[Mesa-dev] nir-to-tgsi review?

[Eric Anholt]

I'd love to get some review on
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3395

I think it's time to make the switch for softpipe.  It improves
performance significantly (~10%) and makes it pass far more piglit
tests than glsl_to_tgsi does.  There are 5 regressions in doubles, but
given how broken doubles were before this I think we don't need to get
every last doubles regression fixed.  There's also one gles31
regression in GSes, but I think that's just an instance of "softpipe
GS handling is broken across the board" as one can see with the
deqp-softpipe-skips.txt list

I also have followup work waiting for this MR -- using it to replace
mesa_to_tgsi (-1200 lines), making i915g use it (letting that driver
compile far more programs than it used to), replacing ATIFS to TGSI
(cutting code and introducing optimization to this ancient shader
path), and eventually pushing it into other TGSI-consuming drivers so
we can delete glsl_to_tgsi (-10kloc) and start cleaning up the program
compile path in mesa/st (my original motivation here).