Re: [patch libquadmath]: Fix build of gfortran using libquadmath for pe-coff targets

[Dave Korn <dave dot korn dot cygwin at gmail dot com>]

On 23/11/2010 19:02, IainS wrote:
> 
> On 23 Nov 2010, at 18:52, Kai Tietz wrote:
> 
>> 2010/11/23 Dave Korn <dave.korn.cygwin@gmail.com>:
>>> On 21/11/2010 19:12, Kai Tietz wrote:
>>>> Hello,
>>>> 
>>>> by recent libquadmath patch there was a bootstrap issue for 
>>>> i686-pc-cygwin, i686-pc-mingw32, and x86_64-w64-mingw32. I was 
>>>> reasoned by some minor nits in configure and makefile AFAICS.
>>> 
>>>> (libquadmath_la_LDFLAGS): Add -no-undefined.
>>> 
>>> Argh.  This can have undesired side-effects on Darwin.  It may have to
>>> only be added for Windows hosts.  (I have to fix this same problem for 
>>> lto-plugin.)
>> 
>> Well, this is the same pattern as used in libssp. So if darwin has an 
>> issue here, we have more then one place in gcc doing this.
> 
> Well, maybe unconditionally overriding flags for all targets is probably 
> prone to gotchas anyway. 

  Well, the idea is for libtool to abstract away all the dependencies and
spare us having to know which flags to set for which host, it should probably
be considered a bug that one option has these two incompatible semantics.  I
can't see any simple way to change it now without breaking existing makefiles
though.  (Well, except if we made -no-undefined *not* mean the same thing as
not giving a -undefined option, but /only/ mean 'build a dll on windows'; but
it would be really illogical if there was a -no- option that wasn't the
opposite of it's plain form.)

> For Darwin, it will cause an issue in the case
> that the linked object needs to resolve symbols dynamically. (that might
> not be relevant here or to libssp - but it probably ought to be fixed there
> too).

  Yep.  Feel free to file me a PR mentioning both that and lto-plugin if you
like, I plan to fix it during this stage3.

    cheers,
      DaveK