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

[Sushant Gokhale via llvm-commits]

================
@@ -1864,6 +1950,63 @@ 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,
+                                      ArrayRef<Instruction *> R1,
+                                      ArrayRef<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[0]->getType(), 1.0), SqrtOp));
+    FDiv->copyMetadata(*I);
+    FDiv->copyFastMathFlags(I);
+    I->replaceAllUsesWith(FDiv);
+  }
+
+  // Although, by value, FSqrt = CI , every instance of R2 may have different
+  // fpmath metadata and fpmath flags. We try to preserve them by cloning the
+  // call instruction per R2 instance.
+  CallInst *FSqrt;
+  for (Instruction *I : R2) {
+    FSqrt = cast<CallInst>(CI->clone());
+    FSqrt->insertBefore(CI);
+    FSqrt->copyFastMathFlags(I);
+    FSqrt->copyMetadata(*I);
+    I->replaceAllUsesWith(FSqrt);
----------------
sushgokh wrote:

agreed

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