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 (('fne', ('fadd', ('b2f', a
), ('fneg', ('b2f', b
))), 0.0), ('ixor', a
, b
)),
177 (('fne', ('b2f', a
) , ('b2f', b
) ), ('ixor', a
, b
)),
178 (('fne', ('fneg', ('b2f', a
)), ('fneg', ('b2f', b
))), ('ixor', a
, b
)),
179 (('feq', ('fadd', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('ior', a
, b
))),
180 (('feq', ('fmax', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('ior', a
, b
))),
181 (('feq', ('bcsel', a
, 1.0, ('b2f', b
)) , 0.0), ('inot', ('ior', a
, b
))),
182 (('feq', ('b2f', a
), ('fneg', ('b2f', b
))), ('inot', ('ior', a
, b
))),
183 (('feq', ('fmul', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('iand', a
, b
))),
184 (('feq', ('fmin', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('iand', a
, b
))),
185 (('feq', ('bcsel', a
, ('b2f', b
), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
187 # -(b2f(a) + b2f(b)) < 0
188 # 0 < b2f(a) + b2f(b)
189 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
191 (('flt', ('fneg', ('fadd', ('b2f', a
), ('b2f', b
))), 0.0), ('ior', a
, b
)),
192 (('flt', 0.0, ('fadd', ('b2f', a
), ('b2f', b
))), ('ior', a
, b
)),
194 # -(b2f(a) + b2f(b)) >= 0
195 # 0 >= b2f(a) + b2f(b)
196 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
198 (('fge', ('fneg', ('fadd', ('b2f', a
), ('b2f', b
))), 0.0), ('inot', ('ior', a
, b
))),
199 (('fge', 0.0, ('fadd', ('b2f', a
), ('b2f', b
))), ('inot', ('ior', a
, b
))),
201 # Some optimizations (below) convert things like (a < b || c < b) into
202 # (min(a, c) < b). However, this interfers with the previous optimizations
203 # that try to remove comparisons with negated sums of b2f. This just
205 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', a
), ('b2f', b
)))), 0.0),
206 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
208 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
209 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
210 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
211 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
213 # Cannot remove the addition from ilt or ige due to overflow.
214 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
215 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
217 # fmin(-b2f(a), b) >= 0.0
218 # -b2f(a) >= 0.0 && b >= 0.0
219 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
220 # b2f(a) == 0.0 && b >= 0.0
221 # a == False && b >= 0.0
224 # The fge in the second replacement is not a typo. I leave the proof that
225 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
227 (('fge', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
228 (('feq', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
230 (('feq', ('b2f', a
), 0.0), ('inot', a
)),
231 (('fne', ('b2f', a
), 0.0), a
),
232 (('ieq', ('b2i', a
), 0), ('inot', a
)),
233 (('ine', ('b2i', a
), 0), a
),
235 (('fne', ('u2f32', a
), 0.0), ('ine', a
, 0)),
236 (('feq', ('u2f32', a
), 0.0), ('ieq', a
, 0)),
237 (('fge', ('u2f32', a
), 0.0), True),
238 (('fge', 0.0, ('u2f32', a
)), ('uge', 0, a
)), # ieq instead?
239 (('flt', ('u2f32', a
), 0.0), False),
240 (('flt', 0.0, ('u2f32', a
)), ('ult', 0, a
)), # ine instead?
241 (('fne', ('i2f32', a
), 0.0), ('ine', a
, 0)),
242 (('feq', ('i2f32', a
), 0.0), ('ieq', a
, 0)),
243 (('fge', ('i2f32', a
), 0.0), ('ige', a
, 0)),
244 (('fge', 0.0, ('i2f32', a
)), ('ige', 0, a
)),
245 (('flt', ('i2f32', a
), 0.0), ('ilt', a
, 0)),
246 (('flt', 0.0, ('i2f32', a
)), ('ilt', 0, a
)),
250 # fabs(a) != 0.0 because fabs(a) must be >= 0
252 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
256 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
259 # 0.0 == fabs(a) because fabs(a) must be >= 0
261 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
265 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
267 (('fmax', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('ior', a
, b
))),
268 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('ior', a
, b
)))),
269 (('fmin', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
270 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('iand', a
, b
)))),
273 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
274 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
275 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
277 # Since b is a constant, constant folding will eliminate the fmin and the
278 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
279 (('fmin', ('b2f', a
), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
281 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
283 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
284 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
285 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
286 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
287 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
288 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
289 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
290 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
291 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
292 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
293 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
294 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
295 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
302 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
303 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
304 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
305 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
306 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
307 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
308 (('fmin', a
, ('fabs', a
)), a
),
309 (('imin', a
, ('iabs', a
)), a
),
310 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
311 (('imax', a
, ('ineg', ('iabs', a
))), a
),
312 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
313 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
314 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
315 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
316 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
317 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
318 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
319 (('fsat', ('fsat', a
)), ('fsat', a
)),
320 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
321 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
322 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
323 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
324 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
325 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
326 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
327 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
328 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
329 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
330 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
331 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
332 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
333 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
334 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
335 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
336 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
337 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
338 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
339 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
340 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
341 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
343 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
344 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
345 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
346 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
347 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
348 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
349 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
350 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
351 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
352 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
353 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
354 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
355 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
356 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
357 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
358 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
360 (('ior', 'a@bool', ('ieq', a
, False)), True),
361 (('ior', 'a@bool', ('inot', a
)), True),
363 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
365 # These patterns can result when (a < b || a < c) => (a < min(b, c))
366 # transformations occur before constant propagation and loop-unrolling.
367 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
368 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
369 (('~fge', a
, ('fmin', b
, a
)), True),
370 (('~fge', ('fmax', a
, b
), a
), True),
371 (('~flt', a
, ('fmin', b
, a
)), False),
372 (('~flt', ('fmax', a
, b
), a
), False),
373 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
374 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
376 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
377 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
378 (('ige', a
, ('imin', b
, a
)), True),
379 (('ige', ('imax', a
, b
), a
), True),
380 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
381 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
382 (('uge', a
, ('umin', b
, a
)), True),
383 (('uge', ('umax', a
, b
), a
), True),
384 (('ilt', a
, ('imin', b
, a
)), False),
385 (('ilt', ('imax', a
, b
), a
), False),
386 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
387 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
388 (('ult', a
, ('umin', b
, a
)), False),
389 (('ult', ('umax', a
, b
), a
), False),
390 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
391 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
393 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
394 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
395 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
396 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
397 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
398 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
399 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
400 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
401 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
402 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
403 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
404 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
405 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
406 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
407 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
408 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
410 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
411 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
412 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
413 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
414 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
415 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
416 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
417 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
418 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
419 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
421 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
422 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
423 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
424 (('iand', 'a@bool', 1.0), ('b2f', a
), '!options->lower_b2f'),
425 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
426 (('ineg', ('b2i@32', a
)), a
),
427 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
428 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
429 # Comparison with the same args. Note that these are not done for
430 # the float versions because NaN always returns false on float
432 (('ilt', a
, a
), False),
433 (('ige', a
, a
), True),
434 (('ieq', a
, a
), True),
435 (('ine', a
, a
), False),
436 (('ult', a
, a
), False),
437 (('uge', a
, a
), True),
438 # Logical and bit operations
439 (('fand', a
, 0.0), 0.0),
441 (('iand', a
, ~
0), a
),
445 (('ior', a
, True), True),
446 (('fxor', a
, a
), 0.0),
449 (('inot', ('inot', a
)), a
),
450 (('ior', ('iand', a
, b
), b
), b
),
451 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
452 (('iand', ('ior', a
, b
), b
), b
),
453 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
455 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
456 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
457 # Shift optimizations
464 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
465 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
466 # Exponential/logarithmic identities
467 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
468 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
469 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
470 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
471 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
472 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
473 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
474 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
475 (('~fpow', a
, 1.0), a
),
476 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
477 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
478 (('~fpow', 2.0, a
), ('fexp2', a
)),
479 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
480 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
481 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
482 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
483 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
484 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
485 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
486 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
487 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
488 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
489 # Division and reciprocal
490 (('~fdiv', 1.0, a
), ('frcp', a
)),
491 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
492 (('~frcp', ('frcp', a
)), a
),
493 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
494 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
495 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
496 # Boolean simplifications
497 (('ieq', 'a@bool', True), a
),
498 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
499 (('ine', 'a@bool', False), a
),
500 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
501 (('bcsel', a
, True, False), a
),
502 (('bcsel', a
, False, True), ('inot', a
)),
503 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
504 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
505 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
506 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
507 (('bcsel', True, b
, c
), b
),
508 (('bcsel', False, b
, c
), c
),
509 (('bcsel', a
, ('b2f(is_used_once)', b
), ('b2f', c
)), ('b2f', ('bcsel', a
, b
, c
))),
510 # The result of this should be hit by constant propagation and, in the
511 # next round of opt_algebraic, get picked up by one of the above two.
512 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
514 (('bcsel', a
, b
, b
), b
),
515 (('fcsel', a
, b
, b
), b
),
518 (('i2b', ('b2i', a
)), a
),
519 (('i2b', 'a@bool'), a
),
520 (('f2i32', ('ftrunc', a
)), ('f2i32', a
)),
521 (('f2u32', ('ftrunc', a
)), ('f2u32', a
)),
522 (('i2b', ('ineg', a
)), ('i2b', a
)),
523 (('i2b', ('iabs', a
)), ('i2b', a
)),
524 (('fabs', ('b2f', a
)), ('b2f', a
)),
525 (('iabs', ('b2i', a
)), ('b2i', a
)),
526 (('inot', ('f2b', a
)), ('feq', a
, 0.0)),
528 # Ironically, mark these as imprecise because removing the conversions may
529 # preserve more precision than doing the conversions (e.g.,
530 # uint(float(0x81818181u)) == 0x81818200).
531 (('~f2i32', ('i2f32', 'a@32')), a
),
532 (('~f2i32', ('u2f32', 'a@32')), a
),
533 (('~f2u32', ('i2f32', 'a@32')), a
),
534 (('~f2u32', ('u2f32', 'a@32')), a
),
536 # Packing and then unpacking does nothing
537 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
538 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
539 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
540 ('unpack_64_2x32_split_y', a
)), a
),
543 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
544 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
545 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
546 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
549 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
550 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
553 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
554 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
555 (('ussub_4x8', a
, 0), a
),
556 (('ussub_4x8', a
, ~
0), 0),
557 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
558 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
559 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
560 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
561 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
562 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
563 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
564 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
566 # Propagate negation up multiplication chains
567 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
568 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
570 # Propagate constants up multiplication chains
571 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
572 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
573 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
574 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
576 # Reassociate constants in add/mul chains so they can be folded together.
577 # For now, we mostly only handle cases where the constants are separated by
578 # a single non-constant. We could do better eventually.
579 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
580 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
581 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
582 (('~fadd', '#a', ('fneg', ('fadd', b
, '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
583 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
586 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
587 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
588 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
590 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
591 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
592 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
594 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
597 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
598 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
599 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
600 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
601 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
603 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
604 ('bcsel', ('ilt', 31, 'bits'), 'insert',
605 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
606 'options->lower_bitfield_insert'),
608 # Alternative lowering that doesn't rely on bfi.
609 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
610 ('bcsel', ('ilt', 31, 'bits'),
613 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
614 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
615 'options->lower_bitfield_insert_to_shifts'),
617 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
618 (('bfm', 'bits', 'offset'),
619 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
620 'options->lower_bfm'),
622 (('ibitfield_extract', 'value', 'offset', 'bits'),
623 ('bcsel', ('ilt', 31, 'bits'), 'value',
624 ('ibfe', 'value', 'offset', 'bits')),
625 'options->lower_bitfield_extract'),
627 (('ubitfield_extract', 'value', 'offset', 'bits'),
628 ('bcsel', ('ult', 31, 'bits'), 'value',
629 ('ubfe', 'value', 'offset', 'bits')),
630 'options->lower_bitfield_extract'),
632 (('ibitfield_extract', 'value', 'offset', 'bits'),
633 ('bcsel', ('ieq', 0, 'bits'),
636 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
637 ('isub', 32, 'bits'))),
638 'options->lower_bitfield_extract_to_shifts'),
640 (('ubitfield_extract', 'value', 'offset', 'bits'),
642 ('ushr', 'value', 'offset'),
643 ('bcsel', ('ieq', 'bits', 32),
645 ('bfm', 'bits', 0))),
646 'options->lower_bitfield_extract_to_shifts'),
648 (('ifind_msb', 'value'),
649 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
650 'options->lower_ifind_msb'),
652 (('find_lsb', 'value'),
653 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
654 'options->lower_find_lsb'),
656 (('extract_i8', a
, 'b@32'),
657 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
658 'options->lower_extract_byte'),
660 (('extract_u8', a
, 'b@32'),
661 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
662 'options->lower_extract_byte'),
664 (('extract_i16', a
, 'b@32'),
665 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
666 'options->lower_extract_word'),
668 (('extract_u16', a
, 'b@32'),
669 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
670 'options->lower_extract_word'),
672 (('pack_unorm_2x16', 'v'),
673 ('pack_uvec2_to_uint',
674 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
675 'options->lower_pack_unorm_2x16'),
677 (('pack_unorm_4x8', 'v'),
678 ('pack_uvec4_to_uint',
679 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
680 'options->lower_pack_unorm_4x8'),
682 (('pack_snorm_2x16', 'v'),
683 ('pack_uvec2_to_uint',
684 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
685 'options->lower_pack_snorm_2x16'),
687 (('pack_snorm_4x8', 'v'),
688 ('pack_uvec4_to_uint',
689 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
690 'options->lower_pack_snorm_4x8'),
692 (('unpack_unorm_2x16', 'v'),
693 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
694 ('extract_u16', 'v', 1))),
696 'options->lower_unpack_unorm_2x16'),
698 (('unpack_unorm_4x8', 'v'),
699 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
700 ('extract_u8', 'v', 1),
701 ('extract_u8', 'v', 2),
702 ('extract_u8', 'v', 3))),
704 'options->lower_unpack_unorm_4x8'),
706 (('unpack_snorm_2x16', 'v'),
707 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec2', ('extract_i16', 'v', 0),
708 ('extract_i16', 'v', 1))),
710 'options->lower_unpack_snorm_2x16'),
712 (('unpack_snorm_4x8', 'v'),
713 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec4', ('extract_i8', 'v', 0),
714 ('extract_i8', 'v', 1),
715 ('extract_i8', 'v', 2),
716 ('extract_i8', 'v', 3))),
718 'options->lower_unpack_snorm_4x8'),
721 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
723 for left
, right
in list(itertools
.combinations(invert
.keys(), 2)) + zip(invert
.keys(), invert
.keys()):
724 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
725 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
726 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
727 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
729 def fexp2i(exp
, bits
):
730 # We assume that exp is already in the right range.
732 return ('ishl', ('iadd', exp
, 127), 23)
734 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
738 def ldexp(f
, exp
, bits
):
739 # First, we clamp exp to a reasonable range. The maximum possible range
740 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
741 # a maximum range of [-149, 127]. This means that we can potentially have
742 # a swing of +-276. If you start with FLT_MAX, you actually have to do
743 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
744 # spec, on the other hand, only requires that we handle an exponent value
745 # in the range [-126, 128]. This implementation is *mostly* correct; it
746 # handles a range on exp of [-252, 254] which allows you to create any
747 # value (including denorms if the hardware supports it) and to adjust the
748 # exponent of any normal value to anything you want.
750 exp
= ('imin', ('imax', exp
, -252), 254)
752 exp
= ('imin', ('imax', exp
, -2044), 2046)
756 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
757 # (We use ishr which isn't the same for -1, but the -1 case still works
758 # since we use exp-exp/2 as the second exponent.) While the spec
759 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
760 # work with denormals and doesn't allow for the full swing in exponents
761 # that you can get with normalized values. Instead, we create two powers
762 # of two and multiply by them each in turn. That way the effective range
763 # of our exponent is doubled.
764 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
765 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
766 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
769 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
770 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
773 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
774 def bitfield_reverse(u
):
775 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
776 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
777 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
778 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
779 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
783 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
785 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
786 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
787 # and, if a is a NaN then the second comparison will fail anyway.
788 for op
in ['flt', 'fge', 'feq']:
790 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
791 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
794 # Add optimizations to handle the case where the result of a ternary is
795 # compared to a constant. This way we can take things like
801 # a ? (0 > 0) : (1 > 0)
803 # which constant folding will eat for lunch. The resulting ternary will
804 # further get cleaned up by the boolean reductions above and we will be
805 # left with just the original variable "a".
806 for op
in ['flt', 'fge', 'feq', 'fne',
807 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
809 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
810 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
811 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
812 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
816 # For example, this converts things like
818 # 1 + mix(0, a - 1, condition)
822 # mix(1, (a-1)+1, condition)
824 # Other optimizations will rearrange the constants.
825 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
827 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
830 # This section contains "late" optimizations that should be run before
831 # creating ffmas and calling regular optimizations for the final time.
832 # Optimizations should go here if they help code generation and conflict
833 # with the regular optimizations.
834 before_ffma_optimizations
= [
835 # Propagate constants down multiplication chains
836 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
837 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
838 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
839 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
841 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
842 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
843 (('~fadd', ('fneg', a
), a
), 0.0),
844 (('iadd', ('ineg', a
), a
), 0),
845 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
846 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
847 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
848 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
851 # This section contains "late" optimizations that should be run after the
852 # regular optimizations have finished. Optimizations should go here if
853 # they help code generation but do not necessarily produce code that is
854 # more easily optimizable.
855 late_optimizations
= [
856 # Most of these optimizations aren't quite safe when you get infinity or
857 # Nan involved but the first one should be fine.
858 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
859 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
860 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
861 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
862 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
863 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
865 (('~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
)))),
867 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
868 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
869 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
870 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
872 (('b2f(is_used_more_than_once)', ('inot', a
)), ('bcsel', a
, 0.0, 1.0)),
873 (('fneg(is_used_more_than_once)', ('b2f', ('inot', a
))), ('bcsel', a
, -0.0, -1.0)),
875 # we do these late so that we don't get in the way of creating ffmas
876 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
877 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
879 # Lowered for backends without a dedicated b2f instruction
880 (('b2f@32', a
), ('iand', a
, 1.0), 'options->lower_b2f'),
883 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
884 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
885 before_ffma_optimizations
).render())
886 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
887 late_optimizations
).render())