From 408330ed48e27bbc1ab7d7aefcdd2f3951ac58ca Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Tue, 3 Jul 2018 12:39:54 -0700 Subject: [PATCH] nir: Don't compare i2f or u2i with zero MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Broadwell and Skylake had similar results. (Skylake shown) total instructions in shared programs: 14277620 -> 14277230 (<.01%) instructions in affected programs: 36905 -> 36515 (-1.06%) helped: 101 HURT: 6 helped stats (abs) min: 1 max: 6 x̄: 4.46 x̃: 6 helped stats (rel) min: 0.32% max: 7.69% x̄: 1.80% x̃: 1.51% HURT stats (abs) min: 1 max: 28 x̄: 10.00 x̃: 1 HURT stats (rel) min: 0.33% max: 1.74% x̄: 0.68% x̃: 0.47% 95% mean confidence interval for instructions value: -4.59 -2.70 95% mean confidence interval for instructions %-change: -1.90% -1.41% Instructions are helped. total cycles in shared programs: 532580716 -> 532577947 (<.01%) cycles in affected programs: 940575 -> 937806 (-0.29%) helped: 92 HURT: 12 helped stats (abs) min: 2 max: 158 x̄: 51.04 x̃: 62 helped stats (rel) min: 0.24% max: 3.99% x̄: 2.14% x̃: 2.41% HURT stats (abs) min: 10 max: 1112 x̄: 160.58 x̃: 63 HURT stats (rel) min: 0.06% max: 21.90% x̄: 4.22% x̃: 0.20% 95% mean confidence interval for cycles value: -50.66 -2.59 95% mean confidence interval for cycles %-change: -2.09% -0.73% Cycles are helped. total spills in shared programs: 8116 -> 8124 (0.10%) spills in affected programs: 200 -> 208 (4.00%) helped: 0 HURT: 2 total fills in shared programs: 11086 -> 11094 (0.07%) fills in affected programs: 436 -> 444 (1.83%) helped: 0 HURT: 2 Ivy Bridge and Haswell had similar results. (Haswell shown) total instructions in shared programs: 12979054 -> 12978067 (<.01%) instructions in affected programs: 33633 -> 32646 (-2.93%) helped: 120 HURT: 2 helped stats (abs) min: 1 max: 13 x̄: 8.53 x̃: 13 helped stats (rel) min: 0.30% max: 16.67% x̄: 4.55% x̃: 3.17% HURT stats (abs) min: 18 max: 18 x̄: 18.00 x̃: 18 HURT stats (rel) min: 1.15% max: 2.84% x̄: 2.00% x̃: 2.00% 95% mean confidence interval for instructions value: -9.19 -6.99 95% mean confidence interval for instructions %-change: -5.27% -3.62% Instructions are helped. total cycles in shared programs: 411212880 -> 411199636 (<.01%) cycles in affected programs: 696441 -> 683197 (-1.90%) helped: 107 HURT: 5 helped stats (abs) min: 2 max: 864 x̄: 124.90 x̃: 146 helped stats (rel) min: 0.03% max: 29.20% x̄: 8.58% x̃: 5.88% HURT stats (abs) min: 2 max: 50 x̄: 24.00 x̃: 22 HURT stats (rel) min: 0.01% max: 5.35% x̄: 1.29% x̃: 0.25% 95% mean confidence interval for cycles value: -136.96 -99.54 95% mean confidence interval for cycles %-change: -9.75% -6.53% Cycles are helped. total spills in shared programs: 78623 -> 78631 (0.01%) spills in affected programs: 66 -> 74 (12.12%) helped: 0 HURT: 2 total fills in shared programs: 80104 -> 80108 (<.01%) fills in affected programs: 133 -> 137 (3.01%) helped: 0 HURT: 2 No changes on Sandy Bridge, Iron Lake, or GM45. Signed-off-by: Ian Romanick Reviewed-by: Thomas Helland --- src/compiler/nir/nir_opt_algebraic.py | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py index 9aeafcb8c46..fdd1af9d177 100644 --- a/src/compiler/nir/nir_opt_algebraic.py +++ b/src/compiler/nir/nir_opt_algebraic.py @@ -193,6 +193,19 @@ optimizations = [ (('ieq', ('b2i', a), 0), ('inot', a)), (('ine', ('b2i', a), 0), a), + (('fne', ('u2f32', a), 0.0), ('ine', a, 0)), + (('feq', ('u2f32', a), 0.0), ('ieq', a, 0)), + (('fge', ('u2f32', a), 0.0), True), + (('fge', 0.0, ('u2f32', a)), ('uge', 0, a)), # ieq instead? + (('flt', ('u2f32', a), 0.0), False), + (('flt', 0.0, ('u2f32', a)), ('ult', 0, a)), # ine instead? + (('fne', ('i2f32', a), 0.0), ('ine', a, 0)), + (('feq', ('i2f32', a), 0.0), ('ieq', a, 0)), + (('fge', ('i2f32', a), 0.0), ('ige', a, 0)), + (('fge', 0.0, ('i2f32', a)), ('ige', 0, a)), + (('flt', ('i2f32', a), 0.0), ('ilt', a, 0)), + (('flt', 0.0, ('i2f32', a)), ('ilt', 0, a)), + # 0.0 < fabs(a) # fabs(a) > 0.0 # fabs(a) != 0.0 because fabs(a) must be >= 0 -- 2.30.2