gcc 4.9 build warnings (was: Re: arm-soc build: 2917 warnings 0 failures (arm-soc/v3.18-rc1-20-g06c0773))

[Thierry Reding]

On Thu, Nov 06, 2014 at 02:08:37PM +0100, Ard Biesheuvel wrote:
> On 6 November 2014 13:59, Thierry Reding <thierry.reding at gmail.com> wrote:
> > On Thu, Nov 06, 2014 at 12:56:27PM +0100, Arnd Bergmann wrote:
> >> On Thursday 06 November 2014 12:49:22 Thierry Reding wrote:
> >> > GCC complains about the format specifier being wrong. %zu/%zd are the
> >> > correct specifiers for variables of type size_t/ssize_t, so wherever a
> >> > size_t or ssize_t is used as parameter it should have a corresponding
> >> > %zu or %zd specifier.
> >> >
> >> > Why not just fix it properly instead of mucking about with the size_t
> >> > typedef?
> >> >
> >>
> >> Yes, but where are %zu and %zd implemented in gcc? I've looked but
> >> couldn't find it. For all I can tell is that gcc's own interpretation
> >> of %z doesn't match its definition of __SIZE_TYPE__ when building for
> >> bare-metal.
> >
> > I think the code you're looking for is gcc/c-family/c-format.c in
> > function format_type_warning() (and its callers). Now my understanding
> > of GCC internals is fairly limited, but what I think is happening is
> > that the matching happens on the exact typedef, so even if size_t is
> > typedef'd to unsigned int and the argument is of type unsigned int the
> > check will still fail and cause the warning.
> >
> 
> Yes, that is what it looks like.
> 
> > Interestingly I can't seem to reproduce these warnings, neither with the
> > native compiler on my system nor a 4.9.0 ARM cross-compiler. I've used
> > the attached source file as a test case (derived from the code at line
> > 1244 in drivers/regmap/regmap.c, which causes the warning in one of your
> 
> The mystery is that arm-linux-gnueabihf-gcc does not produce the
> warning, whereas Olof's arm-none-eabi-gcc does (built from the same
> source version, as far as we can tell)
> 
> > logs). Can you verify that the same warning happens when you run your
> > compiler on that? I've tried compiling with something like this:
> >
> >         $ gcc -Wall -Wextra -Wformat=2 -c test.c
> >
> > I don't think all of these are necessary, plain -Wformat /should/ be
> > enough.
> >
> > It's weird because both operands of the division are size_t, so I'm
> > guessing the compiler looses the type information at some point.
> >
> 
> I have tried this with the BE bare metal compiler I have on my system,
> and it did not produce a warning.
> (gcc-linaro-armeb-none-eabi-4.9-2014.06)

Interesting. Can you still reproduce the build warnings from Olof's logs
if you run that compiler on the kernel? It'd be nice to trigger this in
a minimal test case since it seems to be something that needs to be
fixed in GCC.

Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20141106/5a5a8355/attachment.sig>