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
34 # Convenience variables
41 # Written in the form (<search>, <replace>) where <search> is an expression
42 # and <replace> is either an expression or a value. An expression is
43 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
44 # where each source is either an expression or a value. A value can be
45 # either a numeric constant or a string representing a variable name.
47 # If the opcode in a search expression is prefixed by a '~' character, this
48 # indicates that the operation is inexact. Such operations will only get
49 # applied to SSA values that do not have the exact bit set. This should be
50 # used by by any optimizations that are not bit-for-bit exact. It should not,
51 # however, be used for backend-requested lowering operations as those need to
52 # happen regardless of precision.
54 # Variable names are specified as "[#]name[@type][(cond)][.swiz]" where:
55 # "#" indicates that the given variable will only match constants,
56 # type indicates that the given variable will only match values from ALU
57 # instructions with the given output type,
58 # (cond) specifies an additional condition function (see nir_search_helpers.h),
59 # swiz is a swizzle applied to the variable (only in the <replace> expression)
61 # For constants, you have to be careful to make sure that it is the right
62 # type because python is unaware of the source and destination types of the
65 # All expression types can have a bit-size specified. For opcodes, this
66 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
67 # type and size. In the search half of the expression this indicates that it
68 # should only match that particular bit-size. In the replace half of the
69 # expression this indicates that the constructed value should have that
72 # A special condition "many-comm-expr" can be used with expressions to note
73 # that the expression and its subexpressions have more commutative expressions
74 # than nir_replace_instr can handle. If this special condition is needed with
75 # another condition, the two can be separated by a comma (e.g.,
76 # "(many-comm-expr,is_used_once)").
78 # based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
79 def lowered_sincos(c
):
80 x
= ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi
, a
), c
))), 1.0)
81 x
= ('fmul', ('fsub', x
, ('fmul', x
, ('fabs', x
))), 4.0)
82 return ('ffma', ('ffma', x
, ('fabs', x
), ('fneg', x
)), 0.225, x
)
86 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
87 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitshift'),
88 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitshift'),
90 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
91 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
92 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
93 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
98 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
99 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
100 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
101 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
103 (('fneg', ('fneg', a
)), a
),
104 (('ineg', ('ineg', a
)), a
),
105 (('fabs', ('fabs', a
)), ('fabs', a
)),
106 (('fabs', ('fneg', a
)), ('fabs', a
)),
107 (('fabs', ('u2f', a
)), ('u2f', a
)),
108 (('iabs', ('iabs', a
)), ('iabs', a
)),
109 (('iabs', ('ineg', a
)), ('iabs', a
)),
110 (('f2b', ('fneg', a
)), ('f2b', a
)),
111 (('i2b', ('ineg', a
)), ('i2b', a
)),
112 (('~fadd', a
, 0.0), a
),
114 (('usadd_4x8', a
, 0), a
),
115 (('usadd_4x8', a
, ~
0), ~
0),
116 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
117 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
118 (('~fadd', ('fneg', a
), a
), 0.0),
119 (('iadd', ('ineg', a
), a
), 0),
120 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
121 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
122 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
123 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
124 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
125 (('~fmul', a
, 0.0), 0.0),
127 (('umul_unorm_4x8', a
, 0), 0),
128 (('umul_unorm_4x8', a
, ~
0), a
),
129 (('fmul', a
, 1.0), a
),
131 (('fmul', a
, -1.0), ('fneg', a
)),
132 (('imul', a
, -1), ('ineg', a
)),
133 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
134 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
135 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
136 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
137 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
138 (('~ffma', 0.0, a
, b
), b
),
139 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
140 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
141 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
142 (('~flrp', a
, b
, 0.0), a
),
143 (('~flrp', a
, b
, 1.0), b
),
144 (('~flrp', a
, a
, b
), a
),
145 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
147 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
148 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
149 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
150 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
152 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
153 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
155 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
156 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
158 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
160 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
161 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
162 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
163 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
164 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
165 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
166 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
167 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
168 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
169 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
170 # These are the same as the previous three rules, but it depends on
171 # 1-fsat(x) <=> fsat(1-x). See below.
172 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
173 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
175 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
176 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
177 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
178 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
179 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
181 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
182 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
184 (('fdph', a
, b
), ('fdot4', ('vec4', 'a.x', 'a.y', 'a.z', 1.0), b
), 'options->lower_fdph'),
186 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
), '!options->lower_fdph'),
187 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
188 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
189 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
191 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
192 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
194 (('fdot2', ('vec2', a
, 0.0), b
), ('fmul', a
, b
)),
195 (('fdot2', a
, 1.0), ('fadd', 'a.x', 'a.y')),
197 # Lower fdot to fsum when it is available
198 (('fdot2', a
, b
), ('fsum2', ('fmul', a
, b
)), 'options->lower_fdot'),
199 (('fdot3', a
, b
), ('fsum3', ('fmul', a
, b
)), 'options->lower_fdot'),
200 (('fdot4', a
, b
), ('fsum4', ('fmul', a
, b
)), 'options->lower_fdot'),
201 (('fsum2', a
), ('fadd', 'a.x', 'a.y'), 'options->lower_fdot'),
203 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
204 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
205 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
206 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
208 # 1 - ((1 - a) * (1 - b))
209 # 1 - (1 - a - b + a*b)
210 # 1 - 1 + a + b - a*b
215 (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a
)), ('fadd', 1.0, ('fneg', b
))))),
216 ('flrp', b
, 1.0, a
), '!options->lower_flrp32'),
218 # (a * #b + #c) << #d
219 # ((a * #b) << #d) + (#c << #d)
220 # (a * (#b << #d)) + (#c << #d)
221 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
222 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
226 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
228 # Comparison simplifications
229 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
230 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
231 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
232 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
233 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
234 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
235 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
236 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
237 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
238 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
240 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
241 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
242 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
243 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
244 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
246 # This helps some shaders because, after some optimizations, they end up
247 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
248 # matching would be handled by CSE.
249 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
250 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
251 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
252 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
253 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
254 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
255 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
256 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
257 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
258 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
260 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
261 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
262 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
263 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
264 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
265 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
267 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
268 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
269 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
270 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
274 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
276 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
278 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
280 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
281 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
282 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
283 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
284 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
285 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
286 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
287 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
288 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
289 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
290 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
291 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
292 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
293 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
294 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
295 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
296 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
297 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
298 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
299 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
301 # -(b2f(a) + b2f(b)) < 0
302 # 0 < b2f(a) + b2f(b)
303 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
305 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
306 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
308 # -(b2f(a) + b2f(b)) >= 0
309 # 0 >= b2f(a) + b2f(b)
310 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
312 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
313 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
315 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
316 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
318 # Some optimizations (below) convert things like (a < b || c < b) into
319 # (min(a, c) < b). However, this interfers with the previous optimizations
320 # that try to remove comparisons with negated sums of b2f. This just
322 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
323 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
325 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
326 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
327 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
328 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
330 # Cannot remove the addition from ilt or ige due to overflow.
331 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
332 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
334 # fmin(-b2f(a), b) >= 0.0
335 # -b2f(a) >= 0.0 && b >= 0.0
336 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
337 # b2f(a) == 0.0 && b >= 0.0
338 # a == False && b >= 0.0
341 # The fge in the second replacement is not a typo. I leave the proof that
342 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
344 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
345 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
347 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
348 (('fne', ('b2f', 'a@1'), 0.0), a
),
349 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
350 (('ine', ('b2i', 'a@1'), 0), a
),
352 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
353 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
354 (('fge', ('u2f', a
), 0.0), True),
355 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
356 (('flt', ('u2f', a
), 0.0), False),
357 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
358 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
359 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
360 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
361 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
362 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
363 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
367 # fabs(a) != 0.0 because fabs(a) must be >= 0
369 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
373 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
376 # 0.0 == fabs(a) because fabs(a) must be >= 0
378 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
382 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
384 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
385 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
386 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
387 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
390 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
391 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
392 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
394 # Since b is a constant, constant folding will eliminate the fmin and the
395 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
396 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
398 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
400 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
401 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
402 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
403 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
404 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
405 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
406 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
407 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
408 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
409 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
410 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
411 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
412 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
413 (('bcsel', a
, True, b
), ('ior', a
, b
)),
414 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
415 (('bcsel', a
, b
, False), ('iand', a
, b
)),
416 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
423 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
424 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
425 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
426 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
427 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
428 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
429 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
430 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
431 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
432 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
433 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
434 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
435 (('fmin', a
, ('fabs', a
)), a
),
436 (('imin', a
, ('iabs', a
)), a
),
437 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
438 (('imax', a
, ('ineg', ('iabs', a
))), a
),
439 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
440 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
441 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
442 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
443 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
444 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
445 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
446 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
447 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
448 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
449 (('fsat', ('b2f', a
)), ('b2f', a
)),
450 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
451 (('fsat', ('fsat', a
)), ('fsat', a
)),
452 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_negate && !options->lower_fsat'),
453 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_negate && !options->lower_fsat'),
454 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
455 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
456 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
457 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
458 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
459 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
460 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
461 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
462 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
463 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
464 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
465 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
466 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
467 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
468 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
469 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
470 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
471 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
472 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
473 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
474 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
475 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
476 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
478 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
479 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
480 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
481 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
482 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
483 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
484 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
485 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
486 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
487 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
488 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
489 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
490 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
491 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
492 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
493 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
495 # Common pattern like 'if (i == 0 || i == 1 || ...)'
496 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
497 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
498 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
500 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
501 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
502 # so emit an open-coded version of that.
503 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
504 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
506 (('ior', a
, ('ieq', a
, False)), True),
507 (('ior', a
, ('inot', a
)), -1),
509 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
510 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
512 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
514 # These patterns can result when (a < b || a < c) => (a < min(b, c))
515 # transformations occur before constant propagation and loop-unrolling.
516 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
517 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
518 (('~fge', a
, ('fmin', b
, a
)), True),
519 (('~fge', ('fmax', a
, b
), a
), True),
520 (('~flt', a
, ('fmin', b
, a
)), False),
521 (('~flt', ('fmax', a
, b
), a
), False),
522 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
523 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
525 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
526 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
527 (('ige', a
, ('imin', b
, a
)), True),
528 (('ige', ('imax', a
, b
), a
), True),
529 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
530 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
531 (('uge', a
, ('umin', b
, a
)), True),
532 (('uge', ('umax', a
, b
), a
), True),
533 (('ilt', a
, ('imin', b
, a
)), False),
534 (('ilt', ('imax', a
, b
), a
), False),
535 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
536 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
537 (('ult', a
, ('umin', b
, a
)), False),
538 (('ult', ('umax', a
, b
), a
), False),
539 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
540 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
541 (('ult', a
, ('iand', b
, a
)), False),
542 (('ult', ('ior', a
, b
), a
), False),
543 (('uge', a
, ('iand', b
, a
)), True),
544 (('uge', ('ior', a
, b
), a
), True),
546 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
547 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
548 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
549 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
550 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
551 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
552 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
553 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
554 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
555 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
556 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
557 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
558 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
559 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
560 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
561 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
563 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
565 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
566 ('iabs', ('ishr', a
, b
))),
567 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
569 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
570 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
571 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
572 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
573 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
574 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
575 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
576 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
577 (('seq', ('seq', a
, b
), 1.0), ('seq', a
, b
)),
578 (('seq', ('sne', a
, b
), 1.0), ('sne', a
, b
)),
579 (('seq', ('slt', a
, b
), 1.0), ('slt', a
, b
)),
580 (('seq', ('sge', a
, b
), 1.0), ('sge', a
, b
)),
581 (('sne', ('seq', a
, b
), 0.0), ('seq', a
, b
)),
582 (('sne', ('sne', a
, b
), 0.0), ('sne', a
, b
)),
583 (('sne', ('slt', a
, b
), 0.0), ('slt', a
, b
)),
584 (('sne', ('sge', a
, b
), 0.0), ('sge', a
, b
)),
585 (('seq', ('seq', a
, b
), 0.0), ('sne', a
, b
)),
586 (('seq', ('sne', a
, b
), 0.0), ('seq', a
, b
)),
587 (('seq', ('slt', a
, b
), 0.0), ('sge', a
, b
)),
588 (('seq', ('sge', a
, b
), 0.0), ('slt', a
, b
)),
589 (('sne', ('seq', a
, b
), 1.0), ('sne', a
, b
)),
590 (('sne', ('sne', a
, b
), 1.0), ('seq', a
, b
)),
591 (('sne', ('slt', a
, b
), 1.0), ('sge', a
, b
)),
592 (('sne', ('sge', a
, b
), 1.0), ('slt', a
, b
)),
593 (('fall_equal2', a
, b
), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
594 (('fall_equal3', a
, b
), ('seq', ('fany_nequal3', a
, b
), 0.0), 'options->lower_vector_cmp'),
595 (('fall_equal4', a
, b
), ('seq', ('fany_nequal4', a
, b
), 0.0), 'options->lower_vector_cmp'),
596 (('fany_nequal2', a
, b
), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
597 (('fany_nequal3', a
, b
), ('fsat', ('fdot3', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
598 (('fany_nequal4', a
, b
), ('fsat', ('fdot4', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
599 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
600 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
602 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
603 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
604 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
605 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
606 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
607 (('ineg', ('b2i32', 'a@32')), a
),
608 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
609 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
610 # Comparison with the same args. Note that these are not done for
611 # the float versions because NaN always returns false on float
613 (('ilt', a
, a
), False),
614 (('ige', a
, a
), True),
615 (('ieq', a
, a
), True),
616 (('ine', a
, a
), False),
617 (('ult', a
, a
), False),
618 (('uge', a
, a
), True),
619 # Logical and bit operations
621 (('iand', a
, ~
0), a
),
625 (('ior', a
, True), True),
628 (('inot', ('inot', a
)), a
),
629 (('ior', ('iand', a
, b
), b
), b
),
630 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
631 (('iand', ('ior', a
, b
), b
), b
),
632 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
634 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
635 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
636 # Shift optimizations
643 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
644 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
645 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
646 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
647 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
648 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
649 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
650 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
651 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
652 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
653 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
654 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
655 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
656 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
657 # Exponential/logarithmic identities
658 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
659 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
660 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
661 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
662 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
663 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
664 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
665 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
666 (('~fpow', a
, 1.0), a
),
667 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
668 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
669 (('~fpow', 2.0, a
), ('fexp2', a
)),
670 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
671 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
672 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
673 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
674 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
675 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
676 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
677 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
678 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
679 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
680 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
681 # Division and reciprocal
682 (('~fdiv', 1.0, a
), ('frcp', a
)),
683 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
684 (('~frcp', ('frcp', a
)), a
),
685 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
686 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
687 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
689 (('fsin', a
), lowered_sincos(0.5), 'options->lower_sincos'),
690 (('fcos', a
), lowered_sincos(0.75), 'options->lower_sincos'),
691 # Boolean simplifications
692 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
693 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
694 (('ieq', a
, True), a
),
695 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
696 (('ine', a
, False), a
),
697 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
698 (('bcsel', a
, True, False), a
),
699 (('bcsel', a
, False, True), ('inot', a
)),
700 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
701 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
702 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
703 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
704 (('bcsel', True, b
, c
), b
),
705 (('bcsel', False, b
, c
), c
),
706 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
708 (('bcsel', a
, b
, b
), b
),
709 (('fcsel', a
, b
, b
), b
),
711 # D3D Boolean emulation
712 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
713 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
714 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
715 ('ineg', ('b2i', ('iand', a
, b
)))),
716 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
717 ('ineg', ('b2i', ('ior', a
, b
)))),
718 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
719 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
720 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
721 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
722 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
724 # SM5 32-bit shifts are defined to use the 5 least significant bits
725 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
726 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
727 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
730 (('i2b32', ('b2i', 'a@32')), a
),
731 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
732 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
733 (('i2b', ('ineg', a
)), ('i2b', a
)),
734 (('i2b', ('iabs', a
)), ('i2b', a
)),
735 (('fabs', ('b2f', a
)), ('b2f', a
)),
736 (('iabs', ('b2i', a
)), ('b2i', a
)),
737 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
739 # Ironically, mark these as imprecise because removing the conversions may
740 # preserve more precision than doing the conversions (e.g.,
741 # uint(float(0x81818181u)) == 0x81818200).
742 (('~f2i32', ('i2f', 'a@32')), a
),
743 (('~f2i32', ('u2f', 'a@32')), a
),
744 (('~f2u32', ('i2f', 'a@32')), a
),
745 (('~f2u32', ('u2f', 'a@32')), a
),
747 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
750 # It is undefined to convert a negative floating-point value to an
753 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
754 # some optimizations in the i965 backend to proceed.
755 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
756 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
757 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
758 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
760 (('~fmin', ('fabs', a
), 1.0), ('fsat', ('fabs', a
)), '!options->lower_fsat'),
762 # The result of the multiply must be in [-1, 0], so the result of the ffma
764 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
765 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
766 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
767 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
769 # Packing and then unpacking does nothing
770 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
771 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
772 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
773 ('unpack_64_2x32_split_y', a
)), a
),
775 # Comparing two halves of an unpack separately. While this optimization
776 # should be correct for non-constant values, it's less obvious that it's
777 # useful in that case. For constant values, the pack will fold and we're
778 # guaranteed to reduce the whole tree to one instruction.
779 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
780 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
781 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
784 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
785 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
786 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
787 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
788 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
789 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
790 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
793 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
794 # patterns like those below.
795 for op
in ('ushr', 'ishr'):
796 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
797 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
798 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
800 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
802 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
803 # patterns like those below.
804 for op
in ('extract_u8', 'extract_i8'):
805 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
806 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
807 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
809 optimizations
.extend([
811 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
812 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
813 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
814 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
815 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
818 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
819 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
820 (('ussub_4x8', a
, 0), a
),
821 (('ussub_4x8', a
, ~
0), 0),
822 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
823 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
824 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
825 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
826 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
827 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
828 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
829 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
831 # Propagate negation up multiplication chains
832 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
833 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
835 # Propagate constants up multiplication chains
836 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
837 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
838 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
839 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
841 # Reassociate constants in add/mul chains so they can be folded together.
842 # For now, we mostly only handle cases where the constants are separated by
843 # a single non-constant. We could do better eventually.
844 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
845 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
846 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
847 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
848 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
850 # Drop mul-div by the same value when there's no wrapping.
851 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
854 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
855 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
856 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
858 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
859 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
860 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
862 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
865 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
866 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
867 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
868 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
869 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
871 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
872 ('bcsel', ('ult', 31, 'bits'), 'insert',
873 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
874 'options->lower_bitfield_insert'),
875 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
876 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
877 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
878 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
879 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
880 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
882 # Alternative lowering that doesn't rely on bfi.
883 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
884 ('bcsel', ('ult', 31, 'bits'),
887 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
888 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
889 'options->lower_bitfield_insert_to_shifts'),
891 # Alternative lowering that uses bitfield_select.
892 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
893 ('bcsel', ('ult', 31, 'bits'), 'insert',
894 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
895 'options->lower_bitfield_insert_to_bitfield_select'),
897 (('ibitfield_extract', 'value', 'offset', 'bits'),
898 ('bcsel', ('ult', 31, 'bits'), 'value',
899 ('ibfe', 'value', 'offset', 'bits')),
900 'options->lower_bitfield_extract'),
902 (('ubitfield_extract', 'value', 'offset', 'bits'),
903 ('bcsel', ('ult', 31, 'bits'), 'value',
904 ('ubfe', 'value', 'offset', 'bits')),
905 'options->lower_bitfield_extract'),
907 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
908 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
909 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
910 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
911 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
912 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
913 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
915 (('ibitfield_extract', 'value', 'offset', 'bits'),
916 ('bcsel', ('ieq', 0, 'bits'),
919 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
920 ('isub', 32, 'bits'))),
921 'options->lower_bitfield_extract_to_shifts'),
923 (('ubitfield_extract', 'value', 'offset', 'bits'),
925 ('ushr', 'value', 'offset'),
926 ('bcsel', ('ieq', 'bits', 32),
928 ('isub', ('ishl', 1, 'bits'), 1))),
929 'options->lower_bitfield_extract_to_shifts'),
931 (('ifind_msb', 'value'),
932 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
933 'options->lower_ifind_msb'),
935 (('find_lsb', 'value'),
936 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
937 'options->lower_find_lsb'),
939 (('extract_i8', a
, 'b@32'),
940 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
941 'options->lower_extract_byte'),
943 (('extract_u8', a
, 'b@32'),
944 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
945 'options->lower_extract_byte'),
947 (('extract_i16', a
, 'b@32'),
948 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
949 'options->lower_extract_word'),
951 (('extract_u16', a
, 'b@32'),
952 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
953 'options->lower_extract_word'),
955 (('pack_unorm_2x16', 'v'),
956 ('pack_uvec2_to_uint',
957 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
958 'options->lower_pack_unorm_2x16'),
960 (('pack_unorm_4x8', 'v'),
961 ('pack_uvec4_to_uint',
962 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
963 'options->lower_pack_unorm_4x8'),
965 (('pack_snorm_2x16', 'v'),
966 ('pack_uvec2_to_uint',
967 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
968 'options->lower_pack_snorm_2x16'),
970 (('pack_snorm_4x8', 'v'),
971 ('pack_uvec4_to_uint',
972 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
973 'options->lower_pack_snorm_4x8'),
975 (('unpack_unorm_2x16', 'v'),
976 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
977 ('extract_u16', 'v', 1))),
979 'options->lower_unpack_unorm_2x16'),
981 (('unpack_unorm_4x8', 'v'),
982 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
983 ('extract_u8', 'v', 1),
984 ('extract_u8', 'v', 2),
985 ('extract_u8', 'v', 3))),
987 'options->lower_unpack_unorm_4x8'),
989 (('unpack_snorm_2x16', 'v'),
990 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
991 ('extract_i16', 'v', 1))),
993 'options->lower_unpack_snorm_2x16'),
995 (('unpack_snorm_4x8', 'v'),
996 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
997 ('extract_i8', 'v', 1),
998 ('extract_i8', 'v', 2),
999 ('extract_i8', 'v', 3))),
1001 'options->lower_unpack_snorm_4x8'),
1003 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
1004 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
1007 # bit_size dependent lowerings
1008 for bit_size
in [8, 16, 32, 64]:
1009 # convenience constants
1010 intmax
= (1 << (bit_size
- 1)) - 1
1011 intmin
= 1 << (bit_size
- 1)
1014 (('iadd_sat@' + str(bit_size
), a
, b
),
1015 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
1016 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
1017 (('isub_sat@' + str(bit_size
), a
, b
),
1018 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
1019 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
1022 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
1024 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
1025 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1026 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1027 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1028 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1030 # Optimize x2bN(b2x(x)) -> x
1031 for size
in type_sizes('bool'):
1032 aN
= 'a@' + str(size
)
1033 f2bN
= 'f2b' + str(size
)
1034 i2bN
= 'i2b' + str(size
)
1035 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
1036 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
1038 # Optimize x2yN(b2x(x)) -> b2y
1039 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1040 if x
!= 'f' and y
!= 'f' and x
!= y
:
1043 b2x
= 'b2f' if x
== 'f' else 'b2i'
1044 b2y
= 'b2f' if y
== 'f' else 'b2i'
1045 x2yN
= '{}2{}'.format(x
, y
)
1046 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1048 # Optimize away x2xN(a@N)
1049 for t
in ['int', 'uint', 'float']:
1050 for N
in type_sizes(t
):
1051 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1052 aN
= 'a@{0}'.format(N
)
1053 optimizations
.append(((x2xN
, aN
), a
))
1055 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1056 # In particular, we can optimize away everything except upcast of downcast and
1057 # upcasts where the type differs from the other cast
1058 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1060 # The outer cast is a down-cast. It doesn't matter what the size of the
1061 # argument of the inner cast is because we'll never been in the upcast
1062 # of downcast case. Regardless of types, we'll always end up with y2yN
1064 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1065 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1066 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1067 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1068 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1070 # If the outer cast is an up-cast, we have to be more careful about the
1071 # size of the argument of the inner cast and with types. In this case,
1072 # the type is always the type of type up-cast which is given by the
1074 for P
in type_sizes('uint'):
1075 # We can't optimize away up-cast of down-cast.
1079 # Because we're doing down-cast of down-cast, the types always have
1080 # to match between the two casts
1081 for x
in ['i', 'u']:
1082 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1083 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1084 aP
= 'a@{0}'.format(P
)
1085 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1087 # The N == M case is handled by other optimizations
1090 # Optimize comparisons with up-casts
1091 for t
in ['int', 'uint', 'float']:
1092 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1093 if N
== 1 or N
>= M
:
1096 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1097 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1100 xeq
= 'feq' if t
== 'float' else 'ieq'
1101 xne
= 'fne' if t
== 'float' else 'ine'
1102 xge
= '{0}ge'.format(t
[0])
1103 xlt
= '{0}lt'.format(t
[0])
1105 # Up-casts are lossless so for correctly signed comparisons of
1106 # up-casted values we can do the comparison at the largest of the two
1107 # original sizes and drop one or both of the casts. (We have
1108 # optimizations to drop the no-op casts which this may generate.)
1109 for P
in type_sizes(t
):
1115 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1116 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1117 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1118 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1119 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1120 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1123 # The next bit doesn't work on floats because the range checks would
1124 # get way too complicated.
1125 if t
in ['int', 'uint']:
1127 xN_min
= -(1 << (N
- 1))
1128 xN_max
= (1 << (N
- 1)) - 1
1131 xN_max
= (1 << N
) - 1
1135 # If we're up-casting and comparing to a constant, we can unfold
1136 # the comparison into a comparison with the shrunk down constant
1137 # and a check that the constant fits in the smaller bit size.
1139 ((xeq
, (x2xM
, aN
), '#b'),
1140 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1141 ((xne
, (x2xM
, aN
), '#b'),
1142 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1143 ((xlt
, (x2xM
, aN
), '#b'),
1144 ('iand', (xlt
, xN_min
, b
),
1145 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1146 ((xlt
, '#a', (x2xM
, bN
)),
1147 ('iand', (xlt
, a
, xN_max
),
1148 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1149 ((xge
, (x2xM
, aN
), '#b'),
1150 ('iand', (xge
, xN_max
, b
),
1151 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1152 ((xge
, '#a', (x2xM
, bN
)),
1153 ('iand', (xge
, a
, xN_min
),
1154 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1157 def fexp2i(exp
, bits
):
1158 # We assume that exp is already in the right range.
1160 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1162 return ('ishl', ('iadd', exp
, 127), 23)
1164 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1168 def ldexp(f
, exp
, bits
):
1169 # First, we clamp exp to a reasonable range. The maximum possible range
1170 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1171 # a maximum range of [-149, 127]. This means that we can potentially have
1172 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1173 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1174 # spec, on the other hand, only requires that we handle an exponent value
1175 # in the range [-126, 128]. This implementation is *mostly* correct; it
1176 # handles a range on exp of [-252, 254] which allows you to create any
1177 # value (including denorms if the hardware supports it) and to adjust the
1178 # exponent of any normal value to anything you want.
1180 exp
= ('imin', ('imax', exp
, -28), 30)
1182 exp
= ('imin', ('imax', exp
, -252), 254)
1184 exp
= ('imin', ('imax', exp
, -2044), 2046)
1188 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1189 # (We use ishr which isn't the same for -1, but the -1 case still works
1190 # since we use exp-exp/2 as the second exponent.) While the spec
1191 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1192 # work with denormals and doesn't allow for the full swing in exponents
1193 # that you can get with normalized values. Instead, we create two powers
1194 # of two and multiply by them each in turn. That way the effective range
1195 # of our exponent is doubled.
1196 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1197 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1198 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1201 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1202 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1203 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1206 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1207 def bitfield_reverse(u
):
1208 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1209 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1210 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1211 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1212 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1216 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1218 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1219 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1220 # and, if a is a NaN then the second comparison will fail anyway.
1221 for op
in ['flt', 'fge', 'feq']:
1223 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1224 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1227 # Add optimizations to handle the case where the result of a ternary is
1228 # compared to a constant. This way we can take things like
1234 # a ? (0 > 0) : (1 > 0)
1236 # which constant folding will eat for lunch. The resulting ternary will
1237 # further get cleaned up by the boolean reductions above and we will be
1238 # left with just the original variable "a".
1239 for op
in ['flt', 'fge', 'feq', 'fne',
1240 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1242 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1243 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1244 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1245 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1249 # For example, this converts things like
1251 # 1 + mix(0, a - 1, condition)
1255 # mix(1, (a-1)+1, condition)
1257 # Other optimizations will rearrange the constants.
1258 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1260 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1263 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1266 # If neither layout qualifier is specified, derivatives in compute shaders
1267 # return zero, which is consistent with the handling of built-in texture
1268 # functions like texture() in GLSL 4.50 compute shaders.
1269 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1270 'fddy', 'fddy_fine', 'fddy_coarse']:
1272 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1275 # Some optimizations for ir3-specific instructions.
1277 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1278 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1279 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1280 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1281 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1284 # This section contains "late" optimizations that should be run before
1285 # creating ffmas and calling regular optimizations for the final time.
1286 # Optimizations should go here if they help code generation and conflict
1287 # with the regular optimizations.
1288 before_ffma_optimizations
= [
1289 # Propagate constants down multiplication chains
1290 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1291 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1292 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1293 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1295 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1296 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1297 (('~fadd', ('fneg', a
), a
), 0.0),
1298 (('iadd', ('ineg', a
), a
), 0),
1299 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1300 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1301 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1302 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1304 (('~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
)),
1305 (('~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
)),
1306 (('~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
))),
1309 # This section contains "late" optimizations that should be run after the
1310 # regular optimizations have finished. Optimizations should go here if
1311 # they help code generation but do not necessarily produce code that is
1312 # more easily optimizable.
1313 late_optimizations
= [
1314 # Most of these optimizations aren't quite safe when you get infinity or
1315 # Nan involved but the first one should be fine.
1316 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1317 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1318 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1319 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1320 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1321 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1323 # nir_lower_to_source_mods will collapse this, but its existence during the
1324 # optimization loop can prevent other optimizations.
1325 (('fneg', ('fneg', a
)), a
),
1327 # These are duplicated from the main optimizations table. The late
1328 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1329 # new patterns like these. The patterns that compare with zero are removed
1330 # because they are unlikely to be created in by anything in
1331 # late_optimizations.
1332 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1333 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1334 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1335 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1336 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1337 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1339 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1340 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1342 (('~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
)))),
1344 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1345 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1346 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1347 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1348 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1349 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1350 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1351 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1352 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1353 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1356 (('iand', a
, a
), a
),
1358 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1360 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1361 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1362 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1363 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1365 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1366 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1368 (('~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'),
1369 (('~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'),
1371 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1372 # particular operation is common for expanding values stored in a texture
1373 # from [0,1] to [-1,1].
1374 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1375 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1376 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1377 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1378 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1379 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1380 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1381 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1385 # a + -a*a + a*b (1)
1387 # Option 1: ffma(a, (b-a), a)
1389 # Alternately, after (1):
1395 # Option 2: ffma(a, 2, -(a*a))
1396 # Option 3: ffma(a, 2, (-a)*a)
1397 # Option 4: ffma(a, -a, (2*a)
1398 # Option 5: a * (2 - a)
1400 # There are a lot of other possible combinations.
1401 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1402 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1403 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1404 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1405 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1407 # we do these late so that we don't get in the way of creating ffmas
1408 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1409 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1411 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1413 # Things that look like DPH in the source shader may get expanded to
1414 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1415 # to NIR. After FFMA is generated, this can look like:
1417 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1419 # Reassociate the last addition into the first multiplication.
1420 (('~fadd', ('ffma(is_used_once)', a
, b
, ('ffma', c
, d
, ('fmul', 'e(is_not_const_and_not_fsign)', 'f(is_not_const_and_not_fsign)'))), 'g(is_not_const)'),
1421 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '!options->intel_vec4'),
1422 (('~fadd', ('ffma(is_used_once)', a
, b
, ('fmul', 'e(is_not_const_and_not_fsign)', 'f(is_not_const_and_not_fsign)') ), 'g(is_not_const)'),
1423 ('ffma', a
, b
, ('ffma', e
, 'f', 'g') ), '!options->intel_vec4'),
1426 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1427 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1428 before_ffma_optimizations
).render())
1429 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1430 late_optimizations
).render())