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
, 0.0, b
), b
),
123 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
124 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
125 (('ffma', 1.0, a
, b
), ('fadd', 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
)),
130 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
131 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
132 (('flrp@16', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp16'),
133 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
134 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
135 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
136 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
137 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
138 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
139 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
140 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
141 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
142 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
143 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
144 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
145 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
146 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
147 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
149 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
150 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
152 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
153 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
154 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
155 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
157 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
158 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
160 # (a * #b + #c) << #d
161 # ((a * #b) << #d) + (#c << #d)
162 # (a * (#b << #d)) + (#c << #d)
163 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
164 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
168 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
170 # Comparison simplifications
171 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
172 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
173 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
174 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
175 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
176 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
177 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
178 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
179 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
180 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
184 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
186 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
188 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
190 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
191 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
192 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
193 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
194 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
195 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
196 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
197 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
198 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
199 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
200 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
201 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
202 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
203 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
204 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
205 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
206 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
207 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
208 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
209 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
211 # -(b2f(a) + b2f(b)) < 0
212 # 0 < b2f(a) + b2f(b)
213 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
215 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
216 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
218 # -(b2f(a) + b2f(b)) >= 0
219 # 0 >= b2f(a) + b2f(b)
220 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
222 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
223 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
225 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
226 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
228 # Some optimizations (below) convert things like (a < b || c < b) into
229 # (min(a, c) < b). However, this interfers with the previous optimizations
230 # that try to remove comparisons with negated sums of b2f. This just
232 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
233 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
235 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
236 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
237 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
238 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
240 # Cannot remove the addition from ilt or ige due to overflow.
241 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
242 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
244 # fmin(-b2f(a), b) >= 0.0
245 # -b2f(a) >= 0.0 && b >= 0.0
246 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
247 # b2f(a) == 0.0 && b >= 0.0
248 # a == False && b >= 0.0
251 # The fge in the second replacement is not a typo. I leave the proof that
252 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
254 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
255 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
257 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
258 (('fne', ('b2f', 'a@1'), 0.0), a
),
259 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
260 (('ine', ('b2i', 'a@1'), 0), a
),
262 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
263 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
264 (('fge', ('u2f', a
), 0.0), True),
265 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
266 (('flt', ('u2f', a
), 0.0), False),
267 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
268 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
269 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
270 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
271 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
272 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
273 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
277 # fabs(a) != 0.0 because fabs(a) must be >= 0
279 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
283 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
286 # 0.0 == fabs(a) because fabs(a) must be >= 0
288 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
292 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
294 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
295 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
296 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
297 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
300 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
301 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
302 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
304 # Since b is a constant, constant folding will eliminate the fmin and the
305 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
306 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
308 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
310 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
311 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
312 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
313 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
314 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
315 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
316 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
317 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
318 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
319 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
320 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
321 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
322 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
323 (('bcsel', a
, True, b
), ('ior', a
, b
)),
324 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
325 (('bcsel', a
, b
, False), ('iand', a
, b
)),
326 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
333 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
334 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
335 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
336 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
337 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
338 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
339 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
340 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
341 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
342 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
343 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
344 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
345 (('fmin', a
, ('fabs', a
)), a
),
346 (('imin', a
, ('iabs', a
)), a
),
347 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
348 (('imax', a
, ('ineg', ('iabs', a
))), a
),
349 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
350 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
351 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
352 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
353 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
354 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
355 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
356 (('fsat', ('b2f', a
)), ('b2f', a
)),
357 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
358 (('fsat', ('fsat', a
)), ('fsat', a
)),
359 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
360 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
361 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
362 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
363 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
364 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
365 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
366 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
367 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
368 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
369 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
370 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
371 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
372 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
373 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
374 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
375 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
376 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
377 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
378 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
379 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
380 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
382 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
383 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
384 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
385 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
386 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
387 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
388 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
389 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
390 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
391 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
392 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
393 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
394 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
395 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
396 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
397 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
399 # Common pattern like 'if (i == 0 || i == 1 || ...)'
400 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
401 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
402 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
404 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
405 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
406 # so emit an open-coded version of that.
407 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
408 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
410 (('ior', a
, ('ieq', a
, False)), True),
411 (('ior', a
, ('inot', a
)), -1),
413 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
414 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
416 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
418 # These patterns can result when (a < b || a < c) => (a < min(b, c))
419 # transformations occur before constant propagation and loop-unrolling.
420 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
421 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
422 (('~fge', a
, ('fmin', b
, a
)), True),
423 (('~fge', ('fmax', a
, b
), a
), True),
424 (('~flt', a
, ('fmin', b
, a
)), False),
425 (('~flt', ('fmax', a
, b
), a
), False),
426 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
427 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
429 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
430 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
431 (('ige', a
, ('imin', b
, a
)), True),
432 (('ige', ('imax', a
, b
), a
), True),
433 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
434 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
435 (('uge', a
, ('umin', b
, a
)), True),
436 (('uge', ('umax', a
, b
), a
), True),
437 (('ilt', a
, ('imin', b
, a
)), False),
438 (('ilt', ('imax', a
, b
), a
), False),
439 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
440 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
441 (('ult', a
, ('umin', b
, a
)), False),
442 (('ult', ('umax', a
, b
), a
), False),
443 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
444 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
446 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
447 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
448 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
449 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
450 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
451 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
452 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
453 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
454 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
455 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
456 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
457 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
458 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
459 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
460 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
461 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
463 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
465 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
466 ('iabs', ('ishr', a
, b
))),
467 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
469 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
470 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
471 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
472 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
473 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
474 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
475 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
476 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
477 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
478 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
480 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
481 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
482 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
483 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
484 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
485 (('ineg', ('b2i32', 'a@32')), a
),
486 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
487 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
488 # Comparison with the same args. Note that these are not done for
489 # the float versions because NaN always returns false on float
491 (('ilt', a
, a
), False),
492 (('ige', a
, a
), True),
493 (('ieq', a
, a
), True),
494 (('ine', a
, a
), False),
495 (('ult', a
, a
), False),
496 (('uge', a
, a
), True),
497 # Logical and bit operations
498 (('fand', a
, 0.0), 0.0),
500 (('iand', a
, ~
0), a
),
504 (('ior', a
, True), True),
505 (('fxor', a
, a
), 0.0),
508 (('inot', ('inot', a
)), a
),
509 (('ior', ('iand', a
, b
), b
), b
),
510 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
511 (('iand', ('ior', a
, b
), b
), b
),
512 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
514 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
515 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
516 # Shift optimizations
523 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
524 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
525 # Exponential/logarithmic identities
526 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
527 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
528 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
529 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
530 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
531 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
532 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
533 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
534 (('~fpow', a
, 1.0), a
),
535 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
536 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
537 (('~fpow', 2.0, a
), ('fexp2', a
)),
538 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
539 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
540 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
541 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
542 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
543 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
544 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
545 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
546 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
547 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
548 # Division and reciprocal
549 (('~fdiv', 1.0, a
), ('frcp', a
)),
550 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
551 (('~frcp', ('frcp', a
)), a
),
552 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
553 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
554 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
555 # Boolean simplifications
556 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
557 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
558 (('ieq', a
, True), a
),
559 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
560 (('ine', a
, False), a
),
561 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
562 (('bcsel', a
, True, False), a
),
563 (('bcsel', a
, False, True), ('inot', a
)),
564 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
565 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
566 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
567 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
568 (('bcsel', True, b
, c
), b
),
569 (('bcsel', False, b
, c
), c
),
570 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
571 # The result of this should be hit by constant propagation and, in the
572 # next round of opt_algebraic, get picked up by one of the above two.
573 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
575 (('bcsel', a
, b
, b
), b
),
576 (('fcsel', a
, b
, b
), b
),
578 # D3D Boolean emulation
579 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
580 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
581 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
582 ('ineg', ('b2i', ('iand', a
, b
)))),
583 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
584 ('ineg', ('b2i', ('ior', a
, b
)))),
585 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
586 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
587 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
588 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
589 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
591 # SM5 32-bit shifts are defined to use the 5 least significant bits
592 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
593 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
594 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
597 (('i2b32', ('b2i', 'a@32')), a
),
598 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
599 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
600 (('i2b', ('ineg', a
)), ('i2b', a
)),
601 (('i2b', ('iabs', a
)), ('i2b', a
)),
602 (('fabs', ('b2f', a
)), ('b2f', a
)),
603 (('iabs', ('b2i', a
)), ('b2i', a
)),
604 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
606 # Ironically, mark these as imprecise because removing the conversions may
607 # preserve more precision than doing the conversions (e.g.,
608 # uint(float(0x81818181u)) == 0x81818200).
609 (('~f2i32', ('i2f', 'a@32')), a
),
610 (('~f2i32', ('u2f', 'a@32')), a
),
611 (('~f2u32', ('i2f', 'a@32')), a
),
612 (('~f2u32', ('u2f', 'a@32')), a
),
614 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
617 # It is undefined to convert a negative floating-point value to an
620 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
621 # some optimizations in the i965 backend to proceed.
622 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
623 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
624 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
625 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
627 # Packing and then unpacking does nothing
628 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
629 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
630 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
631 ('unpack_64_2x32_split_y', a
)), a
),
634 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
635 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
636 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
637 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
638 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
639 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
640 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
643 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
644 # patterns like those below.
645 for op
in ('ushr', 'ishr'):
646 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
647 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
648 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
650 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
652 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
653 # patterns like those below.
654 for op
in ('extract_u8', 'extract_i8'):
655 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
656 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
657 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
659 optimizations
.extend([
661 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
662 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
663 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
664 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
665 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
668 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
669 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
670 (('ussub_4x8', a
, 0), a
),
671 (('ussub_4x8', a
, ~
0), 0),
672 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
673 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
674 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
675 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
676 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
677 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
678 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
679 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
681 # Propagate negation up multiplication chains
682 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
683 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
685 # Propagate constants up multiplication chains
686 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
687 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
688 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
689 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
691 # Reassociate constants in add/mul chains so they can be folded together.
692 # For now, we mostly only handle cases where the constants are separated by
693 # a single non-constant. We could do better eventually.
694 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
695 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
696 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
697 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
698 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
701 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
702 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
703 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
705 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
706 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
707 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
709 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
712 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod16'),
713 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
714 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
715 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
716 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
717 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
719 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
720 ('bcsel', ('ilt', 31, 'bits'), 'insert',
721 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
722 'options->lower_bitfield_insert'),
723 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
724 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
725 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
726 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
727 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
728 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
730 # Alternative lowering that doesn't rely on bfi.
731 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
732 ('bcsel', ('ilt', 31, 'bits'),
735 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
736 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
737 'options->lower_bitfield_insert_to_shifts'),
739 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
740 (('bfm', 'bits', 'offset'),
741 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
742 'options->lower_bfm'),
744 (('ibitfield_extract', 'value', 'offset', 'bits'),
745 ('bcsel', ('ilt', 31, 'bits'), 'value',
746 ('ibfe', 'value', 'offset', 'bits')),
747 'options->lower_bitfield_extract'),
749 (('ubitfield_extract', 'value', 'offset', 'bits'),
750 ('bcsel', ('ult', 31, 'bits'), 'value',
751 ('ubfe', 'value', 'offset', 'bits')),
752 'options->lower_bitfield_extract'),
754 (('ibitfield_extract', 'value', 'offset', 'bits'),
755 ('bcsel', ('ieq', 0, 'bits'),
758 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
759 ('isub', 32, 'bits'))),
760 'options->lower_bitfield_extract_to_shifts'),
762 (('ubitfield_extract', 'value', 'offset', 'bits'),
764 ('ushr', 'value', 'offset'),
765 ('bcsel', ('ieq', 'bits', 32),
767 ('bfm', 'bits', 0))),
768 'options->lower_bitfield_extract_to_shifts'),
770 (('ifind_msb', 'value'),
771 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
772 'options->lower_ifind_msb'),
774 (('find_lsb', 'value'),
775 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
776 'options->lower_find_lsb'),
778 (('extract_i8', a
, 'b@32'),
779 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
780 'options->lower_extract_byte'),
782 (('extract_u8', a
, 'b@32'),
783 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
784 'options->lower_extract_byte'),
786 (('extract_i16', a
, 'b@32'),
787 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
788 'options->lower_extract_word'),
790 (('extract_u16', a
, 'b@32'),
791 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
792 'options->lower_extract_word'),
794 (('pack_unorm_2x16', 'v'),
795 ('pack_uvec2_to_uint',
796 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
797 'options->lower_pack_unorm_2x16'),
799 (('pack_unorm_4x8', 'v'),
800 ('pack_uvec4_to_uint',
801 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
802 'options->lower_pack_unorm_4x8'),
804 (('pack_snorm_2x16', 'v'),
805 ('pack_uvec2_to_uint',
806 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
807 'options->lower_pack_snorm_2x16'),
809 (('pack_snorm_4x8', 'v'),
810 ('pack_uvec4_to_uint',
811 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
812 'options->lower_pack_snorm_4x8'),
814 (('unpack_unorm_2x16', 'v'),
815 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
816 ('extract_u16', 'v', 1))),
818 'options->lower_unpack_unorm_2x16'),
820 (('unpack_unorm_4x8', 'v'),
821 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
822 ('extract_u8', 'v', 1),
823 ('extract_u8', 'v', 2),
824 ('extract_u8', 'v', 3))),
826 'options->lower_unpack_unorm_4x8'),
828 (('unpack_snorm_2x16', 'v'),
829 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
830 ('extract_i16', 'v', 1))),
832 'options->lower_unpack_snorm_2x16'),
834 (('unpack_snorm_4x8', 'v'),
835 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
836 ('extract_i8', 'v', 1),
837 ('extract_i8', 'v', 2),
838 ('extract_i8', 'v', 3))),
840 'options->lower_unpack_snorm_4x8'),
842 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
843 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
846 # bit_size dependent lowerings
847 for bit_size
in [8, 16, 32, 64]:
848 # convenience constants
849 intmax
= (1 << (bit_size
- 1)) - 1
850 intmin
= 1 << (bit_size
- 1)
853 (('iadd_sat@' + str(bit_size
), a
, b
),
854 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
855 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
856 (('isub_sat@' + str(bit_size
), a
, b
),
857 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
858 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
861 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
863 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
864 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
865 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
866 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
867 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
869 # Optimize x2bN(b2x(x)) -> x
870 for size
in type_sizes('bool'):
871 aN
= 'a@' + str(size
)
872 f2bN
= 'f2b' + str(size
)
873 i2bN
= 'i2b' + str(size
)
874 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
875 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
877 # Optimize x2yN(b2x(x)) -> b2y
878 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
879 if x
!= 'f' and y
!= 'f' and x
!= y
:
882 b2x
= 'b2f' if x
== 'f' else 'b2i'
883 b2y
= 'b2f' if y
== 'f' else 'b2i'
884 x2yN
= '{}2{}'.format(x
, y
)
885 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
887 # Optimize away x2xN(a@N)
888 for t
in ['int', 'uint', 'float']:
889 for N
in type_sizes(t
):
890 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
891 aN
= 'a@{0}'.format(N
)
892 optimizations
.append(((x2xN
, aN
), a
))
894 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
895 # In particular, we can optimize away everything except upcast of downcast and
896 # upcasts where the type differs from the other cast
897 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
899 # The outer cast is a down-cast. It doesn't matter what the size of the
900 # argument of the inner cast is because we'll never been in the upcast
901 # of downcast case. Regardless of types, we'll always end up with y2yN
903 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
904 x2xN
= '{0}2{0}{1}'.format(x
, N
)
905 y2yM
= '{0}2{0}{1}'.format(y
, M
)
906 y2yN
= '{0}2{0}{1}'.format(y
, N
)
907 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
909 # If the outer cast is an up-cast, we have to be more careful about the
910 # size of the argument of the inner cast and with types. In this case,
911 # the type is always the type of type up-cast which is given by the
913 for P
in type_sizes('uint'):
914 # We can't optimize away up-cast of down-cast.
918 # Because we're doing down-cast of down-cast, the types always have
919 # to match between the two casts
921 x2xN
= '{0}2{0}{1}'.format(x
, N
)
922 x2xM
= '{0}2{0}{1}'.format(x
, M
)
923 aP
= 'a@{0}'.format(P
)
924 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
926 # The N == M case is handled by other optimizations
929 def fexp2i(exp
, bits
):
930 # We assume that exp is already in the right range.
932 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
934 return ('ishl', ('iadd', exp
, 127), 23)
936 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
940 def ldexp(f
, exp
, bits
):
941 # First, we clamp exp to a reasonable range. The maximum possible range
942 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
943 # a maximum range of [-149, 127]. This means that we can potentially have
944 # a swing of +-276. If you start with FLT_MAX, you actually have to do
945 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
946 # spec, on the other hand, only requires that we handle an exponent value
947 # in the range [-126, 128]. This implementation is *mostly* correct; it
948 # handles a range on exp of [-252, 254] which allows you to create any
949 # value (including denorms if the hardware supports it) and to adjust the
950 # exponent of any normal value to anything you want.
952 exp
= ('imin', ('imax', exp
, -28), 30)
954 exp
= ('imin', ('imax', exp
, -252), 254)
956 exp
= ('imin', ('imax', exp
, -2044), 2046)
960 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
961 # (We use ishr which isn't the same for -1, but the -1 case still works
962 # since we use exp-exp/2 as the second exponent.) While the spec
963 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
964 # work with denormals and doesn't allow for the full swing in exponents
965 # that you can get with normalized values. Instead, we create two powers
966 # of two and multiply by them each in turn. That way the effective range
967 # of our exponent is doubled.
968 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
969 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
970 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
973 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
974 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
975 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
978 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
979 def bitfield_reverse(u
):
980 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
981 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
982 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
983 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
984 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
988 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
990 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
991 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
992 # and, if a is a NaN then the second comparison will fail anyway.
993 for op
in ['flt', 'fge', 'feq']:
995 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
996 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
999 # Add optimizations to handle the case where the result of a ternary is
1000 # compared to a constant. This way we can take things like
1006 # a ? (0 > 0) : (1 > 0)
1008 # which constant folding will eat for lunch. The resulting ternary will
1009 # further get cleaned up by the boolean reductions above and we will be
1010 # left with just the original variable "a".
1011 for op
in ['flt', 'fge', 'feq', 'fne',
1012 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1014 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1015 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1016 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1017 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1021 # For example, this converts things like
1023 # 1 + mix(0, a - 1, condition)
1027 # mix(1, (a-1)+1, condition)
1029 # Other optimizations will rearrange the constants.
1030 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1032 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1035 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1038 # If neither layout qualifier is specified, derivatives in compute shaders
1039 # return zero, which is consistent with the handling of built-in texture
1040 # functions like texture() in GLSL 4.50 compute shaders.
1041 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1042 'fddy', 'fddy_fine', 'fddy_coarse']:
1044 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1047 # This section contains "late" optimizations that should be run before
1048 # creating ffmas and calling regular optimizations for the final time.
1049 # Optimizations should go here if they help code generation and conflict
1050 # with the regular optimizations.
1051 before_ffma_optimizations
= [
1052 # Propagate constants down multiplication chains
1053 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1054 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1055 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1056 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1058 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1059 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1060 (('~fadd', ('fneg', a
), a
), 0.0),
1061 (('iadd', ('ineg', a
), a
), 0),
1062 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1063 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1064 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1065 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1068 # This section contains "late" optimizations that should be run after the
1069 # regular optimizations have finished. Optimizations should go here if
1070 # they help code generation but do not necessarily produce code that is
1071 # more easily optimizable.
1072 late_optimizations
= [
1073 # Most of these optimizations aren't quite safe when you get infinity or
1074 # Nan involved but the first one should be fine.
1075 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1076 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1077 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1078 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1079 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1080 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1082 (('~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
)))),
1084 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1085 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1086 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1087 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1089 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a
, 0.0, 1.0)),
1090 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a
, -0.0, -1.0)),
1092 # we do these late so that we don't get in the way of creating ffmas
1093 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1094 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1096 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1099 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1100 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1101 before_ffma_optimizations
).render())
1102 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1103 late_optimizations
).render())