ci: save rust-cache on main only; finish LLBC artifact handoff#209
Open
youknowone wants to merge 2 commits into
Open
ci: save rust-cache on main only; finish LLBC artifact handoff#209youknowone wants to merge 2 commits into
youknowone wants to merge 2 commits into
Conversation
Add save-if: github.ref == 'refs/heads/main' to every Swatinem rust-cache step (prepare, cargo-test, pyre-check, cpython-tests, wasm-build). PR jobs now restore the cargo target cache from main's default-branch entry via restore-keys but no longer persist their own. Each PR previously saved a ~500-600 MB target cache per OS/job; summed across the open PR refs this pushed the repository past GitHub's 10 GB Actions cache budget, and the resulting LRU eviction was thrashing the prepared LLBC/cargo entries. Restoring from main keeps PR builds warm without each ref multiplying the footprint. Assisted-by: Claude
The wasm-build job still restored the LLBC set with actions/cache and hard-failed on a cache miss — the one consumer the earlier artifact migration skipped. Switch it to download-artifact (run-scoped, exempt from the 10 GB cache budget / LRU eviction) and drop .pyre-build/ charon-src from its Charon restore path, matching the other consumers. Assisted-by: Claude
|
Warning Review limit reached
More reviews will be available in 48 minutes and 5 seconds. Learn how PR review limits work. Your organization has used up its prepaid credits, and credit purchases are no longer available. Enable the review add-on in the billing tab to keep reviews running — you're only billed for reviews past your plan's rate limits ($0.25/file). ⌛ How to resolve this issue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based credits. 🚦 How do rate limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan refill rate. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, the refill rate gradually slows as usage increases. The highest same-day bursts are limited more strictly. Please see our Fair Usage Limits Policy for further information. ℹ️ Review info⚙️ Run configurationConfiguration used: Organization UI Review profile: ASSERTIVE Plan: Pro Run ID: 📒 Files selected for processing (1)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
youknowone
added a commit
that referenced
this pull request
Jul 3, 2026
Adds save-if: ${{ github.ref == 'refs/heads/main' }} to all five
Swatinem/rust-cache steps (prepare-charon-llbc, cargo-test, pyre-check,
cpython-tests, wasm-build). PR jobs become restore-only and pull main's
default-branch cache via restore-keys, so each PR no longer persists its
own ~500-600 MB target cache and multiplies the repo past GitHub's 10 GB
cache budget.
This is the still-applicable part of #209; that PR's other commit
(LLBC handoff to wasm-build via artifact) is already superseded on main,
where wasm-build restores LLBC through download-artifact.
Assisted-by: Claude
youknowone
added a commit
that referenced
this pull request
Jul 3, 2026
Adds save-if: ${{ github.ref == 'refs/heads/main' }} to all five
Swatinem/rust-cache steps (prepare-charon-llbc, cargo-test, pyre-check,
cpython-tests, wasm-build). PR jobs become restore-only and pull main's
default-branch cache via restore-keys, so each PR no longer persists its
own ~500-600 MB target cache and multiplies the repo past GitHub's 10 GB
cache budget.
This is the still-applicable part of #209; that PR's other commit
(LLBC handoff to wasm-build via artifact) is already superseded on main,
where wasm-build restores LLBC through download-artifact.
Assisted-by: Claude
youknowone
added a commit
that referenced
this pull request
Jul 3, 2026
Adds save-if: ${{ github.ref == 'refs/heads/main' }} to all five
Swatinem/rust-cache steps (prepare-charon-llbc, cargo-test, pyre-check,
cpython-tests, wasm-build). PR jobs become restore-only and pull main's
default-branch cache via restore-keys, so each PR no longer persists its
own ~500-600 MB target cache and multiplies the repo past GitHub's 10 GB
cache budget.
This is the still-applicable part of #209; that PR's other commit
(LLBC handoff to wasm-build via artifact) is already superseded on main,
where wasm-build restores LLBC through download-artifact.
Assisted-by: Claude
youknowone
added a commit
that referenced
this pull request
Jul 3, 2026
… + CALL_ASSEMBLER default-on) (#347) * jit-trace: exclude bridge walks from terminate_no_replay store-commit `terminate_no_replay` keeps a Terminate walk's eager list stores (journal-commit) so the loop-free function portal can return the walk's result without replaying it. A bridge Terminate walk does not go through that portal: its caller resumes the guarded region through the blackhole, which re-applies the stores. Keeping them here as well double-applied them. Gate the predicate on `!is_bridge_trace` so bridge walks roll their stores back before the blackhole replay. Assisted-by: Claude * builtins: route hasattr/getattr attribute-name check through checkattrname hasattr/getattr validated the name inline with their own "hasattr()/getattr(): attribute name must be string" messages; route both through checkattrname (operation.py:41-45), which raises the unified "attribute name must be string, not '<type>'" message setattr/delattr already use. Assisted-by: Claude * optimizeopt/heap: treat runtime-minted mutable descrs as written by every non-elidable call A descr minted at runtime (ei_index unset, u32::MAX) is outside the compute_bitstrings universe, so no call's write bitstring can ever name it; a bitcheck miss proves nothing. Treating the miss as "not written" let optheap keep a getfield cache across a store residual and re-read a stale value (module-global cell fold: 'acc = acc + a; acc = acc + b' dropped the first term). Field and array descr loops in force_from_effectinfo now conservatively invalidate mutable out-of-universe descrs on every non-elidable call. Assisted-by: Claude * jit-trace: record word-ABI void helpers through a sized descr with caller-supplied effect jit_store_name_to_namespace / jit_list_append are extern "C" helpers whose C signature returns a dummy machine word (-> i64, value ignored) but were recorded as plain void residuals. Add make_call_descr_void_word_abi (result_type=Void, result_size=8, distinct interning key) and TraceCtx::call_void_typed_word_abi so a signature-exact backend lowering (wasm call_indirect) can select the (i64xn) -> i64 type and drop the result. The new entry point takes a caller-supplied EffectInfo — these helpers write the heap (namespace cells, list storage), and the opcode-default empty write set let optheap CSE a getfield across the call; both call sites record MOST_GENERAL (graphanalyze.py:60 analyze_external_call for an unanalyzed external writer). The call_*_typed record tail is factored into record_call_with_descr (invalidate-before-record, pyjitpl.py:2683-2684). Assisted-by: Claude * codewriter: widen BhSizeSpec/BhDescr vtable to u64 The vtable field is an ob_type pointer captured in the producing process (the opcode_descrs.bin build script) and crosses the build->runtime serialization boundary. Declared usize, a 64-bit host pointer failed bincode's width check when deserialized on a wasm32 runtime. Store it as u64 and cast at the consumers; the value is an opaque identity word either way (real vtables are re-resolved via type_id -> gc_cache). Assisted-by: Claude * metainterp: extend MC_DIAG guard-to-bridge tallies; gated fbw bridge walk prints MC_DIAG grows 8 -> 18 slots: compile_bridge entry/InvalidLoop/ retrace_requested/arity-giveup (8-11), start_bridge_tracing entry and its early-return reasons (12-17), and stack_almost_full()==true (5, formerly reserved). Under the existing fbw debug-abort flag, the full-body walker also prints a bridge-walk epilogue line (commit state + outcome kind) and a note when a bridge STORE_SUBSCR specialization declines to the generic residual. Assisted-by: Claude * backend-wasm: bridge chaining and CALL_ASSEMBLER fast paths, on by default Bridge chaining (PYRE_WASM_ENABLE_BRIDGES=0 disables): compiled bridges resolve guard exits in-module through a global fail-index space with nested chaining and a REF_HOME_FLOOR-sized ref-home area; declined bridges fall back to host round-trips. WASM_BRIDGES_ENABLED flips to default-true and the dormant-tracer comment block is dropped. CALL_ASSEMBLER (PYRE_WASM_CA=0 disables, now default-on): - alloc/pop frame helpers lower through the residual (i64xn)->i64 call_indirect type family instead of the jit_call trampoline (residual_max_arity forced >= 2 when the CA arm is emitted). - CA callee frames are libc-allocated jitframes registered with the shadow stack and recycled through a LIFO pool, replacing per-call old-gen allocation that forced back-to-back major collections (gc_majors 1704 -> 0 on fib_recursive). - Frames are sized and gcmapped to FRAME_REF_HOME_FLOOR when bridges are enabled, mirroring execute_token. - CA on a loop that already has bridges (or further chaining on a loop whose CA arm ran) is declined (diag slot 14 decl_ca_chain): composing the two on one recursion produces wrong output on recursion_memo_branch/generator_tree_recursion (task #7). Inline fast paths in generated code: - write barrier: test the TRACK_YOUNG_PTRS header flag inline and only call_indirect the helper when set (genop_discard_cond_call_gc_wb). - nursery allocation (PYRE_WASM_INLINE_ALLOC=0 disables): New/ NewWithVtable and constant-length NewArray/NewArrayClear bump nursery_free inline (header word + array length store) and only call the collecting helper on overflow, gated on the new type_alloc_is_plain trait (no destructor, not a weakref), size below max_nursery_object_size, and gc_stress off. Diagnostics: collection_counts on the GC trait (MiniMarkGC override), pyre_gc_minor/major_collections + pyre_jit_set_inline_alloc guest exports, gc_minors/gc_majors in the runner's [jit-stats] line, runner labels for the new MC_DIAG slots. fib_recursive (fib 0..34) 24s -> 2.9s, fannkuch8 13.8s -> 0.56s; wasm suite 168/169 (remaining fail is the pre-existing nbody libm ULP drift). Assisted-by: Claude * check.py: allow bounded float divergence for the nbody wasm bench only The wasm float tolerance is now opt-in per bench via run_bench(..., wasm_float_tol=True) rather than applied to every wasm bench. Only nbody is marked; all other wasm benches require byte-identical output. The comparison gate becomes `backend == "wasm" and wasm_float_tol`, threaded from run_bench through _run_backend_bench. Corrects the wasm_outputs_match root-cause note: the divergence is libm's arch-specific pow building blocks, not FMA fusion (FMA hypothesis tested and refuted). Assisted-by: Claude * ci: save rust-cache on main only, restore-only on PRs Adds save-if: ${{ github.ref == 'refs/heads/main' }} to all five Swatinem/rust-cache steps (prepare-charon-llbc, cargo-test, pyre-check, cpython-tests, wasm-build). PR jobs become restore-only and pull main's default-branch cache via restore-keys, so each PR no longer persists its own ~500-600 MB target cache and multiplies the repo past GitHub's 10 GB cache budget. This is the still-applicable part of #209; that PR's other commit (LLBC handoff to wasm-build via artifact) is already superseded on main, where wasm-build restores LLBC through download-artifact. Assisted-by: Claude * check.py: add wasm to default backends when its target is installed DEFAULT_BACKENDS gains "wasm" whenever `rustup target list --installed` reports wasm32-unknown-unknown, the sole extra prerequisite for building the wasm backend (the wasmtime runtime is embedded in pyre-wasm-runner, not an external tool). On a machine without that target the probe returns false and the defaults stay dynasm,cranelift, so a bare check.py still runs only the native backends there. Assisted-by: Claude * optimizeopt/heap: gate out-of-universe descr invalidation on compute_bitstrings_has_run The runtime-minted-descr invalidation added in 6091829 fired on any descr whose effect_idx is the u32::MAX sentinel. In the unit tests compute_bitstrings is skipped, so every cached descr carries u32::MAX and an unrelated CallMayForce write bit wrongly invalidated it, breaking test_call_may_force_uses_effectinfo_to_keep_unaffected_cached_fields, test_call_may_force_keeps_unaffected_variable_index_array_cache and test_default_pipeline_escaping_call_arg_flush_is_selective. Gate the clause on compute_bitstrings_has_run() in both the field and array loops: after bitstrings run a u32::MAX effect_idx genuinely means an out-of-universe runtime descr (invalidate conservatively); before that it just means the fixture never stamped the slot, so the existing identity fallback is the correct arbiter. Assisted-by: Claude * ci: install wasm32 target on the Linux pyre/check.py leg only check.py now adds wasm to its default backends when wasm32-unknown-unknown is installed. The three pyre/check.py jobs share one steps anchor, so a step guarded by runner.os == 'Linux' installs the target on the ubuntu leg only — it builds and runs the wasm backend there, while macOS/Windows keep no wasm32 target and stay on dynasm/cranelift. wasm output is platform-independent, so exercising it on one OS is enough. Assisted-by: Claude
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
PR #198's macOS CI jobs failed fast at the Check prepared cache hits guard: the prepared LLBC entry that
prepare-charon-llbcsaved was LRU-evicted before the late-scheduled macOS consumers restored it. The repo's Actions cache had exceeded GitHub's 10 GB limit (~10–12 GB acrossmain+ open PR refs), and the dominant consumers are the per-ref Swatinemrust-cachetarget caches (~500–600 MB each × job × ref).#198 already moved the main consumers' LLBC handoff from
actions/cachetoupload/download-artifact(run-scoped, not evictable). This PR cuts the root footprint and finishes that migration.Changes
rust-cache: save only on
main.save-if: ${{ github.ref == 'refs/heads/main' }}on all five Swatinem steps (prepare, cargo-test, pyre-check, cpython-tests, wasm-build). PR jobs become restore-only and pullmain's default-branch cache viarestore-keys, so each PR no longer persists its own ~500–600 MB target cache.main's baseline each run); deps and unchanged crates still come warm frommain.save-if.wasm-build: LLBC via artifact. The one consumer majit: resized ListRepr port, BUILD_LIST residualization, optheap setfield flush, and truth-value lowering fix #198's migration missed — it still restored LLBC with
actions/cacheand hard-failed on a miss. Switched todownload-artifactand dropped.pyre-build/charon-srcfrom its Charon restore path, matching the other jobs.Effect
Removes the per-ref cache multiplication that pushed the repo past the budget, and makes the LLBC handoff eviction-proof across all consumers.
🤖 Generated with Claude Code