Re: [boost] [type_traits] Variadic function pointer with __fastcall calling convention

Subject: Re: [boost] [type_traits] Variadic function pointer with __fastcall calling convention
From: Edward Diener (eldiener_at_[hidden])
Date: 2013-10-01 22:48:37

On 10/1/2013 9:40 PM, Stephan T. Lavavej wrote:
> I audited VC's <type_traits> for calling convention madness, and I do not believe that custom calling conventions can be applied to old-style varargs.

Nonetheless VC++ ( 8, 9, 10, and 11 ) does not produce a compiler error 
when a __stdcall or __fastcall calling convention is used in a pointer 
to function taking old-style varags. Do you see this as an ongoing bug 
in VC++ ? If so Boost type_traits should not support it AFAICS.

> It's been a few months since I looked at this, but VC sometimes looks at a calling convention that shouldn't be there (this is especially true for x64/ARM) and simply pretends that it's an acceptable calling convention.

Understood.

>
> To figure out all of the criteria for calling conventions, I wrote a templated struct, then attempted to specialize it for all known calling conventions, then removed the ones that triggered errors (due to immediate rejections) or duplicates (due to silent rewrites).
>
> Here is my Matrix Of Doom, transcribed as (mostly) English instead of code:
>
> __cdecl - Always.
> __clrcall - _M_CEE defined.
> __fastcall - _M_IX86 defined and _M_CEE not defined.
> __stdcall - _M_IX86 defined.
> __thiscall - _M_IX86 defined. Member functions only.
> __vectorcall - VC 2013 or newer. _M_CEE must not be defined. Available for _M_X64. Available for _M_IX86 when _M_IX86_FP is at least 2.

The list is appreciated. But were old-style varargs included in your tests ?

>
> To be resistant to the calling convention switches (/Gd, /Gr, /Gz, /Gv) I always mark specializations for free/member function pointers with a calling convention (__cdecl is always present, at a minimum).
> The only exception is the old-style variadics, which I do not mark with any explicit calling convention.
>
> It is (quite) possible that there is a bug in my machinery, but so far it seems to be well-behaved in VC 2013 (the only deficiency is that I need to apply it throughout the STL).

I have not tried VC in VS2013, but the matter should be cleared up so 
that VC++ at least does not at accept the __fastcall calling convention 
with old-style varags if what you say in the first line of your reply is 
true.

>
> STL
>
> -----Original Message-----
> From: Boost [mailto:boost-bounces_at_[hidden]] On Behalf Of Edward Diener
> Sent: Tuesday, October 01, 2013 6:09 PM
> To: boost_at_[hidden]
> Subject: [boost] [type_traits] Variadic function pointer with __fastcall calling convention
>
> In type_traits/detail/is_function_ptr_tester.hpp there are pointer to
> function signatures being tested which have the __fastcall calling
> convention and variadic functions. While MSVC accepts this code clang in
> Windows, which defines _MSC_VER and _MSC_EXTENSIONS, currently gives an
> error. The justification for giving an error is that __fastcall is a
> callee stack cleanup calling convention, so it cannot be used for
> variadic functions.
>
> There is a bug report regarding this in the Bugzilla for llvm at
> http://llvm.org/bugs/show_bug.cgi?id=12535.
>
> Comments have been made to me in e-mail and the clang developer's
> mailing list that it is a type_traits error to allow this, even for
> MSVC-emulation code. Of course we can temporarily fix this for clang on
> Windows, possibly with a BOOST_WORKAROUND solution for clang and see
> if/how clang eventually addresses the bug report. But I would like some
> sort of resolution so that clang on Windows can compile type_traits in
> the current Boost trunk.
>
>
> _______________________________________________
> Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost
>
> _______________________________________________
> Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost
>