fix: LLVM -O2 miscompilation of BEAM unicode/iodata on wasm32#259
Merged
Conversation
- Rewrite ring.erl with correct ring topology (self as counter,
spawned processes as forwarders). Old version had a deadlock bug
where set_next broke the chain.
- Work around BEAM wasm32 ~f format bug: float_to_list/1 uses broken
ryu/snprintf path, use float_to_list/2 with {decimals,N} instead.
- Track thread workers in serve.ts for proper cleanup on exit. BEAM's
halt(0) hangs in pthread_join because threads are blocked on futex
waits. Add halt detection: if stdout goes idle for 2s after output,
force clean exit.
- Remove unused threadChannelOffsets array and keepalive timer.
Ring benchmark results: 1000 processes, 100 rounds = 100K messages
in ~54ms (~1.85M messages/sec).
LLVM's wasm32 backend miscompiles aggregate initialization of structs containing pointers to shadow-stack local arrays at -O2. This causes BEAM's ESTACK macros and iodata traversal in erl_unicode.c to produce garbage, breaking unicode:characters_to_binary/2 and filename conversion. Two fixes: - global.h: explicit field-by-field ESTACK/WSTACK init on wasm32 - Makefile: compile erl_unicode.c at -O1 to avoid the codegen bug Both patches are applied automatically by build-erlang.sh.
brandonpayton
added a commit
that referenced
this pull request
May 29, 2026
## Summary
- Fix LLVM wasm32 backend miscompiling aggregate struct initialization
with shadow-stack pointers at -O2
- Add `global.h` patch: explicit field-by-field ESTACK/WSTACK init under
`#ifdef __wasm32__`
- Add Makefile patch: compile `erl_unicode.c` at `-O1` to avoid iodata
traversal codegen bug
- Both patches applied automatically by `build-erlang.sh`
## Details
BEAM's `unicode:characters_to_binary/2` BIF and
`erts_native_filename_need` traverse iodata using ESTACK macros
(stack-based term traversal). At `-O2`, LLVM's wasm32 backend
miscompiles the aggregate initialization of `ErtsEStack`/`ErtsWStack`
structs that contain pointers to shadow-stack local arrays. This
manifests as:
- `{error, <<>>, {}}` from `unicode:characters_to_binary([[]])`
- EISDIR error opening `start_clean.boot` (filename conversion fails)
The bug is a classic heisenbug — adding `fprintf(stderr, ...)` debug
calls changes code layout enough to prevent miscompilation.
**Root cause**: LLVM optimization passes at `-O2` incorrectly handle
struct fields pointing to `__stack_pointer`-relative local arrays on
wasm32.
**Fix**: Two-pronged approach:
1. `global.h`: Replace aggregate initialization with explicit
field-by-field assignment under `#ifdef __wasm32__` (defensive)
2. `Makefile`: Compile `erl_unicode.c` at `-O1` (fixes the actual
codegen bug)
## Test plan
- [x] BEAM boots and runs: `io:format("Hello from BEAM!~n"), halt().`
- [x] Arithmetic, string, list comprehension, map, binary, atom
operations work
- [x] Cargo tests: 703 passed
- [x] Vitest: 199 passed
- [x] musl libc-test: 0 FAIL (19 XFAIL)
- [x] POSIX test suite: 0 FAIL (1 XFAIL)
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.
Summary
global.hpatch: explicit field-by-field ESTACK/WSTACK init under#ifdef __wasm32__erl_unicode.cat-O1to avoid iodata traversal codegen bugbuild-erlang.shDetails
BEAM's
unicode:characters_to_binary/2BIF anderts_native_filename_needtraverse iodata using ESTACK macros (stack-based term traversal). At-O2, LLVM's wasm32 backend miscompiles the aggregate initialization ofErtsEStack/ErtsWStackstructs that contain pointers to shadow-stack local arrays. This manifests as:{error, <<>>, {}}fromunicode:characters_to_binary([[]])start_clean.boot(filename conversion fails)The bug is a classic heisenbug — adding
fprintf(stderr, ...)debug calls changes code layout enough to prevent miscompilation.Root cause: LLVM optimization passes at
-O2incorrectly handle struct fields pointing to__stack_pointer-relative local arrays on wasm32.Fix: Two-pronged approach:
global.h: Replace aggregate initialization with explicit field-by-field assignment under#ifdef __wasm32__(defensive)Makefile: Compileerl_unicode.cat-O1(fixes the actual codegen bug)Test plan
io:format("Hello from BEAM!~n"), halt().