This web page suggests that if your lex program “has a large number of reserved words, it is more efficient to let lex simply match a string and determine in your own code whether it is a variable or reserved word.”
My question is: More efficient where, and why? If it means compiling the lexer is faster, I don’t really care about that because it is one step removed from the program which uses the lexer to parse input.
It seems to be that lex just uses your description to build a state machine that processes one character at a time. It does not seem logical that increasing the size of the state machine would necessarily cause it to become any slower than using one rule for identifiers and then doing several string comparisons.
Additionally, if it turns out that there is some logical reason for this to make sense as an optimization, what would be considered a large number of reserved words? I have approximately 20, as compared to about 30 other rules for various things. Would that be considered a large number of reserved words? Should I attempt to use the same strategy for some of the other symbols?
I have attempted to google for a result, but the only relevant articles I found stated this strategy as though it were well-known without giving any reason.
In case it is relevant, I am using flex 2.5.35.
Edit: Here is another reference which claims that lex produces an inefficient scanner when asked to match several long literal strings. It also does not give a reason.
According to the flex manual, “[t]he speed of the scanner is independent of the number of rules or … how complicated the rules are with regard to operators such as ‘*’ and ‘|’.”
The main performance losses are due to backtracking. This can be avoided by (among other things) using catch-all rules which will match tokens which “start with” the offending token. For example, if you have a list of reserved words consisting of [a-zA-Z_], and then a rule for matching identifiers of the form [a-zA-Z_][a-zA-Z_0-9]*, the rule for matching identifiers will catch any identifiers which start with the name of a reserved word without having to back up and try to match again.
According to the faq, flex generates a deterministic finite automaton which “does all the matching simultaneously, in parallel.” The result of this is, as was said above, that the speed of the scanner is independent of the number of rules. On the other hand, string comparison is linear in the number of rules.
As a result, the reserved word rules should, in fact, be considerably faster than a lookup table.