/
cgutils.py
1138 lines (932 loc) · 35.6 KB
/
cgutils.py
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
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
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
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
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
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
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
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
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
880
881
882
883
884
885
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
"""
Generic helpers for LLVM code generation.
"""
import collections
from contextlib import contextmanager
import functools
from llvmlite import ir
from numba.core import utils, types, config
import numba.core.datamodel
bool_t = ir.IntType(1)
int8_t = ir.IntType(8)
int32_t = ir.IntType(32)
intp_t = ir.IntType(utils.MACHINE_BITS)
voidptr_t = int8_t.as_pointer()
true_bit = bool_t(1)
false_bit = bool_t(0)
true_byte = int8_t(1)
false_byte = int8_t(0)
def as_bool_bit(builder, value):
return builder.icmp_unsigned('!=', value, value.type(0))
def make_anonymous_struct(builder, values, struct_type=None):
"""
Create an anonymous struct containing the given LLVM *values*.
"""
if struct_type is None:
struct_type = ir.LiteralStructType([v.type for v in values])
struct_val = struct_type(ir.Undefined)
for i, v in enumerate(values):
struct_val = builder.insert_value(struct_val, v, i)
return struct_val
def make_bytearray(buf):
"""
Make a byte array constant from *buf*.
"""
b = bytearray(buf)
n = len(b)
return ir.Constant(ir.ArrayType(ir.IntType(8), n), b)
_struct_proxy_cache = {}
def create_struct_proxy(fe_type, kind='value'):
"""
Returns a specialized StructProxy subclass for the given fe_type.
"""
cache_key = (fe_type, kind)
res = _struct_proxy_cache.get(cache_key)
if res is None:
base = {'value': ValueStructProxy,
'data': DataStructProxy,
}[kind]
clsname = base.__name__ + '_' + str(fe_type)
bases = (base,)
clsmembers = dict(_fe_type=fe_type)
res = type(clsname, bases, clsmembers)
_struct_proxy_cache[cache_key] = res
return res
def copy_struct(dst, src, repl={}):
"""
Copy structure from *src* to *dst* with replacement from *repl*.
"""
repl = repl.copy()
# copy data from src or use those in repl
for k in src._datamodel._fields:
v = repl.pop(k, getattr(src, k))
setattr(dst, k, v)
# use remaining key-values in repl
for k, v in repl.items():
setattr(dst, k, v)
return dst
class _StructProxy(object):
"""
Creates a `Structure` like interface that is constructed with information
from DataModel instance. FE type must have a data model that is a
subclass of StructModel.
"""
# The following class members must be overridden by subclass
_fe_type = None
def __init__(self, context, builder, value=None, ref=None):
self._context = context
self._datamodel = self._context.data_model_manager[self._fe_type]
if not isinstance(self._datamodel, numba.core.datamodel.StructModel):
raise TypeError(
"Not a structure model: {0}".format(self._datamodel))
self._builder = builder
self._be_type = self._get_be_type(self._datamodel)
assert not is_pointer(self._be_type)
outer_ref, ref = self._make_refs(ref)
if ref.type.pointee != self._be_type:
raise AssertionError("bad ref type: expected %s, got %s"
% (self._be_type.as_pointer(), ref.type))
if value is not None:
if value.type != outer_ref.type.pointee:
raise AssertionError("bad value type: expected %s, got %s"
% (outer_ref.type.pointee, value.type))
self._builder.store(value, outer_ref)
self._value = ref
self._outer_ref = outer_ref
def _make_refs(self, ref):
"""
Return an (outer ref, value ref) pair. By default, these are
the same pointers, but a derived class may override this.
"""
if ref is None:
ref = alloca_once(self._builder, self._be_type, zfill=True)
return ref, ref
def _get_be_type(self, datamodel):
raise NotImplementedError
def _cast_member_to_value(self, index, val):
raise NotImplementedError
def _cast_member_from_value(self, index, val):
raise NotImplementedError
def _get_ptr_by_index(self, index):
return gep_inbounds(self._builder, self._value, 0, index)
def _get_ptr_by_name(self, attrname):
index = self._datamodel.get_field_position(attrname)
return self._get_ptr_by_index(index)
def __getattr__(self, field):
"""
Load the LLVM value of the named *field*.
"""
if not field.startswith('_'):
return self[self._datamodel.get_field_position(field)]
else:
raise AttributeError(field)
def __setattr__(self, field, value):
"""
Store the LLVM *value* into the named *field*.
"""
if field.startswith('_'):
return super(_StructProxy, self).__setattr__(field, value)
self[self._datamodel.get_field_position(field)] = value
def __getitem__(self, index):
"""
Load the LLVM value of the field at *index*.
"""
member_val = self._builder.load(self._get_ptr_by_index(index))
return self._cast_member_to_value(index, member_val)
def __setitem__(self, index, value):
"""
Store the LLVM *value* into the field at *index*.
"""
ptr = self._get_ptr_by_index(index)
value = self._cast_member_from_value(index, value)
if value.type != ptr.type.pointee:
if (is_pointer(value.type) and is_pointer(ptr.type.pointee)
and value.type.pointee == ptr.type.pointee.pointee):
# Differ by address-space only
# Auto coerce it
value = self._context.addrspacecast(self._builder,
value,
ptr.type.pointee.addrspace)
else:
raise TypeError("Invalid store of {value.type} to "
"{ptr.type.pointee} in "
"{self._datamodel} "
"(trying to write member #{index})"
.format(value=value, ptr=ptr, self=self,
index=index))
self._builder.store(value, ptr)
def __len__(self):
"""
Return the number of fields.
"""
return self._datamodel.field_count
def _getpointer(self):
"""
Return the LLVM pointer to the underlying structure.
"""
return self._outer_ref
def _getvalue(self):
"""
Load and return the value of the underlying LLVM structure.
"""
return self._builder.load(self._outer_ref)
def _setvalue(self, value):
"""
Store the value in this structure.
"""
assert not is_pointer(value.type)
assert value.type == self._be_type, (value.type, self._be_type)
self._builder.store(value, self._value)
class ValueStructProxy(_StructProxy):
"""
Create a StructProxy suitable for accessing regular values
(e.g. LLVM values or alloca slots).
"""
def _get_be_type(self, datamodel):
return datamodel.get_value_type()
def _cast_member_to_value(self, index, val):
return val
def _cast_member_from_value(self, index, val):
return val
class DataStructProxy(_StructProxy):
"""
Create a StructProxy suitable for accessing data persisted in memory.
"""
def _get_be_type(self, datamodel):
return datamodel.get_data_type()
def _cast_member_to_value(self, index, val):
model = self._datamodel.get_model(index)
return model.from_data(self._builder, val)
def _cast_member_from_value(self, index, val):
model = self._datamodel.get_model(index)
return model.as_data(self._builder, val)
class Structure(object):
"""
A high-level object wrapping a alloca'ed LLVM structure, including
named fields and attribute access.
"""
# XXX Should this warrant several separate constructors?
def __init__(self, context, builder, value=None, ref=None, cast_ref=False):
self._type = context.get_struct_type(self)
self._context = context
self._builder = builder
if ref is None:
self._value = alloca_once(builder, self._type, zfill=True)
if value is not None:
assert not is_pointer(value.type)
assert value.type == self._type, (value.type, self._type)
builder.store(value, self._value)
else:
assert value is None
assert is_pointer(ref.type)
if self._type != ref.type.pointee:
if cast_ref:
ref = builder.bitcast(ref, self._type.as_pointer())
else:
raise TypeError(
"mismatching pointer type: got %s, expected %s"
% (ref.type.pointee, self._type))
self._value = ref
self._namemap = {}
self._fdmap = []
self._typemap = []
base = int32_t(0)
for i, (k, tp) in enumerate(self._fields):
self._namemap[k] = i
self._fdmap.append((base, int32_t(i)))
self._typemap.append(tp)
def _get_ptr_by_index(self, index):
ptr = self._builder.gep(self._value, self._fdmap[index], inbounds=True)
return ptr
def _get_ptr_by_name(self, attrname):
return self._get_ptr_by_index(self._namemap[attrname])
def __getattr__(self, field):
"""
Load the LLVM value of the named *field*.
"""
if not field.startswith('_'):
return self[self._namemap[field]]
else:
raise AttributeError(field)
def __setattr__(self, field, value):
"""
Store the LLVM *value* into the named *field*.
"""
if field.startswith('_'):
return super(Structure, self).__setattr__(field, value)
self[self._namemap[field]] = value
def __getitem__(self, index):
"""
Load the LLVM value of the field at *index*.
"""
return self._builder.load(self._get_ptr_by_index(index))
def __setitem__(self, index, value):
"""
Store the LLVM *value* into the field at *index*.
"""
ptr = self._get_ptr_by_index(index)
if ptr.type.pointee != value.type:
fmt = "Type mismatch: __setitem__(%d, ...) expected %r but got %r"
raise AssertionError(fmt % (index,
str(ptr.type.pointee),
str(value.type)))
self._builder.store(value, ptr)
def __len__(self):
"""
Return the number of fields.
"""
return len(self._namemap)
def _getpointer(self):
"""
Return the LLVM pointer to the underlying structure.
"""
return self._value
def _getvalue(self):
"""
Load and return the value of the underlying LLVM structure.
"""
return self._builder.load(self._value)
def _setvalue(self, value):
"""Store the value in this structure"""
assert not is_pointer(value.type)
assert value.type == self._type, (value.type, self._type)
self._builder.store(value, self._value)
# __iter__ is derived by Python from __len__ and __getitem__
def alloca_once(builder, ty, size=None, name='', zfill=False):
"""Allocate stack memory at the entry block of the current function
pointed by ``builder`` withe llvm type ``ty``. The optional ``size`` arg
set the number of element to allocate. The default is 1. The optional
``name`` arg set the symbol name inside the llvm IR for debugging.
If ``zfill`` is set, fill the memory with zeros at the current
use-site location. Note that the memory is always zero-filled after the
``alloca`` at init-site (the entry block).
"""
if isinstance(size, int):
size = ir.Constant(intp_t, size)
with builder.goto_entry_block():
ptr = builder.alloca(ty, size=size, name=name)
# Always zero-fill at init-site. This is safe.
builder.store(ty(None), ptr)
# Also zero-fill at the use-site
if zfill:
builder.store(ty(None), ptr)
return ptr
def alloca_once_value(builder, value, name=''):
"""
Like alloca_once(), but passing a *value* instead of a type. The
type is inferred and the allocated slot is also initialized with the
given value.
"""
storage = alloca_once(builder, value.type)
builder.store(value, storage)
return storage
def insert_pure_function(module, fnty, name):
"""
Insert a pure function (in the functional programming sense) in the
given module.
"""
fn = module.get_or_insert_function(fnty, name=name)
fn.attributes.add("readonly")
fn.attributes.add("nounwind")
return fn
def terminate(builder, bbend):
bb = builder.basic_block
if bb.terminator is None:
builder.branch(bbend)
def get_null_value(ltype):
return ltype(None)
def is_null(builder, val):
null = get_null_value(val.type)
return builder.icmp_unsigned('==', null, val)
def is_not_null(builder, val):
null = get_null_value(val.type)
return builder.icmp_unsigned('!=', null, val)
def if_unlikely(builder, pred):
return builder.if_then(pred, likely=False)
def if_likely(builder, pred):
return builder.if_then(pred, likely=True)
def ifnot(builder, pred):
return builder.if_then(builder.not_(pred))
def increment_index(builder, val):
"""
Increment an index *val*.
"""
one = val.type(1)
# We pass the "nsw" flag in the hope that LLVM understands the index
# never changes sign. Unfortunately this doesn't always work
# (e.g. ndindex()).
return builder.add(val, one, flags=['nsw'])
Loop = collections.namedtuple('Loop', ('index', 'do_break'))
@contextmanager
def for_range(builder, count, start=None, intp=None):
"""
Generate LLVM IR for a for-loop in [start, count).
*start* is equal to 0 by default.
Yields a Loop namedtuple with the following members:
- `index` is the loop index's value
- `do_break` is a no-argument callable to break out of the loop
"""
if intp is None:
intp = count.type
if start is None:
start = intp(0)
stop = count
bbcond = builder.append_basic_block("for.cond")
bbbody = builder.append_basic_block("for.body")
bbend = builder.append_basic_block("for.end")
def do_break():
builder.branch(bbend)
bbstart = builder.basic_block
builder.branch(bbcond)
with builder.goto_block(bbcond):
index = builder.phi(intp, name="loop.index")
pred = builder.icmp_signed('<', index, stop)
builder.cbranch(pred, bbbody, bbend)
with builder.goto_block(bbbody):
yield Loop(index, do_break)
# Update bbbody as a new basic block may have been activated
bbbody = builder.basic_block
incr = increment_index(builder, index)
terminate(builder, bbcond)
index.add_incoming(start, bbstart)
index.add_incoming(incr, bbbody)
builder.position_at_end(bbend)
@contextmanager
def for_range_slice(builder, start, stop, step, intp=None, inc=True):
"""
Generate LLVM IR for a for-loop based on a slice. Yields a
(index, count) tuple where `index` is the slice index's value
inside the loop, and `count` the iteration count.
Parameters
-------------
builder : object
Builder object
start : int
The beginning value of the slice
stop : int
The end value of the slice
step : int
The step value of the slice
intp :
The data type
inc : boolean, optional
Signals whether the step is positive (True) or negative (False).
Returns
-----------
None
"""
if intp is None:
intp = start.type
bbcond = builder.append_basic_block("for.cond")
bbbody = builder.append_basic_block("for.body")
bbend = builder.append_basic_block("for.end")
bbstart = builder.basic_block
builder.branch(bbcond)
with builder.goto_block(bbcond):
index = builder.phi(intp, name="loop.index")
count = builder.phi(intp, name="loop.count")
if (inc):
pred = builder.icmp_signed('<', index, stop)
else:
pred = builder.icmp_signed('>', index, stop)
builder.cbranch(pred, bbbody, bbend)
with builder.goto_block(bbbody):
yield index, count
bbbody = builder.basic_block
incr = builder.add(index, step)
next_count = increment_index(builder, count)
terminate(builder, bbcond)
index.add_incoming(start, bbstart)
index.add_incoming(incr, bbbody)
count.add_incoming(ir.Constant(intp, 0), bbstart)
count.add_incoming(next_count, bbbody)
builder.position_at_end(bbend)
@contextmanager
def for_range_slice_generic(builder, start, stop, step):
"""
A helper wrapper for for_range_slice(). This is a context manager which
yields two for_range_slice()-alike context managers, the first for
the positive step case, the second for the negative step case.
Use:
with for_range_slice_generic(...) as (pos_range, neg_range):
with pos_range as (idx, count):
...
with neg_range as (idx, count):
...
"""
intp = start.type
is_pos_step = builder.icmp_signed('>=', step, ir.Constant(intp, 0))
pos_for_range = for_range_slice(builder, start, stop, step, intp, inc=True)
neg_for_range = for_range_slice(builder, start, stop, step, intp, inc=False)
@contextmanager
def cm_cond(cond, inner_cm):
with cond:
with inner_cm as value:
yield value
with builder.if_else(is_pos_step, likely=True) as (then, otherwise):
yield cm_cond(then, pos_for_range), cm_cond(otherwise, neg_for_range)
@contextmanager
def loop_nest(builder, shape, intp, order='C'):
"""
Generate a loop nest walking a N-dimensional array.
Yields a tuple of N indices for use in the inner loop body,
iterating over the *shape* space.
If *order* is 'C' (the default), indices are incremented inside-out
(i.e. (0,0), (0,1), (0,2), (1,0) etc.).
If *order* is 'F', they are incremented outside-in
(i.e. (0,0), (1,0), (2,0), (0,1) etc.).
This has performance implications when walking an array as it impacts
the spatial locality of memory accesses.
"""
assert order in 'CF'
if not shape:
# 0-d array
yield ()
else:
if order == 'F':
_swap = lambda x: x[::-1]
else:
_swap = lambda x: x
with _loop_nest(builder, _swap(shape), intp) as indices:
assert len(indices) == len(shape)
yield _swap(indices)
@contextmanager
def _loop_nest(builder, shape, intp):
with for_range(builder, shape[0], intp=intp) as loop:
if len(shape) > 1:
with _loop_nest(builder, shape[1:], intp) as indices:
yield (loop.index,) + indices
else:
yield (loop.index,)
def pack_array(builder, values, ty=None):
"""
Pack a sequence of values in a LLVM array. *ty* should be given
if the array may be empty, in which case the type can't be inferred
from the values.
"""
n = len(values)
if ty is None:
ty = values[0].type
ary = ir.ArrayType(ty, n)(ir.Undefined)
for i, v in enumerate(values):
ary = builder.insert_value(ary, v, i)
return ary
def pack_struct(builder, values):
"""
Pack a sequence of values into a LLVM struct.
"""
structty = ir.LiteralStructType([v.type for v in values])
st = structty(ir.Undefined)
for i, v in enumerate(values):
st = builder.insert_value(st, v, i)
return st
def unpack_tuple(builder, tup, count=None):
"""
Unpack an array or structure of values, return a Python tuple.
"""
if count is None:
# Assuming *tup* is an aggregate
count = len(tup.type.elements)
vals = [builder.extract_value(tup, i)
for i in range(count)]
return vals
def get_item_pointer(context, builder, aryty, ary, inds, wraparound=False,
boundscheck=False):
# Set boundscheck=True for any pointer access that should be
# boundschecked. do_boundscheck() will handle enabling or disabling the
# actual boundschecking based on the user config.
shapes = unpack_tuple(builder, ary.shape, count=aryty.ndim)
strides = unpack_tuple(builder, ary.strides, count=aryty.ndim)
return get_item_pointer2(context, builder, data=ary.data, shape=shapes,
strides=strides, layout=aryty.layout, inds=inds,
wraparound=wraparound, boundscheck=boundscheck)
def do_boundscheck(context, builder, ind, dimlen, axis=None):
def _dbg():
# Remove this when we figure out how to include this information
# in the error message.
if axis is not None:
if isinstance(axis, int):
printf(builder, "debug: IndexError: index %d is out of bounds "
"for axis {} with size %d\n".format(axis), ind, dimlen)
else:
printf(builder, "debug: IndexError: index %d is out of bounds "
"for axis %d with size %d\n", ind, axis,
dimlen)
else:
printf(builder,
"debug: IndexError: index %d is out of bounds for size %d\n",
ind, dimlen)
msg = "index is out of bounds"
out_of_bounds_upper = builder.icmp_signed('>=', ind, dimlen)
with if_unlikely(builder, out_of_bounds_upper):
if config.FULL_TRACEBACKS:
_dbg()
context.call_conv.return_user_exc(builder, IndexError, (msg,))
out_of_bounds_lower = builder.icmp_signed('<', ind, ind.type(0))
with if_unlikely(builder, out_of_bounds_lower):
if config.FULL_TRACEBACKS:
_dbg()
context.call_conv.return_user_exc(builder, IndexError, (msg,))
def get_item_pointer2(context, builder, data, shape, strides, layout, inds,
wraparound=False, boundscheck=False):
# Set boundscheck=True for any pointer access that should be
# boundschecked. do_boundscheck() will handle enabling or disabling the
# actual boundschecking based on the user config.
if wraparound:
# Wraparound
indices = []
for ind, dimlen in zip(inds, shape):
negative = builder.icmp_signed('<', ind, ind.type(0))
wrapped = builder.add(dimlen, ind)
selected = builder.select(negative, wrapped, ind)
indices.append(selected)
else:
indices = inds
if boundscheck:
for axis, (ind, dimlen) in enumerate(zip(indices, shape)):
do_boundscheck(context, builder, ind, dimlen, axis)
if not indices:
# Indexing with empty tuple
return builder.gep(data, [int32_t(0)])
intp = indices[0].type
# Indexing code
if layout in 'CF':
steps = []
# Compute steps for each dimension
if layout == 'C':
# C contiguous
for i in range(len(shape)):
last = intp(1)
for j in shape[i + 1:]:
last = builder.mul(last, j)
steps.append(last)
elif layout == 'F':
# F contiguous
for i in range(len(shape)):
last = intp(1)
for j in shape[:i]:
last = builder.mul(last, j)
steps.append(last)
else:
raise Exception("unreachable")
# Compute index
loc = intp(0)
for i, s in zip(indices, steps):
tmp = builder.mul(i, s)
loc = builder.add(loc, tmp)
ptr = builder.gep(data, [loc])
return ptr
else:
# Any layout
dimoffs = [builder.mul(s, i) for s, i in zip(strides, indices)]
offset = functools.reduce(builder.add, dimoffs)
return pointer_add(builder, data, offset)
def _scalar_pred_against_zero(builder, value, fpred, icond):
nullval = value.type(0)
if isinstance(value.type, (ir.FloatType, ir.DoubleType)):
isnull = fpred(value, nullval)
elif isinstance(value.type, ir.IntType):
isnull = builder.icmp_signed(icond, value, nullval)
else:
raise TypeError("unexpected value type %s" % (value.type,))
return isnull
def is_scalar_zero(builder, value):
"""
Return a predicate representing whether *value* is equal to zero.
"""
return _scalar_pred_against_zero(
builder, value, functools.partial(builder.fcmp_ordered, '=='), '==')
def is_not_scalar_zero(builder, value):
"""
Return a predicate representing whether a *value* is not equal to zero.
(not exactly "not is_scalar_zero" because of nans)
"""
return _scalar_pred_against_zero(
builder, value, functools.partial(builder.fcmp_unordered, '!='), '!=')
def is_scalar_zero_or_nan(builder, value):
"""
Return a predicate representing whether *value* is equal to either zero
or NaN.
"""
return _scalar_pred_against_zero(
builder, value, functools.partial(builder.fcmp_unordered, '=='), '==')
is_true = is_not_scalar_zero
is_false = is_scalar_zero
def is_scalar_neg(builder, value):
"""
Is *value* negative? Assumes *value* is signed.
"""
return _scalar_pred_against_zero(
builder, value, functools.partial(builder.fcmp_ordered, '<'), '<')
def guard_null(context, builder, value, exc_tuple):
"""
Guard against *value* being null or zero.
*exc_tuple* should be a (exception type, arguments...) tuple.
"""
with builder.if_then(is_scalar_zero(builder, value), likely=False):
exc = exc_tuple[0]
exc_args = exc_tuple[1:] or None
context.call_conv.return_user_exc(builder, exc, exc_args)
def guard_memory_error(context, builder, pointer, msg=None):
"""
Guard against *pointer* being NULL (and raise a MemoryError).
"""
assert isinstance(pointer.type, ir.PointerType), pointer.type
exc_args = (msg,) if msg else ()
with builder.if_then(is_null(builder, pointer), likely=False):
context.call_conv.return_user_exc(builder, MemoryError, exc_args)
@contextmanager
def if_zero(builder, value, likely=False):
"""
Execute the given block if the scalar value is zero.
"""
with builder.if_then(is_scalar_zero(builder, value), likely=likely):
yield
guard_zero = guard_null
def is_pointer(ltyp):
"""
Whether the LLVM type *typ* is a struct type.
"""
return isinstance(ltyp, ir.PointerType)
def get_record_member(builder, record, offset, typ):
pval = gep_inbounds(builder, record, 0, offset)
assert not is_pointer(pval.type.pointee)
return builder.bitcast(pval, typ.as_pointer())
def is_neg_int(builder, val):
return builder.icmp_signed('<', val, val.type(0))
def gep_inbounds(builder, ptr, *inds, **kws):
"""
Same as *gep*, but add the `inbounds` keyword.
"""
return gep(builder, ptr, *inds, inbounds=True, **kws)
def gep(builder, ptr, *inds, **kws):
"""
Emit a getelementptr instruction for the given pointer and indices.
The indices can be LLVM values or Python int constants.
"""
name = kws.pop('name', '')
inbounds = kws.pop('inbounds', False)
assert not kws
idx = []
for i in inds:
if isinstance(i, int):
# NOTE: llvm only accepts int32 inside structs, not int64
ind = int32_t(i)
else:
ind = i
idx.append(ind)
return builder.gep(ptr, idx, name=name, inbounds=inbounds)
def pointer_add(builder, ptr, offset, return_type=None):
"""
Add an integral *offset* to pointer *ptr*, and return a pointer
of *return_type* (or, if omitted, the same type as *ptr*).
Note the computation is done in bytes, and ignores the width of
the pointed item type.
"""
intptr = builder.ptrtoint(ptr, intp_t)
if isinstance(offset, int):
offset = intp_t(offset)
intptr = builder.add(intptr, offset)
return builder.inttoptr(intptr, return_type or ptr.type)
def memset(builder, ptr, size, value):
"""
Fill *size* bytes starting from *ptr* with *value*.
"""
fn = builder.module.declare_intrinsic('llvm.memset', (voidptr_t, size.type))
ptr = builder.bitcast(ptr, voidptr_t)
if isinstance(value, int):
value = int8_t(value)
builder.call(fn, [ptr, value, size, bool_t(0)])
def global_constant(builder_or_module, name, value, linkage='internal'):
"""
Get or create a (LLVM module-)global constant with *name* or *value*.
"""
if isinstance(builder_or_module, ir.Module):
module = builder_or_module
else:
module = builder_or_module.module
data = module.add_global_variable(value.type, name=name)
data.linkage = linkage
data.global_constant = True
data.initializer = value
return data
def divmod_by_constant(builder, val, divisor):
"""
Compute the (quotient, remainder) of *val* divided by the constant
positive *divisor*. The semantics reflects those of Python integer
floor division, rather than C's / LLVM's signed division and modulo.
The difference lies with a negative *val*.
"""
assert divisor > 0
divisor = val.type(divisor)
one = val.type(1)
quot = alloca_once(builder, val.type)
with builder.if_else(is_neg_int(builder, val)) as (if_neg, if_pos):
with if_pos:
# quot = val / divisor
quot_val = builder.sdiv(val, divisor)
builder.store(quot_val, quot)
with if_neg:
# quot = -1 + (val + 1) / divisor
val_plus_one = builder.add(val, one)
quot_val = builder.sdiv(val_plus_one, divisor)
builder.store(builder.sub(quot_val, one), quot)
# rem = val - quot * divisor
# (should be slightly faster than a separate modulo operation)
quot_val = builder.load(quot)
rem_val = builder.sub(val, builder.mul(quot_val, divisor))
return quot_val, rem_val
def cbranch_or_continue(builder, cond, bbtrue):
"""
Branch conditionally or continue.
Note: a new block is created and builder is moved to the end of the new
block.
"""
bbcont = builder.append_basic_block('.continue')
builder.cbranch(cond, bbtrue, bbcont)
builder.position_at_end(bbcont)
return bbcont
def memcpy(builder, dst, src, count):
"""
Emit a memcpy to the builder.
Copies each element of dst to src. Unlike the C equivalent, each element
can be any LLVM type.
Assumes
-------
* dst.type == src.type
* count is positive
"""
# Note this does seem to be optimized as a raw memcpy() by LLVM
# whenever possible...
assert dst.type == src.type
with for_range(builder, count, intp=count.type) as loop:
out_ptr = builder.gep(dst, [loop.index])
in_ptr = builder.gep(src, [loop.index])
builder.store(builder.load(in_ptr), out_ptr)
def _raw_memcpy(builder, func_name, dst, src, count, itemsize, align):
size_t = count.type
if isinstance(itemsize, int):
itemsize = ir.Constant(size_t, itemsize)
memcpy = builder.module.declare_intrinsic(func_name,
[voidptr_t, voidptr_t, size_t])
is_volatile = false_bit
builder.call(memcpy, [builder.bitcast(dst, voidptr_t),
builder.bitcast(src, voidptr_t),
builder.mul(count, itemsize),