ports/181838: [patch] x11-servers/xorg-server: fix TLS variables with clang

[Tijl Coosemans]

The following reply was made to PR ports/181838; it has been noted by GNATS.

From: Tijl Coosemans <tijl at FreeBSD.org>
To: Niclas Zeising <zeising at freebsd.org>
Cc: bug-followup at FreeBSD.org
Subject: Re: ports/181838: [patch] x11-servers/xorg-server: fix TLS
 variables with clang
Date: Fri, 6 Sep 2013 23:22:29 +0200

 On Fri, 06 Sep 2013 22:18:27 +0200 Niclas Zeising wrote:
 > Can you please elaborate a bit on this, exactly what happends when
 > setting that configure variable?
 
 The configure script tests how to declare a thread-local variable.  For
 clang and gcc this is done using the __thread keyword like this:
 
 __thread int thread_specific_int_var;
 
 Other compilers use __declspec(thread) instead of __thread.
 
 The code in configure comes from the XORG_TLS call in configure.ac.
 XORG_TLS is defined in m4/xorg-tls.m4. Here are the relevant bits:
 
     AC_MSG_CHECKING(for thread local storage (TLS) support)
     AC_CACHE_VAL(ac_cv_tls, [
         ac_cv_tls=none
         keywords="__thread __declspec(thread)"
         for kw in $keywords ; do
             AC_TRY_COMPILE([int $kw test;], [], ac_cv_tls=$kw)
         done
     ])
     AC_MSG_RESULT($ac_cv_tls)
 
 The for-loop runs over the elements in $keywords and sets ac_cv_tls
 to the keyword if the compiler accepts it.  The problem is that clang
 accepts __declspec(thread) but it doesn't actually work (there's a
 warning about this but no error) so when the for-loop ends ac_cv_tls
 is set to __declspec(thread).  Because the test is inside AC_CACHE_VAL
 it can be skipped by defining ac_cv_tls which is what the patch does.
 
 I found a similar test in pixman but there the for-loop contains a
 break so it ends on the first successful keyword (i.e. __thread).
 
 > Have you tried and ensured that this doesn't break xserver built with
 > gcc?
 
 No, I currently don't have gcc installed, but the keyword for gcc is
 also __thread.  You could search for __thread in a config.log generated
 with gcc to verify this.