[llvm] [InstCombine] Transform high latency, dependent FSQRT/FDIV into FMUL (PR #87474)

[Craig Topper via llvm-commits]

================
@@ -1864,6 +1949,64 @@ static Instruction *foldFDivSqrtDivisor(BinaryOperator &I,
   return BinaryOperator::CreateFMulFMF(Op0, NewSqrt, &I);
 }
 
+// Change
+// X = 1/sqrt(a)
+// R1 = X * X
+// R2 = a * X
+//
+// TO
+//
+// FDiv = 1/a
+// FSqrt = sqrt(a)
+// FMul = FDiv * FSqrt
+// Replace Uses Of R1 With FDiv
+// Replace Uses Of R2 With FSqrt
+// Replace Uses Of X With FMul
+static Value *convertFSqrtDivIntoFMul(CallInst *CI, Instruction *X,
+                                      SmallPtrSetImpl<Instruction *> &R1,
+                                      SmallPtrSetImpl<Instruction *> &R2,
+                                      InstCombiner::BuilderTy &B) {
+
+  B.SetInsertPoint(X);
+
+  // Every instance of R1 may have different fpmath metadata and fpmath flags.
+  // We try to preserve them by having seperate fdiv instruction per R1
+  // instance.
+  Value *SqrtOp = CI->getArgOperand(0);
+  Instruction *FDiv;
+
+  for (Instruction *I : R1) {
+    FDiv = cast<Instruction>(
+        B.CreateFDiv(ConstantFP::get((*R1.begin())->getType(), 1.0), SqrtOp));
----------------
topperc wrote:

Can you get the type from I instead of *R1.begin()?

https://github.com/llvm/llvm-project/pull/87474