3 # Copyright (C) 2014 Intel Corporation
5 # Permission is hereby granted, free of charge, to any person obtaining a
6 # copy of this software and associated documentation files (the "Software"),
7 # to deal in the Software without restriction, including without limitation
8 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 # and/or sell copies of the Software, and to permit persons to whom the
10 # Software is furnished to do so, subject to the following conditions:
12 # The above copyright notice and this permission notice (including the next
13 # paragraph) shall be included in all copies or substantial portions of the
16 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 # Jason Ekstrand (jason@jlekstrand.net)
29 # Convenience variables
35 # Written in the form (<search>, <replace>) where <search> is an expression
36 # and <replace> is either an expression or a value. An expression is
37 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
38 # where each source is either an expression or a value. A value can be
39 # either a numeric constant or a string representing a variable name.
41 # If the opcode in a search expression is prefixed by a '~' character, this
42 # indicates that the operation is inexact. Such operations will only get
43 # applied to SSA values that do not have the exact bit set. This should be
44 # used by by any optimizations that are not bit-for-bit exact. It should not,
45 # however, be used for backend-requested lowering operations as those need to
46 # happen regardless of precision.
48 # Variable names are specified as "[#]name[@type]" where "#" inicates that
49 # the given variable will only match constants and the type indicates that
50 # the given variable will only match values from ALU instructions with the
53 # For constants, you have to be careful to make sure that it is the right
54 # type because python is unaware of the source and destination types of the
58 (('fneg', ('fneg', a
)), a
),
59 (('ineg', ('ineg', a
)), a
),
60 (('fabs', ('fabs', a
)), ('fabs', a
)),
61 (('fabs', ('fneg', a
)), ('fabs', a
)),
62 (('iabs', ('iabs', a
)), ('iabs', a
)),
63 (('iabs', ('ineg', a
)), ('iabs', a
)),
64 (('~fadd', a
, 0.0), a
),
66 (('usadd_4x8', a
, 0), a
),
67 (('usadd_4x8', a
, ~
0), ~
0),
68 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
69 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
70 (('~fadd', ('fneg', a
), a
), 0.0),
71 (('iadd', ('ineg', a
), a
), 0),
72 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
73 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
74 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
75 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
76 (('~fmul', a
, 0.0), 0.0),
78 (('umul_unorm_4x8', a
, 0), 0),
79 (('umul_unorm_4x8', a
, ~
0), a
),
80 (('fmul', a
, 1.0), a
),
82 (('fmul', a
, -1.0), ('fneg', a
)),
83 (('imul', a
, -1), ('ineg', a
)),
84 (('~ffma', 0.0, a
, b
), b
),
85 (('~ffma', a
, 0.0, b
), b
),
86 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
87 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
88 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
89 (('~flrp', a
, b
, 0.0), a
),
90 (('~flrp', a
, b
, 1.0), b
),
91 (('~flrp', a
, a
, b
), a
),
92 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
93 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
94 (('flrp', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp'),
95 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
96 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
97 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp'),
98 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp'),
99 (('~fadd', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp'),
100 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
101 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), '!options->lower_ffma'),
102 # Comparison simplifications
103 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
104 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
105 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
106 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
107 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
108 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
109 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
110 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
114 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
116 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
120 # fabs(a) != 0.0 because fabs(a) must be >= 0
122 (('flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
124 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
125 (('bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
126 (('bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
127 (('bcsel', ('inot', 'a@bool'), b
, c
), ('bcsel', a
, c
, b
)),
128 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
135 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
136 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
137 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
138 (('fsat', ('fsat', a
)), ('fsat', a
)),
139 (('fmin', ('fmax', ('fmin', ('fmax', a
, 0.0), 1.0), 0.0), 1.0), ('fmin', ('fmax', a
, 0.0), 1.0)),
140 (('~ior', ('flt', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
141 (('~ior', ('flt', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
142 (('~ior', ('fge', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
143 (('~ior', ('fge', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
144 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
145 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
146 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
147 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
148 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
149 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
150 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
151 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
152 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
153 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
155 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
156 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
157 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
158 (('iand', 'a@bool', 1.0), ('b2f', a
)),
159 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
160 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
161 # Comparison with the same args. Note that these are not done for
162 # the float versions because NaN always returns false on float
164 (('ilt', a
, a
), False),
165 (('ige', a
, a
), True),
166 (('ieq', a
, a
), True),
167 (('ine', a
, a
), False),
168 (('ult', a
, a
), False),
169 (('uge', a
, a
), True),
170 # Logical and bit operations
171 (('fand', a
, 0.0), 0.0),
173 (('iand', a
, ~
0), a
),
177 (('fxor', a
, a
), 0.0),
180 (('inot', ('inot', a
)), a
),
182 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
183 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
184 # Shift optimizations
191 (('iand', 0xff, ('ushr', a
, 24)), ('ushr', a
, 24)),
192 (('iand', 0xffff, ('ushr', a
, 16)), ('ushr', a
, 16)),
193 # Exponential/logarithmic identities
194 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
195 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
196 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
197 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
198 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
199 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
200 (('~fpow', a
, 1.0), a
),
201 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
202 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
203 (('~fpow', 2.0, a
), ('fexp2', a
)),
204 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
205 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
206 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
207 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
208 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
209 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
210 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
211 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
212 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
213 (('~fadd', ('flog2', a
), ('flog2', b
)), ('flog2', ('fmul', a
, b
))),
214 (('~fadd', ('flog2', a
), ('fneg', ('flog2', b
))), ('flog2', ('fdiv', a
, b
))),
215 (('~fmul', ('fexp2', a
), ('fexp2', b
)), ('fexp2', ('fadd', a
, b
))),
216 # Division and reciprocal
217 (('~fdiv', 1.0, a
), ('frcp', a
)),
218 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
219 (('~frcp', ('frcp', a
)), a
),
220 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
221 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
222 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
223 # Boolean simplifications
224 (('ieq', 'a@bool', True), a
),
225 (('ine', 'a@bool', True), ('inot', a
)),
226 (('ine', 'a@bool', False), a
),
227 (('ieq', 'a@bool', False), ('inot', 'a')),
228 (('bcsel', a
, True, False), ('ine', a
, 0)),
229 (('bcsel', a
, False, True), ('ieq', a
, 0)),
230 (('bcsel', True, b
, c
), b
),
231 (('bcsel', False, b
, c
), c
),
232 # The result of this should be hit by constant propagation and, in the
233 # next round of opt_algebraic, get picked up by one of the above two.
234 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
236 (('bcsel', a
, b
, b
), b
),
237 (('fcsel', a
, b
, b
), b
),
240 (('i2b', ('b2i', a
)), a
),
241 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
242 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
243 (('i2b', ('ineg', a
)), ('i2b', a
)),
244 (('i2b', ('iabs', a
)), ('i2b', a
)),
245 (('fabs', ('b2f', a
)), ('b2f', a
)),
246 (('iabs', ('b2i', a
)), ('b2i', a
)),
249 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
250 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
251 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
252 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
255 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
256 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
259 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
260 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
261 (('ussub_4x8', a
, 0), a
),
262 (('ussub_4x8', a
, ~
0), 0),
263 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
264 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
265 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
266 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
267 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
268 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
269 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
270 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
272 # Propagate negation up multiplication chains
273 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
274 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
277 (('fmod', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
278 (('uadd_carry', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
279 (('usub_borrow', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
281 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
282 ('bcsel', ('ilt', 31, 'bits'), 'insert',
283 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
284 'options->lower_bitfield_insert'),
286 (('ibitfield_extract', 'value', 'offset', 'bits'),
287 ('bcsel', ('ilt', 31, 'bits'), 'value',
288 ('ibfe', 'value', 'offset', 'bits')),
289 'options->lower_bitfield_extract'),
291 (('ubitfield_extract', 'value', 'offset', 'bits'),
292 ('bcsel', ('ult', 31, 'bits'), 'value',
293 ('ubfe', 'value', 'offset', 'bits')),
294 'options->lower_bitfield_extract'),
296 (('extract_i8', a
, b
),
297 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
298 'options->lower_extract_byte'),
300 (('extract_u8', a
, b
),
301 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
302 'options->lower_extract_byte'),
304 (('extract_i16', a
, b
),
305 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
306 'options->lower_extract_word'),
308 (('extract_u16', a
, b
),
309 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
310 'options->lower_extract_word'),
312 (('pack_unorm_2x16', 'v'),
313 ('pack_uvec2_to_uint',
314 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
315 'options->lower_pack_unorm_2x16'),
317 (('pack_unorm_4x8', 'v'),
318 ('pack_uvec4_to_uint',
319 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
320 'options->lower_pack_unorm_4x8'),
322 (('pack_snorm_2x16', 'v'),
323 ('pack_uvec2_to_uint',
324 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
325 'options->lower_pack_snorm_2x16'),
327 (('pack_snorm_4x8', 'v'),
328 ('pack_uvec4_to_uint',
329 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
330 'options->lower_pack_snorm_4x8'),
332 (('unpack_unorm_2x16', 'v'),
333 ('fdiv', ('u2f', ('vec2', ('extract_u16', 'v', 0),
334 ('extract_u16', 'v', 1))),
336 'options->lower_unpack_unorm_2x16'),
338 (('unpack_unorm_4x8', 'v'),
339 ('fdiv', ('u2f', ('vec4', ('extract_u8', 'v', 0),
340 ('extract_u8', 'v', 1),
341 ('extract_u8', 'v', 2),
342 ('extract_u8', 'v', 3))),
344 'options->lower_unpack_unorm_4x8'),
346 (('unpack_snorm_2x16', 'v'),
347 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
348 ('extract_i16', 'v', 1))),
350 'options->lower_unpack_snorm_2x16'),
352 (('unpack_snorm_4x8', 'v'),
353 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
354 ('extract_i8', 'v', 1),
355 ('extract_i8', 'v', 2),
356 ('extract_i8', 'v', 3))),
358 'options->lower_unpack_snorm_4x8'),
361 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
362 def bitfield_reverse(u
):
363 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
364 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
365 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
366 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
367 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
371 optimizations
+= [(bitfield_reverse('x'), ('bitfield_reverse', 'x'))]
374 # Add optimizations to handle the case where the result of a ternary is
375 # compared to a constant. This way we can take things like
381 # a ? (0 > 0) : (1 > 0)
383 # which constant folding will eat for lunch. The resulting ternary will
384 # further get cleaned up by the boolean reductions above and we will be
385 # left with just the original variable "a".
386 for op
in ['flt', 'fge', 'feq', 'fne',
387 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
389 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
390 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
391 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
392 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
395 # This section contains "late" optimizations that should be run after the
396 # regular optimizations have finished. Optimizations should go here if
397 # they help code generation but do not necessarily produce code that is
398 # more easily optimizable.
399 late_optimizations
= [
400 # Most of these optimizations aren't quite safe when you get infinity or
401 # Nan involved but the first one should be fine.
402 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
403 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
404 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
405 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
407 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
408 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
409 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
410 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
413 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()
414 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
415 late_optimizations
).render()