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_bitops'),
87 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
88 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitops'),
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_bitops'),
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 # This is common for address calculations. Reassociating may enable the
229 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
230 # instruction or a constant offset field for in load / store instructions.
231 (('ishl', ('iadd', a
, '#b'), '#c'), ('iadd', ('ishl', a
, c
), ('ishl', b
, c
))),
233 # Comparison simplifications
234 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
235 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
236 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
237 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
238 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
239 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
240 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
241 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
242 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
243 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
245 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
246 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
247 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
248 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
249 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
251 # This helps some shaders because, after some optimizations, they end up
252 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
253 # matching would be handled by CSE.
254 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
255 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
256 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
257 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
258 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
259 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
260 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
261 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
262 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
263 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
265 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
266 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
267 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
268 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
269 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
270 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
272 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
273 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
274 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
275 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
279 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
281 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
283 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
285 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
286 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
287 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
288 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
289 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
290 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
291 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
292 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
293 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
294 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
295 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
296 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
297 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
298 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
299 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
300 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
301 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
302 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
303 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
304 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
306 # -(b2f(a) + b2f(b)) < 0
307 # 0 < b2f(a) + b2f(b)
308 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
310 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
311 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
313 # -(b2f(a) + b2f(b)) >= 0
314 # 0 >= b2f(a) + b2f(b)
315 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
317 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
318 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
320 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
321 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
323 # Some optimizations (below) convert things like (a < b || c < b) into
324 # (min(a, c) < b). However, this interfers with the previous optimizations
325 # that try to remove comparisons with negated sums of b2f. This just
327 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
328 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
330 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
331 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
332 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
333 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
334 (('~flt', ('fadd(is_used_once)', a
, '#b'), '#c'), ('flt', a
, ('fadd', c
, ('fneg', b
)))),
335 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('flt', ('fneg', ('fadd', c
, b
)), a
)),
336 (('~fge', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fge', a
, ('fadd', c
, ('fneg', b
)))),
337 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fge', ('fneg', ('fadd', c
, b
)), a
)),
338 (('~feq', ('fadd(is_used_once)', a
, '#b'), '#c'), ('feq', a
, ('fadd', c
, ('fneg', b
)))),
339 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('feq', ('fneg', ('fadd', c
, b
)), a
)),
340 (('~fne', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fne', a
, ('fadd', c
, ('fneg', b
)))),
341 (('~fne', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fne', ('fneg', ('fadd', c
, b
)), a
)),
343 # Cannot remove the addition from ilt or ige due to overflow.
344 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
345 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
347 # fmin(-b2f(a), b) >= 0.0
348 # -b2f(a) >= 0.0 && b >= 0.0
349 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
350 # b2f(a) == 0.0 && b >= 0.0
351 # a == False && b >= 0.0
354 # The fge in the second replacement is not a typo. I leave the proof that
355 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
357 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
358 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
360 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
361 (('fne', ('b2f', 'a@1'), 0.0), a
),
362 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
363 (('ine', ('b2i', 'a@1'), 0), a
),
365 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
366 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
367 (('fge', ('u2f', a
), 0.0), True),
368 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
369 (('flt', ('u2f', a
), 0.0), False),
370 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
371 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
372 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
373 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
374 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
375 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
376 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
380 # fabs(a) != 0.0 because fabs(a) must be >= 0
382 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
386 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
389 # 0.0 == fabs(a) because fabs(a) must be >= 0
391 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
395 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
397 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
398 (('iand', ('fge', a
, 0.0), ('fge', 1.0, a
)), ('feq', a
, ('fsat', a
)), '!options->lower_fsat'),
400 # (a < 0.0) || (a > 1.0)
401 # !(!(a < 0.0) && !(a > 1.0))
402 # !((a >= 0.0) && (a <= 1.0))
405 (('ior', ('flt', a
, 0.0), ('flt', 1.0, a
)), ('fne', a
, ('fsat', a
)), '!options->lower_fsat'),
407 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
408 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
409 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
410 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
413 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
414 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
415 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
417 # Since b is a constant, constant folding will eliminate the fmin and the
418 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
419 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
421 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
423 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
424 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
425 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
426 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
427 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
428 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
429 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
430 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
431 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
432 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
433 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
434 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
435 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
436 (('bcsel', a
, True, b
), ('ior', a
, b
)),
437 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
438 (('bcsel', a
, b
, False), ('iand', a
, b
)),
439 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
446 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
447 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
448 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
449 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
450 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
451 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
452 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
453 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
454 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
455 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
456 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
457 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
458 (('fmin', a
, ('fabs', a
)), a
),
459 (('imin', a
, ('iabs', a
)), a
),
460 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
461 (('imax', a
, ('ineg', ('iabs', a
))), a
),
462 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
463 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
464 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
465 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
466 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
467 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
468 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
469 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
470 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
471 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
472 (('fsat', ('b2f', a
)), ('b2f', a
)),
473 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
474 (('fsat', ('fsat', a
)), ('fsat', a
)),
475 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_negate && !options->lower_fsat'),
476 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_negate && !options->lower_fsat'),
477 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
478 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
479 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
480 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
481 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
482 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
483 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
484 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
485 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
486 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
487 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
488 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
489 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
490 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
491 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
492 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
493 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
494 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
495 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
496 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
497 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
498 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
499 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
501 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
502 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
503 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
504 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
505 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
506 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
507 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
508 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
509 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
510 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
511 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
512 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
513 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
514 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
515 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
516 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
518 # These derive from the previous patterns with the application of b < 0 <=>
519 # 0 < -b. The transformation should be applied if either comparison is
520 # used once as this ensures that the number of comparisons will not
521 # increase. The sources to the ior and iand are not symmetric, so the
522 # rules have to be duplicated to get this behavior.
523 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
524 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
525 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
526 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
527 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
528 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
529 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
530 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
532 # Common pattern like 'if (i == 0 || i == 1 || ...)'
533 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
534 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
535 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
537 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
538 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
539 # so emit an open-coded version of that.
540 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
541 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
543 (('ior', a
, ('ieq', a
, False)), True),
544 (('ior', a
, ('inot', a
)), -1),
546 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
547 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
549 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0), '!options->lower_bitops'),
551 # These patterns can result when (a < b || a < c) => (a < min(b, c))
552 # transformations occur before constant propagation and loop-unrolling.
553 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
554 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
555 (('~fge', a
, ('fmin', b
, a
)), True),
556 (('~fge', ('fmax', a
, b
), a
), True),
557 (('~flt', a
, ('fmin', b
, a
)), False),
558 (('~flt', ('fmax', a
, b
), a
), False),
559 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
560 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
562 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
563 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
564 (('ige', a
, ('imin', b
, a
)), True),
565 (('ige', ('imax', a
, b
), a
), True),
566 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
567 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
568 (('uge', a
, ('umin', b
, a
)), True),
569 (('uge', ('umax', a
, b
), a
), True),
570 (('ilt', a
, ('imin', b
, a
)), False),
571 (('ilt', ('imax', a
, b
), a
), False),
572 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
573 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
574 (('ult', a
, ('umin', b
, a
)), False),
575 (('ult', ('umax', a
, b
), a
), False),
576 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
577 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
578 (('ult', a
, ('iand', b
, a
)), False),
579 (('ult', ('ior', a
, b
), a
), False),
580 (('uge', a
, ('iand', b
, a
)), True),
581 (('uge', ('ior', a
, b
), a
), True),
583 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
584 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
585 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
586 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
587 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
588 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
589 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
590 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
591 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
592 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
593 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
594 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
595 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
596 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
597 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
598 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
600 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
602 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
603 ('iabs', ('ishr', a
, b
))),
604 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
606 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
607 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
608 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
609 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
610 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
611 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
612 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
613 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
614 (('seq', ('seq', a
, b
), 1.0), ('seq', a
, b
)),
615 (('seq', ('sne', a
, b
), 1.0), ('sne', a
, b
)),
616 (('seq', ('slt', a
, b
), 1.0), ('slt', a
, b
)),
617 (('seq', ('sge', a
, b
), 1.0), ('sge', a
, b
)),
618 (('sne', ('seq', a
, b
), 0.0), ('seq', a
, b
)),
619 (('sne', ('sne', a
, b
), 0.0), ('sne', a
, b
)),
620 (('sne', ('slt', a
, b
), 0.0), ('slt', a
, b
)),
621 (('sne', ('sge', a
, b
), 0.0), ('sge', a
, b
)),
622 (('seq', ('seq', a
, b
), 0.0), ('sne', a
, b
)),
623 (('seq', ('sne', a
, b
), 0.0), ('seq', a
, b
)),
624 (('seq', ('slt', a
, b
), 0.0), ('sge', a
, b
)),
625 (('seq', ('sge', a
, b
), 0.0), ('slt', a
, b
)),
626 (('sne', ('seq', a
, b
), 1.0), ('sne', a
, b
)),
627 (('sne', ('sne', a
, b
), 1.0), ('seq', a
, b
)),
628 (('sne', ('slt', a
, b
), 1.0), ('sge', a
, b
)),
629 (('sne', ('sge', a
, b
), 1.0), ('slt', a
, b
)),
630 (('fall_equal2', a
, b
), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
631 (('fall_equal3', a
, b
), ('seq', ('fany_nequal3', a
, b
), 0.0), 'options->lower_vector_cmp'),
632 (('fall_equal4', a
, b
), ('seq', ('fany_nequal4', a
, b
), 0.0), 'options->lower_vector_cmp'),
633 (('fany_nequal2', a
, b
), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
634 (('fany_nequal3', a
, b
), ('fsat', ('fdot3', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
635 (('fany_nequal4', a
, b
), ('fsat', ('fdot4', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
636 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
637 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
639 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
640 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
641 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
642 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
643 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
644 (('ineg', ('b2i32', 'a@32')), a
),
645 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
646 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
647 # Comparison with the same args. Note that these are not done for
648 # the float versions because NaN always returns false on float
650 (('ilt', a
, a
), False),
651 (('ige', a
, a
), True),
652 (('ieq', a
, a
), True),
653 (('ine', a
, a
), False),
654 (('ult', a
, a
), False),
655 (('uge', a
, a
), True),
656 # Logical and bit operations
658 (('iand', a
, ~
0), a
),
662 (('ior', a
, True), True),
665 (('inot', ('inot', a
)), a
),
666 (('ior', ('iand', a
, b
), b
), b
),
667 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
668 (('iand', ('ior', a
, b
), b
), b
),
669 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
671 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
672 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
673 # Shift optimizations
680 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
681 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
682 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
683 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
684 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
685 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
686 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
687 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
688 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
689 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
690 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
691 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
692 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
693 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
694 # Exponential/logarithmic identities
695 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
696 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
697 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
698 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
699 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
700 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
701 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
702 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
703 (('~fpow', a
, 1.0), a
),
704 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
705 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
706 (('~fpow', 2.0, a
), ('fexp2', a
)),
707 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
708 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
709 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
710 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
711 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
712 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
713 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
714 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
715 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
716 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
717 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
718 # Division and reciprocal
719 (('~fdiv', 1.0, a
), ('frcp', a
)),
720 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
721 (('~frcp', ('frcp', a
)), a
),
722 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
723 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
724 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
726 (('fsin', a
), lowered_sincos(0.5), 'options->lower_sincos'),
727 (('fcos', a
), lowered_sincos(0.75), 'options->lower_sincos'),
728 # Boolean simplifications
729 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
730 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
731 (('ieq', a
, True), a
),
732 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
733 (('ine', a
, False), a
),
734 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
735 (('bcsel', a
, True, False), a
),
736 (('bcsel', a
, False, True), ('inot', a
)),
737 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
738 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
739 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
740 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
741 (('bcsel', True, b
, c
), b
),
742 (('bcsel', False, b
, c
), c
),
743 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
745 (('bcsel', a
, b
, b
), b
),
746 (('fcsel', a
, b
, b
), b
),
748 # D3D Boolean emulation
749 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
750 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
751 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
752 ('ineg', ('b2i', ('iand', a
, b
)))),
753 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
754 ('ineg', ('b2i', ('ior', a
, b
)))),
755 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
756 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
757 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
758 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
759 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
761 # SM5 32-bit shifts are defined to use the 5 least significant bits
762 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
763 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
764 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
767 (('i2b32', ('b2i', 'a@32')), a
),
768 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
769 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
770 (('i2b', ('ineg', a
)), ('i2b', a
)),
771 (('i2b', ('iabs', a
)), ('i2b', a
)),
772 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
774 # Ironically, mark these as imprecise because removing the conversions may
775 # preserve more precision than doing the conversions (e.g.,
776 # uint(float(0x81818181u)) == 0x81818200).
777 (('~f2i32', ('i2f', 'a@32')), a
),
778 (('~f2i32', ('u2f', 'a@32')), a
),
779 (('~f2u32', ('i2f', 'a@32')), a
),
780 (('~f2u32', ('u2f', 'a@32')), a
),
782 (('ffloor', 'a(is_integral)'), a
),
783 (('fceil', 'a(is_integral)'), a
),
784 (('ftrunc', 'a(is_integral)'), a
),
785 (('ffract', 'a(is_integral)'), 0.0),
786 (('fabs', 'a(is_not_negative)'), a
),
787 (('iabs', 'a(is_not_negative)'), a
),
788 (('fsat', 'a(is_not_positive)'), 0.0),
790 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
793 # It is undefined to convert a negative floating-point value to an
796 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
797 # some optimizations in the i965 backend to proceed.
798 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
799 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
800 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
801 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
803 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a
), '!options->lower_fsat'),
805 # The result of the multiply must be in [-1, 0], so the result of the ffma
807 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
808 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
809 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
810 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
812 (('fne', 'a(is_not_zero)', 0.0), True),
813 (('feq', 'a(is_not_zero)', 0.0), False),
815 (('fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
816 (('fge', 'b(is_not_positive)', 'a(is_gt_zero)'), False),
817 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
818 (('fge', 'b(is_not_negative)', 'a(is_not_positive)'), True),
820 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
821 (('flt', 'b(is_not_positive)', 'a(is_gt_zero)'), True),
822 (('flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
823 (('flt', 'b(is_not_negative)', 'a(is_not_positive)'), False),
825 (('ine', 'a(is_not_zero)', 0), True),
826 (('ieq', 'a(is_not_zero)', 0), False),
828 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
829 (('ige', 'b(is_not_positive)', 'a(is_gt_zero)'), False),
830 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
831 (('ige', 'b(is_not_negative)', 'a(is_not_positive)'), True),
833 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
834 (('ilt', 'b(is_not_positive)', 'a(is_gt_zero)'), True),
835 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
836 (('ilt', 'b(is_not_negative)', 'a(is_not_positive)'), False),
838 (('ult', 0, 'a(is_gt_zero)'), True),
840 # Packing and then unpacking does nothing
841 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
842 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
843 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
844 ('unpack_64_2x32_split_y', a
)), a
),
846 # Comparing two halves of an unpack separately. While this optimization
847 # should be correct for non-constant values, it's less obvious that it's
848 # useful in that case. For constant values, the pack will fold and we're
849 # guaranteed to reduce the whole tree to one instruction.
850 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
851 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
852 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
855 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
856 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
857 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
858 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
859 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
860 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
861 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
864 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
865 # patterns like those below.
866 for op
in ('ushr', 'ishr'):
867 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
868 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
869 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
871 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
873 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
874 # patterns like those below.
875 for op
in ('extract_u8', 'extract_i8'):
876 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
877 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
878 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
880 optimizations
.extend([
882 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
883 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
884 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
885 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
886 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
889 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
890 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
891 (('ussub_4x8', a
, 0), a
),
892 (('ussub_4x8', a
, ~
0), 0),
893 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
894 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
895 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
896 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
897 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
898 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
899 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
900 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
902 # Propagate negation up multiplication chains
903 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
904 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
906 # Propagate constants up multiplication chains
907 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
908 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
909 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
910 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
912 # Reassociate constants in add/mul chains so they can be folded together.
913 # For now, we mostly only handle cases where the constants are separated by
914 # a single non-constant. We could do better eventually.
915 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
916 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
917 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
918 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
919 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
921 # Drop mul-div by the same value when there's no wrapping.
922 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
925 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
926 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
927 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
929 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
930 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
931 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
933 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
936 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
937 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
938 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
939 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
940 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
942 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
943 ('bcsel', ('ult', 31, 'bits'), 'insert',
944 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
945 'options->lower_bitfield_insert'),
946 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
947 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
948 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
949 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
950 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
951 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
953 # Alternative lowering that doesn't rely on bfi.
954 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
955 ('bcsel', ('ult', 31, 'bits'),
958 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
959 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
960 'options->lower_bitfield_insert_to_shifts'),
962 # Alternative lowering that uses bitfield_select.
963 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
964 ('bcsel', ('ult', 31, 'bits'), 'insert',
965 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
966 'options->lower_bitfield_insert_to_bitfield_select'),
968 (('ibitfield_extract', 'value', 'offset', 'bits'),
969 ('bcsel', ('ult', 31, 'bits'), 'value',
970 ('ibfe', 'value', 'offset', 'bits')),
971 'options->lower_bitfield_extract'),
973 (('ubitfield_extract', 'value', 'offset', 'bits'),
974 ('bcsel', ('ult', 31, 'bits'), 'value',
975 ('ubfe', 'value', 'offset', 'bits')),
976 'options->lower_bitfield_extract'),
978 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
979 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
980 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
981 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
982 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
983 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
984 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
986 (('ibitfield_extract', 'value', 'offset', 'bits'),
987 ('bcsel', ('ieq', 0, 'bits'),
990 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
991 ('isub', 32, 'bits'))),
992 'options->lower_bitfield_extract_to_shifts'),
994 (('ubitfield_extract', 'value', 'offset', 'bits'),
996 ('ushr', 'value', 'offset'),
997 ('bcsel', ('ieq', 'bits', 32),
999 ('isub', ('ishl', 1, 'bits'), 1))),
1000 'options->lower_bitfield_extract_to_shifts'),
1002 (('ifind_msb', 'value'),
1003 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1004 'options->lower_ifind_msb'),
1006 (('find_lsb', 'value'),
1007 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1008 'options->lower_find_lsb'),
1010 (('extract_i8', a
, 'b@32'),
1011 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
1012 'options->lower_extract_byte'),
1014 (('extract_u8', a
, 'b@32'),
1015 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
1016 'options->lower_extract_byte'),
1018 (('extract_i16', a
, 'b@32'),
1019 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
1020 'options->lower_extract_word'),
1022 (('extract_u16', a
, 'b@32'),
1023 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
1024 'options->lower_extract_word'),
1026 (('pack_unorm_2x16', 'v'),
1027 ('pack_uvec2_to_uint',
1028 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1029 'options->lower_pack_unorm_2x16'),
1031 (('pack_unorm_4x8', 'v'),
1032 ('pack_uvec4_to_uint',
1033 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1034 'options->lower_pack_unorm_4x8'),
1036 (('pack_snorm_2x16', 'v'),
1037 ('pack_uvec2_to_uint',
1038 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1039 'options->lower_pack_snorm_2x16'),
1041 (('pack_snorm_4x8', 'v'),
1042 ('pack_uvec4_to_uint',
1043 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1044 'options->lower_pack_snorm_4x8'),
1046 (('unpack_unorm_2x16', 'v'),
1047 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1048 ('extract_u16', 'v', 1))),
1050 'options->lower_unpack_unorm_2x16'),
1052 (('unpack_unorm_4x8', 'v'),
1053 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1054 ('extract_u8', 'v', 1),
1055 ('extract_u8', 'v', 2),
1056 ('extract_u8', 'v', 3))),
1058 'options->lower_unpack_unorm_4x8'),
1060 (('unpack_snorm_2x16', 'v'),
1061 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1062 ('extract_i16', 'v', 1))),
1064 'options->lower_unpack_snorm_2x16'),
1066 (('unpack_snorm_4x8', 'v'),
1067 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1068 ('extract_i8', 'v', 1),
1069 ('extract_i8', 'v', 2),
1070 ('extract_i8', 'v', 3))),
1072 'options->lower_unpack_snorm_4x8'),
1074 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
1075 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
1078 # bit_size dependent lowerings
1079 for bit_size
in [8, 16, 32, 64]:
1080 # convenience constants
1081 intmax
= (1 << (bit_size
- 1)) - 1
1082 intmin
= 1 << (bit_size
- 1)
1085 (('iadd_sat@' + str(bit_size
), a
, b
),
1086 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
1087 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
1088 (('isub_sat@' + str(bit_size
), a
, b
),
1089 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
1090 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
1093 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
1095 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
1096 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1097 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1098 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1099 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1101 # Optimize x2bN(b2x(x)) -> x
1102 for size
in type_sizes('bool'):
1103 aN
= 'a@' + str(size
)
1104 f2bN
= 'f2b' + str(size
)
1105 i2bN
= 'i2b' + str(size
)
1106 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
1107 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
1109 # Optimize x2yN(b2x(x)) -> b2y
1110 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1111 if x
!= 'f' and y
!= 'f' and x
!= y
:
1114 b2x
= 'b2f' if x
== 'f' else 'b2i'
1115 b2y
= 'b2f' if y
== 'f' else 'b2i'
1116 x2yN
= '{}2{}'.format(x
, y
)
1117 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1119 # Optimize away x2xN(a@N)
1120 for t
in ['int', 'uint', 'float']:
1121 for N
in type_sizes(t
):
1122 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1123 aN
= 'a@{0}'.format(N
)
1124 optimizations
.append(((x2xN
, aN
), a
))
1126 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1127 # In particular, we can optimize away everything except upcast of downcast and
1128 # upcasts where the type differs from the other cast
1129 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1131 # The outer cast is a down-cast. It doesn't matter what the size of the
1132 # argument of the inner cast is because we'll never been in the upcast
1133 # of downcast case. Regardless of types, we'll always end up with y2yN
1135 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1136 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1137 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1138 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1139 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1141 # If the outer cast is an up-cast, we have to be more careful about the
1142 # size of the argument of the inner cast and with types. In this case,
1143 # the type is always the type of type up-cast which is given by the
1145 for P
in type_sizes('uint'):
1146 # We can't optimize away up-cast of down-cast.
1150 # Because we're doing down-cast of down-cast, the types always have
1151 # to match between the two casts
1152 for x
in ['i', 'u']:
1153 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1154 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1155 aP
= 'a@{0}'.format(P
)
1156 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1158 # The N == M case is handled by other optimizations
1161 # Optimize comparisons with up-casts
1162 for t
in ['int', 'uint', 'float']:
1163 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1164 if N
== 1 or N
>= M
:
1167 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1168 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1171 xeq
= 'feq' if t
== 'float' else 'ieq'
1172 xne
= 'fne' if t
== 'float' else 'ine'
1173 xge
= '{0}ge'.format(t
[0])
1174 xlt
= '{0}lt'.format(t
[0])
1176 # Up-casts are lossless so for correctly signed comparisons of
1177 # up-casted values we can do the comparison at the largest of the two
1178 # original sizes and drop one or both of the casts. (We have
1179 # optimizations to drop the no-op casts which this may generate.)
1180 for P
in type_sizes(t
):
1186 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1187 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1188 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1189 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1190 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1191 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1194 # The next bit doesn't work on floats because the range checks would
1195 # get way too complicated.
1196 if t
in ['int', 'uint']:
1198 xN_min
= -(1 << (N
- 1))
1199 xN_max
= (1 << (N
- 1)) - 1
1202 xN_max
= (1 << N
) - 1
1206 # If we're up-casting and comparing to a constant, we can unfold
1207 # the comparison into a comparison with the shrunk down constant
1208 # and a check that the constant fits in the smaller bit size.
1210 ((xeq
, (x2xM
, aN
), '#b'),
1211 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1212 ((xne
, (x2xM
, aN
), '#b'),
1213 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1214 ((xlt
, (x2xM
, aN
), '#b'),
1215 ('iand', (xlt
, xN_min
, b
),
1216 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1217 ((xlt
, '#a', (x2xM
, bN
)),
1218 ('iand', (xlt
, a
, xN_max
),
1219 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1220 ((xge
, (x2xM
, aN
), '#b'),
1221 ('iand', (xge
, xN_max
, b
),
1222 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1223 ((xge
, '#a', (x2xM
, bN
)),
1224 ('iand', (xge
, a
, xN_min
),
1225 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1228 def fexp2i(exp
, bits
):
1229 # We assume that exp is already in the right range.
1231 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1233 return ('ishl', ('iadd', exp
, 127), 23)
1235 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1239 def ldexp(f
, exp
, bits
):
1240 # First, we clamp exp to a reasonable range. The maximum possible range
1241 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1242 # a maximum range of [-149, 127]. This means that we can potentially have
1243 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1244 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1245 # spec, on the other hand, only requires that we handle an exponent value
1246 # in the range [-126, 128]. This implementation is *mostly* correct; it
1247 # handles a range on exp of [-252, 254] which allows you to create any
1248 # value (including denorms if the hardware supports it) and to adjust the
1249 # exponent of any normal value to anything you want.
1251 exp
= ('imin', ('imax', exp
, -28), 30)
1253 exp
= ('imin', ('imax', exp
, -252), 254)
1255 exp
= ('imin', ('imax', exp
, -2044), 2046)
1259 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1260 # (We use ishr which isn't the same for -1, but the -1 case still works
1261 # since we use exp-exp/2 as the second exponent.) While the spec
1262 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1263 # work with denormals and doesn't allow for the full swing in exponents
1264 # that you can get with normalized values. Instead, we create two powers
1265 # of two and multiply by them each in turn. That way the effective range
1266 # of our exponent is doubled.
1267 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1268 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1269 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1272 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1273 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1274 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1277 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1278 def bitfield_reverse(u
):
1279 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1280 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1281 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1282 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1283 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1287 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1289 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1290 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1291 # and, if a is a NaN then the second comparison will fail anyway.
1292 for op
in ['flt', 'fge', 'feq']:
1294 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1295 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1298 # Add optimizations to handle the case where the result of a ternary is
1299 # compared to a constant. This way we can take things like
1305 # a ? (0 > 0) : (1 > 0)
1307 # which constant folding will eat for lunch. The resulting ternary will
1308 # further get cleaned up by the boolean reductions above and we will be
1309 # left with just the original variable "a".
1310 for op
in ['flt', 'fge', 'feq', 'fne',
1311 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1313 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1314 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1315 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1316 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1320 # For example, this converts things like
1322 # 1 + mix(0, a - 1, condition)
1326 # mix(1, (a-1)+1, condition)
1328 # Other optimizations will rearrange the constants.
1329 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1331 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1334 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1337 # If neither layout qualifier is specified, derivatives in compute shaders
1338 # return zero, which is consistent with the handling of built-in texture
1339 # functions like texture() in GLSL 4.50 compute shaders.
1340 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1341 'fddy', 'fddy_fine', 'fddy_coarse']:
1343 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1346 # Some optimizations for ir3-specific instructions.
1348 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1349 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1350 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1351 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1352 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1355 # This section contains "late" optimizations that should be run before
1356 # creating ffmas and calling regular optimizations for the final time.
1357 # Optimizations should go here if they help code generation and conflict
1358 # with the regular optimizations.
1359 before_ffma_optimizations
= [
1360 # Propagate constants down multiplication chains
1361 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1362 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1363 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1364 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1366 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1367 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1368 (('~fadd', ('fneg', a
), a
), 0.0),
1369 (('iadd', ('ineg', a
), a
), 0),
1370 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1371 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1372 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1373 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1375 (('~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
)),
1376 (('~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
)),
1377 (('~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
))),
1380 # This section contains "late" optimizations that should be run after the
1381 # regular optimizations have finished. Optimizations should go here if
1382 # they help code generation but do not necessarily produce code that is
1383 # more easily optimizable.
1384 late_optimizations
= [
1385 # Most of these optimizations aren't quite safe when you get infinity or
1386 # Nan involved but the first one should be fine.
1387 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1388 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1389 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1390 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1391 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1392 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1394 # nir_lower_to_source_mods will collapse this, but its existence during the
1395 # optimization loop can prevent other optimizations.
1396 (('fneg', ('fneg', a
)), a
),
1398 # These are duplicated from the main optimizations table. The late
1399 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1400 # new patterns like these. The patterns that compare with zero are removed
1401 # because they are unlikely to be created in by anything in
1402 # late_optimizations.
1403 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1404 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1405 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1406 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1407 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1408 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1410 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1411 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1413 (('~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
)))),
1415 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1416 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1417 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1418 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1419 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1420 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1421 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1422 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1423 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1424 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1427 (('iand', a
, a
), a
),
1429 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1431 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1432 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1433 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1434 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1436 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1437 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1439 (('~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'),
1440 (('~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'),
1442 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1443 # particular operation is common for expanding values stored in a texture
1444 # from [0,1] to [-1,1].
1445 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1446 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1447 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1448 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1449 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1450 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1451 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1452 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1456 # a + -a*a + a*b (1)
1458 # Option 1: ffma(a, (b-a), a)
1460 # Alternately, after (1):
1466 # Option 2: ffma(a, 2, -(a*a))
1467 # Option 3: ffma(a, 2, (-a)*a)
1468 # Option 4: ffma(a, -a, (2*a)
1469 # Option 5: a * (2 - a)
1471 # There are a lot of other possible combinations.
1472 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1473 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1474 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1475 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1476 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1478 # we do these late so that we don't get in the way of creating ffmas
1479 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1480 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1482 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1484 # Things that look like DPH in the source shader may get expanded to
1485 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1486 # to NIR. After FFMA is generated, this can look like:
1488 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1490 # Reassociate the last addition into the first multiplication.
1491 (('~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)'),
1492 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '!options->intel_vec4'),
1493 (('~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)'),
1494 ('ffma', a
, b
, ('ffma', e
, 'f', 'g') ), '!options->intel_vec4'),
1497 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1498 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1499 before_ffma_optimizations
).render())
1500 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1501 late_optimizations
).render())