svn commit: r310138 - head/lib/libc/stdio

[Warner Losh]

On Fri, Dec 16, 2016 at 3:07 PM, John Baldwin <jhb at freebsd.org> wrote:
> On Friday, December 16, 2016 04:53:04 PM Eric van Gyzen wrote:
>> On 12/16/2016 16:45, John Baldwin wrote:
>> > On Friday, December 16, 2016 08:53:26 PM Dimitry Andric wrote:
>> >> On 16 Dec 2016, at 20:31, Baptiste Daroussin <bapt at FreeBSD.org> wrote:
>> >>>
>> >>> On Fri, Dec 16, 2016 at 01:44:51AM +0000, Conrad E. Meyer wrote:
>> >>>> Author: cem
>> >>>> Date: Fri Dec 16 01:44:50 2016
>> >>>> New Revision: 310138
>> >>>> URL: https://svnweb.freebsd.org/changeset/base/310138
>> >>>>
>> >>>> Log:
>> >>>>  vfprintf(3): Add support for kernel %b format
>> >>>>
>> >>>>  This is a direct port of the kernel %b format.
>> >>>>
>> >>>>  I'm unclear on if (more) non-portable printf extensions will be a
>> >>>>  problem. I think it's desirable to have userspace formats include all
>> >>>>  kernel formats, but there may be competing goals I'm not aware of.
>> >>>>
>> >>>>  Reviewed by:    no one, unfortunately
>> >>>>  Sponsored by:   Dell EMC Isilon
>> >>>>  Differential Revision:  https://reviews.freebsd.org/D8426
>> >>>>
>> >>>
>> >>> I really don't think it is a good idea, if used in userland it would be make
>> >>> more of our code difficult to port elsewhere.
>> >>
>> >> Indeed, this is a bad idea.  These custom format specifiers should be
>> >> eliminated, not multiplied. :-)
>> >>
>> >>
>> >>> Other than that, it makes more difficult to use vanilla gcc with out userland.
>> >>> and it is adding more complexity to be able to build freebsd from a non freebsd
>> >>> system which some people are working on.
>> >>>
>> >>> Personnaly I would prefer to see those extensions removed from the kernel rather
>> >>> than see them available in userland.
>> >>
>> >> Same here.
>> >>
>> >>
>> >>> Can't we use simple helper function instead?
>> >>
>> >> Yes, please.  Just take the snprintb(3) function from NetBSD:
>> >>
>> >> http://netbsd.gw.com/cgi-bin/man-cgi?snprintb+3+NetBSD-current
>> >
>> > In general I agree with something like this instead, but it is quite a bit more
>> > tedious to use as you have to run it once to determine the length, allocate a
>> > buffer, and then run it again.  Calling malloc() for that buffer isn't always
>> > convenient in the kernel (though it should be fine in userland).  Having it live
>> > in printf() itself means the output is generated to the stream without having to
>> > manage a variable-sized intermediate buffer.
>>
>> I imagine most callers can simply use a char[sizeof(fmt)+C] on the stack, where
>> C is some constant that I haven't taken the time to calculate, at the risk of
>> making myself look foolish and unprofessional.
>
> Hmm, that might work, but it is still cumbersome.  Probably to make things readable
> we'd end up with a wrapper:
>
> printb(uint val, const char *fmt)
> {
>    char buf[strlen(fmt) + C];
>
>    snprintb(...);
>    printf("%s", buf);
> }

Sadly this "cure" is worse than the disease.

Warner