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 (('~fmul', a
, 0.0), 0.0),
110 (('umul_unorm_4x8', a
, 0), 0),
111 (('umul_unorm_4x8', a
, ~
0), a
),
112 (('fmul', a
, 1.0), a
),
114 (('fmul', a
, -1.0), ('fneg', a
)),
115 (('imul', a
, -1), ('ineg', a
)),
116 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*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 => 0*0*0 => abs(0)*0
119 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
120 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
121 (('~ffma', 0.0, a
, b
), b
),
122 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
123 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
124 (('~flrp', a
, b
, 0.0), a
),
125 (('~flrp', a
, b
, 1.0), b
),
126 (('~flrp', a
, a
, b
), a
),
127 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
129 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
130 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
131 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
132 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
134 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
135 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
137 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
138 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
140 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
142 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
143 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
144 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
145 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
146 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
147 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
148 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
149 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
150 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
151 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
152 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
153 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
154 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
155 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
156 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
158 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
159 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
161 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
162 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
163 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
164 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
166 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
167 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
169 # (a * #b + #c) << #d
170 # ((a * #b) << #d) + (#c << #d)
171 # (a * (#b << #d)) + (#c << #d)
172 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
173 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
177 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
179 # Comparison simplifications
180 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
181 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
182 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
183 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
184 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
185 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
186 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
187 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
188 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
189 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
193 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
195 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
197 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
199 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
200 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
201 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
202 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
203 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
204 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
205 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
206 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
207 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
208 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
209 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
210 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
211 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
212 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
213 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
214 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
215 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
216 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
217 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
218 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
220 # -(b2f(a) + b2f(b)) < 0
221 # 0 < b2f(a) + b2f(b)
222 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
224 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
225 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
227 # -(b2f(a) + b2f(b)) >= 0
228 # 0 >= b2f(a) + b2f(b)
229 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
231 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
232 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
234 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
235 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
237 # Some optimizations (below) convert things like (a < b || c < b) into
238 # (min(a, c) < b). However, this interfers with the previous optimizations
239 # that try to remove comparisons with negated sums of b2f. This just
241 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
242 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
244 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
245 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
246 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
247 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
249 # Cannot remove the addition from ilt or ige due to overflow.
250 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
251 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
253 # fmin(-b2f(a), b) >= 0.0
254 # -b2f(a) >= 0.0 && b >= 0.0
255 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
256 # b2f(a) == 0.0 && b >= 0.0
257 # a == False && b >= 0.0
260 # The fge in the second replacement is not a typo. I leave the proof that
261 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
263 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
264 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
266 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
267 (('fne', ('b2f', 'a@1'), 0.0), a
),
268 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
269 (('ine', ('b2i', 'a@1'), 0), a
),
271 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
272 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
273 (('fge', ('u2f', a
), 0.0), True),
274 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
275 (('flt', ('u2f', a
), 0.0), False),
276 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
277 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
278 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
279 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
280 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
281 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
282 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
286 # fabs(a) != 0.0 because fabs(a) must be >= 0
288 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
292 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
295 # 0.0 == fabs(a) because fabs(a) must be >= 0
297 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
301 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
303 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
304 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
305 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
306 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
309 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
310 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
311 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
313 # Since b is a constant, constant folding will eliminate the fmin and the
314 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
315 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
317 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
319 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
320 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
321 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
322 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
323 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
324 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
325 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
326 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
327 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
328 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
329 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
330 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
331 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
332 (('bcsel', a
, True, b
), ('ior', a
, b
)),
333 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
334 (('bcsel', a
, b
, False), ('iand', a
, b
)),
335 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
342 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
343 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
344 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
345 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
346 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
347 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
348 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
349 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
350 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
351 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
352 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
353 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
354 (('fmin', a
, ('fabs', a
)), a
),
355 (('imin', a
, ('iabs', a
)), a
),
356 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
357 (('imax', a
, ('ineg', ('iabs', a
))), a
),
358 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
359 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
360 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
361 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
362 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
363 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
364 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
365 (('fsat', ('b2f', a
)), ('b2f', a
)),
366 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
367 (('fsat', ('fsat', a
)), ('fsat', a
)),
368 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
369 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
370 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
371 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
372 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
373 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
374 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
375 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
376 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
377 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
378 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
379 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
380 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
381 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
382 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
383 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
384 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
385 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
386 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
387 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
388 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
389 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
391 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
392 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
393 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
394 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
395 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
396 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
397 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
398 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
399 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
400 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
401 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
402 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
403 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
404 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
405 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
406 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
408 # Common pattern like 'if (i == 0 || i == 1 || ...)'
409 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
410 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
411 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
413 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
414 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
415 # so emit an open-coded version of that.
416 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
417 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
419 (('ior', a
, ('ieq', a
, False)), True),
420 (('ior', a
, ('inot', a
)), -1),
422 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
423 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
425 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
427 # These patterns can result when (a < b || a < c) => (a < min(b, c))
428 # transformations occur before constant propagation and loop-unrolling.
429 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
430 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
431 (('~fge', a
, ('fmin', b
, a
)), True),
432 (('~fge', ('fmax', a
, b
), a
), True),
433 (('~flt', a
, ('fmin', b
, a
)), False),
434 (('~flt', ('fmax', a
, b
), a
), False),
435 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
436 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
438 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
439 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
440 (('ige', a
, ('imin', b
, a
)), True),
441 (('ige', ('imax', a
, b
), a
), True),
442 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
443 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
444 (('uge', a
, ('umin', b
, a
)), True),
445 (('uge', ('umax', a
, b
), a
), True),
446 (('ilt', a
, ('imin', b
, a
)), False),
447 (('ilt', ('imax', a
, b
), a
), False),
448 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
449 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
450 (('ult', a
, ('umin', b
, a
)), False),
451 (('ult', ('umax', a
, b
), a
), False),
452 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
453 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
455 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
456 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
457 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
458 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
459 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
460 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
461 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
462 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
463 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
464 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
465 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
466 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
467 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
468 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
469 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
470 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
472 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
474 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
475 ('iabs', ('ishr', a
, b
))),
476 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
478 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
479 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
480 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
481 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
482 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
483 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
484 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
485 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
486 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
487 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
489 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
490 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
491 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
492 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
493 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
494 (('ineg', ('b2i32', 'a@32')), a
),
495 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
496 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
497 # Comparison with the same args. Note that these are not done for
498 # the float versions because NaN always returns false on float
500 (('ilt', a
, a
), False),
501 (('ige', a
, a
), True),
502 (('ieq', a
, a
), True),
503 (('ine', a
, a
), False),
504 (('ult', a
, a
), False),
505 (('uge', a
, a
), True),
506 # Logical and bit operations
507 (('fand', a
, 0.0), 0.0),
509 (('iand', a
, ~
0), a
),
513 (('ior', a
, True), True),
514 (('fxor', a
, a
), 0.0),
517 (('inot', ('inot', a
)), a
),
518 (('ior', ('iand', a
, b
), b
), b
),
519 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
520 (('iand', ('ior', a
, b
), b
), b
),
521 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
523 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
524 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
525 # Shift optimizations
532 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
533 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
534 # Exponential/logarithmic identities
535 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
536 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
537 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
538 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
539 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
540 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
541 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
542 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
543 (('~fpow', a
, 1.0), a
),
544 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
545 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
546 (('~fpow', 2.0, a
), ('fexp2', a
)),
547 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
548 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
549 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
550 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
551 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
552 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
553 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
554 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
555 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
556 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
557 # Division and reciprocal
558 (('~fdiv', 1.0, a
), ('frcp', a
)),
559 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
560 (('~frcp', ('frcp', a
)), a
),
561 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
562 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
563 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
564 # Boolean simplifications
565 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
566 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
567 (('ieq', a
, True), a
),
568 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
569 (('ine', a
, False), a
),
570 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
571 (('bcsel', a
, True, False), a
),
572 (('bcsel', a
, False, True), ('inot', a
)),
573 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
574 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
575 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
576 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
577 (('bcsel', True, b
, c
), b
),
578 (('bcsel', False, b
, c
), c
),
579 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
580 # The result of this should be hit by constant propagation and, in the
581 # next round of opt_algebraic, get picked up by one of the above two.
582 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
584 (('bcsel', a
, b
, b
), b
),
585 (('fcsel', a
, b
, b
), b
),
587 # D3D Boolean emulation
588 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
589 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
590 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
591 ('ineg', ('b2i', ('iand', a
, b
)))),
592 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
593 ('ineg', ('b2i', ('ior', a
, b
)))),
594 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
595 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
596 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
597 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
598 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
600 # SM5 32-bit shifts are defined to use the 5 least significant bits
601 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
602 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
603 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
606 (('i2b32', ('b2i', 'a@32')), a
),
607 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
608 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
609 (('i2b', ('ineg', a
)), ('i2b', a
)),
610 (('i2b', ('iabs', a
)), ('i2b', a
)),
611 (('fabs', ('b2f', a
)), ('b2f', a
)),
612 (('iabs', ('b2i', a
)), ('b2i', a
)),
613 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
615 # Ironically, mark these as imprecise because removing the conversions may
616 # preserve more precision than doing the conversions (e.g.,
617 # uint(float(0x81818181u)) == 0x81818200).
618 (('~f2i32', ('i2f', 'a@32')), a
),
619 (('~f2i32', ('u2f', 'a@32')), a
),
620 (('~f2u32', ('i2f', 'a@32')), a
),
621 (('~f2u32', ('u2f', 'a@32')), a
),
623 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
626 # It is undefined to convert a negative floating-point value to an
629 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
630 # some optimizations in the i965 backend to proceed.
631 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
632 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
633 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
634 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
636 # Packing and then unpacking does nothing
637 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
638 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
639 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
640 ('unpack_64_2x32_split_y', a
)), a
),
643 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
644 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
645 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
646 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
647 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
648 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
649 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
652 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
653 # patterns like those below.
654 for op
in ('ushr', 'ishr'):
655 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
656 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
657 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
659 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
661 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
662 # patterns like those below.
663 for op
in ('extract_u8', 'extract_i8'):
664 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
665 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
666 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
668 optimizations
.extend([
670 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
671 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
672 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
673 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
674 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
677 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
678 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
679 (('ussub_4x8', a
, 0), a
),
680 (('ussub_4x8', a
, ~
0), 0),
681 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
682 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
683 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
684 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
685 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
686 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
687 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
688 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
690 # Propagate negation up multiplication chains
691 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
692 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
694 # Propagate constants up multiplication chains
695 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
696 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
697 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
698 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
700 # Reassociate constants in add/mul chains so they can be folded together.
701 # For now, we mostly only handle cases where the constants are separated by
702 # a single non-constant. We could do better eventually.
703 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
704 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
705 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
706 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
707 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
710 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
711 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
712 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
714 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
715 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
716 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
718 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
721 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod16'),
722 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
723 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
724 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
725 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
726 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
728 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
729 ('bcsel', ('ilt', 31, 'bits'), 'insert',
730 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
731 'options->lower_bitfield_insert'),
732 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
733 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
734 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
735 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
736 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
737 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
739 # Alternative lowering that doesn't rely on bfi.
740 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
741 ('bcsel', ('ilt', 31, 'bits'),
744 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
745 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
746 'options->lower_bitfield_insert_to_shifts'),
748 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
749 (('bfm', 'bits', 'offset'),
750 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
751 'options->lower_bfm'),
753 (('ibitfield_extract', 'value', 'offset', 'bits'),
754 ('bcsel', ('ilt', 31, 'bits'), 'value',
755 ('ibfe', 'value', 'offset', 'bits')),
756 'options->lower_bitfield_extract'),
758 (('ubitfield_extract', 'value', 'offset', 'bits'),
759 ('bcsel', ('ult', 31, 'bits'), 'value',
760 ('ubfe', 'value', 'offset', 'bits')),
761 'options->lower_bitfield_extract'),
763 (('ibitfield_extract', 'value', 'offset', 'bits'),
764 ('bcsel', ('ieq', 0, 'bits'),
767 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
768 ('isub', 32, 'bits'))),
769 'options->lower_bitfield_extract_to_shifts'),
771 (('ubitfield_extract', 'value', 'offset', 'bits'),
773 ('ushr', 'value', 'offset'),
774 ('bcsel', ('ieq', 'bits', 32),
776 ('bfm', 'bits', 0))),
777 'options->lower_bitfield_extract_to_shifts'),
779 (('ifind_msb', 'value'),
780 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
781 'options->lower_ifind_msb'),
783 (('find_lsb', 'value'),
784 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
785 'options->lower_find_lsb'),
787 (('extract_i8', a
, 'b@32'),
788 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
789 'options->lower_extract_byte'),
791 (('extract_u8', a
, 'b@32'),
792 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
793 'options->lower_extract_byte'),
795 (('extract_i16', a
, 'b@32'),
796 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
797 'options->lower_extract_word'),
799 (('extract_u16', a
, 'b@32'),
800 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
801 'options->lower_extract_word'),
803 (('pack_unorm_2x16', 'v'),
804 ('pack_uvec2_to_uint',
805 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
806 'options->lower_pack_unorm_2x16'),
808 (('pack_unorm_4x8', 'v'),
809 ('pack_uvec4_to_uint',
810 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
811 'options->lower_pack_unorm_4x8'),
813 (('pack_snorm_2x16', 'v'),
814 ('pack_uvec2_to_uint',
815 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
816 'options->lower_pack_snorm_2x16'),
818 (('pack_snorm_4x8', 'v'),
819 ('pack_uvec4_to_uint',
820 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
821 'options->lower_pack_snorm_4x8'),
823 (('unpack_unorm_2x16', 'v'),
824 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
825 ('extract_u16', 'v', 1))),
827 'options->lower_unpack_unorm_2x16'),
829 (('unpack_unorm_4x8', 'v'),
830 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
831 ('extract_u8', 'v', 1),
832 ('extract_u8', 'v', 2),
833 ('extract_u8', 'v', 3))),
835 'options->lower_unpack_unorm_4x8'),
837 (('unpack_snorm_2x16', 'v'),
838 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
839 ('extract_i16', 'v', 1))),
841 'options->lower_unpack_snorm_2x16'),
843 (('unpack_snorm_4x8', 'v'),
844 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
845 ('extract_i8', 'v', 1),
846 ('extract_i8', 'v', 2),
847 ('extract_i8', 'v', 3))),
849 'options->lower_unpack_snorm_4x8'),
851 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
852 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
855 # bit_size dependent lowerings
856 for bit_size
in [8, 16, 32, 64]:
857 # convenience constants
858 intmax
= (1 << (bit_size
- 1)) - 1
859 intmin
= 1 << (bit_size
- 1)
862 (('iadd_sat@' + str(bit_size
), a
, b
),
863 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
864 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
865 (('isub_sat@' + str(bit_size
), a
, b
),
866 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
867 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
870 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
872 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
873 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
874 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
875 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
876 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
878 # Optimize x2bN(b2x(x)) -> x
879 for size
in type_sizes('bool'):
880 aN
= 'a@' + str(size
)
881 f2bN
= 'f2b' + str(size
)
882 i2bN
= 'i2b' + str(size
)
883 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
884 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
886 # Optimize x2yN(b2x(x)) -> b2y
887 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
888 if x
!= 'f' and y
!= 'f' and x
!= y
:
891 b2x
= 'b2f' if x
== 'f' else 'b2i'
892 b2y
= 'b2f' if y
== 'f' else 'b2i'
893 x2yN
= '{}2{}'.format(x
, y
)
894 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
896 # Optimize away x2xN(a@N)
897 for t
in ['int', 'uint', 'float']:
898 for N
in type_sizes(t
):
899 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
900 aN
= 'a@{0}'.format(N
)
901 optimizations
.append(((x2xN
, aN
), a
))
903 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
904 # In particular, we can optimize away everything except upcast of downcast and
905 # upcasts where the type differs from the other cast
906 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
908 # The outer cast is a down-cast. It doesn't matter what the size of the
909 # argument of the inner cast is because we'll never been in the upcast
910 # of downcast case. Regardless of types, we'll always end up with y2yN
912 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
913 x2xN
= '{0}2{0}{1}'.format(x
, N
)
914 y2yM
= '{0}2{0}{1}'.format(y
, M
)
915 y2yN
= '{0}2{0}{1}'.format(y
, N
)
916 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
918 # If the outer cast is an up-cast, we have to be more careful about the
919 # size of the argument of the inner cast and with types. In this case,
920 # the type is always the type of type up-cast which is given by the
922 for P
in type_sizes('uint'):
923 # We can't optimize away up-cast of down-cast.
927 # Because we're doing down-cast of down-cast, the types always have
928 # to match between the two casts
930 x2xN
= '{0}2{0}{1}'.format(x
, N
)
931 x2xM
= '{0}2{0}{1}'.format(x
, M
)
932 aP
= 'a@{0}'.format(P
)
933 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
935 # The N == M case is handled by other optimizations
938 def fexp2i(exp
, bits
):
939 # We assume that exp is already in the right range.
941 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
943 return ('ishl', ('iadd', exp
, 127), 23)
945 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
949 def ldexp(f
, exp
, bits
):
950 # First, we clamp exp to a reasonable range. The maximum possible range
951 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
952 # a maximum range of [-149, 127]. This means that we can potentially have
953 # a swing of +-276. If you start with FLT_MAX, you actually have to do
954 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
955 # spec, on the other hand, only requires that we handle an exponent value
956 # in the range [-126, 128]. This implementation is *mostly* correct; it
957 # handles a range on exp of [-252, 254] which allows you to create any
958 # value (including denorms if the hardware supports it) and to adjust the
959 # exponent of any normal value to anything you want.
961 exp
= ('imin', ('imax', exp
, -28), 30)
963 exp
= ('imin', ('imax', exp
, -252), 254)
965 exp
= ('imin', ('imax', exp
, -2044), 2046)
969 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
970 # (We use ishr which isn't the same for -1, but the -1 case still works
971 # since we use exp-exp/2 as the second exponent.) While the spec
972 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
973 # work with denormals and doesn't allow for the full swing in exponents
974 # that you can get with normalized values. Instead, we create two powers
975 # of two and multiply by them each in turn. That way the effective range
976 # of our exponent is doubled.
977 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
978 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
979 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
982 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
983 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
984 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
987 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
988 def bitfield_reverse(u
):
989 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
990 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
991 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
992 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
993 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
997 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
999 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1000 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1001 # and, if a is a NaN then the second comparison will fail anyway.
1002 for op
in ['flt', 'fge', 'feq']:
1004 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1005 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1008 # Add optimizations to handle the case where the result of a ternary is
1009 # compared to a constant. This way we can take things like
1015 # a ? (0 > 0) : (1 > 0)
1017 # which constant folding will eat for lunch. The resulting ternary will
1018 # further get cleaned up by the boolean reductions above and we will be
1019 # left with just the original variable "a".
1020 for op
in ['flt', 'fge', 'feq', 'fne',
1021 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1023 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1024 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1025 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1026 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1030 # For example, this converts things like
1032 # 1 + mix(0, a - 1, condition)
1036 # mix(1, (a-1)+1, condition)
1038 # Other optimizations will rearrange the constants.
1039 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1041 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1044 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1047 # If neither layout qualifier is specified, derivatives in compute shaders
1048 # return zero, which is consistent with the handling of built-in texture
1049 # functions like texture() in GLSL 4.50 compute shaders.
1050 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1051 'fddy', 'fddy_fine', 'fddy_coarse']:
1053 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1056 # This section contains "late" optimizations that should be run before
1057 # creating ffmas and calling regular optimizations for the final time.
1058 # Optimizations should go here if they help code generation and conflict
1059 # with the regular optimizations.
1060 before_ffma_optimizations
= [
1061 # Propagate constants down multiplication chains
1062 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1063 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1064 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1065 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1067 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1068 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1069 (('~fadd', ('fneg', a
), a
), 0.0),
1070 (('iadd', ('ineg', a
), a
), 0),
1071 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1072 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1073 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1074 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1076 (('~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
)),
1077 (('~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
)),
1078 (('~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
))),
1081 # This section contains "late" optimizations that should be run after the
1082 # regular optimizations have finished. Optimizations should go here if
1083 # they help code generation but do not necessarily produce code that is
1084 # more easily optimizable.
1085 late_optimizations
= [
1086 # Most of these optimizations aren't quite safe when you get infinity or
1087 # Nan involved but the first one should be fine.
1088 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1089 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1090 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1091 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1092 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1093 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1095 (('~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
)))),
1097 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1098 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1099 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1100 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1102 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1103 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1105 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a
, 0.0, 1.0)),
1106 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a
, -0.0, -1.0)),
1108 (('~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'),
1109 (('~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'),
1111 # we do these late so that we don't get in the way of creating ffmas
1112 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1113 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1115 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1118 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1119 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1120 before_ffma_optimizations
).render())
1121 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1122 late_optimizations
).render())