[glean-dev] RFC: dsconfig matching changes

[Brian P. <br...@vm...>]

A coworker found that glean was hitten an assertion failure in 
treadpixperf.cpp when doing results comparisons.  I tracked the issue 
down to a problem with the drawing surface config matching code.

A glean results file contains some number of drawing surface config 
descriptions followed by test-specific results.  When comparing 
results, glean tries to finding matching ds configs in the two files 
to compare.  The DrawingSurfaceConfig::match() function was doing a 
poor job of matching configs so we wound up trying to compare test 
results from (very) different configs and failing.

In particular, this bit of code is used to compare the Z and stencil 
attributes:

    if (z && c.z)
       error += abs(z - c.z);
    if (s && c.s)
       error += abs(s - c.s);

If z = 0 and c.z = 24 the error variable wasn't updated.  I believe 
the correct way to do this is simply:

   error += abs(z - c.z);
   error += abs(s - c.s);

Furthermore, other fields like db and stereo weren't even compared.

Even after I fixed this, however, the matching code sometimes found 
that two configs were matches when there were in fact other, even 
better/exact matches.

To fix this, I modified the match function to first look for an 
_exact_ match where all fields (including X visual ID) were the same. 
  If that fails, then try the error-based matching algorithm. 
Furthermore, I modifed the error-based comparison code to check more 
fields and use different error weights for different fields.

The attached patch implements these changes.  The treadpixperf 
assertion no longer fails and everything else I've tested works properly.

If there are no concerns/objections I'll commit this later.

Be sure you grab the latest code from git.  I just recently committed 
some other fixes (uninitialized memory, etc).

-Brian