[Mlir-commits] [mlir] [MLIR][GPU] Lower remaining Math ops in the XeVM pipeline (PR #213874)

[llvmlistbot at llvm.org]

LouisLu060211 wrote:

> So I tested the exact reproducer from #213583 with #184370, this PR, and both changes combined. With both changes applied,` math.isinf` was successfully lowered inside the device module: `%3 = "llvm.intr.is.fpclass"(%2) <{bit = 516 : i32}> : (f32) -> i1`
> 
> This indicates that the pipeline placement introduced byt the #213874 is necessary, but is not sufficient on the current main without an `math::IsInfOp` conversion pattern. The complete fix, as I understand, requires both this PR and the `IsInfOpLowering `indtroduced by #184370.
> 
> For this PR it might make sense to depend on #184370.

Thanks @maxxlifem, I agree with your analysis. However, I think this PR is necessary but not sufficient without #184370.

https://github.com/llvm/llvm-project/pull/213874