nir/opt_algebraic: add patterns for iand/ior of feq/fneu with 0

Foz-DB Navi21:
Totals from 1245 (0.92% of 134913) affected shaders:
VGPRs: 66232 -> 66248 (+0.02%); split: -0.01%, +0.04%
CodeSize: 5874976 -> 5868168 (-0.12%); split: -0.17%, +0.05%
MaxWaves: 25278 -> 25274 (-0.02%); split: +0.01%, -0.02%
Instrs: 1087502 -> 1085267 (-0.21%); split: -0.21%, +0.00%
Latency: 6531489 -> 6531672 (+0.00%); split: -0.04%, +0.05%
InvThroughput: 1531774 -> 1532327 (+0.04%); split: -0.02%, +0.05%
VClause: 22218 -> 22202 (-0.07%); split: -0.08%, +0.00%
SClause: 45906 -> 45873 (-0.07%); split: -0.08%, +0.01%
Copies: 64004 -> 64102 (+0.15%); split: -0.24%, +0.39%
Branches: 21529 -> 21534 (+0.02%); split: -0.00%, +0.03%
PreSGPRs: 51936 -> 51850 (-0.17%)
PreVGPRs: 55393 -> 55398 (+0.01%); split: -0.02%, +0.03%

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21576>

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 9d2fbb5..60957c8 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -931,6 +931,11 @@ for s in [16, 32, 64]:
        (('~iand', ('fge(is_used_once)', 0.0, 'a@{}'.format(s)), ('fge', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
        (('~iand', ('fge', 0.0, 'a@{}'.format(s)), ('fge(is_used_once)', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
 
+       (('ior', ('feq(is_used_once)', 'a@{}'.format(s), 0.0), ('feq', 'b@{}'.format(s), 0.0)), ('feq', ('fmin', ('fabs', a), ('fabs', b)), 0.0)),
+       (('ior', ('fneu(is_used_once)', 'a@{}'.format(s), 0.0), ('fneu', 'b@{}'.format(s), 0.0)), ('fneu', ('fadd', ('fabs', a), ('fabs', b)), 0.0)),
+       (('iand', ('feq(is_used_once)', 'a@{}'.format(s), 0.0), ('feq', 'b@{}'.format(s), 0.0)), ('feq', ('fadd', ('fabs', a), ('fabs', b)), 0.0)),
+       (('iand', ('fneu(is_used_once)', 'a@{}'.format(s), 0.0), ('fneu', 'b@{}'.format(s), 0.0)), ('fneu', ('fmin', ('fabs', a), ('fabs', b)), 0.0)),
+
        # The (i2f32, ...) part is an open-coded fsign.  When that is combined
        # with the bcsel, it's basically copysign(1.0, a).  There are some
        # behavior differences between this pattern and copysign w.r.t. ±0 and