 Drew T
|
8ba3b82307
|
feat(phase-24): T7 — 3 Fable5-cracked giants ×134 (func_801392FC/8013A530/8013AF20); fleet 65.64->65.75%
- func_801392FC (182, close=2->MATCH): CROSS-BB COMBINE LAW (a backedge-crossing count is a variable,
not a foldable (s16)load; reorg steals the sll into the delay slot) + VOLATILE-FRAME-PARITY (volatile
count-load avoids the cse-common -> combine keep-load -> stale-allocno reload slot). cookbook §36.
- func_8013A530 (204, close=10->MATCH): the RC-6 'reload knot' WAS the pins; the $0-ADD OPAQUE COPY
(iVar7 = fc + zr, zr=$0) assembles to addu $v1,$a1,$zero but doesn't reserve the reg + 3 zero-byte
levers. regalloc.md RC-12/RC-13, §36.
- func_8013AF20 (185, close=15->MATCH): the 'impossible' const-hoist order was libgpu addPrim's 24-bit
BITFIELD store (store_fixed_bit_field emits 0x00ffffff first) + delete the inherited barrier; P_TAG
lifted to engine_types.h (byte-neutral). loop.md, RC-7, §36.
- banked via the §35 recipe (fix_arity_callers --any-proto -> sig_unify -> harvest_verify ->
dedup_propagate --recover). clean fleet check-all 136/136, dedup 1806/0, 0 NON_MATCHING.
- SESSION: 7 giants ×134 total (1 R14-free, 2 Opus-solo, 4 Opus->Fable5); fleet 65.48 -> 65.75%.
gcc-2.7.2 reference tree extended (expmed.c/mips.md/... added). Giant queue (§G) stored for next session.
|
2026-07-04 04:14:43 -06:00 |
|
|
Drew T
|
994368fbcd
|
feat(phase-24): T7b/T5 — 2nd giant func_80138ED0 MATCHED + banked ×134; fleet 65.52→65.56%; §34 idioms
- func_80138ED0 (159 ins, reach-134): Opus applying §32 -> close=21 (all semantics/control-flow/
constants exact, +the giv-init fence lever) -> Fable5Max reading the vanilla gcc-2.7.2 source ->
MATCH (every regalloc/sched class C-reachable, no permuter).
- banked via the standard pipeline (validates it on a real un-reconciled sibling): cast_call_sites
reconciled func_8013914C (u8*,u16*)->(s32,s32)+cast; reconcile_decls a NO-OP (2 data bases
giant-local -> no fleet conflict, confirms no false-positive); harvest_verify ×1 d19c9580 ->
dedup_propagate --recover ×134 (pins/asm body propagates fine). clean fleet check-all 136/136,
dedup 1800->1801/0, 0 NON_MATCHING.
- NEW §31 idioms (cookbook §34 + regalloc.md RC-11 + loop.md): gcc-2.7.2's 3-qty local-alloc SORT
BUG (local-alloc.c:1441 — <=3 qtys allocate in creation not density order; decoy-qty fix) + the
zero-byte asm allocation toolkit (input-only/multi-input/def+use dummies) + the giv-init fence +
gdb-on-cc1. gcc-2.7.2 reference tree completed (18 .c); SETUP §5.6.
- T5 satisfied: reconcile_decls proven on 2 real giants (func_80129CF8 reconcile-path byte-proof +
func_80138ED0 no-op/full-pipeline). CURRENT_PHASE T5 logged.
|
2026-07-03 21:09:38 -06:00 |
|
|
Drew T
|
3b6f2c5f94
|
docs(phase-23): §31 — the gcc-2.7.2 codegen map (4 Fable5 agents read the source)
- docs/gcc-2.7.2-map/{sched,regalloc,loop,cse_expr}.md: source-cited, byte-proven
pass -> residual -> C-lever catalogs (935 lines). cookbook §31 = the index + triage.
- WALLS BROKEN (byte-proven steerable, were "CONFIRMED unsteerable"): §10/§20
hoist-vs-remat (cross-call address-caching), store-vs-load (/s), dbr delay-slot
(D1+S2 fresh-local: func_801770E0 53->49), birthing-boost both directions.
Incidental bank: func_80149374 x134 (fleet 64.86%).
- GENUINE walls -> permuter: S3 chain-priority sink, S11 LUID(x)alloc coupling,
RC-6 pressure-lock, cse 1000-insn table flush.
- CORRECTION (loop agent): tools/reference/gcc-papermario is gcc 2.8.1 NOT 2.7.2
(behavioral biv-elim diff). Vanilla gcc-2.7.2 -> tools/reference/gcc-2.7.2/
(gitignored); SETUP §5.6 + §31 flag it. All banked levers stand (match_one-validated
vs the real cc1). Also corrected: spill-slot = declaration order; §25 4th rank rule.
- the payoff: the cheap tier (Opus agents + local model) can now apply compiler-internal
levers by triage-table lookup, without reading 80k lines of source.
|
2026-07-02 20:12:19 -06:00 |
|