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)][.swiz]" where:
54 # "#" indicates that the given variable will only match constants,
55 # type indicates that the given variable will only match values from ALU
56 # instructions with the given output type,
57 # (cond) specifies an additional condition function (see nir_search_helpers.h),
58 # swiz is a swizzle applied to the variable (only in the <replace> expression)
60 # For constants, you have to be careful to make sure that it is the right
61 # type because python is unaware of the source and destination types of the
64 # All expression types can have a bit-size specified. For opcodes, this
65 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
66 # type and size. In the search half of the expression this indicates that it
67 # should only match that particular bit-size. In the replace half of the
68 # expression this indicates that the constructed value should have that
71 # A special condition "many-comm-expr" can be used with expressions to note
72 # that the expression and its subexpressions have more commutative expressions
73 # than nir_replace_instr can handle. If this special condition is needed with
74 # another condition, the two can be separated by a comma (e.g.,
75 # "(many-comm-expr,is_used_once)").
79 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
80 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitshift'),
81 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitshift'),
83 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
84 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
85 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
86 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
91 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
)), '!options->lower_bitshift'),
92 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
93 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
94 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
96 (('fneg', ('fneg', a
)), a
),
97 (('ineg', ('ineg', a
)), a
),
98 (('fabs', ('fabs', a
)), ('fabs', a
)),
99 (('fabs', ('fneg', a
)), ('fabs', a
)),
100 (('fabs', ('u2f', a
)), ('u2f', a
)),
101 (('iabs', ('iabs', a
)), ('iabs', a
)),
102 (('iabs', ('ineg', a
)), ('iabs', a
)),
103 (('f2b', ('fneg', a
)), ('f2b', a
)),
104 (('i2b', ('ineg', a
)), ('i2b', a
)),
105 (('~fadd', a
, 0.0), a
),
107 (('usadd_4x8', a
, 0), a
),
108 (('usadd_4x8', a
, ~
0), ~
0),
109 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
110 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
111 (('~fadd', ('fneg', a
), a
), 0.0),
112 (('iadd', ('ineg', a
), a
), 0),
113 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
114 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
115 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
116 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
117 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
118 (('~fmul', a
, 0.0), 0.0),
120 (('umul_unorm_4x8', a
, 0), 0),
121 (('umul_unorm_4x8', a
, ~
0), a
),
122 (('fmul', a
, 1.0), a
),
124 (('fmul', a
, -1.0), ('fneg', a
)),
125 (('imul', a
, -1), ('ineg', 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 => 1*a*a => a*a => abs(a)*a
128 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
129 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
130 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
131 (('~ffma', 0.0, a
, b
), b
),
132 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
133 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
134 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
135 (('~flrp', a
, b
, 0.0), a
),
136 (('~flrp', a
, b
, 1.0), b
),
137 (('~flrp', a
, a
, b
), a
),
138 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
140 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
141 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
142 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
143 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
145 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
146 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
148 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
149 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
151 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
153 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
154 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
155 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
156 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
157 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
158 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
159 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
160 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
161 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
162 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
163 # These are the same as the previous three rules, but it depends on
164 # 1-fsat(x) <=> fsat(1-x). See below.
165 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
166 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
168 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
169 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
170 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
171 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
172 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
174 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
175 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
177 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
178 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
179 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
180 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
182 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
183 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
185 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
186 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
187 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
188 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
190 # 1 - ((1 - a) * (1 - b))
191 # 1 - (1 - a - b + a*b)
192 # 1 - 1 + a + b - a*b
197 (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a
)), ('fadd', 1.0, ('fneg', b
))))),
198 ('flrp', b
, 1.0, a
), '!options->lower_flrp32'),
200 # (a * #b + #c) << #d
201 # ((a * #b) << #d) + (#c << #d)
202 # (a * (#b << #d)) + (#c << #d)
203 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
204 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
208 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
210 # Comparison simplifications
211 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
212 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
213 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
214 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
215 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
216 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
217 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
218 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
219 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
220 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
222 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
223 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
224 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
225 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
226 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
228 # This helps some shaders because, after some optimizations, they end up
229 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
230 # matching would be handled by CSE.
231 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
232 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
233 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
234 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
235 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
236 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
237 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
238 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
239 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
240 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
242 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
243 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
244 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
245 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
246 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
247 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
249 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
250 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
251 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
252 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
256 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
258 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
260 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
262 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
263 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
264 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
265 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
266 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
267 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
268 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
269 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
270 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
271 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
272 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
273 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
274 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
275 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
276 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
277 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
278 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
279 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
280 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
281 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', 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 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
288 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
290 # -(b2f(a) + b2f(b)) >= 0
291 # 0 >= b2f(a) + b2f(b)
292 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
294 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
295 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
297 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
298 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
300 # Some optimizations (below) convert things like (a < b || c < b) into
301 # (min(a, c) < b). However, this interfers with the previous optimizations
302 # that try to remove comparisons with negated sums of b2f. This just
304 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
305 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
307 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
308 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
309 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
310 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
312 # Cannot remove the addition from ilt or ige due to overflow.
313 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
314 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
316 # fmin(-b2f(a), b) >= 0.0
317 # -b2f(a) >= 0.0 && b >= 0.0
318 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
319 # b2f(a) == 0.0 && b >= 0.0
320 # a == False && b >= 0.0
323 # The fge in the second replacement is not a typo. I leave the proof that
324 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
326 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
327 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
329 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
330 (('fne', ('b2f', 'a@1'), 0.0), a
),
331 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
332 (('ine', ('b2i', 'a@1'), 0), a
),
334 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
335 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
336 (('fge', ('u2f', a
), 0.0), True),
337 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
338 (('flt', ('u2f', a
), 0.0), False),
339 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
340 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
341 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
342 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
343 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
344 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
345 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
349 # fabs(a) != 0.0 because fabs(a) must be >= 0
351 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
355 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
358 # 0.0 == fabs(a) because fabs(a) must be >= 0
360 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
364 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
366 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
367 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
368 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
369 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
372 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
373 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
374 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
376 # Since b is a constant, constant folding will eliminate the fmin and the
377 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
378 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
380 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
382 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
383 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
384 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
385 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
386 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
387 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
388 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
389 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
390 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
391 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
392 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
393 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
394 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
395 (('bcsel', a
, True, b
), ('ior', a
, b
)),
396 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
397 (('bcsel', a
, b
, False), ('iand', a
, b
)),
398 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
405 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
406 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
407 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
408 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
409 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
410 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
411 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
412 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
413 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
414 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
415 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
416 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
417 (('fmin', a
, ('fabs', a
)), a
),
418 (('imin', a
, ('iabs', a
)), a
),
419 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
420 (('imax', a
, ('ineg', ('iabs', a
))), a
),
421 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
422 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
423 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
424 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
425 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
426 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
427 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
428 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
429 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
430 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
431 (('fsat', ('b2f', a
)), ('b2f', a
)),
432 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
433 (('fsat', ('fsat', a
)), ('fsat', a
)),
434 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_negate && !options->lower_fsat'),
435 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_negate && !options->lower_fsat'),
436 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
437 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
438 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
439 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
440 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
441 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
442 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
443 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
444 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
445 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
446 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
447 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
448 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
449 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
450 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
451 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
452 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
453 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
454 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
455 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
456 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
457 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
458 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
460 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
461 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
462 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
463 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
464 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
465 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
466 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
467 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
468 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
469 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
470 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
471 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
472 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
473 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
474 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
475 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
477 # Common pattern like 'if (i == 0 || i == 1 || ...)'
478 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
479 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
480 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
482 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
483 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
484 # so emit an open-coded version of that.
485 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
486 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
488 (('ior', a
, ('ieq', a
, False)), True),
489 (('ior', a
, ('inot', a
)), -1),
491 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
492 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
494 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
496 # These patterns can result when (a < b || a < c) => (a < min(b, c))
497 # transformations occur before constant propagation and loop-unrolling.
498 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
499 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
500 (('~fge', a
, ('fmin', b
, a
)), True),
501 (('~fge', ('fmax', a
, b
), a
), True),
502 (('~flt', a
, ('fmin', b
, a
)), False),
503 (('~flt', ('fmax', a
, b
), a
), False),
504 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
505 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
507 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
508 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
509 (('ige', a
, ('imin', b
, a
)), True),
510 (('ige', ('imax', a
, b
), a
), True),
511 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
512 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
513 (('uge', a
, ('umin', b
, a
)), True),
514 (('uge', ('umax', a
, b
), a
), True),
515 (('ilt', a
, ('imin', b
, a
)), False),
516 (('ilt', ('imax', a
, b
), a
), False),
517 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
518 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
519 (('ult', a
, ('umin', b
, a
)), False),
520 (('ult', ('umax', a
, b
), a
), False),
521 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
522 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
523 (('ult', a
, ('iand', b
, a
)), False),
524 (('ult', ('ior', a
, b
), a
), False),
525 (('uge', a
, ('iand', b
, a
)), True),
526 (('uge', ('ior', a
, b
), a
), True),
528 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
529 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
530 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
531 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
532 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
533 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
534 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
535 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
536 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
537 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
538 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
539 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
540 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
541 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
542 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
543 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
545 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
547 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
548 ('iabs', ('ishr', a
, b
))),
549 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
551 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
552 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
553 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
554 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
555 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
556 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
557 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
558 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
559 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
560 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
562 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
563 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
564 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
565 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
566 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
567 (('ineg', ('b2i32', 'a@32')), a
),
568 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
569 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
570 # Comparison with the same args. Note that these are not done for
571 # the float versions because NaN always returns false on float
573 (('ilt', a
, a
), False),
574 (('ige', a
, a
), True),
575 (('ieq', a
, a
), True),
576 (('ine', a
, a
), False),
577 (('ult', a
, a
), False),
578 (('uge', a
, a
), True),
579 # Logical and bit operations
581 (('iand', a
, ~
0), a
),
585 (('ior', a
, True), True),
588 (('inot', ('inot', a
)), a
),
589 (('ior', ('iand', a
, b
), b
), b
),
590 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
591 (('iand', ('ior', a
, b
), b
), b
),
592 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
594 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
595 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
596 # Shift optimizations
603 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
604 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
605 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
606 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
607 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
608 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
609 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
610 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
611 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
612 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
613 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
614 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
615 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
616 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
617 # Exponential/logarithmic identities
618 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
619 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
620 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
621 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
622 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
623 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
624 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
625 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
626 (('~fpow', a
, 1.0), a
),
627 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
628 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
629 (('~fpow', 2.0, a
), ('fexp2', a
)),
630 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
631 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
632 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
633 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
634 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
635 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
636 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
637 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
638 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
639 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
640 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
641 # Division and reciprocal
642 (('~fdiv', 1.0, a
), ('frcp', a
)),
643 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
644 (('~frcp', ('frcp', a
)), a
),
645 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
646 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
647 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
648 # Boolean simplifications
649 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
650 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
651 (('ieq', a
, True), a
),
652 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
653 (('ine', a
, False), a
),
654 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
655 (('bcsel', a
, True, False), a
),
656 (('bcsel', a
, False, True), ('inot', a
)),
657 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
658 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
659 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
660 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
661 (('bcsel', True, b
, c
), b
),
662 (('bcsel', False, b
, c
), c
),
663 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
665 (('bcsel', a
, b
, b
), b
),
666 (('fcsel', a
, b
, b
), b
),
668 # D3D Boolean emulation
669 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
670 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
671 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
672 ('ineg', ('b2i', ('iand', a
, b
)))),
673 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
674 ('ineg', ('b2i', ('ior', a
, b
)))),
675 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
676 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
677 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
678 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
679 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
681 # SM5 32-bit shifts are defined to use the 5 least significant bits
682 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
683 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
684 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
687 (('i2b32', ('b2i', 'a@32')), a
),
688 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
689 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
690 (('i2b', ('ineg', a
)), ('i2b', a
)),
691 (('i2b', ('iabs', a
)), ('i2b', a
)),
692 (('fabs', ('b2f', a
)), ('b2f', a
)),
693 (('iabs', ('b2i', a
)), ('b2i', a
)),
694 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
696 # Ironically, mark these as imprecise because removing the conversions may
697 # preserve more precision than doing the conversions (e.g.,
698 # uint(float(0x81818181u)) == 0x81818200).
699 (('~f2i32', ('i2f', 'a@32')), a
),
700 (('~f2i32', ('u2f', 'a@32')), a
),
701 (('~f2u32', ('i2f', 'a@32')), a
),
702 (('~f2u32', ('u2f', 'a@32')), a
),
704 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
707 # It is undefined to convert a negative floating-point value to an
710 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
711 # some optimizations in the i965 backend to proceed.
712 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
713 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
714 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
715 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
717 (('~fmin', ('fabs', a
), 1.0), ('fsat', ('fabs', a
)), '!options->lower_fsat'),
719 # The result of the multiply must be in [-1, 0], so the result of the ffma
721 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
722 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
723 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
724 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
726 # Packing and then unpacking does nothing
727 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
728 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
729 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
730 ('unpack_64_2x32_split_y', a
)), a
),
732 # Comparing two halves of an unpack separately. While this optimization
733 # should be correct for non-constant values, it's less obvious that it's
734 # useful in that case. For constant values, the pack will fold and we're
735 # guaranteed to reduce the whole tree to one instruction.
736 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
737 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
738 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
741 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
742 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
743 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
744 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
745 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
746 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
747 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte')
750 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
751 # patterns like those below.
752 for op
in ('ushr', 'ishr'):
753 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
754 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
755 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
757 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
759 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
760 # patterns like those below.
761 for op
in ('extract_u8', 'extract_i8'):
762 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
763 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
764 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
766 optimizations
.extend([
768 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
769 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
770 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
771 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
772 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
775 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
776 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
777 (('ussub_4x8', a
, 0), a
),
778 (('ussub_4x8', a
, ~
0), 0),
779 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
780 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
781 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
782 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
783 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
784 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
785 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
786 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
788 # Propagate negation up multiplication chains
789 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
790 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
792 # Propagate constants up multiplication chains
793 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
794 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
795 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
796 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
798 # Reassociate constants in add/mul chains so they can be folded together.
799 # For now, we mostly only handle cases where the constants are separated by
800 # a single non-constant. We could do better eventually.
801 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
802 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
803 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
804 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
805 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
807 # Drop mul-div by the same value when there's no wrapping.
808 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
811 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
812 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
813 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
815 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
816 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
817 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
819 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
822 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
823 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
824 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
825 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
826 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
828 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
829 ('bcsel', ('ult', 31, 'bits'), 'insert',
830 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
831 'options->lower_bitfield_insert'),
832 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
833 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
834 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
835 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
836 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
837 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
839 # Alternative lowering that doesn't rely on bfi.
840 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
841 ('bcsel', ('ult', 31, 'bits'),
844 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
845 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
846 'options->lower_bitfield_insert_to_shifts'),
848 # Alternative lowering that uses bitfield_select.
849 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
850 ('bcsel', ('ult', 31, 'bits'), 'insert',
851 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
852 'options->lower_bitfield_insert_to_bitfield_select'),
854 (('ibitfield_extract', 'value', 'offset', 'bits'),
855 ('bcsel', ('ult', 31, 'bits'), 'value',
856 ('ibfe', 'value', 'offset', 'bits')),
857 'options->lower_bitfield_extract'),
859 (('ubitfield_extract', 'value', 'offset', 'bits'),
860 ('bcsel', ('ult', 31, 'bits'), 'value',
861 ('ubfe', 'value', 'offset', 'bits')),
862 'options->lower_bitfield_extract'),
864 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
865 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
866 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
867 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
868 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
869 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
870 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
872 (('ibitfield_extract', 'value', 'offset', 'bits'),
873 ('bcsel', ('ieq', 0, 'bits'),
876 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
877 ('isub', 32, 'bits'))),
878 'options->lower_bitfield_extract_to_shifts'),
880 (('ubitfield_extract', 'value', 'offset', 'bits'),
882 ('ushr', 'value', 'offset'),
883 ('bcsel', ('ieq', 'bits', 32),
885 ('isub', ('ishl', 1, 'bits'), 1))),
886 'options->lower_bitfield_extract_to_shifts'),
888 (('ifind_msb', 'value'),
889 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
890 'options->lower_ifind_msb'),
892 (('find_lsb', 'value'),
893 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
894 'options->lower_find_lsb'),
896 (('extract_i8', a
, 'b@32'),
897 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
898 'options->lower_extract_byte'),
900 (('extract_u8', a
, 'b@32'),
901 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
902 'options->lower_extract_byte'),
904 (('extract_i16', a
, 'b@32'),
905 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
906 'options->lower_extract_word'),
908 (('extract_u16', a
, 'b@32'),
909 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
910 'options->lower_extract_word'),
912 (('pack_unorm_2x16', 'v'),
913 ('pack_uvec2_to_uint',
914 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
915 'options->lower_pack_unorm_2x16'),
917 (('pack_unorm_4x8', 'v'),
918 ('pack_uvec4_to_uint',
919 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
920 'options->lower_pack_unorm_4x8'),
922 (('pack_snorm_2x16', 'v'),
923 ('pack_uvec2_to_uint',
924 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
925 'options->lower_pack_snorm_2x16'),
927 (('pack_snorm_4x8', 'v'),
928 ('pack_uvec4_to_uint',
929 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
930 'options->lower_pack_snorm_4x8'),
932 (('unpack_unorm_2x16', 'v'),
933 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
934 ('extract_u16', 'v', 1))),
936 'options->lower_unpack_unorm_2x16'),
938 (('unpack_unorm_4x8', 'v'),
939 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
940 ('extract_u8', 'v', 1),
941 ('extract_u8', 'v', 2),
942 ('extract_u8', 'v', 3))),
944 'options->lower_unpack_unorm_4x8'),
946 (('unpack_snorm_2x16', 'v'),
947 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
948 ('extract_i16', 'v', 1))),
950 'options->lower_unpack_snorm_2x16'),
952 (('unpack_snorm_4x8', 'v'),
953 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
954 ('extract_i8', 'v', 1),
955 ('extract_i8', 'v', 2),
956 ('extract_i8', 'v', 3))),
958 'options->lower_unpack_snorm_4x8'),
960 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
961 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
964 # bit_size dependent lowerings
965 for bit_size
in [8, 16, 32, 64]:
966 # convenience constants
967 intmax
= (1 << (bit_size
- 1)) - 1
968 intmin
= 1 << (bit_size
- 1)
971 (('iadd_sat@' + str(bit_size
), a
, b
),
972 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
973 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
974 (('isub_sat@' + str(bit_size
), a
, b
),
975 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
976 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
979 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
981 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
982 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
983 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
984 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
985 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
987 # Optimize x2bN(b2x(x)) -> x
988 for size
in type_sizes('bool'):
989 aN
= 'a@' + str(size
)
990 f2bN
= 'f2b' + str(size
)
991 i2bN
= 'i2b' + str(size
)
992 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
993 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
995 # Optimize x2yN(b2x(x)) -> b2y
996 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
997 if x
!= 'f' and y
!= 'f' and x
!= y
:
1000 b2x
= 'b2f' if x
== 'f' else 'b2i'
1001 b2y
= 'b2f' if y
== 'f' else 'b2i'
1002 x2yN
= '{}2{}'.format(x
, y
)
1003 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1005 # Optimize away x2xN(a@N)
1006 for t
in ['int', 'uint', 'float']:
1007 for N
in type_sizes(t
):
1008 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1009 aN
= 'a@{0}'.format(N
)
1010 optimizations
.append(((x2xN
, aN
), a
))
1012 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1013 # In particular, we can optimize away everything except upcast of downcast and
1014 # upcasts where the type differs from the other cast
1015 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1017 # The outer cast is a down-cast. It doesn't matter what the size of the
1018 # argument of the inner cast is because we'll never been in the upcast
1019 # of downcast case. Regardless of types, we'll always end up with y2yN
1021 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1022 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1023 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1024 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1025 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1027 # If the outer cast is an up-cast, we have to be more careful about the
1028 # size of the argument of the inner cast and with types. In this case,
1029 # the type is always the type of type up-cast which is given by the
1031 for P
in type_sizes('uint'):
1032 # We can't optimize away up-cast of down-cast.
1036 # Because we're doing down-cast of down-cast, the types always have
1037 # to match between the two casts
1038 for x
in ['i', 'u']:
1039 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1040 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1041 aP
= 'a@{0}'.format(P
)
1042 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1044 # The N == M case is handled by other optimizations
1047 # Optimize comparisons with up-casts
1048 for t
in ['int', 'uint', 'float']:
1049 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1050 if N
== 1 or N
>= M
:
1053 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1054 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1057 xeq
= 'feq' if t
== 'float' else 'ieq'
1058 xne
= 'fne' if t
== 'float' else 'ine'
1059 xge
= '{0}ge'.format(t
[0])
1060 xlt
= '{0}lt'.format(t
[0])
1062 # Up-casts are lossless so for correctly signed comparisons of
1063 # up-casted values we can do the comparison at the largest of the two
1064 # original sizes and drop one or both of the casts. (We have
1065 # optimizations to drop the no-op casts which this may generate.)
1066 for P
in type_sizes(t
):
1072 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1073 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1074 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1075 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1076 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1077 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1080 # The next bit doesn't work on floats because the range checks would
1081 # get way too complicated.
1082 if t
in ['int', 'uint']:
1084 xN_min
= -(1 << (N
- 1))
1085 xN_max
= (1 << (N
- 1)) - 1
1088 xN_max
= (1 << N
) - 1
1092 # If we're up-casting and comparing to a constant, we can unfold
1093 # the comparison into a comparison with the shrunk down constant
1094 # and a check that the constant fits in the smaller bit size.
1096 ((xeq
, (x2xM
, aN
), '#b'),
1097 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1098 ((xne
, (x2xM
, aN
), '#b'),
1099 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1100 ((xlt
, (x2xM
, aN
), '#b'),
1101 ('iand', (xlt
, xN_min
, b
),
1102 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1103 ((xlt
, '#a', (x2xM
, bN
)),
1104 ('iand', (xlt
, a
, xN_max
),
1105 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1106 ((xge
, (x2xM
, aN
), '#b'),
1107 ('iand', (xge
, xN_max
, b
),
1108 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1109 ((xge
, '#a', (x2xM
, bN
)),
1110 ('iand', (xge
, a
, xN_min
),
1111 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1114 def fexp2i(exp
, bits
):
1115 # We assume that exp is already in the right range.
1117 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1119 return ('ishl', ('iadd', exp
, 127), 23)
1121 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1125 def ldexp(f
, exp
, bits
):
1126 # First, we clamp exp to a reasonable range. The maximum possible range
1127 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1128 # a maximum range of [-149, 127]. This means that we can potentially have
1129 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1130 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1131 # spec, on the other hand, only requires that we handle an exponent value
1132 # in the range [-126, 128]. This implementation is *mostly* correct; it
1133 # handles a range on exp of [-252, 254] which allows you to create any
1134 # value (including denorms if the hardware supports it) and to adjust the
1135 # exponent of any normal value to anything you want.
1137 exp
= ('imin', ('imax', exp
, -28), 30)
1139 exp
= ('imin', ('imax', exp
, -252), 254)
1141 exp
= ('imin', ('imax', exp
, -2044), 2046)
1145 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1146 # (We use ishr which isn't the same for -1, but the -1 case still works
1147 # since we use exp-exp/2 as the second exponent.) While the spec
1148 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1149 # work with denormals and doesn't allow for the full swing in exponents
1150 # that you can get with normalized values. Instead, we create two powers
1151 # of two and multiply by them each in turn. That way the effective range
1152 # of our exponent is doubled.
1153 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1154 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1155 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1158 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1159 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1160 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1163 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1164 def bitfield_reverse(u
):
1165 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1166 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1167 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1168 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1169 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1173 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1175 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1176 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1177 # and, if a is a NaN then the second comparison will fail anyway.
1178 for op
in ['flt', 'fge', 'feq']:
1180 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1181 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1184 # Add optimizations to handle the case where the result of a ternary is
1185 # compared to a constant. This way we can take things like
1191 # a ? (0 > 0) : (1 > 0)
1193 # which constant folding will eat for lunch. The resulting ternary will
1194 # further get cleaned up by the boolean reductions above and we will be
1195 # left with just the original variable "a".
1196 for op
in ['flt', 'fge', 'feq', 'fne',
1197 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1199 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1200 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1201 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1202 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1206 # For example, this converts things like
1208 # 1 + mix(0, a - 1, condition)
1212 # mix(1, (a-1)+1, condition)
1214 # Other optimizations will rearrange the constants.
1215 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1217 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1220 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1223 # If neither layout qualifier is specified, derivatives in compute shaders
1224 # return zero, which is consistent with the handling of built-in texture
1225 # functions like texture() in GLSL 4.50 compute shaders.
1226 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1227 'fddy', 'fddy_fine', 'fddy_coarse']:
1229 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1232 # Some optimizations for ir3-specific instructions.
1234 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1235 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1236 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1237 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1238 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1241 # This section contains "late" optimizations that should be run before
1242 # creating ffmas and calling regular optimizations for the final time.
1243 # Optimizations should go here if they help code generation and conflict
1244 # with the regular optimizations.
1245 before_ffma_optimizations
= [
1246 # Propagate constants down multiplication chains
1247 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1248 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1249 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1250 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1252 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1253 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1254 (('~fadd', ('fneg', a
), a
), 0.0),
1255 (('iadd', ('ineg', a
), a
), 0),
1256 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1257 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1258 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1259 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1261 (('~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
)),
1262 (('~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
)),
1263 (('~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
))),
1266 # This section contains "late" optimizations that should be run after the
1267 # regular optimizations have finished. Optimizations should go here if
1268 # they help code generation but do not necessarily produce code that is
1269 # more easily optimizable.
1270 late_optimizations
= [
1271 # Most of these optimizations aren't quite safe when you get infinity or
1272 # Nan involved but the first one should be fine.
1273 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1274 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1275 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1276 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1277 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1278 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1280 # nir_lower_to_source_mods will collapse this, but its existence during the
1281 # optimization loop can prevent other optimizations.
1282 (('fneg', ('fneg', a
)), a
),
1284 # These are duplicated from the main optimizations table. The late
1285 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1286 # new patterns like these. The patterns that compare with zero are removed
1287 # because they are unlikely to be created in by anything in
1288 # late_optimizations.
1289 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1290 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1291 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1292 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1293 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1294 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1296 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1297 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1299 (('~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
)))),
1301 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1302 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1303 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1304 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1305 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1306 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1307 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1308 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1309 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1310 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1313 (('iand', a
, a
), a
),
1315 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1317 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1318 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1319 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1320 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1322 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1323 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1325 (('~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'),
1326 (('~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'),
1328 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1329 # particular operation is common for expanding values stored in a texture
1330 # from [0,1] to [-1,1].
1331 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1332 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1333 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1334 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1335 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1336 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1337 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1338 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1342 # a + -a*a + a*b (1)
1344 # Option 1: ffma(a, (b-a), a)
1346 # Alternately, after (1):
1352 # Option 2: ffma(a, 2, -(a*a))
1353 # Option 3: ffma(a, 2, (-a)*a)
1354 # Option 4: ffma(a, -a, (2*a)
1355 # Option 5: a * (2 - a)
1357 # There are a lot of other possible combinations.
1358 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1359 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1360 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1361 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1362 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1364 # we do these late so that we don't get in the way of creating ffmas
1365 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1366 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1368 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1370 # Things that look like DPH in the source shader may get expanded to
1371 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1372 # to NIR. After FFMA is generated, this can look like:
1374 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1376 # Reassociate the last addition into the first multiplication.
1377 (('~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)'),
1378 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '!options->intel_vec4'),
1379 (('~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)'),
1380 ('ffma', a
, b
, ('ffma', e
, 'f', 'g') ), '!options->intel_vec4'),
1383 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1384 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1385 before_ffma_optimizations
).render())
1386 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1387 late_optimizations
).render())
projects / mesa.git / blob