[Mlir-commits] [mlir] [mlir][Target] Support Fatbin target for static nvptxcompiler (PR #118044)

[Zichen Lu]

MikaOvO wrote:

> How do you preserve consistency when MLIR_ENABLE_NVPTXCOMPILER_NVFATBIN=OFF?
Isn't libnvptx_compiler.a distributed alongside libnvfatbin_static.a? If yes, then which use-case is this addressing?

My concern is downstream projects that use custom extracted Cuda Toolkit (not extracted libnvfatbin_static.a) will directly cause the original build to fail after this MR if there is no MLIR_ENABLE_NVPTXCOMPILER_NVFATBIN.

Of course, this does not preserve consistency when `MLIR_ENABLE_NVPTXCOMPILER_NVFATBIN=OFF`. 

https://github.com/llvm/llvm-project/pull/118044