• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1#
2# Copyright (C) 2014 Intel Corporation
3#
4# Permission is hereby granted, free of charge, to any person obtaining a
5# copy of this software and associated documentation files (the "Software"),
6# to deal in the Software without restriction, including without limitation
7# the rights to use, copy, modify, merge, publish, distribute, sublicense,
8# and/or sell copies of the Software, and to permit persons to whom the
9# Software is furnished to do so, subject to the following conditions:
10#
11# The above copyright notice and this permission notice (including the next
12# paragraph) shall be included in all copies or substantial portions of the
13# Software.
14#
15# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21# IN THE SOFTWARE.
22#
23# Authors:
24#    Jason Ekstrand (jason@jlekstrand.net)
25
26import nir_algebraic
27
28# Convenience variables
29a = 'a'
30b = 'b'
31c = 'c'
32d = 'd'
33
34# Written in the form (<search>, <replace>) where <search> is an expression
35# and <replace> is either an expression or a value.  An expression is
36# defined as a tuple of the form ([~]<op>, <src0>, <src1>, <src2>, <src3>)
37# where each source is either an expression or a value.  A value can be
38# either a numeric constant or a string representing a variable name.
39#
40# If the opcode in a search expression is prefixed by a '~' character, this
41# indicates that the operation is inexact.  Such operations will only get
42# applied to SSA values that do not have the exact bit set.  This should be
43# used by by any optimizations that are not bit-for-bit exact.  It should not,
44# however, be used for backend-requested lowering operations as those need to
45# happen regardless of precision.
46#
47# Variable names are specified as "[#]name[@type][(cond)]" where "#" inicates
48# that the given variable will only match constants and the type indicates that
49# the given variable will only match values from ALU instructions with the
50# given output type, and (cond) specifies an additional condition function
51# (see nir_search_helpers.h).
52#
53# For constants, you have to be careful to make sure that it is the right
54# type because python is unaware of the source and destination types of the
55# opcodes.
56#
57# All expression types can have a bit-size specified.  For opcodes, this
58# looks like "op@32", for variables it is "a@32" or "a@uint32" to specify a
59# type and size, and for literals, you can write "2.0@32".  In the search half
60# of the expression this indicates that it should only match that particular
61# bit-size.  In the replace half of the expression this indicates that the
62# constructed value should have that bit-size.
63
64optimizations = [
65
66   (('imul', a, '#b@32(is_pos_power_of_two)'), ('ishl', a, ('find_lsb', b))),
67   (('imul', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('ishl', a, ('find_lsb', ('iabs', b))))),
68   (('udiv', a, 1), a),
69   (('idiv', a, 1), a),
70   (('umod', a, 1), 0),
71   (('imod', a, 1), 0),
72   (('udiv', a, '#b@32(is_pos_power_of_two)'), ('ushr', a, ('find_lsb', b))),
73   (('idiv', a, '#b@32(is_pos_power_of_two)'), ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', b))), 'options->lower_idiv'),
74   (('idiv', a, '#b@32(is_neg_power_of_two)'), ('ineg', ('imul', ('isign', a), ('ushr', ('iabs', a), ('find_lsb', ('iabs', b))))), 'options->lower_idiv'),
75   (('umod', a, '#b(is_pos_power_of_two)'),    ('iand', a, ('isub', b, 1))),
76
77   (('fneg', ('fneg', a)), a),
78   (('ineg', ('ineg', a)), a),
79   (('fabs', ('fabs', a)), ('fabs', a)),
80   (('fabs', ('fneg', a)), ('fabs', a)),
81   (('fabs', ('u2f32', a)), ('u2f32', a)),
82   (('iabs', ('iabs', a)), ('iabs', a)),
83   (('iabs', ('ineg', a)), ('iabs', a)),
84   (('~fadd', a, 0.0), a),
85   (('iadd', a, 0), a),
86   (('usadd_4x8', a, 0), a),
87   (('usadd_4x8', a, ~0), ~0),
88   (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
89   (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
90   (('~fadd', ('fneg', a), a), 0.0),
91   (('iadd', ('ineg', a), a), 0),
92   (('iadd', ('ineg', a), ('iadd', a, b)), b),
93   (('iadd', a, ('iadd', ('ineg', a), b)), b),
94   (('~fadd', ('fneg', a), ('fadd', a, b)), b),
95   (('~fadd', a, ('fadd', ('fneg', a), b)), b),
96   (('~fmul', a, 0.0), 0.0),
97   (('imul', a, 0), 0),
98   (('umul_unorm_4x8', a, 0), 0),
99   (('umul_unorm_4x8', a, ~0), a),
100   (('fmul', a, 1.0), a),
101   (('imul', a, 1), a),
102   (('fmul', a, -1.0), ('fneg', a)),
103   (('imul', a, -1), ('ineg', a)),
104   (('~ffma', 0.0, a, b), b),
105   (('~ffma', a, 0.0, b), b),
106   (('~ffma', a, b, 0.0), ('fmul', a, b)),
107   (('ffma', a, 1.0, b), ('fadd', a, b)),
108   (('ffma', 1.0, a, b), ('fadd', a, b)),
109   (('~flrp', a, b, 0.0), a),
110   (('~flrp', a, b, 1.0), b),
111   (('~flrp', a, a, b), a),
112   (('~flrp', 0.0, a, b), ('fmul', a, b)),
113   (('~flrp', a, b, ('b2f', c)), ('bcsel', c, b, a), 'options->lower_flrp32'),
114   (('~flrp', a, 0.0, c), ('fadd', ('fmul', ('fneg', a), c), a)),
115   (('flrp@32', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a), 'options->lower_flrp32'),
116   (('flrp@64', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a), 'options->lower_flrp64'),
117   (('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'),
118   (('~fadd', ('fmul', a, ('fadd', 1.0, ('fneg', ('b2f', c)))), ('fmul', b, ('b2f', c))), ('bcsel', c, b, a), 'options->lower_flrp32'),
119   (('~fadd@32', ('fmul', a, ('fadd', 1.0, ('fneg',         c ))), ('fmul', b,         c )), ('flrp', a, b, c), '!options->lower_flrp32'),
120   (('~fadd@64', ('fmul', a, ('fadd', 1.0, ('fneg',         c ))), ('fmul', b,         c )), ('flrp', a, b, c), '!options->lower_flrp64'),
121   (('~fadd', a, ('fmul', ('b2f', c), ('fadd', b, ('fneg', a)))), ('bcsel', c, b, a), 'options->lower_flrp32'),
122   (('~fadd@32', a, ('fmul',         c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp32'),
123   (('~fadd@64', a, ('fmul',         c , ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp64'),
124   (('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'),
125   (('~fadd', ('fmul', a, b), c), ('ffma', a, b, c), 'options->fuse_ffma'),
126
127   # (a * #b + #c) << #d
128   # ((a * #b) << #d) + (#c << #d)
129   # (a * (#b << #d)) + (#c << #d)
130   (('ishl', ('iadd', ('imul', a, '#b'), '#c'), '#d'),
131    ('iadd', ('imul', a, ('ishl', b, d)), ('ishl', c, d))),
132
133   # (a * #b) << #c
134   # a * (#b << #c)
135   (('ishl', ('imul', a, '#b'), '#c'), ('imul', a, ('ishl', b, c))),
136
137   # Comparison simplifications
138   (('~inot', ('flt', a, b)), ('fge', a, b)),
139   (('~inot', ('fge', a, b)), ('flt', a, b)),
140   (('~inot', ('feq', a, b)), ('fne', a, b)),
141   (('~inot', ('fne', a, b)), ('feq', a, b)),
142   (('inot', ('ilt', a, b)), ('ige', a, b)),
143   (('inot', ('ige', a, b)), ('ilt', a, b)),
144   (('inot', ('ieq', a, b)), ('ine', a, b)),
145   (('inot', ('ine', a, b)), ('ieq', a, b)),
146
147   # 0.0 >= b2f(a)
148   # b2f(a) <= 0.0
149   # b2f(a) == 0.0 because b2f(a) can only be 0 or 1
150   # inot(a)
151   (('fge', 0.0, ('b2f', a)), ('inot', a)),
152
153   (('fge', ('fneg', ('b2f', a)), 0.0), ('inot', a)),
154
155   # 0.0 < fabs(a)
156   # fabs(a) > 0.0
157   # fabs(a) != 0.0 because fabs(a) must be >= 0
158   # a != 0.0
159   (('flt', 0.0, ('fabs', a)), ('fne', a, 0.0)),
160
161   # ignore this opt when the result is used by a bcsel or if so we can make
162   # use of conditional modifiers on supported hardware.
163   (('flt(is_not_used_by_conditional)', ('fadd(is_used_once)', a, ('fneg', b)), 0.0), ('flt', a, b)),
164
165   (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
166   (('bcsel', ('flt', b, a), b, a), ('fmin', a, b)),
167   (('bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
168   (('bcsel', ('inot', a), b, c), ('bcsel', a, c, b)),
169   (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
170   (('bcsel', a, True, 'b@bool'), ('ior', a, b)),
171   (('fmin', a, a), a),
172   (('fmax', a, a), a),
173   (('imin', a, a), a),
174   (('imax', a, a), a),
175   (('umin', a, a), a),
176   (('umax', a, a), a),
177   (('fmin', a, ('fneg', a)), ('fneg', ('fabs', a))),
178   (('imin', a, ('ineg', a)), ('ineg', ('iabs', a))),
179   (('fmin', a, ('fneg', ('fabs', a))), ('fneg', ('fabs', a))),
180   (('imin', a, ('ineg', ('iabs', a))), ('ineg', ('iabs', a))),
181   (('fmin', a, ('fabs', a)), a),
182   (('imin', a, ('iabs', a)), a),
183   (('fmax', a, ('fneg', ('fabs', a))), a),
184   (('imax', a, ('ineg', ('iabs', a))), a),
185   (('fmax', a, ('fabs', a)), ('fabs', a)),
186   (('imax', a, ('iabs', a)), ('iabs', a)),
187   (('fmax', a, ('fneg', a)), ('fabs', a)),
188   (('imax', a, ('ineg', a)), ('iabs', a)),
189   (('~fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'),
190   (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'),
191   (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'),
192   (('fsat', ('fsat', a)), ('fsat', a)),
193   (('fmin', ('fmax', ('fmin', ('fmax', a, b), c), b), c), ('fmin', ('fmax', a, b), c)),
194   (('imin', ('imax', ('imin', ('imax', a, b), c), b), c), ('imin', ('imax', a, b), c)),
195   (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
196   (('fmax', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
197   (('fmin', ('fsat', a), '#b@32(is_zero_to_one)'), ('fsat', ('fmin', a, b))),
198   (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
199   (('~ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
200   (('~ior', ('flt', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
201   (('~ior', ('fge', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
202   (('~ior', ('fge', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
203   (('fabs', ('slt', a, b)), ('slt', a, b)),
204   (('fabs', ('sge', a, b)), ('sge', a, b)),
205   (('fabs', ('seq', a, b)), ('seq', a, b)),
206   (('fabs', ('sne', a, b)), ('sne', a, b)),
207   (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
208   (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
209   (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
210   (('sne', a, b), ('b2f', ('fne', a, b)), 'options->lower_scmp'),
211   (('fne', ('fneg', a), a), ('fne', a, 0.0)),
212   (('feq', ('fneg', a), a), ('feq', a, 0.0)),
213   # Emulating booleans
214   (('imul', ('b2i', a), ('b2i', b)), ('b2i', ('iand', a, b))),
215   (('fmul', ('b2f', a), ('b2f', b)), ('b2f', ('iand', a, b))),
216   (('fsat', ('fadd', ('b2f', a), ('b2f', b))), ('b2f', ('ior', a, b))),
217   (('iand', 'a@bool', 1.0), ('b2f', a)),
218   # True/False are ~0 and 0 in NIR.  b2i of True is 1, and -1 is ~0 (True).
219   (('ineg', ('b2i@32', a)), a),
220   (('flt', ('fneg', ('b2f', a)), 0), a), # Generated by TGSI KILL_IF.
221   (('flt', ('fsub', 0.0, ('b2f', a)), 0), a), # Generated by TGSI KILL_IF.
222   # Comparison with the same args.  Note that these are not done for
223   # the float versions because NaN always returns false on float
224   # inequalities.
225   (('ilt', a, a), False),
226   (('ige', a, a), True),
227   (('ieq', a, a), True),
228   (('ine', a, a), False),
229   (('ult', a, a), False),
230   (('uge', a, a), True),
231   # Logical and bit operations
232   (('fand', a, 0.0), 0.0),
233   (('iand', a, a), a),
234   (('iand', a, ~0), a),
235   (('iand', a, 0), 0),
236   (('ior', a, a), a),
237   (('ior', a, 0), a),
238   (('ior', a, True), True),
239   (('fxor', a, a), 0.0),
240   (('ixor', a, a), 0),
241   (('ixor', a, 0), a),
242   (('inot', ('inot', a)), a),
243   # DeMorgan's Laws
244   (('iand', ('inot', a), ('inot', b)), ('inot', ('ior',  a, b))),
245   (('ior',  ('inot', a), ('inot', b)), ('inot', ('iand', a, b))),
246   # Shift optimizations
247   (('ishl', 0, a), 0),
248   (('ishl', a, 0), a),
249   (('ishr', 0, a), 0),
250   (('ishr', a, 0), a),
251   (('ushr', 0, a), 0),
252   (('ushr', a, 0), a),
253   (('iand', 0xff, ('ushr@32', a, 24)), ('ushr', a, 24)),
254   (('iand', 0xffff, ('ushr@32', a, 16)), ('ushr', a, 16)),
255   # Exponential/logarithmic identities
256   (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
257   (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
258   (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
259   (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
260   (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2', c), d))),
261    ('~fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
262   (('~fpow', a, 1.0), a),
263   (('~fpow', a, 2.0), ('fmul', a, a)),
264   (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
265   (('~fpow', 2.0, a), ('fexp2', a)),
266   (('~fpow', ('fpow', a, 2.2), 0.454545), a),
267   (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
268   (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
269   (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
270   (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
271   (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
272   (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
273   (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
274   (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
275   (('~fmul', ('fexp2', a), ('fexp2', b)), ('fexp2', ('fadd', a, b))),
276   # Division and reciprocal
277   (('~fdiv', 1.0, a), ('frcp', a)),
278   (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
279   (('~frcp', ('frcp', a)), a),
280   (('~frcp', ('fsqrt', a)), ('frsq', a)),
281   (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
282   (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
283   # Boolean simplifications
284   (('ieq', 'a@bool', True), a),
285   (('ine(is_not_used_by_if)', 'a@bool', True), ('inot', a)),
286   (('ine', 'a@bool', False), a),
287   (('ieq(is_not_used_by_if)', 'a@bool', False), ('inot', 'a')),
288   (('bcsel', a, True, False), a),
289   (('bcsel', a, False, True), ('inot', a)),
290   (('bcsel@32', a, 1.0, 0.0), ('b2f', a)),
291   (('bcsel@32', a, 0.0, 1.0), ('b2f', ('inot', a))),
292   (('bcsel@32', a, -1.0, -0.0), ('fneg', ('b2f', a))),
293   (('bcsel@32', a, -0.0, -1.0), ('fneg', ('b2f', ('inot', a)))),
294   (('bcsel', True, b, c), b),
295   (('bcsel', False, b, c), c),
296   # The result of this should be hit by constant propagation and, in the
297   # next round of opt_algebraic, get picked up by one of the above two.
298   (('bcsel', '#a', b, c), ('bcsel', ('ine', 'a', 0), b, c)),
299
300   (('bcsel', a, b, b), b),
301   (('fcsel', a, b, b), b),
302
303   # Conversions
304   (('i2b', ('b2i', a)), a),
305   (('f2i32', ('ftrunc', a)), ('f2i32', a)),
306   (('f2u32', ('ftrunc', a)), ('f2u32', a)),
307   (('i2b', ('ineg', a)), ('i2b', a)),
308   (('i2b', ('iabs', a)), ('i2b', a)),
309   (('fabs', ('b2f', a)), ('b2f', a)),
310   (('iabs', ('b2i', a)), ('b2i', a)),
311
312   # Packing and then unpacking does nothing
313   (('unpack_64_2x32_split_x', ('pack_64_2x32_split', a, b)), a),
314   (('unpack_64_2x32_split_y', ('pack_64_2x32_split', a, b)), b),
315   (('pack_64_2x32_split', ('unpack_64_2x32_split_x', a),
316                           ('unpack_64_2x32_split_y', a)), a),
317
318   # Byte extraction
319   (('ushr', a, 24), ('extract_u8', a, 3), '!options->lower_extract_byte'),
320   (('iand', 0xff, ('ushr', a, 16)), ('extract_u8', a, 2), '!options->lower_extract_byte'),
321   (('iand', 0xff, ('ushr', a,  8)), ('extract_u8', a, 1), '!options->lower_extract_byte'),
322   (('iand', 0xff, a), ('extract_u8', a, 0), '!options->lower_extract_byte'),
323
324    # Word extraction
325   (('ushr', a, 16), ('extract_u16', a, 1), '!options->lower_extract_word'),
326   (('iand', 0xffff, a), ('extract_u16', a, 0), '!options->lower_extract_word'),
327
328   # Subtracts
329   (('~fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)),
330   (('isub', a, ('isub', 0, b)), ('iadd', a, b)),
331   (('ussub_4x8', a, 0), a),
332   (('ussub_4x8', a, ~0), 0),
333   (('fsub', a, b), ('fadd', a, ('fneg', b)), 'options->lower_sub'),
334   (('isub', a, b), ('iadd', a, ('ineg', b)), 'options->lower_sub'),
335   (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
336   (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
337   (('~fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)),
338   (('iadd', a, ('isub', 0, b)), ('isub', a, b)),
339   (('fabs', ('fsub', 0.0, a)), ('fabs', a)),
340   (('iabs', ('isub', 0, a)), ('iabs', a)),
341
342   # Propagate negation up multiplication chains
343   (('fmul', ('fneg', a), b), ('fneg', ('fmul', a, b))),
344   (('imul', ('ineg', a), b), ('ineg', ('imul', a, b))),
345
346   # Propagate constants up multiplication chains
347   (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fmul', ('fmul', a, c), b)),
348   (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('imul', ('imul', a, c), b)),
349   (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('fadd', ('fadd', a, c), b)),
350   (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', 'b(is_not_const)'), '#c'), ('iadd', ('iadd', a, c), b)),
351
352   # Reassociate constants in add/mul chains so they can be folded together.
353   # For now, we mostly only handle cases where the constants are separated by
354   # a single non-constant.  We could do better eventually.
355   (('~fmul', '#a', ('fmul', b, '#c')), ('fmul', ('fmul', a, c), b)),
356   (('imul', '#a', ('imul', b, '#c')), ('imul', ('imul', a, c), b)),
357   (('~fadd', '#a', ('fadd', b, '#c')), ('fadd', ('fadd', a, c), b)),
358   (('iadd', '#a', ('iadd', b, '#c')), ('iadd', ('iadd', a, c), b)),
359
360   # By definition...
361   (('bcsel', ('ige', ('find_lsb', a), 0), ('find_lsb', a), -1), ('find_lsb', a)),
362   (('bcsel', ('ige', ('ifind_msb', a), 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
363   (('bcsel', ('ige', ('ufind_msb', a), 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
364
365   (('bcsel', ('ine', a, 0), ('find_lsb', a), -1), ('find_lsb', a)),
366   (('bcsel', ('ine', a, 0), ('ifind_msb', a), -1), ('ifind_msb', a)),
367   (('bcsel', ('ine', a, 0), ('ufind_msb', a), -1), ('ufind_msb', a)),
368
369   (('bcsel', ('ine', a, -1), ('ifind_msb', a), -1), ('ifind_msb', a)),
370
371   # Misc. lowering
372   (('fmod@32', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod32'),
373   (('fmod@64', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod64'),
374   (('frem', a, b), ('fsub', a, ('fmul', b, ('ftrunc', ('fdiv', a, b)))), 'options->lower_fmod32'),
375   (('uadd_carry@32', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
376   (('usub_borrow@32', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'),
377
378   (('bitfield_insert', 'base', 'insert', 'offset', 'bits'),
379    ('bcsel', ('ilt', 31, 'bits'), 'insert',
380              ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')),
381    'options->lower_bitfield_insert'),
382
383   (('ibitfield_extract', 'value', 'offset', 'bits'),
384    ('bcsel', ('ilt', 31, 'bits'), 'value',
385              ('ibfe', 'value', 'offset', 'bits')),
386    'options->lower_bitfield_extract'),
387
388   (('ubitfield_extract', 'value', 'offset', 'bits'),
389    ('bcsel', ('ult', 31, 'bits'), 'value',
390              ('ubfe', 'value', 'offset', 'bits')),
391    'options->lower_bitfield_extract'),
392
393   (('extract_i8', a, 'b@32'),
394    ('ishr', ('ishl', a, ('imul', ('isub', 3, b), 8)), 24),
395    'options->lower_extract_byte'),
396
397   (('extract_u8', a, 'b@32'),
398    ('iand', ('ushr', a, ('imul', b, 8)), 0xff),
399    'options->lower_extract_byte'),
400
401   (('extract_i16', a, 'b@32'),
402    ('ishr', ('ishl', a, ('imul', ('isub', 1, b), 16)), 16),
403    'options->lower_extract_word'),
404
405   (('extract_u16', a, 'b@32'),
406    ('iand', ('ushr', a, ('imul', b, 16)), 0xffff),
407    'options->lower_extract_word'),
408
409    (('pack_unorm_2x16', 'v'),
410     ('pack_uvec2_to_uint',
411        ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 65535.0)))),
412     'options->lower_pack_unorm_2x16'),
413
414    (('pack_unorm_4x8', 'v'),
415     ('pack_uvec4_to_uint',
416        ('f2u32', ('fround_even', ('fmul', ('fsat', 'v'), 255.0)))),
417     'options->lower_pack_unorm_4x8'),
418
419    (('pack_snorm_2x16', 'v'),
420     ('pack_uvec2_to_uint',
421        ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 32767.0)))),
422     'options->lower_pack_snorm_2x16'),
423
424    (('pack_snorm_4x8', 'v'),
425     ('pack_uvec4_to_uint',
426        ('f2i32', ('fround_even', ('fmul', ('fmin', 1.0, ('fmax', -1.0, 'v')), 127.0)))),
427     'options->lower_pack_snorm_4x8'),
428
429    (('unpack_unorm_2x16', 'v'),
430     ('fdiv', ('u2f32', ('vec2', ('extract_u16', 'v', 0),
431                                 ('extract_u16', 'v', 1))),
432              65535.0),
433     'options->lower_unpack_unorm_2x16'),
434
435    (('unpack_unorm_4x8', 'v'),
436     ('fdiv', ('u2f32', ('vec4', ('extract_u8', 'v', 0),
437                                 ('extract_u8', 'v', 1),
438                                 ('extract_u8', 'v', 2),
439                                 ('extract_u8', 'v', 3))),
440              255.0),
441     'options->lower_unpack_unorm_4x8'),
442
443    (('unpack_snorm_2x16', 'v'),
444     ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec2', ('extract_i16', 'v', 0),
445                                                              ('extract_i16', 'v', 1))),
446                                           32767.0))),
447     'options->lower_unpack_snorm_2x16'),
448
449    (('unpack_snorm_4x8', 'v'),
450     ('fmin', 1.0, ('fmax', -1.0, ('fdiv', ('i2f32', ('vec4', ('extract_i8', 'v', 0),
451                                                              ('extract_i8', 'v', 1),
452                                                              ('extract_i8', 'v', 2),
453                                                              ('extract_i8', 'v', 3))),
454                                           127.0))),
455     'options->lower_unpack_snorm_4x8'),
456]
457
458def fexp2i(exp, bits):
459   # We assume that exp is already in the right range.
460   if bits == 32:
461      return ('ishl', ('iadd', exp, 127), 23)
462   elif bits == 64:
463      return ('pack_64_2x32_split', 0, ('ishl', ('iadd', exp, 1023), 20))
464   else:
465      assert False
466
467def ldexp(f, exp, bits):
468   # First, we clamp exp to a reasonable range.  The maximum possible range
469   # for a normal exponent is [-126, 127] and, throwing in denormals, you get
470   # a maximum range of [-149, 127].  This means that we can potentially have
471   # a swing of +-276.  If you start with FLT_MAX, you actually have to do
472   # ldexp(FLT_MAX, -278) to get it to flush all the way to zero.  The GLSL
473   # spec, on the other hand, only requires that we handle an exponent value
474   # in the range [-126, 128].  This implementation is *mostly* correct; it
475   # handles a range on exp of [-252, 254] which allows you to create any
476   # value (including denorms if the hardware supports it) and to adjust the
477   # exponent of any normal value to anything you want.
478   if bits == 32:
479      exp = ('imin', ('imax', exp, -252), 254)
480   elif bits == 64:
481      exp = ('imin', ('imax', exp, -2044), 2046)
482   else:
483      assert False
484
485   # Now we compute two powers of 2, one for exp/2 and one for exp-exp/2.
486   # (We use ishr which isn't the same for -1, but the -1 case still works
487   # since we use exp-exp/2 as the second exponent.)  While the spec
488   # technically defines ldexp as f * 2.0^exp, simply multiplying once doesn't
489   # work with denormals and doesn't allow for the full swing in exponents
490   # that you can get with normalized values.  Instead, we create two powers
491   # of two and multiply by them each in turn.  That way the effective range
492   # of our exponent is doubled.
493   pow2_1 = fexp2i(('ishr', exp, 1), bits)
494   pow2_2 = fexp2i(('isub', exp, ('ishr', exp, 1)), bits)
495   return ('fmul', ('fmul', f, pow2_1), pow2_2)
496
497optimizations += [
498   (('ldexp@32', 'x', 'exp'), ldexp('x', 'exp', 32)),
499   (('ldexp@64', 'x', 'exp'), ldexp('x', 'exp', 64)),
500]
501
502# Unreal Engine 4 demo applications open-codes bitfieldReverse()
503def bitfield_reverse(u):
504    step1 = ('ior', ('ishl', u, 16), ('ushr', u, 16))
505    step2 = ('ior', ('ishl', ('iand', step1, 0x00ff00ff), 8), ('ushr', ('iand', step1, 0xff00ff00), 8))
506    step3 = ('ior', ('ishl', ('iand', step2, 0x0f0f0f0f), 4), ('ushr', ('iand', step2, 0xf0f0f0f0), 4))
507    step4 = ('ior', ('ishl', ('iand', step3, 0x33333333), 2), ('ushr', ('iand', step3, 0xcccccccc), 2))
508    step5 = ('ior', ('ishl', ('iand', step4, 0x55555555), 1), ('ushr', ('iand', step4, 0xaaaaaaaa), 1))
509
510    return step5
511
512optimizations += [(bitfield_reverse('x@32'), ('bitfield_reverse', 'x'))]
513
514# For any float comparison operation, "cmp", if you have "a == a && a cmp b"
515# then the "a == a" is redundant because it's equivalent to "a is not NaN"
516# and, if a is a NaN then the second comparison will fail anyway.
517for op in ['flt', 'fge', 'feq']:
518   optimizations += [
519      (('iand', ('feq', a, a), (op, a, b)), (op, a, b)),
520      (('iand', ('feq', a, a), (op, b, a)), (op, b, a)),
521   ]
522
523# Add optimizations to handle the case where the result of a ternary is
524# compared to a constant.  This way we can take things like
525#
526# (a ? 0 : 1) > 0
527#
528# and turn it into
529#
530# a ? (0 > 0) : (1 > 0)
531#
532# which constant folding will eat for lunch.  The resulting ternary will
533# further get cleaned up by the boolean reductions above and we will be
534# left with just the original variable "a".
535for op in ['flt', 'fge', 'feq', 'fne',
536           'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']:
537   optimizations += [
538      ((op, ('bcsel', 'a', '#b', '#c'), '#d'),
539       ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))),
540      ((op, '#d', ('bcsel', a, '#b', '#c')),
541       ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
542   ]
543
544# This section contains "late" optimizations that should be run before
545# creating ffmas and calling regular optimizations for the final time.
546# Optimizations should go here if they help code generation and conflict
547# with the regular optimizations.
548before_ffma_optimizations = [
549   # Propagate constants down multiplication chains
550   (('~fmul(is_used_once)', ('fmul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fmul', ('fmul', a, c), b)),
551   (('imul(is_used_once)', ('imul(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('imul', ('imul', a, c), b)),
552   (('~fadd(is_used_once)', ('fadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('fadd', ('fadd', a, c), b)),
553   (('iadd(is_used_once)', ('iadd(is_used_once)', 'a(is_not_const)', '#b'), 'c(is_not_const)'), ('iadd', ('iadd', a, c), b)),
554
555   (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))),
556   (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))),
557   (('~fadd', ('fneg', a), a), 0.0),
558   (('iadd', ('ineg', a), a), 0),
559   (('iadd', ('ineg', a), ('iadd', a, b)), b),
560   (('iadd', a, ('iadd', ('ineg', a), b)), b),
561   (('~fadd', ('fneg', a), ('fadd', a, b)), b),
562   (('~fadd', a, ('fadd', ('fneg', a), b)), b),
563]
564
565# This section contains "late" optimizations that should be run after the
566# regular optimizations have finished.  Optimizations should go here if
567# they help code generation but do not necessarily produce code that is
568# more easily optimizable.
569late_optimizations = [
570   # Most of these optimizations aren't quite safe when you get infinity or
571   # Nan involved but the first one should be fine.
572   (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
573   (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
574   (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
575   (('~fne', ('fadd', a, b), 0.0), ('fne', a, ('fneg', b))),
576
577   (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'),
578   (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'),
579   (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'),
580   (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'),
581
582   (('b2f(is_used_more_than_once)', ('inot', a)), ('bcsel', a, 0.0, 1.0)),
583   (('fneg(is_used_more_than_once)', ('b2f', ('inot', a))), ('bcsel', a, -0.0, -1.0)),
584
585   # we do these late so that we don't get in the way of creating ffmas
586   (('fmin', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmin', a, b))),
587   (('fmax', ('fadd(is_used_once)', '#c', a), ('fadd(is_used_once)', '#c', b)), ('fadd', c, ('fmax', a, b))),
588]
589
590print nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render()
591print nir_algebraic.AlgebraicPass("nir_opt_algebraic_before_ffma",
592                                  before_ffma_optimizations).render()
593print nir_algebraic.AlgebraicPass("nir_opt_algebraic_late",
594                                  late_optimizations).render()
595