2 # Script to analyze experimental results of our branch prediction heuristics
3 # Contributed by Jan Hubicka, SuSE Inc.
4 # Copyright (C) 2001, 2003 Free Software Foundation, Inc.
6 # This file is part of GCC.
8 # GCC is free software; you can redistribute it and/or modify
9 # it under the terms of the GNU General Public License as published by
10 # the Free Software Foundation; either version 2, or (at your option)
13 # GCC is distributed in the hope that it will be useful,
14 # but WITHOUT ANY WARRANTY; without even the implied warranty of
15 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 # GNU General Public License for more details.
18 # You should have received a copy of the GNU General Public License
19 # along with GCC; see the file COPYING. If not, write to
20 # the Free Software Foundation, 59 Temple Place - Suite 330,
21 # Boston, MA 02111-1307, USA.
24 # This script is used to calculate two basic properties of the branch prediction
25 # heuristics - coverage and hitrate. Coverage is number of executions of a given
26 # branch matched by the heuristics and hitrate is probability that once branch is
27 # predicted as taken it is really taken.
29 # These values are useful to determine the quality of given heuristics. Hitrate
30 # may be directly used in predict.c.
33 # Step 1: Compile and profile your program. You need to use -fprofile-arcs
34 # flag to get the profiles
35 # Step 2: Generate log files. The information about given heuristics are
36 # saved into *.bp dumps. You need to pass the -db switch to the compiler as well
37 # as -fbranch-probabilities to get the results of profiling noted in the dumps.
38 # Ensure that there are no "Arc profiling: some edge counts were bad." warnings.
39 # Step 3: Run this script to concatenate all *.life files:
40 # analyze_brprob `find . -name *.life`
41 # the information is collected and print once all files are parsed. This
43 # Note that the script does use bc to perform long arithmetic.
44 # Step 4: Read the results. Basically the following table is printed:
45 # (this is just an example from a very early stage of branch prediction pass
46 # development, so please don't take these numbers seriously)
48 #HEURISTICS BRANCHES (REL) HITRATE COVERAGE (REL)
49 #opcode 2889 83.7% 94.96%/ 97.62% 7516383 75.3%
50 #pointer 246 7.1% 99.69%/ 99.86% 118791 1.2%
51 #loop header 449 13.0% 98.32%/ 99.07% 43553 0.4%
52 #first match 3450 100.0% 89.92%/ 97.27% 9979782 100.0%
53 #loop exit 924 26.8% 88.95%/ 95.58% 9026266 90.4%
54 #error return 150 4.3% 64.48%/ 86.81% 453542 4.5%
55 #call 803 23.3% 51.66%/ 98.61% 3614037 36.2%
56 #loop branch 51 1.5% 99.26%/ 99.27% 26854 0.3%
57 #noreturn call 951 27.6% 100.00%/100.00% 1759809 17.6%
59 # The heuristic called "first match" is a heuristic used by GCC branch
60 # prediction pass and it predicts 89.92% branches correctly.
62 # The quality of heuristics can be rated using both, coverage and hitrate
63 # parameters. For example "loop branch" heuristics (predicting loopback edge
64 # as taken) have both very high hitrate and coverage, so it is very useful.
65 # On the other hand, "exit block" heuristics (predicting exit edges as not
66 # taken) have good hitrate, but poor coverage, so only 3 branches have been
67 # predicted. The "loop header" heuristic has problems, since it tends to
70 # The implementation of this script is somewhat brute force. My awk skills
75 e =
"echo \"scale = 2 ;"e
"\" | bc"
83 /^ .
* heuristics
: .
*.$
/ {
86 sub (/ heuristics
: .
*.$
/,"",name
)
87 if (!
(name in branches
))
99 /^ .
* heuristics
: .
*. exec
[0-9]* hit
[0-9]* (.
*.
)$
/ {
102 sub (/ heuristics
: .
*. exec
[0-9]* hit
[0-9]* (.
*.
)$
/,"",name
)
104 sub (/^ .
* heuristics
: /,"",pred
)
105 sub (/. exec
[0-9]* hit
[0-9]* (.
*.
)$
/,"",pred
)
107 sub (/^ .
* heuristics
: .
*. exec
/,"",count
)
108 sub (/ hit
[0-9]* (.
*.
)$
/,"",count
)
110 sub (/^ .
* heuristics
: .
*. exec
[0-9]* hit
/,"",hit
)
111 sub (/ (.
*.
)$
/,"",hit
)
113 if (int
(pred
) < 50.0)
117 counts
[name
]=counts
[name
] "+" count
118 hits
[name
]=hits
[name
] "+" hit
119 phits
[name
]=phits
[name
] "+(("hit
")<"count
"/2)*("count
"-("hit
"))+(("hit
")>="count
"/2)*("hit
")"
121 #BC crashes on long strings. Irritating.
122 if (length(counts
[name
]) > 2000)
123 counts
[name
] = longeval
(counts
[name
])
124 if (length(hits
[name
]) > 2000)
125 hits
[name
] = longeval
(hits
[name
])
126 if (length(phits
[name
]) > 2000)
127 phits
[name
] = longeval
(phits
[name
])
130 # Heuristics called combined predicts just everything.
131 maxcounts = longeval
(counts
["combined"])
132 maxbranches = branches
["combined"]
133 max = names
["combined"]
134 printf("HEURISTICS BRANCHES (REL) HITRATE COVERAGE (REL)\n")
135 for (i =
0; i
< nnames
; i
++)
138 counts
[name
] = longeval
(counts
[name
])
139 printf ("%-26s %8i %5.1f%% %6s%% / %6s%% %12s %5.1f%%\n",
141 branches
[name
], branches
[name
] * 100 / maxbranches
,
142 longeval
("("hits
[name
]") * 100 /(" counts
[name
]"-0.00001)"),
143 longeval
("("phits
[name
]") * 100 /(" counts
[name
]"-0.00001)"),
144 counts
[name
], longeval
(counts
[name
]" * 100 / ("maxcounts
"-0.00001)"))