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

[Nikita Popov via llvm-commits]

================
@@ -1820,6 +1973,26 @@ 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
+  //
+  // r1 = 1/a
+  // r2 = sqrt(a)
+  // x = r1 * r2
+  SmallSetVector<Instruction *, 2> R1, R2;
+  getFSqrtDivOptPattern(&I, R1, R2);
+  if (!(R1.empty() || R2.empty()) && isFSqrtDivToFMulLegal(&I, R1, R2)) {
+    CallInst *CI = (CallInst *)((&I)->getOperand(1));
----------------
nikic wrote:

```suggestion
    auto *CI = cast<CallInst>(I.getOperand(1));
```

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