2 # Copyright (C) 2014 Intel Corporation
4 # Permission is hereby granted, free of charge, to any person obtaining a
5 # copy of this software and associated documentation files (the "Software"),
6 # to deal in the Software without restriction, including without limitation
7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 # and/or sell copies of the Software, and to permit persons to whom the
9 # Software is furnished to do so, subject to the following conditions:
11 # The above copyright notice and this permission notice (including the next
12 # paragraph) shall be included in all copies or substantial portions of the
15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 # Jason Ekstrand (jason@jlekstrand.net)
26 from __future__
import print_function
28 from collections
import OrderedDict
30 from nir_opcodes
import type_sizes
33 # Convenience variables
40 # Written in the form (<search>, <replace>) where <search> is an expression
41 # and <replace> is either an expression or a value. An expression is
42 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
43 # where each source is either an expression or a value. A value can be
44 # either a numeric constant or a string representing a variable name.
46 # If the opcode in a search expression is prefixed by a '~' character, this
47 # indicates that the operation is inexact. Such operations will only get
48 # applied to SSA values that do not have the exact bit set. This should be
49 # used by by any optimizations that are not bit-for-bit exact. It should not,
50 # however, be used for backend-requested lowering operations as those need to
51 # happen regardless of precision.
53 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
54 # that the given variable will only match constants and the type indicates that
55 # the given variable will only match values from ALU instructions with the
56 # given output type, and (cond) specifies an additional condition function
57 # (see nir_search_helpers.h).
59 # For constants, you have to be careful to make sure that it is the right
60 # type because python is unaware of the source and destination types of the
63 # All expression types can have a bit-size specified. For opcodes, this
64 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
65 # type and size. In the search half of the expression this indicates that it
66 # should only match that particular bit-size. In the replace half of the
67 # expression this indicates that the constructed value should have that
72 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
73 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
74 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
75 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
76 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
77 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
82 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
83 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
84 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
85 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
87 (('fneg', ('fneg', a
)), a
),
88 (('ineg', ('ineg', a
)), a
),
89 (('fabs', ('fabs', a
)), ('fabs', a
)),
90 (('fabs', ('fneg', a
)), ('fabs', a
)),
91 (('fabs', ('u2f', a
)), ('u2f', a
)),
92 (('iabs', ('iabs', a
)), ('iabs', a
)),
93 (('iabs', ('ineg', a
)), ('iabs', a
)),
94 (('f2b', ('fneg', a
)), ('f2b', a
)),
95 (('i2b', ('ineg', a
)), ('i2b', a
)),
96 (('~fadd', a
, 0.0), a
),
98 (('usadd_4x8', a
, 0), a
),
99 (('usadd_4x8', a
, ~
0), ~
0),
100 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
101 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
102 (('~fadd', ('fneg', a
), a
), 0.0),
103 (('iadd', ('ineg', a
), a
), 0),
104 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
105 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
106 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
107 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
108 (('~fmul', a
, 0.0), 0.0),
110 (('umul_unorm_4x8', a
, 0), 0),
111 (('umul_unorm_4x8', a
, ~
0), a
),
112 (('fmul', a
, 1.0), a
),
114 (('fmul', a
, -1.0), ('fneg', a
)),
115 (('imul', a
, -1), ('ineg', a
)),
116 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
117 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
118 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
119 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
120 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
121 (('~ffma', 0.0, a
, b
), b
),
122 (('~ffma', a
, 0.0, b
), b
),
123 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
124 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
125 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
126 (('~flrp', a
, b
, 0.0), a
),
127 (('~flrp', a
, b
, 1.0), b
),
128 (('~flrp', a
, a
, b
), a
),
129 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
131 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
132 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
133 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
134 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
136 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
137 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
139 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
140 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
142 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
144 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
145 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
146 (('flrp@16', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp16'),
147 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
148 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
149 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
150 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
151 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
152 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
153 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
154 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
155 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
156 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
157 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
158 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
159 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
160 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
161 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
163 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
164 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
166 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
167 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
168 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
169 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
171 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
172 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
174 # (a * #b + #c) << #d
175 # ((a * #b) << #d) + (#c << #d)
176 # (a * (#b << #d)) + (#c << #d)
177 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
178 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
182 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
184 # Comparison simplifications
185 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
186 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
187 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
188 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
189 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
190 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
191 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
192 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
193 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
194 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
198 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
200 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
202 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
204 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
205 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
206 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
207 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
208 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
209 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
210 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
211 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
212 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
213 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
214 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
215 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
216 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
217 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
218 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
219 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
220 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
221 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
222 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
223 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
225 # -(b2f(a) + b2f(b)) < 0
226 # 0 < b2f(a) + b2f(b)
227 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
229 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
230 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
232 # -(b2f(a) + b2f(b)) >= 0
233 # 0 >= b2f(a) + b2f(b)
234 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
236 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
237 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
239 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
240 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
242 # Some optimizations (below) convert things like (a < b || c < b) into
243 # (min(a, c) < b). However, this interfers with the previous optimizations
244 # that try to remove comparisons with negated sums of b2f. This just
246 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
247 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
249 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
250 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
251 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
252 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
254 # Cannot remove the addition from ilt or ige due to overflow.
255 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
256 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
258 # fmin(-b2f(a), b) >= 0.0
259 # -b2f(a) >= 0.0 && b >= 0.0
260 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
261 # b2f(a) == 0.0 && b >= 0.0
262 # a == False && b >= 0.0
265 # The fge in the second replacement is not a typo. I leave the proof that
266 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
268 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
269 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
271 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
272 (('fne', ('b2f', 'a@1'), 0.0), a
),
273 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
274 (('ine', ('b2i', 'a@1'), 0), a
),
276 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
277 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
278 (('fge', ('u2f', a
), 0.0), True),
279 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
280 (('flt', ('u2f', a
), 0.0), False),
281 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
282 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
283 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
284 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
285 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
286 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
287 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
291 # fabs(a) != 0.0 because fabs(a) must be >= 0
293 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
297 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
300 # 0.0 == fabs(a) because fabs(a) must be >= 0
302 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
306 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
308 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
309 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
310 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
311 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
314 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
315 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
316 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
318 # Since b is a constant, constant folding will eliminate the fmin and the
319 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
320 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
322 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
324 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
325 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
326 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
327 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
328 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
329 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
330 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
331 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
332 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
333 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
334 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
335 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
336 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
337 (('bcsel', a
, True, b
), ('ior', a
, b
)),
338 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
339 (('bcsel', a
, b
, False), ('iand', a
, b
)),
340 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
347 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
348 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
349 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
350 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
351 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
352 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
353 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
354 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
355 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
356 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
357 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
358 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
359 (('fmin', a
, ('fabs', a
)), a
),
360 (('imin', a
, ('iabs', a
)), a
),
361 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
362 (('imax', a
, ('ineg', ('iabs', a
))), a
),
363 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
364 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
365 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
366 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
367 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
368 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
369 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
370 (('fsat', ('b2f', a
)), ('b2f', a
)),
371 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
372 (('fsat', ('fsat', a
)), ('fsat', a
)),
373 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
374 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
375 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
376 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
377 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
378 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
379 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
380 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
381 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
382 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
383 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
384 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
385 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
386 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
387 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
388 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
389 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
390 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
391 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
392 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
393 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
394 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
396 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
397 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
398 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
399 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
400 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
401 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
402 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
403 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
404 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
405 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
406 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
407 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
408 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
409 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
410 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
411 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
413 # Common pattern like 'if (i == 0 || i == 1 || ...)'
414 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
415 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
416 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
418 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
419 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
420 # so emit an open-coded version of that.
421 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
422 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
424 (('ior', a
, ('ieq', a
, False)), True),
425 (('ior', a
, ('inot', a
)), -1),
427 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
428 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
430 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
432 # These patterns can result when (a < b || a < c) => (a < min(b, c))
433 # transformations occur before constant propagation and loop-unrolling.
434 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
435 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
436 (('~fge', a
, ('fmin', b
, a
)), True),
437 (('~fge', ('fmax', a
, b
), a
), True),
438 (('~flt', a
, ('fmin', b
, a
)), False),
439 (('~flt', ('fmax', a
, b
), a
), False),
440 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
441 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
443 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
444 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
445 (('ige', a
, ('imin', b
, a
)), True),
446 (('ige', ('imax', a
, b
), a
), True),
447 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
448 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
449 (('uge', a
, ('umin', b
, a
)), True),
450 (('uge', ('umax', a
, b
), a
), True),
451 (('ilt', a
, ('imin', b
, a
)), False),
452 (('ilt', ('imax', a
, b
), a
), False),
453 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
454 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
455 (('ult', a
, ('umin', b
, a
)), False),
456 (('ult', ('umax', a
, b
), a
), False),
457 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
458 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
460 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
461 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
462 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
463 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
464 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
465 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
466 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
467 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
468 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
469 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
470 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
471 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
472 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
473 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
474 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
475 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
477 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
479 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
480 ('iabs', ('ishr', a
, b
))),
481 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
483 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
484 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
485 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
486 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
487 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
488 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
489 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
490 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
491 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
492 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
494 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
495 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
496 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
497 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
498 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
499 (('ineg', ('b2i32', 'a@32')), a
),
500 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
501 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
502 # Comparison with the same args. Note that these are not done for
503 # the float versions because NaN always returns false on float
505 (('ilt', a
, a
), False),
506 (('ige', a
, a
), True),
507 (('ieq', a
, a
), True),
508 (('ine', a
, a
), False),
509 (('ult', a
, a
), False),
510 (('uge', a
, a
), True),
511 # Logical and bit operations
512 (('fand', a
, 0.0), 0.0),
514 (('iand', a
, ~
0), a
),
518 (('ior', a
, True), True),
519 (('fxor', a
, a
), 0.0),
522 (('inot', ('inot', a
)), a
),
523 (('ior', ('iand', a
, b
), b
), b
),
524 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
525 (('iand', ('ior', a
, b
), b
), b
),
526 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
528 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
529 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
530 # Shift optimizations
537 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
538 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
539 # Exponential/logarithmic identities
540 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
541 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
542 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
543 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
544 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
545 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
546 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
547 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
548 (('~fpow', a
, 1.0), a
),
549 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
550 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
551 (('~fpow', 2.0, a
), ('fexp2', a
)),
552 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
553 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
554 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
555 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
556 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
557 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
558 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
559 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
560 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
561 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
562 # Division and reciprocal
563 (('~fdiv', 1.0, a
), ('frcp', a
)),
564 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
565 (('~frcp', ('frcp', a
)), a
),
566 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
567 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
568 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
569 # Boolean simplifications
570 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
571 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
572 (('ieq', a
, True), a
),
573 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
574 (('ine', a
, False), a
),
575 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
576 (('bcsel', a
, True, False), a
),
577 (('bcsel', a
, False, True), ('inot', a
)),
578 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
579 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
580 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
581 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
582 (('bcsel', True, b
, c
), b
),
583 (('bcsel', False, b
, c
), c
),
584 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
585 # The result of this should be hit by constant propagation and, in the
586 # next round of opt_algebraic, get picked up by one of the above two.
587 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
589 (('bcsel', a
, b
, b
), b
),
590 (('fcsel', a
, b
, b
), b
),
592 # D3D Boolean emulation
593 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
594 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
595 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
596 ('ineg', ('b2i', ('iand', a
, b
)))),
597 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
598 ('ineg', ('b2i', ('ior', a
, b
)))),
599 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
600 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
601 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
602 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
603 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
605 # SM5 32-bit shifts are defined to use the 5 least significant bits
606 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
607 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
608 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
611 (('i2b32', ('b2i', 'a@32')), a
),
612 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
613 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
614 (('i2b', ('ineg', a
)), ('i2b', a
)),
615 (('i2b', ('iabs', a
)), ('i2b', a
)),
616 (('fabs', ('b2f', a
)), ('b2f', a
)),
617 (('iabs', ('b2i', a
)), ('b2i', a
)),
618 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
620 # Ironically, mark these as imprecise because removing the conversions may
621 # preserve more precision than doing the conversions (e.g.,
622 # uint(float(0x81818181u)) == 0x81818200).
623 (('~f2i32', ('i2f', 'a@32')), a
),
624 (('~f2i32', ('u2f', 'a@32')), a
),
625 (('~f2u32', ('i2f', 'a@32')), a
),
626 (('~f2u32', ('u2f', 'a@32')), a
),
628 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
631 # It is undefined to convert a negative floating-point value to an
634 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
635 # some optimizations in the i965 backend to proceed.
636 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
637 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
638 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
639 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
641 # Packing and then unpacking does nothing
642 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
643 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
644 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
645 ('unpack_64_2x32_split_y', a
)), a
),
648 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
649 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
650 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
651 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
652 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
653 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
654 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
657 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
658 # patterns like those below.
659 for op
in ('ushr', 'ishr'):
660 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
661 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
662 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
664 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
666 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
667 # patterns like those below.
668 for op
in ('extract_u8', 'extract_i8'):
669 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
670 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
671 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
673 optimizations
.extend([
675 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
676 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
677 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
678 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
679 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
682 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
683 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
684 (('ussub_4x8', a
, 0), a
),
685 (('ussub_4x8', a
, ~
0), 0),
686 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
687 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
688 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
689 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
690 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
691 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
692 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
693 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
695 # Propagate negation up multiplication chains
696 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
697 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
699 # Propagate constants up multiplication chains
700 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
701 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
702 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
703 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
705 # Reassociate constants in add/mul chains so they can be folded together.
706 # For now, we mostly only handle cases where the constants are separated by
707 # a single non-constant. We could do better eventually.
708 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
709 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
710 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
711 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
712 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
715 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
716 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
717 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
719 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
720 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
721 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
723 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
726 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod16'),
727 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
728 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
729 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
730 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
731 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
733 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
734 ('bcsel', ('ilt', 31, 'bits'), 'insert',
735 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
736 'options->lower_bitfield_insert'),
737 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
738 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
739 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
740 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
741 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
742 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
744 # Alternative lowering that doesn't rely on bfi.
745 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
746 ('bcsel', ('ilt', 31, 'bits'),
749 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
750 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
751 'options->lower_bitfield_insert_to_shifts'),
753 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
754 (('bfm', 'bits', 'offset'),
755 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
756 'options->lower_bfm'),
758 (('ibitfield_extract', 'value', 'offset', 'bits'),
759 ('bcsel', ('ilt', 31, 'bits'), 'value',
760 ('ibfe', 'value', 'offset', 'bits')),
761 'options->lower_bitfield_extract'),
763 (('ubitfield_extract', 'value', 'offset', 'bits'),
764 ('bcsel', ('ult', 31, 'bits'), 'value',
765 ('ubfe', 'value', 'offset', 'bits')),
766 'options->lower_bitfield_extract'),
768 (('ibitfield_extract', 'value', 'offset', 'bits'),
769 ('bcsel', ('ieq', 0, 'bits'),
772 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
773 ('isub', 32, 'bits'))),
774 'options->lower_bitfield_extract_to_shifts'),
776 (('ubitfield_extract', 'value', 'offset', 'bits'),
778 ('ushr', 'value', 'offset'),
779 ('bcsel', ('ieq', 'bits', 32),
781 ('bfm', 'bits', 0))),
782 'options->lower_bitfield_extract_to_shifts'),
784 (('ifind_msb', 'value'),
785 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
786 'options->lower_ifind_msb'),
788 (('find_lsb', 'value'),
789 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
790 'options->lower_find_lsb'),
792 (('extract_i8', a
, 'b@32'),
793 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
794 'options->lower_extract_byte'),
796 (('extract_u8', a
, 'b@32'),
797 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
798 'options->lower_extract_byte'),
800 (('extract_i16', a
, 'b@32'),
801 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
802 'options->lower_extract_word'),
804 (('extract_u16', a
, 'b@32'),
805 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
806 'options->lower_extract_word'),
808 (('pack_unorm_2x16', 'v'),
809 ('pack_uvec2_to_uint',
810 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
811 'options->lower_pack_unorm_2x16'),
813 (('pack_unorm_4x8', 'v'),
814 ('pack_uvec4_to_uint',
815 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
816 'options->lower_pack_unorm_4x8'),
818 (('pack_snorm_2x16', 'v'),
819 ('pack_uvec2_to_uint',
820 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
821 'options->lower_pack_snorm_2x16'),
823 (('pack_snorm_4x8', 'v'),
824 ('pack_uvec4_to_uint',
825 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
826 'options->lower_pack_snorm_4x8'),
828 (('unpack_unorm_2x16', 'v'),
829 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
830 ('extract_u16', 'v', 1))),
832 'options->lower_unpack_unorm_2x16'),
834 (('unpack_unorm_4x8', 'v'),
835 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
836 ('extract_u8', 'v', 1),
837 ('extract_u8', 'v', 2),
838 ('extract_u8', 'v', 3))),
840 'options->lower_unpack_unorm_4x8'),
842 (('unpack_snorm_2x16', 'v'),
843 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
844 ('extract_i16', 'v', 1))),
846 'options->lower_unpack_snorm_2x16'),
848 (('unpack_snorm_4x8', 'v'),
849 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
850 ('extract_i8', 'v', 1),
851 ('extract_i8', 'v', 2),
852 ('extract_i8', 'v', 3))),
854 'options->lower_unpack_snorm_4x8'),
856 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
857 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
860 # bit_size dependent lowerings
861 for bit_size
in [8, 16, 32, 64]:
862 # convenience constants
863 intmax
= (1 << (bit_size
- 1)) - 1
864 intmin
= 1 << (bit_size
- 1)
867 (('iadd_sat@' + str(bit_size
), a
, b
),
868 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
869 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
870 (('isub_sat@' + str(bit_size
), a
, b
),
871 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
872 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
875 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
877 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
878 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
879 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
880 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
881 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
883 # Optimize x2bN(b2x(x)) -> x
884 for size
in type_sizes('bool'):
885 aN
= 'a@' + str(size
)
886 f2bN
= 'f2b' + str(size
)
887 i2bN
= 'i2b' + str(size
)
888 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
889 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
891 # Optimize x2yN(b2x(x)) -> b2y
892 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
893 if x
!= 'f' and y
!= 'f' and x
!= y
:
896 b2x
= 'b2f' if x
== 'f' else 'b2i'
897 b2y
= 'b2f' if y
== 'f' else 'b2i'
898 x2yN
= '{}2{}'.format(x
, y
)
899 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
901 # Optimize away x2xN(a@N)
902 for t
in ['int', 'uint', 'float']:
903 for N
in type_sizes(t
):
904 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
905 aN
= 'a@{0}'.format(N
)
906 optimizations
.append(((x2xN
, aN
), a
))
908 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
909 # In particular, we can optimize away everything except upcast of downcast and
910 # upcasts where the type differs from the other cast
911 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
913 # The outer cast is a down-cast. It doesn't matter what the size of the
914 # argument of the inner cast is because we'll never been in the upcast
915 # of downcast case. Regardless of types, we'll always end up with y2yN
917 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
918 x2xN
= '{0}2{0}{1}'.format(x
, N
)
919 y2yM
= '{0}2{0}{1}'.format(y
, M
)
920 y2yN
= '{0}2{0}{1}'.format(y
, N
)
921 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
923 # If the outer cast is an up-cast, we have to be more careful about the
924 # size of the argument of the inner cast and with types. In this case,
925 # the type is always the type of type up-cast which is given by the
927 for P
in type_sizes('uint'):
928 # We can't optimize away up-cast of down-cast.
932 # Because we're doing down-cast of down-cast, the types always have
933 # to match between the two casts
935 x2xN
= '{0}2{0}{1}'.format(x
, N
)
936 x2xM
= '{0}2{0}{1}'.format(x
, M
)
937 aP
= 'a@{0}'.format(P
)
938 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
940 # The N == M case is handled by other optimizations
943 def fexp2i(exp
, bits
):
944 # We assume that exp is already in the right range.
946 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
948 return ('ishl', ('iadd', exp
, 127), 23)
950 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
954 def ldexp(f
, exp
, bits
):
955 # First, we clamp exp to a reasonable range. The maximum possible range
956 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
957 # a maximum range of [-149, 127]. This means that we can potentially have
958 # a swing of +-276. If you start with FLT_MAX, you actually have to do
959 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
960 # spec, on the other hand, only requires that we handle an exponent value
961 # in the range [-126, 128]. This implementation is *mostly* correct; it
962 # handles a range on exp of [-252, 254] which allows you to create any
963 # value (including denorms if the hardware supports it) and to adjust the
964 # exponent of any normal value to anything you want.
966 exp
= ('imin', ('imax', exp
, -28), 30)
968 exp
= ('imin', ('imax', exp
, -252), 254)
970 exp
= ('imin', ('imax', exp
, -2044), 2046)
974 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
975 # (We use ishr which isn't the same for -1, but the -1 case still works
976 # since we use exp-exp/2 as the second exponent.) While the spec
977 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
978 # work with denormals and doesn't allow for the full swing in exponents
979 # that you can get with normalized values. Instead, we create two powers
980 # of two and multiply by them each in turn. That way the effective range
981 # of our exponent is doubled.
982 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
983 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
984 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
987 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
988 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
989 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
992 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
993 def bitfield_reverse(u
):
994 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
995 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
996 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
997 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
998 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1002 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1004 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1005 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1006 # and, if a is a NaN then the second comparison will fail anyway.
1007 for op
in ['flt', 'fge', 'feq']:
1009 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1010 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1013 # Add optimizations to handle the case where the result of a ternary is
1014 # compared to a constant. This way we can take things like
1020 # a ? (0 > 0) : (1 > 0)
1022 # which constant folding will eat for lunch. The resulting ternary will
1023 # further get cleaned up by the boolean reductions above and we will be
1024 # left with just the original variable "a".
1025 for op
in ['flt', 'fge', 'feq', 'fne',
1026 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1028 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1029 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1030 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1031 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1035 # For example, this converts things like
1037 # 1 + mix(0, a - 1, condition)
1041 # mix(1, (a-1)+1, condition)
1043 # Other optimizations will rearrange the constants.
1044 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1046 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1049 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1052 # If neither layout qualifier is specified, derivatives in compute shaders
1053 # return zero, which is consistent with the handling of built-in texture
1054 # functions like texture() in GLSL 4.50 compute shaders.
1055 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1056 'fddy', 'fddy_fine', 'fddy_coarse']:
1058 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1061 # This section contains "late" optimizations that should be run before
1062 # creating ffmas and calling regular optimizations for the final time.
1063 # Optimizations should go here if they help code generation and conflict
1064 # with the regular optimizations.
1065 before_ffma_optimizations
= [
1066 # Propagate constants down multiplication chains
1067 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1068 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1069 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1070 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1072 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1073 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1074 (('~fadd', ('fneg', a
), a
), 0.0),
1075 (('iadd', ('ineg', a
), a
), 0),
1076 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1077 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1078 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1079 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1081 (('~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
)),
1082 (('~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
)),
1083 (('~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
))),
1086 # This section contains "late" optimizations that should be run after the
1087 # regular optimizations have finished. Optimizations should go here if
1088 # they help code generation but do not necessarily produce code that is
1089 # more easily optimizable.
1090 late_optimizations
= [
1091 # Most of these optimizations aren't quite safe when you get infinity or
1092 # Nan involved but the first one should be fine.
1093 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1094 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1095 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1096 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1097 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1098 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1100 (('~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
)))),
1102 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1103 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1104 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1105 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1107 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1108 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1110 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a
, 0.0, 1.0)),
1111 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a
, -0.0, -1.0)),
1113 # we do these late so that we don't get in the way of creating ffmas
1114 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1115 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1117 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1120 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1121 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1122 before_ffma_optimizations
).render())
1123 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1124 late_optimizations
).render())
projects / mesa.git / blob