fix(opt): defensive panic on unmapped vreg instead of silent R0 fallback#101
Merged
Conversation
avrabe
force-pushed
the
fix/defensive-vreg-fallback
branch
from
3530719 to
37b9787
Compare
6 tasks
avrabe
force-pushed
the
fix/defensive-vreg-fallback
branch
from
37b9787 to
592fd27
Compare
This was referenced May 11, 2026
avrabe
force-pushed
the
fix/defensive-vreg-fallback
branch
from
592fd27 to
d752201
Compare
6 tasks
avrabe
added a commit
that referenced
this pull request
May 13, 2026
#101 The systematic AAPCS-clobber audit (47 sites + 27 tests) is shippable on its own. The defensive panic in get_arm_reg fires on at least one remaining wasm_to_ir gap (vreg v13 during simple_add.wat fib compile, seen on Bazel CI for genrules test_add_elf / test_add_m7_elf / test_add_riscv_elf). That panic belongs in #101 once the remaining gap is closed. Cargo workspace tests + clippy + fmt clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
avrabe
added a commit
that referenced
this pull request
May 13, 2026
… usages (#108) Sweep of 47 hardcoded R0..R3 sites across the optimizer + selector. Closes the recurring class that PR #86 (i64 const), #97 (i64 extend/wrap), #106 (24 i64 ops), and #107 (CSE MemLoad/Store) each addressed one slice of. Categories audited: * wasm_to_ir slot accounting for i64 arith/extend ops (15 + 3 ops) * wasm_to_ir op-handler gaps (10 wasm ops → 6 new Opcode variants) * CSE pass: missing src-vreg resolution for all i64 opcodes + alias invalidation on re-def * ir_to_arm hardcoded `Reg::R0` / `Reg::R3` destinations (24 handlers fixed via `alloc_i32_scratch`) * select_with_stack i64 sub-word loads hardcoding `R0:R1` dest pair 27 regression tests added across 4 new test files. Defensive panic stays in PR #101 (one remaining v13 vreg case during simple_add.wat fib compilation is being tracked down — silent R0 fallback preserved in this PR).
avrabe
force-pushed
the
fix/defensive-vreg-fallback
branch
from
d752201 to
5a2c798
Compare
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
avrabe
added a commit
that referenced
this pull request
May 14, 2026
Closes the last latent unmapped-vreg gap surfaced by the systematic AAPCS audit (PR #108). `WasmOp::Call` had no wasm_to_ir handler, fell through to `Opcode::Nop`. The call's result vreg was never mapped, and any downstream consumer (`i32.add(call_result, ...)`) resolved its operand via `get_arm_reg`'s silent R0 fallback — miscompiling on every fib-style recursive WAT. Fix: new `Opcode::Call { dest, func_idx }` in synth-opt, mapped in wasm_to_ir, lowered to `BL func_<idx>` with `dest → R0` in ir_to_arm. Narrow scope (does not yet model R0..R3 invalidation across BL — broader call-boundary regalloc rework deferred). +2 regression tests. Unblocks #101 (defensive panic ships next).
Follow-up to issue #93 (silicon-blocking memset bug). The root cause was optimizer_bridge::wasm_to_ir silently dropping three wasm ops (I64ExtendI32U/S, I32WrapI64) — they produced vregs that never got mapped to ARM registers. The downstream lookup in get_arm_reg returned Reg::R0 as a silent fallback, so subsequent i64 shifts read R0 as rm_lo/rm_hi, destroying memset's destination pointer on real silicon. Replace the silent fallback with a panic that names the vreg. Future "wasm op silently dropped" bugs of this class will surface at the boundary instead of producing miscompiled firmware. A compiler crash is strictly better than a hung MCU. Verified: * cargo test --workspace — all pass (no path legitimately hits the fallback — the post-#97 codebase produces mappings for every wasm op it emits IR for) * cargo clippy --workspace --all-targets -- -D warnings — clean * cargo fmt --check — clean Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
avrabe
force-pushed
the
fix/defensive-vreg-fallback
branch
from
5a2c798 to
534735a
Compare
avrabe
added a commit
that referenced
this pull request
May 14, 2026
Two harnesses (i64_lowering_doesnt_clobber_params, encoder_no_panic) keep finding new AAPCS-class bugs at a rate faster than we close them in a single release cycle. Mark them as non-gating (continue-on-error) so the harness infrastructure can ship as part of #100 while the bug-hunting continues in v0.3.x patches. Promotion criteria for moving a harness from exploration → gating: 60s smoke runs report zero findings on main for 2 consecutive weeks. Open bugs tracked: #103 (fixed via #106), #104 (fixed via #107), the class-wide sweep #108, the WasmOp::Call gap #109/#101, the I32WrapI64 preassign #111, and the new #112 (i64-extend chain + Movw R0).
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-up to PR #97 (which fixed the silicon-blocking #93 memset bug).
The root cause of #93 was that
optimizer_bridge::wasm_to_irsilently dropped three wasm ops (I64ExtendI32U/S, I32WrapI64). The downstream lookup helperget_arm_regreturnedReg::R0as a silent fallback when a vreg had no assignment — so subsequent i64 shifts read R0 as theirrm_lo/rm_hi, destroying memset's destination pointer.This PR replaces the silent fallback with a diagnostic panic naming the offending vreg. Any future "wasm op silently dropped" bugs of this class will surface as a compiler crash with a useful error message, instead of producing miscompiled firmware that boots-and-loops on real silicon.
A compiler crash is strictly better than a hung MCU.
Test plan
cargo test --workspace— all pass (the post-fix(opt): i64 lowering miscompiles memset-style loop counter on Cortex-M (#93) #97 codebase produces mappings for every wasm op it emits IR for, so no test legitimately hits the fallback path)cargo clippy --workspace --all-targets -- -D warnings— cleancargo fmt --check— clean🤖 Generated with Claude Code