nir/algebraic: Optimize bfi with odd-valued mask to bitfield_select

shader-db:

Lunar Lake, Meteor Lake, and DG2 had similar results. (Lunar Lake shown)
total instructions in shared programs: 17181254 -> 17181046 (<.01%)
instructions in affected programs: 35834 -> 35626 (-0.58%)
helped: 130 / HURT: 2

total cycles in shared programs: 888543370 -> 888554248 (<.01%)
cycles in affected programs: 7443984 -> 7454862 (0.15%)
helped: 95 / HURT: 87

fossil-db:

Lunar Lake
Totals:
Instrs: 233260196 -> 233259474 (-0.00%); split: -0.00%, +0.00%
Cycle count: 32754567116 -> 32754515890 (-0.00%); split: -0.00%, +0.00%
Max live registers: 71738442 -> 71738398 (-0.00%); split: -0.00%, +0.00%

Totals from 6842 (0.87% of 790721) affected shaders:
Instrs: 5566926 -> 5566204 (-0.01%); split: -0.01%, +0.00%
Cycle count: 512487046 -> 512435820 (-0.01%); split: -0.20%, +0.19%
Max live registers: 1100656 -> 1100612 (-0.00%); split: -0.00%, +0.00%

Meteor Lake and DG2 had similar results. (Meteor Lake shown)
Totals:
Instrs: 264071212 -> 264066944 (-0.00%); split: -0.00%, +0.00%
Cycle count: 26552458051 -> 26553286277 (+0.00%); split: -0.00%, +0.01%
Spill count: 530380 -> 530084 (-0.06%)
Fill count: 613416 -> 612900 (-0.08%)
Scratch Memory Size: 20089856 -> 20075520 (-0.07%)
Max live registers: 46558852 -> 46558811 (-0.00%); split: -0.00%, +0.00%
Max dispatch width: 8034616 -> 8034584 (-0.00%)

Totals from 6653 (0.73% of 905545) affected shaders:
Instrs: 5750844 -> 5746576 (-0.07%); split: -0.08%, +0.00%
Cycle count: 416414845 -> 417243071 (+0.20%); split: -0.20%, +0.40%
Spill count: 1953 -> 1657 (-15.16%)
Fill count: 3556 -> 3040 (-14.51%)
Scratch Memory Size: 92160 -> 77824 (-15.56%)
Max live registers: 566003 -> 565962 (-0.01%); split: -0.01%, +0.00%
Max dispatch width: 55768 -> 55736 (-0.06%)

No shader-db or fossil-db changes on any previous Intel platforms.

Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37186>

src/compiler/nir/nir_opt_algebraic.py

@@ -1771,6 +1771,8 @@ optimizations.extend([
    #              = (b & a)               If a is odd, find_lsb(a) == 0
    (('bfi', '#a(is_odd)', b, 0), ('iand', a, b)),
 
+   (('bfi', '#a(is_odd)', b, c), ('bitfield_select', a, b, c), 'options->has_bitfield_select'),
+
    # Because 'a' is a positive power of two, the result of the bfi is either 0
    # or 'a' depending on whether or not 'b' is odd.  Use 'b&1' for the zero
    # value to help platforms that can't have two constants in a bcsel.