[llvm] [AMDGPU][IGLP] SingleWaveOpt: Don't collect PERMs for generic VALU SG (PR #67759)
via llvm-commits
llvm-commits at lists.llvm.org
Thu Sep 28 18:22:09 PDT 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
<details>
<summary>Changes</summary>
Allowing PERMs to fall into the generic VALU for this mutation is creating edges which block them being classified into the V_PERM SchedGroups. Since other groups in the pipeline depend on V_PERM placement (e.g. DS_WRITE group must be a succ of the V_PERM in stage 2a), not being able to place the V_PERM is a bit catastrophic.
Additionally, remove a bad assert in SharesPredWithPrevNthGroup -- it is possible that we will have an empty cache if OtherGroup has no elements which have a V_PERM pred (possible if the V_PERM SG is empty).
---
Patch is 42.27 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/67759.diff
3 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp (+15-1)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.iglp.opt.single.2b.mir (+67-67)
- (modified) llvm/test/CodeGen/AMDGPU/llvm.amdgcn.iglp.opt.single.2c.mir (+39-39)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp b/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
index f60236080351a2c..d49a439c1675e35 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp
@@ -902,6 +902,20 @@ void MFMASmallGemmOpt::applyIGLPStrategy(
class MFMASmallGemmSingleWaveOpt final : public IGLPStrategy {
private:
+ // Whether or not the instruction is NOT a V_PERM
+ class IsNotPerm final : public InstructionRule {
+ public:
+ bool apply(const SUnit *SU, const ArrayRef<SUnit *> Collection,
+ SmallVectorImpl<SchedGroup> &SyncPipe) override {
+ return SU->getInstr()->getOpcode() != AMDGPU::V_PERM_B32_e64;
+ }
+
+ IsNotPerm(const SIInstrInfo *TII, unsigned SGID,
+ bool NeedsCache = false)
+ : InstructionRule(TII, SGID, NeedsCache) {}
+ };
+
+
// Whether the DS_READ is a predecessor of first four MFMA in region
class EnablesInitialMFMA final : public InstructionRule {
public:
@@ -1078,7 +1092,6 @@ class MFMASmallGemmSingleWaveOpt final : public IGLPStrategy {
}
}
- assert(Cache->size());
auto DAG = SyncPipe[0].DAG;
// Does the previous DS_WRITE share a V_PERM predecessor with this
// VMEM_READ
@@ -1198,6 +1211,7 @@ void MFMASmallGemmSingleWaveOpt::applyIGLPStrategy(
SG = &SyncedSchedGroups[PipelineSyncID].emplace_back(
SchedGroupMask::VALU, 2, PipelineSyncID, DAG, TII);
+ SG->addRule(std::make_shared<IsNotPerm>(TII, SG->getSGID(), false));
SG->initSchedGroup(SyncedInstrs[SG->getSyncID()]);
}
}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.iglp.opt.single.2b.mir b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.iglp.opt.single.2b.mir
index 715a88351bb92fd..bdfa41f9501f615 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.iglp.opt.single.2b.mir
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.iglp.opt.single.2b.mir
@@ -77,103 +77,103 @@ body: |
; GCN-NEXT: [[DS_READ_B128_gfx9_1:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 0, 0, implicit $exec :: (load (s128) from %ir.in4, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[DS_READ_B128_gfx9_2:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF6]], 1040, 0, implicit $exec :: (load (s128) from %ir.in1, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[DS_READ_B128_gfx9_3:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 2064, 0, implicit $exec :: (load (s128) from %ir.in5, !alias.scope !0, addrspace 3)
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub0_sub1, [[DS_READ_B128_gfx9_1]].sub0_sub1, [[V_MFMA_F32_32X32X8F16_mac_e64_]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub0_sub1, [[DS_READ_B128_gfx9_1]].sub0_sub1, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_4:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF6]], 2080, 0, implicit $exec :: (load (s128) from %ir.in2, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[V_ADD_U32_e32_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF8]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_1:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF9]], implicit $exec
- ; GCN-NEXT: dead [[V_MFMA_F32_32X32X8F16_mac_e64_1:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub2_sub3, [[DS_READ_B128_gfx9_1]].sub2_sub3, [[V_MFMA_F32_32X32X8F16_mac_e64_]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: dead [[V_MFMA_F32_32X32X8F16_mac_e64_:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub2_sub3, [[DS_READ_B128_gfx9_1]].sub2_sub3, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_5:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF6]], 3120, 0, implicit $exec :: (load (s128) from %ir.in3, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[V_ADD_U32_e32_2:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF10]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_3:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF11]], implicit $exec
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_2:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_2]].sub0_sub1, [[DS_READ_B128_gfx9_3]].sub0_sub1, [[V_MFMA_F32_32X32X8F16_mac_e64_2]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_2]].sub0_sub1, [[DS_READ_B128_gfx9_3]].sub0_sub1, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_6:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 4128, 0, implicit $exec :: (load (s128) from %ir.in6, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[V_ADD_U32_e32_4:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF12]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_5:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF13]], implicit $exec
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_2:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_2]].sub2_sub3, [[DS_READ_B128_gfx9_3]].sub2_sub3, [[V_MFMA_F32_32X32X8F16_mac_e64_2]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_2]].sub2_sub3, [[DS_READ_B128_gfx9_3]].sub2_sub3, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_7:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 6192, 0, implicit $exec :: (load (s128) from %ir.in7, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[V_ADD_U32_e32_6:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF14]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_7:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF15]], implicit $exec
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_2:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_4]].sub0_sub1, [[DS_READ_B128_gfx9_6]].sub0_sub1, [[V_MFMA_F32_32X32X8F16_mac_e64_2]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_4]].sub0_sub1, [[DS_READ_B128_gfx9_6]].sub0_sub1, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_8:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 1024, 0, implicit $exec :: (load (s128) from %ir.in8, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[V_ADD_U32_e32_8:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF16]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_9:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF17]], implicit $exec
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_2:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_4]].sub2_sub3, [[DS_READ_B128_gfx9_6]].sub2_sub3, [[V_MFMA_F32_32X32X8F16_mac_e64_2]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_4]].sub2_sub3, [[DS_READ_B128_gfx9_6]].sub2_sub3, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_9:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 3088, 0, implicit $exec :: (load (s128) from %ir.in9, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[V_ADD_U32_e32_10:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF18]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_11:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF19]], implicit $exec
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_2:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_5]].sub0_sub1, [[DS_READ_B128_gfx9_7]].sub0_sub1, [[V_MFMA_F32_32X32X8F16_mac_e64_2]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_5]].sub0_sub1, [[DS_READ_B128_gfx9_7]].sub0_sub1, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_10:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 5152, 0, implicit $exec :: (load (s128) from %ir.in10, !alias.scope !0, addrspace 3)
; GCN-NEXT: [[V_ADD_U32_e32_12:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF20]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_13:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF21]], implicit $exec
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_2:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_5]].sub2_sub3, [[DS_READ_B128_gfx9_7]].sub2_sub3, [[V_MFMA_F32_32X32X8F16_mac_e64_2]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_5]].sub2_sub3, [[DS_READ_B128_gfx9_7]].sub2_sub3, [[DEF]], 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[DS_READ_B128_gfx9_11:%[0-9]+]]:av_128_align2 = DS_READ_B128_gfx9 [[DEF7]], 7216, 0, implicit $exec :: (load (s128) from %ir.in11, !alias.scope !0, addrspace 3)
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_2:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub0_sub1, [[DS_READ_B128_gfx9_8]].sub0_sub1, [[V_MFMA_F32_32X32X8F16_mac_e64_2]], 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: undef %63.sub1:vreg_128_align2 = V_PERM_B32_e64 [[DEF28]], [[DEF29]], [[DEF44]], implicit $exec
- ; GCN-NEXT: %63.sub0:vreg_128_align2 = V_PERM_B32_e64 [[DEF30]], [[DEF31]], [[DEF44]], implicit $exec
- ; GCN-NEXT: %63.sub3:vreg_128_align2 = V_PERM_B32_e64 [[DEF24]], [[DEF25]], [[DEF44]], implicit $exec
- ; GCN-NEXT: %63.sub2:vreg_128_align2 = V_PERM_B32_e64 [[DEF26]], [[DEF27]], [[DEF44]], implicit $exec
- ; GCN-NEXT: DS_WRITE_B128_gfx9 [[DEF40]], %63, 0, 0, implicit $exec :: (store (s128) into %ir.in0, !alias.scope !0, addrspace 3)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in12, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN1:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_1]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in13, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN2:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_2]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in14, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN3:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_3]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in15, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_3:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub2_sub3, [[DS_READ_B128_gfx9_8]].sub2_sub3, [[V_MFMA_F32_32X32X8F16_mac_e64_3]], 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN4:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_4]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in16, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN5:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_5]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in17, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN6:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_6]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in18, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN7:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_7]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in19, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_3:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_2]].sub0_sub1, [[DS_READ_B128_gfx9_9]].sub0_sub1, [[V_MFMA_F32_32X32X8F16_mac_e64_3]], 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: undef %64.sub1:vreg_128_align2 = V_PERM_B32_e64 [[DEF36]], [[DEF37]], [[DEF44]], implicit $exec
- ; GCN-NEXT: %64.sub0:vreg_128_align2 = V_PERM_B32_e64 [[DEF38]], [[DEF39]], [[DEF44]], implicit $exec
- ; GCN-NEXT: %64.sub3:vreg_128_align2 = V_PERM_B32_e64 [[DEF32]], [[DEF33]], [[DEF44]], implicit $exec
- ; GCN-NEXT: %64.sub2:vreg_128_align2 = V_PERM_B32_e64 [[DEF34]], [[DEF35]], [[DEF44]], implicit $exec
- ; GCN-NEXT: DS_WRITE_B128_gfx9 [[DEF40]], %64, 1040, 0, implicit $exec :: (store (s128) into %ir.in1, !alias.scope !0, addrspace 3)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN8:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_8]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in20, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN9:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_9]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in21, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN10:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_10]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in22, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN11:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_11]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in23, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_3:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_2]].sub2_sub3, [[DS_READ_B128_gfx9_9]].sub2_sub3, [[V_MFMA_F32_32X32X8F16_mac_e64_3]], 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN12:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_12]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in24, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN13:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_13]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in25, !alias.scope !0, addrspace 7)
; GCN-NEXT: [[V_ADD_U32_e32_14:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF22]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_15:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF42]], [[DEF23]], implicit $exec
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN14:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_14]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in26, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[BUFFER_LOAD_USHORT_OFFEN15:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_15]], [[DEF47]], 0, 0, 0, 0, implicit $exec
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_3:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_4]].sub0_sub1, [[DS_READ_B128_gfx9_10]].sub0_sub1, [[V_MFMA_F32_32X32X8F16_mac_e64_3]], 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: DS_WRITE_B128_gfx9 [[DEF41]], [[DEF2]], 0, 0, implicit $exec :: (store (s128) into %ir.in2, !alias.scope !0, addrspace 3)
+ ; GCN-NEXT: [[DEF1:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub0_sub1, [[DS_READ_B128_gfx9_8]].sub0_sub1, [[DEF1]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: undef [[V_PERM_B32_e64_:%[0-9]+]].sub3:vreg_128_align2 = V_PERM_B32_e64 [[DEF24]], [[DEF25]], [[DEF44]], implicit $exec
+ ; GCN-NEXT: [[V_PERM_B32_e64_:%[0-9]+]].sub2:vreg_128_align2 = V_PERM_B32_e64 [[DEF26]], [[DEF27]], [[DEF44]], implicit $exec
+ ; GCN-NEXT: [[V_PERM_B32_e64_:%[0-9]+]].sub1:vreg_128_align2 = V_PERM_B32_e64 [[DEF28]], [[DEF29]], [[DEF44]], implicit $exec
+ ; GCN-NEXT: [[V_PERM_B32_e64_:%[0-9]+]].sub0:vreg_128_align2 = V_PERM_B32_e64 [[DEF30]], [[DEF31]], [[DEF44]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_16:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 -16, [[DEF45]], implicit $exec
; GCN-NEXT: [[V_ADD_U32_e32_17:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 -16, [[DEF46]], implicit $exec
- ; GCN-NEXT: [[BUFFER_LOAD_DWORDX4_OFFEN:%[0-9]+]]:av_128_align2 = BUFFER_LOAD_DWORDX4_OFFEN [[V_ADD_U32_e32_16]], [[DEF48]], 0, 0, 0, 0, implicit $exec :: (load (s128) from %ir.in26, !alias.scope !0, addrspace 7)
- ; GCN-NEXT: [[V_MFMA_F32_32X32X8F16_mac_e64_3:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_4]].sub2_sub3, [[DS_READ_B128_gfx9_10]].sub2_sub3, [[V_MFMA_F32_32X32X8F16_mac_e64_3]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: DS_WRITE_B128_gfx9 [[DEF40]], [[V_PERM_B32_e64_]], 0, 0, implicit $exec :: (store (s128) into %ir.in0, !alias.scope !0, addrspace 3)
+ ; GCN-NEXT: [[DEF31:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in12, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF30:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_1]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in13, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF29:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_2]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in14, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF28:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_3]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in15, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF1:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_]].sub2_sub3, [[DS_READ_B128_gfx9_8]].sub2_sub3, [[DEF1]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: [[DEF27:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_4]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in16, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF26:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_5]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in17, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF25:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_6]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in18, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF24:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_7]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in19, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF23:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF43]], [[DEF23]], implicit $exec
+ ; GCN-NEXT: [[DEF22:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF43]], [[DEF22]], implicit $exec
+ ; GCN-NEXT: [[DEF1:%[0-9]+]]:areg_512_align2 = contract V_MFMA_F32_32X32X8F16_mac_e64 [[DS_READ_B128_gfx9_2]].sub0_sub1, [[DS_READ_B128_gfx9_9]].sub0_sub1, [[DEF1]], 0, 0, 0, implicit $mode, implicit $exec
+ ; GCN-NEXT: undef [[V_PERM_B32_e64_1:%[0-9]+]].sub3:vreg_128_align2 = V_PERM_B32_e64 [[DEF32]], [[DEF33]], [[DEF44]], implicit $exec
+ ; GCN-NEXT: [[V_PERM_B32_e64_1:%[0-9]+]].sub2:vreg_128_align2 = V_PERM_B32_e64 [[DEF34]], [[DEF35]], [[DEF44]], implicit $exec
+ ; GCN-NEXT: [[V_PERM_B32_e64_1:%[0-9]+]].sub1:vreg_128_align2 = V_PERM_B32_e64 [[DEF36]], [[DEF37]], [[DEF44]], implicit $exec
+ ; GCN-NEXT: [[V_PERM_B32_e64_1:%[0-9]+]].sub0:vreg_128_align2 = V_PERM_B32_e64 [[DEF38]], [[DEF39]], [[DEF44]], implicit $exec
+ ; GCN-NEXT: [[DEF21:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF43]], [[DEF21]], implicit $exec
+ ; GCN-NEXT: [[DEF20:%[0-9]+]]:vgpr_32 = V_ADD_U32_e32 [[DEF43]], [[DEF20]], implicit $exec
+ ; GCN-NEXT: DS_WRITE_B128_gfx9 [[DEF40]], [[V_PERM_B32_e64_1]], 1040, 0, implicit $exec :: (store (s128) into %ir.in1, !alias.scope !0, addrspace 3)
+ ; GCN-NEXT: [[DEF32:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_8]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in20, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF33:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_9]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in21, !alias.scope !0, addrspace 7)
+ ; GCN-NEXT: [[DEF34:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_USHORT_OFFEN [[V_ADD_U32_e32_10]], [[DEF47]], 0, 0, 0, 0, implicit $exec :: (load (s16) from %ir.in22, !alias.scope...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/67759
More information about the llvm-commits
mailing list