[newling]

rocJITsu must distinguish ordinary scalar registers from trap-temporary registers (TTMP0–15). Although both are named by scalar operand selectors, TTMP selectors 108–123 refer to storage owned directly by each wave; they must never be interpreted as offsets into that wave’s ordinary SGPR allocation.

Most of this distinction is already implemented on develop. #9844 introduced wave-private TTMP storage and routed scalar operand, address-calculation, scalar-memory completion, checkpoint, and debugger paths through it. #10345 completed the GFX12 launch-state initialization. Those changes superseded most of the broader implementation proposed in #9578, but two paths still treated a TTMP selector as an ordinary SGPR number.

Before executing s_setpc_b64 or s_swappc_b64, the compute unit checks whether the target is zero. That check decoded the instruction’s raw selector and read sgpr_base + selector directly. For a target held in a TTMP pair, it therefore read unrelated physical SGPR storage and could halt the wave even though the actual target was nonzero. This change reads the already-decoded source operand through the normal register-access interface, which resolves ordinary SGPRs, TTMPs, and other scalar operand kinds according to their architectural meaning.

The race detector had a related bookkeeping problem. It registered every scalar-memory load destination as an ordinary GLOBAL_TO_SGPR event. A load into TTMP does not produce the ordinary-SGPR callbacks needed to consume that event, and selector 108 can lie beyond the event table for a wave with 104 ordinary SGPRs. This change keeps TTMP destinations out of the ordinary-SGPR event table. Modeling pending load hazards for TTMP consumers can be added separately rather than representing them incorrectly as SGPR hazards.

Both regressions were reproduced on unmodified develop: the indirect-PC test followed the wrong zero-target path, while the race-detector test terminated with a segmentation fault. With this change, both pass. Ten related TTMP storage, checkpoint, translation, and GFX12 launch-state tests pass, as do all 3,289 enabled non-benchmark C++ tests; two tests are skipped. The race-detector shared library builds successfully, and formatting checks pass.

This is a standalone follow-up to #9844 and #10345, and it carries forward the remaining useful corrections identified in #9578. The wave-owned observation work in #10030 is related but is not a prerequisite. #10467 currently touches the same indirect-PC pre-check and will need the equivalent selector-aware read if it lands first.

Related: #9577

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 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.

🙋 Wish to Override Policy?

[therock-pr-bot]

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