nir: Use the flrp lowering pass instead of nir_opt_algebraic
[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
33 # Convenience variables
34 a = 'a'
35 b = 'b'
36 c = 'c'
37 d = 'd'
38 e = 'e'
39
40 # Written in the form (<search>, <replace>) where <search> is an expression
41 # and <replace> is either an expression or a value. An expression is
42 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
43 # where each source is either an expression or a value. A value can be
44 # either a numeric constant or a string representing a variable name.
45 #
46 # If the opcode in a search expression is prefixed by a '~' character, this
47 # indicates that the operation is inexact. Such operations will only get
48 # applied to SSA values that do not have the exact bit set. This should be
49 # used by by any optimizations that are not bit-for-bit exact. It should not,
50 # however, be used for backend-requested lowering operations as those need to
51 # happen regardless of precision.
52 #
53 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
54 # that the given variable will only match constants and the type indicates that
55 # the given variable will only match values from ALU instructions with the
56 # given output type, and (cond) specifies an additional condition function
57 # (see nir_search_helpers.h).
58 #
59 # For constants, you have to be careful to make sure that it is the right
60 # type because python is unaware of the source and destination types of the
61 # opcodes.
62 #
63 # All expression types can have a bit-size specified. For opcodes, this
64 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
65 # type and size. In the search half of the expression this indicates that it
66 # should only match that particular bit-size. In the replace half of the
67 # expression this indicates that the constructed value should have that
68 # bit-size.
69
70 optimizations = [
71
72 (('imul', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b))),
73 (('imul', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b))))),
74 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
75 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a, b)), ('imul', a, b)),
76 (('imul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('imul_high', a, b)), 'options->lower_mul_2x32_64'),
77 (('umul_2x32_64', a, b), ('pack_64_2x32_split', ('imul', a, b), ('umul_high', a, b)), 'options->lower_mul_2x32_64'),
78 (('udiv', a, 1), a),
79 (('idiv', a, 1), a),
80 (('umod', a, 1), 0),
81 (('imod', a, 1), 0),
82 (('udiv', a, '#b@32(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b))),
83 (('idiv', a, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), 'options->lower_idiv'),
84 (('idiv', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', ('iabs', b))))), 'options->lower_idiv'),
85 (('umod', a, '#b(is_pos_power_of_two)'), ('iand', a, ('isub', b, 1))),
86
87 (('fneg', ('fneg', a)), a),
88 (('ineg', ('ineg', a)), a),
89 (('fabs', ('fabs', a)), ('fabs', a)),
90 (('fabs', ('fneg', a)), ('fabs', a)),
91 (('fabs', ('u2f', a)), ('u2f', a)),
92 (('iabs', ('iabs', a)), ('iabs', a)),
93 (('iabs', ('ineg', a)), ('iabs', a)),
94 (('f2b', ('fneg', a)), ('f2b', a)),
95 (('i2b', ('ineg', a)), ('i2b', a)),
96 (('~fadd', a, 0.0), a),
97 (('iadd', a, 0), a),
98 (('usadd_4x8', a, 0), a),
99 (('usadd_4x8', a, ~0), ~0),
100 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
101 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
102 (('~fadd', ('fneg', a), a), 0.0),
103 (('iadd', ('ineg', a), a), 0),
104 (('iadd', ('ineg', a), ('iadd', a, b)), b),
105 (('iadd', a, ('iadd', ('ineg', a), b)), b),
106 (('~fadd', ('fneg', a), ('fadd', a, b)), b),
107 (('~fadd', a, ('fadd', ('fneg', a), b)), b),
108 (('~fmul', a, 0.0), 0.0),
109 (('imul', a, 0), 0),
110 (('umul_unorm_4x8', a, 0), 0),
111 (('umul_unorm_4x8', a, ~0), a),
112 (('fmul', a, 1.0), a),
113 (('imul', a, 1), a),
114 (('fmul', a, -1.0), ('fneg', a)),
115 (('imul', a, -1), ('ineg', a)),
116 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
117 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
118 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
119 (('fmul', ('fsign', a), ('fmul', a, a)), ('fmul', ('fabs', a), a)),
120 (('fmul', ('fmul', ('fsign', a), a), a), ('fmul', ('fabs', a), a)),
121 (('~ffma', 0.0, a, b), b),
122 (('~ffma', a, 0.0, b), b),
123 (('~ffma', a, b, 0.0), ('fmul', a, b)),
124 (('ffma', a, 1.0, b), ('fadd', a, b)),
125 (('ffma', 1.0, a, b), ('fadd', a, b)),
126 (('~flrp', a, b, 0.0), a),
127 (('~flrp', a, b, 1.0), b),
128 (('~flrp', a, a, b), a),
129 (('~flrp', 0.0, a, b), ('fmul', a, b)),
130
131 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
132 (('~flrp', a, ('fadd(is_used_once)', a, b), c), ('fadd', ('fmul', b, c), a)),
133 (('~flrp@32', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp32'),
134 (('~flrp@64', a, ('fadd', a, b), c), ('fadd', ('fmul', b, c), a), 'options->lower_flrp64'),
135
136 (('~flrp@32', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp32'),
137 (('~flrp@64', ('fadd', a, b), ('fadd', a, c), d), ('fadd', ('flrp', b, c, d), a), 'options->lower_flrp64'),
138
139 (('~flrp@32', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp32'),
140 (('~flrp@64', a, ('fmul(is_used_once)', a, b), c), ('fmul', ('flrp', 1.0, b, c), a), 'options->lower_flrp64'),
141
142 (('~flrp', ('fmul(is_used_once)', a, b), ('fmul(is_used_once)', a, c), d), ('fmul', ('flrp', b, c, d), a)),
143
144 (('~flrp', a, b, ('b2f', 'c@1')), ('bcsel', c, b, a), 'options->lower_flrp32'),
145 (('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)),
146 (('ftrunc', a), ('bcsel', ('flt', a, 0.0), ('fneg', ('ffloor', ('fabs', a))), ('ffloor', ('fabs', a))), 'options->lower_ftrunc'),
147 (('ffloor', a), ('fsub', a, ('ffract', a)), 'options->lower_ffloor'),
148 (('fadd', a, ('fneg', ('ffract', a))), ('ffloor', a), '!options->lower_ffloor'),
149 (('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'),
150 (('fceil', a), ('fneg', ('ffloor', ('fneg', a))), 'options->lower_fceil'),
151 (('~fadd', ('fmul', a, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b, ('b2f', c))), ('bcsel', c, b, a), 'options->lower_flrp32'),
152 (('~fadd@32', ('fmul', a, ('fadd', 1.0, ('fneg', c ))), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp32'),
153 (('~fadd@64', ('fmul', a, ('fadd', 1.0, ('fneg', c ))), ('fmul', b, c )), ('flrp', a, b, c), '!options->lower_flrp64'),
154 (('~fadd', a, ('fmul', ('b2f', 'c@1'), ('fadd', b, ('fneg', a)))), ('bcsel', c, b, a), 'options->lower_flrp32'),
155 (('~fadd@32', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp32'),
156 (('~fadd@64', a, ('fmul', c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp64'),
157 (('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'),
158 (('~fadd', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma'),
159
160 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b, c)), '#d'), '#e'),
161 ('bcsel', a, ('fmul', ('fadd', ('fmul', b, c), d), e), ('fmul', d, e))),
162
163 (('fdot4', ('vec4', a, b, c, 1.0), d), ('fdph', ('vec3', a, b, c), d)),
164 (('fdot4', ('vec4', a, 0.0, 0.0, 0.0), b), ('fmul', a, b)),
165 (('fdot4', ('vec4', a, b, 0.0, 0.0), c), ('fdot2', ('vec2', a, b), c)),
166 (('fdot4', ('vec4', a, b, c, 0.0), d), ('fdot3', ('vec3', a, b, c), d)),
167
168 (('fdot3', ('vec3', a, 0.0, 0.0), b), ('fmul', a, b)),
169 (('fdot3', ('vec3', a, b, 0.0), c), ('fdot2', ('vec2', a, b), c)),
170
171 # (a * #b + #c) << #d
172 # ((a * #b) << #d) + (#c << #d)
173 # (a * (#b << #d)) + (#c << #d)
174 (('ishl', ('iadd', ('imul', a, '#b'), '#c'), '#d'),
175 ('iadd', ('imul', a, ('ishl', b, d)), ('ishl', c, d))),
176
177 # (a * #b) << #c
178 # a * (#b << #c)
179 (('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))),
180
181 # Comparison simplifications
182 (('~inot', ('flt', a, b)), ('fge', a, b)),
183 (('~inot', ('fge', a, b)), ('flt', a, b)),
184 (('~inot', ('feq', a, b)), ('fne', a, b)),
185 (('~inot', ('fne', a, b)), ('feq', a, b)),
186 (('inot', ('ilt', a, b)), ('ige', a, b)),
187 (('inot', ('ult', a, b)), ('uge', a, b)),
188 (('inot', ('ige', a, b)), ('ilt', a, b)),
189 (('inot', ('uge', a, b)), ('ult', a, b)),
190 (('inot', ('ieq', a, b)), ('ine', a, b)),
191 (('inot', ('ine', a, b)), ('ieq', a, b)),
192
193 # 0.0 >= b2f(a)
194 # b2f(a) <= 0.0
195 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
196 # inot(a)
197 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a)),
198
199 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a)),
200
201 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
202 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a, b)),
203 (('fne', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a, b)),
204 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a, b)),
205 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
206 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a, b)),
207 (('fne', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a, b)),
208 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a, b)),
209 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a, b)),
210 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a, b)),
211 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
212 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a, b))),
213 (('feq', ('bcsel', a, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a, b))),
214 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
215 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
216 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a, b))),
217 (('feq', ('bcsel', a, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a, b))),
218 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a, b)),
219 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a, b)),
220 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a, b)),
221
222 # -(b2f(a) + b2f(b)) < 0
223 # 0 < b2f(a) + b2f(b)
224 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
225 # a || b
226 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a, b)),
227 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a, b)),
228
229 # -(b2f(a) + b2f(b)) >= 0
230 # 0 >= b2f(a) + b2f(b)
231 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
232 # !(a || b)
233 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a, b))),
234 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a, b))),
235
236 (('flt', a, ('fneg', a)), ('flt', a, 0.0)),
237 (('fge', a, ('fneg', a)), ('fge', a, 0.0)),
238
239 # Some optimizations (below) convert things like (a < b || c < b) into
240 # (min(a, c) < b). However, this interfers with the previous optimizations
241 # that try to remove comparisons with negated sums of b2f. This just
242 # breaks that apart.
243 (('flt', ('fmin', c, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
244 ('ior', ('flt', c, 0.0), ('ior', a, b))),
245
246 (('~flt', ('fadd', a, b), a), ('flt', b, 0.0)),
247 (('~fge', ('fadd', a, b), a), ('fge', b, 0.0)),
248 (('~feq', ('fadd', a, b), a), ('feq', b, 0.0)),
249 (('~fne', ('fadd', a, b), a), ('fne', b, 0.0)),
250
251 # Cannot remove the addition from ilt or ige due to overflow.
252 (('ieq', ('iadd', a, b), a), ('ieq', b, 0)),
253 (('ine', ('iadd', a, b), a), ('ine', b, 0)),
254
255 # fmin(-b2f(a), b) >= 0.0
256 # -b2f(a) >= 0.0 && b >= 0.0
257 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
258 # b2f(a) == 0.0 && b >= 0.0
259 # a == False && b >= 0.0
260 # !a && b >= 0.0
261 #
262 # The fge in the second replacement is not a typo. I leave the proof that
263 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
264 # reader.
265 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
266 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a), ('fge', b, 0.0))),
267
268 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a)),
269 (('fne', ('b2f', 'a@1'), 0.0), a),
270 (('ieq', ('b2i', 'a@1'), 0), ('inot', a)),
271 (('ine', ('b2i', 'a@1'), 0), a),
272
273 (('fne', ('u2f', a), 0.0), ('ine', a, 0)),
274 (('feq', ('u2f', a), 0.0), ('ieq', a, 0)),
275 (('fge', ('u2f', a), 0.0), True),
276 (('fge', 0.0, ('u2f', a)), ('uge', 0, a)), # ieq instead?
277 (('flt', ('u2f', a), 0.0), False),
278 (('flt', 0.0, ('u2f', a)), ('ult', 0, a)), # ine instead?
279 (('fne', ('i2f', a), 0.0), ('ine', a, 0)),
280 (('feq', ('i2f', a), 0.0), ('ieq', a, 0)),
281 (('fge', ('i2f', a), 0.0), ('ige', a, 0)),
282 (('fge', 0.0, ('i2f', a)), ('ige', 0, a)),
283 (('flt', ('i2f', a), 0.0), ('ilt', a, 0)),
284 (('flt', 0.0, ('i2f', a)), ('ilt', 0, a)),
285
286 # 0.0 < fabs(a)
287 # fabs(a) > 0.0
288 # fabs(a) != 0.0 because fabs(a) must be >= 0
289 # a != 0.0
290 (('~flt', 0.0, ('fabs', a)), ('fne', a, 0.0)),
291
292 # -fabs(a) < 0.0
293 # fabs(a) > 0.0
294 (('~flt', ('fneg', ('fabs', a)), 0.0), ('fne', a, 0.0)),
295
296 # 0.0 >= fabs(a)
297 # 0.0 == fabs(a) because fabs(a) must be >= 0
298 # 0.0 == a
299 (('fge', 0.0, ('fabs', a)), ('feq', a, 0.0)),
300
301 # -fabs(a) >= 0.0
302 # 0.0 >= fabs(a)
303 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
304
305 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a, b))),
306 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a, b)))),
307 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
308 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a, b)))),
309
310 # fmin(b2f(a), b)
311 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
312 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
313 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
314 #
315 # Since b is a constant, constant folding will eliminate the fmin and the
316 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
317 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a, ('fmin', b, 1.0), ('fmin', b, 0.0))),
318
319 (('flt', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),
320
321 (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
322 (('~bcsel', ('flt', b, a), b, a), ('fmin', a, b)),
323 (('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
324 (('~bcsel', ('fge', a, b), b, a), ('fmin', a, b)),
325 (('~bcsel', ('fge', b, a), b, a), ('fmax', a, b)),
326 (('bcsel', ('i2b', a), b, c), ('bcsel', ('ine', a, 0), b, c)),
327 (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
328 (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
329 (('bcsel', a, b, ('bcsel', a, c, d)), ('bcsel', a, b, d)),
330 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
331 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, c, 'e')), ('bcsel', b, c, ('bcsel', a, d, 'e'))),
332 (('bcsel', a, ('bcsel', b, c, d), ('bcsel(is_used_once)', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
333 (('bcsel', a, ('bcsel(is_used_once)', b, c, d), ('bcsel', b, 'e', d)), ('bcsel', b, ('bcsel', a, c, 'e'), d)),
334 (('bcsel', a, True, b), ('ior', a, b)),
335 (('bcsel', a, a, b), ('ior', a, b)),
336 (('bcsel', a, b, False), ('iand', a, b)),
337 (('bcsel', a, b, a), ('iand', a, b)),
338 (('fmin', a, a), a),
339 (('fmax', a, a), a),
340 (('imin', a, a), a),
341 (('imax', a, a), a),
342 (('umin', a, a), a),
343 (('umax', a, a), a),
344 (('fmax', ('fmax', a, b), b), ('fmax', a, b)),
345 (('umax', ('umax', a, b), b), ('umax', a, b)),
346 (('imax', ('imax', a, b), b), ('imax', a, b)),
347 (('fmin', ('fmin', a, b), b), ('fmin', a, b)),
348 (('umin', ('umin', a, b), b), ('umin', a, b)),
349 (('imin', ('imin', a, b), b), ('imin', a, b)),
350 (('fmax', a, ('fneg', a)), ('fabs', a)),
351 (('imax', a, ('ineg', a)), ('iabs', a)),
352 (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
353 (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
354 (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
355 (('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))),
356 (('fmin', a, ('fabs', a)), a),
357 (('imin', a, ('iabs', a)), a),
358 (('fmax', a, ('fneg', ('fabs', a))), a),
359 (('imax', a, ('ineg', ('iabs', a))), a),
360 (('fmax', a, ('fabs', a)), ('fabs', a)),
361 (('imax', a, ('iabs', a)), ('iabs', a)),
362 (('fmax', a, ('fneg', a)), ('fabs', a)),
363 (('imax', a, ('ineg', a)), ('iabs', a)),
364 (('~fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'),
365 (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'),
366 (('fsat', ('fsign', a)), ('b2f', ('flt', 0.0, a))),
367 (('fsat', ('b2f', a)), ('b2f', a)),
368 (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
369 (('fsat', ('fsat', a)), ('fsat', a)),
370 (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)),
371 (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)),
372 (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
373 (('fmax', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
374 (('fmin', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a, b))),
375 (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
376 (('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
377 (('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
378 (('~ior', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
379 (('~ior', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
380 (('~ior', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmax', b, c))),
381 (('~ior', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmin', a, b), c)),
382 (('~ior', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmin', b, c))),
383 (('~ior', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmax', a, b), c)),
384 (('~iand', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmin', b, c))),
385 (('~iand', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmax', a, b), c)),
386 (('~iand', ('fge(is_used_once)', a, b), ('fge', a, c)), ('fge', a, ('fmax', b, c))),
387 (('~iand', ('fge(is_used_once)', a, c), ('fge', b, c)), ('fge', ('fmin', a, b), c)),
388 (('~iand', ('flt', a, '#b'), ('flt', a, '#c')), ('flt', a, ('fmin', b, c))),
389 (('~iand', ('flt', '#a', c), ('flt', '#b', c)), ('flt', ('fmax', a, b), c)),
390 (('~iand', ('fge', a, '#b'), ('fge', a, '#c')), ('fge', a, ('fmax', b, c))),
391 (('~iand', ('fge', '#a', c), ('fge', '#b', c)), ('fge', ('fmin', a, b), c)),
392
393 (('ior', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imax', b, c))),
394 (('ior', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imin', a, b), c)),
395 (('ior', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imin', b, c))),
396 (('ior', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imax', a, b), c)),
397 (('ior', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umax', b, c))),
398 (('ior', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umin', a, b), c)),
399 (('ior', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umin', b, c))),
400 (('ior', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umax', a, b), c)),
401 (('iand', ('ilt(is_used_once)', a, b), ('ilt', a, c)), ('ilt', a, ('imin', b, c))),
402 (('iand', ('ilt(is_used_once)', a, c), ('ilt', b, c)), ('ilt', ('imax', a, b), c)),
403 (('iand', ('ige(is_used_once)', a, b), ('ige', a, c)), ('ige', a, ('imax', b, c))),
404 (('iand', ('ige(is_used_once)', a, c), ('ige', b, c)), ('ige', ('imin', a, b), c)),
405 (('iand', ('ult(is_used_once)', a, b), ('ult', a, c)), ('ult', a, ('umin', b, c))),
406 (('iand', ('ult(is_used_once)', a, c), ('ult', b, c)), ('ult', ('umax', a, b), c)),
407 (('iand', ('uge(is_used_once)', a, b), ('uge', a, c)), ('uge', a, ('umax', b, c))),
408 (('iand', ('uge(is_used_once)', a, c), ('uge', b, c)), ('uge', ('umin', a, b), c)),
409
410 # Common pattern like 'if (i == 0 || i == 1 || ...)'
411 (('ior', ('ieq', a, 0), ('ieq', a, 1)), ('uge', 1, a)),
412 (('ior', ('uge', 1, a), ('ieq', a, 2)), ('uge', 2, a)),
413 (('ior', ('uge', 2, a), ('ieq', a, 3)), ('uge', 3, a)),
414
415 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
416 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
417 # so emit an open-coded version of that.
418 (('bcsel@32', ('feq', a, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
419 ('ior', 0x3f800000, ('iand', a, 0x80000000))),
420
421 (('ior', a, ('ieq', a, False)), True),
422 (('ior', a, ('inot', a)), -1),
423
424 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a, b)),
425 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a, b))),
426
427 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
428
429 # These patterns can result when (a < b || a < c) => (a < min(b, c))
430 # transformations occur before constant propagation and loop-unrolling.
431 (('~flt', a, ('fmax', b, a)), ('flt', a, b)),
432 (('~flt', ('fmin', a, b), a), ('flt', b, a)),
433 (('~fge', a, ('fmin', b, a)), True),
434 (('~fge', ('fmax', a, b), a), True),
435 (('~flt', a, ('fmin', b, a)), False),
436 (('~flt', ('fmax', a, b), a), False),
437 (('~fge', a, ('fmax', b, a)), ('fge', a, b)),
438 (('~fge', ('fmin', a, b), a), ('fge', b, a)),
439
440 (('ilt', a, ('imax', b, a)), ('ilt', a, b)),
441 (('ilt', ('imin', a, b), a), ('ilt', b, a)),
442 (('ige', a, ('imin', b, a)), True),
443 (('ige', ('imax', a, b), a), True),
444 (('ult', a, ('umax', b, a)), ('ult', a, b)),
445 (('ult', ('umin', a, b), a), ('ult', b, a)),
446 (('uge', a, ('umin', b, a)), True),
447 (('uge', ('umax', a, b), a), True),
448 (('ilt', a, ('imin', b, a)), False),
449 (('ilt', ('imax', a, b), a), False),
450 (('ige', a, ('imax', b, a)), ('ige', a, b)),
451 (('ige', ('imin', a, b), a), ('ige', b, a)),
452 (('ult', a, ('umin', b, a)), False),
453 (('ult', ('umax', a, b), a), False),
454 (('uge', a, ('umax', b, a)), ('uge', a, b)),
455 (('uge', ('umin', a, b), a), ('uge', b, a)),
456
457 (('ilt', '#a', ('imax', '#b', c)), ('ior', ('ilt', a, b), ('ilt', a, c))),
458 (('ilt', ('imin', '#a', b), '#c'), ('ior', ('ilt', a, c), ('ilt', b, c))),
459 (('ige', '#a', ('imin', '#b', c)), ('ior', ('ige', a, b), ('ige', a, c))),
460 (('ige', ('imax', '#a', b), '#c'), ('ior', ('ige', a, c), ('ige', b, c))),
461 (('ult', '#a', ('umax', '#b', c)), ('ior', ('ult', a, b), ('ult', a, c))),
462 (('ult', ('umin', '#a', b), '#c'), ('ior', ('ult', a, c), ('ult', b, c))),
463 (('uge', '#a', ('umin', '#b', c)), ('ior', ('uge', a, b), ('uge', a, c))),
464 (('uge', ('umax', '#a', b), '#c'), ('ior', ('uge', a, c), ('uge', b, c))),
465 (('ilt', '#a', ('imin', '#b', c)), ('iand', ('ilt', a, b), ('ilt', a, c))),
466 (('ilt', ('imax', '#a', b), '#c'), ('iand', ('ilt', a, c), ('ilt', b, c))),
467 (('ige', '#a', ('imax', '#b', c)), ('iand', ('ige', a, b), ('ige', a, c))),
468 (('ige', ('imin', '#a', b), '#c'), ('iand', ('ige', a, c), ('ige', b, c))),
469 (('ult', '#a', ('umin', '#b', c)), ('iand', ('ult', a, b), ('ult', a, c))),
470 (('ult', ('umax', '#a', b), '#c'), ('iand', ('ult', a, c), ('ult', b, c))),
471 (('uge', '#a', ('umax', '#b', c)), ('iand', ('uge', a, b), ('uge', a, c))),
472 (('uge', ('umin', '#a', b), '#c'), ('iand', ('uge', a, c), ('uge', b, c))),
473
474 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
475 # negative.
476 (('bcsel', ('ilt', a, 0), ('ineg', ('ishr', a, b)), ('ishr', a, b)),
477 ('iabs', ('ishr', a, b))),
478 (('iabs', ('ishr', ('iabs', a), b)), ('ishr', ('iabs', a), b)),
479
480 (('fabs', ('slt', a, b)), ('slt', a, b)),
481 (('fabs', ('sge', a, b)), ('sge', a, b)),
482 (('fabs', ('seq', a, b)), ('seq', a, b)),
483 (('fabs', ('sne', a, b)), ('sne', a, b)),
484 (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
485 (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
486 (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
487 (('sne', a, b), ('b2f', ('fne', a, b)), 'options->lower_scmp'),
488 (('fne', ('fneg', a), a), ('fne', a, 0.0)),
489 (('feq', ('fneg', a), a), ('feq', a, 0.0)),
490 # Emulating booleans
491 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a, b))),
492 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a, b))),
493 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a, b))),
494 (('iand', 'a@bool32', 1.0), ('b2f', a)),
495 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
496 (('ineg', ('b2i32', 'a@32')), a),
497 (('flt', ('fneg', ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
498 (('flt', ('fsub', 0.0, ('b2f', 'a@1')), 0), a), # Generated by TGSI KILL_IF.
499 # Comparison with the same args. Note that these are not done for
500 # the float versions because NaN always returns false on float
501 # inequalities.
502 (('ilt', a, a), False),
503 (('ige', a, a), True),
504 (('ieq', a, a), True),
505 (('ine', a, a), False),
506 (('ult', a, a), False),
507 (('uge', a, a), True),
508 # Logical and bit operations
509 (('fand', a, 0.0), 0.0),
510 (('iand', a, a), a),
511 (('iand', a, ~0), a),
512 (('iand', a, 0), 0),
513 (('ior', a, a), a),
514 (('ior', a, 0), a),
515 (('ior', a, True), True),
516 (('fxor', a, a), 0.0),
517 (('ixor', a, a), 0),
518 (('ixor', a, 0), a),
519 (('inot', ('inot', a)), a),
520 (('ior', ('iand', a, b), b), b),
521 (('ior', ('ior', a, b), b), ('ior', a, b)),
522 (('iand', ('ior', a, b), b), b),
523 (('iand', ('iand', a, b), b), ('iand', a, b)),
524 # DeMorgan's Laws
525 (('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))),
526 (('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))),
527 # Shift optimizations
528 (('ishl', 0, a), 0),
529 (('ishl', a, 0), a),
530 (('ishr', 0, a), 0),
531 (('ishr', a, 0), a),
532 (('ushr', 0, a), 0),
533 (('ushr', a, 0), a),
534 (('iand', 0xff, ('ushr@32', a, 24)), ('ushr', a, 24)),
535 (('iand', 0xffff, ('ushr@32', a, 16)), ('ushr', a, 16)),
536 # Exponential/logarithmic identities
537 (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
538 (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
539 (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
540 (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
541 (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
542 ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
543 (('~fexp2', ('fmul', ('flog2', a), 2.0)), ('fmul', a, a)),
544 (('~fexp2', ('fmul', ('flog2', a), 4.0)), ('fmul', ('fmul', a, a), ('fmul', a, a))),
545 (('~fpow', a, 1.0), a),
546 (('~fpow', a, 2.0), ('fmul', a, a)),
547 (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
548 (('~fpow', 2.0, a), ('fexp2', a)),
549 (('~fpow', ('fpow', a, 2.2), 0.454545), a),
550 (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
551 (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
552 (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
553 (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
554 (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
555 (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
556 (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
557 (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
558 (('~fmul', ('fexp2(is_used_once)', a), ('fexp2(is_used_once)', b)), ('fexp2', ('fadd', a, b))),
559 # Division and reciprocal
560 (('~fdiv', 1.0, a), ('frcp', a)),
561 (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
562 (('~frcp', ('frcp', a)), a),
563 (('~frcp', ('fsqrt', a)), ('frsq', a)),
564 (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
565 (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
566 # Boolean simplifications
567 (('i2b32(is_used_by_if)', a), ('ine32', a, 0)),
568 (('i2b1(is_used_by_if)', a), ('ine', a, 0)),
569 (('ieq', a, True), a),
570 (('ine(is_not_used_by_if)', a, True), ('inot', a)),
571 (('ine', a, False), a),
572 (('ieq(is_not_used_by_if)', a, False), ('inot', 'a')),
573 (('bcsel', a, True, False), a),
574 (('bcsel', a, False, True), ('inot', a)),
575 (('bcsel@32', a, 1.0, 0.0), ('b2f', a)),
576 (('bcsel@32', a, 0.0, 1.0), ('b2f', ('inot', a))),
577 (('bcsel@32', a, -1.0, -0.0), ('fneg', ('b2f', a))),
578 (('bcsel@32', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))),
579 (('bcsel', True, b, c), b),
580 (('bcsel', False, b, c), c),
581 (('bcsel', a, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a, b, c))),
582 # The result of this should be hit by constant propagation and, in the
583 # next round of opt_algebraic, get picked up by one of the above two.
584 (('bcsel', '#a', b, c), ('bcsel', ('ine', 'a', 0), b, c)),
585
586 (('bcsel', a, b, b), b),
587 (('fcsel', a, b, b), b),
588
589 # D3D Boolean emulation
590 (('bcsel', a, -1, 0), ('ineg', ('b2i', 'a@1'))),
591 (('bcsel', a, 0, -1), ('ineg', ('b2i', ('inot', a)))),
592 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
593 ('ineg', ('b2i', ('iand', a, b)))),
594 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
595 ('ineg', ('b2i', ('ior', a, b)))),
596 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a)),
597 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a),
598 (('ine', ('ineg', ('b2i', 'a@1')), 0), a),
599 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a)),
600 (('iand', ('ineg', ('b2i', a)), 1.0), ('b2f', a)),
601
602 # SM5 32-bit shifts are defined to use the 5 least significant bits
603 (('ishl', 'a@32', ('iand', 31, b)), ('ishl', a, b)),
604 (('ishr', 'a@32', ('iand', 31, b)), ('ishr', a, b)),
605 (('ushr', 'a@32', ('iand', 31, b)), ('ushr', a, b)),
606
607 # Conversions
608 (('i2b32', ('b2i', 'a@32')), a),
609 (('f2i', ('ftrunc', a)), ('f2i', a)),
610 (('f2u', ('ftrunc', a)), ('f2u', a)),
611 (('i2b', ('ineg', a)), ('i2b', a)),
612 (('i2b', ('iabs', a)), ('i2b', a)),
613 (('fabs', ('b2f', a)), ('b2f', a)),
614 (('iabs', ('b2i', a)), ('b2i', a)),
615 (('inot', ('f2b1', a)), ('feq', a, 0.0)),
616
617 # Ironically, mark these as imprecise because removing the conversions may
618 # preserve more precision than doing the conversions (e.g.,
619 # uint(float(0x81818181u)) == 0x81818200).
620 (('~f2i32', ('i2f', 'a@32')), a),
621 (('~f2i32', ('u2f', 'a@32')), a),
622 (('~f2u32', ('i2f', 'a@32')), a),
623 (('~f2u32', ('u2f', 'a@32')), a),
624
625 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
626 # says:
627 #
628 # It is undefined to convert a negative floating-point value to an
629 # uint.
630 #
631 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
632 # some optimizations in the i965 backend to proceed.
633 (('ige', ('f2u', a), b), ('ige', ('f2i', a), b)),
634 (('ige', b, ('f2u', a)), ('ige', b, ('f2i', a))),
635 (('ilt', ('f2u', a), b), ('ilt', ('f2i', a), b)),
636 (('ilt', b, ('f2u', a)), ('ilt', b, ('f2i', a))),
637
638 # Packing and then unpacking does nothing
639 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a),
640 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b),
641 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a),
642 ('unpack_64_2x32_split_y', a)), a),
643
644 # Byte extraction
645 (('ushr', 'a@16', 8), ('extract_u8', a, 1), '!options->lower_extract_byte'),
646 (('ushr', 'a@32', 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
647 (('ushr', 'a@64', 56), ('extract_u8', a, 7), '!options->lower_extract_byte'),
648 (('ishr', 'a@16', 8), ('extract_i8', a, 1), '!options->lower_extract_byte'),
649 (('ishr', 'a@32', 24), ('extract_i8', a, 3), '!options->lower_extract_byte'),
650 (('ishr', 'a@64', 56), ('extract_i8', a, 7), '!options->lower_extract_byte'),
651 (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte')
652 ]
653
654 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
655 # patterns like those below.
656 for op in ('ushr', 'ishr'):
657 optimizations.extend([(('extract_u8', (op, 'a@16', 8), 0), ('extract_u8', a, 1))])
658 optimizations.extend([(('extract_u8', (op, 'a@32', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 4)])
659 optimizations.extend([(('extract_u8', (op, 'a@64', 8 * i), 0), ('extract_u8', a, i)) for i in range(1, 8)])
660
661 optimizations.extend([(('extract_u8', ('extract_u16', a, 1), 0), ('extract_u8', a, 2))])
662
663 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
664 # patterns like those below.
665 for op in ('extract_u8', 'extract_i8'):
666 optimizations.extend([((op, ('ishl', 'a@16', 8), 1), (op, a, 0))])
667 optimizations.extend([((op, ('ishl', 'a@32', 24 - 8 * i), 3), (op, a, i)) for i in range(2, -1, -1)])
668 optimizations.extend([((op, ('ishl', 'a@64', 56 - 8 * i), 7), (op, a, i)) for i in range(6, -1, -1)])
669
670 optimizations.extend([
671 # Word extraction
672 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a, 0), '!options->lower_extract_word'),
673 (('ushr', 'a@32', 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
674 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a, 0), '!options->lower_extract_word'),
675 (('ishr', 'a@32', 16), ('extract_i16', a, 1), '!options->lower_extract_word'),
676 (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),
677
678 # Subtracts
679 (('~fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)),
680 (('isub', a, ('isub', 0, b)), ('iadd', a, b)),
681 (('ussub_4x8', a, 0), a),
682 (('ussub_4x8', a, ~0), 0),
683 (('fsub', a, b), ('fadd', a, ('fneg', b)), 'options->lower_sub'),
684 (('isub', a, b), ('iadd', a, ('ineg', b)), 'options->lower_sub'),
685 (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
686 (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
687 (('~fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)),
688 (('iadd', a, ('isub', 0, b)), ('isub', a, b)),
689 (('fabs', ('fsub', 0.0, a)), ('fabs', a)),
690 (('iabs', ('isub', 0, a)), ('iabs', a)),
691
692 # Propagate negation up multiplication chains
693 (('fmul', ('fneg', a), b), ('fneg', ('fmul', a, b))),
694 (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),
695
696 # Propagate constants up multiplication chains
697 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)),
698 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)),
699 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)),
700 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)),
701
702 # Reassociate constants in add/mul chains so they can be folded together.
703 # For now, we mostly only handle cases where the constants are separated by
704 # a single non-constant. We could do better eventually.
705 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a, c), b)),
706 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a, c), b)),
707 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a, c), b)),
708 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a, ('fneg', c)), ('fneg', b))),
709 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a, c), b)),
710
711 # By definition...
712 (('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)),
713 (('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
714 (('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
715
716 (('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)),
717 (('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
718 (('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
719
720 (('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),
721
722 # Misc. lowering
723 (('fmod@16', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod16'),
724 (('fmod@32', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod32'),
725 (('fmod@64', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod64'),
726 (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod32'),
727 (('uadd_carry@32', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
728 (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),
729
730 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
731 ('bcsel', ('ilt', 31, 'bits'), 'insert',
732 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
733 'options->lower_bitfield_insert'),
734 (('ihadd', a, b), ('iadd', ('iand', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
735 (('uhadd', a, b), ('iadd', ('iand', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
736 (('irhadd', a, b), ('isub', ('ior', a, b), ('ishr', ('ixor', a, b), 1)), 'options->lower_hadd'),
737 (('urhadd', a, b), ('isub', ('ior', a, b), ('ushr', ('ixor', a, b), 1)), 'options->lower_hadd'),
738 (('uadd_sat', a, b), ('bcsel', ('ult', ('iadd', a, b), a), -1, ('iadd', a, b)), 'options->lower_add_sat'),
739 (('usub_sat', a, b), ('bcsel', ('ult', a, b), 0, ('isub', a, b)), 'options->lower_add_sat'),
740
741 # Alternative lowering that doesn't rely on bfi.
742 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
743 ('bcsel', ('ilt', 31, 'bits'),
744 'insert',
745 ('ior',
746 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
747 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
748 'options->lower_bitfield_insert_to_shifts'),
749
750 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
751 (('bfm', 'bits', 'offset'),
752 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
753 'options->lower_bfm'),
754
755 (('ibitfield_extract', 'value', 'offset', 'bits'),
756 ('bcsel', ('ilt', 31, 'bits'), 'value',
757 ('ibfe', 'value', 'offset', 'bits')),
758 'options->lower_bitfield_extract'),
759
760 (('ubitfield_extract', 'value', 'offset', 'bits'),
761 ('bcsel', ('ult', 31, 'bits'), 'value',
762 ('ubfe', 'value', 'offset', 'bits')),
763 'options->lower_bitfield_extract'),
764
765 (('ibitfield_extract', 'value', 'offset', 'bits'),
766 ('bcsel', ('ieq', 0, 'bits'),
767 0,
768 ('ishr',
769 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
770 ('isub', 32, 'bits'))),
771 'options->lower_bitfield_extract_to_shifts'),
772
773 (('ubitfield_extract', 'value', 'offset', 'bits'),
774 ('iand',
775 ('ushr', 'value', 'offset'),
776 ('bcsel', ('ieq', 'bits', 32),
777 0xffffffff,
778 ('bfm', 'bits', 0))),
779 'options->lower_bitfield_extract_to_shifts'),
780
781 (('ifind_msb', 'value'),
782 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
783 'options->lower_ifind_msb'),
784
785 (('find_lsb', 'value'),
786 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
787 'options->lower_find_lsb'),
788
789 (('extract_i8', a, 'b@32'),
790 ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24),
791 'options->lower_extract_byte'),
792
793 (('extract_u8', a, 'b@32'),
794 ('iand', ('ushr', a, ('imul', b, 8)), 0xff),
795 'options->lower_extract_byte'),
796
797 (('extract_i16', a, 'b@32'),
798 ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16),
799 'options->lower_extract_word'),
800
801 (('extract_u16', a, 'b@32'),
802 ('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
803 'options->lower_extract_word'),
804
805 (('pack_unorm_2x16', 'v'),
806 ('pack_uvec2_to_uint',
807 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
808 'options->lower_pack_unorm_2x16'),
809
810 (('pack_unorm_4x8', 'v'),
811 ('pack_uvec4_to_uint',
812 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
813 'options->lower_pack_unorm_4x8'),
814
815 (('pack_snorm_2x16', 'v'),
816 ('pack_uvec2_to_uint',
817 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
818 'options->lower_pack_snorm_2x16'),
819
820 (('pack_snorm_4x8', 'v'),
821 ('pack_uvec4_to_uint',
822 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
823 'options->lower_pack_snorm_4x8'),
824
825 (('unpack_unorm_2x16', 'v'),
826 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
827 ('extract_u16', 'v', 1))),
828 65535.0),
829 'options->lower_unpack_unorm_2x16'),
830
831 (('unpack_unorm_4x8', 'v'),
832 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
833 ('extract_u8', 'v', 1),
834 ('extract_u8', 'v', 2),
835 ('extract_u8', 'v', 3))),
836 255.0),
837 'options->lower_unpack_unorm_4x8'),
838
839 (('unpack_snorm_2x16', 'v'),
840 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
841 ('extract_i16', 'v', 1))),
842 32767.0))),
843 'options->lower_unpack_snorm_2x16'),
844
845 (('unpack_snorm_4x8', 'v'),
846 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
847 ('extract_i8', 'v', 1),
848 ('extract_i8', 'v', 2),
849 ('extract_i8', 'v', 3))),
850 127.0))),
851 'options->lower_unpack_snorm_4x8'),
852
853 (('isign', a), ('imin', ('imax', a, -1), 1), 'options->lower_isign'),
854 (('fsign', a), ('fsub', ('b2f', ('flt', 0.0, a)), ('b2f', ('flt', a, 0.0))), 'options->lower_fsign'),
855 ])
856
857 # bit_size dependent lowerings
858 for bit_size in [8, 16, 32, 64]:
859 # convenience constants
860 intmax = (1 << (bit_size - 1)) - 1
861 intmin = 1 << (bit_size - 1)
862
863 optimizations += [
864 (('iadd_sat@' + str(bit_size), a, b),
865 ('bcsel', ('ige', b, 1), ('bcsel', ('ilt', ('iadd', a, b), a), intmax, ('iadd', a, b)),
866 ('bcsel', ('ilt', a, ('iadd', a, b)), intmin, ('iadd', a, b))), 'options->lower_add_sat'),
867 (('isub_sat@' + str(bit_size), a, b),
868 ('bcsel', ('ilt', b, 0), ('bcsel', ('ilt', ('isub', a, b), a), intmax, ('isub', a, b)),
869 ('bcsel', ('ilt', a, ('isub', a, b)), intmin, ('isub', a, b))), 'options->lower_add_sat'),
870 ]
871
872 invert = OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
873
874 for left, right in itertools.combinations_with_replacement(invert.keys(), 2):
875 optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))),
876 ('iand', (invert[left], a, b), (invert[right], c, d))))
877 optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))),
878 ('ior', (invert[left], a, b), (invert[right], c, d))))
879
880 # Optimize x2bN(b2x(x)) -> x
881 for size in type_sizes('bool'):
882 aN = 'a@' + str(size)
883 f2bN = 'f2b' + str(size)
884 i2bN = 'i2b' + str(size)
885 optimizations.append(((f2bN, ('b2f', aN)), a))
886 optimizations.append(((i2bN, ('b2i', aN)), a))
887
888 # Optimize x2yN(b2x(x)) -> b2y
889 for x, y in itertools.product(['f', 'u', 'i'], ['f', 'u', 'i']):
890 if x != 'f' and y != 'f' and x != y:
891 continue
892
893 b2x = 'b2f' if x == 'f' else 'b2i'
894 b2y = 'b2f' if y == 'f' else 'b2i'
895 x2yN = '{}2{}'.format(x, y)
896 optimizations.append(((x2yN, (b2x, a)), (b2y, a)))
897
898 # Optimize away x2xN(a@N)
899 for t in ['int', 'uint', 'float']:
900 for N in type_sizes(t):
901 x2xN = '{0}2{0}{1}'.format(t[0], N)
902 aN = 'a@{0}'.format(N)
903 optimizations.append(((x2xN, aN), a))
904
905 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
906 # In particular, we can optimize away everything except upcast of downcast and
907 # upcasts where the type differs from the other cast
908 for N, M in itertools.product(type_sizes('uint'), type_sizes('uint')):
909 if N < M:
910 # The outer cast is a down-cast. It doesn't matter what the size of the
911 # argument of the inner cast is because we'll never been in the upcast
912 # of downcast case. Regardless of types, we'll always end up with y2yN
913 # in the end.
914 for x, y in itertools.product(['i', 'u'], ['i', 'u']):
915 x2xN = '{0}2{0}{1}'.format(x, N)
916 y2yM = '{0}2{0}{1}'.format(y, M)
917 y2yN = '{0}2{0}{1}'.format(y, N)
918 optimizations.append(((x2xN, (y2yM, a)), (y2yN, a)))
919 elif N > M:
920 # If the outer cast is an up-cast, we have to be more careful about the
921 # size of the argument of the inner cast and with types. In this case,
922 # the type is always the type of type up-cast which is given by the
923 # outer cast.
924 for P in type_sizes('uint'):
925 # We can't optimize away up-cast of down-cast.
926 if M < P:
927 continue
928
929 # Because we're doing down-cast of down-cast, the types always have
930 # to match between the two casts
931 for x in ['i', 'u']:
932 x2xN = '{0}2{0}{1}'.format(x, N)
933 x2xM = '{0}2{0}{1}'.format(x, M)
934 aP = 'a@{0}'.format(P)
935 optimizations.append(((x2xN, (x2xM, aP)), (x2xN, a)))
936 else:
937 # The N == M case is handled by other optimizations
938 pass
939
940 def fexp2i(exp, bits):
941 # We assume that exp is already in the right range.
942 if bits == 16:
943 return ('i2i16', ('ishl', ('iadd', exp, 15), 10))
944 elif bits == 32:
945 return ('ishl', ('iadd', exp, 127), 23)
946 elif bits == 64:
947 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
948 else:
949 assert False
950
951 def ldexp(f, exp, bits):
952 # First, we clamp exp to a reasonable range. The maximum possible range
953 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
954 # a maximum range of [-149, 127]. This means that we can potentially have
955 # a swing of +-276. If you start with FLT_MAX, you actually have to do
956 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
957 # spec, on the other hand, only requires that we handle an exponent value
958 # in the range [-126, 128]. This implementation is *mostly* correct; it
959 # handles a range on exp of [-252, 254] which allows you to create any
960 # value (including denorms if the hardware supports it) and to adjust the
961 # exponent of any normal value to anything you want.
962 if bits == 16:
963 exp = ('imin', ('imax', exp, -28), 30)
964 elif bits == 32:
965 exp = ('imin', ('imax', exp, -252), 254)
966 elif bits == 64:
967 exp = ('imin', ('imax', exp, -2044), 2046)
968 else:
969 assert False
970
971 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
972 # (We use ishr which isn't the same for -1, but the -1 case still works
973 # since we use exp-exp/2 as the second exponent.) While the spec
974 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
975 # work with denormals and doesn't allow for the full swing in exponents
976 # that you can get with normalized values. Instead, we create two powers
977 # of two and multiply by them each in turn. That way the effective range
978 # of our exponent is doubled.
979 pow2_1 = fexp2i(('ishr', exp, 1), bits)
980 pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
981 return ('fmul', ('fmul', f, pow2_1), pow2_2)
982
983 optimizations += [
984 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
985 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
986 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
987 ]
988
989 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
990 def bitfield_reverse(u):
991 step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))
992 step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8))
993 step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4))
994 step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2))
995 step5 = ('ior', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1))
996
997 return step5
998
999 optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
1000
1001 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1002 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1003 # and, if a is a NaN then the second comparison will fail anyway.
1004 for op in ['flt', 'fge', 'feq']:
1005 optimizations += [
1006 (('iand', ('feq', a, a), (op, a, b)), (op, a, b)),
1007 (('iand', ('feq', a, a), (op, b, a)), (op, b, a)),
1008 ]
1009
1010 # Add optimizations to handle the case where the result of a ternary is
1011 # compared to a constant. This way we can take things like
1012 #
1013 # (a ? 0 : 1) > 0
1014 #
1015 # and turn it into
1016 #
1017 # a ? (0 > 0) : (1 > 0)
1018 #
1019 # which constant folding will eat for lunch. The resulting ternary will
1020 # further get cleaned up by the boolean reductions above and we will be
1021 # left with just the original variable "a".
1022 for op in ['flt', 'fge', 'feq', 'fne',
1023 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1024 optimizations += [
1025 ((op, ('bcsel', 'a', '#b', '#c'), '#d'),
1026 ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))),
1027 ((op, '#d', ('bcsel', a, '#b', '#c')),
1028 ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
1029 ]
1030
1031
1032 # For example, this converts things like
1033 #
1034 # 1 + mix(0, a - 1, condition)
1035 #
1036 # into
1037 #
1038 # mix(1, (a-1)+1, condition)
1039 #
1040 # Other optimizations will rearrange the constants.
1041 for op in ['fadd', 'fmul', 'iadd', 'imul']:
1042 optimizations += [
1043 ((op, ('bcsel(is_used_once)', a, '#b', c), '#d'), ('bcsel', a, (op, b, d), (op, c, d)))
1044 ]
1045
1046 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1047 # states:
1048 #
1049 # If neither layout qualifier is specified, derivatives in compute shaders
1050 # return zero, which is consistent with the handling of built-in texture
1051 # functions like texture() in GLSL 4.50 compute shaders.
1052 for op in ['fddx', 'fddx_fine', 'fddx_coarse',
1053 'fddy', 'fddy_fine', 'fddy_coarse']:
1054 optimizations += [
1055 ((op, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1056 ]
1057
1058 # This section contains "late" optimizations that should be run before
1059 # creating ffmas and calling regular optimizations for the final time.
1060 # Optimizations should go here if they help code generation and conflict
1061 # with the regular optimizations.
1062 before_ffma_optimizations = [
1063 # Propagate constants down multiplication chains
1064 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)),
1065 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)),
1066 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)),
1067 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)),
1068
1069 (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
1070 (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
1071 (('~fadd', ('fneg', a), a), 0.0),
1072 (('iadd', ('ineg', a), a), 0),
1073 (('iadd', ('ineg', a), ('iadd', a, b)), b),
1074 (('iadd', a, ('iadd', ('ineg', a), b)), b),
1075 (('~fadd', ('fneg', a), ('fadd', a, b)), b),
1076 (('~fadd', a, ('fadd', ('fneg', a), b)), b),
1077
1078 (('~flrp@32', ('fadd(is_used_once)', a, -1.0), ('fadd(is_used_once)', a, 1.0), d), ('fadd', ('flrp', -1.0, 1.0, d), a)),
1079 (('~flrp@32', ('fadd(is_used_once)', a, 1.0), ('fadd(is_used_once)', a, -1.0), d), ('fadd', ('flrp', 1.0, -1.0, d), a)),
1080 (('~flrp@32', ('fadd(is_used_once)', a, '#b'), ('fadd(is_used_once)', a, '#c'), d), ('fadd', ('fmul', d, ('fadd', c, ('fneg', b))), ('fadd', a, b))),
1081 ]
1082
1083 # This section contains "late" optimizations that should be run after the
1084 # regular optimizations have finished. Optimizations should go here if
1085 # they help code generation but do not necessarily produce code that is
1086 # more easily optimizable.
1087 late_optimizations = [
1088 # Most of these optimizations aren't quite safe when you get infinity or
1089 # Nan involved but the first one should be fine.
1090 (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
1091 (('flt', ('fneg', ('fadd', a, b)), 0.0), ('flt', ('fneg', a), b)),
1092 (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
1093 (('~fge', ('fneg', ('fadd', a, b)), 0.0), ('fge', ('fneg', a), b)),
1094 (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
1095 (('~fne', ('fadd', a, b), 0.0), ('fne', a, ('fneg', b))),
1096
1097 (('~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)))),
1098
1099 (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'),
1100 (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'),
1101 (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
1102 (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
1103
1104 (('~flrp@32', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
1105 (('~flrp@64', ('fadd(is_used_once)', a, b), ('fadd(is_used_once)', a, c), d), ('fadd', ('flrp', b, c, d), a)),
1106
1107 (('b2f(is_used_more_than_once)', ('inot', 'a@1')), ('bcsel', a, 0.0, 1.0)),
1108 (('fneg(is_used_more_than_once)', ('b2f', ('inot', 'a@1'))), ('bcsel', a, -0.0, -1.0)),
1109
1110 # we do these late so that we don't get in the way of creating ffmas
1111 (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
1112 (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),
1113
1114 (('bcsel', a, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a, b)))),
1115 ]
1116
1117 print(nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render())
1118 print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_before_ffma",
1119 before_ffma_optimizations).render())
1120 print(nir_algebraic.AlgebraicPass("nir_opt_algebraic_late",
1121 late_optimizations).render())