elk: Use nir_opt_sink and more nir_opt_move

I spent a bunch of time playing around with the various enable bits, and
this was the best I could come up with. Enabling any of
nir_move_comparisons or nir_move_load_ubo in nir_opt_sink helped
instructions quite a bit, but it also caused a large pile of added
spills and fills.

shader-db:

Broadwell
total instructions in shared programs: 18428980 -> 18427957 (<.01%)
instructions in affected programs: 425245 -> 424222 (-0.24%)
helped: 1522 / HURT: 405

total cycles in shared programs: 954756705 -> 953755695 (-0.10%)
cycles in affected programs: 623470486 -> 622469476 (-0.16%)
helped: 17989 / HURT: 21175

total spills in shared programs: 8349 -> 8356 (0.08%)
spills in affected programs: 285 -> 292 (2.46%)
helped: 7 / HURT: 13

total fills in shared programs: 10426 -> 10192 (-2.24%)
fills in affected programs: 675 -> 441 (-34.67%)
helped: 25 / HURT: 1

LOST:   346
GAINED: 554

Haswell
total instructions in shared programs: 16809730 -> 16801634 (-0.05%)
instructions in affected programs: 772251 -> 764155 (-1.05%)
helped: 3055 / HURT: 840

total cycles in shared programs: 945179935 -> 944315696 (-0.09%)
cycles in affected programs: 549177588 -> 548313349 (-0.16%)
helped: 34143 / HURT: 23605

total spills in shared programs: 7699 -> 7666 (-0.43%)
spills in affected programs: 353 -> 320 (-9.35%)
helped: 10 / HURT: 16

total fills in shared programs: 8184 -> 7671 (-6.27%)
fills in affected programs: 1006 -> 493 (-50.99%)
helped: 30 / HURT: 2

total sends in shared programs: 1016676 -> 1016682 (<.01%)
sends in affected programs: 49 -> 55 (12.24%)
helped: 0 / HURT: 6

LOST:   415
GAINED: 441

Ivy Bridge
total instructions in shared programs: 15764955 -> 15757178 (-0.05%)
instructions in affected programs: 707453 -> 699676 (-1.10%)
helped: 2893 / HURT: 547

total cycles in shared programs: 430017934 -> 429720104 (-0.07%)
cycles in affected programs: 251816726 -> 251518896 (-0.12%)
helped: 33110 / HURT: 22056

total spills in shared programs: 1537 -> 1525 (-0.78%)
spills in affected programs: 18 -> 6 (-66.67%)
helped: 6 / HURT: 0

total fills in shared programs: 926 -> 905 (-2.27%)
fills in affected programs: 24 -> 3 (-87.50%)
helped: 6 / HURT: 0

total sends in shared programs: 816646 -> 816652 (<.01%)
sends in affected programs: 49 -> 55 (12.24%)
helped: 0 / HURT: 6

LOST:   332
GAINED: 417

Sandy Bridge
total instructions in shared programs: 14055229 -> 14045281 (-0.07%)
instructions in affected programs: 1436142 -> 1426194 (-0.69%)
helped: 5858 / HURT: 757

total cycles in shared programs: 772123170 -> 813543451 (5.36%)
cycles in affected programs: 521342483 -> 562762764 (7.94%)
helped: 27928 / HURT: 35923

total spills in shared programs: 1742 -> 1741 (-0.06%)
spills in affected programs: 66 -> 65 (-1.52%)
helped: 1 / HURT: 0

total fills in shared programs: 970 -> 967 (-0.31%)
fills in affected programs: 93 -> 90 (-3.23%)
helped: 1 / HURT: 0

total sends in shared programs: 1239222 -> 1238992 (-0.02%)
sends in affected programs: 6137 -> 5907 (-3.75%)
helped: 342 / HURT: 112

LOST:   244
GAINED: 434

Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 8366385 -> 8363954 (-0.03%)
instructions in affected programs: 162761 -> 160330 (-1.49%)
helped: 600 / HURT: 195

total cycles in shared programs: 248992618 -> 252119334 (1.26%)
cycles in affected programs: 50774708 -> 53901424 (6.16%)
helped: 3435 / HURT: 5131

total sends in shared programs: 623693 -> 623681 (<.01%)
sends in affected programs: 351 -> 339 (-3.42%)
helped: 12 / HURT: 0

LOST: 0
GAINED: 6

Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25463>

src/intel/compiler/elk/elk_nir.c

@@ -1565,7 +1565,11 @@ elk_postprocess_nir(nir_shader *nir, const struct elk_compiler *compiler,
 
    OPT(nir_copy_prop);
    OPT(nir_opt_dce);
-   OPT(nir_opt_move, nir_move_comparisons);
+
+   nir_move_options common = nir_move_const_undef | nir_move_load_input |
+                             nir_move_copies | nir_move_load_ssbo;
+   OPT(nir_opt_sink, common);
+   OPT(nir_opt_move, common | nir_move_comparisons | nir_move_load_ubo);
    OPT(nir_opt_dead_cf);
 
    /* TODO: Enable nir_opt_uniform_atomics on Gfx7.x too.