nir: Move the compare-with-zero optimizations to the late section
authorJason Ekstrand <jason.ekstrand@intel.com>
Tue, 24 Mar 2015 00:36:22 +0000 (17:36 -0700)
committerJason Ekstrand <jason.ekstrand@intel.com>
Wed, 1 Apr 2015 19:51:03 +0000 (12:51 -0700)
commite06a3d0282ef018ae4a28b091a632c6a2e8c7c52
tree9198195292e0a8df1d7e80e226056f0017065a9d
parentda294f9b2f666f487001b2a25627c867c40eb3d9
nir: Move the compare-with-zero optimizations to the late section

total instructions in shared programs: 4422307 -> 4422363 (0.00%)
instructions in affected programs:     4230 -> 4286 (1.32%)
helped:                                0
HURT:                                  12

While this does hurt some things, the losses are minor and it prevents the
compare-with-zero optimization from fighting with ffma which is much more
important.

Reviewed-by: Matt Turner <mattst88@gmail.com>
src/glsl/nir/nir_opt_algebraic.py