[flang-commits] [flang] [mlir] [MLIR] Add cpow support in ComplexToROCDLLibraryCalls (PR #153183)

[Slava Zakharin via flang-commits]

================
@@ -1625,15 +1655,19 @@ static constexpr MathOperation mathOperations[] = {
      genFuncType<Ty::Real<16>, Ty::Real<16>, Ty::Integer<8>>,
      genMathOp<mlir::math::FPowIOp>},
     {"pow", RTNAME_STRING(cpowi),
-     genFuncType<Ty::Complex<4>, Ty::Complex<4>, Ty::Integer<4>>, genLibCall},
+     genFuncType<Ty::Complex<4>, Ty::Complex<4>, Ty::Integer<4>>,
+     genComplexPow},
----------------
vzakhari wrote:

Please look at how this case works:
```
    {"pow", RTNAME_STRING(FPow4i),
     genFuncType<Ty::Real<4>, Ty::Real<4>, Ty::Integer<4>>,
     genMathOp<mlir::math::FPowIOp>},
```

Depending on the `mathRuntimeVersion` Flang either generates a call to `_FortranAFPow4i` or an `mlir::math::FPowIOp` operation.  You can do the same for `_FortranAcpowi` vs `complex.pow`, and then handle `complex.pow` any way you wish later in the pipeline. So for AMD GPU you may convert it to the ocml calls, and otherwise you may convert it to `_FortranAcpowi` late in Flang pass pipeline. This way, we get all the benefits of not having a call with side effects at MLIR level, and we can apply folding/canonicalization to `complex.pow`.

https://github.com/llvm/llvm-project/pull/153183