The most likely explanation is the overhead from computing the gl_client_arrays in _mesa_update_state(). Lightsmark makes a lot of calls to glDrawElements() each frame, and changes the vertex arrays between those calls.

That being said, I can't reproduce the regression with r600g. I get approximately 115 FPS with and without the ARB_vertex_attrib_binding changes.

Could you investigate further?

Looking at other test data from relevant time period:

* Why Lighstmark test is marked on HSW D as "blocked"?

* On HSW ULT & Mobile, performance drop is ~30% between 11-10 and 11-11, on IVB D & M it's 37%, on BYT, the performance drop between 11-07 and 11-11 is ~45%.

* On HSW D, x11perf "aa10text" regressed during this interval 13% and "rgb10text" regressed by 68%.  x11perf stuff is CPU bound, but it could be unrelated as on BYT "aa10text" got 45% increase in its performance.

* There's no performance drop in any other tests on any other tested HW.


> Lightsmark makes a lot of calls to glDrawElements() each frame,

Nope.  It's an old program using glBegin/glEnd...

Stuff that Lighstmarks does in every frame:
   44922 glBindTexture (avg = 27.4 calls in each of 1638 frames)
   42369 glEnable (avg = 25.9 calls in each of 1638 frames)
   42336 glActiveTexture (avg = 25.8 calls in each of 1638 frames)
   30908 glVertex2f (avg = 18.9 calls in each of 1638 frames)
   30908 glMultiTexCoord2f (avg = 18.9 calls in each of 1638 frames)
   22780 glGetUniformLocation (avg = 13.9 calls in each of 1638 frames)
   22000 glMatrixMode (avg = 13.4 calls in each of 1637 frames)
   20253 glDisable (avg = 12.4 calls in each of 1638 frames)
   13310 glGetIntegerv (avg = 8.1 calls in each of 1638 frames)
    9737 glUniform1i (avg = 5.9 calls in each of 1638 frames)
    9582 glBlendFunc (avg = 5.9 calls in each of 1637 frames)
    7727 glEnd (avg = 4.7 calls in each of 1638 frames)
    7727 glBegin (avg = 4.7 calls in each of 1638 frames)
    7586 glUseProgram (avg = 4.6 calls in each of 1638 frames)
    7255 glUniform4f (avg = 4.4 calls in each of 1638 frames)
    6714 glPushMatrix (avg = 4.1 calls in each of 1637 frames)
    6714 glPopMatrix (avg = 4.1 calls in each of 1637 frames)
    6639 glLoadMatrixd (avg = 4.1 calls in each of 1637 frames)
    6534 glDepthMask (avg = 4.0 calls in each of 1638 frames)
    4917 glXMakeContextCurrent (avg = 3.0 calls in each of 1638 frames)
    4837 glClear (avg = 3.0 calls in each of 1637 frames)
    4762 glAlphaFunc (avg = 2.9 calls in each of 1637 frames)
    3954 glTexImage2D (avg = 33.5 calls in each of 118 frames)
    3754 glMultMatrixd (avg = 2.3 calls in each of 1637 frames)
    3274 glLoadIdentity (avg = 2.0 calls in each of 1637 frames)
    3267 glGetBooleanv (avg = 2.0 calls in each of 1638 frames)
    2708 glViewport (avg = 1.7 calls in each of 1638 frames)
    2509 glCallList (avg = 1.5 calls in each of 1637 frames)
    2464 glGetFloatv (avg = 1.5 calls in each of 1637 frames)
    1991 glColor3f (avg = 1.2 calls in each of 1637 frames)
    1922 glUniform3f (avg = 1.2 calls in each of 1637 frames)
    1736 glUniform4fv (avg = 1.1 calls in each of 1635 frames)
    1638 glFlush (avg = 1.0 calls in each of 1638 frames)
    1637 glOrtho (avg = 1.0 calls in each of 1637 frames)
    1637 glColor4f (avg = 1.0 calls in each of 1637 frames)

Stuff that Lighstmark does in some frames:
   60558 glClientActiveTexture (avg = 348.0 calls in each of 174 frames)
   60378 glTexCoordPointer (avg = 347.0 calls in each of 174 frames)
   30787 glVertexPointer (avg = 176.9 calls in each of 174 frames)
   30786 glDrawElements (avg = 176.9 calls in each of 174 frames)
   30441 glNormalPointer (avg = 174.9 calls in each of 174 frames)
   29929 glColorPointer (avg = 173.0 calls in each of 173 frames)
   28877 glCullFace (avg = 166.0 calls in each of 174 frames)
    6702 glTexParameteri (avg = 36.8 calls in each of 182 frames)
    3685 glPixelStorei (avg = 31.2 calls in each of 118 frames)
    2347 glBindFramebufferEXT (avg = 3.5 calls in each of 680 frames)
    1583 glFramebufferTexture2DEXT (avg = 2.3 calls in each of 680 frames)
    1583 glEnableClientState (avg = 9.1 calls in each of 174 frames)
    1569 glGenTextures (avg = 65.4 calls in each of 24 frames)
    1360 glColorMask (avg = 2.0 calls in each of 680 frames)
     933 glGetTexLevelParameteriv (avg = 1.4 calls in each of 684 frames)
     883 glDisableClientState (avg = 5.1 calls in each of 174 frames)
     819 glCheckFramebufferStatusEXT (avg = 1.2 calls in each of 680 frames)
     736 glPolygonOffset (avg = 1.1 calls in each of 680 frames)
     680 glClearDepth (avg = 1.0 calls in each of 680 frames)
     218 glUniformMatrix4fv (avg = 2.2 calls in each of 101 frames)
     177 glColor4ub (avg = 1.0 calls in each of 174 frames)
     167 glClearColor (avg = 2.0 calls in each of 84 frames)
     167 glReadBuffer (avg = 2.0 calls in each of 83 frames)
     167 glDrawBuffer (avg = 2.0 calls in each of 83 frames)
      83 glUniform2f (avg = 1.0 calls in each of 83 frames)
      83 glReadPixels (avg = 1.0 calls in each of 83 frames)

Verified it with Mesa git-45a56ce39.

(In reply to comment #3)
> Verified it with Mesa git-45a56ce39.

Can you bisect that?  I don't see anything in the range a594cec..45a56ce39.  The only thing that worries me more than unexplained bugs is unexplained fixes. :(

Also, is the original performance drop reproducible on the 10.0 branch?

Did the kernel change at any point?

(In reply to comment #4)
by bisect the fixed commit is
commit ff353c218a1ab1fd3fb797a4780612ec4b0451d8
Author: Fredrik Höglund <fredrik@kde.org>
Date:   Mon Nov 11 18:54:15 2013 +0100

    mesa: Fix derived vertex state not being updated in glCallList()

    AEcontext::NewState is not always set when the vertex array state
    is changed.

(In reply to comment #4)
It's reproducible on 10.0 branch.

(In reply to comment #6)
> (In reply to comment #4)
> It's reproducible on 10.0 branch.

But ff353c218a1ab1fd3fb797a4780612ec4b0451d8 was cherry-picked
to the 10.0 branch on November 15 according to git.

(In reply to comment #7)
Hi, it's reproducible on 10.0 branch because git-59b01ca also exists on 10.0.
It's also fixed with patch ff353c218 on 10.0.