Marc Espie <[email protected]> writes: > On Mon, Aug 01, 2016 at 01:58:24PM +0200, Jeremie Courreges-Anglas wrote: >> [email protected] (Jeremie Courreges-Anglas) writes: >> >> > +cc espie and jasper >> > >> > Antoine Jacoutot <[email protected]> writes: >> > >> >> On Sun, Jul 31, 2016 at 07:21:39PM +0200, Antoine Jacoutot wrote: >> >>> On July 31, 2016 7:14:21 PM GMT+02:00, [email protected] wrote: >> >>> > >> >>> >Making read(2) return EISDIR for directories breaks two ports, both >> >>> >because they use libtool -bindir. cc(1) gets executed with an unknown >> >>> >option, -bindir, and a path such as /usr/local/bin, which then gets >> >>> >passed to ld(1). ld(1) copes with read(2) returning 0, not with >> >>> >EISDIR. >> >>> >Thanks to Antoine who ran the bulk builds that exposed this problem. >> >>> > >> >>> >-bindir support is meaningless on OpenBSD so handling that option >> >>> >should >> >>> >be easy. The problem is that I don't know how to implement it in >> >>> >libtool(1). GNU libtool recognizes -bindir among cc flags, while our >> >>> >version seems to only handle options passed right after argv[0]. >> >>> > >> >>> >I plan to work around that problem by using GNU libtool for the ports >> >>> >mentioned above, but someone else might want to poke at libtool(1) >> >>> >internals. :) >> >>> > >> >>> >-- >> >>> >jca | PGP: 0x1524E7EE / 5135 92C1 AD36 5293 2BDF DDCC 0DFA 74AE 1524 >> >>> >E7EE >> >>> >> >>> I can have a look at it during g2k16 if no one beats me to it. >> >>> -- >> >>> Antoine >> >> >> >> Hi Jeremie. >> >> >> >> This seems to do the trick for me: >> > >> > Thanks for looking at it. Yup, that works, but I think I know >> > understand why I got confused first. >> > >> > -bindir is just one option among others that should be recognized and >> > ignored in *link* mode. The following diff does just that, I think it >> > fits better in the existing code. I can successfully build >> > devel/libiscsi, -bindir /usr/local/bin doesn't get passed to cc(1). >> > >> > Thoughts? >> > >> > >> > Index: Link.pm >> > =================================================================== >> > RCS file: /cvs/src/usr.bin/libtool/LT/Mode/Link.pm,v >> > retrieving revision 1.31 >> > diff -u -p -p -u -r1.31 Link.pm >> > --- Link.pm 27 Apr 2016 09:50:57 -0000 1.31 >> > +++ Link.pm 1 Aug 2016 11:36:31 -0000 >> > @@ -127,6 +127,7 @@ sub run >> > 'all-static', >> > 'allow-undefined', # we don't care about THAT one >> > 'avoid-version', >> > + 'bindir:', >> > 'dlopen:', >> > 'dlpreopen:', >> > 'export-dynamic', >> > @@ -152,7 +153,7 @@ sub run >> > 'version-info:', >> > 'version-number:'); >> > >> > - # XXX options ignored: dlopen, dlpreopen, no-fast-install, >> > + # XXX options ignored: bindir, dlopen, dlpreopen, no-fast-install, >> > # no-install, no-undefined, precious-files-regex, >> > # shrext, thread-safe, prefer-pic, prefer-non-pic >> >> I'm wondering about the second hunk: is this a mix of options that we >> don't support on purpose and of options that could be useful? >> (eg. -no-undefined.) So is the second hunk of this diff desirable? >> > The main thing about that diff is that it has to go into a successful > bulk.
It did, thanks to Antoine, no breakage. I already got an ok from Antoine. Objections / ok? > As for ignored options, it's informative. It tells us we ignore those > options. Which ones should actually be supported is another story. Fine with me. -- jca | PGP: 0x1524E7EE / 5135 92C1 AD36 5293 2BDF DDCC 0DFA 74AE 1524 E7EE
