0.6.6: lambdas that capture a throwing copy terminate under /Zc:lambda; audit the in-game tree

Open
#247 1 comment 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
46/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
cmake, cpp

Research direction

Start with clean/fsr/src/dllmain.cpp and compare its modelPlaces capture with DeviceRelease in clean/fsr/src/async_create.h; then audit the in-game tree listed in the issue for other throwing lambda captures. Read the contributor notes for the existing code rules and inspect TestDeviceHoldAllocationFailure in clean/fsr/tests/test_frontend.cpp. Done means all identified sites follow the proposed rule, the rule is documented, and the relevant throwing-copy test passes in the CI configuration.

Written by the indexing model from the issue text.

Description

What

With MSVC's /Zc:lambda, a lambda whose capture copy throws terminates the process (std::terminate) instead of unwinding to the caller's handler. /Zc:lambda is on whenever /permissive- is, and the CMake targets of the front end build with /permissive-, so every release build of the in-game code has it on. A capture copy can throw whenever the captured object allocates when it is copied (std::string, std::vector, std::function holding a large callable): the throw is std::bad_alloc, so the case is out of memory, and the process ends in std::terminate instead of the error path.

It was found in the create worker's device hold (#225): [release, device] { release(device); } captured a std::function by copy. A test with a callable whose copy throws aborted with exit code 0xc0000409 in the CI configuration (cmake, Release, /permissive- /W4 /Zi) and not in a hand-made cl build without /permissive-. A stand-alone reproduction narrowed it to this switch: with /Zc:lambda alone the throw terminates, without it the exception reaches the catch; a plain copy of the same std::function (no lambda) is fine in every configuration.

Rule (proposed)

No lambda in code that runs in a game may capture, by copy or by init-capture, an object whose copy can throw. Use one of:

  • a named callable class whose constructor does the copy (that unwinds normally; #225 does this: DeviceRelease in clean/fsr/src/async_create.h), or
  • a capture whose copy cannot throw: a raw pointer, an index, a std::shared_ptr (copy is noexcept), or a std::move into an init-capture of a type with a noexcept move.

Known site

clean/fsr/src/dllmain.cpp: p.modelPlaces = [modelDir = config.modelDir] { ... } init-captures a std::string by copy at start-up. Out of memory at that point ends in std::terminate instead of the normal failure. It is the same class; it is not part of #225.

Asked for (0.6.6)

  1. Audit the whole in-game tree (front end, the D3D12 runtime and its interop, the Ray Reconstruction plugin and runtime, shared code that runs in the game process) for lambdas with a by-copy or init-capture of a type whose copy can throw; fix each as above.
  2. Write the rule in the contributor notes next to the other code rules.
  3. Optional: a small CI check (a grep for init-captures of std::string / std::vector / std::function in the in-game sources) so the rule holds.
  4. Keep a test that drives a throwing copy through any callable the fix touches (see TestDeviceHoldAllocationFailure in clean/fsr/tests/test_frontend.cpp): it must pass in the CI configuration.
Dominant language
C++
Stars
12
Forks
2
Avg merge
3h 10m
Merged PRs (30d)
171

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from phantoslayer21/wronggpu

All issues in phantoslayer21/wronggpu

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.