[Mesa-dev] [PATCH] R600/SI: Add patterns for V_MAD(_LEGACY) instructions.

[Michel Dänzer]

On Die, 2013-01-22 at 16:10 +0100, Tom Stellard wrote: 
> On Tue, Jan 22, 2013 at 04:06:05PM +0100, Michel Dänzer wrote:
> > On Mon, 2013-01-21 at 23:03 +0100, Tom Stellard wrote: 
> > > 
> > > I don't think we emit the int_AMDGPU_mul intrinsic anymore, but it probably
> > > doesn't hurt to keep it around until we sort out all of the legacy vs non-legacy
> > > instruction issues.
> > 
> > Right. BTW, emitting int_AMDGPU_mad for TGSI_OPCODE_MAD doesn't fix the
> > glxgears problem I reported on IRC with current master (in contrast to
> > 0ad1fefd6951aa47ab58a41dc9ee73083cbcf85c, where that problem was
> > introduced). Looks like the intrinsic ends up generating cmp/select
> > pairs instead of V_MAD in some cases now, which breaks somehow.
> > 
> > 
> > In the meantime, I've come to realize that this patch breaks a bunch of
> > piglit tests in the glsl-1.[12]0/execution/variable-indexing groups (I
> > previously thought those tests were broken by the array uniform related
> > changes). I'm attaching a diff of the generated code. Not sure if this
> > is some kind of bad interaction between this change and the control flow
> > code, or another instance of tests which passed accidentally.
> > 
> 
> Do these tests use indirect addressing?  If so, they were probably passing
> by chance.

tgsi_shader_info::indirect_files is 0 for all shaders. So I think this
patch shouldn't go in without more investigation why it breaks these
tests.


-- 
Earthling Michel Dänzer           |                   http://www.amd.com
Libre software enthusiast         |          Debian, X and DRI developer