feat(hotswap): build, package, and sanity-check the hsa-hotswap tool

[lamb-j]

ISSUE ID: #6096

Summary

TheRock-side integration of the relocated hotswap HSA tool. The hotswap HSA_TOOLS_LIB tool has moved out of comgr (libamd_comgr_hotswap_tool.so, removed by ROCm/llvm-project#3007) into rocm-systems projects/hotswap (libhsa-hotswap.so); comgr now provides only the amd_comgr_hotswap_rewrite API. This PR wires TheRock to build, package, and sanity-check the relocated tool.

What this PR does

• compiler/CMakeLists.txt: drop the now-removed HOTSWAP_BUILD_TOOL args; keep COMGR_ENABLE_HOTSWAP_TRANSPILE (comgr keeps the rewrite API).
• core/CMakeLists.txt + core/artifact-core-runtime.toml: declare the hsa-hotswap subproject (rocm-systems projects/hotswap, deps amd-comgr + ROCR-Runtime) and package libhsa-hotswap.so into the core-runtime artifact.
• tests/test_rocm_sanity.py: add test_hotswap_tool_loads — a minimal check that, when hotswap is enabled (libamd_comgr.so exports amd_comgr_hotswap_rewrite), libhsa-hotswap.so is packaged and loads cleanly under ROCr. rocminfo triggers hsa_init, which is when ROCr dlopens HSA_TOOLS_LIB tools; the allowlist is gfx1250 -> gfx1250 only, so the tool stays inert on other targets and rocminfo must still succeed. Skips when hotswap is disabled.

Validation

• ✅ Tool builds, packages, and loads cleanly under ROCr; inert on gfx942 with no regressions (originally verified via a full hip-tests HSA_TOOLS_LIB hookup — now reduced to the lightweight rocminfo sanity check).
• ❌ Actual gfx1250 B0→A0 transpilation is out of scope — there is no gfx1250 runner in CI.

Dependencies

amd-llvm: ✅ satisfied by main. The required ROCm/llvm-project#3007 (remove COMGR hotswap HSA tool) is already in TheRock's compiler/amd-llvm pin (aa451e1f, landed via #6155). No amd-llvm change in this PR.

rocm-systems → temporary pin 75469b9f (removed before merge)

One remaining temp pin. Carries the merged hotswap work not yet in TheRock's rocm-systems pin:

PR	Status	Description
ROCm/rocm-systems#7629	merged to develop	Build/install libhsa-hotswap.so, link hsa-runtime64, re-key HSA_TOOLS_LIB tool name, ISA derivation + tests, OnUnload use-after-free fix, opt-in HSA_HOTSWAP_VERBOSE logging
ROCm/rocm-systems#7715	merged to develop	Restrict HotSwap forwarding to gfx1250 -> gfx1250

Drop this temp pin once the rocm-systems SMP bump brings #7629 + #7715 into TheRock's pin.

Notes

• The hsa-hotswap subproject resolves the amd_comgr + hsa-runtime64 CONFIG packages from its deps; therock_test_validate_shared_lib hard-fails the build if libhsa-hotswap.so isn't produced, surfacing packaging misses.

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

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

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

[therock-pr-bot]

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

[davidd-amd]

What are the kfd tests and is the option wired?

[davidd-amd]

LGTM - just to be clear we are moving existing functionality not adding new functionality. It wasn't clear to me why we need to do this - i.e. what problem it solves but I am not as familiar with these tools.

[lamb-j]

Paused — superseded by an upstream design change.

ROCm/rocm-systems#7921 ("feat(rocr): integrate HotSwap into ROCR loader") pivots the hotswap design: HotSwap becomes native in the ROCR loader (no HSA_TOOLS_LIB), the projects/hotswap plugin and libhsa-hotswap.so are removed, and ROCR lazy-dlopens libamd_comgr.so directly to call amd_comgr_hotswap_rewrite.

That obsoletes this PR's approach:

• The hsa-hotswap subproject + libhsa-hotswap.so packaging + artifact-descriptor capture — no longer applicable (projects/hotswap is deleted upstream).
• The HSA_TOOLS_LIB / rocminfo sanity test — no longer the mechanism.

What still holds (and is already on main): libamd_comgr.so built with the rewrite API (COMGR_ENABLE_HOTSWAP_TRANSPILE, via the #3007 amd-llvm bump #6155) — which is exactly what ROCR dlopens under #7921.

Holding this PR as draft pending #7921's direction. If #7921 lands, this PR will likely be closed (TheRock needs nothing further for the tool side); if #7921 stalls, the plugin approach here remains the fallback.