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
, 0.0), 1.0), 0.0), 1.0), ('fmin', ('fmax', a
, 0.0), 1.0)),
142 (('~ior', ('flt', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
143 (('~ior', ('flt', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
144 (('~ior', ('fge', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
145 (('~ior', ('fge', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
146 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
147 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
148 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
149 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
150 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
151 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
152 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
153 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
154 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
155 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
157 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
158 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
159 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
160 (('iand', 'a@bool', 1.0), ('b2f', a
)),
161 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
162 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
163 # Comparison with the same args. Note that these are not done for
164 # the float versions because NaN always returns false on float
166 (('ilt', a
, a
), False),
167 (('ige', a
, a
), True),
168 (('ieq', a
, a
), True),
169 (('ine', a
, a
), False),
170 (('ult', a
, a
), False),
171 (('uge', a
, a
), True),
172 # Logical and bit operations
173 (('fand', a
, 0.0), 0.0),
175 (('iand', a
, ~
0), a
),
179 (('fxor', a
, a
), 0.0),
182 (('inot', ('inot', a
)), a
),
184 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
185 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
186 # Shift optimizations
193 (('iand', 0xff, ('ushr', a
, 24)), ('ushr', a
, 24)),
194 (('iand', 0xffff, ('ushr', a
, 16)), ('ushr', a
, 16)),
195 # Exponential/logarithmic identities
196 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
197 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
198 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
199 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
200 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
201 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
202 (('~fpow', a
, 1.0), a
),
203 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
204 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
205 (('~fpow', 2.0, a
), ('fexp2', a
)),
206 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
207 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
208 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
209 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
210 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
211 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
212 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
213 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
214 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
215 (('~fadd', ('flog2', a
), ('flog2', b
)), ('flog2', ('fmul', a
, b
))),
216 (('~fadd', ('flog2', a
), ('fneg', ('flog2', b
))), ('flog2', ('fdiv', a
, b
))),
217 (('~fmul', ('fexp2', a
), ('fexp2', b
)), ('fexp2', ('fadd', a
, b
))),
218 # Division and reciprocal
219 (('~fdiv', 1.0, a
), ('frcp', a
)),
220 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
221 (('~frcp', ('frcp', a
)), a
),
222 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
223 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
224 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
225 # Boolean simplifications
226 (('ieq', 'a@bool', True), a
),
227 (('ine', 'a@bool', True), ('inot', a
)),
228 (('ine', 'a@bool', False), a
),
229 (('ieq', 'a@bool', False), ('inot', 'a')),
230 (('bcsel', a
, True, False), ('ine', a
, 0)),
231 (('bcsel', a
, False, True), ('ieq', a
, 0)),
232 (('bcsel', True, b
, c
), b
),
233 (('bcsel', False, b
, c
), c
),
234 # The result of this should be hit by constant propagation and, in the
235 # next round of opt_algebraic, get picked up by one of the above two.
236 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
238 (('bcsel', a
, b
, b
), b
),
239 (('fcsel', a
, b
, b
), b
),
242 (('i2b', ('b2i', a
)), a
),
243 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
244 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
245 (('i2b', ('ineg', a
)), ('i2b', a
)),
246 (('i2b', ('iabs', a
)), ('i2b', a
)),
247 (('fabs', ('b2f', a
)), ('b2f', a
)),
248 (('iabs', ('b2i', a
)), ('b2i', a
)),
251 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
252 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
253 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
254 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
257 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
258 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
261 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
262 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
263 (('ussub_4x8', a
, 0), a
),
264 (('ussub_4x8', a
, ~
0), 0),
265 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
266 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
267 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
268 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
269 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
270 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
271 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
272 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
274 # Propagate negation up multiplication chains
275 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
276 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
279 (('fmod', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
280 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
281 (('uadd_carry', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
282 (('usub_borrow', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
284 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
285 ('bcsel', ('ilt', 31, 'bits'), 'insert',
286 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
287 'options->lower_bitfield_insert'),
289 (('ibitfield_extract', 'value', 'offset', 'bits'),
290 ('bcsel', ('ilt', 31, 'bits'), 'value',
291 ('ibfe', 'value', 'offset', 'bits')),
292 'options->lower_bitfield_extract'),
294 (('ubitfield_extract', 'value', 'offset', 'bits'),
295 ('bcsel', ('ult', 31, 'bits'), 'value',
296 ('ubfe', 'value', 'offset', 'bits')),
297 'options->lower_bitfield_extract'),
299 (('extract_i8', a
, b
),
300 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
301 'options->lower_extract_byte'),
303 (('extract_u8', a
, b
),
304 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
305 'options->lower_extract_byte'),
307 (('extract_i16', a
, b
),
308 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
309 'options->lower_extract_word'),
311 (('extract_u16', a
, b
),
312 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
313 'options->lower_extract_word'),
315 (('pack_unorm_2x16', 'v'),
316 ('pack_uvec2_to_uint',
317 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
318 'options->lower_pack_unorm_2x16'),
320 (('pack_unorm_4x8', 'v'),
321 ('pack_uvec4_to_uint',
322 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
323 'options->lower_pack_unorm_4x8'),
325 (('pack_snorm_2x16', 'v'),
326 ('pack_uvec2_to_uint',
327 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
328 'options->lower_pack_snorm_2x16'),
330 (('pack_snorm_4x8', 'v'),
331 ('pack_uvec4_to_uint',
332 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
333 'options->lower_pack_snorm_4x8'),
335 (('unpack_unorm_2x16', 'v'),
336 ('fdiv', ('u2f', ('vec2', ('extract_u16', 'v', 0),
337 ('extract_u16', 'v', 1))),
339 'options->lower_unpack_unorm_2x16'),
341 (('unpack_unorm_4x8', 'v'),
342 ('fdiv', ('u2f', ('vec4', ('extract_u8', 'v', 0),
343 ('extract_u8', 'v', 1),
344 ('extract_u8', 'v', 2),
345 ('extract_u8', 'v', 3))),
347 'options->lower_unpack_unorm_4x8'),
349 (('unpack_snorm_2x16', 'v'),
350 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
351 ('extract_i16', 'v', 1))),
353 'options->lower_unpack_snorm_2x16'),
355 (('unpack_snorm_4x8', 'v'),
356 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
357 ('extract_i8', 'v', 1),
358 ('extract_i8', 'v', 2),
359 ('extract_i8', 'v', 3))),
361 'options->lower_unpack_snorm_4x8'),
365 # We assume that exp is already in range.
366 return ('ishl', ('iadd', exp
, 127), 23)
369 # First, we clamp exp to a reasonable range. The maximum range that we
370 # need is the largest range for an exponent, ([-127, 128] if you include
371 # inf and 0) plus the number of mantissa bits in either direction to
372 # account for denormals. This means that we need at least a range of
373 # [-150, 151]. For our implementation, however, what we really care
374 # about is that neither exp/2 nor exp-exp/2 go out of the regular range
375 # for floating-point exponents.
376 exp
= ('imin', ('imax', exp
, -252), 254)
378 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
379 # While the spec technically defines ldexp as f * 2.0^exp, simply
380 # multiplying once doesn't work when denormals are involved because
381 # 2.0^exp may not be representable even though ldexp(f, exp) is (see
382 # comments above about range). Instead, we create two powers of two and
383 # multiply by them each in turn. That way the effective range of our
384 # exponent is doubled.
385 pow2_1
= fexp2i(('ishr', exp
, 1))
386 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)))
387 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
389 optimizations
+= [(('ldexp', 'x', 'exp'), ldexp32('x', 'exp'))]
391 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
392 def bitfield_reverse(u
):
393 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
394 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
395 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
396 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
397 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
401 optimizations
+= [(bitfield_reverse('x'), ('bitfield_reverse', 'x'))]
404 # Add optimizations to handle the case where the result of a ternary is
405 # compared to a constant. This way we can take things like
411 # a ? (0 > 0) : (1 > 0)
413 # which constant folding will eat for lunch. The resulting ternary will
414 # further get cleaned up by the boolean reductions above and we will be
415 # left with just the original variable "a".
416 for op
in ['flt', 'fge', 'feq', 'fne',
417 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
419 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
420 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
421 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
422 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
425 # This section contains "late" optimizations that should be run after the
426 # regular optimizations have finished. Optimizations should go here if
427 # they help code generation but do not necessarily produce code that is
428 # more easily optimizable.
429 late_optimizations
= [
430 # Most of these optimizations aren't quite safe when you get infinity or
431 # Nan involved but the first one should be fine.
432 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
433 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
434 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
435 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
437 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
438 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
439 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
440 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
443 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()
444 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
445 late_optimizations
).render()