command_buffer: Fix setting samplers with bound uniforms

gpu/command_buffer/service/program_manager.cc

Index: gpu/command_buffer/service/program_manager.cc
diff --git a/gpu/command_buffer/service/program_manager.cc b/gpu/command_buffer/service/program_manager.cc
index f90c39d50962122ba4444eb8838fd4b0b64d9929..716cfbaa5ebf51f255695db216b93c28bfcdb5dd 100644
--- a/gpu/command_buffer/service/program_manager.cc
+++ b/gpu/command_buffer/service/program_manager.cc
@@ -1410,7 +1410,7 @@ bool Program::SetSamplers(
   }
   size_t location_index =
       GetUniformLocationIndexFromFakeLocation(fake_location);
-  if (location_index >= uniform_infos_.size())
+  if (location_index >= uniform_locations_.size())

[Zhenyao Mo, 2016/01/20 18:41:13]

Correct me if I am wrong.

uniform_locations_ size is defined by the largest client bound location.  If the
client does not bind a uniform location, that location could still be a valid
location - it's just automatically bound to a uniform by the driver.

Also, even if the client binds to a specific location, the underlying uniform
could still be optimized away, or the name does not corresponding to a uniform
at all.

Is that the case?  If yes, then I don't think this change works in general.

[Kimmo Kinnunen, 2016/01/21 12:36:32]

No, automatically bound uniforms get their client-side uniform locations
assigned by the command buffer.
See above (1).
If the client binds an uniform and driver optimizes the uniform away, the
uniform location still exist -- this was the fix while back. The check is below
and above, (2).

     return false;
 
   if (!uniform_locations_[location_index].IsActive())

[Kimmo Kinnunen, 2016/01/21 12:36:32]

(2): This checks if the uniform has been optimized away by the driver.
This should not be executed by glUniform1i, though.
See the other (2).

I'll add an clarification, see the new patch.