feat(#851): aarch64 non-param locals — zero-init stack slots, copy-semantics get/set/tee

[avrabe]

What

The -b aarch64 backend loud-declined every function with a local declaration beyond its params (local.get {i}: non-param locals not yet supported), blocking almost all real functions. This lane (L2 of the v0.51 hub) lowers non-param locals.

Closes part of #851.

How

Lowering (selector.rs):

• Each non-param local (index >= num_params) gets a fixed 8-byte stack slot at [sp, #(idx-num_params)*8]. The frame is sized to a 16-byte SP-aligned multiple and only materialized when the function actually has one or more non-param locals — a params-only function is byte-identical to before (localizing guard).
• Prologue lowers SP and zero-initializes every slot with str xzr (WASM locals default to 0). 64-bit stores clear the whole slot so both i32 and i64 read back intact.
• local.get LOADS the slot into a fresh temp (ldr) — a copy, never the home location — so a later local.set of the same index cannot alias a value already on the value stack. This is the aliasing miscompile the register-by-reference param path would hit.
• local.set stores + pops; local.tee stores without popping.
• Writing a PARAMETER is LOUD-DECLINED (params live in arg registers by reference; homing written params is a later increment) — never a silent write. FP non-param locals fall out honestly via the GP file-check (local types are not threaded to the backend).
• Epilogue restores SP before ret. Interacts correctly with the Feature challenge: add an aarch64 (host-native) backend so output runs and debugs natively on arm64 dev machines #538-cf void blocks: the block End takes the ctrl-pop path (no epilogue), so SP is balanced on every path.

Encoder (encoder.rs): adds str_x_imm/ldr_x_imm (64-bit unsigned-offset load/store) and sub_imm64/add_imm64 (SP adjust), each cross-checked against clang -target aarch64-linux-gnu ground truth.

Gate evidence (RED-first)

scripts/repro/aarch64_locals_851_differential.py compiles aarch64_locals_851.wat with synth compile -b aarch64 and diffs every exported function against wasmtime under two oracles: unicorn (portable/CI) and — on an arm64 host — native MAP_JIT (the real silicon calls the emitted .text through ctypes).

• RED (before): a fix-less binary declines all 8 functions (non-param locals not yet supported / unsupported wasm op: LocalSet/LocalTee); the harness detects the decline and reports RED. Verified.
• GREEN (after): 17/17 cases match wasmtime across all three oracles.

Load-bearing cases (pass a naive lowering, fail a correct one):

• get_set_get_no_alias — local.get must COPY (returns 5, not the 10 a read-by-reference model gives).
• zero_init_read_first — a local read before any write is 0.
• i64_local — the 64-bit slot preserves full width.
• local_across_block — non-param local read/written across a void block + br_if; SP balanced on both paths.

Verification

• cargo test -p synth-backend-aarch64 — 65 passed (incl. 7 new locals tests).
• cargo clippy -p synth-backend-aarch64 --all-targets -- -D warnings — clean.
• cargo fmt — clean.
• Frozen ARM anchors untouched: frozen_codegen_bytes 10/10 (separate backend, ARM codegen has no dep on this crate).
• No regression: aarch64_m2_538_differential.py 182/182 still green.

Coordination

Does NOT touch .github/workflows/ci.yml or scripts/repro/aarch64_matrix.sh (lane L1 owns those — memory + gate vendoring + CI wiring). This lane touches only synth-backend-aarch64 source + its own new differential (aarch64_locals_851.{wat,py}) + doc prose.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L

[codecov]

Codecov Report

❌ Patch coverage is 97.74775% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/synth-backend-aarch64/src/selector.rs	97.38%	5 Missing ⚠️

📢 Thoughts on this report? Let us know!