Re: printf warning with uintmax_t

[Rich Felker <dalias@aerifal.cx>]

On Sat, Feb 09, 2013 at 01:03:19PM +0100, Szabolcs Nagy wrote:
> * Jens Gustedt <jens.gustedt@inria.fr> [2013-02-09 09:35:00 +0100]:
> > Am Freitag, den 08.02.2013, 17:14 -0500 schrieb Rich Felker:
> > > It looks like we're not matching the ABI convention gcc expects, where
> > > [u]intmax_t is the lowest-rank type capable of storing the full
> > > integer range (i.e. long on 64-bit systems).
> > 
> > I didn't check how you do this in musl, but for me gcc without any
> > includes has
> > 
> > #define __INTMAX_TYPE__ long int
> > 
> > so I guess that macro is just what should be taken to match its
> > expectations for printf formats. You still could have a fallback for
> > compilers that don't provide this. I checked for clang and it does,
> > BTW.
> 
> tcc, pcc, cparse (the firm frontend) does not define __INTMAX_TYPE__
> and clang blindly follows gcc
> 
> there may be various possible definitions for intmax_t, assuming
> a given abi, and the compiler has no business knowing which one is
> choosen by the libc
> 
> of course printf format checking, c++ and c11 generics with builtin
> intmax_t functions change that

Indeed. The underlying type of all typedefs is part of the ABI if
you're considering C++.

> i think it's better if the intmax_t definition is fixed for a given
> abi (so libc can define it without consulting the compiler), but
> just like with __WCHAR_TYPE__ it may be ifdefed: if __INTMAX_TYPE__

Actually I believe the current treatment of __WCHAR_TYPE__ is wrong,
isn't it? It should not vary between compilers on a given arch/abi
combination. I suspect the code in i386 alltypes.h now is leftover
cruft from when the type was first fixed, when I didn't understand the
ABI requirements well.

> is defined then use that otherwise a default (however that adds an
> extra branch in the preprocessor for no good reason: it may change
> from compiler to compiler but it should not, the compilers should
> agree on something here)
> 
> once this is sorted out inttypes.h should be fixed as well

I think [u]intmax_t should just be moved back to alltypes.h with the
correct per-arch definitions.

A related issue is the limit values. Right now, UINTMAX_MAX is just
defined as UINT64_MAX, but the type of the latter is wrong (always
ULL). We could either move them to bits, use some #ifdef magic to get
them right based on the assumption that all archs use the lowest-rank
type of the needed size (is this correct?), or if that assumption is
incorrect we could just omit the suffixes like ULL and let the
compiler choose the appropriate type for the hex constants. This may
generate warnings with some bogus gcc versions or modes, though...

Rich