[LV] Preserve GEP nusw when widening memory #160885

artagnon · 2025-09-26T13:39:05Z

No description provided.

llvmbot · 2025-09-26T13:39:39Z

@llvm/pr-subscribers-vectorizers

@llvm/pr-subscribers-llvm-transforms

Author: Ramkumar Ramachandra (artagnon)

Changes

Full diff: https://github.com/llvm/llvm-project/pull/160885.diff

2 Files Affected:

(modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+3-4)
(added) llvm/test/Transforms/LoopVectorize/reverse-induction-gep-nowrap-flags.ll (+189)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index a2d61d689397b..699eb0fb2abd1 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7482,12 +7482,11 @@ VPRecipeBuilder::tryToWidenMemory(Instruction *I, ArrayRef<VPValue *> Operands,
     VPSingleDefRecipe *VectorPtr;
     if (Reverse) {
       // When folding the tail, we may compute an address that we don't in the
-      // original scalar loop and it may not be inbounds. Drop Inbounds in that
-      // case.
+      // original scalar loop: use none in this case.
       GEPNoWrapFlags Flags =
-          (CM.foldTailByMasking() || !GEP || !GEP->isInBounds())
+          CM.foldTailByMasking() || !GEP
               ? GEPNoWrapFlags::none()
-              : GEPNoWrapFlags::inBounds();
+              : GEP->getNoWrapFlags().withoutNoUnsignedWrap();
       VectorPtr =
           new VPVectorEndPointerRecipe(Ptr, &Plan.getVF(), getLoadStoreType(I),
                                        /*Stride*/ -1, Flags, I->getDebugLoc());
diff --git a/llvm/test/Transforms/LoopVectorize/reverse-induction-gep-nowrap-flags.ll b/llvm/test/Transforms/LoopVectorize/reverse-induction-gep-nowrap-flags.ll
new file mode 100644
index 0000000000000..553e56926dcfa
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/reverse-induction-gep-nowrap-flags.ll
@@ -0,0 +1,189 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt < %s -passes=loop-vectorize -force-vector-width=4 -S | FileCheck %s
+
+define i32 @preserve_inbounds(i64 %start, ptr %ptr) {
+; CHECK-LABEL: define i32 @preserve_inbounds(
+; CHECK-SAME: i64 [[START:%.*]], ptr [[PTR:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP4:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = sub i64 [[START]], [[INDEX]]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], -1
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[PTR]], i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 0
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 -3
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP3]], align 4
+; CHECK-NEXT:    [[REVERSE:%.*]] = shufflevector <4 x i32> [[WIDE_LOAD]], <4 x i32> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+; CHECK-NEXT:    [[TMP4]] = add <4 x i32> [[REVERSE]], [[VEC_PHI]]
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
+; CHECK-NEXT:    br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    br label %[[END:.*]]
+; CHECK:       [[SCALAR_PH:.*]]:
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REV_IND:%.*]] = phi i64 [ [[START]], %[[SCALAR_PH]] ], [ [[REV_IND_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REDUX:%.*]] = phi i32 [ 0, %[[SCALAR_PH]] ], [ [[REDUX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REV_IND_NEXT]] = add i64 [[REV_IND]], -1
+; CHECK-NEXT:    [[GEP_PTR_IND:%.*]] = getelementptr inbounds i32, ptr [[PTR]], i64 [[REV_IND_NEXT]]
+; CHECK-NEXT:    [[LD_PTR:%.*]] = load i32, ptr [[GEP_PTR_IND]], align 4
+; CHECK-NEXT:    [[REDUX_NEXT]] = add i32 [[LD_PTR]], [[REDUX]]
+; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 1
+; CHECK-NEXT:    [[EXIT_COND:%.*]] = icmp ne i32 [[IV_NEXT]], 1024
+; CHECK-NEXT:    br i1 [[EXIT_COND]], label %[[LOOP]], label %[[END]]
+; CHECK:       [[END]]:
+; CHECK-NEXT:    [[REDUX_NEXT_LCSSA:%.*]] = phi i32 [ [[REDUX_NEXT]], %[[LOOP]] ], [ [[TMP6]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    ret i32 [[REDUX_NEXT_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
+  %rev.ind = phi i64 [ %start, %entry ], [ %rev.ind.next, %loop ]
+  %redux = phi i32 [ 0, %entry ], [ %redux.next, %loop ]
+  %rev.ind.next = add i64 %rev.ind, -1
+  %gep.ptr.ind = getelementptr inbounds i32, ptr %ptr, i64 %rev.ind.next
+  %ld.ptr = load i32, ptr %gep.ptr.ind, align 4
+  %redux.next = add i32 %ld.ptr, %redux
+  %iv.next = add i32 %iv, 1
+  %exit.cond = icmp ne i32 %iv.next, 1024
+  br i1 %exit.cond, label %loop, label %end
+
+end:
+  ret i32 %redux.next
+}
+
+define i32 @preserve_nusw(i64 %start, ptr %ptr) {
+; CHECK-LABEL: define i32 @preserve_nusw(
+; CHECK-SAME: i64 [[START:%.*]], ptr [[PTR:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP4:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = sub i64 [[START]], [[INDEX]]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], -1
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr nusw i32, ptr [[PTR]], i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr nusw i32, ptr [[TMP1]], i32 0
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr nusw i32, ptr [[TMP2]], i32 -3
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP3]], align 4
+; CHECK-NEXT:    [[REVERSE:%.*]] = shufflevector <4 x i32> [[WIDE_LOAD]], <4 x i32> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+; CHECK-NEXT:    [[TMP4]] = add <4 x i32> [[REVERSE]], [[VEC_PHI]]
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
+; CHECK-NEXT:    br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    br label %[[END:.*]]
+; CHECK:       [[SCALAR_PH:.*]]:
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REV_IND:%.*]] = phi i64 [ [[START]], %[[SCALAR_PH]] ], [ [[REV_IND_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REDUX:%.*]] = phi i32 [ 0, %[[SCALAR_PH]] ], [ [[REDUX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REV_IND_NEXT]] = add i64 [[REV_IND]], -1
+; CHECK-NEXT:    [[GEP_PTR_IND:%.*]] = getelementptr nusw i32, ptr [[PTR]], i64 [[REV_IND_NEXT]]
+; CHECK-NEXT:    [[LD_PTR:%.*]] = load i32, ptr [[GEP_PTR_IND]], align 4
+; CHECK-NEXT:    [[REDUX_NEXT]] = add i32 [[LD_PTR]], [[REDUX]]
+; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 1
+; CHECK-NEXT:    [[EXIT_COND:%.*]] = icmp ne i32 [[IV_NEXT]], 1024
+; CHECK-NEXT:    br i1 [[EXIT_COND]], label %[[LOOP]], label %[[END]]
+; CHECK:       [[END]]:
+; CHECK-NEXT:    [[REDUX_NEXT_LCSSA:%.*]] = phi i32 [ [[REDUX_NEXT]], %[[LOOP]] ], [ [[TMP6]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    ret i32 [[REDUX_NEXT_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
+  %rev.ind = phi i64 [ %start, %entry ], [ %rev.ind.next, %loop ]
+  %redux = phi i32 [ 0, %entry ], [ %redux.next, %loop ]
+  %rev.ind.next = add i64 %rev.ind, -1
+  %gep.ptr.ind = getelementptr nusw i32, ptr %ptr, i64 %rev.ind.next
+  %ld.ptr = load i32, ptr %gep.ptr.ind, align 4
+  %redux.next = add i32 %ld.ptr, %redux
+  %iv.next = add i32 %iv, 1
+  %exit.cond = icmp ne i32 %iv.next, 1024
+  br i1 %exit.cond, label %loop, label %end
+
+end:
+  ret i32 %redux.next
+}
+
+define i32 @drop_nuw(i64 %start, ptr %ptr) {
+; CHECK-LABEL: define i32 @drop_nuw(
+; CHECK-SAME: i64 [[START:%.*]], ptr [[PTR:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP4:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = sub i64 [[START]], [[INDEX]]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], -1
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr nuw i32, ptr [[PTR]], i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr i32, ptr [[TMP1]], i32 0
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr i32, ptr [[TMP2]], i32 -3
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP3]], align 4
+; CHECK-NEXT:    [[REVERSE:%.*]] = shufflevector <4 x i32> [[WIDE_LOAD]], <4 x i32> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+; CHECK-NEXT:    [[TMP4]] = add <4 x i32> [[REVERSE]], [[VEC_PHI]]
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
+; CHECK-NEXT:    br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    br label %[[END:.*]]
+; CHECK:       [[SCALAR_PH:.*]]:
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REV_IND:%.*]] = phi i64 [ [[START]], %[[SCALAR_PH]] ], [ [[REV_IND_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REDUX:%.*]] = phi i32 [ 0, %[[SCALAR_PH]] ], [ [[REDUX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[REV_IND_NEXT]] = add i64 [[REV_IND]], -1
+; CHECK-NEXT:    [[GEP_PTR_IND:%.*]] = getelementptr nuw i32, ptr [[PTR]], i64 [[REV_IND_NEXT]]
+; CHECK-NEXT:    [[LD_PTR:%.*]] = load i32, ptr [[GEP_PTR_IND]], align 4
+; CHECK-NEXT:    [[REDUX_NEXT]] = add i32 [[LD_PTR]], [[REDUX]]
+; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 1
+; CHECK-NEXT:    [[EXIT_COND:%.*]] = icmp ne i32 [[IV_NEXT]], 1024
+; CHECK-NEXT:    br i1 [[EXIT_COND]], label %[[LOOP]], label %[[END]]
+; CHECK:       [[END]]:
+; CHECK-NEXT:    [[REDUX_NEXT_LCSSA:%.*]] = phi i32 [ [[REDUX_NEXT]], %[[LOOP]] ], [ [[TMP6]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    ret i32 [[REDUX_NEXT_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
+  %rev.ind = phi i64 [ %start, %entry ], [ %rev.ind.next, %loop ]
+  %redux = phi i32 [ 0, %entry ], [ %redux.next, %loop ]
+  %rev.ind.next = add i64 %rev.ind, -1
+  %gep.ptr.ind = getelementptr nuw i32, ptr %ptr, i64 %rev.ind.next
+  %ld.ptr = load i32, ptr %gep.ptr.ind, align 4
+  %redux.next = add i32 %ld.ptr, %redux
+  %iv.next = add i32 %iv, 1
+  %exit.cond = icmp ne i32 %iv.next, 1024
+  br i1 %exit.cond, label %loop, label %end
+
+end:
+  ret i32 %redux.next
+}
+;.
+; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]], [[META2]]}
+; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}
+;.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

lukel97

LGTM, but I'm no GEP expert so definitely wait for a second review :)

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

llvm/test/Transforms/LoopVectorize/reverse-induction-gep-nowrap-flags.ll

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

fhahn

LGTM, thanks!

artagnon requested review from fhahn and lukel97 September 26, 2025 13:39

llvmbot added vectorizers llvm:transforms labels Sep 26, 2025

lukel97 reviewed Sep 26, 2025

View reviewed changes

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp Show resolved Hide resolved

lukel97 approved these changes Sep 26, 2025

View reviewed changes

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp Outdated Show resolved Hide resolved

fhahn reviewed Sep 29, 2025

View reviewed changes

llvm/test/Transforms/LoopVectorize/reverse-induction-gep-nowrap-flags.ll Outdated Show resolved Hide resolved

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp Show resolved Hide resolved

artagnon force-pushed the lv-widenmem-gepnw branch from ec76860 to 2a23fca Compare September 29, 2025 11:01

fhahn approved these changes Sep 29, 2025

View reviewed changes

artagnon added 4 commits September 30, 2025 11:06

[LV] Pre-commit reverse-ind-gep-nowrap test

194360c

[LV] Preserve GEP nusw when widening memory

4d12229

[LV] Fix nit in comment

f63a025

[LV] Address review

30f7797

artagnon force-pushed the lv-widenmem-gepnw branch from 2a23fca to 30f7797 Compare September 30, 2025 10:06

artagnon enabled auto-merge (squash) September 30, 2025 10:07

artagnon merged commit 2f7252a into llvm:main Sep 30, 2025
9 checks passed

mahesh-attarde pushed a commit to mahesh-attarde/llvm-project that referenced this pull request Oct 3, 2025

[LV] Preserve GEP nusw when widening memory (llvm#160885)

85591c2

aemerson mentioned this pull request Oct 16, 2025

[clang-reorder-fields] Check for flexible array member (#160262) #163697

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[LV] Preserve GEP nusw when widening memory #160885

[LV] Preserve GEP nusw when widening memory #160885

Uh oh!

artagnon commented Sep 26, 2025

Uh oh!

llvmbot commented Sep 26, 2025 •

edited

Loading

Uh oh!

Uh oh!

lukel97 left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

fhahn left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[LV] Preserve GEP nusw when widening memory #160885

[LV] Preserve GEP nusw when widening memory #160885

Uh oh!

Conversation

artagnon commented Sep 26, 2025

Uh oh!

llvmbot commented Sep 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

lukel97 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

fhahn left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

llvmbot commented Sep 26, 2025 •

edited

Loading