[dileepr1]

Summary

• Required, together with fix(clr): instrument host objects with a ccache-visible ASan flag rocm-systems#14327, to avoid the ASan hipFree() crash when host-ASAN is enabled through ci: declare build variant test triggers as data #8455 and feat(ci): enable focused host-ASAN presubmit for runtime components rocm-systems#10558.
• Host-ASAN passes -Xarch_host -fsanitize=address. ccache drops -Xarch_* unless the compile also has -arch, so hsa-runtime never sees a plain -fsanitize=address and __has_feature(address_sanitizer) stays false.
• SANITIZER_AMDGPU is then off. GPU pool allocations do not remap the first page to host memory, the ASan chunk header in that page is not a coherent host atomic, and hipFree reports attempting free on address which was not malloc()-ed.
• libhsakmt already adds the plain flag for HOST_ASAN. This adds the same flag for the runtime, only when THEROCK_SANITIZER is HOST_ASAN, and only for C/C++. The global -Xarch_host form is unchanged, so device compiles still do not see -fsanitize.
• pre_hook_ROCR-Runtime.cmake was removed in Remove obsolete ROCR LibElf pre-hook #9033 with the obsolete LibElf prefetch. The file is restored with only this flag block, because the subproject loader includes pre_hook_<target>.cmake when it exists.

libamdhip64 itself stays link-only ASan until CLR passes a plain flag that ccache keeps. That change is ROCm/rocm-systems#14327. The local hipFree check passed only with both.

Test plan

• HOST_ASAN build of ROCR compiles hsa-runtime with plain -fsanitize=address and the hsa_amd_memory_pool_allocate trampoline references __asan_init
• A non-HOST_ASAN build does not compile ROCR with -fsanitize=address from this hook
• With the companion CLR change, hipMalloc/hipFree on gfx942 (HSA_XNACK=1) no longer bad-frees

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[shwetagkhatri]

LGTM.
I have a follow up ROCr PR - ROCm/rocm-systems#14925 - after this one lands.

Approving this one - since TheRock HOST_ASAN build silently produces an un-instrumented ROCr and the breakage shows up much later as a bogus - attempting free on address which was not malloc()-ed. This PR restores the flag, which turns SANITIZER_AMDGPU on and activates the head-page swap that was supposed to be running. So this is a corrective fix for TheRock.

[ApurvMishra-amd]

Seems like it should be resolved as #9502 reported this exact ccache bug. It was closed by #9506, which moved the Docker image to ccache 4.12.1, and #9589, which repinned the CI images. Both merged on Oct 29, two days after this PR was opened, and the PR doesn't mention them. The branch is 72 commits behind main.
Please, try to repro and see this issue still exists using TheRock's main branch. If not resolved, go ahead and merge it, else it can be closed in favour of Shweta's rocm-systems#14925.

[dileepr1]

Seems like it should be resolved as #9502 reported this exact ccache bug. It was closed by #9506, which moved the Docker image to ccache 4.12.1, and #9589, which repinned the CI images. Both merged on Oct 29, two days after this PR was opened, and the PR doesn't mention them. The branch is 72 commits behind main. Please, try to repro and see this issue still exists using TheRock's main branch. If not resolved, go ahead and merge it, else it can be closed in favour of Shweta's rocm-systems#14925.

Thank you, looks like the above tickets should take care of the issue.
I'll confirm and close.