The header has the following form:
<header> : <line>+ <terminator>
<terminator> : "samples" <space>* "\n"
<line> : <required> "\n"
| <optional> "\n"
| <unknown> "\n"
<required> : "image" <space>+ <hex_digit>+
| "epoch" <space>+ <time>
| "platform" <space>+ <rest_of_line>
| "event" <space>+ <rest_of_line>
| "period" <space>+ <digit>+
| "tsize" <space>+ <digit>+
| "cpuspeed" <space>+ <digit>+
<optional> : "cpuamask" <space>+ <hex_digit>+
| "cpuimplv" <space>+ <digit>+
| "cpucount" <space>+ <digit>+
| "period" <space>+ <digit>+
| "path" <space>+ <rest_of_line>
<unknown> : <word> <space>+ <rest_of_line>
<major> : <digit>+
<minor> : <digit>+
<time> : YYMMDDHHMM /* Digits indicating UCT time */
| YYYYMMDDHHMMSS /* (Not yet supported) */
<space> : [ \t]
<hex_digit> : <digit> | [a-f] | [A-F]
<digit> : [0-9]
<word> : [^ \t\n]+
<rest_of_line> : [^ \t\n][^\n]*
For each kind of line that falls under <required>, the
profile header should contain exactly one occurrence of that line. To support
future extensions to the file format, all programs that modify profile files
should preserve all <unknown> header lines when writing out the
modified profile.
The format of the binary data is dependent on the major version number
present in the ASCII header. The current version number used by the profiling
tools is 1.01. The documentation below describes the binary data
format for the obsolete versions 0.06 and 0.07. We plan to
document major-version 1 in the near future.
The binary section of the profile begins immediately after the newline
character that terminates the ASCII header. (No whitespace is allowed after
the newline.) This section contains a mapping from each instruction address in
the image, to a count of the number of samples associated with that
address.
The binary section consists of a list of chunks, followed by a fixed length
footer.
<samples> : <chunk>* <footer>
Each chunk contains the sample counts for a contiguous range of instruction
addresses. For each address within this range, the chunk stores the number of
samples associated with this address.
<chunk> : <offset> <number> <count>+
<offset> : <uint32>
<number> : <uint32>
<count> : <uint32>
The beginning of the address range is text_start(image) +
<offset>, where text_start(image) is recorded in the
tstart field in the profile header. The length of the address range
is <number>. The number of <count> entries in
the chunk is <number>. The number of samples associated with
the ith address in the address range is given by the ith
<count> entry in the chunk.
The instruction-address ranges associated with two distinct chunks in one
profile cannot overlap. Furthermore, if two chunks a and b
occur in the same profile and a occurs before b, then the
offset associated with a has to be smaller than the offset associated
with b.
The footer consists of two 4-byte values. The first value is the total
number of instruction addresses in the image that have at least one associated
sample. The second value is the sum of the sample counts associated with all
of the instruction addresses in the image.
<footer> : <total_offsets> <total_samples>
<total_offsets> : <uint32>
<total_samples> : <uint32>
All values in the binary section are stored in little-endian format.
<uint32> /* Unsigned little-endian 32 bit value */
For good performance, the length of the ASCII header should be made a
multiple of four (by adding extra whitespace at the end of the last line of
the ASCII header) so that the binary section is aligned to a 32-bit boundary
relative to the beginning of the profile.
