nir_algebraic.py 49.8 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
#
# Copyright (C) 2014 Intel Corporation
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice (including the next
# paragraph) shall be included in all copies or substantial portions of the
# Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
# IN THE SOFTWARE.
#
# Authors:
#    Jason Ekstrand (jason@jlekstrand.net)

26
from __future__ import print_function
27
import ast
28
from collections import defaultdict
29
30
31
32
import itertools
import struct
import sys
import mako.template
33
import re
34
import traceback
35

36
37
from nir_opcodes import opcodes, type_sizes

38
# This should be the same as NIR_SEARCH_MAX_COMM_OPS in nir_search.c
39
nir_search_max_comm_ops = 8
40

41
42
43
44
45
46
47
48
49
50
# These opcodes are only employed by nir_search.  This provides a mapping from
# opcode to destination type.
conv_opcode_types = {
    'i2f' : 'float',
    'u2f' : 'float',
    'f2f' : 'float',
    'f2u' : 'uint',
    'f2i' : 'int',
    'u2u' : 'uint',
    'i2i' : 'int',
51
52
53
54
    'b2f' : 'float',
    'b2i' : 'int',
    'i2b' : 'bool',
    'f2b' : 'bool',
55
}
56

57
58
59
60
61
62
63
def get_c_opcode(op):
      if op in conv_opcode_types:
         return 'nir_search_op_' + op
      else:
         return 'nir_op_' + op


64
if sys.version_info < (3, 0):
65
    integer_types = (int, long)
66
67
68
    string_type = unicode

else:
69
    integer_types = (int, )
70
71
    string_type = str

72
73
74
75
76
77
78
79
80
81
82
_type_re = re.compile(r"(?P<type>int|uint|bool|float)?(?P<bits>\d+)?")

def type_bits(type_str):
   m = _type_re.match(type_str)
   assert m.group('type')

   if m.group('bits') is None:
      return 0
   else:
      return int(m.group('bits'))

83
84
85
86
87
# Represents a set of variables, each with a unique id
class VarSet(object):
   def __init__(self):
      self.names = {}
      self.ids = itertools.count()
88
      self.immutable = False;
89
90
91

   def __getitem__(self, name):
      if name not in self.names:
92
         assert not self.immutable, "Unknown replacement variable: " + name
Mathieu Bridon's avatar
Mathieu Bridon committed
93
         self.names[name] = next(self.ids)
94
95
96

      return self.names[name]

97
98
99
   def lock(self):
      self.immutable = True

100
101
102
class Value(object):
   @staticmethod
   def create(val, name_base, varset):
103
104
105
      if isinstance(val, bytes):
         val = val.decode('utf-8')

106
107
108
109
      if isinstance(val, tuple):
         return Expression(val, name_base, varset)
      elif isinstance(val, Expression):
         return val
110
      elif isinstance(val, string_type):
111
         return Variable(val, name_base, varset)
112
      elif isinstance(val, (bool, float) + integer_types):
113
114
         return Constant(val, name_base)

115
116
   def __init__(self, val, name, type_str):
      self.in_val = str(val)
117
118
119
      self.name = name
      self.type_str = type_str

120
121
122
   def __str__(self):
      return self.in_val

123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
   def get_bit_size(self):
      """Get the physical bit-size that has been chosen for this value, or if
      there is none, the canonical value which currently represents this
      bit-size class. Variables will be preferred, i.e. if there are any
      variables in the equivalence class, the canonical value will be a
      variable. We do this since we'll need to know which variable each value
      is equivalent to when constructing the replacement expression. This is
      the "find" part of the union-find algorithm.
      """
      bit_size = self

      while isinstance(bit_size, Value):
         if bit_size._bit_size is None:
            break
         bit_size = bit_size._bit_size

      if bit_size is not self:
         self._bit_size = bit_size
      return bit_size

   def set_bit_size(self, other):
      """Make self.get_bit_size() return what other.get_bit_size() return
      before calling this, or just "other" if it's a concrete bit-size. This is
      the "union" part of the union-find algorithm.
      """

      self_bit_size = self.get_bit_size()
      other_bit_size = other if isinstance(other, int) else other.get_bit_size()

      if self_bit_size == other_bit_size:
         return

      self_bit_size._bit_size = other_bit_size

157
158
159
160
161
162
163
164
   @property
   def type_enum(self):
      return "nir_search_value_" + self.type_str

   @property
   def c_type(self):
      return "nir_search_" + self.type_str

165
166
167
168
169
170
171
172
173
174
175
   def __c_name(self, cache):
      if cache is not None and self.name in cache:
         return cache[self.name]
      else:
         return self.name

   def c_value_ptr(self, cache):
      return "&{0}.value".format(self.__c_name(cache))

   def c_ptr(self, cache):
      return "&{0}".format(self.__c_name(cache))
176

177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
   @property
   def c_bit_size(self):
      bit_size = self.get_bit_size()
      if isinstance(bit_size, int):
         return bit_size
      elif isinstance(bit_size, Variable):
         return -bit_size.index - 1
      else:
         # If the bit-size class is neither a variable, nor an actual bit-size, then
         # - If it's in the search expression, we don't need to check anything
         # - If it's in the replace expression, either it's ambiguous (in which
         # case we'd reject it), or it equals the bit-size of the search value
         # We represent these cases with a 0 bit-size.
         return 0

192
   __template = mako.template.Template("""{
193
194
195
196
197
198
199
200
   { ${val.type_enum}, ${val.c_bit_size} },
% if isinstance(val, Constant):
   ${val.type()}, { ${val.hex()} /* ${val.value} */ },
% elif isinstance(val, Variable):
   ${val.index}, /* ${val.var_name} */
   ${'true' if val.is_constant else 'false'},
   ${val.type() or 'nir_type_invalid' },
   ${val.cond if val.cond else 'NULL'},
201
   ${val.swizzle()},
202
203
204
205
% elif isinstance(val, Expression):
   ${'true' if val.inexact else 'false'},
   ${val.comm_expr_idx}, ${val.comm_exprs},
   ${val.c_opcode()},
206
   { ${', '.join(src.c_value_ptr(cache) for src in val.sources)} },
207
208
209
210
   ${val.cond if val.cond else 'NULL'},
% endif
};""")

211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
   def render(self, cache):
      struct_init = self.__template.render(val=self, cache=cache,
                                           Constant=Constant,
                                           Variable=Variable,
                                           Expression=Expression)
      if cache is not None and struct_init in cache:
         # If it's in the cache, register a name remap in the cache and render
         # only a comment saying it's been remapped
         cache[self.name] = cache[struct_init]
         return "/* {} -> {} in the cache */\n".format(self.name,
                                                       cache[struct_init])
      else:
         if cache is not None:
            cache[struct_init] = self.name
         return "static const {} {} = {}\n".format(self.c_type, self.name,
                                                   struct_init)
227

228
_constant_re = re.compile(r"(?P<value>[^@\(]+)(?:@(?P<bits>\d+))?")
229

230
231
class Constant(Value):
   def __init__(self, val, name):
232
      Value.__init__(self, val, name, "constant")
233
234
235
236

      if isinstance(val, (str)):
         m = _constant_re.match(val)
         self.value = ast.literal_eval(m.group('value'))
237
         self._bit_size = int(m.group('bits')) if m.group('bits') else None
238
239
      else:
         self.value = val
240
         self._bit_size = None
241
242

      if isinstance(self.value, bool):
243
244
         assert self._bit_size is None or self._bit_size == 1
         self._bit_size = 1
245

Mathieu Bridon's avatar
Mathieu Bridon committed
246
   def hex(self):
247
248
      if isinstance(self.value, (bool)):
         return 'NIR_TRUE' if self.value else 'NIR_FALSE'
249
      if isinstance(self.value, integer_types):
250
         return hex(self.value)
251
      elif isinstance(self.value, float):
252
253
254
255
256
257
258
259
260
261
         i = struct.unpack('Q', struct.pack('d', self.value))[0]
         h = hex(i)

         # On Python 2 this 'L' suffix is automatically added, but not on Python 3
         # Adding it explicitly makes the generated file identical, regardless
         # of the Python version running this script.
         if h[-1] != 'L' and i > sys.maxsize:
            h += 'L'

         return h
262
263
264
      else:
         assert False

265
266
   def type(self):
      if isinstance(self.value, (bool)):
267
         return "nir_type_bool"
268
      elif isinstance(self.value, integer_types):
269
270
271
272
         return "nir_type_int"
      elif isinstance(self.value, float):
         return "nir_type_float"

273
274
275
276
277
278
279
280
281
282
283
284
285
   def equivalent(self, other):
      """Check that two constants are equivalent.

      This is check is much weaker than equality.  One generally cannot be
      used in place of the other.  Using this implementation for the __eq__
      will break BitSizeValidator.

      """
      if not isinstance(other, type(self)):
         return False

      return self.value == other.value

286
_var_name_re = re.compile(r"(?P<const>#)?(?P<name>\w+)"
287
                          r"(?:@(?P<type>int|uint|bool|float)?(?P<bits>\d+)?)?"
288
289
                          r"(?P<cond>\([^\)]+\))?"
                          r"(?P<swiz>\.[xyzw]+)?")
290

291
292
class Variable(Value):
   def __init__(self, val, name, varset):
293
      Value.__init__(self, val, name, "variable")
294
295
296
297
298

      m = _var_name_re.match(val)
      assert m and m.group('name') is not None

      self.var_name = m.group('name')
299
300
301
302
303
304
305
306

      # Prevent common cases where someone puts quotes around a literal
      # constant.  If we want to support names that have numeric or
      # punctuation characters, we can me the first assertion more flexible.
      assert self.var_name.isalpha()
      assert self.var_name is not 'True'
      assert self.var_name is not 'False'

307
      self.is_constant = m.group('const') is not None
308
      self.cond = m.group('cond')
309
      self.required_type = m.group('type')
310
      self._bit_size = int(m.group('bits')) if m.group('bits') else None
311
      self.swiz = m.group('swiz')
312
313

      if self.required_type == 'bool':
314
315
316
317
         if self._bit_size is not None:
            assert self._bit_size in type_sizes(self.required_type)
         else:
            self._bit_size = 1
318
319

      if self.required_type is not None:
320
         assert self.required_type in ('float', 'bool', 'int', 'uint')
321
322

      self.index = varset[self.var_name]
323

324
325
   def type(self):
      if self.required_type == 'bool':
326
         return "nir_type_bool"
327
      elif self.required_type in ('int', 'uint'):
328
329
330
331
         return "nir_type_int"
      elif self.required_type == 'float':
         return "nir_type_float"

332
333
334
335
336
337
338
339
340
341
342
343
344
   def equivalent(self, other):
      """Check that two variables are equivalent.

      This is check is much weaker than equality.  One generally cannot be
      used in place of the other.  Using this implementation for the __eq__
      will break BitSizeValidator.

      """
      if not isinstance(other, type(self)):
         return False

      return self.index == other.index

345
346
347
348
349
350
   def swizzle(self):
      if self.swiz is not None:
         swizzles = {'x' : 0, 'y' : 1, 'z' : 2, 'w': 3}
         return '{' + ', '.join([str(swizzles[c]) for c in self.swiz[1:]]) + '}'
      return '{0, 1, 2, 3}'

351
352
_opcode_re = re.compile(r"(?P<inexact>~)?(?P<opcode>\w+)(?:@(?P<bits>\d+))?"
                        r"(?P<cond>\([^\)]+\))?")
353

354
355
class Expression(Value):
   def __init__(self, expr, name_base, varset):
356
      Value.__init__(self, expr, name_base, "expression")
357
358
      assert isinstance(expr, tuple)

359
360
361
362
      m = _opcode_re.match(expr[0])
      assert m and m.group('opcode') is not None

      self.opcode = m.group('opcode')
363
      self._bit_size = int(m.group('bits')) if m.group('bits') else None
364
      self.inexact = m.group('inexact') is not None
365
      self.cond = m.group('cond')
366
367
368
369
370
371
372
373
374
375
376
377
378
379

      # "many-comm-expr" isn't really a condition.  It's notification to the
      # generator that this pattern is known to have too many commutative
      # expressions, and an error should not be generated for this case.
      self.many_commutative_expressions = False
      if self.cond and self.cond.find("many-comm-expr") >= 0:
         # Split the condition into a comma-separated list.  Remove
         # "many-comm-expr".  If there is anything left, put it back together.
         c = self.cond[1:-1].split(",")
         c.remove("many-comm-expr")

         self.cond = "({})".format(",".join(c)) if c else None
         self.many_commutative_expressions = True

380
381
382
      self.sources = [ Value.create(src, "{0}_{1}".format(name_base, i), varset)
                       for (i, src) in enumerate(expr[1:]) ]

383
384
385
386
387
      if self.opcode in conv_opcode_types:
         assert self._bit_size is None, \
                'Expression cannot use an unsized conversion opcode with ' \
                'an explicit size; that\'s silly.'

388
389
      self.__index_comm_exprs(0)

390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
   def equivalent(self, other):
      """Check that two variables are equivalent.

      This is check is much weaker than equality.  One generally cannot be
      used in place of the other.  Using this implementation for the __eq__
      will break BitSizeValidator.

      This implementation does not check for equivalence due to commutativity,
      but it could.

      """
      if not isinstance(other, type(self)):
         return False

      if len(self.sources) != len(other.sources):
         return False

      if self.opcode != other.opcode:
         return False

      return all(s.equivalent(o) for s, o in zip(self.sources, other.sources))

412
413
414
415
   def __index_comm_exprs(self, base_idx):
      """Recursively count and index commutative expressions
      """
      self.comm_exprs = 0
416
417
418
419
420

      # A note about the explicit "len(self.sources)" check. The list of
      # sources comes from user input, and that input might be bad.  Check
      # that the expected second source exists before accessing it. Without
      # this check, a unit test that does "('iadd', 'a')" will crash.
421
      if self.opcode not in conv_opcode_types and \
422
423
424
         "2src_commutative" in opcodes[self.opcode].algebraic_properties and \
         len(self.sources) >= 2 and \
         not self.sources[0].equivalent(self.sources[1]):
425
426
427
428
429
430
431
432
433
434
435
         self.comm_expr_idx = base_idx
         self.comm_exprs += 1
      else:
         self.comm_expr_idx = -1

      for s in self.sources:
         if isinstance(s, Expression):
            s.__index_comm_exprs(base_idx + self.comm_exprs)
            self.comm_exprs += s.comm_exprs

      return self.comm_exprs
436
437

   def c_opcode(self):
438
      return get_c_opcode(self.opcode)
439

440
441
442
   def render(self, cache):
      srcs = "\n".join(src.render(cache) for src in self.sources)
      return srcs + super(Expression, self).render(cache)
443

444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
class BitSizeValidator(object):
   """A class for validating bit sizes of expressions.

   NIR supports multiple bit-sizes on expressions in order to handle things
   such as fp64.  The source and destination of every ALU operation is
   assigned a type and that type may or may not specify a bit size.  Sources
   and destinations whose type does not specify a bit size are considered
   "unsized" and automatically take on the bit size of the corresponding
   register or SSA value.  NIR has two simple rules for bit sizes that are
   validated by nir_validator:

    1) A given SSA def or register has a single bit size that is respected by
       everything that reads from it or writes to it.

    2) The bit sizes of all unsized inputs/outputs on any given ALU
       instruction must match.  They need not match the sized inputs or
       outputs but they must match each other.

   In order to keep nir_algebraic relatively simple and easy-to-use,
   nir_search supports a type of bit-size inference based on the two rules
   above.  This is similar to type inference in many common programming
   languages.  If, for instance, you are constructing an add operation and you
   know the second source is 16-bit, then you know that the other source and
   the destination must also be 16-bit.  There are, however, cases where this
   inference can be ambiguous or contradictory.  Consider, for instance, the
   following transformation:

471
   (('usub_borrow', a, b), ('b2i@32', ('ult', a, b)))
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489

   This transformation can potentially cause a problem because usub_borrow is
   well-defined for any bit-size of integer.  However, b2i always generates a
   32-bit result so it could end up replacing a 64-bit expression with one
   that takes two 64-bit values and produces a 32-bit value.  As another
   example, consider this expression:

   (('bcsel', a, b, 0), ('iand', a, b))

   In this case, in the search expression a must be 32-bit but b can
   potentially have any bit size.  If we had a 64-bit b value, we would end up
   trying to and a 32-bit value with a 64-bit value which would be invalid

   This class solves that problem by providing a validation layer that proves
   that a given search-and-replace operation is 100% well-defined before we
   generate any code.  This ensures that bugs are caught at compile time
   rather than at run time.

490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
   Each value maintains a "bit-size class", which is either an actual bit size
   or an equivalence class with other values that must have the same bit size.
   The validator works by combining bit-size classes with each other according
   to the NIR rules outlined above, checking that there are no inconsistencies.
   When doing this for the replacement expression, we make sure to never change
   the equivalence class of any of the search values. We could make the example
   transforms above work by doing some extra run-time checking of the search
   expression, but we make the user specify those constraints themselves, to
   avoid any surprises. Since the replacement bitsizes can only be connected to
   the source bitsize via variables (variables must have the same bitsize in
   the source and replacment expressions) or the roots of the expression (the
   replacement expression must produce the same bit size as the search
   expression), we prevent merging a variable with anything when processing the
   replacement expression, or specializing the search bitsize
   with anything. The former prevents
505

506
   (('bcsel', a, b, 0), ('iand', a, b))
507

508
509
   from being allowed, since we'd have to merge the bitsizes for a and b due to
   the 'iand', while the latter prevents
510

511
   (('usub_borrow', a, b), ('b2i@32', ('ult', a, b)))
512

513
514
515
   from being allowed, since the search expression has the bit size of a and b,
   which can't be specialized to 32 which is the bitsize of the replace
   expression. It also prevents something like:
516

517
   (('b2i', ('i2b', a)), ('ineq', a, 0))
518

519
520
   since the bitsize of 'b2i', which can be anything, can't be specialized to
   the bitsize of a.
521

522
523
524
525
526
527
528
   After doing all this, we check that every subexpression of the replacement
   was assigned a constant bitsize, the bitsize of a variable, or the bitsize
   of the search expresssion, since those are the things that are known when
   constructing the replacement expresssion. Finally, we record the bitsize
   needed in nir_search_value so that we know what to do when building the
   replacement expression.
   """
529

530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
   def __init__(self, varset):
      self._var_classes = [None] * len(varset.names)

   def compare_bitsizes(self, a, b):
      """Determines which bitsize class is a specialization of the other, or
      whether neither is. When we merge two different bitsizes, the
      less-specialized bitsize always points to the more-specialized one, so
      that calling get_bit_size() always gets you the most specialized bitsize.
      The specialization partial order is given by:
      - Physical bitsizes are always the most specialized, and a different
        bitsize can never specialize another.
      - In the search expression, variables can always be specialized to each
        other and to physical bitsizes. In the replace expression, we disallow
        this to avoid adding extra constraints to the search expression that
        the user didn't specify.
      - Expressions and constants without a bitsize can always be specialized to
        each other and variables, but not the other way around.

        We return -1 if a <= b (b can be specialized to a), 0 if a = b, 1 if a >= b,
        and None if they are not comparable (neither a <= b nor b <= a).
      """
      if isinstance(a, int):
         if isinstance(b, int):
            return 0 if a == b else None
         elif isinstance(b, Variable):
            return -1 if self.is_search else None
         else:
            return -1
      elif isinstance(a, Variable):
         if isinstance(b, int):
            return 1 if self.is_search else None
         elif isinstance(b, Variable):
            return 0 if self.is_search or a.index == b.index else None
         else:
            return -1
      else:
         if isinstance(b, int):
            return 1
         elif isinstance(b, Variable):
            return 1
         else:
            return 0

   def unify_bit_size(self, a, b, error_msg):
      """Record that a must have the same bit-size as b. If both
      have been assigned conflicting physical bit-sizes, call "error_msg" with
      the bit-sizes of self and other to get a message and raise an error.
      In the replace expression, disallow merging variables with other
      variables and physical bit-sizes as well.
      """
      a_bit_size = a.get_bit_size()
      b_bit_size = b if isinstance(b, int) else b.get_bit_size()

      cmp_result = self.compare_bitsizes(a_bit_size, b_bit_size)

      assert cmp_result is not None, \
         error_msg(a_bit_size, b_bit_size)

      if cmp_result < 0:
         b_bit_size.set_bit_size(a)
      elif not isinstance(a_bit_size, int):
         a_bit_size.set_bit_size(b)

   def merge_variables(self, val):
      """Perform the first part of type inference by merging all the different
      uses of the same variable. We always do this as if we're in the search
      expression, even if we're actually not, since otherwise we'd get errors
      if the search expression specified some constraint but the replace
      expression didn't, because we'd be merging a variable and a constant.
      """
      if isinstance(val, Variable):
         if self._var_classes[val.index] is None:
            self._var_classes[val.index] = val
         else:
            other = self._var_classes[val.index]
            self.unify_bit_size(other, val,
                  lambda other_bit_size, bit_size:
                     'Variable {} has conflicting bit size requirements: ' \
                     'it must have bit size {} and {}'.format(
                        val.var_name, other_bit_size, bit_size))
610
      elif isinstance(val, Expression):
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
         for src in val.sources:
            self.merge_variables(src)

   def validate_value(self, val):
      """Validate the an expression by performing classic Hindley-Milner
      type inference on bitsizes. This will detect if there are any conflicting
      requirements, and unify variables so that we know which variables must
      have the same bitsize. If we're operating on the replace expression, we
      will refuse to merge different variables together or merge a variable
      with a constant, in order to prevent surprises due to rules unexpectedly
      not matching at runtime.
      """
      if not isinstance(val, Expression):
         return

626
627
628
629
630
631
632
633
634
635
636
      # Generic conversion ops are special in that they have a single unsized
      # source and an unsized destination and the two don't have to match.
      # This means there's no validation or unioning to do here besides the
      # len(val.sources) check.
      if val.opcode in conv_opcode_types:
         assert len(val.sources) == 1, \
            "Expression {} has {} sources, expected 1".format(
               val, len(val.sources))
         self.validate_value(val.sources[0])
         return

637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
      nir_op = opcodes[val.opcode]
      assert len(val.sources) == nir_op.num_inputs, \
         "Expression {} has {} sources, expected {}".format(
            val, len(val.sources), nir_op.num_inputs)

      for src in val.sources:
         self.validate_value(src)

      dst_type_bits = type_bits(nir_op.output_type)

      # First, unify all the sources. That way, an error coming up because two
      # sources have an incompatible bit-size won't produce an error message
      # involving the destination.
      first_unsized_src = None
      for src_type, src in zip(nir_op.input_types, val.sources):
         src_type_bits = type_bits(src_type)
         if src_type_bits == 0:
            if first_unsized_src is None:
               first_unsized_src = src
656
657
               continue

658
659
660
661
662
663
664
            if self.is_search:
               self.unify_bit_size(first_unsized_src, src,
                  lambda first_unsized_src_bit_size, src_bit_size:
                     'Source {} of {} must have bit size {}, while source {} ' \
                     'must have incompatible bit size {}'.format(
                        first_unsized_src, val, first_unsized_src_bit_size,
                        src, src_bit_size))
665
            else:
666
667
668
669
670
671
672
               self.unify_bit_size(first_unsized_src, src,
                  lambda first_unsized_src_bit_size, src_bit_size:
                     'Sources {} (bit size of {}) and {} (bit size of {}) ' \
                     'of {} may not have the same bit size when building the ' \
                     'replacement expression.'.format(
                        first_unsized_src, first_unsized_src_bit_size, src,
                        src_bit_size, val))
673
         else:
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
            if self.is_search:
               self.unify_bit_size(src, src_type_bits,
                  lambda src_bit_size, unused:
                     '{} must have {} bits, but as a source of nir_op_{} '\
                     'it must have {} bits'.format(
                        src, src_bit_size, nir_op.name, src_type_bits))
            else:
               self.unify_bit_size(src, src_type_bits,
                  lambda src_bit_size, unused:
                     '{} has the bit size of {}, but as a source of ' \
                     'nir_op_{} it must have {} bits, which may not be the ' \
                     'same'.format(
                        src, src_bit_size, nir_op.name, src_type_bits))

      if dst_type_bits == 0:
         if first_unsized_src is not None:
            if self.is_search:
               self.unify_bit_size(val, first_unsized_src,
                  lambda val_bit_size, src_bit_size:
                     '{} must have the bit size of {}, while its source {} ' \
                     'must have incompatible bit size {}'.format(
                        val, val_bit_size, first_unsized_src, src_bit_size))
696
            else:
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
               self.unify_bit_size(val, first_unsized_src,
                  lambda val_bit_size, src_bit_size:
                     '{} must have {} bits, but its source {} ' \
                     '(bit size of {}) may not have that bit size ' \
                     'when building the replacement.'.format(
                        val, val_bit_size, first_unsized_src, src_bit_size))
      else:
         self.unify_bit_size(val, dst_type_bits,
            lambda dst_bit_size, unused:
               '{} must have {} bits, but as a destination of nir_op_{} ' \
               'it must have {} bits'.format(
                  val, dst_bit_size, nir_op.name, dst_type_bits))

   def validate_replace(self, val, search):
      bit_size = val.get_bit_size()
      assert isinstance(bit_size, int) or isinstance(bit_size, Variable) or \
            bit_size == search.get_bit_size(), \
            'Ambiguous bit size for replacement value {}: ' \
            'it cannot be deduced from a variable, a fixed bit size ' \
            'somewhere, or the search expression.'.format(val)

      if isinstance(val, Expression):
         for src in val.sources:
            self.validate_replace(src, search)
721

722
723
724
725
726
   def validate(self, search, replace):
      self.is_search = True
      self.merge_variables(search)
      self.merge_variables(replace)
      self.validate_value(search)
727

728
729
      self.is_search = False
      self.validate_value(replace)
730

731
732
733
734
735
      # Check that search is always more specialized than replace. Note that
      # we're doing this in replace mode, disallowing merging variables.
      search_bit_size = search.get_bit_size()
      replace_bit_size = replace.get_bit_size()
      cmp_result = self.compare_bitsizes(search_bit_size, replace_bit_size)
736

737
738
739
740
      assert cmp_result is not None and cmp_result <= 0, \
         'The search expression bit size {} and replace expression ' \
         'bit size {} may not be the same'.format(
               search_bit_size, replace_bit_size)
741

742
743
744
      replace.set_bit_size(search)

      self.validate_replace(replace, search)
745

746
747
_optimization_ids = itertools.count()

748
749
condition_list = ['true']

750
class SearchAndReplace(object):
751
   def __init__(self, transform):
Mathieu Bridon's avatar
Mathieu Bridon committed
752
      self.id = next(_optimization_ids)
753

754
755
756
757
758
759
760
761
762
763
764
      search = transform[0]
      replace = transform[1]
      if len(transform) > 2:
         self.condition = transform[2]
      else:
         self.condition = 'true'

      if self.condition not in condition_list:
         condition_list.append(self.condition)
      self.condition_index = condition_list.index(self.condition)

765
766
767
768
769
770
      varset = VarSet()
      if isinstance(search, Expression):
         self.search = search
      else:
         self.search = Expression(search, "search{0}".format(self.id), varset)

771
772
      varset.lock()

773
774
775
776
777
      if isinstance(replace, Value):
         self.replace = replace
      else:
         self.replace = Value.create(replace, "replace{0}".format(self.id), varset)

778
779
      BitSizeValidator(varset).validate(self.search, self.replace)

780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
class TreeAutomaton(object):
   """This class calculates a bottom-up tree automaton to quickly search for
   the left-hand sides of tranforms. Tree automatons are a generalization of
   classical NFA's and DFA's, where the transition function determines the
   state of the parent node based on the state of its children. We construct a
   deterministic automaton to match patterns, using a similar algorithm to the
   classical NFA to DFA construction. At the moment, it only matches opcodes
   and constants (without checking the actual value), leaving more detailed
   checking to the search function which actually checks the leaves. The
   automaton acts as a quick filter for the search function, requiring only n
   + 1 table lookups for each n-source operation. The implementation is based
   on the theory described in "Tree Automatons: Two Taxonomies and a Toolkit."
   In the language of that reference, this is a frontier-to-root deterministic
   automaton using only symbol filtering. The filtering is crucial to reduce
   both the time taken to generate the tables and the size of the tables.
   """
   def __init__(self, transforms):
      self.patterns = [t.search for t in transforms]
      self._compute_items()
      self._build_table()
      #print('num items: {}'.format(len(set(self.items.values()))))
      #print('num states: {}'.format(len(self.states)))
      #for state, patterns in zip(self.states, self.patterns):
      #   print('{}: num patterns: {}'.format(state, len(patterns)))

   class IndexMap(object):
      """An indexed list of objects, where one can either lookup an object by
      index or find the index associated to an object quickly using a hash
      table. Compared to a list, it has a constant time index(). Compared to a
      set, it provides a stable iteration order.
      """
      def __init__(self, iterable=()):
         self.objects = []
         self.map = {}
         for obj in iterable:
            self.add(obj)

      def __getitem__(self, i):
         return self.objects[i]

      def __contains__(self, obj):
         return obj in self.map

      def __len__(self):
         return len(self.objects)

      def __iter__(self):
         return iter(self.objects)

      def clear(self):
         self.objects = []
         self.map.clear()

      def index(self, obj):
         return self.map[obj]

      def add(self, obj):
         if obj in self.map:
            return self.map[obj]
         else:
            index = len(self.objects)
            self.objects.append(obj)
            self.map[obj] = index
            return index

      def __repr__(self):
         return 'IndexMap([' + ', '.join(repr(e) for e in self.objects) + '])'

   class Item(object):
      """This represents an "item" in the language of "Tree Automatons." This
      is just a subtree of some pattern, which represents a potential partial
      match at runtime. We deduplicate them, so that identical subtrees of
      different patterns share the same object, and store some extra
      information needed for the main algorithm as well.
      """
      def __init__(self, opcode, children):
         self.opcode = opcode
         self.children = children
         # These are the indices of patterns for which this item is the root node.
         self.patterns = []
         # This the set of opcodes for parents of this item. Used to speed up
         # filtering.
         self.parent_ops = set()

      def __str__(self):
         return '(' + ', '.join([self.opcode] + [str(c) for c in self.children]) + ')'

      def __repr__(self):
         return str(self)

   def _compute_items(self):
      """Build a set of all possible items, deduplicating them."""
      # This is a map from (opcode, sources) to item.
      self.items = {}

      # The set of all opcodes used by the patterns. Used later to avoid
      # building and emitting all the tables for opcodes that aren't used.
      self.opcodes = self.IndexMap()

      def get_item(opcode, children, pattern=None):
880
         commutative = len(children) >= 2 \
881
               and "2src_commutative" in opcodes[opcode].algebraic_properties
882
883
884
         item = self.items.setdefault((opcode, children),
                                      self.Item(opcode, children))
         if commutative:
885
            self.items[opcode, (children[1], children[0]) + children[2:]] = item
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
         if pattern is not None:
            item.patterns.append(pattern)
         return item

      self.wildcard = get_item("__wildcard", ())
      self.const = get_item("__const", ())

      def process_subpattern(src, pattern=None):
         if isinstance(src, Constant):
            # Note: we throw away the actual constant value!
            return self.const
         elif isinstance(src, Variable):
            if src.is_constant:
               return self.const
            else:
               # Note: we throw away which variable it is here! This special
               # item is equivalent to nu in "Tree Automatons."
               return self.wildcard
         else:
            assert isinstance(src, Expression)
            opcode = src.opcode
            stripped = opcode.rstrip('0123456789')
            if stripped in conv_opcode_types:
               # Matches that use conversion opcodes with a specific type,
               # like f2b1, are tricky.  Either we construct the automaton to
               # match specific NIR opcodes like nir_op_f2b1, in which case we
               # need to create separate items for each possible NIR opcode
               # for patterns that have a generic opcode like f2b, or we
               # construct it to match the search opcode, in which case we
               # need to map f2b1 to f2b when constructing the automaton. Here
               # we do the latter.
               opcode = stripped
            self.opcodes.add(opcode)
            children = tuple(process_subpattern(c) for c in src.sources)
            item = get_item(opcode, children, pattern)
            for i, child in enumerate(children):
               child.parent_ops.add(opcode)
            return item

      for i, pattern in enumerate(self.patterns):
         process_subpattern(pattern, i)

   def _build_table(self):
      """This is the core algorithm which builds up the transition table. It
      is based off of Algorithm 5.7.38 "Reachability-based tabulation of Cl .
      Comp_a and Filt_{a,i} using integers to identify match sets." It
      simultaneously builds up a list of all possible "match sets" or
      "states", where each match set represents the set of Item's that match a
      given instruction, and builds up the transition table between states.
      """
      # Map from opcode + filtered state indices to transitioned state.
      self.table = defaultdict(dict)
      # Bijection from state to index. q in the original algorithm is
      # len(self.states)
      self.states = self.IndexMap()
      # List of pattern matches for each state index.
      self.state_patterns = []
      # Map from state index to filtered state index for each opcode.
      self.filter = defaultdict(list)
      # Bijections from filtered state to filtered state index for each
      # opcode, called the "representor sets" in the original algorithm.
      # q_{a,j} in the original algorithm is len(self.rep[op]).
      self.rep = defaultdict(self.IndexMap)

      # Everything in self.states with a index at least worklist_index is part
      # of the worklist of newly created states. There is also a worklist of
      # newly fitered states for each opcode, for which worklist_indices
      # serves a similar purpose. worklist_index corresponds to p in the
      # original algorithm, while worklist_indices is p_{a,j} (although since
      # we only filter by opcode/symbol, it's really just p_a).
      self.worklist_index = 0
      worklist_indices = defaultdict(lambda: 0)

      # This is the set of opcodes for which the filtered worklist is non-empty.
      # It's used to avoid scanning opcodes for which there is nothing to
      # process when building the transition table. It corresponds to new_a in
      # the original algorithm.
      new_opcodes = self.IndexMap()

      # Process states on the global worklist, filtering them for each opcode,
      # updating the filter tables, and updating the filtered worklists if any
      # new filtered states are found. Similar to ComputeRepresenterSets() in
      # the original algorithm, although that only processes a single state.
      def process_new_states():
         while self.worklist_index < len(self.states):
            state = self.states[self.worklist_index]

            # Calculate pattern matches for this state. Each pattern is
            # assigned to a unique item, so we don't have to worry about
            # deduplicating them here. However, we do have to sort them so
            # that they're visited at runtime in the order they're specified
            # in the source.
            patterns = list(sorted(p for item in state for p in item.patterns))
            assert len(self.state_patterns) == self.worklist_index
            self.state_patterns.append(patterns)

            # calculate filter table for this state, and update filtered
            # worklists.
            for op in self.opcodes:
               filt = self.filter[op]
               rep = self.rep[op]
               filtered = frozenset(item for item in state if \
                  op in item.parent_ops)
               if filtered in rep:
                  rep_index = rep.index(filtered)
               else:
                  rep_index = rep.add(filtered)
                  new_opcodes.add(op)
               assert len(filt) == self.worklist_index
               filt.append(rep_index)
            self.worklist_index += 1

      # There are two start states: one which can only match as a wildcard,
      # and one which can match as a wildcard or constant. These will be the
      # states of intrinsics/other instructions and load_const instructions,