[Mesa-dev] Target-aware nir_opt_algebraic.

[Eric Anholt]

I'm pretty happy with how it turned out.  What I'm not sure on is that
I've got a couple of places where I replicated another transformation,
like fsat(fsat(a)) -> fsat(a) but with a lowered fsat.  I could see
potentially having only the canonical form appear in nir_opt_algebraic,
and I'd have two optimization loops in my driver: one in canonical form,
and one in the driver's lowered form, with different
nir_shader_compiler_options in the two loops.  I worry a bit, that down
this road lies having at least 3 loops (canonical, then lowered-to-driver
uniform indexing math, then lowered-to-driver operations), and then maybe
I just don't have the creativity right now to see when we're going to need
4 loops, and this seems nuts.

(Note: my goal with a bunch of this series is to rely less on my driver's
internal optimization code, since I'm going to have to more or less throw
it out if I go to doing control flow)