60ee170b43f68de20b22cf0e69b437290b98a378
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
57 # All expression types can have a bit-size specified. For opcodes, this
58 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
59 # type and size, and for literals, you can write "2.0@32". In the search half
60 # of the expression this indicates that it should only match that particular
61 # bit-size. In the replace half of the expression this indicates that the
62 # constructed value should have that bit-size.
65 (('fneg', ('fneg', a
)), a
),
66 (('ineg', ('ineg', a
)), a
),
67 (('fabs', ('fabs', a
)), ('fabs', a
)),
68 (('fabs', ('fneg', a
)), ('fabs', a
)),
69 (('iabs', ('iabs', a
)), ('iabs', a
)),
70 (('iabs', ('ineg', a
)), ('iabs', a
)),
71 (('~fadd', a
, 0.0), a
),
73 (('usadd_4x8', a
, 0), a
),
74 (('usadd_4x8', a
, ~
0), ~
0),
75 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
76 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
77 (('~fadd', ('fneg', a
), a
), 0.0),
78 (('iadd', ('ineg', a
), a
), 0),
79 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
80 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
81 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
82 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
83 (('~fmul', a
, 0.0), 0.0),
85 (('umul_unorm_4x8', a
, 0), 0),
86 (('umul_unorm_4x8', a
, ~
0), a
),
87 (('fmul', a
, 1.0), a
),
89 (('fmul', a
, -1.0), ('fneg', a
)),
90 (('imul', a
, -1), ('ineg', a
)),
91 (('~ffma', 0.0, a
, b
), b
),
92 (('~ffma', a
, 0.0, b
), b
),
93 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
94 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
95 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
96 (('~flrp', a
, b
, 0.0), a
),
97 (('~flrp', a
, b
, 1.0), b
),
98 (('~flrp', a
, a
, b
), a
),
99 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
100 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
101 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
102 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
103 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
104 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
105 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
106 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
107 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
108 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
109 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
110 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
111 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), '!options->lower_ffma'),
112 # Comparison simplifications
113 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
114 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
115 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
116 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
117 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
118 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
119 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
120 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
124 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
126 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
130 # fabs(a) != 0.0 because fabs(a) must be >= 0
132 (('flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
134 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
135 (('bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
136 (('bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
137 (('bcsel', ('inot', 'a@bool'), b
, c
), ('bcsel', a
, c
, b
)),
138 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
139 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
146 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
147 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
148 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
149 (('fsat', ('fsat', a
)), ('fsat', a
)),
150 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
151 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
152 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
153 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
154 (('~ior', ('flt', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
155 (('~ior', ('flt', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
156 (('~ior', ('fge', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
157 (('~ior', ('fge', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
158 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
159 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
160 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
161 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
162 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
163 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
164 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
165 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
166 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
167 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
169 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
170 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
171 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
172 (('iand', 'a@bool', 1.0), ('b2f', a
)),
173 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
174 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
175 # Comparison with the same args. Note that these are not done for
176 # the float versions because NaN always returns false on float
178 (('ilt', a
, a
), False),
179 (('ige', a
, a
), True),
180 (('ieq', a
, a
), True),
181 (('ine', a
, a
), False),
182 (('ult', a
, a
), False),
183 (('uge', a
, a
), True),
184 # Logical and bit operations
185 (('fand', a
, 0.0), 0.0),
187 (('iand', a
, ~
0), a
),
191 (('fxor', a
, a
), 0.0),
194 (('inot', ('inot', a
)), a
),
196 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
197 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
198 # Shift optimizations
205 (('iand', 0xff, ('ushr', a
, 24)), ('ushr', a
, 24)),
206 (('iand', 0xffff, ('ushr', a
, 16)), ('ushr', a
, 16)),
207 # Exponential/logarithmic identities
208 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
209 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
210 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
211 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
212 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
213 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
214 (('~fpow', a
, 1.0), a
),
215 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
216 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
217 (('~fpow', 2.0, a
), ('fexp2', a
)),
218 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
219 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
220 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
221 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
222 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
223 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
224 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
225 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
226 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
227 (('~fadd', ('flog2', a
), ('flog2', b
)), ('flog2', ('fmul', a
, b
))),
228 (('~fadd', ('flog2', a
), ('fneg', ('flog2', b
))), ('flog2', ('fdiv', a
, b
))),
229 (('~fmul', ('fexp2', a
), ('fexp2', b
)), ('fexp2', ('fadd', a
, b
))),
230 # Division and reciprocal
231 (('~fdiv', 1.0, a
), ('frcp', a
)),
232 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
233 (('~frcp', ('frcp', a
)), a
),
234 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
235 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
236 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
237 # Boolean simplifications
238 (('ieq', 'a@bool', True), a
),
239 (('ine', 'a@bool', True), ('inot', a
)),
240 (('ine', 'a@bool', False), a
),
241 (('ieq', 'a@bool', False), ('inot', 'a')),
242 (('bcsel', a
, True, False), ('ine', a
, 0)),
243 (('bcsel', a
, False, True), ('ieq', a
, 0)),
244 (('bcsel', True, b
, c
), b
),
245 (('bcsel', False, b
, c
), c
),
246 # The result of this should be hit by constant propagation and, in the
247 # next round of opt_algebraic, get picked up by one of the above two.
248 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
250 (('bcsel', a
, b
, b
), b
),
251 (('fcsel', a
, b
, b
), b
),
254 (('i2b', ('b2i', a
)), a
),
255 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
256 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
257 (('i2b', ('ineg', a
)), ('i2b', a
)),
258 (('i2b', ('iabs', a
)), ('i2b', a
)),
259 (('fabs', ('b2f', a
)), ('b2f', a
)),
260 (('iabs', ('b2i', a
)), ('b2i', a
)),
263 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
264 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
265 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
266 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
269 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
270 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
273 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
274 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
275 (('ussub_4x8', a
, 0), a
),
276 (('ussub_4x8', a
, ~
0), 0),
277 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
278 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
279 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
280 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
281 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
282 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
283 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
284 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
286 # Propagate negation up multiplication chains
287 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
288 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
290 # Reassociate constants in add/mul chains so they can be folded together.
291 # For now, we only handle cases where the constants are separated by
292 # a single non-constant. We could do better eventually.
293 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
294 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
295 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
296 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
299 (('fmod', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
300 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
301 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
302 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
304 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
305 ('bcsel', ('ilt', 31, 'bits'), 'insert',
306 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
307 'options->lower_bitfield_insert'),
309 (('ibitfield_extract', 'value', 'offset', 'bits'),
310 ('bcsel', ('ilt', 31, 'bits'), 'value',
311 ('ibfe', 'value', 'offset', 'bits')),
312 'options->lower_bitfield_extract'),
314 (('ubitfield_extract', 'value', 'offset', 'bits'),
315 ('bcsel', ('ult', 31, 'bits'), 'value',
316 ('ubfe', 'value', 'offset', 'bits')),
317 'options->lower_bitfield_extract'),
319 (('extract_i8', a
, b
),
320 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
321 'options->lower_extract_byte'),
323 (('extract_u8', a
, b
),
324 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
325 'options->lower_extract_byte'),
327 (('extract_i16', a
, b
),
328 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
329 'options->lower_extract_word'),
331 (('extract_u16', a
, b
),
332 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
333 'options->lower_extract_word'),
335 (('pack_unorm_2x16', 'v'),
336 ('pack_uvec2_to_uint',
337 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
338 'options->lower_pack_unorm_2x16'),
340 (('pack_unorm_4x8', 'v'),
341 ('pack_uvec4_to_uint',
342 ('f2u', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
343 'options->lower_pack_unorm_4x8'),
345 (('pack_snorm_2x16', 'v'),
346 ('pack_uvec2_to_uint',
347 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
348 'options->lower_pack_snorm_2x16'),
350 (('pack_snorm_4x8', 'v'),
351 ('pack_uvec4_to_uint',
352 ('f2i', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
353 'options->lower_pack_snorm_4x8'),
355 (('unpack_unorm_2x16', 'v'),
356 ('fdiv', ('u2f', ('vec2', ('extract_u16', 'v', 0),
357 ('extract_u16', 'v', 1))),
359 'options->lower_unpack_unorm_2x16'),
361 (('unpack_unorm_4x8', 'v'),
362 ('fdiv', ('u2f', ('vec4', ('extract_u8', 'v', 0),
363 ('extract_u8', 'v', 1),
364 ('extract_u8', 'v', 2),
365 ('extract_u8', 'v', 3))),
367 'options->lower_unpack_unorm_4x8'),
369 (('unpack_snorm_2x16', 'v'),
370 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
371 ('extract_i16', 'v', 1))),
373 'options->lower_unpack_snorm_2x16'),
375 (('unpack_snorm_4x8', 'v'),
376 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
377 ('extract_i8', 'v', 1),
378 ('extract_i8', 'v', 2),
379 ('extract_i8', 'v', 3))),
381 'options->lower_unpack_snorm_4x8'),
385 # We assume that exp is already in the range [-126, 127].
386 return ('ishl', ('iadd', exp
, 127), 23)
389 # First, we clamp exp to a reasonable range. The maximum possible range
390 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
391 # a maximum range of [-149, 127]. This means that we can potentially have
392 # a swing of +-276. If you start with FLT_MAX, you actually have to do
393 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
394 # spec, on the other hand, only requires that we handle an exponent value
395 # in the range [-126, 128]. This implementation is *mostly* correct; it
396 # handles a range on exp of [-252, 254] which allows you to create any
397 # value (including denorms if the hardware supports it) and to adjust the
398 # exponent of any normal value to anything you want.
399 exp
= ('imin', ('imax', exp
, -252), 254)
401 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
402 # (We use ishr which isn't the same for -1, but the -1 case still works
403 # since we use exp-exp/2 as the second exponent.) While the spec
404 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
405 # work with denormals and doesn't allow for the full swing in exponents
406 # that you can get with normalized values. Instead, we create two powers
407 # of two and multiply by them each in turn. That way the effective range
408 # of our exponent is doubled.
409 pow2_1
= fexp2i(('ishr', exp
, 1))
410 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)))
411 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
413 optimizations
+= [(('ldexp', 'x', 'exp'), ldexp32('x', 'exp'))]
415 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
416 def bitfield_reverse(u
):
417 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
418 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
419 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
420 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
421 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
425 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
428 # Add optimizations to handle the case where the result of a ternary is
429 # compared to a constant. This way we can take things like
435 # a ? (0 > 0) : (1 > 0)
437 # which constant folding will eat for lunch. The resulting ternary will
438 # further get cleaned up by the boolean reductions above and we will be
439 # left with just the original variable "a".
440 for op
in ['flt', 'fge', 'feq', 'fne',
441 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
443 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
444 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
445 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
446 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
449 # This section contains "late" optimizations that should be run after the
450 # regular optimizations have finished. Optimizations should go here if
451 # they help code generation but do not necessarily produce code that is
452 # more easily optimizable.
453 late_optimizations
= [
454 # Most of these optimizations aren't quite safe when you get infinity or
455 # Nan involved but the first one should be fine.
456 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
457 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
458 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
459 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
461 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
462 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
463 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
464 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
467 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()
468 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
469 late_optimizations
).render()