3 """A shuffle vector fuzz tester.
5 This is a python program to fuzz test the LLVM shufflevector instruction. It
6 generates a function with a random sequnece of shufflevectors, maintaining the
7 element mapping accumulated across the function. It then generates a main
8 function which calls it with a different value in each element and checks that
9 the result matches the expected mapping.
11 Take the output IR printed to stdout, compile it to an executable using whatever
12 set of transforms you want to test, and run the program. If it crashes, it found
16 from __future__
import print_function
26 element_types
= ["i8", "i16", "i32", "i64", "f32", "f64"]
27 parser
= argparse
.ArgumentParser(description
=__doc__
)
29 "-v", "--verbose", action
="store_true", help="Show verbose output"
32 "--seed", default
=str(uuid
.uuid4()), help="A string used to seed the RNG"
35 "--max-shuffle-height",
38 help="Specify a fixed height of shuffle tree to test",
44 help="Include blends of two input vectors",
50 help="Specify a fixed bit width of vector to test",
53 "--fixed-element-type",
54 choices
=element_types
,
55 help="Specify a fixed element type to test",
57 parser
.add_argument("--triple", help="Specify a triple string to include in the IR")
58 args
= parser
.parse_args()
60 random
.seed(args
.seed
)
62 if args
.fixed_element_type
is not None:
63 element_types
= [args
.fixed_element_type
]
65 if args
.fixed_bit_width
is not None:
66 if args
.fixed_bit_width
== 128:
67 width_map
= {"i64": 2, "i32": 4, "i16": 8, "i8": 16, "f64": 2, "f32": 4}
68 (width
, element_type
) = random
.choice(
69 [(width_map
[t
], t
) for t
in element_types
]
71 elif args
.fixed_bit_width
== 256:
72 width_map
= {"i64": 4, "i32": 8, "i16": 16, "i8": 32, "f64": 4, "f32": 8}
73 (width
, element_type
) = random
.choice(
74 [(width_map
[t
], t
) for t
in element_types
]
77 sys
.exit(1) # Checked above by argument parsing.
79 width
= random
.choice([2, 4, 8, 16, 32, 64])
80 element_type
= random
.choice(element_types
)
91 shuffle_range
= (2 * width
) if args
.blends
else width
93 # Because undef (-1) saturates and is indistinguishable when testing the
94 # correctness of a shuffle, we want to bias our fuzz toward having a decent
95 # mixture of non-undef lanes in the end. With a deep shuffle tree, the
96 # probabilies aren't good so we need to bias things. The math here is that if
97 # we uniformly select between -1 and the other inputs, each element of the
98 # result will have the following probability of being undef:
100 # 1 - (shuffle_range/(shuffle_range+1))^max_shuffle_height
102 # More generally, for any probability P of selecting a defined element in
103 # a single shuffle, the end result is:
105 # 1 - P^max_shuffle_height
107 # The power of the shuffle height is the real problem, as we want:
109 # 1 - shuffle_range/(shuffle_range+1)
111 # So we bias the selection of undef at any given node based on the tree
112 # height. Below, let 'A' be 'len(shuffle_range)', 'C' be 'max_shuffle_height',
113 # and 'B' be the bias we use to compensate for
114 # C '((A+1)*A^(1/C))/(A*(A+1)^(1/C))':
116 # 1 - (B * A)/(A + 1)^C = 1 - A/(A + 1)
118 # So at each node we use:
120 # 1 - (B * A)/(A + 1)
121 # = 1 - ((A + 1) * A * A^(1/C))/(A * (A + 1) * (A + 1)^(1/C))
122 # = 1 - ((A + 1) * A^((C + 1)/C))/(A * (A + 1)^((C + 1)/C))
124 # This is the formula we use to select undef lanes in the shuffle.
125 A
= float(shuffle_range
)
126 C
= float(args
.max_shuffle_height
)
128 ((A
+ 1.0) * pow(A
, (C
+ 1.0) / C
)) / (A
* pow(A
+ 1.0, (C
+ 1.0) / C
))
135 if random
.random() <= undef_prob
136 else random
.choice(range(shuffle_range
))
137 for _
in itertools
.repeat(None, width
)
139 for _
in itertools
.repeat(None, args
.max_shuffle_height
- i
)
141 for i
in range(args
.max_shuffle_height
)
145 # Print out the shuffle sequence in a compact form.
148 'Testing shuffle sequence "%s" (v%d%s):'
149 % (args
.seed
, width
, element_type
)
153 for i
, shuffles
in enumerate(shuffle_tree
):
154 print(" tree level %d:" % (i
,), file=sys
.stderr
)
155 for j
, s
in enumerate(shuffles
):
156 print(" shuffle %d: %s" % (j
, s
), file=sys
.stderr
)
157 print("", file=sys
.stderr
)
159 # Symbolically evaluate the shuffle tree.
161 [int(j
% element_modulus
) for j
in range(i
* width
+ 1, (i
+ 1) * width
+ 1)]
162 for i
in range(args
.max_shuffle_height
+ 1)
165 for shuffles
in shuffle_tree
:
169 (results
[i
] if j
< width
else results
[i
+ 1])[j
% width
]
175 for i
, s
in enumerate(shuffles
)
177 if len(results
) != 1:
178 print("ERROR: Bad results: %s" % (results
,), file=sys
.stderr
)
183 print("Which transforms:", file=sys
.stderr
)
184 print(" from: %s" % (inputs
,), file=sys
.stderr
)
185 print(" into: %s" % (result
,), file=sys
.stderr
)
186 print("", file=sys
.stderr
)
188 # The IR uses silly names for floating point types. We also need a same-size
190 integral_element_type
= element_type
191 if element_type
== "f32":
192 integral_element_type
= "i32"
193 element_type
= "float"
194 elif element_type
== "f64":
195 integral_element_type
= "i64"
196 element_type
= "double"
198 # Now we need to generate IR for the shuffle function.
199 subst
= {"N": width
, "T": element_type
, "IT": integral_element_type
}
202 define internal fastcc <%(N)d x %(T)s> @test(%(arguments)s) noinline nounwind {
208 "<%(N)d x %(T)s> %%s.0.%(i)d" % dict(subst
, i
=i
)
209 for i
in range(args
.max_shuffle_height
+ 1)
215 for i
, shuffles
in enumerate(shuffle_tree
):
216 for j
, s
in enumerate(shuffles
):
219 %%s.%(next_i)d.%(j)d = shufflevector <%(N)d x %(T)s> %%s.%(i)d.%(j)d, <%(N)d x %(T)s> %%s.%(i)d.%(next_j)d, <%(N)d x i32> <%(S)s>
230 ["i32 " + (str(si
) if si
!= -1 else "undef") for si
in s
]
237 ret <%(N)d x %(T)s> %%s.%(i)d.0
240 % dict(subst
, i
=len(shuffle_tree
))
243 # Generate some string constants that we can use to report errors.
244 for i
, r
in enumerate(result
):
247 "FAIL(%(seed)s): lane %(lane)d, expected %(result)d, found %%d\n\\0A"
248 % {"seed": args
.seed
, "lane": i
, "result": r
}
250 s
+= "".join(["\\00" for _
in itertools
.repeat(None, 128 - len(s
) + 2)])
253 @error.%(i)d = private unnamed_addr global [128 x i8] c"%(s)s"
258 # Define a wrapper function which is marked 'optnone' to prevent
259 # interprocedural optimizations from deleting the test.
262 define internal fastcc <%(N)d x %(T)s> @test_wrapper(%(arguments)s) optnone noinline {
263 %%result = call fastcc <%(N)d x %(T)s> @test(%(arguments)s)
264 ret <%(N)d x %(T)s> %%result
271 "<%(N)d x %(T)s> %%s.%(i)d" % dict(subst
, i
=i
)
272 for i
in range(args
.max_shuffle_height
+ 1)
278 # Finally, generate a main function which will trap if any lanes are mapped
279 # incorrectly (in an observable way).
284 ; Create a scratch space to print error messages.
285 %%str = alloca [128 x i8]
286 %%str.ptr = getelementptr inbounds [128 x i8], [128 x i8]* %%str, i32 0, i32 0
288 ; Build the input vector and call the test function.
289 %%v = call fastcc <%(N)d x %(T)s> @test_wrapper(%(inputs)s)
290 ; We need to cast this back to an integer type vector to easily check the
292 %%v.cast = bitcast <%(N)d x %(T)s> %%v to <%(N)d x %(IT)s>
300 "<%(N)d x %(T)s> bitcast "
301 "(<%(N)d x %(IT)s> <%(input)s> to <%(N)d x %(T)s>)"
305 ["%(IT)s %(i)d" % dict(subst
, i
=i
) for i
in input]
315 # Test that each non-undef result lane contains the expected value.
316 for i
, r
in enumerate(result
):
321 ; Skip this lane, its value is undef.
322 br label %%test.%(next_i)d
324 % dict(subst
, i
=i
, next_i
=i
+ 1)
330 %%v.%(i)d = extractelement <%(N)d x %(IT)s> %%v.cast, i32 %(i)d
331 %%cmp.%(i)d = icmp ne %(IT)s %%v.%(i)d, %(r)d
332 br i1 %%cmp.%(i)d, label %%die.%(i)d, label %%test.%(next_i)d
335 ; Capture the actual value and print an error message.
336 %%tmp.%(i)d = zext %(IT)s %%v.%(i)d to i2048
337 %%bad.%(i)d = trunc i2048 %%tmp.%(i)d to i32
338 call i32 (i8*, i8*, ...) @sprintf(i8* %%str.ptr, i8* getelementptr inbounds ([128 x i8], [128 x i8]* @error.%(i)d, i32 0, i32 0), i32 %%bad.%(i)d)
339 %%length.%(i)d = call i32 @strlen(i8* %%str.ptr)
340 call i32 @write(i32 2, i8* %%str.ptr, i32 %%length.%(i)d)
341 call void @llvm.trap()
344 % dict(subst
, i
=i
, next_i
=i
+ 1, r
=r
)
353 declare i32 @strlen(i8*)
354 declare i32 @write(i32, i8*, i32)
355 declare i32 @sprintf(i8*, i8*, ...)
356 declare void @llvm.trap() noreturn nounwind
362 if __name__
== "__main__":