[llvm-branch-commits] [llvm] release/23.x: [LV] Invalidate interleave-groups with gaps if for early-exit loops. (#237409) (PR #237895)

Nikita Popov via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Mon Nov 2 01:58:21 PST 2026


https://github.com/nikic created https://github.com/llvm/llvm-project/pull/237895

(cherry picked from commit ccce555818cd682aff8ae4b9d72eb8ed0044331c)

>From 3e9d3a7cf0b7db534ff633a47bc3d7906fae547f Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Fri, 30 Oct 2026 18:48:55 -0700
Subject: [PATCH] [LV] Invalidate interleave-groups with gaps if for early-exit
 loops. (#237409)

If the loop exits from a block other than the latch, the scalar
iteration reserved for the exit may leave the loop before reaching the
grouped accesses.

Invalidate interleave groups in the case, to avoid mis-compile.

Fixes https://github.com/llvm/llvm-project/issues/237192.

(cherry picked from commit ccce555818cd682aff8ae4b9d72eb8ed0044331c)
---
 .../Transforms/Vectorize/LoopVectorize.cpp    |   9 +-
 .../LoopVectorize/interleaved-accesses.ll     | 169 ++++++++++++++++++
 2 files changed, 177 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index a4aae2651ee4c5..d4cc8f3a717e2d 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7953,8 +7953,15 @@ bool LoopVectorizePass::processLoop(Loop *L) {
     UseInterleaved = IsInnerLoop && EnableInterleavedMemAccesses;
 
   // Analyze interleaved memory accesses.
-  if (UseInterleaved)
+  if (UseInterleaved) {
     IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI));
+    // If the loop exits from a block other than the latch, the scalar
+    // iteration reserved for the exit may leave the loop before reaching the
+    // grouped accesses.
+    if (L->getExitingBlock() != L->getLoopLatch() &&
+        !LVL.hasUncountableEarlyExit())
+      IAI.invalidateGroupsRequiringScalarEpilogue();
+  }
 
   if (LVL.hasUncountableEarlyExit()) {
     BasicBlock *LoopLatch = L->getLoopLatch();
diff --git a/llvm/test/Transforms/LoopVectorize/interleaved-accesses.ll b/llvm/test/Transforms/LoopVectorize/interleaved-accesses.ll
index 1a97a87d2edfa7..320ffbb937060b 100644
--- a/llvm/test/Transforms/LoopVectorize/interleaved-accesses.ll
+++ b/llvm/test/Transforms/LoopVectorize/interleaved-accesses.ll
@@ -1546,3 +1546,172 @@ loop:
 end:
   ret void
 }
+
+; The loop exits from the header, so the scalar iteration reserved for the
+; exit does not execute the load. An interleave group with a trailing gap
+; would read past the last accessed element in the final vector iteration.
+define void @gap_header_exit(ptr noalias %dst, ptr noalias %src, i64 %n) {
+; CHECK-LABEL: @gap_header_exit(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[END:%.*]] = getelementptr inbounds nuw [8 x i8], ptr [[SRC:%.*]], i64 [[N:%.*]]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[N]], 1
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 5
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 3
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp eq i64 [[TMP1]], 0
+; CHECK-NEXT:    [[TMP3:%.*]] = select i1 [[TMP2]], i64 4, i64 [[TMP1]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP0]], [[TMP3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = shl i64 [[N_VEC]], 3
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[TMP4]]
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP6:%.*]] = shl i64 [[INDEX]], 3
+; CHECK-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[TMP6]]
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[TMP6]]
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[TMP6]]
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[TMP6]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds nuw i8, ptr [[NEXT_GEP]], i64 4
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP9]], i64 12
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP7]], i64 20
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP8]], i64 28
+; CHECK-NEXT:    [[TMP14:%.*]] = load i32, ptr [[TMP10]], align 4
+; CHECK-NEXT:    [[TMP15:%.*]] = load i32, ptr [[TMP11]], align 4
+; CHECK-NEXT:    [[TMP16:%.*]] = load i32, ptr [[TMP12]], align 4
+; CHECK-NEXT:    [[TMP17:%.*]] = load i32, ptr [[TMP13]], align 4
+; CHECK-NEXT:    [[TMP18:%.*]] = insertelement <4 x i32> poison, i32 [[TMP14]], i64 0
+; CHECK-NEXT:    [[TMP19:%.*]] = insertelement <4 x i32> [[TMP18]], i32 [[TMP15]], i64 1
+; CHECK-NEXT:    [[TMP20:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP16]], i64 2
+; CHECK-NEXT:    [[TMP21:%.*]] = insertelement <4 x i32> [[TMP20]], i32 [[TMP17]], i64 3
+; CHECK-NEXT:    [[TMP22:%.*]] = getelementptr inbounds nuw [4 x i8], ptr [[DST:%.*]], i64 [[INDEX]]
+; CHECK-NEXT:    store <4 x i32> [[TMP21]], ptr [[TMP22]], align 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP23:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP23]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP34:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    br label [[SCALAR_PH]]
+; CHECK:       scalar.ph:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[BC_RESUME_VAL4:%.*]] = phi ptr [ [[TMP5]], [[MIDDLE_BLOCK]] ], [ [[SRC]], [[ENTRY]] ]
+; CHECK-NEXT:    br label [[HEADER:%.*]]
+; CHECK:       header:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[LATCH:%.*]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
+; CHECK-NEXT:    [[PTR:%.*]] = phi ptr [ [[PTR_NEXT:%.*]], [[LATCH]] ], [ [[BC_RESUME_VAL4]], [[SCALAR_PH]] ]
+; CHECK-NEXT:    [[DONE:%.*]] = icmp eq ptr [[PTR]], [[END]]
+; CHECK-NEXT:    br i1 [[DONE]], label [[EXIT:%.*]], label [[LATCH]]
+; CHECK:       latch:
+; CHECK-NEXT:    [[PTR_NEXT]] = getelementptr inbounds nuw i8, ptr [[PTR]], i64 8
+; CHECK-NEXT:    [[GEP_Y:%.*]] = getelementptr inbounds nuw i8, ptr [[PTR]], i64 4
+; CHECK-NEXT:    [[Y:%.*]] = load i32, ptr [[GEP_Y]], align 4
+; CHECK-NEXT:    [[GEP_DST:%.*]] = getelementptr inbounds nuw [4 x i8], ptr [[DST]], i64 [[IV]]
+; CHECK-NEXT:    store i32 [[Y]], ptr [[GEP_DST]], align 4
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    br label [[HEADER]], !llvm.loop [[LOOP35:![0-9]+]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %end = getelementptr inbounds nuw [8 x i8], ptr %src, i64 %n
+  br label %header
+
+header:
+  %iv = phi i64 [ %iv.next, %latch ], [ 0, %entry ]
+  %ptr = phi ptr [ %ptr.next, %latch ], [ %src, %entry ]
+  %done = icmp eq ptr %ptr, %end
+  br i1 %done, label %exit, label %latch
+
+latch:
+  %ptr.next = getelementptr inbounds nuw i8, ptr %ptr, i64 8
+  %gep.y = getelementptr inbounds nuw i8, ptr %ptr, i64 4
+  %y = load i32, ptr %gep.y, align 4
+  %gep.dst = getelementptr inbounds nuw [4 x i8], ptr %dst, i64 %iv
+  store i32 %y, ptr %gep.dst, align 4
+  %iv.next = add nuw nsw i64 %iv, 1
+  br label %header
+
+exit:
+  ret void
+}
+
+; Same as above, but with a countable early exit before the load in addition
+; to the latch exit.
+define void @gap_early_countable_exit(ptr noalias %dst, ptr noalias %src, i64 %n, i64 %m) {
+; CHECK-LABEL: @gap_early_countable_exit(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP1:%.*]] = freeze i64 [[TMP0:%.*]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[TMP1]], -1
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP8]], i64 [[M:%.*]])
+; CHECK-NEXT:    [[TMP3:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP3]], 5
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[TMP4:%.*]] = and i64 [[TMP3]], 3
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[TMP4]], 0
+; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 4, i64 [[TMP4]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP3]], [[TMP6]]
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[DOTSPLIT:%.*]] = getelementptr inbounds nuw [8 x i8], ptr [[SRC:%.*]], i64 [[INDEX]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds nuw i8, ptr [[DOTSPLIT]], i64 4
+; CHECK-NEXT:    [[TMP24:%.*]] = getelementptr inbounds nuw [8 x i8], ptr [[SRC]], i64 [[INDEX]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP24]], i64 12
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds nuw [8 x i8], ptr [[SRC]], i64 [[INDEX]]
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP7]], i64 20
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds nuw [8 x i8], ptr [[SRC]], i64 [[INDEX]]
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP9]], i64 28
+; CHECK-NEXT:    [[TMP14:%.*]] = load i32, ptr [[TMP10]], align 4
+; CHECK-NEXT:    [[TMP15:%.*]] = load i32, ptr [[TMP11]], align 4
+; CHECK-NEXT:    [[TMP16:%.*]] = load i32, ptr [[TMP12]], align 4
+; CHECK-NEXT:    [[TMP17:%.*]] = load i32, ptr [[TMP13]], align 4
+; CHECK-NEXT:    [[TMP18:%.*]] = insertelement <4 x i32> poison, i32 [[TMP14]], i64 0
+; CHECK-NEXT:    [[TMP19:%.*]] = insertelement <4 x i32> [[TMP18]], i32 [[TMP15]], i64 1
+; CHECK-NEXT:    [[TMP20:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP16]], i64 2
+; CHECK-NEXT:    [[TMP21:%.*]] = insertelement <4 x i32> [[TMP20]], i32 [[TMP17]], i64 3
+; CHECK-NEXT:    [[TMP22:%.*]] = getelementptr inbounds nuw [4 x i8], ptr [[DST:%.*]], i64 [[INDEX]]
+; CHECK-NEXT:    store <4 x i32> [[TMP21]], ptr [[TMP22]], align 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP23:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP23]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP36:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    br label [[SCALAR_PH]]
+; CHECK:       scalar.ph:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    br label [[HEADER:%.*]]
+; CHECK:       header:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[LATCH:%.*]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
+; CHECK-NEXT:    [[EC_EARLY:%.*]] = icmp eq i64 [[IV]], [[M]]
+; CHECK-NEXT:    br i1 [[EC_EARLY]], label [[EXIT:%.*]], label [[LATCH]]
+; CHECK:       latch:
+; CHECK-NEXT:    [[GEP_Y_SPLIT:%.*]] = getelementptr inbounds nuw [8 x i8], ptr [[SRC]], i64 [[IV]]
+; CHECK-NEXT:    [[GEP_Y:%.*]] = getelementptr inbounds nuw i8, ptr [[GEP_Y_SPLIT]], i64 4
+; CHECK-NEXT:    [[Y:%.*]] = load i32, ptr [[GEP_Y]], align 4
+; CHECK-NEXT:    [[GEP_DST:%.*]] = getelementptr inbounds nuw [4 x i8], ptr [[DST]], i64 [[IV]]
+; CHECK-NEXT:    store i32 [[Y]], ptr [[GEP_DST]], align 4
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[TMP1]]
+; CHECK-NEXT:    br i1 [[EC]], label [[EXIT]], label [[HEADER]], !llvm.loop [[LOOP37:![0-9]+]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %header
+
+header:
+  %iv = phi i64 [ %iv.next, %latch ], [ 0, %entry ]
+  %ec.early = icmp eq i64 %iv, %m
+  br i1 %ec.early, label %exit, label %latch
+
+latch:
+  %gep.y = getelementptr inbounds nuw [8 x i8], ptr %src, i64 %iv, i64 4
+  %y = load i32, ptr %gep.y, align 4
+  %gep.dst = getelementptr inbounds nuw [4 x i8], ptr %dst, i64 %iv
+  store i32 %y, ptr %gep.dst, align 4
+  %iv.next = add nuw nsw i64 %iv, 1
+  %ec = icmp eq i64 %iv.next, %n
+  br i1 %ec, label %exit, label %header
+
+exit:
+  ret void
+}



More information about the llvm-branch-commits mailing list