mirror of
https://github.com/c64scene-ar/llvm-6502.git
synced 2024-07-22 09:29:31 +00:00
918801162a
a tree of inputs to blend iteratively together. This required a pretty substantial rewrite of the innards. The number of shuffle instructions is now bounded in terms of tree-height. There is a flag to disable blends so that its still possible to test single input shuffles. I've also improved various aspects of how the test program is generated, primarily to simplify the test harness and allow some optimizations to clean up how we actually check the results and build up the inputs. Again, apologies for my likely horrible use of Python... But hey, it works! (Ish?) git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215530 91177308-0d34-0410-b5e6-96231b3b80d8
213 lines
8.0 KiB
Python
Executable File
213 lines
8.0 KiB
Python
Executable File
#!/usr/bin/env python
|
|
|
|
"""A shuffle vector fuzz tester.
|
|
|
|
This is a python program to fuzz test the LLVM shufflevector instruction. It
|
|
generates a function with a random sequnece of shufflevectors, 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.
|
|
"""
|
|
|
|
import argparse
|
|
import itertools
|
|
import random
|
|
import sys
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser(description=__doc__)
|
|
parser.add_argument('seed',
|
|
help='A string used to seed the RNG')
|
|
parser.add_argument('-v', '--verbose', action='store_true',
|
|
help='Show verbose output')
|
|
parser.add_argument('--max-shuffle-height', type=int, default=16,
|
|
help='Specify a fixed height of shuffle tree to test')
|
|
parser.add_argument('--no-blends', dest='blends', action='store_false',
|
|
help='Include blends of two input vectors')
|
|
parser.add_argument('--fixed-bit-width', type=int, choices=[128, 256],
|
|
help='Specify a fixed bit width of vector to test')
|
|
parser.add_argument('--triple',
|
|
help='Specify a triple string to include in the IR')
|
|
args = parser.parse_args()
|
|
|
|
random.seed(args.seed)
|
|
|
|
if args.fixed_bit_width is not None:
|
|
if args.fixed_bit_width == 128:
|
|
(width, element_type) = random.choice(
|
|
[(2, 'i64'), (4, 'i32'), (8, 'i16'), (16, 'i8'),
|
|
(2, 'f64'), (4, 'f32')])
|
|
elif args.fixed_bit_width == 256:
|
|
(width, element_type) = random.choice([
|
|
(4, 'i64'), (8, 'i32'), (16, 'i16'), (32, 'i8'),
|
|
(4, 'f64'), (8, 'f32')])
|
|
else:
|
|
sys.exit(1) # Checked above by argument parsing.
|
|
else:
|
|
width = random.choice([2, 4, 8, 16, 32, 64])
|
|
element_type = random.choice(['i8', 'i16', 'i32', 'i64', 'f32', 'f64'])
|
|
|
|
element_modulus = {
|
|
'i8': 1 << 8, 'i16': 1 << 16, 'i32': 1 << 32, 'i64': 1 << 64,
|
|
'f32': 1 << 32, 'f64': 1 << 64}[element_type]
|
|
|
|
shuffle_range = (2 * width) if args.blends else width
|
|
shuffle_indices = [-1] + range(shuffle_range)
|
|
|
|
shuffle_tree = [[[random.choice(shuffle_indices)
|
|
for _ in itertools.repeat(None, width)]
|
|
for _ in itertools.repeat(None, args.max_shuffle_height - i)]
|
|
for i in xrange(args.max_shuffle_height)]
|
|
|
|
if args.verbose:
|
|
# Print out the shuffle sequence in a compact form.
|
|
print >>sys.stderr, ('Testing shuffle sequence "%s" (v%d%s):' %
|
|
(args.seed, width, element_type))
|
|
for i, shuffles in enumerate(shuffle_tree):
|
|
print >>sys.stderr, ' tree level %d:' % (i,)
|
|
for j, s in enumerate(shuffles):
|
|
print >>sys.stderr, ' shuffle %d: %s' % (j, s)
|
|
print >>sys.stderr, ''
|
|
|
|
# Symbolically evaluate the shuffle tree.
|
|
inputs = [[int(j % element_modulus)
|
|
for j in xrange(i * width + 1, (i + 1) * width + 1)]
|
|
for i in xrange(args.max_shuffle_height + 1)]
|
|
results = inputs
|
|
for shuffles in shuffle_tree:
|
|
results = [[((results[i] if j < width else results[i + 1])[j % width]
|
|
if j != -1 else -1)
|
|
for j in s]
|
|
for i, s in enumerate(shuffles)]
|
|
if len(results) != 1:
|
|
print >>sys.stderr, 'ERROR: Bad results: %s' % (results,)
|
|
sys.exit(1)
|
|
result = results[0]
|
|
|
|
if args.verbose:
|
|
print >>sys.stderr, 'Which transforms:'
|
|
print >>sys.stderr, ' from: %s' % (inputs,)
|
|
print >>sys.stderr, ' into: %s' % (result,)
|
|
print >>sys.stderr, ''
|
|
|
|
# The IR uses silly names for floating point types. We also need a same-size
|
|
# integer type.
|
|
integral_element_type = element_type
|
|
if element_type == 'f32':
|
|
integral_element_type = 'i32'
|
|
element_type = 'float'
|
|
elif element_type == 'f64':
|
|
integral_element_type = 'i64'
|
|
element_type = 'double'
|
|
|
|
# Now we need to generate IR for the shuffle function.
|
|
subst = {'N': width, 'T': element_type, 'IT': integral_element_type}
|
|
print """
|
|
define internal fastcc <%(N)d x %(T)s> @test(%(arguments)s) noinline nounwind {
|
|
entry:""" % dict(subst,
|
|
arguments=', '.join(
|
|
['<%(N)d x %(T)s> %%s.0.%(i)d' % dict(subst, i=i)
|
|
for i in xrange(args.max_shuffle_height + 1)]))
|
|
|
|
for i, shuffles in enumerate(shuffle_tree):
|
|
for j, s in enumerate(shuffles):
|
|
print """
|
|
%%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>
|
|
""".strip('\n') % dict(subst, i=i, next_i=i + 1, j=j, next_j=j + 1,
|
|
S=', '.join(['i32 ' + (str(si) if si != -1 else 'undef')
|
|
for si in s]))
|
|
|
|
print """
|
|
ret <%(N)d x %(T)s> %%s.%(i)d.0
|
|
}
|
|
""" % dict(subst, i=len(shuffle_tree))
|
|
|
|
# Generate some string constants that we can use to report errors.
|
|
for i, r in enumerate(result):
|
|
if r != -1:
|
|
s = ('FAIL(%(seed)s): lane %(lane)d, expected %(result)d, found %%d\\0A' %
|
|
{'seed': args.seed, 'lane': i, 'result': r})
|
|
s += ''.join(['\\00' for _ in itertools.repeat(None, 128 - len(s) + 2)])
|
|
print """
|
|
@error.%(i)d = private unnamed_addr global [128 x i8] c"%(s)s"
|
|
""".strip() % {'i': i, 's': s}
|
|
|
|
# Define a wrapper function which is marked 'optnone' to prevent
|
|
# interprocedural optimizations from deleting the test.
|
|
print """
|
|
define internal fastcc <%(N)d x %(T)s> @test_wrapper(%(arguments)s) optnone noinline {
|
|
%%result = call fastcc <%(N)d x %(T)s> @test(%(arguments)s)
|
|
ret <%(N)d x %(T)s> %%result
|
|
}
|
|
""" % dict(subst,
|
|
arguments=', '.join(['<%(N)d x %(T)s> %%s.%(i)d' % dict(subst, i=i)
|
|
for i in xrange(args.max_shuffle_height + 1)]))
|
|
|
|
# Finally, generate a main function which will trap if any lanes are mapped
|
|
# incorrectly (in an observable way).
|
|
print """
|
|
define i32 @main() {
|
|
entry:
|
|
; Create a scratch space to print error messages.
|
|
%%str = alloca [128 x i8]
|
|
%%str.ptr = getelementptr inbounds [128 x i8]* %%str, i32 0, i32 0
|
|
|
|
; Build the input vector and call the test function.
|
|
%%v = call fastcc <%(N)d x %(T)s> @test_wrapper(%(inputs)s)
|
|
; We need to cast this back to an integer type vector to easily check the
|
|
; result.
|
|
%%v.cast = bitcast <%(N)d x %(T)s> %%v to <%(N)d x %(IT)s>
|
|
br label %%test.0
|
|
""" % dict(subst,
|
|
inputs=', '.join(
|
|
[('<%(N)d x %(T)s> bitcast '
|
|
'(<%(N)d x %(IT)s> <%(input)s> to <%(N)d x %(T)s>)' %
|
|
dict(subst, input=', '.join(['%(IT)s %(i)d' % dict(subst, i=i)
|
|
for i in input])))
|
|
for input in inputs]))
|
|
|
|
# Test that each non-undef result lane contains the expected value.
|
|
for i, r in enumerate(result):
|
|
if r == -1:
|
|
print """
|
|
test.%(i)d:
|
|
; Skip this lane, its value is undef.
|
|
br label %%test.%(next_i)d
|
|
""" % dict(subst, i=i, next_i=i + 1)
|
|
else:
|
|
print """
|
|
test.%(i)d:
|
|
%%v.%(i)d = extractelement <%(N)d x %(IT)s> %%v.cast, i32 %(i)d
|
|
%%cmp.%(i)d = icmp ne %(IT)s %%v.%(i)d, %(r)d
|
|
br i1 %%cmp.%(i)d, label %%die.%(i)d, label %%test.%(next_i)d
|
|
|
|
die.%(i)d:
|
|
; Capture the actual value and print an error message.
|
|
%%tmp.%(i)d = zext %(IT)s %%v.%(i)d to i2048
|
|
%%bad.%(i)d = trunc i2048 %%tmp.%(i)d to i32
|
|
call i32 (i8*, i8*, ...)* @sprintf(i8* %%str.ptr, i8* getelementptr inbounds ([128 x i8]* @error.%(i)d, i32 0, i32 0), i32 %%bad.%(i)d)
|
|
%%length.%(i)d = call i32 @strlen(i8* %%str.ptr)
|
|
%%size.%(i)d = add i32 %%length.%(i)d, 1
|
|
call i32 @write(i32 2, i8* %%str.ptr, i32 %%size.%(i)d)
|
|
call void @llvm.trap()
|
|
unreachable
|
|
""" % dict(subst, i=i, next_i=i + 1, r=r)
|
|
|
|
print """
|
|
test.%d:
|
|
ret i32 0
|
|
}
|
|
|
|
declare i32 @strlen(i8*)
|
|
declare i32 @write(i32, i8*, i32)
|
|
declare i32 @sprintf(i8*, i8*, ...)
|
|
declare void @llvm.trap() noreturn nounwind
|
|
""" % (len(result),)
|
|
|
|
if __name__ == '__main__':
|
|
main()
|