2fe41d02deb12d357ea2045a1118f464741c254e
[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(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
333 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
334 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
335 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
336 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
337 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),
338
339 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
340 (('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
341 (('fge', 0.0, ('fsat(is_used_once)', a)), ('fge', 0.0, a)),
342 (('flt', 0.0, ('fsat(is_used_once)', a)), ('flt', 0.0, a)),
343
344 # 0.0 >= b2f(a)
345 # b2f(a) <= 0.0
346 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
347 # inot(a)
348 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a)),
349
350 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a)),
351
352 (('fneu', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
353 (('fneu', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
354 (('fneu', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a, b)),
355 (('fneu', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a, b)),
356 (('fneu', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
357 (('fneu', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
358 (('fneu', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a, b)),
359 (('fneu', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a, b)),
360 (('fneu', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a, b)),
361 (('fneu', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a, b)),
362 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
363 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
364 (('feq', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a, b))),
365 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
366 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
367 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
368 (('feq', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a, b))),
369 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a, b)),
370 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a, b)),
371 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a, b)),
372
373 # -(b2f(a) + b2f(b)) < 0
374 # 0 < b2f(a) + b2f(b)
375 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
376 # a || b
377 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a, b)),
378 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a, b)),
379
380 # -(b2f(a) + b2f(b)) >= 0
381 # 0 >= b2f(a) + b2f(b)
382 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
383 # !(a || b)
384 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))),
385 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
386
387 (('flt', a, ('fneg', a)), ('flt', a, 0.0)),
388 (('fge', a, ('fneg', a)), ('fge', a, 0.0)),
389
390 # Some optimizations (below) convert things like (a < b || c < b) into
391 # (min(a, c) < b). However, this interfers with the previous optimizations
392 # that try to remove comparisons with negated sums of b2f. This just
393 # breaks that apart.
394 (('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
395 ('ior', ('flt', c, 0.0), ('ior', a, b))),
396
397 (('~flt', ('fadd', a, b), a), ('flt', b, 0.0)),
398 (('~fge', ('fadd', a, b), a), ('fge', b, 0.0)),
399 (('~feq', ('fadd', a, b), a), ('feq', b, 0.0)),
400 (('~fneu', ('fadd', a, b), a), ('fneu', b, 0.0)),
401 (('~flt', ('fadd(is_used_once)', a, '#b'), '#c'), ('flt', a, ('fadd', c, ('fneg', b)))),
402 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('flt', ('fneg', ('fadd', c, b)), a)),
403 (('~fge', ('fadd(is_used_once)', a, '#b'), '#c'), ('fge', a, ('fadd', c, ('fneg', b)))),
404 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fge', ('fneg', ('fadd', c, b)), a)),
405 (('~feq', ('fadd(is_used_once)', a, '#b'), '#c'), ('feq', a, ('fadd', c, ('fneg', b)))),
406 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('feq', ('fneg', ('fadd', c, b)), a)),
407 (('~fneu', ('fadd(is_used_once)', a, '#b'), '#c'), ('fneu', a, ('fadd', c, ('fneg', b)))),
408 (('~fneu', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fneu', ('fneg', ('fadd', c, b)), a)),
409
410 # Cannot remove the addition from ilt or ige due to overflow.
411 (('ieq', ('iadd', a, b), a), ('ieq', b, 0)),
412 (('ine', ('iadd', a, b), a), ('ine', b, 0)),
413
414 # fmin(-b2f(a), b) >= 0.0
415 # -b2f(a) >= 0.0 && b >= 0.0
416 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
417 # b2f(a) == 0.0 && b >= 0.0
418 # a == False && b >= 0.0
419 # !a && b >= 0.0
420 #
421 # The fge in the second replacement is not a typo. I leave the proof that
422 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
423 # reader.
424 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
425 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
426
427 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a)),
428 (('~fneu', ('b2f', 'a@1'), 0.0), a),
429 (('ieq', ('b2i', 'a@1'), 0), ('inot', a)),
430 (('ine', ('b2i', 'a@1'), 0), a),
431
432 (('fneu', ('u2f', a), 0.0), ('ine', a, 0)),
433 (('feq', ('u2f', a), 0.0), ('ieq', a, 0)),
434 (('fge', ('u2f', a), 0.0), True),
435 (('fge', 0.0, ('u2f', a)), ('uge', 0, a)), # ieq instead?
436 (('flt', ('u2f', a), 0.0), False),
437 (('flt', 0.0, ('u2f', a)), ('ult', 0, a)), # ine instead?
438 (('fneu', ('i2f', a), 0.0), ('ine', a, 0)),
439 (('feq', ('i2f', a), 0.0), ('ieq', a, 0)),
440 (('fge', ('i2f', a), 0.0), ('ige', a, 0)),
441 (('fge', 0.0, ('i2f', a)), ('ige', 0, a)),
442 (('flt', ('i2f', a), 0.0), ('ilt', a, 0)),
443 (('flt', 0.0, ('i2f', a)), ('ilt', 0, a)),
444
445 # 0.0 < fabs(a)
446 # fabs(a) > 0.0
447 # fabs(a) != 0.0 because fabs(a) must be >= 0
448 # a != 0.0
449 (('~flt', 0.0, ('fabs', a)), ('fneu', a, 0.0)),
450
451 # -fabs(a) < 0.0
452 # fabs(a) > 0.0
453 (('~flt', ('fneg', ('fabs', a)), 0.0), ('fneu', a, 0.0)),
454
455 # 0.0 >= fabs(a)
456 # 0.0 == fabs(a) because fabs(a) must be >= 0
457 # 0.0 == a
458 (('fge', 0.0, ('fabs', a)), ('feq', a, 0.0)),
459
460 # -fabs(a) >= 0.0
461 # 0.0 >= fabs(a)
462 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
463
464 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
465 (('iand', ('fge', a, 0.0), ('fge', 1.0, a)), ('feq', a, ('fsat', a)), '!options->lower_fsat'),
466
467 # (a < 0.0) || (a > 1.0)
468 # !(!(a < 0.0) && !(a > 1.0))
469 # !((a >= 0.0) && (a <= 1.0))
470 # !(a == fsat(a))
471 # a != fsat(a)
472 (('ior', ('flt', a, 0.0), ('flt', 1.0, a)), ('fneu', a, ('fsat', a)), '!options->lower_fsat'),
473
474 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a, b))),
475 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a, b)))),
476 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
477 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a, b)))),
478
479 # fmin(b2f(a), b)
480 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
481 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
482 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
483 #
484 # Since b is a constant, constant folding will eliminate the fmin and the
485 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
486 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))),
487
488 (('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),
489
490 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
491 (('~bcsel', ('flt', b, a), b, a), ('fmin', a, b)),
492 (('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
493 (('~bcsel', ('fge', a, b), b, a), ('fmin', a, b)),
494 (('~bcsel', ('fge', b, a), b, a), ('fmax', a, b)),
495 (('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)),
496 (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
497 (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
498 (('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)),
499 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
500 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
501 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
502 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
503 (('bcsel', a, True, b), ('ior', a, b)),
504 (('bcsel', a, a, b), ('ior', a, b)),
505 (('bcsel', a, b, False), ('iand', a, b)),
506 (('bcsel', a, b, a), ('iand', a, b)),
507 (('~fmin', a, a), a),
508 (('~fmax', a, a), a),
509 (('imin', a, a), a),
510 (('imax', a, a), a),
511 (('umin', a, a), a),
512 (('umax', a, a), a),
513 (('fmax', ('fmax', a, b), b), ('fmax', a, b)),
514 (('umax', ('umax', a, b), b), ('umax', a, b)),
515 (('imax', ('imax', a, b), b), ('imax', a, b)),
516 (('fmin', ('fmin', a, b), b), ('fmin', a, b)),
517 (('umin', ('umin', a, b), b), ('umin', a, b)),
518 (('imin', ('imin', a, b), b), ('imin', a, b)),
519 (('iand@32', a, ('inot', ('ishr', a, 31))), ('imax', a, 0)),
520
521 # Simplify logic to detect sign of an integer.
522 (('ieq', ('iand', 'a@32', 0x80000000), 0x00000000), ('ige', a, 0)),
523 (('ine', ('iand', 'a@32', 0x80000000), 0x80000000), ('ige', a, 0)),
524 (('ine', ('iand', 'a@32', 0x80000000), 0x00000000), ('ilt', a, 0)),
525 (('ieq', ('iand', 'a@32', 0x80000000), 0x80000000), ('ilt', a, 0)),
526 (('ine', ('ushr', 'a@32', 31), 0), ('ilt', a, 0)),
527 (('ieq', ('ushr', 'a@32', 31), 0), ('ige', a, 0)),
528 (('ieq', ('ushr', 'a@32', 31), 1), ('ilt', a, 0)),
529 (('ine', ('ushr', 'a@32', 31), 1), ('ige', a, 0)),
530 (('ine', ('ishr', 'a@32', 31), 0), ('ilt', a, 0)),
531 (('ieq', ('ishr', 'a@32', 31), 0), ('ige', a, 0)),
532 (('ieq', ('ishr', 'a@32', 31), -1), ('ilt', a, 0)),
533 (('ine', ('ishr', 'a@32', 31), -1), ('ige', a, 0)),
534
535 (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
536 (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
537 (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
538 (('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))),
539 (('~fmin', a, ('fabs', a)), a),
540 (('imin', a, ('iabs', a)), a),
541 (('~fmax', a, ('fneg', ('fabs', a))), a),
542 (('imax', a, ('ineg', ('iabs', a))), a),
543 (('fmax', a, ('fabs', a)), ('fabs', a)),
544 (('imax', a, ('iabs', a)), ('iabs', a)),
545 (('fmax', a, ('fneg', a)), ('fabs', a)),
546 (('imax', a, ('ineg', a)), ('iabs', a), '!options->lower_iabs'),
547 (('~fmax', ('fabs', a), 0.0), ('fabs', a)),
548 (('fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'),
549 # fmax(fmin(a, 1.0), 0.0) is inexact because it returns 1.0 on NaN, while
550 # fsat(a) returns 0.0.
551 (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'),
552 # fmin(fmax(a, -1.0), 0.0) is inexact because it returns -1.0 on NaN, while
553 # fneg(fsat(fneg(a))) returns -0.0 on NaN.
554 (('~fmin', ('fmax', a, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
555 # fmax(fmin(a, 0.0), -1.0) is inexact because it returns 0.0 on NaN, while
556 # fneg(fsat(fneg(a))) returns -0.0 on NaN. This only matters if
557 # SignedZeroInfNanPreserve is set, but we don't currently have any way of
558 # representing this in the optimizations other than the usual ~.
559 (('~fmax', ('fmin', a, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
560 (('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))),
561 (('fsat', ('b2f', a)), ('b2f', a)),
562 (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
563 (('fsat', ('fsat', a)), ('fsat', a)),
564 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b))), ('fsat', ('fadd', ('fneg', a), ('fneg', b))), '!options->lower_fsat'),
565 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fneg', a), b)), '!options->lower_fsat'),
566 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fabs', a), ('fabs', b))), '!options->lower_fsat'),
567 (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)),
568 (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)),
569 (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
570 # Both the left and right patterns are "b" when isnan(a), so this is exact.
571 (('fmax', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
572 # The left pattern is 0.0 when isnan(a) (because fmin(fsat(NaN), b) ->
573 # fmin(0.0, b)) while the right one is "b", so this optimization is inexact.
574 (('~fmin', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a, b))),
575
576 # If a in [0,b] then b-a is also in [0,b]. Since b in [0,1], max(b-a, 0) =
577 # fsat(b-a).
578 #
579 # If a > b, then b-a < 0 and max(b-a, 0) = fsat(b-a) = 0
580 #
581 # This should be NaN safe since max(NaN, 0) = fsat(NaN) = 0.
582 (('fmax', ('fadd(is_used_once)', ('fneg', 'a(is_not_negative)'), '#b@32(is_zero_to_one)'), 0.0),
583 ('fsat', ('fadd', ('fneg', a), b)), '!options->lower_fsat'),
584
585 (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
586 (('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
587 (('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
588 (('~ior', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
589 (('~ior', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
590 (('~ior', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmax', b, c))),
591 (('~ior', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmin', a, b), c)),
592 (('~ior', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmin', b, c))),
593 (('~ior', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmax', a, b), c)),
594 (('~iand', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmin', b, c))),
595 (('~iand', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmax', a, b), c)),
596 (('~iand', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmax', b, c))),
597 (('~iand', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmin', a, b), c)),
598 (('~iand', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmin', b, c))),
599 (('~iand', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmax', a, b), c)),
600 (('~iand', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmax', b, c))),
601 (('~iand', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmin', a, b), c)),
602
603 (('ior', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imax', b, c))),
604 (('ior', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imin', a, b), c)),
605 (('ior', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imin', b, c))),
606 (('ior', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imax', a, b), c)),
607 (('ior', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umax', b, c))),
608 (('ior', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umin', a, b), c)),
609 (('ior', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umin', b, c))),
610 (('ior', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umax', a, b), c)),
611 (('iand', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imin', b, c))),
612 (('iand', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imax', a, b), c)),
613 (('iand', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imax', b, c))),
614 (('iand', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imin', a, b), c)),
615 (('iand', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umin', b, c))),
616 (('iand', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umax', a, b), c)),
617 (('iand', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umax', b, c))),
618 (('iand', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umin', a, b), c)),
619
620 # These derive from the previous patterns with the application of b < 0 <=>
621 # 0 < -b. The transformation should be applied if either comparison is
622 # used once as this ensures that the number of comparisons will not
623 # increase. The sources to the ior and iand are not symmetric, so the
624 # rules have to be duplicated to get this behavior.
625 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
626 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
627 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
628 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
629 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
630 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
631 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
632 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
633
634 # Common pattern like 'if (i == 0 || i == 1 || ...)'
635 (('ior', ('ieq', a, 0), ('ieq', a, 1)), ('uge', 1, a)),
636 (('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)),
637 (('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)),
638
639 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
640 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
641 # so emit an open-coded version of that.
642 (('bcsel@32', ('feq', a, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
643 ('ior', 0x3f800000, ('iand', a, 0x80000000))),
644
645 (('ior', a, ('ieq', a, False)), True),
646 (('ior', a, ('inot', a)), -1),
647
648 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a, b)),
649 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a, b))),
650
651 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umax', a, b), 0)),
652 (('ior', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umin', a, b), 0)),
653 (('iand', ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umin', a, b), 0)),
654 (('ior', ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umax', a, b), 0)),
655
656 # This pattern occurs coutresy of __flt64_nonnan in the soft-fp64 code.
657 # The first part of the iand comes from the !__feq64_nonnan.
658 #
659 # The second pattern is a reformulation of the first based on the relation
660 # (a == 0 || y == 0) <=> umin(a, y) == 0, where b in the first equation
661 # happens to be y == 0.
662 (('iand', ('inot', ('iand', ('ior', ('ieq', a, 0), b), c)), ('ilt', a, 0)),
663 ('iand', ('inot', ('iand', b , c)), ('ilt', a, 0))),
664 (('iand', ('inot', ('iand', ('ieq', ('umin', a, b), 0), c)), ('ilt', a, 0)),
665 ('iand', ('inot', ('iand', ('ieq', b , 0), c)), ('ilt', a, 0))),
666
667 # These patterns can result when (a < b || a < c) => (a < min(b, c))
668 # transformations occur before constant propagation and loop-unrolling.
669 (('~flt', a, ('fmax', b, a)), ('flt', a, b)),
670 (('~flt', ('fmin', a, b), a), ('flt', b, a)),
671 (('~fge', a, ('fmin', b, a)), True),
672 (('~fge', ('fmax', a, b), a), True),
673 (('~flt', a, ('fmin', b, a)), False),
674 (('~flt', ('fmax', a, b), a), False),
675 (('~fge', a, ('fmax', b, a)), ('fge', a, b)),
676 (('~fge', ('fmin', a, b), a), ('fge', b, a)),
677
678 (('ilt', a, ('imax', b, a)), ('ilt', a, b)),
679 (('ilt', ('imin', a, b), a), ('ilt', b, a)),
680 (('ige', a, ('imin', b, a)), True),
681 (('ige', ('imax', a, b), a), True),
682 (('ult', a, ('umax', b, a)), ('ult', a, b)),
683 (('ult', ('umin', a, b), a), ('ult', b, a)),
684 (('uge', a, ('umin', b, a)), True),
685 (('uge', ('umax', a, b), a), True),
686 (('ilt', a, ('imin', b, a)), False),
687 (('ilt', ('imax', a, b), a), False),
688 (('ige', a, ('imax', b, a)), ('ige', a, b)),
689 (('ige', ('imin', a, b), a), ('ige', b, a)),
690 (('ult', a, ('umin', b, a)), False),
691 (('ult', ('umax', a, b), a), False),
692 (('uge', a, ('umax', b, a)), ('uge', a, b)),
693 (('uge', ('umin', a, b), a), ('uge', b, a)),
694 (('ult', a, ('iand', b, a)), False),
695 (('ult', ('ior', a, b), a), False),
696 (('uge', a, ('iand', b, a)), True),
697 (('uge', ('ior', a, b), a), True),
698
699 (('ilt', '#a', ('imax', '#b', c)), ('ior', ('ilt', a, b), ('ilt', a, c))),
700 (('ilt', ('imin', '#a', b), '#c'), ('ior', ('ilt', a, c), ('ilt', b, c))),
701 (('ige', '#a', ('imin', '#b', c)), ('ior', ('ige', a, b), ('ige', a, c))),
702 (('ige', ('imax', '#a', b), '#c'), ('ior', ('ige', a, c), ('ige', b, c))),
703 (('ult', '#a', ('umax', '#b', c)), ('ior', ('ult', a, b), ('ult', a, c))),
704 (('ult', ('umin', '#a', b), '#c'), ('ior', ('ult', a, c), ('ult', b, c))),
705 (('uge', '#a', ('umin', '#b', c)), ('ior', ('uge', a, b), ('uge', a, c))),
706 (('uge', ('umax', '#a', b), '#c'), ('ior', ('uge', a, c), ('uge', b, c))),
707 (('ilt', '#a', ('imin', '#b', c)), ('iand', ('ilt', a, b), ('ilt', a, c))),
708 (('ilt', ('imax', '#a', b), '#c'), ('iand', ('ilt', a, c), ('ilt', b, c))),
709 (('ige', '#a', ('imax', '#b', c)), ('iand', ('ige', a, b), ('ige', a, c))),
710 (('ige', ('imin', '#a', b), '#c'), ('iand', ('ige', a, c), ('ige', b, c))),
711 (('ult', '#a', ('umin', '#b', c)), ('iand', ('ult', a, b), ('ult', a, c))),
712 (('ult', ('umax', '#a', b), '#c'), ('iand', ('ult', a, c), ('ult', b, c))),
713 (('uge', '#a', ('umax', '#b', c)), ('iand', ('uge', a, b), ('uge', a, c))),
714 (('uge', ('umin', '#a', b), '#c'), ('iand', ('uge', a, c), ('uge', b, c))),
715
716 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
717 # negative.
718 (('bcsel', ('ilt', a, 0), ('ineg', ('ishr', a, b)), ('ishr', a, b)),
719 ('iabs', ('ishr', a, b))),
720 (('iabs', ('ishr', ('iabs', a), b)), ('ishr', ('iabs', a), b)),
721
722 (('fabs', ('slt', a, b)), ('slt', a, b)),
723 (('fabs', ('sge', a, b)), ('sge', a, b)),
724 (('fabs', ('seq', a, b)), ('seq', a, b)),
725 (('fabs', ('sne', a, b)), ('sne', a, b)),
726 (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
727 (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
728 (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
729 (('sne', a, b), ('b2f', ('fneu', a, b)), 'options->lower_scmp'),
730 (('seq', ('seq', a, b), 1.0), ('seq', a, b)),
731 (('seq', ('sne', a, b), 1.0), ('sne', a, b)),
732 (('seq', ('slt', a, b), 1.0), ('slt', a, b)),
733 (('seq', ('sge', a, b), 1.0), ('sge', a, b)),
734 (('sne', ('seq', a, b), 0.0), ('seq', a, b)),
735 (('sne', ('sne', a, b), 0.0), ('sne', a, b)),
736 (('sne', ('slt', a, b), 0.0), ('slt', a, b)),
737 (('sne', ('sge', a, b), 0.0), ('sge', a, b)),
738 (('seq', ('seq', a, b), 0.0), ('sne', a, b)),
739 (('seq', ('sne', a, b), 0.0), ('seq', a, b)),
740 (('seq', ('slt', a, b), 0.0), ('sge', a, b)),
741 (('seq', ('sge', a, b), 0.0), ('slt', a, b)),
742 (('sne', ('seq', a, b), 1.0), ('sne', a, b)),
743 (('sne', ('sne', a, b), 1.0), ('seq', a, b)),
744 (('sne', ('slt', a, b), 1.0), ('sge', a, b)),
745 (('sne', ('sge', a, b), 1.0), ('slt', a, b)),
746 (('fall_equal2', a, b), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
747 (('fall_equal3', a, b), ('seq', ('fany_nequal3', a, b), 0.0), 'options->lower_vector_cmp'),
748 (('fall_equal4', a, b), ('seq', ('fany_nequal4', a, b), 0.0), 'options->lower_vector_cmp'),
749 (('fany_nequal2', a, b), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
750 (('fany_nequal3', a, b), ('fsat', ('fdot3', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),
751 (('fany_nequal4', a, b), ('fsat', ('fdot4', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),
752 (('fneu', ('fneg', a), a), ('fneu', a, 0.0)),
753 (('feq', ('fneg', a), a), ('feq', a, 0.0)),
754 # Emulating booleans
755 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
756 (('iand', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
757 (('ior', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('ior', a, b))),
758 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
759 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a, b))),
760 (('iand', 'a@bool32', 1.0), ('b2f', a)),
761 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
762 (('ineg', ('b2i32', 'a@32')), a),
763 (('flt', ('fneg', ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
764 # Comparison with the same args. Note that these are not done for
765 # the float versions because NaN always returns false on float
766 # inequalities.
767 (('ilt', a, a), False),
768 (('ige', a, a), True),
769 (('ieq', a, a), True),
770 (('ine', a, a), False),
771 (('ult', a, a), False),
772 (('uge', a, a), True),
773 # Logical and bit operations
774 (('iand', a, a), a),
775 (('iand', a, ~0), a),
776 (('iand', a, 0), 0),
777 (('ior', a, a), a),
778 (('ior', a, 0), a),
779 (('ior', a, True), True),
780 (('ixor', a, a), 0),
781 (('ixor', a, 0), a),
782 (('inot', ('inot', a)), a),
783 (('ior', ('iand', a, b), b), b),
784 (('ior', ('ior', a, b), b), ('ior', a, b)),
785 (('iand', ('ior', a, b), b), b),
786 (('iand', ('iand', a, b), b), ('iand', a, b)),
787 # DeMorgan's Laws
788 (('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))),
789 (('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))),
790 # Shift optimizations
791 (('ishl', 0, a), 0),
792 (('ishl', a, 0), a),
793 (('ishr', 0, a), 0),
794 (('ishr', a, 0), a),
795 (('ushr', 0, a), 0),
796 (('ushr', a, 0), a),
797 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('iadd', 16, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
798 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('isub', 16, b))), ('urol', a, b), '!options->lower_rotate'),
799 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('iadd', 32, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
800 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('isub', 32, b))), ('urol', a, b), '!options->lower_rotate'),
801 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('iadd', 16, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
802 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('isub', 16, b))), ('uror', a, b), '!options->lower_rotate'),
803 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('iadd', 32, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
804 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('isub', 32, b))), ('uror', a, b), '!options->lower_rotate'),
805 (('urol@16', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 16, b))), 'options->lower_rotate'),
806 (('urol@32', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 32, b))), 'options->lower_rotate'),
807 (('uror@16', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 16, b))), 'options->lower_rotate'),
808 (('uror@32', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 32, b))), 'options->lower_rotate'),
809 # Exponential/logarithmic identities
810 (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
811 (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
812 (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
813 (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
814 (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
815 ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
816 (('~fexp2', ('fmul', ('flog2', a), 0.5)), ('fsqrt', a)),
817 (('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)),
818 (('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))),
819 (('~fpow', a, 1.0), a),
820 (('~fpow', a, 2.0), ('fmul', a, a)),
821 (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
822 (('~fpow', 2.0, a), ('fexp2', a)),
823 (('~fpow', ('fpow', a, 2.2), 0.454545), a),
824 (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
825 (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
826 (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
827 (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
828 (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
829 (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
830 (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
831 (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
832 (('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))),
833 (('bcsel', ('flt', a, 0.0), 0.0, ('fsqrt', a)), ('fsqrt', ('fmax', a, 0.0))),
834 (('~fmul', ('fsqrt', a), ('fsqrt', a)), ('fabs',a)),
835 # Division and reciprocal
836 (('~fdiv', 1.0, a), ('frcp', a)),
837 (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
838 (('~frcp', ('frcp', a)), a),
839 (('~frcp', ('fsqrt', a)), ('frsq', a)),
840 (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
841 (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
842 # Trig
843 (('fsin', a), lowered_sincos(0.5), 'options->lower_sincos'),
844 (('fcos', a), lowered_sincos(0.75), 'options->lower_sincos'),
845 # Boolean simplifications
846 (('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
847 (('i2b1(is_used_by_if)', a), ('ine', a, 0)),
848 (('ieq', a, True), a),
849 (('ine(is_not_used_by_if)', a, True), ('inot', a)),
850 (('ine', a, False), a),
851 (('ieq(is_not_used_by_if)', a, False), ('inot', 'a')),
852 (('bcsel', a, True, False), a),
853 (('bcsel', a, False, True), ('inot', a)),
854 (('bcsel@32', a, 1.0, 0.0), ('b2f', a)),
855 (('bcsel@32', a, 0.0, 1.0), ('b2f', ('inot', a))),
856 (('bcsel@32', a, -1.0, -0.0), ('fneg', ('b2f', a))),
857 (('bcsel@32', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))),
858 (('bcsel', True, b, c), b),
859 (('bcsel', False, b, c), c),
860 (('bcsel', a, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a, b, c))),
861
862 (('bcsel', a, b, b), b),
863 (('~fcsel', a, b, b), b),
864
865 # D3D Boolean emulation
866 (('bcsel', a, -1, 0), ('ineg', ('b2i', 'a@1'))),
867 (('bcsel', a, 0, -1), ('ineg', ('b2i', ('inot', a)))),
868 (('bcsel', a, 1, 0), ('b2i', 'a@1')),
869 (('bcsel', a, 0, 1), ('b2i', ('inot', a))),
870 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
871 ('ineg', ('b2i', ('iand', a, b)))),
872 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
873 ('ineg', ('b2i', ('ior', a, b)))),
874 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a)),
875 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a),
876 (('ine', ('ineg', ('b2i', 'a@1')), 0), a),
877 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a)),
878 (('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)),
879 (('iand', ('ineg', ('b2i', a)), 1), ('b2i', a)),
880
881 # SM5 32-bit shifts are defined to use the 5 least significant bits
882 (('ishl', 'a@32', ('iand', 31, b)), ('ishl', a, b)),
883 (('ishr', 'a@32', ('iand', 31, b)), ('ishr', a, b)),
884 (('ushr', 'a@32', ('iand', 31, b)), ('ushr', a, b)),
885
886 # Conversions
887 (('i2b32', ('b2i', 'a@32')), a),
888 (('f2i', ('ftrunc', a)), ('f2i', a)),
889 (('f2u', ('ftrunc', a)), ('f2u', a)),
890 (('i2b', ('ineg', a)), ('i2b', a)),
891 (('i2b', ('iabs', a)), ('i2b', a)),
892 (('inot', ('f2b1', a)), ('feq', a, 0.0)),
893
894 # The C spec says, "If the value of the integral part cannot be represented
895 # by the integer type, the behavior is undefined." "Undefined" can mean
896 # "the conversion doesn't happen at all."
897 (('~i2f32', ('f2i32', 'a@32')), ('ftrunc', a)),
898
899 # Ironically, mark these as imprecise because removing the conversions may
900 # preserve more precision than doing the conversions (e.g.,
901 # uint(float(0x81818181u)) == 0x81818200).
902 (('~f2i32', ('i2f', 'a@32')), a),
903 (('~f2i32', ('u2f', 'a@32')), a),
904 (('~f2u32', ('i2f', 'a@32')), a),
905 (('~f2u32', ('u2f', 'a@32')), a),
906
907 # Conversions from 16 bits to 32 bits and back can always be removed
908 (('f2f16', ('f2f32', 'a@16')), a),
909 (('f2fmp', ('f2f32', 'a@16')), a),
910 (('i2i16', ('i2i32', 'a@16')), a),
911 (('i2imp', ('i2i32', 'a@16')), a),
912 (('u2u16', ('u2u32', 'a@16')), a),
913 (('u2ump', ('u2u32', 'a@16')), a),
914 (('f2f16', ('b2f32', 'a@1')), ('b2f16', a)),
915 (('f2fmp', ('b2f32', 'a@1')), ('b2f16', a)),
916 (('i2i16', ('b2i32', 'a@1')), ('b2i16', a)),
917 (('i2imp', ('b2i32', 'a@1')), ('b2i16', a)),
918 (('u2u16', ('b2i32', 'a@1')), ('b2i16', a)),
919 (('u2ump', ('b2i32', 'a@1')), ('b2i16', a)),
920 # Conversions to 16 bits would be lossy so they should only be removed if
921 # the instruction was generated by the precision lowering pass.
922 (('f2f32', ('f2fmp', 'a@32')), a),
923 (('i2i32', ('i2imp', 'a@32')), a),
924 (('u2u32', ('u2ump', 'a@32')), a),
925
926 (('ffloor', 'a(is_integral)'), a),
927 (('fceil', 'a(is_integral)'), a),
928 (('ftrunc', 'a(is_integral)'), a),
929 # fract(x) = x - floor(x), so fract(NaN) = NaN
930 (('~ffract', 'a(is_integral)'), 0.0),
931 (('fabs', 'a(is_not_negative)'), a),
932 (('iabs', 'a(is_not_negative)'), a),
933 (('fsat', 'a(is_not_positive)'), 0.0),
934
935 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
936 # says:
937 #
938 # It is undefined to convert a negative floating-point value to an
939 # uint.
940 #
941 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
942 # some optimizations in the i965 backend to proceed.
943 (('ige', ('f2u', a), b), ('ige', ('f2i', a), b)),
944 (('ige', b, ('f2u', a)), ('ige', b, ('f2i', a))),
945 (('ilt', ('f2u', a), b), ('ilt', ('f2i', a), b)),
946 (('ilt', b, ('f2u', a)), ('ilt', b, ('f2i', a))),
947
948 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a), '!options->lower_fsat'),
949
950 # The result of the multiply must be in [-1, 0], so the result of the ffma
951 # must be in [0, 1].
952 (('flt', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), False),
953 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), False),
954 (('fmax', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0)),
955 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0)),
956
957 (('fneu', 'a(is_not_zero)', 0.0), True),
958 (('feq', 'a(is_not_zero)', 0.0), False),
959
960 # In this chart, + means value > 0 and - means value < 0.
961 #
962 # + >= + -> unknown 0 >= + -> false - >= + -> false
963 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
964 # + >= - -> true 0 >= - -> true - >= - -> unknown
965 #
966 # Using grouping conceptually similar to a Karnaugh map...
967 #
968 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
969 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
970 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
971 #
972 # The flt / ilt cases just invert the expected result.
973 #
974 # The results expecting true, must be marked imprecise. The results
975 # expecting false are fine because NaN compared >= or < anything is false.
976
977 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
978 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
979 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
980
981 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
982 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
983 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
984
985 (('ine', 'a(is_not_zero)', 0), True),
986 (('ieq', 'a(is_not_zero)', 0), False),
987
988 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
989 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
990 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
991
992 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
993 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
994 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
995
996 (('ult', 0, 'a(is_gt_zero)'), True),
997 (('ult', a, 0), False),
998
999 # Packing and then unpacking does nothing
1000 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a),
1001 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b),
1002 (('unpack_64_2x32', ('pack_64_2x32_split', a, b)), ('vec2', a, b)),
1003 (('unpack_64_2x32', ('pack_64_2x32', a)), a),
1004 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a),
1005 ('unpack_64_2x32_split_y', a)), a),
1006 (('pack_64_2x32', ('vec2', ('unpack_64_2x32_split_x', a),
1007 ('unpack_64_2x32_split_y', a))), a),
1008 (('pack_64_2x32', ('unpack_64_2x32', a)), a),
1009
1010 # Comparing two halves of an unpack separately. While this optimization
1011 # should be correct for non-constant values, it's less obvious that it's
1012 # useful in that case. For constant values, the pack will fold and we're
1013 # guaranteed to reduce the whole tree to one instruction.
1014 (('iand', ('ieq', ('unpack_32_2x16_split_x', a), '#b'),
1015 ('ieq', ('unpack_32_2x16_split_y', a), '#c')),
1016 ('ieq', a, ('pack_32_2x16_split', b, c))),
1017
1018 # Byte extraction
1019 (('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
1020 (('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
1021 (('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
1022 (('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
1023 (('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
1024 (('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
1025 (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'),
1026
1027 (('ubfe', a, 0, 8), ('extract_u8', a, 0), '!options->lower_extract_byte'),
1028 (('ubfe', a, 8, 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
1029 (('ubfe', a, 16, 8), ('extract_u8', a, 2), '!options->lower_extract_byte'),
1030 (('ubfe', a, 24, 8), ('extract_u8', a, 3), '!options->lower_extract_byte'),
1031 (('ibfe', a, 0, 8), ('extract_i8', a, 0), '!options->lower_extract_byte'),
1032 (('ibfe', a, 8, 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
1033 (('ibfe', a, 16, 8), ('extract_i8', a, 2), '!options->lower_extract_byte'),
1034 (('ibfe', a, 24, 8), ('extract_i8', a, 3), '!options->lower_extract_byte'),
1035
1036 # Word extraction
1037 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
1038 (('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
1039 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
1040 (('ishr', 'a@32', 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
1041 (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),
1042
1043 (('ubfe', a, 0, 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
1044 (('ubfe', a, 16, 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
1045 (('ibfe', a, 0, 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
1046 (('ibfe', a, 16, 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
1047
1048 # Lower pack/unpack
1049 (('pack_64_2x32_split', a, b), ('ior', ('u2u64', a), ('ishl', ('u2u64', b), 32)), 'options->lower_pack_64_2x32_split'),
1050 (('pack_32_2x16_split', a, b), ('ior', ('u2u32', a), ('ishl', ('u2u32', b), 16)), 'options->lower_pack_32_2x16_split'),
1051 (('unpack_64_2x32_split_x', a), ('u2u32', a), 'options->lower_unpack_64_2x32_split'),
1052 (('unpack_64_2x32_split_y', a), ('u2u32', ('ushr', a, 32)), 'options->lower_unpack_64_2x32_split'),
1053 (('unpack_32_2x16_split_x', a), ('u2u16', a), 'options->lower_unpack_32_2x16_split'),
1054 (('unpack_32_2x16_split_y', a), ('u2u16', ('ushr', a, 16)), 'options->lower_unpack_32_2x16_split'),
1055
1056 # Useless masking before unpacking
1057 (('unpack_half_2x16_split_x', ('iand', a, 0xffff)), ('unpack_half_2x16_split_x', a)),
1058 (('unpack_32_2x16_split_x', ('iand', a, 0xffff)), ('unpack_32_2x16_split_x', a)),
1059 (('unpack_64_2x32_split_x', ('iand', a, 0xffffffff)), ('unpack_64_2x32_split_x', a)),
1060 (('unpack_half_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_half_2x16_split_y', a)),
1061 (('unpack_32_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_32_2x16_split_y', a)),
1062 (('unpack_64_2x32_split_y', ('iand', a, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a)),
1063
1064 (('unpack_half_2x16_split_x', ('extract_u16', a, 0)), ('unpack_half_2x16_split_x', a)),
1065 (('unpack_half_2x16_split_x', ('extract_u16', a, 1)), ('unpack_half_2x16_split_y', a)),
1066 (('unpack_32_2x16_split_x', ('extract_u16', a, 0)), ('unpack_32_2x16_split_x', a)),
1067 (('unpack_32_2x16_split_x', ('extract_u16', a, 1)), ('unpack_32_2x16_split_y', a)),
1068
1069 # Optimize half packing
1070 (('ishl', ('pack_half_2x16', ('vec2', a, 0)), 16), ('pack_half_2x16', ('vec2', 0, a))),
1071 (('ushr', ('pack_half_2x16', ('vec2', 0, a)), 16), ('pack_half_2x16', ('vec2', a, 0))),
1072
1073 (('iadd', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
1074 ('pack_half_2x16', ('vec2', a, b))),
1075 (('ior', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
1076 ('pack_half_2x16', ('vec2', a, b))),
1077
1078 (('ishl', ('pack_half_2x16_split', a, 0), 16), ('pack_half_2x16_split', 0, a)),
1079 (('ushr', ('pack_half_2x16_split', 0, a), 16), ('pack_half_2x16_split', a, 0)),
1080 (('extract_u16', ('pack_half_2x16_split', 0, a), 1), ('pack_half_2x16_split', a, 0)),
1081
1082 (('iadd', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
1083 (('ior', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
1084 ])
1085
1086 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
1087 # patterns like those below.
1088 for op in ('ushr', 'ishr'):
1089 optimizations.extend([(('extract_u8', (op, 'a@16', 8), 0), ('extract_u8', a, 1))])
1090 optimizations.extend([(('extract_u8', (op, 'a@32', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)])
1091 optimizations.extend([(('extract_u8', (op, 'a@64', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)])
1092
1093 optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))])
1094
1095 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
1096 # patterns like those below.
1097 for op in ('extract_u8', 'extract_i8'):
1098 optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))])
1099 optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)])
1100 optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)])
1101
1102 optimizations.extend([
1103 # Subtracts
1104 (('ussub_4x8', a, 0), a),
1105 (('ussub_4x8', a, ~0), 0),
1106 # Lower all Subtractions first - they can get recombined later
1107 (('fsub', a, b), ('fadd', a, ('fneg', b))),
1108 (('isub', a, b), ('iadd', a, ('ineg', b))),
1109 (('uabs_usub', a, b), ('bcsel', ('ult', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),
1110 # This is correct. We don't need isub_sat because the result type is unsigned, so it cannot overflow.
1111 (('uabs_isub', a, b), ('bcsel', ('ilt', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),
1112
1113 # Propagate negation up multiplication chains
1114 (('fmul(is_used_by_non_fsat)', ('fneg', a), b), ('fneg', ('fmul', a, b))),
1115 (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),
1116
1117 # Propagate constants up multiplication chains
1118 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)),
1119 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)),
1120 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)),
1121 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)),
1122
1123 # Reassociate constants in add/mul chains so they can be folded together.
1124 # For now, we mostly only handle cases where the constants are separated by
1125 # a single non-constant. We could do better eventually.
1126 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a, c), b)),
1127 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a, c), b)),
1128 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a, c), b)),
1129 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a, ('fneg', c)), ('fneg', b))),
1130 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a, c), b)),
1131 (('iand', '#a', ('iand', 'b(is_not_const)', '#c')), ('iand', ('iand', a, c), b)),
1132 (('ior', '#a', ('ior', 'b(is_not_const)', '#c')), ('ior', ('ior', a, c), b)),
1133 (('ixor', '#a', ('ixor', 'b(is_not_const)', '#c')), ('ixor', ('ixor', a, c), b)),
1134
1135 # Drop mul-div by the same value when there's no wrapping.
1136 (('idiv', ('imul(no_signed_wrap)', a, b), b), a),
1137
1138 # By definition...
1139 (('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)),
1140 (('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
1141 (('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
1142
1143 (('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)),
1144 (('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
1145 (('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
1146
1147 (('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),
1148
1149 (('~fmul', ('bcsel(is_used_once)', c, -1.0, 1.0), b), ('bcsel', c, ('fneg', b), b)),
1150 (('~fmul', ('bcsel(is_used_once)', c, 1.0, -1.0), b), ('bcsel', c, b, ('fneg', b))),
1151 (('~bcsel', ('flt', a, 0.0), ('fneg', a), a), ('fabs', a)),
1152
1153 (('bcsel', a, ('bcsel', b, c, d), d), ('bcsel', ('iand', a, b), c, d)),
1154 (('bcsel', a, b, ('bcsel', c, b, d)), ('bcsel', ('ior', a, c), b, d)),
1155
1156 # Misc. lowering
1157 (('fmod', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'),
1158 (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod'),
1159 (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
1160 (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),
1161
1162 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1163 ('bcsel', ('ult', 31, 'bits'), 'insert',
1164 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
1165 'options->lower_bitfield_insert'),
1166 (('ihadd', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1167 (('uhadd', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1168 (('irhadd', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1169 (('urhadd', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1170 (('ihadd@64', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1171 (('uhadd@64', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1172 (('irhadd@64', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1173 (('urhadd@64', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1174
1175 (('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'),
1176 (('uadd_sat', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat'),
1177 (('usub_sat', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_add_sat'),
1178 (('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'),
1179
1180 # int64_t sum = a + b;
1181 #
1182 # if (a < 0 && b < 0 && a < sum)
1183 # sum = INT64_MIN;
1184 # } else if (a >= 0 && b >= 0 && sum < a)
1185 # sum = INT64_MAX;
1186 # }
1187 #
1188 # A couple optimizations are applied.
1189 #
1190 # 1. a < sum => sum >= 0. This replacement works because it is known that
1191 # a < 0 and b < 0, so sum should also be < 0 unless there was
1192 # underflow.
1193 #
1194 # 2. sum < a => 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 # overflow.
1197 #
1198 # 3. Invert the second if-condition and swap the order of parameters for
1199 # the bcsel. !(a >= 0 && b >= 0 && sum < 0) becomes !(a >= 0) || !(b >=
1200 # 0) || !(sum < 0), and that becomes (a < 0) || (b < 0) || (sum >= 0)
1201 #
1202 # On Intel Gen11, this saves ~11 instructions.
1203 (('iadd_sat@64', a, b), ('bcsel',
1204 ('iand', ('iand', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
1205 0x8000000000000000,
1206 ('bcsel',
1207 ('ior', ('ior', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
1208 ('iadd', a, b),
1209 0x7fffffffffffffff)),
1210 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1211
1212 # int64_t sum = a - b;
1213 #
1214 # if (a < 0 && b >= 0 && a < sum)
1215 # sum = INT64_MIN;
1216 # } else if (a >= 0 && b < 0 && a >= sum)
1217 # sum = INT64_MAX;
1218 # }
1219 #
1220 # Optimizations similar to the iadd_sat case are applied here.
1221 (('isub_sat@64', a, b), ('bcsel',
1222 ('iand', ('iand', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
1223 0x8000000000000000,
1224 ('bcsel',
1225 ('ior', ('ior', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
1226 ('isub', a, b),
1227 0x7fffffffffffffff)),
1228 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1229
1230 # These are done here instead of in the backend because the int64 lowering
1231 # pass will make a mess of the patterns. The first patterns are
1232 # conditioned on nir_lower_minmax64 because it was not clear that it was
1233 # always an improvement on platforms that have real int64 support. No
1234 # shaders in shader-db hit this, so it was hard to say one way or the
1235 # other.
1236 (('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'),
1237 (('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'),
1238 (('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'),
1239 (('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'),
1240 (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1241 (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1242
1243 (('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'),
1244 (('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'),
1245 # 0u < uint(a) <=> uint(a) != 0u
1246 (('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'),
1247
1248 # Alternative lowering that doesn't rely on bfi.
1249 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1250 ('bcsel', ('ult', 31, 'bits'),
1251 'insert',
1252 (('ior',
1253 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1254 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1255 'options->lower_bitfield_insert_to_shifts'),
1256
1257 # Alternative lowering that uses bitfield_select.
1258 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1259 ('bcsel', ('ult', 31, 'bits'), 'insert',
1260 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1261 'options->lower_bitfield_insert_to_bitfield_select'),
1262
1263 (('ibitfield_extract', 'value', 'offset', 'bits'),
1264 ('bcsel', ('ult', 31, 'bits'), 'value',
1265 ('ibfe', 'value', 'offset', 'bits')),
1266 'options->lower_bitfield_extract'),
1267
1268 (('ubitfield_extract', 'value', 'offset', 'bits'),
1269 ('bcsel', ('ult', 31, 'bits'), 'value',
1270 ('ubfe', 'value', 'offset', 'bits')),
1271 'options->lower_bitfield_extract'),
1272
1273 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1274 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1275 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1276 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1277 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1278 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1279 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1280
1281 # Section 8.8 (Integer Functions) of the GLSL 4.60 spec says:
1282 #
1283 # If bits is zero, the result will be zero.
1284 #
1285 # These patterns prevent other patterns from generating invalid results
1286 # when count is zero.
1287 (('ubfe', a, b, 0), 0),
1288 (('ibfe', a, b, 0), 0),
1289
1290 (('ubfe', a, 0, '#b'), ('iand', a, ('ushr', 0xffffffff, ('ineg', b)))),
1291
1292 (('b2i32', ('i2b', ('ubfe', a, b, 1))), ('ubfe', a, b, 1)),
1293 (('b2i32', ('i2b', ('ibfe', a, b, 1))), ('ubfe', a, b, 1)), # ubfe in the replacement is correct
1294 (('ine', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1295 (('ieq', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1296 (('ine', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1297 (('ieq', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1298
1299 (('ibitfield_extract', 'value', 'offset', 'bits'),
1300 ('bcsel', ('ieq', 0, 'bits'),
1301 0,
1302 ('ishr',
1303 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1304 ('isub', 32, 'bits'))),
1305 'options->lower_bitfield_extract_to_shifts'),
1306
1307 (('ubitfield_extract', 'value', 'offset', 'bits'),
1308 ('iand',
1309 ('ushr', 'value', 'offset'),
1310 ('bcsel', ('ieq', 'bits', 32),
1311 0xffffffff,
1312 ('isub', ('ishl', 1, 'bits'), 1))),
1313 'options->lower_bitfield_extract_to_shifts'),
1314
1315 (('ifind_msb', 'value'),
1316 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1317 'options->lower_ifind_msb'),
1318
1319 (('find_lsb', 'value'),
1320 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1321 'options->lower_find_lsb'),
1322
1323 (('extract_i8', a, 'b@32'),
1324 ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24),
1325 'options->lower_extract_byte'),
1326
1327 (('extract_u8', a, 'b@32'),
1328 ('iand', ('ushr', a, ('imul', b, 8)), 0xff),
1329 'options->lower_extract_byte'),
1330
1331 (('extract_i16', a, 'b@32'),
1332 ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16),
1333 'options->lower_extract_word'),
1334
1335 (('extract_u16', a, 'b@32'),
1336 ('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
1337 'options->lower_extract_word'),
1338
1339 (('pack_unorm_2x16', 'v'),
1340 ('pack_uvec2_to_uint',
1341 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1342 'options->lower_pack_unorm_2x16'),
1343
1344 (('pack_unorm_4x8', 'v'),
1345 ('pack_uvec4_to_uint',
1346 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1347 'options->lower_pack_unorm_4x8'),
1348
1349 (('pack_snorm_2x16', 'v'),
1350 ('pack_uvec2_to_uint',
1351 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1352 'options->lower_pack_snorm_2x16'),
1353
1354 (('pack_snorm_4x8', 'v'),
1355 ('pack_uvec4_to_uint',
1356 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1357 'options->lower_pack_snorm_4x8'),
1358
1359 (('unpack_unorm_2x16', 'v'),
1360 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1361 ('extract_u16', 'v', 1))),
1362 65535.0),
1363 'options->lower_unpack_unorm_2x16'),
1364
1365 (('unpack_unorm_4x8', 'v'),
1366 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1367 ('extract_u8', 'v', 1),
1368 ('extract_u8', 'v', 2),
1369 ('extract_u8', 'v', 3))),
1370 255.0),
1371 'options->lower_unpack_unorm_4x8'),
1372
1373 (('unpack_snorm_2x16', 'v'),
1374 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1375 ('extract_i16', 'v', 1))),
1376 32767.0))),
1377 'options->lower_unpack_snorm_2x16'),
1378
1379 (('unpack_snorm_4x8', 'v'),
1380 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1381 ('extract_i8', 'v', 1),
1382 ('extract_i8', 'v', 2),
1383 ('extract_i8', 'v', 3))),
1384 127.0))),
1385 'options->lower_unpack_snorm_4x8'),
1386
1387 (('pack_half_2x16_split', 'a@32', 'b@32'),
1388 ('ior', ('ishl', ('u2u32', ('f2f16', b)), 16), ('u2u32', ('f2f16', a))),
1389 'options->lower_pack_split'),
1390
1391 (('unpack_half_2x16_split_x', 'a@32'),
1392 ('f2f32', ('u2u16', a)),
1393 'options->lower_pack_split'),
1394
1395 (('unpack_half_2x16_split_y', 'a@32'),
1396 ('f2f32', ('u2u16', ('ushr', a, 16))),
1397 'options->lower_pack_split'),
1398
1399 (('pack_32_2x16_split', 'a@16', 'b@16'),
1400 ('ior', ('ishl', ('u2u32', b), 16), ('u2u32', a)),
1401 'options->lower_pack_split'),
1402
1403 (('unpack_32_2x16_split_x', 'a@32'),
1404 ('u2u16', a),
1405 'options->lower_pack_split'),
1406
1407 (('unpack_32_2x16_split_y', 'a@32'),
1408 ('u2u16', ('ushr', 'a', 16)),
1409 'options->lower_pack_split'),
1410
1411 (('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
1412 (('imin', ('imax', a, -1), 1), ('isign', a), '!options->lower_isign'),
1413 (('imax', ('imin', a, 1), -1), ('isign', a), '!options->lower_isign'),
1414 (('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),
1415 (('fadd', ('b2f32', ('flt', 0.0, 'a@32')), ('fneg', ('b2f32', ('flt', 'a@32', 0.0)))), ('fsign', a), '!options->lower_fsign'),
1416 (('iadd', ('b2i32', ('flt', 0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0)))), ('f2i32', ('fsign', a)), '!options->lower_fsign'),
1417
1418 # Address/offset calculations:
1419 # Drivers supporting imul24 should use the nir_lower_amul() pass, this
1420 # rule converts everyone else to imul:
1421 (('amul', a, b), ('imul', a, b), '!options->has_imul24'),
1422
1423 (('umul24', a, b),
1424 ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)),
1425 '!options->has_umul24'),
1426 (('umad24', a, b, c),
1427 ('iadd', ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)), c),
1428 '!options->has_umad24'),
1429
1430 (('imad24_ir3', a, b, 0), ('imul24', a, b)),
1431 (('imad24_ir3', a, 0, c), (c)),
1432 (('imad24_ir3', a, 1, c), ('iadd', a, c)),
1433
1434 # if first two srcs are const, crack apart the imad so constant folding
1435 # can clean up the imul:
1436 # TODO ffma should probably get a similar rule:
1437 (('imad24_ir3', '#a', '#b', c), ('iadd', ('imul', a, b), c)),
1438
1439 # These will turn 24b address/offset calc back into 32b shifts, but
1440 # it should be safe to get back some of the bits of precision that we
1441 # already decided were no necessary:
1442 (('imul24', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'),
1443 (('imul24', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'),
1444 (('imul24', a, 0), (0)),
1445 ])
1446
1447 # bit_size dependent lowerings
1448 for bit_size in [8, 16, 32, 64]:
1449 # convenience constants
1450 intmax = (1 << (bit_size - 1)) - 1
1451 intmin = 1 << (bit_size - 1)
1452
1453 optimizations += [
1454 (('iadd_sat@' + str(bit_size), a, b),
1455 ('bcsel', ('ige', b, 1), ('bcsel', ('ilt', ('iadd', a, b), a), intmax, ('iadd', a, b)),
1456 ('bcsel', ('ilt', a, ('iadd', a, b)), intmin, ('iadd', a, b))), 'options->lower_add_sat'),
1457 (('isub_sat@' + str(bit_size), a, b),
1458 ('bcsel', ('ilt', b, 0), ('bcsel', ('ilt', ('isub', a, b), a), intmax, ('isub', a, b)),
1459 ('bcsel', ('ilt', a, ('isub', a, b)), intmin, ('isub', a, b))), 'options->lower_add_sat'),
1460 ]
1461
1462 invert = OrderedDict([('feq', 'fneu'), ('fneu', 'feq')])
1463
1464 for left, right in itertools.combinations_with_replacement(invert.keys(), 2):
1465 optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))),
1466 ('iand', (invert[left], a, b), (invert[right], c, d))))
1467 optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))),
1468 ('ior', (invert[left], a, b), (invert[right], c, d))))
1469
1470 # Optimize x2bN(b2x(x)) -> x
1471 for size in type_sizes('bool'):
1472 aN = 'a@' + str(size)
1473 f2bN = 'f2b' + str(size)
1474 i2bN = 'i2b' + str(size)
1475 optimizations.append(((f2bN, ('b2f', aN)), a))
1476 optimizations.append(((i2bN, ('b2i', aN)), a))
1477
1478 # Optimize x2yN(b2x(x)) -> b2y
1479 for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1480 if x != 'f' and y != 'f' and x != y:
1481 continue
1482
1483 b2x = 'b2f' if x == 'f' else 'b2i'
1484 b2y = 'b2f' if y == 'f' else 'b2i'
1485 x2yN = '{}2{}'.format(x, y)
1486 optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
1487
1488 # Optimize away x2xN(a@N)
1489 for t in ['int', 'uint', 'float', 'bool']:
1490 for N in type_sizes(t):
1491 x2xN = '{0}2{0}{1}'.format(t[0], N)
1492 aN = 'a@{0}'.format(N)
1493 optimizations.append(((x2xN, aN), a))
1494
1495 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1496 # In particular, we can optimize away everything except upcast of downcast and
1497 # upcasts where the type differs from the other cast
1498 for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
1499 if N < M:
1500 # The outer cast is a down-cast. It doesn't matter what the size of the
1501 # argument of the inner cast is because we'll never been in the upcast
1502 # of downcast case. Regardless of types, we'll always end up with y2yN
1503 # in the end.
1504 for x, y in itertools.product(['i', 'u'], ['i', 'u']):
1505 x2xN = '{0}2{0}{1}'.format(x, N)
1506 y2yM = '{0}2{0}{1}'.format(y, M)
1507 y2yN = '{0}2{0}{1}'.format(y, N)
1508 optimizations.append(((x2xN, (y2yM, a)), (y2yN, a)))
1509 elif N > M:
1510 # If the outer cast is an up-cast, we have to be more careful about the
1511 # size of the argument of the inner cast and with types. In this case,
1512 # the type is always the type of type up-cast which is given by the
1513 # outer cast.
1514 for P in type_sizes('uint'):
1515 # We can't optimize away up-cast of down-cast.
1516 if M < P:
1517 continue
1518
1519 # Because we're doing down-cast of down-cast, the types always have
1520 # to match between the two casts
1521 for x in ['i', 'u']:
1522 x2xN = '{0}2{0}{1}'.format(x, N)
1523 x2xM = '{0}2{0}{1}'.format(x, M)
1524 aP = 'a@{0}'.format(P)
1525 optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a)))
1526 else:
1527 # The N == M case is handled by other optimizations
1528 pass
1529
1530 # Downcast operations should be able to see through pack
1531 for t in ['i', 'u']:
1532 for N in [8, 16, 32]:
1533 x2xN = '{0}2{0}{1}'.format(t, N)
1534 optimizations += [
1535 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
1536 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
1537 ]
1538
1539 # Optimize comparisons with up-casts
1540 for t in ['int', 'uint', 'float']:
1541 for N, M in itertools.product(type_sizes(t), repeat=2):
1542 if N == 1 or N >= M:
1543 continue
1544
1545 cond = 'true'
1546 if N == 8:
1547 cond = 'options->support_8bit_alu'
1548 elif N == 16:
1549 cond = 'options->support_16bit_alu'
1550 x2xM = '{0}2{0}{1}'.format(t[0], M)
1551 x2xN = '{0}2{0}{1}'.format(t[0], N)
1552 aN = 'a@' + str(N)
1553 bN = 'b@' + str(N)
1554 xeq = 'feq' if t == 'float' else 'ieq'
1555 xne = 'fneu' if t == 'float' else 'ine'
1556 xge = '{0}ge'.format(t[0])
1557 xlt = '{0}lt'.format(t[0])
1558
1559 # Up-casts are lossless so for correctly signed comparisons of
1560 # up-casted values we can do the comparison at the largest of the two
1561 # original sizes and drop one or both of the casts. (We have
1562 # optimizations to drop the no-op casts which this may generate.)
1563 for P in type_sizes(t):
1564 if P == 1 or P > N:
1565 continue
1566
1567 bP = 'b@' + str(P)
1568 optimizations += [
1569 ((xeq, (x2xM, aN), (x2xM, bP)), (xeq, a, (x2xN, b)), cond),
1570 ((xne, (x2xM, aN), (x2xM, bP)), (xne, a, (x2xN, b)), cond),
1571 ((xge, (x2xM, aN), (x2xM, bP)), (xge, a, (x2xN, b)), cond),
1572 ((xlt, (x2xM, aN), (x2xM, bP)), (xlt, a, (x2xN, b)), cond),
1573 ((xge, (x2xM, bP), (x2xM, aN)), (xge, (x2xN, b), a), cond),
1574 ((xlt, (x2xM, bP), (x2xM, aN)), (xlt, (x2xN, b), a), cond),
1575 ]
1576
1577 # The next bit doesn't work on floats because the range checks would
1578 # get way too complicated.
1579 if t in ['int', 'uint']:
1580 if t == 'int':
1581 xN_min = -(1 << (N - 1))
1582 xN_max = (1 << (N - 1)) - 1
1583 elif t == 'uint':
1584 xN_min = 0
1585 xN_max = (1 << N) - 1
1586 else:
1587 assert False
1588
1589 # If we're up-casting and comparing to a constant, we can unfold
1590 # the comparison into a comparison with the shrunk down constant
1591 # and a check that the constant fits in the smaller bit size.
1592 optimizations += [
1593 ((xeq, (x2xM, aN), '#b'),
1594 ('iand', (xeq, a, (x2xN, b)), (xeq, (x2xM, (x2xN, b)), b)), cond),
1595 ((xne, (x2xM, aN), '#b'),
1596 ('ior', (xne, a, (x2xN, b)), (xne, (x2xM, (x2xN, b)), b)), cond),
1597 ((xlt, (x2xM, aN), '#b'),
1598 ('iand', (xlt, xN_min, b),
1599 ('ior', (xlt, xN_max, b), (xlt, a, (x2xN, b)))), cond),
1600 ((xlt, '#a', (x2xM, bN)),
1601 ('iand', (xlt, a, xN_max),
1602 ('ior', (xlt, a, xN_min), (xlt, (x2xN, a), b))), cond),
1603 ((xge, (x2xM, aN), '#b'),
1604 ('iand', (xge, xN_max, b),
1605 ('ior', (xge, xN_min, b), (xge, a, (x2xN, b)))), cond),
1606 ((xge, '#a', (x2xM, bN)),
1607 ('iand', (xge, a, xN_min),
1608 ('ior', (xge, a, xN_max), (xge, (x2xN, a), b))), cond),
1609 ]
1610
1611 # Convert masking followed by signed downcast to just unsigned downcast
1612 optimizations += [
1613 (('i2i32', ('iand', 'a@64', 0xffffffff)), ('u2u32', a)),
1614 (('i2i16', ('iand', 'a@32', 0xffff)), ('u2u16', a)),
1615 (('i2i16', ('iand', 'a@64', 0xffff)), ('u2u16', a)),
1616 (('i2i8', ('iand', 'a@16', 0xff)), ('u2u8', a)),
1617 (('i2i8', ('iand', 'a@32', 0xff)), ('u2u8', a)),
1618 (('i2i8', ('iand', 'a@64', 0xff)), ('u2u8', a)),
1619 ]
1620
1621 def fexp2i(exp, bits):
1622 # Generate an expression which constructs value 2.0^exp or 0.0.
1623 #
1624 # We assume that exp is already in a valid range:
1625 #
1626 # * [-15, 15] for 16-bit float
1627 # * [-127, 127] for 32-bit float
1628 # * [-1023, 1023] for 16-bit float
1629 #
1630 # If exp is the lowest value in the valid range, a value of 0.0 is
1631 # constructed. Otherwise, the value 2.0^exp is constructed.
1632 if bits == 16:
1633 return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
1634 elif bits == 32:
1635 return ('ishl', ('iadd', exp, 127), 23)
1636 elif bits == 64:
1637 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
1638 else:
1639 assert False
1640
1641 def ldexp(f, exp, bits):
1642 # The maximum possible range for a normal exponent is [-126, 127] and,
1643 # throwing in denormals, you get a maximum range of [-149, 127]. This
1644 # means that we can potentially have a swing of +-276. If you start with
1645 # FLT_MAX, you actually have to do ldexp(FLT_MAX, -278) to get it to flush
1646 # all the way to zero. The GLSL spec only requires that we handle a subset
1647 # of this range. From version 4.60 of the spec:
1648 #
1649 # "If exp is greater than +128 (single-precision) or +1024
1650 # (double-precision), the value returned is undefined. If exp is less
1651 # than -126 (single-precision) or -1022 (double-precision), the value
1652 # returned may be flushed to zero. Additionally, splitting the value
1653 # into a significand and exponent using frexp() and then reconstructing
1654 # a floating-point value using ldexp() should yield the original input
1655 # for zero and all finite non-denormalized values."
1656 #
1657 # The SPIR-V spec has similar language.
1658 #
1659 # In order to handle the maximum value +128 using the fexp2i() helper
1660 # above, we have to split the exponent in half and do two multiply
1661 # operations.
1662 #
1663 # First, we clamp exp to a reasonable range. Specifically, we clamp to
1664 # twice the full range that is valid for the fexp2i() function above. If
1665 # exp/2 is the bottom value of that range, the fexp2i() expression will
1666 # yield 0.0f which, when multiplied by f, will flush it to zero which is
1667 # allowed by the GLSL and SPIR-V specs for low exponent values. If the
1668 # value is clamped from above, then it must have been above the supported
1669 # range of the GLSL built-in and therefore any return value is acceptable.
1670 if bits == 16:
1671 exp = ('imin', ('imax', exp, -30), 30)
1672 elif bits == 32:
1673 exp = ('imin', ('imax', exp, -254), 254)
1674 elif bits == 64:
1675 exp = ('imin', ('imax', exp, -2046), 2046)
1676 else:
1677 assert False
1678
1679 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1680 # (We use ishr which isn't the same for -1, but the -1 case still works
1681 # since we use exp-exp/2 as the second exponent.) While the spec
1682 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1683 # work with denormals and doesn't allow for the full swing in exponents
1684 # that you can get with normalized values. Instead, we create two powers
1685 # of two and multiply by them each in turn. That way the effective range
1686 # of our exponent is doubled.
1687 pow2_1 = fexp2i(('ishr', exp, 1), bits)
1688 pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
1689 return ('fmul', ('fmul', f, pow2_1), pow2_2)
1690
1691 optimizations += [
1692 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1693 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1694 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1695 ]
1696
1697 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1698 def bitfield_reverse(u):
1699 step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))
1700 step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8))
1701 step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4))
1702 step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2))
1703 step5 = ('ior(many-comm-expr)', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1))
1704
1705 return step5
1706
1707 optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1708
1709 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1710 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1711 # and, if a is a NaN then the second comparison will fail anyway.
1712 for op in ['flt', 'fge', 'feq']:
1713 optimizations += [
1714 (('iand', ('feq', a, a), (op, a, b)), ('!' + op, a, b)),
1715 (('iand', ('feq', a, a), (op, b, a)), ('!' + op, b, a)),
1716 ]
1717
1718 # Add optimizations to handle the case where the result of a ternary is
1719 # compared to a constant. This way we can take things like
1720 #
1721 # (a ? 0 : 1) > 0
1722 #
1723 # and turn it into
1724 #
1725 # a ? (0 > 0) : (1 > 0)
1726 #
1727 # which constant folding will eat for lunch. The resulting ternary will
1728 # further get cleaned up by the boolean reductions above and we will be
1729 # left with just the original variable "a".
1730 for op in ['flt', 'fge', 'feq', 'fneu',
1731 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1732 optimizations += [
1733 ((op, ('bcsel', 'a', '#b', '#c'), '#d'),
1734 ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))),
1735 ((op, '#d', ('bcsel', a, '#b', '#c')),
1736 ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
1737 ]
1738
1739
1740 # For example, this converts things like
1741 #
1742 # 1 + mix(0, a - 1, condition)
1743 #
1744 # into
1745 #
1746 # mix(1, (a-1)+1, condition)
1747 #
1748 # Other optimizations will rearrange the constants.
1749 for op in ['fadd', 'fmul', 'iadd', 'imul']:
1750 optimizations += [
1751 ((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d)))
1752 ]
1753
1754 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1755 # states:
1756 #
1757 # If neither layout qualifier is specified, derivatives in compute shaders
1758 # return zero, which is consistent with the handling of built-in texture
1759 # functions like texture() in GLSL 4.50 compute shaders.
1760 for op in ['fddx', 'fddx_fine', 'fddx_coarse',
1761 'fddy', 'fddy_fine', 'fddy_coarse']:
1762 optimizations += [
1763 ((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1764 ]
1765
1766 # Some optimizations for ir3-specific instructions.
1767 optimizations += [
1768 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1769 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1770 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1771 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1772 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1773 ]
1774
1775 # These kinds of sequences can occur after nir_opt_peephole_select.
1776 #
1777 # NOTE: fadd is not handled here because that gets in the way of ffma
1778 # generation in the i965 driver. Instead, fadd and ffma are handled in
1779 # late_optimizations.
1780
1781 for op in ['flrp']:
1782 optimizations += [
1783 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, c, e)), (op, b, c, ('bcsel', a, d, e))),
1784 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, c, e)), (op, b, c, ('bcsel', a, d, e))),
1785 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, e, d)), (op, b, ('bcsel', a, c, e), d)),
1786 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)),
1787 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, e, c, d)), (op, ('bcsel', a, b, e), c, d)),
1788 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', e, c, d)), (op, ('bcsel', a, b, e), c, d)),
1789 ]
1790
1791 for op in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']:
1792 optimizations += [
1793 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
1794 (('bcsel', a, (op + '(is_used_once)', b, 'c(is_not_const)'), (op, b, d)), (op, b, ('bcsel', a, c, d))),
1795 (('bcsel', a, (op, b, 'c(is_not_const)'), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
1796 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
1797 ]
1798
1799 for op in ['fpow']:
1800 optimizations += [
1801 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, d)), (op, b, ('bcsel', a, c, d))),
1802 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
1803 (('bcsel', a, (op + '(is_used_once)', b, c), (op, d, c)), (op, ('bcsel', a, b, d), c)),
1804 (('bcsel', a, (op, b, c), (op + '(is_used_once)', d, c)), (op, ('bcsel', a, b, d), c)),
1805 ]
1806
1807 for op in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos', 'fneg', 'fabs', 'fsign']:
1808 optimizations += [
1809 (('bcsel', c, (op + '(is_used_once)', a), (op + '(is_used_once)', b)), (op, ('bcsel', c, a, b))),
1810 ]
1811
1812 for op in ['ineg', 'iabs', 'inot', 'isign']:
1813 optimizations += [
1814 ((op, ('bcsel', c, '#a', '#b')), ('bcsel', c, (op, a), (op, b))),
1815 ]
1816
1817 # This section contains optimizations to propagate downsizing conversions of
1818 # constructed vectors into vectors of downsized components. Whether this is
1819 # useful depends on the SIMD semantics of the backend. On a true SIMD machine,
1820 # this reduces the register pressure of the vector itself and often enables the
1821 # conversions to be eliminated via other algebraic rules or constant folding.
1822 # In the worst case on a SIMD architecture, the propagated conversions may be
1823 # revectorized via nir_opt_vectorize so instruction count is minimally
1824 # impacted.
1825 #
1826 # On a machine with SIMD-within-a-register only, this actually
1827 # counterintuitively hurts instruction count. These machines are the same that
1828 # require vectorize_vec2_16bit, so we predicate the optimizations on that flag
1829 # not being set.
1830 #
1831 # Finally for scalar architectures, there should be no difference in generated
1832 # code since it all ends up scalarized at the end, but it might minimally help
1833 # compile-times.
1834
1835 for i in range(2, 4 + 1):
1836 for T in ('f', 'u', 'i'):
1837 vec_inst = ('vec' + str(i),)
1838
1839 indices = ['a', 'b', 'c', 'd']
1840 suffix_in = tuple((indices[j] + '@32') for j in range(i))
1841
1842 to_16 = '{}2{}16'.format(T, T)
1843 to_mp = '{}2{}mp'.format(T, T)
1844
1845 out_16 = tuple((to_16, indices[j]) for j in range(i))
1846 out_mp = tuple((to_mp, indices[j]) for j in range(i))
1847
1848 optimizations += [
1849 ((to_16, vec_inst + suffix_in), vec_inst + out_16, '!options->vectorize_vec2_16bit'),
1850 ((to_mp, vec_inst + suffix_in), vec_inst + out_mp, '!options->vectorize_vec2_16bit')
1851 ]
1852
1853 # This section contains "late" optimizations that should be run before
1854 # creating ffmas and calling regular optimizations for the final time.
1855 # Optimizations should go here if they help code generation and conflict
1856 # with the regular optimizations.
1857 before_ffma_optimizations = [
1858 # Propagate constants down multiplication chains
1859 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)),
1860 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)),
1861 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)),
1862 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)),
1863
1864 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
1865 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
1866 (('~fadd', ('fneg', a), a), 0.0),
1867 (('iadd', ('ineg', a), a), 0),
1868 (('iadd', ('ineg', a), ('iadd', a, b)), b),
1869 (('iadd', a, ('iadd', ('ineg', a), b)), b),
1870 (('~fadd', ('fneg', a), ('fadd', a, b)), b),
1871 (('~fadd', a, ('fadd', ('fneg', a), b)), b),
1872
1873 (('~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)),
1874 (('~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)),
1875 (('~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))),
1876 ]
1877
1878 # This section contains "late" optimizations that should be run after the
1879 # regular optimizations have finished. Optimizations should go here if
1880 # they help code generation but do not necessarily produce code that is
1881 # more easily optimizable.
1882 late_optimizations = [
1883 # Most of these optimizations aren't quite safe when you get infinity or
1884 # Nan involved but the first one should be fine.
1885 (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
1886 (('flt', ('fneg', ('fadd', a, b)), 0.0), ('flt', ('fneg', a), b)),
1887 (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
1888 (('~fge', ('fneg', ('fadd', a, b)), 0.0), ('fge', ('fneg', a), b)),
1889 (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
1890 (('~fneu', ('fadd', a, b), 0.0), ('fneu', a, ('fneg', b))),
1891
1892 # nir_lower_to_source_mods will collapse this, but its existence during the
1893 # optimization loop can prevent other optimizations.
1894 (('fneg', ('fneg', a)), a),
1895
1896 # Subtractions get lowered during optimization, so we need to recombine them
1897 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
1898 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
1899 (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
1900 (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
1901 (('iabs', a), ('imax', a, ('ineg', a)), 'options->lower_iabs'),
1902
1903 # These are duplicated from the main optimizations table. The late
1904 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1905 # new patterns like these. The patterns that compare with zero are removed
1906 # because they are unlikely to be created in by anything in
1907 # late_optimizations.
1908 (('flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
1909 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
1910 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
1911 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
1912 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
1913 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),
1914
1915 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
1916 (('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
1917
1918 (('~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)))),
1919
1920 (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
1921 (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
1922 (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
1923 (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)),
1924 (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
1925 (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
1926 (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
1927 (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
1928 (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)),
1929 (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),
1930
1931 (('ior', a, a), a),
1932 (('iand', a, a), a),
1933
1934 (('iand', ('ine(is_used_once)', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umin', a, b), 0)),
1935 (('ior', ('ieq(is_used_once)', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umin', a, b), 0)),
1936
1937 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),
1938
1939 (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'),
1940 (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'),
1941 (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
1942 (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
1943
1944 (('~flrp@32', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
1945 (('~flrp@64', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
1946
1947 (('~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'),
1948 (('~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'),
1949
1950 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1951 # particular operation is common for expanding values stored in a texture
1952 # from [0,1] to [-1,1].
1953 (('~ffma@32', a, 2.0, -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'),
1954 (('~ffma@32', a, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'),
1955 (('~ffma@32', a, -2.0, 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'),
1956 (('~ffma@32', a, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
1957 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'),
1958 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'),
1959 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'),
1960 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
1961
1962 # flrp(a, b, a)
1963 # a*(1-a) + b*a
1964 # a + -a*a + a*b (1)
1965 # a + a*(b - a)
1966 # Option 1: ffma(a, (b-a), a)
1967 #
1968 # Alternately, after (1):
1969 # a*(1+b) + -a*a
1970 # a*((1+b) + -a)
1971 #
1972 # Let b=1
1973 #
1974 # Option 2: ffma(a, 2, -(a*a))
1975 # Option 3: ffma(a, 2, (-a)*a)
1976 # Option 4: ffma(a, -a, (2*a)
1977 # Option 5: a * (2 - a)
1978 #
1979 # There are a lot of other possible combinations.
1980 (('~ffma@32', ('fadd', b, ('fneg', a)), a, a), ('flrp', a, b, a), '!options->lower_flrp32'),
1981 (('~ffma@32', a, 2.0, ('fneg', ('fmul', a, a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1982 (('~ffma@32', a, 2.0, ('fmul', ('fneg', a), a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1983 (('~ffma@32', a, ('fneg', a), ('fmul', 2.0, a)), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1984 (('~fmul@32', a, ('fadd', 2.0, ('fneg', a))), ('flrp', a, 1.0, a), '!options->lower_flrp32'),
1985
1986 # we do these late so that we don't get in the way of creating ffmas
1987 (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
1988 (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),
1989
1990 (('bcsel', a, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a, b)))),
1991
1992 # Putting this in 'optimizations' interferes with the bcsel(a, op(b, c),
1993 # op(b, d)) => op(b, bcsel(a, c, d)) transformations. I do not know why.
1994 (('bcsel', ('feq', ('fsqrt', 'a(is_not_negative)'), 0.0), intBitsToFloat(0x7f7fffff), ('frsq', a)),
1995 ('fmin', ('frsq', a), intBitsToFloat(0x7f7fffff))),
1996
1997 # Things that look like DPH in the source shader may get expanded to
1998 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1999 # to NIR. After FFMA is generated, this can look like:
2000 #
2001 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
2002 #
2003 # Reassociate the last addition into the first multiplication.
2004 #
2005 # Some shaders do not use 'invariant' in vertex and (possibly) geometry
2006 # shader stages on some outputs that are intended to be invariant. For
2007 # various reasons, this optimization may not be fully applied in all
2008 # shaders used for different rendering passes of the same geometry. This
2009 # can result in Z-fighting artifacts (at best). For now, disable this
2010 # optimization in these stages. See bugzilla #111490. In tessellation
2011 # stages applications seem to use 'precise' when necessary, so allow the
2012 # optimization in those stages.
2013 (('~fadd', ('ffma(is_used_once)', a, b, ('ffma', c, d, ('fmul', 'e(is_not_const_and_not_fsign)', 'f(is_not_const_and_not_fsign)'))), 'g(is_not_const)'),
2014 ('ffma', a, b, ('ffma', c, d, ('ffma', e, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
2015 (('~fadd', ('