glsl: add pass to lower variable array indexing to conditional assignments
authorLuca Barbieri <luca@luca-barbieri.com>
Tue, 7 Sep 2010 23:35:44 +0000 (01:35 +0200)
committerIan Romanick <ian.d.romanick@intel.com>
Fri, 17 Sep 2010 08:58:58 +0000 (10:58 +0200)
commita47539c7a155475de00fa812842721d239abb3f4
treeef930357c3500f4bb490d0ae46d565d9985e87c7
parentdab2a7660a407364a33337327743b56ea9701d9b
glsl: add pass to lower variable array indexing to conditional assignments

Currenly GLSL happily generates indirect addressing of any kind of
arrays.

Unfortunately DirectX 9 GPUs are not guaranteed to support any of them in
general.

This pass fixes that by lowering such constructs to a binary search on the
values, followed at the end by vectorized generation of equality masks, and
4 conditional assignments for each mask generation.

Note that this requires the ir_binop_equal change so that we can emit SEQ
to generate the boolean masks.

Unfortunately, ir_structure_splitting is too dumb to turn the resulting
constant array references to individual variables, so this will need to
be added too before this pass can actually be effective for temps.

Several patches in the glsl2-lower-variable-indexing were squashed
into this commit.  These patches fix bugs in Luca's original
implementation, and the individual patches can be seen in that branch.
This was done to aid bisecting in the future.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
src/glsl/Makefile
src/glsl/Makefile.am
src/glsl/SConscript
src/glsl/ir_optimization.h
src/glsl/lower_variable_index_to_cond_assign.cpp [new file with mode: 0644]