745224e0 gcc-4.9 emmintrin.h build error

The changes in 745224e0 don't seem to build here with gcc-4.9 on linux x64_64. Any idea what's wrong?

    CC credential-store.o
In file included from /usr/lib/.../xmmintrin.h:31:0,
                 from /usr/lib/.../emmintrin.h:31,
                 from git-compat-util.h:708,
                 from cache.h:4,
                 from credential-store.c:1:
/usr/lib/.../mmintrin.h: In function '_mm_cvtsi32_si64':
/usr/lib/.../mmintrin.h:64:3: error: can't convert between vector
values of different size
   return (__m64) __builtin_ia32_vec_init_v2si (__i, 0);
   ^
/usr/lib/.../mmintrin.h: In function '_mm_cvtsi64_si32':
/usr/lib/.../mmintrin.h:107:10: error: incompatible type for argument
1 of '__builtin_ia32_vec_ext_v2si'
   return __builtin_ia32_vec_ext_v2si ((__v2si)__i, 0);
          ^

[...]

In file included from /usr/lib/.../emmintrin.h:31:0,
                 from git-compat-util.h:708,
                 from cache.h:4,
                 from credential-store.c:1:
/usr/lib/.../xmmintrin.h: In function '_mm_add_ss':
/usr/lib/.../xmmintrin.h:127:19: error: incompatible type for argument
1 of '__builtin_ia32_addss'
   return (__m128) __builtin_ia32_addss ((__v4sf)__A, (__v4sf)__B);
                   ^
/usr/lib/.../xmmintrin.h:127:3: note: expected '__vector(4) float' but
argument is of type '__m128'
   return (__m128) __builtin_ia32_addss ((__v4sf)__A, (__v4sf)__B);
   ^
/usr/lib/.../xmmintrin.h:127:19: error: incompatible type for argument
2 of '__builtin_ia32_addss'
   return (__m128) __builtin_ia32_addss ((__v4sf)__A, (__v4sf)__B);

                   ^

[...]

/usr/lib/.../emmintrin.h:1455:3: error: incompatible type for argument
2 of '__builtin_ia32_movntpd'
   __builtin_ia32_movntpd (__A, (__v2df)__B);
   ^
/usr/lib/.../emmintrin.h:1455:3: note: expected '__vector(2) double'
but argument is of type '__m128d'
Makefile:1983: recipe for target 'credential-store.o' failed
make: *** [credential-store.o] Error 1

On Thu, Jul 10, 2014 at 9:59 PM, Tuncer Ayaz wrote:

s/x64_64/x86_64/

Should have written amd64 to avoid the typo :).

On Thu, Jul 10, 2014 at 09:59:37PM +0200, Tuncer Ayaz wrote:

Weird. It compiles fine on my x86_64 box (Debian unstable, gcc 4.9.0-10). Can you tell us more about your environment?

-Peff

On Thu, Jul 10, 2014 at 10:23 PM, Jeff King wrote:

gcc version 4.9.0 20140604 (prerelease)

I normally use a custom config.mak, but I get the error without it too, so it has no direct effect.

Let me know if there's anything to try out for finding the difference.

On Thu, 2014-07-10 at 21:59 +0200, Tuncer Ayaz wrote:

What's in the ...?

Because if you're using headers from a different version of gcc, that might explain it.

On Thu, Jul 10, 2014 at 10:33 PM, David Turner wrote:

/usr/lib/gcc/x86_64-unknown-linux-gnu/4.9.0/include/emmintrin.h

On Thu, 2014-07-10 at 22:44 +0200, Tuncer Ayaz wrote:

That seems fine to me.

It looks like the error messages are coming from inside the system's header files (but this is sometimes misleading). If you just try to compile

#include <emmintrin.h> int main() { }

with whatever options you use for git, does that work? If not, I would say that you have a compiler setup problem.

On Thu, Jul 10, 2014 at 10:53 PM, David Turner wrote:

The above test works on the same machine, so I'll investigate what's going on when building git. Thanks.

Sorry, didn't test properly when I tried with/without config.mak, and PROFILE=BUILD was the problem. I had that in config.mak based on information gathered from INSTALL and Makefile. To be clear, is PROFILE=BUILD (still) supported?

On Fri, 2014-07-11 at 00:12 +0200, Tuncer Ayaz wrote:

For what it's worth, the problem seems to depend on the combination of -DNO_NOTRETURN=1 and -fprofile-use. So I can trigger the same breakage with this: --- #define NO_NORETURN 1 #include "git-compat-util.h"

int main() {} --- gcc -I. -c -fprofile-use=/tmp foo.c -o foo.o

Do we still need NO_NORETURN? Git seems to build without it under GCC 4.6.3 (Ubuntu's version).

On Fri, Jul 11, 2014 at 12:12:55AM +0200, Tuncer Ayaz wrote:

I think none of the regular devs uses PROFILE, and it bit-rotted somewhat. Andi Kleen recently posted a series to fix it[1]. I can reproduce your problem without that series, but compiling with ak/profile-feedback-build merged in seems to work OK.

[1] http://thread.gmane.org/gmane.comp.version-control.git/253005

On Sat, Jul 12, 2014 at 6:56 AM, Jeff King wrote:

Yes, Andi's patch works.

Any idea when ak/profile-feedback-build will land in master?

On Sat, Jul 12, 2014 at 10:47:03AM +0200, Tuncer Ayaz wrote:

Junio posts updates in his "What's Cooking" messages (usually about twice a week). Last Tuesday's listed it as "will merge to next". Unless anybody finds problems with it, I'd expect it in master within a week or two, and probably part of v2.1 when it releases.

-Peff

On Sat, Jul 12, 2014 at 6:56 AM, Jeff King wrote:

By the way, is there no build (CI) server for git.git to regularly test branches on different platforms or at least different build configs on the same platform?

On Sat, Jul 12, 2014 at 10:53:47AM +0200, Tuncer Ayaz wrote:

Junio tests master and next on a few different platforms before pushing out the results of topic integration. Other than that, we rely on people to report problems (and people testing master or even next on their platforms is very much appreciated, as it means we can fix problems before they are cemented in a release).

-Peff