Mesa (master): nir/algebraic: optimize out exact a+0.0 if it's used only as a float

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 614ab26afd05dbb9a99ef6a3cde75ba55279ed32
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=614ab26afd05dbb9a99ef6a3cde75ba55279ed32

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Wed Jul 15 11:16:33 2020 +0100

nir/algebraic: optimize out exact a+0.0 if it's used only as a float

fossil-db (GFX10):
Totals from 133 (0.10% of 139391) affected shaders:
SGPRs: 7864 -> 7856 (-0.10%); split: -0.20%, +0.10%
VGPRs: 4884 -> 4836 (-0.98%)
CodeSize: 288932 -> 287084 (-0.64%)
MaxWaves: 1973 -> 1979 (+0.30%)
Instrs: 53899 -> 53550 (-0.65%)

fossil-db (GFX10.3):
Totals from 133 (0.10% of 139391) affected shaders:
SGPRs: 7832 -> 7835 (+0.04%); split: -0.06%, +0.10%
VGPRs: 5144 -> 5088 (-1.09%)
CodeSize: 318912 -> 316696 (-0.69%)
MaxWaves: 1735 -> 1746 (+0.63%)
Instrs: 65367 -> 64853 (-0.79%)

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Daniel Schürmann <daniel at schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5523>

---

 src/compiler/nir/nir_opt_algebraic.py | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index a506160c537..73c0b3767cf 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -40,6 +40,9 @@ c = 'c'
 d = 'd'
 e = 'e'
 
+signed_zero_inf_nan_preserve_16 = 'nir_is_float_control_signed_zero_inf_nan_preserve(info->float_controls_execution_mode, 16)'
+signed_zero_inf_nan_preserve_32 = 'nir_is_float_control_signed_zero_inf_nan_preserve(info->float_controls_execution_mode, 32)'
+
 # Written in the form (<search>, <replace>) where <search> is an expression
 # and <replace> is either an expression or a value.  An expression is
 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
@@ -124,6 +127,11 @@ optimizations = [
    (('f2b', ('fneg', a)), ('f2b', a)),
    (('i2b', ('ineg', a)), ('i2b', a)),
    (('~fadd', a, 0.0), a),
+   # a+0.0 is 'a' unless 'a' is denormal or -0.0. If it's only used by a
+   # floating point instruction, they should flush any input denormals and we
+   # can replace -0.0 with 0.0 if the float execution mode allows it.
+   (('fadd(is_only_used_as_float)', 'a at 16', 0.0), a, '!'+signed_zero_inf_nan_preserve_16),
+   (('fadd(is_only_used_as_float)', 'a at 32', 0.0), a, '!'+signed_zero_inf_nan_preserve_32),
    (('iadd', a, 0), a),
    (('usadd_4x8', a, 0), a),
    (('usadd_4x8', a, ~0), ~0),
@@ -158,6 +166,8 @@ optimizations = [
    (('fmul', ('fmul', ('fsign', a), a), a), ('fmul', ('fabs', a), a)),
    (('~ffma', 0.0, a, b), b),
    (('~ffma', a, b, 0.0), ('fmul', a, b)),
+   (('ffma at 16', a, b, 0.0), ('fmul', a, b), '!'+signed_zero_inf_nan_preserve_16),
+   (('ffma at 32', a, b, 0.0), ('fmul', a, b), '!'+signed_zero_inf_nan_preserve_32),
    (('ffma', 1.0, a, b), ('fadd', a, b)),
    (('ffma', -1.0, a, b), ('fadd', ('fneg', a), b)),
    (('~flrp', a, b, 0.0), a),