Initial changes picked over from #109833

Author: MacDue

llvm/include/llvm/Analysis/VectorUtils.h

@@ -18,6 +18,7 @@
 #include "llvm/Analysis/LoopAccessAnalysis.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IR/VFABIDemangler.h"
+#include "llvm/IR/VectorUtils.h"
 #include "llvm/Support/CheckedArithmetic.h"
 
 namespace llvm {
@@ -127,18 +128,8 @@ namespace Intrinsic {
 typedef unsigned ID;
 }
 
-/// A helper function for converting Scalar types to vector types. If
-/// the incoming type is void, we return void. If the EC represents a
-/// scalar, we return the scalar type.
-inline Type *ToVectorTy(Type *Scalar, ElementCount EC) {
-  if (Scalar->isVoidTy() || Scalar->isMetadataTy() || EC.isScalar())
-    return Scalar;
-  return VectorType::get(Scalar, EC);
-}
-
-inline Type *ToVectorTy(Type *Scalar, unsigned VF) {
-  return ToVectorTy(Scalar, ElementCount::getFixed(VF));
-}
+/// Returns true if `Ty` can be widened by the loop vectorizer.
+bool canWidenType(Type *Ty);
 
 /// Identify if the intrinsic is trivially vectorizable.
 /// This method returns true if the intrinsic's argument types are all scalars

llvm/include/llvm/IR/VectorUtils.h

@@ -0,0 +1,53 @@
+//===----------- VectorUtils.h -  Vector type utility functions -*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/IR/DerivedTypes.h"
+
+namespace llvm {
+
+/// A helper function for converting Scalar types to vector types. If
+/// the incoming type is void, we return void. If the EC represents a
+/// scalar, we return the scalar type.
+inline Type *ToVectorTy(Type *Scalar, ElementCount EC) {
+  if (Scalar->isVoidTy() || Scalar->isMetadataTy() || EC.isScalar())
+    return Scalar;
+  return VectorType::get(Scalar, EC);
+}
+
+inline Type *ToVectorTy(Type *Scalar, unsigned VF) {
+  return ToVectorTy(Scalar, ElementCount::getFixed(VF));
+}
+
+/// A helper for converting to wider (vector) types. For scalar types, this is
+/// equivalent to calling `ToVectorTy`. For struct types, this returns a new
+/// struct where each element type has been widened to a vector type. Note: Only
+/// unpacked literal struct types are supported.
+Type *ToWideTy(Type *Ty, ElementCount EC);
+
+/// A helper for converting wide types to narrow (non-vector) types. For vector
+/// types, this is equivalent to calling .getScalarType(). For struct types,
+/// this returns a new struct where each element type has been converted to a
+/// scalar type. Note: Only unpacked literal struct types are supported.
+Type *ToNarrowTy(Type *Ty);
+
+/// Returns the types contained in `Ty`. For struct types, it returns the
+/// elements, all other types are returned directly.
+SmallVector<Type *, 2> getContainedTypes(Type *Ty);
+
+/// Returns true if `Ty` is a vector type or a struct of vector types where all
+/// vector types share the same VF.
+bool isWideTy(Type *Ty);
+
+/// Returns the vectorization factor for a widened type.
+inline ElementCount getWideTypeVF(Type *Ty) {
+  assert(isWideTy(Ty) && "expected widened type!");
+  return cast<VectorType>(getContainedTypes(Ty).front())->getElementCount();
+}
+
+} // namespace llvm

llvm/lib/Analysis/VectorUtils.cpp

@@ -39,6 +39,20 @@ static cl::opt<unsigned> MaxInterleaveGroupFactor(
     cl::desc("Maximum factor for an interleaved access group (default = 8)"),
     cl::init(8));
 
+/// Returns true if `Ty` can be widened by the loop vectorizer.
+bool llvm::canWidenType(Type *Ty) {
+  Type *ElTy = Ty;
+  // For now, only allow widening non-packed literal structs where all
+  // element types are the same. This simplifies the cost model and
+  // conversion between scalar and wide types.
+  if (auto *StructTy = dyn_cast<StructType>(Ty);
+      StructTy && !StructTy->isPacked() && StructTy->isLiteral() &&
+      StructTy->containsHomogeneousTypes()) {
+    ElTy = StructTy->elements().front();
+  }
+  return VectorType::isValidElementType(ElTy);
+}
+
 /// Return true if all of the intrinsic's arguments and return type are scalars
 /// for the scalar form of the intrinsic, and vectors for the vector form of the
 /// intrinsic (except operands that are marked as always being scalar by

llvm/lib/IR/CMakeLists.txt

@@ -73,6 +73,7 @@ add_llvm_component_library(LLVMCore
   Value.cpp
   ValueSymbolTable.cpp
   VectorBuilder.cpp
+  VectorUtils.cpp
   Verifier.cpp
   VFABIDemangler.cpp
   RuntimeLibcalls.cpp

llvm/lib/IR/VFABIDemangler.cpp

@@ -11,6 +11,7 @@
 #include "llvm/ADT/SmallString.h"
 #include "llvm/ADT/StringSwitch.h"
 #include "llvm/IR/Module.h"
+#include "llvm/IR/VectorUtils.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
 #include <limits>
@@ -346,12 +347,15 @@ getScalableECFromSignature(const FunctionType *Signature, const VFISAKind ISA,
   // Also check the return type if not void.
   Type *RetTy = Signature->getReturnType();
   if (!RetTy->isVoidTy()) {
-    std::optional<ElementCount> ReturnEC = getElementCountForTy(ISA, RetTy);
-    // If we have an unknown scalar element type we can't find a reasonable VF.
-    if (!ReturnEC)
-      return std::nullopt;
-    if (ElementCount::isKnownLT(*ReturnEC, MinEC))
-      MinEC = *ReturnEC;
+    for (Type *RetTy : getContainedTypes(RetTy)) {
+      std::optional<ElementCount> ReturnEC = getElementCountForTy(ISA, RetTy);
+      // If we have an unknown scalar element type we can't find a reasonable
+      // VF.
+      if (!ReturnEC)
+        return std::nullopt;
+      if (ElementCount::isKnownLT(*ReturnEC, MinEC))
+        MinEC = *ReturnEC;
+    }
   }
 
   // The SVE Vector function call ABI bases the VF on the widest element types
@@ -566,7 +570,7 @@ FunctionType *VFABI::createFunctionType(const VFInfo &Info,
 
   auto *RetTy = ScalarFTy->getReturnType();
   if (!RetTy->isVoidTy())
-    RetTy = VectorType::get(RetTy, VF);
+    RetTy = ToWideTy(RetTy, VF);
   return FunctionType::get(RetTy, VecTypes, false);
 }
 

llvm/lib/IR/VectorUtils.cpp

@@ -0,0 +1,69 @@
+//===----------- VectorUtils.cpp - Vector type utility functions ----------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/IR/VectorUtils.h"
+#include "llvm/ADT/SmallVectorExtras.h"
+
+using namespace llvm;
+
+/// A helper for converting to wider (vector) types. For scalar types, this is
+/// equivalent to calling `ToVectorTy`. For struct types, this returns a new
+/// struct where each element type has been widened to a vector type. Note: Only
+/// unpacked literal struct types are supported.
+Type *llvm::ToWideTy(Type *Ty, ElementCount EC) {
+  if (EC.isScalar())
+    return Ty;
+  auto *StructTy = dyn_cast<StructType>(Ty);
+  if (!StructTy)
+    return ToVectorTy(Ty, EC);
+  assert(StructTy->isLiteral() && !StructTy->isPacked() &&
+         "expected unpacked struct literal");
+  return StructType::get(
+      Ty->getContext(),
+      map_to_vector(StructTy->elements(), [&](Type *ElTy) -> Type * {
+        return VectorType::get(ElTy, EC);
+      }));
+}
+
+/// A helper for converting wide types to narrow (non-vector) types. For vector
+/// types, this is equivalent to calling .getScalarType(). For struct types,
+/// this returns a new struct where each element type has been converted to a
+/// scalar type. Note: Only unpacked literal struct types are supported.
+Type *llvm::ToNarrowTy(Type *Ty) {
+  auto *StructTy = dyn_cast<StructType>(Ty);
+  if (!StructTy)
+    return Ty->getScalarType();
+  assert(StructTy->isLiteral() && !StructTy->isPacked() &&
+         "expected unpacked struct literal");
+  return StructType::get(
+      Ty->getContext(),
+      map_to_vector(StructTy->elements(), [](Type *ElTy) -> Type * {
+        return ElTy->getScalarType();
+      }));
+}
+
+/// Returns the types contained in `Ty`. For struct types, it returns the
+/// elements, all other types are returned directly.
+SmallVector<Type *, 2> llvm::getContainedTypes(Type *Ty) {
+  auto *StructTy = dyn_cast<StructType>(Ty);
+  if (StructTy)
+    return to_vector<2>(StructTy->elements());
+  return {Ty};
+}
+
+/// Returns true if `Ty` is a vector type or a struct of vector types where all
+/// vector types share the same VF.
+bool llvm::isWideTy(Type *Ty) {
+  auto ContainedTys = getContainedTypes(Ty);
+  if (ContainedTys.empty() || !ContainedTys.front()->isVectorTy())
+    return false;
+  ElementCount VF = cast<VectorType>(ContainedTys.front())->getElementCount();
+  return all_of(ContainedTys, [&](Type *Ty) {
+    return Ty->isVectorTy() && cast<VectorType>(Ty)->getElementCount() == VF;
+  });
+}

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

@@ -949,8 +949,8 @@ bool LoopVectorizationLegality::canVectorizeInstrs() {
       // Check that the instruction return type is vectorizable.
       // We can't vectorize casts from vector type to scalar type.
       // Also, we can't vectorize extractelement instructions.
-      if ((!VectorType::isValidElementType(I.getType()) &&
-           !I.getType()->isVoidTy()) ||
+      Type *InstTy = I.getType();
+      if (!(InstTy->isVoidTy() || canWidenType(InstTy)) ||
           (isa<CastInst>(I) &&
            !VectorType::isValidElementType(I.getOperand(0)->getType())) ||
           isa<ExtractElementInst>(I)) {

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -2861,10 +2861,10 @@ LoopVectorizationCostModel::getVectorCallCost(CallInst *CI,
   return ScalarCallCost;
 }
 
-static Type *maybeVectorizeType(Type *Elt, ElementCount VF) {
-  if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy()))
-    return Elt;
-  return VectorType::get(Elt, VF);
+static Type *maybeVectorizeType(Type *Ty, ElementCount VF) {
+  if (VF.isScalar() || !canWidenType(Ty))
+    return Ty;
+  return ToWideTy(Ty, VF);
 }
 
 InstructionCost
@@ -3635,9 +3635,8 @@ void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) {
 
       // ExtractValue instructions must be uniform, because the operands are
       // known to be loop-invariant.
-      if (auto *EVI = dyn_cast<ExtractValueInst>(&I)) {
-        assert(IsOutOfScope(EVI->getAggregateOperand()) &&
-               "Expected aggregate value to be loop invariant");
+      if (auto *EVI = dyn_cast<ExtractValueInst>(&I);
+          EVI && IsOutOfScope(EVI->getAggregateOperand())) {
         AddToWorklistIfAllowed(EVI);
         continue;
       }
@@ -5461,10 +5460,13 @@ InstructionCost LoopVectorizationCostModel::computePredInstDiscount(
     // and phi nodes.
     TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
     if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) {
-      ScalarCost += TTI.getScalarizationOverhead(
-          cast<VectorType>(ToVectorTy(I->getType(), VF)),
-          APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ true,
-          /*Extract*/ false, CostKind);
+      Type *WideTy = ToWideTy(I->getType(), VF);
+      for (Type *VectorTy : getContainedTypes(WideTy)) {
+        ScalarCost += TTI.getScalarizationOverhead(
+            cast<VectorType>(VectorTy), APInt::getAllOnes(VF.getFixedValue()),
+            /*Insert*/ true,
+            /*Extract*/ false, CostKind);
+      }
       ScalarCost +=
           VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind);
     }
@@ -5953,13 +5955,17 @@ InstructionCost LoopVectorizationCostModel::getScalarizationOverhead(
     return 0;
 
   InstructionCost Cost = 0;
-  Type *RetTy = ToVectorTy(I->getType(), VF);
+  Type *RetTy = ToWideTy(I->getType(), VF);
   if (!RetTy->isVoidTy() &&
-      (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore()))
-    Cost += TTI.getScalarizationOverhead(
-        cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()),
-        /*Insert*/ true,
-        /*Extract*/ false, CostKind);
+      (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) {
+
+    for (Type *VectorTy : getContainedTypes(RetTy)) {
+      Cost += TTI.getScalarizationOverhead(
+          cast<VectorType>(VectorTy), APInt::getAllOnes(VF.getKnownMinValue()),
+          /*Insert*/ true,
+          /*Extract*/ false, CostKind);
+    }
+  }
 
   // Some targets keep addresses scalar.
   if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing())
@@ -6219,9 +6225,9 @@ void LoopVectorizationCostModel::setVectorizedCallDecision(ElementCount VF) {
 
       bool MaskRequired = Legal->isMaskRequired(CI);
       // Compute corresponding vector type for return value and arguments.
-      Type *RetTy = ToVectorTy(ScalarRetTy, VF);
+      Type *RetTy = ToWideTy(ScalarRetTy, VF);
       for (Type *ScalarTy : ScalarTys)
-        Tys.push_back(ToVectorTy(ScalarTy, VF));
+        Tys.push_back(ToWideTy(ScalarTy, VF));
 
       // An in-loop reduction using an fmuladd intrinsic is a special case;
       // we don't want the normal cost for that intrinsic.
@@ -6398,7 +6404,7 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I,
            HasSingleCopyAfterVectorization(I, VF));
     VectorTy = RetTy;
   } else
-    VectorTy = ToVectorTy(RetTy, VF);
+    VectorTy = ToWideTy(RetTy, VF);
 
   if (VF.isVector() && VectorTy->isVectorTy() &&
       !TTI.getNumberOfParts(VectorTy))

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -1026,7 +1026,7 @@ InstructionCost VPWidenIntrinsicRecipe::computeCost(ElementCount VF,
     Arguments.push_back(V);
   }
 
-  Type *RetTy = ToVectorTy(Ctx.Types.inferScalarType(this), VF);
+  Type *RetTy = ToWideTy(Ctx.Types.inferScalarType(this), VF);
   SmallVector<Type *> ParamTys;
   for (unsigned I = 0; I != getNumOperands(); ++I)
     ParamTys.push_back(