2 # Copyright (C) 2014 Intel Corporation
4 # Permission is hereby granted, free of charge, to any person obtaining a
5 # copy of this software and associated documentation files (the "Software"),
6 # to deal in the Software without restriction, including without limitation
7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 # and/or sell copies of the Software, and to permit persons to whom the
9 # Software is furnished to do so, subject to the following conditions:
11 # The above copyright notice and this permission notice (including the next
12 # paragraph) shall be included in all copies or substantial portions of the
15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 # Jason Ekstrand (jason@jlekstrand.net)
26 from __future__
import print_function
28 from collections
import OrderedDict
30 from nir_opcodes
import type_sizes
34 # Convenience variables
41 # Written in the form (<search>, <replace>) where <search> is an expression
42 # and <replace> is either an expression or a value. An expression is
43 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
44 # where each source is either an expression or a value. A value can be
45 # either a numeric constant or a string representing a variable name.
47 # If the opcode in a search expression is prefixed by a '~' character, this
48 # indicates that the operation is inexact. Such operations will only get
49 # applied to SSA values that do not have the exact bit set. This should be
50 # used by by any optimizations that are not bit-for-bit exact. It should not,
51 # however, be used for backend-requested lowering operations as those need to
52 # happen regardless of precision.
54 # Variable names are specified as "[#]name[@type][(cond)][.swiz]" where:
55 # "#" indicates that the given variable will only match constants,
56 # type indicates that the given variable will only match values from ALU
57 # instructions with the given output type,
58 # (cond) specifies an additional condition function (see nir_search_helpers.h),
59 # swiz is a swizzle applied to the variable (only in the <replace> expression)
61 # For constants, you have to be careful to make sure that it is the right
62 # type because python is unaware of the source and destination types of the
65 # All expression types can have a bit-size specified. For opcodes, this
66 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
67 # type and size. In the search half of the expression this indicates that it
68 # should only match that particular bit-size. In the replace half of the
69 # expression this indicates that the constructed value should have that
72 # A special condition "many-comm-expr" can be used with expressions to note
73 # that the expression and its subexpressions have more commutative expressions
74 # than nir_replace_instr can handle. If this special condition is needed with
75 # another condition, the two can be separated by a comma (e.g.,
76 # "(many-comm-expr,is_used_once)").
78 # based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
79 def lowered_sincos(c
):
80 x
= ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi
, a
), c
))), 1.0)
81 x
= ('fmul', ('fsub', x
, ('fmul', x
, ('fabs', x
))), 4.0)
82 return ('ffma', ('ffma', x
, ('fabs', x
), ('fneg', x
)), 0.225, x
)
86 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitops'),
87 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
88 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitops'),
90 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
91 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
92 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
93 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
98 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
)), '!options->lower_bitops'),
99 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
100 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
101 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
103 (('~fneg', ('fneg', a
)), a
),
104 (('ineg', ('ineg', a
)), a
),
105 (('fabs', ('fabs', a
)), ('fabs', a
)),
106 (('fabs', ('fneg', a
)), ('fabs', a
)),
107 (('fabs', ('u2f', a
)), ('u2f', a
)),
108 (('iabs', ('iabs', a
)), ('iabs', a
)),
109 (('iabs', ('ineg', a
)), ('iabs', a
)),
110 (('f2b', ('fneg', a
)), ('f2b', a
)),
111 (('i2b', ('ineg', a
)), ('i2b', a
)),
112 (('~fadd', a
, 0.0), a
),
114 (('usadd_4x8', a
, 0), a
),
115 (('usadd_4x8', a
, ~
0), ~
0),
116 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
117 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
118 (('~fadd', ('fneg', a
), a
), 0.0),
119 (('iadd', ('ineg', a
), a
), 0),
120 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
121 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
122 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
123 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
124 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
125 (('~fmul', a
, 0.0), 0.0),
127 (('umul_unorm_4x8', a
, 0), 0),
128 (('umul_unorm_4x8', a
, ~
0), a
),
129 (('~fmul', a
, 1.0), a
),
131 (('fmul', a
, -1.0), ('fneg', a
)),
132 (('imul', a
, -1), ('ineg', a
)),
133 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
134 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
135 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
136 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
137 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
138 (('~ffma', 0.0, a
, b
), b
),
139 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
140 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
141 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
142 (('~flrp', a
, b
, 0.0), a
),
143 (('~flrp', a
, b
, 1.0), b
),
144 (('~flrp', a
, a
, b
), a
),
145 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
147 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
148 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
149 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
150 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
152 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
153 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
155 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
156 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
158 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
160 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
161 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
162 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
163 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
164 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
165 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
166 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
167 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
168 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
169 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
170 # These are the same as the previous three rules, but it depends on
171 # 1-fsat(x) <=> fsat(1-x). See below.
172 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
173 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
175 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
176 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
177 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
178 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
179 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
181 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
182 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
184 (('fdph', a
, b
), ('fdot4', ('vec4', 'a.x', 'a.y', 'a.z', 1.0), b
), 'options->lower_fdph'),
186 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
), '!options->lower_fdph'),
187 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
188 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
189 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
191 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
192 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
194 (('fdot2', ('vec2', a
, 0.0), b
), ('fmul', a
, b
)),
195 (('fdot2', a
, 1.0), ('fadd', 'a.x', 'a.y')),
197 # Lower fdot to fsum when it is available
198 (('fdot2', a
, b
), ('fsum2', ('fmul', a
, b
)), 'options->lower_fdot'),
199 (('fdot3', a
, b
), ('fsum3', ('fmul', a
, b
)), 'options->lower_fdot'),
200 (('fdot4', a
, b
), ('fsum4', ('fmul', a
, b
)), 'options->lower_fdot'),
201 (('fsum2', a
), ('fadd', 'a.x', 'a.y'), 'options->lower_fdot'),
203 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
204 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
205 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
206 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
207 (('~fsub', 1.0, ('fsat', a
)), ('fsat', ('fsub', 1.0, a
))),
209 # 1 - ((1 - a) * (1 - b))
210 # 1 - (1 - a - b + a*b)
211 # 1 - 1 + a + b - a*b
216 (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a
)), ('fadd', 1.0, ('fneg', b
))))),
217 ('flrp', b
, 1.0, a
), '!options->lower_flrp32'),
219 # (a * #b + #c) << #d
220 # ((a * #b) << #d) + (#c << #d)
221 # (a * (#b << #d)) + (#c << #d)
222 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
223 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
227 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
230 # Care must be taken here. Shifts in NIR uses only the lower log2(bitsize)
231 # bits of the second source. These replacements must correctly handle the
232 # case where (b % bitsize) + (c % bitsize) >= bitsize.
233 for s
in [8, 16, 32, 64]:
236 ishl
= "ishl@{}".format(s
)
237 ishr
= "ishr@{}".format(s
)
238 ushr
= "ushr@{}".format(s
)
240 in_bounds
= ('ult', ('iadd', ('iand', b
, mask
), ('iand', c
, mask
)), s
)
242 optimizations
.extend([
243 ((ishl
, (ishl
, a
, '#b'), '#c'), ('bcsel', in_bounds
, (ishl
, a
, ('iadd', b
, c
)), 0)),
244 ((ushr
, (ushr
, a
, '#b'), '#c'), ('bcsel', in_bounds
, (ushr
, a
, ('iadd', b
, c
)), 0)),
246 # To get get -1 for large shifts of negative values, ishr must instead
247 # clamp the shift count to the maximum value.
248 ((ishr
, (ishr
, a
, '#b'), '#c'),
249 (ishr
, a
, ('imin', ('iadd', ('iand', b
, mask
), ('iand', c
, mask
)), s
- 1))),
252 optimizations
.extend([
253 # This is common for address calculations. Reassociating may enable the
254 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
255 # instruction or a constant offset field for in load / store instructions.
256 (('ishl', ('iadd', a
, '#b'), '#c'), ('iadd', ('ishl', a
, c
), ('ishl', b
, c
))),
258 # Comparison simplifications
259 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
260 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
261 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
262 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
263 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
264 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
265 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
266 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
267 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
268 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
270 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
271 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
272 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
273 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
274 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
276 # This helps some shaders because, after some optimizations, they end up
277 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
278 # matching would be handled by CSE.
279 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
280 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
281 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
282 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
283 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
284 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
285 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
286 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
287 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
288 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
290 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
291 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
292 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
293 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
294 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
295 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
297 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
298 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
299 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
300 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
304 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
306 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
308 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
310 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
311 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
312 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
313 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
314 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
315 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
316 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
317 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
318 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
319 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
320 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
321 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
322 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
323 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
324 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
325 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
326 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
327 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
328 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
329 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
331 # -(b2f(a) + b2f(b)) < 0
332 # 0 < b2f(a) + b2f(b)
333 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
335 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
336 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
338 # -(b2f(a) + b2f(b)) >= 0
339 # 0 >= b2f(a) + b2f(b)
340 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
342 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
343 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
345 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
346 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
348 # Some optimizations (below) convert things like (a < b || c < b) into
349 # (min(a, c) < b). However, this interfers with the previous optimizations
350 # that try to remove comparisons with negated sums of b2f. This just
352 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
353 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
355 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
356 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
357 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
358 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
359 (('~flt', ('fadd(is_used_once)', a
, '#b'), '#c'), ('flt', a
, ('fadd', c
, ('fneg', b
)))),
360 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('flt', ('fneg', ('fadd', c
, b
)), a
)),
361 (('~fge', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fge', a
, ('fadd', c
, ('fneg', b
)))),
362 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fge', ('fneg', ('fadd', c
, b
)), a
)),
363 (('~feq', ('fadd(is_used_once)', a
, '#b'), '#c'), ('feq', a
, ('fadd', c
, ('fneg', b
)))),
364 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('feq', ('fneg', ('fadd', c
, b
)), a
)),
365 (('~fne', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fne', a
, ('fadd', c
, ('fneg', b
)))),
366 (('~fne', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fne', ('fneg', ('fadd', c
, b
)), a
)),
368 # Cannot remove the addition from ilt or ige due to overflow.
369 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
370 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
372 # fmin(-b2f(a), b) >= 0.0
373 # -b2f(a) >= 0.0 && b >= 0.0
374 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
375 # b2f(a) == 0.0 && b >= 0.0
376 # a == False && b >= 0.0
379 # The fge in the second replacement is not a typo. I leave the proof that
380 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
382 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
383 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
385 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
386 (('~fne', ('b2f', 'a@1'), 0.0), a
),
387 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
388 (('ine', ('b2i', 'a@1'), 0), a
),
390 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
391 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
392 (('fge', ('u2f', a
), 0.0), True),
393 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
394 (('flt', ('u2f', a
), 0.0), False),
395 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
396 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
397 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
398 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
399 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
400 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
401 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
405 # fabs(a) != 0.0 because fabs(a) must be >= 0
407 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
411 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
414 # 0.0 == fabs(a) because fabs(a) must be >= 0
416 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
420 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
422 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
423 (('iand', ('fge', a
, 0.0), ('fge', 1.0, a
)), ('feq', a
, ('fsat', a
)), '!options->lower_fsat'),
425 # (a < 0.0) || (a > 1.0)
426 # !(!(a < 0.0) && !(a > 1.0))
427 # !((a >= 0.0) && (a <= 1.0))
430 (('ior', ('flt', a
, 0.0), ('flt', 1.0, a
)), ('fne', a
, ('fsat', a
)), '!options->lower_fsat'),
432 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
433 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
434 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
435 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
438 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
439 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
440 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
442 # Since b is a constant, constant folding will eliminate the fmin and the
443 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
444 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
446 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
448 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
449 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
450 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
451 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
452 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
453 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
454 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
455 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
456 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
457 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
458 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
459 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
460 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
461 (('bcsel', a
, True, b
), ('ior', a
, b
)),
462 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
463 (('bcsel', a
, b
, False), ('iand', a
, b
)),
464 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
465 (('~fmin', a
, a
), a
),
466 (('~fmax', a
, a
), a
),
471 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
472 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
473 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
474 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
475 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
476 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
477 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
478 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
479 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
480 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
481 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
482 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
483 (('~fmin', a
, ('fabs', a
)), a
),
484 (('imin', a
, ('iabs', a
)), a
),
485 (('~fmax', a
, ('fneg', ('fabs', a
))), a
),
486 (('imax', a
, ('ineg', ('iabs', a
))), a
),
487 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
488 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
489 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
490 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
491 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
492 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
493 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
494 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
495 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_negate && !options->lower_fsat'),
496 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
497 (('fsat', ('b2f', a
)), ('b2f', a
)),
498 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
499 (('fsat', ('fsat', a
)), ('fsat', a
)),
500 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_negate && !options->lower_fsat'),
501 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_negate && !options->lower_fsat'),
502 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
503 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
504 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
505 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
506 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
507 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
508 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
509 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
510 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
511 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
512 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
513 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
514 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
515 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
516 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
517 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
518 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
519 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
520 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
521 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
522 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
523 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
524 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
526 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
527 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
528 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
529 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
530 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
531 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
532 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
533 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
534 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
535 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
536 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
537 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
538 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
539 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
540 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
541 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
543 # These derive from the previous patterns with the application of b < 0 <=>
544 # 0 < -b. The transformation should be applied if either comparison is
545 # used once as this ensures that the number of comparisons will not
546 # increase. The sources to the ior and iand are not symmetric, so the
547 # rules have to be duplicated to get this behavior.
548 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
549 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
550 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
551 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
552 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
553 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
554 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
555 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
557 # Common pattern like 'if (i == 0 || i == 1 || ...)'
558 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
559 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
560 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
562 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
563 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
564 # so emit an open-coded version of that.
565 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
566 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
568 (('ior', a
, ('ieq', a
, False)), True),
569 (('ior', a
, ('inot', a
)), -1),
571 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
572 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
574 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0), '!options->lower_bitops'),
576 # These patterns can result when (a < b || a < c) => (a < min(b, c))
577 # transformations occur before constant propagation and loop-unrolling.
578 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
579 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
580 (('~fge', a
, ('fmin', b
, a
)), True),
581 (('~fge', ('fmax', a
, b
), a
), True),
582 (('~flt', a
, ('fmin', b
, a
)), False),
583 (('~flt', ('fmax', a
, b
), a
), False),
584 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
585 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
587 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
588 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
589 (('ige', a
, ('imin', b
, a
)), True),
590 (('ige', ('imax', a
, b
), a
), True),
591 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
592 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
593 (('uge', a
, ('umin', b
, a
)), True),
594 (('uge', ('umax', a
, b
), a
), True),
595 (('ilt', a
, ('imin', b
, a
)), False),
596 (('ilt', ('imax', a
, b
), a
), False),
597 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
598 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
599 (('ult', a
, ('umin', b
, a
)), False),
600 (('ult', ('umax', a
, b
), a
), False),
601 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
602 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
603 (('ult', a
, ('iand', b
, a
)), False),
604 (('ult', ('ior', a
, b
), a
), False),
605 (('uge', a
, ('iand', b
, a
)), True),
606 (('uge', ('ior', a
, b
), a
), True),
608 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
609 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
610 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
611 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
612 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
613 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
614 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
615 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
616 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
617 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
618 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
619 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
620 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
621 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
622 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
623 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
625 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
627 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
628 ('iabs', ('ishr', a
, b
))),
629 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
631 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
632 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
633 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
634 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
635 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
636 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
637 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
638 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
639 (('seq', ('seq', a
, b
), 1.0), ('seq', a
, b
)),
640 (('seq', ('sne', a
, b
), 1.0), ('sne', a
, b
)),
641 (('seq', ('slt', a
, b
), 1.0), ('slt', a
, b
)),
642 (('seq', ('sge', a
, b
), 1.0), ('sge', a
, b
)),
643 (('sne', ('seq', a
, b
), 0.0), ('seq', a
, b
)),
644 (('sne', ('sne', a
, b
), 0.0), ('sne', a
, b
)),
645 (('sne', ('slt', a
, b
), 0.0), ('slt', a
, b
)),
646 (('sne', ('sge', a
, b
), 0.0), ('sge', a
, b
)),
647 (('seq', ('seq', a
, b
), 0.0), ('sne', a
, b
)),
648 (('seq', ('sne', a
, b
), 0.0), ('seq', a
, b
)),
649 (('seq', ('slt', a
, b
), 0.0), ('sge', a
, b
)),
650 (('seq', ('sge', a
, b
), 0.0), ('slt', a
, b
)),
651 (('sne', ('seq', a
, b
), 1.0), ('sne', a
, b
)),
652 (('sne', ('sne', a
, b
), 1.0), ('seq', a
, b
)),
653 (('sne', ('slt', a
, b
), 1.0), ('sge', a
, b
)),
654 (('sne', ('sge', a
, b
), 1.0), ('slt', a
, b
)),
655 (('fall_equal2', a
, b
), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
656 (('fall_equal3', a
, b
), ('seq', ('fany_nequal3', a
, b
), 0.0), 'options->lower_vector_cmp'),
657 (('fall_equal4', a
, b
), ('seq', ('fany_nequal4', a
, b
), 0.0), 'options->lower_vector_cmp'),
658 (('fany_nequal2', a
, b
), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
659 (('fany_nequal3', a
, b
), ('fsat', ('fdot3', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
660 (('fany_nequal4', a
, b
), ('fsat', ('fdot4', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
661 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
662 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
664 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
665 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
666 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
667 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
668 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
669 (('ineg', ('b2i32', 'a@32')), a
),
670 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
671 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
672 # Comparison with the same args. Note that these are not done for
673 # the float versions because NaN always returns false on float
675 (('ilt', a
, a
), False),
676 (('ige', a
, a
), True),
677 (('ieq', a
, a
), True),
678 (('ine', a
, a
), False),
679 (('ult', a
, a
), False),
680 (('uge', a
, a
), True),
681 # Logical and bit operations
683 (('iand', a
, ~
0), a
),
687 (('ior', a
, True), True),
690 (('inot', ('inot', a
)), a
),
691 (('ior', ('iand', a
, b
), b
), b
),
692 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
693 (('iand', ('ior', a
, b
), b
), b
),
694 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
696 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
697 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
698 # Shift optimizations
705 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
706 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
707 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
708 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
709 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
710 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
711 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
712 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
713 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
714 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
715 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
716 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
717 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
718 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
719 # Exponential/logarithmic identities
720 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
721 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
722 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
723 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
724 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
725 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
726 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
727 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
728 (('~fpow', a
, 1.0), a
),
729 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
730 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
731 (('~fpow', 2.0, a
), ('fexp2', a
)),
732 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
733 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
734 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
735 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
736 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
737 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
738 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
739 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
740 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
741 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
742 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
743 # Division and reciprocal
744 (('~fdiv', 1.0, a
), ('frcp', a
)),
745 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
746 (('~frcp', ('frcp', a
)), a
),
747 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
748 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
749 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
751 (('fsin', a
), lowered_sincos(0.5), 'options->lower_sincos'),
752 (('fcos', a
), lowered_sincos(0.75), 'options->lower_sincos'),
753 # Boolean simplifications
754 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
755 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
756 (('ieq', a
, True), a
),
757 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
758 (('ine', a
, False), a
),
759 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
760 (('bcsel', a
, True, False), a
),
761 (('bcsel', a
, False, True), ('inot', a
)),
762 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
763 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
764 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
765 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
766 (('bcsel', True, b
, c
), b
),
767 (('bcsel', False, b
, c
), c
),
768 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
770 (('bcsel', a
, b
, b
), b
),
771 (('~fcsel', a
, b
, b
), b
),
773 # D3D Boolean emulation
774 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
775 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
776 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
777 ('ineg', ('b2i', ('iand', a
, b
)))),
778 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
779 ('ineg', ('b2i', ('ior', a
, b
)))),
780 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
781 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
782 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
783 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
784 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
785 (('iand', ('ineg', ('b2i', a
)), 1), ('b2i', a
)),
787 # SM5 32-bit shifts are defined to use the 5 least significant bits
788 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
789 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
790 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
793 (('i2b32', ('b2i', 'a@32')), a
),
794 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
795 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
796 (('i2b', ('ineg', a
)), ('i2b', a
)),
797 (('i2b', ('iabs', a
)), ('i2b', a
)),
798 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
800 # Ironically, mark these as imprecise because removing the conversions may
801 # preserve more precision than doing the conversions (e.g.,
802 # uint(float(0x81818181u)) == 0x81818200).
803 (('~f2i32', ('i2f', 'a@32')), a
),
804 (('~f2i32', ('u2f', 'a@32')), a
),
805 (('~f2u32', ('i2f', 'a@32')), a
),
806 (('~f2u32', ('u2f', 'a@32')), a
),
808 (('ffloor', 'a(is_integral)'), a
),
809 (('fceil', 'a(is_integral)'), a
),
810 (('ftrunc', 'a(is_integral)'), a
),
811 # fract(x) = x - floor(x), so fract(NaN) = NaN
812 (('~ffract', 'a(is_integral)'), 0.0),
813 (('fabs', 'a(is_not_negative)'), a
),
814 (('iabs', 'a(is_not_negative)'), a
),
815 (('fsat', 'a(is_not_positive)'), 0.0),
817 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
820 # It is undefined to convert a negative floating-point value to an
823 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
824 # some optimizations in the i965 backend to proceed.
825 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
826 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
827 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
828 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
830 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a
), '!options->lower_fsat'),
832 # The result of the multiply must be in [-1, 0], so the result of the ffma
834 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
835 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
836 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
837 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
839 (('fne', 'a(is_not_zero)', 0.0), True),
840 (('feq', 'a(is_not_zero)', 0.0), False),
842 # In this chart, + means value > 0 and - means value < 0.
844 # + >= + -> unknown 0 >= + -> false - >= + -> false
845 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
846 # + >= - -> true 0 >= - -> true - >= - -> unknown
848 # Using grouping conceptually similar to a Karnaugh map...
850 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
851 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
852 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
854 # The flt / ilt cases just invert the expected result.
856 # The results expecting true, must be marked imprecise. The results
857 # expecting false are fine because NaN compared >= or < anything is false.
859 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
860 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
861 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
863 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
864 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
865 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
867 (('ine', 'a(is_not_zero)', 0), True),
868 (('ieq', 'a(is_not_zero)', 0), False),
870 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
871 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
872 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
874 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
875 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
876 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
878 (('ult', 0, 'a(is_gt_zero)'), True),
880 # Packing and then unpacking does nothing
881 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
882 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
883 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
884 ('unpack_64_2x32_split_y', a
)), a
),
886 # Comparing two halves of an unpack separately. While this optimization
887 # should be correct for non-constant values, it's less obvious that it's
888 # useful in that case. For constant values, the pack will fold and we're
889 # guaranteed to reduce the whole tree to one instruction.
890 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
891 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
892 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
895 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
896 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
897 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
898 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
899 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
900 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
901 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
903 # Useless masking before unpacking
904 (('unpack_half_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_half_2x16_split_x', a
)),
905 (('unpack_32_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_32_2x16_split_x', a
)),
906 (('unpack_64_2x32_split_x', ('iand', a
, 0xffffffff)), ('unpack_64_2x32_split_x', a
)),
907 (('unpack_half_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_half_2x16_split_y', a
)),
908 (('unpack_32_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_32_2x16_split_y', a
)),
909 (('unpack_64_2x32_split_y', ('iand', a
, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a
)),
912 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
913 # patterns like those below.
914 for op
in ('ushr', 'ishr'):
915 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
916 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
917 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
919 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
921 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
922 # patterns like those below.
923 for op
in ('extract_u8', 'extract_i8'):
924 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
925 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
926 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
928 optimizations
.extend([
930 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
931 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
932 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
933 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
934 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
937 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
938 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
939 (('isub', ('iadd', a
, b
), b
), a
),
940 (('~fsub', ('fadd', a
, b
), b
), a
),
941 (('ussub_4x8', a
, 0), a
),
942 (('ussub_4x8', a
, ~
0), 0),
943 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
944 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
945 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
946 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
947 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
948 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
949 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
950 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
952 # Propagate negation up multiplication chains
953 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
954 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
956 # Propagate constants up multiplication chains
957 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
958 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
959 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
960 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
962 # Reassociate constants in add/mul chains so they can be folded together.
963 # For now, we mostly only handle cases where the constants are separated by
964 # a single non-constant. We could do better eventually.
965 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
966 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
967 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
968 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
969 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
971 # Drop mul-div by the same value when there's no wrapping.
972 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
975 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
976 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
977 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
979 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
980 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
981 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
983 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
986 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
987 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
988 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
989 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
990 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
992 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
993 ('bcsel', ('ult', 31, 'bits'), 'insert',
994 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
995 'options->lower_bitfield_insert'),
996 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
997 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
998 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
999 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1000 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
1001 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
1003 # Alternative lowering that doesn't rely on bfi.
1004 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1005 ('bcsel', ('ult', 31, 'bits'),
1008 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1009 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1010 'options->lower_bitfield_insert_to_shifts'),
1012 # Alternative lowering that uses bitfield_select.
1013 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1014 ('bcsel', ('ult', 31, 'bits'), 'insert',
1015 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1016 'options->lower_bitfield_insert_to_bitfield_select'),
1018 (('ibitfield_extract', 'value', 'offset', 'bits'),
1019 ('bcsel', ('ult', 31, 'bits'), 'value',
1020 ('ibfe', 'value', 'offset', 'bits')),
1021 'options->lower_bitfield_extract'),
1023 (('ubitfield_extract', 'value', 'offset', 'bits'),
1024 ('bcsel', ('ult', 31, 'bits'), 'value',
1025 ('ubfe', 'value', 'offset', 'bits')),
1026 'options->lower_bitfield_extract'),
1028 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1029 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1030 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1031 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1032 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1033 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1034 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1036 (('ibitfield_extract', 'value', 'offset', 'bits'),
1037 ('bcsel', ('ieq', 0, 'bits'),
1040 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1041 ('isub', 32, 'bits'))),
1042 'options->lower_bitfield_extract_to_shifts'),
1044 (('ubitfield_extract', 'value', 'offset', 'bits'),
1046 ('ushr', 'value', 'offset'),
1047 ('bcsel', ('ieq', 'bits', 32),
1049 ('isub', ('ishl', 1, 'bits'), 1))),
1050 'options->lower_bitfield_extract_to_shifts'),
1052 (('ifind_msb', 'value'),
1053 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1054 'options->lower_ifind_msb'),
1056 (('find_lsb', 'value'),
1057 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1058 'options->lower_find_lsb'),
1060 (('extract_i8', a
, 'b@32'),
1061 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
1062 'options->lower_extract_byte'),
1064 (('extract_u8', a
, 'b@32'),
1065 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
1066 'options->lower_extract_byte'),
1068 (('extract_i16', a
, 'b@32'),
1069 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
1070 'options->lower_extract_word'),
1072 (('extract_u16', a
, 'b@32'),
1073 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
1074 'options->lower_extract_word'),
1076 (('pack_unorm_2x16', 'v'),
1077 ('pack_uvec2_to_uint',
1078 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1079 'options->lower_pack_unorm_2x16'),
1081 (('pack_unorm_4x8', 'v'),
1082 ('pack_uvec4_to_uint',
1083 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1084 'options->lower_pack_unorm_4x8'),
1086 (('pack_snorm_2x16', 'v'),
1087 ('pack_uvec2_to_uint',
1088 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1089 'options->lower_pack_snorm_2x16'),
1091 (('pack_snorm_4x8', 'v'),
1092 ('pack_uvec4_to_uint',
1093 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1094 'options->lower_pack_snorm_4x8'),
1096 (('unpack_unorm_2x16', 'v'),
1097 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1098 ('extract_u16', 'v', 1))),
1100 'options->lower_unpack_unorm_2x16'),
1102 (('unpack_unorm_4x8', 'v'),
1103 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1104 ('extract_u8', 'v', 1),
1105 ('extract_u8', 'v', 2),
1106 ('extract_u8', 'v', 3))),
1108 'options->lower_unpack_unorm_4x8'),
1110 (('unpack_snorm_2x16', 'v'),
1111 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1112 ('extract_i16', 'v', 1))),
1114 'options->lower_unpack_snorm_2x16'),
1116 (('unpack_snorm_4x8', 'v'),
1117 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1118 ('extract_i8', 'v', 1),
1119 ('extract_i8', 'v', 2),
1120 ('extract_i8', 'v', 3))),
1122 'options->lower_unpack_snorm_4x8'),
1124 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
1125 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
1128 # bit_size dependent lowerings
1129 for bit_size
in [8, 16, 32, 64]:
1130 # convenience constants
1131 intmax
= (1 << (bit_size
- 1)) - 1
1132 intmin
= 1 << (bit_size
- 1)
1135 (('iadd_sat@' + str(bit_size
), a
, b
),
1136 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
1137 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
1138 (('isub_sat@' + str(bit_size
), a
, b
),
1139 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
1140 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
1143 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
1145 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
1146 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1147 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1148 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1149 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1151 # Optimize x2bN(b2x(x)) -> x
1152 for size
in type_sizes('bool'):
1153 aN
= 'a@' + str(size
)
1154 f2bN
= 'f2b' + str(size
)
1155 i2bN
= 'i2b' + str(size
)
1156 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
1157 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
1159 # Optimize x2yN(b2x(x)) -> b2y
1160 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1161 if x
!= 'f' and y
!= 'f' and x
!= y
:
1164 b2x
= 'b2f' if x
== 'f' else 'b2i'
1165 b2y
= 'b2f' if y
== 'f' else 'b2i'
1166 x2yN
= '{}2{}'.format(x
, y
)
1167 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1169 # Optimize away x2xN(a@N)
1170 for t
in ['int', 'uint', 'float']:
1171 for N
in type_sizes(t
):
1172 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1173 aN
= 'a@{0}'.format(N
)
1174 optimizations
.append(((x2xN
, aN
), a
))
1176 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1177 # In particular, we can optimize away everything except upcast of downcast and
1178 # upcasts where the type differs from the other cast
1179 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1181 # The outer cast is a down-cast. It doesn't matter what the size of the
1182 # argument of the inner cast is because we'll never been in the upcast
1183 # of downcast case. Regardless of types, we'll always end up with y2yN
1185 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1186 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1187 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1188 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1189 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1191 # If the outer cast is an up-cast, we have to be more careful about the
1192 # size of the argument of the inner cast and with types. In this case,
1193 # the type is always the type of type up-cast which is given by the
1195 for P
in type_sizes('uint'):
1196 # We can't optimize away up-cast of down-cast.
1200 # Because we're doing down-cast of down-cast, the types always have
1201 # to match between the two casts
1202 for x
in ['i', 'u']:
1203 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1204 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1205 aP
= 'a@{0}'.format(P
)
1206 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1208 # The N == M case is handled by other optimizations
1211 # Optimize comparisons with up-casts
1212 for t
in ['int', 'uint', 'float']:
1213 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1214 if N
== 1 or N
>= M
:
1217 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1218 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1221 xeq
= 'feq' if t
== 'float' else 'ieq'
1222 xne
= 'fne' if t
== 'float' else 'ine'
1223 xge
= '{0}ge'.format(t
[0])
1224 xlt
= '{0}lt'.format(t
[0])
1226 # Up-casts are lossless so for correctly signed comparisons of
1227 # up-casted values we can do the comparison at the largest of the two
1228 # original sizes and drop one or both of the casts. (We have
1229 # optimizations to drop the no-op casts which this may generate.)
1230 for P
in type_sizes(t
):
1236 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1237 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1238 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1239 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1240 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1241 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1244 # The next bit doesn't work on floats because the range checks would
1245 # get way too complicated.
1246 if t
in ['int', 'uint']:
1248 xN_min
= -(1 << (N
- 1))
1249 xN_max
= (1 << (N
- 1)) - 1
1252 xN_max
= (1 << N
) - 1
1256 # If we're up-casting and comparing to a constant, we can unfold
1257 # the comparison into a comparison with the shrunk down constant
1258 # and a check that the constant fits in the smaller bit size.
1260 ((xeq
, (x2xM
, aN
), '#b'),
1261 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1262 ((xne
, (x2xM
, aN
), '#b'),
1263 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1264 ((xlt
, (x2xM
, aN
), '#b'),
1265 ('iand', (xlt
, xN_min
, b
),
1266 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1267 ((xlt
, '#a', (x2xM
, bN
)),
1268 ('iand', (xlt
, a
, xN_max
),
1269 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1270 ((xge
, (x2xM
, aN
), '#b'),
1271 ('iand', (xge
, xN_max
, b
),
1272 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1273 ((xge
, '#a', (x2xM
, bN
)),
1274 ('iand', (xge
, a
, xN_min
),
1275 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1278 def fexp2i(exp
, bits
):
1279 # We assume that exp is already in the right range.
1281 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1283 return ('ishl', ('iadd', exp
, 127), 23)
1285 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1289 def ldexp(f
, exp
, bits
):
1290 # First, we clamp exp to a reasonable range. The maximum possible range
1291 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1292 # a maximum range of [-149, 127]. This means that we can potentially have
1293 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1294 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1295 # spec, on the other hand, only requires that we handle an exponent value
1296 # in the range [-126, 128]. This implementation is *mostly* correct; it
1297 # handles a range on exp of [-252, 254] which allows you to create any
1298 # value (including denorms if the hardware supports it) and to adjust the
1299 # exponent of any normal value to anything you want.
1301 exp
= ('imin', ('imax', exp
, -28), 30)
1303 exp
= ('imin', ('imax', exp
, -252), 254)
1305 exp
= ('imin', ('imax', exp
, -2044), 2046)
1309 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1310 # (We use ishr which isn't the same for -1, but the -1 case still works
1311 # since we use exp-exp/2 as the second exponent.) While the spec
1312 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1313 # work with denormals and doesn't allow for the full swing in exponents
1314 # that you can get with normalized values. Instead, we create two powers
1315 # of two and multiply by them each in turn. That way the effective range
1316 # of our exponent is doubled.
1317 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1318 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1319 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1322 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1323 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1324 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1327 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1328 def bitfield_reverse(u
):
1329 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1330 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1331 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1332 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1333 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1337 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1339 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1340 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1341 # and, if a is a NaN then the second comparison will fail anyway.
1342 for op
in ['flt', 'fge', 'feq']:
1344 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
1345 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
1348 # Add optimizations to handle the case where the result of a ternary is
1349 # compared to a constant. This way we can take things like
1355 # a ? (0 > 0) : (1 > 0)
1357 # which constant folding will eat for lunch. The resulting ternary will
1358 # further get cleaned up by the boolean reductions above and we will be
1359 # left with just the original variable "a".
1360 for op
in ['flt', 'fge', 'feq', 'fne',
1361 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1363 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1364 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1365 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1366 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1370 # For example, this converts things like
1372 # 1 + mix(0, a - 1, condition)
1376 # mix(1, (a-1)+1, condition)
1378 # Other optimizations will rearrange the constants.
1379 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1381 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1384 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1387 # If neither layout qualifier is specified, derivatives in compute shaders
1388 # return zero, which is consistent with the handling of built-in texture
1389 # functions like texture() in GLSL 4.50 compute shaders.
1390 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1391 'fddy', 'fddy_fine', 'fddy_coarse']:
1393 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1396 # Some optimizations for ir3-specific instructions.
1398 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1399 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1400 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1401 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1402 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1405 # This section contains "late" optimizations that should be run before
1406 # creating ffmas and calling regular optimizations for the final time.
1407 # Optimizations should go here if they help code generation and conflict
1408 # with the regular optimizations.
1409 before_ffma_optimizations
= [
1410 # Propagate constants down multiplication chains
1411 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1412 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1413 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1414 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1416 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1417 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1418 (('~fadd', ('fneg', a
), a
), 0.0),
1419 (('iadd', ('ineg', a
), a
), 0),
1420 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1421 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1422 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1423 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1425 (('~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
)),
1426 (('~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
)),
1427 (('~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
))),
1430 # This section contains "late" optimizations that should be run after the
1431 # regular optimizations have finished. Optimizations should go here if
1432 # they help code generation but do not necessarily produce code that is
1433 # more easily optimizable.
1434 late_optimizations
= [
1435 # Most of these optimizations aren't quite safe when you get infinity or
1436 # Nan involved but the first one should be fine.
1437 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1438 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1439 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1440 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1441 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1442 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1444 # nir_lower_to_source_mods will collapse this, but its existence during the
1445 # optimization loop can prevent other optimizations.
1446 (('fneg', ('fneg', a
)), a
),
1448 # These are duplicated from the main optimizations table. The late
1449 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1450 # new patterns like these. The patterns that compare with zero are removed
1451 # because they are unlikely to be created in by anything in
1452 # late_optimizations.
1453 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1454 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1455 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1456 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1457 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1458 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1460 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1461 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1463 (('~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
)))),
1465 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1466 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1467 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1468 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1469 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1470 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1471 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1472 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1473 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1474 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1477 (('iand', a
, a
), a
),
1479 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1481 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1482 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1483 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1484 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1486 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1487 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1489 (('~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'),
1490 (('~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'),
1492 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1493 # particular operation is common for expanding values stored in a texture
1494 # from [0,1] to [-1,1].
1495 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1496 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1497 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1498 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1499 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1500 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1501 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1502 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1506 # a + -a*a + a*b (1)
1508 # Option 1: ffma(a, (b-a), a)
1510 # Alternately, after (1):
1516 # Option 2: ffma(a, 2, -(a*a))
1517 # Option 3: ffma(a, 2, (-a)*a)
1518 # Option 4: ffma(a, -a, (2*a)
1519 # Option 5: a * (2 - a)
1521 # There are a lot of other possible combinations.
1522 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1523 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1524 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1525 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1526 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1528 # we do these late so that we don't get in the way of creating ffmas
1529 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1530 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1532 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1534 # Things that look like DPH in the source shader may get expanded to
1535 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1536 # to NIR. After FFMA is generated, this can look like:
1538 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1540 # Reassociate the last addition into the first multiplication.
1542 # Some shaders do not use 'invariant' in vertex and (possibly) geometry
1543 # shader stages on some outputs that are intended to be invariant. For
1544 # various reasons, this optimization may not be fully applied in all
1545 # shaders used for different rendering passes of the same geometry. This
1546 # can result in Z-fighting artifacts (at best). For now, disable this
1547 # optimization in these stages. See bugzilla #111490. In tessellation
1548 # stages applications seem to use 'precise' when necessary, so allow the
1549 # optimization in those stages.
1550 (('~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)'),
1551 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1552 (('~fadd', ('ffma(is_used_once)', a
, b
, ('fmul', 'c(is_not_const_and_not_fsign)', 'd(is_not_const_and_not_fsign)') ), 'e(is_not_const)'),
1553 ('ffma', a
, b
, ('ffma', c
, d
, e
)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1556 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1557 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1558 before_ffma_optimizations
).render())
1559 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1560 late_optimizations
).render())
projects / mesa.git / blob