[llvm] 8a115b6 - [LV] Mark checks as never succeeding for high cost cutoff.

[Florian Hahn via llvm-commits]

Author: Florian Hahn
Date: 2025-12-09T20:37:21Z
New Revision: 8a115b6934a90441d77ea54af73e7aaaa1394b38

URL: https://github.com/llvm/llvm-project/commit/8a115b6934a90441d77ea54af73e7aaaa1394b38
DIFF: https://github.com/llvm/llvm-project/commit/8a115b6934a90441d77ea54af73e7aaaa1394b38.diff

LOG: [LV] Mark checks as never succeeding for high cost cutoff.

When GeneratedRTChecks::create bails out due to exceeding the cost
threshold, no runtime checks are generated and we must not proceed
assuming checks have been generated.

Mark the checks as never succeeding, to make sure we don't try to
vectorize assuming the runtime checks hold. This fixes a case where we
previously incorrectly vectorized assuming runtime checks had been
generated when forcing vectorization via metadate.

Fixes the mis-compile mentioned in
https://github.com/llvm/llvm-project/pull/166247#issuecomment-3631471588

Added: 
    

Modified: 
    llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
    llvm/test/Transforms/LoopVectorize/runtime-check-threshold-with-force-metadata.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 15d0fa41bd902..79cdae25e38da 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -1827,8 +1827,12 @@ class GeneratedRTChecks {
     // profile info.
     CostTooHigh =
         LAI.getNumRuntimePointerChecks() > VectorizeMemoryCheckThreshold;
-    if (CostTooHigh)
+    if (CostTooHigh) {
+      // Mark runtime checks as never succeeding when they exceed the threshold.
+      MemRuntimeCheckCond = ConstantInt::getTrue(L->getHeader()->getContext());
+      SCEVCheckCond = ConstantInt::getTrue(L->getHeader()->getContext());
       return;
+    }
 
     BasicBlock *LoopHeader = L->getHeader();
     BasicBlock *Preheader = L->getLoopPreheader();

diff  --git a/llvm/test/Transforms/LoopVectorize/runtime-check-threshold-with-force-metadata.ll b/llvm/test/Transforms/LoopVectorize/runtime-check-threshold-with-force-metadata.ll
index b7d36fe7928e5..5376eb86882b7 100644
--- a/llvm/test/Transforms/LoopVectorize/runtime-check-threshold-with-force-metadata.ll
+++ b/llvm/test/Transforms/LoopVectorize/runtime-check-threshold-with-force-metadata.ll
@@ -2,29 +2,23 @@
 ; RUN: opt -p loop-vectorize -vectorize-memory-check-threshold=0 -S %s | FileCheck --check-prefix=LIMIT0 %s
 ; RUN: opt -p loop-vectorize -vectorize-memory-check-threshold=1 -S %s | FileCheck --check-prefix=LIMIT1 %s
 
-; FIXME: Currently this miscompiles with -vectorize-memory-check-threshold=0;
-; no runtime check is generated even though one is needed and !noalias
-; annotations are added.
+; Make sure we do not incorrectly vectorize with -vectorize-memory-check-threshold=0;
+; no runtime check is generated and the loop should not be vectorized.
 define i16 @runtime_checks_needed(ptr %src, ptr %dst) {
 ; LIMIT0-LABEL: define i16 @runtime_checks_needed(
 ; LIMIT0-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]]) {
-; LIMIT0-NEXT:  [[ENTRY:.*:]]
-; LIMIT0-NEXT:    br label %[[VECTOR_PH:.*]]
-; LIMIT0:       [[VECTOR_PH]]:
-; LIMIT0-NEXT:    [[TMP0:%.*]] = load i16, ptr [[SRC]], align 1, !alias.scope [[META0:![0-9]+]]
-; LIMIT0-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i16> poison, i16 [[TMP0]], i64 0
-; LIMIT0-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i16> [[BROADCAST_SPLATINSERT]], <2 x i16> poison, <2 x i32> zeroinitializer
-; LIMIT0-NEXT:    br label %[[VECTOR_BODY:.*]]
-; LIMIT0:       [[VECTOR_BODY]]:
-; LIMIT0-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; LIMIT0-NEXT:  [[ENTRY:.*]]:
+; LIMIT0-NEXT:    br label %[[LOOP:.*]]
+; LIMIT0:       [[LOOP]]:
+; LIMIT0-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[INDEX_NEXT:%.*]], %[[LOOP]] ]
+; LIMIT0-NEXT:    [[L:%.*]] = load i16, ptr [[SRC]], align 1
 ; LIMIT0-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i16, ptr [[DST]], i64 [[INDEX]]
-; LIMIT0-NEXT:    store <2 x i16> [[BROADCAST_SPLAT]], ptr [[TMP1]], align 1, !alias.scope [[META3:![0-9]+]], !noalias [[META0]]
-; LIMIT0-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; LIMIT0-NEXT:    store i16 [[L]], ptr [[TMP1]], align 1
+; LIMIT0-NEXT:    [[INDEX_NEXT]] = add nuw nsw i64 [[INDEX]], 1
 ; LIMIT0-NEXT:    [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1000
-; LIMIT0-NEXT:    br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
-; LIMIT0:       [[MIDDLE_BLOCK]]:
-; LIMIT0-NEXT:    br label %[[EXIT:.*]]
+; LIMIT0-NEXT:    br i1 [[TMP2]], label %[[EXIT:.*]], label %[[LOOP]], !llvm.loop [[LOOP0:![0-9]+]]
 ; LIMIT0:       [[EXIT]]:
+; LIMIT0-NEXT:    [[TMP0:%.*]] = phi i16 [ [[L]], %[[LOOP]] ]
 ; LIMIT0-NEXT:    ret i16 [[TMP0]]
 ;
 ; LIMIT1-LABEL: define i16 @runtime_checks_needed(
@@ -88,14 +82,9 @@ exit:
 !3 = !{!"llvm.loop.vectorize.enable", i1 true}
 
 ;.
-; LIMIT0: [[META0]] = !{[[META1:![0-9]+]]}
-; LIMIT0: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]]}
-; LIMIT0: [[META2]] = distinct !{[[META2]], !"LVerDomain"}
-; LIMIT0: [[META3]] = !{[[META4:![0-9]+]]}
-; LIMIT0: [[META4]] = distinct !{[[META4]], [[META2]]}
-; LIMIT0: [[LOOP5]] = distinct !{[[LOOP5]], [[META6:![0-9]+]], [[META7:![0-9]+]]}
-; LIMIT0: [[META6]] = !{!"llvm.loop.isvectorized", i32 1}
-; LIMIT0: [[META7]] = !{!"llvm.loop.unroll.runtime.disable"}
+; LIMIT0: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; LIMIT0: [[META1]] = !{!"llvm.loop.vectorize.width", i32 2}
+; LIMIT0: [[META2]] = !{!"llvm.loop.vectorize.enable", i1 true}
 ;.
 ; LIMIT1: [[META0]] = !{[[META1:![0-9]+]]}
 ; LIMIT1: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]]}