Next: , Previous: , Up: flex   [Contents][Index]


17 Performance Considerations

The main design goal of flex is that it generate high-performance scanners. It has been optimized for dealing well with large sets of rules. Aside from the effects on scanner speed of the table compression ‘-C’ options outlined above, there are a number of options/actions which degrade performance. These are, from most expensive to least:

    REJECT
    arbitrary trailing context

    pattern sets that require backing up
    %option yylineno
    %array

    %option interactive
    %option always-interactive

    ^ beginning-of-line operator
    yymore()

with the first two all being quite expensive and the last two being quite cheap. Note also that unput() is implemented as a routine call that potentially does quite a bit of work, while yyless() is a quite-cheap macro. So if you are just putting back some excess text you scanned, use yyless().

REJECT should be avoided at all costs when performance is important. It is a particularly expensive option.

There is one case when %option yylineno can be expensive. That is when your patterns match long tokens that could possibly contain a newline character. There is no performance penalty for rules that can not possibly match newlines, since flex does not need to check them for newlines. In general, you should avoid rules such as [^f]+, which match very long tokens, including newlines, and may possibly match your entire file! A better approach is to separate [^f]+ into two rules:

%option yylineno
%%
    [^f\n]+
    \n+

The above scanner does not incur a performance penalty.

Getting rid of backing up is messy and often may be an enormous amount of work for a complicated scanner. In principal, one begins by using the ‘-b’ flag to generate a lex.backup file. For example, on the input:

    %%
    foo        return TOK_KEYWORD;
    foobar     return TOK_KEYWORD;

the file looks like:

    State #6 is non-accepting -
     associated rule line numbers:
           2       3
     out-transitions: [ o ]
     jam-transitions: EOF [ \001-n  p-\177 ]

    State #8 is non-accepting -
     associated rule line numbers:
           3
     out-transitions: [ a ]
     jam-transitions: EOF [ \001-`  b-\177 ]

    State #9 is non-accepting -
     associated rule line numbers:
           3
     out-transitions: [ r ]
     jam-transitions: EOF [ \001-q  s-\177 ]

    Compressed tables always back up.

The first few lines tell us that there’s a scanner state in which it can make a transition on an ’o’ but not on any other character, and that in that state the currently scanned text does not match any rule. The state occurs when trying to match the rules found at lines 2 and 3 in the input file. If the scanner is in that state and then reads something other than an ’o’, it will have to back up to find a rule which is matched. With a bit of headscratching one can see that this must be the state it’s in when it has seen ‘fo’. When this has happened, if anything other than another ‘o’ is seen, the scanner will have to back up to simply match the ‘f’ (by the default rule).

The comment regarding State #8 indicates there’s a problem when ‘foob’ has been scanned. Indeed, on any character other than an ‘a’, the scanner will have to back up to accept "foo". Similarly, the comment for State #9 concerns when ‘fooba’ has been scanned and an ‘r’ does not follow.

The final comment reminds us that there’s no point going to all the trouble of removing backing up from the rules unless we’re using ‘-Cf’ or ‘-CF’, since there’s no performance gain doing so with compressed scanners.

The way to remove the backing up is to add “error” rules:

    %%
    foo         return TOK_KEYWORD;
    foobar      return TOK_KEYWORD;

    fooba       |
    foob        |
    fo          {
                /* false alarm, not really a keyword */
                return TOK_ID;
                }

Eliminating backing up among a list of keywords can also be done using a “catch-all” rule:

    %%
    foo         return TOK_KEYWORD;
    foobar      return TOK_KEYWORD;

    [a-z]+      return TOK_ID;

This is usually the best solution when appropriate.

Backing up messages tend to cascade. With a complicated set of rules it’s not uncommon to get hundreds of messages. If one can decipher them, though, it often only takes a dozen or so rules to eliminate the backing up (though it’s easy to make a mistake and have an error rule accidentally match a valid token. A possible future flex feature will be to automatically add rules to eliminate backing up).

It’s important to keep in mind that you gain the benefits of eliminating backing up only if you eliminate every instance of backing up. Leaving just one means you gain nothing.

Variable trailing context (where both the leading and trailing parts do not have a fixed length) entails almost the same performance loss as REJECT (i.e., substantial). So when possible a rule like:

    %%
    mouse|rat/(cat|dog)   run();

is better written:

    %%
    mouse/cat|dog         run();
    rat/cat|dog           run();

or as

    %%
    mouse|rat/cat         run();
    mouse|rat/dog         run();

Note that here the special ’|’ action does not provide any savings, and can even make things worse (see Limitations).

Another area where the user can increase a scanner’s performance (and one that’s easier to implement) arises from the fact that the longer the tokens matched, the faster the scanner will run. This is because with long tokens the processing of most input characters takes place in the (short) inner scanning loop, and does not often have to go through the additional work of setting up the scanning environment (e.g., yytext) for the action. Recall the scanner for C comments:

    %x comment
    %%
            int line_num = 1;

    "/*"         BEGIN(comment);

    <comment>[^*\n]*
    <comment>"*"+[^*/\n]*
    <comment>\n             ++line_num;
    <comment>"*"+"/"        BEGIN(INITIAL);

This could be sped up by writing it as:

    %x comment
    %%
            int line_num = 1;

    "/*"         BEGIN(comment);

    <comment>[^*\n]*
    <comment>[^*\n]*\n      ++line_num;
    <comment>"*"+[^*/\n]*
    <comment>"*"+[^*/\n]*\n ++line_num;
    <comment>"*"+"/"        BEGIN(INITIAL);

Now instead of each newline requiring the processing of another action, recognizing the newlines is distributed over the other rules to keep the matched text as long as possible. Note that adding rules does not slow down the scanner! The speed of the scanner is independent of the number of rules or (modulo the considerations given at the beginning of this section) how complicated the rules are with regard to operators such as ‘*’ and ‘|’.

A final example in speeding up a scanner: suppose you want to scan through a file containing identifiers and keywords, one per line and with no other extraneous characters, and recognize all the keywords. A natural first approach is:

    %%
    asm      |
    auto     |
    break    |
    ... etc ...
    volatile |
    while    /* it's a keyword */

    .|\n     /* it's not a keyword */

To eliminate the back-tracking, introduce a catch-all rule:

    %%
    asm      |
    auto     |
    break    |
    ... etc ...
    volatile |
    while    /* it's a keyword */

    [a-z]+   |
    .|\n     /* it's not a keyword */

Now, if it’s guaranteed that there’s exactly one word per line, then we can reduce the total number of matches by a half by merging in the recognition of newlines with that of the other tokens:

    %%
    asm\n    |
    auto\n   |
    break\n  |
    ... etc ...
    volatile\n |
    while\n  /* it's a keyword */

    [a-z]+\n |
    .|\n     /* it's not a keyword */

One has to be careful here, as we have now reintroduced backing up into the scanner. In particular, while we know that there will never be any characters in the input stream other than letters or newlines, flex can’t figure this out, and it will plan for possibly needing to back up when it has scanned a token like ‘auto’ and then the next character is something other than a newline or a letter. Previously it would then just match the ‘auto’ rule and be done, but now it has no ‘auto’ rule, only a ‘auto\n’ rule. To eliminate the possibility of backing up, we could either duplicate all rules but without final newlines, or, since we never expect to encounter such an input and therefore don’t how it’s classified, we can introduce one more catch-all rule, this one which doesn’t include a newline:

    %%
    asm\n    |
    auto\n   |
    break\n  |
    ... etc ...
    volatile\n |
    while\n  /* it's a keyword */

    [a-z]+\n |
    [a-z]+   |
    .|\n     /* it's not a keyword */

Compiled with ‘-Cf’, this is about as fast as one can get a flex scanner to go for this particular problem.

A final note: flex is slow when matching NULs, particularly when a token contains multiple NULs. It’s best to write rules which match short amounts of text if it’s anticipated that the text will often include NULs.

Another final note regarding performance: as mentioned in How the Input Is Matched, dynamically resizing yytext to accommodate huge tokens is a slow process because it presently requires that the (huge) token be rescanned from the beginning. Thus if performance is vital, you should attempt to match “large” quantities of text but not “huge” quantities, where the cutoff between the two is at about 8K characters per token.


Next: Generating C++ Scanners, Previous: Scanner Options, Up: flex   [Contents][Index]