[llvm] [DirectX] Add atan2 intrinsic and expand for DXIL backend (p1) (PR #108865)

[Tex Riddell via llvm-commits]

================
@@ -305,6 +306,48 @@ static Value *expandNormalizeIntrinsic(CallInst *Orig) {
   return Builder.CreateFMul(X, MultiplicandVec);
 }
 
+static Value *expandAtan2Intrinsic(CallInst *Orig) {
+  Value *Y = Orig->getOperand(0);
+  Value *X = Orig->getOperand(1);
+  Type *Ty = X->getType();
+  IRBuilder<> Builder(Orig);
+
+  Value *Tan = Builder.CreateFDiv(Y, X);
+
+  Value *Atan =
+      Builder.CreateIntrinsic(Ty, Intrinsic::atan, {Tan}, nullptr, "Elt.Atan");
+
+  Constant *Pi = ConstantFP::get(Ty, llvm::numbers::pi);
+  Constant *HalfPi = ConstantFP::get(Ty, llvm::numbers::pi / 2);
+  Constant *NegHalfPi = ConstantFP::get(Ty, -llvm::numbers::pi / 2);
+  Constant *Zero = ConstantFP::get(Ty, 0);
+
+  Value *AtanAddPi = Builder.CreateFAdd(Atan, Pi);
+  Value *AtanSubPi = Builder.CreateFSub(Atan, Pi);
+
+  Value *Result = Atan;
+
+  Value *XLt0 = Builder.CreateFCmpOLT(X, Zero);
+  Value *XEq0 = Builder.CreateFCmpOEQ(X, Zero);
+
+  Value *YGe0 = Builder.CreateFCmpOGE(Y, Zero);
+  Value *YLt0 = Builder.CreateFCmpOLT(Y, Zero);
+
+  Value *XLt0AndYGe0 = Builder.CreateAnd(XLt0, YGe0);
+  Result = Builder.CreateSelect(XLt0AndYGe0, AtanAddPi, Result);
+
+  Value *XLt0AndYLt0 = Builder.CreateAnd(XLt0, YLt0);
+  Result = Builder.CreateSelect(XLt0AndYLt0, AtanSubPi, Result);
+
+  Value *XEq0AndYLt0 = Builder.CreateAnd(XEq0, YLt0);
+  Result = Builder.CreateSelect(XEq0AndYLt0, NegHalfPi, Result);
+
+  Value *XEq0AndYGe0 = Builder.CreateAnd(XEq0, YGe0);
+  Result = Builder.CreateSelect(XEq0AndYGe0, HalfPi, Result);
+
+  return Result;
----------------
tex3d wrote:

Looking for that, I see three examples (for log, pow, exp), but this doesn't not seem to be done *most* of the time.  Where it is done, it appears to be only applied to the last intrinsic call, not all of the expanded calls.

Do you know whether these are necessary for our intrinsic expansions?  The thing I would expect to be important would be to preserve fast math flags and FPMathTag, which these don't do either.

For TailCall, it doesn't seem like anything we could use or want for DXIL ops anyway.
For other attributes, what attributes would need to be preserved from the original intrinsic call, and how should these apply to all the instructions and DXIL op calls generated?

https://github.com/llvm/llvm-project/pull/108865