A little analysis shows that the worst-case value for "nr" is 17:
- base_mrf = 2 ... 2
- header present (say gen == 5) ... 4
- aa_dest_stencil_reg (stencil test) ... 5
- SIMD16 mode: += 4 * reg_width ... 13
- source_depth_to_render_target ... 15
- dest_depth_reg ... 17
This resulted in us setting base_mrf to 2 and mlen to 15. In other
words, we'd try to use m2..m16. But m16 doesn't exist pre-Gen6. Also,
the instruction scheduler data structures use arrays of size 16, so this
would cause us to access them out of bounds.
While the debugger system routine may need m0 and m1, we don't use it
today, so the simplest solution is just to move base_mrf back to 1.
That way, our worst case message fits in m1..m15, which is legal.
An alternative would be to fail on SIMD16 in this case, but that seems
a bit unfortunate if there's no real need to reserve m0 and m1.
Fixes new piglit test shaders/depth-test-and-write on Ironlake.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=48218
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
{
this->current_annotation = "FB write header";
bool header_present = true;
- int base_mrf = 2;
+ /* We can potentially have a message length of up to 15, so we have to set
+ * base_mrf to either 0 or 1 in order to fit in m0..m15.
+ */
+ int base_mrf = 1;
int nr = base_mrf;
int reg_width = c->dispatch_width / 8;