2 # -*- encoding: utf-8 -*-
4 # Copyright (C) 2014 Intel Corporation
6 # Permission is hereby granted, free of charge, to any person obtaining a
7 # copy of this software and associated documentation files (the "Software"),
8 # to deal in the Software without restriction, including without limitation
9 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 # and/or sell copies of the Software, and to permit persons to whom the
11 # Software is furnished to do so, subject to the following conditions:
13 # The above copyright notice and this permission notice (including the next
14 # paragraph) shall be included in all copies or substantial portions of the
17 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 # Jason Ekstrand (jason@jlekstrand.net)
30 # Convenience variables
36 # Written in the form (<search>, <replace>) where <search> is an expression
37 # and <replace> is either an expression or a value. An expression is
38 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
39 # where each source is either an expression or a value. A value can be
40 # either a numeric constant or a string representing a variable name.
42 # If the opcode in a search expression is prefixed by a '~' character, this
43 # indicates that the operation is inexact. Such operations will only get
44 # applied to SSA values that do not have the exact bit set. This should be
45 # used by by any optimizations that are not bit-for-bit exact. It should not,
46 # however, be used for backend-requested lowering operations as those need to
47 # happen regardless of precision.
49 # Variable names are specified as "[#]name[@type]" where "#" inicates that
50 # the given variable will only match constants and the type indicates that
51 # the given variable will only match values from ALU instructions with the
54 # For constants, you have to be careful to make sure that it is the right
55 # type because python is unaware of the source and destination types of the
59 (('fneg', ('fneg', a
)), a
),
60 (('ineg', ('ineg', a
)), a
),
61 (('fabs', ('fabs', a
)), ('fabs', a
)),
62 (('fabs', ('fneg', a
)), ('fabs', a
)),
63 (('iabs', ('iabs', a
)), ('iabs', a
)),
64 (('iabs', ('ineg', a
)), ('iabs', a
)),
65 (('~fadd', a
, 0.0), a
),
67 (('usadd_4x8', a
, 0), a
),
68 (('usadd_4x8', a
, ~
0), ~
0),
69 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
70 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
71 (('~fadd', ('fneg', a
), a
), 0.0),
72 (('iadd', ('ineg', a
), a
), 0),
73 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
74 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
75 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
76 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
77 (('~fmul', a
, 0.0), 0.0),
79 (('umul_unorm_4x8', a
, 0), 0),
80 (('umul_unorm_4x8', a
, ~
0), a
),
81 (('fmul', a
, 1.0), a
),
83 (('fmul', a
, -1.0), ('fneg', a
)),
84 (('imul', a
, -1), ('ineg', a
)),
85 (('~ffma', 0.0, a
, b
), b
),
86 (('~ffma', a
, 0.0, b
), b
),
87 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
88 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
89 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
90 (('~flrp', a
, b
, 0.0), a
),
91 (('~flrp', a
, b
, 1.0), b
),
92 (('~flrp', a
, a
, b
), a
),
93 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
94 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
95 (('flrp', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp'),
96 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
97 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
98 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp'),
99 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
100 (('~fadd', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp'),
101 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
102 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), '!options->lower_ffma'),
103 # Comparison simplifications
104 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
105 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
106 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
107 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
108 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
109 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
110 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
111 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
115 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
117 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
121 # fabs(a) != 0.0 because fabs(a) must be >= 0
123 (('flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
125 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
126 (('bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
127 (('bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
128 (('bcsel', ('inot', 'a@bool'), b
, c
), ('bcsel', a
, c
, b
)),
129 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
130 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
137 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
138 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
139 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
140 (('fsat', ('fsat', a
)), ('fsat', a
)),
141 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
142 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
143 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
144 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
145 (('~ior', ('flt', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
146 (('~ior', ('flt', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
147 (('~ior', ('fge', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
148 (('~ior', ('fge', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
149 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
150 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
151 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
152 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
153 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
154 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
155 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
156 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
157 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
158 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
160 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
161 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
162 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
163 (('iand', 'a@bool', 1.0), ('b2f', a
)),
164 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
165 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
166 # Comparison with the same args. Note that these are not done for
167 # the float versions because NaN always returns false on float
169 (('ilt', a
, a
), False),
170 (('ige', a
, a
), True),
171 (('ieq', a
, a
), True),
172 (('ine', a
, a
), False),
173 (('ult', a
, a
), False),
174 (('uge', a
, a
), True),
175 # Logical and bit operations
176 (('fand', a
, 0.0), 0.0),
178 (('iand', a
, ~
0), a
),
182 (('fxor', a
, a
), 0.0),
185 (('inot', ('inot', a
)), a
),
187 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
188 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
189 # Shift optimizations
196 (('iand', 0xff, ('ushr', a
, 24)), ('ushr', a
, 24)),
197 (('iand', 0xffff, ('ushr', a
, 16)), ('ushr', a
, 16)),
198 # Exponential/logarithmic identities
199 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
200 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
201 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
202 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
203 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
204 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
205 (('~fpow', a
, 1.0), a
),
206 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
207 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
208 (('~fpow', 2.0, a
), ('fexp2', a
)),
209 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
210 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
211 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
212 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
213 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
214 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
215 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
216 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
217 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
218 (('~fadd', ('flog2', a
), ('flog2', b
)), ('flog2', ('fmul', a
, b
))),
219 (('~fadd', ('flog2', a
), ('fneg', ('flog2', b
))), ('flog2', ('fdiv', a
, b
))),
220 (('~fmul', ('fexp2', a
), ('fexp2', b
)), ('fexp2', ('fadd', a
, b
))),
221 # Division and reciprocal
222 (('~fdiv', 1.0, a
), ('frcp', a
)),
223 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
224 (('~frcp', ('frcp', a
)), a
),
225 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
226 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
227 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
228 # Boolean simplifications
229 (('ieq', 'a@bool', True), a
),
230 (('ine', 'a@bool', True), ('inot', a
)),
231 (('ine', 'a@bool', False), a
),
232 (('ieq', 'a@bool', False), ('inot', 'a')),
233 (('bcsel', a
, True, False), ('ine', a
, 0)),
234 (('bcsel', a
, False, True), ('ieq', a
, 0)),
235 (('bcsel', True, b
, c
), b
),
236 (('bcsel', False, b
, c
), c
),
237 # The result of this should be hit by constant propagation and, in the
238 # next round of opt_algebraic, get picked up by one of the above two.
239 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
241 (('bcsel', a
, b
, b
), b
),
242 (('fcsel', a
, b
, b
), b
),
245 (('i2b', ('b2i', a
)), a
),
246 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
247 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
248 (('i2b', ('ineg', a
)), ('i2b', a
)),
249 (('i2b', ('iabs', a
)), ('i2b', a
)),
250 (('fabs', ('b2f', a
)), ('b2f', a
)),
251 (('iabs', ('b2i', a
)), ('b2i', a
)),
254 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
255 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
256 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
257 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
260 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
261 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
264 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
265 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
266 (('ussub_4x8', a
, 0), a
),
267 (('ussub_4x8', a
, ~
0), 0),
268 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
269 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
270 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
271 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
272 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
273 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
274 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
275 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
277 # Propagate negation up multiplication chains
278 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
279 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
281 # Reassociate constants in add/mul chains so they can be folded together.
282 # For now, we only handle cases where the constants are separated by
283 # a single non-constant. We could do better eventually.
284 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
285 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
286 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
287 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
290 (('fmod', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
291 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
292 (('uadd_carry', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
293 (('usub_borrow', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
295 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
296 ('bcsel', ('ilt', 31, 'bits'), 'insert',
297 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
298 'options->lower_bitfield_insert'),
300 (('ibitfield_extract', 'value', 'offset', 'bits'),
301 ('bcsel', ('ilt', 31, 'bits'), 'value',
302 ('ibfe', 'value', 'offset', 'bits')),
303 'options->lower_bitfield_extract'),
305 (('ubitfield_extract', 'value', 'offset', 'bits'),
306 ('bcsel', ('ult', 31, 'bits'), 'value',
307 ('ubfe', 'value', 'offset', 'bits')),
308 'options->lower_bitfield_extract'),
310 (('extract_i8', a
, b
),
311 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
312 'options->lower_extract_byte'),
314 (('extract_u8', a
, b
),
315 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
316 'options->lower_extract_byte'),
318 (('extract_i16', a
, b
),
319 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
320 'options->lower_extract_word'),
322 (('extract_u16', a
, b
),
323 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
324 'options->lower_extract_word'),
326 (('pack_unorm_2x16', 'v'),
327 ('pack_uvec2_to_uint',
328 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
329 'options->lower_pack_unorm_2x16'),
331 (('pack_unorm_4x8', 'v'),
332 ('pack_uvec4_to_uint',
333 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
334 'options->lower_pack_unorm_4x8'),
336 (('pack_snorm_2x16', 'v'),
337 ('pack_uvec2_to_uint',
338 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
339 'options->lower_pack_snorm_2x16'),
341 (('pack_snorm_4x8', 'v'),
342 ('pack_uvec4_to_uint',
343 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
344 'options->lower_pack_snorm_4x8'),
346 (('unpack_unorm_2x16', 'v'),
347 ('fdiv', ('u2f', ('vec2', ('extract_u16', 'v', 0),
348 ('extract_u16', 'v', 1))),
350 'options->lower_unpack_unorm_2x16'),
352 (('unpack_unorm_4x8', 'v'),
353 ('fdiv', ('u2f', ('vec4', ('extract_u8', 'v', 0),
354 ('extract_u8', 'v', 1),
355 ('extract_u8', 'v', 2),
356 ('extract_u8', 'v', 3))),
358 'options->lower_unpack_unorm_4x8'),
360 (('unpack_snorm_2x16', 'v'),
361 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
362 ('extract_i16', 'v', 1))),
364 'options->lower_unpack_snorm_2x16'),
366 (('unpack_snorm_4x8', 'v'),
367 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
368 ('extract_i8', 'v', 1),
369 ('extract_i8', 'v', 2),
370 ('extract_i8', 'v', 3))),
372 'options->lower_unpack_snorm_4x8'),
376 # We assume that exp is already in the range [-126, 127].
377 return ('ishl', ('iadd', exp
, 127), 23)
380 # First, we clamp exp to a reasonable range. The maximum possible range
381 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
382 # a maximum range of [-149, 127]. This means that we can potentially have
383 # a swing of +-276. If you start with FLT_MAX, you actually have to do
384 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
385 # spec, on the other hand, only requires that we handle an exponent value
386 # in the range [-126, 128]. This implementation is *mostly* correct; it
387 # handles a range on exp of [-252, 254] which allows you to create any
388 # value (including denorms if the hardware supports it) and to adjust the
389 # exponent of any normal value to anything you want.
390 exp
= ('imin', ('imax', exp
, -252), 254)
392 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
393 # (We use ishr which isn't the same for -1, but the -1 case still works
394 # since we use exp-exp/2 as the second exponent.) While the spec
395 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
396 # work with denormals and doesn't allow for the full swing in exponents
397 # that you can get with normalized values. Instead, we create two powers
398 # of two and multiply by them each in turn. That way the effective range
399 # of our exponent is doubled.
400 pow2_1
= fexp2i(('ishr', exp
, 1))
401 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)))
402 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
404 optimizations
+= [(('ldexp', 'x', 'exp'), ldexp32('x', 'exp'))]
406 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
407 def bitfield_reverse(u
):
408 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
409 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
410 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
411 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
412 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
416 optimizations
+= [(bitfield_reverse('x'), ('bitfield_reverse', 'x'))]
419 # Add optimizations to handle the case where the result of a ternary is
420 # compared to a constant. This way we can take things like
426 # a ? (0 > 0) : (1 > 0)
428 # which constant folding will eat for lunch. The resulting ternary will
429 # further get cleaned up by the boolean reductions above and we will be
430 # left with just the original variable "a".
431 for op
in ['flt', 'fge', 'feq', 'fne',
432 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
434 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
435 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
436 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
437 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
440 # This section contains "late" optimizations that should be run after the
441 # regular optimizations have finished. Optimizations should go here if
442 # they help code generation but do not necessarily produce code that is
443 # more easily optimizable.
444 late_optimizations
= [
445 # Most of these optimizations aren't quite safe when you get infinity or
446 # Nan involved but the first one should be fine.
447 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
448 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
449 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
450 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
452 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
453 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
454 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
455 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
458 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()
459 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
460 late_optimizations
).render()