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
557 (('fand', a
, 0.0), 0.0),
559 (('iand', a
, ~
0), a
),
563 (('ior', a
, True), True),
564 (('fxor', a
, a
), 0.0),
567 (('inot', ('inot', a
)), a
),
568 (('ior', ('iand', a
, b
), b
), b
),
569 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
570 (('iand', ('ior', a
, b
), b
), b
),
571 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
573 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
574 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
575 # Shift optimizations
582 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
583 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
584 # Exponential/logarithmic identities
585 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
586 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
587 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
588 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
589 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
590 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
591 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
592 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
593 (('~fpow', a
, 1.0), a
),
594 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
595 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
596 (('~fpow', 2.0, a
), ('fexp2', a
)),
597 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
598 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
599 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
600 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
601 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
602 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
603 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
604 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
605 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
606 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
607 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
608 # Division and reciprocal
609 (('~fdiv', 1.0, a
), ('frcp', a
)),
610 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
611 (('~frcp', ('frcp', a
)), a
),
612 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
613 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
614 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
615 # Boolean simplifications
616 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
617 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
618 (('ieq', a
, True), a
),
619 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
620 (('ine', a
, False), a
),
621 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
622 (('bcsel', a
, True, False), a
),
623 (('bcsel', a
, False, True), ('inot', a
)),
624 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
625 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
626 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
627 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
628 (('bcsel', True, b
, c
), b
),
629 (('bcsel', False, b
, c
), c
),
630 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
632 (('bcsel', a
, b
, b
), b
),
633 (('fcsel', a
, b
, b
), b
),
635 # D3D Boolean emulation
636 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
637 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
638 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
639 ('ineg', ('b2i', ('iand', a
, b
)))),
640 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
641 ('ineg', ('b2i', ('ior', a
, b
)))),
642 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
643 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
644 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
645 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
646 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
648 # SM5 32-bit shifts are defined to use the 5 least significant bits
649 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
650 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
651 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
654 (('i2b32', ('b2i', 'a@32')), a
),
655 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
656 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
657 (('i2b', ('ineg', a
)), ('i2b', a
)),
658 (('i2b', ('iabs', a
)), ('i2b', a
)),
659 (('fabs', ('b2f', a
)), ('b2f', a
)),
660 (('iabs', ('b2i', a
)), ('b2i', a
)),
661 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
663 # Ironically, mark these as imprecise because removing the conversions may
664 # preserve more precision than doing the conversions (e.g.,
665 # uint(float(0x81818181u)) == 0x81818200).
666 (('~f2i32', ('i2f', 'a@32')), a
),
667 (('~f2i32', ('u2f', 'a@32')), a
),
668 (('~f2u32', ('i2f', 'a@32')), a
),
669 (('~f2u32', ('u2f', 'a@32')), a
),
671 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
674 # It is undefined to convert a negative floating-point value to an
677 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
678 # some optimizations in the i965 backend to proceed.
679 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
680 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
681 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
682 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
684 (('~fmin', ('fabs', a
), 1.0), ('fsat', ('fabs', a
)), '!options->lower_fsat'),
686 # The result of the multiply must be in [-1, 0], so the result of the ffma
688 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
689 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
690 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
691 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
693 # Packing and then unpacking does nothing
694 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
695 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
696 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
697 ('unpack_64_2x32_split_y', a
)), a
),
700 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
701 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
702 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
703 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
704 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
705 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
706 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
709 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
710 # patterns like those below.
711 for op
in ('ushr', 'ishr'):
712 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
713 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
714 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
716 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
718 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
719 # patterns like those below.
720 for op
in ('extract_u8', 'extract_i8'):
721 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
722 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
723 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
725 optimizations
.extend([
727 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
728 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
729 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
730 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
731 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
734 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
735 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
736 (('ussub_4x8', a
, 0), a
),
737 (('ussub_4x8', a
, ~
0), 0),
738 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
739 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
740 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
741 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
742 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
743 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
744 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
745 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
747 # Propagate negation up multiplication chains
748 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
749 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
751 # Propagate constants up multiplication chains
752 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
753 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
754 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
755 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
757 # Reassociate constants in add/mul chains so they can be folded together.
758 # For now, we mostly only handle cases where the constants are separated by
759 # a single non-constant. We could do better eventually.
760 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
761 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
762 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
763 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
764 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
767 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
768 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
769 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
771 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
772 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
773 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
775 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
778 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
779 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
780 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
781 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
782 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
784 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
785 ('bcsel', ('ilt', 31, 'bits'), 'insert',
786 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
787 'options->lower_bitfield_insert'),
788 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
789 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
790 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
791 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
792 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
793 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
795 # Alternative lowering that doesn't rely on bfi.
796 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
797 ('bcsel', ('ilt', 31, 'bits'),
800 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
801 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
802 'options->lower_bitfield_insert_to_shifts'),
804 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
805 (('bfm', 'bits', 'offset'),
806 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
807 'options->lower_bfm'),
809 (('ibitfield_extract', 'value', 'offset', 'bits'),
810 ('bcsel', ('ilt', 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 (('ibitfield_extract', 'value', 'offset', 'bits'),
820 ('bcsel', ('ieq', 0, 'bits'),
823 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
824 ('isub', 32, 'bits'))),
825 'options->lower_bitfield_extract_to_shifts'),
827 (('ubitfield_extract', 'value', 'offset', 'bits'),
829 ('ushr', 'value', 'offset'),
830 ('bcsel', ('ieq', 'bits', 32),
832 ('bfm', 'bits', 0))),
833 'options->lower_bitfield_extract_to_shifts'),
835 (('ifind_msb', 'value'),
836 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
837 'options->lower_ifind_msb'),
839 (('find_lsb', 'value'),
840 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
841 'options->lower_find_lsb'),
843 (('extract_i8', a
, 'b@32'),
844 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
845 'options->lower_extract_byte'),
847 (('extract_u8', a
, 'b@32'),
848 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
849 'options->lower_extract_byte'),
851 (('extract_i16', a
, 'b@32'),
852 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
853 'options->lower_extract_word'),
855 (('extract_u16', a
, 'b@32'),
856 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
857 'options->lower_extract_word'),
859 (('pack_unorm_2x16', 'v'),
860 ('pack_uvec2_to_uint',
861 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
862 'options->lower_pack_unorm_2x16'),
864 (('pack_unorm_4x8', 'v'),
865 ('pack_uvec4_to_uint',
866 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
867 'options->lower_pack_unorm_4x8'),
869 (('pack_snorm_2x16', 'v'),
870 ('pack_uvec2_to_uint',
871 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
872 'options->lower_pack_snorm_2x16'),
874 (('pack_snorm_4x8', 'v'),
875 ('pack_uvec4_to_uint',
876 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
877 'options->lower_pack_snorm_4x8'),
879 (('unpack_unorm_2x16', 'v'),
880 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
881 ('extract_u16', 'v', 1))),
883 'options->lower_unpack_unorm_2x16'),
885 (('unpack_unorm_4x8', 'v'),
886 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
887 ('extract_u8', 'v', 1),
888 ('extract_u8', 'v', 2),
889 ('extract_u8', 'v', 3))),
891 'options->lower_unpack_unorm_4x8'),
893 (('unpack_snorm_2x16', 'v'),
894 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
895 ('extract_i16', 'v', 1))),
897 'options->lower_unpack_snorm_2x16'),
899 (('unpack_snorm_4x8', 'v'),
900 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
901 ('extract_i8', 'v', 1),
902 ('extract_i8', 'v', 2),
903 ('extract_i8', 'v', 3))),
905 'options->lower_unpack_snorm_4x8'),
907 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
908 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
911 # bit_size dependent lowerings
912 for bit_size
in [8, 16, 32, 64]:
913 # convenience constants
914 intmax
= (1 << (bit_size
- 1)) - 1
915 intmin
= 1 << (bit_size
- 1)
918 (('iadd_sat@' + str(bit_size
), a
, b
),
919 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
920 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
921 (('isub_sat@' + str(bit_size
), a
, b
),
922 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
923 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
926 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
928 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
929 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
930 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
931 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
932 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
934 # Optimize x2bN(b2x(x)) -> x
935 for size
in type_sizes('bool'):
936 aN
= 'a@' + str(size
)
937 f2bN
= 'f2b' + str(size
)
938 i2bN
= 'i2b' + str(size
)
939 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
940 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
942 # Optimize x2yN(b2x(x)) -> b2y
943 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
944 if x
!= 'f' and y
!= 'f' and x
!= y
:
947 b2x
= 'b2f' if x
== 'f' else 'b2i'
948 b2y
= 'b2f' if y
== 'f' else 'b2i'
949 x2yN
= '{}2{}'.format(x
, y
)
950 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
952 # Optimize away x2xN(a@N)
953 for t
in ['int', 'uint', 'float']:
954 for N
in type_sizes(t
):
955 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
956 aN
= 'a@{0}'.format(N
)
957 optimizations
.append(((x2xN
, aN
), a
))
959 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
960 # In particular, we can optimize away everything except upcast of downcast and
961 # upcasts where the type differs from the other cast
962 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
964 # The outer cast is a down-cast. It doesn't matter what the size of the
965 # argument of the inner cast is because we'll never been in the upcast
966 # of downcast case. Regardless of types, we'll always end up with y2yN
968 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
969 x2xN
= '{0}2{0}{1}'.format(x
, N
)
970 y2yM
= '{0}2{0}{1}'.format(y
, M
)
971 y2yN
= '{0}2{0}{1}'.format(y
, N
)
972 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
974 # If the outer cast is an up-cast, we have to be more careful about the
975 # size of the argument of the inner cast and with types. In this case,
976 # the type is always the type of type up-cast which is given by the
978 for P
in type_sizes('uint'):
979 # We can't optimize away up-cast of down-cast.
983 # Because we're doing down-cast of down-cast, the types always have
984 # to match between the two casts
986 x2xN
= '{0}2{0}{1}'.format(x
, N
)
987 x2xM
= '{0}2{0}{1}'.format(x
, M
)
988 aP
= 'a@{0}'.format(P
)
989 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
991 # The N == M case is handled by other optimizations
994 def fexp2i(exp
, bits
):
995 # We assume that exp is already in the right range.
997 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
999 return ('ishl', ('iadd', exp
, 127), 23)
1001 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1005 def ldexp(f
, exp
, bits
):
1006 # First, we clamp exp to a reasonable range. The maximum possible range
1007 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1008 # a maximum range of [-149, 127]. This means that we can potentially have
1009 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1010 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1011 # spec, on the other hand, only requires that we handle an exponent value
1012 # in the range [-126, 128]. This implementation is *mostly* correct; it
1013 # handles a range on exp of [-252, 254] which allows you to create any
1014 # value (including denorms if the hardware supports it) and to adjust the
1015 # exponent of any normal value to anything you want.
1017 exp
= ('imin', ('imax', exp
, -28), 30)
1019 exp
= ('imin', ('imax', exp
, -252), 254)
1021 exp
= ('imin', ('imax', exp
, -2044), 2046)
1025 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1026 # (We use ishr which isn't the same for -1, but the -1 case still works
1027 # since we use exp-exp/2 as the second exponent.) While the spec
1028 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1029 # work with denormals and doesn't allow for the full swing in exponents
1030 # that you can get with normalized values. Instead, we create two powers
1031 # of two and multiply by them each in turn. That way the effective range
1032 # of our exponent is doubled.
1033 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1034 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1035 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1038 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1039 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1040 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1043 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1044 def bitfield_reverse(u
):
1045 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1046 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1047 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1048 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1049 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1053 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1055 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1056 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1057 # and, if a is a NaN then the second comparison will fail anyway.
1058 for op
in ['flt', 'fge', 'feq']:
1060 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1061 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1064 # Add optimizations to handle the case where the result of a ternary is
1065 # compared to a constant. This way we can take things like
1071 # a ? (0 > 0) : (1 > 0)
1073 # which constant folding will eat for lunch. The resulting ternary will
1074 # further get cleaned up by the boolean reductions above and we will be
1075 # left with just the original variable "a".
1076 for op
in ['flt', 'fge', 'feq', 'fne',
1077 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1079 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1080 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1081 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1082 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1086 # For example, this converts things like
1088 # 1 + mix(0, a - 1, condition)
1092 # mix(1, (a-1)+1, condition)
1094 # Other optimizations will rearrange the constants.
1095 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1097 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1100 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1103 # If neither layout qualifier is specified, derivatives in compute shaders
1104 # return zero, which is consistent with the handling of built-in texture
1105 # functions like texture() in GLSL 4.50 compute shaders.
1106 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1107 'fddy', 'fddy_fine', 'fddy_coarse']:
1109 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1112 # Some optimizations for ir3-specific instructions.
1114 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1115 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1116 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1117 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1118 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1121 # This section contains "late" optimizations that should be run before
1122 # creating ffmas and calling regular optimizations for the final time.
1123 # Optimizations should go here if they help code generation and conflict
1124 # with the regular optimizations.
1125 before_ffma_optimizations
= [
1126 # Propagate constants down multiplication chains
1127 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1128 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1129 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1130 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1132 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1133 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1134 (('~fadd', ('fneg', a
), a
), 0.0),
1135 (('iadd', ('ineg', a
), a
), 0),
1136 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1137 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1138 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1139 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1141 (('~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
)),
1142 (('~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
)),
1143 (('~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
))),
1146 # This section contains "late" optimizations that should be run after the
1147 # regular optimizations have finished. Optimizations should go here if
1148 # they help code generation but do not necessarily produce code that is
1149 # more easily optimizable.
1150 late_optimizations
= [
1151 # Most of these optimizations aren't quite safe when you get infinity or
1152 # Nan involved but the first one should be fine.
1153 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1154 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1155 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1156 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1157 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1158 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1160 # nir_lower_to_source_mods will collapse this, but its existence during the
1161 # optimization loop can prevent other optimizations.
1162 (('fneg', ('fneg', a
)), a
),
1164 # These are duplicated from the main optimizations table. The late
1165 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1166 # new patterns like these. The patterns that compare with zero are removed
1167 # because they are unlikely to be created in by anything in
1168 # late_optimizations.
1169 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1170 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1171 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1172 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1173 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1174 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1176 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1177 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1179 (('~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
)))),
1181 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1182 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1183 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1184 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1185 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1186 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1187 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1188 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1189 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1190 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1193 (('iand', a
, a
), a
),
1195 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1197 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1198 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1199 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1200 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1202 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1203 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1205 (('~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'),
1206 (('~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'),
1208 # we do these late so that we don't get in the way of creating ffmas
1209 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1210 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1212 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1215 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1216 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1217 before_ffma_optimizations
).render())
1218 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1219 late_optimizations
).render())