This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: functions statically linked on x86_64 while i386 uses glibc functions


Andi Kleen <ak@suse.de> wrote:
> Hello,
>
> I was doing some binary size comparisons between -m32 and -m64 builds on x86_64
> Linux (SUSE 10.2 with gcc 4.1.2). This is the same build environment, same glibc, etc.
> just one configured for i386-linux and built with -m32 and the other for x86_64-linux
>
> One thing I noticed is that a few executable sizes differ widely:
>
> (.text sizes)   64bit   32bit
> ./src/csplit    95481   27704
> ./src/nl        87879   19551
> ./src/ptx       98316   28623
> ./src/expr      88373   20731
> ./src/tac       83726   15935
>
> The other differences are in the expected range of a few percent.
>
> On some investigation it looks like 64bit contains the regexpr functions and some
> other library functions, while the 32bit versions use the one from glibc.
> I assume that must be a (harmless) bug since the glibc version (2.5) is identical
> between 32bit and 64bit.

Hi Andi,

Thanks for reporting that.  glibc-2.5's regex code probably still suffers
from the bug that makes it misbehave on 64-bit systems for strings
longer than 2^31 bytes.  This has been fixed in gnulib's regex code since
glibc-2.3.5.  So, coreutils works around it, along with any other project
that uses gnulib's regex module, by detecting the offending condition and
providing a working replacement.  Here's the relevant part of m4/regex.m4:

	    /* Reject hosts whose regoff_t values are too narrow.
	       These include glibc 2.3.5 on hosts with 64-bit ptrdiff_t
	       and 32-bit int.  */
	    if (sizeof (regoff_t) < sizeof (ptrdiff_t)
		|| sizeof (regoff_t) < sizeof (ssize_t))
	      exit (1);

When that configure-run test program fails, those programs link with
the replacement regex functions from gnulib.

If you want to save space in those executables, and you don't think
there's a significant risk of malfunction[*], you can make coreutils
use the library code like this:

  gl_cv_func_re_compile_pattern_working=yes ./configure

[*] Can you even get a 2GB-long string into expr?
I rarely use the regexp code in tac and nl, and use ptx even less
frequently.  Personally, I'm not worried about csplit failing if/when
I run it on a file with such a long line -- I'll have other, more
immediate problems with the file, like getting that line into memory :)


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]