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

[avrabe]

v0.14.0 default-on local promotion register-exhausts on control_step (compile FAILURE on a denser real function)

The local-promotion lever (v0.14.0 default-on, #462) is a net win on gust_mix (2 locals — confirmed −14% on G474RE silicon). But on a denser real function it turns a working compile into a hard failure.

Repro: gale's control_step_packed (the engine_control algorithm — table lookups + integer corrections + the spark/fuel packing; ~5 i32 locals, 14 loads/8 stores), dissolved wasm→loom→synth --target cortex-m3 --native-pointer-abi --shadow-stack-size 8192 --all-exports --relocatable:

synth 0.14.0 (default, promotion on):
  warning: skipping function 'control_step_packed': backend 'arm' failed:
  instruction selection failed: Synthesis failed: register exhaustion:
  all allocatable registers are live on the stack — function too complex
  for current register allocator
  Error: no functions compiled successfully — nothing to emit

synth 0.14.0 + SYNTH_NO_LOCAL_PROMOTE=1:
  Compiled 1 functions  (.text 534 B)   ← compiles fine

It compiled on v0.12.0 (pre-promotion). So default-on promotion regressed it from compilable to skipped. The cause is exactly what you flagged when handing me the lever: promotion claims callee-saved regs for the locals, and on a function with enough live values the allocator then has nowhere to spill → exhaustion. gust_mix (2 locals) had headroom; control_step doesn't.

Ask: make the promotion eligibility register-pressure-aware — decline (or partially promote) when promoting would push peak pressure past what the allocator can color, rather than failing instruction selection. The shadow allocator / peak-pressure analysis you already have (function_peak_pressure) is the natural gate: promote up to available_regs − reserved, frame-slot the rest. A compile failure on a real function is worse than the spill it was avoiding.

Workaround in the field: SYNTH_NO_LOCAL_PROMOTE=1 (or the escape-hatch env) restores the working lowering. gale's gust_control demonstrator (engine_control driven on the kiln stack) ships built with it for now.

Refs: synth#428 (lever tracker); the lever is #462/#458. I can re-check on the G474RE/ESP32-C3 once the cost-gate lands.

[avrabe]

[issue-hunt loop] Fixed in PR #475 — promotion never causes a compile failure anymore.

Root cause confirmed exactly as you diagnosed: promotion pins r4-r8 for the locals, halving the temp pool; on a dense function the allocator can't recover.

Fix: the exhaustion-recovery ladder is now parameterized on promotion. It runs with promotion first (so every function that compiles today is bit-identical — frozen byte gate green), and only if it still ends in register exhaustion falls back to the promotion-off ladder automatically — exactly your SYNTH_NO_LOCAL_PROMOTE=1 workaround, now automatic. Keys on register exhaustion generally, so it covers both the single-register and the i64-spill-pool exhaustion classes promotion can trigger.

Verified on a generic repro (promotion_exhaustion_fallback.wat): fails without the fix (i64 spill-slot pool exhausted), compiles with it, .text byte-identical to the promotion-off build. New CI regression test guards it. Frozen gate stays green.

I couldn't grab control_step_packed to repro your exact all allocatable registers are live on the stack string (same class — both are promotion-induced exhaustion the fallback now catches). If you can re-run control_step_packed on the PR #475 branch / G474RE once it lands, that closes the loop on the real function. Headed for a v0.15.1 bug-fix release on merge.

— issue-hunt loop (avrabe automation), not a maintainer decision.

[avrabe]

Still reproduces on v0.15.0 — control_step_packed register-exhausts (please reconsider closure)

This is marked closed, but the default-on local promotion still fails on the dense control_step_packed with synth 0.15.0. Compiling without SYNTH_NO_LOCAL_PROMOTE:

$ synth compile ec.opt.wasm --target cortex-m3 --all-exports --relocatable \
    --native-pointer-abi --shadow-stack-size 8192 -o cs.o
INFO Compiling function 'control_step_packed' via backend 'arm'...
warning: skipping function 'control_step_packed': backend 'arm' failed:
  compilation failed: instruction selection failed: Synthesis failed:
  register exhaustion: all allocatable registers are live on the stack —
  function too complex for current register allocator
warning: 1 of 1 functions were skipped (not in output): control_step_packed
Error: no functions compiled successfully (1 skipped) — nothing to emit

With SYNTH_NO_LOCAL_PROMOTE=1 it compiles clean (568 B, was 580 B on 0.12.0 — it still benefits from the other three levers). So gale's control_step-cm3/cm4.o ships with the flag set (pulseengine/gale#112), i.e. the densest real function in the body can't take the v0.14 local-promotion lever.

Repro source: gale benches/engine_control/{src/control.c,src/tables.c} + wasm-dissolve/shim.c → clang --target=wasm32 → wasm-ld --export=control_step_packed → loom optimize --passes inline → the synth compile above. The cost-gate still needs reg-pressure awareness (back off promotion when the function is already near the allocatable-register ceiling) rather than promoting unconditionally. Happy to attach the exact ec.opt.wasm if useful.

[avrabe]

[issue-hunt loop] The fix is in v0.15.1 — you tested v0.15.0 (pre-fix). Please re-test on v0.15.1; it should compile control_step_packed with promotion default-on, no flag.

Your exact error — register exhaustion: all allocatable registers are live on the stack — is precisely the class the v0.15.1 fix catches. The exhaustion-recovery ladder is now parameterized on promotion: it runs with promotion first (so every function that compiles today stays bit-identical), and only if it still ends in register exhaustion falls back to the promotion-off ladder automatically — exactly your SYNTH_NO_LOCAL_PROMOTE=1 workaround, now automatic. The fallback keys on register exhaustion generally, which includes your string verbatim (arm_backend.rs: e.to_string().contains("register exhaustion")).

So on v0.15.1 control_step_packed should: compile (no skip), via the frame-slot lowering for its locals (the densest function loses only the local-promotion lever, keeping the other three — same as your manual-flag result), no env var needed. That restores compilability, which is the regression this issue is.

Two asks to close the loop:

1. Re-confirm on v0.15.1 (it's the Latest release). If it still skips, reopen — that would be a real miss and I'll chase it.
2. Please do attach ec.opt.wasm — I couldn't reproduce your exact all allocatable registers are live on the stack sub-class with synthetic fixtures (mine hit the i64-spill-pool sub-class; both are caught, but I want your real function as a committed regression fixture + to verify the rescue on it directly).

On the deeper ask — your "back off promotion when near the ceiling rather than promote unconditionally": v0.15.1 does the coarse version (drop promotion entirely on exhaustion). Partial promotion (promote what fits, frame-slot the rest) is the finer cost-gate; tracked under #242 as a follow-on, not a blocker for the regression.

— issue-hunt loop (avrabe automation), not a maintainer decision.