[LV] Add support for partial reductions without a binary op

Consider IR such as this:

for.body:
  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
  %accum = phi i32 [ 0, %entry ], [ %add, %for.body ]
  %gep.a = getelementptr i8, ptr %a, i64 %iv
  %load.a = load i8, ptr %gep.a, align 1
  %ext.a = zext i8 %load.a to i32
  %add = add i32 %ext.a, %accum
  %iv.next = add i64 %iv, 1
  %exitcond.not = icmp eq i64 %iv.next, 1025
  br i1 %exitcond.not, label %for.exit, label %for.body

Conceptually we can vectorise this using partial reductions too,
although the current loop vectoriser implementation requires the
accumulation of a multiply. For AArch64 this is easily done with
a udot or sdot with an identity operand, i.e. a vector of (i16 1).

In order to do this I had to teach getScaledReductions that the
accumulated value may come from a unary op, hence there is only
one extension to consider. Similarly, I updated the vplan and
AArch64 TTI cost model to understand the possible unary op.

Author: david-arm

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -1325,9 +1325,21 @@ class TargetTransformInfo {
   /// \return The cost of a partial reduction, which is a reduction from a
   /// vector to another vector with fewer elements of larger size. They are
   /// represented by the llvm.experimental.partial.reduce.add intrinsic, which
-  /// takes an accumulator and a binary operation operand that itself is fed by
-  /// two extends. An example of an operation that uses a partial reduction is a
-  /// dot product, which reduces two vectors to another of 4 times fewer and 4
+  /// takes an accumulator of type \p AccumType and a second vector operand to
+  /// be accumulated, whose element count is specified by \p VF. The type of
+  /// reduction is specified by \p Opcode. The second operand passed to the
+  /// intrinsic could be the result of an extend, such as sext or zext. In
+  /// this case \p BinOp is nullopt, \p InputTypeA represents the type being
+  /// extended and \p OpAExtend the operation, i.e. sign- or zero-extend.
+  /// Also, \p InputTypeB should be nullptr and OpBExtend should be None.
+  /// Alternatively, the second operand could be the result of a binary
+  /// operation performed on two extends, i.e.
+  ///   mul(zext i8 %a -> i32, zext i8 %b -> i32).
+  /// In this case \p BinOp may specify the opcode of the binary operation,
+  /// \p InputTypeA and \p InputTypeB the types being extended, and
+  /// \p OpAExtend, \p OpBExtend the form of extensions. An example of an
+  /// operation that uses a partial reduction is a dot product, which reduces
+  /// two vectors in binary mul operation to another of 4 times fewer and 4
   /// times larger elements.
   LLVM_ABI InstructionCost getPartialReductionCost(
       unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType,

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -5402,11 +5402,21 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
 
   // Sub opcodes currently only occur in chained cases.
   // Independent partial reduction subtractions are still costed as an add
-  if (Opcode != Instruction::Add && Opcode != Instruction::Sub)
+  if ((Opcode != Instruction::Add && Opcode != Instruction::Sub) ||
+      OpAExtend == TTI::PR_None)
     return Invalid;
 
-  if (InputTypeA != InputTypeB)
+  // We only support multiply binary operations for now, and for muls we
+  // require the types being extended to be the same.
+  // NOTE: For muls AArch64 supports lowering mixed extensions to a usdot but
+  // only if the i8mm or sve/streaming features are available.
+  if (BinOp && (*BinOp != Instruction::Mul || InputTypeA != InputTypeB ||
+                OpBExtend == TTI::PR_None ||
+                (OpAExtend != OpBExtend && !ST->hasMatMulInt8() &&
+                 !ST->isSVEorStreamingSVEAvailable())))
     return Invalid;
+  assert((BinOp || (OpBExtend == TTI::PR_None && !InputTypeB)) &&
+         "Unexpected values for OpBExtend or InputTypeB");
 
   EVT InputEVT = EVT::getEVT(InputTypeA);
   EVT AccumEVT = EVT::getEVT(AccumType);
@@ -5453,15 +5463,6 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
   } else
     return Invalid;
 
-  // AArch64 supports lowering mixed fixed-width extensions to a usdot but only
-  // if the i8mm feature is available.
-  if (OpAExtend == TTI::PR_None || OpBExtend == TTI::PR_None ||
-      (OpAExtend != OpBExtend && !ST->hasMatMulInt8()))
-    return Invalid;
-
-  if (!BinOp || *BinOp != Instruction::Mul)
-    return Invalid;
-
   return Cost;
 }
 

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -8150,15 +8150,15 @@ void VPRecipeBuilder::collectScaledReductions(VFRange &Range) {
   // something that isn't another partial reduction. This is because the
   // extends are intended to be lowered along with the reduction itself.
 
-  // Build up a set of partial reduction bin ops for efficient use checking.
-  SmallSet<User *, 4> PartialReductionBinOps;
+  // Build up a set of partial reduction ops for efficient use checking.
+  SmallSet<User *, 4> PartialReductionOps;
   for (const auto &[PartialRdx, _] : PartialReductionChains)
-    PartialReductionBinOps.insert(PartialRdx.BinOp);
+    PartialReductionOps.insert(PartialRdx.ExtendUser);
 
   auto ExtendIsOnlyUsedByPartialReductions =
-      [&PartialReductionBinOps](Instruction *Extend) {
+      [&PartialReductionOps](Instruction *Extend) {
         return all_of(Extend->users(), [&](const User *U) {
-          return PartialReductionBinOps.contains(U);
+          return PartialReductionOps.contains(U);
         });
       };
 
@@ -8167,15 +8167,14 @@ void VPRecipeBuilder::collectScaledReductions(VFRange &Range) {
   for (auto Pair : PartialReductionChains) {
     PartialReductionChain Chain = Pair.first;
     if (ExtendIsOnlyUsedByPartialReductions(Chain.ExtendA) &&
-        ExtendIsOnlyUsedByPartialReductions(Chain.ExtendB))
+        (!Chain.ExtendB || ExtendIsOnlyUsedByPartialReductions(Chain.ExtendB)))
       ScaledReductionMap.try_emplace(Chain.Reduction, Pair.second);
   }
 }
 
 bool VPRecipeBuilder::getScaledReductions(
     Instruction *PHI, Instruction *RdxExitInstr, VFRange &Range,
     SmallVectorImpl<std::pair<PartialReductionChain, unsigned>> &Chains) {
-
   if (!CM.TheLoop->contains(RdxExitInstr))
     return false;
 
@@ -8204,43 +8203,75 @@ bool VPRecipeBuilder::getScaledReductions(
   if (PhiOp != PHI)
     return false;
 
-  auto *BinOp = dyn_cast<BinaryOperator>(Op);
-  if (!BinOp || !BinOp->hasOneUse())
-    return false;
-
   using namespace llvm::PatternMatch;
-  // Use the side-effect of match to replace BinOp only if the pattern is
-  // matched, we don't care at this point whether it actually matched.
-  match(BinOp, m_Neg(m_BinOp(BinOp)));
 
-  Value *A, *B;
-  if (!match(BinOp->getOperand(0), m_ZExtOrSExt(m_Value(A))) ||
-      !match(BinOp->getOperand(1), m_ZExtOrSExt(m_Value(B))))
-    return false;
+  // If the update is a binary operator, check both of its operands to see if
+  // they are extends. Otherwise, see if the update comes directly from an
+  // extend.
+  Instruction *Exts[2] = {nullptr};
+  BinaryOperator *ExtendUser = dyn_cast<BinaryOperator>(Op);
+  std::optional<unsigned> BinOpc;
+  Type *ExtOpTypes[2] = {nullptr};
+
+  auto CollectExtInfo = [&Exts,
+                         &ExtOpTypes](SmallVectorImpl<Value *> &Ops) -> bool {
+    unsigned I = 0;
+    for (Value *OpI : Ops) {
+      Value *ExtOp;
+      if (!match(OpI, m_ZExtOrSExt(m_Value(ExtOp))))
+        return false;
+      Exts[I] = cast<Instruction>(OpI);
+      ExtOpTypes[I] = ExtOp->getType();
+      I++;
+    }
+    return true;
+  };
+
+  if (ExtendUser) {
+    if (!ExtendUser->hasOneUse())
+      return false;
 
-  Instruction *ExtA = cast<Instruction>(BinOp->getOperand(0));
-  Instruction *ExtB = cast<Instruction>(BinOp->getOperand(1));
+    // Use the side-effect of match to replace BinOp only if the pattern is
+    // matched, we don't care at this point whether it actually matched.
+    match(ExtendUser, m_Neg(m_BinOp(ExtendUser)));
+
+    SmallVector<Value *> Ops(ExtendUser->operands());
+    if (!CollectExtInfo(Ops))
+      return false;
+
+    BinOpc = std::make_optional(ExtendUser->getOpcode());
+  } else if (match(Update, m_Add(m_Value(), m_Value()))) {
+    // We already know the operands for Update are Op and PhiOp.
+    SmallVector<Value *> Ops({Op});
+    if (!CollectExtInfo(Ops))
+      return false;
+
+    ExtendUser = Update;
+    BinOpc = std::nullopt;
+  } else
+    return false;
 
   TTI::PartialReductionExtendKind OpAExtend =
-      TargetTransformInfo::getPartialReductionExtendKind(ExtA);
+      TargetTransformInfo::getPartialReductionExtendKind(Exts[0]);
   TTI::PartialReductionExtendKind OpBExtend =
-      TargetTransformInfo::getPartialReductionExtendKind(ExtB);
-
-  PartialReductionChain Chain(RdxExitInstr, ExtA, ExtB, BinOp);
+      Exts[1] ? TargetTransformInfo::getPartialReductionExtendKind(Exts[1])
+              : TargetTransformInfo::PR_None;
+  PartialReductionChain Chain(RdxExitInstr, Exts[0], Exts[1], ExtendUser);
 
   TypeSize PHISize = PHI->getType()->getPrimitiveSizeInBits();
-  TypeSize ASize = A->getType()->getPrimitiveSizeInBits();
-
+  TypeSize ASize = ExtOpTypes[0]->getPrimitiveSizeInBits();
   if (!PHISize.hasKnownScalarFactor(ASize))
     return false;
 
-  unsigned TargetScaleFactor = PHISize.getKnownScalarFactor(ASize);
+  unsigned TargetScaleFactor =
+      PHI->getType()->getPrimitiveSizeInBits().getKnownScalarFactor(
+          ExtOpTypes[0]->getPrimitiveSizeInBits());
 
   if (LoopVectorizationPlanner::getDecisionAndClampRange(
           [&](ElementCount VF) {
             InstructionCost Cost = TTI->getPartialReductionCost(
-                Update->getOpcode(), A->getType(), B->getType(), PHI->getType(),
-                VF, OpAExtend, OpBExtend, BinOp->getOpcode());
+                Update->getOpcode(), ExtOpTypes[0], ExtOpTypes[1],
+                PHI->getType(), VF, OpAExtend, OpBExtend, BinOpc);
             return Cost.isValid();
           },
           Range)) {

llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h

@@ -24,22 +24,23 @@ struct HistogramInfo;
 struct VFRange;
 
 /// A chain of instructions that form a partial reduction.
-/// Designed to match: reduction_bin_op (bin_op (extend (A), (extend (B))),
-/// accumulator).
+/// Designed to match either:
+///   reduction_bin_op (extend (A), accumulator), or
+///   reduction_bin_op (bin_op (extend (A), (extend (B))), accumulator).
 struct PartialReductionChain {
   PartialReductionChain(Instruction *Reduction, Instruction *ExtendA,
-                        Instruction *ExtendB, Instruction *BinOp)
-      : Reduction(Reduction), ExtendA(ExtendA), ExtendB(ExtendB), BinOp(BinOp) {
-  }
+                        Instruction *ExtendB, Instruction *ExtendUser)
+      : Reduction(Reduction), ExtendA(ExtendA), ExtendB(ExtendB),
+        ExtendUser(ExtendUser) {}
   /// The top-level binary operation that forms the reduction to a scalar
   /// after the loop body.
   Instruction *Reduction;
   /// The extension of each of the inner binary operation's operands.
   Instruction *ExtendA;
   Instruction *ExtendB;
 
-  /// The binary operation using the extends that is then reduced.
-  Instruction *BinOp;
+  /// The user of the extend that is then reduced.
+  Instruction *ExtendUser;
 };
 
 /// Helper class to create VPRecipies from IR instructions.

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -296,34 +296,21 @@ bool VPRecipeBase::isScalarCast() const {
 InstructionCost
 VPPartialReductionRecipe::computeCost(ElementCount VF,
                                       VPCostContext &Ctx) const {
-  std::optional<unsigned> Opcode = std::nullopt;
-  VPValue *BinOp = getOperand(1);
+  std::optional<unsigned> Opcode;
+  VPValue *Op = getOperand(0);
+  VPRecipeBase *OpR = Op->getDefiningRecipe();
 
-  // If the partial reduction is predicated, a select will be operand 0 rather
-  // than the binary op
+  // If the partial reduction is predicated, a select will be operand 0
   using namespace llvm::VPlanPatternMatch;
-  if (match(getOperand(1), m_Select(m_VPValue(), m_VPValue(), m_VPValue())))
-    BinOp = BinOp->getDefiningRecipe()->getOperand(1);
-
-  // If BinOp is a negation, use the side effect of match to assign the actual
-  // binary operation to BinOp
-  match(BinOp, m_Binary<Instruction::Sub>(m_SpecificInt(0), m_VPValue(BinOp)));
-  VPRecipeBase *BinOpR = BinOp->getDefiningRecipe();
-
-  if (auto *WidenR = dyn_cast<VPWidenRecipe>(BinOpR))
-    Opcode = std::make_optional(WidenR->getOpcode());
-
-  VPRecipeBase *ExtAR = BinOpR->getOperand(0)->getDefiningRecipe();
-  VPRecipeBase *ExtBR = BinOpR->getOperand(1)->getDefiningRecipe();
+  if (match(getOperand(1), m_Select(m_VPValue(), m_VPValue(Op), m_VPValue()))) {
+    OpR = Op->getDefiningRecipe();
+  }
 
-  auto *PhiType = Ctx.Types.inferScalarType(getOperand(1));
-  auto *InputTypeA = Ctx.Types.inferScalarType(ExtAR ? ExtAR->getOperand(0)
-                                                     : BinOpR->getOperand(0));
-  auto *InputTypeB = Ctx.Types.inferScalarType(ExtBR ? ExtBR->getOperand(0)
-                                                     : BinOpR->getOperand(1));
+  Type *InputTypeA = nullptr, *InputTypeB = nullptr;
+  TTI::PartialReductionExtendKind ExtAType = TargetTransformInfo::PR_None,
+                                  ExtBType = TargetTransformInfo::PR_None;
 
   auto GetExtendKind = [](VPRecipeBase *R) {
-    // The extend could come from outside the plan.
     if (!R)
       return TargetTransformInfo::PR_None;
     auto *WidenCastR = dyn_cast<VPWidenCastRecipe>(R);
@@ -336,9 +323,43 @@ VPPartialReductionRecipe::computeCost(ElementCount VF,
     return TargetTransformInfo::PR_None;
   };
 
+  // Pick out opcode, type/ext information and use sub side effects from a widen recipe.
+  auto HandleWiden = [&](VPWidenRecipe* Widen){
+    if (match(Widen,
+              m_Binary<Instruction::Sub>(m_SpecificInt(0), m_VPValue(Op)))) {
+      Widen = dyn_cast<VPWidenRecipe>(Op->getDefiningRecipe());
+    }
+    Opcode = Widen->getOpcode();
+    VPRecipeBase *ExtAR = Widen->getOperand(0)->getDefiningRecipe();
+    VPRecipeBase *ExtBR = Widen->getOperand(1)->getDefiningRecipe();
+    InputTypeA = Ctx.Types.inferScalarType(ExtAR ? ExtAR->getOperand(0)
+                                                  : Widen->getOperand(0));
+    InputTypeB = Ctx.Types.inferScalarType(ExtBR ? ExtBR->getOperand(0)
+                                                  : Widen->getOperand(1));
+    ExtAType = GetExtendKind(ExtAR);
+    ExtBType = GetExtendKind(ExtBR);
+  };
+
+  if (isa<VPWidenCastRecipe>(OpR)) {
+    InputTypeA = Ctx.Types.inferScalarType(OpR->getOperand(0));
+    ExtAType = GetExtendKind(OpR);
+  } else if (isa<VPReductionPHIRecipe>(OpR)) {
+    auto RedPhiOp1R = getOperand(1)->getDefiningRecipe();
+    if (isa<VPWidenCastRecipe>(RedPhiOp1R)) {
+      InputTypeA = Ctx.Types.inferScalarType(RedPhiOp1R->getOperand(0));
+      ExtAType = GetExtendKind(RedPhiOp1R);
+    } else if (auto Widen = dyn_cast<VPWidenRecipe>(RedPhiOp1R))
+      HandleWiden(Widen);
+  } else if (auto Widen = dyn_cast<VPWidenRecipe>(OpR)) {
+    HandleWiden(Widen);
+  } else if (auto Reduction = dyn_cast<VPPartialReductionRecipe>(OpR)) {
+    return Reduction->computeCost(VF, Ctx);
+  }
+
+  auto *PhiType = Ctx.Types.inferScalarType(getOperand(1));
   return Ctx.TTI.getPartialReductionCost(getOpcode(), InputTypeA, InputTypeB,
-                                         PhiType, VF, GetExtendKind(ExtAR),
-                                         GetExtendKind(ExtBR), Opcode);
+                                         PhiType, VF, ExtAType, ExtBType,
+                                         Opcode);
 }
 
 void VPPartialReductionRecipe::execute(VPTransformState &State) {

llvm/test/CodeGen/AArch64/neon-partial-reduce-dot-product.ll

@@ -1233,4 +1233,3 @@ entry:
     %partial.reduce = tail call <2 x i64> @llvm.experimental.vector.partial.reduce.add(<2 x i64> %acc, <8 x i64> %input.wide)
     ret <2 x i64> %partial.reduce
 }
-