When checking output VS in glsl-mat-from-int-ctor-03 piglit, I got the following
(part of) code.
mov(8) g19<1>.xyzF g6<4,4,1>.xyzzD { align16
1Q };
dp4(8) g115<1>.wF g4<4,4,1>F g2.4<0,4,1>F { align16
NoDDChk 1Q };
cmp.nz.f0(8) null<1>F g11<4,4,1>.xyzzF g19<4,4,1>.xyzzF { align16
1Q switch };
cmp.nz.f0(8) null<1>D g7<4,4,1>D 0D { align16
1Q switch };
(+f0.any4h) mov(8) g21<1>.xUD 0xffffffffUD { align16
1Q };
Clearly the first cmp can be removed because the result is overwritten by the
second one.
Investigating why this is not happening, saw that in brw_vec4, after running
opt_vector_float(), we are running optimizations just once, instead of in a loop
until no progress happens.
Not sure if there is a reason to keep it separated from the previous
loop. Tracking back the code seems that originally it wasn't added because
opt_vector_float() wasn't written as an optimization, and no optimizations were
done after running it. But later someone suggested to run some optimizations if
opt_vector_float() success, which made to add a conditional. In final commits,
that opt_vector_float() was converted in a true optimiztaion, but still kept out
of the loop. At this point I'm not sure if there was a good reason (no
explanation found) or just to make the less changes in code. Maybe someone can
bring light here.
So merged those optimizations inside the previous loop (second commit). But this
made some piglit tests to never end. Checking about this, saw that some
optimizations were been reverted by others (specifically CSE reverted by
copy-propagation). So I added a minor change in CSE (first commit) that prevents
apply it when the common expression is just an immediate (it saves nothing, and
adds a new instruction).
The free shaders in shader-db are not improved (nor hurted). But testing against
the non-free shaders we get a small improvement:
total instructions in shared programs: 6819828 -> 6819468 (-0.01%)
instructions in affected programs: 30516 -> 30156 (-1.18%)
total loops in shared programs: 1971 -> 1971 (0.00%)
helped: 154
HURT: 0
GAINED: 0
LOST: 0
Juan A. Suarez Romero (2):
i965: Do not apply CSE opt to MOV immediate
i965: run brw_vec4 optimizations in loop
src/mesa/drivers/dri/i965/brw_vec4.cpp | 10 +++-------
src/mesa/drivers/dri/i965/brw_vec4_cse.cpp | 13 ++++++++++++-
2 files changed, 15 insertions(+), 8 deletions(-)
--
2.5.0
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
