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 # If the opcode in a replacement expression is prefixed by a '!' character,
73 # this indicated that the new expression will be marked exact.
75 # A special condition "many-comm-expr" can be used with expressions to note
76 # that the expression and its subexpressions have more commutative expressions
77 # than nir_replace_instr can handle. If this special condition is needed with
78 # another condition, the two can be separated by a comma (e.g.,
79 # "(many-comm-expr,is_used_once)").
81 # based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
82 def lowered_sincos(c
):
83 x
= ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi
, a
), c
))), 1.0)
84 x
= ('fmul', ('fsub', x
, ('fmul', x
, ('fabs', x
))), 4.0)
85 return ('ffma', ('ffma', x
, ('fabs', x
), ('fneg', x
)), 0.225, x
)
89 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitops'),
90 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
91 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitops'),
93 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
94 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
95 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
96 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
101 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
)), '!options->lower_bitops'),
102 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
103 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
104 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
106 (('~fneg', ('fneg', a
)), a
),
107 (('ineg', ('ineg', a
)), a
),
108 (('fabs', ('fabs', a
)), ('fabs', a
)),
109 (('fabs', ('fneg', a
)), ('fabs', a
)),
110 (('fabs', ('u2f', a
)), ('u2f', a
)),
111 (('iabs', ('iabs', a
)), ('iabs', a
)),
112 (('iabs', ('ineg', a
)), ('iabs', a
)),
113 (('f2b', ('fneg', a
)), ('f2b', a
)),
114 (('i2b', ('ineg', a
)), ('i2b', a
)),
115 (('~fadd', a
, 0.0), a
),
117 (('usadd_4x8', a
, 0), a
),
118 (('usadd_4x8', a
, ~
0), ~
0),
119 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
120 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
121 (('~fadd', ('fneg', a
), a
), 0.0),
122 (('iadd', ('ineg', a
), a
), 0),
123 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
124 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
125 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
126 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
127 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
128 (('~fmul', a
, 0.0), 0.0),
130 (('umul_unorm_4x8', a
, 0), 0),
131 (('umul_unorm_4x8', a
, ~
0), a
),
132 (('~fmul', a
, 1.0), a
),
134 (('fmul', a
, -1.0), ('fneg', a
)),
135 (('imul', a
, -1), ('ineg', a
)),
136 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
137 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
138 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
139 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
140 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
141 (('~ffma', 0.0, a
, b
), b
),
142 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
143 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
144 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
145 (('~flrp', a
, b
, 0.0), a
),
146 (('~flrp', a
, b
, 1.0), b
),
147 (('~flrp', a
, a
, b
), a
),
148 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
150 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
151 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
152 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
153 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
155 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
156 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
158 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
159 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
161 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
163 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
164 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
165 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
166 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
167 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
168 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
169 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
170 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
171 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
172 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
173 # These are the same as the previous three rules, but it depends on
174 # 1-fsat(x) <=> fsat(1-x). See below.
175 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
176 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
178 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
179 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
180 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
181 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
182 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
184 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
185 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
187 (('fdph', a
, b
), ('fdot4', ('vec4', 'a.x', 'a.y', 'a.z', 1.0), b
), 'options->lower_fdph'),
189 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
), '!options->lower_fdph'),
190 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
191 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
192 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
194 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
195 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
197 (('fdot2', ('vec2', a
, 0.0), b
), ('fmul', a
, b
)),
198 (('fdot2', a
, 1.0), ('fadd', 'a.x', 'a.y')),
200 # Lower fdot to fsum when it is available
201 (('fdot2', a
, b
), ('fsum2', ('fmul', a
, b
)), 'options->lower_fdot'),
202 (('fdot3', a
, b
), ('fsum3', ('fmul', a
, b
)), 'options->lower_fdot'),
203 (('fdot4', a
, b
), ('fsum4', ('fmul', a
, b
)), 'options->lower_fdot'),
204 (('fsum2', a
), ('fadd', 'a.x', 'a.y'), 'options->lower_fdot'),
206 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
207 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
208 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
209 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
211 # 1 - ((1 - a) * (1 - b))
212 # 1 - (1 - a - b + a*b)
213 # 1 - 1 + a + b - a*b
218 (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a
)), ('fadd', 1.0, ('fneg', b
))))),
219 ('flrp', b
, 1.0, a
), '!options->lower_flrp32'),
221 # (a * #b + #c) << #d
222 # ((a * #b) << #d) + (#c << #d)
223 # (a * (#b << #d)) + (#c << #d)
224 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
225 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
229 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
232 # Care must be taken here. Shifts in NIR uses only the lower log2(bitsize)
233 # bits of the second source. These replacements must correctly handle the
234 # case where (b % bitsize) + (c % bitsize) >= bitsize.
235 for s
in [8, 16, 32, 64]:
238 ishl
= "ishl@{}".format(s
)
239 ishr
= "ishr@{}".format(s
)
240 ushr
= "ushr@{}".format(s
)
242 in_bounds
= ('ult', ('iadd', ('iand', b
, mask
), ('iand', c
, mask
)), s
)
244 optimizations
.extend([
245 ((ishl
, (ishl
, a
, '#b'), '#c'), ('bcsel', in_bounds
, (ishl
, a
, ('iadd', b
, c
)), 0)),
246 ((ushr
, (ushr
, a
, '#b'), '#c'), ('bcsel', in_bounds
, (ushr
, a
, ('iadd', b
, c
)), 0)),
248 # To get get -1 for large shifts of negative values, ishr must instead
249 # clamp the shift count to the maximum value.
250 ((ishr
, (ishr
, a
, '#b'), '#c'),
251 (ishr
, a
, ('imin', ('iadd', ('iand', b
, mask
), ('iand', c
, mask
)), s
- 1))),
254 optimizations
.extend([
255 # This is common for address calculations. Reassociating may enable the
256 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
257 # instruction or a constant offset field for in load / store instructions.
258 (('ishl', ('iadd', a
, '#b'), '#c'), ('iadd', ('ishl', a
, c
), ('ishl', b
, c
))),
260 # Comparison simplifications
261 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
262 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
263 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
264 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
265 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
266 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
267 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
268 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
269 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
270 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
272 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
273 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
274 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
275 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
276 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
278 # This helps some shaders because, after some optimizations, they end up
279 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
280 # matching would be handled by CSE.
281 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
282 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
283 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
284 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
285 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
286 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
287 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
288 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
289 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
290 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
292 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
293 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
294 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
295 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
296 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
297 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
299 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
300 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
301 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
302 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
306 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
308 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
310 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
312 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
313 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
314 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
315 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
316 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
317 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
318 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
319 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
320 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
321 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
322 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
323 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
324 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
325 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
326 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
327 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
328 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
329 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
330 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
331 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
333 # -(b2f(a) + b2f(b)) < 0
334 # 0 < b2f(a) + b2f(b)
335 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
337 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
338 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
340 # -(b2f(a) + b2f(b)) >= 0
341 # 0 >= b2f(a) + b2f(b)
342 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
344 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
345 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
347 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
348 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
350 # Some optimizations (below) convert things like (a < b || c < b) into
351 # (min(a, c) < b). However, this interfers with the previous optimizations
352 # that try to remove comparisons with negated sums of b2f. This just
354 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
355 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
357 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
358 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
359 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
360 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
361 (('~flt', ('fadd(is_used_once)', a
, '#b'), '#c'), ('flt', a
, ('fadd', c
, ('fneg', b
)))),
362 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('flt', ('fneg', ('fadd', c
, b
)), a
)),
363 (('~fge', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fge', a
, ('fadd', c
, ('fneg', b
)))),
364 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fge', ('fneg', ('fadd', c
, b
)), a
)),
365 (('~feq', ('fadd(is_used_once)', a
, '#b'), '#c'), ('feq', a
, ('fadd', c
, ('fneg', b
)))),
366 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('feq', ('fneg', ('fadd', c
, b
)), a
)),
367 (('~fne', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fne', a
, ('fadd', c
, ('fneg', b
)))),
368 (('~fne', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fne', ('fneg', ('fadd', c
, b
)), a
)),
370 # Cannot remove the addition from ilt or ige due to overflow.
371 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
372 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
374 # fmin(-b2f(a), b) >= 0.0
375 # -b2f(a) >= 0.0 && b >= 0.0
376 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
377 # b2f(a) == 0.0 && b >= 0.0
378 # a == False && b >= 0.0
381 # The fge in the second replacement is not a typo. I leave the proof that
382 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
384 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
385 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
387 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
388 (('~fne', ('b2f', 'a@1'), 0.0), a
),
389 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
390 (('ine', ('b2i', 'a@1'), 0), a
),
392 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
393 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
394 (('fge', ('u2f', a
), 0.0), True),
395 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
396 (('flt', ('u2f', a
), 0.0), False),
397 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
398 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
399 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
400 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
401 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
402 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
403 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
407 # fabs(a) != 0.0 because fabs(a) must be >= 0
409 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
413 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
416 # 0.0 == fabs(a) because fabs(a) must be >= 0
418 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
422 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
424 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
425 (('iand', ('fge', a
, 0.0), ('fge', 1.0, a
)), ('feq', a
, ('fsat', a
)), '!options->lower_fsat'),
427 # (a < 0.0) || (a > 1.0)
428 # !(!(a < 0.0) && !(a > 1.0))
429 # !((a >= 0.0) && (a <= 1.0))
432 (('ior', ('flt', a
, 0.0), ('flt', 1.0, a
)), ('fne', a
, ('fsat', a
)), '!options->lower_fsat'),
434 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
435 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
436 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
437 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
440 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
441 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
442 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
444 # Since b is a constant, constant folding will eliminate the fmin and the
445 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
446 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
448 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
450 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
451 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
452 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
453 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
454 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
455 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
456 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
457 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
458 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
459 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
460 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
461 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
462 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
463 (('bcsel', a
, True, b
), ('ior', a
, b
)),
464 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
465 (('bcsel', a
, b
, False), ('iand', a
, b
)),
466 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
467 (('~fmin', a
, a
), a
),
468 (('~fmax', a
, a
), a
),
473 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
474 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
475 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
476 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
477 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
478 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
479 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
480 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
481 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
482 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
483 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
484 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
485 (('~fmin', a
, ('fabs', a
)), a
),
486 (('imin', a
, ('iabs', a
)), a
),
487 (('~fmax', a
, ('fneg', ('fabs', a
))), a
),
488 (('imax', a
, ('ineg', ('iabs', a
))), a
),
489 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
490 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
491 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
492 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
493 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
494 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
495 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
496 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
497 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
498 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
499 (('fsat', ('b2f', a
)), ('b2f', a
)),
500 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
501 (('fsat', ('fsat', a
)), ('fsat', a
)),
502 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_fsat'),
503 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_fsat'),
504 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
505 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
506 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
507 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
508 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
509 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
510 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
511 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
512 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
513 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
514 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
515 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
516 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
517 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
518 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
519 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
520 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
521 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
522 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
523 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
524 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
525 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
526 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
528 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
529 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
530 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
531 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
532 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
533 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
534 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
535 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
536 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
537 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
538 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
539 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
540 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
541 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
542 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
543 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
545 # These derive from the previous patterns with the application of b < 0 <=>
546 # 0 < -b. The transformation should be applied if either comparison is
547 # used once as this ensures that the number of comparisons will not
548 # increase. The sources to the ior and iand are not symmetric, so the
549 # rules have to be duplicated to get this behavior.
550 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
551 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
552 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
553 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
554 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
555 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
556 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
557 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
559 # Common pattern like 'if (i == 0 || i == 1 || ...)'
560 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
561 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
562 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
564 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
565 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
566 # so emit an open-coded version of that.
567 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
568 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
570 (('ior', a
, ('ieq', a
, False)), True),
571 (('ior', a
, ('inot', a
)), -1),
573 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
574 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
576 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0), '!options->lower_bitops'),
578 # These patterns can result when (a < b || a < c) => (a < min(b, c))
579 # transformations occur before constant propagation and loop-unrolling.
580 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
581 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
582 (('~fge', a
, ('fmin', b
, a
)), True),
583 (('~fge', ('fmax', a
, b
), a
), True),
584 (('~flt', a
, ('fmin', b
, a
)), False),
585 (('~flt', ('fmax', a
, b
), a
), False),
586 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
587 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
589 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
590 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
591 (('ige', a
, ('imin', b
, a
)), True),
592 (('ige', ('imax', a
, b
), a
), True),
593 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
594 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
595 (('uge', a
, ('umin', b
, a
)), True),
596 (('uge', ('umax', a
, b
), a
), True),
597 (('ilt', a
, ('imin', b
, a
)), False),
598 (('ilt', ('imax', a
, b
), a
), False),
599 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
600 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
601 (('ult', a
, ('umin', b
, a
)), False),
602 (('ult', ('umax', a
, b
), a
), False),
603 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
604 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
605 (('ult', a
, ('iand', b
, a
)), False),
606 (('ult', ('ior', a
, b
), a
), False),
607 (('uge', a
, ('iand', b
, a
)), True),
608 (('uge', ('ior', a
, b
), a
), True),
610 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
611 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
612 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
613 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
614 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
615 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
616 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
617 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
618 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
619 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
620 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
621 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
622 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
623 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
624 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
625 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
627 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
629 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
630 ('iabs', ('ishr', a
, b
))),
631 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
633 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
634 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
635 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
636 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
637 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
638 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
639 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
640 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
641 (('seq', ('seq', a
, b
), 1.0), ('seq', a
, b
)),
642 (('seq', ('sne', a
, b
), 1.0), ('sne', a
, b
)),
643 (('seq', ('slt', a
, b
), 1.0), ('slt', a
, b
)),
644 (('seq', ('sge', a
, b
), 1.0), ('sge', a
, b
)),
645 (('sne', ('seq', a
, b
), 0.0), ('seq', a
, b
)),
646 (('sne', ('sne', a
, b
), 0.0), ('sne', a
, b
)),
647 (('sne', ('slt', a
, b
), 0.0), ('slt', a
, b
)),
648 (('sne', ('sge', a
, b
), 0.0), ('sge', a
, b
)),
649 (('seq', ('seq', a
, b
), 0.0), ('sne', a
, b
)),
650 (('seq', ('sne', a
, b
), 0.0), ('seq', a
, b
)),
651 (('seq', ('slt', a
, b
), 0.0), ('sge', a
, b
)),
652 (('seq', ('sge', a
, b
), 0.0), ('slt', a
, b
)),
653 (('sne', ('seq', a
, b
), 1.0), ('sne', a
, b
)),
654 (('sne', ('sne', a
, b
), 1.0), ('seq', a
, b
)),
655 (('sne', ('slt', a
, b
), 1.0), ('sge', a
, b
)),
656 (('sne', ('sge', a
, b
), 1.0), ('slt', a
, b
)),
657 (('fall_equal2', a
, b
), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
658 (('fall_equal3', a
, b
), ('seq', ('fany_nequal3', a
, b
), 0.0), 'options->lower_vector_cmp'),
659 (('fall_equal4', a
, b
), ('seq', ('fany_nequal4', a
, b
), 0.0), 'options->lower_vector_cmp'),
660 (('fany_nequal2', a
, b
), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
661 (('fany_nequal3', a
, b
), ('fsat', ('fdot3', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
662 (('fany_nequal4', a
, b
), ('fsat', ('fdot4', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
663 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
664 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
666 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
667 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
668 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
669 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
670 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
671 (('ineg', ('b2i32', 'a@32')), a
),
672 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
673 # Comparison with the same args. Note that these are not done for
674 # the float versions because NaN always returns false on float
676 (('ilt', a
, a
), False),
677 (('ige', a
, a
), True),
678 (('ieq', a
, a
), True),
679 (('ine', a
, a
), False),
680 (('ult', a
, a
), False),
681 (('uge', a
, a
), True),
682 # Logical and bit operations
684 (('iand', a
, ~
0), a
),
688 (('ior', a
, True), True),
691 (('inot', ('inot', a
)), a
),
692 (('ior', ('iand', a
, b
), b
), b
),
693 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
694 (('iand', ('ior', a
, b
), b
), b
),
695 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
697 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
698 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
699 # Shift optimizations
706 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
707 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
708 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
709 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
710 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
711 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
712 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
713 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
714 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
715 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
716 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
717 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
718 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
719 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
720 # Exponential/logarithmic identities
721 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
722 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
723 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
724 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
725 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
726 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
727 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
728 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
729 (('~fpow', a
, 1.0), a
),
730 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
731 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
732 (('~fpow', 2.0, a
), ('fexp2', a
)),
733 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
734 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
735 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
736 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
737 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
738 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
739 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
740 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
741 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
742 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
743 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
744 # Division and reciprocal
745 (('~fdiv', 1.0, a
), ('frcp', a
)),
746 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
747 (('~frcp', ('frcp', a
)), a
),
748 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
749 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
750 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
752 (('fsin', a
), lowered_sincos(0.5), 'options->lower_sincos'),
753 (('fcos', a
), lowered_sincos(0.75), 'options->lower_sincos'),
754 # Boolean simplifications
755 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
756 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
757 (('ieq', a
, True), a
),
758 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
759 (('ine', a
, False), a
),
760 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
761 (('bcsel', a
, True, False), a
),
762 (('bcsel', a
, False, True), ('inot', a
)),
763 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
764 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
765 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
766 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
767 (('bcsel', True, b
, c
), b
),
768 (('bcsel', False, b
, c
), c
),
769 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
771 (('bcsel', a
, b
, b
), b
),
772 (('~fcsel', a
, b
, b
), b
),
774 # D3D Boolean emulation
775 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
776 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
777 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
778 ('ineg', ('b2i', ('iand', a
, b
)))),
779 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
780 ('ineg', ('b2i', ('ior', a
, b
)))),
781 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
782 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
783 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
784 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
785 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
786 (('iand', ('ineg', ('b2i', a
)), 1), ('b2i', a
)),
788 # SM5 32-bit shifts are defined to use the 5 least significant bits
789 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
790 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
791 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
794 (('i2b32', ('b2i', 'a@32')), a
),
795 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
796 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
797 (('i2b', ('ineg', a
)), ('i2b', a
)),
798 (('i2b', ('iabs', a
)), ('i2b', a
)),
799 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
801 # Ironically, mark these as imprecise because removing the conversions may
802 # preserve more precision than doing the conversions (e.g.,
803 # uint(float(0x81818181u)) == 0x81818200).
804 (('~f2i32', ('i2f', 'a@32')), a
),
805 (('~f2i32', ('u2f', 'a@32')), a
),
806 (('~f2u32', ('i2f', 'a@32')), a
),
807 (('~f2u32', ('u2f', 'a@32')), a
),
809 (('ffloor', 'a(is_integral)'), a
),
810 (('fceil', 'a(is_integral)'), a
),
811 (('ftrunc', 'a(is_integral)'), a
),
812 # fract(x) = x - floor(x), so fract(NaN) = NaN
813 (('~ffract', 'a(is_integral)'), 0.0),
814 (('fabs', 'a(is_not_negative)'), a
),
815 (('iabs', 'a(is_not_negative)'), a
),
816 (('fsat', 'a(is_not_positive)'), 0.0),
818 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
821 # It is undefined to convert a negative floating-point value to an
824 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
825 # some optimizations in the i965 backend to proceed.
826 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
827 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
828 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
829 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
831 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a
), '!options->lower_fsat'),
833 # The result of the multiply must be in [-1, 0], so the result of the ffma
835 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
836 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
837 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
838 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
840 (('fne', 'a(is_not_zero)', 0.0), True),
841 (('feq', 'a(is_not_zero)', 0.0), False),
843 # In this chart, + means value > 0 and - means value < 0.
845 # + >= + -> unknown 0 >= + -> false - >= + -> false
846 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
847 # + >= - -> true 0 >= - -> true - >= - -> unknown
849 # Using grouping conceptually similar to a Karnaugh map...
851 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
852 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
853 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
855 # The flt / ilt cases just invert the expected result.
857 # The results expecting true, must be marked imprecise. The results
858 # expecting false are fine because NaN compared >= or < anything is false.
860 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
861 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
862 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
864 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
865 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
866 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
868 (('ine', 'a(is_not_zero)', 0), True),
869 (('ieq', 'a(is_not_zero)', 0), False),
871 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
872 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
873 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
875 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
876 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
877 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
879 (('ult', 0, 'a(is_gt_zero)'), True),
881 # Packing and then unpacking does nothing
882 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
883 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
884 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
885 ('unpack_64_2x32_split_y', a
)), a
),
887 # Comparing two halves of an unpack separately. While this optimization
888 # should be correct for non-constant values, it's less obvious that it's
889 # useful in that case. For constant values, the pack will fold and we're
890 # guaranteed to reduce the whole tree to one instruction.
891 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
892 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
893 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
896 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
897 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
898 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
899 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
900 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
901 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
902 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
904 # Useless masking before unpacking
905 (('unpack_half_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_half_2x16_split_x', a
)),
906 (('unpack_32_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_32_2x16_split_x', a
)),
907 (('unpack_64_2x32_split_x', ('iand', a
, 0xffffffff)), ('unpack_64_2x32_split_x', a
)),
908 (('unpack_half_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_half_2x16_split_y', a
)),
909 (('unpack_32_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_32_2x16_split_y', a
)),
910 (('unpack_64_2x32_split_y', ('iand', a
, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a
)),
913 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
914 # patterns like those below.
915 for op
in ('ushr', 'ishr'):
916 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
917 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
918 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
920 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
922 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
923 # patterns like those below.
924 for op
in ('extract_u8', 'extract_i8'):
925 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
926 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
927 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
929 optimizations
.extend([
931 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
932 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
933 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
934 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
935 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
938 (('ussub_4x8', a
, 0), a
),
939 (('ussub_4x8', a
, ~
0), 0),
940 # Lower all Subtractions first - they can get recombined later
941 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
))),
942 (('isub', a
, b
), ('iadd', a
, ('ineg', b
))),
944 # Propagate negation up multiplication chains
945 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
946 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
948 # Propagate constants up multiplication chains
949 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
950 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
951 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
952 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
954 # Reassociate constants in add/mul chains so they can be folded together.
955 # For now, we mostly only handle cases where the constants are separated by
956 # a single non-constant. We could do better eventually.
957 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
958 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
959 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
960 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
961 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
963 # Drop mul-div by the same value when there's no wrapping.
964 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
967 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
968 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
969 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
971 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
972 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
973 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
975 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
978 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
979 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
980 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
981 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
982 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
984 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
985 ('bcsel', ('ult', 31, 'bits'), 'insert',
986 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
987 'options->lower_bitfield_insert'),
988 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
989 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
990 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
991 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
992 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
993 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
995 # Alternative lowering that doesn't rely on bfi.
996 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
997 ('bcsel', ('ult', 31, 'bits'),
1000 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1001 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1002 'options->lower_bitfield_insert_to_shifts'),
1004 # Alternative lowering that uses bitfield_select.
1005 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1006 ('bcsel', ('ult', 31, 'bits'), 'insert',
1007 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1008 'options->lower_bitfield_insert_to_bitfield_select'),
1010 (('ibitfield_extract', 'value', 'offset', 'bits'),
1011 ('bcsel', ('ult', 31, 'bits'), 'value',
1012 ('ibfe', 'value', 'offset', 'bits')),
1013 'options->lower_bitfield_extract'),
1015 (('ubitfield_extract', 'value', 'offset', 'bits'),
1016 ('bcsel', ('ult', 31, 'bits'), 'value',
1017 ('ubfe', 'value', 'offset', 'bits')),
1018 'options->lower_bitfield_extract'),
1020 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1021 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1022 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1023 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1024 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1025 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1026 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1028 (('ibitfield_extract', 'value', 'offset', 'bits'),
1029 ('bcsel', ('ieq', 0, 'bits'),
1032 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1033 ('isub', 32, 'bits'))),
1034 'options->lower_bitfield_extract_to_shifts'),
1036 (('ubitfield_extract', 'value', 'offset', 'bits'),
1038 ('ushr', 'value', 'offset'),
1039 ('bcsel', ('ieq', 'bits', 32),
1041 ('isub', ('ishl', 1, 'bits'), 1))),
1042 'options->lower_bitfield_extract_to_shifts'),
1044 (('ifind_msb', 'value'),
1045 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1046 'options->lower_ifind_msb'),
1048 (('find_lsb', 'value'),
1049 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1050 'options->lower_find_lsb'),
1052 (('extract_i8', a
, 'b@32'),
1053 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
1054 'options->lower_extract_byte'),
1056 (('extract_u8', a
, 'b@32'),
1057 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
1058 'options->lower_extract_byte'),
1060 (('extract_i16', a
, 'b@32'),
1061 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
1062 'options->lower_extract_word'),
1064 (('extract_u16', a
, 'b@32'),
1065 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
1066 'options->lower_extract_word'),
1068 (('pack_unorm_2x16', 'v'),
1069 ('pack_uvec2_to_uint',
1070 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1071 'options->lower_pack_unorm_2x16'),
1073 (('pack_unorm_4x8', 'v'),
1074 ('pack_uvec4_to_uint',
1075 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1076 'options->lower_pack_unorm_4x8'),
1078 (('pack_snorm_2x16', 'v'),
1079 ('pack_uvec2_to_uint',
1080 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1081 'options->lower_pack_snorm_2x16'),
1083 (('pack_snorm_4x8', 'v'),
1084 ('pack_uvec4_to_uint',
1085 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1086 'options->lower_pack_snorm_4x8'),
1088 (('unpack_unorm_2x16', 'v'),
1089 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1090 ('extract_u16', 'v', 1))),
1092 'options->lower_unpack_unorm_2x16'),
1094 (('unpack_unorm_4x8', 'v'),
1095 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1096 ('extract_u8', 'v', 1),
1097 ('extract_u8', 'v', 2),
1098 ('extract_u8', 'v', 3))),
1100 'options->lower_unpack_unorm_4x8'),
1102 (('unpack_snorm_2x16', 'v'),
1103 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1104 ('extract_i16', 'v', 1))),
1106 'options->lower_unpack_snorm_2x16'),
1108 (('unpack_snorm_4x8', 'v'),
1109 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1110 ('extract_i8', 'v', 1),
1111 ('extract_i8', 'v', 2),
1112 ('extract_i8', 'v', 3))),
1114 'options->lower_unpack_snorm_4x8'),
1116 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
1117 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
1119 # Address/offset calculations:
1120 # Drivers supporting imul24 should use the nir_lower_amul() pass, this
1121 # rule converts everyone else to imul:
1122 (('amul', a
, b
), ('imul', a
, b
), '!options->has_imul24'),
1124 (('imad24_ir3', a
, b
, 0), ('imul24', a
, b
)),
1125 (('imad24_ir3', a
, 0, c
), (c
)),
1126 (('imad24_ir3', a
, 1, c
), ('iadd', a
, c
)),
1128 # if first two srcs are const, crack apart the imad so constant folding
1129 # can clean up the imul:
1130 # TODO ffma should probably get a similar rule:
1131 (('imad24_ir3', '#a', '#b', c
), ('iadd', ('imul', a
, b
), c
)),
1133 # These will turn 24b address/offset calc back into 32b shifts, but
1134 # it should be safe to get back some of the bits of precision that we
1135 # already decided were no necessary:
1136 (('imul24', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitops'),
1137 (('imul24', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
1138 (('imul24', a
, 0), (0)),
1141 # bit_size dependent lowerings
1142 for bit_size
in [8, 16, 32, 64]:
1143 # convenience constants
1144 intmax
= (1 << (bit_size
- 1)) - 1
1145 intmin
= 1 << (bit_size
- 1)
1148 (('iadd_sat@' + str(bit_size
), a
, b
),
1149 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
1150 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
1151 (('isub_sat@' + str(bit_size
), a
, b
),
1152 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
1153 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
1156 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
1158 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
1159 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1160 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1161 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1162 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1164 # Optimize x2bN(b2x(x)) -> x
1165 for size
in type_sizes('bool'):
1166 aN
= 'a@' + str(size
)
1167 f2bN
= 'f2b' + str(size
)
1168 i2bN
= 'i2b' + str(size
)
1169 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
1170 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
1172 # Optimize x2yN(b2x(x)) -> b2y
1173 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1174 if x
!= 'f' and y
!= 'f' and x
!= y
:
1177 b2x
= 'b2f' if x
== 'f' else 'b2i'
1178 b2y
= 'b2f' if y
== 'f' else 'b2i'
1179 x2yN
= '{}2{}'.format(x
, y
)
1180 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1182 # Optimize away x2xN(a@N)
1183 for t
in ['int', 'uint', 'float']:
1184 for N
in type_sizes(t
):
1185 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1186 aN
= 'a@{0}'.format(N
)
1187 optimizations
.append(((x2xN
, aN
), a
))
1189 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1190 # In particular, we can optimize away everything except upcast of downcast and
1191 # upcasts where the type differs from the other cast
1192 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1194 # The outer cast is a down-cast. It doesn't matter what the size of the
1195 # argument of the inner cast is because we'll never been in the upcast
1196 # of downcast case. Regardless of types, we'll always end up with y2yN
1198 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1199 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1200 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1201 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1202 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1204 # If the outer cast is an up-cast, we have to be more careful about the
1205 # size of the argument of the inner cast and with types. In this case,
1206 # the type is always the type of type up-cast which is given by the
1208 for P
in type_sizes('uint'):
1209 # We can't optimize away up-cast of down-cast.
1213 # Because we're doing down-cast of down-cast, the types always have
1214 # to match between the two casts
1215 for x
in ['i', 'u']:
1216 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1217 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1218 aP
= 'a@{0}'.format(P
)
1219 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1221 # The N == M case is handled by other optimizations
1224 # Optimize comparisons with up-casts
1225 for t
in ['int', 'uint', 'float']:
1226 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1227 if N
== 1 or N
>= M
:
1230 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1231 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1234 xeq
= 'feq' if t
== 'float' else 'ieq'
1235 xne
= 'fne' if t
== 'float' else 'ine'
1236 xge
= '{0}ge'.format(t
[0])
1237 xlt
= '{0}lt'.format(t
[0])
1239 # Up-casts are lossless so for correctly signed comparisons of
1240 # up-casted values we can do the comparison at the largest of the two
1241 # original sizes and drop one or both of the casts. (We have
1242 # optimizations to drop the no-op casts which this may generate.)
1243 for P
in type_sizes(t
):
1249 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1250 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1251 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1252 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1253 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1254 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1257 # The next bit doesn't work on floats because the range checks would
1258 # get way too complicated.
1259 if t
in ['int', 'uint']:
1261 xN_min
= -(1 << (N
- 1))
1262 xN_max
= (1 << (N
- 1)) - 1
1265 xN_max
= (1 << N
) - 1
1269 # If we're up-casting and comparing to a constant, we can unfold
1270 # the comparison into a comparison with the shrunk down constant
1271 # and a check that the constant fits in the smaller bit size.
1273 ((xeq
, (x2xM
, aN
), '#b'),
1274 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1275 ((xne
, (x2xM
, aN
), '#b'),
1276 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1277 ((xlt
, (x2xM
, aN
), '#b'),
1278 ('iand', (xlt
, xN_min
, b
),
1279 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1280 ((xlt
, '#a', (x2xM
, bN
)),
1281 ('iand', (xlt
, a
, xN_max
),
1282 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1283 ((xge
, (x2xM
, aN
), '#b'),
1284 ('iand', (xge
, xN_max
, b
),
1285 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1286 ((xge
, '#a', (x2xM
, bN
)),
1287 ('iand', (xge
, a
, xN_min
),
1288 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1291 def fexp2i(exp
, bits
):
1292 # We assume that exp is already in the right range.
1294 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1296 return ('ishl', ('iadd', exp
, 127), 23)
1298 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1302 def ldexp(f
, exp
, bits
):
1303 # First, we clamp exp to a reasonable range. The maximum possible range
1304 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1305 # a maximum range of [-149, 127]. This means that we can potentially have
1306 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1307 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1308 # spec, on the other hand, only requires that we handle an exponent value
1309 # in the range [-126, 128]. This implementation is *mostly* correct; it
1310 # handles a range on exp of [-252, 254] which allows you to create any
1311 # value (including denorms if the hardware supports it) and to adjust the
1312 # exponent of any normal value to anything you want.
1314 exp
= ('imin', ('imax', exp
, -28), 30)
1316 exp
= ('imin', ('imax', exp
, -252), 254)
1318 exp
= ('imin', ('imax', exp
, -2044), 2046)
1322 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1323 # (We use ishr which isn't the same for -1, but the -1 case still works
1324 # since we use exp-exp/2 as the second exponent.) While the spec
1325 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1326 # work with denormals and doesn't allow for the full swing in exponents
1327 # that you can get with normalized values. Instead, we create two powers
1328 # of two and multiply by them each in turn. That way the effective range
1329 # of our exponent is doubled.
1330 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1331 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1332 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1335 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1336 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1337 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1340 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1341 def bitfield_reverse(u
):
1342 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1343 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1344 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1345 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1346 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1350 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1352 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1353 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1354 # and, if a is a NaN then the second comparison will fail anyway.
1355 for op
in ['flt', 'fge', 'feq']:
1357 (('iand', ('feq', a
, a
), (op
, a
, b
)), ('!' + op
, a
, b
)),
1358 (('iand', ('feq', a
, a
), (op
, b
, a
)), ('!' + op
, b
, a
)),
1361 # Add optimizations to handle the case where the result of a ternary is
1362 # compared to a constant. This way we can take things like
1368 # a ? (0 > 0) : (1 > 0)
1370 # which constant folding will eat for lunch. The resulting ternary will
1371 # further get cleaned up by the boolean reductions above and we will be
1372 # left with just the original variable "a".
1373 for op
in ['flt', 'fge', 'feq', 'fne',
1374 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1376 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1377 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1378 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1379 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1383 # For example, this converts things like
1385 # 1 + mix(0, a - 1, condition)
1389 # mix(1, (a-1)+1, condition)
1391 # Other optimizations will rearrange the constants.
1392 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1394 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1397 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1400 # If neither layout qualifier is specified, derivatives in compute shaders
1401 # return zero, which is consistent with the handling of built-in texture
1402 # functions like texture() in GLSL 4.50 compute shaders.
1403 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1404 'fddy', 'fddy_fine', 'fddy_coarse']:
1406 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1409 # Some optimizations for ir3-specific instructions.
1411 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1412 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1413 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1414 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1415 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1418 # This section contains "late" optimizations that should be run before
1419 # creating ffmas and calling regular optimizations for the final time.
1420 # Optimizations should go here if they help code generation and conflict
1421 # with the regular optimizations.
1422 before_ffma_optimizations
= [
1423 # Propagate constants down multiplication chains
1424 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1425 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1426 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1427 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1429 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1430 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1431 (('~fadd', ('fneg', a
), a
), 0.0),
1432 (('iadd', ('ineg', a
), a
), 0),
1433 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1434 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1435 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1436 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1438 (('~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
)),
1439 (('~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
)),
1440 (('~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
))),
1443 # This section contains "late" optimizations that should be run after the
1444 # regular optimizations have finished. Optimizations should go here if
1445 # they help code generation but do not necessarily produce code that is
1446 # more easily optimizable.
1447 late_optimizations
= [
1448 # Most of these optimizations aren't quite safe when you get infinity or
1449 # Nan involved but the first one should be fine.
1450 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1451 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1452 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1453 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1454 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1455 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1457 # nir_lower_to_source_mods will collapse this, but its existence during the
1458 # optimization loop can prevent other optimizations.
1459 (('fneg', ('fneg', a
)), a
),
1461 # Subtractions get lowered during optimization, so we need to recombine them
1462 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
1463 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
1464 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
1465 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
1467 # These are duplicated from the main optimizations table. The late
1468 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1469 # new patterns like these. The patterns that compare with zero are removed
1470 # because they are unlikely to be created in by anything in
1471 # late_optimizations.
1472 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1473 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1474 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1475 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1476 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1477 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1479 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1480 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1482 (('~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
)))),
1484 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1485 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1486 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1487 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1488 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1489 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1490 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1491 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1492 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1493 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1496 (('iand', a
, a
), a
),
1498 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1500 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1501 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1502 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1503 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1505 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1506 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1508 (('~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'),
1509 (('~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'),
1511 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1512 # particular operation is common for expanding values stored in a texture
1513 # from [0,1] to [-1,1].
1514 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1515 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1516 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1517 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1518 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1519 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1520 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1521 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1525 # a + -a*a + a*b (1)
1527 # Option 1: ffma(a, (b-a), a)
1529 # Alternately, after (1):
1535 # Option 2: ffma(a, 2, -(a*a))
1536 # Option 3: ffma(a, 2, (-a)*a)
1537 # Option 4: ffma(a, -a, (2*a)
1538 # Option 5: a * (2 - a)
1540 # There are a lot of other possible combinations.
1541 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1542 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1543 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1544 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1545 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1547 # we do these late so that we don't get in the way of creating ffmas
1548 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1549 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1551 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1553 # Things that look like DPH in the source shader may get expanded to
1554 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1555 # to NIR. After FFMA is generated, this can look like:
1557 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1559 # Reassociate the last addition into the first multiplication.
1561 # Some shaders do not use 'invariant' in vertex and (possibly) geometry
1562 # shader stages on some outputs that are intended to be invariant. For
1563 # various reasons, this optimization may not be fully applied in all
1564 # shaders used for different rendering passes of the same geometry. This
1565 # can result in Z-fighting artifacts (at best). For now, disable this
1566 # optimization in these stages. See bugzilla #111490. In tessellation
1567 # stages applications seem to use 'precise' when necessary, so allow the
1568 # optimization in those stages.
1569 (('~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)'),
1570 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1571 (('~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)'),
1572 ('ffma', a
, b
, ('ffma', c
, d
, e
)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1575 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1576 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1577 before_ffma_optimizations
).render())
1578 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1579 late_optimizations
).render())