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)
26 from __future__
import print_function
28 from collections
import OrderedDict
30 from nir_opcodes
import type_sizes
35 # Convenience variables
42 # Written in the form (<search>, <replace>) where <search> is an expression
43 # and <replace> is either an expression or a value. An expression is
44 # defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
45 # where each source is either an expression or a value. A value can be
46 # either a numeric constant or a string representing a variable name.
48 # If the opcode in a search expression is prefixed by a '~' character, this
49 # indicates that the operation is inexact. Such operations will only get
50 # applied to SSA values that do not have the exact bit set. This should be
51 # used by by any optimizations that are not bit-for-bit exact. It should not,
52 # however, be used for backend-requested lowering operations as those need to
53 # happen regardless of precision.
55 # Variable names are specified as "[#]name[@type][(cond)][.swiz]" where:
56 # "#" indicates that the given variable will only match constants,
57 # type indicates that the given variable will only match values from ALU
58 # instructions with the given output type,
59 # (cond) specifies an additional condition function (see nir_search_helpers.h),
60 # swiz is a swizzle applied to the variable (only in the <replace> expression)
62 # For constants, you have to be careful to make sure that it is the right
63 # type because python is unaware of the source and destination types of the
66 # All expression types can have a bit-size specified. For opcodes, this
67 # looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
68 # type and size. In the search half of the expression this indicates that it
69 # should only match that particular bit-size. In the replace half of the
70 # expression this indicates that the constructed value should have that
73 # If the opcode in a replacement expression is prefixed by a '!' character,
74 # this indicated that the new expression will be marked exact.
76 # A special condition "many-comm-expr" can be used with expressions to note
77 # that the expression and its subexpressions have more commutative expressions
78 # than nir_replace_instr can handle. If this special condition is needed with
79 # another condition, the two can be separated by a comma (e.g.,
80 # "(many-comm-expr,is_used_once)").
82 # based on https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
83 def lowered_sincos(c
):
84 x
= ('fsub', ('fmul', 2.0, ('ffract', ('fadd', ('fmul', 0.5 / pi
, a
), c
))), 1.0)
85 x
= ('fmul', ('fsub', x
, ('fmul', x
, ('fabs', x
))), 4.0)
86 return ('ffma', ('ffma', x
, ('fabs', x
), ('fneg', x
)), 0.225, x
)
88 def intBitsToFloat(i
):
89 return struct
.unpack('!f', struct
.pack('!I', i
))[0]
93 (('imul', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitops'),
94 (('imul', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
95 (('ishl', a
, '#b@32'), ('imul', a
, ('ishl', 1, b
)), 'options->lower_bitops'),
97 (('unpack_64_2x32_split_x', ('imul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
98 (('unpack_64_2x32_split_x', ('umul_2x32_64(is_used_once)', a
, b
)), ('imul', a
, b
)),
99 (('imul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('imul_high', a
, b
)), 'options->lower_mul_2x32_64'),
100 (('umul_2x32_64', a
, b
), ('pack_64_2x32_split', ('imul', a
, b
), ('umul_high', a
, b
)), 'options->lower_mul_2x32_64'),
105 (('udiv', a
, '#b@32(is_pos_power_of_two)'), ('ushr', a
, ('find_lsb', b
)), '!options->lower_bitops'),
106 (('idiv', a
, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', b
))), 'options->lower_idiv'),
107 (('idiv', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a
), ('ushr', ('iabs', a
), ('find_lsb', ('iabs', b
))))), 'options->lower_idiv'),
108 (('umod', a
, '#b(is_pos_power_of_two)'), ('iand', a
, ('isub', b
, 1))),
110 (('~fneg', ('fneg', a
)), a
),
111 (('ineg', ('ineg', a
)), a
),
112 (('fabs', ('fabs', a
)), ('fabs', a
)),
113 (('fabs', ('fneg', a
)), ('fabs', a
)),
114 (('fabs', ('u2f', a
)), ('u2f', a
)),
115 (('iabs', ('iabs', a
)), ('iabs', a
)),
116 (('iabs', ('ineg', a
)), ('iabs', a
)),
117 (('f2b', ('fneg', a
)), ('f2b', a
)),
118 (('i2b', ('ineg', a
)), ('i2b', a
)),
119 (('~fadd', a
, 0.0), a
),
121 (('usadd_4x8', a
, 0), a
),
122 (('usadd_4x8', a
, ~
0), ~
0),
123 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
124 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
125 (('~fadd', ('fneg', a
), a
), 0.0),
126 (('iadd', ('ineg', a
), a
), 0),
127 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
128 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
129 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
130 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
131 (('fadd', ('fsat', a
), ('fsat', ('fneg', a
))), ('fsat', ('fabs', a
))),
132 (('~fmul', a
, 0.0), 0.0),
134 (('umul_unorm_4x8', a
, 0), 0),
135 (('umul_unorm_4x8', a
, ~
0), a
),
136 (('~fmul', a
, 1.0), a
),
138 (('fmul', a
, -1.0), ('fneg', a
)),
139 (('imul', a
, -1), ('ineg', a
)),
140 # If a < 0: fsign(a)*a*a => -1*a*a => -a*a => abs(a)*a
141 # If a > 0: fsign(a)*a*a => 1*a*a => a*a => abs(a)*a
142 # If a == 0: fsign(a)*a*a => 0*0*0 => abs(0)*0
143 (('fmul', ('fsign', a
), ('fmul', a
, a
)), ('fmul', ('fabs', a
), a
)),
144 (('fmul', ('fmul', ('fsign', a
), a
), a
), ('fmul', ('fabs', a
), a
)),
145 (('~ffma', 0.0, a
, b
), b
),
146 (('~ffma', a
, b
, 0.0), ('fmul', a
, b
)),
147 (('ffma', 1.0, a
, b
), ('fadd', a
, b
)),
148 (('ffma', -1.0, a
, b
), ('fadd', ('fneg', a
), b
)),
149 (('~flrp', a
, b
, 0.0), a
),
150 (('~flrp', a
, b
, 1.0), b
),
151 (('~flrp', a
, a
, b
), a
),
152 (('~flrp', 0.0, a
, b
), ('fmul', a
, b
)),
154 # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
155 (('~flrp', a
, ('fadd(is_used_once)', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
)),
156 (('~flrp@32', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp32'),
157 (('~flrp@64', a
, ('fadd', a
, b
), c
), ('fadd', ('fmul', b
, c
), a
), 'options->lower_flrp64'),
159 (('~flrp@32', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp32'),
160 (('~flrp@64', ('fadd', a
, b
), ('fadd', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
), 'options->lower_flrp64'),
162 (('~flrp@32', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp32'),
163 (('~flrp@64', a
, ('fmul(is_used_once)', a
, b
), c
), ('fmul', ('flrp', 1.0, b
, c
), a
), 'options->lower_flrp64'),
165 (('~flrp', ('fmul(is_used_once)', a
, b
), ('fmul(is_used_once)', a
, c
), d
), ('fmul', ('flrp', b
, c
, d
), a
)),
167 (('~flrp', a
, b
, ('b2f', 'c@1')), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
168 (('~flrp', a
, 0.0, c
), ('fadd', ('fmul', ('fneg', a
), c
), a
)),
169 (('ftrunc', a
), ('bcsel', ('flt', a
, 0.0), ('fneg', ('ffloor', ('fabs', a
))), ('ffloor', ('fabs', a
))), 'options->lower_ftrunc'),
170 (('ffloor', a
), ('fsub', a
, ('ffract', a
)), 'options->lower_ffloor'),
171 (('fadd', a
, ('fneg', ('ffract', a
))), ('ffloor', a
), '!options->lower_ffloor'),
172 (('ffract', a
), ('fsub', a
, ('ffloor', a
)), 'options->lower_ffract'),
173 (('fceil', a
), ('fneg', ('ffloor', ('fneg', a
))), 'options->lower_fceil'),
174 (('~fadd', ('fmul', a
, ('fadd', 1.0, ('fneg', ('b2f', 'c@1')))), ('fmul', b
, ('b2f', c
))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
175 (('~fadd@32', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
176 (('~fadd@64', ('fmul', a
, ('fadd', 1.0, ('fneg', c
) )), ('fmul', b
, c
)), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
177 # These are the same as the previous three rules, but it depends on
178 # 1-fsat(x) <=> fsat(1-x). See below.
179 (('~fadd@32', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp32'),
180 (('~fadd@64', ('fmul', a
, ('fsat', ('fadd', 1.0, ('fneg', c
)))), ('fmul', b
, ('fsat', c
))), ('flrp', a
, b
, ('fsat', c
)), '!options->lower_flrp64'),
182 (('~fadd', a
, ('fmul', ('b2f', 'c@1'), ('fadd', b
, ('fneg', a
)))), ('bcsel', c
, b
, a
), 'options->lower_flrp32'),
183 (('~fadd@32', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp32'),
184 (('~fadd@64', a
, ('fmul', c
, ('fadd', b
, ('fneg', a
)))), ('flrp', a
, b
, c
), '!options->lower_flrp64'),
185 (('ffma', a
, b
, c
), ('fadd', ('fmul', a
, b
), c
), 'options->lower_ffma'),
186 (('~fadd', ('fmul', a
, b
), c
), ('ffma', a
, b
, c
), 'options->fuse_ffma'),
188 (('~fmul', ('fadd', ('iand', ('ineg', ('b2i32', 'a@bool')), ('fmul', b
, c
)), '#d'), '#e'),
189 ('bcsel', a
, ('fmul', ('fadd', ('fmul', b
, c
), d
), e
), ('fmul', d
, e
))),
191 (('fdph', a
, b
), ('fdot4', ('vec4', 'a.x', 'a.y', 'a.z', 1.0), b
), 'options->lower_fdph'),
193 (('fdot4', ('vec4', a
, b
, c
, 1.0), d
), ('fdph', ('vec3', a
, b
, c
), d
), '!options->lower_fdph'),
194 (('fdot4', ('vec4', a
, 0.0, 0.0, 0.0), b
), ('fmul', a
, b
)),
195 (('fdot4', ('vec4', a
, b
, 0.0, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
196 (('fdot4', ('vec4', a
, b
, c
, 0.0), d
), ('fdot3', ('vec3', a
, b
, c
), d
)),
198 (('fdot3', ('vec3', a
, 0.0, 0.0), b
), ('fmul', a
, b
)),
199 (('fdot3', ('vec3', a
, b
, 0.0), c
), ('fdot2', ('vec2', a
, b
), c
)),
201 (('fdot2', ('vec2', a
, 0.0), b
), ('fmul', a
, b
)),
202 (('fdot2', a
, 1.0), ('fadd', 'a.x', 'a.y')),
204 # Lower fdot to fsum when it is available
205 (('fdot2', a
, b
), ('fsum2', ('fmul', a
, b
)), 'options->lower_fdot'),
206 (('fdot3', a
, b
), ('fsum3', ('fmul', a
, b
)), 'options->lower_fdot'),
207 (('fdot4', a
, b
), ('fsum4', ('fmul', a
, b
)), 'options->lower_fdot'),
208 (('fsum2', a
), ('fadd', 'a.x', 'a.y'), 'options->lower_fdot'),
210 # If x >= 0 and x <= 1: fsat(1 - x) == 1 - fsat(x) trivially
211 # If x < 0: 1 - fsat(x) => 1 - 0 => 1 and fsat(1 - x) => fsat(> 1) => 1
212 # If x > 1: 1 - fsat(x) => 1 - 1 => 0 and fsat(1 - x) => fsat(< 0) => 0
213 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
215 # 1 - ((1 - a) * (1 - b))
216 # 1 - (1 - a - b + a*b)
217 # 1 - 1 + a + b - a*b
222 (('~fadd@32', 1.0, ('fneg', ('fmul', ('fadd', 1.0, ('fneg', a
)), ('fadd', 1.0, ('fneg', b
))))),
223 ('flrp', b
, 1.0, a
), '!options->lower_flrp32'),
225 # (a * #b + #c) << #d
226 # ((a * #b) << #d) + (#c << #d)
227 # (a * (#b << #d)) + (#c << #d)
228 (('ishl', ('iadd', ('imul', a
, '#b'), '#c'), '#d'),
229 ('iadd', ('imul', a
, ('ishl', b
, d
)), ('ishl', c
, d
))),
233 (('ishl', ('imul', a
, '#b'), '#c'), ('imul', a
, ('ishl', b
, c
))),
236 # Care must be taken here. Shifts in NIR uses only the lower log2(bitsize)
237 # bits of the second source. These replacements must correctly handle the
238 # case where (b % bitsize) + (c % bitsize) >= bitsize.
239 for s
in [8, 16, 32, 64]:
242 ishl
= "ishl@{}".format(s
)
243 ishr
= "ishr@{}".format(s
)
244 ushr
= "ushr@{}".format(s
)
246 in_bounds
= ('ult', ('iadd', ('iand', b
, mask
), ('iand', c
, mask
)), s
)
248 optimizations
.extend([
249 ((ishl
, (ishl
, a
, '#b'), '#c'), ('bcsel', in_bounds
, (ishl
, a
, ('iadd', b
, c
)), 0)),
250 ((ushr
, (ushr
, a
, '#b'), '#c'), ('bcsel', in_bounds
, (ushr
, a
, ('iadd', b
, c
)), 0)),
252 # To get get -1 for large shifts of negative values, ishr must instead
253 # clamp the shift count to the maximum value.
254 ((ishr
, (ishr
, a
, '#b'), '#c'),
255 (ishr
, a
, ('imin', ('iadd', ('iand', b
, mask
), ('iand', c
, mask
)), s
- 1))),
258 optimizations
.extend([
259 # This is common for address calculations. Reassociating may enable the
260 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
261 # instruction or a constant offset field for in load / store instructions.
262 (('ishl', ('iadd', a
, '#b'), '#c'), ('iadd', ('ishl', a
, c
), ('ishl', b
, c
))),
264 # Comparison simplifications
265 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
266 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
267 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
268 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
269 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
270 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
271 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
272 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
273 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
274 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
276 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
277 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
278 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
279 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
280 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
282 # This helps some shaders because, after some optimizations, they end up
283 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
284 # matching would be handled by CSE.
285 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
286 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
287 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
288 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
289 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
290 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
291 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
292 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
293 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
294 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
296 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
297 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
298 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
299 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
300 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
301 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
303 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
304 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
305 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
306 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
310 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
312 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
314 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
316 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
317 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
318 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
319 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
320 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
321 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
322 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
323 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
324 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
325 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
326 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
327 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
328 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
329 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
330 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
331 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
332 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
333 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
334 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
335 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
337 # -(b2f(a) + b2f(b)) < 0
338 # 0 < b2f(a) + b2f(b)
339 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
341 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
342 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
344 # -(b2f(a) + b2f(b)) >= 0
345 # 0 >= b2f(a) + b2f(b)
346 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
348 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
349 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
351 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
352 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
354 # Some optimizations (below) convert things like (a < b || c < b) into
355 # (min(a, c) < b). However, this interfers with the previous optimizations
356 # that try to remove comparisons with negated sums of b2f. This just
358 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
359 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
361 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
362 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
363 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
364 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
365 (('~flt', ('fadd(is_used_once)', a
, '#b'), '#c'), ('flt', a
, ('fadd', c
, ('fneg', b
)))),
366 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('flt', ('fneg', ('fadd', c
, b
)), a
)),
367 (('~fge', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fge', a
, ('fadd', c
, ('fneg', b
)))),
368 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fge', ('fneg', ('fadd', c
, b
)), a
)),
369 (('~feq', ('fadd(is_used_once)', a
, '#b'), '#c'), ('feq', a
, ('fadd', c
, ('fneg', b
)))),
370 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('feq', ('fneg', ('fadd', c
, b
)), a
)),
371 (('~fne', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fne', a
, ('fadd', c
, ('fneg', b
)))),
372 (('~fne', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fne', ('fneg', ('fadd', c
, b
)), a
)),
374 # Cannot remove the addition from ilt or ige due to overflow.
375 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
376 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
378 # fmin(-b2f(a), b) >= 0.0
379 # -b2f(a) >= 0.0 && b >= 0.0
380 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
381 # b2f(a) == 0.0 && b >= 0.0
382 # a == False && b >= 0.0
385 # The fge in the second replacement is not a typo. I leave the proof that
386 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
388 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
389 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
391 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
392 (('~fne', ('b2f', 'a@1'), 0.0), a
),
393 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
394 (('ine', ('b2i', 'a@1'), 0), a
),
396 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
397 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
398 (('fge', ('u2f', a
), 0.0), True),
399 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
400 (('flt', ('u2f', a
), 0.0), False),
401 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
402 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
403 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
404 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
405 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
406 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
407 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
411 # fabs(a) != 0.0 because fabs(a) must be >= 0
413 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
417 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
420 # 0.0 == fabs(a) because fabs(a) must be >= 0
422 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
426 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
428 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
429 (('iand', ('fge', a
, 0.0), ('fge', 1.0, a
)), ('feq', a
, ('fsat', a
)), '!options->lower_fsat'),
431 # (a < 0.0) || (a > 1.0)
432 # !(!(a < 0.0) && !(a > 1.0))
433 # !((a >= 0.0) && (a <= 1.0))
436 (('ior', ('flt', a
, 0.0), ('flt', 1.0, a
)), ('fne', a
, ('fsat', a
)), '!options->lower_fsat'),
438 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
439 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
440 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
441 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
444 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
445 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
446 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
448 # Since b is a constant, constant folding will eliminate the fmin and the
449 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
450 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
452 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
454 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
455 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
456 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
457 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
458 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
459 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
460 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
461 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
462 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
463 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
464 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
465 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
466 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
467 (('bcsel', a
, True, b
), ('ior', a
, b
)),
468 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
469 (('bcsel', a
, b
, False), ('iand', a
, b
)),
470 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
471 (('~fmin', a
, a
), a
),
472 (('~fmax', a
, a
), a
),
477 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
478 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
479 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
480 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
481 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
482 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
483 (('iand@32', a
, ('inot', ('ishr', a
, 31))), ('imax', a
, 0)),
484 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
485 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
486 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
487 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
488 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
489 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
490 (('~fmin', a
, ('fabs', a
)), a
),
491 (('imin', a
, ('iabs', a
)), a
),
492 (('~fmax', a
, ('fneg', ('fabs', a
))), a
),
493 (('imax', a
, ('ineg', ('iabs', a
))), a
),
494 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
495 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
496 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
497 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
498 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
499 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
500 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
501 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
502 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
503 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
504 (('fsat', ('b2f', a
)), ('b2f', a
)),
505 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
506 (('fsat', ('fsat', a
)), ('fsat', a
)),
507 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_fsat'),
508 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_fsat'),
509 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
510 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
511 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
512 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
513 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
514 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
515 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
516 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
517 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
518 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
519 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
520 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
521 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
522 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
523 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
524 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
525 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
526 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
527 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
528 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
529 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
530 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
531 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
533 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
534 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
535 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
536 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
537 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
538 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
539 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
540 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
541 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
542 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
543 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
544 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
545 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
546 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
547 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
548 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
550 # These derive from the previous patterns with the application of b < 0 <=>
551 # 0 < -b. The transformation should be applied if either comparison is
552 # used once as this ensures that the number of comparisons will not
553 # increase. The sources to the ior and iand are not symmetric, so the
554 # rules have to be duplicated to get this behavior.
555 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
556 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
557 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
558 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
559 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
560 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
561 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
562 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
564 # Common pattern like 'if (i == 0 || i == 1 || ...)'
565 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
566 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
567 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
569 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
570 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
571 # so emit an open-coded version of that.
572 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
573 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
575 (('ior', a
, ('ieq', a
, False)), True),
576 (('ior', a
, ('inot', a
)), -1),
578 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
579 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
581 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0), '!options->lower_bitops'),
583 # These patterns can result when (a < b || a < c) => (a < min(b, c))
584 # transformations occur before constant propagation and loop-unrolling.
585 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
586 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
587 (('~fge', a
, ('fmin', b
, a
)), True),
588 (('~fge', ('fmax', a
, b
), a
), True),
589 (('~flt', a
, ('fmin', b
, a
)), False),
590 (('~flt', ('fmax', a
, b
), a
), False),
591 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
592 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
594 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
595 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
596 (('ige', a
, ('imin', b
, a
)), True),
597 (('ige', ('imax', a
, b
), a
), True),
598 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
599 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
600 (('uge', a
, ('umin', b
, a
)), True),
601 (('uge', ('umax', a
, b
), a
), True),
602 (('ilt', a
, ('imin', b
, a
)), False),
603 (('ilt', ('imax', a
, b
), a
), False),
604 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
605 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
606 (('ult', a
, ('umin', b
, a
)), False),
607 (('ult', ('umax', a
, b
), a
), False),
608 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
609 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
610 (('ult', a
, ('iand', b
, a
)), False),
611 (('ult', ('ior', a
, b
), a
), False),
612 (('uge', a
, ('iand', b
, a
)), True),
613 (('uge', ('ior', a
, b
), a
), True),
615 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
616 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
617 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
618 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
619 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
620 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
621 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
622 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
623 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
624 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
625 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
626 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
627 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
628 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
629 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
630 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
632 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
634 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
635 ('iabs', ('ishr', a
, b
))),
636 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
638 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
639 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
640 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
641 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
642 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
643 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
644 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
645 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
646 (('seq', ('seq', a
, b
), 1.0), ('seq', a
, b
)),
647 (('seq', ('sne', a
, b
), 1.0), ('sne', a
, b
)),
648 (('seq', ('slt', a
, b
), 1.0), ('slt', a
, b
)),
649 (('seq', ('sge', a
, b
), 1.0), ('sge', a
, b
)),
650 (('sne', ('seq', a
, b
), 0.0), ('seq', a
, b
)),
651 (('sne', ('sne', a
, b
), 0.0), ('sne', a
, b
)),
652 (('sne', ('slt', a
, b
), 0.0), ('slt', a
, b
)),
653 (('sne', ('sge', a
, b
), 0.0), ('sge', a
, b
)),
654 (('seq', ('seq', a
, b
), 0.0), ('sne', a
, b
)),
655 (('seq', ('sne', a
, b
), 0.0), ('seq', a
, b
)),
656 (('seq', ('slt', a
, b
), 0.0), ('sge', a
, b
)),
657 (('seq', ('sge', a
, b
), 0.0), ('slt', a
, b
)),
658 (('sne', ('seq', a
, b
), 1.0), ('sne', a
, b
)),
659 (('sne', ('sne', a
, b
), 1.0), ('seq', a
, b
)),
660 (('sne', ('slt', a
, b
), 1.0), ('sge', a
, b
)),
661 (('sne', ('sge', a
, b
), 1.0), ('slt', a
, b
)),
662 (('fall_equal2', a
, b
), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
663 (('fall_equal3', a
, b
), ('seq', ('fany_nequal3', a
, b
), 0.0), 'options->lower_vector_cmp'),
664 (('fall_equal4', a
, b
), ('seq', ('fany_nequal4', a
, b
), 0.0), 'options->lower_vector_cmp'),
665 (('fany_nequal2', a
, b
), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
666 (('fany_nequal3', a
, b
), ('fsat', ('fdot3', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
667 (('fany_nequal4', a
, b
), ('fsat', ('fdot4', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
668 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
669 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
671 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
672 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
673 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
674 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
675 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
676 (('ineg', ('b2i32', 'a@32')), a
),
677 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
678 # Comparison with the same args. Note that these are not done for
679 # the float versions because NaN always returns false on float
681 (('ilt', a
, a
), False),
682 (('ige', a
, a
), True),
683 (('ieq', a
, a
), True),
684 (('ine', a
, a
), False),
685 (('ult', a
, a
), False),
686 (('uge', a
, a
), True),
687 # Logical and bit operations
689 (('iand', a
, ~
0), a
),
693 (('ior', a
, True), True),
696 (('inot', ('inot', a
)), a
),
697 (('ior', ('iand', a
, b
), b
), b
),
698 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
699 (('iand', ('ior', a
, b
), b
), b
),
700 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
702 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
703 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
704 # Shift optimizations
711 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
712 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
713 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
714 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
715 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
716 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
717 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
718 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
719 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
720 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
721 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
722 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
723 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
724 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
725 # Exponential/logarithmic identities
726 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
727 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
728 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
729 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
730 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
731 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
732 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
733 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
734 (('~fpow', a
, 1.0), a
),
735 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
736 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
737 (('~fpow', 2.0, a
), ('fexp2', a
)),
738 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
739 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
740 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
741 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
742 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
743 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
744 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
745 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
746 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
747 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
748 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
749 (('~fmul', ('fsqrt', a
), ('fsqrt', a
)), ('fabs',a
)),
750 # Division and reciprocal
751 (('~fdiv', 1.0, a
), ('frcp', a
)),
752 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
753 (('~frcp', ('frcp', a
)), a
),
754 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
755 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
756 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
758 (('fsin', a
), lowered_sincos(0.5), 'options->lower_sincos'),
759 (('fcos', a
), lowered_sincos(0.75), 'options->lower_sincos'),
760 # Boolean simplifications
761 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
762 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
763 (('ieq', a
, True), a
),
764 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
765 (('ine', a
, False), a
),
766 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
767 (('bcsel', a
, True, False), a
),
768 (('bcsel', a
, False, True), ('inot', a
)),
769 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
770 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
771 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
772 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
773 (('bcsel', True, b
, c
), b
),
774 (('bcsel', False, b
, c
), c
),
775 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
777 (('bcsel', a
, b
, b
), b
),
778 (('~fcsel', a
, b
, b
), b
),
780 # D3D Boolean emulation
781 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
782 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
783 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
784 ('ineg', ('b2i', ('iand', a
, b
)))),
785 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
786 ('ineg', ('b2i', ('ior', a
, b
)))),
787 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
788 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
789 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
790 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
791 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
792 (('iand', ('ineg', ('b2i', a
)), 1), ('b2i', a
)),
794 # SM5 32-bit shifts are defined to use the 5 least significant bits
795 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
796 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
797 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
800 (('i2b32', ('b2i', 'a@32')), a
),
801 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
802 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
803 (('i2b', ('ineg', a
)), ('i2b', a
)),
804 (('i2b', ('iabs', a
)), ('i2b', a
)),
805 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
807 # The C spec says, "If the value of the integral part cannot be represented
808 # by the integer type, the behavior is undefined." "Undefined" can mean
809 # "the conversion doesn't happen at all."
810 (('~i2f32', ('f2i32', 'a@32')), ('ftrunc', a
)),
812 # Ironically, mark these as imprecise because removing the conversions may
813 # preserve more precision than doing the conversions (e.g.,
814 # uint(float(0x81818181u)) == 0x81818200).
815 (('~f2i32', ('i2f', 'a@32')), a
),
816 (('~f2i32', ('u2f', 'a@32')), a
),
817 (('~f2u32', ('i2f', 'a@32')), a
),
818 (('~f2u32', ('u2f', 'a@32')), a
),
820 (('ffloor', 'a(is_integral)'), a
),
821 (('fceil', 'a(is_integral)'), a
),
822 (('ftrunc', 'a(is_integral)'), a
),
823 # fract(x) = x - floor(x), so fract(NaN) = NaN
824 (('~ffract', 'a(is_integral)'), 0.0),
825 (('fabs', 'a(is_not_negative)'), a
),
826 (('iabs', 'a(is_not_negative)'), a
),
827 (('fsat', 'a(is_not_positive)'), 0.0),
829 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
832 # It is undefined to convert a negative floating-point value to an
835 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
836 # some optimizations in the i965 backend to proceed.
837 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
838 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
839 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
840 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
842 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a
), '!options->lower_fsat'),
844 # The result of the multiply must be in [-1, 0], so the result of the ffma
846 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
847 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
848 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
849 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
851 (('fne', 'a(is_not_zero)', 0.0), True),
852 (('feq', 'a(is_not_zero)', 0.0), False),
854 # In this chart, + means value > 0 and - means value < 0.
856 # + >= + -> unknown 0 >= + -> false - >= + -> false
857 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
858 # + >= - -> true 0 >= - -> true - >= - -> unknown
860 # Using grouping conceptually similar to a Karnaugh map...
862 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
863 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
864 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
866 # The flt / ilt cases just invert the expected result.
868 # The results expecting true, must be marked imprecise. The results
869 # expecting false are fine because NaN compared >= or < anything is false.
871 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
872 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
873 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
875 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
876 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
877 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
879 (('ine', 'a(is_not_zero)', 0), True),
880 (('ieq', 'a(is_not_zero)', 0), False),
882 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
883 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
884 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
886 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
887 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
888 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
890 (('ult', 0, 'a(is_gt_zero)'), True),
892 # Packing and then unpacking does nothing
893 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
894 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
895 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
896 ('unpack_64_2x32_split_y', a
)), a
),
898 # Comparing two halves of an unpack separately. While this optimization
899 # should be correct for non-constant values, it's less obvious that it's
900 # useful in that case. For constant values, the pack will fold and we're
901 # guaranteed to reduce the whole tree to one instruction.
902 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
903 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
904 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
907 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
908 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
909 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
910 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
911 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
912 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
913 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
915 # Useless masking before unpacking
916 (('unpack_half_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_half_2x16_split_x', a
)),
917 (('unpack_32_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_32_2x16_split_x', a
)),
918 (('unpack_64_2x32_split_x', ('iand', a
, 0xffffffff)), ('unpack_64_2x32_split_x', a
)),
919 (('unpack_half_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_half_2x16_split_y', a
)),
920 (('unpack_32_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_32_2x16_split_y', a
)),
921 (('unpack_64_2x32_split_y', ('iand', a
, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a
)),
924 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
925 # patterns like those below.
926 for op
in ('ushr', 'ishr'):
927 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
928 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
929 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
931 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
933 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
934 # patterns like those below.
935 for op
in ('extract_u8', 'extract_i8'):
936 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
937 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
938 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
940 optimizations
.extend([
942 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
943 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
944 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
945 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
946 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
949 (('ussub_4x8', a
, 0), a
),
950 (('ussub_4x8', a
, ~
0), 0),
951 # Lower all Subtractions first - they can get recombined later
952 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
))),
953 (('isub', a
, b
), ('iadd', a
, ('ineg', b
))),
954 (('uabs_usub', a
, b
), ('bcsel', ('ult', a
, b
), ('ineg', ('isub', a
, b
)), ('isub', a
, b
))),
955 # This is correct. We don't need isub_sat because the result type is unsigned, so it cannot overflow.
956 (('uabs_isub', a
, b
), ('bcsel', ('ilt', a
, b
), ('ineg', ('isub', a
, b
)), ('isub', a
, b
))),
958 # Propagate negation up multiplication chains
959 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
960 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
962 # Propagate constants up multiplication chains
963 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
964 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
965 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
966 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
968 # Reassociate constants in add/mul chains so they can be folded together.
969 # For now, we mostly only handle cases where the constants are separated by
970 # a single non-constant. We could do better eventually.
971 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
972 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
973 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
974 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
975 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
977 # Drop mul-div by the same value when there's no wrapping.
978 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
981 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
982 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
983 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
985 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
986 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
987 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
989 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
992 (('fmod@16', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
993 (('fmod@32', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
994 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
995 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
996 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
998 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
999 ('bcsel', ('ult', 31, 'bits'), 'insert',
1000 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
1001 'options->lower_bitfield_insert'),
1002 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1003 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1004 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1005 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1006 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
1007 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
1009 # Alternative lowering that doesn't rely on bfi.
1010 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1011 ('bcsel', ('ult', 31, 'bits'),
1014 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1015 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1016 'options->lower_bitfield_insert_to_shifts'),
1018 # Alternative lowering that uses bitfield_select.
1019 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1020 ('bcsel', ('ult', 31, 'bits'), 'insert',
1021 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1022 'options->lower_bitfield_insert_to_bitfield_select'),
1024 (('ibitfield_extract', 'value', 'offset', 'bits'),
1025 ('bcsel', ('ult', 31, 'bits'), 'value',
1026 ('ibfe', 'value', 'offset', 'bits')),
1027 'options->lower_bitfield_extract'),
1029 (('ubitfield_extract', 'value', 'offset', 'bits'),
1030 ('bcsel', ('ult', 31, 'bits'), 'value',
1031 ('ubfe', 'value', 'offset', 'bits')),
1032 'options->lower_bitfield_extract'),
1034 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1035 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1036 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1037 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1038 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1039 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1040 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1042 (('ibitfield_extract', 'value', 'offset', 'bits'),
1043 ('bcsel', ('ieq', 0, 'bits'),
1046 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1047 ('isub', 32, 'bits'))),
1048 'options->lower_bitfield_extract_to_shifts'),
1050 (('ubitfield_extract', 'value', 'offset', 'bits'),
1052 ('ushr', 'value', 'offset'),
1053 ('bcsel', ('ieq', 'bits', 32),
1055 ('isub', ('ishl', 1, 'bits'), 1))),
1056 'options->lower_bitfield_extract_to_shifts'),
1058 (('ifind_msb', 'value'),
1059 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1060 'options->lower_ifind_msb'),
1062 (('find_lsb', 'value'),
1063 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1064 'options->lower_find_lsb'),
1066 (('extract_i8', a
, 'b@32'),
1067 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
1068 'options->lower_extract_byte'),
1070 (('extract_u8', a
, 'b@32'),
1071 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
1072 'options->lower_extract_byte'),
1074 (('extract_i16', a
, 'b@32'),
1075 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
1076 'options->lower_extract_word'),
1078 (('extract_u16', a
, 'b@32'),
1079 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
1080 'options->lower_extract_word'),
1082 (('pack_unorm_2x16', 'v'),
1083 ('pack_uvec2_to_uint',
1084 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1085 'options->lower_pack_unorm_2x16'),
1087 (('pack_unorm_4x8', 'v'),
1088 ('pack_uvec4_to_uint',
1089 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1090 'options->lower_pack_unorm_4x8'),
1092 (('pack_snorm_2x16', 'v'),
1093 ('pack_uvec2_to_uint',
1094 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1095 'options->lower_pack_snorm_2x16'),
1097 (('pack_snorm_4x8', 'v'),
1098 ('pack_uvec4_to_uint',
1099 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1100 'options->lower_pack_snorm_4x8'),
1102 (('unpack_unorm_2x16', 'v'),
1103 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1104 ('extract_u16', 'v', 1))),
1106 'options->lower_unpack_unorm_2x16'),
1108 (('unpack_unorm_4x8', 'v'),
1109 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1110 ('extract_u8', 'v', 1),
1111 ('extract_u8', 'v', 2),
1112 ('extract_u8', 'v', 3))),
1114 'options->lower_unpack_unorm_4x8'),
1116 (('unpack_snorm_2x16', 'v'),
1117 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1118 ('extract_i16', 'v', 1))),
1120 'options->lower_unpack_snorm_2x16'),
1122 (('unpack_snorm_4x8', 'v'),
1123 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1124 ('extract_i8', 'v', 1),
1125 ('extract_i8', 'v', 2),
1126 ('extract_i8', 'v', 3))),
1128 'options->lower_unpack_snorm_4x8'),
1130 (('pack_half_2x16_split', 'a@32', 'b@32'),
1131 ('ior', ('ishl', ('u2u32', ('f2f16', b
)), 16), ('u2u32', ('f2f16', a
))),
1132 'options->lower_pack_half_2x16_split'),
1134 (('unpack_half_2x16_split_x', 'a@32'),
1135 ('f2f32', ('u2u16', a
)),
1136 'options->lower_unpack_half_2x16_split'),
1138 (('unpack_half_2x16_split_y', 'a@32'),
1139 ('f2f32', ('u2u16', ('ushr', a
, 16))),
1140 'options->lower_unpack_half_2x16_split'),
1142 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
1143 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
1145 # Address/offset calculations:
1146 # Drivers supporting imul24 should use the nir_lower_amul() pass, this
1147 # rule converts everyone else to imul:
1148 (('amul', a
, b
), ('imul', a
, b
), '!options->has_imul24'),
1150 (('imad24_ir3', a
, b
, 0), ('imul24', a
, b
)),
1151 (('imad24_ir3', a
, 0, c
), (c
)),
1152 (('imad24_ir3', a
, 1, c
), ('iadd', a
, c
)),
1154 # if first two srcs are const, crack apart the imad so constant folding
1155 # can clean up the imul:
1156 # TODO ffma should probably get a similar rule:
1157 (('imad24_ir3', '#a', '#b', c
), ('iadd', ('imul', a
, b
), c
)),
1159 # These will turn 24b address/offset calc back into 32b shifts, but
1160 # it should be safe to get back some of the bits of precision that we
1161 # already decided were no necessary:
1162 (('imul24', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitops'),
1163 (('imul24', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
1164 (('imul24', a
, 0), (0)),
1167 # bit_size dependent lowerings
1168 for bit_size
in [8, 16, 32, 64]:
1169 # convenience constants
1170 intmax
= (1 << (bit_size
- 1)) - 1
1171 intmin
= 1 << (bit_size
- 1)
1174 (('iadd_sat@' + str(bit_size
), a
, b
),
1175 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
1176 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
1177 (('isub_sat@' + str(bit_size
), a
, b
),
1178 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
1179 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
1182 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq'), ('fge', 'flt'), ('flt', 'fge')])
1184 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
1185 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1186 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1187 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1188 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1190 # Optimize x2bN(b2x(x)) -> x
1191 for size
in type_sizes('bool'):
1192 aN
= 'a@' + str(size
)
1193 f2bN
= 'f2b' + str(size
)
1194 i2bN
= 'i2b' + str(size
)
1195 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
1196 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
1198 # Optimize x2yN(b2x(x)) -> b2y
1199 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1200 if x
!= 'f' and y
!= 'f' and x
!= y
:
1203 b2x
= 'b2f' if x
== 'f' else 'b2i'
1204 b2y
= 'b2f' if y
== 'f' else 'b2i'
1205 x2yN
= '{}2{}'.format(x
, y
)
1206 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1208 # Optimize away x2xN(a@N)
1209 for t
in ['int', 'uint', 'float']:
1210 for N
in type_sizes(t
):
1211 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1212 aN
= 'a@{0}'.format(N
)
1213 optimizations
.append(((x2xN
, aN
), a
))
1215 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1216 # In particular, we can optimize away everything except upcast of downcast and
1217 # upcasts where the type differs from the other cast
1218 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1220 # The outer cast is a down-cast. It doesn't matter what the size of the
1221 # argument of the inner cast is because we'll never been in the upcast
1222 # of downcast case. Regardless of types, we'll always end up with y2yN
1224 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1225 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1226 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1227 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1228 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1230 # If the outer cast is an up-cast, we have to be more careful about the
1231 # size of the argument of the inner cast and with types. In this case,
1232 # the type is always the type of type up-cast which is given by the
1234 for P
in type_sizes('uint'):
1235 # We can't optimize away up-cast of down-cast.
1239 # Because we're doing down-cast of down-cast, the types always have
1240 # to match between the two casts
1241 for x
in ['i', 'u']:
1242 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1243 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1244 aP
= 'a@{0}'.format(P
)
1245 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1247 # The N == M case is handled by other optimizations
1250 # Optimize comparisons with up-casts
1251 for t
in ['int', 'uint', 'float']:
1252 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1253 if N
== 1 or N
>= M
:
1256 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1257 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1260 xeq
= 'feq' if t
== 'float' else 'ieq'
1261 xne
= 'fne' if t
== 'float' else 'ine'
1262 xge
= '{0}ge'.format(t
[0])
1263 xlt
= '{0}lt'.format(t
[0])
1265 # Up-casts are lossless so for correctly signed comparisons of
1266 # up-casted values we can do the comparison at the largest of the two
1267 # original sizes and drop one or both of the casts. (We have
1268 # optimizations to drop the no-op casts which this may generate.)
1269 for P
in type_sizes(t
):
1275 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1276 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1277 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1278 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1279 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1280 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1283 # The next bit doesn't work on floats because the range checks would
1284 # get way too complicated.
1285 if t
in ['int', 'uint']:
1287 xN_min
= -(1 << (N
- 1))
1288 xN_max
= (1 << (N
- 1)) - 1
1291 xN_max
= (1 << N
) - 1
1295 # If we're up-casting and comparing to a constant, we can unfold
1296 # the comparison into a comparison with the shrunk down constant
1297 # and a check that the constant fits in the smaller bit size.
1299 ((xeq
, (x2xM
, aN
), '#b'),
1300 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1301 ((xne
, (x2xM
, aN
), '#b'),
1302 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1303 ((xlt
, (x2xM
, aN
), '#b'),
1304 ('iand', (xlt
, xN_min
, b
),
1305 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1306 ((xlt
, '#a', (x2xM
, bN
)),
1307 ('iand', (xlt
, a
, xN_max
),
1308 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1309 ((xge
, (x2xM
, aN
), '#b'),
1310 ('iand', (xge
, xN_max
, b
),
1311 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1312 ((xge
, '#a', (x2xM
, bN
)),
1313 ('iand', (xge
, a
, xN_min
),
1314 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1317 def fexp2i(exp
, bits
):
1318 # We assume that exp is already in the right range.
1320 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1322 return ('ishl', ('iadd', exp
, 127), 23)
1324 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1328 def ldexp(f
, exp
, bits
):
1329 # First, we clamp exp to a reasonable range. The maximum possible range
1330 # for a normal exponent is [-126, 127] and, throwing in denormals, you get
1331 # a maximum range of [-149, 127]. This means that we can potentially have
1332 # a swing of +-276. If you start with FLT_MAX, you actually have to do
1333 # ldexp(FLT_MAX, -278) to get it to flush all the way to zero. The GLSL
1334 # spec, on the other hand, only requires that we handle an exponent value
1335 # in the range [-126, 128]. This implementation is *mostly* correct; it
1336 # handles a range on exp of [-252, 254] which allows you to create any
1337 # value (including denorms if the hardware supports it) and to adjust the
1338 # exponent of any normal value to anything you want.
1340 exp
= ('imin', ('imax', exp
, -28), 30)
1342 exp
= ('imin', ('imax', exp
, -252), 254)
1344 exp
= ('imin', ('imax', exp
, -2044), 2046)
1348 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1349 # (We use ishr which isn't the same for -1, but the -1 case still works
1350 # since we use exp-exp/2 as the second exponent.) While the spec
1351 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1352 # work with denormals and doesn't allow for the full swing in exponents
1353 # that you can get with normalized values. Instead, we create two powers
1354 # of two and multiply by them each in turn. That way the effective range
1355 # of our exponent is doubled.
1356 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1357 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1358 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1361 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1362 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1363 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1366 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1367 def bitfield_reverse(u
):
1368 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1369 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1370 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1371 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1372 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1376 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1378 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1379 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1380 # and, if a is a NaN then the second comparison will fail anyway.
1381 for op
in ['flt', 'fge', 'feq']:
1383 (('iand', ('feq', a
, a
), (op
, a
, b
)), ('!' + op
, a
, b
)),
1384 (('iand', ('feq', a
, a
), (op
, b
, a
)), ('!' + op
, b
, a
)),
1387 # Add optimizations to handle the case where the result of a ternary is
1388 # compared to a constant. This way we can take things like
1394 # a ? (0 > 0) : (1 > 0)
1396 # which constant folding will eat for lunch. The resulting ternary will
1397 # further get cleaned up by the boolean reductions above and we will be
1398 # left with just the original variable "a".
1399 for op
in ['flt', 'fge', 'feq', 'fne',
1400 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1402 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1403 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1404 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1405 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1409 # For example, this converts things like
1411 # 1 + mix(0, a - 1, condition)
1415 # mix(1, (a-1)+1, condition)
1417 # Other optimizations will rearrange the constants.
1418 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1420 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1423 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1426 # If neither layout qualifier is specified, derivatives in compute shaders
1427 # return zero, which is consistent with the handling of built-in texture
1428 # functions like texture() in GLSL 4.50 compute shaders.
1429 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1430 'fddy', 'fddy_fine', 'fddy_coarse']:
1432 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1435 # Some optimizations for ir3-specific instructions.
1437 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1438 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1439 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1440 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1441 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1444 # These kinds of sequences can occur after nir_opt_peephole_select.
1446 # NOTE: fadd is not handled here because that gets in the way of ffma
1447 # generation in the i965 driver. Instead, fadd and ffma are handled in
1448 # late_optimizations.
1452 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1453 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1454 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1455 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1456 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, e
, c
, d
)), (op
, ('bcsel', a
, b
, e
), c
, d
)),
1457 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', e
, c
, d
)), (op
, ('bcsel', a
, b
, e
), c
, d
)),
1460 for op
in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']:
1462 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, 'd(is_not_const)')), (op
, b
, ('bcsel', a
, c
, d
))),
1463 (('bcsel', a
, (op
+ '(is_used_once)', b
, 'c(is_not_const)'), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1464 (('bcsel', a
, (op
, b
, 'c(is_not_const)'), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1465 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, 'd(is_not_const)')), (op
, b
, ('bcsel', a
, c
, d
))),
1470 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1471 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1472 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, d
, c
)), (op
, ('bcsel', a
, b
, d
), c
)),
1473 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', d
, c
)), (op
, ('bcsel', a
, b
, d
), c
)),
1476 for op
in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos']:
1478 (('bcsel', a
, (op
+ '(is_used_once)', b
), (op
, c
)), (op
, ('bcsel', a
, b
, c
))),
1479 (('bcsel', a
, (op
, b
), (op
+ '(is_used_once)', c
)), (op
, ('bcsel', a
, b
, c
))),
1482 # This section contains "late" optimizations that should be run before
1483 # creating ffmas and calling regular optimizations for the final time.
1484 # Optimizations should go here if they help code generation and conflict
1485 # with the regular optimizations.
1486 before_ffma_optimizations
= [
1487 # Propagate constants down multiplication chains
1488 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1489 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1490 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1491 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1493 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1494 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1495 (('~fadd', ('fneg', a
), a
), 0.0),
1496 (('iadd', ('ineg', a
), a
), 0),
1497 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1498 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1499 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1500 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1502 (('~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
)),
1503 (('~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
)),
1504 (('~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
))),
1507 # This section contains "late" optimizations that should be run after the
1508 # regular optimizations have finished. Optimizations should go here if
1509 # they help code generation but do not necessarily produce code that is
1510 # more easily optimizable.
1511 late_optimizations
= [
1512 # Most of these optimizations aren't quite safe when you get infinity or
1513 # Nan involved but the first one should be fine.
1514 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1515 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1516 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1517 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1518 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1519 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1521 # nir_lower_to_source_mods will collapse this, but its existence during the
1522 # optimization loop can prevent other optimizations.
1523 (('fneg', ('fneg', a
)), a
),
1525 # Subtractions get lowered during optimization, so we need to recombine them
1526 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
1527 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
1528 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
1529 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
1531 # These are duplicated from the main optimizations table. The late
1532 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1533 # new patterns like these. The patterns that compare with zero are removed
1534 # because they are unlikely to be created in by anything in
1535 # late_optimizations.
1536 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1537 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1538 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1539 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1540 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1541 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1543 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1544 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1546 (('~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
)))),
1548 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1549 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1550 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1551 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1552 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1553 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1554 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1555 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1556 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1557 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1560 (('iand', a
, a
), a
),
1562 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1564 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1565 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1566 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1567 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1569 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1570 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1572 (('~fadd@32', 1.0, ('fmul(is_used_once)', c
, ('fadd', b
, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c
)), ('fmul', b
, c
)), 'options->lower_flrp32'),
1573 (('~fadd@64', 1.0, ('fmul(is_used_once)', c
, ('fadd', b
, -1.0 ))), ('fadd', ('fadd', 1.0, ('fneg', c
)), ('fmul', b
, c
)), 'options->lower_flrp64'),
1575 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1576 # particular operation is common for expanding values stored in a texture
1577 # from [0,1] to [-1,1].
1578 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1579 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1580 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1581 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1582 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1583 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1584 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1585 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1589 # a + -a*a + a*b (1)
1591 # Option 1: ffma(a, (b-a), a)
1593 # Alternately, after (1):
1599 # Option 2: ffma(a, 2, -(a*a))
1600 # Option 3: ffma(a, 2, (-a)*a)
1601 # Option 4: ffma(a, -a, (2*a)
1602 # Option 5: a * (2 - a)
1604 # There are a lot of other possible combinations.
1605 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1606 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1607 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1608 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1609 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1611 # we do these late so that we don't get in the way of creating ffmas
1612 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1613 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1615 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1617 # Putting this in 'optimizations' interferes with the bcsel(a, op(b, c),
1618 # op(b, d)) => op(b, bcsel(a, c, d)) transformations. I do not know why.
1619 (('bcsel', ('feq', ('fsqrt', 'a(is_not_negative)'), 0.0), intBitsToFloat(0x7f7fffff), ('frsq', a
)),
1620 ('fmin', ('frsq', a
), intBitsToFloat(0x7f7fffff))),
1622 # Things that look like DPH in the source shader may get expanded to
1623 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1624 # to NIR. After FFMA is generated, this can look like:
1626 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1628 # Reassociate the last addition into the first multiplication.
1630 # Some shaders do not use 'invariant' in vertex and (possibly) geometry
1631 # shader stages on some outputs that are intended to be invariant. For
1632 # various reasons, this optimization may not be fully applied in all
1633 # shaders used for different rendering passes of the same geometry. This
1634 # can result in Z-fighting artifacts (at best). For now, disable this
1635 # optimization in these stages. See bugzilla #111490. In tessellation
1636 # stages applications seem to use 'precise' when necessary, so allow the
1637 # optimization in those stages.
1638 (('~fadd', ('ffma(is_used_once)', a
, b
, ('ffma', c
, d
, ('fmul', 'e(is_not_const_and_not_fsign)', 'f(is_not_const_and_not_fsign)'))), 'g(is_not_const)'),
1639 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1640 (('~fadd', ('ffma(is_used_once)', a
, b
, ('fmul', 'c(is_not_const_and_not_fsign)', 'd(is_not_const_and_not_fsign)') ), 'e(is_not_const)'),
1641 ('ffma', a
, b
, ('ffma', c
, d
, e
)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1645 late_optimizations
+= [
1646 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1647 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1651 late_optimizations
+= [
1652 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1653 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1655 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1656 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1659 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1660 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1661 before_ffma_optimizations
).render())
1662 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1663 late_optimizations
).render())