2 # Copyright (C) 2014 Intel Corporation
4 # Permission is hereby granted, free of charge, to any person obtaining a
5 # copy of this software and associated documentation files (the "Software"),
6 # to deal in the Software without restriction, including without limitation
7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 # and/or sell copies of the Software, and to permit persons to whom the
9 # Software is furnished to do so, subject to the following conditions:
11 # The above copyright notice and this permission notice (including the next
12 # paragraph) shall be included in all copies or substantial portions of the
15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 # Jason Ekstrand (jason@jlekstrand.net)
26 from __future__
import print_function
28 from collections
import OrderedDict
30 from nir_opcodes
import type_sizes
33 # Convenience variables
40 # Written in the form (<search>, <replace>) where <search> is an expression
41 # and <replace> is either an expression or a value. An expression is
42 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
43 # where each source is either an expression or a value. A value can be
44 # either a numeric constant or a string representing a variable name.
46 # If the opcode in a search expression is prefixed by a '~' character, this
47 # indicates that the operation is inexact. Such operations will only get
48 # applied to SSA values that do not have the exact bit set. This should be
49 # used by by any optimizations that are not bit-for-bit exact. It should not,
50 # however, be used for backend-requested lowering operations as those need to
51 # happen regardless of precision.
53 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
54 # that the given variable will only match constants and the type indicates that
55 # the given variable will only match values from ALU instructions with the
56 # given output type, and (cond) specifies an additional condition function
57 # (see nir_search_helpers.h).
59 # For constants, you have to be careful to make sure that it is the right
60 # type because python is unaware of the source and destination types of the
63 # All expression types can have a bit-size specified. For opcodes, this
64 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
65 # type and size. In the search half of the expression this indicates that it
66 # should only match that particular bit-size. In the replace half of the
67 # expression this indicates that the constructed value should have that
72 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
73 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
74 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
75 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
76 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
77 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
82 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
83 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
84 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
85 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
87 (('fneg', ('fneg', a
)), a
),
88 (('ineg', ('ineg', a
)), a
),
89 (('fabs', ('fabs', a
)), ('fabs', a
)),
90 (('fabs', ('fneg', a
)), ('fabs', a
)),
91 (('fabs', ('u2f', a
)), ('u2f', a
)),
92 (('iabs', ('iabs', a
)), ('iabs', a
)),
93 (('iabs', ('ineg', a
)), ('iabs', a
)),
94 (('f2b', ('fneg', a
)), ('f2b', a
)),
95 (('i2b', ('ineg', a
)), ('i2b', a
)),
96 (('~fadd', a
, 0.0), a
),
98 (('usadd_4x8', a
, 0), a
),
99 (('usadd_4x8', a
, ~
0), ~
0),
100 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
101 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
102 (('~fadd', ('fneg', a
), a
), 0.0),
103 (('iadd', ('ineg', a
), a
), 0),
104 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
105 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
106 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
107 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
108 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
109 (('~fmul', a
, 0.0), 0.0),
111 (('umul_unorm_4x8', a
, 0), 0),
112 (('umul_unorm_4x8', a
, ~
0), a
),
113 (('fmul', a
, 1.0), a
),
115 (('fmul', a
, -1.0), ('fneg', a
)),
116 (('imul', a
, -1), ('ineg', a
)),
117 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
118 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
119 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
120 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
121 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
122 (('~ffma', 0.0, a
, b
), b
),
123 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
124 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
125 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
126 (('~flrp', a
, b
, 0.0), a
),
127 (('~flrp', a
, b
, 1.0), b
),
128 (('~flrp', a
, a
, b
), a
),
129 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
131 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
132 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
133 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
134 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
136 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
137 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
139 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
140 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
142 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
144 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
145 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
146 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
147 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
148 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
149 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
150 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
151 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
152 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
153 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
154 # These are the same as the previous three rules, but it depends on
155 # 1-fsat(x) <=> fsat(1-x):
157 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
158 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
159 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
160 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
161 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
163 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
164 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
165 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
166 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
167 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
169 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
170 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
172 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
173 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
174 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
175 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
177 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
178 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
180 # (a * #b + #c) << #d
181 # ((a * #b) << #d) + (#c << #d)
182 # (a * (#b << #d)) + (#c << #d)
183 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
184 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
188 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
190 # Comparison simplifications
191 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
192 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
193 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
194 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
195 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
196 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
197 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
198 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
199 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
200 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
204 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
206 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
208 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
210 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
211 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
212 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
213 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
214 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
215 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
216 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
217 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
218 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
219 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
220 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
221 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
222 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
223 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
224 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
225 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
226 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
227 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
228 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
229 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
231 # -(b2f(a) + b2f(b)) < 0
232 # 0 < b2f(a) + b2f(b)
233 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
235 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
236 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
238 # -(b2f(a) + b2f(b)) >= 0
239 # 0 >= b2f(a) + b2f(b)
240 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
242 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
243 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
245 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
246 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
248 # Some optimizations (below) convert things like (a < b || c < b) into
249 # (min(a, c) < b). However, this interfers with the previous optimizations
250 # that try to remove comparisons with negated sums of b2f. This just
252 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
253 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
255 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
256 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
257 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
258 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
260 # Cannot remove the addition from ilt or ige due to overflow.
261 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
262 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
264 # fmin(-b2f(a), b) >= 0.0
265 # -b2f(a) >= 0.0 && b >= 0.0
266 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
267 # b2f(a) == 0.0 && b >= 0.0
268 # a == False && b >= 0.0
271 # The fge in the second replacement is not a typo. I leave the proof that
272 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
274 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
275 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
277 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
278 (('fne', ('b2f', 'a@1'), 0.0), a
),
279 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
280 (('ine', ('b2i', 'a@1'), 0), a
),
282 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
283 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
284 (('fge', ('u2f', a
), 0.0), True),
285 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
286 (('flt', ('u2f', a
), 0.0), False),
287 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
288 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
289 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
290 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
291 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
292 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
293 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
297 # fabs(a) != 0.0 because fabs(a) must be >= 0
299 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
303 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
306 # 0.0 == fabs(a) because fabs(a) must be >= 0
308 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
312 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
314 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
315 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
316 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
317 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
320 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
321 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
322 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
324 # Since b is a constant, constant folding will eliminate the fmin and the
325 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
326 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
328 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
330 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
331 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
332 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
333 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
334 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
335 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
336 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
337 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
338 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
339 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
340 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
341 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
342 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
343 (('bcsel', a
, True, b
), ('ior', a
, b
)),
344 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
345 (('bcsel', a
, b
, False), ('iand', a
, b
)),
346 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
353 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
354 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
355 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
356 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
357 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
358 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
359 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
360 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
361 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
362 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
363 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
364 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
365 (('fmin', a
, ('fabs', a
)), a
),
366 (('imin', a
, ('iabs', a
)), a
),
367 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
368 (('imax', a
, ('ineg', ('iabs', a
))), a
),
369 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
370 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
371 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
372 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
373 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
374 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
375 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
376 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
377 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
378 (('fsat', ('b2f', a
)), ('b2f', a
)),
379 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
380 (('fsat', ('fsat', a
)), ('fsat', a
)),
381 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_negate && !options->lower_fsat'),
382 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_negate && !options->lower_fsat'),
383 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
384 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
385 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
386 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
387 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
388 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
389 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
390 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
391 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
392 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
393 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
394 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
395 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
396 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
397 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
398 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
399 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
400 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
401 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
402 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
403 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
404 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
405 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
407 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
408 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
409 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
410 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
411 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
412 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
413 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
414 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
415 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
416 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
417 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
418 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
419 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
420 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
421 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
422 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
424 # Common pattern like 'if (i == 0 || i == 1 || ...)'
425 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
426 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
427 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
429 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
430 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
431 # so emit an open-coded version of that.
432 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
433 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
435 (('ior', a
, ('ieq', a
, False)), True),
436 (('ior', a
, ('inot', a
)), -1),
438 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
439 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
441 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
443 # These patterns can result when (a < b || a < c) => (a < min(b, c))
444 # transformations occur before constant propagation and loop-unrolling.
445 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
446 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
447 (('~fge', a
, ('fmin', b
, a
)), True),
448 (('~fge', ('fmax', a
, b
), a
), True),
449 (('~flt', a
, ('fmin', b
, a
)), False),
450 (('~flt', ('fmax', a
, b
), a
), False),
451 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
452 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
454 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
455 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
456 (('ige', a
, ('imin', b
, a
)), True),
457 (('ige', ('imax', a
, b
), a
), True),
458 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
459 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
460 (('uge', a
, ('umin', b
, a
)), True),
461 (('uge', ('umax', a
, b
), a
), True),
462 (('ilt', a
, ('imin', b
, a
)), False),
463 (('ilt', ('imax', a
, b
), a
), False),
464 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
465 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
466 (('ult', a
, ('umin', b
, a
)), False),
467 (('ult', ('umax', a
, b
), a
), False),
468 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
469 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
471 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
472 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
473 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
474 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
475 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
476 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
477 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
478 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
479 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
480 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
481 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
482 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
483 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
484 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
485 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
486 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
488 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
490 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
491 ('iabs', ('ishr', a
, b
))),
492 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
494 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
495 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
496 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
497 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
498 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
499 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
500 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
501 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
502 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
503 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
505 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
506 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
507 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
508 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
509 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
510 (('ineg', ('b2i32', 'a@32')), a
),
511 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
512 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
513 # Comparison with the same args. Note that these are not done for
514 # the float versions because NaN always returns false on float
516 (('ilt', a
, a
), False),
517 (('ige', a
, a
), True),
518 (('ieq', a
, a
), True),
519 (('ine', a
, a
), False),
520 (('ult', a
, a
), False),
521 (('uge', a
, a
), True),
522 # Logical and bit operations
523 (('fand', a
, 0.0), 0.0),
525 (('iand', a
, ~
0), a
),
529 (('ior', a
, True), True),
530 (('fxor', a
, a
), 0.0),
533 (('inot', ('inot', a
)), a
),
534 (('ior', ('iand', a
, b
), b
), b
),
535 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
536 (('iand', ('ior', a
, b
), b
), b
),
537 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
539 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
540 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
541 # Shift optimizations
548 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
549 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
550 # Exponential/logarithmic identities
551 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
552 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
553 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
554 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
555 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
556 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
557 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
558 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
559 (('~fpow', a
, 1.0), a
),
560 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
561 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
562 (('~fpow', 2.0, a
), ('fexp2', a
)),
563 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
564 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
565 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
566 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
567 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
568 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
569 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
570 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
571 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
572 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
573 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
574 # Division and reciprocal
575 (('~fdiv', 1.0, a
), ('frcp', a
)),
576 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
577 (('~frcp', ('frcp', a
)), a
),
578 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
579 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
580 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
581 # Boolean simplifications
582 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
583 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
584 (('ieq', a
, True), a
),
585 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
586 (('ine', a
, False), a
),
587 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
588 (('bcsel', a
, True, False), a
),
589 (('bcsel', a
, False, True), ('inot', a
)),
590 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
591 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
592 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
593 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
594 (('bcsel', True, b
, c
), b
),
595 (('bcsel', False, b
, c
), c
),
596 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
597 # The result of this should be hit by constant propagation and, in the
598 # next round of opt_algebraic, get picked up by one of the above two.
599 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
601 (('bcsel', a
, b
, b
), b
),
602 (('fcsel', a
, b
, b
), b
),
604 # D3D Boolean emulation
605 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
606 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
607 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
608 ('ineg', ('b2i', ('iand', a
, b
)))),
609 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
610 ('ineg', ('b2i', ('ior', a
, b
)))),
611 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
612 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
613 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
614 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
615 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
617 # SM5 32-bit shifts are defined to use the 5 least significant bits
618 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
619 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
620 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
623 (('i2b32', ('b2i', 'a@32')), a
),
624 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
625 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
626 (('i2b', ('ineg', a
)), ('i2b', a
)),
627 (('i2b', ('iabs', a
)), ('i2b', a
)),
628 (('fabs', ('b2f', a
)), ('b2f', a
)),
629 (('iabs', ('b2i', a
)), ('b2i', a
)),
630 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
632 # Ironically, mark these as imprecise because removing the conversions may
633 # preserve more precision than doing the conversions (e.g.,
634 # uint(float(0x81818181u)) == 0x81818200).
635 (('~f2i32', ('i2f', 'a@32')), a
),
636 (('~f2i32', ('u2f', 'a@32')), a
),
637 (('~f2u32', ('i2f', 'a@32')), a
),
638 (('~f2u32', ('u2f', 'a@32')), a
),
640 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
643 # It is undefined to convert a negative floating-point value to an
646 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
647 # some optimizations in the i965 backend to proceed.
648 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
649 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
650 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
651 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
653 (('~fmin', ('fabs', a
), 1.0), ('fsat', ('fabs', a
)), '!options->lower_fsat'),
655 # The result of the multiply must be in [-1, 0], so the result of the ffma
657 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
658 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
659 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
660 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
662 # Packing and then unpacking does nothing
663 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
664 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
665 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
666 ('unpack_64_2x32_split_y', a
)), a
),
669 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
670 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
671 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
672 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
673 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
674 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
675 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
678 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
679 # patterns like those below.
680 for op
in ('ushr', 'ishr'):
681 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
682 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
683 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
685 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
687 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
688 # patterns like those below.
689 for op
in ('extract_u8', 'extract_i8'):
690 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
691 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
692 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
694 optimizations
.extend([
696 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
697 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
698 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
699 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
700 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
703 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
704 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
705 (('ussub_4x8', a
, 0), a
),
706 (('ussub_4x8', a
, ~
0), 0),
707 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
708 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
709 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
710 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
711 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
712 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
713 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
714 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
716 # Propagate negation up multiplication chains
717 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
718 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
720 # Propagate constants up multiplication chains
721 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
722 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
723 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
724 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
726 # Reassociate constants in add/mul chains so they can be folded together.
727 # For now, we mostly only handle cases where the constants are separated by
728 # a single non-constant. We could do better eventually.
729 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
730 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
731 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
732 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
733 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
736 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
737 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
738 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
740 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
741 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
742 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
744 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
747 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod16'),
748 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
749 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
750 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
751 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
752 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
754 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
755 ('bcsel', ('ilt', 31, 'bits'), 'insert',
756 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
757 'options->lower_bitfield_insert'),
758 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
759 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
760 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
761 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
762 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
763 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
765 # Alternative lowering that doesn't rely on bfi.
766 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
767 ('bcsel', ('ilt', 31, 'bits'),
770 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
771 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
772 'options->lower_bitfield_insert_to_shifts'),
774 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
775 (('bfm', 'bits', 'offset'),
776 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
777 'options->lower_bfm'),
779 (('ibitfield_extract', 'value', 'offset', 'bits'),
780 ('bcsel', ('ilt', 31, 'bits'), 'value',
781 ('ibfe', 'value', 'offset', 'bits')),
782 'options->lower_bitfield_extract'),
784 (('ubitfield_extract', 'value', 'offset', 'bits'),
785 ('bcsel', ('ult', 31, 'bits'), 'value',
786 ('ubfe', 'value', 'offset', 'bits')),
787 'options->lower_bitfield_extract'),
789 (('ibitfield_extract', 'value', 'offset', 'bits'),
790 ('bcsel', ('ieq', 0, 'bits'),
793 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
794 ('isub', 32, 'bits'))),
795 'options->lower_bitfield_extract_to_shifts'),
797 (('ubitfield_extract', 'value', 'offset', 'bits'),
799 ('ushr', 'value', 'offset'),
800 ('bcsel', ('ieq', 'bits', 32),
802 ('bfm', 'bits', 0))),
803 'options->lower_bitfield_extract_to_shifts'),
805 (('ifind_msb', 'value'),
806 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
807 'options->lower_ifind_msb'),
809 (('find_lsb', 'value'),
810 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
811 'options->lower_find_lsb'),
813 (('extract_i8', a
, 'b@32'),
814 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
815 'options->lower_extract_byte'),
817 (('extract_u8', a
, 'b@32'),
818 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
819 'options->lower_extract_byte'),
821 (('extract_i16', a
, 'b@32'),
822 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
823 'options->lower_extract_word'),
825 (('extract_u16', a
, 'b@32'),
826 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
827 'options->lower_extract_word'),
829 (('pack_unorm_2x16', 'v'),
830 ('pack_uvec2_to_uint',
831 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
832 'options->lower_pack_unorm_2x16'),
834 (('pack_unorm_4x8', 'v'),
835 ('pack_uvec4_to_uint',
836 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
837 'options->lower_pack_unorm_4x8'),
839 (('pack_snorm_2x16', 'v'),
840 ('pack_uvec2_to_uint',
841 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
842 'options->lower_pack_snorm_2x16'),
844 (('pack_snorm_4x8', 'v'),
845 ('pack_uvec4_to_uint',
846 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
847 'options->lower_pack_snorm_4x8'),
849 (('unpack_unorm_2x16', 'v'),
850 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
851 ('extract_u16', 'v', 1))),
853 'options->lower_unpack_unorm_2x16'),
855 (('unpack_unorm_4x8', 'v'),
856 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
857 ('extract_u8', 'v', 1),
858 ('extract_u8', 'v', 2),
859 ('extract_u8', 'v', 3))),
861 'options->lower_unpack_unorm_4x8'),
863 (('unpack_snorm_2x16', 'v'),
864 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
865 ('extract_i16', 'v', 1))),
867 'options->lower_unpack_snorm_2x16'),
869 (('unpack_snorm_4x8', 'v'),
870 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
871 ('extract_i8', 'v', 1),
872 ('extract_i8', 'v', 2),
873 ('extract_i8', 'v', 3))),
875 'options->lower_unpack_snorm_4x8'),
877 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
878 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
881 # bit_size dependent lowerings
882 for bit_size
in [8, 16, 32, 64]:
883 # convenience constants
884 intmax
= (1 << (bit_size
- 1)) - 1
885 intmin
= 1 << (bit_size
- 1)
888 (('iadd_sat@' + str(bit_size
), a
, b
),
889 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
890 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
891 (('isub_sat@' + str(bit_size
), a
, b
),
892 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
893 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
896 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
898 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
899 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
900 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
901 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
902 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
904 # Optimize x2bN(b2x(x)) -> x
905 for size
in type_sizes('bool'):
906 aN
= 'a@' + str(size
)
907 f2bN
= 'f2b' + str(size
)
908 i2bN
= 'i2b' + str(size
)
909 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
910 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
912 # Optimize x2yN(b2x(x)) -> b2y
913 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
914 if x
!= 'f' and y
!= 'f' and x
!= y
:
917 b2x
= 'b2f' if x
== 'f' else 'b2i'
918 b2y
= 'b2f' if y
== 'f' else 'b2i'
919 x2yN
= '{}2{}'.format(x
, y
)
920 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
922 # Optimize away x2xN(a@N)
923 for t
in ['int', 'uint', 'float']:
924 for N
in type_sizes(t
):
925 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
926 aN
= 'a@{0}'.format(N
)
927 optimizations
.append(((x2xN
, aN
), a
))
929 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
930 # In particular, we can optimize away everything except upcast of downcast and
931 # upcasts where the type differs from the other cast
932 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
934 # The outer cast is a down-cast. It doesn't matter what the size of the
935 # argument of the inner cast is because we'll never been in the upcast
936 # of downcast case. Regardless of types, we'll always end up with y2yN
938 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
939 x2xN
= '{0}2{0}{1}'.format(x
, N
)
940 y2yM
= '{0}2{0}{1}'.format(y
, M
)
941 y2yN
= '{0}2{0}{1}'.format(y
, N
)
942 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
944 # If the outer cast is an up-cast, we have to be more careful about the
945 # size of the argument of the inner cast and with types. In this case,
946 # the type is always the type of type up-cast which is given by the
948 for P
in type_sizes('uint'):
949 # We can't optimize away up-cast of down-cast.
953 # Because we're doing down-cast of down-cast, the types always have
954 # to match between the two casts
956 x2xN
= '{0}2{0}{1}'.format(x
, N
)
957 x2xM
= '{0}2{0}{1}'.format(x
, M
)
958 aP
= 'a@{0}'.format(P
)
959 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
961 # The N == M case is handled by other optimizations
964 def fexp2i(exp
, bits
):
965 # We assume that exp is already in the right range.
967 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
969 return ('ishl', ('iadd', exp
, 127), 23)
971 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
975 def ldexp(f
, exp
, bits
):
976 # First, we clamp exp to a reasonable range. The maximum possible range
977 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
978 # a maximum range of [-149, 127]. This means that we can potentially have
979 # a swing of +-276. If you start with FLT_MAX, you actually have to do
980 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
981 # spec, on the other hand, only requires that we handle an exponent value
982 # in the range [-126, 128]. This implementation is *mostly* correct; it
983 # handles a range on exp of [-252, 254] which allows you to create any
984 # value (including denorms if the hardware supports it) and to adjust the
985 # exponent of any normal value to anything you want.
987 exp
= ('imin', ('imax', exp
, -28), 30)
989 exp
= ('imin', ('imax', exp
, -252), 254)
991 exp
= ('imin', ('imax', exp
, -2044), 2046)
995 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
996 # (We use ishr which isn't the same for -1, but the -1 case still works
997 # since we use exp-exp/2 as the second exponent.) While the spec
998 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
999 # work with denormals and doesn't allow for the full swing in exponents
1000 # that you can get with normalized values. Instead, we create two powers
1001 # of two and multiply by them each in turn. That way the effective range
1002 # of our exponent is doubled.
1003 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1004 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1005 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1008 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1009 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1010 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1013 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1014 def bitfield_reverse(u
):
1015 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1016 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1017 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1018 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1019 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1023 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1025 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1026 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1027 # and, if a is a NaN then the second comparison will fail anyway.
1028 for op
in ['flt', 'fge', 'feq']:
1030 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1031 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1034 # Add optimizations to handle the case where the result of a ternary is
1035 # compared to a constant. This way we can take things like
1041 # a ? (0 > 0) : (1 > 0)
1043 # which constant folding will eat for lunch. The resulting ternary will
1044 # further get cleaned up by the boolean reductions above and we will be
1045 # left with just the original variable "a".
1046 for op
in ['flt', 'fge', 'feq', 'fne',
1047 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1049 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1050 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1051 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1052 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1056 # For example, this converts things like
1058 # 1 + mix(0, a - 1, condition)
1062 # mix(1, (a-1)+1, condition)
1064 # Other optimizations will rearrange the constants.
1065 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1067 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1070 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1073 # If neither layout qualifier is specified, derivatives in compute shaders
1074 # return zero, which is consistent with the handling of built-in texture
1075 # functions like texture() in GLSL 4.50 compute shaders.
1076 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1077 'fddy', 'fddy_fine', 'fddy_coarse']:
1079 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1082 # This section contains "late" optimizations that should be run before
1083 # creating ffmas and calling regular optimizations for the final time.
1084 # Optimizations should go here if they help code generation and conflict
1085 # with the regular optimizations.
1086 before_ffma_optimizations
= [
1087 # Propagate constants down multiplication chains
1088 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1089 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1090 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1091 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1093 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1094 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1095 (('~fadd', ('fneg', a
), a
), 0.0),
1096 (('iadd', ('ineg', a
), a
), 0),
1097 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1098 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1099 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1100 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1102 (('~flrp@32', ('fadd(is_used_once)', a
, -1.0), ('fadd(is_used_once)', a
, 1.0), d
), ('fadd', ('flrp', -1.0, 1.0, d
), a
)),
1103 (('~flrp@32', ('fadd(is_used_once)', a
, 1.0), ('fadd(is_used_once)', a
, -1.0), d
), ('fadd', ('flrp', 1.0, -1.0, d
), a
)),
1104 (('~flrp@32', ('fadd(is_used_once)', a
, '#b'), ('fadd(is_used_once)', a
, '#c'), d
), ('fadd', ('fmul', d
, ('fadd', c
, ('fneg', b
))), ('fadd', a
, b
))),
1107 # This section contains "late" optimizations that should be run after the
1108 # regular optimizations have finished. Optimizations should go here if
1109 # they help code generation but do not necessarily produce code that is
1110 # more easily optimizable.
1111 late_optimizations
= [
1112 # Most of these optimizations aren't quite safe when you get infinity or
1113 # Nan involved but the first one should be fine.
1114 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1115 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1116 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1117 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1118 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1119 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1121 (('~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
)))),
1123 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1124 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1125 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1126 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1128 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1129 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1131 (('~fadd@32', 1.0, ('fmul(is_used_once)', c
, ('fadd', b
, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c
)), ('fmul', b
, c
)), 'options->lower_flrp32'),
1132 (('~fadd@64', 1.0, ('fmul(is_used_once)', c
, ('fadd', b
, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c
)), ('fmul', b
, c
)), 'options->lower_flrp64'),
1134 # we do these late so that we don't get in the way of creating ffmas
1135 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1136 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1138 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1141 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1142 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1143 before_ffma_optimizations
).render())
1144 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1145 late_optimizations
).render())