r600/sb: fix rotated register in while loop
authorGert Wollny <gw.fossdev@gmail.com>
Thu, 31 May 2018 21:25:09 +0000 (23:25 +0200)
committerDave Airlie <airlied@redhat.com>
Mon, 25 Jun 2018 04:39:41 +0000 (05:39 +0100)
commitcd7db0ab0a871b34d6d07a1db9d3b8b4b8342b35
tree6c6b7ad2c69fb0d3c2c8359fb52d79421b8703f0
parent1977e92ee3a9d1a1ba9c5afec1eb3bd908c3f297
r600/sb: fix rotated register in while loop

This patch is based on
https://lists.freedesktop.org/archives/mesa-dev/2018-February/185805.html

Dave Airlie:

 "A bunch of CTS tests led me to write
  tests/shaders/ssa/fs-while-loop-rotate-value.shader_test
  which r600/sb always fell over on.

  GCM seems to move some of the copies into other basic blocks,
  if we don't allow this to happen then it doesn't seem to schedule
  them badly.

  Everything I've read on SSA/phi copies say they have to happen
  in parallel, so keeping them in the same basic block seems like
  a good way to keep some of that property."

This patch differs from the one proposed by Dave in that it only adds
the NF_DONT_MOVE flag to copy_move instructions that are created by split_phi*
and that are located in loops.

Fixes piglit: tests/shaders/ssa/fs-while-loop-rotate-value.shader_test
(no regressions in the shader set). It also fixes all failing tests from

  dEQP-GLES3.functional.shaders.loops.*

Signed-off-by: Gert Wollny <gw.fossdev@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
src/gallium/drivers/r600/sb/sb_ra_init.cpp