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)
26 from __future__
import print_function
28 from collections
import OrderedDict
32 # Convenience variables
38 # Written in the form (<search>, <replace>) where <search> is an expression
39 # and <replace> is either an expression or a value. An expression is
40 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
41 # where each source is either an expression or a value. A value can be
42 # either a numeric constant or a string representing a variable name.
44 # If the opcode in a search expression is prefixed by a '~' character, this
45 # indicates that the operation is inexact. Such operations will only get
46 # applied to SSA values that do not have the exact bit set. This should be
47 # used by by any optimizations that are not bit-for-bit exact. It should not,
48 # however, be used for backend-requested lowering operations as those need to
49 # happen regardless of precision.
51 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
52 # that the given variable will only match constants and the type indicates that
53 # the given variable will only match values from ALU instructions with the
54 # given output type, and (cond) specifies an additional condition function
55 # (see nir_search_helpers.h).
57 # For constants, you have to be careful to make sure that it is the right
58 # type because python is unaware of the source and destination types of the
61 # All expression types can have a bit-size specified. For opcodes, this
62 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
63 # type and size, and for literals, you can write "2.0@32". In the search half
64 # of the expression this indicates that it should only match that particular
65 # bit-size. In the replace half of the expression this indicates that the
66 # constructed value should have that bit-size.
70 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
71 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
76 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
77 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
78 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
79 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
81 (('fneg', ('fneg', a
)), a
),
82 (('ineg', ('ineg', a
)), a
),
83 (('fabs', ('fabs', a
)), ('fabs', a
)),
84 (('fabs', ('fneg', a
)), ('fabs', a
)),
85 (('fabs', ('u2f32', a
)), ('u2f32', a
)),
86 (('iabs', ('iabs', a
)), ('iabs', a
)),
87 (('iabs', ('ineg', a
)), ('iabs', a
)),
88 (('~fadd', a
, 0.0), a
),
90 (('usadd_4x8', a
, 0), a
),
91 (('usadd_4x8', a
, ~
0), ~
0),
92 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
93 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
94 (('~fadd', ('fneg', a
), a
), 0.0),
95 (('iadd', ('ineg', a
), a
), 0),
96 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
97 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
98 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
99 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
100 (('~fmul', a
, 0.0), 0.0),
102 (('umul_unorm_4x8', a
, 0), 0),
103 (('umul_unorm_4x8', a
, ~
0), a
),
104 (('fmul', a
, 1.0), a
),
106 (('fmul', a
, -1.0), ('fneg', a
)),
107 (('imul', a
, -1), ('ineg', a
)),
108 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
109 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
110 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
111 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
112 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
113 (('~ffma', 0.0, a
, b
), b
),
114 (('~ffma', a
, 0.0, b
), b
),
115 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
116 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
117 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
118 (('~flrp', a
, b
, 0.0), a
),
119 (('~flrp', a
, b
, 1.0), b
),
120 (('~flrp', a
, a
, b
), a
),
121 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
122 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
123 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
124 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
125 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
126 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
127 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
128 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
129 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
130 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
131 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
132 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
133 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
134 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
136 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
137 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
138 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
139 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
141 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
142 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
144 # (a * #b + #c) << #d
145 # ((a * #b) << #d) + (#c << #d)
146 # (a * (#b << #d)) + (#c << #d)
147 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
148 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
152 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
154 # Comparison simplifications
155 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
156 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
157 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
158 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
159 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
160 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
161 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
162 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
163 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
164 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
168 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
170 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
172 (('fge', ('fneg', ('b2f', a
)), 0.0), ('inot', a
)),
174 (('fne', ('fadd', ('b2f', a
), ('b2f', b
)), 0.0), ('ior', a
, b
)),
175 (('fne', ('fmax', ('b2f', a
), ('b2f', b
)), 0.0), ('ior', a
, b
)),
176 (('fne', ('bcsel', a
, 1.0, ('b2f', b
)) , 0.0), ('ior', a
, b
)),
177 (('fne', ('b2f', a
), ('fneg', ('b2f', b
))), ('ior', a
, b
)),
178 (('fne', ('fmul', ('b2f', a
), ('b2f', b
)), 0.0), ('iand', a
, b
)),
179 (('fne', ('fmin', ('b2f', a
), ('b2f', b
)), 0.0), ('iand', a
, b
)),
180 (('fne', ('bcsel', a
, ('b2f', b
), 0.0) , 0.0), ('iand', a
, b
)),
181 (('fne', ('fadd', ('b2f', a
), ('fneg', ('b2f', b
))), 0.0), ('ixor', a
, b
)),
182 (('fne', ('b2f', a
) , ('b2f', b
) ), ('ixor', a
, b
)),
183 (('fne', ('fneg', ('b2f', a
)), ('fneg', ('b2f', b
))), ('ixor', a
, b
)),
184 (('feq', ('fadd', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('ior', a
, b
))),
185 (('feq', ('fmax', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('ior', a
, b
))),
186 (('feq', ('bcsel', a
, 1.0, ('b2f', b
)) , 0.0), ('inot', ('ior', a
, b
))),
187 (('feq', ('b2f', a
), ('fneg', ('b2f', b
))), ('inot', ('ior', a
, b
))),
188 (('feq', ('fmul', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('iand', a
, b
))),
189 (('feq', ('fmin', ('b2f', a
), ('b2f', b
)), 0.0), ('inot', ('iand', a
, b
))),
190 (('feq', ('bcsel', a
, ('b2f', b
), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
191 (('feq', ('fadd', ('b2f', a
), ('fneg', ('b2f', b
))), 0.0), ('ieq', a
, b
)),
192 (('feq', ('b2f', a
) , ('b2f', b
) ), ('ieq', a
, b
)),
193 (('feq', ('fneg', ('b2f', a
)), ('fneg', ('b2f', b
))), ('ieq', a
, b
)),
195 # -(b2f(a) + b2f(b)) < 0
196 # 0 < b2f(a) + b2f(b)
197 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
199 (('flt', ('fneg', ('fadd', ('b2f', a
), ('b2f', b
))), 0.0), ('ior', a
, b
)),
200 (('flt', 0.0, ('fadd', ('b2f', a
), ('b2f', b
))), ('ior', a
, b
)),
202 # -(b2f(a) + b2f(b)) >= 0
203 # 0 >= b2f(a) + b2f(b)
204 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
206 (('fge', ('fneg', ('fadd', ('b2f', a
), ('b2f', b
))), 0.0), ('inot', ('ior', a
, b
))),
207 (('fge', 0.0, ('fadd', ('b2f', a
), ('b2f', b
))), ('inot', ('ior', a
, b
))),
209 # Some optimizations (below) convert things like (a < b || c < b) into
210 # (min(a, c) < b). However, this interfers with the previous optimizations
211 # that try to remove comparisons with negated sums of b2f. This just
213 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', a
), ('b2f', b
)))), 0.0),
214 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
216 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
217 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
218 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
219 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
221 # Cannot remove the addition from ilt or ige due to overflow.
222 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
223 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
225 # fmin(-b2f(a), b) >= 0.0
226 # -b2f(a) >= 0.0 && b >= 0.0
227 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
228 # b2f(a) == 0.0 && b >= 0.0
229 # a == False && b >= 0.0
232 # The fge in the second replacement is not a typo. I leave the proof that
233 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
235 (('fge', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
236 (('feq', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
238 (('feq', ('b2f', a
), 0.0), ('inot', a
)),
239 (('fne', ('b2f', a
), 0.0), a
),
240 (('ieq', ('b2i', a
), 0), ('inot', a
)),
241 (('ine', ('b2i', a
), 0), a
),
243 (('fne', ('u2f32', a
), 0.0), ('ine', a
, 0)),
244 (('feq', ('u2f32', a
), 0.0), ('ieq', a
, 0)),
245 (('fge', ('u2f32', a
), 0.0), True),
246 (('fge', 0.0, ('u2f32', a
)), ('uge', 0, a
)), # ieq instead?
247 (('flt', ('u2f32', a
), 0.0), False),
248 (('flt', 0.0, ('u2f32', a
)), ('ult', 0, a
)), # ine instead?
249 (('fne', ('i2f32', a
), 0.0), ('ine', a
, 0)),
250 (('feq', ('i2f32', a
), 0.0), ('ieq', a
, 0)),
251 (('fge', ('i2f32', a
), 0.0), ('ige', a
, 0)),
252 (('fge', 0.0, ('i2f32', a
)), ('ige', 0, a
)),
253 (('flt', ('i2f32', a
), 0.0), ('ilt', a
, 0)),
254 (('flt', 0.0, ('i2f32', a
)), ('ilt', 0, a
)),
258 # fabs(a) != 0.0 because fabs(a) must be >= 0
260 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
264 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
267 # 0.0 == fabs(a) because fabs(a) must be >= 0
269 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
273 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
275 (('fmax', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('ior', a
, b
))),
276 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('ior', a
, b
)))),
277 (('fmin', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
278 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('iand', a
, b
)))),
281 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
282 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
283 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
285 # Since b is a constant, constant folding will eliminate the fmin and the
286 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
287 (('fmin', ('b2f', a
), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
289 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
291 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
292 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
293 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
294 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
295 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
296 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
297 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
298 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
299 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
300 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
301 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
302 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
303 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
310 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
311 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
312 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
313 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
314 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
315 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
316 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
317 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
318 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
319 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
320 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
321 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
322 (('fmin', a
, ('fabs', a
)), a
),
323 (('imin', a
, ('iabs', a
)), a
),
324 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
325 (('imax', a
, ('ineg', ('iabs', a
))), a
),
326 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
327 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
328 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
329 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
330 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
331 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
332 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
333 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
334 (('fsat', ('fsat', a
)), ('fsat', a
)),
335 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
336 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
337 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
338 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
339 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
340 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
341 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
342 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
343 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
344 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
345 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
346 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
347 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
348 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
349 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
350 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
351 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
352 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
353 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
354 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
355 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
356 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
358 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
359 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
360 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
361 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
362 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
363 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
364 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
365 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
366 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
367 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
368 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
369 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
370 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
371 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
372 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
373 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
375 (('ior', 'a@bool', ('ieq', a
, False)), True),
376 (('ior', 'a@bool', ('inot', a
)), True),
378 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
380 # These patterns can result when (a < b || a < c) => (a < min(b, c))
381 # transformations occur before constant propagation and loop-unrolling.
382 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
383 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
384 (('~fge', a
, ('fmin', b
, a
)), True),
385 (('~fge', ('fmax', a
, b
), a
), True),
386 (('~flt', a
, ('fmin', b
, a
)), False),
387 (('~flt', ('fmax', a
, b
), a
), False),
388 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
389 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
391 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
392 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
393 (('ige', a
, ('imin', b
, a
)), True),
394 (('ige', ('imax', a
, b
), a
), True),
395 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
396 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
397 (('uge', a
, ('umin', b
, a
)), True),
398 (('uge', ('umax', a
, b
), a
), True),
399 (('ilt', a
, ('imin', b
, a
)), False),
400 (('ilt', ('imax', a
, b
), a
), False),
401 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
402 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
403 (('ult', a
, ('umin', b
, a
)), False),
404 (('ult', ('umax', a
, b
), a
), False),
405 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
406 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
408 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
409 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
410 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
411 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
412 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
413 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
414 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
415 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
416 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
417 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
418 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
419 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
420 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
421 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
422 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
423 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
425 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
426 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
427 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
428 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
429 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
430 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
431 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
432 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
433 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
434 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
436 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
437 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
438 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
439 (('iand', 'a@bool', 1.0), ('b2f', a
), '!options->lower_b2f'),
440 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
441 (('ineg', ('b2i@32', a
)), a
),
442 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
443 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
444 # Comparison with the same args. Note that these are not done for
445 # the float versions because NaN always returns false on float
447 (('ilt', a
, a
), False),
448 (('ige', a
, a
), True),
449 (('ieq', a
, a
), True),
450 (('ine', a
, a
), False),
451 (('ult', a
, a
), False),
452 (('uge', a
, a
), True),
453 # Logical and bit operations
454 (('fand', a
, 0.0), 0.0),
456 (('iand', a
, ~
0), a
),
460 (('ior', a
, True), True),
461 (('fxor', a
, a
), 0.0),
464 (('inot', ('inot', a
)), a
),
465 (('ior', ('iand', a
, b
), b
), b
),
466 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
467 (('iand', ('ior', a
, b
), b
), b
),
468 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
470 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
471 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
472 # Shift optimizations
479 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
480 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
481 # Exponential/logarithmic identities
482 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
483 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
484 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
485 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
486 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
487 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
488 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
489 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
490 (('~fpow', a
, 1.0), a
),
491 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
492 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
493 (('~fpow', 2.0, a
), ('fexp2', a
)),
494 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
495 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
496 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
497 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
498 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
499 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
500 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
501 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
502 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
503 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
504 # Division and reciprocal
505 (('~fdiv', 1.0, a
), ('frcp', a
)),
506 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
507 (('~frcp', ('frcp', a
)), a
),
508 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
509 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
510 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
511 # Boolean simplifications
512 (('ieq', 'a@bool', True), a
),
513 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
514 (('ine', 'a@bool', False), a
),
515 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
516 (('bcsel', a
, True, False), a
),
517 (('bcsel', a
, False, True), ('inot', a
)),
518 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
519 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
520 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
521 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
522 (('bcsel', True, b
, c
), b
),
523 (('bcsel', False, b
, c
), c
),
524 (('bcsel', a
, ('b2f(is_used_once)', b
), ('b2f', c
)), ('b2f', ('bcsel', a
, b
, c
))),
525 # The result of this should be hit by constant propagation and, in the
526 # next round of opt_algebraic, get picked up by one of the above two.
527 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
529 (('bcsel', a
, b
, b
), b
),
530 (('fcsel', a
, b
, b
), b
),
533 (('i2b', ('b2i', a
)), a
),
534 (('i2b', 'a@bool'), a
),
535 (('f2i32', ('ftrunc', a
)), ('f2i32', a
)),
536 (('f2u32', ('ftrunc', a
)), ('f2u32', a
)),
537 (('i2b', ('ineg', a
)), ('i2b', a
)),
538 (('i2b', ('iabs', a
)), ('i2b', a
)),
539 (('fabs', ('b2f', a
)), ('b2f', a
)),
540 (('iabs', ('b2i', a
)), ('b2i', a
)),
541 (('inot', ('f2b', a
)), ('feq', a
, 0.0)),
543 # Ironically, mark these as imprecise because removing the conversions may
544 # preserve more precision than doing the conversions (e.g.,
545 # uint(float(0x81818181u)) == 0x81818200).
546 (('~f2i32', ('i2f32', 'a@32')), a
),
547 (('~f2i32', ('u2f32', 'a@32')), a
),
548 (('~f2u32', ('i2f32', 'a@32')), a
),
549 (('~f2u32', ('u2f32', 'a@32')), a
),
551 # Packing and then unpacking does nothing
552 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
553 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
554 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
555 ('unpack_64_2x32_split_y', a
)), a
),
558 (('ushr', ('ishl', 'a@32', 24), 24), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
559 (('ushr', ('ishl', 'a@32', 16), 24), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
560 (('ushr', ('ishl', 'a@32', 8), 24), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
561 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
562 (('ishr', ('ishl', 'a@32', 24), 24), ('extract_i8', a
, 0), '!options->lower_extract_byte'),
563 (('ishr', ('ishl', 'a@32', 16), 24), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
564 (('ishr', ('ishl', 'a@32', 8), 24), ('extract_i8', a
, 2), '!options->lower_extract_byte'),
565 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
566 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
567 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
568 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
571 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
572 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
573 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
574 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
575 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
578 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
579 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
580 (('ussub_4x8', a
, 0), a
),
581 (('ussub_4x8', a
, ~
0), 0),
582 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
583 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
584 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
585 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
586 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
587 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
588 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
589 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
591 # Propagate negation up multiplication chains
592 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
593 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
595 # Propagate constants up multiplication chains
596 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
597 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
598 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
599 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
601 # Reassociate constants in add/mul chains so they can be folded together.
602 # For now, we mostly only handle cases where the constants are separated by
603 # a single non-constant. We could do better eventually.
604 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
605 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
606 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
607 (('~fadd', '#a', ('fneg', ('fadd', b
, '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
608 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
611 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
612 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
613 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
615 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
616 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
617 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
619 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
622 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
623 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
624 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
625 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
626 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
628 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
629 ('bcsel', ('ilt', 31, 'bits'), 'insert',
630 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
631 'options->lower_bitfield_insert'),
633 # Alternative lowering that doesn't rely on bfi.
634 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
635 ('bcsel', ('ilt', 31, 'bits'),
638 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
639 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
640 'options->lower_bitfield_insert_to_shifts'),
642 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
643 (('bfm', 'bits', 'offset'),
644 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
645 'options->lower_bfm'),
647 (('ibitfield_extract', 'value', 'offset', 'bits'),
648 ('bcsel', ('ilt', 31, 'bits'), 'value',
649 ('ibfe', 'value', 'offset', 'bits')),
650 'options->lower_bitfield_extract'),
652 (('ubitfield_extract', 'value', 'offset', 'bits'),
653 ('bcsel', ('ult', 31, 'bits'), 'value',
654 ('ubfe', 'value', 'offset', 'bits')),
655 'options->lower_bitfield_extract'),
657 (('ibitfield_extract', 'value', 'offset', 'bits'),
658 ('bcsel', ('ieq', 0, 'bits'),
661 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
662 ('isub', 32, 'bits'))),
663 'options->lower_bitfield_extract_to_shifts'),
665 (('ubitfield_extract', 'value', 'offset', 'bits'),
667 ('ushr', 'value', 'offset'),
668 ('bcsel', ('ieq', 'bits', 32),
670 ('bfm', 'bits', 0))),
671 'options->lower_bitfield_extract_to_shifts'),
673 (('ifind_msb', 'value'),
674 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
675 'options->lower_ifind_msb'),
677 (('find_lsb', 'value'),
678 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
679 'options->lower_find_lsb'),
681 (('extract_i8', a
, 'b@32'),
682 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
683 'options->lower_extract_byte'),
685 (('extract_u8', a
, 'b@32'),
686 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
687 'options->lower_extract_byte'),
689 (('extract_i16', a
, 'b@32'),
690 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
691 'options->lower_extract_word'),
693 (('extract_u16', a
, 'b@32'),
694 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
695 'options->lower_extract_word'),
697 (('pack_unorm_2x16', 'v'),
698 ('pack_uvec2_to_uint',
699 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
700 'options->lower_pack_unorm_2x16'),
702 (('pack_unorm_4x8', 'v'),
703 ('pack_uvec4_to_uint',
704 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
705 'options->lower_pack_unorm_4x8'),
707 (('pack_snorm_2x16', 'v'),
708 ('pack_uvec2_to_uint',
709 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
710 'options->lower_pack_snorm_2x16'),
712 (('pack_snorm_4x8', 'v'),
713 ('pack_uvec4_to_uint',
714 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
715 'options->lower_pack_snorm_4x8'),
717 (('unpack_unorm_2x16', 'v'),
718 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
719 ('extract_u16', 'v', 1))),
721 'options->lower_unpack_unorm_2x16'),
723 (('unpack_unorm_4x8', 'v'),
724 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
725 ('extract_u8', 'v', 1),
726 ('extract_u8', 'v', 2),
727 ('extract_u8', 'v', 3))),
729 'options->lower_unpack_unorm_4x8'),
731 (('unpack_snorm_2x16', 'v'),
732 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec2', ('extract_i16', 'v', 0),
733 ('extract_i16', 'v', 1))),
735 'options->lower_unpack_snorm_2x16'),
737 (('unpack_snorm_4x8', 'v'),
738 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec4', ('extract_i8', 'v', 0),
739 ('extract_i8', 'v', 1),
740 ('extract_i8', 'v', 2),
741 ('extract_i8', 'v', 3))),
743 'options->lower_unpack_snorm_4x8'),
746 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
748 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
749 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
750 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
751 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
752 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
754 def fexp2i(exp
, bits
):
755 # We assume that exp is already in the right range.
757 return ('ishl', ('iadd', exp
, 127), 23)
759 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
763 def ldexp(f
, exp
, bits
):
764 # First, we clamp exp to a reasonable range. The maximum possible range
765 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
766 # a maximum range of [-149, 127]. This means that we can potentially have
767 # a swing of +-276. If you start with FLT_MAX, you actually have to do
768 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
769 # spec, on the other hand, only requires that we handle an exponent value
770 # in the range [-126, 128]. This implementation is *mostly* correct; it
771 # handles a range on exp of [-252, 254] which allows you to create any
772 # value (including denorms if the hardware supports it) and to adjust the
773 # exponent of any normal value to anything you want.
775 exp
= ('imin', ('imax', exp
, -252), 254)
777 exp
= ('imin', ('imax', exp
, -2044), 2046)
781 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
782 # (We use ishr which isn't the same for -1, but the -1 case still works
783 # since we use exp-exp/2 as the second exponent.) While the spec
784 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
785 # work with denormals and doesn't allow for the full swing in exponents
786 # that you can get with normalized values. Instead, we create two powers
787 # of two and multiply by them each in turn. That way the effective range
788 # of our exponent is doubled.
789 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
790 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
791 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
794 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
795 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
798 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
799 def bitfield_reverse(u
):
800 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
801 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
802 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
803 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
804 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
808 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
810 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
811 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
812 # and, if a is a NaN then the second comparison will fail anyway.
813 for op
in ['flt', 'fge', 'feq']:
815 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
816 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
819 # Add optimizations to handle the case where the result of a ternary is
820 # compared to a constant. This way we can take things like
826 # a ? (0 > 0) : (1 > 0)
828 # which constant folding will eat for lunch. The resulting ternary will
829 # further get cleaned up by the boolean reductions above and we will be
830 # left with just the original variable "a".
831 for op
in ['flt', 'fge', 'feq', 'fne',
832 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
834 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
835 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
836 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
837 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
841 # For example, this converts things like
843 # 1 + mix(0, a - 1, condition)
847 # mix(1, (a-1)+1, condition)
849 # Other optimizations will rearrange the constants.
850 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
852 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
855 # This section contains "late" optimizations that should be run before
856 # creating ffmas and calling regular optimizations for the final time.
857 # Optimizations should go here if they help code generation and conflict
858 # with the regular optimizations.
859 before_ffma_optimizations
= [
860 # Propagate constants down multiplication chains
861 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
862 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
863 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
864 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
866 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
867 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
868 (('~fadd', ('fneg', a
), a
), 0.0),
869 (('iadd', ('ineg', a
), a
), 0),
870 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
871 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
872 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
873 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
876 # This section contains "late" optimizations that should be run after the
877 # regular optimizations have finished. Optimizations should go here if
878 # they help code generation but do not necessarily produce code that is
879 # more easily optimizable.
880 late_optimizations
= [
881 # Most of these optimizations aren't quite safe when you get infinity or
882 # Nan involved but the first one should be fine.
883 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
884 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
885 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
886 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
887 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
888 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
890 (('~fge', ('fmin(is_used_once)', ('fadd(is_used_once)', a
, b
), ('fadd', c
, d
)), 0.0), ('iand', ('fge', a
, ('fneg', b
)), ('fge', c
, ('fneg', d
)))),
892 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
893 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
894 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
895 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
897 (('b2f(is_used_more_than_once)', ('inot', a
)), ('bcsel', a
, 0.0, 1.0)),
898 (('fneg(is_used_more_than_once)', ('b2f', ('inot', a
))), ('bcsel', a
, -0.0, -1.0)),
900 # we do these late so that we don't get in the way of creating ffmas
901 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
902 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
904 # Lowered for backends without a dedicated b2f instruction
905 (('b2f@32', a
), ('iand', a
, 1.0), 'options->lower_b2f'),
908 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
909 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
910 before_ffma_optimizations
).render())
911 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
912 late_optimizations
).render())