[llvm] [Exegesis] CPU selection, when native arch and target mismatch (PR #131014)

[via llvm-commits]

AnastasiyaChernikova wrote:

> > In this case, we are talking about a situation where the user does not set any mcpu. In this scenario, exegesis itself puts the mcpu in native.
> 
> Right. I understand that part.
> 
> > And already during the processing process, the basic mcpu of the architecture on which the assembly is taking place is substituted. I believe that when the mcpu is not set by the user, it is assumed that just the basic one will be used, as it is done in the case of non-cross compilation. That's the functionality I'm offering here.
> > Regarding the first item from the list - I agree. Replaced with the choice of generic.
> 
> I'm not sure basing how `llvm-exegesis` should behave on how the `-mcpu` flag is treated in `clang` is a good idea. Building a scheduling model requires pretty exact knowledge of the CPU it is being run on, which is why I presume the default was `-mcpu=native` rather than `-mcpu=generic` or something else.
> 
> What would be the drawback of simply erroring out if we detect cross compilation with `-mcpu=native` and error out, telling the user to specify a specific CPU?

It seems to me that the logic would be incorrect. In general, the mcpu flag is not required for exegesis assembly in the general case. Making this mandatory for cross-compilation seems strange to me. After all, the main idea of setting this flag is to choose a specific feature.

https://github.com/llvm/llvm-project/pull/131014