tooling(sourcetrail): code-graph setup + asyncify allocator-suspend plan

sourcetrail/: Sourcetrail 2021.4.19 indexing pipeline for the wasm port —
compile-db transform (rsp expansion, PCH strip, libc++-11 pinning, path
rewrites), removelist-candidate analysis over the indexed call graph, README
with regen steps and the hard-won tricks (relative --project-file hang,
SDK-header poisoning). Heavy artifacts (.srctrldb, compile dbs, libcxx
headers) stay untracked via the folder's .gitignore.

docs/asyncify-allocator-suspend/plan.md: verified plan for the nanosleep-shim
zero-duration guard (mimalloc mi_atomic_yield=sleep(0) can suspend malloc on
the main thread), the red/green contention test, and the adversarially
verified removelist additions with measured payoff (80.9k -> 60.0k
instrumented functions). Planned, not yet executed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P2NHhbzuEHqP2JmcrSD96D
This commit is contained in:
Viktor Vaczi 2026-08-11 13:53:01 +02:00
commit 33e23e0a65
7 changed files with 571 additions and 0 deletions

View file

@ -0,0 +1,131 @@
<!-- STATUS: PLANNED, NOT EXECUTED (saved 2026-08-10). Verification findings herein are real
(measured against the Aug 6 kicad_editor build, emsdk 4.0.2); the shim guard, the test,
and the removelist additions have NOT been applied yet. -->
# Nanosleep shim zero-duration guard + mimalloc contention red/green test
## Context
The asyncify investigation established that **memory allocation can suspend on the main thread**: emscripten's vendored mimalloc uses `sleep(0)` as its spin-wait yield (`mi_atomic_yield` fallback — wasm32-emscripten matches no arch case in `include/mimalloc/atomic.h`), and `wasm/shims/nanosleep_yield.c` unconditionally converts main-thread `nanosleep` into an Asyncify event-loop yield. Under cross-thread delayed-free contention, malloc's slow path can unwind mid-allocation — a latent reentrancy hazard that also blocks the planned allocator removelist entries (~14K-function instrumentation win).
Adversarial verification (real kicad_editor binary + emsdk 4.0.2 sources) proved: the `sleep(0)` chain is the allocator's **only** suspend path; all four mimalloc spin sites sleep constant 0; mimalloc is the module's **only zero-duration sleeper** (all other nanosleep callers are ≥1 ms constants); `std::this_thread::yield`/`sched_yield` don't route through nanosleep. Hence an `ms == 0` early return in the shim severs the path completely and affects nothing else.
**This task:** (1) the zero-duration guard in the shim; (2) a standalone red/green C++ wasm test that exercises the mimalloc contention path and detects main-thread event-loop turns — RED before the fix, GREEN after; (3) add the verified-safe memory-touching entry families to `scripts/common/asyncify-removelist.txt` (user-proposed list, adversarially verified against the real module first).
Out of scope (explicit follow-ups): full kicad_editor rebuild + 3D e2e suite run with the new removelist (the new entries only take effect at the next app build's post-process anyway).
## Part 1 — shim fix
**File:** `pcbjam/wasm/shims/nanosleep_yield.c` (52 lines).
In `nanosleep()`: after computing `ms`, wrap the yield/sleep branch in `if (ms > 0.0)` — zero-duration requests return 0 immediately on both main thread and workers (a 0 ms blocking sleep is a no-op anyway; `sleep(0)` arrives as exactly `{0,0}` per musl, so the guard fires deterministically). Extend the header comment: mimalloc's `mi_atomic_yield` is `sleep(0)` (spin-politeness hint, must NOT become an event-loop yield mid-malloc); a zero-duration sleep never promised an event-loop turn; verified 2026-08-10 that mimalloc is the only zero-duration caller in the module.
Rebuild consequences: only binaries that link the shim — future KiCad app builds, `pthread-ondemand`, and the new test. Already-built `output/*.wasm` unchanged. (CI: touching `wasm/**` busts the testapps cache — expected.)
## Part 2 — the test
### App: `tests/apps/standalone/mimalloc-storm/mimalloc_storm_test.cpp`
Plain `int main()` app (non-wx — template: `standalone/coroutine-pthread/main_repro.cpp`), pthreads + asyncify + **`-sMALLOC=mimalloc`** (used nowhere in the test tree today) + the nanosleep shim.
**Storm** (turns the nanosecond `MI_DELAYED_FREEING` windows into a hit-rate game we control):
- Main thread allocates batches of small same-size blocks (64 B → pages fill fast → full pages enter delayed-free mode; ~200k blocks ≈ ~200 full pages/round).
- W worker `std::thread`s free the previous round's blocks in tight loops (every cross-thread free of a full page runs the two-CAS `DELAYED_FREEING` bracket).
- Main concurrently churns alloc/free in the same size class and calls `mi_collect(true)` each round (`extern "C" void mi_collect(bool)` — public API, headers not on include path; reaches `_mi_heap_delayed_free_all`, the spin site).
- Bounded by ITERATION counts (no wall-clock). Early-exit once ≥50 turns observed (keeps pre-fix RED runs fast); parameters tuned empirically at the red-validation step.
**Detector** (did main return to the event loop mid-storm?):
- `EM_ASM` arms a self-re-arming `setTimeout(0)` that increments `globalThis.__stormTurns`; a synchronous C++ storm can only let it run if an Asyncify unwind happened inside. Main checks the counter via `EM_ASM_INT` every k iterations.
- Phase 2 assertion (shim's load-bearing behavior unchanged): re-arm marker, `nanosleep(5 ms)` on main → the marker MUST have run (nonzero sleeps still yield).
- Console contract (repo idiom — `EM_ASM` console.log markers, self-terminating, documented in the file header):
- `[MIMALLOC_STORM] START threads=W blocks=B rounds=R`
- `[MIMALLOC_STORM] SUMMARY stormTurns=N sleepTurned=0|1 completed=1`
### Build rules: `tests/apps/Makefile.wasm`
- `LDFLAGS_MIMALLOC_STORM` modeled on `LDFLAGS_COROUTINE_PTHREAD_NOWX` (line ~875: EH_FLAGS, ALLOW_MEMORY_GROWTH, ASYNCIFY=1 + stack size, DYNCALLS, -pthread, pool=`navigator.hardwareConcurrency`, STRICT=0, no wx) **plus `-sMALLOC=mimalloc`**.
- `.o` rule for the app; `nanosleep_yield.o` reuse per the `pthread-ondemand` pattern (`Makefile.wasm:681`, compiled `-c -pthread`); `.html` rule; `.PHONY: mimalloc-storm`; `all:` accumulation line.
- No `mallinfo_stub.c` (no OCC). No new suspending import (`env.__asyncjs__*` already in `scripts/common/asyncify-imports.txt`). Post-link asyncify (`apply-asyncify.sh --no-removelist`) + dyncall injection happen automatically via `build-wasm-test.sh`'s `find -newer` fan-out — no driver-script changes.
### Spec: `tests/e2e/mimalloc-storm.spec.ts`
- Placement in `tests/e2e/` → runs under the existing `wx-chromium` project; **zero playwright-config/CI edits** and passes `lint:ci-coverage`.
- Pattern copied from `tests/e2e/coroutine-pthread.spec.ts` (the non-wx precedent): `testLogger` fixture, best-effort `tryLoadApp(...).catch(() => {})` with the documented marker comment, then `expect.poll` for the `SUMMARY` line; parse it; assert `stormTurns === 0`, `sleepTurned === 1`, `completed === 1`, and no page errors (favicon-filtered).
- Determinism compliance: no `waitForTimeout`, no inline retries, bounded C++ iterations; the app always emits a terminal SUMMARY (self-capping), so the poll is bounded.
## Part 3 — asyncify-removelist additions (VERIFIED)
Adversarial verification against the real kicad_editor module is complete: **all proposed families SAFE**, conditional on the Part 1 shim guard shipping first. Add to `scripts/common/asyncify-removelist.txt` (exact patterns — see traps below):
```
# --- Memory-touching families (safe ONLY with the nanosleep zero-duration guard:
# --- mimalloc's mi_atomic_yield is sleep(0); with the guard, no allocation path suspends.
# --- Also conditional: no one registers mi_register_deferred_free/output/error hooks or a
# --- suspending std::new_handler (all unregistered as of 2026-08-10 audit).
std::__2::basic_string<*>*
std::__2::char_traits<*>*
std::__2::vector<*>*
std::__2::__tree*
std::__2::to_string*
std::__2::__itoa*
std::__2::__split_buffer<*>*
std::__2::__shared_ptr_emplace<*>*
std::__2::__shared_ptr_pointer<*>*
std::__2::__shared_count*
std::__2::__shared_weak_count*
std::__2::deque<*>*
std::__2::__hash_table<*>*
std::__2::map<*>*
std::__2::__list_imp<*>*
std::__2::allocator*
boost::uuids::*
operator new*
operator delete*
aligned_alloc
mi_*
_mi_*
sbrk
# Deliberately NOT std::__2::__function* — type-erased std::function invocation (operator())
# is how tool/dialog callbacks run; 111 __func::operator() bodies reach startModal/wxMilliSleep
# directly (e.g. ShowPreferences, file dialogs), so __function frames must stay instrumented.
# Container entries above are still safe: containers only touch callables via their
# clone/destroy lifecycle ops, none of which reach a suspend (verified 2026-08-10).
```
**Authoring traps (verified the hard way):**
- **Bracket balance:** Binaryen's list parser tracks `<>()[]{}` nesting — an unbalanced pattern like `std::__2::basic_string<*` **aborts the whole asyncify pass** (`Fatal: failed to parse lists`). Template patterns must close the bracket: `basic_string<*>*`.
- **Breadth anchoring:** bare `basic_string*` sweeps in `basic_string_view`/`basic_stringbuf`/`basic_stringstream` (31 extra names); the `<`-anchored form excludes them. `__tree*` deliberately includes `__tree_node_base`/iterators (verified). Bare `vector*`/`char_traits*` (no namespace) are wrong forms.
**Key evidence:** zero direct suspend paths post-guard in every family; `__tree` comparators are monomorphized (no `call_indirect` in e.g. `__find_equal` with `std::less`); container-of-`std::function` ops (e.g. TOOL_MANAGER's `deque<function<void()>>::push_back`) only touch the functor's clone/move/destroy lifecycle slot — of 12,880 `__function` lifecycle ops in the module, zero reach a suspend; all 111 suspend-reaching `__function` members are `operator()` invocation bodies (the excluded family). `boost::uuids` (KIID) seeds `mt19937` once via synchronous `getentropy` (wasi import, not asyncify-relevant).
**Measured payoff** (single-threaded verbose runs, real hoisted module):
| Run | Removelist | Instrumented | Module size |
|---|---|---|---|
| A | current file | 80,871 (74%) | 253 MB |
| B | + allocator family | 66,499 (60%) | 218 MB |
| **C** | **+ all families above** | **60,047 (55%)** | **204 MB** |
Run C is a strict subset of B (sanity: zero additions); only non-matching warnings are the 5 pre-existing OCC entries (expected — OCC absent from this app). Audit validity: this binary / emsdk 4.0.2; re-run on emsdk bump or libc++ ABI-namespace change (artifacts in `$W`: `analysis.pkl`, `logC.txt`, `runC.wasm`).
## Part 4 — execution sequence
1. **Build the test against the UNPATCHED shim**: `./scripts/build-wasm-test.sh mimalloc-storm` (targeted — the default `all` target is currently broken by pre-existing working-tree deletions, see caveat). Run the spec: `cd tests && npx playwright test --project=wx-chromium e2e/mimalloc-storm.spec.ts`. Expect the spec to FAIL with `stormTurns > 0` — that failure IS the red validation (proves the storm reaches the mimalloc yield). Record the observed turn count; tune B/R/W if turns are marginal (<~10).
2. **Apply the shim fix** (Part 1), rebuild the same target (make tracks the shim as an explicit prerequisite), re-run the spec → GREEN: `stormTurns === 0`, storm completes (no-deadlock proof), `sleepTurned === 1`.
3. `cd tests && npm run lint:determinism` (spec must pass the linter).
4. **Removelist update:** add the verified entries to `scripts/common/asyncify-removelist.txt` with a documented block: the guard dependency (entries safe ONLY with the zero-duration nanosleep guard), the dormant-slot condition (deferred-free/new_handler/output hooks unregistered), and the `std::__2::__function*` exclusion rationale. Sanity-check the new list against the scratch module: one `wasm-opt --asyncify` run on `$W/hoisted.wasm` with the updated file confirms the expected instrumented-function count and that no entry matches zero functions unexpectedly.
5. Stop before committing — present the diff; commit on main via the user's `/git-feature-commit` flow on request (trunk-based repo).
## Caveats found during exploration
- **Pre-existing working-tree deletions** (not ours; do not touch): `tests/apps/standalone/asyncify-races/races_test.cpp`, `tests/asyncify/asyncify-races.spec.ts`, `tests/web/eeschema-fp-selector.spec.ts` are tracked but deleted, which breaks `Makefile.wasm`'s `all` target and full `npm run test:e2e`. We build only our target and run only our spec; flag the state to the user at the end.
- `build-wasm-test.sh` requires `build-wasm/wxwidgets/wx-config` (wx libs built) and stubs the emsdk in-link asyncify — the emsdk lives at `tools/emsdk/` (installed via `scripts/setup-emsdk.sh` if absent).
- Reference precedents: `races_test.cpp` (via `git show HEAD:...`) for Asyncify-state probing from C++ if the setTimeout detector needs corroboration; `pthread_ondemand_test.cpp` for the watchdog/marker idiom.
## Verification summary
- RED observed pre-fix (spec fails on `stormTurns > 0`), GREEN post-fix on identical parameters — the core deliverable.
- GREEN also proves: spin completes without the yield (no deadlock), 5 ms sleep still turns the event loop (worker-boot behavior preserved).
- `lint:determinism` clean.
- Removelist sanity run on `$W/hoisted.wasm` reproduces run C's numbers (60,047 instrumented / 204 MB pre-O1) with no unexpected non-matching-pattern warnings.
- Follow-ups NOT here: full editor rebuild + 3D e2e suite with the new removelist (entries take effect at the next app build's host post-process).

9
sourcetrail/.gitignore vendored Normal file
View file

@ -0,0 +1,9 @@
# Generated/machine-local artifacts — only the README, scripts, and project
# files are tracked. Everything below is rebuilt by the steps in README.md.
*.srctrldb
*.srctrldb_tmp
*.srctrlbm
compile_commands.json
sample_compile_commands.json
libcxx-11/
libcxx-11.src.tar.xz

113
sourcetrail/README.md Normal file
View file

@ -0,0 +1,113 @@
# Sourcetrail code graph for the KiCad WASM port
Interactive symbol-level code graph (classes, calls, includes, inheritance) of the
merged `kicad_editor` build — all 2,176 TUs: pcbnew, eeschema, common, 3D viewer,
and the `wasm/` port layer — indexed by [Sourcetrail](https://github.com/CoatiSoftware/Sourcetrail).
Lives in `pcbjam/sourcetrail/`: README, scripts and project files are tracked; the heavy
artifacts (index db, compile db, libc++ headers) are ignored via the local `.gitignore`
and rebuilt with the steps below.
## What's in this folder
| File | Purpose |
|---|---|
| `kicad-wasm.srctrlprj` / `.srctrldb` | Sourcetrail project + indexed database (~440 MB) |
| `compile_commands.json` | Transformed compile db the indexer consumes |
| `transform_compile_db.py` | Turns the container's emscripten compile db into the above |
| `libcxx-11/` | Pinned libc++ 11.1.0 headers (see Tricks) |
| `asyncify_candidates.py` | Ranks subsystems safe for the asyncify removelist |
| `sample.srctrlprj` + `sample_compile_commands.json` | 6-file smoke test for the pipeline |
## Install
- **Sourcetrail 2021.4.19** — the last free release (the maintained fork paywalls binaries):
<https://github.com/CoatiSoftware/Sourcetrail/releases/tag/2021.4.19>, macOS dmg → `/Applications`,
then `xattr -dr com.apple.quarantine /Applications/Sourcetrail.app`. It's x86_64 — needs Rosetta 2
on Apple Silicon (works fine).
- Nothing else: python3 stdlib only; header mirrors are exported from the Docker build.
## Run
```bash
# browse (GUI)
open -a Sourcetrail /Users/V/IdeaProjects/pcbjam-private/pcbjam/sourcetrail/kicad-wasm.srctrlprj
# (re)index from the terminal — ABSOLUTE project path, see Tricks
/Applications/Sourcetrail.app/Contents/MacOS/Sourcetrail index \
--project-file /Users/V/IdeaProjects/pcbjam-private/pcbjam/sourcetrail/kicad-wasm.srctrlprj
```
Full index ≈ 12 min all-cores. In the GUI there is no single whole-project graph: search a
symbol (Cmd+F — e.g. `BOARD`, `KIPLATFORM`) and click it; the graph pane centers on it and
expands as you click nodes/edges.
### Refreshing after a KiCad rebuild
The compile db comes from the `main`-branch Docker build cache (KiCad's CMake exports it
by default — no reconfigure needed). From `pcbjam/`:
```bash
source scripts/common/versions.sh
COMPOSE_PROJECT_NAME=kicad-wasm-main docker compose -f docker/docker-compose.yml up -d --build
# compile db + generated sources/headers + .rsp files + deps/wx includes
docker compose -p kicad-wasm-main -f docker/docker-compose.yml exec -T kicad-wasm-builder bash -c \
"cd /workspace && { find build-wasm/kicad-kicad_editor \( -name '*.h' -o -name '*.hpp' -o -name '*.hxx' \
-o -name '*.hh' -o -name '*.inc' -o -name '*.rsp' -o -name '*.cc' -o -name '*.cpp' -o -name '*.cxx' \
-o -name 'compile_commands.json' \) -type f; echo build-wasm/sysroot/include; \
echo build-wasm/wxwidgets/lib/wx/include; } | tar -cf - -T -" | tar -xf - -C .
# emscripten sysroot headers -> tools/emsdk mirror (only after an emsdk bump)
docker compose -p kicad-wasm-main -f docker/docker-compose.yml exec -T kicad-wasm-builder \
tar -cf - -C / emsdk/upstream/emscripten/cache/sysroot/include | tar -xf - -C tools/
cd sourcetrail
python3 transform_compile_db.py ../kicad-kicad_editor/compile_commands.json compile_commands.json
# then the `Sourcetrail index` command above
```
## Tricks (why this isn't just "point Sourcetrail at the cdb")
- **The bundled clang is ~LLVM 11.** Modern libc++ (emsdk 4.x's or the macOS SDK's) does not
parse under it. The transform pins `libcxx-11/` via `-nostdinc++` and takes C headers from the
emsdk musl sysroot mirror (`tools/emsdk/`). Expected residue: ~14 errors, all in the
libc++11/musl locale seam (`_CTYPE_*`, `strtoull_l`, one fatal `xlocale.h`) — harmless to the graph.
- **Transform surgery:** `@CMakeFiles/*.rsp` response files are expanded inline (old clang's cdb
loader can't); PCH is stripped (`-Xclang -include-pch` of clang-20 `.pch` binaries) and replaced
with `-include cmake_pch.hxx`; emscripten-only flags (`-sFOO`, `-fwasm-exceptions`) dropped,
`--target=wasm32-unknown-emscripten` + `-fexceptions` added; paths rewritten
`/workspace``pcbjam/`, `/emsdk``pcbjam/tools/emsdk/`.
- **CLI hangs on relative `--project-file` paths.** Silently — idle event loop, log stops after
"Maven executable path detection". Always pass absolute paths.
- **Global header paths are deliberately empty** in
`~/Library/Application Support/Sourcetrail/ApplicationSettings.xml`. First launch auto-filled
macOS-26-SDK paths, which poison every parse (see clang-11 point). Don't re-run header path
detection from Preferences; `has_prefilled_header_search_paths=1` keeps it from coming back.
- **"N files (126 complete)" undersells the index.** A file counts as complete only if *every* TU
touching it had zero errors; the 14 std-header errors are included nearly everywhere, so the
flag cascades. The symbols/references themselves are all recorded.
## Asyncify removelist candidates
```bash
python3 asyncify_candidates.py
```
Computes, over the indexed call graph (187K call edges + override pseudo-edges for virtual
dispatch), which functions can NEVER reach a suspend point, aggregated per module. Seeds =
every `ShowModal`/`Yield`/`Sleep`/progress-dialog/`COROUTINE` function (**strict**), plus
`ProcessEvent`-style synchronous dispatch (**lenient** — a dispatched handler may suspend and
unwind through the dispatcher). Functions clean under *lenient* are candidates for
`scripts/common/asyncify-removelist.txt` (matching rules are documented in that file:
one prefix wildcard per symbol, e.g. `SHAPE_POLY_SET::*`).
Headline results from the 2026-08-07 index: `kiapi` generated protobuf (15.2K funcs),
`libs/kimath` (7.3K), `clipper2` (6.3K), `nlohmann_json`/`fmt`/`pegtl`/`zint` are 100% clean;
`pcbnew/router` is 1,820/1,827 clean (the 7 are the `Wait()` tool-integration layer — the PNS
shove/optimizer core never suspends).
**Caveats:** the C++ graph can't see calls through `std::function`, event tables, or raw function
pointers, and asyncify operates on the post-inlining *wasm* call graph, not C++ symbols. Before
shipping an entry: ground-truth with Binaryen's asyncify verbose/advise output in
`apply-asyncify.sh`, and rely on e2e — a wrong removal traps loudly (`unreachable`) at the first
unwind through it. Tweak seeds / path depths at the top of the script.

View file

@ -0,0 +1,125 @@
#!/usr/bin/env python3
"""Rank subsystems by asyncify-removelist safety using the Sourcetrail db.
Taint = "a suspend point is reachable from this function" (reverse BFS over
call edges from suspend primitives, with base->override pseudo-edges for
virtual dispatch). Two levels:
strict direct suspend primitives only (Yield/Sleep/ShowModal/progress/
coroutine/fiber)
lenient strict + synchronous event dispatch (ProcessEvent & co), since a
dispatched handler may suspend and unwind through the dispatcher
Functions never tainted even in lenient mode are removelist candidates.
Blind spots (validate with ASYNCIFY_ADVISE before shipping): std::function /
event-table indirection, function pointers, wx/libc internals outside the cdb.
"""
import re
import sqlite3
from collections import defaultdict, deque
DB = "kicad-wasm.srctrldb"
PCBJAM = "/Users/V/IdeaProjects/pcbjam-private/pcbjam/"
FUNC_TYPES = (4096, 8192)
STRICT_PATTERNS = [
r"\tnShowModal", # every ShowModal incl. DIALOG_SHIM/KIDIALOG wrappers
r"\tnShowWindowModal",
r"\tnShowQuasiModal",
r"mwxYield", r"mwxSafeYield", r"mwxYieldIfNeeded",
r"\tnYield", r"\tnYieldFor", r"\tnDoYieldFor", r"\tnSafeYield",
r"\tnSleep", r"mwxSleep", r"mwxMilliSleep", r"mwxMicroSleep",
r"sleep_for", r"sleep_until", r"nanosleep", r"emscripten_sleep",
r"emscripten_fiber", r"__asyncjs__",
r"mwxGenericProgressDialog", r"mwxProgressDialog",
r"mwxMessageBox", r"mwxExecute",
r"mCOROUTINE<",
]
LENIENT_EXTRA = [
r"\tnProcessEvent\t", r"\tnSafelyProcessEvent", r"\tnProcessPendingEvents",
r"\tnHandleEvent\t", r"\tnProcessEventLocally",
]
def main():
con = sqlite3.connect(DB)
funcs = {} # id -> serialized_name
for nid, name in con.execute(
f"SELECT id, serialized_name FROM node WHERE type IN {FUNC_TYPES}"):
funcs[nid] = name
# callee -> callers (reverse call graph)
rev = defaultdict(list)
for s, t in con.execute("SELECT source_node_id, target_node_id FROM edge WHERE type=8"):
rev[t].append(s)
# virtual dispatch: caller of Base::f may land in Derived::f, so taint of
# Derived::f must flow to callers of Base::f -> pseudo-edge callee=Derived,
# caller-side=Base is wrong; we need: if Derived tainted then Base tainted
# is NOT true. Correct: call to Base::f can dispatch to Derived::f, so if
# Derived::f suspends, callers of Base::f suspend. Model: rev[Derived] gets
# nothing; instead treat Base::f as a caller of every override Derived::f.
for s, t in con.execute("SELECT source_node_id, target_node_id FROM edge WHERE type=32"):
# override edge: Derived::f (source) -> Base::f (target)
rev[s].append(t) # taint flows Derived -> Base -> Base's callers
def seeds_for(patterns):
pats = [re.compile(p) for p in patterns]
return {nid for nid, name in funcs.items() if any(p.search(name) for p in pats)}
def closure(seed_ids):
seen = set(seed_ids)
q = deque(seed_ids)
while q:
n = q.popleft()
for caller in rev.get(n, ()):
if caller not in seen:
seen.add(caller)
q.append(caller)
return seen
strict_seeds = seeds_for(STRICT_PATTERNS)
lenient_seeds = strict_seeds | seeds_for(STRICT_PATTERNS + LENIENT_EXTRA)
print(f"functions: {len(funcs)} strict seeds: {len(strict_seeds)} "
f"lenient seeds: {len(lenient_seeds)}")
tainted_strict = closure(strict_seeds)
tainted_lenient = closure(lenient_seeds)
# function -> file (prefer definition scope locations, type=1)
loc = {}
for typ in (1, 0):
for eid, path in con.execute(
"SELECT o.element_id, f.path FROM occurrence o "
"JOIN source_location sl ON sl.id=o.source_location_id "
"JOIN file f ON f.id=sl.file_node_id WHERE sl.type=?", (typ,)):
if eid in funcs and eid not in loc:
loc[eid] = path
def module(path):
if not path.startswith(PCBJAM):
return None # std/emsdk/deps headers — not our code
rel = path[len(PCBJAM):]
parts = rel.split("/")
depth = 3 if parts[0] in ("kicad", "build-wasm") else 2
return "/".join(parts[:depth]) if len(parts) > depth else "/".join(parts[:-1])
stats = defaultdict(lambda: [0, 0, 0]) # module -> [total, strict, lenient]
for nid in funcs:
m = module(loc.get(nid, ""))
if m is None:
continue
stats[m][0] += 1
if nid in tainted_strict:
stats[m][1] += 1
if nid in tainted_lenient:
stats[m][2] += 1
rows = [(m, t, s, l, t - l) for m, (t, s, l) in stats.items() if t >= 20]
rows.sort(key=lambda r: -r[4])
print(f"\n{'module':<44}{'funcs':>7}{'strict✗':>9}{'lenient✗':>9}{'clean':>7}{'clean%':>8}")
for m, t, s, l, clean in rows[:45]:
print(f"{m:<44}{t:>7}{s:>9}{l:>9}{clean:>7}{100*clean//t:>7}%")
if __name__ == "__main__":
main()

View file

@ -0,0 +1,21 @@
<?xml version="1.0" encoding="utf-8" ?>
<config>
<source_groups>
<source_group_9a1b2c3d-0000-4000-8000-kicadwasm0001>
<build_file_path>
<compilation_db_path>./compile_commands.json</compilation_db_path>
</build_file_path>
<indexed_header_paths>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/kicad</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/wasm</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/wxwidgets/include</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/build-wasm/kicad-kicad_editor</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/build-wasm/wxwidgets/lib/wx/include</indexed_header_path>
</indexed_header_paths>
<name>KiCad WASM</name>
<status>enabled</status>
<type>C/C++ from Compilation Database</type>
</source_group_9a1b2c3d-0000-4000-8000-kicadwasm0001>
</source_groups>
<version>8</version>
</config>

View file

@ -0,0 +1,21 @@
<?xml version="1.0" encoding="utf-8" ?>
<config>
<source_groups>
<source_group_9a1b2c3d-0000-4000-8000-kicadwasm0001>
<build_file_path>
<compilation_db_path>./sample_compile_commands.json</compilation_db_path>
</build_file_path>
<indexed_header_paths>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/kicad</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/wasm</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/wxwidgets/include</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/build-wasm/kicad-kicad_editor</indexed_header_path>
<indexed_header_path>/Users/V/IdeaProjects/pcbjam-private/pcbjam/build-wasm/wxwidgets/lib/wx/include</indexed_header_path>
</indexed_header_paths>
<name>KiCad WASM sample</name>
<status>enabled</status>
<type>C/C++ from Compilation Database</type>
</source_group_9a1b2c3d-0000-4000-8000-kicadwasm0001>
</source_groups>
<version>8</version>
</config>

View file

@ -0,0 +1,151 @@
#!/usr/bin/env python3
"""Transform the in-container emscripten compile_commands.json into one
Sourcetrail 2021.4.19 (bundled clang ~11) can index on the host.
- expands @CMakeFiles/....rsp response files inline
- strips emscripten-only and PCH flags (old clang can't load clang-20 .pch)
- replaces the PCH with `-include cmake_pch.hxx` so those headers still parse
- injects libc++ 11 headers (-nostdinc++) + emscripten sysroot includes
- rewrites /workspace -> host pcbjam, /emsdk -> host tools/emsdk mirror
Usage: transform_compile_db.py [--no-wasm-target] <in.json> <out.json>
"""
import json
import os
import re
import shlex
import sys
PCBJAM = "/Users/V/IdeaProjects/pcbjam-private/pcbjam"
ST_DIR = f"{PCBJAM}/sourcetrail"
EMSDK_HOST = f"{PCBJAM}/tools/emsdk"
PATH_MAP = [("/workspace/", PCBJAM + "/"), ("/emsdk/", EMSDK_HOST + "/")]
DROP_EXACT = {
"-fwasm-exceptions",
"-Winvalid-pch",
"--emit-symbol-map",
}
DROP_PREFIX_RE = re.compile(r"^-s[A-Z][A-Z_0-9]*(=.*)?$|^-gseparate-dwarf")
def host_path(p: str) -> str:
for src, dst in PATH_MAP:
if p.startswith(src):
return dst + p[len(src):]
return p
def rewrite_tok(tok: str) -> str:
for src, dst in PATH_MAP:
tok = tok.replace(src, dst)
return tok
def expand_rsp(tokens, directory):
out = []
for tok in tokens:
if tok.startswith("@"):
rsp = tok[1:]
if not os.path.isabs(rsp):
rsp = os.path.join(directory, rsp)
rsp_host = host_path(rsp)
with open(rsp_host) as f:
out.extend(shlex.split(f.read()))
else:
out.append(tok)
return out
def transform(tokens, directory, wasm_target=True):
tokens = expand_rsp(tokens, directory)
argv0 = tokens[0]
lang_cxx = argv0.endswith("++")
out = []
i = 1
pch_headers = []
while i < len(tokens):
tok = tokens[i]
if tok == "-Xclang" and i + 1 < len(tokens):
nxt = tokens[i + 1]
if nxt == "-fno-pch-timestamp":
i += 2
continue
if nxt == "-include-pch":
# -Xclang -include-pch -Xclang <path.pch>
if i + 3 < len(tokens) and tokens[i + 2] == "-Xclang":
pch = tokens[i + 3]
hdr = pch[:-4] if pch.endswith(".pch") else pch
if os.path.exists(host_path(hdr)):
pch_headers.append(hdr)
i += 4
continue
i += 2
continue
out.extend([tok, nxt])
i += 2
continue
if tok in DROP_EXACT or DROP_PREFIX_RE.match(tok):
i += 1
continue
if not wasm_target and tok in ("-matomics", "-mbulk-memory", "-pthread"):
i += 1
continue
out.append(tok)
i += 1
for hdr in pch_headers:
out.extend(["-include", hdr])
# Rewrite container paths in the ORIGINAL tokens only, then prepend the
# host-path injections — rewriting after injection would re-fire on the
# /emsdk/ substring inside the host tools/emsdk mirror path.
out = [rewrite_tok(t) for t in out]
inject = ["-isystem", f"{EMSDK_HOST}/upstream/emscripten/cache/sysroot/include",
"-fexceptions"]
if lang_cxx:
inject = ["-nostdinc++", "-isystem", f"{ST_DIR}/libcxx-11/include"] + inject
if wasm_target:
inject = ["--target=wasm32-unknown-emscripten"] + inject
return ["clang++" if lang_cxx else "clang"] + inject + out
def main():
args = sys.argv[1:]
wasm_target = True
if args and args[0] == "--no-wasm-target":
wasm_target = False
args = args[1:]
src, dst = args
with open(src) as f:
db = json.load(f)
out_db = []
missing = 0
for e in db:
tokens = shlex.split(e["command"])
directory = e["directory"]
new_tokens = transform(tokens, directory, wasm_target)
file_host = host_path(e["file"])
if not os.path.exists(file_host):
missing += 1
continue
out_db.append({
"directory": host_path(directory),
"command": " ".join(shlex.quote(t) for t in new_tokens),
"file": file_host,
})
with open(dst, "w") as f:
json.dump(out_db, f, indent=1)
print(f"wrote {len(out_db)} entries to {dst} ({missing} skipped: file missing on host)")
if __name__ == "__main__":
main()