feat(selector): i64 stack params + pair spill-slot growth — direct-path completeness (#503, #587)

[avrabe]

Closes the two honest-skip classes that kept falcon's i64-heavy helpers out of the ELF on the direct/ARM path. Per-class red→green below; frozen anchors bit-identical.

Class 1 — #503-i64: 64-bit params AAPCS-passed on the stack (Closes #503)

Red (pinned): s_mix/s_align/s_p3/s_call/s_wr in scripts/repro/i64_stack_param_503.wat all loud-skipped on v0.28.0 main with #503: function has N params including a 64-bit (i64/f64) param… / #518/#503: an i64/f64 param is AAPCS-passed past R3 (falcon func_58/func_163/func_164 class).

Fix: new aapcs_param_layout — a full NCRN+NSAA walk over declared widths. Every param is either register-resident (even-aligned pair for wide) or gets a caller-frame NSAA offset (8-byte-aligned for wide). compute_local_layout.incoming_params is now width-aware (offset, is_i64); LocalGet/LocalSet/LocalTee read/write wide slots via I64Ldr/I64Str [sp, #frame+24+nsaa] (12-bit guard covers off+4 for the hi half). param_slots backs only register-resident params. Wide stack params in has-call functions lower too (the caller's frame trivially survives the BL); the #518 decline is narrowed to REGISTER-resident i64 params in frame-backing functions — still a loud skip, still cited by d_call.

Found along the way (silent miscompile, now fixed on BOTH paths): per AAPCS C.5 there is no register back-fill after a stack spill, so p3 of (i64 i32 i32 i32) is stack-passed — the old index_to_reg fallback read it from R3 (= p2) and compiled without warning (s_nar in the fixture). The optimized path had the same width-naive homing (its #518 decline only fires on i64-param reads), so wide-param signatures are now routed to the direct selector in arm_backend.

Green: scripts/repro/i64_stack_param_503_differential.py — 8 shapes × 2 paths (direct --relocatable incl. THM_CALL patching, and default) all match wasmtime under unicorn, exit 0.

Still declined (honest, tracked): outgoing i64 call args past R3 (pop_call_args refusal, cites #503) and register-resident i64 params + call (#518) — neither is a falcon signature case.

Class 2 — #587 direct-path remainder: i64 spill-slot pool exhaustion

Red (pinned): scripts/repro/i64_spill_pool_587.wat (20 simultaneously-live i64 consts, ~16 concurrent pair spills vs the fixed 8-slot pool) loud-skipped with exactly falcon func_60/func_73's message: register exhaustion: i64 spill-slot pool exhausted — function too complex for current register allocator. (#587 itself is closed; this is its direct-path remainder, per the issue's func_60/func_73 bullet.)

Fix: the pool is growable — SpillState is Vec-backed, compute_local_layout reserves i64_spill_slots × 8 bytes, and a new pool-grow recovery retry in arm_backend::select_direct reruns the entire existing sequence (base → spill → param-backing rungs, promotion-on then the #474 promotion-off fallback) with the pool sized from a conservative operand-stack-depth bound (new synth_core::wasm_stack_check::max_depth_bound, +4 transient slots, clamped to 120 = 960 B, 12-bit-friendly). Deliberately last resort: any function that compiled yesterday is produced by exactly yesterday's path — first draft had the rung inside the promote-on ladder and promotion_never_causes_compile_failure_474 caught it changing a compiles-today function's bytes; final ordering keeps it green.

Green: scripts/repro/i64_spill_pool_587_differential.py — hp20 compiles via rung=pool-grow result=ok and matches wasmtime on 4 args × 2 paths, exit 0.

Gates

• cargo test -p synth-synthesis -p synth-cli — all green (46 test binaries, 0 failures), including the frozen byte gates (frozen_fixtures_text_is_bit_identical_oracle_001 etc.) and v0.14.0 default-on local promotion register-exhausts on denser functions (control_step) — compile FAILURE, was fine on 0.12.0; cost-gate needs reg-pressure awareness #474 bit-identity.
• cargo fmt --check + cargo clippy --workspace --exclude synth-verify --all-targets -- -D warnings — clean.
• New unit tests: AAPCS NSAA offset matrix, i64-stack-param lowers, narrow-after-wide reads stack not R3, wide-stack+call lowers / wide-reg+call still declines, pool-grow rescues 20-live-i64, max_depth_bound.
• New CLI integration test i64_completeness_503_587.rs (no skips + pool-grow rung non-vacuity).
• Updated pinned artifacts: test_359_i64_stack_param_errs → test_503_i64_stack_param_lowers; i64_param_518_differential.py moves d_past_r3 from the decline contract to emitted+executes-correctly (full run exit 0); adjacent oracles (stack_args_503, spill_on_exhaust_242, spill_rung_581, high_pressure_i64, i64_load_store_372) all exit 0.
• CI: new isolated i64-completeness-503-587-oracle job runs both differentials.

Refs #587 (direct-path remainder), #518, #242 (VCR-RA), #494.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 97.42647% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/synth-synthesis/src/instruction_selector.rs	96.50%	7 Missing ⚠️

📢 Thoughts on this report? Let us know!