[Mlir-commits] [mlir] [MLIR] Fix `ComplexToStandard` lowering of `complex::MulOp` (PR #119591)

[Jordan Rupprecht]

rupprecht wrote:

I found some tests in TF that fail due to this change. See cases MulComplex64SpecialCases and MulComplex128SpecialCases: https://github.com/tensorflow/tensorflow/blob/913caa8102bc4a7940f1aa205803e1188737c301/tensorflow/core/kernels/mlir_generated/gpu_binary_ops_test.cc#L1052

The failures are all inf/nan types of differences, so it seem expected based on this patch, e.g.

```
tensorflow/core/framework/tensor_testutil.cc:213
Value of: IsClose(Tx[i], Ty[i], typed_atol, typed_rtol)
  Actual: false ((-inf,-nan) not close to (nan,nan))
Expected: true
i = 65 Tx[i] = (-inf,-nan) Ty[i] = (nan,nan)
```

I expect @akuegel is best suited to explain if we still need this behavior, but may be on vacation right now.

https://github.com/llvm/llvm-project/pull/119591