[VectorCombine] Use getInstructionCost to cost Shuffle. #122068
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This allows it to produce a more accurate cost for the shuffle, using the more accurate calls to getShuffleCost in getInstructionCost.
Member
|
@llvm/pr-subscribers-vectorizers Author: David Green (davemgreen) ChangesThis allows it to produce a more accurate cost for the shuffle, using the more accurate calls to getShuffleCost in getInstructionCost. It helps fix some of the regressions from vector combine a little while ago, now that we have better subvector extract costs. Full diff: https://github.com/llvm/llvm-project/pull/122068.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
index 120eafae8c5ac5..1a669b5058e799 100644
--- a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
+++ b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
@@ -2023,9 +2023,7 @@ bool VectorCombine::foldShuffleOfShuffles(Instruction &I) {
if (Match1)
InnerCost1 = TTI.getInstructionCost(cast<Instruction>(OuterV1), CostKind);
- InstructionCost OuterCost = TTI.getShuffleCost(
- TargetTransformInfo::SK_PermuteTwoSrc, ShuffleImmTy, OuterMask, CostKind,
- 0, nullptr, {OuterV0, OuterV1}, &I);
+ InstructionCost OuterCost = TTI.getInstructionCost(&I, CostKind);
InstructionCost OldCost = InnerCost0 + InnerCost1 + OuterCost;
diff --git a/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll b/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll
index 7d9524420286d6..6a08402fed1db4 100644
--- a/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll
+++ b/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll
@@ -416,21 +416,23 @@ define internal noundef <8 x i16> @_ZL24cmplx_mul_combined_re_im11__Int16x8_t20c
; CHECK-NEXT: [[SCALE_SROA_0_0_EXTRACT_TRUNC:%.*]] = trunc i64 [[SCALE_COERCE]] to i16
; CHECK-NEXT: [[SCALE_SROA_2_0_EXTRACT_SHIFT36:%.*]] = lshr i64 [[SCALE_COERCE]], 16
; CHECK-NEXT: [[SCALE_SROA_2_0_EXTRACT_TRUNC:%.*]] = trunc i64 [[SCALE_SROA_2_0_EXTRACT_SHIFT36]] to i16
+; CHECK-NEXT: [[SHUFFLE_I:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <8 x i32> <i32 1, i32 0, i32 3, i32 2, i32 5, i32 4, i32 7, i32 6>
; CHECK-NEXT: [[VECINIT_I19:%.*]] = insertelement <8 x i16> poison, i16 [[SCALE_SROA_0_0_EXTRACT_TRUNC]], i64 0
; CHECK-NEXT: [[VECINIT_I:%.*]] = insertelement <8 x i16> poison, i16 [[SCALE_SROA_2_0_EXTRACT_TRUNC]], i64 0
; CHECK-NEXT: [[VECINIT7_I:%.*]] = shufflevector <8 x i16> [[VECINIT_I]], <8 x i16> poison, <8 x i32> zeroinitializer
; CHECK-NEXT: [[VQNEGQ_V1_I:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.sqneg.v8i16(<8 x i16> [[VECINIT7_I]])
+; CHECK-NEXT: [[VBSL5_I:%.*]] = shufflevector <8 x i16> [[VQNEGQ_V1_I]], <8 x i16> [[VECINIT_I]], <8 x i32> <i32 0, i32 8, i32 2, i32 8, i32 4, i32 8, i32 6, i32 8>
; CHECK-NEXT: [[SHUFFLE_I85:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
; CHECK-NEXT: [[SHUFFLE_I82:%.*]] = shufflevector <8 x i16> [[VECINIT_I19]], <8 x i16> poison, <4 x i32> zeroinitializer
; CHECK-NEXT: [[VQDMULL_V2_I72:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I85]], <4 x i16> [[SHUFFLE_I82]])
; CHECK-NEXT: [[SHUFFLE_I97:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
; CHECK-NEXT: [[VQDMULL_V2_I:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I97]], <4 x i16> [[SHUFFLE_I82]])
-; CHECK-NEXT: [[SHUFFLE_I79:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 1, i32 0, i32 3, i32 2>
-; CHECK-NEXT: [[SHUFFLE_I76:%.*]] = shufflevector <8 x i16> [[VQNEGQ_V1_I]], <8 x i16> [[VECINIT_I]], <4 x i32> <i32 0, i32 8, i32 2, i32 8>
+; CHECK-NEXT: [[SHUFFLE_I79:%.*]] = shufflevector <8 x i16> [[SHUFFLE_I]], <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT: [[SHUFFLE_I76:%.*]] = shufflevector <8 x i16> [[VBSL5_I]], <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
; CHECK-NEXT: [[VQDMLAL2_I106:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I79]], <4 x i16> [[SHUFFLE_I76]])
; CHECK-NEXT: [[VQDMLAL_V3_I107:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqadd.v4i32(<4 x i32> [[VQDMULL_V2_I72]], <4 x i32> [[VQDMLAL2_I106]])
-; CHECK-NEXT: [[SHUFFLE_I91:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 5, i32 4, i32 7, i32 6>
-; CHECK-NEXT: [[SHUFFLE_I88:%.*]] = shufflevector <8 x i16> [[VQNEGQ_V1_I]], <8 x i16> [[VECINIT_I]], <4 x i32> <i32 4, i32 8, i32 6, i32 8>
+; CHECK-NEXT: [[SHUFFLE_I91:%.*]] = shufflevector <8 x i16> [[SHUFFLE_I]], <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; CHECK-NEXT: [[SHUFFLE_I88:%.*]] = shufflevector <8 x i16> [[VBSL5_I]], <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
; CHECK-NEXT: [[VQDMLAL2_I:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I91]], <4 x i16> [[SHUFFLE_I88]])
; CHECK-NEXT: [[VQDMLAL_V3_I:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqadd.v4i32(<4 x i32> [[VQDMULL_V2_I]], <4 x i32> [[VQDMLAL2_I]])
; CHECK-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VQDMLAL_V3_I107]] to <8 x i16>
|
Member
|
@llvm/pr-subscribers-llvm-transforms Author: David Green (davemgreen) ChangesThis allows it to produce a more accurate cost for the shuffle, using the more accurate calls to getShuffleCost in getInstructionCost. It helps fix some of the regressions from vector combine a little while ago, now that we have better subvector extract costs. Full diff: https://github.com/llvm/llvm-project/pull/122068.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
index 120eafae8c5ac5..1a669b5058e799 100644
--- a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
+++ b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
@@ -2023,9 +2023,7 @@ bool VectorCombine::foldShuffleOfShuffles(Instruction &I) {
if (Match1)
InnerCost1 = TTI.getInstructionCost(cast<Instruction>(OuterV1), CostKind);
- InstructionCost OuterCost = TTI.getShuffleCost(
- TargetTransformInfo::SK_PermuteTwoSrc, ShuffleImmTy, OuterMask, CostKind,
- 0, nullptr, {OuterV0, OuterV1}, &I);
+ InstructionCost OuterCost = TTI.getInstructionCost(&I, CostKind);
InstructionCost OldCost = InnerCost0 + InnerCost1 + OuterCost;
diff --git a/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll b/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll
index 7d9524420286d6..6a08402fed1db4 100644
--- a/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll
+++ b/llvm/test/Transforms/PhaseOrdering/AArch64/block_scaling_decompr_8bit.ll
@@ -416,21 +416,23 @@ define internal noundef <8 x i16> @_ZL24cmplx_mul_combined_re_im11__Int16x8_t20c
; CHECK-NEXT: [[SCALE_SROA_0_0_EXTRACT_TRUNC:%.*]] = trunc i64 [[SCALE_COERCE]] to i16
; CHECK-NEXT: [[SCALE_SROA_2_0_EXTRACT_SHIFT36:%.*]] = lshr i64 [[SCALE_COERCE]], 16
; CHECK-NEXT: [[SCALE_SROA_2_0_EXTRACT_TRUNC:%.*]] = trunc i64 [[SCALE_SROA_2_0_EXTRACT_SHIFT36]] to i16
+; CHECK-NEXT: [[SHUFFLE_I:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <8 x i32> <i32 1, i32 0, i32 3, i32 2, i32 5, i32 4, i32 7, i32 6>
; CHECK-NEXT: [[VECINIT_I19:%.*]] = insertelement <8 x i16> poison, i16 [[SCALE_SROA_0_0_EXTRACT_TRUNC]], i64 0
; CHECK-NEXT: [[VECINIT_I:%.*]] = insertelement <8 x i16> poison, i16 [[SCALE_SROA_2_0_EXTRACT_TRUNC]], i64 0
; CHECK-NEXT: [[VECINIT7_I:%.*]] = shufflevector <8 x i16> [[VECINIT_I]], <8 x i16> poison, <8 x i32> zeroinitializer
; CHECK-NEXT: [[VQNEGQ_V1_I:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.sqneg.v8i16(<8 x i16> [[VECINIT7_I]])
+; CHECK-NEXT: [[VBSL5_I:%.*]] = shufflevector <8 x i16> [[VQNEGQ_V1_I]], <8 x i16> [[VECINIT_I]], <8 x i32> <i32 0, i32 8, i32 2, i32 8, i32 4, i32 8, i32 6, i32 8>
; CHECK-NEXT: [[SHUFFLE_I85:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
; CHECK-NEXT: [[SHUFFLE_I82:%.*]] = shufflevector <8 x i16> [[VECINIT_I19]], <8 x i16> poison, <4 x i32> zeroinitializer
; CHECK-NEXT: [[VQDMULL_V2_I72:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I85]], <4 x i16> [[SHUFFLE_I82]])
; CHECK-NEXT: [[SHUFFLE_I97:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
; CHECK-NEXT: [[VQDMULL_V2_I:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I97]], <4 x i16> [[SHUFFLE_I82]])
-; CHECK-NEXT: [[SHUFFLE_I79:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 1, i32 0, i32 3, i32 2>
-; CHECK-NEXT: [[SHUFFLE_I76:%.*]] = shufflevector <8 x i16> [[VQNEGQ_V1_I]], <8 x i16> [[VECINIT_I]], <4 x i32> <i32 0, i32 8, i32 2, i32 8>
+; CHECK-NEXT: [[SHUFFLE_I79:%.*]] = shufflevector <8 x i16> [[SHUFFLE_I]], <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT: [[SHUFFLE_I76:%.*]] = shufflevector <8 x i16> [[VBSL5_I]], <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
; CHECK-NEXT: [[VQDMLAL2_I106:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I79]], <4 x i16> [[SHUFFLE_I76]])
; CHECK-NEXT: [[VQDMLAL_V3_I107:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqadd.v4i32(<4 x i32> [[VQDMULL_V2_I72]], <4 x i32> [[VQDMLAL2_I106]])
-; CHECK-NEXT: [[SHUFFLE_I91:%.*]] = shufflevector <8 x i16> [[A]], <8 x i16> poison, <4 x i32> <i32 5, i32 4, i32 7, i32 6>
-; CHECK-NEXT: [[SHUFFLE_I88:%.*]] = shufflevector <8 x i16> [[VQNEGQ_V1_I]], <8 x i16> [[VECINIT_I]], <4 x i32> <i32 4, i32 8, i32 6, i32 8>
+; CHECK-NEXT: [[SHUFFLE_I91:%.*]] = shufflevector <8 x i16> [[SHUFFLE_I]], <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; CHECK-NEXT: [[SHUFFLE_I88:%.*]] = shufflevector <8 x i16> [[VBSL5_I]], <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
; CHECK-NEXT: [[VQDMLAL2_I:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[SHUFFLE_I91]], <4 x i16> [[SHUFFLE_I88]])
; CHECK-NEXT: [[VQDMLAL_V3_I:%.*]] = tail call <4 x i32> @llvm.aarch64.neon.sqadd.v4i32(<4 x i32> [[VQDMULL_V2_I]], <4 x i32> [[VQDMLAL2_I]])
; CHECK-NEXT: [[TMP0:%.*]] = bitcast <4 x i32> [[VQDMLAL_V3_I107]] to <8 x i16>
|
RKSimon
approved these changes
Jan 8, 2025
Collaborator
RKSimon
left a comment
RKSimon
left a comment
There was a problem hiding this comment.
LGTM - some of the other 'OldCost' calculations could probably benefit from something similar.
Contributor
|
Hi, after this patch we're seeing a 6% improvement on RISC-V with one of the SPEC CPU 2017 benchmarks because we're able to fold away a particularly expensive shuffle: https://lnt.lukelau.me/db_default/v4/nts/profile/13/91/88 So just want to say thanks, nice find! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This allows it to produce a more accurate cost for the shuffle, using the more accurate calls to getShuffleCost in getInstructionCost. It helps fix some of the regressions from vector combine a little while ago, now that we have better subvector extract costs.