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

[Matt Arsenault via llvm-commits]

================
@@ -1796,6 +1941,30 @@ static Instruction *foldFDivSqrtDivisor(BinaryOperator &I,
   return BinaryOperator::CreateFMulFMF(Op0, NewSqrt, &I);
 }
 
+Value *convertFSqrtDivIntoFMul(CallInst *CI, Instruction *X,
+                               SmallSetVector<Instruction *, 2> &R1,
+                               SmallSetVector<Instruction *, 2> &R2,
+                               Value *SqrtOp, InstCombiner::BuilderTy &B) {
+
+  // 1. synthesize tmp1 = 1/a and replace uses of r1
+  B.SetInsertPoint(X);
+  Value *Tmp1 =
+      B.CreateFDivFMF(ConstantFP::get(R1[0]->getType(), 1.0), SqrtOp, R1[0]);
----------------
arsenm wrote:

You do not have any tests showing !fpmath metadata is preserved. This is separate from the FastMathFlags (and I believe is getting dropped in this code as written). I would like to see preservation where there is fpmath metadata set on each of the sqrt, div, and both 

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