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 # Optimize a pattern of address calculation created by DXVK where the offset is
259 # divided by 4 and then multipled by 4. This can be turned into an iand and the
260 # additions before can be reassociated to CSE the iand instruction.
261 for log2
in range(1, 7): # powers of two from 2 to 64
263 mask
= 0xffffffff & ~
(v
- 1)
264 b_is_multiple
= '#b(is_unsigned_multiple_of_{})'.format(v
)
266 optimizations
.extend([
267 # 'a >> #b << #b' -> 'a & ~((1 << #b) - 1)'
268 (('ishl@32', ('ushr@32', a
, log2
), log2
), ('iand', a
, mask
)),
270 # Reassociate for improved CSE
271 (('iand@32', ('iadd@32', a
, b_is_multiple
), mask
), ('iadd', ('iand', a
, mask
), b
)),
274 # To save space in the state tables, reduce to the set that is known to help.
275 # Previously, this was range(1, 32). In addition, a couple rules inside the
276 # loop are commented out. Revisit someday, probably after mesa/#2635 has some
278 for i
in [1, 2, 16, 24]:
279 lo_mask
= 0xffffffff >> i
280 hi_mask
= (0xffffffff << i
) & 0xffffffff
282 optimizations
.extend([
283 # This pattern seems to only help in the soft-fp64 code.
284 (('ishl@32', ('iand', 'a@32', lo_mask
), i
), ('ishl', a
, i
)),
285 # (('ushr@32', ('iand', 'a@32', hi_mask), i), ('ushr', a, i)),
286 # (('ishr@32', ('iand', 'a@32', hi_mask), i), ('ishr', a, i)),
288 (('iand', ('ishl', 'a@32', i
), hi_mask
), ('ishl', a
, i
)),
289 (('iand', ('ushr', 'a@32', i
), lo_mask
), ('ushr', a
, i
)),
290 # (('iand', ('ishr', 'a@32', i), lo_mask), ('ushr', a, i)), # Yes, ushr is correct
293 optimizations
.extend([
294 # This is common for address calculations. Reassociating may enable the
295 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
296 # instruction or a constant offset field for in load / store instructions.
297 (('ishl', ('iadd', a
, '#b'), '#c'), ('iadd', ('ishl', a
, c
), ('ishl', b
, c
))),
299 # Comparison simplifications
300 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
301 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
302 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
303 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
304 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
305 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
306 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
307 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
308 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
309 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
311 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
312 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
313 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
314 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
315 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
317 # This helps some shaders because, after some optimizations, they end up
318 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
319 # matching would be handled by CSE.
320 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
321 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
322 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
323 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
324 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
325 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
326 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
327 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
328 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
329 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
331 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
332 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
333 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
334 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
335 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
336 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
338 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
339 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
340 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
341 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
345 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
347 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
349 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
351 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
352 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
353 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
354 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
355 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
356 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
357 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
358 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
359 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
360 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
361 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
362 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
363 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
364 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
365 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
366 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
367 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
368 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
369 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
370 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
372 # -(b2f(a) + b2f(b)) < 0
373 # 0 < b2f(a) + b2f(b)
374 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
376 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
377 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
379 # -(b2f(a) + b2f(b)) >= 0
380 # 0 >= b2f(a) + b2f(b)
381 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
383 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
384 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
386 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
387 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
389 # Some optimizations (below) convert things like (a < b || c < b) into
390 # (min(a, c) < b). However, this interfers with the previous optimizations
391 # that try to remove comparisons with negated sums of b2f. This just
393 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
394 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
396 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
397 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
398 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
399 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
400 (('~flt', ('fadd(is_used_once)', a
, '#b'), '#c'), ('flt', a
, ('fadd', c
, ('fneg', b
)))),
401 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('flt', ('fneg', ('fadd', c
, b
)), a
)),
402 (('~fge', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fge', a
, ('fadd', c
, ('fneg', b
)))),
403 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fge', ('fneg', ('fadd', c
, b
)), a
)),
404 (('~feq', ('fadd(is_used_once)', a
, '#b'), '#c'), ('feq', a
, ('fadd', c
, ('fneg', b
)))),
405 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('feq', ('fneg', ('fadd', c
, b
)), a
)),
406 (('~fne', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fne', a
, ('fadd', c
, ('fneg', b
)))),
407 (('~fne', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fne', ('fneg', ('fadd', c
, b
)), a
)),
409 # Cannot remove the addition from ilt or ige due to overflow.
410 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
411 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
413 # fmin(-b2f(a), b) >= 0.0
414 # -b2f(a) >= 0.0 && b >= 0.0
415 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
416 # b2f(a) == 0.0 && b >= 0.0
417 # a == False && b >= 0.0
420 # The fge in the second replacement is not a typo. I leave the proof that
421 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
423 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
424 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
426 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
427 (('~fne', ('b2f', 'a@1'), 0.0), a
),
428 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
429 (('ine', ('b2i', 'a@1'), 0), a
),
431 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
432 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
433 (('fge', ('u2f', a
), 0.0), True),
434 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
435 (('flt', ('u2f', a
), 0.0), False),
436 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
437 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
438 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
439 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
440 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
441 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
442 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
446 # fabs(a) != 0.0 because fabs(a) must be >= 0
448 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
452 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
455 # 0.0 == fabs(a) because fabs(a) must be >= 0
457 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
461 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
463 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
464 (('iand', ('fge', a
, 0.0), ('fge', 1.0, a
)), ('feq', a
, ('fsat', a
)), '!options->lower_fsat'),
466 # (a < 0.0) || (a > 1.0)
467 # !(!(a < 0.0) && !(a > 1.0))
468 # !((a >= 0.0) && (a <= 1.0))
471 (('ior', ('flt', a
, 0.0), ('flt', 1.0, a
)), ('fne', a
, ('fsat', a
)), '!options->lower_fsat'),
473 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
474 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
475 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
476 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
479 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
480 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
481 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
483 # Since b is a constant, constant folding will eliminate the fmin and the
484 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
485 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
487 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
489 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
490 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
491 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
492 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
493 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
494 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
495 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
496 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
497 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
498 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
499 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
500 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
501 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
502 (('bcsel', a
, True, b
), ('ior', a
, b
)),
503 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
504 (('bcsel', a
, b
, False), ('iand', a
, b
)),
505 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
506 (('~fmin', a
, a
), a
),
507 (('~fmax', a
, a
), a
),
512 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
513 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
514 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
515 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
516 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
517 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
518 (('iand@32', a
, ('inot', ('ishr', a
, 31))), ('imax', a
, 0)),
520 # Simplify logic to detect sign of an integer.
521 (('ieq', ('iand', a
, 0x80000000), 0x00000000), ('ige', a
, 0)),
522 (('ine', ('iand', a
, 0x80000000), 0x80000000), ('ige', a
, 0)),
523 (('ine', ('iand', a
, 0x80000000), 0x00000000), ('ilt', a
, 0)),
524 (('ieq', ('iand', a
, 0x80000000), 0x80000000), ('ilt', a
, 0)),
525 (('ine', ('ushr', 'a@32', 31), 0), ('ilt', a
, 0)),
526 (('ieq', ('ushr', 'a@32', 31), 0), ('ige', a
, 0)),
527 (('ieq', ('ushr', 'a@32', 31), 1), ('ilt', a
, 0)),
528 (('ine', ('ushr', 'a@32', 31), 1), ('ige', a
, 0)),
529 (('ine', ('ishr', 'a@32', 31), 0), ('ilt', a
, 0)),
530 (('ieq', ('ishr', 'a@32', 31), 0), ('ige', a
, 0)),
531 (('ieq', ('ishr', 'a@32', 31), -1), ('ilt', a
, 0)),
532 (('ine', ('ishr', 'a@32', 31), -1), ('ige', a
, 0)),
534 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
535 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
536 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
537 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
538 (('~fmin', a
, ('fabs', a
)), a
),
539 (('imin', a
, ('iabs', a
)), a
),
540 (('~fmax', a
, ('fneg', ('fabs', a
))), a
),
541 (('imax', a
, ('ineg', ('iabs', a
))), a
),
542 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
543 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
544 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
545 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
546 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
547 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
548 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
549 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
550 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
551 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
552 (('fsat', ('b2f', a
)), ('b2f', a
)),
553 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
554 (('fsat', ('fsat', a
)), ('fsat', a
)),
555 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_fsat'),
556 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_fsat'),
557 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
558 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
559 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
560 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
561 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
562 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
563 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
564 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
565 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
566 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
567 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
568 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
569 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
570 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
571 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
572 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
573 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
574 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
575 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
576 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
577 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
578 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
579 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
581 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
582 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
583 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
584 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
585 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
586 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
587 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
588 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
589 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
590 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
591 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
592 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
593 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
594 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
595 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
596 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
598 # These derive from the previous patterns with the application of b < 0 <=>
599 # 0 < -b. The transformation should be applied if either comparison is
600 # used once as this ensures that the number of comparisons will not
601 # increase. The sources to the ior and iand are not symmetric, so the
602 # rules have to be duplicated to get this behavior.
603 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
604 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
605 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
606 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
607 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
608 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
609 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
610 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
612 # Common pattern like 'if (i == 0 || i == 1 || ...)'
613 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
614 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
615 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
617 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
618 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
619 # so emit an open-coded version of that.
620 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
621 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
623 (('ior', a
, ('ieq', a
, False)), True),
624 (('ior', a
, ('inot', a
)), -1),
626 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
627 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
629 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', a
, b
), 0), '!options->lower_bitops'),
630 (('ior', ('ine', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('ior', a
, b
), 0), '!options->lower_bitops'),
632 # This pattern occurs coutresy of __flt64_nonnan in the soft-fp64 code.
633 # The first part of the iand comes from the !__feq64_nonnan.
635 # The second pattern is a reformulation of the first based on the relation
636 # (a == 0 || y == 0) <=> umin(a, y) == 0, where b in the first equation
637 # happens to be y == 0.
638 (('iand', ('inot', ('iand', ('ior', ('ieq', a
, 0), b
), c
)), ('ilt', a
, 0)),
639 ('iand', ('inot', ('iand', b
, c
)), ('ilt', a
, 0))),
640 (('iand', ('inot', ('iand', ('ieq', ('umin', a
, b
), 0), c
)), ('ilt', a
, 0)),
641 ('iand', ('inot', ('iand', ('ieq', b
, 0), c
)), ('ilt', a
, 0))),
643 # These patterns can result when (a < b || a < c) => (a < min(b, c))
644 # transformations occur before constant propagation and loop-unrolling.
645 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
646 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
647 (('~fge', a
, ('fmin', b
, a
)), True),
648 (('~fge', ('fmax', a
, b
), a
), True),
649 (('~flt', a
, ('fmin', b
, a
)), False),
650 (('~flt', ('fmax', a
, b
), a
), False),
651 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
652 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
654 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
655 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
656 (('ige', a
, ('imin', b
, a
)), True),
657 (('ige', ('imax', a
, b
), a
), True),
658 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
659 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
660 (('uge', a
, ('umin', b
, a
)), True),
661 (('uge', ('umax', a
, b
), a
), True),
662 (('ilt', a
, ('imin', b
, a
)), False),
663 (('ilt', ('imax', a
, b
), a
), False),
664 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
665 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
666 (('ult', a
, ('umin', b
, a
)), False),
667 (('ult', ('umax', a
, b
), a
), False),
668 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
669 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
670 (('ult', a
, ('iand', b
, a
)), False),
671 (('ult', ('ior', a
, b
), a
), False),
672 (('uge', a
, ('iand', b
, a
)), True),
673 (('uge', ('ior', a
, b
), a
), True),
675 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
676 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
677 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
678 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
679 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
680 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
681 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
682 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
683 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
684 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
685 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
686 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
687 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
688 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
689 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
690 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
692 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
694 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
695 ('iabs', ('ishr', a
, b
))),
696 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
698 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
699 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
700 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
701 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
702 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
703 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
704 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
705 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
706 (('seq', ('seq', a
, b
), 1.0), ('seq', a
, b
)),
707 (('seq', ('sne', a
, b
), 1.0), ('sne', a
, b
)),
708 (('seq', ('slt', a
, b
), 1.0), ('slt', a
, b
)),
709 (('seq', ('sge', a
, b
), 1.0), ('sge', a
, b
)),
710 (('sne', ('seq', a
, b
), 0.0), ('seq', a
, b
)),
711 (('sne', ('sne', a
, b
), 0.0), ('sne', a
, b
)),
712 (('sne', ('slt', a
, b
), 0.0), ('slt', a
, b
)),
713 (('sne', ('sge', a
, b
), 0.0), ('sge', a
, b
)),
714 (('seq', ('seq', a
, b
), 0.0), ('sne', a
, b
)),
715 (('seq', ('sne', a
, b
), 0.0), ('seq', a
, b
)),
716 (('seq', ('slt', a
, b
), 0.0), ('sge', a
, b
)),
717 (('seq', ('sge', a
, b
), 0.0), ('slt', a
, b
)),
718 (('sne', ('seq', a
, b
), 1.0), ('sne', a
, b
)),
719 (('sne', ('sne', a
, b
), 1.0), ('seq', a
, b
)),
720 (('sne', ('slt', a
, b
), 1.0), ('sge', a
, b
)),
721 (('sne', ('sge', a
, b
), 1.0), ('slt', a
, b
)),
722 (('fall_equal2', a
, b
), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
723 (('fall_equal3', a
, b
), ('seq', ('fany_nequal3', a
, b
), 0.0), 'options->lower_vector_cmp'),
724 (('fall_equal4', a
, b
), ('seq', ('fany_nequal4', a
, b
), 0.0), 'options->lower_vector_cmp'),
725 (('fany_nequal2', a
, b
), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
726 (('fany_nequal3', a
, b
), ('fsat', ('fdot3', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
727 (('fany_nequal4', a
, b
), ('fsat', ('fdot4', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
728 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
729 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
731 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
732 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
733 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
734 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
735 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
736 (('ineg', ('b2i32', 'a@32')), a
),
737 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
738 # Comparison with the same args. Note that these are not done for
739 # the float versions because NaN always returns false on float
741 (('ilt', a
, a
), False),
742 (('ige', a
, a
), True),
743 (('ieq', a
, a
), True),
744 (('ine', a
, a
), False),
745 (('ult', a
, a
), False),
746 (('uge', a
, a
), True),
747 # Logical and bit operations
749 (('iand', a
, ~
0), a
),
753 (('ior', a
, True), True),
756 (('inot', ('inot', a
)), a
),
757 (('ior', ('iand', a
, b
), b
), b
),
758 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
759 (('iand', ('ior', a
, b
), b
), b
),
760 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
762 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
763 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
764 # Shift optimizations
771 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
772 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
773 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
774 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
775 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
776 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
777 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
778 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
779 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
780 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
781 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
782 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
783 # Exponential/logarithmic identities
784 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
785 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
786 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
787 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
788 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
789 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
790 (('~fexp2', ('fmul', ('flog2', a
), 0.5)), ('fsqrt', a
)),
791 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
792 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
793 (('~fpow', a
, 1.0), a
),
794 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
795 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
796 (('~fpow', 2.0, a
), ('fexp2', a
)),
797 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
798 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
799 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
800 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
801 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
802 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
803 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
804 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
805 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
806 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
807 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
808 (('~fmul', ('fsqrt', a
), ('fsqrt', a
)), ('fabs',a
)),
809 # Division and reciprocal
810 (('~fdiv', 1.0, a
), ('frcp', a
)),
811 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
812 (('~frcp', ('frcp', a
)), a
),
813 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
814 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
815 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
817 (('fsin', a
), lowered_sincos(0.5), 'options->lower_sincos'),
818 (('fcos', a
), lowered_sincos(0.75), 'options->lower_sincos'),
819 # Boolean simplifications
820 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
821 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
822 (('ieq', a
, True), a
),
823 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
824 (('ine', a
, False), a
),
825 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
826 (('bcsel', a
, True, False), a
),
827 (('bcsel', a
, False, True), ('inot', a
)),
828 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
829 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
830 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
831 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
832 (('bcsel', True, b
, c
), b
),
833 (('bcsel', False, b
, c
), c
),
834 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
836 (('bcsel', a
, b
, b
), b
),
837 (('~fcsel', a
, b
, b
), b
),
839 # D3D Boolean emulation
840 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
841 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
842 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
843 ('ineg', ('b2i', ('iand', a
, b
)))),
844 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
845 ('ineg', ('b2i', ('ior', a
, b
)))),
846 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
847 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
848 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
849 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
850 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
851 (('iand', ('ineg', ('b2i', a
)), 1), ('b2i', a
)),
853 # SM5 32-bit shifts are defined to use the 5 least significant bits
854 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
855 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
856 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
859 (('i2b32', ('b2i', 'a@32')), a
),
860 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
861 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
862 (('i2b', ('ineg', a
)), ('i2b', a
)),
863 (('i2b', ('iabs', a
)), ('i2b', a
)),
864 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
866 # The C spec says, "If the value of the integral part cannot be represented
867 # by the integer type, the behavior is undefined." "Undefined" can mean
868 # "the conversion doesn't happen at all."
869 (('~i2f32', ('f2i32', 'a@32')), ('ftrunc', a
)),
871 # Ironically, mark these as imprecise because removing the conversions may
872 # preserve more precision than doing the conversions (e.g.,
873 # uint(float(0x81818181u)) == 0x81818200).
874 (('~f2i32', ('i2f', 'a@32')), a
),
875 (('~f2i32', ('u2f', 'a@32')), a
),
876 (('~f2u32', ('i2f', 'a@32')), a
),
877 (('~f2u32', ('u2f', 'a@32')), a
),
879 # Conversions from float16 to float32 and back can always be removed
880 (('f2f16', ('f2f32', 'a@16')), a
),
881 (('f2fmp', ('f2f32', 'a@16')), a
),
882 # Conversions to float16 would be lossy so they should only be removed if
883 # the instruction was generated by the precision lowering pass.
884 (('f2f32', ('f2fmp', 'a@32')), a
),
886 (('ffloor', 'a(is_integral)'), a
),
887 (('fceil', 'a(is_integral)'), a
),
888 (('ftrunc', 'a(is_integral)'), a
),
889 # fract(x) = x - floor(x), so fract(NaN) = NaN
890 (('~ffract', 'a(is_integral)'), 0.0),
891 (('fabs', 'a(is_not_negative)'), a
),
892 (('iabs', 'a(is_not_negative)'), a
),
893 (('fsat', 'a(is_not_positive)'), 0.0),
895 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
898 # It is undefined to convert a negative floating-point value to an
901 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
902 # some optimizations in the i965 backend to proceed.
903 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
904 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
905 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
906 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
908 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a
), '!options->lower_fsat'),
910 # The result of the multiply must be in [-1, 0], so the result of the ffma
912 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
913 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
914 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
915 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
917 (('fne', 'a(is_not_zero)', 0.0), True),
918 (('feq', 'a(is_not_zero)', 0.0), False),
920 # In this chart, + means value > 0 and - means value < 0.
922 # + >= + -> unknown 0 >= + -> false - >= + -> false
923 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
924 # + >= - -> true 0 >= - -> true - >= - -> unknown
926 # Using grouping conceptually similar to a Karnaugh map...
928 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
929 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
930 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
932 # The flt / ilt cases just invert the expected result.
934 # The results expecting true, must be marked imprecise. The results
935 # expecting false are fine because NaN compared >= or < anything is false.
937 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
938 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
939 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
941 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
942 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
943 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
945 (('ine', 'a(is_not_zero)', 0), True),
946 (('ieq', 'a(is_not_zero)', 0), False),
948 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
949 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
950 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
952 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
953 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
954 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
956 (('ult', 0, 'a(is_gt_zero)'), True),
957 (('ult', a
, 0), False),
959 # Packing and then unpacking does nothing
960 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
961 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
962 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
963 ('unpack_64_2x32_split_y', a
)), a
),
965 # Comparing two halves of an unpack separately. While this optimization
966 # should be correct for non-constant values, it's less obvious that it's
967 # useful in that case. For constant values, the pack will fold and we're
968 # guaranteed to reduce the whole tree to one instruction.
969 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
970 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
971 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
974 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
975 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
976 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
977 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
978 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
979 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
980 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
982 # Useless masking before unpacking
983 (('unpack_half_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_half_2x16_split_x', a
)),
984 (('unpack_32_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_32_2x16_split_x', a
)),
985 (('unpack_64_2x32_split_x', ('iand', a
, 0xffffffff)), ('unpack_64_2x32_split_x', a
)),
986 (('unpack_half_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_half_2x16_split_y', a
)),
987 (('unpack_32_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_32_2x16_split_y', a
)),
988 (('unpack_64_2x32_split_y', ('iand', a
, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a
)),
990 # Optimize half packing
991 (('ishl', ('pack_half_2x16', ('vec2', a
, 0)), 16), ('pack_half_2x16', ('vec2', 0, a
))),
992 (('ishr', ('pack_half_2x16', ('vec2', 0, a
)), 16), ('pack_half_2x16', ('vec2', a
, 0))),
994 (('iadd', ('pack_half_2x16', ('vec2', a
, 0)), ('pack_half_2x16', ('vec2', 0, b
))),
995 ('pack_half_2x16', ('vec2', a
, b
))),
996 (('ior', ('pack_half_2x16', ('vec2', a
, 0)), ('pack_half_2x16', ('vec2', 0, b
))),
997 ('pack_half_2x16', ('vec2', a
, b
))),
1000 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
1001 # patterns like those below.
1002 for op
in ('ushr', 'ishr'):
1003 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
1004 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
1005 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
1007 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
1009 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
1010 # patterns like those below.
1011 for op
in ('extract_u8', 'extract_i8'):
1012 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
1013 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
1014 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
1016 optimizations
.extend([
1018 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
1019 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
1020 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
1021 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
1022 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
1025 (('ussub_4x8', a
, 0), a
),
1026 (('ussub_4x8', a
, ~
0), 0),
1027 # Lower all Subtractions first - they can get recombined later
1028 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
))),
1029 (('isub', a
, b
), ('iadd', a
, ('ineg', b
))),
1030 (('uabs_usub', a
, b
), ('bcsel', ('ult', a
, b
), ('ineg', ('isub', a
, b
)), ('isub', a
, b
))),
1031 # This is correct. We don't need isub_sat because the result type is unsigned, so it cannot overflow.
1032 (('uabs_isub', a
, b
), ('bcsel', ('ilt', a
, b
), ('ineg', ('isub', a
, b
)), ('isub', a
, b
))),
1034 # Propagate negation up multiplication chains
1035 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
1036 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
1038 # Propagate constants up multiplication chains
1039 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
1040 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
1041 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
1042 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
1044 # Reassociate constants in add/mul chains so they can be folded together.
1045 # For now, we mostly only handle cases where the constants are separated by
1046 # a single non-constant. We could do better eventually.
1047 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
1048 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
1049 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
1050 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
1051 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
1052 (('iand', '#a', ('iand', 'b(is_not_const)', '#c')), ('iand', ('iand', a
, c
), b
)),
1053 (('ior', '#a', ('ior', 'b(is_not_const)', '#c')), ('ior', ('ior', a
, c
), b
)),
1054 (('ixor', '#a', ('ixor', 'b(is_not_const)', '#c')), ('ixor', ('ixor', a
, c
), b
)),
1056 # Drop mul-div by the same value when there's no wrapping.
1057 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
1060 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
1061 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
1062 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
1064 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
1065 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
1066 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
1068 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
1071 (('fmod', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
1072 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
1073 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
1074 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
1076 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1077 ('bcsel', ('ult', 31, 'bits'), 'insert',
1078 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
1079 'options->lower_bitfield_insert'),
1080 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1081 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1082 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1083 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1084 (('ihadd@64', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1085 (('uhadd@64', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1086 (('irhadd@64', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1087 (('urhadd@64', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1089 (('uadd_sat@64', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1090 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
1091 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
1092 (('usub_sat@64', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_usub_sat64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1094 # int64_t sum = a + b;
1096 # if (a < 0 && b < 0 && a < sum)
1098 # } else if (a >= 0 && b >= 0 && sum < a)
1102 # A couple optimizations are applied.
1104 # 1. a < sum => sum >= 0. This replacement works because it is known that
1105 # a < 0 and b < 0, so sum should also be < 0 unless there was
1108 # 2. sum < a => sum < 0. This replacement works because it is known that
1109 # a >= 0 and b >= 0, so sum should also be >= 0 unless there was
1112 # 3. Invert the second if-condition and swap the order of parameters for
1113 # the bcsel. !(a >= 0 && b >= 0 && sum < 0) becomes !(a >= 0) || !(b >=
1114 # 0) || !(sum < 0), and that becomes (a < 0) || (b < 0) || (sum >= 0)
1116 # On Intel Gen11, this saves ~11 instructions.
1117 (('iadd_sat@64', a
, b
), ('bcsel',
1118 ('iand', ('iand', ('ilt', a
, 0), ('ilt', b
, 0)), ('ige', ('iadd', a
, b
), 0)),
1121 ('ior', ('ior', ('ilt', a
, 0), ('ilt', b
, 0)), ('ige', ('iadd', a
, b
), 0)),
1123 0x7fffffffffffffff)),
1124 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1126 # int64_t sum = a - b;
1128 # if (a < 0 && b >= 0 && a < sum)
1130 # } else if (a >= 0 && b < 0 && a >= sum)
1134 # Optimizations similar to the iadd_sat case are applied here.
1135 (('isub_sat@64', a
, b
), ('bcsel',
1136 ('iand', ('iand', ('ilt', a
, 0), ('ige', b
, 0)), ('ige', ('isub', a
, b
), 0)),
1139 ('ior', ('ior', ('ilt', a
, 0), ('ige', b
, 0)), ('ige', ('isub', a
, b
), 0)),
1141 0x7fffffffffffffff)),
1142 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1144 # These are done here instead of in the backend because the int64 lowering
1145 # pass will make a mess of the patterns. The first patterns are
1146 # conditioned on nir_lower_minmax64 because it was not clear that it was
1147 # always an improvement on platforms that have real int64 support. No
1148 # shaders in shader-db hit this, so it was hard to say one way or the
1150 (('ilt', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imax', ('unpack_64_2x32_split_y', a
), ('unpack_64_2x32_split_y', b
)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1151 (('ilt', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ilt', ('imin', ('unpack_64_2x32_split_y', a
), ('unpack_64_2x32_split_y', b
)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1152 (('ige', ('imax(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imax', ('unpack_64_2x32_split_y', a
), ('unpack_64_2x32_split_y', b
)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1153 (('ige', ('imin(is_used_once)', 'a@64', 'b@64'), 0), ('ige', ('imin', ('unpack_64_2x32_split_y', a
), ('unpack_64_2x32_split_y', b
)), 0), '(options->lower_int64_options & nir_lower_minmax64) != 0'),
1154 (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a
), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1155 (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a
), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1157 (('ine', 'a@64', 0), ('ine', ('ior', ('unpack_64_2x32_split_x', a
), ('unpack_64_2x32_split_y', a
)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1158 (('ieq', 'a@64', 0), ('ieq', ('ior', ('unpack_64_2x32_split_x', a
), ('unpack_64_2x32_split_y', a
)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1159 # 0u < uint(a) <=> uint(a) != 0u
1160 (('ult', 0, 'a@64'), ('ine', ('ior', ('unpack_64_2x32_split_x', a
), ('unpack_64_2x32_split_y', a
)), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1162 # Alternative lowering that doesn't rely on bfi.
1163 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1164 ('bcsel', ('ult', 31, 'bits'),
1167 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1168 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1169 'options->lower_bitfield_insert_to_shifts'),
1171 # Alternative lowering that uses bitfield_select.
1172 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1173 ('bcsel', ('ult', 31, 'bits'), 'insert',
1174 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1175 'options->lower_bitfield_insert_to_bitfield_select'),
1177 (('ibitfield_extract', 'value', 'offset', 'bits'),
1178 ('bcsel', ('ult', 31, 'bits'), 'value',
1179 ('ibfe', 'value', 'offset', 'bits')),
1180 'options->lower_bitfield_extract'),
1182 (('ubitfield_extract', 'value', 'offset', 'bits'),
1183 ('bcsel', ('ult', 31, 'bits'), 'value',
1184 ('ubfe', 'value', 'offset', 'bits')),
1185 'options->lower_bitfield_extract'),
1187 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1188 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1189 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1190 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1191 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1192 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1193 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1195 (('ibitfield_extract', 'value', 'offset', 'bits'),
1196 ('bcsel', ('ieq', 0, 'bits'),
1199 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1200 ('isub', 32, 'bits'))),
1201 'options->lower_bitfield_extract_to_shifts'),
1203 (('ubitfield_extract', 'value', 'offset', 'bits'),
1205 ('ushr', 'value', 'offset'),
1206 ('bcsel', ('ieq', 'bits', 32),
1208 ('isub', ('ishl', 1, 'bits'), 1))),
1209 'options->lower_bitfield_extract_to_shifts'),
1211 (('ifind_msb', 'value'),
1212 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1213 'options->lower_ifind_msb'),
1215 (('find_lsb', 'value'),
1216 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1217 'options->lower_find_lsb'),
1219 (('extract_i8', a
, 'b@32'),
1220 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
1221 'options->lower_extract_byte'),
1223 (('extract_u8', a
, 'b@32'),
1224 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
1225 'options->lower_extract_byte'),
1227 (('extract_i16', a
, 'b@32'),
1228 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
1229 'options->lower_extract_word'),
1231 (('extract_u16', a
, 'b@32'),
1232 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
1233 'options->lower_extract_word'),
1235 (('pack_unorm_2x16', 'v'),
1236 ('pack_uvec2_to_uint',
1237 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1238 'options->lower_pack_unorm_2x16'),
1240 (('pack_unorm_4x8', 'v'),
1241 ('pack_uvec4_to_uint',
1242 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1243 'options->lower_pack_unorm_4x8'),
1245 (('pack_snorm_2x16', 'v'),
1246 ('pack_uvec2_to_uint',
1247 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1248 'options->lower_pack_snorm_2x16'),
1250 (('pack_snorm_4x8', 'v'),
1251 ('pack_uvec4_to_uint',
1252 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1253 'options->lower_pack_snorm_4x8'),
1255 (('unpack_unorm_2x16', 'v'),
1256 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1257 ('extract_u16', 'v', 1))),
1259 'options->lower_unpack_unorm_2x16'),
1261 (('unpack_unorm_4x8', 'v'),
1262 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1263 ('extract_u8', 'v', 1),
1264 ('extract_u8', 'v', 2),
1265 ('extract_u8', 'v', 3))),
1267 'options->lower_unpack_unorm_4x8'),
1269 (('unpack_snorm_2x16', 'v'),
1270 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1271 ('extract_i16', 'v', 1))),
1273 'options->lower_unpack_snorm_2x16'),
1275 (('unpack_snorm_4x8', 'v'),
1276 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1277 ('extract_i8', 'v', 1),
1278 ('extract_i8', 'v', 2),
1279 ('extract_i8', 'v', 3))),
1281 'options->lower_unpack_snorm_4x8'),
1283 (('pack_half_2x16_split', 'a@32', 'b@32'),
1284 ('ior', ('ishl', ('u2u32', ('f2f16', b
)), 16), ('u2u32', ('f2f16', a
))),
1285 'options->lower_pack_half_2x16_split'),
1287 (('unpack_half_2x16_split_x', 'a@32'),
1288 ('f2f32', ('u2u16', a
)),
1289 'options->lower_unpack_half_2x16_split'),
1291 (('unpack_half_2x16_split_y', 'a@32'),
1292 ('f2f32', ('u2u16', ('ushr', a
, 16))),
1293 'options->lower_unpack_half_2x16_split'),
1295 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
1296 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
1298 # Address/offset calculations:
1299 # Drivers supporting imul24 should use the nir_lower_amul() pass, this
1300 # rule converts everyone else to imul:
1301 (('amul', a
, b
), ('imul', a
, b
), '!options->has_imul24'),
1303 (('imad24_ir3', a
, b
, 0), ('imul24', a
, b
)),
1304 (('imad24_ir3', a
, 0, c
), (c
)),
1305 (('imad24_ir3', a
, 1, c
), ('iadd', a
, c
)),
1307 # if first two srcs are const, crack apart the imad so constant folding
1308 # can clean up the imul:
1309 # TODO ffma should probably get a similar rule:
1310 (('imad24_ir3', '#a', '#b', c
), ('iadd', ('imul', a
, b
), c
)),
1312 # These will turn 24b address/offset calc back into 32b shifts, but
1313 # it should be safe to get back some of the bits of precision that we
1314 # already decided were no necessary:
1315 (('imul24', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitops'),
1316 (('imul24', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
1317 (('imul24', a
, 0), (0)),
1320 # bit_size dependent lowerings
1321 for bit_size
in [8, 16, 32, 64]:
1322 # convenience constants
1323 intmax
= (1 << (bit_size
- 1)) - 1
1324 intmin
= 1 << (bit_size
- 1)
1327 (('iadd_sat@' + str(bit_size
), a
, b
),
1328 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
1329 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
1330 (('isub_sat@' + str(bit_size
), a
, b
),
1331 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
1332 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
1335 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq')])
1337 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
1338 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1339 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1340 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1341 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1343 # Optimize x2bN(b2x(x)) -> x
1344 for size
in type_sizes('bool'):
1345 aN
= 'a@' + str(size
)
1346 f2bN
= 'f2b' + str(size
)
1347 i2bN
= 'i2b' + str(size
)
1348 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
1349 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
1351 # Optimize x2yN(b2x(x)) -> b2y
1352 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1353 if x
!= 'f' and y
!= 'f' and x
!= y
:
1356 b2x
= 'b2f' if x
== 'f' else 'b2i'
1357 b2y
= 'b2f' if y
== 'f' else 'b2i'
1358 x2yN
= '{}2{}'.format(x
, y
)
1359 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1361 # Optimize away x2xN(a@N)
1362 for t
in ['int', 'uint', 'float', 'bool']:
1363 for N
in type_sizes(t
):
1364 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1365 aN
= 'a@{0}'.format(N
)
1366 optimizations
.append(((x2xN
, aN
), a
))
1368 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1369 # In particular, we can optimize away everything except upcast of downcast and
1370 # upcasts where the type differs from the other cast
1371 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1373 # The outer cast is a down-cast. It doesn't matter what the size of the
1374 # argument of the inner cast is because we'll never been in the upcast
1375 # of downcast case. Regardless of types, we'll always end up with y2yN
1377 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1378 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1379 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1380 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1381 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1383 # If the outer cast is an up-cast, we have to be more careful about the
1384 # size of the argument of the inner cast and with types. In this case,
1385 # the type is always the type of type up-cast which is given by the
1387 for P
in type_sizes('uint'):
1388 # We can't optimize away up-cast of down-cast.
1392 # Because we're doing down-cast of down-cast, the types always have
1393 # to match between the two casts
1394 for x
in ['i', 'u']:
1395 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1396 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1397 aP
= 'a@{0}'.format(P
)
1398 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1400 # The N == M case is handled by other optimizations
1403 # Downcast operations should be able to see through pack
1404 for t
in ['i', 'u']:
1405 for N
in [8, 16, 32]:
1406 x2xN
= '{0}2{0}{1}'.format(t
, N
)
1408 ((x2xN
, ('pack_64_2x32_split', a
, b
)), (x2xN
, a
)),
1409 ((x2xN
, ('pack_64_2x32_split', a
, b
)), (x2xN
, a
)),
1412 # Optimize comparisons with up-casts
1413 for t
in ['int', 'uint', 'float']:
1414 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1415 if N
== 1 or N
>= M
:
1418 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1419 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1422 xeq
= 'feq' if t
== 'float' else 'ieq'
1423 xne
= 'fne' if t
== 'float' else 'ine'
1424 xge
= '{0}ge'.format(t
[0])
1425 xlt
= '{0}lt'.format(t
[0])
1427 # Up-casts are lossless so for correctly signed comparisons of
1428 # up-casted values we can do the comparison at the largest of the two
1429 # original sizes and drop one or both of the casts. (We have
1430 # optimizations to drop the no-op casts which this may generate.)
1431 for P
in type_sizes(t
):
1437 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1438 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1439 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1440 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1441 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1442 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1445 # The next bit doesn't work on floats because the range checks would
1446 # get way too complicated.
1447 if t
in ['int', 'uint']:
1449 xN_min
= -(1 << (N
- 1))
1450 xN_max
= (1 << (N
- 1)) - 1
1453 xN_max
= (1 << N
) - 1
1457 # If we're up-casting and comparing to a constant, we can unfold
1458 # the comparison into a comparison with the shrunk down constant
1459 # and a check that the constant fits in the smaller bit size.
1461 ((xeq
, (x2xM
, aN
), '#b'),
1462 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1463 ((xne
, (x2xM
, aN
), '#b'),
1464 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1465 ((xlt
, (x2xM
, aN
), '#b'),
1466 ('iand', (xlt
, xN_min
, b
),
1467 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1468 ((xlt
, '#a', (x2xM
, bN
)),
1469 ('iand', (xlt
, a
, xN_max
),
1470 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1471 ((xge
, (x2xM
, aN
), '#b'),
1472 ('iand', (xge
, xN_max
, b
),
1473 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1474 ((xge
, '#a', (x2xM
, bN
)),
1475 ('iand', (xge
, a
, xN_min
),
1476 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1479 def fexp2i(exp
, bits
):
1480 # Generate an expression which constructs value 2.0^exp or 0.0.
1482 # We assume that exp is already in a valid range:
1484 # * [-15, 15] for 16-bit float
1485 # * [-127, 127] for 32-bit float
1486 # * [-1023, 1023] for 16-bit float
1488 # If exp is the lowest value in the valid range, a value of 0.0 is
1489 # constructed. Otherwise, the value 2.0^exp is constructed.
1491 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1493 return ('ishl', ('iadd', exp
, 127), 23)
1495 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1499 def ldexp(f
, exp
, bits
):
1500 # The maximum possible range for a normal exponent is [-126, 127] and,
1501 # throwing in denormals, you get a maximum range of [-149, 127]. This
1502 # means that we can potentially have a swing of +-276. If you start with
1503 # FLT_MAX, you actually have to do ldexp(FLT_MAX, -278) to get it to flush
1504 # all the way to zero. The GLSL spec only requires that we handle a subset
1505 # of this range. From version 4.60 of the spec:
1507 # "If exp is greater than +128 (single-precision) or +1024
1508 # (double-precision), the value returned is undefined. If exp is less
1509 # than -126 (single-precision) or -1022 (double-precision), the value
1510 # returned may be flushed to zero. Additionally, splitting the value
1511 # into a significand and exponent using frexp() and then reconstructing
1512 # a floating-point value using ldexp() should yield the original input
1513 # for zero and all finite non-denormalized values."
1515 # The SPIR-V spec has similar language.
1517 # In order to handle the maximum value +128 using the fexp2i() helper
1518 # above, we have to split the exponent in half and do two multiply
1521 # First, we clamp exp to a reasonable range. Specifically, we clamp to
1522 # twice the full range that is valid for the fexp2i() function above. If
1523 # exp/2 is the bottom value of that range, the fexp2i() expression will
1524 # yield 0.0f which, when multiplied by f, will flush it to zero which is
1525 # allowed by the GLSL and SPIR-V specs for low exponent values. If the
1526 # value is clamped from above, then it must have been above the supported
1527 # range of the GLSL built-in and therefore any return value is acceptable.
1529 exp
= ('imin', ('imax', exp
, -30), 30)
1531 exp
= ('imin', ('imax', exp
, -254), 254)
1533 exp
= ('imin', ('imax', exp
, -2046), 2046)
1537 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1538 # (We use ishr which isn't the same for -1, but the -1 case still works
1539 # since we use exp-exp/2 as the second exponent.) While the spec
1540 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1541 # work with denormals and doesn't allow for the full swing in exponents
1542 # that you can get with normalized values. Instead, we create two powers
1543 # of two and multiply by them each in turn. That way the effective range
1544 # of our exponent is doubled.
1545 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1546 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1547 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1550 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1551 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1552 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1555 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1556 def bitfield_reverse(u
):
1557 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1558 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1559 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1560 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1561 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1565 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1567 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1568 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1569 # and, if a is a NaN then the second comparison will fail anyway.
1570 for op
in ['flt', 'fge', 'feq']:
1572 (('iand', ('feq', a
, a
), (op
, a
, b
)), ('!' + op
, a
, b
)),
1573 (('iand', ('feq', a
, a
), (op
, b
, a
)), ('!' + op
, b
, a
)),
1576 # Add optimizations to handle the case where the result of a ternary is
1577 # compared to a constant. This way we can take things like
1583 # a ? (0 > 0) : (1 > 0)
1585 # which constant folding will eat for lunch. The resulting ternary will
1586 # further get cleaned up by the boolean reductions above and we will be
1587 # left with just the original variable "a".
1588 for op
in ['flt', 'fge', 'feq', 'fne',
1589 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1591 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1592 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1593 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1594 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1598 # For example, this converts things like
1600 # 1 + mix(0, a - 1, condition)
1604 # mix(1, (a-1)+1, condition)
1606 # Other optimizations will rearrange the constants.
1607 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1609 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1612 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1615 # If neither layout qualifier is specified, derivatives in compute shaders
1616 # return zero, which is consistent with the handling of built-in texture
1617 # functions like texture() in GLSL 4.50 compute shaders.
1618 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1619 'fddy', 'fddy_fine', 'fddy_coarse']:
1621 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1624 # Some optimizations for ir3-specific instructions.
1626 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1627 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1628 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1629 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1630 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1633 # These kinds of sequences can occur after nir_opt_peephole_select.
1635 # NOTE: fadd is not handled here because that gets in the way of ffma
1636 # generation in the i965 driver. Instead, fadd and ffma are handled in
1637 # late_optimizations.
1641 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1642 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1643 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1644 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1645 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, e
, c
, d
)), (op
, ('bcsel', a
, b
, e
), c
, d
)),
1646 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', e
, c
, d
)), (op
, ('bcsel', a
, b
, e
), c
, d
)),
1649 for op
in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']:
1651 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, 'd(is_not_const)')), (op
, b
, ('bcsel', a
, c
, d
))),
1652 (('bcsel', a
, (op
+ '(is_used_once)', b
, 'c(is_not_const)'), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1653 (('bcsel', a
, (op
, b
, 'c(is_not_const)'), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1654 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, 'd(is_not_const)')), (op
, b
, ('bcsel', a
, c
, d
))),
1659 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1660 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1661 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, d
, c
)), (op
, ('bcsel', a
, b
, d
), c
)),
1662 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', d
, c
)), (op
, ('bcsel', a
, b
, d
), c
)),
1665 for op
in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos']:
1667 (('bcsel', a
, (op
+ '(is_used_once)', b
), (op
, c
)), (op
, ('bcsel', a
, b
, c
))),
1668 (('bcsel', a
, (op
, b
), (op
+ '(is_used_once)', c
)), (op
, ('bcsel', a
, b
, c
))),
1671 # This section contains "late" optimizations that should be run before
1672 # creating ffmas and calling regular optimizations for the final time.
1673 # Optimizations should go here if they help code generation and conflict
1674 # with the regular optimizations.
1675 before_ffma_optimizations
= [
1676 # Propagate constants down multiplication chains
1677 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1678 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1679 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1680 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1682 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1683 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1684 (('~fadd', ('fneg', a
), a
), 0.0),
1685 (('iadd', ('ineg', a
), a
), 0),
1686 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1687 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1688 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1689 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1691 (('~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
)),
1692 (('~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
)),
1693 (('~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
))),
1696 # This section contains "late" optimizations that should be run after the
1697 # regular optimizations have finished. Optimizations should go here if
1698 # they help code generation but do not necessarily produce code that is
1699 # more easily optimizable.
1700 late_optimizations
= [
1701 # Most of these optimizations aren't quite safe when you get infinity or
1702 # Nan involved but the first one should be fine.
1703 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1704 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1705 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1706 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1707 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1708 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1710 # nir_lower_to_source_mods will collapse this, but its existence during the
1711 # optimization loop can prevent other optimizations.
1712 (('fneg', ('fneg', a
)), a
),
1714 # Subtractions get lowered during optimization, so we need to recombine them
1715 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
1716 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
1717 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
1718 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
1720 # These are duplicated from the main optimizations table. The late
1721 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1722 # new patterns like these. The patterns that compare with zero are removed
1723 # because they are unlikely to be created in by anything in
1724 # late_optimizations.
1725 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1726 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1727 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1728 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1729 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1730 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1732 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1733 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1735 (('~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
)))),
1737 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1738 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1739 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1740 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1741 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1742 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1743 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1744 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1745 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1746 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1749 (('iand', a
, a
), a
),
1751 (('iand', ('ine(is_used_once)', 'a@32', 0), ('ine', 'b@32', 0)), ('ine', ('umin', a
, b
), 0)),
1752 (('ior', ('ieq(is_used_once)', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('umin', a
, b
), 0)),
1754 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1756 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1757 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1758 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1759 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1761 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1762 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1764 (('~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'),
1765 (('~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'),
1767 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1768 # particular operation is common for expanding values stored in a texture
1769 # from [0,1] to [-1,1].
1770 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1771 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1772 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1773 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1774 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1775 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1776 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1777 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1781 # a + -a*a + a*b (1)
1783 # Option 1: ffma(a, (b-a), a)
1785 # Alternately, after (1):
1791 # Option 2: ffma(a, 2, -(a*a))
1792 # Option 3: ffma(a, 2, (-a)*a)
1793 # Option 4: ffma(a, -a, (2*a)
1794 # Option 5: a * (2 - a)
1796 # There are a lot of other possible combinations.
1797 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1798 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1799 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1800 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1801 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1803 # we do these late so that we don't get in the way of creating ffmas
1804 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1805 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1807 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1809 # Putting this in 'optimizations' interferes with the bcsel(a, op(b, c),
1810 # op(b, d)) => op(b, bcsel(a, c, d)) transformations. I do not know why.
1811 (('bcsel', ('feq', ('fsqrt', 'a(is_not_negative)'), 0.0), intBitsToFloat(0x7f7fffff), ('frsq', a
)),
1812 ('fmin', ('frsq', a
), intBitsToFloat(0x7f7fffff))),
1814 # Things that look like DPH in the source shader may get expanded to
1815 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1816 # to NIR. After FFMA is generated, this can look like:
1818 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1820 # Reassociate the last addition into the first multiplication.
1822 # Some shaders do not use 'invariant' in vertex and (possibly) geometry
1823 # shader stages on some outputs that are intended to be invariant. For
1824 # various reasons, this optimization may not be fully applied in all
1825 # shaders used for different rendering passes of the same geometry. This
1826 # can result in Z-fighting artifacts (at best). For now, disable this
1827 # optimization in these stages. See bugzilla #111490. In tessellation
1828 # stages applications seem to use 'precise' when necessary, so allow the
1829 # optimization in those stages.
1830 (('~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)'),
1831 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1832 (('~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)'),
1833 ('ffma', a
, b
, ('ffma', c
, d
, e
)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1835 # Convert f2fmp instructions to concrete f2f16 instructions. At this point
1836 # any conversions that could have been removed will have been removed in
1837 # nir_opt_algebraic so any remaining ones are required.
1838 (('f2fmp', a
), ('f2f16', a
)),
1842 late_optimizations
+= [
1843 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1844 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1848 late_optimizations
+= [
1849 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1850 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1852 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1853 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1856 distribute_src_mods
= [
1857 # Try to remove some spurious negations rather than pushing them down.
1858 (('fmul', ('fneg', a
), ('fneg', b
)), ('fmul', a
, b
)),
1859 (('ffma', ('fneg', a
), ('fneg', b
), c
), ('ffma', a
, b
, c
)),
1860 (('fdot_replicated2', ('fneg', a
), ('fneg', b
)), ('fdot_replicated2', a
, b
)),
1861 (('fdot_replicated3', ('fneg', a
), ('fneg', b
)), ('fdot_replicated3', a
, b
)),
1862 (('fdot_replicated4', ('fneg', a
), ('fneg', b
)), ('fdot_replicated4', a
, b
)),
1863 (('fneg', ('fneg', a
)), a
),
1865 (('fneg', ('ffma(is_used_once)', a
, b
, c
)), ('ffma', ('fneg', a
), b
, ('fneg', c
))),
1866 (('fneg', ('flrp(is_used_once)', a
, b
, c
)), ('flrp', ('fneg', a
), ('fneg', b
), c
)),
1867 (('fneg', ('fadd(is_used_once)', a
, b
)), ('fadd', ('fneg', a
), ('fneg', b
))),
1869 # Note that fmin <-> fmax. I don't think there is a way to distribute
1870 # fabs() into fmin or fmax.
1871 (('fneg', ('fmin(is_used_once)', a
, b
)), ('fmax', ('fneg', a
), ('fneg', b
))),
1872 (('fneg', ('fmax(is_used_once)', a
, b
)), ('fmin', ('fneg', a
), ('fneg', b
))),
1874 # fdph works mostly like fdot, but to get the correct result, the negation
1875 # must be applied to the second source.
1876 (('fneg', ('fdph_replicated(is_used_once)', a
, b
)), ('fdph_replicated', a
, ('fneg', b
))),
1877 (('fabs', ('fdph_replicated(is_used_once)', a
, b
)), ('fdph_replicated', ('fabs', a
), ('fabs', b
))),
1879 (('fneg', ('fsign(is_used_once)', a
)), ('fsign', ('fneg', a
))),
1880 (('fabs', ('fsign(is_used_once)', a
)), ('fsign', ('fabs', a
))),
1883 for op
in ['fmul', 'fdot_replicated2', 'fdot_replicated3', 'fdot_replicated4']:
1884 distribute_src_mods
.extend([
1885 (('fneg', (op
+ '(is_used_once)', a
, b
)), (op
, ('fneg', a
), b
)),
1886 (('fabs', (op
+ '(is_used_once)', a
, b
)), (op
, ('fabs', a
), ('fabs', b
))),
1889 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1890 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1891 before_ffma_optimizations
).render())
1892 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1893 late_optimizations
).render())
1894 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_distribute_src_mods",
1895 distribute_src_mods
).render())
projects / mesa.git / blob