Remove extend cost for fdot and fix for Invalid costs

Author: dheaton-arm

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -5844,6 +5844,14 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
       return Cost;
   }
 
+  // f16 -> f32 is natively supported for fdot
+  if (ST->isSVEorStreamingSVEAvailable() && ST->hasSVE2p1() &&
+      Opcode == Instruction::FAdd) {
+    if (AccumLT.second.getScalarType() == MVT::f32 &&
+        InputLT.second.getScalarType() == MVT::f16)
+      return Cost;
+  }
+
   // Add additional cost for the extends that would need to be inserted.
   return Cost + 2;
 }

llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h

@@ -498,6 +498,12 @@ m_c_Mul(const Op0_t &Op0, const Op1_t &Op1) {
   return m_c_Binary<Instruction::Mul, Op0_t, Op1_t>(Op0, Op1);
 }
 
+template <typename Op0_t, typename Op1_t>
+inline AllRecipe_match<Instruction::FMul, Op0_t, Op1_t>
+m_FMul(const Op0_t &Op0, const Op1_t &Op1) {
+  return m_Binary<Instruction::FMul, Op0_t, Op1_t>(Op0, Op1);
+}
+
 /// Match a binary AND operation.
 template <typename Op0_t, typename Op1_t>
 inline AllRecipe_commutative_match<Instruction::And, Op0_t, Op1_t>

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -317,7 +317,8 @@ VPPartialReductionRecipe::computeCost(ElementCount VF,
   // FIXME: Replace the entire function with this once all partial reduction
   // variants are bundled into VPExpressionRecipe.
   if (!match(Op, m_Select(m_VPValue(), m_VPValue(Op), m_VPValue())) &&
-      !match(Op, m_Mul(m_VPValue(), m_ConstantInt(MulConst)))) {
+      !match(Op, m_Mul(m_VPValue(), m_ConstantInt(MulConst))) &&
+      !match(Op, m_FMul(m_VPValue(), m_VPValue()))) {
     auto *PhiType = Ctx.Types.inferScalarType(getChainOp());
     auto *InputType = Ctx.Types.inferScalarType(getVecOp());
     return Ctx.TTI.getPartialReductionCost(getOpcode(), InputType, InputType,