[codex] Roll bring-up closure evidence onto merged submodule pins#115
Conversation
The validated runtime closure lane now spans three repos: LLVM accepts the final-symbol TPCREL low relocations, QEMU patches the relocated branch sites and preserves halfword L.BSTART decode units, and the canonical QEMU skill records the new gate and recurring triage rule. This superproject pin captures the tested submodule SHAs and refreshes the skills-evolution record so future runs can reproduce the same closure point without reconstructing the session history. Constraint: The superproject must pin only the verified submodule heads and explicit skills-evolution evidence from this lane Rejected: Leave the pinning implicit in detached submodule worktrees | the validated state would not be reproducible from the superproject alone Confidence: high Scope-risk: moderate Reversibility: clean Directive: Re-run the call/ret runtime contract before moving any of these three pins independently Tested: bash /Users/zhoubot/linx-isa/tools/ci/check_linx_callret_crossstack.sh; CLANG=/Users/zhoubot/linx-isa/compiler/llvm/build-linxisa-clang/bin/clang TARGET=linx64-linx-none-elf OUT_DIR=/Users/zhoubot/linx-isa/avs/compiler/linx-llvm/tests/out-linx64 /Users/zhoubot/linx-isa/avs/compiler/linx-llvm/tests/run.sh; python3 /Users/zhoubot/linx-isa/avs/qemu/run_callret_contract.py; direct execution of the four touched lld RUN lines Not-tested: Full top-level dirty worktree outside the repinned submodules Co-authored-by: OmX <omx@oh-my-codex.dev>
The submodule fixes for LLVM, QEMU, Linux, and linx-model are now merged, and the superproject needs to pin those canonical SHAs while carrying the matching AVS, checklist, and documentation updates that describe the current bring-up state. This also keeps the QEMU system closure edits and the doc-site refresh in one reproducible superproject rollup. Constraint: Superproject gate truth must pin merged upstream SHAs rather than temporary feature-branch tips Rejected: Leave the superproject on pre-merge pins | breaks reproducibility and separates docs from the merged code state Confidence: medium Scope-risk: broad Directive: Re-run strict cross-repo closure after every repin before promoting any gate-status doc to canonical truth Tested: python3 /Users/zhoubot/linx-isa/avs/qemu/run_tests.py --suite system --timeout 20 -v; bash /Users/zhoubot/linx-isa/avs/qemu/check_system_strict.sh Not-tested: Fresh full strict cross-repo closure after this merged-pin rollup
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request implements a cross-stack transition to fused direct-call source syntax (BSTART.STD CALL, ..., ra=...) and advances the scalar closure lane for LinxISA. Key changes include updating the LLVM backend tests, QEMU contract gates, and runtime startup assembly to enforce the new fused header policy, alongside significant documentation updates and the addition of an automated ISA reference generator. Feedback highlights a fragile string check in the assembly parser, an I/O inefficiency in the test runner due to removed caching, and the need for architectural atomicity in the freestanding atomic builtins using native AMO instructions.
| continue | ||
| if in_file_scope_inline_asm: | ||
| continue | ||
| if "CALL" not in line or "ICALL" in line or "BSTART" not in line: |
There was a problem hiding this comment.
The fast-path check "ICALL" in line is fragile because it will cause valid BSTART CALL instructions to be skipped if the string "ICALL" appears anywhere on the line, such as in a comment. Since ASM_CALL_RE is designed to specifically match the desired BSTART CALL patterns and naturally excludes ICALL due to its word boundaries and prefixes, this manual check is redundant and potentially incorrect.
| if "CALL" not in line or "ICALL" in line or "BSTART" not in line: | |
| if "CALL" not in line or "BSTART" not in line: |
|
|
||
| def _pto_kernel_src(name: str) -> str: | ||
| rel_path = PTO_KERNEL_PATHS.get(name) | ||
| rel_path = _load_pto_kernel_catalog().get(name) |
There was a problem hiding this comment.
This change removes the module-level cache for the kernel catalog, resulting in the catalog being reloaded from disk on every call to _pto_kernel_src. This is inefficient, especially when called within loops or list comprehensions (e.g., in _extra_sources_for_suite). It is recommended to cache the result of _load_pto_kernel_catalog() to avoid repeated I/O.
| (void)failure_memorder; | ||
|
|
||
| uint32_t old = *ptr; | ||
| linx_u32 old = *ptr; |
There was a problem hiding this comment.
The implementation of __atomic_compare_exchange_4 (and other builtins in this file) is not atomic. In a multi-core or multi-threaded environment, this will lead to race conditions. Since the LinxISA supports Atomic Memory Operations (AMOs) like LR.W/SC.W and SWAPW, these builtins should be implemented using those instructions via inline assembly to ensure architectural atomicity.
1 participant
What changed
Why
The submodule fixes are merged upstream, so the superproject now needs the canonical pins and matching bring-up evidence in one reproducible rollup.
Impact
This makes the current bring-up state reproducible from the superproject again and moves the next blocker to the remaining strict-closure failures rather than local QEMU system regressions.
Linked submodule PRs
Validation
python3 avs/qemu/run_tests.py --suite system --timeout 20 -vbash avs/qemu/check_system_strict.sh