2 # Copyright (C) 2014 Intel Corporation
4 # Permission is hereby granted, free of charge, to any person obtaining a
5 # copy of this software and associated documentation files (the "Software"),
6 # to deal in the Software without restriction, including without limitation
7 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 # and/or sell copies of the Software, and to permit persons to whom the
9 # Software is furnished to do so, subject to the following conditions:
11 # The above copyright notice and this permission notice (including the next
12 # paragraph) shall be included in all copies or substantial portions of the
15 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 # Jason Ekstrand (jason@jlekstrand.net)
29 # Convenience variables
35 # Written in the form (<search>, <replace>) where <search> is an expression
36 # and <replace> is either an expression or a value. An expression is
37 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
38 # where each source is either an expression or a value. A value can be
39 # either a numeric constant or a string representing a variable name.
41 # If the opcode in a search expression is prefixed by a '~' character, this
42 # indicates that the operation is inexact. Such operations will only get
43 # applied to SSA values that do not have the exact bit set. This should be
44 # used by by any optimizations that are not bit-for-bit exact. It should not,
45 # however, be used for backend-requested lowering operations as those need to
46 # happen regardless of precision.
48 # Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
49 # that the given variable will only match constants and the type indicates that
50 # the given variable will only match values from ALU instructions with the
51 # given output type, and (cond) specifies an additional condition function
52 # (see nir_search_helpers.h).
54 # For constants, you have to be careful to make sure that it is the right
55 # type because python is unaware of the source and destination types of the
58 # All expression types can have a bit-size specified. For opcodes, this
59 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
60 # type and size, and for literals, you can write "2.0@32". In the search half
61 # of the expression this indicates that it should only match that particular
62 # bit-size. In the replace half of the expression this indicates that the
63 # constructed value should have that bit-size.
67 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
))),
68 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
))))),
73 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
))),
74 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
75 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
76 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
78 (('fneg', ('fneg', a
)), a
),
79 (('ineg', ('ineg', a
)), a
),
80 (('fabs', ('fabs', a
)), ('fabs', a
)),
81 (('fabs', ('fneg', a
)), ('fabs', a
)),
82 (('fabs', ('u2f32', a
)), ('u2f32', a
)),
83 (('iabs', ('iabs', a
)), ('iabs', a
)),
84 (('iabs', ('ineg', a
)), ('iabs', a
)),
85 (('~fadd', a
, 0.0), a
),
87 (('usadd_4x8', a
, 0), a
),
88 (('usadd_4x8', a
, ~
0), ~
0),
89 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
90 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
91 (('~fadd', ('fneg', a
), a
), 0.0),
92 (('iadd', ('ineg', a
), a
), 0),
93 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
94 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
95 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
96 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
97 (('~fmul', a
, 0.0), 0.0),
99 (('umul_unorm_4x8', a
, 0), 0),
100 (('umul_unorm_4x8', a
, ~
0), a
),
101 (('fmul', a
, 1.0), a
),
103 (('fmul', a
, -1.0), ('fneg', a
)),
104 (('imul', a
, -1), ('ineg', a
)),
105 (('~ffma', 0.0, a
, b
), b
),
106 (('~ffma', a
, 0.0, b
), b
),
107 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
108 (('ffma', a
, 1.0, b
), ('fadd', a
, b
)),
109 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
110 (('~flrp', a
, b
, 0.0), a
),
111 (('~flrp', a
, b
, 1.0), b
),
112 (('~flrp', a
, a
, b
), a
),
113 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
114 (('~flrp', a
, b
, ('b2f', c
)), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
115 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
116 (('flrp@32', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp32'),
117 (('flrp@64', a
, b
, c
), ('fadd', ('fmul', c
, ('fsub', b
, a
)), a
), 'options->lower_flrp64'),
118 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
119 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', c
)))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
120 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
121 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
))), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
122 (('~fadd', a
, ('fmul', ('b2f', c
), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
123 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
124 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
125 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
126 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
128 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
)),
129 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
130 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
131 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
133 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
134 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
136 # (a * #b + #c) << #d
137 # ((a * #b) << #d) + (#c << #d)
138 # (a * (#b << #d)) + (#c << #d)
139 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
140 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
144 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
146 # Comparison simplifications
147 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
148 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
149 (('~inot', ('feq', a
, b
)), ('fne', a
, b
)),
150 (('~inot', ('fne', a
, b
)), ('feq', a
, b
)),
151 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
152 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
153 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
154 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
155 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
156 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
160 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
162 (('fge', 0.0, ('b2f', a
)), ('inot', a
)),
164 (('fge', ('fneg', ('b2f', a
)), 0.0), ('inot', a
)),
166 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
167 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
168 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
169 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
171 # Cannot remove the addition from ilt or ige due to overflow.
172 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
173 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
175 # fmin(-b2f(a), b) >= 0.0
176 # -b2f(a) >= 0.0 && b >= 0.0
177 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
178 # b2f(a) == 0.0 && b >= 0.0
179 # a == False && b >= 0.0
182 # The fge in the second replacement is not a typo. I leave the proof that
183 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
185 (('fge', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
186 (('feq', ('fmin', ('fneg', ('b2f', a
)), b
), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
188 (('feq', ('b2f', a
), 0.0), ('inot', a
)),
189 (('fne', ('b2f', a
), 0.0), a
),
190 (('ieq', ('b2i', a
), 0), ('inot', a
)),
191 (('ine', ('b2i', a
), 0), a
),
195 # fabs(a) != 0.0 because fabs(a) must be >= 0
197 (('flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
199 (('fmax', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('ior', a
, b
))),
200 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('ior', a
, b
)))),
201 (('fmin', ('b2f(is_used_once)', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
202 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', a
)), ('fneg', ('b2f', b
))), ('fneg', ('b2f', ('iand', a
, b
)))),
205 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
206 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
207 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
209 # Since b is a constant, constant folding will eliminate the fmin and the
210 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
211 (('fmin', ('b2f', a
), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
213 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
215 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
216 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
217 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
218 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
219 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
220 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
221 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
222 (('bcsel', a
, True, 'b@bool'), ('ior', a
, b
)),
229 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
230 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
231 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
232 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
233 (('fmin', a
, ('fabs', a
)), a
),
234 (('imin', a
, ('iabs', a
)), a
),
235 (('fmax', a
, ('fneg', ('fabs', a
))), a
),
236 (('imax', a
, ('ineg', ('iabs', a
))), a
),
237 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
238 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
239 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
240 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
241 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
242 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
243 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
244 (('fsat', ('fsat', a
)), ('fsat', a
)),
245 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
246 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
247 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
248 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
249 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
250 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
251 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
252 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
253 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
254 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
255 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
256 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
257 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
258 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
259 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
260 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
261 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
262 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
263 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
264 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
265 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
266 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
268 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
269 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
270 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
271 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
272 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
273 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
274 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
275 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
276 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
277 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
278 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
279 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
280 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
281 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
282 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
283 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
285 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0)),
287 # These patterns can result when (a < b || a < c) => (a < min(b, c))
288 # transformations occur before constant propagation and loop-unrolling.
289 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
290 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
291 (('~fge', a
, ('fmin', b
, a
)), True),
292 (('~fge', ('fmax', a
, b
), a
), True),
293 (('~flt', a
, ('fmin', b
, a
)), False),
294 (('~flt', ('fmax', a
, b
), a
), False),
295 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
296 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
298 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
299 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
300 (('ige', a
, ('imin', b
, a
)), True),
301 (('ige', ('imax', a
, b
), a
), True),
302 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
303 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
304 (('uge', a
, ('umin', b
, a
)), True),
305 (('uge', ('umax', a
, b
), a
), True),
306 (('ilt', a
, ('imin', b
, a
)), False),
307 (('ilt', ('imax', a
, b
), a
), False),
308 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
309 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
310 (('ult', a
, ('umin', b
, a
)), False),
311 (('ult', ('umax', a
, b
), a
), False),
312 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
313 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
315 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
316 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
317 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
318 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
319 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
320 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
321 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
322 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
323 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
324 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
325 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
326 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
327 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
328 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
329 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
330 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
332 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
333 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
334 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
335 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
336 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
337 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
338 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
339 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
340 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
341 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
343 (('imul', ('b2i', a
), ('b2i', b
)), ('b2i', ('iand', a
, b
))),
344 (('fmul', ('b2f', a
), ('b2f', b
)), ('b2f', ('iand', a
, b
))),
345 (('fsat', ('fadd', ('b2f', a
), ('b2f', b
))), ('b2f', ('ior', a
, b
))),
346 (('iand', 'a@bool', 1.0), ('b2f', a
), '!options->lower_b2f'),
347 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
348 (('ineg', ('b2i@32', a
)), a
),
349 (('flt', ('fneg', ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
350 (('flt', ('fsub', 0.0, ('b2f', a
)), 0), a
), # Generated by TGSI KILL_IF.
351 # Comparison with the same args. Note that these are not done for
352 # the float versions because NaN always returns false on float
354 (('ilt', a
, a
), False),
355 (('ige', a
, a
), True),
356 (('ieq', a
, a
), True),
357 (('ine', a
, a
), False),
358 (('ult', a
, a
), False),
359 (('uge', a
, a
), True),
360 # Logical and bit operations
361 (('fand', a
, 0.0), 0.0),
363 (('iand', a
, ~
0), a
),
367 (('ior', a
, True), True),
368 (('fxor', a
, a
), 0.0),
371 (('inot', ('inot', a
)), a
),
373 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
374 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
375 # Shift optimizations
382 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
383 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
384 # Exponential/logarithmic identities
385 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
386 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
387 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
388 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
389 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
390 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
391 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
392 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
393 (('~fpow', a
, 1.0), a
),
394 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
395 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
396 (('~fpow', 2.0, a
), ('fexp2', a
)),
397 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
398 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
399 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
400 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
401 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
402 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
403 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
404 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
405 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
406 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
407 # Division and reciprocal
408 (('~fdiv', 1.0, a
), ('frcp', a
)),
409 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
410 (('~frcp', ('frcp', a
)), a
),
411 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
412 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
413 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
414 # Boolean simplifications
415 (('ieq', 'a@bool', True), a
),
416 (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a
)),
417 (('ine', 'a@bool', False), a
),
418 (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
419 (('bcsel', a
, True, False), a
),
420 (('bcsel', a
, False, True), ('inot', a
)),
421 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
422 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
423 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
424 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
425 (('bcsel', True, b
, c
), b
),
426 (('bcsel', False, b
, c
), c
),
427 (('bcsel', a
, ('b2f(is_used_once)', b
), ('b2f', c
)), ('b2f', ('bcsel', a
, b
, c
))),
428 # The result of this should be hit by constant propagation and, in the
429 # next round of opt_algebraic, get picked up by one of the above two.
430 (('bcsel', '#a', b
, c
), ('bcsel', ('ine', 'a', 0), b
, c
)),
432 (('bcsel', a
, b
, b
), b
),
433 (('fcsel', a
, b
, b
), b
),
436 (('i2b', ('b2i', a
)), a
),
437 (('i2b', 'a@bool'), a
),
438 (('f2i32', ('ftrunc', a
)), ('f2i32', a
)),
439 (('f2u32', ('ftrunc', a
)), ('f2u32', a
)),
440 (('i2b', ('ineg', a
)), ('i2b', a
)),
441 (('i2b', ('iabs', a
)), ('i2b', a
)),
442 (('fabs', ('b2f', a
)), ('b2f', a
)),
443 (('iabs', ('b2i', a
)), ('b2i', a
)),
444 (('inot', ('f2b', a
)), ('feq', a
, 0.0)),
446 # Packing and then unpacking does nothing
447 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
448 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
449 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
450 ('unpack_64_2x32_split_y', a
)), a
),
453 (('ushr', a
, 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
454 (('iand', 0xff, ('ushr', a
, 16)), ('extract_u8', a
, 2), '!options->lower_extract_byte'),
455 (('iand', 0xff, ('ushr', a
, 8)), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
456 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
459 (('ushr', a
, 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
460 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
463 (('~fsub', a
, ('fsub', 0.0, b
)), ('fadd', a
, b
)),
464 (('isub', a
, ('isub', 0, b
)), ('iadd', a
, b
)),
465 (('ussub_4x8', a
, 0), a
),
466 (('ussub_4x8', a
, ~
0), 0),
467 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
)), 'options->lower_sub'),
468 (('isub', a
, b
), ('iadd', a
, ('ineg', b
)), 'options->lower_sub'),
469 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
470 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
471 (('~fadd', a
, ('fsub', 0.0, b
)), ('fsub', a
, b
)),
472 (('iadd', a
, ('isub', 0, b
)), ('isub', a
, b
)),
473 (('fabs', ('fsub', 0.0, a
)), ('fabs', a
)),
474 (('iabs', ('isub', 0, a
)), ('iabs', a
)),
476 # Propagate negation up multiplication chains
477 (('fmul', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
478 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
480 # Propagate constants up multiplication chains
481 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
482 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
483 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
484 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
486 # Reassociate constants in add/mul chains so they can be folded together.
487 # For now, we mostly only handle cases where the constants are separated by
488 # a single non-constant. We could do better eventually.
489 (('~fmul', '#a', ('fmul', b
, '#c')), ('fmul', ('fmul', a
, c
), b
)),
490 (('imul', '#a', ('imul', b
, '#c')), ('imul', ('imul', a
, c
), b
)),
491 (('~fadd', '#a', ('fadd', b
, '#c')), ('fadd', ('fadd', a
, c
), b
)),
492 (('~fadd', '#a', ('fneg', ('fadd', b
, '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
493 (('iadd', '#a', ('iadd', b
, '#c')), ('iadd', ('iadd', a
, c
), b
)),
496 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
497 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
498 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
500 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
501 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
502 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
504 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
507 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
508 (('fmod@64', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod64'),
509 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod32'),
510 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
511 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
513 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
514 ('bcsel', ('ilt', 31, 'bits'), 'insert',
515 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
516 'options->lower_bitfield_insert'),
518 # Alternative lowering that doesn't rely on bfi.
519 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
520 ('bcsel', ('ilt', 31, 'bits'),
523 ('iand', 'base', ('inot', ('bfm', 'bits', 'offset'))),
524 ('iand', ('ishl', 'insert', 'offset'), ('bfm', 'bits', 'offset')))),
525 'options->lower_bitfield_insert_to_shifts'),
527 # bfm lowering -- note that the NIR opcode is undefined if either arg is 32.
528 (('bfm', 'bits', 'offset'),
529 ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'),
530 'options->lower_bfm'),
532 (('ibitfield_extract', 'value', 'offset', 'bits'),
533 ('bcsel', ('ilt', 31, 'bits'), 'value',
534 ('ibfe', 'value', 'offset', 'bits')),
535 'options->lower_bitfield_extract'),
537 (('ubitfield_extract', 'value', 'offset', 'bits'),
538 ('bcsel', ('ult', 31, 'bits'), 'value',
539 ('ubfe', 'value', 'offset', 'bits')),
540 'options->lower_bitfield_extract'),
542 (('ibitfield_extract', 'value', 'offset', 'bits'),
543 ('bcsel', ('ieq', 0, 'bits'),
546 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
547 ('isub', 32, 'bits'))),
548 'options->lower_bitfield_extract_to_shifts'),
550 (('ubitfield_extract', 'value', 'offset', 'bits'),
552 ('ushr', 'value', 'offset'),
553 ('bcsel', ('ieq', 'bits', 32),
555 ('bfm', 'bits', 0))),
556 'options->lower_bitfield_extract_to_shifts'),
558 (('ifind_msb', 'value'),
559 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
560 'options->lower_ifind_msb'),
562 (('extract_i8', a
, 'b@32'),
563 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
564 'options->lower_extract_byte'),
566 (('extract_u8', a
, 'b@32'),
567 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
568 'options->lower_extract_byte'),
570 (('extract_i16', a
, 'b@32'),
571 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
572 'options->lower_extract_word'),
574 (('extract_u16', a
, 'b@32'),
575 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
576 'options->lower_extract_word'),
578 (('pack_unorm_2x16', 'v'),
579 ('pack_uvec2_to_uint',
580 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
581 'options->lower_pack_unorm_2x16'),
583 (('pack_unorm_4x8', 'v'),
584 ('pack_uvec4_to_uint',
585 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
586 'options->lower_pack_unorm_4x8'),
588 (('pack_snorm_2x16', 'v'),
589 ('pack_uvec2_to_uint',
590 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
591 'options->lower_pack_snorm_2x16'),
593 (('pack_snorm_4x8', 'v'),
594 ('pack_uvec4_to_uint',
595 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
596 'options->lower_pack_snorm_4x8'),
598 (('unpack_unorm_2x16', 'v'),
599 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
600 ('extract_u16', 'v', 1))),
602 'options->lower_unpack_unorm_2x16'),
604 (('unpack_unorm_4x8', 'v'),
605 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
606 ('extract_u8', 'v', 1),
607 ('extract_u8', 'v', 2),
608 ('extract_u8', 'v', 3))),
610 'options->lower_unpack_unorm_4x8'),
612 (('unpack_snorm_2x16', 'v'),
613 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec2', ('extract_i16', 'v', 0),
614 ('extract_i16', 'v', 1))),
616 'options->lower_unpack_snorm_2x16'),
618 (('unpack_snorm_4x8', 'v'),
619 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec4', ('extract_i8', 'v', 0),
620 ('extract_i8', 'v', 1),
621 ('extract_i8', 'v', 2),
622 ('extract_i8', 'v', 3))),
624 'options->lower_unpack_snorm_4x8'),
627 invert
= {'feq': 'fne', 'fne': 'feq', 'fge': 'flt', 'flt': 'fge' }
629 for left
, right
in list(itertools
.combinations(invert
.keys(), 2)) + zip(invert
.keys(), invert
.keys()):
630 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
631 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
632 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
633 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
635 def fexp2i(exp
, bits
):
636 # We assume that exp is already in the right range.
638 return ('ishl', ('iadd', exp
, 127), 23)
640 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
644 def ldexp(f
, exp
, bits
):
645 # First, we clamp exp to a reasonable range. The maximum possible range
646 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
647 # a maximum range of [-149, 127]. This means that we can potentially have
648 # a swing of +-276. If you start with FLT_MAX, you actually have to do
649 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
650 # spec, on the other hand, only requires that we handle an exponent value
651 # in the range [-126, 128]. This implementation is *mostly* correct; it
652 # handles a range on exp of [-252, 254] which allows you to create any
653 # value (including denorms if the hardware supports it) and to adjust the
654 # exponent of any normal value to anything you want.
656 exp
= ('imin', ('imax', exp
, -252), 254)
658 exp
= ('imin', ('imax', exp
, -2044), 2046)
662 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
663 # (We use ishr which isn't the same for -1, but the -1 case still works
664 # since we use exp-exp/2 as the second exponent.) While the spec
665 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
666 # work with denormals and doesn't allow for the full swing in exponents
667 # that you can get with normalized values. Instead, we create two powers
668 # of two and multiply by them each in turn. That way the effective range
669 # of our exponent is doubled.
670 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
671 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
672 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
675 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
676 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
679 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
680 def bitfield_reverse(u
):
681 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
682 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
683 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
684 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
685 step5
= ('ior', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
689 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
691 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
692 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
693 # and, if a is a NaN then the second comparison will fail anyway.
694 for op
in ['flt', 'fge', 'feq']:
696 (('iand', ('feq', a
, a
), (op
, a
, b
)), (op
, a
, b
)),
697 (('iand', ('feq', a
, a
), (op
, b
, a
)), (op
, b
, a
)),
700 # Add optimizations to handle the case where the result of a ternary is
701 # compared to a constant. This way we can take things like
707 # a ? (0 > 0) : (1 > 0)
709 # which constant folding will eat for lunch. The resulting ternary will
710 # further get cleaned up by the boolean reductions above and we will be
711 # left with just the original variable "a".
712 for op
in ['flt', 'fge', 'feq', 'fne',
713 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
715 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
716 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
717 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
718 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
722 # For example, this converts things like
724 # 1 + mix(0, a - 1, condition)
728 # mix(1, (a-1)+1, condition)
730 # Other optimizations will rearrange the constants.
731 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
733 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
736 # This section contains "late" optimizations that should be run before
737 # creating ffmas and calling regular optimizations for the final time.
738 # Optimizations should go here if they help code generation and conflict
739 # with the regular optimizations.
740 before_ffma_optimizations
= [
741 # Propagate constants down multiplication chains
742 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
743 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
744 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
745 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
747 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
748 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
749 (('~fadd', ('fneg', a
), a
), 0.0),
750 (('iadd', ('ineg', a
), a
), 0),
751 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
752 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
753 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
754 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
757 # This section contains "late" optimizations that should be run after the
758 # regular optimizations have finished. Optimizations should go here if
759 # they help code generation but do not necessarily produce code that is
760 # more easily optimizable.
761 late_optimizations
= [
762 # Most of these optimizations aren't quite safe when you get infinity or
763 # Nan involved but the first one should be fine.
764 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
765 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
766 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
767 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
768 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
769 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
771 (('~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
)))),
773 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
774 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
775 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
776 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
778 (('b2f(is_used_more_than_once)', ('inot', a
)), ('bcsel', a
, 0.0, 1.0)),
779 (('fneg(is_used_more_than_once)', ('b2f', ('inot', a
))), ('bcsel', a
, -0.0, -1.0)),
781 # we do these late so that we don't get in the way of creating ffmas
782 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
783 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
785 # Lowered for backends without a dedicated b2f instruction
786 (('b2f@32', a
), ('iand', a
, 1.0), 'options->lower_b2f'),
789 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render()
790 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
791 before_ffma_optimizations
).render()
792 print nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
793 late_optimizations
).render()