2 # Copyright (C) 2014 Intel Corporation
4 # Permission is hereby granted, free of charge, to any person obtaining a
5 # copy of this software and associated documentation files (the "Software"),
6 # to deal in the Software without restriction, including without limitation
7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 # and/or sell copies of the Software, and to permit persons to whom the
9 # Software is furnished to do so, subject to the following conditions:
11 # The above copyright notice and this permission notice (including the next
12 # paragraph) shall be included in all copies or substantial portions of the
15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 # Jason Ekstrand (jason@jlekstrand.net)
26 from __future__
import print_function
28 from collections
import OrderedDict
30 from nir_opcodes
import type_sizes
33 # Convenience variables
40 # Written in the form (<search>, <replace>) where <search> is an expression
41 # and <replace> is either an expression or a value. An expression is
42 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
43 # where each source is either an expression or a value. A value can be
44 # either a numeric constant or a string representing a variable name.
46 # If the opcode in a search expression is prefixed by a '~' character, this
47 # indicates that the operation is inexact. Such operations will only get
48 # applied to SSA values that do not have the exact bit set. This should be
49 # used by by any optimizations that are not bit-for-bit exact. It should not,
50 # however, be used for backend-requested lowering operations as those need to
51 # happen regardless of precision.
53 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
54 # that the given variable will only match constants and the type indicates that
55 # the given variable will only match values from ALU instructions with the
56 # given output type, and (cond) specifies an additional condition function
57 # (see nir_search_helpers.h).
59 # For constants, you have to be careful to make sure that it is the right
60 # type because python is unaware of the source and destination types of the
63 # All expression types can have a bit-size specified. For opcodes, this
64 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
65 # type and size. In the search half of the expression this indicates that it
66 # should only match that particular bit-size. In the replace half of the
67 # expression this indicates that the constructed value should have that
70 # A special condition "many-comm-expr" can be used with expressions to note
71 # that the expression and its subexpressions have more commutative expressions
72 # than nir_replace_instr can handle. If this special condition is needed with
73 # another condition, the two can be separated by a comma (e.g.,
74 # "(many-comm-expr,is_used_once)").
78 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
79 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitshift'),
80 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitshift'),
82 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
83 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
84 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
85 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
90 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
91 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
92 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
93 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
95 (('fneg', ('fneg', a
)), a
),
96 (('ineg', ('ineg', a
)), a
),
97 (('fabs', ('fabs', a
)), ('fabs', a
)),
98 (('fabs', ('fneg', a
)), ('fabs', a
)),
99 (('fabs', ('u2f', a
)), ('u2f', a
)),
100 (('iabs', ('iabs', a
)), ('iabs', a
)),
101 (('iabs', ('ineg', a
)), ('iabs', a
)),
102 (('f2b', ('fneg', a
)), ('f2b', a
)),
103 (('i2b', ('ineg', a
)), ('i2b', a
)),
104 (('~fadd', a
, 0.0), a
),
106 (('usadd_4x8', a
, 0), a
),
107 (('usadd_4x8', a
, ~
0), ~
0),
108 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
109 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
110 (('~fadd', ('fneg', a
), a
), 0.0),
111 (('iadd', ('ineg', a
), a
), 0),
112 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
113 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
114 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
115 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
116 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
117 (('~fmul', a
, 0.0), 0.0),
119 (('umul_unorm_4x8', a
, 0), 0),
120 (('umul_unorm_4x8', a
, ~
0), a
),
121 (('fmul', a
, 1.0), a
),
123 (('fmul', a
, -1.0), ('fneg', a
)),
124 (('imul', a
, -1), ('ineg', a
)),
125 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
126 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
127 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
128 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
129 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
130 (('~ffma', 0.0, a
, b
), b
),
131 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
132 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
133 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
134 (('~flrp', a
, b
, 0.0), a
),
135 (('~flrp', a
, b
, 1.0), b
),
136 (('~flrp', a
, a
, b
), a
),
137 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
139 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
140 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
141 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
142 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
144 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
145 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
147 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
148 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
150 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
152 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
153 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
154 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
155 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
156 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
157 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
158 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
159 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
160 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
161 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
162 # These are the same as the previous three rules, but it depends on
163 # 1-fsat(x) <=> fsat(1-x). See below.
164 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
165 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
167 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
168 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
169 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
170 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
171 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
173 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
174 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
176 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
177 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
178 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
179 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
181 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
182 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
184 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
185 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
186 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
187 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
189 # 1 - ((1 - a) * (1 - b))
190 # 1 - (1 - a - b + a*b)
191 # 1 - 1 + a + b - a*b
196 (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a
)), ('fadd', 1.0, ('fneg', b
))))),
197 ('flrp', b
, 1.0, a
), '!options->lower_flrp32'),
199 # (a * #b + #c) << #d
200 # ((a * #b) << #d) + (#c << #d)
201 # (a * (#b << #d)) + (#c << #d)
202 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
203 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
207 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
209 # Comparison simplifications
210 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
211 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
212 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
213 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
214 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
215 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
216 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
217 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
218 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
219 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
221 # This helps some shaders because, after some optimizations, they end up
222 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
223 # matching would be handled by CSE.
224 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
225 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
226 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
227 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
228 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
229 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
230 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
231 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
232 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
233 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
235 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
236 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
237 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
238 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
239 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
240 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
242 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
243 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
244 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
245 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
249 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
251 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
253 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
255 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
256 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
257 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
258 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
259 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
260 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
261 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
262 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
263 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
264 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
265 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
266 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
267 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
268 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
269 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
270 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
271 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
272 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
273 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
274 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
276 # -(b2f(a) + b2f(b)) < 0
277 # 0 < b2f(a) + b2f(b)
278 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
280 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
281 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
283 # -(b2f(a) + b2f(b)) >= 0
284 # 0 >= b2f(a) + b2f(b)
285 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
287 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
288 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
290 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
291 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
293 # Some optimizations (below) convert things like (a < b || c < b) into
294 # (min(a, c) < b). However, this interfers with the previous optimizations
295 # that try to remove comparisons with negated sums of b2f. This just
297 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
298 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
300 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
301 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
302 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
303 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
305 # Cannot remove the addition from ilt or ige due to overflow.
306 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
307 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
309 # fmin(-b2f(a), b) >= 0.0
310 # -b2f(a) >= 0.0 && b >= 0.0
311 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
312 # b2f(a) == 0.0 && b >= 0.0
313 # a == False && b >= 0.0
316 # The fge in the second replacement is not a typo. I leave the proof that
317 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
319 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
320 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
322 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
323 (('fne', ('b2f', 'a@1'), 0.0), a
),
324 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
325 (('ine', ('b2i', 'a@1'), 0), a
),
327 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
328 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
329 (('fge', ('u2f', a
), 0.0), True),
330 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
331 (('flt', ('u2f', a
), 0.0), False),
332 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
333 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
334 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
335 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
336 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
337 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
338 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
342 # fabs(a) != 0.0 because fabs(a) must be >= 0
344 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
348 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
351 # 0.0 == fabs(a) because fabs(a) must be >= 0
353 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
357 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
359 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
360 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
361 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
362 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
365 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
366 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
367 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
369 # Since b is a constant, constant folding will eliminate the fmin and the
370 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
371 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
373 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
375 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
376 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
377 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
378 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
379 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
380 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
381 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
382 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
383 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
384 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
385 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
386 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
387 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
388 (('bcsel', a
, True, b
), ('ior', a
, b
)),
389 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
390 (('bcsel', a
, b
, False), ('iand', a
, b
)),
391 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
398 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
399 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
400 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
401 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
402 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
403 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
404 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
405 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
406 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
407 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
408 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
409 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
410 (('fmin', a
, ('fabs', a
)), a
),
411 (('imin', a
, ('iabs', a
)), a
),
412 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
413 (('imax', a
, ('ineg', ('iabs', a
))), a
),
414 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
415 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
416 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
417 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
418 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
419 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
420 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
421 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
422 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
423 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
424 (('fsat', ('b2f', a
)), ('b2f', a
)),
425 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
426 (('fsat', ('fsat', a
)), ('fsat', a
)),
427 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_negate && !options->lower_fsat'),
428 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_negate && !options->lower_fsat'),
429 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
430 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
431 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
432 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
433 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
434 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
435 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
436 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
437 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
438 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
439 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
440 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
441 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
442 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
443 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
444 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
445 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
446 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
447 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
448 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
449 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
450 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
451 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
453 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
454 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
455 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
456 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
457 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
458 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
459 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
460 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
461 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
462 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
463 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
464 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
465 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
466 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
467 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
468 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
470 # Common pattern like 'if (i == 0 || i == 1 || ...)'
471 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
472 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
473 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
475 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
476 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
477 # so emit an open-coded version of that.
478 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
479 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
481 (('ior', a
, ('ieq', a
, False)), True),
482 (('ior', a
, ('inot', a
)), -1),
484 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
485 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
487 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
489 # These patterns can result when (a < b || a < c) => (a < min(b, c))
490 # transformations occur before constant propagation and loop-unrolling.
491 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
492 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
493 (('~fge', a
, ('fmin', b
, a
)), True),
494 (('~fge', ('fmax', a
, b
), a
), True),
495 (('~flt', a
, ('fmin', b
, a
)), False),
496 (('~flt', ('fmax', a
, b
), a
), False),
497 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
498 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
500 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
501 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
502 (('ige', a
, ('imin', b
, a
)), True),
503 (('ige', ('imax', a
, b
), a
), True),
504 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
505 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
506 (('uge', a
, ('umin', b
, a
)), True),
507 (('uge', ('umax', a
, b
), a
), True),
508 (('ilt', a
, ('imin', b
, a
)), False),
509 (('ilt', ('imax', a
, b
), a
), False),
510 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
511 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
512 (('ult', a
, ('umin', b
, a
)), False),
513 (('ult', ('umax', a
, b
), a
), False),
514 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
515 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
516 (('ult', a
, ('iand', b
, a
)), False),
517 (('ult', ('ior', a
, b
), a
), False),
518 (('uge', a
, ('iand', b
, a
)), True),
519 (('uge', ('ior', a
, b
), a
), True),
521 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
522 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
523 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
524 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
525 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
526 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
527 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
528 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
529 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
530 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
531 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
532 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
533 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
534 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
535 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
536 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
538 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
540 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
541 ('iabs', ('ishr', a
, b
))),
542 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
544 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
545 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
546 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
547 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
548 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
549 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
550 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
551 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
552 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
553 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
555 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
556 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
557 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
558 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
559 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
560 (('ineg', ('b2i32', 'a@32')), a
),
561 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
562 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
563 # Comparison with the same args. Note that these are not done for
564 # the float versions because NaN always returns false on float
566 (('ilt', a
, a
), False),
567 (('ige', a
, a
), True),
568 (('ieq', a
, a
), True),
569 (('ine', a
, a
), False),
570 (('ult', a
, a
), False),
571 (('uge', a
, a
), True),
572 # Logical and bit operations
574 (('iand', a
, ~
0), a
),
578 (('ior', a
, True), True),
581 (('inot', ('inot', a
)), a
),
582 (('ior', ('iand', a
, b
), b
), b
),
583 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
584 (('iand', ('ior', a
, b
), b
), b
),
585 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
587 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
588 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
589 # Shift optimizations
596 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
597 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
598 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
599 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
600 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
601 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
602 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
603 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
604 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
605 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
606 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
607 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
608 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
609 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
610 # Exponential/logarithmic identities
611 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
612 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
613 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
614 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
615 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
616 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
617 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
618 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
619 (('~fpow', a
, 1.0), a
),
620 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
621 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
622 (('~fpow', 2.0, a
), ('fexp2', a
)),
623 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
624 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
625 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
626 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
627 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
628 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
629 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
630 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
631 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
632 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
633 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
634 # Division and reciprocal
635 (('~fdiv', 1.0, a
), ('frcp', a
)),
636 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
637 (('~frcp', ('frcp', a
)), a
),
638 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
639 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
640 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
641 # Boolean simplifications
642 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
643 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
644 (('ieq', a
, True), a
),
645 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
646 (('ine', a
, False), a
),
647 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
648 (('bcsel', a
, True, False), a
),
649 (('bcsel', a
, False, True), ('inot', a
)),
650 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
651 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
652 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
653 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
654 (('bcsel', True, b
, c
), b
),
655 (('bcsel', False, b
, c
), c
),
656 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
658 (('bcsel', a
, b
, b
), b
),
659 (('fcsel', a
, b
, b
), b
),
661 # D3D Boolean emulation
662 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
663 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
664 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
665 ('ineg', ('b2i', ('iand', a
, b
)))),
666 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
667 ('ineg', ('b2i', ('ior', a
, b
)))),
668 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
669 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
670 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
671 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
672 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
674 # SM5 32-bit shifts are defined to use the 5 least significant bits
675 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
676 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
677 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
680 (('i2b32', ('b2i', 'a@32')), a
),
681 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
682 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
683 (('i2b', ('ineg', a
)), ('i2b', a
)),
684 (('i2b', ('iabs', a
)), ('i2b', a
)),
685 (('fabs', ('b2f', a
)), ('b2f', a
)),
686 (('iabs', ('b2i', a
)), ('b2i', a
)),
687 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
689 # Ironically, mark these as imprecise because removing the conversions may
690 # preserve more precision than doing the conversions (e.g.,
691 # uint(float(0x81818181u)) == 0x81818200).
692 (('~f2i32', ('i2f', 'a@32')), a
),
693 (('~f2i32', ('u2f', 'a@32')), a
),
694 (('~f2u32', ('i2f', 'a@32')), a
),
695 (('~f2u32', ('u2f', 'a@32')), a
),
697 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
700 # It is undefined to convert a negative floating-point value to an
703 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
704 # some optimizations in the i965 backend to proceed.
705 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
706 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
707 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
708 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
710 (('~fmin', ('fabs', a
), 1.0), ('fsat', ('fabs', a
)), '!options->lower_fsat'),
712 # The result of the multiply must be in [-1, 0], so the result of the ffma
714 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
715 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
716 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
717 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
719 # Packing and then unpacking does nothing
720 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
721 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
722 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
723 ('unpack_64_2x32_split_y', a
)), a
),
725 # Comparing two halves of an unpack separately. While this optimization
726 # should be correct for non-constant values, it's less obvious that it's
727 # useful in that case. For constant values, the pack will fold and we're
728 # guaranteed to reduce the whole tree to one instruction.
729 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
730 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
731 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
734 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
735 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
736 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
737 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
738 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
739 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
740 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
743 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
744 # patterns like those below.
745 for op
in ('ushr', 'ishr'):
746 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
747 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
748 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
750 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
752 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
753 # patterns like those below.
754 for op
in ('extract_u8', 'extract_i8'):
755 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
756 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
757 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
759 optimizations
.extend([
761 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
762 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
763 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
764 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
765 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
768 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
769 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
770 (('ussub_4x8', a
, 0), a
),
771 (('ussub_4x8', a
, ~
0), 0),
772 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
773 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
774 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
775 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
776 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
777 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
778 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
779 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
781 # Propagate negation up multiplication chains
782 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
783 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
785 # Propagate constants up multiplication chains
786 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
787 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
788 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
789 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
791 # Reassociate constants in add/mul chains so they can be folded together.
792 # For now, we mostly only handle cases where the constants are separated by
793 # a single non-constant. We could do better eventually.
794 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
795 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
796 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
797 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
798 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
800 # Drop mul-div by the same value when there's no wrapping.
801 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
804 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
805 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
806 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
808 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
809 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
810 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
812 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
815 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
816 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
817 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
818 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
819 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
821 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
822 ('bcsel', ('ult', 31, 'bits'), 'insert',
823 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
824 'options->lower_bitfield_insert'),
825 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
826 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
827 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
828 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
829 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
830 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
832 # Alternative lowering that doesn't rely on bfi.
833 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
834 ('bcsel', ('ult', 31, 'bits'),
837 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
838 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
839 'options->lower_bitfield_insert_to_shifts'),
841 # Alternative lowering that uses bitfield_select.
842 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
843 ('bcsel', ('ult', 31, 'bits'), 'insert',
844 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
845 'options->lower_bitfield_insert_to_bitfield_select'),
847 (('ibitfield_extract', 'value', 'offset', 'bits'),
848 ('bcsel', ('ult', 31, 'bits'), 'value',
849 ('ibfe', 'value', 'offset', 'bits')),
850 'options->lower_bitfield_extract'),
852 (('ubitfield_extract', 'value', 'offset', 'bits'),
853 ('bcsel', ('ult', 31, 'bits'), 'value',
854 ('ubfe', 'value', 'offset', 'bits')),
855 'options->lower_bitfield_extract'),
857 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
858 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
859 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
860 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
861 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
862 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
863 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
865 (('ibitfield_extract', 'value', 'offset', 'bits'),
866 ('bcsel', ('ieq', 0, 'bits'),
869 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
870 ('isub', 32, 'bits'))),
871 'options->lower_bitfield_extract_to_shifts'),
873 (('ubitfield_extract', 'value', 'offset', 'bits'),
875 ('ushr', 'value', 'offset'),
876 ('bcsel', ('ieq', 'bits', 32),
878 ('isub', ('ishl', 1, 'bits'), 1))),
879 'options->lower_bitfield_extract_to_shifts'),
881 (('ifind_msb', 'value'),
882 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
883 'options->lower_ifind_msb'),
885 (('find_lsb', 'value'),
886 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
887 'options->lower_find_lsb'),
889 (('extract_i8', a
, 'b@32'),
890 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
891 'options->lower_extract_byte'),
893 (('extract_u8', a
, 'b@32'),
894 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
895 'options->lower_extract_byte'),
897 (('extract_i16', a
, 'b@32'),
898 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
899 'options->lower_extract_word'),
901 (('extract_u16', a
, 'b@32'),
902 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
903 'options->lower_extract_word'),
905 (('pack_unorm_2x16', 'v'),
906 ('pack_uvec2_to_uint',
907 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
908 'options->lower_pack_unorm_2x16'),
910 (('pack_unorm_4x8', 'v'),
911 ('pack_uvec4_to_uint',
912 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
913 'options->lower_pack_unorm_4x8'),
915 (('pack_snorm_2x16', 'v'),
916 ('pack_uvec2_to_uint',
917 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
918 'options->lower_pack_snorm_2x16'),
920 (('pack_snorm_4x8', 'v'),
921 ('pack_uvec4_to_uint',
922 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
923 'options->lower_pack_snorm_4x8'),
925 (('unpack_unorm_2x16', 'v'),
926 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
927 ('extract_u16', 'v', 1))),
929 'options->lower_unpack_unorm_2x16'),
931 (('unpack_unorm_4x8', 'v'),
932 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
933 ('extract_u8', 'v', 1),
934 ('extract_u8', 'v', 2),
935 ('extract_u8', 'v', 3))),
937 'options->lower_unpack_unorm_4x8'),
939 (('unpack_snorm_2x16', 'v'),
940 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
941 ('extract_i16', 'v', 1))),
943 'options->lower_unpack_snorm_2x16'),
945 (('unpack_snorm_4x8', 'v'),
946 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
947 ('extract_i8', 'v', 1),
948 ('extract_i8', 'v', 2),
949 ('extract_i8', 'v', 3))),
951 'options->lower_unpack_snorm_4x8'),
953 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
954 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
957 # bit_size dependent lowerings
958 for bit_size
in [8, 16, 32, 64]:
959 # convenience constants
960 intmax
= (1 << (bit_size
- 1)) - 1
961 intmin
= 1 << (bit_size
- 1)
964 (('iadd_sat@' + str(bit_size
), a
, b
),
965 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
966 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
967 (('isub_sat@' + str(bit_size
), a
, b
),
968 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
969 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
972 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
974 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
975 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
976 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
977 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
978 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
980 # Optimize x2bN(b2x(x)) -> x
981 for size
in type_sizes('bool'):
982 aN
= 'a@' + str(size
)
983 f2bN
= 'f2b' + str(size
)
984 i2bN
= 'i2b' + str(size
)
985 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
986 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
988 # Optimize x2yN(b2x(x)) -> b2y
989 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
990 if x
!= 'f' and y
!= 'f' and x
!= y
:
993 b2x
= 'b2f' if x
== 'f' else 'b2i'
994 b2y
= 'b2f' if y
== 'f' else 'b2i'
995 x2yN
= '{}2{}'.format(x
, y
)
996 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
998 # Optimize away x2xN(a@N)
999 for t
in ['int', 'uint', 'float']:
1000 for N
in type_sizes(t
):
1001 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1002 aN
= 'a@{0}'.format(N
)
1003 optimizations
.append(((x2xN
, aN
), a
))
1005 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1006 # In particular, we can optimize away everything except upcast of downcast and
1007 # upcasts where the type differs from the other cast
1008 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1010 # The outer cast is a down-cast. It doesn't matter what the size of the
1011 # argument of the inner cast is because we'll never been in the upcast
1012 # of downcast case. Regardless of types, we'll always end up with y2yN
1014 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1015 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1016 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1017 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1018 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1020 # If the outer cast is an up-cast, we have to be more careful about the
1021 # size of the argument of the inner cast and with types. In this case,
1022 # the type is always the type of type up-cast which is given by the
1024 for P
in type_sizes('uint'):
1025 # We can't optimize away up-cast of down-cast.
1029 # Because we're doing down-cast of down-cast, the types always have
1030 # to match between the two casts
1031 for x
in ['i', 'u']:
1032 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1033 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1034 aP
= 'a@{0}'.format(P
)
1035 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1037 # The N == M case is handled by other optimizations
1040 def fexp2i(exp
, bits
):
1041 # We assume that exp is already in the right range.
1043 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1045 return ('ishl', ('iadd', exp
, 127), 23)
1047 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1051 def ldexp(f
, exp
, bits
):
1052 # First, we clamp exp to a reasonable range. The maximum possible range
1053 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1054 # a maximum range of [-149, 127]. This means that we can potentially have
1055 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1056 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1057 # spec, on the other hand, only requires that we handle an exponent value
1058 # in the range [-126, 128]. This implementation is *mostly* correct; it
1059 # handles a range on exp of [-252, 254] which allows you to create any
1060 # value (including denorms if the hardware supports it) and to adjust the
1061 # exponent of any normal value to anything you want.
1063 exp
= ('imin', ('imax', exp
, -28), 30)
1065 exp
= ('imin', ('imax', exp
, -252), 254)
1067 exp
= ('imin', ('imax', exp
, -2044), 2046)
1071 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1072 # (We use ishr which isn't the same for -1, but the -1 case still works
1073 # since we use exp-exp/2 as the second exponent.) While the spec
1074 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1075 # work with denormals and doesn't allow for the full swing in exponents
1076 # that you can get with normalized values. Instead, we create two powers
1077 # of two and multiply by them each in turn. That way the effective range
1078 # of our exponent is doubled.
1079 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1080 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1081 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1084 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1085 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1086 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1089 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1090 def bitfield_reverse(u
):
1091 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1092 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1093 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1094 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1095 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1099 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1101 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1102 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1103 # and, if a is a NaN then the second comparison will fail anyway.
1104 for op
in ['flt', 'fge', 'feq']:
1106 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1107 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1110 # Add optimizations to handle the case where the result of a ternary is
1111 # compared to a constant. This way we can take things like
1117 # a ? (0 > 0) : (1 > 0)
1119 # which constant folding will eat for lunch. The resulting ternary will
1120 # further get cleaned up by the boolean reductions above and we will be
1121 # left with just the original variable "a".
1122 for op
in ['flt', 'fge', 'feq', 'fne',
1123 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1125 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1126 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1127 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1128 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1132 # For example, this converts things like
1134 # 1 + mix(0, a - 1, condition)
1138 # mix(1, (a-1)+1, condition)
1140 # Other optimizations will rearrange the constants.
1141 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1143 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1146 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1149 # If neither layout qualifier is specified, derivatives in compute shaders
1150 # return zero, which is consistent with the handling of built-in texture
1151 # functions like texture() in GLSL 4.50 compute shaders.
1152 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1153 'fddy', 'fddy_fine', 'fddy_coarse']:
1155 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1158 # Some optimizations for ir3-specific instructions.
1160 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1161 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1162 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1163 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1164 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1167 # This section contains "late" optimizations that should be run before
1168 # creating ffmas and calling regular optimizations for the final time.
1169 # Optimizations should go here if they help code generation and conflict
1170 # with the regular optimizations.
1171 before_ffma_optimizations
= [
1172 # Propagate constants down multiplication chains
1173 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1174 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1175 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1176 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1178 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1179 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1180 (('~fadd', ('fneg', a
), a
), 0.0),
1181 (('iadd', ('ineg', a
), a
), 0),
1182 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1183 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1184 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1185 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1187 (('~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
)),
1188 (('~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
)),
1189 (('~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
))),
1192 # This section contains "late" optimizations that should be run after the
1193 # regular optimizations have finished. Optimizations should go here if
1194 # they help code generation but do not necessarily produce code that is
1195 # more easily optimizable.
1196 late_optimizations
= [
1197 # Most of these optimizations aren't quite safe when you get infinity or
1198 # Nan involved but the first one should be fine.
1199 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1200 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1201 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1202 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1203 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1204 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1206 # nir_lower_to_source_mods will collapse this, but its existence during the
1207 # optimization loop can prevent other optimizations.
1208 (('fneg', ('fneg', a
)), a
),
1210 # These are duplicated from the main optimizations table. The late
1211 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1212 # new patterns like these. The patterns that compare with zero are removed
1213 # because they are unlikely to be created in by anything in
1214 # late_optimizations.
1215 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1216 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1217 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1218 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1219 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1220 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1222 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1223 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1225 (('~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
)))),
1227 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1228 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1229 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1230 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1231 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1232 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1233 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1234 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1235 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1236 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1239 (('iand', a
, a
), a
),
1241 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1243 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1244 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1245 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1246 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1248 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1249 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1251 (('~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'),
1252 (('~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'),
1254 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1255 # particular operation is common for expanding values stored in a texture
1256 # from [0,1] to [-1,1].
1257 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1258 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1259 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1260 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1261 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1262 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1263 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1264 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1268 # a + -a*a + a*b (1)
1270 # Option 1: ffma(a, (b-a), a)
1272 # Alternately, after (1):
1278 # Option 2: ffma(a, 2, -(a*a))
1279 # Option 3: ffma(a, 2, (-a)*a)
1280 # Option 4: ffma(a, -a, (2*a)
1281 # Option 5: a * (2 - a)
1283 # There are a lot of other possible combinations.
1284 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1285 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1286 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1287 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1288 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1290 # we do these late so that we don't get in the way of creating ffmas
1291 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1292 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1294 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1296 # Things that look like DPH in the source shader may get expanded to
1297 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1298 # to NIR. After FFMA is generated, this can look like:
1300 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1302 # Reassociate the last addition into the first multiplication.
1303 (('~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)'),
1304 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '!options->intel_vec4'),
1305 (('~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)'),
1306 ('ffma', a
, b
, ('ffma', e
, 'f', 'g') ), '!options->intel_vec4'),
1309 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1310 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1311 before_ffma_optimizations
).render())
1312 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1313 late_optimizations
).render())