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

[via llvm-commits]

================
@@ -1820,6 +1989,30 @@ Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) {
     return R;
 
   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+  // Convert
+  // x = 1.0/sqrt(a)
+  // r1 = x * x;
+  // r2 = a * x;
+  //
+  // TO
+  //
+  // tmp1 = 1.0 / a
+  // tmp2 = sqrt (a)
+  // tmp3 = tmp1 * tmp2
+  // x = tmp3
+  // r1 = tmp1
+  // r2 = tmp2
----------------
sushgokh wrote:

@yashssh As said above, this might give the reader an incorrect picture that we are hoisting the r1/r2 above x. We are just hoisting the operations corresponding to r1/r2 above x.
If you and @arsenm still suggest to modify this, will change it

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