[SLP]Fix PR89988: do extra analysis of the icmp args to correctly han…

…dle signed/unsigned comparison. If operands of icmp has different signedness, need to consider extending unsigned operands to correctly handle comparison with the signed operands.
llvm · Apr 25, 2024 · f758bb6 · f758bb6
1 parent bef6687
commit f758bb6
Show file tree

Hide file tree

Showing 2 changed files with 11 additions and 4 deletions.
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -15072,11 +15072,16 @@ void BoUpSLP::computeMinimumValueSizes() {
       IsSignedCmp =
           NodeIdx < VectorizableTree.size() &&
           any_of(VectorizableTree[NodeIdx]->UserTreeIndices,
-                 [](const EdgeInfo &EI) {
+                 [&](const EdgeInfo &EI) {
                    return EI.UserTE->getOpcode() == Instruction::ICmp &&
-                          any_of(EI.UserTE->Scalars, [](Value *V) {
+                          any_of(EI.UserTE->Scalars, [&](Value *V) {
                             auto *IC = dyn_cast<ICmpInst>(V);
-                            return IC && IC->isSigned();
+                            return IC &&
+                                   (IC->isSigned() ||
+                                    !isKnownNonNegative(IC->getOperand(0),
+                                                        SimplifyQuery(*DL)) ||
+                                    !isKnownNonNegative(IC->getOperand(1),
+                                                        SimplifyQuery(*DL)));
                           });
                  });
     }

diff --git a/llvm/test/Transforms/SLPVectorizer/RISCV/unsigned-icmp-signed-op.ll b/llvm/test/Transforms/SLPVectorizer/RISCV/unsigned-icmp-signed-op.ll
@@ -8,7 +8,9 @@ define i32 @test(ptr %f, i16 %0) {
 ; CHECK-NEXT:    [[TMP1:%.*]] = load i16, ptr [[F]], align 2
 ; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <4 x i16> <i16 0, i16 poison, i16 0, i16 0>, i16 [[TMP0]], i32 1
 ; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <4 x i16> <i16 0, i16 poison, i16 0, i16 0>, i16 [[TMP1]], i32 1
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp ule <4 x i16> [[TMP3]], [[TMP2]]
+; CHECK-NEXT:    [[TMP6:%.*]] = zext <4 x i16> [[TMP3]] to <4 x i32>
+; CHECK-NEXT:    [[TMP7:%.*]] = sext <4 x i16> [[TMP2]] to <4 x i32>
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp ule <4 x i32> [[TMP6]], [[TMP7]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = call i1 @llvm.vector.reduce.and.v4i1(<4 x i1> [[TMP4]])
 ; CHECK-NEXT:    [[ZEXT_4:%.*]] = zext i1 [[TMP5]] to i32
 ; CHECK-NEXT:    ret i32 [[ZEXT_4]]