When you call pcre[16|32]_exec(), it makes use of an internal
function called match(). This calls itself recursively at branch
points in the pattern, in order to remember the state of the
match so that it can back up and try a different alternative if
the first one fails. As matching proceeds deeper and deeper into
the tree of possibilities, the recursion depth increases. The
match() function is also called in other circumstances, for
example, whenever a parenthesized sub-pattern is entered, and in
certain cases of repetition.
Not all calls of match() increase the recursion depth; for an
item such as a* it may be called several times at the same level,
after matching different numbers of a's. Furthermore, in a number
of cases where the result of the recursive call would immediately
be passed back as the result of the current call (a "tail
recursion"), the function is just restarted instead.
The above comments apply when pcre[16|32]_exec() is run in its
normal interpretive manner. If the pattern was studied with the
PCRE_STUDY_JIT_COMPILE option, and just-in-time compiling was
successful, and the options passed to pcre[16|32]_exec() were not
incompatible, the matching process uses the JIT-compiled code
instead of the match() function. In this case, the memory
requirements are handled entirely differently. See the pcrejit
documentation for details.
The pcre[16|32]_dfa_exec() function operates in an entirely
different way, and uses recursion only when there is a regular
expression recursion or subroutine call in the pattern. This
includes the processing of assertion and "once-only" subpatterns,
which are handled like subroutine calls. Normally, these are
never very deep, and the limit on the complexity of
pcre[16|32]_dfa_exec() is controlled by the amount of workspace
it is given. However, it is possible to write patterns with
runaway infinite recursions; such patterns will cause
pcre[16|32]_dfa_exec() to run out of stack. At present, there is
no protection against this.
The comments that follow do NOT apply to pcre[16|32]_dfa_exec();
they are relevant only for pcre[16|32]_exec() without the JIT
optimization.
Reducing pcre[16|32]_exec()'s stack usage
Each time that match() is actually called recursively, it uses
memory from the process stack. For certain kinds of pattern and
data, very large amounts of stack may be needed, despite the
recognition of "tail recursion". You can often reduce the amount
of recursion, and therefore the amount of stack used, by
modifying the pattern that is being matched. Consider, for
example, this pattern:
([^<]|<(?!inet))+
It matches from wherever it starts until it encounters "<inet" or
the end of the data, and is the kind of pattern that might be
used when processing an XML file. Each iteration of the outer
parentheses matches either one character that is not "<" or a "<"
that is not followed by "inet". However, each time a parenthesis
is processed, a recursion occurs, so this formulation uses a
stack frame for each matched character. For a long string, a lot
of stack is required. Consider now this rewritten pattern, which
matches exactly the same strings:
([^<]++|<(?!inet))+
This uses very much less stack, because runs of characters that
do not contain "<" are "swallowed" in one item inside the
parentheses. Recursion happens only when a "<" character that is
not followed by "inet" is encountered (and we assume this is
relatively rare). A possessive quantifier is used to stop any
backtracking into the runs of non-"<" characters, but that is not
related to stack usage.
This example shows that one way of avoiding stack problems when
matching long subject strings is to write repeated parenthesized
subpatterns to match more than one character whenever possible.
Compiling PCRE to use heap instead of stack for pcre[16|32]_exec()
In environments where stack memory is constrained, you might want
to compile PCRE to use heap memory instead of stack for
remembering back-up points when pcre[16|32]_exec() is running.
This makes it run a lot more slowly, however. Details of how to
do this are given in the pcrebuild documentation. When built in
this way, instead of using the stack, PCRE obtains and frees
memory by calling the functions that are pointed to by the
pcre[16|32]_stack_malloc and pcre[16|32]_stack_free variables. By
default, these point to malloc() and free(), but you can replace
the pointers to cause PCRE to use your own functions. Since the
block sizes are always the same, and are always freed in reverse
order, it may be possible to implement customized memory handlers
that are more efficient than the standard functions.
Limiting pcre[16|32]_exec()'s stack usage
You can set limits on the number of times that match() is called,
both in total and recursively. If a limit is exceeded,
pcre[16|32]_exec() returns an error code. Setting suitable limits
should prevent it from running out of stack. The default values
of the limits are very large, and unlikely ever to operate. They
can be changed when PCRE is built, and they can also be set when
pcre[16|32]_exec() is called. For details of these interfaces,
see the pcrebuild documentation and the section on extra data for
pcre[16|32]_exec() in the pcreapi documentation.
As a very rough rule of thumb, you should reckon on about 500
bytes per recursion. Thus, if you want to limit your stack usage
to 8Mb, you should set the limit at 16000 recursions. A 64Mb
stack, on the other hand, can support around 128000 recursions.
In Unix-like environments, the pcretest test program has a
command line option (-S) that can be used to increase the size of
its stack. As long as the stack is large enough, another option
(-M) can be used to find the smallest limits that allow a
particular pattern to match a given subject string. This is done
by calling pcre[16|32]_exec() repeatedly with different limits.
Obtaining an estimate of stack usage
The actual amount of stack used per recursion can vary quite a
lot, depending on the compiler that was used to build PCRE and
the optimization or debugging options that were set for it. The
rule of thumb value of 500 bytes mentioned above may be larger or
smaller than what is actually needed. A better approximation can
be obtained by running this command:
pcretest -m -C
The -C option causes pcretest to output information about the
options with which PCRE was compiled. When -m is also given
(before -C), information about stack use is given in a line like
this:
Match recursion uses stack: approximate frame size = 640 bytes
The value is approximate because some recursions need a bit more
(up to perhaps 16 more bytes).
If the above command is given when PCRE is compiled to use the
heap instead of the stack for recursion, the value that is output
is the size of each block that is obtained from the heap.
Changing stack size in Unix-like systems
In Unix-like environments, there is not often a problem with the
stack unless very long strings are involved, though the default
limit on stack size varies from system to system. Values from 8Mb
to 64Mb are common. You can find your default limit by running
the command:
ulimit -s
Unfortunately, the effect of running out of stack is often
SIGSEGV, though sometimes a more explicit error message is given.
You can normally increase the limit on stack size by code such as
this:
struct rlimit rlim;
getrlimit(RLIMIT_STACK, &rlim);
rlim.rlim_cur = 100*1024*1024;
setrlimit(RLIMIT_STACK, &rlim);
This reads the current limits (soft and hard) using getrlimit(),
then attempts to increase the soft limit to 100Mb using
setrlimit(). You must do this before calling pcre[16|32]_exec().
Changing stack size in Mac OS X
Using setrlimit(), as described above, should also work on Mac OS
X. It is also possible to set a stack size when linking a
program. There is a discussion about stack sizes in Mac OS X at
this web site: http://developer.apple.com/qa/qa2005/qa1419.html.
