Bug 71506

17c94de33baf66ad5c264b7a046394c651bc6126 is the first bad commit
commit 17c94de33baf66ad5c264b7a046394c651bc6126
Author: Ian Romanick <ian.d.romanick@intel.com>
Date:   Mon Sep 10 17:11:33 2012 +0300

    mesa/dri: Add basic plumbing for GLX_ARB_robustness reset notification strategy
    
    No drivers advertise the DRI2 extension yet, so no driver should ever
    see a value other than false for notify_reset.
    
    The changes in nouveau use tabs because nouveau seems to have it's own
    indentation rules.
    
    Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
    Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

:040000 040000 76b2737e9b48d9c042d4bc05020803c7328d605a dba600f5c9a4c2e147e30e60fbd5b7f286aaa17a M	src
bisect run success

That is bizarre.  I always 'make check' as part of my build, and I haven't seen that failure.  There must be something different about your configure flags that allows this to happen.  Something is happening that causes src/glx/indirect_glx.o go get pulled in when linking glx-test, and it should not.

Can you attach the output of 'nm -u src/glx/tests/*.o'?

Created attachment 89058 [details]
nm -u src/glx/tests/*.o

Hrm... diffing that with mine doesn't show anything obvious.  How about 'nm -u src/glx/.libs/create_context.o'?  I get:

[idr@mumford-wire master-64]$ nm -u src/glx/.libs/create_context.o 
                 U __assert_fail
                 U free
                 U GetGLXScreenConfigs
                 U _GLOBAL_OFFSET_TABLE_
                 U __glXSendErrorForXcb
                 U indirect_create_context_attribs
                 U xcb_generate_id
                 U xcb_glx_create_context_attribs_arb_checked
                 U xcb_request_check
                 U XGetXCBConnection

I see the same output.

$ nm -u src/glx/.libs/create_context.o
                 U __assert_fail
                 U free
                 U GetGLXScreenConfigs
                 U _GLOBAL_OFFSET_TABLE_
                 U __glXSendErrorForXcb
                 U indirect_create_context_attribs
                 U xcb_generate_id
                 U xcb_glx_create_context_attribs_arb_checked
                 U xcb_request_check
                 U XGetXCBConnection

I'm really confused now. :(  You're sure that bisect is good?  There's nothing in that commit that modifies any file under src/glx.

I'm also not able to reproduce this even using your configure options.

If you revert changes to Makefile.am made by these two commits, test builds and passess fine:

http://cgit.freedesktop.org/mesa/mesa/commit/?id=8c5330226f391a7a29b6538851090b0ef730a239

http://cgit.freedesktop.org/mesa/mesa/commit/?id=0cce5538673148ffcd7aa9479f6b88cf9a641352

Even to be more precise, (manually) reverting this commit fixes the issue, at least here:

http://cgit.freedesktop.org/mesa/mesa/commit/?id=0cce5538673148ffcd7aa9479f6b88cf9a641352

Reproducible with GCC 4.8.2 and Binutils 2.23.2.

Hm... does building with --enable-glx-tls help?  I wonder if the added __glXGetCurrentContext in fake_glx_screen.cpp is to blame.

It builds for me if query_renderer_unittest.cpp is removed from list of source files without any modifications to fake_glx_screen.cpp

And indeed, using --enable-glx-tls seems to fix the build.

Created attachment 89110 [details] [review]
Patch to glx/tests to provide a stub __glXGetCurrentContext() function when $(DEFINES) are such that is not a macro

(In reply to comment #9)
> Hm... does building with --enable-glx-tls help?  I wonder if the added
> __glXGetCurrentContext in fake_glx_screen.cpp is to blame.

Apparently so.  Here's the relevant output doing a build with LDFLAGS=-Wl,--print-map

> Archive member included because of file (symbol)
> 
> ../../../src/glx/.libs/libglx.a(clientinfo.o)
>                               clientinfo_unittest.o (__glX_send_client_info)
> ../../../src/glx/.libs/libglx.a(create_context.o)
>                               create_context_unittest.o (glXCreateContextAttribsARB)
> ../../../src/glx/.libs/libglx.a(indirect_size.o)
>                               enum_sizes.o (__glCallLists_size)
> ../../../src/glx/.libs/libglx.a(indirect_init.o)
>                               indirect_api.o (__glXNewIndirectAPI)
> ../../../src/glx/.libs/libglx.a(query_renderer.o)
>                               query_renderer_unittest.o (glXQueryRendererIntegerMESA)
> ../../../src/glx/.libs/libglx.a(glxcurrent.o)
>                               ../../../src/glx/.libs/libglx.a(query_renderer.o) (__glXGetCurrentContext)
> ../../../src/glx/.libs/libglx.a(glxext.o)
>                               ../../../src/glx/.libs/libglx.a(glxcurrent.o) (__glXSetupForCommand)
> ../../../src/glx/.libs/libglx.a(glxconfig.o)
>                               ../../../src/glx/.libs/libglx.a(glxext.o) (glx_config_destroy_list)
> ../../../src/glx/.libs/libglx.a(glx_query.o)
>                               ../../../src/glx/.libs/libglx.a(glxext.o) (__glXQueryServerString)
> ../../../src/glx/.libs/libglx.a(glxhash.o)
>                               ../../../src/glx/.libs/libglx.a(glxext.o) (__glxHashCreate)
> ../../../src/glx/.libs/libglx.a(indirect_glx.o)
>                               ../../../src/glx/.libs/libglx.a(glxext.o) (indirect_create_screen)

So it seems that that the removal of the __glXGetCurrentContext stub in commit 8c5330226f391a7a29b6538851090b0ef730a239 "After adding $(DEFINES) to AM_CPPFLAGS, the __glXGetCurrentContext wrapper function is no longer needed and causes compile errors. Using the correct defines causes it to be a macro!" is not always correct.

Reading src/glx/glxclient.h, __glXGetCurrentContext is not a macro if HAVE_PTHREAD && !GLX_USE_TLS, so the stub should probably be provided in that case.  Patch attached.

To ssh://git.freedesktop.org/git/mesa/mesa
   21ae513..5ab59e5  master -> master

Might be a good idea to pick this up for the 10.0 branch ?