Skip to content

release-watch: synth v0.51.0 on real fused falcon v1.128 — float gate unchanged (aarch64 release) - #170

Merged
avrabe merged 2 commits into
mainfrom
release-watch/synth051-float-gate
Jul 28, 2026
Merged

release-watch: synth v0.51.0 on real fused falcon v1.128 — float gate unchanged (aarch64 release)#170
avrabe merged 2 commits into
mainfrom
release-watch/synth051-float-gate

Conversation

@avrabe

@avrabe avrabe commented Jul 28, 2026

Copy link
Copy Markdown
Contributor

STEP-1 release-watch. synth v0.51.0 shipped (2026-07-23). Tested against the real fused falcon v1.128 (sha 54020c5e) per the AFD-035 "verify against the real build" discipline.

Pipeline (full DD-018 chain): meld 0.39 fuse → loom 1.2 optimize → synth --relocatable -t cortex-m7dp, controlled A/B on the identical fused+loom'd input.

Result: synth 0.50.0 and 0.51.0 emit a byte-identical 9-skip set — same functions, same reasons:

  • func_29/31/42F32ConvertI64S (i64→f32 conversion, arm backend can't lower) ×3
  • func_36/78 — spill-slot pool exhausted ×2
  • func_88 — no free callee-saved register ×1
  • func_69/75/172 — arm backend compilation-failed ×3

Why unchanged is correct: v0.51.0 is an aarch64 backend release (#851 — aarch64 non-param locals + linmem load/store, "runs real WASM modules") + differential-oracle hardening (#850/#853). It does not touch the ARM/cortex-m float lowering path that holds the falcon skips. aarch64 is orthogonal to the RT1176 (M7/M4/M3 — all cortex-m), so no falcon-lane movement was on offer. No regression, no progress, no upstream issue (v0.51 didn't target the gate).

Dominant residual blocker family stays F32ConvertI64S + register-allocation exhaustion. Gate STILL OPEN. AFD-035 progression log updated. (Count note: the 9 here is the loom-optimized full chain; the historical "13" was pre-loom-optimize — loom clears ~4 skippable funcs.)

rivet validate → PASS.

🤖 Generated with Claude Code

avrabe and others added 2 commits July 28, 2026 20:49
…at gate UNCHANGED (aarch64 release, ARM path untouched)

STEP-1 release-watch: synth v0.51.0 shipped. Tested against the REAL fused falcon
v1.128 (sha 54020c5e) on the full DD-018 chain (meld 0.39 -> loom 1.2 -> synth
--relocatable -t cortex-m7dp), controlled 0.50.0-vs-0.51.0 on identical input:
BYTE-IDENTICAL 9-skip set (F32ConvertI64S x3, spill-exhaust x2, no-free-callee,
arm-compile-fail x3). v0.51 is an aarch64 backend release (#851) + oracle hardening
(#850) - it does not touch the ARM/cortex-m float lowering path, so the unchanged
gate is the correct, expected result (no regression, no progress on offer). AFD-035
progression log updated; no upstream issue (v0.51 didn't target the gate).

rivet validate PASS.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ild-matrix rung)

synth v0.51.0 on the same fused+loom falcon v1.128, swept per target ISA:
m7dp=9, m7=9, m4f(estimator)=9, m4=30, m3(F100-class)=30. Three reads:
(1) F32ConvertI64S=3 on every target = target-independent irreducible float-gate core;
(2) M7 + M4-estimator share the same 9-skip VFP floor (no separate estimator pole);
(3) soft-float explodes RA pressure 6->27 -> independently confirms F100 runs the small
gale VER-OS-HM-001 failsafe FSM, NOT a lowered full falcon (validates DD-025 partitioning
+ DD-026 decision 4). rivet validate PASS.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@avrabe
avrabe merged commit b9180f9 into main Jul 28, 2026
5 checks passed
@avrabe
avrabe deleted the release-watch/synth051-float-gate branch July 28, 2026 18:56
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant