[llvm] [CostModel][AArch64] Make extractelement, with fmul user, free whenev… (PR #111479)

[David Green via llvm-commits]

================
@@ -3226,6 +3227,130 @@ InstructionCost AArch64TTIImpl::getVectorInstrCostHelper(const Instruction *I,
     // compile-time considerations.
   }
 
+  // In case of Neon, if there exists extractelement from lane != 0 such that
+  // 1. extractelement does not necessitate a move from vector_reg -> GPR.
+  // 2. extractelement result feeds into fmul.
+  // 3. Other operand of fmul is an extractelement from lane 0 or lane
+  // equivalent to 0.
+  // then the extractelement can be merged with fmul in the backend and it
+  // incurs no cost.
+  // e.g.
+  // define double @foo(<2 x double> %a) {
+  //   %1 = extractelement <2 x double> %a, i32 0
+  //   %2 = extractelement <2 x double> %a, i32 1
+  //   %res = fmul double %1, %2
+  //   ret double %res
+  // }
+  // %2 and %res can be merged in the backend to generate fmul d0, d0, v1.d[1]
+  auto ExtractCanFuseWithFmul = [&]() {
+    // We bail out if the extract is from lane 0.
+    if (Index == 0)
+      return false;
+
+    // Check if the scalar element type of the vector operand of ExtractElement
+    // instruction is one of the allowed types.
+    auto IsAllowedScalarTy = [&](const Type *T) {
+      return T->isFloatTy() || T->isDoubleTy() ||
+             (T->isHalfTy() && ST->hasFullFP16());
+    };
+
+    // Check if the extractelement user is scalar fmul.
+    auto IsUserFMulScalarTy = [](const Value *EEUser) {
+      // Check if the user is scalar fmul.
+      const auto *BO = dyn_cast_if_present<BinaryOperator>(EEUser);
+      return BO && BO->getOpcode() == BinaryOperator::FMul &&
+             !BO->getType()->isVectorTy();
+    };
+
+    // Check if the type constraints on input vector type and result scalar type
+    // of extractelement instruction are satisfied.
+    auto TypeConstraintsOnEESatisfied =
+        [&IsAllowedScalarTy](const Type *VectorTy, const Type *ScalarTy) {
+          return isa<FixedVectorType>(VectorTy) && IsAllowedScalarTy(ScalarTy);
+        };
+
+    // Check if the extract index is from lane 0 or lane equivalent to 0 for a
+    // certain scalar type and a certain vector register width.
+    auto IsExtractLaneEquivalentToZero = [&](const unsigned &Idx,
+                                             const unsigned &EltSz) {
+      auto RegWidth =
+          getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
+              .getFixedValue();
+      return (Idx == 0 || (Idx * EltSz) % RegWidth == 0);
+    };
+
+    if (Opcode.has_value()) {
+      if (!TypeConstraintsOnEESatisfied(Val, Val->getScalarType()))
+        return false;
----------------
davemgreen wrote:

Can you explain what you mean?

In this case I didn't mean to remove the checking of the types, just to move it up to above the `if (Scalar) {`. Then remove the version that uses `EE->getVectorOperand()->getType()` below, as that should always be equal to `Val` if it exists.

https://github.com/llvm/llvm-project/pull/111479