i965: Make brw_reg_from_fs_reg() halve exec_size when compressed.
authorKenneth Graunke <kenneth@whitecape.org>
Fri, 13 May 2016 06:04:10 +0000 (23:04 -0700)
committerKenneth Graunke <kenneth@whitecape.org>
Tue, 17 May 2016 21:40:37 +0000 (14:40 -0700)
commitdc657a8201fe942fce2c68835f1ee52b348e52f0
tree662f0ee106c27129930a54a9bb25f16a4157733a
parent062ad8166936d18a4b7ae11f5d39e8f39a9ae6c3
i965: Make brw_reg_from_fs_reg() halve exec_size when compressed.

In a5d7e144eaf43fee37e6ff9e2de194407087632b, Connor generalized the
exec_size halving code to handle more cases.  As part of this, he made
it not halve anything if the region accessed falls completely in a
single register.

Unfortunately, it started producing some invalid regions:

-add(16)  g6<1>F  g10<8,8,1>UW    -g1<0,1,0>F    { align1 compr };
-add(16)  g8<1>F  g12<8,8,1>UW    -g1.1<0,1,0>F  { align1 compr };
+add(16)  g6<1>F  g10<16,16,1>UW  -g1<0,1,0>F    { align1 compr };
+add(16)  g8<1>F  g12<16,16,1>UW  -g1.1<0,1,0>F  { align1 compr };

Here, the UW source region completely fits within a register.  However,
we have to use instruction compression because the destination region
spans two registers.  <16,16,1> is invalid because it's compressed.

To handle this, skip the "everything fits in one register" case and
fall through to the exec_size halving case when compressed.

Fixes hundreds of Piglit regressions on GM965.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95370
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
src/mesa/drivers/dri/i965/brw_fs_generator.cpp