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
40 # Written in the form (<search>, <replace>) where <search> is an expression
41 # and <replace> is either an expression or a value. An expression is
42 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
43 # where each source is either an expression or a value. A value can be
44 # either a numeric constant or a string representing a variable name.
46 # If the opcode in a search expression is prefixed by a '~' character, this
47 # indicates that the operation is inexact. Such operations will only get
48 # applied to SSA values that do not have the exact bit set. This should be
49 # used by by any optimizations that are not bit-for-bit exact. It should not,
50 # however, be used for backend-requested lowering operations as those need to
51 # happen regardless of precision.
53 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
54 # that the given variable will only match constants and the type indicates that
55 # the given variable will only match values from ALU instructions with the
56 # given output type, and (cond) specifies an additional condition function
57 # (see nir_search_helpers.h).
59 # For constants, you have to be careful to make sure that it is the right
60 # type because python is unaware of the source and destination types of the
63 # All expression types can have a bit-size specified. For opcodes, this
64 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
65 # type and size. In the search half of the expression this indicates that it
66 # should only match that particular bit-size. In the replace half of the
67 # expression this indicates that the constructed value should have that
72 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
73 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitshift'),
74 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitshift'),
76 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
77 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
78 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
79 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
84 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
85 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
86 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
87 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
89 (('fneg', ('fneg', a
)), a
),
90 (('ineg', ('ineg', a
)), a
),
91 (('fabs', ('fabs', a
)), ('fabs', a
)),
92 (('fabs', ('fneg', a
)), ('fabs', a
)),
93 (('fabs', ('u2f', a
)), ('u2f', a
)),
94 (('iabs', ('iabs', a
)), ('iabs', a
)),
95 (('iabs', ('ineg', a
)), ('iabs', a
)),
96 (('f2b', ('fneg', a
)), ('f2b', a
)),
97 (('i2b', ('ineg', a
)), ('i2b', a
)),
98 (('~fadd', a
, 0.0), a
),
100 (('usadd_4x8', a
, 0), a
),
101 (('usadd_4x8', a
, ~
0), ~
0),
102 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
103 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
104 (('~fadd', ('fneg', a
), a
), 0.0),
105 (('iadd', ('ineg', a
), a
), 0),
106 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
107 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
108 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
109 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
110 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
111 (('~fmul', a
, 0.0), 0.0),
113 (('umul_unorm_4x8', a
, 0), 0),
114 (('umul_unorm_4x8', a
, ~
0), a
),
115 (('fmul', a
, 1.0), a
),
117 (('fmul', a
, -1.0), ('fneg', a
)),
118 (('imul', a
, -1), ('ineg', a
)),
119 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
120 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
121 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
122 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
123 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
124 (('~ffma', 0.0, a
, b
), b
),
125 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
126 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
127 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
128 (('~flrp', a
, b
, 0.0), a
),
129 (('~flrp', a
, b
, 1.0), b
),
130 (('~flrp', a
, a
, b
), a
),
131 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
133 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
134 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
135 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
136 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
138 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
139 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
141 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
142 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
144 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
146 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
147 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
148 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
149 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
150 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
151 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
152 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
153 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
154 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
155 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
156 # These are the same as the previous three rules, but it depends on
157 # 1-fsat(x) <=> fsat(1-x). See below.
158 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
159 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
161 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
162 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
163 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
164 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
165 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
167 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
168 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
170 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
171 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
172 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
173 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
175 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
176 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
178 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
179 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
180 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
181 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
183 # (a * #b + #c) << #d
184 # ((a * #b) << #d) + (#c << #d)
185 # (a * (#b << #d)) + (#c << #d)
186 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
187 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
191 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
193 # Comparison simplifications
194 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
195 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
196 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
197 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
198 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
199 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
200 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
201 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
202 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
203 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
205 # This helps some shaders because, after some optimizations, they end up
206 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
207 # matching would be handled by CSE.
208 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
209 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
210 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
211 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
212 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
213 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
214 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
215 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
216 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
217 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
219 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
220 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
221 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
222 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
223 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
224 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
226 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
227 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
228 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
229 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
233 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
235 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
237 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
239 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
240 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
241 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
242 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
243 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
244 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
245 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
246 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
247 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
248 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
249 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
250 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
251 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
252 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
253 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
254 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
255 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
256 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
257 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
258 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
260 # -(b2f(a) + b2f(b)) < 0
261 # 0 < b2f(a) + b2f(b)
262 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
264 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
265 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
267 # -(b2f(a) + b2f(b)) >= 0
268 # 0 >= b2f(a) + b2f(b)
269 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
271 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
272 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
274 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
275 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
277 # Some optimizations (below) convert things like (a < b || c < b) into
278 # (min(a, c) < b). However, this interfers with the previous optimizations
279 # that try to remove comparisons with negated sums of b2f. This just
281 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
282 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
284 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
285 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
286 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
287 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
289 # Cannot remove the addition from ilt or ige due to overflow.
290 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
291 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
293 # fmin(-b2f(a), b) >= 0.0
294 # -b2f(a) >= 0.0 && b >= 0.0
295 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
296 # b2f(a) == 0.0 && b >= 0.0
297 # a == False && b >= 0.0
300 # The fge in the second replacement is not a typo. I leave the proof that
301 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
303 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
304 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
306 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
307 (('fne', ('b2f', 'a@1'), 0.0), a
),
308 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
309 (('ine', ('b2i', 'a@1'), 0), a
),
311 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
312 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
313 (('fge', ('u2f', a
), 0.0), True),
314 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
315 (('flt', ('u2f', a
), 0.0), False),
316 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
317 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
318 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
319 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
320 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
321 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
322 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
326 # fabs(a) != 0.0 because fabs(a) must be >= 0
328 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
332 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
335 # 0.0 == fabs(a) because fabs(a) must be >= 0
337 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
341 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
343 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
344 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
345 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
346 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
349 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
350 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
351 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
353 # Since b is a constant, constant folding will eliminate the fmin and the
354 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
355 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
357 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
359 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
360 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
361 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
362 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
363 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
364 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
365 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
366 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
367 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
368 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
369 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
370 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
371 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
372 (('bcsel', a
, True, b
), ('ior', a
, b
)),
373 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
374 (('bcsel', a
, b
, False), ('iand', a
, b
)),
375 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
382 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
383 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
384 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
385 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
386 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
387 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
388 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
389 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
390 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
391 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
392 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
393 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
394 (('fmin', a
, ('fabs', a
)), a
),
395 (('imin', a
, ('iabs', a
)), a
),
396 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
397 (('imax', a
, ('ineg', ('iabs', a
))), a
),
398 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
399 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
400 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
401 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
402 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
403 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
404 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
405 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
406 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
407 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
408 (('fsat', ('b2f', a
)), ('b2f', a
)),
409 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
410 (('fsat', ('fsat', a
)), ('fsat', a
)),
411 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_negate && !options->lower_fsat'),
412 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_negate && !options->lower_fsat'),
413 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
414 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
415 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
416 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
417 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
418 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
419 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
420 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
421 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
422 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
423 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
424 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
425 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
426 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
427 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
428 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
429 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
430 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
431 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
432 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
433 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
434 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
435 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
437 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
438 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
439 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
440 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
441 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
442 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
443 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
444 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
445 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
446 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
447 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
448 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
449 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
450 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
451 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
452 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
454 # Common pattern like 'if (i == 0 || i == 1 || ...)'
455 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
456 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
457 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
459 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
460 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
461 # so emit an open-coded version of that.
462 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
463 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
465 (('ior', a
, ('ieq', a
, False)), True),
466 (('ior', a
, ('inot', a
)), -1),
468 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
469 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
471 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
473 # These patterns can result when (a < b || a < c) => (a < min(b, c))
474 # transformations occur before constant propagation and loop-unrolling.
475 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
476 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
477 (('~fge', a
, ('fmin', b
, a
)), True),
478 (('~fge', ('fmax', a
, b
), a
), True),
479 (('~flt', a
, ('fmin', b
, a
)), False),
480 (('~flt', ('fmax', a
, b
), a
), False),
481 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
482 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
484 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
485 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
486 (('ige', a
, ('imin', b
, a
)), True),
487 (('ige', ('imax', a
, b
), a
), True),
488 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
489 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
490 (('uge', a
, ('umin', b
, a
)), True),
491 (('uge', ('umax', a
, b
), a
), True),
492 (('ilt', a
, ('imin', b
, a
)), False),
493 (('ilt', ('imax', a
, b
), a
), False),
494 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
495 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
496 (('ult', a
, ('umin', b
, a
)), False),
497 (('ult', ('umax', a
, b
), a
), False),
498 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
499 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
500 (('ult', a
, ('iand', b
, a
)), False),
501 (('ult', ('ior', a
, b
), a
), False),
502 (('uge', a
, ('iand', b
, a
)), True),
503 (('uge', ('ior', a
, b
), a
), True),
505 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
506 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
507 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
508 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
509 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
510 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
511 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
512 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
513 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
514 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
515 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
516 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
517 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
518 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
519 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
520 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
522 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
524 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
525 ('iabs', ('ishr', a
, b
))),
526 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
528 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
529 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
530 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
531 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
532 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
533 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
534 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
535 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
536 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
537 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
539 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
540 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
541 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
542 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
543 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
544 (('ineg', ('b2i32', 'a@32')), a
),
545 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
546 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
547 # Comparison with the same args. Note that these are not done for
548 # the float versions because NaN always returns false on float
550 (('ilt', a
, a
), False),
551 (('ige', a
, a
), True),
552 (('ieq', a
, a
), True),
553 (('ine', a
, a
), False),
554 (('ult', a
, a
), False),
555 (('uge', a
, a
), True),
556 # Logical and bit operations
558 (('iand', a
, ~
0), a
),
562 (('ior', a
, True), True),
565 (('inot', ('inot', a
)), a
),
566 (('ior', ('iand', a
, b
), b
), b
),
567 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
568 (('iand', ('ior', a
, b
), b
), b
),
569 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
571 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
572 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
573 # Shift optimizations
580 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
581 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
582 # Exponential/logarithmic identities
583 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
584 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
585 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
586 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
587 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
588 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
589 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
590 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
591 (('~fpow', a
, 1.0), a
),
592 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
593 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
594 (('~fpow', 2.0, a
), ('fexp2', a
)),
595 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
596 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
597 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
598 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
599 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
600 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
601 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
602 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
603 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
604 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
605 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
606 # Division and reciprocal
607 (('~fdiv', 1.0, a
), ('frcp', a
)),
608 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
609 (('~frcp', ('frcp', a
)), a
),
610 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
611 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
612 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
613 # Boolean simplifications
614 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
615 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
616 (('ieq', a
, True), a
),
617 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
618 (('ine', a
, False), a
),
619 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
620 (('bcsel', a
, True, False), a
),
621 (('bcsel', a
, False, True), ('inot', a
)),
622 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
623 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
624 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
625 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
626 (('bcsel', True, b
, c
), b
),
627 (('bcsel', False, b
, c
), c
),
628 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
630 (('bcsel', a
, b
, b
), b
),
631 (('fcsel', a
, b
, b
), b
),
633 # D3D Boolean emulation
634 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
635 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
636 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
637 ('ineg', ('b2i', ('iand', a
, b
)))),
638 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
639 ('ineg', ('b2i', ('ior', a
, b
)))),
640 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
641 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
642 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
643 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
644 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
646 # SM5 32-bit shifts are defined to use the 5 least significant bits
647 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
648 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
649 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
652 (('i2b32', ('b2i', 'a@32')), a
),
653 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
654 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
655 (('i2b', ('ineg', a
)), ('i2b', a
)),
656 (('i2b', ('iabs', a
)), ('i2b', a
)),
657 (('fabs', ('b2f', a
)), ('b2f', a
)),
658 (('iabs', ('b2i', a
)), ('b2i', a
)),
659 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
661 # Ironically, mark these as imprecise because removing the conversions may
662 # preserve more precision than doing the conversions (e.g.,
663 # uint(float(0x81818181u)) == 0x81818200).
664 (('~f2i32', ('i2f', 'a@32')), a
),
665 (('~f2i32', ('u2f', 'a@32')), a
),
666 (('~f2u32', ('i2f', 'a@32')), a
),
667 (('~f2u32', ('u2f', 'a@32')), a
),
669 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
672 # It is undefined to convert a negative floating-point value to an
675 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
676 # some optimizations in the i965 backend to proceed.
677 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
678 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
679 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
680 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
682 (('~fmin', ('fabs', a
), 1.0), ('fsat', ('fabs', a
)), '!options->lower_fsat'),
684 # The result of the multiply must be in [-1, 0], so the result of the ffma
686 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
687 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
688 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
689 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
691 # Packing and then unpacking does nothing
692 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
693 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
694 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
695 ('unpack_64_2x32_split_y', a
)), a
),
698 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
699 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
700 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
701 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
702 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
703 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
704 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
707 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
708 # patterns like those below.
709 for op
in ('ushr', 'ishr'):
710 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
711 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
712 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
714 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
716 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
717 # patterns like those below.
718 for op
in ('extract_u8', 'extract_i8'):
719 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
720 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
721 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
723 optimizations
.extend([
725 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
726 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
727 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
728 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
729 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
732 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
733 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
734 (('ussub_4x8', a
, 0), a
),
735 (('ussub_4x8', a
, ~
0), 0),
736 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
737 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
738 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
739 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
740 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
741 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
742 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
743 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
745 # Propagate negation up multiplication chains
746 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
747 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
749 # Propagate constants up multiplication chains
750 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
751 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
752 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
753 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
755 # Reassociate constants in add/mul chains so they can be folded together.
756 # For now, we mostly only handle cases where the constants are separated by
757 # a single non-constant. We could do better eventually.
758 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
759 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
760 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
761 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
762 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
764 # Drop mul-div by the same value when there's no wrapping.
765 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
768 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
769 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
770 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
772 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
773 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
774 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
776 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
779 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
780 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
781 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
782 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
783 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
785 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
786 ('bcsel', ('ult', 31, 'bits'), 'insert',
787 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
788 'options->lower_bitfield_insert'),
789 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
790 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
791 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
792 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
793 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
794 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
796 # Alternative lowering that doesn't rely on bfi.
797 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
799 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
800 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
801 'options->lower_bitfield_insert_to_shifts'),
803 # Alternative lowering that uses bitfield_select.
804 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
805 ('bcsel', ('ult', 31, 'bits'), 'insert',
806 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
807 'options->lower_bitfield_insert_to_bitfield_select'),
809 (('ibitfield_extract', 'value', 'offset', 'bits'),
810 ('bcsel', ('ult', 31, 'bits'), 'value',
811 ('ibfe', 'value', 'offset', 'bits')),
812 'options->lower_bitfield_extract'),
814 (('ubitfield_extract', 'value', 'offset', 'bits'),
815 ('bcsel', ('ult', 31, 'bits'), 'value',
816 ('ubfe', 'value', 'offset', 'bits')),
817 'options->lower_bitfield_extract'),
819 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
820 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
821 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
822 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
823 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
824 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
825 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
827 (('ibitfield_extract', 'value', 'offset', 'bits'),
828 ('bcsel', ('ieq', 0, 'bits'),
831 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
832 ('isub', 32, 'bits'))),
833 'options->lower_bitfield_extract_to_shifts'),
835 (('ubitfield_extract', 'value', 'offset', 'bits'),
837 ('ushr', 'value', 'offset'),
838 ('bcsel', ('ieq', 'bits', 32),
840 ('isub', ('ishl', 1, 'bits'), 1))),
841 'options->lower_bitfield_extract_to_shifts'),
843 (('ifind_msb', 'value'),
844 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
845 'options->lower_ifind_msb'),
847 (('find_lsb', 'value'),
848 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
849 'options->lower_find_lsb'),
851 (('extract_i8', a
, 'b@32'),
852 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
853 'options->lower_extract_byte'),
855 (('extract_u8', a
, 'b@32'),
856 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
857 'options->lower_extract_byte'),
859 (('extract_i16', a
, 'b@32'),
860 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
861 'options->lower_extract_word'),
863 (('extract_u16', a
, 'b@32'),
864 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
865 'options->lower_extract_word'),
867 (('pack_unorm_2x16', 'v'),
868 ('pack_uvec2_to_uint',
869 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
870 'options->lower_pack_unorm_2x16'),
872 (('pack_unorm_4x8', 'v'),
873 ('pack_uvec4_to_uint',
874 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
875 'options->lower_pack_unorm_4x8'),
877 (('pack_snorm_2x16', 'v'),
878 ('pack_uvec2_to_uint',
879 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
880 'options->lower_pack_snorm_2x16'),
882 (('pack_snorm_4x8', 'v'),
883 ('pack_uvec4_to_uint',
884 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
885 'options->lower_pack_snorm_4x8'),
887 (('unpack_unorm_2x16', 'v'),
888 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
889 ('extract_u16', 'v', 1))),
891 'options->lower_unpack_unorm_2x16'),
893 (('unpack_unorm_4x8', 'v'),
894 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
895 ('extract_u8', 'v', 1),
896 ('extract_u8', 'v', 2),
897 ('extract_u8', 'v', 3))),
899 'options->lower_unpack_unorm_4x8'),
901 (('unpack_snorm_2x16', 'v'),
902 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
903 ('extract_i16', 'v', 1))),
905 'options->lower_unpack_snorm_2x16'),
907 (('unpack_snorm_4x8', 'v'),
908 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
909 ('extract_i8', 'v', 1),
910 ('extract_i8', 'v', 2),
911 ('extract_i8', 'v', 3))),
913 'options->lower_unpack_snorm_4x8'),
915 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
916 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
919 # bit_size dependent lowerings
920 for bit_size
in [8, 16, 32, 64]:
921 # convenience constants
922 intmax
= (1 << (bit_size
- 1)) - 1
923 intmin
= 1 << (bit_size
- 1)
926 (('iadd_sat@' + str(bit_size
), a
, b
),
927 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
928 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
929 (('isub_sat@' + str(bit_size
), a
, b
),
930 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
931 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
934 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
936 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
937 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
938 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
939 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
940 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
942 # Optimize x2bN(b2x(x)) -> x
943 for size
in type_sizes('bool'):
944 aN
= 'a@' + str(size
)
945 f2bN
= 'f2b' + str(size
)
946 i2bN
= 'i2b' + str(size
)
947 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
948 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
950 # Optimize x2yN(b2x(x)) -> b2y
951 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
952 if x
!= 'f' and y
!= 'f' and x
!= y
:
955 b2x
= 'b2f' if x
== 'f' else 'b2i'
956 b2y
= 'b2f' if y
== 'f' else 'b2i'
957 x2yN
= '{}2{}'.format(x
, y
)
958 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
960 # Optimize away x2xN(a@N)
961 for t
in ['int', 'uint', 'float']:
962 for N
in type_sizes(t
):
963 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
964 aN
= 'a@{0}'.format(N
)
965 optimizations
.append(((x2xN
, aN
), a
))
967 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
968 # In particular, we can optimize away everything except upcast of downcast and
969 # upcasts where the type differs from the other cast
970 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
972 # The outer cast is a down-cast. It doesn't matter what the size of the
973 # argument of the inner cast is because we'll never been in the upcast
974 # of downcast case. Regardless of types, we'll always end up with y2yN
976 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
977 x2xN
= '{0}2{0}{1}'.format(x
, N
)
978 y2yM
= '{0}2{0}{1}'.format(y
, M
)
979 y2yN
= '{0}2{0}{1}'.format(y
, N
)
980 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
982 # If the outer cast is an up-cast, we have to be more careful about the
983 # size of the argument of the inner cast and with types. In this case,
984 # the type is always the type of type up-cast which is given by the
986 for P
in type_sizes('uint'):
987 # We can't optimize away up-cast of down-cast.
991 # Because we're doing down-cast of down-cast, the types always have
992 # to match between the two casts
994 x2xN
= '{0}2{0}{1}'.format(x
, N
)
995 x2xM
= '{0}2{0}{1}'.format(x
, M
)
996 aP
= 'a@{0}'.format(P
)
997 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
999 # The N == M case is handled by other optimizations
1002 def fexp2i(exp
, bits
):
1003 # We assume that exp is already in the right range.
1005 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1007 return ('ishl', ('iadd', exp
, 127), 23)
1009 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1013 def ldexp(f
, exp
, bits
):
1014 # First, we clamp exp to a reasonable range. The maximum possible range
1015 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1016 # a maximum range of [-149, 127]. This means that we can potentially have
1017 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1018 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1019 # spec, on the other hand, only requires that we handle an exponent value
1020 # in the range [-126, 128]. This implementation is *mostly* correct; it
1021 # handles a range on exp of [-252, 254] which allows you to create any
1022 # value (including denorms if the hardware supports it) and to adjust the
1023 # exponent of any normal value to anything you want.
1025 exp
= ('imin', ('imax', exp
, -28), 30)
1027 exp
= ('imin', ('imax', exp
, -252), 254)
1029 exp
= ('imin', ('imax', exp
, -2044), 2046)
1033 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1034 # (We use ishr which isn't the same for -1, but the -1 case still works
1035 # since we use exp-exp/2 as the second exponent.) While the spec
1036 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1037 # work with denormals and doesn't allow for the full swing in exponents
1038 # that you can get with normalized values. Instead, we create two powers
1039 # of two and multiply by them each in turn. That way the effective range
1040 # of our exponent is doubled.
1041 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1042 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1043 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1046 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1047 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1048 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1051 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1052 def bitfield_reverse(u
):
1053 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1054 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1055 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1056 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1057 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1061 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1063 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1064 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1065 # and, if a is a NaN then the second comparison will fail anyway.
1066 for op
in ['flt', 'fge', 'feq']:
1068 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1069 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1072 # Add optimizations to handle the case where the result of a ternary is
1073 # compared to a constant. This way we can take things like
1079 # a ? (0 > 0) : (1 > 0)
1081 # which constant folding will eat for lunch. The resulting ternary will
1082 # further get cleaned up by the boolean reductions above and we will be
1083 # left with just the original variable "a".
1084 for op
in ['flt', 'fge', 'feq', 'fne',
1085 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1087 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1088 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1089 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1090 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1094 # For example, this converts things like
1096 # 1 + mix(0, a - 1, condition)
1100 # mix(1, (a-1)+1, condition)
1102 # Other optimizations will rearrange the constants.
1103 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1105 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1108 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1111 # If neither layout qualifier is specified, derivatives in compute shaders
1112 # return zero, which is consistent with the handling of built-in texture
1113 # functions like texture() in GLSL 4.50 compute shaders.
1114 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1115 'fddy', 'fddy_fine', 'fddy_coarse']:
1117 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1120 # Some optimizations for ir3-specific instructions.
1122 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1123 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1124 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1125 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1126 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1129 # This section contains "late" optimizations that should be run before
1130 # creating ffmas and calling regular optimizations for the final time.
1131 # Optimizations should go here if they help code generation and conflict
1132 # with the regular optimizations.
1133 before_ffma_optimizations
= [
1134 # Propagate constants down multiplication chains
1135 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1136 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1137 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1138 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1140 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1141 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1142 (('~fadd', ('fneg', a
), a
), 0.0),
1143 (('iadd', ('ineg', a
), a
), 0),
1144 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1145 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1146 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1147 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1149 (('~flrp@32', ('fadd(is_used_once)', a
, -1.0), ('fadd(is_used_once)', a
, 1.0), d
), ('fadd', ('flrp', -1.0, 1.0, d
), a
)),
1150 (('~flrp@32', ('fadd(is_used_once)', a
, 1.0), ('fadd(is_used_once)', a
, -1.0), d
), ('fadd', ('flrp', 1.0, -1.0, d
), a
)),
1151 (('~flrp@32', ('fadd(is_used_once)', a
, '#b'), ('fadd(is_used_once)', a
, '#c'), d
), ('fadd', ('fmul', d
, ('fadd', c
, ('fneg', b
))), ('fadd', a
, b
))),
1154 # This section contains "late" optimizations that should be run after the
1155 # regular optimizations have finished. Optimizations should go here if
1156 # they help code generation but do not necessarily produce code that is
1157 # more easily optimizable.
1158 late_optimizations
= [
1159 # Most of these optimizations aren't quite safe when you get infinity or
1160 # Nan involved but the first one should be fine.
1161 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1162 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1163 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1164 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1165 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1166 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1168 # nir_lower_to_source_mods will collapse this, but its existence during the
1169 # optimization loop can prevent other optimizations.
1170 (('fneg', ('fneg', a
)), a
),
1172 # These are duplicated from the main optimizations table. The late
1173 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1174 # new patterns like these. The patterns that compare with zero are removed
1175 # because they are unlikely to be created in by anything in
1176 # late_optimizations.
1177 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1178 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1179 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1180 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1181 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1182 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1184 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1185 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1187 (('~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
)))),
1189 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1190 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1191 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1192 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1193 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1194 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1195 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1196 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1197 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1198 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1201 (('iand', a
, a
), a
),
1203 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1205 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1206 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1207 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1208 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1210 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1211 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1213 (('~fadd@32', 1.0, ('fmul(is_used_once)', c
, ('fadd', b
, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c
)), ('fmul', b
, c
)), 'options->lower_flrp32'),
1214 (('~fadd@64', 1.0, ('fmul(is_used_once)', c
, ('fadd', b
, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c
)), ('fmul', b
, c
)), 'options->lower_flrp64'),
1216 # we do these late so that we don't get in the way of creating ffmas
1217 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1218 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1220 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1223 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1224 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1225 before_ffma_optimizations
).render())
1226 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1227 late_optimizations
).render())