[X86] combineEXTRACT_SUBVECTOR - fold extract_subvector(subv_broadcast_load(ptr),0) -> load(ptr) #126523
Conversation
This file contains 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
…t_load(ptr),0) -> load(ptr) This can typically be handled by SimplifyDemandedVectorElts, but this will fail when there are multiple uses of the subv_broadcast_load node, but if there's just one use of the load result (and the rest are uses of the memory chain), we can still replace with a load and update the chain accordingly. Noticed on llvm#126517
|
@llvm/pr-subscribers-backend-x86 Author: Simon Pilgrim (RKSimon) ChangesThis is typically handled by SimplifyDemandedVectorElts, but this will fail when there are multiple uses of the subv_broadcast_load node, but if there's just one use of the load result (and the rest are uses of the memory chain), we can still replace with a load and update the chain accordingly. Noticed on #126517 Full diff: https://github.com/llvm/llvm-project/pull/126523.diff 3 Files Affected:
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 995b4de12ce12c..601bf7bb6b817d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -58477,10 +58477,26 @@ static SDValue combineEXTRACT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
DAG.isSplatValue(InVec, /*AllowUndefs*/ false)))
return extractSubVector(InVec, 0, DAG, DL, SizeInBits);
- // If we're extracting a broadcasted subvector, just use the lowest subvector.
- if (IdxVal != 0 && InVec.getOpcode() == X86ISD::SUBV_BROADCAST_LOAD &&
- cast<MemIntrinsicSDNode>(InVec)->getMemoryVT() == VT)
- return extractSubVector(InVec, 0, DAG, DL, SizeInBits);
+ // Check if we're extracting a whole broadcasted subvector.
+ if (InVec.getOpcode() == X86ISD::SUBV_BROADCAST_LOAD) {
+ auto *MemIntr = cast<MemIntrinsicSDNode>(InVec);
+ EVT MemVT = MemIntr->getMemoryVT();
+ if (MemVT == VT) {
+ // Just use the lowest subvector.
+ if (IdxVal != 0)
+ return extractSubVector(InVec, 0, DAG, DL, SizeInBits);
+ // If this is the only use, we can replace with a regular load (this may
+ // have been missed by SimplifyDemandedVectorElts due to extra uses of the
+ // memory chain).
+ if (InVec.hasOneUse()) {
+ SDValue Ld =
+ DAG.getLoad(MemVT, DL, MemIntr->getChain(), MemIntr->getBasePtr(),
+ MemIntr->getMemOperand());
+ DAG.makeEquivalentMemoryOrdering(SDValue(MemIntr, 1), Ld.getValue(1));
+ return Ld;
+ }
+ }
+ }
// Attempt to extract from the source of a shuffle vector.
if ((InSizeInBits % SizeInBits) == 0 && (IdxVal % NumSubElts) == 0) {
diff --git a/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast_from_memory.ll b/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast_from_memory.ll
index 3d72319f59ca9e..e47a9ac3a0c0bf 100644
--- a/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast_from_memory.ll
+++ b/llvm/test/CodeGen/X86/any_extend_vector_inreg_of_broadcast_from_memory.ll
@@ -3634,19 +3634,18 @@ define void @vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3(ptr %i
;
; AVX-LABEL: vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3:
; AVX: # %bb.0:
-; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
-; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0],mem[1,2,3]
+; AVX-NEXT: vmovdqa (%rdi), %xmm0
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm0[0,1],mem[2,3,4,5,6,7]
; AVX-NEXT: vpaddb (%rsi), %xmm1, %xmm1
; AVX-NEXT: vmovdqa (%rdi), %xmm2
; AVX-NEXT: vmovdqa 16(%rdi), %xmm3
; AVX-NEXT: vpaddb 48(%rsi), %xmm3, %xmm3
-; AVX-NEXT: vpaddb 16(%rsi), %xmm0, %xmm0
; AVX-NEXT: vpaddb 32(%rsi), %xmm2, %xmm2
-; AVX-NEXT: vmovdqa %xmm2, 32(%rdx)
+; AVX-NEXT: vpaddb 16(%rsi), %xmm0, %xmm0
; AVX-NEXT: vmovdqa %xmm0, 16(%rdx)
+; AVX-NEXT: vmovdqa %xmm2, 32(%rdx)
; AVX-NEXT: vmovdqa %xmm3, 48(%rdx)
; AVX-NEXT: vmovdqa %xmm1, (%rdx)
-; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3:
@@ -3820,19 +3819,18 @@ define void @vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3(ptr %i
;
; AVX-LABEL: vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3:
; AVX: # %bb.0:
-; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
-; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],mem[2,3]
+; AVX-NEXT: vmovdqa (%rdi), %xmm0
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm0[0,1,2,3],mem[4,5,6,7]
; AVX-NEXT: vpaddb (%rsi), %xmm1, %xmm1
; AVX-NEXT: vmovdqa (%rdi), %xmm2
; AVX-NEXT: vmovdqa 16(%rdi), %xmm3
; AVX-NEXT: vpaddb 48(%rsi), %xmm3, %xmm3
-; AVX-NEXT: vpaddb 16(%rsi), %xmm0, %xmm0
; AVX-NEXT: vpaddb 32(%rsi), %xmm2, %xmm2
-; AVX-NEXT: vmovdqa %xmm2, 32(%rdx)
+; AVX-NEXT: vpaddb 16(%rsi), %xmm0, %xmm0
; AVX-NEXT: vmovdqa %xmm0, 16(%rdx)
+; AVX-NEXT: vmovdqa %xmm2, 32(%rdx)
; AVX-NEXT: vmovdqa %xmm3, 48(%rdx)
; AVX-NEXT: vmovdqa %xmm1, (%rdx)
-; AVX-NEXT: vzeroupper
; AVX-NEXT: retq
;
; AVX2-LABEL: vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3:
diff --git a/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast_from_memory.ll b/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast_from_memory.ll
index acedcf42639066..a3f134922ba3cc 100644
--- a/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast_from_memory.ll
+++ b/llvm/test/CodeGen/X86/zero_extend_vector_inreg_of_broadcast_from_memory.ll
@@ -4044,18 +4044,17 @@ define void @vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3(ptr %i
;
; AVX-LABEL: vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3:
; AVX: # %bb.0:
-; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
-; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0],mem[1,2,3]
+; AVX-NEXT: vmovdqa (%rdi), %xmm0
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm0[0,1],mem[2,3,4,5,6,7]
; AVX-NEXT: vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; AVX-NEXT: vxorps %xmm3, %xmm3, %xmm3
-; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm3[1,2,3]
-; AVX-NEXT: vpaddb 16(%rsi), %xmm0, %xmm0
; AVX-NEXT: vpaddb (%rsi), %xmm1, %xmm1
+; AVX-NEXT: vpxor %xmm3, %xmm3, %xmm3
+; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm3[2,3,4,5,6,7]
+; AVX-NEXT: vpaddb 16(%rsi), %xmm0, %xmm0
; AVX-NEXT: vpaddb 32(%rsi), %xmm2, %xmm2
; AVX-NEXT: vmovdqa %xmm2, 32(%rdx)
-; AVX-NEXT: vmovdqa %xmm1, (%rdx)
; AVX-NEXT: vmovdqa %xmm0, 16(%rdx)
-; AVX-NEXT: vzeroupper
+; AVX-NEXT: vmovdqa %xmm1, (%rdx)
; AVX-NEXT: retq
;
; AVX2-LABEL: vec384_i32_widen_to_i128_factor4_broadcast_to_v3i128_factor3:
@@ -4263,17 +4262,16 @@ define void @vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3(ptr %i
;
; AVX-LABEL: vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3:
; AVX: # %bb.0:
-; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
-; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],mem[2,3]
+; AVX-NEXT: vmovdqa (%rdi), %xmm0
+; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm0[0,1,2,3],mem[4,5,6,7]
; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
+; AVX-NEXT: vpaddb (%rsi), %xmm1, %xmm1
; AVX-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero
; AVX-NEXT: vpaddb 16(%rsi), %xmm0, %xmm0
-; AVX-NEXT: vpaddb (%rsi), %xmm1, %xmm1
; AVX-NEXT: vpaddb 32(%rsi), %xmm2, %xmm2
; AVX-NEXT: vmovdqa %xmm2, 32(%rdx)
-; AVX-NEXT: vmovdqa %xmm1, (%rdx)
; AVX-NEXT: vmovdqa %xmm0, 16(%rdx)
-; AVX-NEXT: vzeroupper
+; AVX-NEXT: vmovdqa %xmm1, (%rdx)
; AVX-NEXT: retq
;
; AVX2-LABEL: vec384_i64_widen_to_i128_factor2_broadcast_to_v3i128_factor3:
|
phoebewang
reviewed
Feb 11, 2025
alejandro-alvarez-sonarsource
pushed a commit
to alejandro-alvarez-sonarsource/llvm-project
that referenced
this pull request
Feb 12, 2025
…t_load(ptr),0) -> load(ptr) (llvm#126523) This is typically handled by SimplifyDemandedVectorElts, but this will fail when there are multiple uses of the subv_broadcast_load node, but if there's just one use of the load result (and the rest are uses of the memory chain), we can still replace with a load and update the chain accordingly. Noticed on llvm#126517
flovent
pushed a commit
to flovent/llvm-project
that referenced
this pull request
Feb 13, 2025
…t_load(ptr),0) -> load(ptr) (llvm#126523) This is typically handled by SimplifyDemandedVectorElts, but this will fail when there are multiple uses of the subv_broadcast_load node, but if there's just one use of the load result (and the rest are uses of the memory chain), we can still replace with a load and update the chain accordingly. Noticed on llvm#126517
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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 is typically handled by SimplifyDemandedVectorElts, but this will fail when there are multiple uses of the subv_broadcast_load node, but if there's just one use of the load result (and the rest are uses of the memory chain), we can still replace with a load and update the chain accordingly.
Noticed on #126517