Repository navigation
[0.32][Backend] Evaluate register retention and helper scratch reuse #39
Description
Follow-up to #34 and D5. Scalar stack-slot reuse is implemented in 0.31; register retention and helper-scratch reuse are separate work. The maintainer promoted this finite evaluation into the 0.32 completion scope on 2026-10-04. Candidate adoption remains conditional on measured evidence; rejecting a measured candidate and retaining the existing backend is a valid evaluation outcome. The October 3 independent-review waiver covers this cost-research work, but does not waive frozen budgets, checker agreement or required gates.
Choose a bounded local register-retention experiment before considering a general allocator. Treat helper-scratch ownership as an independent decision.
Acceptance:
- Define live-register sets, helper/call clobbers, address-exposed exclusions and conservative fallbacks.
- Test joins, backedges, overlapping operands, calls, wide fixed bytes and outgoing arguments.
- Measure static loads/stores and executed traffic where instrumented; report ELF, cycles and stack independently.
- Preserve source-map/checker claims without deriving lifetimes from offsets alone.
- Record tradeoffs explicitly, preserve fixed budgets, and run backend and CI gates.
Activity
[-][Post-0.31][Backend] Evaluate register retention and helper scratch reuse[/-][+][0.32][Backend] Evaluate register retention and helper scratch reuse[/+]a19q3 commentedon Oct 5, 2026
0.32 evaluation progress; keeping this issue open until the additional replay archive is published.
The retained research baseline already records the first binary-left retention candidate: 1,707 unchanged executions. A second, separately archived local experiment now exercises adjacent scalar
LoadConst/LoadVarproducers and right-operand reuse, with explicit live-register avoid sets. It invalidates retention at calls/unclassified instructions and block boundaries, and preserves stores, parameter spills, wide/resource/borrow/address exclusions and outgoing arguments.Results: 35 identical-IR probe pairs passed (25 save one cycle, ten tie); the complete 6,782-metric comparator has zero regressions; 1,707 paired scheduler traces have four one-cycle improvements and 1,703 ties. The changed executions are valid/rejection paths for
scalar-overlap-64andscalar-overlap-260: one fewer executedLD, one extra register-moveADDI, and unchanged ELF bytes, stores and observed stack.Decision: reject this candidate for the 0.32 default because that measured benefit does not justify another cross-instruction register contract. Retain conservative helper-scratch ownership. This does not establish that broader retention is useless, and no general allocator or scratch reuse is implemented by this evaluation. The candidate is isolated from the implementation branch; no full candidate backend/CI pass is claimed.
The pending archive contains both exact dirty-source replay patches, baseline/candidate main and multi-Script reports, full attribution traces, comparisons and hashes. Publication and final evidence links remain outstanding; this comment is a progress record, not a closure or release claim.
a19q3 commentedon Oct 10, 2026
Completed within the 0.32 implementation scope. This closes the bounded register-retention and helper-scratch evaluation with explicit measured reject/retain decisions; it is not a release or public-deployment claim, and it does not implement a general allocator, dead-store elimination, parameter-spill coalescing or scratch reuse.
t0between adjacent instructions in one basic block, in functions already admitted by conservative scalar-slot allocation above optimization level zero. Calls and unclassified instructions invalidate the fact; block-local state cannot cross joins/backedges; explicit avoid sets protect a retained operand during large-offset loads. Wide/resource/borrow/address-exposed exclusions, parameter spills, all stores and outgoing arguments are unchanged. The pre-implementation contractdocs/CELLSCRIPT_SCALAR_RETENTION_EXPERIMENT.mdis embedded in both replay patches.scalar-overlap-64andscalar-overlap-260. Changed rows show one fewer executedLDand one extraADDImove with unchanged ELF bytes, static stores and observed stack. ELF, cycles and stack are reported separately; per-case values are not frequency-weighted savings.dev, strict-fullbackend(with pinned CKB stateful acceptance) andciat clean6a453d3015784784a27ba1aa0a4b6b1a73db79a8(validation.json); the resumed replay re-verified all twenty archive files and replayed the native comparator at6dba1768with no regression.a19q3 commentedon Oct 10, 2026
Closing as completed: both bounded register-retention candidates measured and rejected, helper-scratch ownership retained, archive published and replay-verified. Evidence: #issuecomment-6099388124.
docs: record #39/#40 closure evidence links in 0.32 implementation map