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)
28 # Convenience variables
34 # Written in the form (<search>, <replace>) where <search> is an expression
35 # and <replace> is either an expression or a value. An expression is
36 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
37 # where each source is either an expression or a value. A value can be
38 # either a numeric constant or a string representing a variable name.
40 # If the opcode in a search expression is prefixed by a '~' character, this
41 # indicates that the operation is inexact. Such operations will only get
42 # applied to SSA values that do not have the exact bit set. This should be
43 # used by by any optimizations that are not bit-for-bit exact. It should not,
44 # however, be used for backend-requested lowering operations as those need to
45 # happen regardless of precision.
47 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
48 # that the given variable will only match constants and the type indicates that
49 # the given variable will only match values from ALU instructions with the
50 # given output type, and (cond) specifies an additional condition function
51 # (see nir_search_helpers.h).
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.
66 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
67 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
72 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
73 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
74 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
75 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
77 (('fneg', ('fneg', a
)), a
),
78 (('ineg', ('ineg', a
)), a
),
79 (('fabs', ('fabs', a
)), ('fabs', a
)),
80 (('fabs', ('fneg', a
)), ('fabs', a
)),
81 (('fabs', ('u2f32', a
)), ('u2f32', a
)),
82 (('iabs', ('iabs', a
)), ('iabs', a
)),
83 (('iabs', ('ineg', a
)), ('iabs', a
)),
84 (('~fadd', a
, 0.0), a
),
86 (('usadd_4x8', a
, 0), a
),
87 (('usadd_4x8', a
, ~
0), ~
0),
88 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
89 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
90 (('~fadd', ('fneg', a
), a
), 0.0),
91 (('iadd', ('ineg', a
), a
), 0),
92 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
93 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
94 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
95 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
96 (('~fmul', a
, 0.0), 0.0),
98 (('umul_unorm_4x8', a
, 0), 0),
99 (('umul_unorm_4x8', a
, ~
0), a
),
100 (('fmul', a
, 1.0), a
),
102 (('fmul', a
, -1.0), ('fneg', a
)),
103 (('imul', a
, -1), ('ineg', a
)),
104 (('~ffma', 0.0, a
, b
), b
),
105 (('~ffma', a
, 0.0, b
), b
),
106 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
107 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
108 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
109 (('~flrp', a
, b
, 0.0), a
),
110 (('~flrp', a
, b
, 1.0), b
),
111 (('~flrp', a
, a
, b
), a
),
112 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
113 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
114 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
115 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
116 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
117 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
118 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
119 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
120 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
121 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
122 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
123 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
124 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
125 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
127 # (a * #b + #c) << #d
128 # ((a * #b) << #d) + (#c << #d)
129 # (a * (#b << #d)) + (#c << #d)
130 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
131 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
135 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
137 # Comparison simplifications
138 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
139 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
140 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
141 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
142 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
143 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
144 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
145 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
149 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
151 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
153 (('fge', ('fneg', ('b2f', a
)), 0.0), ('inot', a
)),
157 # fabs(a) != 0.0 because fabs(a) must be >= 0
159 (('flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
161 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
162 (('bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
163 (('bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
164 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
165 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
166 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
173 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
174 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
175 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
176 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
177 (('fmin', a
, ('fabs', a
)), a
),
178 (('imin', a
, ('iabs', a
)), a
),
179 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
180 (('imax', a
, ('ineg', ('iabs', a
))), a
),
181 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
182 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
183 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
184 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
185 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
186 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
187 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
188 (('fsat', ('fsat', a
)), ('fsat', a
)),
189 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
190 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
191 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
192 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
193 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
194 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
195 (('~ior', ('flt', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
196 (('~ior', ('flt', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
197 (('~ior', ('fge', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
198 (('~ior', ('fge', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
199 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
200 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
201 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
202 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
203 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
204 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
205 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
206 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
207 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
208 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
210 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
211 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
212 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
213 (('iand', 'a@bool', 1.0), ('b2f', a
)),
214 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
215 (('ineg', ('b2i@32', a
)), a
),
216 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
217 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
218 # Comparison with the same args. Note that these are not done for
219 # the float versions because NaN always returns false on float
221 (('ilt', a
, a
), False),
222 (('ige', a
, a
), True),
223 (('ieq', a
, a
), True),
224 (('ine', a
, a
), False),
225 (('ult', a
, a
), False),
226 (('uge', a
, a
), True),
227 # Logical and bit operations
228 (('fand', a
, 0.0), 0.0),
230 (('iand', a
, ~
0), a
),
234 (('ior', a
, True), True),
235 (('fxor', a
, a
), 0.0),
238 (('inot', ('inot', a
)), a
),
240 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
241 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
242 # Shift optimizations
249 (('iand', 0xff, ('ushr', a
, 24)), ('ushr', a
, 24)),
250 (('iand', 0xffff, ('ushr', a
, 16)), ('ushr', a
, 16)),
251 # Exponential/logarithmic identities
252 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
253 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
254 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
255 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
256 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
257 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
258 (('~fpow', a
, 1.0), a
),
259 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
260 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
261 (('~fpow', 2.0, a
), ('fexp2', a
)),
262 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
263 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
264 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
265 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
266 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
267 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
268 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
269 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
270 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
271 (('~fmul', ('fexp2', a
), ('fexp2', b
)), ('fexp2', ('fadd', a
, b
))),
272 # Division and reciprocal
273 (('~fdiv', 1.0, a
), ('frcp', a
)),
274 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
275 (('~frcp', ('frcp', a
)), a
),
276 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
277 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
278 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
279 # Boolean simplifications
280 (('ieq', 'a@bool', True), a
),
281 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
282 (('ine', 'a@bool', False), a
),
283 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
284 (('bcsel', a
, True, False), a
),
285 (('bcsel', a
, False, True), ('inot', a
)),
286 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
287 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
288 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
289 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
290 (('bcsel', True, b
, c
), b
),
291 (('bcsel', False, b
, c
), c
),
292 # The result of this should be hit by constant propagation and, in the
293 # next round of opt_algebraic, get picked up by one of the above two.
294 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
296 (('bcsel', a
, b
, b
), b
),
297 (('fcsel', a
, b
, b
), b
),
300 (('i2b', ('b2i', a
)), a
),
301 (('f2i32', ('ftrunc', a
)), ('f2i32', a
)),
302 (('f2u32', ('ftrunc', a
)), ('f2u32', a
)),
303 (('i2b', ('ineg', a
)), ('i2b', a
)),
304 (('i2b', ('iabs', a
)), ('i2b', a
)),
305 (('fabs', ('b2f', a
)), ('b2f', a
)),
306 (('iabs', ('b2i', a
)), ('b2i', a
)),
308 # Packing and then unpacking does nothing
309 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
310 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
311 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
312 ('unpack_64_2x32_split_y', a
)), a
),
315 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
316 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
317 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
318 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
321 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
322 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
325 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
326 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
327 (('ussub_4x8', a
, 0), a
),
328 (('ussub_4x8', a
, ~
0), 0),
329 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
330 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
331 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
332 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
333 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
334 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
335 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
336 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
338 # Propagate negation up multiplication chains
339 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
340 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
342 # Reassociate constants in add/mul chains so they can be folded together.
343 # For now, we only handle cases where the constants are separated by
344 # a single non-constant. We could do better eventually.
345 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
346 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
347 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
348 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
351 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
352 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
353 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
354 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
355 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
357 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
358 ('bcsel', ('ilt', 31, 'bits'), 'insert',
359 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
360 'options->lower_bitfield_insert'),
362 (('ibitfield_extract', 'value', 'offset', 'bits'),
363 ('bcsel', ('ilt', 31, 'bits'), 'value',
364 ('ibfe', 'value', 'offset', 'bits')),
365 'options->lower_bitfield_extract'),
367 (('ubitfield_extract', 'value', 'offset', 'bits'),
368 ('bcsel', ('ult', 31, 'bits'), 'value',
369 ('ubfe', 'value', 'offset', 'bits')),
370 'options->lower_bitfield_extract'),
372 (('extract_i8', a
, 'b@32'),
373 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
374 'options->lower_extract_byte'),
376 (('extract_u8', a
, 'b@32'),
377 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
378 'options->lower_extract_byte'),
380 (('extract_i16', a
, 'b@32'),
381 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
382 'options->lower_extract_word'),
384 (('extract_u16', a
, 'b@32'),
385 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
386 'options->lower_extract_word'),
388 (('pack_unorm_2x16', 'v'),
389 ('pack_uvec2_to_uint',
390 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
391 'options->lower_pack_unorm_2x16'),
393 (('pack_unorm_4x8', 'v'),
394 ('pack_uvec4_to_uint',
395 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
396 'options->lower_pack_unorm_4x8'),
398 (('pack_snorm_2x16', 'v'),
399 ('pack_uvec2_to_uint',
400 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
401 'options->lower_pack_snorm_2x16'),
403 (('pack_snorm_4x8', 'v'),
404 ('pack_uvec4_to_uint',
405 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
406 'options->lower_pack_snorm_4x8'),
408 (('unpack_unorm_2x16', 'v'),
409 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
410 ('extract_u16', 'v', 1))),
412 'options->lower_unpack_unorm_2x16'),
414 (('unpack_unorm_4x8', 'v'),
415 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
416 ('extract_u8', 'v', 1),
417 ('extract_u8', 'v', 2),
418 ('extract_u8', 'v', 3))),
420 'options->lower_unpack_unorm_4x8'),
422 (('unpack_snorm_2x16', 'v'),
423 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec2', ('extract_i16', 'v', 0),
424 ('extract_i16', 'v', 1))),
426 'options->lower_unpack_snorm_2x16'),
428 (('unpack_snorm_4x8', 'v'),
429 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec4', ('extract_i8', 'v', 0),
430 ('extract_i8', 'v', 1),
431 ('extract_i8', 'v', 2),
432 ('extract_i8', 'v', 3))),
434 'options->lower_unpack_snorm_4x8'),
437 def fexp2i(exp
, bits
):
438 # We assume that exp is already in the right range.
440 return ('ishl', ('iadd', exp
, 127), 23)
442 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
446 def ldexp(f
, exp
, bits
):
447 # First, we clamp exp to a reasonable range. The maximum possible range
448 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
449 # a maximum range of [-149, 127]. This means that we can potentially have
450 # a swing of +-276. If you start with FLT_MAX, you actually have to do
451 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
452 # spec, on the other hand, only requires that we handle an exponent value
453 # in the range [-126, 128]. This implementation is *mostly* correct; it
454 # handles a range on exp of [-252, 254] which allows you to create any
455 # value (including denorms if the hardware supports it) and to adjust the
456 # exponent of any normal value to anything you want.
458 exp
= ('imin', ('imax', exp
, -252), 254)
460 exp
= ('imin', ('imax', exp
, -2044), 2046)
464 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
465 # (We use ishr which isn't the same for -1, but the -1 case still works
466 # since we use exp-exp/2 as the second exponent.) While the spec
467 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
468 # work with denormals and doesn't allow for the full swing in exponents
469 # that you can get with normalized values. Instead, we create two powers
470 # of two and multiply by them each in turn. That way the effective range
471 # of our exponent is doubled.
472 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
473 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
474 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
477 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32)),
478 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64)),
481 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
482 def bitfield_reverse(u
):
483 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
484 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
485 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
486 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
487 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
491 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
493 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
494 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
495 # and, if a is a NaN then the second comparison will fail anyway.
496 for op
in ['flt', 'fge', 'feq']:
498 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
499 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
502 # Add optimizations to handle the case where the result of a ternary is
503 # compared to a constant. This way we can take things like
509 # a ? (0 > 0) : (1 > 0)
511 # which constant folding will eat for lunch. The resulting ternary will
512 # further get cleaned up by the boolean reductions above and we will be
513 # left with just the original variable "a".
514 for op
in ['flt', 'fge', 'feq', 'fne',
515 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
517 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
518 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
519 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
520 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
523 # This section contains "late" optimizations that should be run after the
524 # regular optimizations have finished. Optimizations should go here if
525 # they help code generation but do not necessarily produce code that is
526 # more easily optimizable.
527 late_optimizations
= [
528 # Most of these optimizations aren't quite safe when you get infinity or
529 # Nan involved but the first one should be fine.
530 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
531 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
532 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
533 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
535 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
536 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
537 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
538 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
540 (('b2f(is_used_more_than_once)', ('inot', a
)), ('bcsel', a
, 0.0, 1.0)),
541 (('fneg(is_used_more_than_once)', ('b2f', ('inot', a
))), ('bcsel', a
, -0.0, -1.0)),
543 # we do these late so that we don't get in the way of creating ffmas
544 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
545 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
548 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()
549 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
550 late_optimizations
).render()