[iOS] EXC_BAD_ACCESS in bridgeless mode - runtime captured by reference in async lambda

[matos-claudio]

Environment

• react-native-fast-tflite: 2.0.0
• React Native: 0.84 (bridgeless mode)
• iOS

Problem

TensorflowPlugin.cpp captures runtime by reference ([=, &runtime])
in async lambdas passed to callInvoker->invokeAsync. When the lambda
executes asynchronously, the runtime reference is dangling — causing
EXC_BAD_ACCESS.

Two occurrences:

1. callInvoker->invokeAsync([=, &runtime]() { ← loadTensorflowModel
2. this->_callInvoker->invokeAsync([=, &runtime]() { ← run()

Fix

Capture runtime by pointer instead:

// BEFORE
callInvoker->invokeAsync([=, &runtime]() {
  auto result = jsi::Object::createFromHostObject(runtime, plugin);
  promise->resolve(std::move(result));
});

// AFTER  
auto* runtimePtr = &runtime;
callInvoker->invokeAsync([=]() {
  auto result = jsi::Object::createFromHostObject(*runtimePtr, plugin);
  promise->resolve(std::move(result));
});

Same fix for the second occurrence in run().

Workaround

Use react-native-nitro-tflite (unofficial drop-in replacement).

[maintenance-hans]

Guten Tag, Hans here.

Note

New features, bugfixes, updates and other improvements are all handled mostly by @mrousavy in his free time.
To support @mrousavy, please consider 💖 sponsoring him on GitHub 💖.
Sponsored issues will be prioritized.

[mrousavy]

Capture runtime by pointer instead:

This is not a fix for the issue and has exactly the same semantics. Reference and pointer are both not strong or shared ref-counted references.

[jslok]

react-native-nitro-tflite works great for me.