在 2020/5/6 18:48, Martin Storsjö 写道:
> 
> This is generally the risk of this kind of commit - regardless of how
> right/wrong the status quo is, there's _a lot_ of code that relies on it
> behaving in a specific way, and changing that will certainly run into
> minor issues in a lot of places.
> 

As it was I that committed the patch - I thought it should be safe,
because 0) G++ has been hardcoding `__USE_MINGW_ANSI_STDIO` for years
(see 'bits/os_defines.h'), and 1) MSYS2 has it too (see
'makepkg_mingw64.conf' which affects all packages in C or C++).

It was however an oversight that during bootstrapping GCC doesn't use
the CPPFLAGS from the build environment, which I think is unusual, as
all MSYS2 packages have `__USE_MINGW_ANSI_STDIO`. The hard-coded (?)
`printf` attribute expects the MS behavior, which shouldn't be the case,
if we were building something that calls functions from libgomp.h
directly, despite the fact that it isn't a public header.


>> 2. Make GCC treat `format(printf)` as `format(gnu_printf)` if C11 or
>>   C++11 is selected, or
> 
> I'm not very keen on that. The compiler should ideally not assume to
> much about what the platform headers do in detail, because it limits
> what we can change. (If we'd make that change in the compiler, and then
> for another reason end up reverting the commit, we'd have the same issue
> in reverse.)
> 

Yes I agree. It's pretty bad that the C (but not C++) standard allows
users to declare standard functions without including appropriate
headers. If someone declare `printf` themselves then they will end up in
calling it from MSVCRT, even in C11 mode.


>> 3. Replace `format(printf)` with `format(gnu_printf)` in libgomp source.
> 
> This also kind of hardcodes too much; libgomp in general can't and
> shouldn't assume too much about which format kind it uses, unless
> libgomp itself hardcodes __USE_MINGW_ANSI_STDIO. Also, the whole
> gnu_printf format is something that only GCC supports, not Clang, but I
> guess libgomp is rather GCC specific anyway.
> 

As long as 'libgomp.h' is a private header, it might be fine to change
or remove this attribute.


> However we do have a define that should expand to the right thing - just
> like inttypes.h PRIu64 follows __USE_MINGW_ANSI_STDIO - we have
> __MINGW_PRINTF_FORMAT.
> 
> So something like this should work:
> 
> #ifdef __MINGW32__
> #define PRINTF_FORMAT __MINGW_PRINTF_FORMAT
> #else
> #define PRINTF_FORMAT printf
> #endif
> 
> __attribute__((format(PRINTF_FORMAT)))
> 
> Not very pretty, but should work without hardcoding any assumptions
> about which format actually is used.
> 

So we end up in patching libgomp source anyway. I will try this tomorrow.



-- 
Best regards,
LH_Mouse