i965: Don't add barrier deps for FB write messages.
authorKenneth Graunke <kenneth@whitecape.org>
Thu, 4 Feb 2016 16:10:02 +0000 (08:10 -0800)
committerKenneth Graunke <kenneth@whitecape.org>
Tue, 9 Feb 2016 00:59:35 +0000 (16:59 -0800)
commitd0e1d6b7e27bf5f05436e47080d326d7daa63af2
tree8ab70cb8c75d8745ef03c1a7685a66ecf2956e6b
parent6502b3f60e193b314bd20261a8290709a4a56674
i965: Don't add barrier deps for FB write messages.

There are never render target reads, so there are no scheduling hazards.

Giving the extra flexibility to the scheduler makes it possible to do
FB writes as soon as their sources are available, reducing register
pressure.  It also makes it possible to do the payload setup for more
than one FB write message at a time, which could better hide latency.

shader-db results on Skylake:

total instructions in shared programs: 9110254 -> 9110211 (-0.00%)
instructions in affected programs: 2898 -> 2855 (-1.48%)
helped: 3
HURT:   0
LOST:   0
GAINED: 1

A reduction in instruction counts is surprising, but legitimate:
the three shaders helped were spilling, and reducing register
pressure allowed us to issue fewer spills/fills.

total cycles in shared programs: 69035108 -> 68928820 (-0.15%)
cycles in affected programs: 4412402 -> 4306114 (-2.41%)
helped: 4457
HURT: 213

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
src/mesa/drivers/dri/i965/brw_schedule_instructions.cpp