[llvm] 7a5e2c9 - [LV] Add test with threshold=0 and metadata forcing vectorization.
Florian Hahn via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 9 12:06:56 PST 2025
Author: Florian Hahn
Date: 2025-12-09T20:06:38Z
New Revision: 7a5e2c9358eabff3d9eb66141590ac453a2e0e08
URL: https://github.com/llvm/llvm-project/commit/7a5e2c9358eabff3d9eb66141590ac453a2e0e08
DIFF: https://github.com/llvm/llvm-project/commit/7a5e2c9358eabff3d9eb66141590ac453a2e0e08.diff
LOG: [LV] Add test with threshold=0 and metadata forcing vectorization.
Test case for the mis-compile mentioned in
https://github.com/llvm/llvm-project/pull/166247#issuecomment-3631471588
The issue is that we don't generate a runtime check even though it is
required to vectorize.
Added:
llvm/test/Transforms/LoopVectorize/runtime-check-threshold-with-force-metadata.ll
Modified:
Removed:
################################################################################
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
new file mode 100644
index 0000000000000..b7d36fe7928e5
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/runtime-check-threshold-with-force-metadata.ll
@@ -0,0 +1,109 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; 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.
+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: [[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: [[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: [[EXIT]]:
+; LIMIT0-NEXT: ret i16 [[TMP0]]
+;
+; LIMIT1-LABEL: define i16 @runtime_checks_needed(
+; LIMIT1-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]]) {
+; LIMIT1-NEXT: [[ENTRY:.*:]]
+; LIMIT1-NEXT: br label %[[VECTOR_MEMCHECK:.*]]
+; LIMIT1: [[VECTOR_MEMCHECK]]:
+; LIMIT1-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[DST]], i64 2000
+; LIMIT1-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[SRC]], i64 2
+; LIMIT1-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[DST]], [[SCEVGEP1]]
+; LIMIT1-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[SRC]], [[SCEVGEP]]
+; LIMIT1-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
+; LIMIT1-NEXT: br i1 [[FOUND_CONFLICT]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; LIMIT1: [[VECTOR_PH]]:
+; LIMIT1-NEXT: [[TMP0:%.*]] = load i16, ptr [[SRC]], align 1, !alias.scope [[META0:![0-9]+]]
+; LIMIT1-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i16> poison, i16 [[TMP0]], i64 0
+; LIMIT1-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i16> [[BROADCAST_SPLATINSERT]], <2 x i16> poison, <2 x i32> zeroinitializer
+; LIMIT1-NEXT: br label %[[VECTOR_BODY:.*]]
+; LIMIT1: [[VECTOR_BODY]]:
+; LIMIT1-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; LIMIT1-NEXT: [[TMP1:%.*]] = getelementptr inbounds i16, ptr [[DST]], i64 [[INDEX]]
+; LIMIT1-NEXT: store <2 x i16> [[BROADCAST_SPLAT]], ptr [[TMP1]], align 1, !alias.scope [[META3:![0-9]+]], !noalias [[META0]]
+; LIMIT1-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; LIMIT1-NEXT: [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1000
+; LIMIT1-NEXT: br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
+; LIMIT1: [[MIDDLE_BLOCK]]:
+; LIMIT1-NEXT: br label %[[EXIT:.*]]
+; LIMIT1: [[SCALAR_PH]]:
+; LIMIT1-NEXT: br label %[[LOOP:.*]]
+; LIMIT1: [[LOOP]]:
+; LIMIT1-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; LIMIT1-NEXT: [[L:%.*]] = load i16, ptr [[SRC]], align 1
+; LIMIT1-NEXT: [[GEP_DST:%.*]] = getelementptr inbounds i16, ptr [[DST]], i64 [[IV]]
+; LIMIT1-NEXT: store i16 [[L]], ptr [[GEP_DST]], align 1
+; LIMIT1-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; LIMIT1-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], 1000
+; LIMIT1-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP8:![0-9]+]]
+; LIMIT1: [[EXIT]]:
+; LIMIT1-NEXT: [[L_LCSSA:%.*]] = phi i16 [ [[L]], %[[LOOP]] ], [ [[TMP0]], %[[MIDDLE_BLOCK]] ]
+; LIMIT1-NEXT: ret i16 [[L_LCSSA]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %l = load i16, ptr %src, align 1
+ %gep.dst = getelementptr inbounds i16, ptr %dst, i64 %iv
+ store i16 %l, ptr %gep.dst, align 1
+ %iv.next = add nsw nuw i64 %iv, 1
+ %ec = icmp eq i64 %iv.next, 1000
+ br i1 %ec, label %exit, label %loop, !llvm.loop !0
+
+exit:
+ ret i16 %l
+}
+
+!0 = distinct !{!0, !2, !3}
+!1 = !{!"llvm.loop.mustprogress"}
+!2 = !{!"llvm.loop.vectorize.width", i32 2}
+!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"}
+;.
+; LIMIT1: [[META0]] = !{[[META1:![0-9]+]]}
+; LIMIT1: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]]}
+; LIMIT1: [[META2]] = distinct !{[[META2]], !"LVerDomain"}
+; LIMIT1: [[META3]] = !{[[META4:![0-9]+]]}
+; LIMIT1: [[META4]] = distinct !{[[META4]], [[META2]]}
+; LIMIT1: [[LOOP5]] = distinct !{[[LOOP5]], [[META6:![0-9]+]], [[META7:![0-9]+]]}
+; LIMIT1: [[META6]] = !{!"llvm.loop.isvectorized", i32 1}
+; LIMIT1: [[META7]] = !{!"llvm.loop.unroll.runtime.disable"}
+; LIMIT1: [[LOOP8]] = distinct !{[[LOOP8]], [[META6]]}
+;.
More information about the llvm-commits
mailing list