Windows SDK 22000+ missing __analysis_noreturn SAL

[riverar]

I suspect this is what led to a few APIs slipping under the radar when implementing noreturn support in metadata.

19042 winnt.h

NTSYSAPI
__analysis_noreturn
VOID
NTAPI
RtlRaiseException(
    _In_ PEXCEPTION_RECORD ExceptionRecord
    );

22000.0 and 22000.120 winnt.h

NTSYSAPI
VOID
NTAPI
RtlRaiseException(
    _In_ PEXCEPTION_RECORD ExceptionRecord
    );

@mikebattista Can we get an internal bug tracking this? Feel free to do what you wish with this issue afterwards. Thanks in advance!

[sotteson1]

I looked up the change, and it was changed on purpose. The bug title was:

[EH] RaiseException can return but is decorated with __analysis_noreturn

The commit said:
Yes, it is perfectly legitimate for (Rtl)RaiseException to return.

So I guess it's possible for this function to return, although I don't know the specifics. I wonder if we should remove the attribute from the metadata to be consistent.

[kennykerr]

That's surprising. Adding @lzybkr who was also expecting it not to return. And @JamesMcNellis who knows all things about exceptions.

[KalleOlaviNiemitalo]

I suppose RaiseException will return normally if an exception filter returns EXCEPTION_CONTINUE_EXECUTION and dwExceptionFlags does not include EXCEPTION_NONCONTINUABLE.

[lzybkr]

Under normal conditions, RtlRaiseException finishes by setting the thread (register) context - not by returning. If something goes wrong in setting the context, RtlRaiseException raises a different non-continuable exception, so - in theory it shouldn't reach the return instruction, but maybe it's not impossible?

That said, @KalleOlaviNiemitalo points out a scenario where that may look like an ordinary return. To make this concrete - this following example prints Hello world.

#include <windows.h>
#include <stdio.h>

int main() {
    __try {
        RaiseException(0x12345678, 0, 0, 0);
        printf("Hello ");
    } __except(EXCEPTION_CONTINUE_EXECUTION) {
    }

    printf("world\n");
}

Based on this example, I'd agree that the prototype should not include noreturn, even knowing that RtlRaiseException is not actually returning.