Aargh. I'm further in getting my open_memstream implementation going (I've now verified that my temporary-file backup works on Cygwin 1.5.x, as well as on Linux with configure primed with ac_cv_func_open_memstream=no; while mingw encountered some C++ issues; separate messages to come soon). But in the process of testing it, I discovered that Solaris 10 __fpurge() is broken - the sequence fpurge()/fputc()/ftell() gives the wrong answer. It looks like none of our existing tests directly cover this situation, since the gnulib unit tests for fpurge passed when run in isolation (Solaris did not need a replacement for fclose in that situation). But once my new open_memstream forced unconditional overrides for fflush and fclose, then test-fpurge is now calling rpl_fclose which calls ftell() in exactly that scenario. As proof that the bug exists independently of gnulib:
$ cat foo.c
#include <stdio.h>
#include <stdio_ext.h>
int main() {
FILE *fp = fopen ("file", "r+");
fseek (fp, -1, SEEK_END);
if (getc (fp) != 'h') return 1;
if (getc (fp) != EOF) return 2;
__fpurge (fp);
if (putc ('!', fp) != '!') return 3;
printf ("%d\n", (int) ftell (fp));
if (fclose (fp) != 0) return 4;
return 0;
}
$ printf foogarsh > file
$ ./foo
-8183
I guess we need to beef up the unit tests to expose the Solaris bug,
then work on replacing broken __fpurge with a version that works
correctly on read-write streams.
--
Eric Blake ebl...@redhat.com +1-801-349-2682
Libvirt virtualization library http://libvirt.org
signature.asc
Description: OpenPGP digital signature
