feat: const-CSE rematerialization-avoidance — sound, measures the residency gap (#209/#242)

[avrabe]

The const-CSE pass, built and validated — and the measurement that pins the real blocker

apply_const_cse drops a movw rd,#v that re-materializes a constant already resident in ra and retargets rd's reads to ra. Every rewrite is proven by the liveness analysis (rename_use rewrites uses never defs; safe_cse_uses requires ra still holds v through every read and rd's value to be segment-local), and it only removes materializations (pressure never rises) — so unlike the mla fusion (#277) it cannot regress on-target.

Validated sound

Behind SYNTH_CONST_CSE=1 (default-off). With it on, all three differential fixtures are result-identical: control_step 0x00210A55, flight_seam 0x07FDF307, div_const 338/338.

The measured finding (the point)

It fires zero times on the real flat_flight — 170→170 insns, 26→26 movw, byte-identical. Because the greedy selector already destroyed the residency it needs: by the second movw #0x7e, the register holding the first was reused for another value, so there's no resident copy to fold.

This is the third time measurement redirected the work (dead-store #246 no-op → mla regressed #277 → const-CSE fires 0). The consistent lesson: const-CSE is an allocator capability, not a post-pass. It only fires after re-allocation gives each value a stable register.

So

The pass stays flag-gated infra — differential-validated and de-risked for the post-re-allocation pipeline stage — not default-on (0 delta on greedy code, which the audit rule forbids shipping). The genuine prerequisite is the re-allocation that restores residency; that's the next and final piece.

3 unit tests (fold a clamp bound; decline on clobbered-resident; decline on maybe-live-out). clippy/fmt clean; 322 lib tests pass.

🤖 Generated with Claude Code

[avrabe]

On-target validation across the acceptance suite with SYNTH_CONST_CSE=1 (#281) — it's sound, a clean win where it fires, and pins the residency gap per-bench:

bench	const-CSE off	on (SYNTH_CONST_CSE=1)	result
control_step	158	157	−1 cyc, seam 2165333 ✅ — 1 opportunity survives
flat_flight	255	255 (byte-identical)	0 fires — residency gap (its ~14 need vreg output)
filter_axis	37	37	0 fires
controller_step	162	162	0 fires

So your "sound but measures the residency gap" framing is exactly right on silicon: const-CSE is correct and strictly non-regressing (the one control_step site that stays resident folds to a real −1 cyc), and the 0-fires on flat_flight is the measured proof that the residency comes from the vreg output, not the CSE pass. The pass is ready; it's starved of resident constants until the selector stops overloading registers.

This is the clean half of the #277 lesson playing out: const-CSE can't regress (removes only), so unlike mla there's no on-target risk to flipping it on once residency exists — the 0 → 14 on flat_flight will land entirely from the vreg output feeding this same pass. I'll re-run the full suite (shadow + silicon) the moment that's in; expect control_step 157 to hold and flat_flight to take its first real drop. Nice — the pass being non-regressing by construction is the right property.

[codecov]

Codecov Report

❌ Patch coverage is 46.85864% with 203 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/synth-synthesis/src/liveness.rs	46.70%	202 Missing ⚠️
crates/synth-backend/src/arm_backend.rs	66.66%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!