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

[llvmlistbot at llvm.org]

maxxlife wrote:

I tested this PR at commit 9a0508ed3b13 using the exact input and full reproduction command from #213583 
The pipeline still fails during kernel serialization:
`error: cannot be converted to LLVM IR: missing 'LLVMTranslationDialectInterface' registration for dialect for op: math.isinf
error: An error happened while serializing module.`

The dumped `gpu.module` shows that the surrounding operations have already been converted to llvm.*m while math.isinf remains uncoverted:
`%3 = "math.isinf"(%2) : (f32) -> i1`

I also tested #184370 at commit 56e8d0f1a6e3. Its direct MathToLLVM test lowers math.isinf to llvm.intr.is.fpclass, and the full #213583 reproduction command completes with exit code 0 on that PR's older base revision.

It therefore appears that adding `createConvertMathToLLVMPass()` to the XeVM pipeline is necessary but not sufficient on the current main, because current MathToLLVM pattern set does not include `math::IsInfOp`. The complete fix appears to require the `IsInfOpLowering` from #184370 together with the pipeline placement introduced by this PR.

I will try to apply lowering from #184370 to the #213874 and will rebuild `mlir-opt`. I think, it will work.

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