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
| #!/usr/bin/env python
"""A shuffle-select vector fuzz tester.
This is a python program to fuzz test the LLVM shufflevector and select
instructions. It generates a function with a random sequnece of shufflevectors
while optionally attaching it with a select instruction (regular or zero merge),
maintaining the element mapping accumulated across the function. It then
generates a main function which calls it with a different value in each element
and checks that the result matches the expected mapping.
Take the output IR printed to stdout, compile it to an executable using whatever
set of transforms you want to test, and run the program. If it crashes, it found
a bug (an error message with the expected and actual result is printed).
"""
from __future__ import print_function
import random
import uuid
import argparse
# Possibility of one undef index in generated mask for shufflevector instruction
SHUF_UNDEF_POS = 0.15
# Possibility of one undef index in generated mask for select instruction
SEL_UNDEF_POS = 0.15
# Possibility of adding a select instruction to the result of a shufflevector
ADD_SEL_POS = 0.4
# If we are adding a select instruction, this is the possibility of a
# merge-select instruction (1 - MERGE_SEL_POS = possibility of zero-merge-select
# instruction.
MERGE_SEL_POS = 0.5
test_template = r'''
define internal fastcc {ty} @test({inputs}) noinline nounwind {{
entry:
{instructions}
ret {ty} {last_name}
}}
'''
error_template = r'''@error.{lane} = private unnamed_addr global [64 x i8] c"FAIL: lane {lane}, expected {exp}, found %d\0A{padding}"'''
main_template = r'''
define i32 @main() {{
entry:
; Create a scratch space to print error messages.
%str = alloca [64 x i8]
%str.ptr = getelementptr inbounds [64 x i8], [64 x i8]* %str, i32 0, i32 0
; Build the input vector and call the test function.
%v = call fastcc {ty} @test({inputs})
br label %test.0
{check_die}
}}
declare i32 @strlen(i8*)
declare i32 @write(i32, i8*, i32)
declare i32 @sprintf(i8*, i8*, ...)
declare void @llvm.trap() noreturn nounwind
'''
check_template = r'''
test.{lane}:
%v.{lane} = extractelement {ty} %v, i32 {lane}
%cmp.{lane} = {i_f}cmp {ordered}ne {scalar_ty} %v.{lane}, {exp}
br i1 %cmp.{lane}, label %die.{lane}, label %test.{n_lane}
'''
undef_check_template = r'''
test.{lane}:
; Skip this lane, its value is undef.
br label %test.{n_lane}
'''
die_template = r'''
die.{lane}:
; Capture the actual value and print an error message.
call i32 (i8*, i8*, ...) @sprintf(i8* %str.ptr, i8* getelementptr inbounds ([64 x i8], [64 x i8]* @error.{lane}, i32 0, i32 0), {scalar_ty} %v.{lane})
%length.{lane} = call i32 @strlen(i8* %str.ptr)
call i32 @write(i32 2, i8* %str.ptr, i32 %length.{lane})
call void @llvm.trap()
unreachable
'''
class Type:
def __init__(self, is_float, elt_width, elt_num):
self.is_float = is_float # Boolean
self.elt_width = elt_width # Integer
self.elt_num = elt_num # Integer
def dump(self):
if self.is_float:
str_elt = 'float' if self.elt_width == 32 else 'double'
else:
str_elt = 'i' + str(self.elt_width)
if self.elt_num == 1:
return str_elt
else:
return '<' + str(self.elt_num) + ' x ' + str_elt + '>'
def get_scalar_type(self):
return Type(self.is_float, self.elt_width, 1)
# Class to represent any value (variable) that can be used.
class Value:
def __init__(self, name, ty, value = None):
self.ty = ty # Type
self.name = name # String
self.value = value # list of integers or floating points
# Class to represent an IR instruction (shuffle/select).
class Instruction(Value):
def __init__(self, name, ty, op0, op1, mask):
Value.__init__(self, name, ty)
self.op0 = op0 # Value
self.op1 = op1 # Value
self.mask = mask # list of integers
def dump(self): pass
def calc_value(self): pass
# Class to represent an IR shuffle instruction
class ShufInstr(Instruction):
shuf_template = ' {name} = shufflevector {ty} {op0}, {ty} {op1}, <{num} x i32> {mask}\n'
def __init__(self, name, ty, op0, op1, mask):
Instruction.__init__(self, '%shuf' + name, ty, op0, op1, mask)
def dump(self):
str_mask = [('i32 ' + str(idx)) if idx != -1 else 'i32 undef' for idx in self.mask]
str_mask = '<' + (', ').join(str_mask) + '>'
return self.shuf_template.format(name = self.name, ty = self.ty.dump(), op0 = self.op0.name,
op1 = self.op1.name, num = self.ty.elt_num, mask = str_mask)
def calc_value(self):
if self.value != None:
print('Trying to calculate the value of a shuffle instruction twice')
exit(1)
result = []
for i in range(len(self.mask)):
index = self.mask[i]
if index < self.ty.elt_num and index >= 0:
result.append(self.op0.value[index])
elif index >= self.ty.elt_num:
index = index % self.ty.elt_num
result.append(self.op1.value[index])
else: # -1 => undef
result.append(-1)
self.value = result
# Class to represent an IR select instruction
class SelectInstr(Instruction):
sel_template = ' {name} = select <{num} x i1> {mask}, {ty} {op0}, {ty} {op1}\n'
def __init__(self, name, ty, op0, op1, mask):
Instruction.__init__(self, '%sel' + name, ty, op0, op1, mask)
def dump(self):
str_mask = [('i1 ' + str(idx)) if idx != -1 else 'i1 undef' for idx in self.mask]
str_mask = '<' + (', ').join(str_mask) + '>'
return self.sel_template.format(name = self.name, ty = self.ty.dump(), op0 = self.op0.name,
op1 = self.op1.name, num = self.ty.elt_num, mask = str_mask)
def calc_value(self):
if self.value != None:
print('Trying to calculate the value of a select instruction twice')
exit(1)
result = []
for i in range(len(self.mask)):
index = self.mask[i]
if index == 1:
result.append(self.op0.value[i])
elif index == 0:
result.append(self.op1.value[i])
else: # -1 => undef
result.append(-1)
self.value = result
# Returns a list of Values initialized with actual numbers according to the
# provided type
def gen_inputs(ty, num):
inputs = []
for i in range(num):
inp = []
for j in range(ty.elt_num):
if ty.is_float:
inp.append(float(i*ty.elt_num + j))
else:
inp.append((i*ty.elt_num + j) % (1 << ty.elt_width))
inputs.append(Value('%inp' + str(i), ty, inp))
return inputs
# Returns a random vector type to be tested
# In case one of the dimensions (scalar type/number of elements) is provided,
# fill the blank dimension and return appropriate Type object.
def get_random_type(ty, num_elts):
if ty != None:
if ty == 'i8':
is_float = False
width = 8
elif ty == 'i16':
is_float = False
width = 16
elif ty == 'i32':
is_float = False
width = 32
elif ty == 'i64':
is_float = False
width = 64
elif ty == 'f32':
is_float = True
width = 32
elif ty == 'f64':
is_float = True
width = 64
int_elt_widths = [8, 16, 32, 64]
float_elt_widths = [32, 64]
if num_elts == None:
num_elts = random.choice(range(2, 65))
if ty == None:
# 1 for integer type, 0 for floating-point
if random.randint(0,1):
is_float = False
width = random.choice(int_elt_widths)
else:
is_float = True
width = random.choice(float_elt_widths)
return Type(is_float, width, num_elts)
# Generate mask for shufflevector IR instruction, with SHUF_UNDEF_POS possibility
# of one undef index.
def gen_shuf_mask(ty):
mask = []
for i in range(ty.elt_num):
if SHUF_UNDEF_POS/ty.elt_num > random.random():
mask.append(-1)
else:
mask.append(random.randint(0, ty.elt_num*2 - 1))
return mask
# Generate mask for select IR instruction, with SEL_UNDEF_POS possibility
# of one undef index.
def gen_sel_mask(ty):
mask = []
for i in range(ty.elt_num):
if SEL_UNDEF_POS/ty.elt_num > random.random():
mask.append(-1)
else:
mask.append(random.randint(0, 1))
return mask
# Generate shuffle instructions with optional select instruction after.
def gen_insts(inputs, ty):
int_zero_init = Value('zeroinitializer', ty, [0]*ty.elt_num)
float_zero_init = Value('zeroinitializer', ty, [0.0]*ty.elt_num)
insts = []
name_idx = 0
while len(inputs) > 1:
# Choose 2 available Values - remove them from inputs list.
[idx0, idx1] = sorted(random.sample(range(len(inputs)), 2))
op0 = inputs[idx0]
op1 = inputs[idx1]
# Create the shuffle instruction.
shuf_mask = gen_shuf_mask(ty)
shuf_inst = ShufInstr(str(name_idx), ty, op0, op1, shuf_mask)
shuf_inst.calc_value()
# Add the new shuffle instruction to the list of instructions.
insts.append(shuf_inst)
# Optionally, add select instruction with the result of the previous shuffle.
if random.random() < ADD_SEL_POS:
# Either blending with a random Value or with an all-zero vector.
if random.random() < MERGE_SEL_POS:
op2 = random.choice(inputs)
else:
op2 = float_zero_init if ty.is_float else int_zero_init
select_mask = gen_sel_mask(ty)
select_inst = SelectInstr(str(name_idx), ty, shuf_inst, op2, select_mask)
select_inst.calc_value()
# Add the select instructions to the list of instructions and to the available Values.
insts.append(select_inst)
inputs.append(select_inst)
else:
# If the shuffle instruction is not followed by select, add it to the available Values.
inputs.append(shuf_inst)
del inputs[idx1]
del inputs[idx0]
name_idx += 1
return insts
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--seed', default=str(uuid.uuid4()),
help='A string used to seed the RNG')
parser.add_argument('--max-num-inputs', type=int, default=20,
help='Specify the maximum number of vector inputs for the test. (default: 20)')
parser.add_argument('--min-num-inputs', type=int, default=10,
help='Specify the minimum number of vector inputs for the test. (default: 10)')
parser.add_argument('--type', default=None,
help='''
Choose specific type to be tested.
i8, i16, i32, i64, f32 or f64.
(default: random)''')
parser.add_argument('--num-elts', default=None, type=int,
help='Choose specific number of vector elements to be tested. (default: random)')
args = parser.parse_args()
print('; The seed used for this test is ' + args.seed)
assert args.min_num_inputs < args.max_num_inputs , "Minimum value greater than maximum."
assert args.type in [None, 'i8', 'i16', 'i32', 'i64', 'f32', 'f64'], "Illegal type."
assert args.num_elts == None or args.num_elts > 0, "num_elts must be a positive integer."
random.seed(args.seed)
ty = get_random_type(args.type, args.num_elts)
inputs = gen_inputs(ty, random.randint(args.min_num_inputs, args.max_num_inputs))
inputs_str = (', ').join([inp.ty.dump() + ' ' + inp.name for inp in inputs])
inputs_values = [inp.value for inp in inputs]
insts = gen_insts(inputs, ty)
assert len(inputs) == 1, "Only one value should be left after generating phase"
res = inputs[0]
# print the actual test function by dumping the generated instructions.
insts_str = ''.join([inst.dump() for inst in insts])
print(test_template.format(ty = ty.dump(), inputs = inputs_str,
instructions = insts_str, last_name = res.name))
# Print the error message templates as global strings
for i in range(len(res.value)):
pad = ''.join(['\\00']*(31 - len(str(i)) - len(str(res.value[i]))))
print(error_template.format(lane = str(i), exp = str(res.value[i]),
padding = pad))
# Prepare the runtime checks and failure handlers.
scalar_ty = ty.get_scalar_type()
check_die = ''
i_f = 'f' if ty.is_float else 'i'
ordered = 'o' if ty.is_float else ''
for i in range(len(res.value)):
if res.value[i] != -1:
# Emit runtime check for each non-undef expected value.
check_die += check_template.format(lane = str(i), n_lane = str(i+1),
ty = ty.dump(), i_f = i_f, scalar_ty = scalar_ty.dump(),
exp = str(res.value[i]), ordered = ordered)
# Emit failure handler for each runtime check with proper error message
check_die += die_template.format(lane = str(i), scalar_ty = scalar_ty.dump())
else:
# Ignore lanes with undef result
check_die += undef_check_template.format(lane = str(i), n_lane = str(i+1))
check_die += '\ntest.' + str(len(res.value)) + ':\n'
check_die += ' ret i32 0'
# Prepare the input values passed to the test function.
inputs_values = [', '.join([scalar_ty.dump() + ' ' + str(i) for i in inp]) for inp in inputs_values]
inputs = ', '.join([ty.dump() + ' <' + inp + '>' for inp in inputs_values])
print(main_template.format(ty = ty.dump(), inputs = inputs, check_die = check_die))
if __name__ == '__main__':
main()
|