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
30 from nir_opcodes
import type_sizes
33 # Convenience variables
39 # Written in the form (<search>, <replace>) where <search> is an expression
40 # and <replace> is either an expression or a value. An expression is
41 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
42 # where each source is either an expression or a value. A value can be
43 # either a numeric constant or a string representing a variable name.
45 # If the opcode in a search expression is prefixed by a '~' character, this
46 # indicates that the operation is inexact. Such operations will only get
47 # applied to SSA values that do not have the exact bit set. This should be
48 # used by by any optimizations that are not bit-for-bit exact. It should not,
49 # however, be used for backend-requested lowering operations as those need to
50 # happen regardless of precision.
52 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
53 # that the given variable will only match constants and the type indicates that
54 # the given variable will only match values from ALU instructions with the
55 # given output type, and (cond) specifies an additional condition function
56 # (see nir_search_helpers.h).
58 # For constants, you have to be careful to make sure that it is the right
59 # type because python is unaware of the source and destination types of the
62 # All expression types can have a bit-size specified. For opcodes, this
63 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
64 # type and size. In the search half of the expression this indicates that it
65 # should only match that particular bit-size. In the replace half of the
66 # expression this indicates that the constructed value should have that
71 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
72 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
77 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
78 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
79 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
80 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
82 (('fneg', ('fneg', a
)), a
),
83 (('ineg', ('ineg', a
)), a
),
84 (('fabs', ('fabs', a
)), ('fabs', a
)),
85 (('fabs', ('fneg', a
)), ('fabs', a
)),
86 (('fabs', ('u2f', a
)), ('u2f', a
)),
87 (('iabs', ('iabs', a
)), ('iabs', a
)),
88 (('iabs', ('ineg', a
)), ('iabs', a
)),
89 (('f2b', ('fneg', a
)), ('f2b', a
)),
90 (('i2b', ('ineg', a
)), ('i2b', a
)),
91 (('~fadd', a
, 0.0), a
),
93 (('usadd_4x8', a
, 0), a
),
94 (('usadd_4x8', a
, ~
0), ~
0),
95 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
96 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
97 (('~fadd', ('fneg', a
), a
), 0.0),
98 (('iadd', ('ineg', a
), a
), 0),
99 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
100 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
101 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
102 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
103 (('~fmul', a
, 0.0), 0.0),
105 (('umul_unorm_4x8', a
, 0), 0),
106 (('umul_unorm_4x8', a
, ~
0), a
),
107 (('fmul', a
, 1.0), a
),
109 (('fmul', a
, -1.0), ('fneg', a
)),
110 (('imul', a
, -1), ('ineg', a
)),
111 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
112 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
113 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
114 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
115 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
116 (('~ffma', 0.0, a
, b
), b
),
117 (('~ffma', a
, 0.0, b
), b
),
118 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
119 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
120 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
121 (('~flrp', a
, b
, 0.0), a
),
122 (('~flrp', a
, b
, 1.0), b
),
123 (('~flrp', a
, a
, b
), a
),
124 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
125 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
126 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
127 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
128 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
129 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
130 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
131 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
132 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
133 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
134 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
135 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
136 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
137 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
138 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
139 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
141 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
142 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
143 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
144 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
146 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
147 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
149 # (a * #b + #c) << #d
150 # ((a * #b) << #d) + (#c << #d)
151 # (a * (#b << #d)) + (#c << #d)
152 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
153 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
157 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
159 # Comparison simplifications
160 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
161 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
162 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
163 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
164 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
165 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
166 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
167 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
168 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
169 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
173 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
175 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
177 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
179 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
180 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
181 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
182 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
183 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
184 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
185 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
186 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
187 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
188 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
189 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
190 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
191 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
192 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
193 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
194 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
195 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
196 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
197 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
198 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
200 # -(b2f(a) + b2f(b)) < 0
201 # 0 < b2f(a) + b2f(b)
202 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
204 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
205 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
207 # -(b2f(a) + b2f(b)) >= 0
208 # 0 >= b2f(a) + b2f(b)
209 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
211 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
212 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
214 # Some optimizations (below) convert things like (a < b || c < b) into
215 # (min(a, c) < b). However, this interfers with the previous optimizations
216 # that try to remove comparisons with negated sums of b2f. This just
218 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
219 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
221 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
222 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
223 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
224 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
226 # Cannot remove the addition from ilt or ige due to overflow.
227 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
228 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
230 # fmin(-b2f(a), b) >= 0.0
231 # -b2f(a) >= 0.0 && b >= 0.0
232 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
233 # b2f(a) == 0.0 && b >= 0.0
234 # a == False && b >= 0.0
237 # The fge in the second replacement is not a typo. I leave the proof that
238 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
240 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
241 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
243 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
244 (('fne', ('b2f', 'a@1'), 0.0), a
),
245 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
246 (('ine', ('b2i', 'a@1'), 0), a
),
248 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
249 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
250 (('fge', ('u2f', a
), 0.0), True),
251 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
252 (('flt', ('u2f', a
), 0.0), False),
253 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
254 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
255 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
256 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
257 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
258 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
259 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
263 # fabs(a) != 0.0 because fabs(a) must be >= 0
265 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
269 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
272 # 0.0 == fabs(a) because fabs(a) must be >= 0
274 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
278 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
280 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
281 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
282 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
283 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
286 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
287 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
288 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
290 # Since b is a constant, constant folding will eliminate the fmin and the
291 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
292 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
294 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
296 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
297 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
298 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
299 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
300 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
301 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
302 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
303 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
304 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
305 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
306 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
307 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
308 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
315 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
316 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
317 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
318 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
319 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
320 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
321 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
322 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
323 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
324 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
325 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
326 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
327 (('fmin', a
, ('fabs', a
)), a
),
328 (('imin', a
, ('iabs', a
)), a
),
329 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
330 (('imax', a
, ('ineg', ('iabs', a
))), a
),
331 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
332 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
333 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
334 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
335 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
336 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
337 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
338 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
339 (('fsat', ('fsat', a
)), ('fsat', a
)),
340 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
341 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
342 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
343 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
344 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
345 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
346 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
347 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
348 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
349 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
350 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
351 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
352 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
353 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
354 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
355 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
356 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
357 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
358 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
359 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
360 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
361 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
363 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
364 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
365 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
366 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
367 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
368 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
369 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
370 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
371 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
372 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
373 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
374 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
375 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
376 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
377 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
378 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
380 (('ior', 'a@bool', ('ieq', a
, False)), True),
381 (('ior', a
, ('inot', a
)), -1),
383 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
385 # These patterns can result when (a < b || a < c) => (a < min(b, c))
386 # transformations occur before constant propagation and loop-unrolling.
387 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
388 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
389 (('~fge', a
, ('fmin', b
, a
)), True),
390 (('~fge', ('fmax', a
, b
), a
), True),
391 (('~flt', a
, ('fmin', b
, a
)), False),
392 (('~flt', ('fmax', a
, b
), a
), False),
393 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
394 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
396 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
397 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
398 (('ige', a
, ('imin', b
, a
)), True),
399 (('ige', ('imax', a
, b
), a
), True),
400 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
401 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
402 (('uge', a
, ('umin', b
, a
)), True),
403 (('uge', ('umax', a
, b
), a
), True),
404 (('ilt', a
, ('imin', b
, a
)), False),
405 (('ilt', ('imax', a
, b
), a
), False),
406 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
407 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
408 (('ult', a
, ('umin', b
, a
)), False),
409 (('ult', ('umax', a
, b
), a
), False),
410 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
411 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
413 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
414 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
415 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
416 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
417 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
418 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
419 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
420 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
421 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
422 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
423 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
424 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
425 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
426 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
427 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
428 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
430 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
431 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
432 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
433 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
434 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
435 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
436 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
437 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
438 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
439 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
441 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
442 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
443 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
444 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
445 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
446 (('ineg', ('b2i32', 'a@32')), a
),
447 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
448 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
449 # Comparison with the same args. Note that these are not done for
450 # the float versions because NaN always returns false on float
452 (('ilt', a
, a
), False),
453 (('ige', a
, a
), True),
454 (('ieq', a
, a
), True),
455 (('ine', a
, a
), False),
456 (('ult', a
, a
), False),
457 (('uge', a
, a
), True),
458 # Logical and bit operations
459 (('fand', a
, 0.0), 0.0),
461 (('iand', a
, ~
0), a
),
465 (('ior', a
, True), True),
466 (('fxor', a
, a
), 0.0),
469 (('inot', ('inot', a
)), a
),
470 (('ior', ('iand', a
, b
), b
), b
),
471 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
472 (('iand', ('ior', a
, b
), b
), b
),
473 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
475 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
476 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
477 # Shift optimizations
484 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
485 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
486 # Exponential/logarithmic identities
487 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
488 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
489 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
490 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
491 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
492 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
493 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
494 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
495 (('~fpow', a
, 1.0), a
),
496 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
497 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
498 (('~fpow', 2.0, a
), ('fexp2', a
)),
499 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
500 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
501 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
502 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
503 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
504 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
505 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
506 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
507 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
508 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
509 # Division and reciprocal
510 (('~fdiv', 1.0, a
), ('frcp', a
)),
511 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
512 (('~frcp', ('frcp', a
)), a
),
513 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
514 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
515 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
516 # Boolean simplifications
517 (('ieq', 'a@bool', True), a
),
518 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
519 (('ine', 'a@bool', False), a
),
520 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
521 (('bcsel', a
, True, False), a
),
522 (('bcsel', a
, False, True), ('inot', a
)),
523 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
524 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
525 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
526 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
527 (('bcsel', True, b
, c
), b
),
528 (('bcsel', False, b
, c
), c
),
529 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
530 # The result of this should be hit by constant propagation and, in the
531 # next round of opt_algebraic, get picked up by one of the above two.
532 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
534 (('bcsel', a
, b
, b
), b
),
535 (('fcsel', a
, b
, b
), b
),
537 # D3D Boolean emulation
538 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
539 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
540 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
541 ('ineg', ('b2i', ('iand', a
, b
)))),
542 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
543 ('ineg', ('b2i', ('ior', a
, b
)))),
544 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
545 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
546 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
547 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
548 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
551 (('i2b32', ('b2i', 'a@32')), a
),
552 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
553 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
554 (('i2b', ('ineg', a
)), ('i2b', a
)),
555 (('i2b', ('iabs', a
)), ('i2b', a
)),
556 (('fabs', ('b2f', a
)), ('b2f', a
)),
557 (('iabs', ('b2i', a
)), ('b2i', a
)),
558 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
560 # Ironically, mark these as imprecise because removing the conversions may
561 # preserve more precision than doing the conversions (e.g.,
562 # uint(float(0x81818181u)) == 0x81818200).
563 (('~f2i32', ('i2f', 'a@32')), a
),
564 (('~f2i32', ('u2f', 'a@32')), a
),
565 (('~f2u32', ('i2f', 'a@32')), a
),
566 (('~f2u32', ('u2f', 'a@32')), a
),
568 # Packing and then unpacking does nothing
569 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
570 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
571 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
572 ('unpack_64_2x32_split_y', a
)), a
),
575 (('ushr', ('ishl', 'a@32', 24), 24), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
576 (('ushr', ('ishl', 'a@32', 16), 24), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
577 (('ushr', ('ishl', 'a@32', 8), 24), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
578 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
579 (('ishr', ('ishl', 'a@32', 24), 24), ('extract_i8', a
, 0), '!options->lower_extract_byte'),
580 (('ishr', ('ishl', 'a@32', 16), 24), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
581 (('ishr', ('ishl', 'a@32', 8), 24), ('extract_i8', a
, 2), '!options->lower_extract_byte'),
582 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
583 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
584 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
585 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
588 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
589 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
590 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
591 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
592 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
595 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
596 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
597 (('ussub_4x8', a
, 0), a
),
598 (('ussub_4x8', a
, ~
0), 0),
599 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
600 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
601 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
602 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
603 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
604 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
605 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
606 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
608 # Propagate negation up multiplication chains
609 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
610 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
612 # Propagate constants up multiplication chains
613 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
614 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
615 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
616 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
618 # Reassociate constants in add/mul chains so they can be folded together.
619 # For now, we mostly only handle cases where the constants are separated by
620 # a single non-constant. We could do better eventually.
621 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
622 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
623 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
624 (('~fadd', '#a', ('fneg', ('fadd', b
, '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
625 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
628 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
629 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
630 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
632 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
633 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
634 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
636 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
639 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
640 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
641 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
642 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
643 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
645 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
646 ('bcsel', ('ilt', 31, 'bits'), 'insert',
647 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
648 'options->lower_bitfield_insert'),
650 # Alternative lowering that doesn't rely on bfi.
651 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
652 ('bcsel', ('ilt', 31, 'bits'),
655 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
656 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
657 'options->lower_bitfield_insert_to_shifts'),
659 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
660 (('bfm', 'bits', 'offset'),
661 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
662 'options->lower_bfm'),
664 (('ibitfield_extract', 'value', 'offset', 'bits'),
665 ('bcsel', ('ilt', 31, 'bits'), 'value',
666 ('ibfe', 'value', 'offset', 'bits')),
667 'options->lower_bitfield_extract'),
669 (('ubitfield_extract', 'value', 'offset', 'bits'),
670 ('bcsel', ('ult', 31, 'bits'), 'value',
671 ('ubfe', 'value', 'offset', 'bits')),
672 'options->lower_bitfield_extract'),
674 (('ibitfield_extract', 'value', 'offset', 'bits'),
675 ('bcsel', ('ieq', 0, 'bits'),
678 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
679 ('isub', 32, 'bits'))),
680 'options->lower_bitfield_extract_to_shifts'),
682 (('ubitfield_extract', 'value', 'offset', 'bits'),
684 ('ushr', 'value', 'offset'),
685 ('bcsel', ('ieq', 'bits', 32),
687 ('bfm', 'bits', 0))),
688 'options->lower_bitfield_extract_to_shifts'),
690 (('ifind_msb', 'value'),
691 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
692 'options->lower_ifind_msb'),
694 (('find_lsb', 'value'),
695 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
696 'options->lower_find_lsb'),
698 (('extract_i8', a
, 'b@32'),
699 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
700 'options->lower_extract_byte'),
702 (('extract_u8', a
, 'b@32'),
703 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
704 'options->lower_extract_byte'),
706 (('extract_i16', a
, 'b@32'),
707 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
708 'options->lower_extract_word'),
710 (('extract_u16', a
, 'b@32'),
711 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
712 'options->lower_extract_word'),
714 (('pack_unorm_2x16', 'v'),
715 ('pack_uvec2_to_uint',
716 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
717 'options->lower_pack_unorm_2x16'),
719 (('pack_unorm_4x8', 'v'),
720 ('pack_uvec4_to_uint',
721 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
722 'options->lower_pack_unorm_4x8'),
724 (('pack_snorm_2x16', 'v'),
725 ('pack_uvec2_to_uint',
726 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
727 'options->lower_pack_snorm_2x16'),
729 (('pack_snorm_4x8', 'v'),
730 ('pack_uvec4_to_uint',
731 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
732 'options->lower_pack_snorm_4x8'),
734 (('unpack_unorm_2x16', 'v'),
735 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
736 ('extract_u16', 'v', 1))),
738 'options->lower_unpack_unorm_2x16'),
740 (('unpack_unorm_4x8', 'v'),
741 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
742 ('extract_u8', 'v', 1),
743 ('extract_u8', 'v', 2),
744 ('extract_u8', 'v', 3))),
746 'options->lower_unpack_unorm_4x8'),
748 (('unpack_snorm_2x16', 'v'),
749 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
750 ('extract_i16', 'v', 1))),
752 'options->lower_unpack_snorm_2x16'),
754 (('unpack_snorm_4x8', 'v'),
755 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
756 ('extract_i8', 'v', 1),
757 ('extract_i8', 'v', 2),
758 ('extract_i8', 'v', 3))),
760 'options->lower_unpack_snorm_4x8'),
762 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
765 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
767 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
768 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
769 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
770 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
771 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
773 # Optimize x2bN(b2x(x)) -> x
774 for size
in type_sizes('bool'):
775 aN
= 'a@' + str(size
)
776 f2bN
= 'f2b' + str(size
)
777 i2bN
= 'i2b' + str(size
)
778 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
779 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
781 # Optimize x2yN(b2x(x)) -> b2y
782 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
783 if x
!= 'f' and y
!= 'f' and x
!= y
:
786 b2x
= 'b2f' if x
== 'f' else 'b2i'
787 b2y
= 'b2f' if y
== 'f' else 'b2i'
788 x2yN
= '{}2{}'.format(x
, y
)
789 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
791 def fexp2i(exp
, bits
):
792 # We assume that exp is already in the right range.
794 return ('ishl', ('iadd', exp
, 127), 23)
796 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
800 def ldexp(f
, exp
, bits
):
801 # First, we clamp exp to a reasonable range. The maximum possible range
802 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
803 # a maximum range of [-149, 127]. This means that we can potentially have
804 # a swing of +-276. If you start with FLT_MAX, you actually have to do
805 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
806 # spec, on the other hand, only requires that we handle an exponent value
807 # in the range [-126, 128]. This implementation is *mostly* correct; it
808 # handles a range on exp of [-252, 254] which allows you to create any
809 # value (including denorms if the hardware supports it) and to adjust the
810 # exponent of any normal value to anything you want.
812 exp
= ('imin', ('imax', exp
, -252), 254)
814 exp
= ('imin', ('imax', exp
, -2044), 2046)
818 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
819 # (We use ishr which isn't the same for -1, but the -1 case still works
820 # since we use exp-exp/2 as the second exponent.) While the spec
821 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
822 # work with denormals and doesn't allow for the full swing in exponents
823 # that you can get with normalized values. Instead, we create two powers
824 # of two and multiply by them each in turn. That way the effective range
825 # of our exponent is doubled.
826 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
827 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
828 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
831 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
832 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
835 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
836 def bitfield_reverse(u
):
837 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
838 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
839 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
840 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
841 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
845 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
847 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
848 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
849 # and, if a is a NaN then the second comparison will fail anyway.
850 for op
in ['flt', 'fge', 'feq']:
852 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
853 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
856 # Add optimizations to handle the case where the result of a ternary is
857 # compared to a constant. This way we can take things like
863 # a ? (0 > 0) : (1 > 0)
865 # which constant folding will eat for lunch. The resulting ternary will
866 # further get cleaned up by the boolean reductions above and we will be
867 # left with just the original variable "a".
868 for op
in ['flt', 'fge', 'feq', 'fne',
869 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
871 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
872 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
873 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
874 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
878 # For example, this converts things like
880 # 1 + mix(0, a - 1, condition)
884 # mix(1, (a-1)+1, condition)
886 # Other optimizations will rearrange the constants.
887 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
889 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
892 # This section contains "late" optimizations that should be run before
893 # creating ffmas and calling regular optimizations for the final time.
894 # Optimizations should go here if they help code generation and conflict
895 # with the regular optimizations.
896 before_ffma_optimizations
= [
897 # Propagate constants down multiplication chains
898 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
899 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
900 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
901 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
903 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
904 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
905 (('~fadd', ('fneg', a
), a
), 0.0),
906 (('iadd', ('ineg', a
), a
), 0),
907 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
908 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
909 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
910 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
913 # This section contains "late" optimizations that should be run after the
914 # regular optimizations have finished. Optimizations should go here if
915 # they help code generation but do not necessarily produce code that is
916 # more easily optimizable.
917 late_optimizations
= [
918 # Most of these optimizations aren't quite safe when you get infinity or
919 # Nan involved but the first one should be fine.
920 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
921 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
922 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
923 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
924 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
925 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
927 (('~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
)))),
929 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
930 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
931 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
932 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
934 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a
, 0.0, 1.0)),
935 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a
, -0.0, -1.0)),
937 # we do these late so that we don't get in the way of creating ffmas
938 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
939 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
942 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
943 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
944 before_ffma_optimizations
).render())
945 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
946 late_optimizations
).render())