111
|
1 #!/usr/bin/env python3
|
|
2 #
|
|
3 # Script to analyze results of our branch prediction heuristics
|
|
4 #
|
|
5 # This file is part of GCC.
|
|
6 #
|
|
7 # GCC is free software; you can redistribute it and/or modify it under
|
|
8 # the terms of the GNU General Public License as published by the Free
|
|
9 # Software Foundation; either version 3, or (at your option) any later
|
|
10 # version.
|
|
11 #
|
|
12 # GCC is distributed in the hope that it will be useful, but WITHOUT ANY
|
|
13 # WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
14 # FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
15 # for more details.
|
|
16 #
|
|
17 # You should have received a copy of the GNU General Public License
|
|
18 # along with GCC; see the file COPYING3. If not see
|
|
19 # <http://www.gnu.org/licenses/>. */
|
|
20 #
|
|
21 #
|
|
22 #
|
|
23 # This script is used to calculate two basic properties of the branch prediction
|
|
24 # heuristics - coverage and hitrate. Coverage is number of executions
|
|
25 # of a given branch matched by the heuristics and hitrate is probability
|
|
26 # that once branch is predicted as taken it is really taken.
|
|
27 #
|
|
28 # These values are useful to determine the quality of given heuristics.
|
|
29 # Hitrate may be directly used in predict.def.
|
|
30 #
|
|
31 # Usage:
|
|
32 # Step 1: Compile and profile your program. You need to use -fprofile-generate
|
|
33 # flag to get the profiles.
|
|
34 # Step 2: Make a reference run of the intrumented application.
|
|
35 # Step 3: Compile the program with collected profile and dump IPA profiles
|
|
36 # (-fprofile-use -fdump-ipa-profile-details)
|
|
37 # Step 4: Collect all generated dump files:
|
|
38 # find . -name '*.profile' | xargs cat > dump_file
|
|
39 # Step 5: Run the script:
|
|
40 # ./analyze_brprob.py dump_file
|
|
41 # and read results. Basically the following table is printed:
|
|
42 #
|
|
43 # HEURISTICS BRANCHES (REL) HITRATE COVERAGE (REL)
|
|
44 # early return (on trees) 3 0.2% 35.83% / 93.64% 66360 0.0%
|
|
45 # guess loop iv compare 8 0.6% 53.35% / 53.73% 11183344 0.0%
|
|
46 # call 18 1.4% 31.95% / 69.95% 51880179 0.2%
|
|
47 # loop guard 23 1.8% 84.13% / 84.85% 13749065956 42.2%
|
|
48 # opcode values positive (on trees) 42 3.3% 15.71% / 84.81% 6771097902 20.8%
|
|
49 # opcode values nonequal (on trees) 226 17.6% 72.48% / 72.84% 844753864 2.6%
|
|
50 # loop exit 231 18.0% 86.97% / 86.98% 8952666897 27.5%
|
|
51 # loop iterations 239 18.6% 91.10% / 91.10% 3062707264 9.4%
|
|
52 # DS theory 281 21.9% 82.08% / 83.39% 7787264075 23.9%
|
|
53 # no prediction 293 22.9% 46.92% / 70.70% 2293267840 7.0%
|
|
54 # guessed loop iterations 313 24.4% 76.41% / 76.41% 10782750177 33.1%
|
|
55 # first match 708 55.2% 82.30% / 82.31% 22489588691 69.0%
|
|
56 # combined 1282 100.0% 79.76% / 81.75% 32570120606 100.0%
|
|
57 #
|
|
58 #
|
|
59 # The heuristics called "first match" is a heuristics used by GCC branch
|
|
60 # prediction pass and it predicts 55.2% branches correctly. As you can,
|
|
61 # the heuristics has very good covertage (69.05%). On the other hand,
|
|
62 # "opcode values nonequal (on trees)" heuristics has good hirate, but poor
|
|
63 # coverage.
|
|
64
|
|
65 import sys
|
|
66 import os
|
|
67 import re
|
|
68 import argparse
|
|
69
|
|
70 from math import *
|
|
71
|
|
72 counter_aggregates = set(['combined', 'first match', 'DS theory',
|
|
73 'no prediction'])
|
131
|
74 hot_threshold = 10
|
111
|
75
|
|
76 def percentage(a, b):
|
|
77 return 100.0 * a / b
|
|
78
|
|
79 def average(values):
|
|
80 return 1.0 * sum(values) / len(values)
|
|
81
|
|
82 def average_cutoff(values, cut):
|
|
83 l = len(values)
|
|
84 skip = floor(l * cut / 2)
|
|
85 if skip > 0:
|
|
86 values.sort()
|
|
87 values = values[skip:-skip]
|
|
88 return average(values)
|
|
89
|
|
90 def median(values):
|
|
91 values.sort()
|
|
92 return values[int(len(values) / 2)]
|
|
93
|
|
94 class PredictDefFile:
|
|
95 def __init__(self, path):
|
|
96 self.path = path
|
|
97 self.predictors = {}
|
|
98
|
|
99 def parse_and_modify(self, heuristics, write_def_file):
|
|
100 lines = [x.rstrip() for x in open(self.path).readlines()]
|
|
101
|
|
102 p = None
|
|
103 modified_lines = []
|
|
104 for l in lines:
|
|
105 if l.startswith('DEF_PREDICTOR'):
|
|
106 m = re.match('.*"(.*)".*', l)
|
|
107 p = m.group(1)
|
|
108 elif l == '':
|
|
109 p = None
|
|
110
|
|
111 if p != None:
|
|
112 heuristic = [x for x in heuristics if x.name == p]
|
|
113 heuristic = heuristic[0] if len(heuristic) == 1 else None
|
|
114
|
|
115 m = re.match('.*HITRATE \(([^)]*)\).*', l)
|
|
116 if (m != None):
|
|
117 self.predictors[p] = int(m.group(1))
|
|
118
|
|
119 # modify the line
|
|
120 if heuristic != None:
|
|
121 new_line = (l[:m.start(1)]
|
|
122 + str(round(heuristic.get_hitrate()))
|
|
123 + l[m.end(1):])
|
|
124 l = new_line
|
|
125 p = None
|
|
126 elif 'PROB_VERY_LIKELY' in l:
|
|
127 self.predictors[p] = 100
|
|
128 modified_lines.append(l)
|
|
129
|
|
130 # save the file
|
|
131 if write_def_file:
|
|
132 with open(self.path, 'w+') as f:
|
|
133 for l in modified_lines:
|
|
134 f.write(l + '\n')
|
131
|
135 class Heuristics:
|
|
136 def __init__(self, count, hits, fits):
|
|
137 self.count = count
|
|
138 self.hits = hits
|
|
139 self.fits = fits
|
111
|
140
|
|
141 class Summary:
|
|
142 def __init__(self, name):
|
|
143 self.name = name
|
131
|
144 self.edges= []
|
|
145
|
|
146 def branches(self):
|
|
147 return len(self.edges)
|
|
148
|
|
149 def hits(self):
|
|
150 return sum([x.hits for x in self.edges])
|
|
151
|
|
152 def fits(self):
|
|
153 return sum([x.fits for x in self.edges])
|
|
154
|
|
155 def count(self):
|
|
156 return sum([x.count for x in self.edges])
|
|
157
|
|
158 def successfull_branches(self):
|
|
159 return len([x for x in self.edges if 2 * x.hits >= x.count])
|
111
|
160
|
|
161 def get_hitrate(self):
|
131
|
162 return 100.0 * self.hits() / self.count()
|
111
|
163
|
|
164 def get_branch_hitrate(self):
|
131
|
165 return 100.0 * self.successfull_branches() / self.branches()
|
111
|
166
|
|
167 def count_formatted(self):
|
131
|
168 v = self.count()
|
|
169 for unit in ['', 'k', 'M', 'G', 'T', 'P', 'E', 'Z', 'Y']:
|
111
|
170 if v < 1000:
|
|
171 return "%3.2f%s" % (v, unit)
|
|
172 v /= 1000.0
|
|
173 return "%.1f%s" % (v, 'Y')
|
|
174
|
131
|
175 def count(self):
|
|
176 return sum([x.count for x in self.edges])
|
|
177
|
111
|
178 def print(self, branches_max, count_max, predict_def):
|
131
|
179 # filter out most hot edges (if requested)
|
|
180 self.edges = sorted(self.edges, reverse = True, key = lambda x: x.count)
|
|
181 if args.coverage_threshold != None:
|
|
182 threshold = args.coverage_threshold * self.count() / 100
|
|
183 edges = [x for x in self.edges if x.count < threshold]
|
|
184 if len(edges) != 0:
|
|
185 self.edges = edges
|
|
186
|
111
|
187 predicted_as = None
|
|
188 if predict_def != None and self.name in predict_def.predictors:
|
|
189 predicted_as = predict_def.predictors[self.name]
|
|
190
|
|
191 print('%-40s %8i %5.1f%% %11.2f%% %7.2f%% / %6.2f%% %14i %8s %5.1f%%' %
|
131
|
192 (self.name, self.branches(),
|
|
193 percentage(self.branches(), branches_max),
|
111
|
194 self.get_branch_hitrate(),
|
|
195 self.get_hitrate(),
|
131
|
196 percentage(self.fits(), self.count()),
|
|
197 self.count(), self.count_formatted(),
|
|
198 percentage(self.count(), count_max)), end = '')
|
111
|
199
|
|
200 if predicted_as != None:
|
|
201 print('%12i%% %5.1f%%' % (predicted_as,
|
|
202 self.get_hitrate() - predicted_as), end = '')
|
131
|
203 else:
|
|
204 print(' ' * 20, end = '')
|
|
205
|
|
206 # print details about the most important edges
|
|
207 if args.coverage_threshold == None:
|
|
208 edges = [x for x in self.edges[:100] if x.count * hot_threshold > self.count()]
|
|
209 if args.verbose:
|
|
210 for c in edges:
|
|
211 r = 100.0 * c.count / self.count()
|
|
212 print(' %.0f%%:%d' % (r, c.count), end = '')
|
|
213 elif len(edges) > 0:
|
|
214 print(' %0.0f%%:%d' % (100.0 * sum([x.count for x in edges]) / self.count(), len(edges)), end = '')
|
|
215
|
111
|
216 print()
|
|
217
|
|
218 class Profile:
|
|
219 def __init__(self, filename):
|
|
220 self.filename = filename
|
|
221 self.heuristics = {}
|
|
222 self.niter_vector = []
|
|
223
|
|
224 def add(self, name, prediction, count, hits):
|
|
225 if not name in self.heuristics:
|
|
226 self.heuristics[name] = Summary(name)
|
|
227
|
|
228 s = self.heuristics[name]
|
|
229
|
|
230 if prediction < 50:
|
|
231 hits = count - hits
|
|
232 remaining = count - hits
|
131
|
233 fits = max(hits, remaining)
|
111
|
234
|
131
|
235 s.edges.append(Heuristics(count, hits, fits))
|
111
|
236
|
|
237 def add_loop_niter(self, niter):
|
|
238 if niter > 0:
|
|
239 self.niter_vector.append(niter)
|
|
240
|
|
241 def branches_max(self):
|
131
|
242 return max([v.branches() for k, v in self.heuristics.items()])
|
111
|
243
|
|
244 def count_max(self):
|
131
|
245 return max([v.count() for k, v in self.heuristics.items()])
|
111
|
246
|
|
247 def print_group(self, sorting, group_name, heuristics, predict_def):
|
|
248 count_max = self.count_max()
|
|
249 branches_max = self.branches_max()
|
|
250
|
131
|
251 sorter = lambda x: x.branches()
|
111
|
252 if sorting == 'branch-hitrate':
|
|
253 sorter = lambda x: x.get_branch_hitrate()
|
|
254 elif sorting == 'hitrate':
|
|
255 sorter = lambda x: x.get_hitrate()
|
|
256 elif sorting == 'coverage':
|
|
257 sorter = lambda x: x.count
|
|
258 elif sorting == 'name':
|
|
259 sorter = lambda x: x.name.lower()
|
|
260
|
131
|
261 print('%-40s %8s %6s %12s %18s %14s %8s %6s %12s %6s %s' %
|
111
|
262 ('HEURISTICS', 'BRANCHES', '(REL)',
|
|
263 'BR. HITRATE', 'HITRATE', 'COVERAGE', 'COVERAGE', '(REL)',
|
131
|
264 'predict.def', '(REL)', 'HOT branches (>%d%%)' % hot_threshold))
|
111
|
265 for h in sorted(heuristics, key = sorter):
|
|
266 h.print(branches_max, count_max, predict_def)
|
|
267
|
|
268 def dump(self, sorting):
|
|
269 heuristics = self.heuristics.values()
|
|
270 if len(heuristics) == 0:
|
|
271 print('No heuristics available')
|
|
272 return
|
|
273
|
|
274 predict_def = None
|
|
275 if args.def_file != None:
|
|
276 predict_def = PredictDefFile(args.def_file)
|
|
277 predict_def.parse_and_modify(heuristics, args.write_def_file)
|
|
278
|
|
279 special = list(filter(lambda x: x.name in counter_aggregates,
|
|
280 heuristics))
|
|
281 normal = list(filter(lambda x: x.name not in counter_aggregates,
|
|
282 heuristics))
|
|
283
|
|
284 self.print_group(sorting, 'HEURISTICS', normal, predict_def)
|
|
285 print()
|
|
286 self.print_group(sorting, 'HEURISTIC AGGREGATES', special, predict_def)
|
|
287
|
|
288 if len(self.niter_vector) > 0:
|
|
289 print ('\nLoop count: %d' % len(self.niter_vector)),
|
|
290 print(' avg. # of iter: %.2f' % average(self.niter_vector))
|
|
291 print(' median # of iter: %.2f' % median(self.niter_vector))
|
|
292 for v in [1, 5, 10, 20, 30]:
|
|
293 cut = 0.01 * v
|
|
294 print(' avg. (%d%% cutoff) # of iter: %.2f'
|
|
295 % (v, average_cutoff(self.niter_vector, cut)))
|
|
296
|
|
297 parser = argparse.ArgumentParser()
|
|
298 parser.add_argument('dump_file', metavar = 'dump_file',
|
|
299 help = 'IPA profile dump file')
|
|
300 parser.add_argument('-s', '--sorting', dest = 'sorting',
|
|
301 choices = ['branches', 'branch-hitrate', 'hitrate', 'coverage', 'name'],
|
|
302 default = 'branches')
|
|
303 parser.add_argument('-d', '--def-file', help = 'path to predict.def')
|
|
304 parser.add_argument('-w', '--write-def-file', action = 'store_true',
|
|
305 help = 'Modify predict.def file in order to set new numbers')
|
131
|
306 parser.add_argument('-c', '--coverage-threshold', type = int,
|
|
307 help = 'Ignore edges that have percentage coverage >= coverage-threshold')
|
|
308 parser.add_argument('-v', '--verbose', action = 'store_true', help = 'Print verbose informations')
|
111
|
309
|
|
310 args = parser.parse_args()
|
|
311
|
|
312 profile = Profile(args.dump_file)
|
|
313 loop_niter_str = ';; profile-based iteration count: '
|
131
|
314
|
111
|
315 for l in open(args.dump_file):
|
131
|
316 if l.startswith(';;heuristics;'):
|
|
317 parts = l.strip().split(';')
|
|
318 assert len(parts) == 8
|
|
319 name = parts[3]
|
|
320 prediction = float(parts[6])
|
|
321 count = int(parts[4])
|
|
322 hits = int(parts[5])
|
111
|
323
|
|
324 profile.add(name, prediction, count, hits)
|
|
325 elif l.startswith(loop_niter_str):
|
|
326 v = int(l[len(loop_niter_str):])
|
|
327 profile.add_loop_niter(v)
|
|
328
|
|
329 profile.dump(args.sorting)
|