Adjustments for (upcoming) LLVM-23 and Clang's new offload driver

[linehill]

This patch updates chipStar to work on non-forked, upstream LLVM-23. The recent Clang has switched to use its new offload driver by default for HIP compilation and chipStar is adjusted to it. There is a still an option to use the old while it still around in Clang.

That said, chipStar doesn't yet work on the recent Clang until the following upstream patch lands:
llvm/llvm-project#187655.

[colleeneb]

/run-aurora-ci

[colleeneb]

/run-aurora-ci

[pvelesko]

We already have #1194

This PR is missing CI for llvm 23 @linehill anything wrong with the existing PR?

[linehill]

We already have #1194

This PR is missing CI for llvm 23 @linehill anything wrong with the existing PR?

That one haven't seen progress for a while and its changes are based on an old LLVM revision that wouldn't work on the more recent one.

[pvelesko]

That one haven't seen progress for a while and its changes are based on an old LLVM revision that wouldn't work on the more recent one.

Yeah I got it to the stage where things were working but since LLMV is still getting commits in I didn't see the point in doing further work until LLVM side PRs are merged.

[linehill]

Yeah I got it to the stage where things were working but since LLMV is still getting commits in I didn't see the point in doing further work until LLVM side PRs are merged.

Fair enough. We (PoCL) have a goal to have the chipStar working on non-forked Clang before the LLVM-23 release. Given your on-going LLVM upstream efforts (MacOS and SPIR-V BE support for HIPSPV and maybe more?), I'm concerned if we are going to miss that set goal if we need to wait them to be merged first.

[pvelesko]

@linehill sorry I missed this comment!
I don't think any of my upstream LLVM PRs are blocking anything? Or am I missing something?

[linehill]

@linehill sorry I missed this comment! I don't think any of my upstream LLVM PRs are blocking anything? Or am I missing something?

No, this PR is based on the recent state of the LLVM main branch - last time tested on a LLVM revision from the last week.

[pvelesko]

llvm/llvm-project#183991 MacOS support has been merged.

We don't have to wait on the others I don't want to block your pocl progress

[linehill]

Rebased.

llvm/llvm-project#183991 MacOS support has been merged.

I tested this PR on Ubuntu 24.04/x86_64 on the recent LLVM which includes this LLVM PR - all seems to be good.

We don't have to wait on the others I don't want to block your pocl progress

If the CIs turn green, would this PR be good for landing?

[pvelesko]

@linehill Can you check if your rebase includes llvm/llvm-project#190443?

[pvelesko]

LGTM

[linehill]

@linehill Can you check if your rebase includes llvm/llvm-project#190443?

No, but I'll check it.

[pvelesko]

Shouldn't affect anything. I'll make a new release candidate

[linehill]

Thanks for the review and merge.

@linehill Can you check if your rebase includes llvm/llvm-project#190443?

No, but I'll check it.

Checked on the LLVM featuring the linked PR locally: all seems to be good.

[pvelesko]

Yeah I got it to the stage where things were working but since LLMV is still getting commits in I didn't see the point in doing further work until LLVM side PRs are merged.

Fair enough. We (PoCL) have a goal to have the chipStar working on non-forked Clang before the LLVM-23 release. Given your on-going LLVM upstream efforts (MacOS and SPIR-V BE support for HIPSPV and maybe more?), I'm concerned if we are going to miss that set goal if we need to wait them to be merged first.

btw, we can use the non-forked clang since LLVM 22.0 + Native SPIRV backend. All tests pass.

@linehill