feat: def-site anchored copies + bitfield member recovery

[macabeus]

Fourth round on kleod:UpdateHUDCounterDisplay:agbcc. diff:18 → MATCH (m2c: noncompile). Board: asmlift 360 vs m2c 342, 0 lost, 6 gained, 0 other flips over 743 rows.

The 18 remaining rows split into exactly two missing capabilities, and the row falls out of building both — hand-edited variants attributed every row up front (pre-init/placement ~12, bitfield signedness ~6; all three edits combined compiled byte-exact before any code was written).

Capability 1 — def-site-anchored constant merge copies (/defsite)

SSA destruction places a merge copy where its CFG edge is, but the asm often materialized the constant earlier: movs r9, #0 at entry ahead of a single-armed overwrite, movs r5, #1 at the top of an arm ahead of a nested if. The new anchorConstCopies structuring option emits such a copy at the const op's original program position and suppresses the edge copies it replaces; the existing empty-then peephole then yields the single-armed positive if the source wrote — polarity included.

Sound by refusal: unnamed-const args only; the def's block must dominate every edge source; ANY in-loop shape declines outright (block dominance is not per-iteration precedence — the /preinit sticky-arm class from #13); the merge variable must name no other value; a const whose anchored sibling could clobber it on a path to its edge keeps the edge placement; and a switch-tree test block carrying an anchored write is never consumed as a discarded test (it re-roots the chain, so the write is emitted before the dispatch). Off by default, enumerated as the differ-refereed /defsite axis crossed with branch sense — it can never cost a match.

This axis alone flipped five synthetic rows carrying the same shape (fcmp, iszero, lor, notb, signum — the v=0; if(...) v=1 boolean-result idiom), on top of taking this row 18→6.

Capability 2 — bitfield members from the symbol map

The map now carries bitfield members (bitWidth + bitOffset, LSB-first), and the (x << a) >> b extract of a struct global's loaded bytes is spelled gSym.field when the map declares a bitfield at exactly those bits. The payoff is typing, not cosmetics: the declared u32 dreamStones : 7 promotes to signed int, so sdiv renders / and recompiles to __divsi3 where the raw-shift spelling stayed u32 (__udivsi3, bls-for-ble — the row's whole 6-point family).

Exactness gates: position, width, and signedness must all match; the window must lie inside the loaded bytes; a signless field never matches; a volatile container refuses the fold; declaredFields seats members by bit cursor and drops anything the u32 f : n no-straddle allocation model cannot reproduce — verified byte-faithful against agbcc (extract shifts, promotion-driven __divsi3, and a cross-byte field all reproduce from the synthesized declaration alone). The provider emits bitfield facts only for little-endian ELFs (EI_DATA), and core enforces the same boundary via capabilities.endianness — its first consumer.

Adversarial rounds (two passes, as this repo's history demands)

Pass 1 found two CRITICALs, both fixed and pinned by tests:

• The fold moved a captured load past stores/calls — g.field re-reads memory at the render position where the asm captured bits at the load's. Fixed with an ordering gate.
• An anchored write in a discarded switch test block was silently dropped while its edge copy stayed suppressed. Fixed in PRE4.

Pass 2 then broke the first fix: fn.blocks is ADDRESS order, not topological order, so a store laid out after the render but executing between load and render slipped past the linear-position scan. The gate is now path-based, reusing the materialization model's own emitPos/memWriteBetween machinery (newly exported from analysis.ts) — the same cycle-aware discipline every other memory read already obeys. Also from the audits: the /no-bitfield fallback axis (the named read recompiles at the declaration's access width; where that diverges the honest shifts are what match, and the differ referees), and the width-8/16 CAST_PATTERNS shadowing is documented in both files and pinned.

Verification

pnpm bench run + regression → 0 lost, 0 missing, 6 gained, 0 other flips (743 rows); stale-check fresh, provenance 63f02b6, dirty: false. pnpm test:offline 697 passing (57 files). pnpm typecheck, pnpm lint (0 errors), pnpm format:check clean. Maps re-vendored: kleod +40 bitfield members, pokeemerald +5099, sa3 +5; the big-endian projects carry none by construction.

🤖 Generated with Claude Code