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@16', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp16'),
132 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
133 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
134 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
135 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
136 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
137 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
138 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
139 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
140 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
141 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
142 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
143 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
144 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
145 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
147 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
148 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
149 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
150 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
152 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
153 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
155 # (a * #b + #c) << #d
156 # ((a * #b) << #d) + (#c << #d)
157 # (a * (#b << #d)) + (#c << #d)
158 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
159 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
163 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
165 # Comparison simplifications
166 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
167 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
168 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
169 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
170 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
171 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
172 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
173 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
174 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
175 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
179 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
181 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
183 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
185 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
186 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
187 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
188 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
189 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
190 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
191 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
192 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
193 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
194 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
195 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
196 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
197 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
198 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
199 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
200 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
201 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
202 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
203 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
204 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
206 # -(b2f(a) + b2f(b)) < 0
207 # 0 < b2f(a) + b2f(b)
208 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
210 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
211 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
213 # -(b2f(a) + b2f(b)) >= 0
214 # 0 >= b2f(a) + b2f(b)
215 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
217 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
218 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
220 # Some optimizations (below) convert things like (a < b || c < b) into
221 # (min(a, c) < b). However, this interfers with the previous optimizations
222 # that try to remove comparisons with negated sums of b2f. This just
224 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
225 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
227 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
228 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
229 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
230 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
232 # Cannot remove the addition from ilt or ige due to overflow.
233 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
234 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
236 # fmin(-b2f(a), b) >= 0.0
237 # -b2f(a) >= 0.0 && b >= 0.0
238 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
239 # b2f(a) == 0.0 && b >= 0.0
240 # a == False && b >= 0.0
243 # The fge in the second replacement is not a typo. I leave the proof that
244 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
246 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
247 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
249 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
250 (('fne', ('b2f', 'a@1'), 0.0), a
),
251 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
252 (('ine', ('b2i', 'a@1'), 0), a
),
254 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
255 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
256 (('fge', ('u2f', a
), 0.0), True),
257 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
258 (('flt', ('u2f', a
), 0.0), False),
259 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
260 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
261 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
262 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
263 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
264 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
265 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
269 # fabs(a) != 0.0 because fabs(a) must be >= 0
271 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
275 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
278 # 0.0 == fabs(a) because fabs(a) must be >= 0
280 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
284 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
286 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
287 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
288 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
289 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
292 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
293 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
294 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
296 # Since b is a constant, constant folding will eliminate the fmin and the
297 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
298 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
300 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
302 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
303 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
304 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
305 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
306 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
307 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
308 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
309 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
310 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
311 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
312 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
313 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
314 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
315 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
322 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
323 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
324 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
325 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
326 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
327 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
328 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
329 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
330 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
331 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
332 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
333 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
334 (('fmin', a
, ('fabs', a
)), a
),
335 (('imin', a
, ('iabs', a
)), a
),
336 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
337 (('imax', a
, ('ineg', ('iabs', a
))), a
),
338 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
339 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
340 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
341 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
342 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
343 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
344 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
345 (('fsat', ('b2f', a
)), ('b2f', a
)),
346 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
347 (('fsat', ('fsat', a
)), ('fsat', a
)),
348 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
349 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
350 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
351 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
352 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
353 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
354 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
355 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
356 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
357 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
358 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
359 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
360 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
361 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
362 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
363 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
364 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
365 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
366 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
367 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
368 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
369 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
371 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
372 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
373 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
374 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
375 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
376 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
377 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
378 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
379 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
380 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
381 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
382 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
383 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
384 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
385 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
386 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
388 # Common pattern like 'if (i == 0 || i == 1 || ...)'
389 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
390 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
391 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
393 (('ior', 'a@bool', ('ieq', a
, False)), True),
394 (('ior', a
, ('inot', a
)), -1),
396 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
398 # These patterns can result when (a < b || a < c) => (a < min(b, c))
399 # transformations occur before constant propagation and loop-unrolling.
400 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
401 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
402 (('~fge', a
, ('fmin', b
, a
)), True),
403 (('~fge', ('fmax', a
, b
), a
), True),
404 (('~flt', a
, ('fmin', b
, a
)), False),
405 (('~flt', ('fmax', a
, b
), a
), False),
406 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
407 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
409 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
410 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
411 (('ige', a
, ('imin', b
, a
)), True),
412 (('ige', ('imax', a
, b
), a
), True),
413 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
414 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
415 (('uge', a
, ('umin', b
, a
)), True),
416 (('uge', ('umax', a
, b
), a
), True),
417 (('ilt', a
, ('imin', b
, a
)), False),
418 (('ilt', ('imax', a
, b
), a
), False),
419 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
420 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
421 (('ult', a
, ('umin', b
, a
)), False),
422 (('ult', ('umax', a
, b
), a
), False),
423 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
424 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
426 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
427 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
428 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
429 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
430 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
431 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
432 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
433 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
434 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
435 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
436 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
437 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
438 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
439 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
440 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
441 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
443 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
445 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
446 ('iabs', ('ishr', a
, b
))),
447 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
449 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
450 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
451 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
452 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
453 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
454 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
455 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
456 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
457 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
458 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
460 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
461 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
462 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
463 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
464 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
465 (('ineg', ('b2i32', 'a@32')), a
),
466 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
467 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
468 # Comparison with the same args. Note that these are not done for
469 # the float versions because NaN always returns false on float
471 (('ilt', a
, a
), False),
472 (('ige', a
, a
), True),
473 (('ieq', a
, a
), True),
474 (('ine', a
, a
), False),
475 (('ult', a
, a
), False),
476 (('uge', a
, a
), True),
477 # Logical and bit operations
478 (('fand', a
, 0.0), 0.0),
480 (('iand', a
, ~
0), a
),
484 (('ior', a
, True), True),
485 (('fxor', a
, a
), 0.0),
488 (('inot', ('inot', a
)), a
),
489 (('ior', ('iand', a
, b
), b
), b
),
490 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
491 (('iand', ('ior', a
, b
), b
), b
),
492 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
494 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
495 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
496 # Shift optimizations
503 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
504 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
505 # Exponential/logarithmic identities
506 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
507 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
508 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
509 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
510 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
511 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
512 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
513 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
514 (('~fpow', a
, 1.0), a
),
515 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
516 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
517 (('~fpow', 2.0, a
), ('fexp2', a
)),
518 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
519 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
520 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
521 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
522 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
523 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
524 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
525 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
526 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
527 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
528 # Division and reciprocal
529 (('~fdiv', 1.0, a
), ('frcp', a
)),
530 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
531 (('~frcp', ('frcp', a
)), a
),
532 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
533 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
534 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
535 # Boolean simplifications
536 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
537 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
538 (('ieq', 'a@bool', True), a
),
539 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
540 (('ine', 'a@bool', False), a
),
541 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
542 (('bcsel', a
, True, False), a
),
543 (('bcsel', a
, False, True), ('inot', a
)),
544 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
545 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
546 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
547 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
548 (('bcsel', True, b
, c
), b
),
549 (('bcsel', False, b
, c
), c
),
550 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
551 # The result of this should be hit by constant propagation and, in the
552 # next round of opt_algebraic, get picked up by one of the above two.
553 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
555 (('bcsel', a
, b
, b
), b
),
556 (('fcsel', a
, b
, b
), b
),
558 # D3D Boolean emulation
559 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
560 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
561 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
562 ('ineg', ('b2i', ('iand', a
, b
)))),
563 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
564 ('ineg', ('b2i', ('ior', a
, b
)))),
565 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
566 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
567 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
568 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
569 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
571 # SM5 32-bit shifts are defined to use the 5 least significant bits
572 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
573 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
574 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
577 (('i2b32', ('b2i', 'a@32')), a
),
578 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
579 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
580 (('i2b', ('ineg', a
)), ('i2b', a
)),
581 (('i2b', ('iabs', a
)), ('i2b', a
)),
582 (('fabs', ('b2f', a
)), ('b2f', a
)),
583 (('iabs', ('b2i', a
)), ('b2i', a
)),
584 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
586 # Ironically, mark these as imprecise because removing the conversions may
587 # preserve more precision than doing the conversions (e.g.,
588 # uint(float(0x81818181u)) == 0x81818200).
589 (('~f2i32', ('i2f', 'a@32')), a
),
590 (('~f2i32', ('u2f', 'a@32')), a
),
591 (('~f2u32', ('i2f', 'a@32')), a
),
592 (('~f2u32', ('u2f', 'a@32')), a
),
594 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
597 # It is undefined to convert a negative floating-point value to an
600 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
601 # some optimizations in the i965 backend to proceed.
602 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
603 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
604 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
605 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
607 # Packing and then unpacking does nothing
608 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
609 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
610 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
611 ('unpack_64_2x32_split_y', a
)), a
),
614 (('ushr', ('ishl', 'a@32', 24), 24), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
615 (('ushr', ('ishl', 'a@32', 16), 24), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
616 (('ushr', ('ishl', 'a@32', 8), 24), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
617 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
618 (('ishr', ('ishl', 'a@32', 24), 24), ('extract_i8', a
, 0), '!options->lower_extract_byte'),
619 (('ishr', ('ishl', 'a@32', 16), 24), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
620 (('ishr', ('ishl', 'a@32', 8), 24), ('extract_i8', a
, 2), '!options->lower_extract_byte'),
621 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
622 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
623 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
624 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
627 # The ('extract_u8', a, 0) pattern, above, can trigger in cases where the
628 # shift count is based on a loop induction variable. Once the loop is
629 # unrolled, constant folding will generate patterns like those below.
630 for op
in ('ushr', 'ishr'):
631 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
632 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
633 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
635 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
637 # The ('extract_[iu]8', a, 3) patterns, above, can trigger in cases where the
638 # shift count is based on a loop induction variable. Once the loop is
639 # unrolled, constant folding will generate patterns like those below.
640 for op
in ('extract_u8', 'extract_i8'):
641 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
643 optimizations
.extend([
645 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
646 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
647 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
648 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
649 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
652 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
653 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
654 (('ussub_4x8', a
, 0), a
),
655 (('ussub_4x8', a
, ~
0), 0),
656 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
657 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
658 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
659 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
660 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
661 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
662 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
663 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
665 # Propagate negation up multiplication chains
666 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
667 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
669 # Propagate constants up multiplication chains
670 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
671 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
672 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
673 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
675 # Reassociate constants in add/mul chains so they can be folded together.
676 # For now, we mostly only handle cases where the constants are separated by
677 # a single non-constant. We could do better eventually.
678 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
679 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
680 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
681 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
682 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
685 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
686 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
687 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
689 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
690 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
691 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
693 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
696 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod16'),
697 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
698 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
699 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
700 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
701 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
703 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
704 ('bcsel', ('ilt', 31, 'bits'), 'insert',
705 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
706 'options->lower_bitfield_insert'),
707 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
708 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
709 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
710 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
711 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
712 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
714 # Alternative lowering that doesn't rely on bfi.
715 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
716 ('bcsel', ('ilt', 31, 'bits'),
719 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
720 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
721 'options->lower_bitfield_insert_to_shifts'),
723 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
724 (('bfm', 'bits', 'offset'),
725 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
726 'options->lower_bfm'),
728 (('ibitfield_extract', 'value', 'offset', 'bits'),
729 ('bcsel', ('ilt', 31, 'bits'), 'value',
730 ('ibfe', 'value', 'offset', 'bits')),
731 'options->lower_bitfield_extract'),
733 (('ubitfield_extract', 'value', 'offset', 'bits'),
734 ('bcsel', ('ult', 31, 'bits'), 'value',
735 ('ubfe', 'value', 'offset', 'bits')),
736 'options->lower_bitfield_extract'),
738 (('ibitfield_extract', 'value', 'offset', 'bits'),
739 ('bcsel', ('ieq', 0, 'bits'),
742 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
743 ('isub', 32, 'bits'))),
744 'options->lower_bitfield_extract_to_shifts'),
746 (('ubitfield_extract', 'value', 'offset', 'bits'),
748 ('ushr', 'value', 'offset'),
749 ('bcsel', ('ieq', 'bits', 32),
751 ('bfm', 'bits', 0))),
752 'options->lower_bitfield_extract_to_shifts'),
754 (('ifind_msb', 'value'),
755 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
756 'options->lower_ifind_msb'),
758 (('find_lsb', 'value'),
759 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
760 'options->lower_find_lsb'),
762 (('extract_i8', a
, 'b@32'),
763 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
764 'options->lower_extract_byte'),
766 (('extract_u8', a
, 'b@32'),
767 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
768 'options->lower_extract_byte'),
770 (('extract_i16', a
, 'b@32'),
771 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
772 'options->lower_extract_word'),
774 (('extract_u16', a
, 'b@32'),
775 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
776 'options->lower_extract_word'),
778 (('pack_unorm_2x16', 'v'),
779 ('pack_uvec2_to_uint',
780 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
781 'options->lower_pack_unorm_2x16'),
783 (('pack_unorm_4x8', 'v'),
784 ('pack_uvec4_to_uint',
785 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
786 'options->lower_pack_unorm_4x8'),
788 (('pack_snorm_2x16', 'v'),
789 ('pack_uvec2_to_uint',
790 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
791 'options->lower_pack_snorm_2x16'),
793 (('pack_snorm_4x8', 'v'),
794 ('pack_uvec4_to_uint',
795 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
796 'options->lower_pack_snorm_4x8'),
798 (('unpack_unorm_2x16', 'v'),
799 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
800 ('extract_u16', 'v', 1))),
802 'options->lower_unpack_unorm_2x16'),
804 (('unpack_unorm_4x8', 'v'),
805 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
806 ('extract_u8', 'v', 1),
807 ('extract_u8', 'v', 2),
808 ('extract_u8', 'v', 3))),
810 'options->lower_unpack_unorm_4x8'),
812 (('unpack_snorm_2x16', 'v'),
813 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
814 ('extract_i16', 'v', 1))),
816 'options->lower_unpack_snorm_2x16'),
818 (('unpack_snorm_4x8', 'v'),
819 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
820 ('extract_i8', 'v', 1),
821 ('extract_i8', 'v', 2),
822 ('extract_i8', 'v', 3))),
824 'options->lower_unpack_snorm_4x8'),
826 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
829 # bit_size dependent lowerings
830 for bit_size
in [8, 16, 32, 64]:
831 # convenience constants
832 intmax
= (1 << (bit_size
- 1)) - 1
833 intmin
= 1 << (bit_size
- 1)
836 (('iadd_sat@' + str(bit_size
), a
, b
),
837 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
838 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
839 (('isub_sat@' + str(bit_size
), a
, b
),
840 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
841 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
844 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
846 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
847 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
848 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
849 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
850 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
852 # Optimize x2bN(b2x(x)) -> x
853 for size
in type_sizes('bool'):
854 aN
= 'a@' + str(size
)
855 f2bN
= 'f2b' + str(size
)
856 i2bN
= 'i2b' + str(size
)
857 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
858 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
860 # Optimize x2yN(b2x(x)) -> b2y
861 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
862 if x
!= 'f' and y
!= 'f' and x
!= y
:
865 b2x
= 'b2f' if x
== 'f' else 'b2i'
866 b2y
= 'b2f' if y
== 'f' else 'b2i'
867 x2yN
= '{}2{}'.format(x
, y
)
868 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
870 def fexp2i(exp
, bits
):
871 # We assume that exp is already in the right range.
873 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
875 return ('ishl', ('iadd', exp
, 127), 23)
877 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
881 def ldexp(f
, exp
, bits
):
882 # First, we clamp exp to a reasonable range. The maximum possible range
883 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
884 # a maximum range of [-149, 127]. This means that we can potentially have
885 # a swing of +-276. If you start with FLT_MAX, you actually have to do
886 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
887 # spec, on the other hand, only requires that we handle an exponent value
888 # in the range [-126, 128]. This implementation is *mostly* correct; it
889 # handles a range on exp of [-252, 254] which allows you to create any
890 # value (including denorms if the hardware supports it) and to adjust the
891 # exponent of any normal value to anything you want.
893 exp
= ('imin', ('imax', exp
, -28), 30)
895 exp
= ('imin', ('imax', exp
, -252), 254)
897 exp
= ('imin', ('imax', exp
, -2044), 2046)
901 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
902 # (We use ishr which isn't the same for -1, but the -1 case still works
903 # since we use exp-exp/2 as the second exponent.) While the spec
904 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
905 # work with denormals and doesn't allow for the full swing in exponents
906 # that you can get with normalized values. Instead, we create two powers
907 # of two and multiply by them each in turn. That way the effective range
908 # of our exponent is doubled.
909 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
910 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
911 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
914 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
915 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
916 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
919 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
920 def bitfield_reverse(u
):
921 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
922 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
923 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
924 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
925 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
929 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
931 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
932 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
933 # and, if a is a NaN then the second comparison will fail anyway.
934 for op
in ['flt', 'fge', 'feq']:
936 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
937 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
940 # Add optimizations to handle the case where the result of a ternary is
941 # compared to a constant. This way we can take things like
947 # a ? (0 > 0) : (1 > 0)
949 # which constant folding will eat for lunch. The resulting ternary will
950 # further get cleaned up by the boolean reductions above and we will be
951 # left with just the original variable "a".
952 for op
in ['flt', 'fge', 'feq', 'fne',
953 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
955 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
956 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
957 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
958 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
962 # For example, this converts things like
964 # 1 + mix(0, a - 1, condition)
968 # mix(1, (a-1)+1, condition)
970 # Other optimizations will rearrange the constants.
971 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
973 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
976 # This section contains "late" optimizations that should be run before
977 # creating ffmas and calling regular optimizations for the final time.
978 # Optimizations should go here if they help code generation and conflict
979 # with the regular optimizations.
980 before_ffma_optimizations
= [
981 # Propagate constants down multiplication chains
982 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
983 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
984 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
985 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
987 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
988 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
989 (('~fadd', ('fneg', a
), a
), 0.0),
990 (('iadd', ('ineg', a
), a
), 0),
991 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
992 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
993 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
994 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
997 # This section contains "late" optimizations that should be run after the
998 # regular optimizations have finished. Optimizations should go here if
999 # they help code generation but do not necessarily produce code that is
1000 # more easily optimizable.
1001 late_optimizations
= [
1002 # Most of these optimizations aren't quite safe when you get infinity or
1003 # Nan involved but the first one should be fine.
1004 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1005 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1006 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1007 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1008 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1009 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1011 (('~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
)))),
1013 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1014 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1015 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1016 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1018 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a
, 0.0, 1.0)),
1019 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a
, -0.0, -1.0)),
1021 # we do these late so that we don't get in the way of creating ffmas
1022 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1023 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1025 (('bcsel', 'a@bool', 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1028 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1029 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1030 before_ffma_optimizations
).render())
1031 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1032 late_optimizations
).render())
projects / mesa.git / blob