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
))))),
73 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
74 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
75 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
76 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
81 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
82 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
83 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
84 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
86 (('fneg', ('fneg', a
)), a
),
87 (('ineg', ('ineg', a
)), a
),
88 (('fabs', ('fabs', a
)), ('fabs', a
)),
89 (('fabs', ('fneg', a
)), ('fabs', a
)),
90 (('fabs', ('u2f', a
)), ('u2f', a
)),
91 (('iabs', ('iabs', a
)), ('iabs', a
)),
92 (('iabs', ('ineg', a
)), ('iabs', a
)),
93 (('f2b', ('fneg', a
)), ('f2b', a
)),
94 (('i2b', ('ineg', a
)), ('i2b', a
)),
95 (('~fadd', a
, 0.0), a
),
97 (('usadd_4x8', a
, 0), a
),
98 (('usadd_4x8', a
, ~
0), ~
0),
99 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
100 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
101 (('~fadd', ('fneg', a
), a
), 0.0),
102 (('iadd', ('ineg', a
), a
), 0),
103 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
104 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
105 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
106 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
107 (('~fmul', a
, 0.0), 0.0),
109 (('umul_unorm_4x8', a
, 0), 0),
110 (('umul_unorm_4x8', a
, ~
0), a
),
111 (('fmul', a
, 1.0), a
),
113 (('fmul', a
, -1.0), ('fneg', a
)),
114 (('imul', a
, -1), ('ineg', a
)),
115 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
116 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
117 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
118 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
119 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
120 (('~ffma', 0.0, a
, b
), b
),
121 (('~ffma', a
, 0.0, b
), b
),
122 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
123 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
124 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
125 (('~flrp', a
, b
, 0.0), a
),
126 (('~flrp', a
, b
, 1.0), b
),
127 (('~flrp', a
, a
, b
), a
),
128 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
129 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
130 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
131 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
132 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
133 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
134 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
135 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
136 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
137 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
138 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
139 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
140 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
141 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
142 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
143 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
144 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
146 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
147 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
148 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
149 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
151 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
152 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
154 # (a * #b + #c) << #d
155 # ((a * #b) << #d) + (#c << #d)
156 # (a * (#b << #d)) + (#c << #d)
157 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
158 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
162 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
164 # Comparison simplifications
165 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
166 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
167 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
168 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
169 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
170 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
171 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
172 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
173 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
174 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
178 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
180 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
182 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
184 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
185 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
186 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
187 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
188 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
189 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
190 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
191 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
192 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
193 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
194 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
195 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
196 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
197 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
198 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
199 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
200 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
201 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
202 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
203 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
205 # -(b2f(a) + b2f(b)) < 0
206 # 0 < b2f(a) + b2f(b)
207 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
209 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
210 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
212 # -(b2f(a) + b2f(b)) >= 0
213 # 0 >= b2f(a) + b2f(b)
214 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
216 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
217 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
219 # Some optimizations (below) convert things like (a < b || c < b) into
220 # (min(a, c) < b). However, this interfers with the previous optimizations
221 # that try to remove comparisons with negated sums of b2f. This just
223 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
224 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
226 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
227 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
228 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
229 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
231 # Cannot remove the addition from ilt or ige due to overflow.
232 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
233 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
235 # fmin(-b2f(a), b) >= 0.0
236 # -b2f(a) >= 0.0 && b >= 0.0
237 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
238 # b2f(a) == 0.0 && b >= 0.0
239 # a == False && b >= 0.0
242 # The fge in the second replacement is not a typo. I leave the proof that
243 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
245 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
246 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
248 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
249 (('fne', ('b2f', 'a@1'), 0.0), a
),
250 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
251 (('ine', ('b2i', 'a@1'), 0), a
),
253 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
254 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
255 (('fge', ('u2f', a
), 0.0), True),
256 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
257 (('flt', ('u2f', a
), 0.0), False),
258 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
259 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
260 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
261 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
262 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
263 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
264 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
268 # fabs(a) != 0.0 because fabs(a) must be >= 0
270 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
274 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
277 # 0.0 == fabs(a) because fabs(a) must be >= 0
279 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
283 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
285 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
286 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
287 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
288 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
291 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
292 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
293 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
295 # Since b is a constant, constant folding will eliminate the fmin and the
296 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
297 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
299 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
301 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
302 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
303 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
304 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
305 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
306 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
307 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
308 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
309 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
310 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
311 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
312 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
313 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
314 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
321 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
322 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
323 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
324 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
325 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
326 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
327 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
328 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
329 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
330 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
331 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
332 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
333 (('fmin', a
, ('fabs', a
)), a
),
334 (('imin', a
, ('iabs', a
)), a
),
335 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
336 (('imax', a
, ('ineg', ('iabs', a
))), a
),
337 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
338 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
339 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
340 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
341 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
342 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
343 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
344 (('fsat', ('b2f', a
)), ('b2f', a
)),
345 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
346 (('fsat', ('fsat', a
)), ('fsat', a
)),
347 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
348 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
349 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
350 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
351 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
352 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
353 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
354 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
355 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
356 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
357 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
358 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
359 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
360 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
361 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
362 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
363 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
364 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
365 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
366 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
367 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
368 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
370 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
371 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
372 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
373 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
374 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
375 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
376 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
377 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
378 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
379 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
380 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
381 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
382 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
383 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
384 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
385 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
387 # Common pattern like 'if (i == 0 || i == 1 || ...)'
388 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
389 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
390 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
392 (('ior', 'a@bool', ('ieq', a
, False)), True),
393 (('ior', a
, ('inot', a
)), -1),
395 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
397 # These patterns can result when (a < b || a < c) => (a < min(b, c))
398 # transformations occur before constant propagation and loop-unrolling.
399 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
400 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
401 (('~fge', a
, ('fmin', b
, a
)), True),
402 (('~fge', ('fmax', a
, b
), a
), True),
403 (('~flt', a
, ('fmin', b
, a
)), False),
404 (('~flt', ('fmax', a
, b
), a
), False),
405 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
406 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
408 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
409 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
410 (('ige', a
, ('imin', b
, a
)), True),
411 (('ige', ('imax', a
, b
), a
), True),
412 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
413 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
414 (('uge', a
, ('umin', b
, a
)), True),
415 (('uge', ('umax', a
, b
), a
), True),
416 (('ilt', a
, ('imin', b
, a
)), False),
417 (('ilt', ('imax', a
, b
), a
), False),
418 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
419 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
420 (('ult', a
, ('umin', b
, a
)), False),
421 (('ult', ('umax', a
, b
), a
), False),
422 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
423 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
425 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
426 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
427 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
428 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
429 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
430 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
431 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
432 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
433 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
434 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
435 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
436 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
437 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
438 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
439 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
440 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
442 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
443 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
444 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
445 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
446 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
447 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
448 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
449 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
450 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
451 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
453 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
454 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
455 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
456 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
457 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
458 (('ineg', ('b2i32', 'a@32')), a
),
459 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
460 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
461 # Comparison with the same args. Note that these are not done for
462 # the float versions because NaN always returns false on float
464 (('ilt', a
, a
), False),
465 (('ige', a
, a
), True),
466 (('ieq', a
, a
), True),
467 (('ine', a
, a
), False),
468 (('ult', a
, a
), False),
469 (('uge', a
, a
), True),
470 # Logical and bit operations
471 (('fand', a
, 0.0), 0.0),
473 (('iand', a
, ~
0), a
),
477 (('ior', a
, True), True),
478 (('fxor', a
, a
), 0.0),
481 (('inot', ('inot', a
)), a
),
482 (('ior', ('iand', a
, b
), b
), b
),
483 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
484 (('iand', ('ior', a
, b
), b
), b
),
485 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
487 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
488 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
489 # Shift optimizations
496 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
497 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
498 # Exponential/logarithmic identities
499 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
500 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
501 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
502 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
503 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
504 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
505 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
506 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
507 (('~fpow', a
, 1.0), a
),
508 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
509 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
510 (('~fpow', 2.0, a
), ('fexp2', a
)),
511 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
512 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
513 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
514 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
515 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
516 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
517 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
518 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
519 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
520 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
521 # Division and reciprocal
522 (('~fdiv', 1.0, a
), ('frcp', a
)),
523 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
524 (('~frcp', ('frcp', a
)), a
),
525 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
526 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
527 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
528 # Boolean simplifications
529 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
530 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
531 (('ieq', 'a@bool', True), a
),
532 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
533 (('ine', 'a@bool', False), a
),
534 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
535 (('bcsel', a
, True, False), a
),
536 (('bcsel', a
, False, True), ('inot', a
)),
537 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
538 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
539 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
540 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
541 (('bcsel', True, b
, c
), b
),
542 (('bcsel', False, b
, c
), c
),
543 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
544 # The result of this should be hit by constant propagation and, in the
545 # next round of opt_algebraic, get picked up by one of the above two.
546 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
548 (('bcsel', a
, b
, b
), b
),
549 (('fcsel', a
, b
, b
), b
),
551 # D3D Boolean emulation
552 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
553 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
554 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
555 ('ineg', ('b2i', ('iand', a
, b
)))),
556 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
557 ('ineg', ('b2i', ('ior', a
, b
)))),
558 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
559 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
560 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
561 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
562 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
564 # SM5 32-bit shifts are defined to use the 5 least significant bits
565 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
566 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
567 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
570 (('i2b32', ('b2i', 'a@32')), a
),
571 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
572 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
573 (('i2b', ('ineg', a
)), ('i2b', a
)),
574 (('i2b', ('iabs', a
)), ('i2b', a
)),
575 (('fabs', ('b2f', a
)), ('b2f', a
)),
576 (('iabs', ('b2i', a
)), ('b2i', a
)),
577 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
579 # Ironically, mark these as imprecise because removing the conversions may
580 # preserve more precision than doing the conversions (e.g.,
581 # uint(float(0x81818181u)) == 0x81818200).
582 (('~f2i32', ('i2f', 'a@32')), a
),
583 (('~f2i32', ('u2f', 'a@32')), a
),
584 (('~f2u32', ('i2f', 'a@32')), a
),
585 (('~f2u32', ('u2f', 'a@32')), a
),
587 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
590 # It is undefined to convert a negative floating-point value to an
593 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
594 # some optimizations in the i965 backend to proceed.
595 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
596 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
597 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
598 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
600 # Packing and then unpacking does nothing
601 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
602 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
603 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
604 ('unpack_64_2x32_split_y', a
)), a
),
607 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
608 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
609 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
610 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
611 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
614 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
615 # patterns like those below.
616 for op
in ('ushr', 'ishr'):
617 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
618 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
619 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
621 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
623 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
624 # patterns like those below.
625 for op
in ('extract_u8', 'extract_i8'):
626 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), ('extract_u8', a
, i
)) for i
in range(2, -1, -1)])
627 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), ('extract_u8', a
, i
)) for i
in range(6, -1, -1)])
629 optimizations
.extend([
631 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
632 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
633 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
634 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
635 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
638 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
639 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
640 (('ussub_4x8', a
, 0), a
),
641 (('ussub_4x8', a
, ~
0), 0),
642 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
643 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
644 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
645 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
646 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
647 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
648 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
649 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
651 # Propagate negation up multiplication chains
652 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
653 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
655 # Propagate constants up multiplication chains
656 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
657 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
658 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
659 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
661 # Reassociate constants in add/mul chains so they can be folded together.
662 # For now, we mostly only handle cases where the constants are separated by
663 # a single non-constant. We could do better eventually.
664 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
665 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
666 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
667 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
668 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
671 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
672 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
673 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
675 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
676 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
677 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
679 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
682 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
683 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
684 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
685 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
686 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
688 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
689 ('bcsel', ('ilt', 31, 'bits'), 'insert',
690 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
691 'options->lower_bitfield_insert'),
692 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
693 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
694 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
695 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
696 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
697 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
699 # Alternative lowering that doesn't rely on bfi.
700 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
701 ('bcsel', ('ilt', 31, 'bits'),
704 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
705 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
706 'options->lower_bitfield_insert_to_shifts'),
708 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
709 (('bfm', 'bits', 'offset'),
710 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
711 'options->lower_bfm'),
713 (('ibitfield_extract', 'value', 'offset', 'bits'),
714 ('bcsel', ('ilt', 31, 'bits'), 'value',
715 ('ibfe', 'value', 'offset', 'bits')),
716 'options->lower_bitfield_extract'),
718 (('ubitfield_extract', 'value', 'offset', 'bits'),
719 ('bcsel', ('ult', 31, 'bits'), 'value',
720 ('ubfe', 'value', 'offset', 'bits')),
721 'options->lower_bitfield_extract'),
723 (('ibitfield_extract', 'value', 'offset', 'bits'),
724 ('bcsel', ('ieq', 0, 'bits'),
727 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
728 ('isub', 32, 'bits'))),
729 'options->lower_bitfield_extract_to_shifts'),
731 (('ubitfield_extract', 'value', 'offset', 'bits'),
733 ('ushr', 'value', 'offset'),
734 ('bcsel', ('ieq', 'bits', 32),
736 ('bfm', 'bits', 0))),
737 'options->lower_bitfield_extract_to_shifts'),
739 (('ifind_msb', 'value'),
740 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
741 'options->lower_ifind_msb'),
743 (('find_lsb', 'value'),
744 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
745 'options->lower_find_lsb'),
747 (('extract_i8', a
, 'b@32'),
748 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
749 'options->lower_extract_byte'),
751 (('extract_u8', a
, 'b@32'),
752 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
753 'options->lower_extract_byte'),
755 (('extract_i16', a
, 'b@32'),
756 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
757 'options->lower_extract_word'),
759 (('extract_u16', a
, 'b@32'),
760 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
761 'options->lower_extract_word'),
763 (('pack_unorm_2x16', 'v'),
764 ('pack_uvec2_to_uint',
765 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
766 'options->lower_pack_unorm_2x16'),
768 (('pack_unorm_4x8', 'v'),
769 ('pack_uvec4_to_uint',
770 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
771 'options->lower_pack_unorm_4x8'),
773 (('pack_snorm_2x16', 'v'),
774 ('pack_uvec2_to_uint',
775 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
776 'options->lower_pack_snorm_2x16'),
778 (('pack_snorm_4x8', 'v'),
779 ('pack_uvec4_to_uint',
780 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
781 'options->lower_pack_snorm_4x8'),
783 (('unpack_unorm_2x16', 'v'),
784 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
785 ('extract_u16', 'v', 1))),
787 'options->lower_unpack_unorm_2x16'),
789 (('unpack_unorm_4x8', 'v'),
790 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
791 ('extract_u8', 'v', 1),
792 ('extract_u8', 'v', 2),
793 ('extract_u8', 'v', 3))),
795 'options->lower_unpack_unorm_4x8'),
797 (('unpack_snorm_2x16', 'v'),
798 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
799 ('extract_i16', 'v', 1))),
801 'options->lower_unpack_snorm_2x16'),
803 (('unpack_snorm_4x8', 'v'),
804 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
805 ('extract_i8', 'v', 1),
806 ('extract_i8', 'v', 2),
807 ('extract_i8', 'v', 3))),
809 'options->lower_unpack_snorm_4x8'),
811 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
814 # bit_size dependent lowerings
815 for bit_size
in [8, 16, 32, 64]:
816 # convenience constants
817 intmax
= (1 << (bit_size
- 1)) - 1
818 intmin
= 1 << (bit_size
- 1)
821 (('iadd_sat@' + str(bit_size
), a
, b
),
822 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
823 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
824 (('isub_sat@' + str(bit_size
), a
, b
),
825 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
826 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
829 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
831 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
832 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
833 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
834 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
835 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
837 # Optimize x2bN(b2x(x)) -> x
838 for size
in type_sizes('bool'):
839 aN
= 'a@' + str(size
)
840 f2bN
= 'f2b' + str(size
)
841 i2bN
= 'i2b' + str(size
)
842 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
843 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
845 # Optimize x2yN(b2x(x)) -> b2y
846 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
847 if x
!= 'f' and y
!= 'f' and x
!= y
:
850 b2x
= 'b2f' if x
== 'f' else 'b2i'
851 b2y
= 'b2f' if y
== 'f' else 'b2i'
852 x2yN
= '{}2{}'.format(x
, y
)
853 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
855 def fexp2i(exp
, bits
):
856 # We assume that exp is already in the right range.
858 return ('ishl', ('iadd', exp
, 127), 23)
860 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
864 def ldexp(f
, exp
, bits
):
865 # First, we clamp exp to a reasonable range. The maximum possible range
866 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
867 # a maximum range of [-149, 127]. This means that we can potentially have
868 # a swing of +-276. If you start with FLT_MAX, you actually have to do
869 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
870 # spec, on the other hand, only requires that we handle an exponent value
871 # in the range [-126, 128]. This implementation is *mostly* correct; it
872 # handles a range on exp of [-252, 254] which allows you to create any
873 # value (including denorms if the hardware supports it) and to adjust the
874 # exponent of any normal value to anything you want.
876 exp
= ('imin', ('imax', exp
, -252), 254)
878 exp
= ('imin', ('imax', exp
, -2044), 2046)
882 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
883 # (We use ishr which isn't the same for -1, but the -1 case still works
884 # since we use exp-exp/2 as the second exponent.) While the spec
885 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
886 # work with denormals and doesn't allow for the full swing in exponents
887 # that you can get with normalized values. Instead, we create two powers
888 # of two and multiply by them each in turn. That way the effective range
889 # of our exponent is doubled.
890 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
891 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
892 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
895 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
896 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
899 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
900 def bitfield_reverse(u
):
901 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
902 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
903 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
904 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
905 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
909 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
911 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
912 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
913 # and, if a is a NaN then the second comparison will fail anyway.
914 for op
in ['flt', 'fge', 'feq']:
916 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
917 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
920 # Add optimizations to handle the case where the result of a ternary is
921 # compared to a constant. This way we can take things like
927 # a ? (0 > 0) : (1 > 0)
929 # which constant folding will eat for lunch. The resulting ternary will
930 # further get cleaned up by the boolean reductions above and we will be
931 # left with just the original variable "a".
932 for op
in ['flt', 'fge', 'feq', 'fne',
933 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
935 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
936 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
937 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
938 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
942 # For example, this converts things like
944 # 1 + mix(0, a - 1, condition)
948 # mix(1, (a-1)+1, condition)
950 # Other optimizations will rearrange the constants.
951 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
953 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
956 # This section contains "late" optimizations that should be run before
957 # creating ffmas and calling regular optimizations for the final time.
958 # Optimizations should go here if they help code generation and conflict
959 # with the regular optimizations.
960 before_ffma_optimizations
= [
961 # Propagate constants down multiplication chains
962 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
963 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
964 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
965 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
967 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
968 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
969 (('~fadd', ('fneg', a
), a
), 0.0),
970 (('iadd', ('ineg', a
), a
), 0),
971 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
972 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
973 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
974 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
977 # This section contains "late" optimizations that should be run after the
978 # regular optimizations have finished. Optimizations should go here if
979 # they help code generation but do not necessarily produce code that is
980 # more easily optimizable.
981 late_optimizations
= [
982 # Most of these optimizations aren't quite safe when you get infinity or
983 # Nan involved but the first one should be fine.
984 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
985 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
986 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
987 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
988 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
989 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
991 (('~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
)))),
993 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
994 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
995 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
996 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
998 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a
, 0.0, 1.0)),
999 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a
, -0.0, -1.0)),
1001 # we do these late so that we don't get in the way of creating ffmas
1002 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1003 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1005 (('bcsel', 'a@bool', 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1008 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1009 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1010 before_ffma_optimizations
).render())
1011 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1012 late_optimizations
).render())
projects / mesa.git / blob