[SLP][REVEC] Make SLP support revectorization (-slp-revec) and add simple test. (#98269)

This PR will make SLP support revectorization. Add an option -slp-revec
to control the functionality.

reference:

https://discourse.llvm.org/t/rfc-make-slp-vectorizer-revectorize-vector-instructions/79436

Author: HanKuanChen

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -113,6 +113,10 @@ static cl::opt<bool>
     RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden,
                         cl::desc("Run the SLP vectorization passes"));
 
+static cl::opt<bool>
+    SLPReVec("slp-revec", cl::init(false), cl::Hidden,
+             cl::desc("Enable vectorization for wider vector utilization"));
+
 static cl::opt<int>
     SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
                      cl::desc("Only vectorize if you gain more than this "
@@ -227,13 +231,26 @@ static const unsigned MaxPHINumOperands = 128;
 /// avoids spending time checking the cost model and realizing that they will
 /// be inevitably scalarized.
 static bool isValidElementType(Type *Ty) {
+  // TODO: Support ScalableVectorType.
+  if (SLPReVec && isa<FixedVectorType>(Ty))
+    Ty = Ty->getScalarType();
   return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() &&
          !Ty->isPPC_FP128Ty();
 }
 
+/// \returns the number of elements for Ty.
+static unsigned getNumElements(Type *Ty) {
+  assert(!isa<ScalableVectorType>(Ty) &&
+         "ScalableVectorType is not supported.");
+  if (auto *VecTy = dyn_cast<FixedVectorType>(Ty))
+    return VecTy->getNumElements();
+  return 1;
+}
+
 /// \returns the vector type of ScalarTy based on vectorization factor.
 static FixedVectorType *getWidenedType(Type *ScalarTy, unsigned VF) {
-  return FixedVectorType::get(ScalarTy, VF);
+  return FixedVectorType::get(ScalarTy->getScalarType(),
+                              VF * getNumElements(ScalarTy));
 }
 
 /// \returns True if the value is a constant (but not globals/constant
@@ -6779,15 +6796,15 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
   }
 
   // Don't handle vectors.
-  if (S.OpValue->getType()->isVectorTy() &&
+  if (!SLPReVec && S.OpValue->getType()->isVectorTy() &&
       !isa<InsertElementInst>(S.OpValue)) {
     LLVM_DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
     newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx);
     return;
   }
 
   if (StoreInst *SI = dyn_cast<StoreInst>(S.OpValue))
-    if (SI->getValueOperand()->getType()->isVectorTy()) {
+    if (!SLPReVec && SI->getValueOperand()->getType()->isVectorTy()) {
       LLVM_DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
       newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx);
       return;
@@ -11833,10 +11850,12 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
   Value *castToScalarTyElem(Value *V,
                             std::optional<bool> IsSigned = std::nullopt) {
     auto *VecTy = cast<VectorType>(V->getType());
-    if (VecTy->getElementType() == ScalarTy)
+    assert(getNumElements(ScalarTy) < getNumElements(VecTy) &&
+           (getNumElements(VecTy) % getNumElements(ScalarTy) == 0));
+    if (VecTy->getElementType() == ScalarTy->getScalarType())
       return V;
     return Builder.CreateIntCast(
-        V, VectorType::get(ScalarTy, VecTy->getElementCount()),
+        V, VectorType::get(ScalarTy->getScalarType(), VecTy->getElementCount()),
         IsSigned.value_or(!isKnownNonNegative(V, SimplifyQuery(*R.DL))));
   }
 
@@ -12221,7 +12240,8 @@ Value *BoUpSLP::vectorizeOperand(TreeEntry *E, unsigned NodeIdx,
         return ShuffleBuilder.finalize(std::nullopt);
       };
       Value *V = vectorizeTree(VE, PostponedPHIs);
-      if (VF != cast<FixedVectorType>(V->getType())->getNumElements()) {
+      if (VF * getNumElements(VL[0]->getType()) !=
+          cast<FixedVectorType>(V->getType())->getNumElements()) {
         if (!VE->ReuseShuffleIndices.empty()) {
           // Reshuffle to get only unique values.
           // If some of the scalars are duplicated in the vectorization

llvm/test/Transforms/SLPVectorizer/revec.ll

@@ -0,0 +1,40 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes=slp-vectorizer -S -slp-revec -slp-max-reg-size=1024 -slp-threshold=-100 %s | FileCheck %s
+
+define void @test1(ptr %a, ptr %b, ptr %c) {
+; CHECK-LABEL: @test1(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load <16 x i32>, ptr [[A:%.*]], align 4
+; CHECK-NEXT:    [[TMP1:%.*]] = load <16 x i32>, ptr [[B:%.*]], align 4
+; CHECK-NEXT:    [[TMP2:%.*]] = add <16 x i32> [[TMP1]], [[TMP0]]
+; CHECK-NEXT:    store <16 x i32> [[TMP2]], ptr [[C:%.*]], align 4
+; CHECK-NEXT:    ret void
+;
+entry:
+  %arrayidx3 = getelementptr inbounds i32, ptr %a, i64 4
+  %arrayidx7 = getelementptr inbounds i32, ptr %a, i64 8
+  %arrayidx11 = getelementptr inbounds i32, ptr %a, i64 12
+  %0 = load <4 x i32>, ptr %a, align 4
+  %1 = load <4 x i32>, ptr %arrayidx3, align 4
+  %2 = load <4 x i32>, ptr %arrayidx7, align 4
+  %3 = load <4 x i32>, ptr %arrayidx11, align 4
+  %arrayidx19 = getelementptr inbounds i32, ptr %b, i64 4
+  %arrayidx23 = getelementptr inbounds i32, ptr %b, i64 8
+  %arrayidx27 = getelementptr inbounds i32, ptr %b, i64 12
+  %4 = load <4 x i32>, ptr %b, align 4
+  %5 = load <4 x i32>, ptr %arrayidx19, align 4
+  %6 = load <4 x i32>, ptr %arrayidx23, align 4
+  %7 = load <4 x i32>, ptr %arrayidx27, align 4
+  %add.i = add <4 x i32> %4, %0
+  %add.i63 = add <4 x i32> %5, %1
+  %add.i64 = add <4 x i32> %6, %2
+  %add.i65 = add <4 x i32> %7, %3
+  %arrayidx36 = getelementptr inbounds i32, ptr %c, i64 4
+  %arrayidx39 = getelementptr inbounds i32, ptr %c, i64 8
+  %arrayidx42 = getelementptr inbounds i32, ptr %c, i64 12
+  store <4 x i32> %add.i, ptr %c, align 4
+  store <4 x i32> %add.i63, ptr %arrayidx36, align 4
+  store <4 x i32> %add.i64, ptr %arrayidx39, align 4
+  store <4 x i32> %add.i65, ptr %arrayidx42, align 4
+  ret void
+}