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

[Sushant Gokhale via llvm-commits]

================
@@ -1796,6 +1878,64 @@ static Instruction *foldFDivSqrtDivisor(BinaryOperator &I,
   return BinaryOperator::CreateFMulFMF(Op0, NewSqrt, &I);
 }
 
+// Change
+// X = 1/sqrt(a)
+// R1 = X * X
+// R2 = a * X
+//
+// TO
+//
+// Tmp1 = 1/a
+// Tmp2 = sqrt(a)
+// Tmp3 = Tmp1 * Tmp2
+// Replace Uses Of R1 With Tmp1
+// Replace Uses Of R2 With Tmp2
+// Replace Uses Of X With Tmp3
+static Value *convertFSqrtDivIntoFMul(CallInst *CI, Instruction *X,
+                                      ArrayRef<Instruction *> R1,
+                                      ArrayRef<Instruction *> R2, Value *SqrtOp,
+                                      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.
+  Instruction *Tmp1;
----------------
sushgokh wrote:

Tmp1/Tmp2 is being used for synthesizing Tmp3. So, I need to keep it the way as it is now. But if there is better way of doing this, sure I can do it

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