Windows/iOS Crash During Deep JSON Encoding/Decoding

[jevonmao]

A Windows-specific crash was revealed by the test case added in PR #1242, which fixes a recursion depth mismatch between JSONDecoder and JSONEncoder.

On macOS and Linux, all tests including the deeply nested JSON case pass. However, on Windows, the process crashes during JSONEncoderTests.test_JSONPassTests, exiting without any test result line, which suggests a potential stack overflow caused by deep recursion.

Reproduction

1. On Windows, run JSONEncoderTests.test_JSONPassTests from the test suite introduced in Fix #1176 JSON Decoder and Encoder limit disagreement #1242.

Relevant links

• PR Fixing Encoder/Decoder Mismatch: Fix #1176 JSON Decoder and Encoder limit disagreement #1242
• Original bug report: JSONEncoder can't encode what JSONDecoder decoded (they disagree on limits...) #1176

[jmschonfeld]

I added more info to #1306 but I suspect this is because Windows/iOS (and other embedded platforms) have a significantly smaller default stack size (1MB vs 8MB) which leads to hitting the stack guard with fewer recursive frames. I suspect we need a lower threshold for these platforms (512 seems to hit the guard during decoding, 400 seems to hit the guard during encoding but decoding is fine on iOS, so perhaps 300 or 256?). Additionally, the recursive limit on iOS doesn't seem to be respected during encoding (as encoding an object of 400 nested arrays crashed even with a recursive limit of 5) so I suspect there's a few latent bugs here that need to be resolved before adding a test that nears the recursion limit

[YOCKOW]

Additionally, the recursive limit on iOS doesn't seem to be respected during encoding (as encoding an object of 400 nested arrays crashed even with a recursive limit of 5)

Is it perhaps related to the fact that __JSONEncoder has no limits for depth? I'm unsure whether or not wrapGeneric induces recursion on specific platforms though...