Stl: unable to override functions in recent exception / xmemory changes

[angelfor3v3r]

previously, this function: https://github.com/microsoft/STL/blob/379e61781a38eafcff2bd2bc8a099ef5591594ba/stl/inc/xmemory#L28.. returned (size_t)( -1 ).

within the last few months, the function now calls into _Throw_bad_array_new_length which uses the _THROW macro and will result in code generation importing functions from MSVC libraries.

previously, a user was able to suppress generation of these exceptions by disabling them with compiler flags and specifying #define _HAS_EXCEPTIONS 0.

since these "new" throw functions already have a body defined, a user cannot override them with their own implementation. the other exception functions have a definition here: https://github.com/microsoft/STL/blob/53cdb9f8a85c2500077ccf0676619110340c03bf/stl/inc/xutility#L5092... which allow a user to provide their own implementation (and, resulting in strings and exception imports to be omitted).

are there any changes planned for this? a new way to get add your own implementation for exception handling? or have exception / xmemory actually listen to the _HAS_EXCEPTIONS preprocessor definition?

[sylveon]

FYI, defining/overriding anything but a limited set of things (like std::hash specializations) in namespace std is UB.

In this particular situation, it seems that the solution is to write your own allocator.

[angelfor3v3r]

FYI, defining/overriding anything but a limited set of things (like std::hash specializations) in namespace std is UB.

In this particular situation, it seems that the solution is to write your own allocator.

if a user wants exceptions disabled, i feel as if they should be _disabled_ and not _somewhat disabled_.

is there even a way to fully overwrite the default allocator that standard containers use (like std::string, std::vector, etc) or must a user define these types manually with the new allocator?

[StephanTLavavej]

Except as mandated by the Standard (for operator new etc.), we don't support replacing separately compiled functions in the STL.

At this time, _HAS_EXCEPTIONS=0 is mostly unsupported. We try to avoid breaking it gratuitously, but it's undocumented, not officially supported, and not really tested. We might overhaul it in the future to be a supported mode, but we have no concrete plans yet. Ultimately I think we want _THROW to become fastfail in _HAS_EXCEPTIONS=0 mode instead of whatever it does today.

In this case, the STL contains a number of header-only mentions of _THROW, in addition to our separately compiled helpers. (The separately compiled helpers were originally added to avoid circular header dependencies, not for user replacements.) Therefore, I don't believe that this change is a gratuitous break. The bad_array_new_length was done as part of implementing LWG-3038, although it mandated a length_error. As I explained to the Library Evolution Working Group:

An implementer veto is when the Working Paper tells us to do a thing, and then we don't do that thing, but instead do some other thing. We don't do this lightly - we reserve it only for the most dire of circumstances - e.g. when the WP's text is self-contradictory, unusably broken, or impossible to implement given our compatibility requirements.

In this case, we view throwing a non-bad_alloc-derived exception from allocate() as unusably broken, especially given the existence of a bad_alloc-derived exception that was more or less meant for this scenario. Our pre-existing behavior was that we clamped integer overflow to static_cast<size_t>(-1) and let that trigger plain bad_alloc; changing that to length_error would have been a runtime regression in functionality. Upgrading bad_alloc to bad_array_new_length is a compatible change.

This is now tracked by LWG-3237.

As your request is non-Standard, and I don't see anything else we can do here aside from overhauling _HAS_EXCEPTIONS=0 (which we may not be able to do in the VS 2019 release series due to binary compatibility), I believe this issue should be resolved as wontfix.

(Of course, because we're open-source now, you're free to take our codebase and modify it to suit your purposes.)