[SYCL] Do module cleanup in sycl-post-link even if there were no splits

[sarnex]

This allows for unused symbol removal through GlobalDCE that would happen if there were splits. Right now we have a mismatch where split modules have unused symbols removed but copied modules (no splits) do not.

[MrSidims]

I'm not really an expert in the modified code. Can we have ; CHECK: _ZTSZ4mainEUlT_E0_ to give me a hint, that the change won't affect kernels?

[maksimsab]

@sarnex Could you please give a clue why internal linkage is replaced with weak_odr?

[sarnex]

@maksimsab sure, it's because those globals are now getting optimized out by globalDCE. the point of those tests is to test something else, so i changed the linkage as to maintain the purpose of the test. note that if there were any splits (like if they used fp64), they would already be optimized out in the splits, so it's just because there happen to be no splits that they weren't optimized.

@MrSidims sure, i will add this

[sarnex]

@AlexeySachkov Hey, do you mind taking a look at this one to make sure it's not insane? Thanks

[asudarsa]

Thanks Nick. LGTM. Seems like something that was missed earlier.
One question: if we are now doing cleanup in ALL cases, can we just call the cleanup routine in one site (in sycl-post-link, right after calling getDeviceCodeSplitter)?
I will approve nevertheless. This is just something that you can address here if possible.

Thanks

[sarnex]

@asudarsa We should be able to do that but I will do it in a follow up change if that's okay.

[sarnex]

@intel/llvm-gatekeepers can we merge please? the original cuda failure is not related and i tried to rerun it but it seems no runners picked it up

This request was automatically failed because there were no enabled runners online to process the request for more than 1 days.

[sarnex]

@intel/llvm-gatekeepers CUDA passed on rerun so all CI is passing now, can we merge? Thanks for rerunning the CI!

[sarnex]

@asudarsa It looks like even if we move the cleanup call outside of ModuleSplitter, we will still need to call it in multiple sites due to the special ESIMD handling in handleESIMD, so I'm not sure we get any code cleanup benefit, so I'm planning to leave the code as is if that's okay.

[asudarsa]

Ah. Ok.....Thanks for considering it Nick. It was again not a PR blocker. Thanks again for fixing this.