Cause a debug break when a panic occurs on Windows.

[retep998]

In order to aid debugging it would be really useful if Rust would trigger a debug break whenever a panic occurs on Windows. Ideally using the __debugbreak intrinsic which is equivalent to int 3 on x86 platforms.

[brson]

Why just on windows?

There used to be an instruction on the panic path that triggered the debugger on Unix at least. I recall it could be annoying.

[Diggsey]

When debugging C++ programs on windows, any uncaught exceptions will trigger the just-in-time debugger. The program will pause before unwinding if the exception would have been uncaught, and you're able to enter a debugging session as though the exception has not yet happened (such that the next instruction would cause the exception).

It would be really nice if the same experience was possible with rust.

[vadimcn]

In order to aid debugging it would be really useful if Rust would trigger a debug break whenever a panic occurs on Windows.

But in that case programs won't be able to unwind from panic without dropping into debugger. When debugger is not attached, int 3crashes the program.
You might be thinking of if (IsDebuggerPresent()) DebugBreak(); sort of thing, but even that is annoying, IMO, because the debugger will always do a breakpoint stop, even if you were debugging something else.
What's wrong with setting a breakpoint on rust_panic?

[Diggsey]

Having looked into it a little, I think the best option would be to remove the top-level exception handler which rustc installs, when compiling for windows.

When a panic is uncaught, it should then automatically trigger the windows JIT debugging feature, or if running in a debugger, it would break into the debugger. Setting a breakpoint on rust_panic only works if you've already got a debugger attached.

[cyplo]

Would you expect different behaviour for programs compiled in release mode in comparison to those compiled in debug mode with regards to the above ?

[dtolnay]

Seems reasonable. I would like to see an implementation of this in a PR, along with a brief writeup of how this improves the debugging experience and how it impacts non-debug use cases, and I think the libs team will be able to arrive at a discussion in the PR.

[ArtemGr]

How about using an environment variable, similar to RUST_BACKTRACE, to make this optional?