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 optimizations
.extend([
275 # This is common for address calculations. Reassociating may enable the
276 # 'a<<c' to be CSE'd. It also helps architectures that have an ISHLADD
277 # instruction or a constant offset field for in load / store instructions.
278 (('ishl', ('iadd', a
, '#b'), '#c'), ('iadd', ('ishl', a
, c
), ('ishl', b
, c
))),
280 # Comparison simplifications
281 (('~inot', ('flt', a
, b
)), ('fge', a
, b
)),
282 (('~inot', ('fge', a
, b
)), ('flt', a
, b
)),
283 (('inot', ('feq', a
, b
)), ('fne', a
, b
)),
284 (('inot', ('fne', a
, b
)), ('feq', a
, b
)),
285 (('inot', ('ilt', a
, b
)), ('ige', a
, b
)),
286 (('inot', ('ult', a
, b
)), ('uge', a
, b
)),
287 (('inot', ('ige', a
, b
)), ('ilt', a
, b
)),
288 (('inot', ('uge', a
, b
)), ('ult', a
, b
)),
289 (('inot', ('ieq', a
, b
)), ('ine', a
, b
)),
290 (('inot', ('ine', a
, b
)), ('ieq', a
, b
)),
292 (('iand', ('feq', a
, b
), ('fne', a
, b
)), False),
293 (('iand', ('flt', a
, b
), ('flt', b
, a
)), False),
294 (('iand', ('ieq', a
, b
), ('ine', a
, b
)), False),
295 (('iand', ('ilt', a
, b
), ('ilt', b
, a
)), False),
296 (('iand', ('ult', a
, b
), ('ult', b
, a
)), False),
298 # This helps some shaders because, after some optimizations, they end up
299 # with patterns like (-a < -b) || (b < a). In an ideal world, this sort of
300 # matching would be handled by CSE.
301 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
302 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
303 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
304 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
305 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
306 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
307 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
308 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
309 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
310 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
312 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
313 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
314 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
315 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
316 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
317 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
319 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
320 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
321 (('fge', 0.0, ('fsat(is_used_once)', a
)), ('fge', 0.0, a
)),
322 (('flt', 0.0, ('fsat(is_used_once)', a
)), ('flt', 0.0, a
)),
326 # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
328 (('fge', 0.0, ('b2f', 'a@1')), ('inot', a
)),
330 (('fge', ('fneg', ('b2f', 'a@1')), 0.0), ('inot', a
)),
332 (('fne', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
333 (('fne', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('ior', a
, b
)),
334 (('fne', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('ior', a
, b
)),
335 (('fne', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('ior', a
, b
)),
336 (('fne', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
337 (('fne', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('iand', a
, b
)),
338 (('fne', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('iand', a
, b
)),
339 (('fne', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ixor', a
, b
)),
340 (('fne', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ixor', a
, b
)),
341 (('fne', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ixor', a
, b
)),
342 (('feq', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
343 (('feq', ('fmax', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('ior', a
, b
))),
344 (('feq', ('bcsel', a
, 1.0, ('b2f', 'b@1')) , 0.0), ('inot', ('ior', a
, b
))),
345 (('feq', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
346 (('feq', ('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
347 (('feq', ('fmin', ('b2f', 'a@1'), ('b2f', 'b@1')), 0.0), ('inot', ('iand', a
, b
))),
348 (('feq', ('bcsel', a
, ('b2f', 'b@1'), 0.0) , 0.0), ('inot', ('iand', a
, b
))),
349 (('feq', ('fadd', ('b2f', 'a@1'), ('fneg', ('b2f', 'b@1'))), 0.0), ('ieq', a
, b
)),
350 (('feq', ('b2f', 'a@1') , ('b2f', 'b@1') ), ('ieq', a
, b
)),
351 (('feq', ('fneg', ('b2f', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('ieq', a
, b
)),
353 # -(b2f(a) + b2f(b)) < 0
354 # 0 < b2f(a) + b2f(b)
355 # 0 != b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
357 (('flt', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('ior', a
, b
)),
358 (('flt', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('ior', a
, b
)),
360 # -(b2f(a) + b2f(b)) >= 0
361 # 0 >= b2f(a) + b2f(b)
362 # 0 == b2f(a) + b2f(b) b2f must be 0 or 1, so the sum is non-negative
364 (('fge', ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), 0.0), ('inot', ('ior', a
, b
))),
365 (('fge', 0.0, ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('inot', ('ior', a
, b
))),
367 (('flt', a
, ('fneg', a
)), ('flt', a
, 0.0)),
368 (('fge', a
, ('fneg', a
)), ('fge', a
, 0.0)),
370 # Some optimizations (below) convert things like (a < b || c < b) into
371 # (min(a, c) < b). However, this interfers with the previous optimizations
372 # that try to remove comparisons with negated sums of b2f. This just
374 (('flt', ('fmin', c
, ('fneg', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1')))), 0.0),
375 ('ior', ('flt', c
, 0.0), ('ior', a
, b
))),
377 (('~flt', ('fadd', a
, b
), a
), ('flt', b
, 0.0)),
378 (('~fge', ('fadd', a
, b
), a
), ('fge', b
, 0.0)),
379 (('~feq', ('fadd', a
, b
), a
), ('feq', b
, 0.0)),
380 (('~fne', ('fadd', a
, b
), a
), ('fne', b
, 0.0)),
381 (('~flt', ('fadd(is_used_once)', a
, '#b'), '#c'), ('flt', a
, ('fadd', c
, ('fneg', b
)))),
382 (('~flt', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('flt', ('fneg', ('fadd', c
, b
)), a
)),
383 (('~fge', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fge', a
, ('fadd', c
, ('fneg', b
)))),
384 (('~fge', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fge', ('fneg', ('fadd', c
, b
)), a
)),
385 (('~feq', ('fadd(is_used_once)', a
, '#b'), '#c'), ('feq', a
, ('fadd', c
, ('fneg', b
)))),
386 (('~feq', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('feq', ('fneg', ('fadd', c
, b
)), a
)),
387 (('~fne', ('fadd(is_used_once)', a
, '#b'), '#c'), ('fne', a
, ('fadd', c
, ('fneg', b
)))),
388 (('~fne', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, '#b')), '#c'), ('fne', ('fneg', ('fadd', c
, b
)), a
)),
390 # Cannot remove the addition from ilt or ige due to overflow.
391 (('ieq', ('iadd', a
, b
), a
), ('ieq', b
, 0)),
392 (('ine', ('iadd', a
, b
), a
), ('ine', b
, 0)),
394 # fmin(-b2f(a), b) >= 0.0
395 # -b2f(a) >= 0.0 && b >= 0.0
396 # -b2f(a) == 0.0 && b >= 0.0 -b2f can only be 0 or -1, never >0
397 # b2f(a) == 0.0 && b >= 0.0
398 # a == False && b >= 0.0
401 # The fge in the second replacement is not a typo. I leave the proof that
402 # "fmin(-b2f(a), b) >= 0 <=> fmin(-b2f(a), b) == 0" as an exercise for the
404 (('fge', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
405 (('feq', ('fmin', ('fneg', ('b2f', 'a@1')), 'b@1'), 0.0), ('iand', ('inot', a
), ('fge', b
, 0.0))),
407 (('feq', ('b2f', 'a@1'), 0.0), ('inot', a
)),
408 (('~fne', ('b2f', 'a@1'), 0.0), a
),
409 (('ieq', ('b2i', 'a@1'), 0), ('inot', a
)),
410 (('ine', ('b2i', 'a@1'), 0), a
),
412 (('fne', ('u2f', a
), 0.0), ('ine', a
, 0)),
413 (('feq', ('u2f', a
), 0.0), ('ieq', a
, 0)),
414 (('fge', ('u2f', a
), 0.0), True),
415 (('fge', 0.0, ('u2f', a
)), ('uge', 0, a
)), # ieq instead?
416 (('flt', ('u2f', a
), 0.0), False),
417 (('flt', 0.0, ('u2f', a
)), ('ult', 0, a
)), # ine instead?
418 (('fne', ('i2f', a
), 0.0), ('ine', a
, 0)),
419 (('feq', ('i2f', a
), 0.0), ('ieq', a
, 0)),
420 (('fge', ('i2f', a
), 0.0), ('ige', a
, 0)),
421 (('fge', 0.0, ('i2f', a
)), ('ige', 0, a
)),
422 (('flt', ('i2f', a
), 0.0), ('ilt', a
, 0)),
423 (('flt', 0.0, ('i2f', a
)), ('ilt', 0, a
)),
427 # fabs(a) != 0.0 because fabs(a) must be >= 0
429 (('~flt', 0.0, ('fabs', a
)), ('fne', a
, 0.0)),
433 (('~flt', ('fneg', ('fabs', a
)), 0.0), ('fne', a
, 0.0)),
436 # 0.0 == fabs(a) because fabs(a) must be >= 0
438 (('fge', 0.0, ('fabs', a
)), ('feq', a
, 0.0)),
442 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
444 # (a >= 0.0) && (a <= 1.0) -> fsat(a) == a
445 (('iand', ('fge', a
, 0.0), ('fge', 1.0, a
)), ('feq', a
, ('fsat', a
)), '!options->lower_fsat'),
447 # (a < 0.0) || (a > 1.0)
448 # !(!(a < 0.0) && !(a > 1.0))
449 # !((a >= 0.0) && (a <= 1.0))
452 (('ior', ('flt', a
, 0.0), ('flt', 1.0, a
)), ('fne', a
, ('fsat', a
)), '!options->lower_fsat'),
454 (('fmax', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('ior', a
, b
))),
455 (('fmax', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('ior', a
, b
)))),
456 (('fmin', ('b2f(is_used_once)', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
457 (('fmin', ('fneg(is_used_once)', ('b2f(is_used_once)', 'a@1')), ('fneg', ('b2f', 'b@1'))), ('fneg', ('b2f', ('iand', a
, b
)))),
460 # bcsel(a, fmin(b2f(a), b), fmin(b2f(a), b))
461 # bcsel(a, fmin(b2f(True), b), fmin(b2f(False), b))
462 # bcsel(a, fmin(1.0, b), fmin(0.0, b))
464 # Since b is a constant, constant folding will eliminate the fmin and the
465 # fmax. If b is > 1.0, the bcsel will be replaced with a b2f.
466 (('fmin', ('b2f', 'a@1'), '#b'), ('bcsel', a
, ('fmin', b
, 1.0), ('fmin', b
, 0.0))),
468 (('flt', ('fadd(is_used_once)', a
, ('fneg', b
)), 0.0), ('flt', a
, b
)),
470 (('fge', ('fneg', ('fabs', a
)), 0.0), ('feq', a
, 0.0)),
471 (('~bcsel', ('flt', b
, a
), b
, a
), ('fmin', a
, b
)),
472 (('~bcsel', ('flt', a
, b
), b
, a
), ('fmax', a
, b
)),
473 (('~bcsel', ('fge', a
, b
), b
, a
), ('fmin', a
, b
)),
474 (('~bcsel', ('fge', b
, a
), b
, a
), ('fmax', a
, b
)),
475 (('bcsel', ('i2b', a
), b
, c
), ('bcsel', ('ine', a
, 0), b
, c
)),
476 (('bcsel', ('inot', a
), b
, c
), ('bcsel', a
, c
, b
)),
477 (('bcsel', a
, ('bcsel', a
, b
, c
), d
), ('bcsel', a
, b
, d
)),
478 (('bcsel', a
, b
, ('bcsel', a
, c
, d
)), ('bcsel', a
, b
, d
)),
479 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
480 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, c
, 'e')), ('bcsel', b
, c
, ('bcsel', a
, d
, 'e'))),
481 (('bcsel', a
, ('bcsel', b
, c
, d
), ('bcsel(is_used_once)', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
482 (('bcsel', a
, ('bcsel(is_used_once)', b
, c
, d
), ('bcsel', b
, 'e', d
)), ('bcsel', b
, ('bcsel', a
, c
, 'e'), d
)),
483 (('bcsel', a
, True, b
), ('ior', a
, b
)),
484 (('bcsel', a
, a
, b
), ('ior', a
, b
)),
485 (('bcsel', a
, b
, False), ('iand', a
, b
)),
486 (('bcsel', a
, b
, a
), ('iand', a
, b
)),
487 (('~fmin', a
, a
), a
),
488 (('~fmax', a
, a
), a
),
493 (('fmax', ('fmax', a
, b
), b
), ('fmax', a
, b
)),
494 (('umax', ('umax', a
, b
), b
), ('umax', a
, b
)),
495 (('imax', ('imax', a
, b
), b
), ('imax', a
, b
)),
496 (('fmin', ('fmin', a
, b
), b
), ('fmin', a
, b
)),
497 (('umin', ('umin', a
, b
), b
), ('umin', a
, b
)),
498 (('imin', ('imin', a
, b
), b
), ('imin', a
, b
)),
499 (('iand@32', a
, ('inot', ('ishr', a
, 31))), ('imax', a
, 0)),
500 (('fmin', a
, ('fneg', a
)), ('fneg', ('fabs', a
))),
501 (('imin', a
, ('ineg', a
)), ('ineg', ('iabs', a
))),
502 (('fmin', a
, ('fneg', ('fabs', a
))), ('fneg', ('fabs', a
))),
503 (('imin', a
, ('ineg', ('iabs', a
))), ('ineg', ('iabs', a
))),
504 (('~fmin', a
, ('fabs', a
)), a
),
505 (('imin', a
, ('iabs', a
)), a
),
506 (('~fmax', a
, ('fneg', ('fabs', a
))), a
),
507 (('imax', a
, ('ineg', ('iabs', a
))), a
),
508 (('fmax', a
, ('fabs', a
)), ('fabs', a
)),
509 (('imax', a
, ('iabs', a
)), ('iabs', a
)),
510 (('fmax', a
, ('fneg', a
)), ('fabs', a
)),
511 (('imax', a
, ('ineg', a
)), ('iabs', a
)),
512 (('~fmax', ('fabs', a
), 0.0), ('fabs', a
)),
513 (('~fmin', ('fmax', a
, 0.0), 1.0), ('fsat', a
), '!options->lower_fsat'),
514 (('~fmax', ('fmin', a
, 1.0), 0.0), ('fsat', a
), '!options->lower_fsat'),
515 (('~fmin', ('fmax', a
, -1.0), 0.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
516 (('~fmax', ('fmin', a
, 0.0), -1.0), ('fneg', ('fsat', ('fneg', a
))), '!options->lower_fsat'),
517 (('fsat', ('fsign', a
)), ('b2f', ('flt', 0.0, a
))),
518 (('fsat', ('b2f', a
)), ('b2f', a
)),
519 (('fsat', a
), ('fmin', ('fmax', a
, 0.0), 1.0), 'options->lower_fsat'),
520 (('fsat', ('fsat', a
)), ('fsat', a
)),
521 (('fsat', ('fneg(is_used_once)', ('fadd(is_used_once)', a
, b
))), ('fsat', ('fadd', ('fneg', a
), ('fneg', b
))), '!options->lower_fsat'),
522 (('fsat', ('fneg(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fneg', a
), b
)), '!options->lower_fsat'),
523 (('fsat', ('fabs(is_used_once)', ('fmul(is_used_once)', a
, b
))), ('fsat', ('fmul', ('fabs', a
), ('fabs', b
))), '!options->lower_fsat'),
524 (('fmin', ('fmax', ('fmin', ('fmax', a
, b
), c
), b
), c
), ('fmin', ('fmax', a
, b
), c
)),
525 (('imin', ('imax', ('imin', ('imax', a
, b
), c
), b
), c
), ('imin', ('imax', a
, b
), c
)),
526 (('umin', ('umax', ('umin', ('umax', a
, b
), c
), b
), c
), ('umin', ('umax', a
, b
), c
)),
527 (('fmax', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a
, b
))),
528 (('fmin', ('fsat', a
), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a
, b
))),
529 (('extract_u8', ('imin', ('imax', a
, 0), 0xff), 0), ('imin', ('imax', a
, 0), 0xff)),
530 (('~ior', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmax', b
, c
))),
531 (('~ior', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmin', a
, b
), c
)),
532 (('~ior', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmin', b
, c
))),
533 (('~ior', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmax', a
, b
), c
)),
534 (('~ior', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmax', b
, c
))),
535 (('~ior', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmin', a
, b
), c
)),
536 (('~ior', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmin', b
, c
))),
537 (('~ior', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmax', a
, b
), c
)),
538 (('~iand', ('flt(is_used_once)', a
, b
), ('flt', a
, c
)), ('flt', a
, ('fmin', b
, c
))),
539 (('~iand', ('flt(is_used_once)', a
, c
), ('flt', b
, c
)), ('flt', ('fmax', a
, b
), c
)),
540 (('~iand', ('fge(is_used_once)', a
, b
), ('fge', a
, c
)), ('fge', a
, ('fmax', b
, c
))),
541 (('~iand', ('fge(is_used_once)', a
, c
), ('fge', b
, c
)), ('fge', ('fmin', a
, b
), c
)),
542 (('~iand', ('flt', a
, '#b'), ('flt', a
, '#c')), ('flt', a
, ('fmin', b
, c
))),
543 (('~iand', ('flt', '#a', c
), ('flt', '#b', c
)), ('flt', ('fmax', a
, b
), c
)),
544 (('~iand', ('fge', a
, '#b'), ('fge', a
, '#c')), ('fge', a
, ('fmax', b
, c
))),
545 (('~iand', ('fge', '#a', c
), ('fge', '#b', c
)), ('fge', ('fmin', a
, b
), c
)),
547 (('ior', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imax', b
, c
))),
548 (('ior', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imin', a
, b
), c
)),
549 (('ior', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imin', b
, c
))),
550 (('ior', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imax', a
, b
), c
)),
551 (('ior', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umax', b
, c
))),
552 (('ior', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umin', a
, b
), c
)),
553 (('ior', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umin', b
, c
))),
554 (('ior', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umax', a
, b
), c
)),
555 (('iand', ('ilt(is_used_once)', a
, b
), ('ilt', a
, c
)), ('ilt', a
, ('imin', b
, c
))),
556 (('iand', ('ilt(is_used_once)', a
, c
), ('ilt', b
, c
)), ('ilt', ('imax', a
, b
), c
)),
557 (('iand', ('ige(is_used_once)', a
, b
), ('ige', a
, c
)), ('ige', a
, ('imax', b
, c
))),
558 (('iand', ('ige(is_used_once)', a
, c
), ('ige', b
, c
)), ('ige', ('imin', a
, b
), c
)),
559 (('iand', ('ult(is_used_once)', a
, b
), ('ult', a
, c
)), ('ult', a
, ('umin', b
, c
))),
560 (('iand', ('ult(is_used_once)', a
, c
), ('ult', b
, c
)), ('ult', ('umax', a
, b
), c
)),
561 (('iand', ('uge(is_used_once)', a
, b
), ('uge', a
, c
)), ('uge', a
, ('umax', b
, c
))),
562 (('iand', ('uge(is_used_once)', a
, c
), ('uge', b
, c
)), ('uge', ('umin', a
, b
), c
)),
564 # These derive from the previous patterns with the application of b < 0 <=>
565 # 0 < -b. The transformation should be applied if either comparison is
566 # used once as this ensures that the number of comparisons will not
567 # increase. The sources to the ior and iand are not symmetric, so the
568 # rules have to be duplicated to get this behavior.
569 (('~ior', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
570 (('~ior', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmax', a
, ('fneg', b
)))),
571 (('~ior', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
572 (('~ior', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmin', a
, ('fneg', b
)))),
573 (('~iand', ('flt(is_used_once)', 0.0, 'a@32'), ('flt', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
574 (('~iand', ('flt', 0.0, 'a@32'), ('flt(is_used_once)', 'b@32', 0.0)), ('flt', 0.0, ('fmin', a
, ('fneg', b
)))),
575 (('~iand', ('fge(is_used_once)', 0.0, 'a@32'), ('fge', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
576 (('~iand', ('fge', 0.0, 'a@32'), ('fge(is_used_once)', 'b@32', 0.0)), ('fge', 0.0, ('fmax', a
, ('fneg', b
)))),
578 # Common pattern like 'if (i == 0 || i == 1 || ...)'
579 (('ior', ('ieq', a
, 0), ('ieq', a
, 1)), ('uge', 1, a
)),
580 (('ior', ('uge', 1, a
), ('ieq', a
, 2)), ('uge', 2, a
)),
581 (('ior', ('uge', 2, a
), ('ieq', a
, 3)), ('uge', 3, a
)),
583 # The (i2f32, ...) part is an open-coded fsign. When that is combined with
584 # the bcsel, it's basically copysign(1.0, a). There is no copysign in NIR,
585 # so emit an open-coded version of that.
586 (('bcsel@32', ('feq', a
, 0.0), 1.0, ('i2f32', ('iadd', ('b2i32', ('flt', 0.0, 'a@32')), ('ineg', ('b2i32', ('flt', 'a@32', 0.0)))))),
587 ('ior', 0x3f800000, ('iand', a
, 0x80000000))),
589 (('ior', a
, ('ieq', a
, False)), True),
590 (('ior', a
, ('inot', a
)), -1),
592 (('ine', ('ineg', ('b2i32', 'a@1')), ('ineg', ('b2i32', 'b@1'))), ('ine', a
, b
)),
593 (('b2i32', ('ine', 'a@1', 'b@1')), ('b2i32', ('ixor', a
, b
))),
595 (('iand', ('ieq', 'a@32', 0), ('ieq', 'b@32', 0)), ('ieq', ('ior', 'a@32', 'b@32'), 0), '!options->lower_bitops'),
597 # These patterns can result when (a < b || a < c) => (a < min(b, c))
598 # transformations occur before constant propagation and loop-unrolling.
599 (('~flt', a
, ('fmax', b
, a
)), ('flt', a
, b
)),
600 (('~flt', ('fmin', a
, b
), a
), ('flt', b
, a
)),
601 (('~fge', a
, ('fmin', b
, a
)), True),
602 (('~fge', ('fmax', a
, b
), a
), True),
603 (('~flt', a
, ('fmin', b
, a
)), False),
604 (('~flt', ('fmax', a
, b
), a
), False),
605 (('~fge', a
, ('fmax', b
, a
)), ('fge', a
, b
)),
606 (('~fge', ('fmin', a
, b
), a
), ('fge', b
, a
)),
608 (('ilt', a
, ('imax', b
, a
)), ('ilt', a
, b
)),
609 (('ilt', ('imin', a
, b
), a
), ('ilt', b
, a
)),
610 (('ige', a
, ('imin', b
, a
)), True),
611 (('ige', ('imax', a
, b
), a
), True),
612 (('ult', a
, ('umax', b
, a
)), ('ult', a
, b
)),
613 (('ult', ('umin', a
, b
), a
), ('ult', b
, a
)),
614 (('uge', a
, ('umin', b
, a
)), True),
615 (('uge', ('umax', a
, b
), a
), True),
616 (('ilt', a
, ('imin', b
, a
)), False),
617 (('ilt', ('imax', a
, b
), a
), False),
618 (('ige', a
, ('imax', b
, a
)), ('ige', a
, b
)),
619 (('ige', ('imin', a
, b
), a
), ('ige', b
, a
)),
620 (('ult', a
, ('umin', b
, a
)), False),
621 (('ult', ('umax', a
, b
), a
), False),
622 (('uge', a
, ('umax', b
, a
)), ('uge', a
, b
)),
623 (('uge', ('umin', a
, b
), a
), ('uge', b
, a
)),
624 (('ult', a
, ('iand', b
, a
)), False),
625 (('ult', ('ior', a
, b
), a
), False),
626 (('uge', a
, ('iand', b
, a
)), True),
627 (('uge', ('ior', a
, b
), a
), True),
629 (('ilt', '#a', ('imax', '#b', c
)), ('ior', ('ilt', a
, b
), ('ilt', a
, c
))),
630 (('ilt', ('imin', '#a', b
), '#c'), ('ior', ('ilt', a
, c
), ('ilt', b
, c
))),
631 (('ige', '#a', ('imin', '#b', c
)), ('ior', ('ige', a
, b
), ('ige', a
, c
))),
632 (('ige', ('imax', '#a', b
), '#c'), ('ior', ('ige', a
, c
), ('ige', b
, c
))),
633 (('ult', '#a', ('umax', '#b', c
)), ('ior', ('ult', a
, b
), ('ult', a
, c
))),
634 (('ult', ('umin', '#a', b
), '#c'), ('ior', ('ult', a
, c
), ('ult', b
, c
))),
635 (('uge', '#a', ('umin', '#b', c
)), ('ior', ('uge', a
, b
), ('uge', a
, c
))),
636 (('uge', ('umax', '#a', b
), '#c'), ('ior', ('uge', a
, c
), ('uge', b
, c
))),
637 (('ilt', '#a', ('imin', '#b', c
)), ('iand', ('ilt', a
, b
), ('ilt', a
, c
))),
638 (('ilt', ('imax', '#a', b
), '#c'), ('iand', ('ilt', a
, c
), ('ilt', b
, c
))),
639 (('ige', '#a', ('imax', '#b', c
)), ('iand', ('ige', a
, b
), ('ige', a
, c
))),
640 (('ige', ('imin', '#a', b
), '#c'), ('iand', ('ige', a
, c
), ('ige', b
, c
))),
641 (('ult', '#a', ('umin', '#b', c
)), ('iand', ('ult', a
, b
), ('ult', a
, c
))),
642 (('ult', ('umax', '#a', b
), '#c'), ('iand', ('ult', a
, c
), ('ult', b
, c
))),
643 (('uge', '#a', ('umax', '#b', c
)), ('iand', ('uge', a
, b
), ('uge', a
, c
))),
644 (('uge', ('umin', '#a', b
), '#c'), ('iand', ('uge', a
, c
), ('uge', b
, c
))),
646 # Thanks to sign extension, the ishr(a, b) is negative if and only if a is
648 (('bcsel', ('ilt', a
, 0), ('ineg', ('ishr', a
, b
)), ('ishr', a
, b
)),
649 ('iabs', ('ishr', a
, b
))),
650 (('iabs', ('ishr', ('iabs', a
), b
)), ('ishr', ('iabs', a
), b
)),
652 (('fabs', ('slt', a
, b
)), ('slt', a
, b
)),
653 (('fabs', ('sge', a
, b
)), ('sge', a
, b
)),
654 (('fabs', ('seq', a
, b
)), ('seq', a
, b
)),
655 (('fabs', ('sne', a
, b
)), ('sne', a
, b
)),
656 (('slt', a
, b
), ('b2f', ('flt', a
, b
)), 'options->lower_scmp'),
657 (('sge', a
, b
), ('b2f', ('fge', a
, b
)), 'options->lower_scmp'),
658 (('seq', a
, b
), ('b2f', ('feq', a
, b
)), 'options->lower_scmp'),
659 (('sne', a
, b
), ('b2f', ('fne', a
, b
)), 'options->lower_scmp'),
660 (('seq', ('seq', a
, b
), 1.0), ('seq', a
, b
)),
661 (('seq', ('sne', a
, b
), 1.0), ('sne', a
, b
)),
662 (('seq', ('slt', a
, b
), 1.0), ('slt', a
, b
)),
663 (('seq', ('sge', a
, b
), 1.0), ('sge', a
, b
)),
664 (('sne', ('seq', a
, b
), 0.0), ('seq', a
, b
)),
665 (('sne', ('sne', a
, b
), 0.0), ('sne', a
, b
)),
666 (('sne', ('slt', a
, b
), 0.0), ('slt', a
, b
)),
667 (('sne', ('sge', a
, b
), 0.0), ('sge', a
, b
)),
668 (('seq', ('seq', a
, b
), 0.0), ('sne', a
, b
)),
669 (('seq', ('sne', a
, b
), 0.0), ('seq', a
, b
)),
670 (('seq', ('slt', a
, b
), 0.0), ('sge', a
, b
)),
671 (('seq', ('sge', a
, b
), 0.0), ('slt', a
, b
)),
672 (('sne', ('seq', a
, b
), 1.0), ('sne', a
, b
)),
673 (('sne', ('sne', a
, b
), 1.0), ('seq', a
, b
)),
674 (('sne', ('slt', a
, b
), 1.0), ('sge', a
, b
)),
675 (('sne', ('sge', a
, b
), 1.0), ('slt', a
, b
)),
676 (('fall_equal2', a
, b
), ('fmin', ('seq', 'a.x', 'b.x'), ('seq', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
677 (('fall_equal3', a
, b
), ('seq', ('fany_nequal3', a
, b
), 0.0), 'options->lower_vector_cmp'),
678 (('fall_equal4', a
, b
), ('seq', ('fany_nequal4', a
, b
), 0.0), 'options->lower_vector_cmp'),
679 (('fany_nequal2', a
, b
), ('fmax', ('sne', 'a.x', 'b.x'), ('sne', 'a.y', 'b.y')), 'options->lower_vector_cmp'),
680 (('fany_nequal3', a
, b
), ('fsat', ('fdot3', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
681 (('fany_nequal4', a
, b
), ('fsat', ('fdot4', ('sne', a
, b
), ('sne', a
, b
))), 'options->lower_vector_cmp'),
682 (('fne', ('fneg', a
), a
), ('fne', a
, 0.0)),
683 (('feq', ('fneg', a
), a
), ('feq', a
, 0.0)),
685 (('imul', ('b2i', 'a@1'), ('b2i', 'b@1')), ('b2i', ('iand', a
, b
))),
686 (('fmul', ('b2f', 'a@1'), ('b2f', 'b@1')), ('b2f', ('iand', a
, b
))),
687 (('fsat', ('fadd', ('b2f', 'a@1'), ('b2f', 'b@1'))), ('b2f', ('ior', a
, b
))),
688 (('iand', 'a@bool32', 1.0), ('b2f', a
)),
689 # True/False are ~0 and 0 in NIR. b2i of True is 1, and -1 is ~0 (True).
690 (('ineg', ('b2i32', 'a@32')), a
),
691 (('flt', ('fneg', ('b2f', 'a@1')), 0), a
), # Generated by TGSI KILL_IF.
692 # Comparison with the same args. Note that these are not done for
693 # the float versions because NaN always returns false on float
695 (('ilt', a
, a
), False),
696 (('ige', a
, a
), True),
697 (('ieq', a
, a
), True),
698 (('ine', a
, a
), False),
699 (('ult', a
, a
), False),
700 (('uge', a
, a
), True),
701 # Logical and bit operations
703 (('iand', a
, ~
0), a
),
707 (('ior', a
, True), True),
710 (('inot', ('inot', a
)), a
),
711 (('ior', ('iand', a
, b
), b
), b
),
712 (('ior', ('ior', a
, b
), b
), ('ior', a
, b
)),
713 (('iand', ('ior', a
, b
), b
), b
),
714 (('iand', ('iand', a
, b
), b
), ('iand', a
, b
)),
716 (('iand', ('inot', a
), ('inot', b
)), ('inot', ('ior', a
, b
))),
717 (('ior', ('inot', a
), ('inot', b
)), ('inot', ('iand', a
, b
))),
718 # Shift optimizations
725 (('iand', 0xff, ('ushr@32', a
, 24)), ('ushr', a
, 24)),
726 (('iand', 0xffff, ('ushr@32', a
, 16)), ('ushr', a
, 16)),
727 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('iadd', 16, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
728 (('ior', ('ishl@16', a
, b
), ('ushr@16', a
, ('isub', 16, b
))), ('urol', a
, b
), '!options->lower_rotate'),
729 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('iadd', 32, ('ineg', b
)))), ('urol', a
, b
), '!options->lower_rotate'),
730 (('ior', ('ishl@32', a
, b
), ('ushr@32', a
, ('isub', 32, b
))), ('urol', a
, b
), '!options->lower_rotate'),
731 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('iadd', 16, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
732 (('ior', ('ushr@16', a
, b
), ('ishl@16', a
, ('isub', 16, b
))), ('uror', a
, b
), '!options->lower_rotate'),
733 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('iadd', 32, ('ineg', b
)))), ('uror', a
, b
), '!options->lower_rotate'),
734 (('ior', ('ushr@32', a
, b
), ('ishl@32', a
, ('isub', 32, b
))), ('uror', a
, b
), '!options->lower_rotate'),
735 (('urol@16', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 16, b
))), 'options->lower_rotate'),
736 (('urol@32', a
, b
), ('ior', ('ishl', a
, b
), ('ushr', a
, ('isub', 32, b
))), 'options->lower_rotate'),
737 (('uror@16', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 16, b
))), 'options->lower_rotate'),
738 (('uror@32', a
, b
), ('ior', ('ushr', a
, b
), ('ishl', a
, ('isub', 32, b
))), 'options->lower_rotate'),
739 # Exponential/logarithmic identities
740 (('~fexp2', ('flog2', a
)), a
), # 2^lg2(a) = a
741 (('~flog2', ('fexp2', a
)), a
), # lg2(2^a) = a
742 (('fpow', a
, b
), ('fexp2', ('fmul', ('flog2', a
), b
)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
743 (('~fexp2', ('fmul', ('flog2', a
), b
)), ('fpow', a
, b
), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
744 (('~fexp2', ('fadd', ('fmul', ('flog2', a
), b
), ('fmul', ('flog2', c
), d
))),
745 ('~fmul', ('fpow', a
, b
), ('fpow', c
, d
)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
746 (('~fexp2', ('fmul', ('flog2', a
), 2.0)), ('fmul', a
, a
)),
747 (('~fexp2', ('fmul', ('flog2', a
), 4.0)), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
748 (('~fpow', a
, 1.0), a
),
749 (('~fpow', a
, 2.0), ('fmul', a
, a
)),
750 (('~fpow', a
, 4.0), ('fmul', ('fmul', a
, a
), ('fmul', a
, a
))),
751 (('~fpow', 2.0, a
), ('fexp2', a
)),
752 (('~fpow', ('fpow', a
, 2.2), 0.454545), a
),
753 (('~fpow', ('fabs', ('fpow', a
, 2.2)), 0.454545), ('fabs', a
)),
754 (('~fsqrt', ('fexp2', a
)), ('fexp2', ('fmul', 0.5, a
))),
755 (('~frcp', ('fexp2', a
)), ('fexp2', ('fneg', a
))),
756 (('~frsq', ('fexp2', a
)), ('fexp2', ('fmul', -0.5, a
))),
757 (('~flog2', ('fsqrt', a
)), ('fmul', 0.5, ('flog2', a
))),
758 (('~flog2', ('frcp', a
)), ('fneg', ('flog2', a
))),
759 (('~flog2', ('frsq', a
)), ('fmul', -0.5, ('flog2', a
))),
760 (('~flog2', ('fpow', a
, b
)), ('fmul', b
, ('flog2', a
))),
761 (('~fmul', ('fexp2(is_used_once)', a
), ('fexp2(is_used_once)', b
)), ('fexp2', ('fadd', a
, b
))),
762 (('bcsel', ('flt', a
, 0.0), 0.0, ('fsqrt', a
)), ('fsqrt', ('fmax', a
, 0.0))),
763 (('~fmul', ('fsqrt', a
), ('fsqrt', a
)), ('fabs',a
)),
764 # Division and reciprocal
765 (('~fdiv', 1.0, a
), ('frcp', a
)),
766 (('fdiv', a
, b
), ('fmul', a
, ('frcp', b
)), 'options->lower_fdiv'),
767 (('~frcp', ('frcp', a
)), a
),
768 (('~frcp', ('fsqrt', a
)), ('frsq', a
)),
769 (('fsqrt', a
), ('frcp', ('frsq', a
)), 'options->lower_fsqrt'),
770 (('~frcp', ('frsq', a
)), ('fsqrt', a
), '!options->lower_fsqrt'),
772 (('fsin', a
), lowered_sincos(0.5), 'options->lower_sincos'),
773 (('fcos', a
), lowered_sincos(0.75), 'options->lower_sincos'),
774 # Boolean simplifications
775 (('i2b32(is_used_by_if)', a
), ('ine32', a
, 0)),
776 (('i2b1(is_used_by_if)', a
), ('ine', a
, 0)),
777 (('ieq', a
, True), a
),
778 (('ine(is_not_used_by_if)', a
, True), ('inot', a
)),
779 (('ine', a
, False), a
),
780 (('ieq(is_not_used_by_if)', a
, False), ('inot', 'a')),
781 (('bcsel', a
, True, False), a
),
782 (('bcsel', a
, False, True), ('inot', a
)),
783 (('bcsel@32', a
, 1.0, 0.0), ('b2f', a
)),
784 (('bcsel@32', a
, 0.0, 1.0), ('b2f', ('inot', a
))),
785 (('bcsel@32', a
, -1.0, -0.0), ('fneg', ('b2f', a
))),
786 (('bcsel@32', a
, -0.0, -1.0), ('fneg', ('b2f', ('inot', a
)))),
787 (('bcsel', True, b
, c
), b
),
788 (('bcsel', False, b
, c
), c
),
789 (('bcsel', a
, ('b2f(is_used_once)', 'b@32'), ('b2f', 'c@32')), ('b2f', ('bcsel', a
, b
, c
))),
791 (('bcsel', a
, b
, b
), b
),
792 (('~fcsel', a
, b
, b
), b
),
794 # D3D Boolean emulation
795 (('bcsel', a
, -1, 0), ('ineg', ('b2i', 'a@1'))),
796 (('bcsel', a
, 0, -1), ('ineg', ('b2i', ('inot', a
)))),
797 (('iand', ('ineg', ('b2i', 'a@1')), ('ineg', ('b2i', 'b@1'))),
798 ('ineg', ('b2i', ('iand', a
, b
)))),
799 (('ior', ('ineg', ('b2i','a@1')), ('ineg', ('b2i', 'b@1'))),
800 ('ineg', ('b2i', ('ior', a
, b
)))),
801 (('ieq', ('ineg', ('b2i', 'a@1')), 0), ('inot', a
)),
802 (('ieq', ('ineg', ('b2i', 'a@1')), -1), a
),
803 (('ine', ('ineg', ('b2i', 'a@1')), 0), a
),
804 (('ine', ('ineg', ('b2i', 'a@1')), -1), ('inot', a
)),
805 (('iand', ('ineg', ('b2i', a
)), 1.0), ('b2f', a
)),
806 (('iand', ('ineg', ('b2i', a
)), 1), ('b2i', a
)),
808 # SM5 32-bit shifts are defined to use the 5 least significant bits
809 (('ishl', 'a@32', ('iand', 31, b
)), ('ishl', a
, b
)),
810 (('ishr', 'a@32', ('iand', 31, b
)), ('ishr', a
, b
)),
811 (('ushr', 'a@32', ('iand', 31, b
)), ('ushr', a
, b
)),
814 (('i2b32', ('b2i', 'a@32')), a
),
815 (('f2i', ('ftrunc', a
)), ('f2i', a
)),
816 (('f2u', ('ftrunc', a
)), ('f2u', a
)),
817 (('i2b', ('ineg', a
)), ('i2b', a
)),
818 (('i2b', ('iabs', a
)), ('i2b', a
)),
819 (('inot', ('f2b1', a
)), ('feq', a
, 0.0)),
821 # The C spec says, "If the value of the integral part cannot be represented
822 # by the integer type, the behavior is undefined." "Undefined" can mean
823 # "the conversion doesn't happen at all."
824 (('~i2f32', ('f2i32', 'a@32')), ('ftrunc', a
)),
826 # Ironically, mark these as imprecise because removing the conversions may
827 # preserve more precision than doing the conversions (e.g.,
828 # uint(float(0x81818181u)) == 0x81818200).
829 (('~f2i32', ('i2f', 'a@32')), a
),
830 (('~f2i32', ('u2f', 'a@32')), a
),
831 (('~f2u32', ('i2f', 'a@32')), a
),
832 (('~f2u32', ('u2f', 'a@32')), a
),
834 # Conversions from float16 to float32 and back can always be removed
835 (('f2f16', ('f2f32', 'a@16')), a
),
836 (('f2fmp', ('f2f32', 'a@16')), a
),
837 # Conversions to float16 would be lossy so they should only be removed if
838 # the instruction was generated by the precision lowering pass.
839 (('f2f32', ('f2fmp', 'a@32')), a
),
841 (('ffloor', 'a(is_integral)'), a
),
842 (('fceil', 'a(is_integral)'), a
),
843 (('ftrunc', 'a(is_integral)'), a
),
844 # fract(x) = x - floor(x), so fract(NaN) = NaN
845 (('~ffract', 'a(is_integral)'), 0.0),
846 (('fabs', 'a(is_not_negative)'), a
),
847 (('iabs', 'a(is_not_negative)'), a
),
848 (('fsat', 'a(is_not_positive)'), 0.0),
850 # Section 5.4.1 (Conversion and Scalar Constructors) of the GLSL 4.60 spec
853 # It is undefined to convert a negative floating-point value to an
856 # Assuming that (uint)some_float behaves like (uint)(int)some_float allows
857 # some optimizations in the i965 backend to proceed.
858 (('ige', ('f2u', a
), b
), ('ige', ('f2i', a
), b
)),
859 (('ige', b
, ('f2u', a
)), ('ige', b
, ('f2i', a
))),
860 (('ilt', ('f2u', a
), b
), ('ilt', ('f2i', a
), b
)),
861 (('ilt', b
, ('f2u', a
)), ('ilt', b
, ('f2i', a
))),
863 (('~fmin', 'a(is_not_negative)', 1.0), ('fsat', a
), '!options->lower_fsat'),
865 # The result of the multiply must be in [-1, 0], so the result of the ffma
867 (('flt', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), False),
868 (('flt', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), False),
869 (('fmax', ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0), 0.0), ('fadd', ('fmul', ('fsat', a
), ('fneg', ('fsat', a
))), 1.0)),
870 (('fmax', ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0), 0.0), ('fadd', ('fneg', ('fmul', ('fsat', a
), ('fsat', a
))), 1.0)),
872 (('fne', 'a(is_not_zero)', 0.0), True),
873 (('feq', 'a(is_not_zero)', 0.0), False),
875 # In this chart, + means value > 0 and - means value < 0.
877 # + >= + -> unknown 0 >= + -> false - >= + -> false
878 # + >= 0 -> true 0 >= 0 -> true - >= 0 -> false
879 # + >= - -> true 0 >= - -> true - >= - -> unknown
881 # Using grouping conceptually similar to a Karnaugh map...
883 # (+ >= 0, + >= -, 0 >= 0, 0 >= -) == (is_not_negative >= is_not_positive) -> true
884 # (0 >= +, - >= +) == (is_not_positive >= gt_zero) -> false
885 # (- >= +, - >= 0) == (lt_zero >= is_not_negative) -> false
887 # The flt / ilt cases just invert the expected result.
889 # The results expecting true, must be marked imprecise. The results
890 # expecting false are fine because NaN compared >= or < anything is false.
892 (('~fge', 'a(is_not_negative)', 'b(is_not_positive)'), True),
893 (('fge', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
894 (('fge', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
896 (('flt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
897 (('~flt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
898 (('~flt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
900 (('ine', 'a(is_not_zero)', 0), True),
901 (('ieq', 'a(is_not_zero)', 0), False),
903 (('ige', 'a(is_not_negative)', 'b(is_not_positive)'), True),
904 (('ige', 'a(is_not_positive)', 'b(is_gt_zero)'), False),
905 (('ige', 'a(is_lt_zero)', 'b(is_not_negative)'), False),
907 (('ilt', 'a(is_not_negative)', 'b(is_not_positive)'), False),
908 (('ilt', 'a(is_not_positive)', 'b(is_gt_zero)'), True),
909 (('ilt', 'a(is_lt_zero)', 'b(is_not_negative)'), True),
911 (('ult', 0, 'a(is_gt_zero)'), True),
913 # Packing and then unpacking does nothing
914 (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a
, b
)), a
),
915 (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a
, b
)), b
),
916 (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a
),
917 ('unpack_64_2x32_split_y', a
)), a
),
919 # Comparing two halves of an unpack separately. While this optimization
920 # should be correct for non-constant values, it's less obvious that it's
921 # useful in that case. For constant values, the pack will fold and we're
922 # guaranteed to reduce the whole tree to one instruction.
923 (('iand', ('ieq', ('unpack_32_2x16_split_x', a
), '#b'),
924 ('ieq', ('unpack_32_2x16_split_y', a
), '#c')),
925 ('ieq', a
, ('pack_32_2x16_split', b
, c
))),
928 (('ushr', 'a@16', 8), ('extract_u8', a
, 1), '!options->lower_extract_byte'),
929 (('ushr', 'a@32', 24), ('extract_u8', a
, 3), '!options->lower_extract_byte'),
930 (('ushr', 'a@64', 56), ('extract_u8', a
, 7), '!options->lower_extract_byte'),
931 (('ishr', 'a@16', 8), ('extract_i8', a
, 1), '!options->lower_extract_byte'),
932 (('ishr', 'a@32', 24), ('extract_i8', a
, 3), '!options->lower_extract_byte'),
933 (('ishr', 'a@64', 56), ('extract_i8', a
, 7), '!options->lower_extract_byte'),
934 (('iand', 0xff, a
), ('extract_u8', a
, 0), '!options->lower_extract_byte'),
936 # Useless masking before unpacking
937 (('unpack_half_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_half_2x16_split_x', a
)),
938 (('unpack_32_2x16_split_x', ('iand', a
, 0xffff)), ('unpack_32_2x16_split_x', a
)),
939 (('unpack_64_2x32_split_x', ('iand', a
, 0xffffffff)), ('unpack_64_2x32_split_x', a
)),
940 (('unpack_half_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_half_2x16_split_y', a
)),
941 (('unpack_32_2x16_split_y', ('iand', a
, 0xffff0000)), ('unpack_32_2x16_split_y', a
)),
942 (('unpack_64_2x32_split_y', ('iand', a
, 0xffffffff00000000)), ('unpack_64_2x32_split_y', a
)),
944 # Optimize half packing
945 (('ishl', ('pack_half_2x16', ('vec2', a
, 0)), 16), ('pack_half_2x16', ('vec2', 0, a
))),
946 (('ishr', ('pack_half_2x16', ('vec2', 0, a
)), 16), ('pack_half_2x16', ('vec2', a
, 0))),
948 (('iadd', ('pack_half_2x16', ('vec2', a
, 0)), ('pack_half_2x16', ('vec2', 0, b
))),
949 ('pack_half_2x16', ('vec2', a
, b
))),
950 (('ior', ('pack_half_2x16', ('vec2', a
, 0)), ('pack_half_2x16', ('vec2', 0, b
))),
951 ('pack_half_2x16', ('vec2', a
, b
))),
954 # After the ('extract_u8', a, 0) pattern, above, triggers, there will be
955 # patterns like those below.
956 for op
in ('ushr', 'ishr'):
957 optimizations
.extend([(('extract_u8', (op
, 'a@16', 8), 0), ('extract_u8', a
, 1))])
958 optimizations
.extend([(('extract_u8', (op
, 'a@32', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 4)])
959 optimizations
.extend([(('extract_u8', (op
, 'a@64', 8 * i
), 0), ('extract_u8', a
, i
)) for i
in range(1, 8)])
961 optimizations
.extend([(('extract_u8', ('extract_u16', a
, 1), 0), ('extract_u8', a
, 2))])
963 # After the ('extract_[iu]8', a, 3) patterns, above, trigger, there will be
964 # patterns like those below.
965 for op
in ('extract_u8', 'extract_i8'):
966 optimizations
.extend([((op
, ('ishl', 'a@16', 8), 1), (op
, a
, 0))])
967 optimizations
.extend([((op
, ('ishl', 'a@32', 24 - 8 * i
), 3), (op
, a
, i
)) for i
in range(2, -1, -1)])
968 optimizations
.extend([((op
, ('ishl', 'a@64', 56 - 8 * i
), 7), (op
, a
, i
)) for i
in range(6, -1, -1)])
970 optimizations
.extend([
972 (('ushr', ('ishl', 'a@32', 16), 16), ('extract_u16', a
, 0), '!options->lower_extract_word'),
973 (('ushr', 'a@32', 16), ('extract_u16', a
, 1), '!options->lower_extract_word'),
974 (('ishr', ('ishl', 'a@32', 16), 16), ('extract_i16', a
, 0), '!options->lower_extract_word'),
975 (('ishr', 'a@32', 16), ('extract_i16', a
, 1), '!options->lower_extract_word'),
976 (('iand', 0xffff, a
), ('extract_u16', a
, 0), '!options->lower_extract_word'),
979 (('ussub_4x8', a
, 0), a
),
980 (('ussub_4x8', a
, ~
0), 0),
981 # Lower all Subtractions first - they can get recombined later
982 (('fsub', a
, b
), ('fadd', a
, ('fneg', b
))),
983 (('isub', a
, b
), ('iadd', a
, ('ineg', b
))),
984 (('uabs_usub', a
, b
), ('bcsel', ('ult', a
, b
), ('ineg', ('isub', a
, b
)), ('isub', a
, b
))),
985 # This is correct. We don't need isub_sat because the result type is unsigned, so it cannot overflow.
986 (('uabs_isub', a
, b
), ('bcsel', ('ilt', a
, b
), ('ineg', ('isub', a
, b
)), ('isub', a
, b
))),
988 # Propagate negation up multiplication chains
989 (('fmul(is_used_by_non_fsat)', ('fneg', a
), b
), ('fneg', ('fmul', a
, b
))),
990 (('imul', ('ineg', a
), b
), ('ineg', ('imul', a
, b
))),
992 # Propagate constants up multiplication chains
993 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a
, c
), b
)),
994 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a
, c
), b
)),
995 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a
, c
), b
)),
996 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a
, c
), b
)),
998 # Reassociate constants in add/mul chains so they can be folded together.
999 # For now, we mostly only handle cases where the constants are separated by
1000 # a single non-constant. We could do better eventually.
1001 (('~fmul', '#a', ('fmul', 'b(is_not_const)', '#c')), ('fmul', ('fmul', a
, c
), b
)),
1002 (('imul', '#a', ('imul', 'b(is_not_const)', '#c')), ('imul', ('imul', a
, c
), b
)),
1003 (('~fadd', '#a', ('fadd', 'b(is_not_const)', '#c')), ('fadd', ('fadd', a
, c
), b
)),
1004 (('~fadd', '#a', ('fneg', ('fadd', 'b(is_not_const)', '#c'))), ('fadd', ('fadd', a
, ('fneg', c
)), ('fneg', b
))),
1005 (('iadd', '#a', ('iadd', 'b(is_not_const)', '#c')), ('iadd', ('iadd', a
, c
), b
)),
1007 # Drop mul-div by the same value when there's no wrapping.
1008 (('idiv', ('imul(no_signed_wrap)', a
, b
), b
), a
),
1011 (('bcsel', ('ige', ('find_lsb', a
), 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
1012 (('bcsel', ('ige', ('ifind_msb', a
), 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
1013 (('bcsel', ('ige', ('ufind_msb', a
), 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
1015 (('bcsel', ('ine', a
, 0), ('find_lsb', a
), -1), ('find_lsb', a
)),
1016 (('bcsel', ('ine', a
, 0), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
1017 (('bcsel', ('ine', a
, 0), ('ufind_msb', a
), -1), ('ufind_msb', a
)),
1019 (('bcsel', ('ine', a
, -1), ('ifind_msb', a
), -1), ('ifind_msb', a
)),
1022 (('fmod', a
, b
), ('fsub', a
, ('fmul', b
, ('ffloor', ('fdiv', a
, b
)))), 'options->lower_fmod'),
1023 (('frem', a
, b
), ('fsub', a
, ('fmul', b
, ('ftrunc', ('fdiv', a
, b
)))), 'options->lower_fmod'),
1024 (('uadd_carry@32', a
, b
), ('b2i', ('ult', ('iadd', a
, b
), a
)), 'options->lower_uadd_carry'),
1025 (('usub_borrow@32', a
, b
), ('b2i', ('ult', a
, b
)), 'options->lower_usub_borrow'),
1027 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1028 ('bcsel', ('ult', 31, 'bits'), 'insert',
1029 ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
1030 'options->lower_bitfield_insert'),
1031 (('ihadd', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1032 (('uhadd', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1033 (('irhadd', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1034 (('urhadd', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd'),
1035 (('ihadd@64', a
, b
), ('iadd', ('iand', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1036 (('uhadd@64', a
, b
), ('iadd', ('iand', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1037 (('irhadd@64', a
, b
), ('isub', ('ior', a
, b
), ('ishr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1038 (('urhadd@64', a
, b
), ('isub', ('ior', a
, b
), ('ushr', ('ixor', a
, b
), 1)), 'options->lower_hadd64 || (options->lower_int64_options & nir_lower_iadd64) != 0'),
1040 (('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'),
1041 (('uadd_sat', a
, b
), ('bcsel', ('ult', ('iadd', a
, b
), a
), -1, ('iadd', a
, b
)), 'options->lower_add_sat'),
1042 (('usub_sat', a
, b
), ('bcsel', ('ult', a
, b
), 0, ('isub', a
, b
)), 'options->lower_add_sat'),
1043 (('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'),
1045 # int64_t sum = a + b;
1047 # if (a < 0 && b < 0 && a < sum)
1049 # } else if (a >= 0 && b >= 0 && sum < a)
1053 # A couple optimizations are applied.
1055 # 1. a < sum => sum >= 0. This replacement works because it is known that
1056 # a < 0 and b < 0, so sum should also be < 0 unless there was
1059 # 2. sum < a => sum < 0. This replacement works because it is known that
1060 # a >= 0 and b >= 0, so sum should also be >= 0 unless there was
1063 # 3. Invert the second if-condition and swap the order of parameters for
1064 # the bcsel. !(a >= 0 && b >= 0 && sum < 0) becomes !(a >= 0) || !(b >=
1065 # 0) || !(sum < 0), and that becomes (a < 0) || (b < 0) || (sum >= 0)
1067 # On Intel Gen11, this saves ~11 instructions.
1068 (('iadd_sat@64', a
, b
), ('bcsel',
1069 ('iand', ('iand', ('ilt', a
, 0), ('ilt', b
, 0)), ('ige', ('iadd', a
, b
), 0)),
1072 ('ior', ('ior', ('ilt', a
, 0), ('ilt', b
, 0)), ('ige', ('iadd', a
, b
), 0)),
1074 0x7fffffffffffffff)),
1075 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1077 # int64_t sum = a - b;
1079 # if (a < 0 && b >= 0 && a < sum)
1081 # } else if (a >= 0 && b < 0 && a >= sum)
1085 # Optimizations similar to the iadd_sat case are applied here.
1086 (('isub_sat@64', a
, b
), ('bcsel',
1087 ('iand', ('iand', ('ilt', a
, 0), ('ige', b
, 0)), ('ige', ('isub', a
, b
), 0)),
1090 ('ior', ('ior', ('ilt', a
, 0), ('ige', b
, 0)), ('ige', ('isub', a
, b
), 0)),
1092 0x7fffffffffffffff)),
1093 '(options->lower_int64_options & nir_lower_iadd64) != 0'),
1095 # These are done here instead of in the backend because the int64 lowering
1096 # pass will make a mess of the patterns. The first patterns are
1097 # conditioned on nir_lower_minmax64 because it was not clear that it was
1098 # always an improvement on platforms that have real int64 support. No
1099 # shaders in shader-db hit this, so it was hard to say one way or the
1101 (('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'),
1102 (('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'),
1103 (('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'),
1104 (('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'),
1105 (('ilt', 'a@64', 0), ('ilt', ('unpack_64_2x32_split_y', a
), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1106 (('ige', 'a@64', 0), ('ige', ('unpack_64_2x32_split_y', a
), 0), '(options->lower_int64_options & nir_lower_icmp64) != 0'),
1108 (('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'),
1109 (('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'),
1110 # 0u < uint(a) <=> uint(a) != 0u
1111 (('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'),
1113 # Alternative lowering that doesn't rely on bfi.
1114 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1115 ('bcsel', ('ult', 31, 'bits'),
1118 ('iand', 'base', ('inot', ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))),
1119 ('iand', ('ishl', 'insert', 'offset'), ('ishl', ('isub', ('ishl', 1, 'bits'), 1), 'offset'))))),
1120 'options->lower_bitfield_insert_to_shifts'),
1122 # Alternative lowering that uses bitfield_select.
1123 (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
1124 ('bcsel', ('ult', 31, 'bits'), 'insert',
1125 ('bitfield_select', ('bfm', 'bits', 'offset'), ('ishl', 'insert', 'offset'), 'base')),
1126 'options->lower_bitfield_insert_to_bitfield_select'),
1128 (('ibitfield_extract', 'value', 'offset', 'bits'),
1129 ('bcsel', ('ult', 31, 'bits'), 'value',
1130 ('ibfe', 'value', 'offset', 'bits')),
1131 'options->lower_bitfield_extract'),
1133 (('ubitfield_extract', 'value', 'offset', 'bits'),
1134 ('bcsel', ('ult', 31, 'bits'), 'value',
1135 ('ubfe', 'value', 'offset', 'bits')),
1136 'options->lower_bitfield_extract'),
1138 # Note that these opcodes are defined to only use the five least significant bits of 'offset' and 'bits'
1139 (('ubfe', 'value', 'offset', ('iand', 31, 'bits')), ('ubfe', 'value', 'offset', 'bits')),
1140 (('ubfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ubfe', 'value', 'offset', 'bits')),
1141 (('ibfe', 'value', 'offset', ('iand', 31, 'bits')), ('ibfe', 'value', 'offset', 'bits')),
1142 (('ibfe', 'value', ('iand', 31, 'offset'), 'bits'), ('ibfe', 'value', 'offset', 'bits')),
1143 (('bfm', 'bits', ('iand', 31, 'offset')), ('bfm', 'bits', 'offset')),
1144 (('bfm', ('iand', 31, 'bits'), 'offset'), ('bfm', 'bits', 'offset')),
1146 (('ibitfield_extract', 'value', 'offset', 'bits'),
1147 ('bcsel', ('ieq', 0, 'bits'),
1150 ('ishl', 'value', ('isub', ('isub', 32, 'bits'), 'offset')),
1151 ('isub', 32, 'bits'))),
1152 'options->lower_bitfield_extract_to_shifts'),
1154 (('ubitfield_extract', 'value', 'offset', 'bits'),
1156 ('ushr', 'value', 'offset'),
1157 ('bcsel', ('ieq', 'bits', 32),
1159 ('isub', ('ishl', 1, 'bits'), 1))),
1160 'options->lower_bitfield_extract_to_shifts'),
1162 (('ifind_msb', 'value'),
1163 ('ufind_msb', ('bcsel', ('ilt', 'value', 0), ('inot', 'value'), 'value')),
1164 'options->lower_ifind_msb'),
1166 (('find_lsb', 'value'),
1167 ('ufind_msb', ('iand', 'value', ('ineg', 'value'))),
1168 'options->lower_find_lsb'),
1170 (('extract_i8', a
, 'b@32'),
1171 ('ishr', ('ishl', a
, ('imul', ('isub', 3, b
), 8)), 24),
1172 'options->lower_extract_byte'),
1174 (('extract_u8', a
, 'b@32'),
1175 ('iand', ('ushr', a
, ('imul', b
, 8)), 0xff),
1176 'options->lower_extract_byte'),
1178 (('extract_i16', a
, 'b@32'),
1179 ('ishr', ('ishl', a
, ('imul', ('isub', 1, b
), 16)), 16),
1180 'options->lower_extract_word'),
1182 (('extract_u16', a
, 'b@32'),
1183 ('iand', ('ushr', a
, ('imul', b
, 16)), 0xffff),
1184 'options->lower_extract_word'),
1186 (('pack_unorm_2x16', 'v'),
1187 ('pack_uvec2_to_uint',
1188 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
1189 'options->lower_pack_unorm_2x16'),
1191 (('pack_unorm_4x8', 'v'),
1192 ('pack_uvec4_to_uint',
1193 ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
1194 'options->lower_pack_unorm_4x8'),
1196 (('pack_snorm_2x16', 'v'),
1197 ('pack_uvec2_to_uint',
1198 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
1199 'options->lower_pack_snorm_2x16'),
1201 (('pack_snorm_4x8', 'v'),
1202 ('pack_uvec4_to_uint',
1203 ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
1204 'options->lower_pack_snorm_4x8'),
1206 (('unpack_unorm_2x16', 'v'),
1207 ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
1208 ('extract_u16', 'v', 1))),
1210 'options->lower_unpack_unorm_2x16'),
1212 (('unpack_unorm_4x8', 'v'),
1213 ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
1214 ('extract_u8', 'v', 1),
1215 ('extract_u8', 'v', 2),
1216 ('extract_u8', 'v', 3))),
1218 'options->lower_unpack_unorm_4x8'),
1220 (('unpack_snorm_2x16', 'v'),
1221 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec2', ('extract_i16', 'v', 0),
1222 ('extract_i16', 'v', 1))),
1224 'options->lower_unpack_snorm_2x16'),
1226 (('unpack_snorm_4x8', 'v'),
1227 ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f', ('vec4', ('extract_i8', 'v', 0),
1228 ('extract_i8', 'v', 1),
1229 ('extract_i8', 'v', 2),
1230 ('extract_i8', 'v', 3))),
1232 'options->lower_unpack_snorm_4x8'),
1234 (('pack_half_2x16_split', 'a@32', 'b@32'),
1235 ('ior', ('ishl', ('u2u32', ('f2f16', b
)), 16), ('u2u32', ('f2f16', a
))),
1236 'options->lower_pack_half_2x16_split'),
1238 (('unpack_half_2x16_split_x', 'a@32'),
1239 ('f2f32', ('u2u16', a
)),
1240 'options->lower_unpack_half_2x16_split'),
1242 (('unpack_half_2x16_split_y', 'a@32'),
1243 ('f2f32', ('u2u16', ('ushr', a
, 16))),
1244 'options->lower_unpack_half_2x16_split'),
1246 (('isign', a
), ('imin', ('imax', a
, -1), 1), 'options->lower_isign'),
1247 (('fsign', a
), ('fsub', ('b2f', ('flt', 0.0, a
)), ('b2f', ('flt', a
, 0.0))), 'options->lower_fsign'),
1249 # Address/offset calculations:
1250 # Drivers supporting imul24 should use the nir_lower_amul() pass, this
1251 # rule converts everyone else to imul:
1252 (('amul', a
, b
), ('imul', a
, b
), '!options->has_imul24'),
1254 (('imad24_ir3', a
, b
, 0), ('imul24', a
, b
)),
1255 (('imad24_ir3', a
, 0, c
), (c
)),
1256 (('imad24_ir3', a
, 1, c
), ('iadd', a
, c
)),
1258 # if first two srcs are const, crack apart the imad so constant folding
1259 # can clean up the imul:
1260 # TODO ffma should probably get a similar rule:
1261 (('imad24_ir3', '#a', '#b', c
), ('iadd', ('imul', a
, b
), c
)),
1263 # These will turn 24b address/offset calc back into 32b shifts, but
1264 # it should be safe to get back some of the bits of precision that we
1265 # already decided were no necessary:
1266 (('imul24', a
, '#b@32(is_pos_power_of_two)'), ('ishl', a
, ('find_lsb', b
)), '!options->lower_bitops'),
1267 (('imul24', a
, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a
, ('find_lsb', ('iabs', b
)))), '!options->lower_bitops'),
1268 (('imul24', a
, 0), (0)),
1271 # bit_size dependent lowerings
1272 for bit_size
in [8, 16, 32, 64]:
1273 # convenience constants
1274 intmax
= (1 << (bit_size
- 1)) - 1
1275 intmin
= 1 << (bit_size
- 1)
1278 (('iadd_sat@' + str(bit_size
), a
, b
),
1279 ('bcsel', ('ige', b
, 1), ('bcsel', ('ilt', ('iadd', a
, b
), a
), intmax
, ('iadd', a
, b
)),
1280 ('bcsel', ('ilt', a
, ('iadd', a
, b
)), intmin
, ('iadd', a
, b
))), 'options->lower_add_sat'),
1281 (('isub_sat@' + str(bit_size
), a
, b
),
1282 ('bcsel', ('ilt', b
, 0), ('bcsel', ('ilt', ('isub', a
, b
), a
), intmax
, ('isub', a
, b
)),
1283 ('bcsel', ('ilt', a
, ('isub', a
, b
)), intmin
, ('isub', a
, b
))), 'options->lower_add_sat'),
1286 invert
= OrderedDict([('feq', 'fne'), ('fne', 'feq')])
1288 for left
, right
in itertools
.combinations_with_replacement(invert
.keys(), 2):
1289 optimizations
.append((('inot', ('ior(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1290 ('iand', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1291 optimizations
.append((('inot', ('iand(is_used_once)', (left
, a
, b
), (right
, c
, d
))),
1292 ('ior', (invert
[left
], a
, b
), (invert
[right
], c
, d
))))
1294 # Optimize x2bN(b2x(x)) -> x
1295 for size
in type_sizes('bool'):
1296 aN
= 'a@' + str(size
)
1297 f2bN
= 'f2b' + str(size
)
1298 i2bN
= 'i2b' + str(size
)
1299 optimizations
.append(((f2bN
, ('b2f', aN
)), a
))
1300 optimizations
.append(((i2bN
, ('b2i', aN
)), a
))
1302 # Optimize x2yN(b2x(x)) -> b2y
1303 for x
, y
in itertools
.product(['f', 'u', 'i'], ['f', 'u', 'i']):
1304 if x
!= 'f' and y
!= 'f' and x
!= y
:
1307 b2x
= 'b2f' if x
== 'f' else 'b2i'
1308 b2y
= 'b2f' if y
== 'f' else 'b2i'
1309 x2yN
= '{}2{}'.format(x
, y
)
1310 optimizations
.append(((x2yN
, (b2x
, a
)), (b2y
, a
)))
1312 # Optimize away x2xN(a@N)
1313 for t
in ['int', 'uint', 'float']:
1314 for N
in type_sizes(t
):
1315 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1316 aN
= 'a@{0}'.format(N
)
1317 optimizations
.append(((x2xN
, aN
), a
))
1319 # Optimize x2xN(y2yM(a@P)) -> y2yN(a) for integers
1320 # In particular, we can optimize away everything except upcast of downcast and
1321 # upcasts where the type differs from the other cast
1322 for N
, M
in itertools
.product(type_sizes('uint'), type_sizes('uint')):
1324 # The outer cast is a down-cast. It doesn't matter what the size of the
1325 # argument of the inner cast is because we'll never been in the upcast
1326 # of downcast case. Regardless of types, we'll always end up with y2yN
1328 for x
, y
in itertools
.product(['i', 'u'], ['i', 'u']):
1329 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1330 y2yM
= '{0}2{0}{1}'.format(y
, M
)
1331 y2yN
= '{0}2{0}{1}'.format(y
, N
)
1332 optimizations
.append(((x2xN
, (y2yM
, a
)), (y2yN
, a
)))
1334 # If the outer cast is an up-cast, we have to be more careful about the
1335 # size of the argument of the inner cast and with types. In this case,
1336 # the type is always the type of type up-cast which is given by the
1338 for P
in type_sizes('uint'):
1339 # We can't optimize away up-cast of down-cast.
1343 # Because we're doing down-cast of down-cast, the types always have
1344 # to match between the two casts
1345 for x
in ['i', 'u']:
1346 x2xN
= '{0}2{0}{1}'.format(x
, N
)
1347 x2xM
= '{0}2{0}{1}'.format(x
, M
)
1348 aP
= 'a@{0}'.format(P
)
1349 optimizations
.append(((x2xN
, (x2xM
, aP
)), (x2xN
, a
)))
1351 # The N == M case is handled by other optimizations
1354 # Optimize comparisons with up-casts
1355 for t
in ['int', 'uint', 'float']:
1356 for N
, M
in itertools
.product(type_sizes(t
), repeat
=2):
1357 if N
== 1 or N
>= M
:
1360 x2xM
= '{0}2{0}{1}'.format(t
[0], M
)
1361 x2xN
= '{0}2{0}{1}'.format(t
[0], N
)
1364 xeq
= 'feq' if t
== 'float' else 'ieq'
1365 xne
= 'fne' if t
== 'float' else 'ine'
1366 xge
= '{0}ge'.format(t
[0])
1367 xlt
= '{0}lt'.format(t
[0])
1369 # Up-casts are lossless so for correctly signed comparisons of
1370 # up-casted values we can do the comparison at the largest of the two
1371 # original sizes and drop one or both of the casts. (We have
1372 # optimizations to drop the no-op casts which this may generate.)
1373 for P
in type_sizes(t
):
1379 ((xeq
, (x2xM
, aN
), (x2xM
, bP
)), (xeq
, a
, (x2xN
, b
))),
1380 ((xne
, (x2xM
, aN
), (x2xM
, bP
)), (xne
, a
, (x2xN
, b
))),
1381 ((xge
, (x2xM
, aN
), (x2xM
, bP
)), (xge
, a
, (x2xN
, b
))),
1382 ((xlt
, (x2xM
, aN
), (x2xM
, bP
)), (xlt
, a
, (x2xN
, b
))),
1383 ((xge
, (x2xM
, bP
), (x2xM
, aN
)), (xge
, (x2xN
, b
), a
)),
1384 ((xlt
, (x2xM
, bP
), (x2xM
, aN
)), (xlt
, (x2xN
, b
), a
)),
1387 # The next bit doesn't work on floats because the range checks would
1388 # get way too complicated.
1389 if t
in ['int', 'uint']:
1391 xN_min
= -(1 << (N
- 1))
1392 xN_max
= (1 << (N
- 1)) - 1
1395 xN_max
= (1 << N
) - 1
1399 # If we're up-casting and comparing to a constant, we can unfold
1400 # the comparison into a comparison with the shrunk down constant
1401 # and a check that the constant fits in the smaller bit size.
1403 ((xeq
, (x2xM
, aN
), '#b'),
1404 ('iand', (xeq
, a
, (x2xN
, b
)), (xeq
, (x2xM
, (x2xN
, b
)), b
))),
1405 ((xne
, (x2xM
, aN
), '#b'),
1406 ('ior', (xne
, a
, (x2xN
, b
)), (xne
, (x2xM
, (x2xN
, b
)), b
))),
1407 ((xlt
, (x2xM
, aN
), '#b'),
1408 ('iand', (xlt
, xN_min
, b
),
1409 ('ior', (xlt
, xN_max
, b
), (xlt
, a
, (x2xN
, b
))))),
1410 ((xlt
, '#a', (x2xM
, bN
)),
1411 ('iand', (xlt
, a
, xN_max
),
1412 ('ior', (xlt
, a
, xN_min
), (xlt
, (x2xN
, a
), b
)))),
1413 ((xge
, (x2xM
, aN
), '#b'),
1414 ('iand', (xge
, xN_max
, b
),
1415 ('ior', (xge
, xN_min
, b
), (xge
, a
, (x2xN
, b
))))),
1416 ((xge
, '#a', (x2xM
, bN
)),
1417 ('iand', (xge
, a
, xN_min
),
1418 ('ior', (xge
, a
, xN_max
), (xge
, (x2xN
, a
), b
)))),
1421 def fexp2i(exp
, bits
):
1422 # Generate an expression which constructs value 2.0^exp or 0.0.
1424 # We assume that exp is already in a valid range:
1426 # * [-15, 15] for 16-bit float
1427 # * [-127, 127] for 32-bit float
1428 # * [-1023, 1023] for 16-bit float
1430 # If exp is the lowest value in the valid range, a value of 0.0 is
1431 # constructed. Otherwise, the value 2.0^exp is constructed.
1433 return ('i2i16', ('ishl', ('iadd', exp
, 15), 10))
1435 return ('ishl', ('iadd', exp
, 127), 23)
1437 return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp
, 1023), 20))
1441 def ldexp(f
, exp
, bits
):
1442 # The maximum possible range for a normal exponent is [-126, 127] and,
1443 # throwing in denormals, you get a maximum range of [-149, 127]. This
1444 # means that we can potentially have a swing of +-276. If you start with
1445 # FLT_MAX, you actually have to do ldexp(FLT_MAX, -278) to get it to flush
1446 # all the way to zero. The GLSL spec only requires that we handle a subset
1447 # of this range. From version 4.60 of the spec:
1449 # "If exp is greater than +128 (single-precision) or +1024
1450 # (double-precision), the value returned is undefined. If exp is less
1451 # than -126 (single-precision) or -1022 (double-precision), the value
1452 # returned may be flushed to zero. Additionally, splitting the value
1453 # into a significand and exponent using frexp() and then reconstructing
1454 # a floating-point value using ldexp() should yield the original input
1455 # for zero and all finite non-denormalized values."
1457 # The SPIR-V spec has similar language.
1459 # In order to handle the maximum value +128 using the fexp2i() helper
1460 # above, we have to split the exponent in half and do two multiply
1463 # First, we clamp exp to a reasonable range. Specifically, we clamp to
1464 # twice the full range that is valid for the fexp2i() function above. If
1465 # exp/2 is the bottom value of that range, the fexp2i() expression will
1466 # yield 0.0f which, when multiplied by f, will flush it to zero which is
1467 # allowed by the GLSL and SPIR-V specs for low exponent values. If the
1468 # value is clamped from above, then it must have been above the supported
1469 # range of the GLSL built-in and therefore any return value is acceptable.
1471 exp
= ('imin', ('imax', exp
, -30), 30)
1473 exp
= ('imin', ('imax', exp
, -254), 254)
1475 exp
= ('imin', ('imax', exp
, -2046), 2046)
1479 # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
1480 # (We use ishr which isn't the same for -1, but the -1 case still works
1481 # since we use exp-exp/2 as the second exponent.) While the spec
1482 # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
1483 # work with denormals and doesn't allow for the full swing in exponents
1484 # that you can get with normalized values. Instead, we create two powers
1485 # of two and multiply by them each in turn. That way the effective range
1486 # of our exponent is doubled.
1487 pow2_1
= fexp2i(('ishr', exp
, 1), bits
)
1488 pow2_2
= fexp2i(('isub', exp
, ('ishr', exp
, 1)), bits
)
1489 return ('fmul', ('fmul', f
, pow2_1
), pow2_2
)
1492 (('ldexp@16', 'x', 'exp'), ldexp('x', 'exp', 16), 'options->lower_ldexp'),
1493 (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32), 'options->lower_ldexp'),
1494 (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64), 'options->lower_ldexp'),
1497 # Unreal Engine 4 demo applications open-codes bitfieldReverse()
1498 def bitfield_reverse(u
):
1499 step1
= ('ior', ('ishl', u
, 16), ('ushr', u
, 16))
1500 step2
= ('ior', ('ishl', ('iand', step1
, 0x00ff00ff), 8), ('ushr', ('iand', step1
, 0xff00ff00), 8))
1501 step3
= ('ior', ('ishl', ('iand', step2
, 0x0f0f0f0f), 4), ('ushr', ('iand', step2
, 0xf0f0f0f0), 4))
1502 step4
= ('ior', ('ishl', ('iand', step3
, 0x33333333), 2), ('ushr', ('iand', step3
, 0xcccccccc), 2))
1503 step5
= ('ior(many-comm-expr)', ('ishl', ('iand', step4
, 0x55555555), 1), ('ushr', ('iand', step4
, 0xaaaaaaaa), 1))
1507 optimizations
+= [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'), '!options->lower_bitfield_reverse')]
1509 # For any float comparison operation, "cmp", if you have "a == a && a cmp b"
1510 # then the "a == a" is redundant because it's equivalent to "a is not NaN"
1511 # and, if a is a NaN then the second comparison will fail anyway.
1512 for op
in ['flt', 'fge', 'feq']:
1514 (('iand', ('feq', a
, a
), (op
, a
, b
)), ('!' + op
, a
, b
)),
1515 (('iand', ('feq', a
, a
), (op
, b
, a
)), ('!' + op
, b
, a
)),
1518 # Add optimizations to handle the case where the result of a ternary is
1519 # compared to a constant. This way we can take things like
1525 # a ? (0 > 0) : (1 > 0)
1527 # which constant folding will eat for lunch. The resulting ternary will
1528 # further get cleaned up by the boolean reductions above and we will be
1529 # left with just the original variable "a".
1530 for op
in ['flt', 'fge', 'feq', 'fne',
1531 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
1533 ((op
, ('bcsel', 'a', '#b', '#c'), '#d'),
1534 ('bcsel', 'a', (op
, 'b', 'd'), (op
, 'c', 'd'))),
1535 ((op
, '#d', ('bcsel', a
, '#b', '#c')),
1536 ('bcsel', 'a', (op
, 'd', 'b'), (op
, 'd', 'c'))),
1540 # For example, this converts things like
1542 # 1 + mix(0, a - 1, condition)
1546 # mix(1, (a-1)+1, condition)
1548 # Other optimizations will rearrange the constants.
1549 for op
in ['fadd', 'fmul', 'iadd', 'imul']:
1551 ((op
, ('bcsel(is_used_once)', a
, '#b', c
), '#d'), ('bcsel', a
, (op
, b
, d
), (op
, c
, d
)))
1554 # For derivatives in compute shaders, GLSL_NV_compute_shader_derivatives
1557 # If neither layout qualifier is specified, derivatives in compute shaders
1558 # return zero, which is consistent with the handling of built-in texture
1559 # functions like texture() in GLSL 4.50 compute shaders.
1560 for op
in ['fddx', 'fddx_fine', 'fddx_coarse',
1561 'fddy', 'fddy_fine', 'fddy_coarse']:
1563 ((op
, 'a'), 0.0, 'info->stage == MESA_SHADER_COMPUTE && info->cs.derivative_group == DERIVATIVE_GROUP_NONE')
1566 # Some optimizations for ir3-specific instructions.
1568 # 'al * bl': If either 'al' or 'bl' is zero, return zero.
1569 (('umul_low', '#a(is_lower_half_zero)', 'b'), (0)),
1570 # '(ah * bl) << 16 + c': If either 'ah' or 'bl' is zero, return 'c'.
1571 (('imadsh_mix16', '#a@32(is_lower_half_zero)', 'b@32', 'c@32'), ('c')),
1572 (('imadsh_mix16', 'a@32', '#b@32(is_upper_half_zero)', 'c@32'), ('c')),
1575 # These kinds of sequences can occur after nir_opt_peephole_select.
1577 # NOTE: fadd is not handled here because that gets in the way of ffma
1578 # generation in the i965 driver. Instead, fadd and ffma are handled in
1579 # late_optimizations.
1583 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1584 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1585 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1586 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1587 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, e
, c
, d
)), (op
, ('bcsel', a
, b
, e
), c
, d
)),
1588 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', e
, c
, d
)), (op
, ('bcsel', a
, b
, e
), c
, d
)),
1591 for op
in ['fmul', 'iadd', 'imul', 'iand', 'ior', 'ixor', 'fmin', 'fmax', 'imin', 'imax', 'umin', 'umax']:
1593 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, 'd(is_not_const)')), (op
, b
, ('bcsel', a
, c
, d
))),
1594 (('bcsel', a
, (op
+ '(is_used_once)', b
, 'c(is_not_const)'), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1595 (('bcsel', a
, (op
, b
, 'c(is_not_const)'), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1596 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, 'd(is_not_const)')), (op
, b
, ('bcsel', a
, c
, d
))),
1601 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1602 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1603 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, d
, c
)), (op
, ('bcsel', a
, b
, d
), c
)),
1604 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', d
, c
)), (op
, ('bcsel', a
, b
, d
), c
)),
1607 for op
in ['frcp', 'frsq', 'fsqrt', 'fexp2', 'flog2', 'fsign', 'fsin', 'fcos']:
1609 (('bcsel', a
, (op
+ '(is_used_once)', b
), (op
, c
)), (op
, ('bcsel', a
, b
, c
))),
1610 (('bcsel', a
, (op
, b
), (op
+ '(is_used_once)', c
)), (op
, ('bcsel', a
, b
, c
))),
1613 # This section contains "late" optimizations that should be run before
1614 # creating ffmas and calling regular optimizations for the final time.
1615 # Optimizations should go here if they help code generation and conflict
1616 # with the regular optimizations.
1617 before_ffma_optimizations
= [
1618 # Propagate constants down multiplication chains
1619 (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a
, c
), b
)),
1620 (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a
, c
), b
)),
1621 (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a
, c
), b
)),
1622 (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a
, c
), b
)),
1624 (('~fadd', ('fmul', a
, b
), ('fmul', a
, c
)), ('fmul', a
, ('fadd', b
, c
))),
1625 (('iadd', ('imul', a
, b
), ('imul', a
, c
)), ('imul', a
, ('iadd', b
, c
))),
1626 (('~fadd', ('fneg', a
), a
), 0.0),
1627 (('iadd', ('ineg', a
), a
), 0),
1628 (('iadd', ('ineg', a
), ('iadd', a
, b
)), b
),
1629 (('iadd', a
, ('iadd', ('ineg', a
), b
)), b
),
1630 (('~fadd', ('fneg', a
), ('fadd', a
, b
)), b
),
1631 (('~fadd', a
, ('fadd', ('fneg', a
), b
)), b
),
1633 (('~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
)),
1634 (('~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
)),
1635 (('~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
))),
1638 # This section contains "late" optimizations that should be run after the
1639 # regular optimizations have finished. Optimizations should go here if
1640 # they help code generation but do not necessarily produce code that is
1641 # more easily optimizable.
1642 late_optimizations
= [
1643 # Most of these optimizations aren't quite safe when you get infinity or
1644 # Nan involved but the first one should be fine.
1645 (('flt', ('fadd', a
, b
), 0.0), ('flt', a
, ('fneg', b
))),
1646 (('flt', ('fneg', ('fadd', a
, b
)), 0.0), ('flt', ('fneg', a
), b
)),
1647 (('~fge', ('fadd', a
, b
), 0.0), ('fge', a
, ('fneg', b
))),
1648 (('~fge', ('fneg', ('fadd', a
, b
)), 0.0), ('fge', ('fneg', a
), b
)),
1649 (('~feq', ('fadd', a
, b
), 0.0), ('feq', a
, ('fneg', b
))),
1650 (('~fne', ('fadd', a
, b
), 0.0), ('fne', a
, ('fneg', b
))),
1652 # nir_lower_to_source_mods will collapse this, but its existence during the
1653 # optimization loop can prevent other optimizations.
1654 (('fneg', ('fneg', a
)), a
),
1656 # Subtractions get lowered during optimization, so we need to recombine them
1657 (('fadd', 'a', ('fneg', 'b')), ('fsub', 'a', 'b'), '!options->lower_sub'),
1658 (('iadd', 'a', ('ineg', 'b')), ('isub', 'a', 'b'), '!options->lower_sub'),
1659 (('fneg', a
), ('fsub', 0.0, a
), 'options->lower_negate'),
1660 (('ineg', a
), ('isub', 0, a
), 'options->lower_negate'),
1662 # These are duplicated from the main optimizations table. The late
1663 # patterns that rearrange expressions like x - .5 < 0 to x < .5 can create
1664 # new patterns like these. The patterns that compare with zero are removed
1665 # because they are unlikely to be created in by anything in
1666 # late_optimizations.
1667 (('flt', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('flt', a
, b
)),
1668 (('flt', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('flt', b
, a
)),
1669 (('fge', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fge', a
, b
)),
1670 (('fge', '#b(is_gt_0_and_lt_1)', ('fsat(is_used_once)', a
)), ('fge', b
, a
)),
1671 (('feq', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('feq', a
, b
)),
1672 (('fne', ('fsat(is_used_once)', a
), '#b(is_gt_0_and_lt_1)'), ('fne', a
, b
)),
1674 (('fge', ('fsat(is_used_once)', a
), 1.0), ('fge', a
, 1.0)),
1675 (('flt', ('fsat(is_used_once)', a
), 1.0), ('flt', a
, 1.0)),
1677 (('~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
)))),
1679 (('flt', ('fneg', a
), ('fneg', b
)), ('flt', b
, a
)),
1680 (('fge', ('fneg', a
), ('fneg', b
)), ('fge', b
, a
)),
1681 (('feq', ('fneg', a
), ('fneg', b
)), ('feq', b
, a
)),
1682 (('fne', ('fneg', a
), ('fneg', b
)), ('fne', b
, a
)),
1683 (('flt', ('fneg', a
), -1.0), ('flt', 1.0, a
)),
1684 (('flt', -1.0, ('fneg', a
)), ('flt', a
, 1.0)),
1685 (('fge', ('fneg', a
), -1.0), ('fge', 1.0, a
)),
1686 (('fge', -1.0, ('fneg', a
)), ('fge', a
, 1.0)),
1687 (('fne', ('fneg', a
), -1.0), ('fne', 1.0, a
)),
1688 (('feq', -1.0, ('fneg', a
)), ('feq', a
, 1.0)),
1691 (('iand', a
, a
), a
),
1693 (('~fadd', ('fneg(is_used_once)', ('fsat(is_used_once)', 'a(is_not_fmul)')), 1.0), ('fsat', ('fadd', 1.0, ('fneg', a
)))),
1695 (('fdot2', a
, b
), ('fdot_replicated2', a
, b
), 'options->fdot_replicates'),
1696 (('fdot3', a
, b
), ('fdot_replicated3', a
, b
), 'options->fdot_replicates'),
1697 (('fdot4', a
, b
), ('fdot_replicated4', a
, b
), 'options->fdot_replicates'),
1698 (('fdph', a
, b
), ('fdph_replicated', a
, b
), 'options->fdot_replicates'),
1700 (('~flrp@32', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1701 (('~flrp@64', ('fadd(is_used_once)', a
, b
), ('fadd(is_used_once)', a
, c
), d
), ('fadd', ('flrp', b
, c
, d
), a
)),
1703 (('~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'),
1704 (('~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'),
1706 # A similar operation could apply to any ffma(#a, b, #(-a/2)), but this
1707 # particular operation is common for expanding values stored in a texture
1708 # from [0,1] to [-1,1].
1709 (('~ffma@32', a
, 2.0, -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1710 (('~ffma@32', a
, -2.0, -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1711 (('~ffma@32', a
, -2.0, 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1712 (('~ffma@32', a
, 2.0, 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1713 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), -1.0), ('flrp', -1.0, 1.0, a
), '!options->lower_flrp32'),
1714 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), -1.0), ('flrp', -1.0, 1.0, ('fneg', a
)), '!options->lower_flrp32'),
1715 (('~fadd@32', ('fmul(is_used_once)', -2.0, a
), 1.0), ('flrp', 1.0, -1.0, a
), '!options->lower_flrp32'),
1716 (('~fadd@32', ('fmul(is_used_once)', 2.0, a
), 1.0), ('flrp', 1.0, -1.0, ('fneg', a
)), '!options->lower_flrp32'),
1720 # a + -a*a + a*b (1)
1722 # Option 1: ffma(a, (b-a), a)
1724 # Alternately, after (1):
1730 # Option 2: ffma(a, 2, -(a*a))
1731 # Option 3: ffma(a, 2, (-a)*a)
1732 # Option 4: ffma(a, -a, (2*a)
1733 # Option 5: a * (2 - a)
1735 # There are a lot of other possible combinations.
1736 (('~ffma@32', ('fadd', b
, ('fneg', a
)), a
, a
), ('flrp', a
, b
, a
), '!options->lower_flrp32'),
1737 (('~ffma@32', a
, 2.0, ('fneg', ('fmul', a
, a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1738 (('~ffma@32', a
, 2.0, ('fmul', ('fneg', a
), a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1739 (('~ffma@32', a
, ('fneg', a
), ('fmul', 2.0, a
)), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1740 (('~fmul@32', a
, ('fadd', 2.0, ('fneg', a
))), ('flrp', a
, 1.0, a
), '!options->lower_flrp32'),
1742 # we do these late so that we don't get in the way of creating ffmas
1743 (('fmin', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmin', a
, b
))),
1744 (('fmax', ('fadd(is_used_once)', '#c', a
), ('fadd(is_used_once)', '#c', b
)), ('fadd', c
, ('fmax', a
, b
))),
1746 (('bcsel', a
, 0, ('b2f32', ('inot', 'b@bool'))), ('b2f32', ('inot', ('ior', a
, b
)))),
1748 # Putting this in 'optimizations' interferes with the bcsel(a, op(b, c),
1749 # op(b, d)) => op(b, bcsel(a, c, d)) transformations. I do not know why.
1750 (('bcsel', ('feq', ('fsqrt', 'a(is_not_negative)'), 0.0), intBitsToFloat(0x7f7fffff), ('frsq', a
)),
1751 ('fmin', ('frsq', a
), intBitsToFloat(0x7f7fffff))),
1753 # Things that look like DPH in the source shader may get expanded to
1754 # something that looks like dot(v1.xyz, v2.xyz) + v1.w by the time it gets
1755 # to NIR. After FFMA is generated, this can look like:
1757 # fadd(ffma(v1.z, v2.z, ffma(v1.y, v2.y, fmul(v1.x, v2.x))), v1.w)
1759 # Reassociate the last addition into the first multiplication.
1761 # Some shaders do not use 'invariant' in vertex and (possibly) geometry
1762 # shader stages on some outputs that are intended to be invariant. For
1763 # various reasons, this optimization may not be fully applied in all
1764 # shaders used for different rendering passes of the same geometry. This
1765 # can result in Z-fighting artifacts (at best). For now, disable this
1766 # optimization in these stages. See bugzilla #111490. In tessellation
1767 # stages applications seem to use 'precise' when necessary, so allow the
1768 # optimization in those stages.
1769 (('~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)'),
1770 ('ffma', a
, b
, ('ffma', c
, d
, ('ffma', e
, 'f', 'g'))), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1771 (('~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)'),
1772 ('ffma', a
, b
, ('ffma', c
, d
, e
)), '(info->stage != MESA_SHADER_VERTEX && info->stage != MESA_SHADER_GEOMETRY) && !options->intel_vec4'),
1774 # Convert f2fmp instructions to concrete f2f16 instructions. At this point
1775 # any conversions that could have been removed will have been removed in
1776 # nir_opt_algebraic so any remaining ones are required.
1777 (('f2fmp', a
), ('f2f16', a
)),
1781 late_optimizations
+= [
1782 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
), (op
, b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1783 (('bcsel', a
, (op
, b
, c
), (op
+ '(is_used_once)', b
, d
)), (op
, b
, ('bcsel', a
, c
, d
))),
1787 late_optimizations
+= [
1788 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1789 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, c
, e
)), (op
, b
, c
, ('bcsel', a
, d
, e
))),
1791 (('bcsel', a
, (op
+ '(is_used_once)', b
, c
, d
), (op
, b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1792 (('bcsel', a
, (op
, b
, c
, d
), (op
+ '(is_used_once)', b
, e
, d
)), (op
, b
, ('bcsel', a
, c
, e
), d
)),
1795 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic", optimizations
).render())
1796 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_before_ffma",
1797 before_ffma_optimizations
).render())
1798 print(nir_algebraic
.AlgebraicPass("nir_opt_algebraic_late",
1799 late_optimizations
).render())