[SLP][REVEC] Fix type comparison and mask transformation for REVEC.

When REVEC is enabled, ScalarTy may be a FixedVectorType. Compare its
element type to decide if casting is needed. Also apply mask
transformation accordingly.

Author: HanKuanChen

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -16998,7 +16998,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
         });
       return IsSigned;
     };
-    if (cast<VectorType>(Op1->getType())->getElementType() != ScalarTy) {
+    if (cast<VectorType>(Op1->getType())->getElementType() !=
+        ScalarTy->getScalarType()) {
       assert(ScalarTy->isIntegerTy() && "Expected item in MinBWs.");
       Op1 = Builder.CreateIntCast(
           Op1,
@@ -17007,7 +17008,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
               cast<FixedVectorType>(Op1->getType())->getNumElements()),
           GetOperandSignedness(&OpTE1));
     }
-    if (cast<VectorType>(Op2->getType())->getElementType() != ScalarTy) {
+    if (cast<VectorType>(Op2->getType())->getElementType() !=
+        ScalarTy->getScalarType()) {
       assert(ScalarTy->isIntegerTy() && "Expected item in MinBWs.");
       Op2 = Builder.CreateIntCast(
           Op2,
@@ -17022,9 +17024,15 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
           Mask.begin(),
           std::next(Mask.begin(), E->CombinedEntriesWithIndices.back().second),
           0);
+      unsigned ScalarTyNumElements = getNumElements(ScalarTy);
+      if (ScalarTyNumElements != 1) {
+        assert(SLPReVec && "Only supported by REVEC.");
+        transformScalarShuffleIndiciesToVector(ScalarTyNumElements, Mask);
+      }
       Value *Vec = Builder.CreateShuffleVector(Op1, Mask);
       Vec = createInsertVector(Builder, Vec, Op2,
-                               E->CombinedEntriesWithIndices.back().second);
+                               E->CombinedEntriesWithIndices.back().second *
+                                   ScalarTyNumElements);
       E->VectorizedValue = Vec;
       return Vec;
     }

llvm/test/Transforms/SLPVectorizer/X86/revec-SplitVectorize.ll

@@ -2,6 +2,31 @@
 ; RUN: opt -mtriple=x86_64-unknown-linux-gnu -mattr=+avx -passes=slp-vectorizer -S -slp-revec < %s | FileCheck %s
 
 define void @test() {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> poison, <4 x i32> zeroinitializer, i64 0)
+; CHECK-NEXT:    [[TMP1:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> [[TMP0]], <4 x i32> zeroinitializer, i64 4)
+; CHECK-NEXT:    [[TMP2:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> [[TMP1]], <4 x i32> zeroinitializer, i64 8)
+; CHECK-NEXT:    [[TMP3:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> [[TMP2]], <4 x i32> zeroinitializer, i64 12)
+; CHECK-NEXT:    [[TMP4:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> [[TMP3]], <4 x i32> zeroinitializer, i64 16)
+; CHECK-NEXT:    [[TMP5:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> [[TMP4]], <4 x i32> zeroinitializer, i64 20)
+; CHECK-NEXT:    [[TMP6:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> [[TMP5]], <4 x i32> zeroinitializer, i64 24)
+; CHECK-NEXT:    [[TMP7:%.*]] = call <32 x i32> @llvm.vector.insert.v32i32.v4i32(<32 x i32> [[TMP6]], <4 x i32> zeroinitializer, i64 28)
+; CHECK-NEXT:    [[TMP8:%.*]] = trunc <32 x i32> [[TMP7]] to <32 x i1>
+; CHECK-NEXT:    [[TMP9:%.*]] = call <16 x i1> @llvm.vector.insert.v16i1.v4i1(<16 x i1> poison, <4 x i1> zeroinitializer, i64 0)
+; CHECK-NEXT:    [[TMP10:%.*]] = call <16 x i1> @llvm.vector.insert.v16i1.v4i1(<16 x i1> [[TMP9]], <4 x i1> zeroinitializer, i64 4)
+; CHECK-NEXT:    [[TMP11:%.*]] = call <16 x i1> @llvm.vector.insert.v16i1.v4i1(<16 x i1> [[TMP10]], <4 x i1> zeroinitializer, i64 8)
+; CHECK-NEXT:    [[TMP12:%.*]] = call <16 x i1> @llvm.vector.insert.v16i1.v4i1(<16 x i1> [[TMP11]], <4 x i1> zeroinitializer, i64 12)
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[TMP13:%.*]] = phi <32 x i1> [ [[TMP8]], [[ENTRY:%.*]] ], [ [[TMP18:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP14:%.*]] = select <16 x i1> [[TMP12]], <16 x i1> [[TMP12]], <16 x i1> [[TMP12]]
+; CHECK-NEXT:    [[TMP15:%.*]] = xor <16 x i1> [[TMP12]], [[TMP12]]
+; CHECK-NEXT:    [[TMP16:%.*]] = shufflevector <16 x i1> [[TMP14]], <16 x i1> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP17:%.*]] = call <32 x i1> @llvm.vector.insert.v32i1.v16i1(<32 x i1> [[TMP16]], <16 x i1> [[TMP15]], i64 16)
+; CHECK-NEXT:    [[TMP18]] = or <32 x i1> [[TMP13]], [[TMP17]]
+; CHECK-NEXT:    br label [[VECTOR_BODY]]
+;
 entry:
   br label %vector.body