d775a23fceec574d8c56ac77095aca36894823e5
[mesa.git] / src / compiler / nir / nir_opt_algebraic.py
1 #
2 # Copyright (C) 2014 Intel Corporation
3 #
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:
10 #
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
13 # Software.
14 #
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
21 # IN THE SOFTWARE.
22 #
23 # Authors:
24 # Jason Ekstrand (jason@jlekstrand.net)
25
26 from __future__ import print_function
27
28 from collections import OrderedDict
29 import nir_algebraic
30 from nir_opcodes import type_sizes
31 import itertools
32 import struct
33 from math import pi
34
35 # Convenience variables
36 a = 'a'
37 b = 'b'
38 c = 'c'
39 d = 'd'
40 e = 'e'
41
42 # Written in the form (<search>, <replace>) where <search> is an expression
43 # and <replace> is either an expression or a value. An expression is
44 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
45 # where each source is either an expression or a value. A value can be
46 # either a numeric constant or a string representing a variable name.
47 #
48 # If the opcode in a search expression is prefixed by a '~' character, this
49 # indicates that the operation is inexact. Such operations will only get
50 # applied to SSA values that do not have the exact bit set. This should be
51 # used by by any optimizations that are not bit-for-bit exact. It should not,
52 # however, be used for backend-requested lowering operations as those need to
53 # happen regardless of precision.
54 #
55 # Variable names are specified as "[#]name[@type][(cond)][.swiz]" where:
56 # "#" indicates that the given variable will only match constants,
57 # type indicates that the given variable will only match values from ALU
58 # instructions with the given output type,
59 # (cond) specifies an additional condition function (see nir_search_helpers.h),
60 # swiz is a swizzle applied to the variable (only in the <replace> expression)
61 #
62 # For constants, you have to be careful to make sure that it is the right
63 # type because python is unaware of the source and destination types of the
64 # opcodes.
65 #
66 # All expression types can have a bit-size specified. For opcodes, this
67 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
68 # type and size. In the search half of the expression this indicates that it
69 # should only match that particular bit-size. In the replace half of the
70 # expression this indicates that the constructed value should have that
71 # bit-size.
72 #
73 # If the opcode in a replacement expression is prefixed by a '!' character,
74 # this indicated that the new expression will be marked exact.
75 #
76 # A special condition "many-comm-expr" can be used with expressions to note
77 # that the expression and its subexpressions have more commutative expressions
78 # than nir_replace_instr can handle. If this special condition is needed with
79 # another condition, the two can be separated by a comma (e.g.,
80 # "(many-comm-expr,is_used_once)").
81
82 # based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
83 def lowered_sincos(c):
84 x = ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi, a), c))), 1.0)
85 x = ('fmul', ('fsub', x, ('fmul', x, ('fabs', x))), 4.0)
86 return ('ffma', ('ffma', x, ('fabs', x), ('fneg', x)), 0.225, x)
87
88 def intBitsToFloat(i):
89 return struct.unpack('!f', struct.pack('!I', i))[0]
90
91 optimizations = [
92
93 (('imul', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'),
94 (('imul', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'),
95 (('ishl', a, '#b@32'), ('imul', a, ('ishl', 1, b)), 'options->lower_bitops'),
96
97 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
98 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
99 (('imul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('imul_high', a, b)), 'options->lower_mul_2x32_64'),
100 (('umul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('umul_high', a, b)), 'options->lower_mul_2x32_64'),
101 (('udiv', a, 1), a),
102 (('idiv', a, 1), a),
103 (('umod', a, 1), 0),
104 (('imod', a, 1), 0),
105 (('udiv', a, '#b@32(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b)), '!options->lower_bitops'),
106 (('idiv', a, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), 'options->lower_idiv'),
107 (('idiv', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', ('iabs', b))))), 'options->lower_idiv'),
108 (('umod', a, '#b(is_pos_power_of_two)'), ('iand', a, ('isub', b, 1))),
109
110 (('~fneg', ('fneg', a)), a),
111 (('ineg', ('ineg', a)), a),
112 (('fabs', ('fneg', a)), ('fabs', a)),
113 (('fabs', ('u2f', a)), ('u2f', a)),
114 (('iabs', ('iabs', a)), ('iabs', a)),
115 (('iabs', ('ineg', a)), ('iabs', a)),
116 (('f2b', ('fneg', a)), ('f2b', a)),
117 (('i2b', ('ineg', a)), ('i2b', a)),
118 (('~fadd', a, 0.0), a),
119 (('iadd', a, 0), a),
120 (('usadd_4x8', a, 0), a),
121 (('usadd_4x8', a, ~0), ~0),
122 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
123 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
124 (('iand', ('ior', a, b), ('ior', a, c)), ('ior', a, ('iand', b, c))),
125 (('ior', ('iand', a, b), ('iand', a, c)), ('iand', a, ('ior', b, c))),
126 (('~fadd', ('fneg', a), a), 0.0),
127 (('iadd', ('ineg', a), a), 0),
128 (('iadd', ('ineg', a), ('iadd', a, b)), b),
129 (('iadd', a, ('iadd', ('ineg', a), b)), b),
130 (('~fadd', ('fneg', a), ('fadd', a, b)), b),
131 (('~fadd', a, ('fadd', ('fneg', a), b)), b),
132 (('fadd', ('fsat', a), ('fsat', ('fneg', a))), ('fsat', ('fabs', a))),
133 (('~fmul', a, 0.0), 0.0),
134 (('imul', a, 0), 0),
135 (('umul_unorm_4x8', a, 0), 0),
136 (('umul_unorm_4x8', a, ~0), a),
137 (('~fmul', a, 1.0), a),
138 (('imul', a, 1), a),
139 (('fmul', a, -1.0), ('fneg', a)),
140 (('imul', a, -1), ('ineg', a)),
141 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
142 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
143 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
144 (('fmul', ('fsign', a), ('fmul', a, a)), ('fmul', ('fabs', a), a)),
145 (('fmul', ('fmul', ('fsign', a), a), a), ('fmul', ('fabs', a), a)),
146 (('~ffma', 0.0, a, b), b),
147 (('~ffma', a, b, 0.0), ('fmul', a, b)),
148 (('ffma', 1.0, a, b), ('fadd', a, b)),
149 (('ffma', -1.0, a, b), ('fadd', ('fneg', a), b)),
150 (('~flrp', a, b, 0.0), a),
151 (('~flrp', a, b, 1.0), b),
152 (('~flrp', a, a, b), a),
153 (('~flrp', 0.0, a, b), ('fmul', a, b)),
154
155 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
156 (('~flrp', a, ('fadd(is_used_once)', a, b), c), ('fadd', ('fmul', b, c), a)),
157 (('~flrp@32', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp32'),
158 (('~flrp@64', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp64'),
159
160 (('~flrp@32', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp32'),
161 (('~flrp@64', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp64'),
162
163 (('~flrp@32', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp32'),
164 (('~flrp@64', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp64'),
165
166 (('~flrp', ('fmul(is_used_once)', a, b), ('fmul(is_used_once)', a, c), d), ('fmul', ('flrp', b, c, d), a)),
167
168 (('~flrp', a, b, ('b2f', 'c@1')), ('bcsel', c, b, a), 'options->lower_flrp32'),
169 (('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)),
170 (('ftrunc', a), ('bcsel', ('flt', a, 0.0), ('fneg', ('ffloor', ('fabs', a))), ('ffloor', ('fabs', a))), 'options->lower_ftrunc'),
171 (('ffloor', a), ('fsub', a, ('ffract', a)), 'options->lower_ffloor'),
172 (('fadd', a, ('fneg', ('ffract', a))), ('ffloor', a), '!options->lower_ffloor'),
173 (('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'),
174 (('fceil', a), ('fneg', ('ffloor', ('fneg', a))), 'options->lower_fceil'),
175 (('~fadd', ('fmul', a, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b, ('b2f', c))), ('bcsel', c, b, a), 'options->lower_flrp32'),
176 (('~fadd@32', ('fmul', a, ('fadd', 1.0, ('fneg', c ) )), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp32'),
177 (('~fadd@64', ('fmul', a, ('fadd', 1.0, ('fneg', c ) )), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp64'),
178 # These are the same as the previous three rules, but it depends on
179 # 1-fsat(x) <=> fsat(1-x). See below.
180 (('~fadd@32', ('fmul', a, ('fsat', ('fadd', 1.0, ('fneg', c )))), ('fmul', b, ('fsat', c))), ('flrp', a, b, ('fsat', c)), '!options->lower_flrp32'),
181 (('~fadd@64', ('fmul', a, ('fsat', ('fadd', 1.0, ('fneg', c )))), ('fmul', b, ('fsat', c))), ('flrp', a, b, ('fsat', c)), '!options->lower_flrp64'),
182
183 (('~fadd', a, ('fmul', ('b2f', 'c@1'), ('fadd', b, ('fneg', a)))), ('bcsel', c, b, a), 'options->lower_flrp32'),
184 (('~fadd@32', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp32'),
185 (('~fadd@64', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp64'),
186 (('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'),
187 (('~fadd', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma'),
188
189 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b, c)), '#d'), '#e'),
190 ('bcsel', a, ('fmul', ('fadd', ('fmul', b, c), d), e), ('fmul', d, e))),
191
192 (('fdph', a, b), ('fdot4', ('vec4', 'a.x', 'a.y', 'a.z', 1.0), b), 'options->lower_fdph'),
193
194 (('fdot4', ('vec4', a, b, c, 1.0), d), ('fdph', ('vec3', a, b, c), d), '!options->lower_fdph'),
195 (('fdot4', ('vec4', a, 0.0, 0.0, 0.0), b), ('fmul', a, b)),
196 (('fdot4', ('vec4', a, b, 0.0, 0.0), c), ('fdot2', ('vec2', a, b), c)),
197 (('fdot4', ('vec4', a, b, c, 0.0), d), ('fdot3', ('vec3', a, b, c), d)),
198
199 (('fdot3', ('vec3', a, 0.0, 0.0), b), ('fmul', a, b)),
200 (('fdot3', ('vec3', a, b, 0.0), c), ('fdot2', ('vec2', a, b), c)),
201
202 (('fdot2', ('vec2', a, 0.0), b), ('fmul', a, b)),
203 (('fdot2', a, 1.0), ('fadd', 'a.x', 'a.y')),
204
205 # Lower fdot to fsum when it is available
206 (('fdot2', a, b), ('fsum2', ('fmul', a, b)), 'options->lower_fdot'),
207 (('fdot3', a, b), ('fsum3', ('fmul', a, b)), 'options->lower_fdot'),
208 (('fdot4', a, b), ('fsum4', ('fmul', a, b)), 'options->lower_fdot'),
209 (('fsum2', a), ('fadd', 'a.x', 'a.y'), 'options->lower_fdot'),
210
211 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
212 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
213 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
214 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),
215
216 # 1 - ((1 - a) * (1 - b))
217 # 1 - (1 - a - b + a*b)
218 # 1 - 1 + a + b - a*b
219 # a + b - a*b
220 # a + b*(1 - a)
221 # b*(1 - a) + 1*a
222 # flrp(b, 1, a)
223 (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a)), ('fadd', 1.0, ('fneg', b))))),
224 ('flrp', b, 1.0, a), '!options->lower_flrp32'),
225
226 # (a * #b + #c) << #d
227 # ((a * #b) << #d) + (#c << #d)
228 # (a * (#b << #d)) + (#c << #d)
229 (('ishl', ('iadd', ('imul', a, '#b'), '#c'), '#d'),
230 ('iadd', ('imul', a, ('ishl', b, d)), ('ishl', c, d))),
231
232 # (a * #b) << #c
233 # a * (#b << #c)
234 (('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))),
235 ]
236
237 # Care must be taken here. Shifts in NIR uses only the lower log2(bitsize)
238 # bits of the second source. These replacements must correctly handle the
239 # case where (b % bitsize) + (c % bitsize) >= bitsize.
240 for s in [8, 16, 32, 64]:
241 mask = (1 << s) - 1
242
243 ishl = "ishl@{}".format(s)
244 ishr = "ishr@{}".format(s)
245 ushr = "ushr@{}".format(s)
246
247 in_bounds = ('ult', ('iadd', ('iand', b, mask), ('iand', c, mask)), s)
248
249 optimizations.extend([
250 ((ishl, (ishl, a, '#b'), '#c'), ('bcsel', in_bounds, (ishl, a, ('iadd', b, c)), 0)),
251 ((ushr, (ushr, a, '#b'), '#c'), ('bcsel', in_bounds, (ushr, a, ('iadd', b, c)), 0)),
252
253 # To get get -1 for large shifts of negative values, ishr must instead
254 # clamp the shift count to the maximum value.
255 ((ishr, (ishr, a, '#b'), '#c'),
256 (ishr, a, ('imin', ('iadd', ('iand', b, mask), ('iand', c, mask)), s - 1))),
257 ])
258
259 # Optimize a pattern of address calculation created by DXVK where the offset is
260 # divided by 4 and then multipled by 4. This can be turned into an iand and the
261 # additions before can be reassociated to CSE the iand instruction.
262 for log2 in range(1, 7): # powers of two from 2 to 64
263 v = 1 << log2
264 mask = 0xffffffff & ~(v - 1)
265 b_is_multiple = '#b(is_unsigned_multiple_of_{})'.format(v)
266
267 optimizations.extend([
268 # 'a >> #b << #b' -> 'a & ~((1 << #b) - 1)'
269 (('ishl@32', ('ushr@32', a, log2), log2), ('iand', a, mask)),
270
271 # Reassociate for improved CSE
272 (('iand@32', ('iadd@32', a, b_is_multiple), mask), ('iadd', ('iand', a, mask), b)),
273 ])
274
275 # To save space in the state tables, reduce to the set that is known to help.
276 # Previously, this was range(1, 32). In addition, a couple rules inside the
277 # loop are commented out. Revisit someday, probably after mesa/#2635 has some
278 # resolution.
279 for i in [1, 2, 16, 24]:
280 lo_mask = 0xffffffff >> i
281 hi_mask = (0xffffffff << i) & 0xffffffff
282
283 optimizations.extend([
284 # This pattern seems to only help in the soft-fp64 code.
285 (('ishl@32', ('iand', 'a@32', lo_mask), i), ('ishl', a, i)),
286 # (('ushr@32', ('iand', 'a@32', hi_mask), i), ('ushr', a, i)),
287 # (('ishr@32', ('iand', 'a@32', hi_mask), i), ('ishr', a, i)),
288
289 (('iand', ('ishl', 'a@32', i), hi_mask), ('ishl', a, i)),
290 (('iand', ('ushr', 'a@32', i), lo_mask), ('ushr', a, i)),
291 # (('iand', ('ishr', 'a@32', i), lo_mask), ('ushr', a, i)), # Yes, ushr is correct
292 ])
293
294 optimizations.extend([
295 # This is common for address calculations. Reassociating may enable the
296 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
297 # instruction or a constant offset field for in load / store instructions.
298 (('ishl', ('iadd', a, '#b'), '#c'), ('iadd', ('ishl', a, c), ('ishl', b, c))),
299
300 # Comparison simplifications
301 (('~inot', ('flt', a, b)), ('fge', a, b)),
302 (('~inot', ('fge', a, b)), ('flt', a, b)),
303 (('inot', ('feq', a, b)), ('fneu', a, b)),
304 (('inot', ('fneu', a, b)), ('feq', a, b)),
305 (('inot', ('ilt', a, b)), ('ige', a, b)),
306 (('inot', ('ult', a, b)), ('uge', a, b)),
307 (('inot', ('ige', a, b)), ('ilt', a, b)),
308 (('inot', ('uge', a, b)), ('ult', a, b)),
309 (('inot', ('ieq', a, b)), ('ine', a, b)),
310 (('inot', ('ine', a, b)), ('ieq', a, b)),
311
312 (('iand', ('feq', a, b), ('fneu', a, b)), False),
313 (('iand', ('flt', a, b), ('flt', b, a)), False),
314 (('iand', ('ieq', a, b), ('ine', a, b)), False),
315 (('iand', ('ilt', a, b), ('ilt', b, a)), False),
316 (('iand', ('ult', a, b), ('ult', b, a)), False),
317
318 # This helps some shaders because, after some optimizations, they end up
319 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
320 # matching would be handled by CSE.
321 (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
322 (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
323 (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
324 (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)),
325 (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
326 (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
327 (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
328 (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
329 (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)),
330 (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),
331
332 # flt(fsat(a), b > 0 && b < 1) is inexact if a is NaN (fsat(NaN) is 0)
333 # because it returns True while flt(a, b) always returns False.
334 (('~flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
335 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
336 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
337 # fge(b > 0 && b < 1, fsat(a)) is inexact if a is NaN (fsat(NaN) is 0)
338 # because it returns True while fge(b, a) always returns False.
339 (('~fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
340 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
341 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),
342
343 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
344 (('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
345 (('fge', 0.0, ('fsat(is_used_once)', a)), ('fge', 0.0, a)),
346 (('flt', 0.0, ('fsat(is_used_once)', a)), ('flt', 0.0, a)),
347
348 # 0.0 >= b2f(a)
349 # b2f(a) <= 0.0
350 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
351 # inot(a)
352 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a)),
353
354 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a)),
355
356 (('fneu', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
357 (('fneu', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
358 (('fneu', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a, b)),
359 (('fneu', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a, b)),
360 (('fneu', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
361 (('fneu', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
362 (('fneu', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a, b)),
363 (('fneu', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a, b)),
364 (('fneu', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a, b)),
365 (('fneu', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a, b)),
366 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
367 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
368 (('feq', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a, b))),
369 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
370 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
371 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
372 (('feq', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a, b))),
373 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a, b)),
374 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a, b)),
375 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a, b)),
376
377 # -(b2f(a) + b2f(b)) < 0
378 # 0 < b2f(a) + b2f(b)
379 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
380 # a || b
381 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a, b)),
382 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a, b)),
383
384 # -(b2f(a) + b2f(b)) >= 0
385 # 0 >= b2f(a) + b2f(b)
386 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
387 # !(a || b)
388 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))),
389 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
390
391 (('flt', a, ('fneg', a)), ('flt', a, 0.0)),
392 (('fge', a, ('fneg', a)), ('fge', a, 0.0)),
393
394 # Some optimizations (below) convert things like (a < b || c < b) into
395 # (min(a, c) < b). However, this interfers with the previous optimizations
396 # that try to remove comparisons with negated sums of b2f. This just
397 # breaks that apart.
398 (('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
399 ('ior', ('flt', c, 0.0), ('ior', a, b))),
400
401 (('~flt', ('fadd', a, b), a), ('flt', b, 0.0)),
402 (('~fge', ('fadd', a, b), a), ('fge', b, 0.0)),
403 (('~feq', ('fadd', a, b), a), ('feq', b, 0.0)),
404 (('~fneu', ('fadd', a, b), a), ('fneu', b, 0.0)),
405 (('~flt', ('fadd(is_used_once)', a, '#b'), '#c'), ('flt', a, ('fadd', c, ('fneg', b)))),
406 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('flt', ('fneg', ('fadd', c, b)), a)),
407 (('~fge', ('fadd(is_used_once)', a, '#b'), '#c'), ('fge', a, ('fadd', c, ('fneg', b)))),
408 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fge', ('fneg', ('fadd', c, b)), a)),
409 (('~feq', ('fadd(is_used_once)', a, '#b'), '#c'), ('feq', a, ('fadd', c, ('fneg', b)))),
410 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('feq', ('fneg', ('fadd', c, b)), a)),
411 (('~fneu', ('fadd(is_used_once)', a, '#b'), '#c'), ('fneu', a, ('fadd', c, ('fneg', b)))),
412 (('~fneu', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fneu', ('fneg', ('fadd', c, b)), a)),
413
414 # Cannot remove the addition from ilt or ige due to overflow.
415 (('ieq', ('iadd', a, b), a), ('ieq', b, 0)),
416 (('ine', ('iadd', a, b), a), ('ine', b, 0)),
417
418 # fmin(-b2f(a), b) >= 0.0
419 # -b2f(a) >= 0.0 && b >= 0.0
420 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
421 # b2f(a) == 0.0 && b >= 0.0
422 # a == False && b >= 0.0
423 # !a && b >= 0.0
424 #
425 # The fge in the second replacement is not a typo. I leave the proof that
426 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
427 # reader.
428 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
429 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
430
431 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a)),
432 (('~fneu', ('b2f', 'a@1'), 0.0), a),
433 (('ieq', ('b2i', 'a@1'), 0), ('inot', a)),
434 (('ine', ('b2i', 'a@1'), 0), a),
435
436 (('fneu', ('u2f', a), 0.0), ('ine', a, 0)),
437 (('feq', ('u2f', a), 0.0), ('ieq', a, 0)),
438 (('fge', ('u2f', a), 0.0), True),
439 (('fge', 0.0, ('u2f', a)), ('uge', 0, a)), # ieq instead?
440 (('flt', ('u2f', a), 0.0), False),
441 (('flt', 0.0, ('u2f', a)), ('ult', 0, a)), # ine instead?
442 (('fneu', ('i2f', a), 0.0), ('ine', a, 0)),
443 (('feq', ('i2f', a), 0.0), ('ieq', a, 0)),
444 (('fge', ('i2f', a), 0.0), ('ige', a, 0)),
445 (('fge', 0.0, ('i2f', a)), ('ige', 0, a)),
446 (('flt', ('i2f', a), 0.0), ('ilt', a, 0)),
447 (('flt', 0.0, ('i2f', a)), ('ilt', 0, a)),
448
449 # 0.0 < fabs(a)
450 # fabs(a) > 0.0
451 # fabs(a) != 0.0 because fabs(a) must be >= 0
452 # a != 0.0
453 (('~flt', 0.0, ('fabs', a)), ('fneu', a, 0.0)),
454
455 # -fabs(a) < 0.0
456 # fabs(a) > 0.0
457 (('~flt', ('fneg', ('fabs', a)), 0.0), ('fneu', a, 0.0)),
458
459 # 0.0 >= fabs(a)
460 # 0.0 == fabs(a) because fabs(a) must be >= 0
461 # 0.0 == a
462 (('fge', 0.0, ('fabs', a)), ('feq', a, 0.0)),
463
464 # -fabs(a) >= 0.0
465 # 0.0 >= fabs(a)
466 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
467
468 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
469 (('iand', ('fge', a, 0.0), ('fge', 1.0, a)), ('feq', a, ('fsat', a)), '!options->lower_fsat'),
470
471 # (a < 0.0) || (a > 1.0)
472 # !(!(a < 0.0) && !(a > 1.0))
473 # !((a >= 0.0) && (a <= 1.0))
474 # !(a == fsat(a))
475 # a != fsat(a)
476 (('ior', ('flt', a, 0.0), ('flt', 1.0, a)), ('fneu', a, ('fsat', a)), '!options->lower_fsat'),
477
478 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a, b))),
479 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a, b)))),
480 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
481 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a, b)))),
482
483 # fmin(b2f(a), b)
484 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
485 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
486 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
487 #
488 # Since b is a constant, constant folding will eliminate the fmin and the
489 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
490 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))),
491
492 (('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),
493
494 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
495 (('~bcsel', ('flt', b, a), b, a), ('fmin', a, b)),
496 (('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
497 (('~bcsel', ('fge', a, b), b, a), ('fmin', a, b)),
498 (('~bcsel', ('fge', b, a), b, a), ('fmax', a, b)),
499 (('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)),
500 (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
501 (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
502 (('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)),
503 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
504 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
505 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
506 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
507 (('bcsel', a, True, b), ('ior', a, b)),
508 (('bcsel', a, a, b), ('ior', a, b)),
509 (('bcsel', a, b, False), ('iand', a, b)),
510 (('bcsel', a, b, a), ('iand', a, b)),
511 (('~fmin', a, a), a),
512 (('~fmax', a, a), a),
513 (('imin', a, a), a),
514 (('imax', a, a), a),
515 (('umin', a, a), a),
516 (('umax', a, a), a),
517 (('fmax', ('fmax', a, b), b), ('fmax', a, b)),
518 (('umax', ('umax', a, b), b), ('umax', a, b)),
519 (('imax', ('imax', a, b), b), ('imax', a, b)),
520 (('fmin', ('fmin', a, b), b), ('fmin', a, b)),
521 (('umin', ('umin', a, b), b), ('umin', a, b)),
522 (('imin', ('imin', a, b), b), ('imin', a, b)),
523 (('iand@32', a, ('inot', ('ishr', a, 31))), ('imax', a, 0)),
524
525 # Simplify logic to detect sign of an integer.
526 (('ieq', ('iand', 'a@32', 0x80000000), 0x00000000), ('ige', a, 0)),
527 (('ine', ('iand', 'a@32', 0x80000000), 0x80000000), ('ige', a, 0)),
528 (('ine', ('iand', 'a@32', 0x80000000), 0x00000000), ('ilt', a, 0)),
529 (('ieq', ('iand', 'a@32', 0x80000000), 0x80000000), ('ilt', a, 0)),
530 (('ine', ('ushr', 'a@32', 31), 0), ('ilt', a, 0)),
531 (('ieq', ('ushr', 'a@32', 31), 0), ('ige', a, 0)),
532 (('ieq', ('ushr', 'a@32', 31), 1), ('ilt', a, 0)),
533 (('ine', ('ushr', 'a@32', 31), 1), ('ige', a, 0)),
534 (('ine', ('ishr', 'a@32', 31), 0), ('ilt', a, 0)),
535 (('ieq', ('ishr', 'a@32', 31), 0), ('ige', a, 0)),
536 (('ieq', ('ishr', 'a@32', 31), -1), ('ilt', a, 0)),
537 (('ine', ('ishr', 'a@32', 31), -1), ('ige', a, 0)),
538
539 (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
540 (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
541 (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
542 (('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))),
543 (('~fmin', a, ('fabs', a)), a),
544 (('imin', a, ('iabs', a)), a),
545 (('~fmax', a, ('fneg', ('fabs', a))), a),
546 (('imax', a, ('ineg', ('iabs', a))), a),
547 (('fmax', a, ('fabs', a)), ('fabs', a)),
548 (('imax', a, ('iabs', a)), ('iabs', a)),
549 (('fmax', a, ('fneg', a)), ('fabs', a)),
550 (('imax', a, ('ineg', a)), ('iabs', a), '!options->lower_iabs'),
551 (('~fmax', ('fabs', a), 0.0), ('fabs', a)),
552 (('fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'),
553 # fmax(fmin(a, 1.0), 0.0) is inexact because it returns 1.0 on NaN, while
554 # fsat(a) returns 0.0.
555 (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'),
556 # fmin(fmax(a, -1.0), 0.0) is inexact because it returns -1.0 on NaN, while
557 # fneg(fsat(fneg(a))) returns -0.0 on NaN.
558 (('~fmin', ('fmax', a, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
559 # fmax(fmin(a, 0.0), -1.0) is inexact because it returns 0.0 on NaN, while
560 # fneg(fsat(fneg(a))) returns -0.0 on NaN. This only matters if
561 # SignedZeroInfNanPreserve is set, but we don't currently have any way of
562 # representing this in the optimizations other than the usual ~.
563 (('~fmax', ('fmin', a, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
564 (('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))),
565 (('fsat', ('b2f', a)), ('b2f', a)),
566 (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
567 (('fsat', ('fsat', a)), ('fsat', a)),
568 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b))), ('fsat', ('fadd', ('fneg', a), ('fneg', b))), '!options->lower_fsat'),
569 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fneg', a), b)), '!options->lower_fsat'),
570 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fabs', a), ('fabs', b))), '!options->lower_fsat'),
571 (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)),
572 (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)),
573 (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
574 # Both the left and right patterns are "b" when isnan(a), so this is exact.
575 (('fmax', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
576 # The left pattern is 0.0 when isnan(a) (because fmin(fsat(NaN), b) ->
577 # fmin(0.0, b)) while the right one is "b", so this optimization is inexact.
578 (('~fmin', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a, b))),
579
580 # If a in [0,b] then b-a is also in [0,b]. Since b in [0,1], max(b-a, 0) =
581 # fsat(b-a).
582 #
583 # If a > b, then b-a < 0 and max(b-a, 0) = fsat(b-a) = 0
584 #
585 # This should be NaN safe since max(NaN, 0) = fsat(NaN) = 0.
586 (('fmax', ('fadd(is_used_once)', ('fneg', 'a(is_not_negative)'), '#b@32(is_zero_to_one)'), 0.0),
587 ('fsat', ('fadd', ('fneg', a), b)), '!options->lower_fsat'),
588
589 (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
590 (('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
591 (('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
592 (('~ior', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
593 (('~ior', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
594 (('~ior', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmax', b, c))),
595 (('~ior', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmin', a, b), c)),
596 (('~ior', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmin', b, c))),
597 (('~ior', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmax', a, b), c)),
598 (('~iand', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmin', b, c))),
599 (('~iand', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmax', a, b), c)),
600 (('~iand', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmax', b, c))),
601 (('~iand', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmin', a, b), c)),
602 (('~iand', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmin', b, c))),
603 (('~iand', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmax', a, b), c)),
604 (('~iand', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmax', b, c))),
605 (('~iand', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmin', a, b), c)),
606
607 (('ior', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imax', b, c))),
608 (('ior', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imin', a, b), c)),
609 (('ior', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imin', b, c))),
610 (('ior', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imax', a, b), c)),
611 (('ior', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umax', b, c))),
612 (('ior', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umin', a, b), c)),
613 (('ior', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umin', b, c))),
614 (('ior', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umax', a, b), c)),
615 (('iand', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imin', b, c))),
616 (('iand', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imax', a, b), c)),
617 (('iand', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imax', b, c))),
618 (('iand', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imin', a, b), c)),
619 (('iand', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umin', b, c))),
620 (('iand', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umax', a, b), c)),
621 (('iand', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umax', b, c))),
622 (('iand', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umin', a, b), c)),
623
624 # These derive from the previous patterns with the application of b < 0 <=>
625 # 0 < -b. The transformation should be applied if either comparison is
626 # used once as this ensures that the number of comparisons will not
627 # increase. The sources to the ior and iand are not symmetric, so the
628 # rules have to be duplicated to get this behavior.
629 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
630 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
631 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
632 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
633 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
634 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
635 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
636 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
637
638 # Common pattern like 'if (i == 0 || i == 1 || ...)'
639 (('ior', ('ieq', a, 0), ('ieq', a, 1)), ('uge', 1, a)),
640 (('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)),
641 (('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)),
642
643 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
644 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
645 # so emit an open-coded version of that.
646 (('bcsel@32', ('feq', a, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
647 ('ior', 0x3f800000, ('iand', a, 0x80000000))),
648
649 (('ior', a, ('ieq', a, False)), True),
650 (('ior', a, ('inot', a)), -1),
651
652 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a, b)),
653 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a, b))),
654
655 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umax', a, b), 0)),
656 (('ior', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umin', a, b), 0)),
657 (('iand', ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umin', a, b), 0)),
658 (('ior', ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umax', a, b), 0)),
659
660 # This pattern occurs coutresy of __flt64_nonnan in the soft-fp64 code.
661 # The first part of the iand comes from the !__feq64_nonnan.
662 #
663 # The second pattern is a reformulation of the first based on the relation
664 # (a == 0 || y == 0) <=> umin(a, y) == 0, where b in the first equation
665 # happens to be y == 0.
666 (('iand', ('inot', ('iand', ('ior', ('ieq', a, 0), b), c)), ('ilt', a, 0)),
667 ('iand', ('inot', ('iand', b , c)), ('ilt', a, 0))),
668 (('iand', ('inot', ('iand', ('ieq', ('umin', a, b), 0), c)), ('ilt', a, 0)),
669 ('iand', ('inot', ('iand', ('ieq', b , 0), c)), ('ilt', a, 0))),
670
671 # These patterns can result when (a < b || a < c) => (a < min(b, c))
672 # transformations occur before constant propagation and loop-unrolling.
673 (('~flt', a, ('fmax', b, a)), ('flt', a, b)),
674 (('~flt', ('fmin', a, b), a), ('flt', b, a)),
675 (('~fge', a, ('fmin', b, a)), True),
676 (('~fge', ('fmax', a, b), a), True),
677 (('~flt', a, ('fmin', b, a)), False),
678 (('~flt', ('fmax', a, b), a), False),
679 (('~fge', a, ('fmax', b, a)), ('fge', a, b)),
680 (('~fge', ('fmin', a, b), a), ('fge', b, a)),
681
682 (('ilt', a, ('imax', b, a)), ('ilt', a, b)),
683 (('ilt', ('imin', a, b), a), ('ilt', b, a)),
684 (('ige', a, ('imin', b, a)), True),
685 (('ige', ('imax', a, b), a), True),
686 (('ult', a, ('umax', b, a)), ('ult', a, b)),
687 (('ult', ('umin', a, b), a), ('ult', b, a)),
688 (('uge', a, ('umin', b, a)), True),
689 (('uge', ('umax', a, b), a), True),
690 (('ilt', a, ('imin', b, a)), False),
691 (('ilt', ('imax', a, b), a), False),
692 (('ige', a, ('imax', b, a)), ('ige', a, b)),
693 (('ige', ('imin', a, b), a), ('ige', b, a)),
694 (('ult', a, ('umin', b, a)), False),
695 (('ult', ('umax', a, b), a), False),
696 (('uge', a, ('umax', b, a)), ('uge', a, b)),
697 (('uge', ('umin', a, b), a), ('uge', b, a)),
698 (('ult', a, ('iand', b, a)), False),
699 (('ult', ('ior', a, b), a), False),
700 (('uge', a, ('iand', b, a)), True),
701 (('uge', ('ior', a, b), a), True),
702
703 (('ilt', '#a', ('imax', '#b', c)), ('ior', ('ilt', a, b), ('ilt', a, c))),
704 (('ilt', ('imin', '#a', b), '#c'), ('ior', ('ilt', a, c), ('ilt', b, c))),
705 (('ige', '#a', ('imin', '#b', c)), ('ior', ('ige', a, b), ('ige', a, c))),
706 (('ige', ('imax', '#a', b), '#c'), ('ior', ('ige', a, c), ('ige', b, c))),
707 (('ult', '#a', ('umax', '#b', c)), ('ior', ('ult', a, b), ('ult', a, c))),
708 (('ult', ('umin', '#a', b), '#c'), ('ior', ('ult', a, c), ('ult', b, c))),
709 (('uge', '#a', ('umin', '#b', c)), ('ior', ('uge', a, b), ('uge', a, c))),
710 (('uge', ('umax', '#a', b), '#c'), ('ior', ('uge', a, c), ('uge', b, c))),
711 (('ilt', '#a', ('imin', '#b', c)), ('iand', ('ilt', a, b), ('ilt', a, c))),
712 (('ilt', ('imax', '#a', b), '#c'), ('iand', ('ilt', a, c), ('ilt', b, c))),
713 (('ige', '#a', ('imax', '#b', c)), ('iand', ('ige', a, b), ('ige', a, c))),
714 (('ige', ('imin', '#a', b), '#c'), ('iand', ('ige', a, c), ('ige', b, c))),
715 (('ult', '#a', ('umin', '#b', c)), ('iand', ('ult', a, b), ('ult', a, c))),
716 (('ult', ('umax', '#a', b), '#c'), ('iand', ('ult', a, c), ('ult', b, c))),
717 (('uge', '#a', ('umax', '#b', c)), ('iand', ('uge', a, b), ('uge', a, c))),
718 (('uge', ('umin', '#a', b), '#c'), ('iand', ('uge', a, c), ('uge', b, c))),
719
720 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
721 # negative.
722 (('bcsel', ('ilt', a, 0), ('ineg', ('ishr', a, b)), ('ishr', a, b)),
723 ('iabs', ('ishr', a, b))),
724 (('iabs', ('ishr', ('iabs', a), b)), ('ishr', ('iabs', a), b)),
725
726 (('fabs', ('slt', a, b)), ('slt', a, b)),
727 (('fabs', ('sge', a, b)), ('sge', a, b)),
728 (('fabs', ('seq', a, b)), ('seq', a, b)),
729 (('fabs', ('sne', a, b)), ('sne', a, b)),
730 (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
731 (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
732 (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
733 (('sne', a, b), ('b2f', ('fneu', a, b)), 'options->lower_scmp'),
734 (('seq', ('seq', a, b), 1.0), ('seq', a, b)),
735 (('seq', ('sne', a, b), 1.0), ('sne', a, b)),
736 (('seq', ('slt', a, b), 1.0), ('slt', a, b)),
737 (('seq', ('sge', a, b), 1.0), ('sge', a, b)),
738 (('sne', ('seq', a, b), 0.0), ('seq', a, b)),
739 (('sne', ('sne', a, b), 0.0), ('sne', a, b)),
740 (('sne', ('slt', a, b), 0.0), ('slt', a, b)),
741 (('sne', ('sge', a, b), 0.0), ('sge', a, b)),
742 (('seq', ('seq', a, b), 0.0), ('sne', a, b)),
743 (('seq', ('sne', a, b), 0.0), ('seq', a, b)),
744 (('seq', ('slt', a, b), 0.0), ('sge', a, b)),
745 (('seq', ('sge', a, b), 0.0), ('slt', a, b)),
746 (('sne', ('seq', a, b), 1.0), ('sne', a, b)),
747 (('sne', ('sne', a, b), 1.0), ('seq', a, b)),
748 (('sne', ('slt', a, b), 1.0), ('sge', a, b)),
749 (('sne', ('sge', a, b), 1.0), ('slt', a, b)),
750 (('fall_equal2', a, b), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
751 (('fall_equal3', a, b), ('seq', ('fany_nequal3', a, b), 0.0), 'options->lower_vector_cmp'),
752 (('fall_equal4', a, b), ('seq', ('fany_nequal4', a, b), 0.0), 'options->lower_vector_cmp'),
753 (('fany_nequal2', a, b), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
754 (('fany_nequal3', a, b), ('fsat', ('fdot3', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),
755 (('fany_nequal4', a, b), ('fsat', ('fdot4', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),
756 (('fneu', ('fneg', a), a), ('fneu', a, 0.0)),
757 (('feq', ('fneg', a), a), ('feq', a, 0.0)),
758 # Emulating booleans
759 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
760 (('iand', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
761 (('ior', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('ior', a, b))),
762 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
763 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a, b))),
764 (('iand', 'a@bool32', 1.0), ('b2f', a)),
765 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
766 (('ineg', ('b2i32', 'a@32')), a),
767 (('flt', ('fneg', ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
768 # Comparison with the same args. Note that these are not done for
769 # the float versions because NaN always returns false on float
770 # inequalities.
771 (('ilt', a, a), False),
772 (('ige', a, a), True),
773 (('ieq', a, a), True),
774 (('ine', a, a), False),
775 (('ult', a, a), False),
776 (('uge', a, a), True),
777 # Logical and bit operations
778 (('iand', a, a), a),
779 (('iand', a, ~0), a),
780 (('iand', a, 0), 0),
781 (('ior', a, a), a),
782 (('ior', a, 0), a),
783 (('ior', a, True), True),
784 (('ixor', a, a), 0),
785 (('ixor', a, 0), a),
786 (('inot', ('inot', a)), a),
787 (('ior', ('iand', a, b), b), b),
788 (('ior', ('ior', a, b), b), ('ior', a, b)),
789 (('iand', ('ior', a, b), b), b),
790 (('iand', ('iand', a, b), b), ('iand', a, b)),
791 # DeMorgan's Laws
792 (('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))),
793 (('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))),
794 # Shift optimizations
795 (('ishl', 0, a), 0),
796 (('ishl', a, 0), a),
797 (('ishr', 0, a), 0),
798 (('ishr', a, 0), a),
799 (('ushr', 0, a), 0),
800 (('ushr', a, 0), a),
801 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('iadd', 16, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
802 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('isub', 16, b))), ('urol', a, b), '!options->lower_rotate'),
803 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('iadd', 32, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
804 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('isub', 32, b))), ('urol', a, b), '!options->lower_rotate'),
805 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('iadd', 16, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
806 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('isub', 16, b))), ('uror', a, b), '!options->lower_rotate'),
807 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('iadd', 32, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
808 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('isub', 32, b))), ('uror', a, b), '!options->lower_rotate'),
809 (('urol@16', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 16, b))), 'options->lower_rotate'),
810 (('urol@32', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 32, b))), 'options->lower_rotate'),
811 (('uror@16', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 16, b))), 'options->lower_rotate'),
812 (('uror@32', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 32, b))), 'options->lower_rotate'),
813 # Exponential/logarithmic identities
814 (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
815 (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
816 (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
817 (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
818 (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
819 ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
820 (('~fexp2', ('fmul', ('flog2', a), 0.5)), ('fsqrt', a)),
821 (('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)),
822 (('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))),
823 (('~fpow', a, 1.0), a),
824 (('~fpow', a, 2.0), ('fmul', a, a)),
825 (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
826 (('~fpow', 2.0, a), ('fexp2', a)),
827 (('~fpow', ('fpow', a, 2.2), 0.454545), a),
828 (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
829 (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
830 (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
831 (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
832 (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
833 (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
834 (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
835 (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
836 (('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))),
837 (('bcsel', ('flt', a, 0.0), 0.0, ('fsqrt', a)), ('fsqrt', ('fmax', a, 0.0))),
838 (('~fmul', ('fsqrt', a), ('fsqrt', a)), ('fabs',a)),
839 # Division and reciprocal
840 (('~fdiv', 1.0, a), ('frcp', a)),
841 (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
842 (('~frcp', ('frcp', a)), a),
843 (('~frcp', ('fsqrt', a)), ('frsq', a)),
844 (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
845 (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
846 # Trig
847 (('fsin', a), lowered_sincos(0.5), 'options->lower_sincos'),
848 (('fcos', a), lowered_sincos(0.75), 'options->lower_sincos'),
849 # Boolean simplifications
850 (('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
851 (('i2b1(is_used_by_if)', a), ('ine', a, 0)),
852 (('ieq', a, True), a),
853 (('ine(is_not_used_by_if)', a, True), ('inot', a)),
854 (('ine', a, False), a),
855 (('ieq(is_not_used_by_if)', a, False), ('inot', 'a')),
856 (('bcsel', a, True, False), a),
857 (('bcsel', a, False, True), ('inot', a)),
858 (('bcsel@32', a, 1.0, 0.0), ('b2f', a)),
859 (('bcsel@32', a, 0.0, 1.0), ('b2f', ('inot', a))),
860 (('bcsel@32', a, -1.0, -0.0), ('fneg', ('b2f', a))),
861 (('bcsel@32', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))),
862 (('bcsel', True, b, c), b),
863 (('bcsel', False, b, c), c),
864 (('bcsel', a, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a, b, c))),
865
866 (('bcsel', a, b, b), b),
867 (('~fcsel', a, b, b), b),
868
869 # D3D Boolean emulation
870 (('bcsel', a, -1, 0), ('ineg', ('b2i', 'a@1'))),
871 (('bcsel', a, 0, -1), ('ineg', ('b2i', ('inot', a)))),
872 (('bcsel', a, 1, 0), ('b2i', 'a@1')),
873 (('bcsel', a, 0, 1), ('b2i', ('inot', a))),
874 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
875 ('ineg', ('b2i', ('iand', a, b)))),
876 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
877 ('ineg', ('b2i', ('ior', a, b)))),
878 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a)),
879 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a),
880 (('ine', ('ineg', ('b2i', 'a@1')), 0), a),
881 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a)),
882 (('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)),
883 (('iand', ('ineg', ('b2i', a)), 1), ('b2i', a)),
884
885 # SM5 32-bit shifts are defined to use the 5 least significant bits
886 (('ishl', 'a@32', ('iand', 31, b)), ('ishl', a, b)),
887 (('ishr', 'a@32', ('iand', 31, b)), ('ishr', a, b)),
888 (('ushr', 'a@32', ('iand', 31, b)), ('ushr', a, b)),
889
890 # Conversions
891 (('i2b32', ('b2i', 'a@32')), a),
892 (('f2i', ('ftrunc', a)), ('f2i', a)),
893 (('f2u', ('ftrunc', a)), ('f2u', a)),
894 (('i2b', ('ineg', a)), ('i2b', a)),
895 (('i2b', ('iabs', a)), ('i2b', a)),
896 (('inot', ('f2b1', a)), ('feq', a, 0.0)),
897
898 # The C spec says, "If the value of the integral part cannot be represented
899 # by the integer type, the behavior is undefined." "Undefined" can mean
900 # "the conversion doesn't happen at all."
901 (('~i2f32', ('f2i32', 'a@32')), ('ftrunc', a)),
902
903 # Ironically, mark these as imprecise because removing the conversions may
904 # preserve more precision than doing the conversions (e.g.,
905 # uint(float(0x81818181u)) == 0x81818200).
906 (('~f2i32', ('i2f', 'a@32')), a),
907 (('~f2i32', ('u2f', 'a@32')), a),
908 (('~f2u32', ('i2f', 'a@32')), a),
909 (('~f2u32', ('u2f', 'a@32')), a),
910
911 # Conversions from 16 bits to 32 bits and back can always be removed
912 (('f2f16', ('f2f32', 'a@16')), a),
913 (('f2fmp', ('f2f32', 'a@16')), a),
914 (('i2i16', ('i2i32', 'a@16')), a),
915 (('i2imp', ('i2i32', 'a@16')), a),
916 (('u2u16', ('u2u32', 'a@16')), a),
917 (('u2ump', ('u2u32', 'a@16')), a),
918 (('f2f16', ('b2f32', 'a@1')), ('b2f16', a)),
919 (('f2fmp', ('b2f32', 'a@1')), ('b2f16', a)),
920 (('i2i16', ('b2i32', 'a@1')), ('b2i16', a)),
921 (('i2imp', ('b2i32', 'a@1')), ('b2i16', a)),
922 (('u2u16', ('b2i32', 'a@1')), ('b2i16', a)),
923 (('u2ump', ('b2i32', 'a@1')), ('b2i16', a)),
924 # Conversions to 16 bits would be lossy so they should only be removed if
925 # the instruction was generated by the precision lowering pass.
926 (('f2f32', ('f2fmp', 'a@32')), a),
927 (('i2i32', ('i2imp', 'a@32')), a),
928 (('u2u32', ('u2ump', 'a@32')), a),
929
930 (('ffloor', 'a(is_integral)'), a),
931 (('fceil', 'a(is_integral)'), a),
932 (('ftrunc', 'a(is_integral)'), a),
933 # fract(x) = x - floor(x), so fract(NaN) = NaN
934 (('~ffract', 'a(is_integral)'), 0.0),
935 (('fabs', 'a(is_not_negative)'), a),
936 (('iabs', 'a(is_not_negative)'), a),
937 (('fsat', 'a(is_not_positive)'), 0.0),
938
939 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
940 # says:
941 #
942 # It is undefined to convert a negative floating-point value to an
943 # uint.
944 #
945 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
946 # some optimizations in the i965 backend to proceed.
947 (('ige', ('f2u', a), b), ('ige', ('f2i', a), b)),
948 (('ige', b, ('f2u', a)), ('ige', b, ('f2i', a))),
949 (('ilt', ('f2u', a), b), ('ilt', ('f2i', a), b)),
950 (('ilt', b, ('f2u', a)), ('ilt', b, ('f2i', a))),
951
952 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a), '!options->lower_fsat'),
953
954 # The result of the multiply must be in [-1, 0], so the result of the ffma
955 # must be in [0, 1].
956 (('flt', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), False),
957 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), False),
958 (('fmax', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0)),
959 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0)),
960
961 (('fneu', 'a(is_not_zero)', 0.0), True),
962 (('feq', 'a(is_not_zero)', 0.0), False),
963
964 # In this chart, + means value > 0 and - means value < 0.
965 #
966 # + >= + -> unknown 0 >= + -> false - >= + -> false
967 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
968 # + >= - -> true 0 >= - -> true - >= - -> unknown
969 #
970 # Using grouping conceptually similar to a Karnaugh map...
971 #
972 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
973 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
974 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
975 #
976 # The flt / ilt cases just invert the expected result.
977 #
978 # The results expecting true, must be marked imprecise. The results
979 # expecting false are fine because NaN compared >= or < anything is false.
980
981 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
982 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
983 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
984
985 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
986 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
987 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
988
989 (('ine', 'a(is_not_zero)', 0), True),
990 (('ieq', 'a(is_not_zero)', 0), False),
991
992 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
993 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
994 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
995
996 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
997 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
998 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
999
1000 (('ult', 0, 'a(is_gt_zero)'), True),
1001 (('ult', a, 0), False),
1002
1003 # Packing and then unpacking does nothing
1004 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a),
1005 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b),
1006 (('unpack_64_2x32', ('pack_64_2x32_split', a, b)), ('vec2', a, b)),
1007 (('unpack_64_2x32', ('pack_64_2x32', a)), a),
1008 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a),
1009 ('unpack_64_2x32_split_y', a)), a),
1010 (('pack_64_2x32', ('vec2', ('unpack_64_2x32_split_x', a),
1011 ('unpack_64_2x32_split_y', a))), a),
1012 (('pack_64_2x32', ('unpack_64_2x32', a)), a),
1013
1014 # Comparing two halves of an unpack separately. While this optimization
1015 # should be correct for non-constant values, it's less obvious that it's
1016 # useful in that case. For constant values, the pack will fold and we're
1017 # guaranteed to reduce the whole tree to one instruction.
1018 (('iand', ('ieq', ('unpack_32_2x16_split_x', a), '#b'),
1019 ('ieq', ('unpack_32_2x16_split_y', a), '#c')),
1020 ('ieq', a, ('pack_32_2x16_split', b, c))),
1021
1022 # Byte extraction
1023 (('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
1024 (('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
1025 (('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
1026 (('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
1027 (('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
1028 (('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
1029 (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'),
1030
1031 (('ubfe', a, 0, 8), ('extract_u8', a, 0), '!options->lower_extract_byte'),
1032 (('ubfe', a, 8, 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
1033 (('ubfe', a, 16, 8), ('extract_u8', a, 2), '!options->lower_extract_byte'),
1034 (('ubfe', a, 24, 8), ('extract_u8', a, 3), '!options->lower_extract_byte'),
1035 (('ibfe', a, 0, 8), ('extract_i8', a, 0), '!options->lower_extract_byte'),
1036 (('ibfe', a, 8, 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
1037 (('ibfe', a, 16, 8), ('extract_i8', a, 2), '!options->lower_extract_byte'),
1038 (('ibfe', a, 24, 8), ('extract_i8', a, 3), '!options->lower_extract_byte'),
1039
1040 # Word extraction
1041 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
1042 (('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
1043 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
1044 (('ishr', 'a@32', 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
1045 (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),
1046
1047 (('ubfe', a, 0, 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
1048 (('ubfe', a, 16, 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
1049 (('ibfe', a, 0, 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
1050 (('ibfe', a, 16, 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
1051
1052 # Lower pack/unpack
1053 (('pack_64_2x32_split', a, b), ('ior', ('u2u64', a), ('ishl', ('u2u64', b), 32)), 'options->lower_pack_64_2x32_split'),
1054 (('pack_32_2x16_split', a, b), ('ior', ('u2u32', a), ('ishl', ('u2u32', b), 16)), 'options->lower_pack_32_2x16_split'),
1055 (('unpack_64_2x32_split_x', a), ('u2u32', a), 'options->lower_unpack_64_2x32_split'),
1056 (('unpack_64_2x32_split_y', a), ('u2u32', ('ushr', a, 32)), 'options->lower_unpack_64_2x32_split'),
1057 (('unpack_32_2x16_split_x', a), ('u2u16', a), 'options->lower_unpack_32_2x16_split'),
1058 (('unpack_32_2x16_split_y', a), ('u2u16', ('ushr', a, 16)), 'options->lower_unpack_32_2x16_split'),
1059
1060 # Useless masking before unpacking
1061 (('unpack_half_2x16_split_x', ('iand', a, 0xffff)), ('unpack_half_2x16_split_x', a)),
1062 (('unpack_32_2x16_split_x', ('iand', a, 0xffff)), ('unpack_32_2x16_split_x', a)),
1063 (('unpack_64_2x32_split_x', ('iand', a, 0xffffffff)), ('unpack_64_2x32_split_x', a)),
1064 (('unpack_half_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_half_2x16_split_y', a)),
1065 (('unpack_32_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_32_2x16_split_y', a)),
1066 (('unpack_64_2x32_split_y', ('iand', a, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a)),
1067
1068 (('unpack_half_2x16_split_x', ('extract_u16', a, 0)), ('unpack_half_2x16_split_x', a)),
1069 (('unpack_half_2x16_split_x', ('extract_u16', a, 1)), ('unpack_half_2x16_split_y', a)),
1070 (('unpack_32_2x16_split_x', ('extract_u16', a, 0)), ('unpack_32_2x16_split_x', a)),
1071 (('unpack_32_2x16_split_x', ('extract_u16', a, 1)), ('unpack_32_2x16_split_y', a)),
1072
1073 # Optimize half packing
1074 (('ishl', ('pack_half_2x16', ('vec2', a, 0)), 16), ('pack_half_2x16', ('vec2', 0, a))),
1075 (('ushr', ('pack_half_2x16', ('vec2', 0, a)), 16), ('pack_half_2x16', ('vec2', a, 0))),
1076
1077 (('iadd', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
1078 ('pack_half_2x16', ('vec2', a, b))),
1079 (('ior', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
1080 ('pack_half_2x16', ('vec2', a, b))),
1081
1082 (('ishl', ('pack_half_2x16_split', a, 0), 16), ('pack_half_2x16_split', 0, a)),
1083 (('ushr', ('pack_half_2x16_split', 0, a), 16), ('pack_half_2x16_split', a, 0)),
1084 (('extract_u16', ('pack_half_2x16_split', 0, a), 1), ('pack_half_2x16_split', a, 0)),
1085
1086 (('iadd', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
1087 (('ior', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
1088 ])
1089
1090 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
1091 # patterns like those below.
1092 for op in ('ushr', 'ishr'):
1093 optimizations.extend([(('extract_u8', (op, 'a@16', 8), 0), ('extract_u8', a, 1))])
1094 optimizations.extend([(('extract_u8', (op, 'a@32', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)])
1095 optimizations.extend([(('extract_u8', (op, 'a@64', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)])
1096
1097 optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))])
1098
1099 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
1100 # patterns like those below.
1101 for op in ('extract_u8', 'extract_i8'):
1102 optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))])
1103 optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)])
1104 optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)])
1105
1106 optimizations.extend([
1107 # Subtracts
1108 (('ussub_4x8', a, 0), a),
1109 (('ussub_4x8', a, ~0), 0),
1110 # Lower all Subtractions first - they can get recombined later
1111 (('fsub', a, b), ('fadd', a, ('fneg', b))),
1112 (('isub', a, b), ('iadd', a, ('ineg', b))),
1113 (('uabs_usub', a, b), ('bcsel', ('ult', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),
1114 # This is correct. We don't need isub_sat because the result type is unsigned, so it cannot overflow.
1115 (('uabs_isub', a, b), ('bcsel', ('ilt', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),
1116
1117 # Propagate negation up multiplication chains
1118 (('fmul(is_used_by_non_fsat)', ('fneg', a), b), ('fneg', ('fmul', a, b))),
1119 (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),
1120
1121 # Propagate constants up multiplication chains
1122 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)),
1123 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)),
1124 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)),
1125 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)),
1126
1127 # Reassociate constants in add/mul chains so they can be folded together.
1128 # For now, we mostly only handle cases where the constants are separated by
1129 # a single non-constant. We could do better eventually.
1130 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a, c), b)),
1131 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a, c), b)),
1132 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a, c), b)),
1133 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a, ('fneg', c)), ('fneg', b))),
1134 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a, c), b)),
1135 (('iand', '#a', ('iand', 'b(is_not_const)', '#c')), ('iand', ('iand', a, c), b)),
1136 (('ior', '#a', ('ior', 'b(is_not_const)', '#c')), ('ior', ('ior', a, c), b)),
1137 (('ixor', '#a', ('ixor', 'b(is_not_const)', '#c')), ('ixor', ('ixor', a, c), b)),
1138
1139 # Drop mul-div by the same value when there's no wrapping.
1140 (('idiv', ('imul(no_signed_wrap)', a, b), b), a),
1141
1142 # By definition...
1143 (('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)),
1144 (('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
1145 (('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
1146
1147 (('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)),
1148 (('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
1149 (('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
1150
1151 (('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),
1152
1153 (('~fmul', ('bcsel(is_used_once)', c, -1.0, 1.0), b), ('bcsel', c, ('fneg', b), b)),
1154 (('~fmul', ('bcsel(is_used_once)', c, 1.0, -1.0), b), ('bcsel', c, b, ('fneg', b))),
1155 (('~bcsel', ('flt', a, 0.0), ('fneg', a), a), ('fabs', a)),
1156
1157 (('bcsel', a, ('bcsel', b, c, d), d), ('bcsel', ('iand', a, b), c, d)),
1158 (('bcsel', a, b, ('bcsel', c, b, d)), ('bcsel', ('ior', a, c), b, d)),
1159
1160 # Misc. lowering
1161 (('fmod', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'),
1162 (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod'),
1163 (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
1164 (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),
1165
1166 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1167 ('bcsel', ('ult', 31, 'bits'), 'insert',
1168 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
1169 'options->lower_bitfield_insert'),
1170 (('ihadd', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1171 (('uhadd', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1172 (('irhadd', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1173 (('urhadd', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1174 (('ihadd@64', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1175 (('uhadd@64', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1176 (('irhadd@64', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1177 (('urhadd@64', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1178
1179 (('uadd_sat@64', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1180 (('uadd_sat', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat'),
1181 (('usub_sat', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_add_sat'),
1182 (('usub_sat@64', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_usub_sat64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1183
1184 # int64_t sum = a + b;
1185 #
1186 # if (a < 0 && b < 0 && a < sum)
1187 # sum = INT64_MIN;
1188 # } else if (a >= 0 && b >= 0 && sum < a)
1189 # sum = INT64_MAX;
1190 # }
1191 #
1192 # A couple optimizations are applied.
1193 #
1194 # 1. a < sum => sum >= 0. This replacement works because it is known that
1195 # a < 0 and b < 0, so sum should also be < 0 unless there was
1196 # underflow.
1197 #
1198 # 2. sum < a => sum < 0. This replacement works because it is known that
1199 # a >= 0 and b >= 0, so sum should also be >= 0 unless there was
1200 # overflow.
1201 #
1202 # 3. Invert the second if-condition and swap the order of parameters for
1203 # the bcsel. !(a >= 0 && b >= 0 && sum < 0) becomes !(a >= 0) || !(b >=
1204 # 0) || !(sum < 0), and that becomes (a < 0) || (b < 0) || (sum >= 0)
1205 #
1206 # On Intel Gen11, this saves ~11 instructions.
1207 (('iadd_sat@64', a, b), ('bcsel',
1208 ('iand', ('iand', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
1209 0x8000000000000000,
1210 ('bcsel',
1211 ('ior', ('ior', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
1212 ('iadd', a, b),
1213 0x7fffffffffffffff)),
1214 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1215
1216 # int64_t sum = a - b;
1217 #
1218 # if (a < 0 && b >= 0 && a < sum)
1219 # sum = INT64_MIN;
1220 # } else if (a >= 0 && b < 0 && a >= sum)
1221 # sum = INT64_MAX;
1222 # }
1223 #
1224 # Optimizations similar to the iadd_sat case are applied here.
1225 (('isub_sat@64', a, b), ('bcsel',
1226 ('iand', ('iand', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
1227 0x8000000000000000,
1228 ('bcsel',
1229 ('ior', ('ior', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
1230 ('isub', a, b),
1231 0x7fffffffffffffff)),
1232 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1233
1234 # These are done here instead of in the backend because the int64 lowering
1235 # pass will make a mess of the patterns. The first patterns are
1236 # conditioned on nir_lower_minmax64 because it was not clear that it was
1237 # always an improvement on platforms that have real int64 support. No
1238 # shaders in shader-db hit this, so it was hard to say one way or the
1239 # other.
1240 (('ilt', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1241 (('ilt', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1242 (('ige', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imax', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1243 (('ige', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imin', ('unpack_64_2x32_split_y', a), ('unpack_64_2x32_split_y', b)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1244 (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1245 (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1246
1247 (('ine', 'a@64', 0), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1248 (('ieq', 'a@64', 0), ('ieq', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1249 # 0u < uint(a) <=> uint(a) != 0u
1250 (('ult', 0, 'a@64'), ('ine', ('ior', ('unpack_64_2x32_split_x', a), ('unpack_64_2x32_split_y', a)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1251
1252 # Alternative lowering that doesn't rely on bfi.
1253 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1254 ('bcsel', ('ult', 31, 'bits'),
1255 'insert',
1256 (('ior',
1257 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1258 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1259 'options->lower_bitfield_insert_to_shifts'),
1260
1261 # Alternative lowering that uses bitfield_select.
1262 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1263 ('bcsel', ('ult', 31, 'bits'), 'insert',
1264 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1265 'options->lower_bitfield_insert_to_bitfield_select'),
1266
1267 (('ibitfield_extract', 'value', 'offset', 'bits'),
1268 ('bcsel', ('ult', 31, 'bits'), 'value',
1269 ('ibfe', 'value', 'offset', 'bits')),
1270 'options->lower_bitfield_extract'),
1271
1272 (('ubitfield_extract', 'value', 'offset', 'bits'),
1273 ('bcsel', ('ult', 31, 'bits'), 'value',
1274 ('ubfe', 'value', 'offset', 'bits')),
1275 'options->lower_bitfield_extract'),
1276
1277 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1278 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1279 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1280 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1281 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1282 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1283 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1284
1285 # Section 8.8 (Integer Functions) of the GLSL 4.60 spec says:
1286 #
1287 # If bits is zero, the result will be zero.
1288 #
1289 # These patterns prevent other patterns from generating invalid results
1290 # when count is zero.
1291 (('ubfe', a, b, 0), 0),
1292 (('ibfe', a, b, 0), 0),
1293
1294 (('ubfe', a, 0, '#b'), ('iand', a, ('ushr', 0xffffffff, ('ineg', b)))),
1295
1296 (('b2i32', ('i2b', ('ubfe', a, b, 1))), ('ubfe', a, b, 1)),
1297 (('b2i32', ('i2b', ('ibfe', a, b, 1))), ('ubfe', a, b, 1)), # ubfe in the replacement is correct
1298 (('ine', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1299 (('ieq', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1300 (('ine', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1301 (('ieq', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1302
1303 (('ibitfield_extract', 'value', 'offset', 'bits'),
1304 ('bcsel', ('ieq', 0, 'bits'),
1305 0,
1306 ('ishr',
1307 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1308 ('isub', 32, 'bits'))),
1309 'options->lower_bitfield_extract_to_shifts'),
1310
1311 (('ubitfield_extract', 'value', 'offset', 'bits'),
1312 ('iand',
1313 ('ushr', 'value', 'offset'),
1314 ('bcsel', ('ieq', 'bits', 32),
1315 0xffffffff,
1316 ('isub', ('ishl', 1, 'bits'), 1))),
1317 'options->lower_bitfield_extract_to_shifts'),
1318
1319 (('ifind_msb', 'value'),
1320 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1321 'options->lower_ifind_msb'),
1322
1323 (('find_lsb', 'value'),
1324 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1325 'options->lower_find_lsb'),
1326
1327 (('extract_i8', a, 'b@32'),
1328 ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24),
1329 'options->lower_extract_byte'),
1330
1331 (('extract_u8', a, 'b@32'),
1332 ('iand', ('ushr', a, ('imul', b, 8)), 0xff),
1333 'options->lower_extract_byte'),
1334
1335 (('extract_i16', a, 'b@32'),
1336 ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16),
1337 'options->lower_extract_word'),
1338
1339 (('extract_u16', a, 'b@32'),
1340 ('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
1341 'options->lower_extract_word'),
1342
1343 (('pack_unorm_2x16', 'v'),
1344 ('pack_uvec2_to_uint',
1345 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1346 'options->lower_pack_unorm_2x16'),
1347
1348 (('pack_unorm_4x8', 'v'),
1349 ('pack_uvec4_to_uint',
1350 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1351 'options->lower_pack_unorm_4x8'),
1352
1353 (('pack_snorm_2x16', 'v'),
1354 ('pack_uvec2_to_uint',
1355 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1356 'options->lower_pack_snorm_2x16'),
1357
1358 (('pack_snorm_4x8', 'v'),
1359 ('pack_uvec4_to_uint',
1360 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1361 'options->lower_pack_snorm_4x8'),
1362
1363 (('unpack_unorm_2x16', 'v'),
1364 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1365 ('extract_u16', 'v', 1))),
1366 65535.0),
1367 'options->lower_unpack_unorm_2x16'),
1368
1369 (('unpack_unorm_4x8', 'v'),
1370 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1371 ('extract_u8', 'v', 1),
1372 ('extract_u8', 'v', 2),
1373 ('extract_u8', 'v', 3))),
1374 255.0),
1375 'options->lower_unpack_unorm_4x8'),
1376
1377 (('unpack_snorm_2x16', 'v'),
1378 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1379 ('extract_i16', 'v', 1))),
1380 32767.0))),
1381 'options->lower_unpack_snorm_2x16'),
1382
1383 (('unpack_snorm_4x8', 'v'),
1384 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1385 ('extract_i8', 'v', 1),
1386 ('extract_i8', 'v', 2),
1387 ('extract_i8', 'v', 3))),
1388 127.0))),
1389 'options->lower_unpack_snorm_4x8'),
1390
1391 (('pack_half_2x16_split', 'a@32', 'b@32'),
1392 ('ior', ('ishl', ('u2u32', ('f2f16', b)), 16), ('u2u32', ('f2f16', a))),
1393 'options->lower_pack_split'),
1394
1395 (('unpack_half_2x16_split_x', 'a@32'),
1396 ('f2f32', ('u2u16', a)),
1397 'options->lower_pack_split'),
1398
1399 (('unpack_half_2x16_split_y', 'a@32'),
1400 ('f2f32', ('u2u16', ('ushr', a, 16))),
1401 'options->lower_pack_split'),
1402
1403 (('pack_32_2x16_split', 'a@16', 'b@16'),
1404 ('ior', ('ishl', ('u2u32', b), 16), ('u2u32', a)),
1405 'options->lower_pack_split'),
1406
1407 (('unpack_32_2x16_split_x', 'a@32'),
1408 ('u2u16', a),
1409 'options->lower_pack_split'),
1410
1411 (('unpack_32_2x16_split_y', 'a@32'),
1412 ('u2u16', ('ushr', 'a', 16)),
1413 'options->lower_pack_split'),
1414
1415 (('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
1416 (('imin', ('imax', a, -1), 1), ('isign', a), '!options->lower_isign'),
1417 (('imax', ('imin', a, 1), -1), ('isign', a), '!options->lower_isign'),
1418 (('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),
1419 (('fadd', ('b2f32', ('flt', 0.0, 'a@32')), ('fneg', ('b2f32', ('flt', 'a@32', 0.0)))), ('fsign', a), '!options->lower_fsign'),
1420 (('iadd', ('b2i32', ('flt', 0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0)))), ('f2i32', ('fsign', a)), '!options->lower_fsign'),
1421
1422 # Address/offset calculations:
1423 # Drivers supporting imul24 should use the nir_lower_amul() pass, this
1424 # rule converts everyone else to imul:
1425 (('amul', a, b), ('imul', a, b), '!options->has_imul24'),
1426
1427 (('umul24', a, b),
1428 ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)),
1429 '!options->has_umul24'),
1430 (('umad24', a, b, c),
1431 ('iadd', ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)), c),
1432 '!options->has_umad24'),
1433
1434 (('imad24_ir3', a, b, 0), ('imul24', a, b)),
1435 (('imad24_ir3', a, 0, c), (c)),
1436 (('imad24_ir3', a, 1, c), ('iadd', a, c)),
1437
1438 # if first two srcs are const, crack apart the imad so constant folding
1439 # can clean up the imul:
1440 # TODO ffma should probably get a similar rule:
1441 (('imad24_ir3', '#a', '#b', c), ('iadd', ('imul', a, b), c)),
1442
1443 # These will turn 24b address/offset calc back into 32b shifts, but
1444 # it should be safe to get back some of the bits of precision that we
1445 # already decided were no necessary:
1446 (('imul24', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'),
1447 (('imul24', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'),
1448 (('imul24', a, 0), (0)),
1449 ])
1450
1451 # bit_size dependent lowerings
1452 for bit_size in [8, 16, 32, 64]:
1453 # convenience constants
1454 intmax = (1 << (bit_size - 1)) - 1
1455 intmin = 1 << (bit_size - 1)
1456
1457 optimizations += [
1458 (('iadd_sat@' + str(bit_size), a, b),
1459 ('bcsel', ('ige', b, 1), ('bcsel', ('ilt', ('iadd', a, b), a), intmax, ('iadd', a, b)),
1460 ('bcsel', ('ilt', a, ('iadd', a, b)), intmin, ('iadd', a, b))), 'options->lower_add_sat'),
1461 (('isub_sat@' + str(bit_size), a, b),
1462 ('bcsel', ('ilt', b, 0), ('bcsel', ('ilt', ('isub', a, b), a), intmax, ('isub', a, b)),
1463 ('bcsel', ('ilt', a, ('isub', a, b)), intmin, ('isub', a, b))), 'options->lower_add_sat'),
1464 ]
1465
1466 invert = OrderedDict([('feq', 'fneu'), ('fneu', 'feq')])
1467
1468 for left, right in itertools.combinations_with_replacement(invert.keys(), 2):
1469 optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))),
1470 ('iand', (invert[left], a, b), (invert[right], c, d))))
1471 optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))),
1472 ('ior', (invert[left], a, b), (invert[right], c, d))))
1473
1474 # Optimize x2bN(b2x(x)) -> x
1475 for size in type_sizes('bool'):
1476 aN = 'a@' + str(size)
1477 f2bN = 'f2b' + str(size)
1478 i2bN = 'i2b' + str(size)
1479 optimizations.append(((f2bN, ('b2f', aN)), a))
1480 optimizations.append(((i2bN, ('b2i', aN)), a))
1481
1482 # Optimize x2yN(b2x(x)) -> b2y
1483 for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1484 if x != 'f' and y != 'f' and x != y:
1485 continue
1486
1487 b2x = 'b2f' if x == 'f' else 'b2i'
1488 b2y = 'b2f' if y == 'f' else 'b2i'
1489 x2yN = '{}2{}'.format(x, y)
1490 optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
1491
1492 # Optimize away x2xN(a@N)
1493 for t in ['int', 'uint', 'float', 'bool']:
1494 for N in type_sizes(t):
1495 x2xN = '{0}2{0}{1}'.format(t[0], N)
1496 aN = 'a@{0}'.format(N)
1497 optimizations.append(((x2xN, aN), a))
1498
1499 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1500 # In particular, we can optimize away everything except upcast of downcast and
1501 # upcasts where the type differs from the other cast
1502 for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
1503 if N < M:
1504 # The outer cast is a down-cast. It doesn't matter what the size of the
1505 # argument of the inner cast is because we'll never been in the upcast
1506 # of downcast case. Regardless of types, we'll always end up with y2yN
1507 # in the end.
1508 for x, y in itertools.product(['i', 'u'], ['i', 'u']):
1509 x2xN = '{0}2{0}{1}'.format(x, N)
1510 y2yM = '{0}2{0}{1}'.format(y, M)
1511 y2yN = '{0}2{0}{1}'.format(y, N)
1512 optimizations.append(((x2xN, (y2yM, a)), (y2yN, a)))
1513 elif N > M:
1514 # If the outer cast is an up-cast, we have to be more careful about the
1515 # size of the argument of the inner cast and with types. In this case,
1516 # the type is always the type of type up-cast which is given by the
1517 # outer cast.
1518 for P in type_sizes('uint'):
1519 # We can't optimize away up-cast of down-cast.
1520 if M < P:
1521 continue
1522
1523 # Because we're doing down-cast of down-cast, the types always have
1524 # to match between the two casts
1525 for x in ['i', 'u']:
1526 x2xN = '{0}2{0}{1}'.format(x, N)
1527 x2xM = '{0}2{0}{1}'.format(x, M)
1528 aP = 'a@{0}'.format(P)
1529 optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a)))
1530 else:
1531 # The N == M case is handled by other optimizations
1532 pass
1533
1534 # Downcast operations should be able to see through pack
1535 for t in ['i', 'u']:
1536 for N in [8, 16, 32]:
1537 x2xN = '{0}2{0}{1}'.format(t, N)
1538 optimizations += [
1539 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
1540 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
1541 ]
1542
1543 # Optimize comparisons with up-casts
1544 for t in ['int', 'uint', 'float']:
1545 for N, M in itertools.product(type_sizes(t), repeat=2):
1546 if N == 1 or N >= M:
1547 continue
1548
1549 cond = 'true'
1550 if N == 8:
1551 cond = 'options->support_8bit_alu'
1552 elif N == 16:
1553 cond = 'options->support_16bit_alu'
1554 x2xM = '{0}2{0}{1}'.format(t[0], M)
1555 x2xN = '{0}2{0}{1}'.format(t[0], N)
1556 aN = 'a@' + str(N)
1557 bN = 'b@' + str(N)
1558 xeq = 'feq' if t == 'float' else 'ieq'
1559 xne = 'fneu' if t == 'float' else 'ine'
1560 xge = '{0}ge'.format(t[0])
1561 xlt = '{0}lt'.format(t[0])
1562
1563 # Up-casts are lossless so for correctly signed comparisons of
1564 # up-casted values we can do the comparison at the largest of the two
1565 # original sizes and drop one or both of the casts. (We have
1566 # optimizations to drop the no-op casts which this may generate.)
1567 for P in type_sizes(t):
1568 if P == 1 or P > N:
1569 continue
1570
1571 bP = 'b@' + str(P)
1572 optimizations += [
1573 ((xeq, (x2xM, aN), (x2xM, bP)), (xeq, a, (x2xN, b)), cond),
1574 ((xne, (x2xM, aN), (x2xM, bP)), (xne, a, (x2xN, b)), cond),
1575 ((xge, (x2xM, aN), (x2xM, bP)), (xge, a, (x2xN, b)), cond),
1576 ((xlt, (x2xM, aN), (x2xM, bP)), (xlt, a, (x2xN, b)), cond),
1577 ((xge, (x2xM, bP), (x2xM, aN)), (xge, (x2xN, b), a), cond),
1578 ((xlt, (x2xM, bP), (x2xM, aN)), (xlt, (x2xN, b), a), cond),
1579 ]
1580
1581 # The next bit doesn't work on floats because the range checks would
1582 # get way too complicated.
1583 if t in ['int', 'uint']:
1584 if t == 'int':
1585 xN_min = -(1 << (N - 1))
1586 xN_max = (1 << (N - 1)) - 1
1587 elif t == 'uint':
1588 xN_min = 0
1589 xN_max = (1 << N) - 1
1590 else:
1591 assert False
1592
1593 # If we're up-casting and comparing to a constant, we can unfold
1594 # the comparison into a comparison with the shrunk down constant
1595 # and a check that the constant fits in the smaller bit size.
1596 optimizations += [
1597 ((xeq, (x2xM, aN), '#b'),
1598 ('iand', (xeq, a, (x2xN, b)), (xeq, (x2xM, (x2xN, b)), b)), cond),
1599 ((xne, (x2xM, aN), '#b'),
1600 ('ior', (xne, a, (x2xN, b)), (xne, (x2xM, (x2xN, b)), b)), cond),
1601 ((xlt, (x2xM, aN), '#b'),
1602 ('iand', (xlt, xN_min, b),
1603 ('ior', (xlt, xN_max, b), (xlt, a, (x2xN, b)))), cond),
1604 ((xlt, '#a', (x2xM, bN)),
1605 ('iand', (xlt, a, xN_max),
1606 ('ior', (xlt, a, xN_min), (xlt, (x2xN, a), b))), cond),
1607 ((xge, (x2xM, aN), '#b'),
1608 ('iand', (xge, xN_max, b),
1609 ('ior', (xge, xN_min, b), (xge, a, (x2xN, b)))), cond),
1610 ((xge, '#a', (x2xM, bN)),
1611 ('iand', (xge, a, xN_min),
1612 ('ior', (xge, a, xN_max), (xge, (x2xN, a), b))), cond),
1613 ]
1614
1615 # Convert masking followed by signed downcast to just unsigned downcast
1616 optimizations += [
1617 (('i2i32', ('iand', 'a@64', 0xffffffff)), ('u2u32', a)),
1618 (('i2i16', ('iand', 'a@32', 0xffff)), ('u2u16', a)),
1619 (('i2i16', ('iand', 'a@64', 0xffff)), ('u2u16', a)),
1620 (('i2i8', ('iand', 'a@16', 0xff)), ('u2u8', a)),
1621 (('i2i8', ('iand', 'a@32', 0xff)), ('u2u8', a)),
1622 (('i2i8', ('iand', 'a@64', 0xff)), ('u2u8', a)),
1623 ]
1624
1625 def fexp2i(exp, bits):
1626 # Generate an expression which constructs value 2.0^exp or 0.0.
1627 #
1628 # We assume that exp is already in a valid range:
1629 #
1630 # * [-15, 15] for 16-bit float
1631 # * [-127, 127] for 32-bit float
1632 # * [-1023, 1023] for 16-bit float
1633 #
1634 # If exp is the lowest value in the valid range, a value of 0.0 is
1635 # constructed. Otherwise, the value 2.0^exp is constructed.
1636 if bits == 16:
1637 return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
1638 elif bits == 32:
1639 return ('ishl', ('iadd', exp, 127), 23)
1640 elif bits == 64:
1641 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
1642 else:
1643 assert False
1644
1645 def ldexp(f, exp, bits):
1646 # The maximum possible range for a normal exponent is [-126, 127] and,
1647 # throwing in denormals, you get a maximum range of [-149, 127]. This
1648 # means that we can potentially have a swing of +-276. If you start with
1649 # FLT_MAX, you actually have to do ldexp(FLT_MAX, -278) to get it to flush
1650 # all the way to zero. The GLSL spec only requires that we handle a subset
1651 # of this range. From version 4.60 of the spec:
1652 #
1653 # "If exp is greater than +128 (single-precision) or +1024
1654 # (double-precision), the value returned is undefined. If exp is less
1655 # than -126 (single-precision) or -1022 (double-precision), the value
1656 # returned may be flushed to zero. Additionally, splitting the value
1657 # into a significand and exponent using frexp() and then reconstructing
1658 # a floating-point value using ldexp() should yield the original input
1659 # for zero and all finite non-denormalized values."
1660 #
1661 # The SPIR-V spec has similar language.
1662 #
1663 # In order to handle the maximum value +128 using the fexp2i() helper
1664 # above, we have to split the exponent in half and do two multiply
1665 # operations.
1666 #
1667 # First, we clamp exp to a reasonable range. Specifically, we clamp to
1668 # twice the full range that is valid for the fexp2i() function above. If
1669 # exp/2 is the bottom value of that range, the fexp2i() expression will
1670 # yield 0.0f which, when multiplied by f, will flush it to zero which is
1671 # allowed by the GLSL and SPIR-V specs for low exponent values. If the
1672 # value is clamped from above, then it must have been above the supported
1673 # range of the GLSL built-in and therefore any return value is acceptable.
1674 if bits == 16:
1675 exp = ('imin', ('imax', exp, -30), 30)
1676 elif bits == 32:
1677 exp = ('imin', ('imax', exp, -254), 254)
1678 elif bits == 64:
1679 exp = ('imin', ('imax', exp, -2046), 2046)
1680 else:
1681 assert False
1682
1683 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1684 # (We use ishr which isn't the same for -1, but the -1 case still works
1685 # since we use exp-exp/2 as the second exponent.) While the spec
1686 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1687 # work with denormals and doesn't allow for the full swing in exponents
1688 # that you can get with normalized values. Instead, we create two powers
1689 # of two and multiply by them each in turn. That way the effective range
1690 # of our exponent is doubled.
1691 pow2_1 = fexp2i(('ishr', exp, 1), bits)
1692 pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
1693 return ('fmul', ('fmul', f, pow2_1), pow2_2)
1694
1695 optimizations += [
1696 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1697 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1698 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1699 ]
1700
1701 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1702 def bitfield_reverse(u):
1703 step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))
1704 step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8))
1705 step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4))
1706 step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2))
1707 step5 = ('ior(many-comm-expr)', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1))
1708
1709 return step5
1710
1711 optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1712
1713 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1714 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1715 # and, if a is a NaN then the second comparison will fail anyway.
1716 for op in ['flt', 'fge', 'feq']:
1717 optimizations += [
1718 (('iand', ('feq', a, a), (op, a, b)), ('!' + op, a, b)),
1719 (('iand', ('feq', a, a), (op, b, a)), ('!' + op, b, a)),
1720 ]
1721
1722 # Add optimizations to handle the case where the result of a ternary is
1723 # compared to a constant. This way we can take things like
1724 #
1725 # (a ? 0 : 1) > 0
1726 #
1727 # and turn it into
1728 #
1729 # a ? (0 > 0) : (1 > 0)
1730 #
1731 # which constant folding will eat for lunch. The resulting ternary will
1732 # further get cleaned up by the boolean reductions above and we will be
1733 # left with just the original variable "a".
1734 for op in ['flt', 'fge', 'feq', 'fneu',
1735 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1736 optimizations += [
1737 ((op, ('bcsel', 'a', '#b', '#c'), '#d'),
1738 ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))),
1739 ((op, '#d', ('bcsel', a, '#b', '#c')),
1740 ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
1741 ]
1742
1743
1744 # For example, this converts things like
1745 #
1746 # 1 + mix(0, a - 1, condition)
1747 #
1748 # into
1749 #
1750 # mix(1, (a-1)+1, condition)
1751 #
1752 # Other optimizations will rearrange the constants.
1753 for op in ['fadd', 'fmul', 'iadd', 'imul']:
1754 optimizations += [
1755 ((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d)))
1756 ]
1757
1758 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1759 # states:
1760 #
1761 # If neither layout qualifier is specified, derivatives in compute shaders
1762 # return zero, which is consistent with the handling of built-in texture
1763 # functions like texture() in GLSL 4.50 compute shaders.
1764 for op in ['fddx', 'fddx_fine', 'fddx_coarse',
1765 'fddy', 'fddy_fine', 'fddy_coarse']:
1766 optimizations += [
1767 ((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1768 ]
1769
1770 # Some optimizations for ir3-specific instructions.
1771 optimizations += [
1772 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1773 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1774 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1775 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1776 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1777 ]
1778
1779 # These kinds of sequences can occur after nir_opt_peephole_select.
1780 #
1781 # NOTE: fadd is not handled here because that gets in the way of ffma
1782 # generation in the i965 driver. Instead, fadd and ffma are handled in
1783 # late_optimizations.
1784
1785 for op in ['flrp']:
1786 optimizations += [
1787 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, c, e)), (op, b, c, ('bcsel', a, d, e))),
1788 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, c, e)), (op, b, c, ('bcsel', a, d, e))),
1789 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, e, d)), (op, b, ('bcsel', a, c, e), d)),
1790 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)),
1791 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, e, c, d)), (op, ('bcsel', a, b, e), c, d)),
1792 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', e, c, d)), (op, ('bcsel', a, b, e), c, d)),
1793 ]
1794
1795 for op in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']:
1796 optimizations += [
1797 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
1798 (('bcsel', a, (op + '(is_used_once)', b, 'c(is_not_const)'), (op, b, d)), (op, b, ('bcsel', a, c, d))),
1799 (('bcsel', a, (op, b, 'c(is_not_const)'), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
1800 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
1801 ]
1802
1803 for op in ['fpow']:
1804 optimizations += [
1805 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, d)), (op, b, ('bcsel', a, c, d))),
1806 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
1807 (('bcsel', a, (op + '(is_used_once)', b, c), (op, d, c)), (op, ('bcsel', a, b, d), c)),
1808 (('bcsel', a, (op, b, c), (op + '(is_used_once)', d, c)), (op, ('bcsel', a, b, d), c)),
1809 ]
1810
1811 for op in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos', 'fneg', 'fabs', 'fsign']:
1812 optimizations += [
1813 (('bcsel', c, (op + '(is_used_once)', a), (op + '(is_used_once)', b)), (op, ('bcsel', c, a, b))),
1814 ]
1815
1816 for op in ['ineg', 'iabs', 'inot', 'isign']:
1817 optimizations += [
1818 ((op, ('bcsel', c, '#a', '#b')), ('bcsel', c, (op, a), (op, b))),
1819 ]
1820
1821 # This section contains optimizations to propagate downsizing conversions of
1822 # constructed vectors into vectors of downsized components. Whether this is
1823 # useful depends on the SIMD semantics of the backend. On a true SIMD machine,
1824 # this reduces the register pressure of the vector itself and often enables the
1825 # conversions to be eliminated via other algebraic rules or constant folding.
1826 # In the worst case on a SIMD architecture, the propagated conversions may be
1827 # revectorized via nir_opt_vectorize so instruction count is minimally
1828 # impacted.
1829 #
1830 # On a machine with SIMD-within-a-register only, this actually
1831 # counterintuitively hurts instruction count. These machines are the same that
1832 # require vectorize_vec2_16bit, so we predicate the optimizations on that flag
1833 # not being set.
1834 #
1835 # Finally for scalar architectures, there should be no difference in generated
1836 # code since it all ends up scalarized at the end, but it might minimally help
1837 # compile-times.
1838
1839 for i in range(2, 4 + 1):
1840 for T in ('f', 'u', 'i'):
1841 vec_inst = ('vec' + str(i),)
1842
1843 indices = ['a', 'b', 'c', 'd']
1844 suffix_in = tuple((indices[j] + '@32') for j in range(i))
1845
1846 to_16 = '{}2{}16'.format(T, T)
1847 to_mp = '{}2{}mp'.format(T, T)
1848
1849 out_16 = tuple((to_16, indices[j]) for j in range(i))
1850 out_mp = tuple((to_mp, indices[j]) for j in range(i))
1851
1852 optimizations += [
1853 ((to_16, vec_inst + suffix_in), vec_inst + out_16, '!options->vectorize_vec2_16bit'),
1854 ((to_mp, vec_inst + suffix_in), vec_inst + out_mp, '!options->vectorize_vec2_16bit')
1855 ]
1856
1857 # This section contains "late" optimizations that should be run before
1858 # creating ffmas and calling regular optimizations for the final time.
1859 # Optimizations should go here if they help code generation and conflict
1860 # with the regular optimizations.
1861 before_ffma_optimizations = [
1862 # Propagate constants down multiplication chains
1863 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)),
1864 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)),
1865 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)),
1866 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)),
1867
1868 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
1869 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
1870 (('~fadd', ('fneg', a), a), 0.0),
1871 (('iadd', ('ineg', a), a), 0),
1872 (('iadd', ('ineg', a), ('iadd', a, b)), b),
1873 (('iadd', a, ('iadd', ('ineg', a), b)), b),
1874 (('~fadd', ('fneg', a), ('fadd', a, b)), b),
1875 (('~fadd', a, ('fadd', ('fneg', a), b)), b),
1876
1877 (('~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)),
1878 (('~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)),
1879 (('~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))),
1880 ]
1881
1882 # This section contains "late" optimizations that should be run after the
1883 # regular optimizations have finished. Optimizations should go here if
1884 # they help code generation but do not necessarily produce code that is
1885 # more easily optimizable.
1886 late_optimizations = [
1887 # Most of these optimizations aren't quite safe when you get infinity or
1888 # Nan involved but the first one should be fine.
1889 (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
1890 (('flt', ('fneg', ('fadd', a, b)), 0.0), ('flt', ('fneg', a), b)),
1891 (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
1892 (('~fge', ('fneg', ('fadd', a, b)), 0.0), ('fge', ('fneg', a), b)),
1893 (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
1894 (('~fneu', ('fadd', a, b), 0.0), ('fneu', a, ('fneg', b))),
1895
1896 # nir_lower_to_source_mods will collapse this, but its existence during the
1897 # optimization loop can prevent other optimizations.
1898 (('fneg', ('fneg', a)), a),
1899
1900 # Subtractions get lowered during optimization, so we need to recombine them
1901 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
1902 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
1903 (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
1904 (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
1905 (('iabs', a), ('imax', a, ('ineg', a)), 'options->lower_iabs'),
1906
1907 # These are duplicated from the main optimizations table. The late
1908 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1909 # new patterns like these. The patterns that compare with zero are removed
1910 # because they are unlikely to be created in by anything in
1911 # late_optimizations.
1912 (('flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
1913 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
1914 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
1915 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
1916 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
1917 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),
1918
1919 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
1920 # flt(fsat(a), 1.0) is inexact because it returns True if a is NaN
1921 # (fsat(NaN) is 0), while flt(a, 1.0) always returns FALSE.
1922 (('~flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
1923
1924 (('~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)))),
1925
1926 (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
1927 (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
1928 (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
1929 (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)),
1930 (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
1931 (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
1932 (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
1933 (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
1934 (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)),
1935 (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),
1936
1937 (('ior', a, a), a),
1938 (('iand', a, a), a),
1939
1940 (('iand', ('ine(is_used_once)', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umin', a, b), 0)),
1941 (('ior', ('ieq(is_used_once)', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umin', a, b), 0)),
1942
1943 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),
1944
1945 (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'),
1946 (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'),
1947 (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
1948 (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
1949
1950 (('~flrp@32', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
1951 (('~flrp@64', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
1952
1953 (('~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'),
1954 (('~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'),
1955
1956 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1957 # particular operation is common for expanding values stored in a texture
1958 # from [0,1] to [-1,1].
1959 (('~ffma@32', a, 2.0, -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'),
1960 (('~ffma@32', a, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'),
1961 (('~ffma@32', a, -2.0, 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'),
1962 (('~ffma@32', a, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
1963 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'),
1964 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'),
1965 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'),
1966 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
1967
1968 # flrp(a, b, a)
1969 # a*(1-a) + b*a
1970 # a + -a*a + a*b (1)
1971 # a + a*(b - a)
1972 # Option 1: ffma(a, (b-a), a)
1973 #
1974 # Alternately, after (1):
1975 # a*(1+b) + -a*a
1976 # a*((1+b) + -a)
1977 #
1978 # Let b=1
1979 #
1980 # Option 2: ffma(a, 2, -(a*a))
1981 # Option 3: ffma(a, 2, (-a)*a)
1982 # Option 4: ffma(a, -a, (2*a)
1983 # Option 5: a * (2 - a)
1984 #
1985 # There are a lot of other possible combinations.
1986 (('~ffma@32', ('fadd', b, ('fneg', a)), a, a), ('flrp', a, b, a), '!options->lower_flrp32'),
1987 (('~ffma@32', a, 2.0, ('fneg', ('fmul', a, a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1988 (('~ffma@32', a, 2.0, ('fmul', ('fneg', a), a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1989 (('~ffma@32', a, ('fneg', a), ('fmul', 2.0, a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1990 (('~fmul@32', a, ('fadd', 2.0, ('fneg', a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1991
1992 # we do these late so that we don't get in the way of creating ffmas
1993 (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
1994 (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),
1995
1996 (('bcsel', a, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a, b)))),
1997
1998 # Putting this in 'optimizations' interferes with the bcsel(a, op(b, c),
1999 # op(b, d)) => op(b, bcsel(a, c, d)) transformations. I do not know why.
2000 (('bcsel', ('feq', ('fsqrt', 'a(is_not_negative)'), 0.0), intBitsToFloat(0x7f7fffff), ('frsq', a)),
2001 ('fmin', ('frsq', a), intBitsToFloat(0x7f7fffff))),
2002
2003 # Things that look like DPH in the source shader may get expanded to
2004 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
2005 # to NIR. After FFMA is generated, this can look like:
2006 #
2007 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
2008 #
2009 # Reassociate the last addition into the first multiplication.
2010 #
2011 # Some shaders do not use 'invariant' in vertex and (possibly) geometry
2012 # shader stages on some outputs that are intended to be invariant. For
2013 # various reasons, this optimization may not be fully applied in all
2014 # shaders used for different rendering passes of the same geometry. This
2015 # can result in Z-fighting artifacts (at best). For now, disable this
2016 # optimization in these stages. See bugzilla #111490. In tessellation
2017 # stages applications seem to use 'precise' when necessary, so allow the
2018 # optimization in those stages.
2019 (('~fadd', ('ffma(is_use