1# -*- coding: utf-8 -*- 2# 3# Copyright (C) 2014 Intel Corporation 4# 5# Permission is hereby granted, free of charge, to any person obtaining a 6# copy of this software and associated documentation files (the "Software"), 7# to deal in the Software without restriction, including without limitation 8# the rights to use, copy, modify, merge, publish, distribute, sublicense, 9# and/or sell copies of the Software, and to permit persons to whom the 10# Software is furnished to do so, subject to the following conditions: 11# 12# The above copyright notice and this permission notice (including the next 13# paragraph) shall be included in all copies or substantial portions of the 14# Software. 15# 16# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 21# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 22# IN THE SOFTWARE. 23# 24# Authors: 25# Jason Ekstrand (jason@jlekstrand.net) 26 27from __future__ import print_function 28 29from collections import OrderedDict 30import nir_algebraic 31from nir_opcodes import type_sizes 32import itertools 33import struct 34from math import pi 35 36# Convenience variables 37a = 'a' 38b = 'b' 39c = 'c' 40d = 'd' 41e = 'e' 42 43# Written in the form (<search>, <replace>) where <search> is an expression 44# and <replace> is either an expression or a value. An expression is 45# defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>) 46# where each source is either an expression or a value. A value can be 47# either a numeric constant or a string representing a variable name. 48# 49# If the opcode in a search expression is prefixed by a '~' character, this 50# indicates that the operation is inexact. Such operations will only get 51# applied to SSA values that do not have the exact bit set. This should be 52# used by by any optimizations that are not bit-for-bit exact. It should not, 53# however, be used for backend-requested lowering operations as those need to 54# happen regardless of precision. 55# 56# Variable names are specified as "[#]name[@type][(cond)][.swiz]" where: 57# "#" indicates that the given variable will only match constants, 58# type indicates that the given variable will only match values from ALU 59# instructions with the given output type, 60# (cond) specifies an additional condition function (see nir_search_helpers.h), 61# swiz is a swizzle applied to the variable (only in the <replace> expression) 62# 63# For constants, you have to be careful to make sure that it is the right 64# type because python is unaware of the source and destination types of the 65# opcodes. 66# 67# All expression types can have a bit-size specified. For opcodes, this 68# looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a 69# type and size. In the search half of the expression this indicates that it 70# should only match that particular bit-size. In the replace half of the 71# expression this indicates that the constructed value should have that 72# bit-size. 73# 74# If the opcode in a replacement expression is prefixed by a '!' character, 75# this indicated that the new expression will be marked exact. 76# 77# A special condition "many-comm-expr" can be used with expressions to note 78# that the expression and its subexpressions have more commutative expressions 79# than nir_replace_instr can handle. If this special condition is needed with 80# another condition, the two can be separated by a comma (e.g., 81# "(many-comm-expr,is_used_once)"). 82 83# based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648 84def lowered_sincos(c): 85 x = ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi, a), c))), 1.0) 86 x = ('fmul', ('fsub', x, ('fmul', x, ('fabs', x))), 4.0) 87 return ('ffma', ('ffma', x, ('fabs', x), ('fneg', x)), 0.225, x) 88 89def intBitsToFloat(i): 90 return struct.unpack('!f', struct.pack('!I', i))[0] 91 92optimizations = [ 93 94 (('imul', a, '#b(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'), 95 (('imul', a, '#b(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'), 96 (('ishl', a, '#b'), ('imul', a, ('ishl', 1, b)), 'options->lower_bitops'), 97 98 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a, b)), ('imul', a, b)), 99 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a, b)), ('imul', a, b)), 100 (('imul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('imul_high', a, b)), 'options->lower_mul_2x32_64'), 101 (('umul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('umul_high', a, b)), 'options->lower_mul_2x32_64'), 102 (('udiv', a, 1), a), 103 (('idiv', a, 1), a), 104 (('umod', a, 1), 0), 105 (('imod', a, 1), 0), 106 (('udiv', a, '#b(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b)), '!options->lower_bitops'), 107 (('idiv', a, '#b(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), '!options->lower_bitops'), 108 (('idiv', a, '#b(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', ('iabs', b))))), '!options->lower_bitops'), 109 (('umod', a, '#b(is_pos_power_of_two)'), ('iand', a, ('isub', b, 1))), 110 111 (('~fneg', ('fneg', a)), a), 112 (('ineg', ('ineg', a)), a), 113 (('fabs', ('fneg', a)), ('fabs', a)), 114 (('fabs', ('u2f', a)), ('u2f', a)), 115 (('iabs', ('iabs', a)), ('iabs', a)), 116 (('iabs', ('ineg', a)), ('iabs', a)), 117 (('f2b', ('fneg', a)), ('f2b', a)), 118 (('i2b', ('ineg', a)), ('i2b', a)), 119 (('~fadd', a, 0.0), a), 120 (('iadd', a, 0), a), 121 (('usadd_4x8', a, 0), a), 122 (('usadd_4x8', a, ~0), ~0), 123 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))), 124 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))), 125 (('iand', ('ior', a, b), ('ior', a, c)), ('ior', a, ('iand', b, c))), 126 (('ior', ('iand', a, b), ('iand', a, c)), ('iand', a, ('ior', b, c))), 127 (('~fadd', ('fneg', a), a), 0.0), 128 (('iadd', ('ineg', a), a), 0), 129 (('iadd', ('ineg', a), ('iadd', a, b)), b), 130 (('iadd', a, ('iadd', ('ineg', a), b)), b), 131 (('~fadd', ('fneg', a), ('fadd', a, b)), b), 132 (('~fadd', a, ('fadd', ('fneg', a), b)), b), 133 (('fadd', ('fsat', a), ('fsat', ('fneg', a))), ('fsat', ('fabs', a))), 134 (('~fmul', a, 0.0), 0.0), 135 (('imul', a, 0), 0), 136 (('umul_unorm_4x8', a, 0), 0), 137 (('umul_unorm_4x8', a, ~0), a), 138 (('~fmul', a, 1.0), a), 139 (('imul', a, 1), a), 140 (('fmul', a, -1.0), ('fneg', a)), 141 (('imul', a, -1), ('ineg', a)), 142 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a 143 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a 144 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0 145 (('fmul', ('fsign', a), ('fmul', a, a)), ('fmul', ('fabs', a), a)), 146 (('fmul', ('fmul', ('fsign', a), a), a), ('fmul', ('fabs', a), a)), 147 (('~ffma', 0.0, a, b), b), 148 (('~ffma', a, b, 0.0), ('fmul', a, b)), 149 (('ffma', 1.0, a, b), ('fadd', a, b)), 150 (('ffma', -1.0, a, b), ('fadd', ('fneg', a), b)), 151 (('~flrp', a, b, 0.0), a), 152 (('~flrp', a, b, 1.0), b), 153 (('~flrp', a, a, b), a), 154 (('~flrp', 0.0, a, b), ('fmul', a, b)), 155 156 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c) 157 (('~flrp', a, ('fadd(is_used_once)', a, b), c), ('fadd', ('fmul', b, c), a)), 158] 159 160# Float sizes 161for s in [16, 32, 64]: 162 optimizations.extend([ 163 (('~flrp@{}'.format(s), a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp{}'.format(s)), 164 (('~flrp@{}'.format(s), ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp{}'.format(s)), 165 (('~flrp@{}'.format(s), a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp{}'.format(s)), 166 167 (('~flrp@{}'.format(s), a, b, ('b2f', 'c@1')), ('bcsel', c, b, a), 'options->lower_flrp{}'.format(s)), 168 169 (('~fadd@{}'.format(s), ('fmul', a, ('fadd', 1.0, ('fneg', c))), ('fmul', b, c)), ('flrp', a, b, c), '!options->lower_flrp{}'.format(s)), 170 # These are the same as the previous three rules, but it depends on 171 # 1-fsat(x) <=> fsat(1-x). See below. 172 (('~fadd@{}'.format(s), ('fmul', a, ('fsat', ('fadd', 1.0, ('fneg', c)))), ('fmul', b, ('fsat', c))), ('flrp', a, b, ('fsat', c)), '!options->lower_flrp{}'.format(s)), 173 (('~fadd@{}'.format(s), a, ('fmul', c, ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp{}'.format(s)), 174 175 (('~fadd@{}'.format(s), ('fmul', a, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b, ('b2f', c))), ('bcsel', c, b, a), 'options->lower_flrp{}'.format(s)), 176 (('~fadd@{}'.format(s), a, ('fmul', ('b2f', 'c@1'), ('fadd', b, ('fneg', a)))), ('bcsel', c, b, a), 'options->lower_flrp{}'.format(s)), 177 178 # 1 - ((1 - a) * (1 - b)) 179 # 1 - (1 - a - b + a*b) 180 # 1 - 1 + a + b - a*b 181 # a + b - a*b 182 # a + b*(1 - a) 183 # b*(1 - a) + 1*a 184 # flrp(b, 1, a) 185 (('~fadd@{}'.format(s), 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a)), ('fadd', 1.0, ('fneg', b))))), ('flrp', b, 1.0, a), '!options->lower_flrp{}'.format(s)), 186 ]) 187 188optimizations.extend([ 189 (('~flrp', ('fmul(is_used_once)', a, b), ('fmul(is_used_once)', a, c), d), ('fmul', ('flrp', b, c, d), a)), 190 191 (('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)), 192 (('ftrunc', a), ('bcsel', ('flt', a, 0.0), ('fneg', ('ffloor', ('fabs', a))), ('ffloor', ('fabs', a))), 'options->lower_ftrunc'), 193 (('ffloor', a), ('fsub', a, ('ffract', a)), 'options->lower_ffloor'), 194 (('fadd', a, ('fneg', ('ffract', a))), ('ffloor', a), '!options->lower_ffloor'), 195 (('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'), 196 (('fceil', a), ('fneg', ('ffloor', ('fneg', a))), 'options->lower_fceil'), 197 (('ffma@16', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma16'), 198 (('ffma@32', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma32'), 199 (('ffma@64', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma64'), 200 # Always lower inexact ffma, because it will be fused back by late optimizations (nir_opt_algebraic_late). 201 (('~ffma@16', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma16'), 202 (('~ffma@32', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma32'), 203 (('~ffma@64', a, b, c), ('fadd', ('fmul', a, b), c), 'options->fuse_ffma64'), 204 205 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i', 'a@bool')), ('fmul', b, c)), '#d'), '#e'), 206 ('bcsel', a, ('fmul', ('fadd', ('fmul', b, c), d), e), ('fmul', d, e))), 207 208 (('fdph', a, b), ('fdot4', ('vec4', 'a.x', 'a.y', 'a.z', 1.0), b), 'options->lower_fdph'), 209 210 (('fdot4', ('vec4', a, b, c, 1.0), d), ('fdph', ('vec3', a, b, c), d), '!options->lower_fdph'), 211 (('fdot4', ('vec4', a, 0.0, 0.0, 0.0), b), ('fmul', a, b)), 212 (('fdot4', ('vec4', a, b, 0.0, 0.0), c), ('fdot2', ('vec2', a, b), c)), 213 (('fdot4', ('vec4', a, b, c, 0.0), d), ('fdot3', ('vec3', a, b, c), d)), 214 215 (('fdot3', ('vec3', a, 0.0, 0.0), b), ('fmul', a, b)), 216 (('fdot3', ('vec3', a, b, 0.0), c), ('fdot2', ('vec2', a, b), c)), 217 218 (('fdot2', ('vec2', a, 0.0), b), ('fmul', a, b)), 219 (('fdot2', a, 1.0), ('fadd', 'a.x', 'a.y')), 220 221 # Lower fdot to fsum when it is available 222 (('fdot2', a, b), ('fsum2', ('fmul', a, b)), 'options->lower_fdot'), 223 (('fdot3', a, b), ('fsum3', ('fmul', a, b)), 'options->lower_fdot'), 224 (('fdot4', a, b), ('fsum4', ('fmul', a, b)), 'options->lower_fdot'), 225 (('fsum2', a), ('fadd', 'a.x', 'a.y'), 'options->lower_fdot'), 226 227 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially 228 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1 229 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0 230 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))), 231 232 # (a * #b + #c) << #d 233 # ((a * #b) << #d) + (#c << #d) 234 # (a * (#b << #d)) + (#c << #d) 235 (('ishl', ('iadd', ('imul', a, '#b'), '#c'), '#d'), 236 ('iadd', ('imul', a, ('ishl', b, d)), ('ishl', c, d))), 237 238 # (a * #b) << #c 239 # a * (#b << #c) 240 (('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))), 241]) 242 243# Care must be taken here. Shifts in NIR uses only the lower log2(bitsize) 244# bits of the second source. These replacements must correctly handle the 245# case where (b % bitsize) + (c % bitsize) >= bitsize. 246for s in [8, 16, 32, 64]: 247 mask = (1 << s) - 1 248 249 ishl = "ishl@{}".format(s) 250 ishr = "ishr@{}".format(s) 251 ushr = "ushr@{}".format(s) 252 253 in_bounds = ('ult', ('iadd', ('iand', b, mask), ('iand', c, mask)), s) 254 255 optimizations.extend([ 256 ((ishl, (ishl, a, '#b'), '#c'), ('bcsel', in_bounds, (ishl, a, ('iadd', b, c)), 0)), 257 ((ushr, (ushr, a, '#b'), '#c'), ('bcsel', in_bounds, (ushr, a, ('iadd', b, c)), 0)), 258 259 # To get get -1 for large shifts of negative values, ishr must instead 260 # clamp the shift count to the maximum value. 261 ((ishr, (ishr, a, '#b'), '#c'), 262 (ishr, a, ('imin', ('iadd', ('iand', b, mask), ('iand', c, mask)), s - 1))), 263 ]) 264 265# Optimize a pattern of address calculation created by DXVK where the offset is 266# divided by 4 and then multipled by 4. This can be turned into an iand and the 267# additions before can be reassociated to CSE the iand instruction. 268for log2 in range(1, 7): # powers of two from 2 to 64 269 v = 1 << log2 270 mask = 0xffffffff & ~(v - 1) 271 b_is_multiple = '#b(is_unsigned_multiple_of_{})'.format(v) 272 273 optimizations.extend([ 274 # 'a >> #b << #b' -> 'a & ~((1 << #b) - 1)' 275 (('ishl', ('ushr', a, log2), log2), ('iand', a, mask)), 276 277 # Reassociate for improved CSE 278 (('iand', ('iadd', a, b_is_multiple), mask), ('iadd', ('iand', a, mask), b)), 279 ]) 280 281# To save space in the state tables, reduce to the set that is known to help. 282# Previously, this was range(1, 32). In addition, a couple rules inside the 283# loop are commented out. Revisit someday, probably after mesa/#2635 has some 284# resolution. 285for i in [1, 2, 16, 24]: 286 lo_mask = 0xffffffff >> i 287 hi_mask = (0xffffffff << i) & 0xffffffff 288 289 optimizations.extend([ 290 # This pattern seems to only help in the soft-fp64 code. 291 (('ishl@32', ('iand', 'a@32', lo_mask), i), ('ishl', a, i)), 292# (('ushr@32', ('iand', 'a@32', hi_mask), i), ('ushr', a, i)), 293# (('ishr@32', ('iand', 'a@32', hi_mask), i), ('ishr', a, i)), 294 295 (('iand', ('ishl', 'a@32', i), hi_mask), ('ishl', a, i)), 296 (('iand', ('ushr', 'a@32', i), lo_mask), ('ushr', a, i)), 297# (('iand', ('ishr', 'a@32', i), lo_mask), ('ushr', a, i)), # Yes, ushr is correct 298 ]) 299 300optimizations.extend([ 301 # This is common for address calculations. Reassociating may enable the 302 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD 303 # instruction or a constant offset field for in load / store instructions. 304 (('ishl', ('iadd', a, '#b'), '#c'), ('iadd', ('ishl', a, c), ('ishl', b, c))), 305 306 # (a + #b) * #c 307 (('imul', ('iadd(is_used_once)', a, '#b'), '#c'), ('iadd', ('imul', a, c), ('imul', b, c))), 308 309 # Comparison simplifications 310 (('~inot', ('flt', a, b)), ('fge', a, b)), 311 (('~inot', ('fge', a, b)), ('flt', a, b)), 312 (('inot', ('feq', a, b)), ('fneu', a, b)), 313 (('inot', ('fneu', a, b)), ('feq', a, b)), 314 (('inot', ('ilt', a, b)), ('ige', a, b)), 315 (('inot', ('ult', a, b)), ('uge', a, b)), 316 (('inot', ('ige', a, b)), ('ilt', a, b)), 317 (('inot', ('uge', a, b)), ('ult', a, b)), 318 (('inot', ('ieq', a, b)), ('ine', a, b)), 319 (('inot', ('ine', a, b)), ('ieq', a, b)), 320 321 (('iand', ('feq', a, b), ('fneu', a, b)), False), 322 (('iand', ('flt', a, b), ('flt', b, a)), False), 323 (('iand', ('ieq', a, b), ('ine', a, b)), False), 324 (('iand', ('ilt', a, b), ('ilt', b, a)), False), 325 (('iand', ('ult', a, b), ('ult', b, a)), False), 326 327 # This helps some shaders because, after some optimizations, they end up 328 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of 329 # matching would be handled by CSE. 330 (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)), 331 (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)), 332 (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)), 333 (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)), 334 (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)), 335 (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)), 336 (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)), 337 (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)), 338 (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)), 339 (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)), 340 341 # flt(fsat(a), b > 0 && b < 1) is inexact if a is NaN (fsat(NaN) is 0) 342 # because it returns True while flt(a, b) always returns False. 343 (('~flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)), 344 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)), 345 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)), 346 # fge(b > 0 && b < 1, fsat(a)) is inexact if a is NaN (fsat(NaN) is 0) 347 # because it returns True while fge(b, a) always returns False. 348 (('~fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)), 349 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)), 350 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)), 351 352 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)), 353 (('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)), 354 (('fge', 0.0, ('fsat(is_used_once)', a)), ('fge', 0.0, a)), 355 (('flt', 0.0, ('fsat(is_used_once)', a)), ('flt', 0.0, a)), 356 357 # 0.0 >= b2f(a) 358 # b2f(a) <= 0.0 359 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1 360 # inot(a) 361 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a)), 362 363 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a)), 364 365 (('fneu', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)), 366 (('fneu', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)), 367 (('fneu', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a, b)), 368 (('fneu', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a, b)), 369 (('fneu', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)), 370 (('fneu', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)), 371 (('fneu', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a, b)), 372 (('fneu', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a, b)), 373 (('fneu', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a, b)), 374 (('fneu', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a, b)), 375 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))), 376 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))), 377 (('feq', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a, b))), 378 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a, b))), 379 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))), 380 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))), 381 (('feq', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a, b))), 382 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a, b)), 383 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a, b)), 384 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a, b)), 385 386 # -(b2f(a) + b2f(b)) < 0 387 # 0 < b2f(a) + b2f(b) 388 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative 389 # a || b 390 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a, b)), 391 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a, b)), 392 393 # -(b2f(a) + b2f(b)) >= 0 394 # 0 >= b2f(a) + b2f(b) 395 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative 396 # !(a || b) 397 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))), 398 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))), 399 400 (('flt', a, ('fneg', a)), ('flt', a, 0.0)), 401 (('fge', a, ('fneg', a)), ('fge', a, 0.0)), 402 403 # Some optimizations (below) convert things like (a < b || c < b) into 404 # (min(a, c) < b). However, this interfers with the previous optimizations 405 # that try to remove comparisons with negated sums of b2f. This just 406 # breaks that apart. 407 (('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0), 408 ('ior', ('flt', c, 0.0), ('ior', a, b))), 409 410 (('~flt', ('fadd', a, b), a), ('flt', b, 0.0)), 411 (('~fge', ('fadd', a, b), a), ('fge', b, 0.0)), 412 (('~feq', ('fadd', a, b), a), ('feq', b, 0.0)), 413 (('~fneu', ('fadd', a, b), a), ('fneu', b, 0.0)), 414 (('~flt', ('fadd(is_used_once)', a, '#b'), '#c'), ('flt', a, ('fadd', c, ('fneg', b)))), 415 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('flt', ('fneg', ('fadd', c, b)), a)), 416 (('~fge', ('fadd(is_used_once)', a, '#b'), '#c'), ('fge', a, ('fadd', c, ('fneg', b)))), 417 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fge', ('fneg', ('fadd', c, b)), a)), 418 (('~feq', ('fadd(is_used_once)', a, '#b'), '#c'), ('feq', a, ('fadd', c, ('fneg', b)))), 419 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('feq', ('fneg', ('fadd', c, b)), a)), 420 (('~fneu', ('fadd(is_used_once)', a, '#b'), '#c'), ('fneu', a, ('fadd', c, ('fneg', b)))), 421 (('~fneu', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fneu', ('fneg', ('fadd', c, b)), a)), 422 423 # Cannot remove the addition from ilt or ige due to overflow. 424 (('ieq', ('iadd', a, b), a), ('ieq', b, 0)), 425 (('ine', ('iadd', a, b), a), ('ine', b, 0)), 426 427 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a)), 428 (('~fneu', ('b2f', 'a@1'), 0.0), a), 429 (('ieq', ('b2i', 'a@1'), 0), ('inot', a)), 430 (('ine', ('b2i', 'a@1'), 0), a), 431 432 (('fneu', ('u2f', a), 0.0), ('ine', a, 0)), 433 (('feq', ('u2f', a), 0.0), ('ieq', a, 0)), 434 (('fge', ('u2f', a), 0.0), True), 435 (('fge', 0.0, ('u2f', a)), ('uge', 0, a)), # ieq instead? 436 (('flt', ('u2f', a), 0.0), False), 437 (('flt', 0.0, ('u2f', a)), ('ult', 0, a)), # ine instead? 438 (('fneu', ('i2f', a), 0.0), ('ine', a, 0)), 439 (('feq', ('i2f', a), 0.0), ('ieq', a, 0)), 440 (('fge', ('i2f', a), 0.0), ('ige', a, 0)), 441 (('fge', 0.0, ('i2f', a)), ('ige', 0, a)), 442 (('flt', ('i2f', a), 0.0), ('ilt', a, 0)), 443 (('flt', 0.0, ('i2f', a)), ('ilt', 0, a)), 444 445 # 0.0 < fabs(a) 446 # fabs(a) > 0.0 447 # fabs(a) != 0.0 because fabs(a) must be >= 0 448 # a != 0.0 449 (('~flt', 0.0, ('fabs', a)), ('fneu', a, 0.0)), 450 451 # -fabs(a) < 0.0 452 # fabs(a) > 0.0 453 (('~flt', ('fneg', ('fabs', a)), 0.0), ('fneu', a, 0.0)), 454 455 # 0.0 >= fabs(a) 456 # 0.0 == fabs(a) because fabs(a) must be >= 0 457 # 0.0 == a 458 (('fge', 0.0, ('fabs', a)), ('feq', a, 0.0)), 459 460 # -fabs(a) >= 0.0 461 # 0.0 >= fabs(a) 462 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)), 463 464 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a 465 (('iand', ('fge', a, 0.0), ('fge', 1.0, a)), ('feq', a, ('fsat', a)), '!options->lower_fsat'), 466 467 # (a < 0.0) || (a > 1.0) 468 # !(!(a < 0.0) && !(a > 1.0)) 469 # !((a >= 0.0) && (a <= 1.0)) 470 # !(a == fsat(a)) 471 # a != fsat(a) 472 (('ior', ('flt', a, 0.0), ('flt', 1.0, a)), ('fneu', a, ('fsat', a)), '!options->lower_fsat'), 473 474 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a, b))), 475 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a, b)))), 476 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))), 477 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a, b)))), 478 479 # fmin(b2f(a), b) 480 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b)) 481 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b)) 482 # bcsel(a, fmin(1.0, b), fmin(0.0, b)) 483 # 484 # Since b is a constant, constant folding will eliminate the fmin and the 485 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f. 486 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))), 487 488 (('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)), 489 490 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)), 491 (('~bcsel', ('flt', b, a), b, a), ('fmin', a, b)), 492 (('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)), 493 (('~bcsel', ('fge', a, b), b, a), ('fmin', a, b)), 494 (('~bcsel', ('fge', b, a), b, a), ('fmax', a, b)), 495 (('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)), 496 (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)), 497 (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)), 498 (('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)), 499 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))), 500 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))), 501 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)), 502 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)), 503 (('bcsel', a, True, b), ('ior', a, b)), 504 (('bcsel', a, a, b), ('ior', a, b)), 505 (('bcsel', a, b, False), ('iand', a, b)), 506 (('bcsel', a, b, a), ('iand', a, b)), 507 (('~fmin', a, a), a), 508 (('~fmax', a, a), a), 509 (('imin', a, a), a), 510 (('imax', a, a), a), 511 (('umin', a, a), a), 512 (('umax', a, a), a), 513 (('fmax', ('fmax', a, b), b), ('fmax', a, b)), 514 (('umax', ('umax', a, b), b), ('umax', a, b)), 515 (('imax', ('imax', a, b), b), ('imax', a, b)), 516 (('fmin', ('fmin', a, b), b), ('fmin', a, b)), 517 (('umin', ('umin', a, b), b), ('umin', a, b)), 518 (('imin', ('imin', a, b), b), ('imin', a, b)), 519]) 520 521# Integer sizes 522for s in [8, 16, 32, 64]: 523 optimizations.extend([ 524 (('iand@{}'.format(s), a, ('inot', ('ishr', a, s - 1))), ('imax', a, 0)), 525 526 # Simplify logic to detect sign of an integer. 527 (('ieq', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 0), ('ige', a, 0)), 528 (('ine', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 1 << (s - 1)), ('ige', a, 0)), 529 (('ine', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 0), ('ilt', a, 0)), 530 (('ieq', ('iand', 'a@{}'.format(s), 1 << (s - 1)), 1 << (s - 1)), ('ilt', a, 0)), 531 (('ine', ('ushr', 'a@{}'.format(s), s - 1), 0), ('ilt', a, 0)), 532 (('ieq', ('ushr', 'a@{}'.format(s), s - 1), 0), ('ige', a, 0)), 533 (('ieq', ('ushr', 'a@{}'.format(s), s - 1), 1), ('ilt', a, 0)), 534 (('ine', ('ushr', 'a@{}'.format(s), s - 1), 1), ('ige', a, 0)), 535 (('ine', ('ishr', 'a@{}'.format(s), s - 1), 0), ('ilt', a, 0)), 536 (('ieq', ('ishr', 'a@{}'.format(s), s - 1), 0), ('ige', a, 0)), 537 (('ieq', ('ishr', 'a@{}'.format(s), s - 1), -1), ('ilt', a, 0)), 538 (('ine', ('ishr', 'a@{}'.format(s), s - 1), -1), ('ige', a, 0)), 539 ]) 540 541optimizations.extend([ 542 (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))), 543 (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))), 544 (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))), 545 (('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))), 546 (('~fmin', a, ('fabs', a)), a), 547 (('imin', a, ('iabs', a)), a), 548 (('~fmax', a, ('fneg', ('fabs', a))), a), 549 (('imax', a, ('ineg', ('iabs', a))), a), 550 (('fmax', a, ('fabs', a)), ('fabs', a)), 551 (('imax', a, ('iabs', a)), ('iabs', a)), 552 (('fmax', a, ('fneg', a)), ('fabs', a)), 553 (('imax', a, ('ineg', a)), ('iabs', a), '!options->lower_iabs'), 554 (('~fmax', ('fabs', a), 0.0), ('fabs', a)), 555 (('fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'), 556 # fmax(fmin(a, 1.0), 0.0) is inexact because it returns 1.0 on NaN, while 557 # fsat(a) returns 0.0. 558 (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'), 559 # fmin(fmax(a, -1.0), 0.0) is inexact because it returns -1.0 on NaN, while 560 # fneg(fsat(fneg(a))) returns -0.0 on NaN. 561 (('~fmin', ('fmax', a, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'), 562 # fmax(fmin(a, 0.0), -1.0) is inexact because it returns 0.0 on NaN, while 563 # fneg(fsat(fneg(a))) returns -0.0 on NaN. This only matters if 564 # SignedZeroInfNanPreserve is set, but we don't currently have any way of 565 # representing this in the optimizations other than the usual ~. 566 (('~fmax', ('fmin', a, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'), 567 (('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))), 568 (('fsat', ('b2f', a)), ('b2f', a)), 569 (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'), 570 (('fsat', ('fsat', a)), ('fsat', a)), 571 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b))), ('fsat', ('fadd', ('fneg', a), ('fneg', b))), '!options->lower_fsat'), 572 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fneg', a), b)), '!options->lower_fsat'), 573 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fabs', a), ('fabs', b))), '!options->lower_fsat'), 574 (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)), 575 (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)), 576 (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)), 577 # Both the left and right patterns are "b" when isnan(a), so this is exact. 578 (('fmax', ('fsat', a), '#b(is_zero_to_one)'), ('fsat', ('fmax', a, b))), 579 # The left pattern is 0.0 when isnan(a) (because fmin(fsat(NaN), b) -> 580 # fmin(0.0, b)) while the right one is "b", so this optimization is inexact. 581 (('~fmin', ('fsat', a), '#b(is_zero_to_one)'), ('fsat', ('fmin', a, b))), 582 583 # If a in [0,b] then b-a is also in [0,b]. Since b in [0,1], max(b-a, 0) = 584 # fsat(b-a). 585 # 586 # If a > b, then b-a < 0 and max(b-a, 0) = fsat(b-a) = 0 587 # 588 # This should be NaN safe since max(NaN, 0) = fsat(NaN) = 0. 589 (('fmax', ('fadd(is_used_once)', ('fneg', 'a(is_not_negative)'), '#b(is_zero_to_one)'), 0.0), 590 ('fsat', ('fadd', ('fneg', a), b)), '!options->lower_fsat'), 591 592 (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)), 593 (('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))), 594 (('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)), 595 (('~ior', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))), 596 (('~ior', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)), 597 (('~ior', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmax', b, c))), 598 (('~ior', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmin', a, b), c)), 599 (('~ior', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmin', b, c))), 600 (('~ior', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmax', a, b), c)), 601 (('~iand', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmin', b, c))), 602 (('~iand', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmax', a, b), c)), 603 (('~iand', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmax', b, c))), 604 (('~iand', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmin', a, b), c)), 605 (('~iand', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmin', b, c))), 606 (('~iand', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmax', a, b), c)), 607 (('~iand', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmax', b, c))), 608 (('~iand', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmin', a, b), c)), 609 610 (('ior', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imax', b, c))), 611 (('ior', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imin', a, b), c)), 612 (('ior', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imin', b, c))), 613 (('ior', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imax', a, b), c)), 614 (('ior', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umax', b, c))), 615 (('ior', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umin', a, b), c)), 616 (('ior', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umin', b, c))), 617 (('ior', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umax', a, b), c)), 618 (('iand', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imin', b, c))), 619 (('iand', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imax', a, b), c)), 620 (('iand', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imax', b, c))), 621 (('iand', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imin', a, b), c)), 622 (('iand', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umin', b, c))), 623 (('iand', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umax', a, b), c)), 624 (('iand', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umax', b, c))), 625 (('iand', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umin', a, b), c)), 626]) 627 628# Float sizes 629for s in [16, 32, 64]: 630 optimizations.extend([ 631 # These derive from the previous patterns with the application of b < 0 <=> 632 # 0 < -b. The transformation should be applied if either comparison is 633 # used once as this ensures that the number of comparisons will not 634 # increase. The sources to the ior and iand are not symmetric, so the 635 # rules have to be duplicated to get this behavior. 636 (('~ior', ('flt(is_used_once)', 0.0, 'a@{}'.format(s)), ('flt', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))), 637 (('~ior', ('flt', 0.0, 'a@{}'.format(s)), ('flt(is_used_once)', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))), 638 (('~ior', ('fge(is_used_once)', 0.0, 'a@{}'.format(s)), ('fge', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))), 639 (('~ior', ('fge', 0.0, 'a@{}'.format(s)), ('fge(is_used_once)', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))), 640 (('~iand', ('flt(is_used_once)', 0.0, 'a@{}'.format(s)), ('flt', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))), 641 (('~iand', ('flt', 0.0, 'a@{}'.format(s)), ('flt(is_used_once)', 'b@{}'.format(s), 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))), 642 (('~iand', ('fge(is_used_once)', 0.0, 'a@{}'.format(s)), ('fge', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))), 643 (('~iand', ('fge', 0.0, 'a@{}'.format(s)), ('fge(is_used_once)', 'b@{}'.format(s), 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))), 644 645 # The (i2f32, ...) part is an open-coded fsign. When that is combined 646 # with the bcsel, it's basically copysign(1.0, a). There are some 647 # behavior differences between this pattern and copysign w.r.t. ±0 and 648 # NaN. copysign(x, y) blindly takes the sign bit from y and applies it 649 # to x, regardless of whether either or both values are NaN. 650 # 651 # If a != a: bcsel(False, 1.0, i2f(b2i(False) - b2i(False))) = 0, 652 # int(NaN >= 0.0) - int(NaN < 0.0) = 0 - 0 = 0 653 # If a == ±0: bcsel(True, 1.0, ...) = 1.0, 654 # int(±0.0 >= 0.0) - int(±0.0 < 0.0) = 1 - 0 = 1 655 # 656 # For all other values of 'a', the original and replacement behave as 657 # copysign. 658 # 659 # Marking the replacement comparisons as precise prevents any future 660 # optimizations from replacing either of the comparisons with the 661 # logical-not of the other. 662 # 663 # Note: Use b2i32 in the replacement because some platforms that 664 # support fp16 don't support int16. 665 (('bcsel@{}'.format(s), ('feq', a, 0.0), 1.0, ('i2f{}'.format(s), ('iadd', ('b2i{}'.format(s), ('flt', 0.0, 'a@{}'.format(s))), ('ineg', ('b2i{}'.format(s), ('flt', 'a@{}'.format(s), 0.0)))))), 666 ('i2f{}'.format(s), ('iadd', ('b2i32', ('!fge', a, 0.0)), ('ineg', ('b2i32', ('!flt', a, 0.0)))))), 667 668 (('bcsel', a, ('b2f(is_used_once)', 'b@{}'.format(s)), ('b2f', 'c@{}'.format(s))), ('b2f', ('bcsel', a, b, c))), 669 670 # The C spec says, "If the value of the integral part cannot be represented 671 # by the integer type, the behavior is undefined." "Undefined" can mean 672 # "the conversion doesn't happen at all." 673 (('~i2f{}'.format(s), ('f2i', 'a@{}'.format(s))), ('ftrunc', a)), 674 675 # Ironically, mark these as imprecise because removing the conversions may 676 # preserve more precision than doing the conversions (e.g., 677 # uint(float(0x81818181u)) == 0x81818200). 678 (('~f2i{}'.format(s), ('i2f', 'a@{}'.format(s))), a), 679 (('~f2i{}'.format(s), ('u2f', 'a@{}'.format(s))), a), 680 (('~f2u{}'.format(s), ('i2f', 'a@{}'.format(s))), a), 681 (('~f2u{}'.format(s), ('u2f', 'a@{}'.format(s))), a), 682 683 (('fadd', ('b2f{}'.format(s), ('flt', 0.0, 'a@{}'.format(s))), ('fneg', ('b2f{}'.format(s), ('flt', 'a@{}'.format(s), 0.0)))), ('fsign', a), '!options->lower_fsign'), 684 (('iadd', ('b2i{}'.format(s), ('flt', 0, 'a@{}'.format(s))), ('ineg', ('b2i{}'.format(s), ('flt', 'a@{}'.format(s), 0)))), ('f2i{}'.format(s), ('fsign', a)), '!options->lower_fsign'), 685 ]) 686 687 # float? -> float? -> floatS ==> float? -> floatS 688 (('~f2f{}'.format(s), ('f2f', a)), ('f2f{}'.format(s), a)), 689 690 # int? -> float? -> floatS ==> int? -> floatS 691 (('~f2f{}'.format(s), ('u2f', a)), ('u2f{}'.format(s), a)), 692 (('~f2f{}'.format(s), ('i2f', a)), ('i2f{}'.format(s), a)), 693 694 # float? -> float? -> intS ==> float? -> intS 695 (('~f2u{}'.format(s), ('f2f', a)), ('f2u{}'.format(s), a)), 696 (('~f2i{}'.format(s), ('f2f', a)), ('f2i{}'.format(s), a)), 697 698 for B in [32, 64]: 699 if s < B: 700 optimizations.extend([ 701 # S = smaller, B = bigger 702 # typeS -> typeB -> typeS ==> identity 703 (('f2f{}'.format(s), ('f2f{}'.format(B), 'a@{}'.format(s))), a), 704 (('i2i{}'.format(s), ('i2i{}'.format(B), 'a@{}'.format(s))), a), 705 (('u2u{}'.format(s), ('u2u{}'.format(B), 'a@{}'.format(s))), a), 706 707 # bool1 -> typeB -> typeS ==> bool1 -> typeS 708 (('f2f{}'.format(s), ('b2f{}'.format(B), 'a@1')), ('b2f{}'.format(s), a)), 709 (('i2i{}'.format(s), ('b2i{}'.format(B), 'a@1')), ('b2i{}'.format(s), a)), 710 (('u2u{}'.format(s), ('b2i{}'.format(B), 'a@1')), ('b2i{}'.format(s), a)), 711 712 # floatS -> floatB -> intB ==> floatS -> intB 713 (('f2u{}'.format(B), ('f2f{}'.format(B), 'a@{}'.format(s))), ('f2u{}'.format(B), a)), 714 (('f2i{}'.format(B), ('f2f{}'.format(B), 'a@{}'.format(s))), ('f2i{}'.format(B), a)), 715 716 # int? -> floatB -> floatS ==> int? -> floatS 717 (('f2f{}'.format(s), ('u2f{}'.format(B), a)), ('u2f{}'.format(s), a)), 718 (('f2f{}'.format(s), ('i2f{}'.format(B), a)), ('i2f{}'.format(s), a)), 719 720 # intS -> intB -> floatB ==> intS -> floatB 721 (('u2f{}'.format(B), ('u2u{}'.format(B), 'a@{}'.format(s))), ('u2f{}'.format(B), a)), 722 (('i2f{}'.format(B), ('i2i{}'.format(B), 'a@{}'.format(s))), ('i2f{}'.format(B), a)), 723 ]) 724 725# mediump variants of the above 726optimizations.extend([ 727 # int32 -> float32 -> float16 ==> int32 -> float16 728 (('f2fmp', ('u2f32', 'a@32')), ('u2fmp', a)), 729 (('f2fmp', ('i2f32', 'a@32')), ('i2fmp', a)), 730 731 # float32 -> float16 -> int16 ==> float32 -> int16 732 (('f2u16', ('f2fmp', 'a@32')), ('f2u16', a)), 733 (('f2i16', ('f2fmp', 'a@32')), ('f2i16', a)), 734 735 # float32 -> int32 -> int16 ==> float32 -> int16 736 (('i2imp', ('f2u32', 'a@32')), ('f2ump', a)), 737 (('i2imp', ('f2i32', 'a@32')), ('f2imp', a)), 738 739 # int32 -> int16 -> float16 ==> int32 -> float16 740 (('u2f16', ('i2imp', 'a@32')), ('u2f16', a)), 741 (('i2f16', ('i2imp', 'a@32')), ('i2f16', a)), 742]) 743 744# Integer sizes 745for s in [8, 16, 32, 64]: 746 optimizations.extend([ 747 (('iand', ('ieq', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('ior', a, b), 0), 'options->lower_umax'), 748 (('ior', ('ine', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('ior', a, b), 0), 'options->lower_umin'), 749 (('iand', ('ieq', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('umax', a, b), 0), '!options->lower_umax'), 750 (('ior', ('ieq', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('umin', a, b), 0), '!options->lower_umin'), 751 (('iand', ('ine', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('umin', a, b), 0), '!options->lower_umin'), 752 (('ior', ('ine', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('umax', a, b), 0), '!options->lower_umax'), 753 754 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True). 755 (('ineg', ('b2i{}'.format(s), 'a@{}'.format(s))), a), 756 757 # SM5 32-bit shifts are defined to use the 5 least significant bits (or 4 bits for 16 bits) 758 (('ishl', 'a@{}'.format(s), ('iand', s - 1, b)), ('ishl', a, b)), 759 (('ishr', 'a@{}'.format(s), ('iand', s - 1, b)), ('ishr', a, b)), 760 (('ushr', 'a@{}'.format(s), ('iand', s - 1, b)), ('ushr', a, b)), 761 ]) 762 763optimizations.extend([ 764 # Common pattern like 'if (i == 0 || i == 1 || ...)' 765 (('ior', ('ieq', a, 0), ('ieq', a, 1)), ('uge', 1, a)), 766 (('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)), 767 (('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)), 768 769 (('ior', a, ('ieq', a, False)), True), 770 (('ior', a, ('inot', a)), -1), 771 772 (('ine', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))), ('ine', a, b)), 773 (('b2i', ('ine', 'a@1', 'b@1')), ('b2i', ('ixor', a, b))), 774 775 # This pattern occurs coutresy of __flt64_nonnan in the soft-fp64 code. 776 # The first part of the iand comes from the !__feq64_nonnan. 777 # 778 # The second pattern is a reformulation of the first based on the relation 779 # (a == 0 || y == 0) <=> umin(a, y) == 0, where b in the first equation 780 # happens to be y == 0. 781 (('iand', ('inot', ('iand', ('ior', ('ieq', a, 0), b), c)), ('ilt', a, 0)), 782 ('iand', ('inot', ('iand', b , c)), ('ilt', a, 0))), 783 (('iand', ('inot', ('iand', ('ieq', ('umin', a, b), 0), c)), ('ilt', a, 0)), 784 ('iand', ('inot', ('iand', ('ieq', b , 0), c)), ('ilt', a, 0))), 785 786 # These patterns can result when (a < b || a < c) => (a < min(b, c)) 787 # transformations occur before constant propagation and loop-unrolling. 788 (('~flt', a, ('fmax', b, a)), ('flt', a, b)), 789 (('~flt', ('fmin', a, b), a), ('flt', b, a)), 790 (('~fge', a, ('fmin', b, a)), True), 791 (('~fge', ('fmax', a, b), a), True), 792 (('~flt', a, ('fmin', b, a)), False), 793 (('~flt', ('fmax', a, b), a), False), 794 (('~fge', a, ('fmax', b, a)), ('fge', a, b)), 795 (('~fge', ('fmin', a, b), a), ('fge', b, a)), 796 797 (('ilt', a, ('imax', b, a)), ('ilt', a, b)), 798 (('ilt', ('imin', a, b), a), ('ilt', b, a)), 799 (('ige', a, ('imin', b, a)), True), 800 (('ige', ('imax', a, b), a), True), 801 (('ult', a, ('umax', b, a)), ('ult', a, b)), 802 (('ult', ('umin', a, b), a), ('ult', b, a)), 803 (('uge', a, ('umin', b, a)), True), 804 (('uge', ('umax', a, b), a), True), 805 (('ilt', a, ('imin', b, a)), False), 806 (('ilt', ('imax', a, b), a), False), 807 (('ige', a, ('imax', b, a)), ('ige', a, b)), 808 (('ige', ('imin', a, b), a), ('ige', b, a)), 809 (('ult', a, ('umin', b, a)), False), 810 (('ult', ('umax', a, b), a), False), 811 (('uge', a, ('umax', b, a)), ('uge', a, b)), 812 (('uge', ('umin', a, b), a), ('uge', b, a)), 813 (('ult', a, ('iand', b, a)), False), 814 (('ult', ('ior', a, b), a), False), 815 (('uge', a, ('iand', b, a)), True), 816 (('uge', ('ior', a, b), a), True), 817 818 (('ilt', '#a', ('imax', '#b', c)), ('ior', ('ilt', a, b), ('ilt', a, c))), 819 (('ilt', ('imin', '#a', b), '#c'), ('ior', ('ilt', a, c), ('ilt', b, c))), 820 (('ige', '#a', ('imin', '#b', c)), ('ior', ('ige', a, b), ('ige', a, c))), 821 (('ige', ('imax', '#a', b), '#c'), ('ior', ('ige', a, c), ('ige', b, c))), 822 (('ult', '#a', ('umax', '#b', c)), ('ior', ('ult', a, b), ('ult', a, c))), 823 (('ult', ('umin', '#a', b), '#c'), ('ior', ('ult', a, c), ('ult', b, c))), 824 (('uge', '#a', ('umin', '#b', c)), ('ior', ('uge', a, b), ('uge', a, c))), 825 (('uge', ('umax', '#a', b), '#c'), ('ior', ('uge', a, c), ('uge', b, c))), 826 (('ilt', '#a', ('imin', '#b', c)), ('iand', ('ilt', a, b), ('ilt', a, c))), 827 (('ilt', ('imax', '#a', b), '#c'), ('iand', ('ilt', a, c), ('ilt', b, c))), 828 (('ige', '#a', ('imax', '#b', c)), ('iand', ('ige', a, b), ('ige', a, c))), 829 (('ige', ('imin', '#a', b), '#c'), ('iand', ('ige', a, c), ('ige', b, c))), 830 (('ult', '#a', ('umin', '#b', c)), ('iand', ('ult', a, b), ('ult', a, c))), 831 (('ult', ('umax', '#a', b), '#c'), ('iand', ('ult', a, c), ('ult', b, c))), 832 (('uge', '#a', ('umax', '#b', c)), ('iand', ('uge', a, b), ('uge', a, c))), 833 (('uge', ('umin', '#a', b), '#c'), ('iand', ('uge', a, c), ('uge', b, c))), 834 835 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is 836 # negative. 837 (('bcsel', ('ilt', a, 0), ('ineg', ('ishr', a, b)), ('ishr', a, b)), 838 ('iabs', ('ishr', a, b))), 839 (('iabs', ('ishr', ('iabs', a), b)), ('ishr', ('iabs', a), b)), 840 841 (('fabs', ('slt', a, b)), ('slt', a, b)), 842 (('fabs', ('sge', a, b)), ('sge', a, b)), 843 (('fabs', ('seq', a, b)), ('seq', a, b)), 844 (('fabs', ('sne', a, b)), ('sne', a, b)), 845 (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'), 846 (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'), 847 (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'), 848 (('sne', a, b), ('b2f', ('fneu', a, b)), 'options->lower_scmp'), 849 (('seq', ('seq', a, b), 1.0), ('seq', a, b)), 850 (('seq', ('sne', a, b), 1.0), ('sne', a, b)), 851 (('seq', ('slt', a, b), 1.0), ('slt', a, b)), 852 (('seq', ('sge', a, b), 1.0), ('sge', a, b)), 853 (('sne', ('seq', a, b), 0.0), ('seq', a, b)), 854 (('sne', ('sne', a, b), 0.0), ('sne', a, b)), 855 (('sne', ('slt', a, b), 0.0), ('slt', a, b)), 856 (('sne', ('sge', a, b), 0.0), ('sge', a, b)), 857 (('seq', ('seq', a, b), 0.0), ('sne', a, b)), 858 (('seq', ('sne', a, b), 0.0), ('seq', a, b)), 859 (('seq', ('slt', a, b), 0.0), ('sge', a, b)), 860 (('seq', ('sge', a, b), 0.0), ('slt', a, b)), 861 (('sne', ('seq', a, b), 1.0), ('sne', a, b)), 862 (('sne', ('sne', a, b), 1.0), ('seq', a, b)), 863 (('sne', ('slt', a, b), 1.0), ('sge', a, b)), 864 (('sne', ('sge', a, b), 1.0), ('slt', a, b)), 865 (('fall_equal2', a, b), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'), 866 (('fall_equal3', a, b), ('seq', ('fany_nequal3', a, b), 0.0), 'options->lower_vector_cmp'), 867 (('fall_equal4', a, b), ('seq', ('fany_nequal4', a, b), 0.0), 'options->lower_vector_cmp'), 868 (('fany_nequal2', a, b), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'), 869 (('fany_nequal3', a, b), ('fsat', ('fdot3', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'), 870 (('fany_nequal4', a, b), ('fsat', ('fdot4', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'), 871 872 (('ball_iequal2', a, b), ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), 'options->lower_vector_cmp'), 873 (('ball_iequal3', a, b), ('iand', ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), ('ieq', 'a.z', 'b.z')), 'options->lower_vector_cmp'), 874 (('ball_iequal4', a, b), ('iand', ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), ('iand', ('ieq', 'a.z', 'b.z'), ('ieq', 'a.w', 'b.w'))), 'options->lower_vector_cmp'), 875 876 (('bany_inequal2', a, b), ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), 'options->lower_vector_cmp'), 877 (('bany_inequal3', a, b), ('ior', ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), ('ine', 'a.z', 'b.z')), 'options->lower_vector_cmp'), 878 (('bany_inequal4', a, b), ('ior', ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), ('ior', ('ine', 'a.z', 'b.z'), ('ine', 'a.w', 'b.w'))), 'options->lower_vector_cmp'), 879 880 (('ball_fequal2', a, b), ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), 'options->lower_vector_cmp'), 881 (('ball_fequal3', a, b), ('iand', ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), ('feq', 'a.z', 'b.z')), 'options->lower_vector_cmp'), 882 (('ball_fequal4', a, b), ('iand', ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), ('iand', ('feq', 'a.z', 'b.z'), ('feq', 'a.w', 'b.w'))), 'options->lower_vector_cmp'), 883 884 (('bany_fnequal2', a, b), ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), 'options->lower_vector_cmp'), 885 (('bany_fnequal3', a, b), ('ior', ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), ('fneu', 'a.z', 'b.z')), 'options->lower_vector_cmp'), 886 (('bany_fnequal4', a, b), ('ior', ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), ('ior', ('fneu', 'a.z', 'b.z'), ('fneu', 'a.w', 'b.w'))), 'options->lower_vector_cmp'), 887 888 (('fneu', ('fneg', a), a), ('fneu', a, 0.0)), 889 (('feq', ('fneg', a), a), ('feq', a, 0.0)), 890 # Emulating booleans 891 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))), 892 (('iand', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))), 893 (('ior', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('ior', a, b))), 894 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))), 895 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a, b))), 896 (('iand', 'a@bool16', 1.0), ('b2f', a)), 897 (('iand', 'a@bool32', 1.0), ('b2f', a)), 898 (('flt', ('fneg', ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF. 899 # Comparison with the same args. Note that these are not done for 900 # the float versions because NaN always returns false on float 901 # inequalities. 902 (('ilt', a, a), False), 903 (('ige', a, a), True), 904 (('ieq', a, a), True), 905 (('ine', a, a), False), 906 (('ult', a, a), False), 907 (('uge', a, a), True), 908 # Logical and bit operations 909 (('iand', a, a), a), 910 (('iand', a, ~0), a), 911 (('iand', a, 0), 0), 912 (('ior', a, a), a), 913 (('ior', a, 0), a), 914 (('ior', a, True), True), 915 (('ixor', a, a), 0), 916 (('ixor', a, 0), a), 917 (('inot', ('inot', a)), a), 918 (('ior', ('iand', a, b), b), b), 919 (('ior', ('ior', a, b), b), ('ior', a, b)), 920 (('iand', ('ior', a, b), b), b), 921 (('iand', ('iand', a, b), b), ('iand', a, b)), 922 # DeMorgan's Laws 923 (('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))), 924 (('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))), 925 # Shift optimizations 926 (('ishl', 0, a), 0), 927 (('ishl', a, 0), a), 928 (('ishr', 0, a), 0), 929 (('ishr', a, 0), a), 930 (('ushr', 0, a), 0), 931 (('ushr', a, 0), a), 932 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('iadd', 16, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'), 933 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('isub', 16, b))), ('urol', a, b), '!options->lower_rotate'), 934 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('iadd', 32, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'), 935 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('isub', 32, b))), ('urol', a, b), '!options->lower_rotate'), 936 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('iadd', 16, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'), 937 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('isub', 16, b))), ('uror', a, b), '!options->lower_rotate'), 938 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('iadd', 32, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'), 939 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('isub', 32, b))), ('uror', a, b), '!options->lower_rotate'), 940 (('urol@16', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 16, b))), 'options->lower_rotate'), 941 (('urol@32', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 32, b))), 'options->lower_rotate'), 942 (('uror@16', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 16, b))), 'options->lower_rotate'), 943 (('uror@32', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 32, b))), 'options->lower_rotate'), 944 # Exponential/logarithmic identities 945 (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a 946 (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a 947 (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b) 948 (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b 949 (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))), 950 ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d 951 (('~fexp2', ('fmul', ('flog2', a), 0.5)), ('fsqrt', a)), 952 (('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)), 953 (('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))), 954 (('~fpow', a, 1.0), a), 955 (('~fpow', a, 2.0), ('fmul', a, a)), 956 (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))), 957 (('~fpow', 2.0, a), ('fexp2', a)), 958 (('~fpow', ('fpow', a, 2.2), 0.454545), a), 959 (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)), 960 (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))), 961 (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))), 962 (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))), 963 (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))), 964 (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))), 965 (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))), 966 (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))), 967 (('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))), 968 (('bcsel', ('flt', a, 0.0), 0.0, ('fsqrt', a)), ('fsqrt', ('fmax', a, 0.0))), 969 (('~fmul', ('fsqrt', a), ('fsqrt', a)), ('fabs',a)), 970 # Division and reciprocal 971 (('~fdiv', 1.0, a), ('frcp', a)), 972 (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'), 973 (('~frcp', ('frcp', a)), a), 974 (('~frcp', ('fsqrt', a)), ('frsq', a)), 975 (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'), 976 (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'), 977 # Trig 978 (('fsin', a), lowered_sincos(0.5), 'options->lower_sincos'), 979 (('fcos', a), lowered_sincos(0.75), 'options->lower_sincos'), 980 # Boolean simplifications 981 (('i2b16(is_used_by_if)', a), ('ine16', a, 0)), 982 (('i2b32(is_used_by_if)', a), ('ine32', a, 0)), 983 (('i2b1(is_used_by_if)', a), ('ine', a, 0)), 984 (('ieq', a, True), a), 985 (('ine(is_not_used_by_if)', a, True), ('inot', a)), 986 (('ine', a, False), a), 987 (('ieq(is_not_used_by_if)', a, False), ('inot', 'a')), 988 (('bcsel', a, True, False), a), 989 (('bcsel', a, False, True), ('inot', a)), 990 (('bcsel', True, b, c), b), 991 (('bcsel', False, b, c), c), 992 993 (('bcsel@16', a, 1.0, 0.0), ('b2f', a)), 994 (('bcsel@16', a, 0.0, 1.0), ('b2f', ('inot', a))), 995 (('bcsel@16', a, -1.0, -0.0), ('fneg', ('b2f', a))), 996 (('bcsel@16', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))), 997 (('bcsel@32', a, 1.0, 0.0), ('b2f', a)), 998 (('bcsel@32', a, 0.0, 1.0), ('b2f', ('inot', a))), 999 (('bcsel@32', a, -1.0, -0.0), ('fneg', ('b2f', a))), 1000 (('bcsel@32', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))), 1001 (('bcsel@64', a, 1.0, 0.0), ('b2f', a), '!(options->lower_doubles_options & nir_lower_fp64_full_software)'), 1002 (('bcsel@64', a, 0.0, 1.0), ('b2f', ('inot', a)), '!(options->lower_doubles_options & nir_lower_fp64_full_software)'), 1003 (('bcsel@64', a, -1.0, -0.0), ('fneg', ('b2f', a)), '!(options->lower_doubles_options & nir_lower_fp64_full_software)'), 1004 (('bcsel@64', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a))), '!(options->lower_doubles_options & nir_lower_fp64_full_software)'), 1005 1006 (('bcsel', a, b, b), b), 1007 (('~fcsel', a, b, b), b), 1008 1009 # D3D Boolean emulation 1010 (('bcsel', a, -1, 0), ('ineg', ('b2i', 'a@1'))), 1011 (('bcsel', a, 0, -1), ('ineg', ('b2i', ('inot', a)))), 1012 (('bcsel', a, 1, 0), ('b2i', 'a@1')), 1013 (('bcsel', a, 0, 1), ('b2i', ('inot', a))), 1014 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))), 1015 ('ineg', ('b2i', ('iand', a, b)))), 1016 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))), 1017 ('ineg', ('b2i', ('ior', a, b)))), 1018 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a)), 1019 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a), 1020 (('ine', ('ineg', ('b2i', 'a@1')), 0), a), 1021 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a)), 1022 (('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)), 1023 (('iand', ('ineg', ('b2i', a)), 1), ('b2i', a)), 1024 1025 # Conversions 1026 (('i2b16', ('b2i', 'a@16')), a), 1027 (('i2b32', ('b2i', 'a@32')), a), 1028 (('f2i', ('ftrunc', a)), ('f2i', a)), 1029 (('f2u', ('ftrunc', a)), ('f2u', a)), 1030 (('i2b', ('ineg', a)), ('i2b', a)), 1031 (('i2b', ('iabs', a)), ('i2b', a)), 1032 (('inot', ('f2b1', a)), ('feq', a, 0.0)), 1033 1034 # Conversions from 16 bits to 32 bits and back can always be removed 1035 (('f2fmp', ('f2f32', 'a@16')), a), 1036 (('i2imp', ('i2i32', 'a@16')), a), 1037 (('i2imp', ('u2u32', 'a@16')), a), 1038 1039 (('f2imp', ('f2f32', 'a@16')), ('f2i16', a)), 1040 (('f2ump', ('f2f32', 'a@16')), ('f2u16', a)), 1041 (('i2fmp', ('i2i32', 'a@16')), ('i2f16', a)), 1042 (('u2fmp', ('u2u32', 'a@16')), ('u2f16', a)), 1043 1044 (('f2fmp', ('b2f32', 'a@1')), ('b2f16', a)), 1045 (('i2imp', ('b2i32', 'a@1')), ('b2i16', a)), 1046 (('i2imp', ('b2i32', 'a@1')), ('b2i16', a)), 1047 1048 (('f2imp', ('b2f32', 'a@1')), ('b2i16', a)), 1049 (('f2ump', ('b2f32', 'a@1')), ('b2i16', a)), 1050 (('i2fmp', ('b2i32', 'a@1')), ('b2f16', a)), 1051 (('u2fmp', ('b2i32', 'a@1')), ('b2f16', a)), 1052 1053 # Conversions to 16 bits would be lossy so they should only be removed if 1054 # the instruction was generated by the precision lowering pass. 1055 (('f2f32', ('f2fmp', 'a@32')), a), 1056 (('i2i32', ('i2imp', 'a@32')), a), 1057 (('u2u32', ('i2imp', 'a@32')), a), 1058 1059 (('i2i32', ('f2imp', 'a@32')), ('f2i32', a)), 1060 (('u2u32', ('f2ump', 'a@32')), ('f2u32', a)), 1061 (('f2f32', ('i2fmp', 'a@32')), ('i2f32', a)), 1062 (('f2f32', ('u2fmp', 'a@32')), ('u2f32', a)), 1063 1064 (('ffloor', 'a(is_integral)'), a), 1065 (('fceil', 'a(is_integral)'), a), 1066 (('ftrunc', 'a(is_integral)'), a), 1067 # fract(x) = x - floor(x), so fract(NaN) = NaN 1068 (('~ffract', 'a(is_integral)'), 0.0), 1069 (('fabs', 'a(is_not_negative)'), a), 1070 (('iabs', 'a(is_not_negative)'), a), 1071 (('fsat', 'a(is_not_positive)'), 0.0), 1072 1073 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec 1074 # says: 1075 # 1076 # It is undefined to convert a negative floating-point value to an 1077 # uint. 1078 # 1079 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows 1080 # some optimizations in the i965 backend to proceed. 1081 (('ige', ('f2u', a), b), ('ige', ('f2i', a), b)), 1082 (('ige', b, ('f2u', a)), ('ige', b, ('f2i', a))), 1083 (('ilt', ('f2u', a), b), ('ilt', ('f2i', a), b)), 1084 (('ilt', b, ('f2u', a)), ('ilt', b, ('f2i', a))), 1085 1086 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a), '!options->lower_fsat'), 1087 1088 # The result of the multiply must be in [-1, 0], so the result of the ffma 1089 # must be in [0, 1]. 1090 (('flt', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), False), 1091 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), False), 1092 (('fmax', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0)), 1093 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0)), 1094 1095 (('fneu', 'a(is_not_zero)', 0.0), True), 1096 (('feq', 'a(is_not_zero)', 0.0), False), 1097 1098 # In this chart, + means value > 0 and - means value < 0. 1099 # 1100 # + >= + -> unknown 0 >= + -> false - >= + -> false 1101 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false 1102 # + >= - -> true 0 >= - -> true - >= - -> unknown 1103 # 1104 # Using grouping conceptually similar to a Karnaugh map... 1105 # 1106 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true 1107 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false 1108 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false 1109 # 1110 # The flt / ilt cases just invert the expected result. 1111 # 1112 # The results expecting true, must be marked imprecise. The results 1113 # expecting false are fine because NaN compared >= or < anything is false. 1114 1115 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True), 1116 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False), 1117 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False), 1118 1119 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False), 1120 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True), 1121 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True), 1122 1123 (('ine', 'a(is_not_zero)', 0), True), 1124 (('ieq', 'a(is_not_zero)', 0), False), 1125 1126 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True), 1127 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False), 1128 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False), 1129 1130 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False), 1131 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True), 1132 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True), 1133 1134 (('ult', 0, 'a(is_gt_zero)'), True), 1135 (('ult', a, 0), False), 1136 1137 # Packing and then unpacking does nothing 1138 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a), 1139 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b), 1140 (('unpack_64_2x32', ('pack_64_2x32_split', a, b)), ('vec2', a, b)), 1141 (('unpack_64_2x32', ('pack_64_2x32', a)), a), 1142 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a), 1143 ('unpack_64_2x32_split_y', a)), a), 1144 (('pack_64_2x32', ('vec2', ('unpack_64_2x32_split_x', a), 1145 ('unpack_64_2x32_split_y', a))), a), 1146 (('pack_64_2x32', ('unpack_64_2x32', a)), a), 1147 1148 # Comparing two halves of an unpack separately. While this optimization 1149 # should be correct for non-constant values, it's less obvious that it's 1150 # useful in that case. For constant values, the pack will fold and we're 1151 # guaranteed to reduce the whole tree to one instruction. 1152 (('iand', ('ieq', ('unpack_32_2x16_split_x', a), '#b'), 1153 ('ieq', ('unpack_32_2x16_split_y', a), '#c')), 1154 ('ieq', a, ('pack_32_2x16_split', b, c))), 1155 1156 # Byte extraction 1157 (('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'), 1158 (('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'), 1159 (('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'), 1160 (('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'), 1161 (('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'), 1162 (('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'), 1163 (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'), 1164 1165 (('ubfe', a, 0, 8), ('extract_u8', a, 0), '!options->lower_extract_byte'), 1166 (('ubfe', a, 8, 8), ('extract_u8', a, 1), '!options->lower_extract_byte'), 1167 (('ubfe', a, 16, 8), ('extract_u8', a, 2), '!options->lower_extract_byte'), 1168 (('ubfe', a, 24, 8), ('extract_u8', a, 3), '!options->lower_extract_byte'), 1169 (('ibfe', a, 0, 8), ('extract_i8', a, 0), '!options->lower_extract_byte'), 1170 (('ibfe', a, 8, 8), ('extract_i8', a, 1), '!options->lower_extract_byte'), 1171 (('ibfe', a, 16, 8), ('extract_i8', a, 2), '!options->lower_extract_byte'), 1172 (('ibfe', a, 24, 8), ('extract_i8', a, 3), '!options->lower_extract_byte'), 1173 1174 # Word extraction 1175 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'), 1176 (('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'), 1177 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a, 0), '!options->lower_extract_word'), 1178 (('ishr', 'a@32', 16), ('extract_i16', a, 1), '!options->lower_extract_word'), 1179 (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'), 1180 1181 (('ubfe', a, 0, 16), ('extract_u16', a, 0), '!options->lower_extract_word'), 1182 (('ubfe', a, 16, 16), ('extract_u16', a, 1), '!options->lower_extract_word'), 1183 (('ibfe', a, 0, 16), ('extract_i16', a, 0), '!options->lower_extract_word'), 1184 (('ibfe', a, 16, 16), ('extract_i16', a, 1), '!options->lower_extract_word'), 1185 1186 # Lower pack/unpack 1187 (('pack_64_2x32_split', a, b), ('ior', ('u2u64', a), ('ishl', ('u2u64', b), 32)), 'options->lower_pack_64_2x32_split'), 1188 (('pack_32_2x16_split', a, b), ('ior', ('u2u32', a), ('ishl', ('u2u32', b), 16)), 'options->lower_pack_32_2x16_split'), 1189 (('unpack_64_2x32_split_x', a), ('u2u32', a), 'options->lower_unpack_64_2x32_split'), 1190 (('unpack_64_2x32_split_y', a), ('u2u32', ('ushr', a, 32)), 'options->lower_unpack_64_2x32_split'), 1191 (('unpack_32_2x16_split_x', a), ('u2u16', a), 'options->lower_unpack_32_2x16_split'), 1192 (('unpack_32_2x16_split_y', a), ('u2u16', ('ushr', a, 16)), 'options->lower_unpack_32_2x16_split'), 1193 1194 # Useless masking before unpacking 1195 (('unpack_half_2x16_split_x', ('iand', a, 0xffff)), ('unpack_half_2x16_split_x', a)), 1196 (('unpack_32_2x16_split_x', ('iand', a, 0xffff)), ('unpack_32_2x16_split_x', a)), 1197 (('unpack_64_2x32_split_x', ('iand', a, 0xffffffff)), ('unpack_64_2x32_split_x', a)), 1198 (('unpack_half_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_half_2x16_split_y', a)), 1199 (('unpack_32_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_32_2x16_split_y', a)), 1200 (('unpack_64_2x32_split_y', ('iand', a, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a)), 1201 1202 (('unpack_half_2x16_split_x', ('extract_u16', a, 0)), ('unpack_half_2x16_split_x', a)), 1203 (('unpack_half_2x16_split_x', ('extract_u16', a, 1)), ('unpack_half_2x16_split_y', a)), 1204 (('unpack_half_2x16_split_x', ('ushr', a, 16)), ('unpack_half_2x16_split_y', a)), 1205 (('unpack_32_2x16_split_x', ('extract_u16', a, 0)), ('unpack_32_2x16_split_x', a)), 1206 (('unpack_32_2x16_split_x', ('extract_u16', a, 1)), ('unpack_32_2x16_split_y', a)), 1207 1208 # Optimize half packing 1209 (('ishl', ('pack_half_2x16', ('vec2', a, 0)), 16), ('pack_half_2x16', ('vec2', 0, a))), 1210 (('ushr', ('pack_half_2x16', ('vec2', 0, a)), 16), ('pack_half_2x16', ('vec2', a, 0))), 1211 1212 (('iadd', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))), 1213 ('pack_half_2x16', ('vec2', a, b))), 1214 (('ior', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))), 1215 ('pack_half_2x16', ('vec2', a, b))), 1216 1217 (('ishl', ('pack_half_2x16_split', a, 0), 16), ('pack_half_2x16_split', 0, a)), 1218 (('ushr', ('pack_half_2x16_split', 0, a), 16), ('pack_half_2x16_split', a, 0)), 1219 (('extract_u16', ('pack_half_2x16_split', 0, a), 1), ('pack_half_2x16_split', a, 0)), 1220 1221 (('iadd', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)), 1222 (('ior', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)), 1223]) 1224 1225# After the ('extract_u8', a, 0) pattern, above, triggers, there will be 1226# patterns like those below. 1227for op in ('ushr', 'ishr'): 1228 optimizations.extend([(('extract_u8', (op, 'a@16', 8), 0), ('extract_u8', a, 1))]) 1229 optimizations.extend([(('extract_u8', (op, 'a@32', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)]) 1230 optimizations.extend([(('extract_u8', (op, 'a@64', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)]) 1231 1232optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))]) 1233 1234# After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be 1235# patterns like those below. 1236for op in ('extract_u8', 'extract_i8'): 1237 optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))]) 1238 optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)]) 1239 optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)]) 1240 1241optimizations.extend([ 1242 # Subtracts 1243 (('ussub_4x8', a, 0), a), 1244 (('ussub_4x8', a, ~0), 0), 1245 # Lower all Subtractions first - they can get recombined later 1246 (('fsub', a, b), ('fadd', a, ('fneg', b))), 1247 (('isub', a, b), ('iadd', a, ('ineg', b))), 1248 (('uabs_usub', a, b), ('bcsel', ('ult', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))), 1249 # This is correct. We don't need isub_sat because the result type is unsigned, so it cannot overflow. 1250 (('uabs_isub', a, b), ('bcsel', ('ilt', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))), 1251 1252 # Propagate negation up multiplication chains 1253 (('fmul(is_used_by_non_fsat)', ('fneg', a), b), ('fneg', ('fmul', a, b))), 1254 (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))), 1255 1256 # Propagate constants up multiplication chains 1257 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)), 1258 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)), 1259 # Prefer moving out a multiplication for more MAD/FMA-friendly code 1260 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_fmul)'), '#c'), ('fadd', ('fadd', a, c), b)), 1261 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)), 1262 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)), 1263 1264 # Reassociate constants in add/mul chains so they can be folded together. 1265 # For now, we mostly only handle cases where the constants are separated by 1266 # a single non-constant. We could do better eventually. 1267 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a, c), b)), 1268 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a, c), b)), 1269 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a, c), b)), 1270 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a, ('fneg', c)), ('fneg', b))), 1271 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a, c), b)), 1272 (('iand', '#a', ('iand', 'b(is_not_const)', '#c')), ('iand', ('iand', a, c), b)), 1273 (('ior', '#a', ('ior', 'b(is_not_const)', '#c')), ('ior', ('ior', a, c), b)), 1274 (('ixor', '#a', ('ixor', 'b(is_not_const)', '#c')), ('ixor', ('ixor', a, c), b)), 1275 1276 # Drop mul-div by the same value when there's no wrapping. 1277 (('idiv', ('imul(no_signed_wrap)', a, b), b), a), 1278 1279 # By definition... 1280 (('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)), 1281 (('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)), 1282 (('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)), 1283 1284 (('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)), 1285 (('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)), 1286 (('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)), 1287 1288 (('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)), 1289 1290 (('~fmul', ('bcsel(is_used_once)', c, -1.0, 1.0), b), ('bcsel', c, ('fneg', b), b)), 1291 (('~fmul', ('bcsel(is_used_once)', c, 1.0, -1.0), b), ('bcsel', c, b, ('fneg', b))), 1292 (('~bcsel', ('flt', a, 0.0), ('fneg', a), a), ('fabs', a)), 1293 1294 (('bcsel', a, ('bcsel', b, c, d), d), ('bcsel', ('iand', a, b), c, d)), 1295 (('bcsel', a, b, ('bcsel', c, b, d)), ('bcsel', ('ior', a, c), b, d)), 1296 1297 # Misc. lowering 1298 (('fmod', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'), 1299 (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod'), 1300 (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'), 1301 (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'), 1302 1303 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'), 1304 ('bcsel', ('ult', 31, 'bits'), 'insert', 1305 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')), 1306 'options->lower_bitfield_insert'), 1307 (('ihadd', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'), 1308 (('uhadd', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'), 1309 (('irhadd', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'), 1310 (('urhadd', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'), 1311 (('ihadd@64', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'), 1312 (('uhadd@64', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'), 1313 (('irhadd@64', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'), 1314 (('urhadd@64', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'), 1315 1316 (('uadd_sat@64', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat || (options->lower_int64_options & nir_lower_iadd64) != 0'), 1317 (('uadd_sat', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat'), 1318 (('usub_sat', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_add_sat'), 1319 (('usub_sat@64', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_usub_sat64 || (options->lower_int64_options & nir_lower_iadd64) != 0'), 1320 1321 # int64_t sum = a + b; 1322 # 1323 # if (a < 0 && b < 0 && a < sum) 1324 # sum = INT64_MIN; 1325 # } else if (a >= 0 && b >= 0 && sum < a) 1326 # sum = INT64_MAX; 1327 # } 1328 # 1329 # A couple optimizations are applied. 1330 # 1331 # 1. a < sum => sum >= 0. This replacement works because it is known that 1332 # a < 0 and b < 0, so sum should also be < 0 unless there was 1333 # underflow. 1334 # 1335 # 2. sum < a => sum < 0. This replacement works because it is known that 1336 # a >= 0 and b >= 0, so sum should also be >= 0 unless there was 1337 # overflow. 1338 # 1339 # 3. Invert the second if-condition and swap the order of parameters for 1340 # the bcsel. !(a >= 0 && b >= 0 && sum < 0) becomes !(a >= 0) || !(b >= 1341 # 0) || !(sum < 0), and that becomes (a < 0) || (b < 0) || (sum >= 0) 1342 # 1343 # On Intel Gen11, this saves ~11 instructions. 1344 (('iadd_sat@64', a, b), ('bcsel', 1345 ('iand', ('iand', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)), 1346 0x8000000000000000, 1347 ('bcsel', 1348 ('ior', ('ior', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)), 1349 ('iadd', a, b), 1350 0x7fffffffffffffff)), 1351 '(options->lower_int64_options & nir_lower_iadd64) != 0'), 1352 1353 # int64_t sum = a - b; 1354 # 1355 # if (a < 0 && b >= 0 && a < sum) 1356 # sum = INT64_MIN; 1357 # } else if (a >= 0 && b < 0 && a >= sum) 1358 # sum = INT64_MAX; 1359 # } 1360 # 1361 # Optimizations similar to the iadd_sat case are applied here. 1362 (('isub_sat@64', a, b), ('bcsel', 1363 ('iand', ('iand', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)), 1364 0x8000000000000000, 1365 ('bcsel', 1366 ('ior', ('ior', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)), 1367 ('isub', a, b), 1368 0x7fffffffffffffff)), 1369 '(options->lower_int64_options & nir_lower_iadd64) != 0'), 1370 1371 # These are done here instead of in the backend because the int64 lowering 1372 # pass will make a mess of the patterns. The first patterns are 1373 # conditioned on nir_lower_minmax64 because it was not clear that it was 1374 # always an improvement on platforms that have real int64 support. No 1375 # shaders in shader-db hit this, so it was hard to say one way or the 1376 # other. 1377 (('ilt', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'), 1378 (('ilt', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'), 1379 (('ige', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'), 1380 (('ige', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'), 1381 (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'), 1382 (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'), 1383 1384 (('ine', 'a@64', 0), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'), 1385 (('ieq', 'a@64', 0), ('ieq', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'), 1386 # 0u < uint(a) <=> uint(a) != 0u 1387 (('ult', 0, 'a@64'), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'), 1388 1389 # Alternative lowering that doesn't rely on bfi. 1390 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'), 1391 ('bcsel', ('ult', 31, 'bits'), 1392 'insert', 1393 (('ior', 1394 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))), 1395 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))), 1396 'options->lower_bitfield_insert_to_shifts'), 1397 1398 # Alternative lowering that uses bitfield_select. 1399 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'), 1400 ('bcsel', ('ult', 31, 'bits'), 'insert', 1401 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')), 1402 'options->lower_bitfield_insert_to_bitfield_select'), 1403 1404 (('ibitfield_extract', 'value', 'offset', 'bits'), 1405 ('bcsel', ('ult', 31, 'bits'), 'value', 1406 ('ibfe', 'value', 'offset', 'bits')), 1407 'options->lower_bitfield_extract'), 1408 1409 (('ubitfield_extract', 'value', 'offset', 'bits'), 1410 ('bcsel', ('ult', 31, 'bits'), 'value', 1411 ('ubfe', 'value', 'offset', 'bits')), 1412 'options->lower_bitfield_extract'), 1413 1414 # (src0 & src1) | (~src0 & src2). Constant fold if src2 is 0. 1415 (('bitfield_select', a, b, 0), ('iand', a, b)), 1416 1417 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits' 1418 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')), 1419 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')), 1420 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')), 1421 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')), 1422 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')), 1423 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')), 1424 1425 # Section 8.8 (Integer Functions) of the GLSL 4.60 spec says: 1426 # 1427 # If bits is zero, the result will be zero. 1428 # 1429 # These patterns prevent other patterns from generating invalid results 1430 # when count is zero. 1431 (('ubfe', a, b, 0), 0), 1432 (('ibfe', a, b, 0), 0), 1433 1434 (('ubfe', a, 0, '#b'), ('iand', a, ('ushr', 0xffffffff, ('ineg', b)))), 1435 1436 (('b2i32', ('i2b', ('ubfe', a, b, 1))), ('ubfe', a, b, 1)), 1437 (('b2i32', ('i2b', ('ibfe', a, b, 1))), ('ubfe', a, b, 1)), # ubfe in the replacement is correct 1438 (('ine', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)), 1439 (('ieq', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)), 1440 (('ine', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)), 1441 (('ieq', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)), 1442 1443 (('ibitfield_extract', 'value', 'offset', 'bits'), 1444 ('bcsel', ('ieq', 0, 'bits'), 1445 0, 1446 ('ishr', 1447 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')), 1448 ('isub', 32, 'bits'))), 1449 'options->lower_bitfield_extract_to_shifts'), 1450 1451 (('ubitfield_extract', 'value', 'offset', 'bits'), 1452 ('iand', 1453 ('ushr', 'value', 'offset'), 1454 ('bcsel', ('ieq', 'bits', 32), 1455 0xffffffff, 1456 ('isub', ('ishl', 1, 'bits'), 1))), 1457 'options->lower_bitfield_extract_to_shifts'), 1458 1459 (('ifind_msb', 'value'), 1460 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')), 1461 'options->lower_ifind_msb'), 1462 1463 (('find_lsb', 'value'), 1464 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))), 1465 'options->lower_find_lsb'), 1466 1467 (('extract_i8', a, 'b@32'), 1468 ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24), 1469 'options->lower_extract_byte'), 1470 1471 (('extract_u8', a, 'b@32'), 1472 ('iand', ('ushr', a, ('imul', b, 8)), 0xff), 1473 'options->lower_extract_byte'), 1474 1475 (('extract_i16', a, 'b@32'), 1476 ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16), 1477 'options->lower_extract_word'), 1478 1479 (('extract_u16', a, 'b@32'), 1480 ('iand', ('ushr', a, ('imul', b, 16)), 0xffff), 1481 'options->lower_extract_word'), 1482 1483 (('pack_unorm_2x16', 'v'), 1484 ('pack_uvec2_to_uint', 1485 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))), 1486 'options->lower_pack_unorm_2x16'), 1487 1488 (('pack_unorm_4x8', 'v'), 1489 ('pack_uvec4_to_uint', 1490 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))), 1491 'options->lower_pack_unorm_4x8'), 1492 1493 (('pack_snorm_2x16', 'v'), 1494 ('pack_uvec2_to_uint', 1495 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))), 1496 'options->lower_pack_snorm_2x16'), 1497 1498 (('pack_snorm_4x8', 'v'), 1499 ('pack_uvec4_to_uint', 1500 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))), 1501 'options->lower_pack_snorm_4x8'), 1502 1503 (('unpack_unorm_2x16', 'v'), 1504 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0), 1505 ('extract_u16', 'v', 1))), 1506 65535.0), 1507 'options->lower_unpack_unorm_2x16'), 1508 1509 (('unpack_unorm_4x8', 'v'), 1510 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0), 1511 ('extract_u8', 'v', 1), 1512 ('extract_u8', 'v', 2), 1513 ('extract_u8', 'v', 3))), 1514 255.0), 1515 'options->lower_unpack_unorm_4x8'), 1516 1517 (('unpack_snorm_2x16', 'v'), 1518 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0), 1519 ('extract_i16', 'v', 1))), 1520 32767.0))), 1521 'options->lower_unpack_snorm_2x16'), 1522 1523 (('unpack_snorm_4x8', 'v'), 1524 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0), 1525 ('extract_i8', 'v', 1), 1526 ('extract_i8', 'v', 2), 1527 ('extract_i8', 'v', 3))), 1528 127.0))), 1529 'options->lower_unpack_snorm_4x8'), 1530 1531 (('pack_half_2x16_split', 'a@32', 'b@32'), 1532 ('ior', ('ishl', ('u2u32', ('f2f16', b)), 16), ('u2u32', ('f2f16', a))), 1533 'options->lower_pack_split'), 1534 1535 (('unpack_half_2x16_split_x', 'a@32'), 1536 ('f2f32', ('u2u16', a)), 1537 'options->lower_pack_split'), 1538 1539 (('unpack_half_2x16_split_y', 'a@32'), 1540 ('f2f32', ('u2u16', ('ushr', a, 16))), 1541 'options->lower_pack_split'), 1542 1543 (('pack_32_2x16_split', 'a@16', 'b@16'), 1544 ('ior', ('ishl', ('u2u32', b), 16), ('u2u32', a)), 1545 'options->lower_pack_split'), 1546 1547 (('unpack_32_2x16_split_x', 'a@32'), 1548 ('u2u16', a), 1549 'options->lower_pack_split'), 1550 1551 (('unpack_32_2x16_split_y', 'a@32'), 1552 ('u2u16', ('ushr', 'a', 16)), 1553 'options->lower_pack_split'), 1554 1555 (('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'), 1556 (('imin', ('imax', a, -1), 1), ('isign', a), '!options->lower_isign'), 1557 (('imax', ('imin', a, 1), -1), ('isign', a), '!options->lower_isign'), 1558 (('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'), 1559 1560 # Address/offset calculations: 1561 # Drivers supporting imul24 should use the nir_lower_amul() pass, this 1562 # rule converts everyone else to imul: 1563 (('amul', a, b), ('imul', a, b), '!options->has_imul24'), 1564 1565 (('umul24', a, b), 1566 ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)), 1567 '!options->has_umul24'), 1568 (('umad24', a, b, c), 1569 ('iadd', ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)), c), 1570 '!options->has_umad24'), 1571 1572 (('imad24_ir3', a, b, 0), ('imul24', a, b)), 1573 (('imad24_ir3', a, 0, c), (c)), 1574 (('imad24_ir3', a, 1, c), ('iadd', a, c)), 1575 1576 # if first two srcs are const, crack apart the imad so constant folding 1577 # can clean up the imul: 1578 # TODO ffma should probably get a similar rule: 1579 (('imad24_ir3', '#a', '#b', c), ('iadd', ('imul', a, b), c)), 1580 1581 # These will turn 24b address/offset calc back into 32b shifts, but 1582 # it should be safe to get back some of the bits of precision that we 1583 # already decided were no necessary: 1584 (('imul24', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'), 1585 (('imul24', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'), 1586 (('imul24', a, 0), (0)), 1587]) 1588 1589# bit_size dependent lowerings 1590for bit_size in [8, 16, 32, 64]: 1591 # convenience constants 1592 intmax = (1 << (bit_size - 1)) - 1 1593 intmin = 1 << (bit_size - 1) 1594 1595 optimizations += [ 1596 (('iadd_sat@' + str(bit_size), a, b), 1597 ('bcsel', ('ige', b, 1), ('bcsel', ('ilt', ('iadd', a, b), a), intmax, ('iadd', a, b)), 1598 ('bcsel', ('ilt', a, ('iadd', a, b)), intmin, ('iadd', a, b))), 'options->lower_add_sat'), 1599 (('isub_sat@' + str(bit_size), a, b), 1600 ('bcsel', ('ilt', b, 0), ('bcsel', ('ilt', ('isub', a, b), a), intmax, ('isub', a, b)), 1601 ('bcsel', ('ilt', a, ('isub', a, b)), intmin, ('isub', a, b))), 'options->lower_add_sat'), 1602 ] 1603 1604invert = OrderedDict([('feq', 'fneu'), ('fneu', 'feq')]) 1605 1606for left, right in itertools.combinations_with_replacement(invert.keys(), 2): 1607 optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))), 1608 ('iand', (invert[left], a, b), (invert[right], c, d)))) 1609 optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))), 1610 ('ior', (invert[left], a, b), (invert[right], c, d)))) 1611 1612# Optimize x2bN(b2x(x)) -> x 1613for size in type_sizes('bool'): 1614 aN = 'a@' + str(size) 1615 f2bN = 'f2b' + str(size) 1616 i2bN = 'i2b' + str(size) 1617 optimizations.append(((f2bN, ('b2f', aN)), a)) 1618 optimizations.append(((i2bN, ('b2i', aN)), a)) 1619 1620# Optimize x2yN(b2x(x)) -> b2y 1621for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']): 1622 if x != 'f' and y != 'f' and x != y: 1623 continue 1624 1625 b2x = 'b2f' if x == 'f' else 'b2i' 1626 b2y = 'b2f' if y == 'f' else 'b2i' 1627 x2yN = '{}2{}'.format(x, y) 1628 optimizations.append(((x2yN, (b2x, a)), (b2y, a))) 1629 1630# Optimize away x2xN(a@N) 1631for t in ['int', 'uint', 'float', 'bool']: 1632 for N in type_sizes(t): 1633 x2xN = '{0}2{0}{1}'.format(t[0], N) 1634 aN = 'a@{0}'.format(N) 1635 optimizations.append(((x2xN, aN), a)) 1636 1637# Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers 1638# In particular, we can optimize away everything except upcast of downcast and 1639# upcasts where the type differs from the other cast 1640for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')): 1641 if N < M: 1642 # The outer cast is a down-cast. It doesn't matter what the size of the 1643 # argument of the inner cast is because we'll never been in the upcast 1644 # of downcast case. Regardless of types, we'll always end up with y2yN 1645 # in the end. 1646 for x, y in itertools.product(['i', 'u'], ['i', 'u']): 1647 x2xN = '{0}2{0}{1}'.format(x, N) 1648 y2yM = '{0}2{0}{1}'.format(y, M) 1649 y2yN = '{0}2{0}{1}'.format(y, N) 1650 optimizations.append(((x2xN, (y2yM, a)), (y2yN, a))) 1651 elif N > M: 1652 # If the outer cast is an up-cast, we have to be more careful about the 1653 # size of the argument of the inner cast and with types. In this case, 1654 # the type is always the type of type up-cast which is given by the 1655 # outer cast. 1656 for P in type_sizes('uint'): 1657 # We can't optimize away up-cast of down-cast. 1658 if M < P: 1659 continue 1660 1661 # Because we're doing down-cast of down-cast, the types always have 1662 # to match between the two casts 1663 for x in ['i', 'u']: 1664 x2xN = '{0}2{0}{1}'.format(x, N) 1665 x2xM = '{0}2{0}{1}'.format(x, M) 1666 aP = 'a@{0}'.format(P) 1667 optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a))) 1668 else: 1669 # The N == M case is handled by other optimizations 1670 pass 1671 1672# Downcast operations should be able to see through pack 1673for t in ['i', 'u']: 1674 for N in [8, 16, 32]: 1675 x2xN = '{0}2{0}{1}'.format(t, N) 1676 optimizations += [ 1677 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)), 1678 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)), 1679 ] 1680 1681# Optimize comparisons with up-casts 1682for t in ['int', 'uint', 'float']: 1683 for N, M in itertools.product(type_sizes(t), repeat=2): 1684 if N == 1 or N >= M: 1685 continue 1686 1687 cond = 'true' 1688 if N == 8: 1689 cond = 'options->support_8bit_alu' 1690 elif N == 16: 1691 cond = 'options->support_16bit_alu' 1692 x2xM = '{0}2{0}{1}'.format(t[0], M) 1693 x2xN = '{0}2{0}{1}'.format(t[0], N) 1694 aN = 'a@' + str(N) 1695 bN = 'b@' + str(N) 1696 xeq = 'feq' if t == 'float' else 'ieq' 1697 xne = 'fneu' if t == 'float' else 'ine' 1698 xge = '{0}ge'.format(t[0]) 1699 xlt = '{0}lt'.format(t[0]) 1700 1701 # Up-casts are lossless so for correctly signed comparisons of 1702 # up-casted values we can do the comparison at the largest of the two 1703 # original sizes and drop one or both of the casts. (We have 1704 # optimizations to drop the no-op casts which this may generate.) 1705 for P in type_sizes(t): 1706 if P == 1 or P > N: 1707 continue 1708 1709 bP = 'b@' + str(P) 1710 optimizations += [ 1711 ((xeq, (x2xM, aN), (x2xM, bP)), (xeq, a, (x2xN, b)), cond), 1712 ((xne, (x2xM, aN), (x2xM, bP)), (xne, a, (x2xN, b)), cond), 1713 ((xge, (x2xM, aN), (x2xM, bP)), (xge, a, (x2xN, b)), cond), 1714 ((xlt, (x2xM, aN), (x2xM, bP)), (xlt, a, (x2xN, b)), cond), 1715 ((xge, (x2xM, bP), (x2xM, aN)), (xge, (x2xN, b), a), cond), 1716 ((xlt, (x2xM, bP), (x2xM, aN)), (xlt, (x2xN, b), a), cond), 1717 ] 1718 1719 # The next bit doesn't work on floats because the range checks would 1720 # get way too complicated. 1721 if t in ['int', 'uint']: 1722 if t == 'int': 1723 xN_min = -(1 << (N - 1)) 1724 xN_max = (1 << (N - 1)) - 1 1725 elif t == 'uint': 1726 xN_min = 0 1727 xN_max = (1 << N) - 1 1728 else: 1729 assert False 1730 1731 # If we're up-casting and comparing to a constant, we can unfold 1732 # the comparison into a comparison with the shrunk down constant 1733 # and a check that the constant fits in the smaller bit size. 1734 optimizations += [ 1735 ((xeq, (x2xM, aN), '#b'), 1736 ('iand', (xeq, a, (x2xN, b)), (xeq, (x2xM, (x2xN, b)), b)), cond), 1737 ((xne, (x2xM, aN), '#b'), 1738 ('ior', (xne, a, (x2xN, b)), (xne, (x2xM, (x2xN, b)), b)), cond), 1739 ((xlt, (x2xM, aN), '#b'), 1740 ('iand', (xlt, xN_min, b), 1741 ('ior', (xlt, xN_max, b), (xlt, a, (x2xN, b)))), cond), 1742 ((xlt, '#a', (x2xM, bN)), 1743 ('iand', (xlt, a, xN_max), 1744 ('ior', (xlt, a, xN_min), (xlt, (x2xN, a), b))), cond), 1745 ((xge, (x2xM, aN), '#b'), 1746 ('iand', (xge, xN_max, b), 1747 ('ior', (xge, xN_min, b), (xge, a, (x2xN, b)))), cond), 1748 ((xge, '#a', (x2xM, bN)), 1749 ('iand', (xge, a, xN_min), 1750 ('ior', (xge, a, xN_max), (xge, (x2xN, a), b))), cond), 1751 ] 1752 1753# Convert masking followed by signed downcast to just unsigned downcast 1754optimizations += [ 1755 (('i2i32', ('iand', 'a@64', 0xffffffff)), ('u2u32', a)), 1756 (('i2i16', ('iand', 'a@32', 0xffff)), ('u2u16', a)), 1757 (('i2i16', ('iand', 'a@64', 0xffff)), ('u2u16', a)), 1758 (('i2i8', ('iand', 'a@16', 0xff)), ('u2u8', a)), 1759 (('i2i8', ('iand', 'a@32', 0xff)), ('u2u8', a)), 1760 (('i2i8', ('iand', 'a@64', 0xff)), ('u2u8', a)), 1761] 1762 1763def fexp2i(exp, bits): 1764 # Generate an expression which constructs value 2.0^exp or 0.0. 1765 # 1766 # We assume that exp is already in a valid range: 1767 # 1768 # * [-15, 15] for 16-bit float 1769 # * [-127, 127] for 32-bit float 1770 # * [-1023, 1023] for 16-bit float 1771 # 1772 # If exp is the lowest value in the valid range, a value of 0.0 is 1773 # constructed. Otherwise, the value 2.0^exp is constructed. 1774 if bits == 16: 1775 return ('i2i16', ('ishl', ('iadd', exp, 15), 10)) 1776 elif bits == 32: 1777 return ('ishl', ('iadd', exp, 127), 23) 1778 elif bits == 64: 1779 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20)) 1780 else: 1781 assert False 1782 1783def ldexp(f, exp, bits): 1784 # The maximum possible range for a normal exponent is [-126, 127] and, 1785 # throwing in denormals, you get a maximum range of [-149, 127]. This 1786 # means that we can potentially have a swing of +-276. If you start with 1787 # FLT_MAX, you actually have to do ldexp(FLT_MAX, -278) to get it to flush 1788 # all the way to zero. The GLSL spec only requires that we handle a subset 1789 # of this range. From version 4.60 of the spec: 1790 # 1791 # "If exp is greater than +128 (single-precision) or +1024 1792 # (double-precision), the value returned is undefined. If exp is less 1793 # than -126 (single-precision) or -1022 (double-precision), the value 1794 # returned may be flushed to zero. Additionally, splitting the value 1795 # into a significand and exponent using frexp() and then reconstructing 1796 # a floating-point value using ldexp() should yield the original input 1797 # for zero and all finite non-denormalized values." 1798 # 1799 # The SPIR-V spec has similar language. 1800 # 1801 # In order to handle the maximum value +128 using the fexp2i() helper 1802 # above, we have to split the exponent in half and do two multiply 1803 # operations. 1804 # 1805 # First, we clamp exp to a reasonable range. Specifically, we clamp to 1806 # twice the full range that is valid for the fexp2i() function above. If 1807 # exp/2 is the bottom value of that range, the fexp2i() expression will 1808 # yield 0.0f which, when multiplied by f, will flush it to zero which is 1809 # allowed by the GLSL and SPIR-V specs for low exponent values. If the 1810 # value is clamped from above, then it must have been above the supported 1811 # range of the GLSL built-in and therefore any return value is acceptable. 1812 if bits == 16: 1813 exp = ('imin', ('imax', exp, -30), 30) 1814 elif bits == 32: 1815 exp = ('imin', ('imax', exp, -254), 254) 1816 elif bits == 64: 1817 exp = ('imin', ('imax', exp, -2046), 2046) 1818 else: 1819 assert False 1820 1821 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2. 1822 # (We use ishr which isn't the same for -1, but the -1 case still works 1823 # since we use exp-exp/2 as the second exponent.) While the spec 1824 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't 1825 # work with denormals and doesn't allow for the full swing in exponents 1826 # that you can get with normalized values. Instead, we create two powers 1827 # of two and multiply by them each in turn. That way the effective range 1828 # of our exponent is doubled. 1829 pow2_1 = fexp2i(('ishr', exp, 1), bits) 1830 pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits) 1831 return ('fmul', ('fmul', f, pow2_1), pow2_2) 1832 1833optimizations += [ 1834 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'), 1835 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'), 1836 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'), 1837] 1838 1839# Unreal Engine 4 demo applications open-codes bitfieldReverse() 1840def bitfield_reverse(u): 1841 step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16)) 1842 step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8)) 1843 step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4)) 1844 step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2)) 1845 step5 = ('ior(many-comm-expr)', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1)) 1846 1847 return step5 1848 1849optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')] 1850 1851# "all_equal(eq(a, b), vec(~0))" is the same as "all_equal(a, b)" 1852# "any_nequal(neq(a, b), vec(0))" is the same as "any_nequal(a, b)" 1853for ncomp in [2, 3, 4, 8, 16]: 1854 optimizations += [ 1855 (('ball_iequal' + str(ncomp), ('ieq', a, b), ~0), ('ball_iequal' + str(ncomp), a, b)), 1856 (('ball_iequal' + str(ncomp), ('feq', a, b), ~0), ('ball_fequal' + str(ncomp), a, b)), 1857 (('bany_inequal' + str(ncomp), ('ine', a, b), 0), ('bany_inequal' + str(ncomp), a, b)), 1858 (('bany_inequal' + str(ncomp), ('fneu', a, b), 0), ('bany_fnequal' + str(ncomp), a, b)), 1859 ] 1860 1861# For any float comparison operation, "cmp", if you have "a == a && a cmp b" 1862# then the "a == a" is redundant because it's equivalent to "a is not NaN" 1863# and, if a is a NaN then the second comparison will fail anyway. 1864for op in ['flt', 'fge', 'feq']: 1865 optimizations += [ 1866 (('iand', ('feq', a, a), (op, a, b)), ('!' + op, a, b)), 1867 (('iand', ('feq', a, a), (op, b, a)), ('!' + op, b, a)), 1868 ] 1869 1870# Add optimizations to handle the case where the result of a ternary is 1871# compared to a constant. This way we can take things like 1872# 1873# (a ? 0 : 1) > 0 1874# 1875# and turn it into 1876# 1877# a ? (0 > 0) : (1 > 0) 1878# 1879# which constant folding will eat for lunch. The resulting ternary will 1880# further get cleaned up by the boolean reductions above and we will be 1881# left with just the original variable "a". 1882for op in ['flt', 'fge', 'feq', 'fneu', 1883 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']: 1884 optimizations += [ 1885 ((op, ('bcsel', 'a', '#b', '#c'), '#d'), 1886 ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))), 1887 ((op, '#d', ('bcsel', a, '#b', '#c')), 1888 ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))), 1889 ] 1890 1891 1892# For example, this converts things like 1893# 1894# 1 + mix(0, a - 1, condition) 1895# 1896# into 1897# 1898# mix(1, (a-1)+1, condition) 1899# 1900# Other optimizations will rearrange the constants. 1901for op in ['fadd', 'fmul', 'iadd', 'imul']: 1902 optimizations += [ 1903 ((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d))) 1904 ] 1905 1906# For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives 1907# states: 1908# 1909# If neither layout qualifier is specified, derivatives in compute shaders 1910# return zero, which is consistent with the handling of built-in texture 1911# functions like texture() in GLSL 4.50 compute shaders. 1912for op in ['fddx', 'fddx_fine', 'fddx_coarse', 1913 'fddy', 'fddy_fine', 'fddy_coarse']: 1914 optimizations += [ 1915 ((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE') 1916] 1917 1918# Some optimizations for ir3-specific instructions. 1919optimizations += [ 1920 # 'al * bl': If either 'al' or 'bl' is zero, return zero. 1921 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)), 1922 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'. 1923 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')), 1924 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')), 1925] 1926 1927# These kinds of sequences can occur after nir_opt_peephole_select. 1928# 1929# NOTE: fadd is not handled here because that gets in the way of ffma 1930# generation in the i965 driver. Instead, fadd and ffma are handled in 1931# late_optimizations. 1932 1933for op in ['flrp']: 1934 optimizations += [ 1935 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, c, e)), (op, b, c, ('bcsel', a, d, e))), 1936 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, c, e)), (op, b, c, ('bcsel', a, d, e))), 1937 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, e, d)), (op, b, ('bcsel', a, c, e), d)), 1938 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)), 1939 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, e, c, d)), (op, ('bcsel', a, b, e), c, d)), 1940 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', e, c, d)), (op, ('bcsel', a, b, e), c, d)), 1941 ] 1942 1943for op in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']: 1944 optimizations += [ 1945 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))), 1946 (('bcsel', a, (op + '(is_used_once)', b, 'c(is_not_const)'), (op, b, d)), (op, b, ('bcsel', a, c, d))), 1947 (('bcsel', a, (op, b, 'c(is_not_const)'), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))), 1948 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))), 1949 ] 1950 1951for op in ['fpow']: 1952 optimizations += [ 1953 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, d)), (op, b, ('bcsel', a, c, d))), 1954 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))), 1955 (('bcsel', a, (op + '(is_used_once)', b, c), (op, d, c)), (op, ('bcsel', a, b, d), c)), 1956 (('bcsel', a, (op, b, c), (op + '(is_used_once)', d, c)), (op, ('bcsel', a, b, d), c)), 1957 ] 1958 1959for op in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos', 'fneg', 'fabs', 'fsign']: 1960 optimizations += [ 1961 (('bcsel', c, (op + '(is_used_once)', a), (op + '(is_used_once)', b)), (op, ('bcsel', c, a, b))), 1962 ] 1963 1964for op in ['ineg', 'iabs', 'inot', 'isign']: 1965 optimizations += [ 1966 ((op, ('bcsel', c, '#a', '#b')), ('bcsel', c, (op, a), (op, b))), 1967 ] 1968 1969# This section contains optimizations to propagate downsizing conversions of 1970# constructed vectors into vectors of downsized components. Whether this is 1971# useful depends on the SIMD semantics of the backend. On a true SIMD machine, 1972# this reduces the register pressure of the vector itself and often enables the 1973# conversions to be eliminated via other algebraic rules or constant folding. 1974# In the worst case on a SIMD architecture, the propagated conversions may be 1975# revectorized via nir_opt_vectorize so instruction count is minimally 1976# impacted. 1977# 1978# On a machine with SIMD-within-a-register only, this actually 1979# counterintuitively hurts instruction count. These machines are the same that 1980# require vectorize_vec2_16bit, so we predicate the optimizations on that flag 1981# not being set. 1982# 1983# Finally for scalar architectures, there should be no difference in generated 1984# code since it all ends up scalarized at the end, but it might minimally help 1985# compile-times. 1986 1987for i in range(2, 4 + 1): 1988 for T in ('f', 'u', 'i'): 1989 vec_inst = ('vec' + str(i),) 1990 1991 indices = ['a', 'b', 'c', 'd'] 1992 suffix_in = tuple((indices[j] + '@32') for j in range(i)) 1993 1994 to_16 = '{}2{}16'.format(T, T) 1995 to_mp = '{}2{}mp'.format(T, T) 1996 1997 out_16 = tuple((to_16, indices[j]) for j in range(i)) 1998 out_mp = tuple((to_mp, indices[j]) for j in range(i)) 1999 2000 optimizations += [ 2001 ((to_16, vec_inst + suffix_in), vec_inst + out_16, '!options->vectorize_vec2_16bit'), 2002 ] 2003 # u2ump doesn't exist, because it's equal to i2imp 2004 if T in ['f', 'i']: 2005 optimizations += [ 2006 ((to_mp, vec_inst + suffix_in), vec_inst + out_mp, '!options->vectorize_vec2_16bit') 2007 ] 2008 2009# This section contains "late" optimizations that should be run before 2010# creating ffmas and calling regular optimizations for the final time. 2011# Optimizations should go here if they help code generation and conflict 2012# with the regular optimizations. 2013before_ffma_optimizations = [ 2014 # Propagate constants down multiplication chains 2015 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)), 2016 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)), 2017 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)), 2018 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)), 2019 2020 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))), 2021 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))), 2022 (('~fadd', ('fneg', a), a), 0.0), 2023 (('iadd', ('ineg', a), a), 0), 2024 (('iadd', ('ineg', a), ('iadd', a, b)), b), 2025 (('iadd', a, ('iadd', ('ineg', a), b)), b), 2026 (('~fadd', ('fneg', a), ('fadd', a, b)), b), 2027 (('~fadd', a, ('fadd', ('fneg', a), b)), b), 2028 2029 (('~flrp', ('fadd(is_used_once)', a, -1.0), ('fadd(is_used_once)', a, 1.0), d), ('fadd', ('flrp', -1.0, 1.0, d), a)), 2030 (('~flrp', ('fadd(is_used_once)', a, 1.0), ('fadd(is_used_once)', a, -1.0), d), ('fadd', ('flrp', 1.0, -1.0, d), a)), 2031 (('~flrp', ('fadd(is_used_once)', a, '#b'), ('fadd(is_used_once)', a, '#c'), d), ('fadd', ('fmul', d, ('fadd', c, ('fneg', b))), ('fadd', a, b))), 2032] 2033 2034# This section contains "late" optimizations that should be run after the 2035# regular optimizations have finished. Optimizations should go here if 2036# they help code generation but do not necessarily produce code that is 2037# more easily optimizable. 2038late_optimizations = [ 2039 # Most of these optimizations aren't quite safe when you get infinity or 2040 # Nan involved but the first one should be fine. 2041 (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))), 2042 (('flt', ('fneg', ('fadd', a, b)), 0.0), ('flt', ('fneg', a), b)), 2043 (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))), 2044 (('~fge', ('fneg', ('fadd', a, b)), 0.0), ('fge', ('fneg', a), b)), 2045 (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))), 2046 (('~fneu', ('fadd', a, b), 0.0), ('fneu', a, ('fneg', b))), 2047 2048 # nir_lower_to_source_mods will collapse this, but its existence during the 2049 # optimization loop can prevent other optimizations. 2050 (('fneg', ('fneg', a)), a), 2051 2052 # Subtractions get lowered during optimization, so we need to recombine them 2053 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'), 2054 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'), 2055 (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'), 2056 (('ineg', a), ('isub', 0, a), 'options->lower_negate'), 2057 (('iabs', a), ('imax', a, ('ineg', a)), 'options->lower_iabs'), 2058 (('~fadd@16', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma16'), 2059 (('~fadd@32', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma32'), 2060 (('~fadd@64', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma64'), 2061 2062 # These are duplicated from the main optimizations table. The late 2063 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create 2064 # new patterns like these. The patterns that compare with zero are removed 2065 # because they are unlikely to be created in by anything in 2066 # late_optimizations. 2067 (('flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)), 2068 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)), 2069 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)), 2070 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)), 2071 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)), 2072 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)), 2073 2074 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)), 2075 # flt(fsat(a), 1.0) is inexact because it returns True if a is NaN 2076 # (fsat(NaN) is 0), while flt(a, 1.0) always returns FALSE. 2077 (('~flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)), 2078 2079 (('~fge', ('fmin(is_used_once)', ('fadd(is_used_once)', a, b), ('fadd', c, d)), 0.0), ('iand', ('fge', a, ('fneg', b)), ('fge', c, ('fneg', d)))), 2080 2081 (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)), 2082 (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)), 2083 (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)), 2084 (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)), 2085 (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)), 2086 (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)), 2087 (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)), 2088 (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)), 2089 (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)), 2090 (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)), 2091 2092 (('ior', a, a), a), 2093 (('iand', a, a), a), 2094 2095 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))), 2096 2097 (('fdot2', a, b), ('fdot2_replicated', a, b), 'options->fdot_replicates'), 2098 (('fdot3', a, b), ('fdot3_replicated', a, b), 'options->fdot_replicates'), 2099 (('fdot4', a, b), ('fdot4_replicated', a, b), 'options->fdot_replicates'), 2100 (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'), 2101 2102 (('~flrp', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)), 2103 2104 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this 2105 # particular operation is common for expanding values stored in a texture 2106 # from [0,1] to [-1,1]. 2107 (('~ffma@32', a, 2.0, -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'), 2108 (('~ffma@32', a, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'), 2109 (('~ffma@32', a, -2.0, 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'), 2110 (('~ffma@32', a, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'), 2111 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'), 2112 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'), 2113 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'), 2114 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'), 2115 2116 # flrp(a, b, a) 2117 # a*(1-a) + b*a 2118 # a + -a*a + a*b (1) 2119 # a + a*(b - a) 2120 # Option 1: ffma(a, (b-a), a) 2121 # 2122 # Alternately, after (1): 2123 # a*(1+b) + -a*a 2124 # a*((1+b) + -a) 2125 # 2126 # Let b=1 2127 # 2128 # Option 2: ffma(a, 2, -(a*a)) 2129 # Option 3: ffma(a, 2, (-a)*a) 2130 # Option 4: ffma(a, -a, (2*a) 2131 # Option 5: a * (2 - a) 2132 # 2133 # There are a lot of other possible combinations. 2134 (('~ffma@32', ('fadd', b, ('fneg', a)), a, a), ('flrp', a, b, a), '!options->lower_flrp32'), 2135 (('~ffma@32', a, 2.0, ('fneg', ('fmul', a, a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'), 2136 (('~ffma@32', a, 2.0, ('fmul', ('fneg', a), a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'), 2137 (('~ffma@32', a, ('fneg', a), ('fmul', 2.0, a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'), 2138 (('~fmul@32', a, ('fadd', 2.0, ('fneg', a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'), 2139 2140 # we do these late so that we don't get in the way of creating ffmas 2141 (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))), 2142 (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))), 2143 2144 # Putting this in 'optimizations' interferes with the bcsel(a, op(b, c), 2145 # op(b, d)) => op(b, bcsel(a, c, d)) transformations. I do not know why. 2146 (('bcsel', ('feq', ('fsqrt', 'a(is_not_negative)'), 0.0), intBitsToFloat(0x7f7fffff), ('frsq', a)), 2147 ('fmin', ('frsq', a), intBitsToFloat(0x7f7fffff))), 2148 2149 # Things that look like DPH in the source shader may get expanded to 2150 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets 2151 # to NIR. After FFMA is generated, this can look like: 2152 # 2153 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w) 2154 # 2155 # Reassociate the last addition into the first multiplication. 2156 # 2157 # Some shaders do not use 'invariant' in vertex and (possibly) geometry 2158 # shader stages on some outputs that are intended to be invariant. For 2159 # various reasons, this optimization may not be fully applied in all 2160 # shaders used for different rendering passes of the same geometry. This 2161 # can result in Z-fighting artifacts (at best). For now, disable this 2162 # optimization in these stages. See bugzilla #111490. In tessellation 2163 # stages applications seem to use 'precise' when necessary, so allow the 2164 # optimization in those stages. 2165 (('~fadd', ('ffma(is_used_once)', a, b, ('ffma', c, d, ('fmul', 'e(is_not_const_and_not_fsign)', 'f(is_not_const_and_not_fsign)'))), 'g(is_not_const)'), 2166 ('ffma', a, b, ('ffma', c, d, ('ffma', e, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'), 2167 (('~fadd', ('ffma(is_used_once)', a, b, ('fmul', 'c(is_not_const_and_not_fsign)', 'd(is_not_const_and_not_fsign)') ), 'e(is_not_const)'), 2168 ('ffma', a, b, ('ffma', c, d, e)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'), 2169 2170 # Section 8.8 (Integer Functions) of the GLSL 4.60 spec says: 2171 # 2172 # If bits is zero, the result will be zero. 2173 # 2174 # These prevent the next two lowerings generating incorrect results when 2175 # count is zero. 2176 (('ubfe', a, b, 0), 0), 2177 (('ibfe', a, b, 0), 0), 2178 2179 # On Intel GPUs, BFE is a 3-source instruction. Like all 3-source 2180 # instructions on Intel GPUs, it cannot have an immediate values as 2181 # sources. There are also limitations on source register strides. As a 2182 # result, it is very easy for 3-source instruction combined with either 2183 # loads of immediate values or copies from weird register strides to be 2184 # more expensive than the primitive instructions it represents. 2185 (('ubfe', a, '#b', '#c'), ('iand', ('ushr', 0xffffffff, ('ineg', c)), ('ushr', a, b)), 'options->lower_bfe_with_two_constants'), 2186 2187 # b is the lowest order bit to be extracted and c is the number of bits to 2188 # extract. The inner shift removes the bits above b + c by shifting left 2189 # 32 - (b + c). ishl only sees the low 5 bits of the shift count, which is 2190 # -(b + c). The outer shift moves the bit that was at b to bit zero. 2191 # After the first shift, that bit is now at b + (32 - (b + c)) or 32 - c. 2192 # This means that it must be shifted right by 32 - c or -c bits. 2193 (('ibfe', a, '#b', '#c'), ('ishr', ('ishl', a, ('ineg', ('iadd', b, c))), ('ineg', c)), 'options->lower_bfe_with_two_constants'), 2194 2195 # Clean up no-op shifts that may result from the bfe lowerings. 2196 (('ishl', a, 0), a), 2197 (('ishl', a, -32), a), 2198 (('ishr', a, 0), a), 2199 (('ishr', a, -32), a), 2200 (('ushr', a, 0), a), 2201] 2202 2203# Integer sizes 2204for s in [8, 16, 32, 64]: 2205 late_optimizations.extend([ 2206 (('iand', ('ine(is_used_once)', 'a@{}'.format(s), 0), ('ine', 'b@{}'.format(s), 0)), ('ine', ('umin', a, b), 0)), 2207 (('ior', ('ieq(is_used_once)', 'a@{}'.format(s), 0), ('ieq', 'b@{}'.format(s), 0)), ('ieq', ('umin', a, b), 0)), 2208 ]) 2209 2210# Float sizes 2211for s in [16, 32, 64]: 2212 late_optimizations.extend([ 2213 (('~fadd@{}'.format(s), 1.0, ('fmul(is_used_once)', c , ('fadd', b, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c)), ('fmul', b, c)), 'options->lower_flrp{}'.format(s)), 2214 (('bcsel', a, 0, ('b2f{}'.format(s), ('inot', 'b@bool'))), ('b2f{}'.format(s), ('inot', ('ior', a, b)))), 2215 ]) 2216 2217for op in ['fadd']: 2218 late_optimizations += [ 2219 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, d)), (op, b, ('bcsel', a, c, d))), 2220 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))), 2221 ] 2222 2223for op in ['ffma']: 2224 late_optimizations += [ 2225 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, c, e)), (op, b, c, ('bcsel', a, d, e))), 2226 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, c, e)), (op, b, c, ('bcsel', a, d, e))), 2227 2228 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, e, d)), (op, b, ('bcsel', a, c, e), d)), 2229 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)), 2230 ] 2231 2232# mediump: If an opcode is surrounded by conversions, remove the conversions. 2233# The rationale is that type conversions + the low precision opcode are more 2234# expensive that the same arithmetic opcode at higher precision. 2235# 2236# This must be done in late optimizations, because we need normal optimizations to 2237# first eliminate temporary up-conversions such as in op1(f2fmp(f2f32(op2()))). 2238# 2239# Unary opcodes 2240for op in ['fabs', 'fceil', 'fcos', 'fddx', 'fddx_coarse', 'fddx_fine', 'fddy', 2241 'fddy_coarse', 'fddy_fine', 'fexp2', 'ffloor', 'ffract', 'flog2', 'fneg', 2242 'frcp', 'fround_even', 'frsq', 'fsat', 'fsign', 'fsin', 'fsqrt']: 2243 late_optimizations += [(('~f2f32', (op, ('f2fmp', a))), (op, a))] 2244 2245# Binary opcodes 2246for op in ['fadd', 'fdiv', 'fmax', 'fmin', 'fmod', 'fmul', 'fpow', 'frem']: 2247 late_optimizations += [(('~f2f32', (op, ('f2fmp', a), ('f2fmp', b))), (op, a, b))] 2248 2249# Ternary opcodes 2250for op in ['ffma', 'flrp']: 2251 late_optimizations += [(('~f2f32', (op, ('f2fmp', a), ('f2fmp', b), ('f2fmp', c))), (op, a, b, c))] 2252 2253# Comparison opcodes 2254for op in ['feq', 'fge', 'flt', 'fneu']: 2255 late_optimizations += [(('~' + op, ('f2fmp', a), ('f2fmp', b)), (op, a, b))] 2256 2257# Do this last, so that the f2fmp patterns above have effect. 2258late_optimizations += [ 2259 # Convert *2*mp instructions to concrete *2*16 instructions. At this point 2260 # any conversions that could have been removed will have been removed in 2261 # nir_opt_algebraic so any remaining ones are required. 2262 (('f2fmp', a), ('f2f16', a)), 2263 (('f2imp', a), ('f2i16', a)), 2264 (('f2ump', a), ('f2u16', a)), 2265 (('i2imp', a), ('i2i16', a)), 2266 (('i2fmp', a), ('i2f16', a)), 2267 (('i2imp', a), ('u2u16', a)), 2268 (('u2fmp', a), ('u2f16', a)), 2269] 2270 2271distribute_src_mods = [ 2272 # Try to remove some spurious negations rather than pushing them down. 2273 (('fmul', ('fneg', a), ('fneg', b)), ('fmul', a, b)), 2274 (('ffma', ('fneg', a), ('fneg', b), c), ('ffma', a, b, c)), 2275 (('fdot2_replicated', ('fneg', a), ('fneg', b)), ('fdot2_replicated', a, b)), 2276 (('fdot3_replicated', ('fneg', a), ('fneg', b)), ('fdot3_replicated', a, b)), 2277 (('fdot4_replicated', ('fneg', a), ('fneg', b)), ('fdot4_replicated', a, b)), 2278 (('fneg', ('fneg', a)), a), 2279 2280 (('fneg', ('fmul(is_used_once)', a, b)), ('fmul', ('fneg', a), b)), 2281 (('fabs', ('fmul(is_used_once)', a, b)), ('fmul', ('fabs', a), ('fabs', b))), 2282 2283 (('fneg', ('ffma(is_used_once)', a, b, c)), ('ffma', ('fneg', a), b, ('fneg', c))), 2284 (('fneg', ('flrp(is_used_once)', a, b, c)), ('flrp', ('fneg', a), ('fneg', b), c)), 2285 (('fneg', ('fadd(is_used_once)', a, b)), ('fadd', ('fneg', a), ('fneg', b))), 2286 2287 # Note that fmin <-> fmax. I don't think there is a way to distribute 2288 # fabs() into fmin or fmax. 2289 (('fneg', ('fmin(is_used_once)', a, b)), ('fmax', ('fneg', a), ('fneg', b))), 2290 (('fneg', ('fmax(is_used_once)', a, b)), ('fmin', ('fneg', a), ('fneg', b))), 2291 2292 (('fneg', ('fdot2_replicated(is_used_once)', a, b)), ('fdot2_replicated', ('fneg', a), b)), 2293 (('fneg', ('fdot3_replicated(is_used_once)', a, b)), ('fdot3_replicated', ('fneg', a), b)), 2294 (('fneg', ('fdot4_replicated(is_used_once)', a, b)), ('fdot4_replicated', ('fneg', a), b)), 2295 2296 # fdph works mostly like fdot, but to get the correct result, the negation 2297 # must be applied to the second source. 2298 (('fneg', ('fdph_replicated(is_used_once)', a, b)), ('fdph_replicated', a, ('fneg', b))), 2299 2300 (('fneg', ('fsign(is_used_once)', a)), ('fsign', ('fneg', a))), 2301 (('fabs', ('fsign(is_used_once)', a)), ('fsign', ('fabs', a))), 2302] 2303 2304print(nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render()) 2305print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_before_ffma", 2306 before_ffma_optimizations).render()) 2307print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_late", 2308 late_optimizations).render()) 2309print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_distribute_src_mods", 2310 distribute_src_mods).render()) 2311