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 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
131 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
132 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
133 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
134 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
135 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
136 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
137 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
138 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
139 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
140 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
142 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
143 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
144 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
145 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
147 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
148 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
150 # (a * #b + #c) << #d
151 # ((a * #b) << #d) + (#c << #d)
152 # (a * (#b << #d)) + (#c << #d)
153 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
154 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
158 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
160 # Comparison simplifications
161 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
162 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
163 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
164 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
165 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
166 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
167 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
168 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
169 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
170 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
174 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
176 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
178 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
180 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
181 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
182 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
183 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
184 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
185 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
186 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
187 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
188 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
189 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
190 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
191 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
192 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
193 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
194 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
195 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
196 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
197 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
198 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
199 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
201 # -(b2f(a) + b2f(b)) < 0
202 # 0 < b2f(a) + b2f(b)
203 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
205 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
206 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
208 # -(b2f(a) + b2f(b)) >= 0
209 # 0 >= b2f(a) + b2f(b)
210 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
212 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
213 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
215 # Some optimizations (below) convert things like (a < b || c < b) into
216 # (min(a, c) < b). However, this interfers with the previous optimizations
217 # that try to remove comparisons with negated sums of b2f. This just
219 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
220 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
222 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
223 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
224 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
225 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
227 # Cannot remove the addition from ilt or ige due to overflow.
228 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
229 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
231 # fmin(-b2f(a), b) >= 0.0
232 # -b2f(a) >= 0.0 && b >= 0.0
233 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
234 # b2f(a) == 0.0 && b >= 0.0
235 # a == False && b >= 0.0
238 # The fge in the second replacement is not a typo. I leave the proof that
239 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
241 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
242 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
244 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
245 (('fne', ('b2f', 'a@1'), 0.0), a
),
246 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
247 (('ine', ('b2i', 'a@1'), 0), a
),
249 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
250 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
251 (('fge', ('u2f', a
), 0.0), True),
252 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
253 (('flt', ('u2f', a
), 0.0), False),
254 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
255 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
256 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
257 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
258 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
259 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
260 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
264 # fabs(a) != 0.0 because fabs(a) must be >= 0
266 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
270 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
273 # 0.0 == fabs(a) because fabs(a) must be >= 0
275 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
279 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
281 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
282 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
283 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
284 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
287 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
288 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
289 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
291 # Since b is a constant, constant folding will eliminate the fmin and the
292 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
293 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
295 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
297 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
298 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
299 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
300 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
301 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
302 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
303 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
304 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
305 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
306 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
307 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
308 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
309 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
310 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
317 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
318 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
319 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
320 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
321 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
322 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
323 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
324 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
325 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
326 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
327 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
328 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
329 (('fmin', a
, ('fabs', a
)), a
),
330 (('imin', a
, ('iabs', a
)), a
),
331 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
332 (('imax', a
, ('ineg', ('iabs', a
))), a
),
333 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
334 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
335 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
336 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
337 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
338 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
339 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
340 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
341 (('fsat', ('fsat', a
)), ('fsat', a
)),
342 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
343 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
344 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
345 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
346 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
347 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
348 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
349 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
350 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
351 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
352 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
353 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
354 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
355 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
356 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
357 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
358 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
359 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
360 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
361 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
362 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
363 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
365 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
366 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
367 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
368 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
369 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
370 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
371 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
372 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
373 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
374 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
375 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
376 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
377 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
378 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
379 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
380 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
382 # Common pattern like 'if (i == 0 || i == 1 || ...)'
383 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
384 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
385 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
387 (('ior', 'a@bool', ('ieq', a
, False)), True),
388 (('ior', a
, ('inot', a
)), -1),
390 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
392 # These patterns can result when (a < b || a < c) => (a < min(b, c))
393 # transformations occur before constant propagation and loop-unrolling.
394 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
395 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
396 (('~fge', a
, ('fmin', b
, a
)), True),
397 (('~fge', ('fmax', a
, b
), a
), True),
398 (('~flt', a
, ('fmin', b
, a
)), False),
399 (('~flt', ('fmax', a
, b
), a
), False),
400 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
401 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
403 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
404 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
405 (('ige', a
, ('imin', b
, a
)), True),
406 (('ige', ('imax', a
, b
), a
), True),
407 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
408 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
409 (('uge', a
, ('umin', b
, a
)), True),
410 (('uge', ('umax', a
, b
), a
), True),
411 (('ilt', a
, ('imin', b
, a
)), False),
412 (('ilt', ('imax', a
, b
), a
), False),
413 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
414 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
415 (('ult', a
, ('umin', b
, a
)), False),
416 (('ult', ('umax', a
, b
), a
), False),
417 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
418 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
420 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
421 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
422 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
423 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
424 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
425 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
426 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
427 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
428 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
429 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
430 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
431 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
432 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
433 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
434 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
435 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
437 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
438 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
439 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
440 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
441 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
442 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
443 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
444 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
445 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
446 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
448 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
449 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
450 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
451 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
452 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
453 (('ineg', ('b2i32', 'a@32')), a
),
454 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
455 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
456 # Comparison with the same args. Note that these are not done for
457 # the float versions because NaN always returns false on float
459 (('ilt', a
, a
), False),
460 (('ige', a
, a
), True),
461 (('ieq', a
, a
), True),
462 (('ine', a
, a
), False),
463 (('ult', a
, a
), False),
464 (('uge', a
, a
), True),
465 # Logical and bit operations
466 (('fand', a
, 0.0), 0.0),
468 (('iand', a
, ~
0), a
),
472 (('ior', a
, True), True),
473 (('fxor', a
, a
), 0.0),
476 (('inot', ('inot', a
)), a
),
477 (('ior', ('iand', a
, b
), b
), b
),
478 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
479 (('iand', ('ior', a
, b
), b
), b
),
480 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
482 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
483 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
484 # Shift optimizations
491 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
492 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
493 # Exponential/logarithmic identities
494 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
495 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
496 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
497 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
498 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
499 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
500 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
501 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
502 (('~fpow', a
, 1.0), a
),
503 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
504 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
505 (('~fpow', 2.0, a
), ('fexp2', a
)),
506 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
507 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
508 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
509 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
510 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
511 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
512 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
513 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
514 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
515 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
516 # Division and reciprocal
517 (('~fdiv', 1.0, a
), ('frcp', a
)),
518 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
519 (('~frcp', ('frcp', a
)), a
),
520 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
521 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
522 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
523 # Boolean simplifications
524 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
525 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
526 (('ieq', 'a@bool', True), a
),
527 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
528 (('ine', 'a@bool', False), a
),
529 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
530 (('bcsel', a
, True, False), a
),
531 (('bcsel', a
, False, True), ('inot', a
)),
532 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
533 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
534 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
535 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
536 (('bcsel', True, b
, c
), b
),
537 (('bcsel', False, b
, c
), c
),
538 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
539 # The result of this should be hit by constant propagation and, in the
540 # next round of opt_algebraic, get picked up by one of the above two.
541 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
543 (('bcsel', a
, b
, b
), b
),
544 (('fcsel', a
, b
, b
), b
),
546 # D3D Boolean emulation
547 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
548 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
549 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
550 ('ineg', ('b2i', ('iand', a
, b
)))),
551 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
552 ('ineg', ('b2i', ('ior', a
, b
)))),
553 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
554 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
555 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
556 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
557 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
559 # SM5 32-bit shifts are defined to use the 5 least significant bits
560 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
561 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
562 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
565 (('i2b32', ('b2i', 'a@32')), a
),
566 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
567 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
568 (('i2b', ('ineg', a
)), ('i2b', a
)),
569 (('i2b', ('iabs', a
)), ('i2b', a
)),
570 (('fabs', ('b2f', a
)), ('b2f', a
)),
571 (('iabs', ('b2i', a
)), ('b2i', a
)),
572 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
574 # Ironically, mark these as imprecise because removing the conversions may
575 # preserve more precision than doing the conversions (e.g.,
576 # uint(float(0x81818181u)) == 0x81818200).
577 (('~f2i32', ('i2f', 'a@32')), a
),
578 (('~f2i32', ('u2f', 'a@32')), a
),
579 (('~f2u32', ('i2f', 'a@32')), a
),
580 (('~f2u32', ('u2f', 'a@32')), a
),
582 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
585 # It is undefined to convert a negative floating-point value to an
588 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
589 # some optimizations in the i965 backend to proceed.
590 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
591 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
592 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
593 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
595 # Packing and then unpacking does nothing
596 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
597 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
598 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
599 ('unpack_64_2x32_split_y', a
)), a
),
602 (('ushr', ('ishl', 'a@32', 24), 24), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
603 (('ushr', ('ishl', 'a@32', 16), 24), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
604 (('ushr', ('ishl', 'a@32', 8), 24), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
605 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
606 (('ishr', ('ishl', 'a@32', 24), 24), ('extract_i8', a
, 0), '!options->lower_extract_byte'),
607 (('ishr', ('ishl', 'a@32', 16), 24), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
608 (('ishr', ('ishl', 'a@32', 8), 24), ('extract_i8', a
, 2), '!options->lower_extract_byte'),
609 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
610 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
611 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
612 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
615 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
616 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
617 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
618 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
619 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
622 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
623 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
624 (('ussub_4x8', a
, 0), a
),
625 (('ussub_4x8', a
, ~
0), 0),
626 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
627 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
628 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
629 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
630 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
631 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
632 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
633 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
635 # Propagate negation up multiplication chains
636 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
637 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
639 # Propagate constants up multiplication chains
640 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
641 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
642 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
643 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
645 # Reassociate constants in add/mul chains so they can be folded together.
646 # For now, we mostly only handle cases where the constants are separated by
647 # a single non-constant. We could do better eventually.
648 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
649 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
650 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
651 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
652 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
655 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
656 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
657 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
659 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
660 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
661 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
663 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
666 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
667 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
668 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
669 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
670 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
672 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
673 ('bcsel', ('ilt', 31, 'bits'), 'insert',
674 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
675 'options->lower_bitfield_insert'),
677 # Alternative lowering that doesn't rely on bfi.
678 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
679 ('bcsel', ('ilt', 31, 'bits'),
682 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
683 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
684 'options->lower_bitfield_insert_to_shifts'),
686 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
687 (('bfm', 'bits', 'offset'),
688 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
689 'options->lower_bfm'),
691 (('ibitfield_extract', 'value', 'offset', 'bits'),
692 ('bcsel', ('ilt', 31, 'bits'), 'value',
693 ('ibfe', 'value', 'offset', 'bits')),
694 'options->lower_bitfield_extract'),
696 (('ubitfield_extract', 'value', 'offset', 'bits'),
697 ('bcsel', ('ult', 31, 'bits'), 'value',
698 ('ubfe', 'value', 'offset', 'bits')),
699 'options->lower_bitfield_extract'),
701 (('ibitfield_extract', 'value', 'offset', 'bits'),
702 ('bcsel', ('ieq', 0, 'bits'),
705 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
706 ('isub', 32, 'bits'))),
707 'options->lower_bitfield_extract_to_shifts'),
709 (('ubitfield_extract', 'value', 'offset', 'bits'),
711 ('ushr', 'value', 'offset'),
712 ('bcsel', ('ieq', 'bits', 32),
714 ('bfm', 'bits', 0))),
715 'options->lower_bitfield_extract_to_shifts'),
717 (('ifind_msb', 'value'),
718 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
719 'options->lower_ifind_msb'),
721 (('find_lsb', 'value'),
722 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
723 'options->lower_find_lsb'),
725 (('extract_i8', a
, 'b@32'),
726 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
727 'options->lower_extract_byte'),
729 (('extract_u8', a
, 'b@32'),
730 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
731 'options->lower_extract_byte'),
733 (('extract_i16', a
, 'b@32'),
734 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
735 'options->lower_extract_word'),
737 (('extract_u16', a
, 'b@32'),
738 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
739 'options->lower_extract_word'),
741 (('pack_unorm_2x16', 'v'),
742 ('pack_uvec2_to_uint',
743 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
744 'options->lower_pack_unorm_2x16'),
746 (('pack_unorm_4x8', 'v'),
747 ('pack_uvec4_to_uint',
748 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
749 'options->lower_pack_unorm_4x8'),
751 (('pack_snorm_2x16', 'v'),
752 ('pack_uvec2_to_uint',
753 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
754 'options->lower_pack_snorm_2x16'),
756 (('pack_snorm_4x8', 'v'),
757 ('pack_uvec4_to_uint',
758 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
759 'options->lower_pack_snorm_4x8'),
761 (('unpack_unorm_2x16', 'v'),
762 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
763 ('extract_u16', 'v', 1))),
765 'options->lower_unpack_unorm_2x16'),
767 (('unpack_unorm_4x8', 'v'),
768 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
769 ('extract_u8', 'v', 1),
770 ('extract_u8', 'v', 2),
771 ('extract_u8', 'v', 3))),
773 'options->lower_unpack_unorm_4x8'),
775 (('unpack_snorm_2x16', 'v'),
776 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
777 ('extract_i16', 'v', 1))),
779 'options->lower_unpack_snorm_2x16'),
781 (('unpack_snorm_4x8', 'v'),
782 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
783 ('extract_i8', 'v', 1),
784 ('extract_i8', 'v', 2),
785 ('extract_i8', 'v', 3))),
787 'options->lower_unpack_snorm_4x8'),
789 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
792 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
794 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
795 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
796 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
797 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
798 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
800 # Optimize x2bN(b2x(x)) -> x
801 for size
in type_sizes('bool'):
802 aN
= 'a@' + str(size
)
803 f2bN
= 'f2b' + str(size
)
804 i2bN
= 'i2b' + str(size
)
805 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
806 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
808 # Optimize x2yN(b2x(x)) -> b2y
809 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
810 if x
!= 'f' and y
!= 'f' and x
!= y
:
813 b2x
= 'b2f' if x
== 'f' else 'b2i'
814 b2y
= 'b2f' if y
== 'f' else 'b2i'
815 x2yN
= '{}2{}'.format(x
, y
)
816 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
818 def fexp2i(exp
, bits
):
819 # We assume that exp is already in the right range.
821 return ('ishl', ('iadd', exp
, 127), 23)
823 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
827 def ldexp(f
, exp
, bits
):
828 # First, we clamp exp to a reasonable range. The maximum possible range
829 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
830 # a maximum range of [-149, 127]. This means that we can potentially have
831 # a swing of +-276. If you start with FLT_MAX, you actually have to do
832 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
833 # spec, on the other hand, only requires that we handle an exponent value
834 # in the range [-126, 128]. This implementation is *mostly* correct; it
835 # handles a range on exp of [-252, 254] which allows you to create any
836 # value (including denorms if the hardware supports it) and to adjust the
837 # exponent of any normal value to anything you want.
839 exp
= ('imin', ('imax', exp
, -252), 254)
841 exp
= ('imin', ('imax', exp
, -2044), 2046)
845 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
846 # (We use ishr which isn't the same for -1, but the -1 case still works
847 # since we use exp-exp/2 as the second exponent.) While the spec
848 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
849 # work with denormals and doesn't allow for the full swing in exponents
850 # that you can get with normalized values. Instead, we create two powers
851 # of two and multiply by them each in turn. That way the effective range
852 # of our exponent is doubled.
853 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
854 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
855 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
858 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
859 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
862 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
863 def bitfield_reverse(u
):
864 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
865 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
866 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
867 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
868 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
872 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
874 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
875 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
876 # and, if a is a NaN then the second comparison will fail anyway.
877 for op
in ['flt', 'fge', 'feq']:
879 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
880 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
883 # Add optimizations to handle the case where the result of a ternary is
884 # compared to a constant. This way we can take things like
890 # a ? (0 > 0) : (1 > 0)
892 # which constant folding will eat for lunch. The resulting ternary will
893 # further get cleaned up by the boolean reductions above and we will be
894 # left with just the original variable "a".
895 for op
in ['flt', 'fge', 'feq', 'fne',
896 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
898 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
899 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
900 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
901 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
905 # For example, this converts things like
907 # 1 + mix(0, a - 1, condition)
911 # mix(1, (a-1)+1, condition)
913 # Other optimizations will rearrange the constants.
914 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
916 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
919 # This section contains "late" optimizations that should be run before
920 # creating ffmas and calling regular optimizations for the final time.
921 # Optimizations should go here if they help code generation and conflict
922 # with the regular optimizations.
923 before_ffma_optimizations
= [
924 # Propagate constants down multiplication chains
925 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
926 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
927 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
928 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
930 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
931 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
932 (('~fadd', ('fneg', a
), a
), 0.0),
933 (('iadd', ('ineg', a
), a
), 0),
934 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
935 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
936 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
937 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
940 # This section contains "late" optimizations that should be run after the
941 # regular optimizations have finished. Optimizations should go here if
942 # they help code generation but do not necessarily produce code that is
943 # more easily optimizable.
944 late_optimizations
= [
945 # Most of these optimizations aren't quite safe when you get infinity or
946 # Nan involved but the first one should be fine.
947 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
948 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
949 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
950 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
951 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
952 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
954 (('~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
)))),
956 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
957 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
958 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
959 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
961 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a
, 0.0, 1.0)),
962 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a
, -0.0, -1.0)),
964 # we do these late so that we don't get in the way of creating ffmas
965 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
966 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
968 (('bcsel', 'a@bool', 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
971 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
972 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
973 before_ffma_optimizations
).render())
974 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
975 late_optimizations
).render())