[llvm] 2ae4f5a - [SLP][NFC] Pre-commit tests for build-vector stores in store chains (#195531)
via llvm-commits
llvm-commits at lists.llvm.org
Sun May 3 08:49:38 PDT 2026
Author: Yaxun (Sam) Liu
Date: 2026-05-03T11:49:33-04:00
New Revision: 2ae4f5a702947c05cb2b3cf1c0cb3f953506ae8e
URL: https://github.com/llvm/llvm-project/commit/2ae4f5a702947c05cb2b3cf1c0cb3f953506ae8e
DIFF: https://github.com/llvm/llvm-project/commit/2ae4f5a702947c05cb2b3cf1c0cb3f953506ae8e.diff
LOG: [SLP][NFC] Pre-commit tests for build-vector stores in store chains (#195531)
These tests exercise SLP store-chain handling when an explicit
build-vector store (a chain of insertelements stored to memory) is
interleaved with scalar stores at adjacent addresses. The CHECK lines
reflect current behavior:
- buildvector_store_middle: vector store sits in the middle of the
chain.
SLP currently leaves it intact and packs the surrounding scalars into
a narrower <2 x float> + scalar tail mix.
- buildvector_store_start / buildvector_store_end: vector store sits at
the start/end of the chain. SLP already combines the adjacent scalar
stores into a clean <4 x float> store, so these are regression cases
that should remain unchanged.
- buildvector_store_duplicate_offset: scalar stores at the same address
bracket the vector store; SLP keeps the chain split.
A follow-up patch updates the CHECK lines for the cases that change.
Added:
llvm/test/Transforms/SLPVectorizer/X86/buildvector-store-chains.ll
Modified:
Removed:
################################################################################
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/buildvector-store-chains.ll b/llvm/test/Transforms/SLPVectorizer/X86/buildvector-store-chains.ll
new file mode 100644
index 0000000000000..ad55b6dd445c3
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/buildvector-store-chains.ll
@@ -0,0 +1,191 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s
+
+; Tests for store chains containing both scalar stores and a build-vector store
+; at adjacent addresses. The CHECK lines below capture current SLP behavior,
+; which leaves the build-vector store unchanged and packs the surrounding
+; scalars into a narrower mix. A follow-up patch updates these checks to
+; retile the chains as full vector stores.
+
+define void @buildvector_store_middle(ptr %p, float %a0, float %a1, float %a2, float %a3, float %a4, float %a5, float %a6, float %a7) {
+; CHECK-LABEL: define void @buildvector_store_middle(
+; CHECK-SAME: ptr [[P:%.*]], float [[A0:%.*]], float [[A1:%.*]], float [[A2:%.*]], float [[A3:%.*]], float [[A4:%.*]], float [[A5:%.*]], float [[A6:%.*]], float [[A7:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x float> poison, float [[A0]], i32 0
+; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x float> [[TMP0]], float [[A1]], i32 1
+; CHECK-NEXT: [[TMP2:%.*]] = fadd <2 x float> [[TMP1]], splat (float 1.000000e+00)
+; CHECK-NEXT: [[V2:%.*]] = fadd float [[A2]], 1.000000e+00
+; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> poison, float [[A3]], i32 0
+; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x float> [[TMP3]], float [[A4]], i32 1
+; CHECK-NEXT: [[TMP5:%.*]] = insertelement <4 x float> [[TMP4]], float [[A5]], i32 2
+; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x float> [[TMP5]], float [[A6]], i32 3
+; CHECK-NEXT: [[TMP7:%.*]] = fadd <4 x float> [[TMP6]], splat (float 1.000000e+00)
+; CHECK-NEXT: [[V7:%.*]] = fadd float [[A7]], 1.000000e+00
+; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[P]], align 4
+; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds float, ptr [[P]], i64 2
+; CHECK-NEXT: store float [[V2]], ptr [[P2]], align 4
+; CHECK-NEXT: [[P3:%.*]] = getelementptr inbounds float, ptr [[P]], i64 3
+; CHECK-NEXT: store <4 x float> [[TMP7]], ptr [[P3]], align 4
+; CHECK-NEXT: [[P7:%.*]] = getelementptr inbounds float, ptr [[P]], i64 7
+; CHECK-NEXT: store float [[V7]], ptr [[P7]], align 4
+; CHECK-NEXT: ret void
+;
+entry:
+ %v0 = fadd float %a0, 1.000000e+00
+ %v1 = fadd float %a1, 1.000000e+00
+ %v2 = fadd float %a2, 1.000000e+00
+ %v3 = fadd float %a3, 1.000000e+00
+ %v4 = fadd float %a4, 1.000000e+00
+ %v5 = fadd float %a5, 1.000000e+00
+ %v6 = fadd float %a6, 1.000000e+00
+ %v7 = fadd float %a7, 1.000000e+00
+ store float %v0, ptr %p, align 4
+ %p1 = getelementptr inbounds float, ptr %p, i64 1
+ store float %v1, ptr %p1, align 4
+ %p2 = getelementptr inbounds float, ptr %p, i64 2
+ store float %v2, ptr %p2, align 4
+ %p3 = getelementptr inbounds float, ptr %p, i64 3
+ %b0 = insertelement <4 x float> poison, float %v3, i32 0
+ %b1 = insertelement <4 x float> %b0, float %v4, i32 1
+ %b2 = insertelement <4 x float> %b1, float %v5, i32 2
+ %b3 = insertelement <4 x float> %b2, float %v6, i32 3
+ store <4 x float> %b3, ptr %p3, align 4
+ %p7 = getelementptr inbounds float, ptr %p, i64 7
+ store float %v7, ptr %p7, align 4
+ ret void
+}
+
+define void @buildvector_store_start(ptr %p, float %a0, float %a1, float %a2, float %a3, float %a4, float %a5, float %a6, float %a7) {
+; CHECK-LABEL: define void @buildvector_store_start(
+; CHECK-SAME: ptr [[P:%.*]], float [[A0:%.*]], float [[A1:%.*]], float [[A2:%.*]], float [[A3:%.*]], float [[A4:%.*]], float [[A5:%.*]], float [[A6:%.*]], float [[A7:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[TMP0:%.*]] = insertelement <4 x float> poison, float [[A0]], i32 0
+; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x float> [[TMP0]], float [[A1]], i32 1
+; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x float> [[TMP1]], float [[A2]], i32 2
+; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> [[TMP2]], float [[A3]], i32 3
+; CHECK-NEXT: [[TMP4:%.*]] = fadd <4 x float> [[TMP3]], splat (float 1.000000e+00)
+; CHECK-NEXT: [[TMP5:%.*]] = insertelement <4 x float> poison, float [[A4]], i32 0
+; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x float> [[TMP5]], float [[A5]], i32 1
+; CHECK-NEXT: [[TMP7:%.*]] = insertelement <4 x float> [[TMP6]], float [[A6]], i32 2
+; CHECK-NEXT: [[TMP8:%.*]] = insertelement <4 x float> [[TMP7]], float [[A7]], i32 3
+; CHECK-NEXT: [[TMP9:%.*]] = fadd <4 x float> [[TMP8]], splat (float 1.000000e+00)
+; CHECK-NEXT: store <4 x float> [[TMP4]], ptr [[P]], align 4
+; CHECK-NEXT: [[P4:%.*]] = getelementptr inbounds float, ptr [[P]], i64 4
+; CHECK-NEXT: store <4 x float> [[TMP9]], ptr [[P4]], align 4
+; CHECK-NEXT: ret void
+;
+entry:
+ %v0 = fadd float %a0, 1.000000e+00
+ %v1 = fadd float %a1, 1.000000e+00
+ %v2 = fadd float %a2, 1.000000e+00
+ %v3 = fadd float %a3, 1.000000e+00
+ %v4 = fadd float %a4, 1.000000e+00
+ %v5 = fadd float %a5, 1.000000e+00
+ %v6 = fadd float %a6, 1.000000e+00
+ %v7 = fadd float %a7, 1.000000e+00
+ %b0 = insertelement <4 x float> poison, float %v0, i32 0
+ %b1 = insertelement <4 x float> %b0, float %v1, i32 1
+ %b2 = insertelement <4 x float> %b1, float %v2, i32 2
+ %b3 = insertelement <4 x float> %b2, float %v3, i32 3
+ store <4 x float> %b3, ptr %p, align 4
+ %p4 = getelementptr inbounds float, ptr %p, i64 4
+ store float %v4, ptr %p4, align 4
+ %p5 = getelementptr inbounds float, ptr %p, i64 5
+ store float %v5, ptr %p5, align 4
+ %p6 = getelementptr inbounds float, ptr %p, i64 6
+ store float %v6, ptr %p6, align 4
+ %p7 = getelementptr inbounds float, ptr %p, i64 7
+ store float %v7, ptr %p7, align 4
+ ret void
+}
+
+define void @buildvector_store_end(ptr %p, float %a0, float %a1, float %a2, float %a3, float %a4, float %a5, float %a6, float %a7) {
+; CHECK-LABEL: define void @buildvector_store_end(
+; CHECK-SAME: ptr [[P:%.*]], float [[A0:%.*]], float [[A1:%.*]], float [[A2:%.*]], float [[A3:%.*]], float [[A4:%.*]], float [[A5:%.*]], float [[A6:%.*]], float [[A7:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[TMP0:%.*]] = insertelement <4 x float> poison, float [[A0]], i32 0
+; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x float> [[TMP0]], float [[A1]], i32 1
+; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x float> [[TMP1]], float [[A2]], i32 2
+; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> [[TMP2]], float [[A3]], i32 3
+; CHECK-NEXT: [[TMP4:%.*]] = fadd <4 x float> [[TMP3]], splat (float 1.000000e+00)
+; CHECK-NEXT: [[TMP5:%.*]] = insertelement <4 x float> poison, float [[A4]], i32 0
+; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x float> [[TMP5]], float [[A5]], i32 1
+; CHECK-NEXT: [[TMP7:%.*]] = insertelement <4 x float> [[TMP6]], float [[A6]], i32 2
+; CHECK-NEXT: [[TMP8:%.*]] = insertelement <4 x float> [[TMP7]], float [[A7]], i32 3
+; CHECK-NEXT: [[TMP9:%.*]] = fadd <4 x float> [[TMP8]], splat (float 1.000000e+00)
+; CHECK-NEXT: store <4 x float> [[TMP4]], ptr [[P]], align 4
+; CHECK-NEXT: [[P4:%.*]] = getelementptr inbounds float, ptr [[P]], i64 4
+; CHECK-NEXT: store <4 x float> [[TMP9]], ptr [[P4]], align 4
+; CHECK-NEXT: ret void
+;
+entry:
+ %v0 = fadd float %a0, 1.000000e+00
+ %v1 = fadd float %a1, 1.000000e+00
+ %v2 = fadd float %a2, 1.000000e+00
+ %v3 = fadd float %a3, 1.000000e+00
+ %v4 = fadd float %a4, 1.000000e+00
+ %v5 = fadd float %a5, 1.000000e+00
+ %v6 = fadd float %a6, 1.000000e+00
+ %v7 = fadd float %a7, 1.000000e+00
+ store float %v0, ptr %p, align 4
+ %p1 = getelementptr inbounds float, ptr %p, i64 1
+ store float %v1, ptr %p1, align 4
+ %p2 = getelementptr inbounds float, ptr %p, i64 2
+ store float %v2, ptr %p2, align 4
+ %p3 = getelementptr inbounds float, ptr %p, i64 3
+ store float %v3, ptr %p3, align 4
+ %p4 = getelementptr inbounds float, ptr %p, i64 4
+ %b0 = insertelement <4 x float> poison, float %v4, i32 0
+ %b1 = insertelement <4 x float> %b0, float %v5, i32 1
+ %b2 = insertelement <4 x float> %b1, float %v6, i32 2
+ %b3 = insertelement <4 x float> %b2, float %v7, i32 3
+ store <4 x float> %b3, ptr %p4, align 4
+ ret void
+}
+
+define void @buildvector_store_duplicate_offset(ptr %p, float %a0, float %a1, float %a2, float %a3, float %a4, float %a5, float %a6, float %a7) {
+; CHECK-LABEL: define void @buildvector_store_duplicate_offset(
+; CHECK-SAME: ptr [[P:%.*]], float [[A0:%.*]], float [[A1:%.*]], float [[A2:%.*]], float [[A3:%.*]], float [[A4:%.*]], float [[A5:%.*]], float [[A6:%.*]], float [[A7:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x float> poison, float [[A0]], i32 0
+; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x float> [[TMP0]], float [[A1]], i32 1
+; CHECK-NEXT: [[TMP2:%.*]] = fadd <2 x float> [[TMP1]], splat (float 1.000000e+00)
+; CHECK-NEXT: [[V2:%.*]] = fadd float [[A2]], 1.000000e+00
+; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> poison, float [[A3]], i32 0
+; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x float> [[TMP3]], float [[A4]], i32 1
+; CHECK-NEXT: [[TMP5:%.*]] = insertelement <4 x float> [[TMP4]], float [[A5]], i32 2
+; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x float> [[TMP5]], float [[A6]], i32 3
+; CHECK-NEXT: [[TMP7:%.*]] = fadd <4 x float> [[TMP6]], splat (float 1.000000e+00)
+; CHECK-NEXT: [[V7:%.*]] = fadd float [[A7]], 1.000000e+00
+; CHECK-NEXT: [[P1:%.*]] = getelementptr inbounds float, ptr [[P]], i64 1
+; CHECK-NEXT: store <2 x float> [[TMP2]], ptr [[P]], align 4
+; CHECK-NEXT: store float [[V2]], ptr [[P1]], align 4
+; CHECK-NEXT: [[P2:%.*]] = getelementptr inbounds float, ptr [[P]], i64 2
+; CHECK-NEXT: store <4 x float> [[TMP7]], ptr [[P2]], align 4
+; CHECK-NEXT: [[P6:%.*]] = getelementptr inbounds float, ptr [[P]], i64 6
+; CHECK-NEXT: store float [[V7]], ptr [[P6]], align 4
+; CHECK-NEXT: ret void
+;
+entry:
+ %v0 = fadd float %a0, 1.000000e+00
+ %v1 = fadd float %a1, 1.000000e+00
+ %v2 = fadd float %a2, 1.000000e+00
+ %v3 = fadd float %a3, 1.000000e+00
+ %v4 = fadd float %a4, 1.000000e+00
+ %v5 = fadd float %a5, 1.000000e+00
+ %v6 = fadd float %a6, 1.000000e+00
+ %v7 = fadd float %a7, 1.000000e+00
+ store float %v0, ptr %p, align 4
+ %p1 = getelementptr inbounds float, ptr %p, i64 1
+ store float %v1, ptr %p1, align 4
+ store float %v2, ptr %p1, align 4
+ %p2 = getelementptr inbounds float, ptr %p, i64 2
+ %b0 = insertelement <4 x float> poison, float %v3, i32 0
+ %b1 = insertelement <4 x float> %b0, float %v4, i32 1
+ %b2 = insertelement <4 x float> %b1, float %v5, i32 2
+ %b3 = insertelement <4 x float> %b2, float %v6, i32 3
+ store <4 x float> %b3, ptr %p2, align 4
+ %p6 = getelementptr inbounds float, ptr %p, i64 6
+ store float %v7, ptr %p6, align 4
+ ret void
+}
More information about the llvm-commits
mailing list