamable $vgpr0_vgpr1
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:vreg_64_align2 = V_PK_FMA_F32 8, %0, 8, %0, 11, %0, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
%2:vreg_64_align2 = V_PK_FMA_F32 8, %0, 8, %0, 11, %0, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
%3:vreg_64_align2 = V_PK_FMA_F32 8, %0, 8, %0, 11, %0, 0, 0, 0, 0, 0, implicit $mode, implicit $exec
S_NOP 0, implicit %1
S_NOP 0, implicit %2
S_NOP 0, implicit %3
S_ENDPGM 0, implicit %0
...
---
name: test_remat_v_pk_mul_f32
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GCN-LABEL: name: test_remat_v_pk_mul_f32
; GCN: liveins: $vgpr0_vgpr1
; GCN-NEXT: {{ $}}
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_MUL_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_MUL_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_MUL_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: S_ENDPGM 0, implicit killed renamable $vgpr0_vgpr1
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:vreg_64_align2 = nofpexcept V_PK_MUL_F32 8, %0, 8, %0, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
%2:vreg_64_align2 = nofpexcept V_PK_MUL_F32 8, %0, 8, %0, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
%3:vreg_64_align2 = nofpexcept V_PK_MUL_F32 8, %0, 8, %0, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
S_NOP 0, implicit %1
S_NOP 0, implicit %2
S_NOP 0, implicit %3
S_ENDPGM 0, implicit %0
...
---
name: test_remat_v_pk_add_f32
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GCN-LABEL: name: test_remat_v_pk_add_f32
; GCN: liveins: $vgpr0_vgpr1
; GCN-NEXT: {{ $}}
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_ADD_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_ADD_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_ADD_F32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: S_ENDPGM 0, implicit killed renamable $vgpr0_vgpr1
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:vreg_64_align2 = nofpexcept V_PK_ADD_F32 8, %0, 8, %0, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
%2:vreg_64_align2 = nofpexcept V_PK_ADD_F32 8, %0, 8, %0, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
%3:vreg_64_align2 = nofpexcept V_PK_ADD_F32 8, %0, 8, %0, 11, 0, 0, 0, 0, implicit $mode, implicit $exec
S_NOP 0, implicit %1
S_NOP 0, implicit %2
S_NOP 0, implicit %3
S_ENDPGM 0, implicit %0
...
---
name: test_remat_v_pk_mov_b32
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GCN-LABEL: name: test_remat_v_pk_mov_b32
; GCN: liveins: $vgpr0_vgpr1
; GCN-NEXT: {{ $}}
; GCN-NEXT: renamable $vgpr2_vgpr3 = V_PK_MOV_B32 8, $vgpr0_vgpr1, 8, $vgpr0_vgpr1, 11, 0, 0, 0, 0, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = V_PK_MOV_B32 9, $vgpr0_vgpr1, 9, $vgpr0_vgpr1, 12, 0, 0, 0, 0, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: renamable $vgpr2_vgpr3 = V_PK_MOV_B32 10, $vgpr0_vgpr1, 10, $vgpr0_vgpr1, 13, 0, 0, 0, 0, implicit $exec
; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr2_vgpr3
; GCN-NEXT: S_ENDPGM 0, implicit killed renamable $vgpr0_vgpr1
%0:vreg_64_align2 = COPY $vgpr0_vgpr1
%1:vreg_64_align2 = V_PK_MOV_B32 8, %0, 8, %0, 11, 0, 0, 0, 0, implicit $exec
%2:vreg_64_align2 = V_PK_MOV_B32 9, %0, 9, %0, 12, 0, 0, 0, 0, implicit $exec
%3:vreg_64_align2 = V_PK_MOV_B32 10, %0, 10, %0, 13, 0, 0, 0, 0, implicit $exec
S_NOP 0, implicit %1
S_NOP 0, implicit %2
S_NOP 0, implicit %3
S_ENDPGM 0, implicit %0
...