swjs_call_function_no_catch corrupts an Emscripten callee's WASM function table on the second invocation

Führe die angehängte viz-repro.html aus und vergleiche den fehlschlagenden H-Pfad mit den erfolgreichen Fällen J und K. Untersuche dann Package/runtime.js im Umfeld von swjs_call_function_no_catch bei Zeile 573, mit Fokus auf die Verarbeitung von Argumenten und Rückgabewerten. Als erledigt gilt die Identifizierung des Bridge-Zustands oder der Interaktion, die dazu führt, dass der zweite Aufruf einen Trap auslöst, sowie die Dokumentation oder Validierung eines geeigneten Fixes.

Summary

When Swift WASM (via JavaScriptKit) calls a JavaScript function that internally invokes an Emscripten-built WASM module (@viz-js/viz), the second such call on the same Emscripten instance traps inside the Emscripten callee with RuntimeError: table index is out of bounds. The trap fires inside the callee's WASM, but only when the caller is JavaScriptKit; the same JS function called the same number of times from plain JavaScript works perfectly. Yielding to the JS event loop between calls does not reliably help — the corruption is per-Emscripten-instance, not just timing-dependent. Allocating a fresh Emscripten instance for each Swift-originated call avoids the trap entirely.

This is reproducible with an 11-test differential matrix that isolates the trigger to the swjs_call_function_no_catch path specifically.

Environment

• JavaScriptKit: as bundled with Swift 6.3 / BridgeJS plugin (current)
• @viz-js/viz: 3.26.0 (Graphviz 14.1.5, Emscripten-built)
• Browser: Chrome (latest), macOS 15.4
• runtime.js referenced below is BridgeJS-generated Package/runtime.js

Failing stack

RuntimeError: table index is out of bounds
    at wasm://wasm/...:wasm-function[1006]:0x5742f
    at wasm://wasm/...:wasm-function[160]:0x9cad
    at wasm://wasm/...:wasm-function[1097]:0x619e2
    at wasm://wasm/...:wasm-function[2197]:0xc45bc
    at Object.ccall (viz.js:9)
    at renderInput (viz.js:135)
    at Viz.render (viz.js:308)
    at Viz.renderString (viz.js:318)
    at window.graphvizLayoutJSON (vgraph-v1.0.1.js:218)
    at swjs_call_function_no_catch (Package/runtime.js:573)

The trap is inside viz-js's WASM linear memory / function table, but it only fires when the call frame above is swjs_call_function_no_catch. Identical JavaScript code, identical viz-js instance, identical input DOT — fails when called from Swift WASM, passes when called from JS.

Differential test matrix

A single HTML page (viz-repro.html, source linked below) runs eleven variants of "render a small graph" against the same viz-js instance, varying only how the call reaches it.

Test	What it varies	Result
A	Two viz.renderString calls from plain JS, same DOT	PASS
B	Two calls from JS, second DOT has shuffled attribute order	PASS
C	Two calls from JS, DOT contains fontname (no font in WASM)	PASS
D	Two calls from JS, with a parallel instance() allocation in flight	PASS
E	Two calls from JS in one synchronous turn (no microtask between)	PASS
F	Same as E, but routed through wrapper code that dispatches via window.graphvizLayoutJSON	PASS
G	Full Swift WASM bundle initialised (BridgeJS exports loaded), then JS-only viz calls	PASS
H	Two Swift→JS→viz calls on the same instance, same DOT, same turn	FAIL
I	Same as H + await new Promise(r => setTimeout(r, 0)) between calls	UNRELIABLE — depends on whether a background instance() allocation happens to resolve in time
J	One Swift→viz call, then one JS-direct viz call on the same instance	PASS — confirms the viz instance still answers JS callers correctly
K	Two Swift→viz calls with 500 ms yield (ensures a fresh instance is swapped in)	PASS — confirms the bug is per-instance, not per-call-accumulating

The matrix isolates the trigger to the second swjs_call_function_no_catch invocation that ends up calling the same Emscripten instance. JS callers cannot reproduce it under any input or timing variation; Swift callers reproduce it deterministically.

J is particularly informative: after a Swift→viz call, the same viz instance still serves a JS-direct call correctly. So the viz-js heap is not globally corrupted — only the Swift-bridge call path on that instance is broken. K confirms that allocating a fresh viz instance per Swift call avoids the trap entirely.

Hypothesis

I don't have enough JavaScriptKit internals knowledge to call the cause. What we observe is consistent with swjs_call_function_no_catch leaving residue in JS-side state that the JS function's next execution path through the Emscripten callee interacts with — possibly something around how the JSObject reference table is reused, how return values from JS-into-Emscripten are held, or how the underlying WebAssembly.Table imports are shared between modules. The fact that JS-direct calls don't reproduce, while two Swift bridge calls do, suggests it's specific to the bridge mechanics rather than the Emscripten callee's own state machine.

Reproducer

viz-repro.zip

Self-contained ZIP attached: viz-repro.zip (1.9 MB).

To run:

unzip viz-repro.zip
cd viz-repro-bundle
python3 -m http.server 8000

Then open http://localhost:8000/viz-repro.html and click each lettered button. H reliably fails; A–G, J, K reliably pass; I is timing-dependent.

The ZIP includes the pre-built Swift WASM bundle (Package/VGraphWasm.wasm + runtime.js) so no Swift toolchain or build step is needed. @viz-js/viz and @bjorn3/browser_wasi_shim are loaded from jsDelivr.

I have not tested whether other Emscripten-built WASM modules (e.g. @hpcc-js/wasm) exhibit the same behaviour. Happy to test if it would help narrow scope.

Workaround

Maintain a pool of pre-allocated viz-js instances and consume a fresh one per Swift→viz call (refill async). Pool size N tolerates up to N Swift→viz calls per event-loop turn. This is shipping in our codebase now, but it's a heavy hammer — eagerly allocating 5+ Emscripten WASM instances at startup for a workaround that should not be necessary.

Ask

1. Is this a known interaction pattern between swjs_call_function_no_catch and Emscripten-built callees, or a fresh report?
2. If the call mechanism is leaving recoverable state in the Emscripten module, is there a way to flush / reset it from the Swift side without disposing the whole instance?
3. Would you accept a PR adding an "after-call-cleanup hook" in JavaScriptKit's bridge, or is the right answer something Emscripten-side?

Happy to dig further into the JavaScriptKit runtime if you can point me at the most likely files. The smoking gun is somewhere in how swjs_call_function_no_catch (runtime.js:573) marshals arguments and return values when the called JS function calls into another WASM module.