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
32 # Convenience variables
38 # Written in the form (<search>, <replace>) where <search> is an expression
39 # and <replace> is either an expression or a value. An expression is
40 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
41 # where each source is either an expression or a value. A value can be
42 # either a numeric constant or a string representing a variable name.
44 # If the opcode in a search expression is prefixed by a '~' character, this
45 # indicates that the operation is inexact. Such operations will only get
46 # applied to SSA values that do not have the exact bit set. This should be
47 # used by by any optimizations that are not bit-for-bit exact. It should not,
48 # however, be used for backend-requested lowering operations as those need to
49 # happen regardless of precision.
51 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
52 # that the given variable will only match constants and the type indicates that
53 # the given variable will only match values from ALU instructions with the
54 # given output type, and (cond) specifies an additional condition function
55 # (see nir_search_helpers.h).
57 # For constants, you have to be careful to make sure that it is the right
58 # type because python is unaware of the source and destination types of the
61 # All expression types can have a bit-size specified. For opcodes, this
62 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
63 # type and size, and for literals, you can write "2.0@32". In the search half
64 # of the expression this indicates that it should only match that particular
65 # bit-size. In the replace half of the expression this indicates that the
66 # constructed value should have that bit-size.
70 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
71 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
76 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
77 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
78 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
79 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
81 (('fneg', ('fneg', a
)), a
),
82 (('ineg', ('ineg', a
)), a
),
83 (('fabs', ('fabs', a
)), ('fabs', a
)),
84 (('fabs', ('fneg', a
)), ('fabs', a
)),
85 (('fabs', ('u2f32', a
)), ('u2f32', a
)),
86 (('iabs', ('iabs', a
)), ('iabs', a
)),
87 (('iabs', ('ineg', a
)), ('iabs', a
)),
88 (('~fadd', a
, 0.0), a
),
90 (('usadd_4x8', a
, 0), a
),
91 (('usadd_4x8', a
, ~
0), ~
0),
92 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
93 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
94 (('~fadd', ('fneg', a
), a
), 0.0),
95 (('iadd', ('ineg', a
), a
), 0),
96 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
97 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
98 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
99 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
100 (('~fmul', a
, 0.0), 0.0),
102 (('umul_unorm_4x8', a
, 0), 0),
103 (('umul_unorm_4x8', a
, ~
0), a
),
104 (('fmul', a
, 1.0), a
),
106 (('fmul', a
, -1.0), ('fneg', a
)),
107 (('imul', a
, -1), ('ineg', a
)),
108 (('~ffma', 0.0, a
, b
), b
),
109 (('~ffma', a
, 0.0, b
), b
),
110 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
111 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
112 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
113 (('~flrp', a
, b
, 0.0), a
),
114 (('~flrp', a
, b
, 1.0), b
),
115 (('~flrp', a
, a
, b
), a
),
116 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
117 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
118 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
119 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
120 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
121 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
122 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
123 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
124 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
125 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
126 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
127 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
128 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
129 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
131 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
132 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
133 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
134 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
136 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
137 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
139 # (a * #b + #c) << #d
140 # ((a * #b) << #d) + (#c << #d)
141 # (a * (#b << #d)) + (#c << #d)
142 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
143 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
147 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
149 # Comparison simplifications
150 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
151 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
152 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
153 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
154 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
155 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
156 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
157 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
158 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
159 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
163 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
165 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
167 (('fge', ('fneg', ('b2f', a
)), 0.0), ('inot', a
)),
169 (('fne', ('fadd', ('b2f', a
), ('b2f', b
)), 0.0), ('ior', a
, b
)),
170 (('fne', ('fmax', ('b2f', a
), ('b2f', b
)), 0.0), ('ior', a
, b
)),
171 (('fne', ('bcsel', a
, 1.0, ('b2f', b
)) , 0.0), ('ior', a
, b
)),
172 (('fne', ('b2f', a
), ('fneg', ('b2f', b
))), ('ior', a
, b
)),
173 (('fne', ('fmul', ('b2f', a
), ('b2f', b
)), 0.0), ('iand', a
, b
)),
174 (('fne', ('fmin', ('b2f', a
), ('b2f', b
)), 0.0), ('iand', a
, b
)),
175 (('fne', ('bcsel', a
, ('b2f', b
), 0.0) , 0.0), ('iand', a
, b
)),
176 (('feq', ('fadd', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('ior', a
, b
))),
177 (('feq', ('fmax', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('ior', a
, b
))),
178 (('feq', ('bcsel', a
, 1.0, ('b2f', b
)) , 0.0), ('inot', ('ior', a
, b
))),
179 (('feq', ('b2f', a
), ('fneg', ('b2f', b
))), ('inot', ('ior', a
, b
))),
181 # -(b2f(a) + b2f(b)) < 0
182 # 0 < b2f(a) + b2f(b)
183 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
185 (('flt', ('fneg', ('fadd', ('b2f', a
), ('b2f', b
))), 0.0), ('ior', a
, b
)),
186 (('flt', 0.0, ('fadd', ('b2f', a
), ('b2f', b
))), ('ior', a
, b
)),
188 # -(b2f(a) + b2f(b)) >= 0
189 # 0 >= b2f(a) + b2f(b)
190 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
192 (('fge', ('fneg', ('fadd', ('b2f', a
), ('b2f', b
))), 0.0), ('inot', ('ior', a
, b
))),
193 (('fge', 0.0, ('fadd', ('b2f', a
), ('b2f', b
))), ('inot', ('ior', a
, b
))),
195 # Some optimizations (below) convert things like (a < b || c < b) into
196 # (min(a, c) < b). However, this interfers with the previous optimizations
197 # that try to remove comparisons with negated sums of b2f. This just
199 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', a
), ('b2f', b
)))), 0.0),
200 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
202 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
203 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
204 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
205 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
207 # Cannot remove the addition from ilt or ige due to overflow.
208 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
209 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
211 # fmin(-b2f(a), b) >= 0.0
212 # -b2f(a) >= 0.0 && b >= 0.0
213 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
214 # b2f(a) == 0.0 && b >= 0.0
215 # a == False && b >= 0.0
218 # The fge in the second replacement is not a typo. I leave the proof that
219 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
221 (('fge', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
222 (('feq', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
224 (('feq', ('b2f', a
), 0.0), ('inot', a
)),
225 (('fne', ('b2f', a
), 0.0), a
),
226 (('ieq', ('b2i', a
), 0), ('inot', a
)),
227 (('ine', ('b2i', a
), 0), a
),
229 (('fne', ('u2f32', a
), 0.0), ('ine', a
, 0)),
230 (('feq', ('u2f32', a
), 0.0), ('ieq', a
, 0)),
231 (('fge', ('u2f32', a
), 0.0), True),
232 (('fge', 0.0, ('u2f32', a
)), ('uge', 0, a
)), # ieq instead?
233 (('flt', ('u2f32', a
), 0.0), False),
234 (('flt', 0.0, ('u2f32', a
)), ('ult', 0, a
)), # ine instead?
235 (('fne', ('i2f32', a
), 0.0), ('ine', a
, 0)),
236 (('feq', ('i2f32', a
), 0.0), ('ieq', a
, 0)),
237 (('fge', ('i2f32', a
), 0.0), ('ige', a
, 0)),
238 (('fge', 0.0, ('i2f32', a
)), ('ige', 0, a
)),
239 (('flt', ('i2f32', a
), 0.0), ('ilt', a
, 0)),
240 (('flt', 0.0, ('i2f32', a
)), ('ilt', 0, a
)),
244 # fabs(a) != 0.0 because fabs(a) must be >= 0
246 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
250 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
253 # 0.0 == fabs(a) because fabs(a) must be >= 0
255 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
259 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
261 (('fmax', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('ior', a
, b
))),
262 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('ior', a
, b
)))),
263 (('fmin', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
264 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('iand', a
, b
)))),
267 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
268 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
269 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
271 # Since b is a constant, constant folding will eliminate the fmin and the
272 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
273 (('fmin', ('b2f', a
), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
275 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
277 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
278 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
279 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
280 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
281 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
282 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
283 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
284 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
285 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
286 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
287 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
288 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
289 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
296 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
297 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
298 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
299 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
300 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
301 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
302 (('fmin', a
, ('fabs', a
)), a
),
303 (('imin', a
, ('iabs', a
)), a
),
304 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
305 (('imax', a
, ('ineg', ('iabs', a
))), a
),
306 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
307 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
308 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
309 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
310 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
311 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
312 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
313 (('fsat', ('fsat', a
)), ('fsat', a
)),
314 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
315 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
316 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
317 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
318 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
319 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
320 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
321 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
322 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
323 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
324 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
325 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
326 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
327 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
328 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
329 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
330 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
331 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
332 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
333 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
334 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
335 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
337 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
338 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
339 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
340 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
341 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
342 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
343 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
344 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
345 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
346 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
347 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
348 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
349 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
350 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
351 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
352 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
354 (('ior', 'a@bool', ('ieq', a
, False)), True),
355 (('ior', 'a@bool', ('inot', a
)), True),
357 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
359 # These patterns can result when (a < b || a < c) => (a < min(b, c))
360 # transformations occur before constant propagation and loop-unrolling.
361 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
362 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
363 (('~fge', a
, ('fmin', b
, a
)), True),
364 (('~fge', ('fmax', a
, b
), a
), True),
365 (('~flt', a
, ('fmin', b
, a
)), False),
366 (('~flt', ('fmax', a
, b
), a
), False),
367 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
368 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
370 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
371 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
372 (('ige', a
, ('imin', b
, a
)), True),
373 (('ige', ('imax', a
, b
), a
), True),
374 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
375 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
376 (('uge', a
, ('umin', b
, a
)), True),
377 (('uge', ('umax', a
, b
), a
), True),
378 (('ilt', a
, ('imin', b
, a
)), False),
379 (('ilt', ('imax', a
, b
), a
), False),
380 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
381 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
382 (('ult', a
, ('umin', b
, a
)), False),
383 (('ult', ('umax', a
, b
), a
), False),
384 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
385 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
387 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
388 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
389 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
390 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
391 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
392 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
393 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
394 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
395 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
396 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
397 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
398 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
399 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
400 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
401 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
402 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
404 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
405 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
406 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
407 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
408 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
409 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
410 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
411 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
412 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
413 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
415 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
416 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
417 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
418 (('iand', 'a@bool', 1.0), ('b2f', a
), '!options->lower_b2f'),
419 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
420 (('ineg', ('b2i@32', a
)), a
),
421 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
422 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
423 # Comparison with the same args. Note that these are not done for
424 # the float versions because NaN always returns false on float
426 (('ilt', a
, a
), False),
427 (('ige', a
, a
), True),
428 (('ieq', a
, a
), True),
429 (('ine', a
, a
), False),
430 (('ult', a
, a
), False),
431 (('uge', a
, a
), True),
432 # Logical and bit operations
433 (('fand', a
, 0.0), 0.0),
435 (('iand', a
, ~
0), a
),
439 (('ior', a
, True), True),
440 (('fxor', a
, a
), 0.0),
443 (('inot', ('inot', a
)), a
),
444 (('ior', ('iand', a
, b
), b
), b
),
445 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
446 (('iand', ('ior', a
, b
), b
), b
),
447 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
449 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
450 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
451 # Shift optimizations
458 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
459 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
460 # Exponential/logarithmic identities
461 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
462 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
463 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
464 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
465 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
466 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
467 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
468 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
469 (('~fpow', a
, 1.0), a
),
470 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
471 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
472 (('~fpow', 2.0, a
), ('fexp2', a
)),
473 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
474 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
475 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
476 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
477 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
478 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
479 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
480 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
481 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
482 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
483 # Division and reciprocal
484 (('~fdiv', 1.0, a
), ('frcp', a
)),
485 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
486 (('~frcp', ('frcp', a
)), a
),
487 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
488 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
489 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
490 # Boolean simplifications
491 (('ieq', 'a@bool', True), a
),
492 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
493 (('ine', 'a@bool', False), a
),
494 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
495 (('bcsel', a
, True, False), a
),
496 (('bcsel', a
, False, True), ('inot', a
)),
497 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
498 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
499 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
500 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
501 (('bcsel', True, b
, c
), b
),
502 (('bcsel', False, b
, c
), c
),
503 (('bcsel', a
, ('b2f(is_used_once)', b
), ('b2f', c
)), ('b2f', ('bcsel', a
, b
, c
))),
504 # The result of this should be hit by constant propagation and, in the
505 # next round of opt_algebraic, get picked up by one of the above two.
506 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
508 (('bcsel', a
, b
, b
), b
),
509 (('fcsel', a
, b
, b
), b
),
512 (('i2b', ('b2i', a
)), a
),
513 (('i2b', 'a@bool'), a
),
514 (('f2i32', ('ftrunc', a
)), ('f2i32', a
)),
515 (('f2u32', ('ftrunc', a
)), ('f2u32', a
)),
516 (('i2b', ('ineg', a
)), ('i2b', a
)),
517 (('i2b', ('iabs', a
)), ('i2b', a
)),
518 (('fabs', ('b2f', a
)), ('b2f', a
)),
519 (('iabs', ('b2i', a
)), ('b2i', a
)),
520 (('inot', ('f2b', a
)), ('feq', a
, 0.0)),
522 # Ironically, mark these as imprecise because removing the conversions may
523 # preserve more precision than doing the conversions (e.g.,
524 # uint(float(0x81818181u)) == 0x81818200).
525 (('~f2i32', ('i2f32', 'a@32')), a
),
526 (('~f2i32', ('u2f32', 'a@32')), a
),
527 (('~f2u32', ('i2f32', 'a@32')), a
),
528 (('~f2u32', ('u2f32', 'a@32')), a
),
530 # Packing and then unpacking does nothing
531 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
532 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
533 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
534 ('unpack_64_2x32_split_y', a
)), a
),
537 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
538 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
539 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
540 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
543 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
544 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
547 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
548 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
549 (('ussub_4x8', a
, 0), a
),
550 (('ussub_4x8', a
, ~
0), 0),
551 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
552 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
553 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
554 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
555 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
556 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
557 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
558 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
560 # Propagate negation up multiplication chains
561 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
562 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
564 # Propagate constants up multiplication chains
565 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
566 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
567 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
568 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
570 # Reassociate constants in add/mul chains so they can be folded together.
571 # For now, we mostly only handle cases where the constants are separated by
572 # a single non-constant. We could do better eventually.
573 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
574 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
575 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
576 (('~fadd', '#a', ('fneg', ('fadd', b
, '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
577 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
580 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
581 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
582 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
584 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
585 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
586 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
588 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
591 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
592 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
593 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
594 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
595 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
597 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
598 ('bcsel', ('ilt', 31, 'bits'), 'insert',
599 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
600 'options->lower_bitfield_insert'),
602 # Alternative lowering that doesn't rely on bfi.
603 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
604 ('bcsel', ('ilt', 31, 'bits'),
607 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
608 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
609 'options->lower_bitfield_insert_to_shifts'),
611 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
612 (('bfm', 'bits', 'offset'),
613 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
614 'options->lower_bfm'),
616 (('ibitfield_extract', 'value', 'offset', 'bits'),
617 ('bcsel', ('ilt', 31, 'bits'), 'value',
618 ('ibfe', 'value', 'offset', 'bits')),
619 'options->lower_bitfield_extract'),
621 (('ubitfield_extract', 'value', 'offset', 'bits'),
622 ('bcsel', ('ult', 31, 'bits'), 'value',
623 ('ubfe', 'value', 'offset', 'bits')),
624 'options->lower_bitfield_extract'),
626 (('ibitfield_extract', 'value', 'offset', 'bits'),
627 ('bcsel', ('ieq', 0, 'bits'),
630 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
631 ('isub', 32, 'bits'))),
632 'options->lower_bitfield_extract_to_shifts'),
634 (('ubitfield_extract', 'value', 'offset', 'bits'),
636 ('ushr', 'value', 'offset'),
637 ('bcsel', ('ieq', 'bits', 32),
639 ('bfm', 'bits', 0))),
640 'options->lower_bitfield_extract_to_shifts'),
642 (('ifind_msb', 'value'),
643 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
644 'options->lower_ifind_msb'),
646 (('find_lsb', 'value'),
647 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
648 'options->lower_find_lsb'),
650 (('extract_i8', a
, 'b@32'),
651 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
652 'options->lower_extract_byte'),
654 (('extract_u8', a
, 'b@32'),
655 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
656 'options->lower_extract_byte'),
658 (('extract_i16', a
, 'b@32'),
659 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
660 'options->lower_extract_word'),
662 (('extract_u16', a
, 'b@32'),
663 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
664 'options->lower_extract_word'),
666 (('pack_unorm_2x16', 'v'),
667 ('pack_uvec2_to_uint',
668 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
669 'options->lower_pack_unorm_2x16'),
671 (('pack_unorm_4x8', 'v'),
672 ('pack_uvec4_to_uint',
673 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
674 'options->lower_pack_unorm_4x8'),
676 (('pack_snorm_2x16', 'v'),
677 ('pack_uvec2_to_uint',
678 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
679 'options->lower_pack_snorm_2x16'),
681 (('pack_snorm_4x8', 'v'),
682 ('pack_uvec4_to_uint',
683 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
684 'options->lower_pack_snorm_4x8'),
686 (('unpack_unorm_2x16', 'v'),
687 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
688 ('extract_u16', 'v', 1))),
690 'options->lower_unpack_unorm_2x16'),
692 (('unpack_unorm_4x8', 'v'),
693 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
694 ('extract_u8', 'v', 1),
695 ('extract_u8', 'v', 2),
696 ('extract_u8', 'v', 3))),
698 'options->lower_unpack_unorm_4x8'),
700 (('unpack_snorm_2x16', 'v'),
701 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec2', ('extract_i16', 'v', 0),
702 ('extract_i16', 'v', 1))),
704 'options->lower_unpack_snorm_2x16'),
706 (('unpack_snorm_4x8', 'v'),
707 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec4', ('extract_i8', 'v', 0),
708 ('extract_i8', 'v', 1),
709 ('extract_i8', 'v', 2),
710 ('extract_i8', 'v', 3))),
712 'options->lower_unpack_snorm_4x8'),
715 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
717 for left
, right
in list(itertools
.combinations(invert
.keys(), 2)) + zip(invert
.keys(), invert
.keys()):
718 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
719 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
720 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
721 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
723 def fexp2i(exp
, bits
):
724 # We assume that exp is already in the right range.
726 return ('ishl', ('iadd', exp
, 127), 23)
728 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
732 def ldexp(f
, exp
, bits
):
733 # First, we clamp exp to a reasonable range. The maximum possible range
734 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
735 # a maximum range of [-149, 127]. This means that we can potentially have
736 # a swing of +-276. If you start with FLT_MAX, you actually have to do
737 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
738 # spec, on the other hand, only requires that we handle an exponent value
739 # in the range [-126, 128]. This implementation is *mostly* correct; it
740 # handles a range on exp of [-252, 254] which allows you to create any
741 # value (including denorms if the hardware supports it) and to adjust the
742 # exponent of any normal value to anything you want.
744 exp
= ('imin', ('imax', exp
, -252), 254)
746 exp
= ('imin', ('imax', exp
, -2044), 2046)
750 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
751 # (We use ishr which isn't the same for -1, but the -1 case still works
752 # since we use exp-exp/2 as the second exponent.) While the spec
753 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
754 # work with denormals and doesn't allow for the full swing in exponents
755 # that you can get with normalized values. Instead, we create two powers
756 # of two and multiply by them each in turn. That way the effective range
757 # of our exponent is doubled.
758 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
759 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
760 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
763 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
764 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
767 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
768 def bitfield_reverse(u
):
769 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
770 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
771 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
772 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
773 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
777 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
779 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
780 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
781 # and, if a is a NaN then the second comparison will fail anyway.
782 for op
in ['flt', 'fge', 'feq']:
784 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
785 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
788 # Add optimizations to handle the case where the result of a ternary is
789 # compared to a constant. This way we can take things like
795 # a ? (0 > 0) : (1 > 0)
797 # which constant folding will eat for lunch. The resulting ternary will
798 # further get cleaned up by the boolean reductions above and we will be
799 # left with just the original variable "a".
800 for op
in ['flt', 'fge', 'feq', 'fne',
801 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
803 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
804 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
805 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
806 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
810 # For example, this converts things like
812 # 1 + mix(0, a - 1, condition)
816 # mix(1, (a-1)+1, condition)
818 # Other optimizations will rearrange the constants.
819 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
821 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
824 # This section contains "late" optimizations that should be run before
825 # creating ffmas and calling regular optimizations for the final time.
826 # Optimizations should go here if they help code generation and conflict
827 # with the regular optimizations.
828 before_ffma_optimizations
= [
829 # Propagate constants down multiplication chains
830 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
831 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
832 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
833 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
835 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
836 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
837 (('~fadd', ('fneg', a
), a
), 0.0),
838 (('iadd', ('ineg', a
), a
), 0),
839 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
840 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
841 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
842 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
845 # This section contains "late" optimizations that should be run after the
846 # regular optimizations have finished. Optimizations should go here if
847 # they help code generation but do not necessarily produce code that is
848 # more easily optimizable.
849 late_optimizations
= [
850 # Most of these optimizations aren't quite safe when you get infinity or
851 # Nan involved but the first one should be fine.
852 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
853 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
854 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
855 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
856 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
857 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
859 (('~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
)))),
861 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
862 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
863 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
864 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
866 (('b2f(is_used_more_than_once)', ('inot', a
)), ('bcsel', a
, 0.0, 1.0)),
867 (('fneg(is_used_more_than_once)', ('b2f', ('inot', a
))), ('bcsel', a
, -0.0, -1.0)),
869 # we do these late so that we don't get in the way of creating ffmas
870 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
871 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
873 # Lowered for backends without a dedicated b2f instruction
874 (('b2f@32', a
), ('iand', a
, 1.0), 'options->lower_b2f'),
877 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
878 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
879 before_ffma_optimizations
).render())
880 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
881 late_optimizations
).render())
projects / mesa.git / blob