usr/src/usr.bin/gprof/PSD.doc/equations.me

\"      @(#)equations.me        1.1 %G%
.EQ
delim ##
.EN
.sh 1 "Notation"
.pp
Given a program, #P#, and an ordered list of instructions, #I#,
composing #P#,
.EQ
P ~ = ~ {I sub 1} ^ {I sub 2} ^ ... ^ {I sub m}
.EN
define the address of an instruction to be its index in the program.
Given a namelist, #N#,  that associates names with distinguished
addresses, called entry points, sorted by increasing entry point values,
.EQ
N ~ = ~ "{" ( n sub 1 , ~ e sub 1 ) ,..., ( n sub r , ~ e sub r ) "}"
.EN
#N# partitions the addresses of #P# into ranges, called routines,
such that
.EQ
R sub 0 ~ mark = ~ "{" 1 ,..., {e sub 1} - 1 "}"
.EN C
.EQ
R sub i ~ lineup = ~ "{" e sub i ,..., {e sub i} - 1 "}" ~
for ~ 1 <= i < r
.EN C
.EQ
R sub r ~ lineup = ~ "{" e sub r ,..., m "}"
.EN
.(q
<<<watch out.  we have only one entry point per routine.  not
everybody does.>>>
.)q
A profilable program consists of the instructions and a namelist.
.pp
Given a histogram of address samples, #H#, ordered such that the
#i sup th# sample is the count for the #i sup th#
instruction, define the #selftime#, #S#, of each of the #r# routines
as
.EQ
S sub i ~ = ~ sum from {j ~ \(mo ~ {R sub i}} {H sub j}
.EN
.pp
Given a set of arcs, #A#, gathered during the profiled execution of
the program, where #A# consists of ordered triples,
#(tail, ~ head, ~ count)#, giving the address of the
tail and head of the arc, and a count of the number of times the
execution of the program traversed this arc from instruction
#I sub tail# to instruction #I sub head#,
the number of calls to the #i sup th# routine can be calculated
as
.EQ
C sub i ~ = ~ sum from { ( tail , ~ head , ~ count ) ~ \(mo ~ A }
                       "{" count | head ~ \(mo ~ R sub i "}"
.EN
and the number of calls to the #i sup th# routine from the
#j sup th# routine as
.EQ
C sub i sup j ~ = ~ sum from { ( tail , ~ head , ~ count ) ~ \(mo ~ A }
                                  "{" count | tail ~ \(mo ~ R sub i ,
                                  head ~ \(mo ~ R sub j "}"
.EN
.pp
If there is an arc with #tail ~ \(mo ~ R sub i# and
#head ~ \(mo ~ R sub j#, then #R sub j# is a child of
#R sub i#, denoted #R sub i ~ Child ~ R sub j#.
#R sub i# has as a child #R sub j#.
The transitive closure of the #Child# relation is called the
#Descendants# relation.
The inverse of the #Child# relation is the #Parent# relation.
For example, if #R sub i ~ Child ~ R sub j# then
#R sub j ~ Parent ~ R sub i#.
The transitive closure of the #Parent# relation is called the
#Ancestor# relation.
.pp
We are interested in computing the following recurrence, which
gives the time for a routine based on its selftime and an
appropriate fraction of the time of its descendants.
.EQ
T sub i ~ = ~ S sub i + sum from {i ~ Child ~ j}
                        {T sub j \(mu {C sub i sup j} over {C sub i}}
.EN
This recurrence can be trivially solved for leaves of the call
graph, e.g. those with no children.
In the general case, however, it is necessary to order the nodes
of the graph so that the times for all children of a routine
are known before computing the time for the routine itself.
Such an order could be imposed by a depth-first numbering of the
routines with a single traversal of each of the arcs in #A#.
<<<this falls off here because i don't know if we want to continue
with this, and it's hard to type.>>>