2 # Copyright (C) 2014 Intel Corporation
4 # Permission is hereby granted, free of charge, to any person obtaining a
5 # copy of this software and associated documentation files (the "Software"),
6 # to deal in the Software without restriction, including without limitation
7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 # and/or sell copies of the Software, and to permit persons to whom the
9 # Software is furnished to do so, subject to the following conditions:
11 # The above copyright notice and this permission notice (including the next
12 # paragraph) shall be included in all copies or substantial portions of the
15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 # Jason Ekstrand (jason@jlekstrand.net)
26 from __future__
import print_function
28 from collections
import OrderedDict
32 # Convenience variables
38 # Written in the form (<search>, <replace>) where <search> is an expression
39 # and <replace> is either an expression or a value. An expression is
40 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
41 # where each source is either an expression or a value. A value can be
42 # either a numeric constant or a string representing a variable name.
44 # If the opcode in a search expression is prefixed by a '~' character, this
45 # indicates that the operation is inexact. Such operations will only get
46 # applied to SSA values that do not have the exact bit set. This should be
47 # used by by any optimizations that are not bit-for-bit exact. It should not,
48 # however, be used for backend-requested lowering operations as those need to
49 # happen regardless of precision.
51 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
52 # that the given variable will only match constants and the type indicates that
53 # the given variable will only match values from ALU instructions with the
54 # given output type, and (cond) specifies an additional condition function
55 # (see nir_search_helpers.h).
57 # For constants, you have to be careful to make sure that it is the right
58 # type because python is unaware of the source and destination types of the
61 # All expression types can have a bit-size specified. For opcodes, this
62 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
63 # type and size, and for literals, you can write "2.0@32". In the search half
64 # of the expression this indicates that it should only match that particular
65 # bit-size. In the replace half of the expression this indicates that the
66 # constructed value should have that bit-size.
70 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
71 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
76 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
77 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
78 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
79 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
81 (('fneg', ('fneg', a
)), a
),
82 (('ineg', ('ineg', a
)), a
),
83 (('fabs', ('fabs', a
)), ('fabs', a
)),
84 (('fabs', ('fneg', a
)), ('fabs', a
)),
85 (('fabs', ('u2f32', a
)), ('u2f32', a
)),
86 (('iabs', ('iabs', a
)), ('iabs', a
)),
87 (('iabs', ('ineg', a
)), ('iabs', a
)),
88 (('~fadd', a
, 0.0), a
),
90 (('usadd_4x8', a
, 0), a
),
91 (('usadd_4x8', a
, ~
0), ~
0),
92 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
93 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
94 (('~fadd', ('fneg', a
), a
), 0.0),
95 (('iadd', ('ineg', a
), a
), 0),
96 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
97 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
98 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
99 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
100 (('~fmul', a
, 0.0), 0.0),
102 (('umul_unorm_4x8', a
, 0), 0),
103 (('umul_unorm_4x8', a
, ~
0), a
),
104 (('fmul', a
, 1.0), a
),
106 (('fmul', a
, -1.0), ('fneg', a
)),
107 (('imul', a
, -1), ('ineg', a
)),
108 (('~ffma', 0.0, a
, b
), b
),
109 (('~ffma', a
, 0.0, b
), b
),
110 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
111 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
112 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
113 (('~flrp', a
, b
, 0.0), a
),
114 (('~flrp', a
, b
, 1.0), b
),
115 (('~flrp', a
, a
, b
), a
),
116 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
117 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
118 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
119 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
120 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
121 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
122 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
123 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
124 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
125 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
126 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
127 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
128 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
129 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
131 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
132 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
133 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
134 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
136 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
137 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
139 # (a * #b + #c) << #d
140 # ((a * #b) << #d) + (#c << #d)
141 # (a * (#b << #d)) + (#c << #d)
142 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
143 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
147 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
149 # Comparison simplifications
150 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
151 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
152 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
153 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
154 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
155 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
156 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
157 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
158 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
159 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
163 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
165 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
167 (('fge', ('fneg', ('b2f', a
)), 0.0), ('inot', a
)),
169 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
170 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
171 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
172 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
174 # Cannot remove the addition from ilt or ige due to overflow.
175 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
176 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
178 # fmin(-b2f(a), b) >= 0.0
179 # -b2f(a) >= 0.0 && b >= 0.0
180 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
181 # b2f(a) == 0.0 && b >= 0.0
182 # a == False && b >= 0.0
185 # The fge in the second replacement is not a typo. I leave the proof that
186 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
188 (('fge', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
189 (('feq', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
191 (('feq', ('b2f', a
), 0.0), ('inot', a
)),
192 (('fne', ('b2f', a
), 0.0), a
),
193 (('ieq', ('b2i', a
), 0), ('inot', a
)),
194 (('ine', ('b2i', a
), 0), a
),
196 (('fne', ('u2f32', a
), 0.0), ('ine', a
, 0)),
197 (('feq', ('u2f32', a
), 0.0), ('ieq', a
, 0)),
198 (('fge', ('u2f32', a
), 0.0), True),
199 (('fge', 0.0, ('u2f32', a
)), ('uge', 0, a
)), # ieq instead?
200 (('flt', ('u2f32', a
), 0.0), False),
201 (('flt', 0.0, ('u2f32', a
)), ('ult', 0, a
)), # ine instead?
202 (('fne', ('i2f32', a
), 0.0), ('ine', a
, 0)),
203 (('feq', ('i2f32', a
), 0.0), ('ieq', a
, 0)),
204 (('fge', ('i2f32', a
), 0.0), ('ige', a
, 0)),
205 (('fge', 0.0, ('i2f32', a
)), ('ige', 0, a
)),
206 (('flt', ('i2f32', a
), 0.0), ('ilt', a
, 0)),
207 (('flt', 0.0, ('i2f32', a
)), ('ilt', 0, a
)),
211 # fabs(a) != 0.0 because fabs(a) must be >= 0
213 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
215 (('fmax', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('ior', a
, b
))),
216 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('ior', a
, b
)))),
217 (('fmin', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
218 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('iand', a
, b
)))),
221 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
222 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
223 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
225 # Since b is a constant, constant folding will eliminate the fmin and the
226 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
227 (('fmin', ('b2f', a
), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
229 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
231 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
232 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
233 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
234 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
235 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
236 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
237 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
238 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
239 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
246 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
247 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
248 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
249 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
250 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
251 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
252 (('fmin', a
, ('fabs', a
)), a
),
253 (('imin', a
, ('iabs', a
)), a
),
254 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
255 (('imax', a
, ('ineg', ('iabs', a
))), a
),
256 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
257 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
258 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
259 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
260 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
261 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
262 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
263 (('fsat', ('fsat', a
)), ('fsat', a
)),
264 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
265 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
266 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
267 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
268 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
269 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
270 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
271 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
272 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
273 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
274 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
275 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
276 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
277 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
278 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
279 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
280 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
281 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
282 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
283 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
284 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
285 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
287 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
288 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
289 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
290 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
291 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
292 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
293 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
294 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
295 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
296 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
297 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
298 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
299 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
300 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
301 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
302 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
304 (('ior', 'a@bool', ('ieq', a
, False)), True),
305 (('ior', 'a@bool', ('inot', a
)), True),
307 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
309 # These patterns can result when (a < b || a < c) => (a < min(b, c))
310 # transformations occur before constant propagation and loop-unrolling.
311 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
312 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
313 (('~fge', a
, ('fmin', b
, a
)), True),
314 (('~fge', ('fmax', a
, b
), a
), True),
315 (('~flt', a
, ('fmin', b
, a
)), False),
316 (('~flt', ('fmax', a
, b
), a
), False),
317 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
318 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
320 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
321 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
322 (('ige', a
, ('imin', b
, a
)), True),
323 (('ige', ('imax', a
, b
), a
), True),
324 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
325 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
326 (('uge', a
, ('umin', b
, a
)), True),
327 (('uge', ('umax', a
, b
), a
), True),
328 (('ilt', a
, ('imin', b
, a
)), False),
329 (('ilt', ('imax', a
, b
), a
), False),
330 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
331 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
332 (('ult', a
, ('umin', b
, a
)), False),
333 (('ult', ('umax', a
, b
), a
), False),
334 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
335 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
337 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
338 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
339 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
340 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
341 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
342 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
343 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
344 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
345 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
346 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
347 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
348 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
349 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
350 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
351 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
352 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
354 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
355 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
356 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
357 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
358 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
359 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
360 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
361 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
362 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
363 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
365 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
366 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
367 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
368 (('iand', 'a@bool', 1.0), ('b2f', a
), '!options->lower_b2f'),
369 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
370 (('ineg', ('b2i@32', a
)), a
),
371 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
372 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
373 # Comparison with the same args. Note that these are not done for
374 # the float versions because NaN always returns false on float
376 (('ilt', a
, a
), False),
377 (('ige', a
, a
), True),
378 (('ieq', a
, a
), True),
379 (('ine', a
, a
), False),
380 (('ult', a
, a
), False),
381 (('uge', a
, a
), True),
382 # Logical and bit operations
383 (('fand', a
, 0.0), 0.0),
385 (('iand', a
, ~
0), a
),
389 (('ior', a
, True), True),
390 (('fxor', a
, a
), 0.0),
393 (('inot', ('inot', a
)), a
),
394 (('ior', ('iand', a
, b
), b
), b
),
395 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
396 (('iand', ('ior', a
, b
), b
), b
),
397 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
399 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
400 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
401 # Shift optimizations
408 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
409 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
410 # Exponential/logarithmic identities
411 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
412 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
413 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
414 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
415 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
416 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
417 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
418 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
419 (('~fpow', a
, 1.0), a
),
420 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
421 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
422 (('~fpow', 2.0, a
), ('fexp2', a
)),
423 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
424 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
425 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
426 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
427 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
428 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
429 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
430 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
431 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
432 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
433 # Division and reciprocal
434 (('~fdiv', 1.0, a
), ('frcp', a
)),
435 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
436 (('~frcp', ('frcp', a
)), a
),
437 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
438 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
439 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
440 # Boolean simplifications
441 (('ieq', 'a@bool', True), a
),
442 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
443 (('ine', 'a@bool', False), a
),
444 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
445 (('bcsel', a
, True, False), a
),
446 (('bcsel', a
, False, True), ('inot', a
)),
447 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
448 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
449 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
450 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
451 (('bcsel', True, b
, c
), b
),
452 (('bcsel', False, b
, c
), c
),
453 (('bcsel', a
, ('b2f(is_used_once)', b
), ('b2f', c
)), ('b2f', ('bcsel', a
, b
, c
))),
454 # The result of this should be hit by constant propagation and, in the
455 # next round of opt_algebraic, get picked up by one of the above two.
456 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
458 (('bcsel', a
, b
, b
), b
),
459 (('fcsel', a
, b
, b
), b
),
462 (('i2b', ('b2i', a
)), a
),
463 (('i2b', 'a@bool'), a
),
464 (('f2i32', ('ftrunc', a
)), ('f2i32', a
)),
465 (('f2u32', ('ftrunc', a
)), ('f2u32', a
)),
466 (('i2b', ('ineg', a
)), ('i2b', a
)),
467 (('i2b', ('iabs', a
)), ('i2b', a
)),
468 (('fabs', ('b2f', a
)), ('b2f', a
)),
469 (('iabs', ('b2i', a
)), ('b2i', a
)),
470 (('inot', ('f2b', a
)), ('feq', a
, 0.0)),
472 # Ironically, mark these as imprecise because removing the conversions may
473 # preserve more precision than doing the conversions (e.g.,
474 # uint(float(0x81818181u)) == 0x81818200).
475 (('~f2i32', ('i2f32', 'a@32')), a
),
476 (('~f2i32', ('u2f32', 'a@32')), a
),
477 (('~f2u32', ('i2f32', 'a@32')), a
),
478 (('~f2u32', ('u2f32', 'a@32')), a
),
480 # Packing and then unpacking does nothing
481 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
482 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
483 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
484 ('unpack_64_2x32_split_y', a
)), a
),
487 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
488 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
489 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
490 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
493 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
494 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
497 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
498 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
499 (('ussub_4x8', a
, 0), a
),
500 (('ussub_4x8', a
, ~
0), 0),
501 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
502 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
503 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
504 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
505 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
506 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
507 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
508 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
510 # Propagate negation up multiplication chains
511 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
512 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
514 # Propagate constants up multiplication chains
515 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
516 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
517 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
518 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
520 # Reassociate constants in add/mul chains so they can be folded together.
521 # For now, we mostly only handle cases where the constants are separated by
522 # a single non-constant. We could do better eventually.
523 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
524 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
525 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
526 (('~fadd', '#a', ('fneg', ('fadd', b
, '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
527 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
530 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
531 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
532 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
534 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
535 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
536 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
538 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
541 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
542 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
543 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
544 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
545 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
547 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
548 ('bcsel', ('ilt', 31, 'bits'), 'insert',
549 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
550 'options->lower_bitfield_insert'),
552 # Alternative lowering that doesn't rely on bfi.
553 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
554 ('bcsel', ('ilt', 31, 'bits'),
557 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
558 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
559 'options->lower_bitfield_insert_to_shifts'),
561 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
562 (('bfm', 'bits', 'offset'),
563 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
564 'options->lower_bfm'),
566 (('ibitfield_extract', 'value', 'offset', 'bits'),
567 ('bcsel', ('ilt', 31, 'bits'), 'value',
568 ('ibfe', 'value', 'offset', 'bits')),
569 'options->lower_bitfield_extract'),
571 (('ubitfield_extract', 'value', 'offset', 'bits'),
572 ('bcsel', ('ult', 31, 'bits'), 'value',
573 ('ubfe', 'value', 'offset', 'bits')),
574 'options->lower_bitfield_extract'),
576 (('ibitfield_extract', 'value', 'offset', 'bits'),
577 ('bcsel', ('ieq', 0, 'bits'),
580 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
581 ('isub', 32, 'bits'))),
582 'options->lower_bitfield_extract_to_shifts'),
584 (('ubitfield_extract', 'value', 'offset', 'bits'),
586 ('ushr', 'value', 'offset'),
587 ('bcsel', ('ieq', 'bits', 32),
589 ('bfm', 'bits', 0))),
590 'options->lower_bitfield_extract_to_shifts'),
592 (('ifind_msb', 'value'),
593 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
594 'options->lower_ifind_msb'),
596 (('find_lsb', 'value'),
597 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
598 'options->lower_find_lsb'),
600 (('extract_i8', a
, 'b@32'),
601 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
602 'options->lower_extract_byte'),
604 (('extract_u8', a
, 'b@32'),
605 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
606 'options->lower_extract_byte'),
608 (('extract_i16', a
, 'b@32'),
609 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
610 'options->lower_extract_word'),
612 (('extract_u16', a
, 'b@32'),
613 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
614 'options->lower_extract_word'),
616 (('pack_unorm_2x16', 'v'),
617 ('pack_uvec2_to_uint',
618 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
619 'options->lower_pack_unorm_2x16'),
621 (('pack_unorm_4x8', 'v'),
622 ('pack_uvec4_to_uint',
623 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
624 'options->lower_pack_unorm_4x8'),
626 (('pack_snorm_2x16', 'v'),
627 ('pack_uvec2_to_uint',
628 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
629 'options->lower_pack_snorm_2x16'),
631 (('pack_snorm_4x8', 'v'),
632 ('pack_uvec4_to_uint',
633 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
634 'options->lower_pack_snorm_4x8'),
636 (('unpack_unorm_2x16', 'v'),
637 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
638 ('extract_u16', 'v', 1))),
640 'options->lower_unpack_unorm_2x16'),
642 (('unpack_unorm_4x8', 'v'),
643 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
644 ('extract_u8', 'v', 1),
645 ('extract_u8', 'v', 2),
646 ('extract_u8', 'v', 3))),
648 'options->lower_unpack_unorm_4x8'),
650 (('unpack_snorm_2x16', 'v'),
651 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec2', ('extract_i16', 'v', 0),
652 ('extract_i16', 'v', 1))),
654 'options->lower_unpack_snorm_2x16'),
656 (('unpack_snorm_4x8', 'v'),
657 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec4', ('extract_i8', 'v', 0),
658 ('extract_i8', 'v', 1),
659 ('extract_i8', 'v', 2),
660 ('extract_i8', 'v', 3))),
662 'options->lower_unpack_snorm_4x8'),
665 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
667 for left
, right
in list(itertools
.combinations(invert
.keys(), 2)) + zip(invert
.keys(), invert
.keys()):
668 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
669 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
670 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
671 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
673 def fexp2i(exp
, bits
):
674 # We assume that exp is already in the right range.
676 return ('ishl', ('iadd', exp
, 127), 23)
678 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
682 def ldexp(f
, exp
, bits
):
683 # First, we clamp exp to a reasonable range. The maximum possible range
684 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
685 # a maximum range of [-149, 127]. This means that we can potentially have
686 # a swing of +-276. If you start with FLT_MAX, you actually have to do
687 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
688 # spec, on the other hand, only requires that we handle an exponent value
689 # in the range [-126, 128]. This implementation is *mostly* correct; it
690 # handles a range on exp of [-252, 254] which allows you to create any
691 # value (including denorms if the hardware supports it) and to adjust the
692 # exponent of any normal value to anything you want.
694 exp
= ('imin', ('imax', exp
, -252), 254)
696 exp
= ('imin', ('imax', exp
, -2044), 2046)
700 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
701 # (We use ishr which isn't the same for -1, but the -1 case still works
702 # since we use exp-exp/2 as the second exponent.) While the spec
703 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
704 # work with denormals and doesn't allow for the full swing in exponents
705 # that you can get with normalized values. Instead, we create two powers
706 # of two and multiply by them each in turn. That way the effective range
707 # of our exponent is doubled.
708 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
709 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
710 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
713 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
714 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
717 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
718 def bitfield_reverse(u
):
719 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
720 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
721 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
722 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
723 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
727 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
729 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
730 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
731 # and, if a is a NaN then the second comparison will fail anyway.
732 for op
in ['flt', 'fge', 'feq']:
734 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
735 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
738 # Add optimizations to handle the case where the result of a ternary is
739 # compared to a constant. This way we can take things like
745 # a ? (0 > 0) : (1 > 0)
747 # which constant folding will eat for lunch. The resulting ternary will
748 # further get cleaned up by the boolean reductions above and we will be
749 # left with just the original variable "a".
750 for op
in ['flt', 'fge', 'feq', 'fne',
751 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
753 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
754 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
755 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
756 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
760 # For example, this converts things like
762 # 1 + mix(0, a - 1, condition)
766 # mix(1, (a-1)+1, condition)
768 # Other optimizations will rearrange the constants.
769 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
771 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
774 # This section contains "late" optimizations that should be run before
775 # creating ffmas and calling regular optimizations for the final time.
776 # Optimizations should go here if they help code generation and conflict
777 # with the regular optimizations.
778 before_ffma_optimizations
= [
779 # Propagate constants down multiplication chains
780 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
781 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
782 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
783 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
785 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
786 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
787 (('~fadd', ('fneg', a
), a
), 0.0),
788 (('iadd', ('ineg', a
), a
), 0),
789 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
790 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
791 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
792 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
795 # This section contains "late" optimizations that should be run after the
796 # regular optimizations have finished. Optimizations should go here if
797 # they help code generation but do not necessarily produce code that is
798 # more easily optimizable.
799 late_optimizations
= [
800 # Most of these optimizations aren't quite safe when you get infinity or
801 # Nan involved but the first one should be fine.
802 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
803 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
804 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
805 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
806 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
807 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
809 (('~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
)))),
811 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
812 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
813 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
814 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
816 (('b2f(is_used_more_than_once)', ('inot', a
)), ('bcsel', a
, 0.0, 1.0)),
817 (('fneg(is_used_more_than_once)', ('b2f', ('inot', a
))), ('bcsel', a
, -0.0, -1.0)),
819 # we do these late so that we don't get in the way of creating ffmas
820 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
821 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
823 # Lowered for backends without a dedicated b2f instruction
824 (('b2f@32', a
), ('iand', a
, 1.0), 'options->lower_b2f'),
827 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
828 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
829 before_ffma_optimizations
).render())
830 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
831 late_optimizations
).render())