[SLP][REVEC] Fix CreateInsertElement does not use the correct result if MinBWs applied. (#104558)

Author: HanKuanChen

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -11977,7 +11977,7 @@ Value *BoUpSLP::gather(ArrayRef<Value *> VL, Value *Root, Type *ScalarTy) {
     if (auto *VecTy = dyn_cast<FixedVectorType>(Scalar->getType())) {
       assert(SLPReVec && "FixedVectorType is not expected.");
       Vec = InsElt = Builder.CreateInsertVector(
-          Vec->getType(), Vec, V,
+          Vec->getType(), Vec, Scalar,
           Builder.getInt64(Pos * VecTy->getNumElements()));
       auto *II = dyn_cast<IntrinsicInst>(InsElt);
       if (!II || II->getIntrinsicID() != Intrinsic::vector_insert)

llvm/test/Transforms/SLPVectorizer/revec.ll

@@ -336,3 +336,31 @@ entry:
   store <4 x i32> %vmovl.i108, ptr %add.ptr35, align 4
   ret void
 }
+
+define void @test11(<2 x i64> %0, i64 %1, <2 x i64> %2) {
+; CHECK-LABEL: @test11(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <2 x i64> [[TMP0:%.*]], i64 [[TMP1:%.*]], i32 1
+; CHECK-NEXT:    [[TMP4:%.*]] = add <2 x i64> <i64 5, i64 0>, [[TMP2:%.*]]
+; CHECK-NEXT:    [[TMP5:%.*]] = trunc <2 x i64> [[TMP4]] to <2 x i16>
+; CHECK-NEXT:    [[TMP6:%.*]] = call <4 x i16> @llvm.vector.insert.v4i16.v2i16(<4 x i16> poison, <2 x i16> [[TMP5]], i64 0)
+; CHECK-NEXT:    [[TMP7:%.*]] = trunc <2 x i64> [[TMP3]] to <2 x i16>
+; CHECK-NEXT:    [[TMP8:%.*]] = call <4 x i16> @llvm.vector.insert.v4i16.v2i16(<4 x i16> [[TMP6]], <2 x i16> [[TMP7]], i64 2)
+; CHECK-NEXT:    [[TMP9:%.*]] = trunc <4 x i16> [[TMP8]] to <4 x i8>
+; CHECK-NEXT:    [[TMP10:%.*]] = call <4 x i8> @llvm.vector.insert.v4i8.v2i8(<4 x i8> poison, <2 x i8> zeroinitializer, i64 0)
+; CHECK-NEXT:    [[TMP11:%.*]] = call <4 x i8> @llvm.vector.insert.v4i8.v2i8(<4 x i8> [[TMP10]], <2 x i8> zeroinitializer, i64 2)
+; CHECK-NEXT:    [[TMP12:%.*]] = urem <4 x i8> [[TMP9]], [[TMP11]]
+; CHECK-NEXT:    [[TMP13:%.*]] = icmp ne <4 x i8> [[TMP12]], [[TMP11]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %3 = insertelement <2 x i64> %0, i64 %1, i32 1
+  %4 = add <2 x i64> <i64 5, i64 0>, %2
+  %5 = trunc <2 x i64> %3 to <2 x i8>
+  %6 = trunc <2 x i64> %4 to <2 x i8>
+  %7 = urem <2 x i8> %5, zeroinitializer
+  %8 = urem <2 x i8> %6, zeroinitializer
+  %9 = icmp ne <2 x i8> %7, zeroinitializer
+  %10 = icmp ne <2 x i8> %8, zeroinitializer
+  ret void
+}