nir/algebraic: Recognize open-coded flrp(a, b, a)

No shader-db changes Ice Lake, Iron Lake, or GM45 as these platforms
lack a LRP instruction.

v2: Remove flrp@64 cases.  Since Gen11 removes flrp@32, it seems
unlikely that we'll ever have a flrp@64.  Should that occur, the cases
can be added back.

All Gen6-Gen9 platforms had similar results. (Skylake shown)
total instructions in shared programs: 15041996 -> 15041184 (<.01%)
instructions in affected programs: 71776 -> 70964 (-1.13%)
helped: 312
HURT: 0
helped stats (abs) min: 2 max: 3 x̄: 2.60 x̃: 3
helped stats (rel) min: 0.36% max: 4.55% x̄: 1.75% x̃: 1.28%
95% mean confidence interval for instructions value: -2.66 -2.55
95% mean confidence interval for instructions %-change: -1.89% -1.61%
Instructions are helped.

total cycles in shared programs: 354303333 -> 354301807 (<.01%)
cycles in affected programs: 433742 -> 432216 (-0.35%)
helped: 206
HURT: 78
helped stats (abs) min: 2 max: 244 x̄: 21.02 x̃: 8
helped stats (rel) min: 0.06% max: 19.59% x̄: 1.72% x̃: 0.82%
HURT stats (abs)   min: 1 max: 220 x̄: 35.95 x̃: 10
HURT stats (rel)   min: 0.07% max: 30.48% x̄: 2.53% x̃: 0.56%
95% mean confidence interval for cycles value: -10.68 -0.06
95% mean confidence interval for cycles %-change: -0.99% -0.12%
Cycles are helped.

Reviewed-by: Matt Turner <mattst88@gmail.com>

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 8fcf0a1cbffe697765331a6ca20925229a597892..6ae937651e0a1ca9d64f56c18b197570671d8857 100644 (file)
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -1255,6 +1255,30 @@ late_optimizations = [
    (('~fadd@32', ('fmul(is_used_once)', -2.0, a),  1.0), ('flrp',  1.0, -1.0,          a ), '!options->lower_flrp32'),
    (('~fadd@32', ('fmul(is_used_once)',  2.0, a),  1.0), ('flrp',  1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
 
+    # flrp(a, b, a)
+    # a*(1-a) + b*a
+    # a + -a*a + a*b    (1)
+    # a + a*(b - a)
+    # Option 1: ffma(a, (b-a), a)
+    #
+    # Alternately, after (1):
+    # a*(1+b) + -a*a
+    # a*((1+b) + -a)
+    #
+    # Let b=1
+    #
+    # Option 2: ffma(a, 2, -(a*a))
+    # Option 3: ffma(a, 2, (-a)*a)
+    # Option 4: ffma(a, -a, (2*a)
+    # Option 5: a * (2 - a)
+    #
+    # There are a lot of other possible combinations.
+   (('~ffma@32', ('fadd', b, ('fneg', a)), a, a), ('flrp', a, b, a), '!options->lower_flrp32'),
+   (('~ffma@32', a, 2.0, ('fneg', ('fmul', a, a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
+   (('~ffma@32', a, 2.0, ('fmul', ('fneg', a), a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
+   (('~ffma@32', a, ('fneg', a), ('fmul', 2.0, a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
+   (('~fmul@32', a, ('fadd', 2.0, ('fneg', a))),    ('flrp', a, 1.0, a), '!options->lower_flrp32'),
+
    # we do these late so that we don't get in the way of creating ffmas
    (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
    (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),