[LoopVectorizer][AArch64] Add support for partial reduce subtraction

Instead of implementing a new intrinsic for subtracting partial
reductions, generate a negation instruction for the second operand
of the partial reduction.

Author: NickGuy-Arm

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -4673,7 +4673,7 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
   InstructionCost Invalid = InstructionCost::getInvalid();
   InstructionCost Cost(TTI::TCC_Basic);
 
-  if (Opcode != Instruction::Add)
+  if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
     return Invalid;
 
   if (InputTypeA != InputTypeB)

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -8697,8 +8697,9 @@ void VPRecipeBuilder::collectScaledReductions(VFRange &Range) {
 
   // Build up a set of partial reduction bin ops for efficient use checking.
   SmallSet<User *, 4> PartialReductionBinOps;
-  for (const auto &[PartialRdx, _] : PartialReductionChains)
+  for (const auto &[PartialRdx, _] : PartialReductionChains) {
     PartialReductionBinOps.insert(PartialRdx.BinOp);
+  }
 
   auto ExtendIsOnlyUsedByPartialReductions =
       [&PartialReductionBinOps](Instruction *Extend) {
@@ -8761,20 +8762,23 @@ bool VPRecipeBuilder::getScaledReductions(
     return false;
 
   using namespace llvm::PatternMatch;
+  BinaryOperator *ExtendedBinOp = BinOp;
+  match(BinOp, m_Neg(m_BinOp(ExtendedBinOp)));
+
   Value *A, *B;
-  if (!match(BinOp->getOperand(0), m_ZExtOrSExt(m_Value(A))) ||
-      !match(BinOp->getOperand(1), m_ZExtOrSExt(m_Value(B))))
+  if (!match(ExtendedBinOp->getOperand(0), m_ZExtOrSExt(m_Value(A))) ||
+      !match(ExtendedBinOp->getOperand(1), m_ZExtOrSExt(m_Value(B))))
     return false;
 
-  Instruction *ExtA = cast<Instruction>(BinOp->getOperand(0));
-  Instruction *ExtB = cast<Instruction>(BinOp->getOperand(1));
+  Instruction *ExtA = cast<Instruction>(ExtendedBinOp->getOperand(0));
+  Instruction *ExtB = cast<Instruction>(ExtendedBinOp->getOperand(1));
 
   TTI::PartialReductionExtendKind OpAExtend =
       TargetTransformInfo::getPartialReductionExtendKind(ExtA);
   TTI::PartialReductionExtendKind OpBExtend =
       TargetTransformInfo::getPartialReductionExtendKind(ExtB);
 
-  PartialReductionChain Chain(RdxExitInstr, ExtA, ExtB, BinOp);
+  PartialReductionChain Chain(RdxExitInstr, ExtA, ExtB, ExtendedBinOp);
 
   unsigned TargetScaleFactor =
       PHI->getType()->getPrimitiveSizeInBits().getKnownScalarFactor(
@@ -8785,7 +8789,7 @@ bool VPRecipeBuilder::getScaledReductions(
             InstructionCost Cost = TTI->getPartialReductionCost(
                 Update->getOpcode(), A->getType(), B->getType(), PHI->getType(),
                 VF, OpAExtend, OpBExtend,
-                std::make_optional(BinOp->getOpcode()));
+                std::make_optional(ExtendedBinOp->getOpcode()));
             return Cost.isValid();
           },
           Range)) {

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -25,6 +25,7 @@
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/PatternMatch.h"
 #include "llvm/IR/Type.h"
 #include "llvm/IR/Value.h"
 #include "llvm/IR/VectorBuilder.h"
@@ -282,7 +283,20 @@ InstructionCost
 VPPartialReductionRecipe::computeCost(ElementCount VF,
                                       VPCostContext &Ctx) const {
   std::optional<unsigned> Opcode = std::nullopt;
-  VPRecipeBase *BinOpR = getOperand(0)->getDefiningRecipe();
+  VPValue *BinOp = getOperand(0);
+  VPRecipeBase *BinOpR = BinOp->getDefiningRecipe();
+
+  using namespace llvm::PatternMatch;
+  if (auto *UnderInst =
+          dyn_cast_if_present<Instruction>(BinOp->getUnderlyingValue())) {
+    if (match(UnderInst, m_Neg(m_BinOp()))) {
+      BinOpR = BinOpR->getOperand(1)->getDefiningRecipe();
+    }
+  }
+  // BinOp is never used again, any further interaction should be via the
+  // defining recipe `BinOpR`
+  BinOp = nullptr;
+
   if (auto *WidenR = dyn_cast<VPWidenRecipe>(BinOpR))
     Opcode = std::make_optional(WidenR->getOpcode());
 
@@ -318,13 +332,20 @@ void VPPartialReductionRecipe::execute(VPTransformState &State) {
   State.setDebugLocFrom(getDebugLoc());
   auto &Builder = State.Builder;
 
-  assert(getOpcode() == Instruction::Add &&
-         "Unhandled partial reduction opcode");
-
   Value *BinOpVal = State.get(getOperand(0));
   Value *PhiVal = State.get(getOperand(1));
   assert(PhiVal && BinOpVal && "Phi and Mul must be set");
 
+  unsigned Opcode = getOpcode();
+
+  if (Opcode == Instruction::Sub) {
+    bool HasNSW = cast<Instruction>(BinOpVal)->hasNoSignedWrap();
+    BinOpVal = Builder.CreateNeg(BinOpVal, "", HasNSW);
+    Opcode = Instruction::Add;
+  }
+
+  assert(Opcode == Instruction::Add && "Unhandled partial reduction opcode");
+
   Type *RetTy = PhiVal->getType();
 
   CallInst *V = Builder.CreateIntrinsic(