nir/algebraic: trivially enable existing 32-bit patterns for all bit sizes
[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(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'),
94 (('imul', a, '#b(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'),
95 (('ishl', a, '#b'), ('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(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b)), '!options->lower_bitops'),
106 (('idiv', a, '#b(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), 'options->lower_idiv'),
107 (('idiv', a, '#b(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', ('b2i', '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', ('ushr', a, log2), log2), ('iand', a, mask)),
270
271 # Reassociate for improved CSE
272 (('iand', ('iadd', a, b_is_multiple), mask), ('iadd', ('iand', a, mask), b)),
273 ])
274
275 # To save space in the state tables, reduce to the set that is known to help.
276 # Previously, this was range(1, 32). In addition, a couple rules inside the
277 # loop are commented out. Revisit someday, probably after mesa/#2635 has some
278 # resolution.
279 for i in [1, 2, 16, 24]:
280 lo_mask = 0xffffffff >> i
281 hi_mask = (0xffffffff << i) & 0xffffffff
282
283 optimizations.extend([
284 # This pattern seems to only help in the soft-fp64 code.
285 (('ishl@32', ('iand', 'a@32', lo_mask), i), ('ishl', a, i)),
286 # (('ushr@32', ('iand', 'a@32', hi_mask), i), ('ushr', a, i)),
287 # (('ishr@32', ('iand', 'a@32', hi_mask), i), ('ishr', a, i)),
288
289 (('iand', ('ishl', 'a@32', i), hi_mask), ('ishl', a, i)),
290 (('iand', ('ushr', 'a@32', i), lo_mask), ('ushr', a, i)),
291 # (('iand', ('ishr', 'a@32', i), lo_mask), ('ushr', a, i)), # Yes, ushr is correct
292 ])
293
294 optimizations.extend([
295 # This is common for address calculations. Reassociating may enable the
296 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
297 # instruction or a constant offset field for in load / store instructions.
298 (('ishl', ('iadd', a, '#b'), '#c'), ('iadd', ('ishl', a, c), ('ishl', b, c))),
299
300 # Comparison simplifications
301 (('~inot', ('flt', a, b)), ('fge', a, b)),
302 (('~inot', ('fge', a, b)), ('flt', a, b)),
303 (('inot', ('feq', a, b)), ('fneu', a, b)),
304 (('inot', ('fneu', a, b)), ('feq', a, b)),
305 (('inot', ('ilt', a, b)), ('ige', a, b)),
306 (('inot', ('ult', a, b)), ('uge', a, b)),
307 (('inot', ('ige', a, b)), ('ilt', a, b)),
308 (('inot', ('uge', a, b)), ('ult', a, b)),
309 (('inot', ('ieq', a, b)), ('ine', a, b)),
310 (('inot', ('ine', a, b)), ('ieq', a, b)),
311
312 (('iand', ('feq', a, b), ('fneu', a, b)), False),
313 (('iand', ('flt', a, b), ('flt', b, a)), False),
314 (('iand', ('ieq', a, b), ('ine', a, b)), False),
315 (('iand', ('ilt', a, b), ('ilt', b, a)), False),
316 (('iand', ('ult', a, b), ('ult', b, a)), False),
317
318 # This helps some shaders because, after some optimizations, they end up
319 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
320 # matching would be handled by CSE.
321 (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
322 (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
323 (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
324 (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)),
325 (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
326 (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
327 (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
328 (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
329 (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)),
330 (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),
331
332 # flt(fsat(a), b > 0 && b < 1) is inexact if a is NaN (fsat(NaN) is 0)
333 # because it returns True while flt(a, b) always returns False.
334 (('~flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
335 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
336 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
337 # fge(b > 0 && b < 1, fsat(a)) is inexact if a is NaN (fsat(NaN) is 0)
338 # because it returns True while fge(b, a) always returns False.
339 (('~fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
340 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
341 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),
342
343 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
344 (('flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
345 (('fge', 0.0, ('fsat(is_used_once)', a)), ('fge', 0.0, a)),
346 (('flt', 0.0, ('fsat(is_used_once)', a)), ('flt', 0.0, a)),
347
348 # 0.0 >= b2f(a)
349 # b2f(a) <= 0.0
350 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
351 # inot(a)
352 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a)),
353
354 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a)),
355
356 (('fneu', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
357 (('fneu', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
358 (('fneu', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a, b)),
359 (('fneu', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a, b)),
360 (('fneu', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
361 (('fneu', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
362 (('fneu', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a, b)),
363 (('fneu', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a, b)),
364 (('fneu', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a, b)),
365 (('fneu', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a, b)),
366 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
367 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
368 (('feq', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a, b))),
369 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
370 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
371 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
372 (('feq', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a, b))),
373 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a, b)),
374 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a, b)),
375 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a, b)),
376
377 # -(b2f(a) + b2f(b)) < 0
378 # 0 < b2f(a) + b2f(b)
379 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
380 # a || b
381 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a, b)),
382 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a, b)),
383
384 # -(b2f(a) + b2f(b)) >= 0
385 # 0 >= b2f(a) + b2f(b)
386 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
387 # !(a || b)
388 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))),
389 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
390
391 (('flt', a, ('fneg', a)), ('flt', a, 0.0)),
392 (('fge', a, ('fneg', a)), ('fge', a, 0.0)),
393
394 # Some optimizations (below) convert things like (a < b || c < b) into
395 # (min(a, c) < b). However, this interfers with the previous optimizations
396 # that try to remove comparisons with negated sums of b2f. This just
397 # breaks that apart.
398 (('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
399 ('ior', ('flt', c, 0.0), ('ior', a, b))),
400
401 (('~flt', ('fadd', a, b), a), ('flt', b, 0.0)),
402 (('~fge', ('fadd', a, b), a), ('fge', b, 0.0)),
403 (('~feq', ('fadd', a, b), a), ('feq', b, 0.0)),
404 (('~fneu', ('fadd', a, b), a), ('fneu', b, 0.0)),
405 (('~flt', ('fadd(is_used_once)', a, '#b'), '#c'), ('flt', a, ('fadd', c, ('fneg', b)))),
406 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('flt', ('fneg', ('fadd', c, b)), a)),
407 (('~fge', ('fadd(is_used_once)', a, '#b'), '#c'), ('fge', a, ('fadd', c, ('fneg', b)))),
408 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fge', ('fneg', ('fadd', c, b)), a)),
409 (('~feq', ('fadd(is_used_once)', a, '#b'), '#c'), ('feq', a, ('fadd', c, ('fneg', b)))),
410 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('feq', ('fneg', ('fadd', c, b)), a)),
411 (('~fneu', ('fadd(is_used_once)', a, '#b'), '#c'), ('fneu', a, ('fadd', c, ('fneg', b)))),
412 (('~fneu', ('fneg(is_used_once)', ('fadd(is_used_once)', a, '#b')), '#c'), ('fneu', ('fneg', ('fadd', c, b)), a)),
413
414 # Cannot remove the addition from ilt or ige due to overflow.
415 (('ieq', ('iadd', a, b), a), ('ieq', b, 0)),
416 (('ine', ('iadd', a, b), a), ('ine', b, 0)),
417
418 # fmin(-b2f(a), b) >= 0.0
419 # -b2f(a) >= 0.0 && b >= 0.0
420 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
421 # b2f(a) == 0.0 && b >= 0.0
422 # a == False && b >= 0.0
423 # !a && b >= 0.0
424 #
425 # The fge in the second replacement is not a typo. I leave the proof that
426 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
427 # reader.
428 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
429 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
430
431 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a)),
432 (('~fneu', ('b2f', 'a@1'), 0.0), a),
433 (('ieq', ('b2i', 'a@1'), 0), ('inot', a)),
434 (('ine', ('b2i', 'a@1'), 0), a),
435
436 (('fneu', ('u2f', a), 0.0), ('ine', a, 0)),
437 (('feq', ('u2f', a), 0.0), ('ieq', a, 0)),
438 (('fge', ('u2f', a), 0.0), True),
439 (('fge', 0.0, ('u2f', a)), ('uge', 0, a)), # ieq instead?
440 (('flt', ('u2f', a), 0.0), False),
441 (('flt', 0.0, ('u2f', a)), ('ult', 0, a)), # ine instead?
442 (('fneu', ('i2f', a), 0.0), ('ine', a, 0)),
443 (('feq', ('i2f', a), 0.0), ('ieq', a, 0)),
444 (('fge', ('i2f', a), 0.0), ('ige', a, 0)),
445 (('fge', 0.0, ('i2f', a)), ('ige', 0, a)),
446 (('flt', ('i2f', a), 0.0), ('ilt', a, 0)),
447 (('flt', 0.0, ('i2f', a)), ('ilt', 0, a)),
448
449 # 0.0 < fabs(a)
450 # fabs(a) > 0.0
451 # fabs(a) != 0.0 because fabs(a) must be >= 0
452 # a != 0.0
453 (('~flt', 0.0, ('fabs', a)), ('fneu', a, 0.0)),
454
455 # -fabs(a) < 0.0
456 # fabs(a) > 0.0
457 (('~flt', ('fneg', ('fabs', a)), 0.0), ('fneu', a, 0.0)),
458
459 # 0.0 >= fabs(a)
460 # 0.0 == fabs(a) because fabs(a) must be >= 0
461 # 0.0 == a
462 (('fge', 0.0, ('fabs', a)), ('feq', a, 0.0)),
463
464 # -fabs(a) >= 0.0
465 # 0.0 >= fabs(a)
466 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
467
468 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
469 (('iand', ('fge', a, 0.0), ('fge', 1.0, a)), ('feq', a, ('fsat', a)), '!options->lower_fsat'),
470
471 # (a < 0.0) || (a > 1.0)
472 # !(!(a < 0.0) && !(a > 1.0))
473 # !((a >= 0.0) && (a <= 1.0))
474 # !(a == fsat(a))
475 # a != fsat(a)
476 (('ior', ('flt', a, 0.0), ('flt', 1.0, a)), ('fneu', a, ('fsat', a)), '!options->lower_fsat'),
477
478 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a, b))),
479 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a, b)))),
480 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
481 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a, b)))),
482
483 # fmin(b2f(a), b)
484 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
485 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
486 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
487 #
488 # Since b is a constant, constant folding will eliminate the fmin and the
489 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
490 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))),
491
492 (('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),
493
494 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
495 (('~bcsel', ('flt', b, a), b, a), ('fmin', a, b)),
496 (('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
497 (('~bcsel', ('fge', a, b), b, a), ('fmin', a, b)),
498 (('~bcsel', ('fge', b, a), b, a), ('fmax', a, b)),
499 (('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)),
500 (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
501 (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
502 (('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)),
503 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
504 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
505 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
506 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
507 (('bcsel', a, True, b), ('ior', a, b)),
508 (('bcsel', a, a, b), ('ior', a, b)),
509 (('bcsel', a, b, False), ('iand', a, b)),
510 (('bcsel', a, b, a), ('iand', a, b)),
511 (('~fmin', a, a), a),
512 (('~fmax', a, a), a),
513 (('imin', a, a), a),
514 (('imax', a, a), a),
515 (('umin', a, a), a),
516 (('umax', a, a), a),
517 (('fmax', ('fmax', a, b), b), ('fmax', a, b)),
518 (('umax', ('umax', a, b), b), ('umax', a, b)),
519 (('imax', ('imax', a, b), b), ('imax', a, b)),
520 (('fmin', ('fmin', a, b), b), ('fmin', a, b)),
521 (('umin', ('umin', a, b), b), ('umin', a, b)),
522 (('imin', ('imin', a, b), b), ('imin', a, b)),
523 (('iand@32', a, ('inot', ('ishr', a, 31))), ('imax', a, 0)),
524
525 # Simplify logic to detect sign of an integer.
526 (('ieq', ('iand', 'a@32', 0x80000000), 0x00000000), ('ige', a, 0)),
527 (('ine', ('iand', 'a@32', 0x80000000), 0x80000000), ('ige', a, 0)),
528 (('ine', ('iand', 'a@32', 0x80000000), 0x00000000), ('ilt', a, 0)),
529 (('ieq', ('iand', 'a@32', 0x80000000), 0x80000000), ('ilt', a, 0)),
530 (('ine', ('ushr', 'a@32', 31), 0), ('ilt', a, 0)),
531 (('ieq', ('ushr', 'a@32', 31), 0), ('ige', a, 0)),
532 (('ieq', ('ushr', 'a@32', 31), 1), ('ilt', a, 0)),
533 (('ine', ('ushr', 'a@32', 31), 1), ('ige', a, 0)),
534 (('ine', ('ishr', 'a@32', 31), 0), ('ilt', a, 0)),
535 (('ieq', ('ishr', 'a@32', 31), 0), ('ige', a, 0)),
536 (('ieq', ('ishr', 'a@32', 31), -1), ('ilt', a, 0)),
537 (('ine', ('ishr', 'a@32', 31), -1), ('ige', a, 0)),
538
539 (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
540 (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
541 (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
542 (('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))),
543 (('~fmin', a, ('fabs', a)), a),
544 (('imin', a, ('iabs', a)), a),
545 (('~fmax', a, ('fneg', ('fabs', a))), a),
546 (('imax', a, ('ineg', ('iabs', a))), a),
547 (('fmax', a, ('fabs', a)), ('fabs', a)),
548 (('imax', a, ('iabs', a)), ('iabs', a)),
549 (('fmax', a, ('fneg', a)), ('fabs', a)),
550 (('imax', a, ('ineg', a)), ('iabs', a), '!options->lower_iabs'),
551 (('~fmax', ('fabs', a), 0.0), ('fabs', a)),
552 (('fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'),
553 # fmax(fmin(a, 1.0), 0.0) is inexact because it returns 1.0 on NaN, while
554 # fsat(a) returns 0.0.
555 (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'),
556 # fmin(fmax(a, -1.0), 0.0) is inexact because it returns -1.0 on NaN, while
557 # fneg(fsat(fneg(a))) returns -0.0 on NaN.
558 (('~fmin', ('fmax', a, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
559 # fmax(fmin(a, 0.0), -1.0) is inexact because it returns 0.0 on NaN, while
560 # fneg(fsat(fneg(a))) returns -0.0 on NaN. This only matters if
561 # SignedZeroInfNanPreserve is set, but we don't currently have any way of
562 # representing this in the optimizations other than the usual ~.
563 (('~fmax', ('fmin', a, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a))), '!options->lower_fsat'),
564 (('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))),
565 (('fsat', ('b2f', a)), ('b2f', a)),
566 (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
567 (('fsat', ('fsat', a)), ('fsat', a)),
568 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a, b))), ('fsat', ('fadd', ('fneg', a), ('fneg', b))), '!options->lower_fsat'),
569 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fneg', a), b)), '!options->lower_fsat'),
570 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a, b))), ('fsat', ('fmul', ('fabs', a), ('fabs', b))), '!options->lower_fsat'),
571 (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)),
572 (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)),
573 (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
574 # Both the left and right patterns are "b" when isnan(a), so this is exact.
575 (('fmax', ('fsat', a), '#b(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
576 # The left pattern is 0.0 when isnan(a) (because fmin(fsat(NaN), b) ->
577 # fmin(0.0, b)) while the right one is "b", so this optimization is inexact.
578 (('~fmin', ('fsat', a), '#b(is_zero_to_one)'), ('fsat', ('fmin', a, b))),
579
580 # If a in [0,b] then b-a is also in [0,b]. Since b in [0,1], max(b-a, 0) =
581 # fsat(b-a).
582 #
583 # If a > b, then b-a < 0 and max(b-a, 0) = fsat(b-a) = 0
584 #
585 # This should be NaN safe since max(NaN, 0) = fsat(NaN) = 0.
586 (('fmax', ('fadd(is_used_once)', ('fneg', 'a(is_not_negative)'), '#b(is_zero_to_one)'), 0.0),
587 ('fsat', ('fadd', ('fneg', a), b)), '!options->lower_fsat'),
588
589 (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
590 (('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
591 (('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
592 (('~ior', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
593 (('~ior', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
594 (('~ior', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmax', b, c))),
595 (('~ior', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmin', a, b), c)),
596 (('~ior', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmin', b, c))),
597 (('~ior', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmax', a, b), c)),
598 (('~iand', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmin', b, c))),
599 (('~iand', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmax', a, b), c)),
600 (('~iand', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmax', b, c))),
601 (('~iand', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmin', a, b), c)),
602 (('~iand', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmin', b, c))),
603 (('~iand', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmax', a, b), c)),
604 (('~iand', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmax', b, c))),
605 (('~iand', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmin', a, b), c)),
606
607 (('ior', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imax', b, c))),
608 (('ior', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imin', a, b), c)),
609 (('ior', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imin', b, c))),
610 (('ior', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imax', a, b), c)),
611 (('ior', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umax', b, c))),
612 (('ior', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umin', a, b), c)),
613 (('ior', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umin', b, c))),
614 (('ior', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umax', a, b), c)),
615 (('iand', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imin', b, c))),
616 (('iand', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imax', a, b), c)),
617 (('iand', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imax', b, c))),
618 (('iand', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imin', a, b), c)),
619 (('iand', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umin', b, c))),
620 (('iand', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umax', a, b), c)),
621 (('iand', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umax', b, c))),
622 (('iand', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umin', a, b), c)),
623
624 # These derive from the previous patterns with the application of b < 0 <=>
625 # 0 < -b. The transformation should be applied if either comparison is
626 # used once as this ensures that the number of comparisons will not
627 # increase. The sources to the ior and iand are not symmetric, so the
628 # rules have to be duplicated to get this behavior.
629 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
630 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a, ('fneg', b)))),
631 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
632 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a, ('fneg', b)))),
633 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
634 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a, ('fneg', b)))),
635 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
636 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a, ('fneg', b)))),
637
638 # Common pattern like 'if (i == 0 || i == 1 || ...)'
639 (('ior', ('ieq', a, 0), ('ieq', a, 1)), ('uge', 1, a)),
640 (('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)),
641 (('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)),
642
643 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
644 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
645 # so emit an open-coded version of that.
646 (('bcsel@32', ('feq', a, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
647 ('ior', 0x3f800000, ('iand', a, 0x80000000))),
648
649 (('ior', a, ('ieq', a, False)), True),
650 (('ior', a, ('inot', a)), -1),
651
652 (('ine', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))), ('ine', a, b)),
653 (('b2i', ('ine', 'a@1', 'b@1')), ('b2i', ('ixor', a, b))),
654
655 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umax', a, b), 0)),
656 (('ior', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umin', a, b), 0)),
657 (('iand', ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umin', a, b), 0)),
658 (('ior', ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umax', a, b), 0)),
659
660 # This pattern occurs coutresy of __flt64_nonnan in the soft-fp64 code.
661 # The first part of the iand comes from the !__feq64_nonnan.
662 #
663 # The second pattern is a reformulation of the first based on the relation
664 # (a == 0 || y == 0) <=> umin(a, y) == 0, where b in the first equation
665 # happens to be y == 0.
666 (('iand', ('inot', ('iand', ('ior', ('ieq', a, 0), b), c)), ('ilt', a, 0)),
667 ('iand', ('inot', ('iand', b , c)), ('ilt', a, 0))),
668 (('iand', ('inot', ('iand', ('ieq', ('umin', a, b), 0), c)), ('ilt', a, 0)),
669 ('iand', ('inot', ('iand', ('ieq', b , 0), c)), ('ilt', a, 0))),
670
671 # These patterns can result when (a < b || a < c) => (a < min(b, c))
672 # transformations occur before constant propagation and loop-unrolling.
673 (('~flt', a, ('fmax', b, a)), ('flt', a, b)),
674 (('~flt', ('fmin', a, b), a), ('flt', b, a)),
675 (('~fge', a, ('fmin', b, a)), True),
676 (('~fge', ('fmax', a, b), a), True),
677 (('~flt', a, ('fmin', b, a)), False),
678 (('~flt', ('fmax', a, b), a), False),
679 (('~fge', a, ('fmax', b, a)), ('fge', a, b)),
680 (('~fge', ('fmin', a, b), a), ('fge', b, a)),
681
682 (('ilt', a, ('imax', b, a)), ('ilt', a, b)),
683 (('ilt', ('imin', a, b), a), ('ilt', b, a)),
684 (('ige', a, ('imin', b, a)), True),
685 (('ige', ('imax', a, b), a), True),
686 (('ult', a, ('umax', b, a)), ('ult', a, b)),
687 (('ult', ('umin', a, b), a), ('ult', b, a)),
688 (('uge', a, ('umin', b, a)), True),
689 (('uge', ('umax', a, b), a), True),
690 (('ilt', a, ('imin', b, a)), False),
691 (('ilt', ('imax', a, b), a), False),
692 (('ige', a, ('imax', b, a)), ('ige', a, b)),
693 (('ige', ('imin', a, b), a), ('ige', b, a)),
694 (('ult', a, ('umin', b, a)), False),
695 (('ult', ('umax', a, b), a), False),
696 (('uge', a, ('umax', b, a)), ('uge', a, b)),
697 (('uge', ('umin', a, b), a), ('uge', b, a)),
698 (('ult', a, ('iand', b, a)), False),
699 (('ult', ('ior', a, b), a), False),
700 (('uge', a, ('iand', b, a)), True),
701 (('uge', ('ior', a, b), a), True),
702
703 (('ilt', '#a', ('imax', '#b', c)), ('ior', ('ilt', a, b), ('ilt', a, c))),
704 (('ilt', ('imin', '#a', b), '#c'), ('ior', ('ilt', a, c), ('ilt', b, c))),
705 (('ige', '#a', ('imin', '#b', c)), ('ior', ('ige', a, b), ('ige', a, c))),
706 (('ige', ('imax', '#a', b), '#c'), ('ior', ('ige', a, c), ('ige', b, c))),
707 (('ult', '#a', ('umax', '#b', c)), ('ior', ('ult', a, b), ('ult', a, c))),
708 (('ult', ('umin', '#a', b), '#c'), ('ior', ('ult', a, c), ('ult', b, c))),
709 (('uge', '#a', ('umin', '#b', c)), ('ior', ('uge', a, b), ('uge', a, c))),
710 (('uge', ('umax', '#a', b), '#c'), ('ior', ('uge', a, c), ('uge', b, c))),
711 (('ilt', '#a', ('imin', '#b', c)), ('iand', ('ilt', a, b), ('ilt', a, c))),
712 (('ilt', ('imax', '#a', b), '#c'), ('iand', ('ilt', a, c), ('ilt', b, c))),
713 (('ige', '#a', ('imax', '#b', c)), ('iand', ('ige', a, b), ('ige', a, c))),
714 (('ige', ('imin', '#a', b), '#c'), ('iand', ('ige', a, c), ('ige', b, c))),
715 (('ult', '#a', ('umin', '#b', c)), ('iand', ('ult', a, b), ('ult', a, c))),
716 (('ult', ('umax', '#a', b), '#c'), ('iand', ('ult', a, c), ('ult', b, c))),
717 (('uge', '#a', ('umax', '#b', c)), ('iand', ('uge', a, b), ('uge', a, c))),
718 (('uge', ('umin', '#a', b), '#c'), ('iand', ('uge', a, c), ('uge', b, c))),
719
720 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
721 # negative.
722 (('bcsel', ('ilt', a, 0), ('ineg', ('ishr', a, b)), ('ishr', a, b)),
723 ('iabs', ('ishr', a, b))),
724 (('iabs', ('ishr', ('iabs', a), b)), ('ishr', ('iabs', a), b)),
725
726 (('fabs', ('slt', a, b)), ('slt', a, b)),
727 (('fabs', ('sge', a, b)), ('sge', a, b)),
728 (('fabs', ('seq', a, b)), ('seq', a, b)),
729 (('fabs', ('sne', a, b)), ('sne', a, b)),
730 (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
731 (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
732 (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
733 (('sne', a, b), ('b2f', ('fneu', a, b)), 'options->lower_scmp'),
734 (('seq', ('seq', a, b), 1.0), ('seq', a, b)),
735 (('seq', ('sne', a, b), 1.0), ('sne', a, b)),
736 (('seq', ('slt', a, b), 1.0), ('slt', a, b)),
737 (('seq', ('sge', a, b), 1.0), ('sge', a, b)),
738 (('sne', ('seq', a, b), 0.0), ('seq', a, b)),
739 (('sne', ('sne', a, b), 0.0), ('sne', a, b)),
740 (('sne', ('slt', a, b), 0.0), ('slt', a, b)),
741 (('sne', ('sge', a, b), 0.0), ('sge', a, b)),
742 (('seq', ('seq', a, b), 0.0), ('sne', a, b)),
743 (('seq', ('sne', a, b), 0.0), ('seq', a, b)),
744 (('seq', ('slt', a, b), 0.0), ('sge', a, b)),
745 (('seq', ('sge', a, b), 0.0), ('slt', a, b)),
746 (('sne', ('seq', a, b), 1.0), ('sne', a, b)),
747 (('sne', ('sne', a, b), 1.0), ('seq', a, b)),
748 (('sne', ('slt', a, b), 1.0), ('sge', a, b)),
749 (('sne', ('sge', a, b), 1.0), ('slt', a, b)),
750 (('fall_equal2', a, b), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
751 (('fall_equal3', a, b), ('seq', ('fany_nequal3', a, b), 0.0), 'options->lower_vector_cmp'),
752 (('fall_equal4', a, b), ('seq', ('fany_nequal4', a, b), 0.0), 'options->lower_vector_cmp'),
753 (('fany_nequal2', a, b), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
754 (('fany_nequal3', a, b), ('fsat', ('fdot3', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),
755 (('fany_nequal4', a, b), ('fsat', ('fdot4', ('sne', a, b), ('sne', a, b))), 'options->lower_vector_cmp'),
756
757 (('ball_iequal2', a, b), ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
758 (('ball_iequal3', a, b), ('iand', ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), ('ieq', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
759 (('ball_iequal4', a, b), ('iand', ('iand', ('ieq', 'a.x', 'b.x'), ('ieq', 'a.y', 'b.y')), ('iand', ('ieq', 'a.z', 'b.z'), ('ieq', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),
760
761 (('bany_inequal2', a, b), ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
762 (('bany_inequal3', a, b), ('ior', ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), ('ine', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
763 (('bany_inequal4', a, b), ('ior', ('ior', ('ine', 'a.x', 'b.x'), ('ine', 'a.y', 'b.y')), ('ior', ('ine', 'a.z', 'b.z'), ('ine', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),
764
765 (('ball_fequal2', a, b), ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
766 (('ball_fequal3', a, b), ('iand', ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), ('feq', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
767 (('ball_fequal4', a, b), ('iand', ('iand', ('feq', 'a.x', 'b.x'), ('feq', 'a.y', 'b.y')), ('iand', ('feq', 'a.z', 'b.z'), ('feq', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),
768
769 (('bany_fnequal2', a, b), ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
770 (('bany_fnequal3', a, b), ('ior', ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), ('fneu', 'a.z', 'b.z')), 'options->lower_vector_cmp'),
771 (('bany_fnequal4', a, b), ('ior', ('ior', ('fneu', 'a.x', 'b.x'), ('fneu', 'a.y', 'b.y')), ('ior', ('fneu', 'a.z', 'b.z'), ('fneu', 'a.w', 'b.w'))), 'options->lower_vector_cmp'),
772
773 (('fneu', ('fneg', a), a), ('fneu', a, 0.0)),
774 (('feq', ('fneg', a), a), ('feq', a, 0.0)),
775 # Emulating booleans
776 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
777 (('iand', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
778 (('ior', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('ior', a, b))),
779 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
780 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a, b))),
781 (('iand', 'a@bool32', 1.0), ('b2f', a)),
782 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
783 (('ineg', ('b2i32', 'a@32')), a),
784 (('flt', ('fneg', ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
785 # Comparison with the same args. Note that these are not done for
786 # the float versions because NaN always returns false on float
787 # inequalities.
788 (('ilt', a, a), False),
789 (('ige', a, a), True),
790 (('ieq', a, a), True),
791 (('ine', a, a), False),
792 (('ult', a, a), False),
793 (('uge', a, a), True),
794 # Logical and bit operations
795 (('iand', a, a), a),
796 (('iand', a, ~0), a),
797 (('iand', a, 0), 0),
798 (('ior', a, a), a),
799 (('ior', a, 0), a),
800 (('ior', a, True), True),
801 (('ixor', a, a), 0),
802 (('ixor', a, 0), a),
803 (('inot', ('inot', a)), a),
804 (('ior', ('iand', a, b), b), b),
805 (('ior', ('ior', a, b), b), ('ior', a, b)),
806 (('iand', ('ior', a, b), b), b),
807 (('iand', ('iand', a, b), b), ('iand', a, b)),
808 # DeMorgan's Laws
809 (('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))),
810 (('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))),
811 # Shift optimizations
812 (('ishl', 0, a), 0),
813 (('ishl', a, 0), a),
814 (('ishr', 0, a), 0),
815 (('ishr', a, 0), a),
816 (('ushr', 0, a), 0),
817 (('ushr', a, 0), a),
818 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('iadd', 16, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
819 (('ior', ('ishl@16', a, b), ('ushr@16', a, ('isub', 16, b))), ('urol', a, b), '!options->lower_rotate'),
820 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('iadd', 32, ('ineg', b)))), ('urol', a, b), '!options->lower_rotate'),
821 (('ior', ('ishl@32', a, b), ('ushr@32', a, ('isub', 32, b))), ('urol', a, b), '!options->lower_rotate'),
822 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('iadd', 16, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
823 (('ior', ('ushr@16', a, b), ('ishl@16', a, ('isub', 16, b))), ('uror', a, b), '!options->lower_rotate'),
824 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('iadd', 32, ('ineg', b)))), ('uror', a, b), '!options->lower_rotate'),
825 (('ior', ('ushr@32', a, b), ('ishl@32', a, ('isub', 32, b))), ('uror', a, b), '!options->lower_rotate'),
826 (('urol@16', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 16, b))), 'options->lower_rotate'),
827 (('urol@32', a, b), ('ior', ('ishl', a, b), ('ushr', a, ('isub', 32, b))), 'options->lower_rotate'),
828 (('uror@16', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 16, b))), 'options->lower_rotate'),
829 (('uror@32', a, b), ('ior', ('ushr', a, b), ('ishl', a, ('isub', 32, b))), 'options->lower_rotate'),
830 # Exponential/logarithmic identities
831 (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
832 (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
833 (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
834 (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
835 (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
836 ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
837 (('~fexp2', ('fmul', ('flog2', a), 0.5)), ('fsqrt', a)),
838 (('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)),
839 (('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))),
840 (('~fpow', a, 1.0), a),
841 (('~fpow', a, 2.0), ('fmul', a, a)),
842 (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
843 (('~fpow', 2.0, a), ('fexp2', a)),
844 (('~fpow', ('fpow', a, 2.2), 0.454545), a),
845 (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
846 (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
847 (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
848 (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
849 (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
850 (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
851 (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
852 (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
853 (('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))),
854 (('bcsel', ('flt', a, 0.0), 0.0, ('fsqrt', a)), ('fsqrt', ('fmax', a, 0.0))),
855 (('~fmul', ('fsqrt', a), ('fsqrt', a)), ('fabs',a)),
856 # Division and reciprocal
857 (('~fdiv', 1.0, a), ('frcp', a)),
858 (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
859 (('~frcp', ('frcp', a)), a),
860 (('~frcp', ('fsqrt', a)), ('frsq', a)),
861 (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
862 (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
863 # Trig
864 (('fsin', a), lowered_sincos(0.5), 'options->lower_sincos'),
865 (('fcos', a), lowered_sincos(0.75), 'options->lower_sincos'),
866 # Boolean simplifications
867 (('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
868 (('i2b1(is_used_by_if)', a), ('ine', a, 0)),
869 (('ieq', a, True), a),
870 (('ine(is_not_used_by_if)', a, True), ('inot', a)),
871 (('ine', a, False), a),
872 (('ieq(is_not_used_by_if)', a, False), ('inot', 'a')),
873 (('bcsel', a, True, False), a),
874 (('bcsel', a, False, True), ('inot', a)),
875 (('bcsel', a, 1.0, 0.0), ('b2f', a)),
876 (('bcsel', a, 0.0, 1.0), ('b2f', ('inot', a))),
877 (('bcsel', a, -1.0, -0.0), ('fneg', ('b2f', a))),
878 (('bcsel', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))),
879 (('bcsel', True, b, c), b),
880 (('bcsel', False, b, c), c),
881 (('bcsel', a, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a, b, c))),
882
883 (('bcsel', a, b, b), b),
884 (('~fcsel', a, b, b), b),
885
886 # D3D Boolean emulation
887 (('bcsel', a, -1, 0), ('ineg', ('b2i', 'a@1'))),
888 (('bcsel', a, 0, -1), ('ineg', ('b2i', ('inot', a)))),
889 (('bcsel', a, 1, 0), ('b2i', 'a@1')),
890 (('bcsel', a, 0, 1), ('b2i', ('inot', a))),
891 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
892 ('ineg', ('b2i', ('iand', a, b)))),
893 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
894 ('ineg', ('b2i', ('ior', a, b)))),
895 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a)),
896 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a),
897 (('ine', ('ineg', ('b2i', 'a@1')), 0), a),
898 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a)),
899 (('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)),
900 (('iand', ('ineg', ('b2i', a)), 1), ('b2i', a)),
901
902 # SM5 32-bit shifts are defined to use the 5 least significant bits
903 (('ishl', 'a@32', ('iand', 31, b)), ('ishl', a, b)),
904 (('ishr', 'a@32', ('iand', 31, b)), ('ishr', a, b)),
905 (('ushr', 'a@32', ('iand', 31, b)), ('ushr', a, b)),
906
907 # Conversions
908 (('i2b32', ('b2i', 'a@32')), a),
909 (('f2i', ('ftrunc', a)), ('f2i', a)),
910 (('f2u', ('ftrunc', a)), ('f2u', a)),
911 (('i2b', ('ineg', a)), ('i2b', a)),
912 (('i2b', ('iabs', a)), ('i2b', a)),
913 (('inot', ('f2b1', a)), ('feq', a, 0.0)),
914
915 # The C spec says, "If the value of the integral part cannot be represented
916 # by the integer type, the behavior is undefined." "Undefined" can mean
917 # "the conversion doesn't happen at all."
918 (('~i2f32', ('f2i32', 'a@32')), ('ftrunc', a)),
919
920 # Ironically, mark these as imprecise because removing the conversions may
921 # preserve more precision than doing the conversions (e.g.,
922 # uint(float(0x81818181u)) == 0x81818200).
923 (('~f2i32', ('i2f', 'a@32')), a),
924 (('~f2i32', ('u2f', 'a@32')), a),
925 (('~f2u32', ('i2f', 'a@32')), a),
926 (('~f2u32', ('u2f', 'a@32')), a),
927
928 # Conversions from 16 bits to 32 bits and back can always be removed
929 (('f2f16', ('f2f32', 'a@16')), a),
930 (('f2fmp', ('f2f32', 'a@16')), a),
931 (('i2i16', ('i2i32', 'a@16')), a),
932 (('i2imp', ('i2i32', 'a@16')), a),
933 (('u2u16', ('u2u32', 'a@16')), a),
934 (('u2ump', ('u2u32', 'a@16')), a),
935 (('f2f16', ('b2f32', 'a@1')), ('b2f16', a)),
936 (('f2fmp', ('b2f32', 'a@1')), ('b2f16', a)),
937 (('i2i16', ('b2i32', 'a@1')), ('b2i16', a)),
938 (('i2imp', ('b2i32', 'a@1')), ('b2i16', a)),
939 (('u2u16', ('b2i32', 'a@1')), ('b2i16', a)),
940 (('u2ump', ('b2i32', 'a@1')), ('b2i16', a)),
941 # Conversions to 16 bits would be lossy so they should only be removed if
942 # the instruction was generated by the precision lowering pass.
943 (('f2f32', ('f2fmp', 'a@32')), a),
944 (('i2i32', ('i2imp', 'a@32')), a),
945 (('u2u32', ('u2ump', 'a@32')), a),
946
947 (('ffloor', 'a(is_integral)'), a),
948 (('fceil', 'a(is_integral)'), a),
949 (('ftrunc', 'a(is_integral)'), a),
950 # fract(x) = x - floor(x), so fract(NaN) = NaN
951 (('~ffract', 'a(is_integral)'), 0.0),
952 (('fabs', 'a(is_not_negative)'), a),
953 (('iabs', 'a(is_not_negative)'), a),
954 (('fsat', 'a(is_not_positive)'), 0.0),
955
956 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
957 # says:
958 #
959 # It is undefined to convert a negative floating-point value to an
960 # uint.
961 #
962 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
963 # some optimizations in the i965 backend to proceed.
964 (('ige', ('f2u', a), b), ('ige', ('f2i', a), b)),
965 (('ige', b, ('f2u', a)), ('ige', b, ('f2i', a))),
966 (('ilt', ('f2u', a), b), ('ilt', ('f2i', a), b)),
967 (('ilt', b, ('f2u', a)), ('ilt', b, ('f2i', a))),
968
969 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a), '!options->lower_fsat'),
970
971 # The result of the multiply must be in [-1, 0], so the result of the ffma
972 # must be in [0, 1].
973 (('flt', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), False),
974 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), False),
975 (('fmax', ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a), ('fneg', ('fsat', a))), 1.0)),
976 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a), ('fsat', a))), 1.0)),
977
978 (('fneu', 'a(is_not_zero)', 0.0), True),
979 (('feq', 'a(is_not_zero)', 0.0), False),
980
981 # In this chart, + means value > 0 and - means value < 0.
982 #
983 # + >= + -> unknown 0 >= + -> false - >= + -> false
984 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
985 # + >= - -> true 0 >= - -> true - >= - -> unknown
986 #
987 # Using grouping conceptually similar to a Karnaugh map...
988 #
989 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
990 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
991 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
992 #
993 # The flt / ilt cases just invert the expected result.
994 #
995 # The results expecting true, must be marked imprecise. The results
996 # expecting false are fine because NaN compared >= or < anything is false.
997
998 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
999 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
1000 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
1001
1002 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
1003 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
1004 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
1005
1006 (('ine', 'a(is_not_zero)', 0), True),
1007 (('ieq', 'a(is_not_zero)', 0), False),
1008
1009 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
1010 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
1011 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
1012
1013 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
1014 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
1015 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
1016
1017 (('ult', 0, 'a(is_gt_zero)'), True),
1018 (('ult', a, 0), False),
1019
1020 # Packing and then unpacking does nothing
1021 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a),
1022 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b),
1023 (('unpack_64_2x32', ('pack_64_2x32_split', a, b)), ('vec2', a, b)),
1024 (('unpack_64_2x32', ('pack_64_2x32', a)), a),
1025 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a),
1026 ('unpack_64_2x32_split_y', a)), a),
1027 (('pack_64_2x32', ('vec2', ('unpack_64_2x32_split_x', a),
1028 ('unpack_64_2x32_split_y', a))), a),
1029 (('pack_64_2x32', ('unpack_64_2x32', a)), a),
1030
1031 # Comparing two halves of an unpack separately. While this optimization
1032 # should be correct for non-constant values, it's less obvious that it's
1033 # useful in that case. For constant values, the pack will fold and we're
1034 # guaranteed to reduce the whole tree to one instruction.
1035 (('iand', ('ieq', ('unpack_32_2x16_split_x', a), '#b'),
1036 ('ieq', ('unpack_32_2x16_split_y', a), '#c')),
1037 ('ieq', a, ('pack_32_2x16_split', b, c))),
1038
1039 # Byte extraction
1040 (('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
1041 (('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
1042 (('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
1043 (('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
1044 (('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
1045 (('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
1046 (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'),
1047
1048 (('ubfe', a, 0, 8), ('extract_u8', a, 0), '!options->lower_extract_byte'),
1049 (('ubfe', a, 8, 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
1050 (('ubfe', a, 16, 8), ('extract_u8', a, 2), '!options->lower_extract_byte'),
1051 (('ubfe', a, 24, 8), ('extract_u8', a, 3), '!options->lower_extract_byte'),
1052 (('ibfe', a, 0, 8), ('extract_i8', a, 0), '!options->lower_extract_byte'),
1053 (('ibfe', a, 8, 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
1054 (('ibfe', a, 16, 8), ('extract_i8', a, 2), '!options->lower_extract_byte'),
1055 (('ibfe', a, 24, 8), ('extract_i8', a, 3), '!options->lower_extract_byte'),
1056
1057 # Word extraction
1058 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
1059 (('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
1060 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
1061 (('ishr', 'a@32', 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
1062 (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),
1063
1064 (('ubfe', a, 0, 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
1065 (('ubfe', a, 16, 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
1066 (('ibfe', a, 0, 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
1067 (('ibfe', a, 16, 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
1068
1069 # Lower pack/unpack
1070 (('pack_64_2x32_split', a, b), ('ior', ('u2u64', a), ('ishl', ('u2u64', b), 32)), 'options->lower_pack_64_2x32_split'),
1071 (('pack_32_2x16_split', a, b), ('ior', ('u2u32', a), ('ishl', ('u2u32', b), 16)), 'options->lower_pack_32_2x16_split'),
1072 (('unpack_64_2x32_split_x', a), ('u2u32', a), 'options->lower_unpack_64_2x32_split'),
1073 (('unpack_64_2x32_split_y', a), ('u2u32', ('ushr', a, 32)), 'options->lower_unpack_64_2x32_split'),
1074 (('unpack_32_2x16_split_x', a), ('u2u16', a), 'options->lower_unpack_32_2x16_split'),
1075 (('unpack_32_2x16_split_y', a), ('u2u16', ('ushr', a, 16)), 'options->lower_unpack_32_2x16_split'),
1076
1077 # Useless masking before unpacking
1078 (('unpack_half_2x16_split_x', ('iand', a, 0xffff)), ('unpack_half_2x16_split_x', a)),
1079 (('unpack_32_2x16_split_x', ('iand', a, 0xffff)), ('unpack_32_2x16_split_x', a)),
1080 (('unpack_64_2x32_split_x', ('iand', a, 0xffffffff)), ('unpack_64_2x32_split_x', a)),
1081 (('unpack_half_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_half_2x16_split_y', a)),
1082 (('unpack_32_2x16_split_y', ('iand', a, 0xffff0000)), ('unpack_32_2x16_split_y', a)),
1083 (('unpack_64_2x32_split_y', ('iand', a, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a)),
1084
1085 (('unpack_half_2x16_split_x', ('extract_u16', a, 0)), ('unpack_half_2x16_split_x', a)),
1086 (('unpack_half_2x16_split_x', ('extract_u16', a, 1)), ('unpack_half_2x16_split_y', a)),
1087 (('unpack_32_2x16_split_x', ('extract_u16', a, 0)), ('unpack_32_2x16_split_x', a)),
1088 (('unpack_32_2x16_split_x', ('extract_u16', a, 1)), ('unpack_32_2x16_split_y', a)),
1089
1090 # Optimize half packing
1091 (('ishl', ('pack_half_2x16', ('vec2', a, 0)), 16), ('pack_half_2x16', ('vec2', 0, a))),
1092 (('ushr', ('pack_half_2x16', ('vec2', 0, a)), 16), ('pack_half_2x16', ('vec2', a, 0))),
1093
1094 (('iadd', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
1095 ('pack_half_2x16', ('vec2', a, b))),
1096 (('ior', ('pack_half_2x16', ('vec2', a, 0)), ('pack_half_2x16', ('vec2', 0, b))),
1097 ('pack_half_2x16', ('vec2', a, b))),
1098
1099 (('ishl', ('pack_half_2x16_split', a, 0), 16), ('pack_half_2x16_split', 0, a)),
1100 (('ushr', ('pack_half_2x16_split', 0, a), 16), ('pack_half_2x16_split', a, 0)),
1101 (('extract_u16', ('pack_half_2x16_split', 0, a), 1), ('pack_half_2x16_split', a, 0)),
1102
1103 (('iadd', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
1104 (('ior', ('pack_half_2x16_split', a, 0), ('pack_half_2x16_split', 0, b)), ('pack_half_2x16_split', a, b)),
1105 ])
1106
1107 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
1108 # patterns like those below.
1109 for op in ('ushr', 'ishr'):
1110 optimizations.extend([(('extract_u8', (op, 'a@16', 8), 0), ('extract_u8', a, 1))])
1111 optimizations.extend([(('extract_u8', (op, 'a@32', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)])
1112 optimizations.extend([(('extract_u8', (op, 'a@64', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)])
1113
1114 optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))])
1115
1116 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
1117 # patterns like those below.
1118 for op in ('extract_u8', 'extract_i8'):
1119 optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))])
1120 optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)])
1121 optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)])
1122
1123 optimizations.extend([
1124 # Subtracts
1125 (('ussub_4x8', a, 0), a),
1126 (('ussub_4x8', a, ~0), 0),
1127 # Lower all Subtractions first - they can get recombined later
1128 (('fsub', a, b), ('fadd', a, ('fneg', b))),
1129 (('isub', a, b), ('iadd', a, ('ineg', b))),
1130 (('uabs_usub', a, b), ('bcsel', ('ult', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),
1131 # This is correct. We don't need isub_sat because the result type is unsigned, so it cannot overflow.
1132 (('uabs_isub', a, b), ('bcsel', ('ilt', a, b), ('ineg', ('isub', a, b)), ('isub', a, b))),
1133
1134 # Propagate negation up multiplication chains
1135 (('fmul(is_used_by_non_fsat)', ('fneg', a), b), ('fneg', ('fmul', a, b))),
1136 (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),
1137
1138 # Propagate constants up multiplication chains
1139 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)),
1140 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)),
1141 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)),
1142 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)),
1143
1144 # Reassociate constants in add/mul chains so they can be folded together.
1145 # For now, we mostly only handle cases where the constants are separated by
1146 # a single non-constant. We could do better eventually.
1147 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a, c), b)),
1148 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a, c), b)),
1149 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a, c), b)),
1150 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a, ('fneg', c)), ('fneg', b))),
1151 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a, c), b)),
1152 (('iand', '#a', ('iand', 'b(is_not_const)', '#c')), ('iand', ('iand', a, c), b)),
1153 (('ior', '#a', ('ior', 'b(is_not_const)', '#c')), ('ior', ('ior', a, c), b)),
1154 (('ixor', '#a', ('ixor', 'b(is_not_const)', '#c')), ('ixor', ('ixor', a, c), b)),
1155
1156 # Drop mul-div by the same value when there's no wrapping.
1157 (('idiv', ('imul(no_signed_wrap)', a, b), b), a),
1158
1159 # By definition...
1160 (('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)),
1161 (('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
1162 (('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
1163
1164 (('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)),
1165 (('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
1166 (('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
1167
1168 (('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),
1169
1170 (('~fmul', ('bcsel(is_used_once)', c, -1.0, 1.0), b), ('bcsel', c, ('fneg', b), b)),
1171 (('~fmul', ('bcsel(is_used_once)', c, 1.0, -1.0), b), ('bcsel', c, b, ('fneg', b))),
1172 (('~bcsel', ('flt', a, 0.0), ('fneg', a), a), ('fabs', a)),
1173
1174 (('bcsel', a, ('bcsel', b, c, d), d), ('bcsel', ('iand', a, b), c, d)),
1175 (('bcsel', a, b, ('bcsel', c, b, d)), ('bcsel', ('ior', a, c), b, d)),
1176
1177 # Misc. lowering
1178 (('fmod', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'),
1179 (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod'),
1180 (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
1181 (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),
1182
1183 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1184 ('bcsel', ('ult', 31, 'bits'), 'insert',
1185 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
1186 'options->lower_bitfield_insert'),
1187 (('ihadd', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1188 (('uhadd', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1189 (('irhadd', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1190 (('urhadd', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
1191 (('ihadd@64', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1192 (('uhadd@64', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1193 (('irhadd@64', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1194 (('urhadd@64', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1195
1196 (('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'),
1197 (('uadd_sat', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat'),
1198 (('usub_sat', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_add_sat'),
1199 (('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'),
1200
1201 # int64_t sum = a + b;
1202 #
1203 # if (a < 0 && b < 0 && a < sum)
1204 # sum = INT64_MIN;
1205 # } else if (a >= 0 && b >= 0 && sum < a)
1206 # sum = INT64_MAX;
1207 # }
1208 #
1209 # A couple optimizations are applied.
1210 #
1211 # 1. a < sum => sum >= 0. This replacement works because it is known that
1212 # a < 0 and b < 0, so sum should also be < 0 unless there was
1213 # underflow.
1214 #
1215 # 2. sum < a => sum < 0. This replacement works because it is known that
1216 # a >= 0 and b >= 0, so sum should also be >= 0 unless there was
1217 # overflow.
1218 #
1219 # 3. Invert the second if-condition and swap the order of parameters for
1220 # the bcsel. !(a >= 0 && b >= 0 && sum < 0) becomes !(a >= 0) || !(b >=
1221 # 0) || !(sum < 0), and that becomes (a < 0) || (b < 0) || (sum >= 0)
1222 #
1223 # On Intel Gen11, this saves ~11 instructions.
1224 (('iadd_sat@64', a, b), ('bcsel',
1225 ('iand', ('iand', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
1226 0x8000000000000000,
1227 ('bcsel',
1228 ('ior', ('ior', ('ilt', a, 0), ('ilt', b, 0)), ('ige', ('iadd', a, b), 0)),
1229 ('iadd', a, b),
1230 0x7fffffffffffffff)),
1231 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1232
1233 # int64_t sum = a - b;
1234 #
1235 # if (a < 0 && b >= 0 && a < sum)
1236 # sum = INT64_MIN;
1237 # } else if (a >= 0 && b < 0 && a >= sum)
1238 # sum = INT64_MAX;
1239 # }
1240 #
1241 # Optimizations similar to the iadd_sat case are applied here.
1242 (('isub_sat@64', a, b), ('bcsel',
1243 ('iand', ('iand', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
1244 0x8000000000000000,
1245 ('bcsel',
1246 ('ior', ('ior', ('ilt', a, 0), ('ige', b, 0)), ('ige', ('isub', a, b), 0)),
1247 ('isub', a, b),
1248 0x7fffffffffffffff)),
1249 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1250
1251 # These are done here instead of in the backend because the int64 lowering
1252 # pass will make a mess of the patterns. The first patterns are
1253 # conditioned on nir_lower_minmax64 because it was not clear that it was
1254 # always an improvement on platforms that have real int64 support. No
1255 # shaders in shader-db hit this, so it was hard to say one way or the
1256 # other.
1257 (('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'),
1258 (('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'),
1259 (('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'),
1260 (('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'),
1261 (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1262 (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1263
1264 (('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'),
1265 (('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'),
1266 # 0u < uint(a) <=> uint(a) != 0u
1267 (('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'),
1268
1269 # Alternative lowering that doesn't rely on bfi.
1270 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1271 ('bcsel', ('ult', 31, 'bits'),
1272 'insert',
1273 (('ior',
1274 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1275 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1276 'options->lower_bitfield_insert_to_shifts'),
1277
1278 # Alternative lowering that uses bitfield_select.
1279 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1280 ('bcsel', ('ult', 31, 'bits'), 'insert',
1281 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1282 'options->lower_bitfield_insert_to_bitfield_select'),
1283
1284 (('ibitfield_extract', 'value', 'offset', 'bits'),
1285 ('bcsel', ('ult', 31, 'bits'), 'value',
1286 ('ibfe', 'value', 'offset', 'bits')),
1287 'options->lower_bitfield_extract'),
1288
1289 (('ubitfield_extract', 'value', 'offset', 'bits'),
1290 ('bcsel', ('ult', 31, 'bits'), 'value',
1291 ('ubfe', 'value', 'offset', 'bits')),
1292 'options->lower_bitfield_extract'),
1293
1294 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1295 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1296 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1297 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1298 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1299 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1300 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1301
1302 # Section 8.8 (Integer Functions) of the GLSL 4.60 spec says:
1303 #
1304 # If bits is zero, the result will be zero.
1305 #
1306 # These patterns prevent other patterns from generating invalid results
1307 # when count is zero.
1308 (('ubfe', a, b, 0), 0),
1309 (('ibfe', a, b, 0), 0),
1310
1311 (('ubfe', a, 0, '#b'), ('iand', a, ('ushr', 0xffffffff, ('ineg', b)))),
1312
1313 (('b2i32', ('i2b', ('ubfe', a, b, 1))), ('ubfe', a, b, 1)),
1314 (('b2i32', ('i2b', ('ibfe', a, b, 1))), ('ubfe', a, b, 1)), # ubfe in the replacement is correct
1315 (('ine', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1316 (('ieq', ('ibfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1317 (('ine', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ine', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1318 (('ieq', ('ubfe(is_used_once)', a, '#b', '#c'), 0), ('ieq', ('iand', a, ('ishl', ('ushr', 0xffffffff, ('ineg', c)), b)), 0)),
1319
1320 (('ibitfield_extract', 'value', 'offset', 'bits'),
1321 ('bcsel', ('ieq', 0, 'bits'),
1322 0,
1323 ('ishr',
1324 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1325 ('isub', 32, 'bits'))),
1326 'options->lower_bitfield_extract_to_shifts'),
1327
1328 (('ubitfield_extract', 'value', 'offset', 'bits'),
1329 ('iand',
1330 ('ushr', 'value', 'offset'),
1331 ('bcsel', ('ieq', 'bits', 32),
1332 0xffffffff,
1333 ('isub', ('ishl', 1, 'bits'), 1))),
1334 'options->lower_bitfield_extract_to_shifts'),
1335
1336 (('ifind_msb', 'value'),
1337 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1338 'options->lower_ifind_msb'),
1339
1340 (('find_lsb', 'value'),
1341 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1342 'options->lower_find_lsb'),
1343
1344 (('extract_i8', a, 'b@32'),
1345 ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24),
1346 'options->lower_extract_byte'),
1347
1348 (('extract_u8', a, 'b@32'),
1349 ('iand', ('ushr', a, ('imul', b, 8)), 0xff),
1350 'options->lower_extract_byte'),
1351
1352 (('extract_i16', a, 'b@32'),
1353 ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16),
1354 'options->lower_extract_word'),
1355
1356 (('extract_u16', a, 'b@32'),
1357 ('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
1358 'options->lower_extract_word'),
1359
1360 (('pack_unorm_2x16', 'v'),
1361 ('pack_uvec2_to_uint',
1362 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1363 'options->lower_pack_unorm_2x16'),
1364
1365 (('pack_unorm_4x8', 'v'),
1366 ('pack_uvec4_to_uint',
1367 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1368 'options->lower_pack_unorm_4x8'),
1369
1370 (('pack_snorm_2x16', 'v'),
1371 ('pack_uvec2_to_uint',
1372 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1373 'options->lower_pack_snorm_2x16'),
1374
1375 (('pack_snorm_4x8', 'v'),
1376 ('pack_uvec4_to_uint',
1377 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1378 'options->lower_pack_snorm_4x8'),
1379
1380 (('unpack_unorm_2x16', 'v'),
1381 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1382 ('extract_u16', 'v', 1))),
1383 65535.0),
1384 'options->lower_unpack_unorm_2x16'),
1385
1386 (('unpack_unorm_4x8', 'v'),
1387 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1388 ('extract_u8', 'v', 1),
1389 ('extract_u8', 'v', 2),
1390 ('extract_u8', 'v', 3))),
1391 255.0),
1392 'options->lower_unpack_unorm_4x8'),
1393
1394 (('unpack_snorm_2x16', 'v'),
1395 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1396 ('extract_i16', 'v', 1))),
1397 32767.0))),
1398 'options->lower_unpack_snorm_2x16'),
1399
1400 (('unpack_snorm_4x8', 'v'),
1401 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1402 ('extract_i8', 'v', 1),
1403 ('extract_i8', 'v', 2),
1404 ('extract_i8', 'v', 3))),
1405 127.0))),
1406 'options->lower_unpack_snorm_4x8'),
1407
1408 (('pack_half_2x16_split', 'a@32', 'b@32'),
1409 ('ior', ('ishl', ('u2u32', ('f2f16', b)), 16), ('u2u32', ('f2f16', a))),
1410 'options->lower_pack_split'),
1411
1412 (('unpack_half_2x16_split_x', 'a@32'),
1413 ('f2f32', ('u2u16', a)),
1414 'options->lower_pack_split'),
1415
1416 (('unpack_half_2x16_split_y', 'a@32'),
1417 ('f2f32', ('u2u16', ('ushr', a, 16))),
1418 'options->lower_pack_split'),
1419
1420 (('pack_32_2x16_split', 'a@16', 'b@16'),
1421 ('ior', ('ishl', ('u2u32', b), 16), ('u2u32', a)),
1422 'options->lower_pack_split'),
1423
1424 (('unpack_32_2x16_split_x', 'a@32'),
1425 ('u2u16', a),
1426 'options->lower_pack_split'),
1427
1428 (('unpack_32_2x16_split_y', 'a@32'),
1429 ('u2u16', ('ushr', 'a', 16)),
1430 'options->lower_pack_split'),
1431
1432 (('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
1433 (('imin', ('imax', a, -1), 1), ('isign', a), '!options->lower_isign'),
1434 (('imax', ('imin', a, 1), -1), ('isign', a), '!options->lower_isign'),
1435 (('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),
1436 (('fadd', ('b2f32', ('flt', 0.0, 'a@32')), ('fneg', ('b2f32', ('flt', 'a@32', 0.0)))), ('fsign', a), '!options->lower_fsign'),
1437 (('iadd', ('b2i32', ('flt', 0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0)))), ('f2i32', ('fsign', a)), '!options->lower_fsign'),
1438
1439 # Address/offset calculations:
1440 # Drivers supporting imul24 should use the nir_lower_amul() pass, this
1441 # rule converts everyone else to imul:
1442 (('amul', a, b), ('imul', a, b), '!options->has_imul24'),
1443
1444 (('umul24', a, b),
1445 ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)),
1446 '!options->has_umul24'),
1447 (('umad24', a, b, c),
1448 ('iadd', ('imul', ('iand', a, 0xffffff), ('iand', b, 0xffffff)), c),
1449 '!options->has_umad24'),
1450
1451 (('imad24_ir3', a, b, 0), ('imul24', a, b)),
1452 (('imad24_ir3', a, 0, c), (c)),
1453 (('imad24_ir3', a, 1, c), ('iadd', a, c)),
1454
1455 # if first two srcs are const, crack apart the imad so constant folding
1456 # can clean up the imul:
1457 # TODO ffma should probably get a similar rule:
1458 (('imad24_ir3', '#a', '#b', c), ('iadd', ('imul', a, b), c)),
1459
1460 # These will turn 24b address/offset calc back into 32b shifts, but
1461 # it should be safe to get back some of the bits of precision that we
1462 # already decided were no necessary:
1463 (('imul24', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b)), '!options->lower_bitops'),
1464 (('imul24', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b)))), '!options->lower_bitops'),
1465 (('imul24', a, 0), (0)),
1466 ])
1467
1468 # bit_size dependent lowerings
1469 for bit_size in [8, 16, 32, 64]:
1470 # convenience constants
1471 intmax = (1 << (bit_size - 1)) - 1
1472 intmin = 1 << (bit_size - 1)
1473
1474 optimizations += [
1475 (('iadd_sat@' + str(bit_size), a, b),
1476 ('bcsel', ('ige', b, 1), ('bcsel', ('ilt', ('iadd', a, b), a), intmax, ('iadd', a, b)),
1477 ('bcsel', ('ilt', a, ('iadd', a, b)), intmin, ('iadd', a, b))), 'options->lower_add_sat'),
1478 (('isub_sat@' + str(bit_size), a, b),
1479 ('bcsel', ('ilt', b, 0), ('bcsel', ('ilt', ('isub', a, b), a), intmax, ('isub', a, b)),
1480 ('bcsel', ('ilt', a, ('isub', a, b)), intmin, ('isub', a, b))), 'options->lower_add_sat'),
1481 ]
1482
1483 invert = OrderedDict([('feq', 'fneu'), ('fneu', 'feq')])
1484
1485 for left, right in itertools.combinations_with_replacement(invert.keys(), 2):
1486 optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))),
1487 ('iand', (invert[left], a, b), (invert[right], c, d))))
1488 optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))),
1489 ('ior', (invert[left], a, b), (invert[right], c, d))))
1490
1491 # Optimize x2bN(b2x(x)) -> x
1492 for size in type_sizes('bool'):
1493 aN = 'a@' + str(size)
1494 f2bN = 'f2b' + str(size)
1495 i2bN = 'i2b' + str(size)
1496 optimizations.append(((f2bN, ('b2f', aN)), a))
1497 optimizations.append(((i2bN, ('b2i', aN)), a))
1498
1499 # Optimize x2yN(b2x(x)) -> b2y
1500 for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1501 if x != 'f' and y != 'f' and x != y:
1502 continue
1503
1504 b2x = 'b2f' if x == 'f' else 'b2i'
1505 b2y = 'b2f' if y == 'f' else 'b2i'
1506 x2yN = '{}2{}'.format(x, y)
1507 optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
1508
1509 # Optimize away x2xN(a@N)
1510 for t in ['int', 'uint', 'float', 'bool']:
1511 for N in type_sizes(t):
1512 x2xN = '{0}2{0}{1}'.format(t[0], N)
1513 aN = 'a@{0}'.format(N)
1514 optimizations.append(((x2xN, aN), a))
1515
1516 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1517 # In particular, we can optimize away everything except upcast of downcast and
1518 # upcasts where the type differs from the other cast
1519 for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
1520 if N < M:
1521 # The outer cast is a down-cast. It doesn't matter what the size of the
1522 # argument of the inner cast is because we'll never been in the upcast
1523 # of downcast case. Regardless of types, we'll always end up with y2yN
1524 # in the end.
1525 for x, y in itertools.product(['i', 'u'], ['i', 'u']):
1526 x2xN = '{0}2{0}{1}'.format(x, N)
1527 y2yM = '{0}2{0}{1}'.format(y, M)
1528 y2yN = '{0}2{0}{1}'.format(y, N)
1529 optimizations.append(((x2xN, (y2yM, a)), (y2yN, a)))
1530 elif N > M:
1531 # If the outer cast is an up-cast, we have to be more careful about the
1532 # size of the argument of the inner cast and with types. In this case,
1533 # the type is always the type of type up-cast which is given by the
1534 # outer cast.
1535 for P in type_sizes('uint'):
1536 # We can't optimize away up-cast of down-cast.
1537 if M < P:
1538 continue
1539
1540 # Because we're doing down-cast of down-cast, the types always have
1541 # to match between the two casts
1542 for x in ['i', 'u']:
1543 x2xN = '{0}2{0}{1}'.format(x, N)
1544 x2xM = '{0}2{0}{1}'.format(x, M)
1545 aP = 'a@{0}'.format(P)
1546 optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a)))
1547 else:
1548 # The N == M case is handled by other optimizations
1549 pass
1550
1551 # Downcast operations should be able to see through pack
1552 for t in ['i', 'u']:
1553 for N in [8, 16, 32]:
1554 x2xN = '{0}2{0}{1}'.format(t, N)
1555 optimizations += [
1556 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
1557 ((x2xN, ('pack_64_2x32_split', a, b)), (x2xN, a)),
1558 ]
1559
1560 # Optimize comparisons with up-casts
1561 for t in ['int', 'uint', 'float']:
1562 for N, M in itertools.product(type_sizes(t), repeat=2):
1563 if N == 1 or N >= M:
1564 continue
1565
1566 cond = 'true'
1567 if N == 8:
1568 cond = 'options->support_8bit_alu'
1569 elif N == 16:
1570 cond = 'options->support_16bit_alu'
1571 x2xM = '{0}2{0}{1}'.format(t[0], M)
1572 x2xN = '{0}2{0}{1}'.format(t[0], N)
1573 aN = 'a@' + str(N)
1574 bN = 'b@' + str(N)
1575 xeq = 'feq' if t == 'float' else 'ieq'
1576 xne = 'fneu' if t == 'float' else 'ine'
1577 xge = '{0}ge'.format(t[0])
1578 xlt = '{0}lt'.format(t[0])
1579
1580 # Up-casts are lossless so for correctly signed comparisons of
1581 # up-casted values we can do the comparison at the largest of the two
1582 # original sizes and drop one or both of the casts. (We have
1583 # optimizations to drop the no-op casts which this may generate.)
1584 for P in type_sizes(t):
1585 if P == 1 or P > N:
1586 continue
1587
1588 bP = 'b@' + str(P)
1589 optimizations += [
1590 ((xeq, (x2xM, aN), (x2xM, bP)), (xeq, a, (x2xN, b)), cond),
1591 ((xne, (x2xM, aN), (x2xM, bP)), (xne, a, (x2xN, b)), cond),
1592 ((xge, (x2xM, aN), (x2xM, bP)), (xge, a, (x2xN, b)), cond),
1593 ((xlt, (x2xM, aN), (x2xM, bP)), (xlt, a, (x2xN, b)), cond),
1594 ((xge, (x2xM, bP), (x2xM, aN)), (xge, (x2xN, b), a), cond),
1595 ((xlt, (x2xM, bP), (x2xM, aN)), (xlt, (x2xN, b), a), cond),
1596 ]
1597
1598 # The next bit doesn't work on floats because the range checks would
1599 # get way too complicated.
1600 if t in ['int', 'uint']:
1601 if t == 'int':
1602 xN_min = -(1 << (N - 1))
1603 xN_max = (1 << (N - 1)) - 1
1604 elif t == 'uint':
1605 xN_min = 0
1606 xN_max = (1 << N) - 1
1607 else:
1608 assert False
1609
1610 # If we're up-casting and comparing to a constant, we can unfold
1611 # the comparison into a comparison with the shrunk down constant
1612 # and a check that the constant fits in the smaller bit size.
1613 optimizations += [
1614 ((xeq, (x2xM, aN), '#b'),
1615 ('iand', (xeq, a, (x2xN, b)), (xeq, (x2xM, (x2xN, b)), b)), cond),
1616 ((xne, (x2xM, aN), '#b'),
1617 ('ior', (xne, a, (x2xN, b)), (xne, (x2xM, (x2xN, b)), b)), cond),
1618 ((xlt, (x2xM, aN), '#b'),
1619 ('iand', (xlt, xN_min, b),
1620 ('ior', (xlt, xN_max, b), (xlt, a, (x2xN, b)))), cond),
1621 ((xlt, '#a', (x2xM, bN)),
1622 ('iand', (xlt, a, xN_max),
1623 ('ior', (xlt, a, xN_min), (xlt, (x2xN, a), b))), cond),
1624 ((xge, (x2xM, aN), '#b'),
1625 ('iand', (xge, xN_max, b),
1626 ('ior', (xge, xN_min, b), (xge, a, (x2xN, b)))), cond),
1627 ((xge, '#a', (x2xM, bN)),
1628 ('iand', (xge, a, xN_min),
1629 ('ior', (xge, a, xN_max), (xge, (x2xN, a), b))), cond),
1630 ]
1631
1632 # Convert masking followed by signed downcast to just unsigned downcast
1633 optimizations += [
1634 (('i2i32', ('iand', 'a@64', 0xffffffff)), ('u2u32', a)),
1635 (('i2i16', ('iand', 'a@32', 0xffff)), ('u2u16', a)),
1636 (('i2i16', ('iand', 'a@64', 0xffff)), ('u2u16', a)),
1637 (('i2i8', ('iand', 'a@16', 0xff)), ('u2u8', a)),
1638 (('i2i8', ('iand', 'a@32', 0xff)), ('u2u8', a)),
1639 (('i2i8', ('iand', 'a@64', 0xff)), ('u2u8', a)),
1640 ]
1641
1642 def fexp2i(exp, bits):
1643 # Generate an expression which constructs value 2.0^exp or 0.0.
1644 #
1645 # We assume that exp is already in a valid range:
1646 #
1647 # * [-15, 15] for 16-bit float
1648 # * [-127, 127] for 32-bit float
1649 # * [-1023, 1023] for 16-bit float
1650 #
1651 # If exp is the lowest value in the valid range, a value of 0.0 is
1652 # constructed. Otherwise, the value 2.0^exp is constructed.
1653 if bits == 16:
1654 return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
1655 elif bits == 32:
1656 return ('ishl', ('iadd', exp, 127), 23)
1657 elif bits == 64:
1658 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
1659 else:
1660 assert False
1661
1662 def ldexp(f, exp, bits):
1663 # The maximum possible range for a normal exponent is [-126, 127] and,
1664 # throwing in denormals, you get a maximum range of [-149, 127]. This
1665 # means that we can potentially have a swing of +-276. If you start with
1666 # FLT_MAX, you actually have to do ldexp(FLT_MAX, -278) to get it to flush
1667 # all the way to zero. The GLSL spec only requires that we handle a subset
1668 # of this range. From version 4.60 of the spec:
1669 #
1670 # "If exp is greater than +128 (single-precision) or +1024
1671 # (double-precision), the value returned is undefined. If exp is less
1672 # than -126 (single-precision) or -1022 (double-precision), the value
1673 # returned may be flushed to zero. Additionally, splitting the value
1674 # into a significand and exponent using frexp() and then reconstructing
1675 # a floating-point value using ldexp() should yield the original input
1676 # for zero and all finite non-denormalized values."
1677 #
1678 # The SPIR-V spec has similar language.
1679 #
1680 # In order to handle the maximum value +128 using the fexp2i() helper
1681 # above, we have to split the exponent in half and do two multiply
1682 # operations.
1683 #
1684 # First, we clamp exp to a reasonable range. Specifically, we clamp to
1685 # twice the full range that is valid for the fexp2i() function above. If
1686 # exp/2 is the bottom value of that range, the fexp2i() expression will
1687 # yield 0.0f which, when multiplied by f, will flush it to zero which is
1688 # allowed by the GLSL and SPIR-V specs for low exponent values. If the
1689 # value is clamped from above, then it must have been above the supported
1690 # range of the GLSL built-in and therefore any return value is acceptable.
1691 if bits == 16:
1692 exp = ('imin', ('imax', exp, -30), 30)
1693 elif bits == 32:
1694 exp = ('imin', ('imax', exp, -254), 254)
1695 elif bits == 64:
1696 exp = ('imin', ('imax', exp, -2046), 2046)
1697 else:
1698 assert False
1699
1700 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1701 # (We use ishr which isn't the same for -1, but the -1 case still works
1702 # since we use exp-exp/2 as the second exponent.) While the spec
1703 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1704 # work with denormals and doesn't allow for the full swing in exponents
1705 # that you can get with normalized values. Instead, we create two powers
1706 # of two and multiply by them each in turn. That way the effective range
1707 # of our exponent is doubled.
1708 pow2_1 = fexp2i(('ishr', exp, 1), bits)
1709 pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
1710 return ('fmul', ('fmul', f, pow2_1), pow2_2)
1711
1712 optimizations += [
1713 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1714 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1715 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1716 ]
1717
1718 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1719 def bitfield_reverse(u):
1720 step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))
1721 step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8))
1722 step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4))
1723 step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2))
1724 step5 = ('ior(many-comm-expr)', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1))
1725
1726 return step5
1727
1728 optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1729
1730 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1731 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1732 # and, if a is a NaN then the second comparison will fail anyway.
1733 for op in ['flt', 'fge', 'feq']:
1734 optimizations += [
1735 (('iand', ('feq', a, a), (op, a, b)), ('!' + op, a, b)),
1736 (('iand', ('feq', a, a), (op, b, a)), ('!' + op, b, a)),
1737 ]
1738
1739 # Add optimizations to handle the case where the result of a ternary is
1740 # compared to a constant. This way we can take things like
1741 #
1742 # (a ? 0 : 1) > 0
1743 #
1744 # and turn it into
1745 #
1746 # a ? (0 > 0) : (1 > 0)
1747 #
1748 # which constant folding will eat for lunch. The resulting ternary will
1749 # further get cleaned up by the boolean reductions above and we will be
1750 # left with just the original variable "a".
1751 for op in ['flt', 'fge', 'feq', 'fneu',
1752 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1753 optimizations += [
1754 ((op, ('bcsel', 'a', '#b', '#c'), '#d'),
1755 ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))),
1756 ((op, '#d', ('bcsel', a, '#b', '#c')),
1757 ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
1758 ]
1759
1760
1761 # For example, this converts things like
1762 #
1763 # 1 + mix(0, a - 1, condition)
1764 #
1765 # into
1766 #
1767 # mix(1, (a-1)+1, condition)
1768 #
1769 # Other optimizations will rearrange the constants.
1770 for op in ['fadd', 'fmul', 'iadd', 'imul']:
1771 optimizations += [
1772 ((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d)))
1773 ]
1774
1775 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1776 # states:
1777 #
1778 # If neither layout qualifier is specified, derivatives in compute shaders
1779 # return zero, which is consistent with the handling of built-in texture
1780 # functions like texture() in GLSL 4.50 compute shaders.
1781 for op in ['fddx', 'fddx_fine', 'fddx_coarse',
1782 'fddy', 'fddy_fine', 'fddy_coarse']:
1783 optimizations += [
1784 ((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1785 ]
1786
1787 # Some optimizations for ir3-specific instructions.
1788 optimizations += [
1789 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1790 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1791 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1792 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1793 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1794 ]
1795
1796 # These kinds of sequences can occur after nir_opt_peephole_select.
1797 #
1798 # NOTE: fadd is not handled here because that gets in the way of ffma
1799 # generation in the i965 driver. Instead, fadd and ffma are handled in
1800 # late_optimizations.
1801
1802 for op in ['flrp']:
1803 optimizations += [
1804 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, c, e)), (op, b, c, ('bcsel', a, d, e))),
1805 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, c, e)), (op, b, c, ('bcsel', a, d, e))),
1806 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, b, e, d)), (op, b, ('bcsel', a, c, e), d)),
1807 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', b, e, d)), (op, b, ('bcsel', a, c, e), d)),
1808 (('bcsel', a, (op + '(is_used_once)', b, c, d), (op, e, c, d)), (op, ('bcsel', a, b, e), c, d)),
1809 (('bcsel', a, (op, b, c, d), (op + '(is_used_once)', e, c, d)), (op, ('bcsel', a, b, e), c, d)),
1810 ]
1811
1812 for op in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']:
1813 optimizations += [
1814 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
1815 (('bcsel', a, (op + '(is_used_once)', b, 'c(is_not_const)'), (op, b, d)), (op, b, ('bcsel', a, c, d))),
1816 (('bcsel', a, (op, b, 'c(is_not_const)'), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
1817 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, 'd(is_not_const)')), (op, b, ('bcsel', a, c, d))),
1818 ]
1819
1820 for op in ['fpow']:
1821 optimizations += [
1822 (('bcsel', a, (op + '(is_used_once)', b, c), (op, b, d)), (op, b, ('bcsel', a, c, d))),
1823 (('bcsel', a, (op, b, c), (op + '(is_used_once)', b, d)), (op, b, ('bcsel', a, c, d))),
1824 (('bcsel', a, (op + '(is_used_once)', b, c), (op, d, c)), (op, ('bcsel', a, b, d), c)),
1825 (('bcsel', a, (op, b, c), (op + '(is_used_once)', d, c)), (op, ('bcsel', a, b, d), c)),
1826 ]
1827
1828 for op in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos', 'fneg', 'fabs', 'fsign']:
1829 optimizations += [
1830 (('bcsel', c, (op + '(is_used_once)', a), (op + '(is_used_once)', b)), (op, ('bcsel', c, a, b))),
1831 ]
1832
1833 for op in ['ineg', 'iabs', 'inot', 'isign']:
1834 optimizations += [
1835 ((op, ('bcsel', c, '#a', '#b')), ('bcsel', c, (op, a), (op, b))),
1836 ]
1837
1838 # This section contains optimizations to propagate downsizing conversions of
1839 # constructed vectors into vectors of downsized components. Whether this is
1840 # useful depends on the SIMD semantics of the backend. On a true SIMD machine,
1841 # this reduces the register pressure of the vector itself and often enables the
1842 # conversions to be eliminated via other algebraic rules or constant folding.
1843 # In the worst case on a SIMD architecture, the propagated conversions may be
1844 # revectorized via nir_opt_vectorize so instruction count is minimally
1845 # impacted.
1846 #
1847 # On a machine with SIMD-within-a-register only, this actually
1848 # counterintuitively hurts instruction count. These machines are the same that
1849 # require vectorize_vec2_16bit, so we predicate the optimizations on that flag
1850 # not being set.
1851 #
1852 # Finally for scalar architectures, there should be no difference in generated
1853 # code since it all ends up scalarized at the end, but it might minimally help
1854 # compile-times.
1855
1856 for i in range(2, 4 + 1):
1857 for T in ('f', 'u', 'i'):
1858 vec_inst = ('vec' + str(i),)
1859
1860 indices = ['a', 'b', 'c', 'd']
1861 suffix_in = tuple((indices[j] + '@32') for j in range(i))
1862
1863 to_16 = '{}2{}16'.format(T, T)
1864 to_mp = '{}2{}mp'.format(T, T)
1865
1866 out_16 = tuple((to_16, indices[j]) for j in range(i))
1867 out_mp = tuple((to_mp, indices[j]) for j in range(i))
1868
1869 optimizations += [
1870 ((to_16, vec_inst + suffix_in), vec_inst + out_16, '!options->vectorize_vec2_16bit'),
1871 ((to_mp, vec_inst + suffix_in), vec_inst + out_mp, '!options->vectorize_vec2_16bit')
1872 ]
1873
1874 # This section contains "late" optimizations that should be run before
1875 # creating ffmas and calling regular optimizations for the final time.
1876 # Optimizations should go here if they help code generation and conflict
1877 # with the regular optimizations.
1878 before_ffma_optimizations = [
1879 # Propagate constants down multiplication chains
1880 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)),
1881 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)),
1882 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)),
1883 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)),
1884
1885 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
1886 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
1887 (('~fadd', ('fneg', a), a), 0.0),
1888 (('iadd', ('ineg', a), a), 0),
1889 (('iadd', ('ineg', a), ('iadd', a, b)), b),
1890 (('iadd', a, ('iadd', ('ineg', a), b)), b),
1891 (('~fadd', ('fneg', a), ('fadd', a, b)), b),
1892 (('~fadd', a, ('fadd', ('fneg', a), b)), b),
1893
1894 (('~flrp', ('fadd(is_used_once)', a, -1.0), ('fadd(is_used_once)', a, 1.0), d), ('fadd', ('flrp', -1.0, 1.0, d), a)),
1895 (('~flrp', ('fadd(is_used_once)', a, 1.0), ('fadd(is_used_once)', a, -1.0), d), ('fadd', ('flrp', 1.0, -1.0, d), a)),
1896 (('~flrp', ('fadd(is_used_once)', a, '#b'), ('fadd(is_used_once)', a, '#c'), d), ('fadd', ('fmul', d, ('fadd', c, ('fneg', b))), ('fadd', a, b))),
1897 ]
1898
1899 # This section contains "late" optimizations that should be run after the
1900 # regular optimizations have finished. Optimizations should go here if
1901 # they help code generation but do not necessarily produce code that is
1902 # more easily optimizable.
1903 late_optimizations = [
1904 # Most of these optimizations aren't quite safe when you get infinity or
1905 # Nan involved but the first one should be fine.
1906 (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
1907 (('flt', ('fneg', ('fadd', a, b)), 0.0), ('flt', ('fneg', a), b)),
1908 (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
1909 (('~fge', ('fneg', ('fadd', a, b)), 0.0), ('fge', ('fneg', a), b)),
1910 (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
1911 (('~fneu', ('fadd', a, b), 0.0), ('fneu', a, ('fneg', b))),
1912
1913 # nir_lower_to_source_mods will collapse this, but its existence during the
1914 # optimization loop can prevent other optimizations.
1915 (('fneg', ('fneg', a)), a),
1916
1917 # Subtractions get lowered during optimization, so we need to recombine them
1918 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
1919 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
1920 (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
1921 (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
1922 (('iabs', a), ('imax', a, ('ineg', a)), 'options->lower_iabs'),
1923
1924 # These are duplicated from the main optimizations table. The late
1925 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1926 # new patterns like these. The patterns that compare with zero are removed
1927 # because they are unlikely to be created in by anything in
1928 # late_optimizations.
1929 (('flt', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('flt', a, b)),
1930 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('flt', b, a)),
1931 (('fge', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fge', a, b)),
1932 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a)), ('fge', b, a)),
1933 (('feq', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('feq', a, b)),
1934 (('fneu', ('fsat(is_used_once)', a), '#b(is_gt_0_and_lt_1)'), ('fneu', a, b)),
1935
1936 (('fge', ('fsat(is_used_once)', a), 1.0), ('fge', a, 1.0)),
1937 # flt(fsat(a), 1.0) is inexact because it returns True if a is NaN
1938 # (fsat(NaN) is 0), while flt(a, 1.0) always returns FALSE.
1939 (('~flt', ('fsat(is_used_once)', a), 1.0), ('flt', a, 1.0)),
1940
1941 (('~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)))),
1942
1943 (('flt', ('fneg', a), ('fneg', b)), ('flt', b, a)),
1944 (('fge', ('fneg', a), ('fneg', b)), ('fge', b, a)),
1945 (('feq', ('fneg', a), ('fneg', b)), ('feq', b, a)),
1946 (('fneu', ('fneg', a), ('fneg', b)), ('fneu', b, a)),
1947 (('flt', ('fneg', a), -1.0), ('flt', 1.0, a)),
1948 (('flt', -1.0, ('fneg', a)), ('flt', a, 1.0)),
1949 (('fge', ('fneg', a), -1.0), ('fge', 1.0, a)),
1950 (('fge', -1.0, ('fneg', a)), ('fge', a, 1.0)),
1951 (('fneu', ('fneg', a), -1.0), ('fneu', 1.0, a)),
1952 (('feq', -1.0, ('fneg', a)), ('feq', a, 1.0)),
1953
1954 (('ior', a, a), a),
1955 (('iand', a, a), a),
1956
1957 (('iand', ('ine(is_used_once)', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umin', a, b), 0)),
1958 (('ior', ('ieq(is_used_once)', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umin', a, b), 0)),
1959
1960 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a)))),
1961
1962 (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'),
1963 (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'),
1964 (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
1965 (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
1966
1967 (('~flrp', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
1968
1969 (('~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'),
1970 (('~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'),
1971
1972 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1973 # particular operation is common for expanding values stored in a texture
1974 # from [0,1] to [-1,1].
1975 (('~ffma@32', a, 2.0, -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'),
1976 (('~ffma@32', a, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'),
1977 (('~ffma@32', a, -2.0, 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'),
1978 (('~ffma@32', a, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
1979 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), -1.0), ('flrp', -1.0, 1.0, a ), '!options->lower_flrp32'),
1980 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), -1.0), ('flrp', -1.0, 1.0, ('fneg', a)), '!options->lower_flrp32'),
1981 (('~fadd@32', ('fmul(is_used_once)', -2.0, a), 1.0), ('flrp', 1.0, -1.0, a ), '!options->lower_flrp32'),
1982 (('~fadd@32', ('fmul(is_used_once)', 2.0, a), 1.0), ('flrp', 1.0, -1.0, ('fneg', a)), '!options->lower_flrp32'),
1983
1984 # flrp(a, b, a)
1985 # a*(1-a) + b*a
1986 # a + -a*a + a*b (1)
1987 # a + a*(b - a)
1988 # Option 1: ffma(a, (b-a), a)
1989 #
1990 # Alternately, after (1):
1991 # a*(1+b) + -a*a
1992 # a*((1+b) + -a)
1993 #
1994 # Let b=1
1995 #
1996 # Option 2: ffma(a, 2, -(a*a))
1997 # Option 3: ffma(a, 2, (-a)*a)
1998 # Option 4: ffma(a, -a, (2*a)