ILC: Add property for target processor tuning

[sweemer]

Description

Currently there is the IlcInstructionSet property that works (I assume) in a similar way to the -march option for GCC and Clang, but I can't find any ILC property that corresponds to the -mtune option. Is target processor tuning functionality available in ILC? If so, can a property be created to configure it? If not, can it be added?

Also, is there a property that can be set to show all the ILC args used for compilation such as --instruction-set? I tried setting IlcVerboseLogging but it doesn't print out the ILC args despite printing out a lot of other information.

Reproduction Steps

N/A

Expected behavior

Target processor tuning is available

Actual behavior

Only target processor architecture selection is available

Regression?

No response

Known Workarounds

No response

Configuration

No response

Other information

No response

[dotnet-policy-service]

Tagging subscribers to this area: @agocke, @sbomer, @vitek-karas
See info in area-owners.md if you want to be subscribed.

[MichalStrehovsky]

Native AOT can only add this if the underlying code generator supports this. I don't think RyuJIT supports this. I'll move this to the codegen area path so that relevant feature team can consider/decide this.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[EgorBo]

We generally intentionally avoid any kind of processor-specific tuning as it complicates test matrix a lot for questionable benefits. So yeah, JIT doesn't have -mtune. The only thing we do in cpu detection is to disable AVX512 for old Intels where it leads to throttling.

[sweemer]

@EgorBo Thanks for the explanation. Does this mean that RyuJIT doesn't consider the current processor's cache hierarchy, branch prediction capabilities, etc. when generating machine code? If so then this could be one potential area for improvement (which I'm sure I'm not the first one to suggest).

@MichalStrehovsky Would you be the right person to answer my second question about how to see all the ILC args used for compilation?

[EgorBo]

@EgorBo Thanks for the explanation. Does this mean that RyuJIT doesn't consider the current processor's cache hierarchy, branch prediction capabilities, etc. when generating machine code? If so then this could be one potential area for improvement (which I'm sure I'm not the first one to suggest).

Correct. We might eventually rely on some CPU-specific details when (if) we implement partial loop unrolling etc. but for now I can't think of a case where any of that could help us. E.g. for the "branch prediction capabilities" we have a more profitable Dynamic PGO.

[jkotas]

how to see all the ILC args used for compilation?

ILC does not have an option to print its own command line. You can find the full command line passed to ILC in the <your project name>.ilc.rsp file under your obj directory.

[sweemer]

Correct. We might eventually rely on some CPU-specific details when (if) we implement partial loop unrolling etc. but for now I can't think of a case where any of that could help us. E.g. for the "branch prediction capabilities" we have a more profitable Dynamic PGO.

So I guess the best approach would be to leverage Dynamic PGO within Native AOT somehow. Looks like there is already issue #94052 about it but no activity for the past year. Is it somewhere on the long-term roadmap?

[JulieLeeMSFT]

Please follow up in #94052 directly. Closing this issue.