What are the differences between PEGs and CFGs? - regex

From this wikipedia page:
The fundamental difference between
context-free grammars and parsing
expression grammars is that the PEG's
choice operator is ordered. If the
first alternative succeeds, the second
alternative is ignored. Thus ordered
choice is not commutative, unlike
unordered choice as in context-free
grammars and regular expressions.
Ordered choice is analogous to soft
cut operators available in some logic
programming languages.
Why does PEG's choice operator short circuits the matching? Is it because to minimize memory usage (due to memoization)?
I'm not sure what the choice operator is in regular expressions but let's suppose it is this: /[aeiou]/ to match a vowel. So this regex is commutative because I could have written it in any of the 5! (five factorial) permutations of the vowel characters? i.e. /[aeiou]/ behaves the same as /[eiaou]/. What is the advantage of it being commutative? (c.f. PEG's non-commutativity)
The consequence is that if a CFG is
transliterated directly to a PEG, any
ambiguity in the former is resolved by
deterministically picking one parse
tree from the possible parses. By
carefully choosing the order in which
the grammar alternatives are
specified, a programmer has a great
deal of control over which parse tree
is selected.
Is this saying that PEG's grammar is superior to CFG's?

A CFG grammar is non-deterministic, meaning that some input could result in two or more possible parse-trees. Though most CFG-based parser-generators have restrictions on the determinability of the grammar. It will give a warning or error if it has two or more choices.
A PEG grammar is deterministic, meaning that any input can only be parsed one way.
To take a classic example; The grammar
if_statement := "if" "(" expr ")" statement "else" statement
| "if" "(" expr ")" statement;
applied to the input
if (x1) if (x2) y1 else y2
could either be parsed as
if_statement(x1, if_statement(x2, y1, y2))
or
if_statement(x1, if_statement(x2, y1), y2)
A CFG-parser would generate a Shift/Reduce-conflict, since it can't decide if it should shift (read another token), or reduce (complete the node), when reaching the "else" keyword. Of course, there are ways to get around this problem.
A PEG-parser would always pick the first choice.
Which one is better is for you to decide. My opinion is that often PEG-grammars is easier to write, and CFG grammars easier to analyze.

I think you're confusing CFG with LR and with ambiguity. Grammars are not deterministic/nondeterministic, though their parsers may be. An ambiguous grammar is still CFG if it complies with the definition, and a deterministic parser can be built for it doing what PEG does.

PEGs and CFGs are two different ways of specifying a language. If you write a parser by hand, chances are very good that you will write a so-called recursive descent parser. A recursive descent parser will automatically resolve any ambiguities in your grammar, but does so silently and likely not in the way you would have wanted. The problem with this is that you never find out that there were ambiguities that got automatically resolved, unless you thoroughly test your parser. PEGs are basically a formalization of recursive descent parsers, and so come with this problem. For examples of this problem see How does backtracking affect the language recognized by a parser?, and https://cs.stackexchange.com/questions/143480/dragon-book-4-4-5-exercise/143975.
CFGs have a lot of theory to back them up, but PEGs not so much. The set of languages that can be encoded by CFG and those that can be encoded by PEG partially overlap, but neither encompasses the other.
For a more thorough review of this I recommend the excellent essay Which Parsing Approach?

Related

Rules & Actions for Parser Generator, and

I am trying to wrap my head around an assignment question, therefore I would very highly appreciate any help in the right direction (and not necessarily a complete answer). I am being asked to write the grammar specification for this parser. The specification for the grammar that I must implement can be found here:
http://anoopsarkar.github.io/compilers-class/decafspec.html
Although the documentation is there, I do not understand a few things, such as how to write (in my .y file) things such as
{ identifier },+
I understand that this would mean a comma-separated list of 1 (or more) occurrences of an identifier, however when I write it as such, the compiler displays an error of unrecognized symbols '+' and ',', being mistaken as whitespace. I tried '{' identifier "},+", but I haven't the slightest clue whether that is correct or not.
I have written the lexical analyzer portion (as it was from the previous segment of the assignment) which returns tokens (T_ID, T_PLUS, etc.) accordingly, however there is this new notion that I must assign 'yylval' to be the value of the token itself. To my understanding, this is only necessary if I am in need of the actual value of the token, therefore I would need the value of an identifier token T_ID, but not necessarily the value of T_PLUS, being '+'. This is done by creating a %union in the parser generator file, which I have done, and have provided the tokens that I currently believe would require the literal token value with the proper yylval assignment.
Here is my lexical analysis code (I could not get it to format properly, I apologize): https://pastebin.com/XMZwvWCK
Here is my parser file decafast.y: https://pastebin.com/2jvaBFQh
And here is the final piece of code supplied to me, the C++ code to build an abstract syntax tree at the end:
https://pastebin.com/ELy53VrW?fbclid=IwAR2cFT_-pGKlVZ2liC-zAe3Fw0BWDlGjrrayqEGV4JuJq1_7nKoe9-TLTlA
To finalize my question, I do not know if I am creating my grammar rules correctly. I have tried my best to follow the specification in the above website, but I can't help but feel that what I am writing is completely wrong. My compiler is spitting out nothing but "warning: rule useless in grammar" for almost every (if not every) rule.
If anyone could help me out and point me in the right direction on how to make any progress, I would highly, highly appreciate it.
The decaf specification is written in (an) Extended Backus Naur Form (EBNF), which includes a number of convenience operators for repetition, optionality and grouping. These are not part of the bison/yacc syntax, which is pretty well limited to BNF. (Bison/yacc do allow the alternation operator |, but since there is no way to group subpatterns, alteration can only be used at the top-level, to combine two productions for the same non-terminal.)
The short section at the beginning of the specification which describes EBNF includes a grammar for the particular variety of EBNF that is being used. (Since this grammar is itself recursively written in the same EBNF, there is a need to apply a bit of inductive reasoning.) When it says, for example,
CommaList = "{" Expression "}+," .
it is not saying that "}+," is the peculiar spelling of a comma-repetition operator. What it is saying is that when you see something in the Decaf grammar surrounded by { and }+,, that should be interpreted as describing a comma-separated list.
For example, the Decaf grammar includes:
FieldDecl = var { identifier }+, Type ";" .
That means that a FieldDecl can be (amongst other possibilities) the token var followed by a comma-separated list of identifier tokens and then followed by a Type and finally a semicolon.
As I said, bison/yacc don't implement the EBNF operators, so you have to find an equivalent yourself. Since BNF doesn't allow any form of grouping -- and a list is a grouped subexpression -- we need to rewrite the subexpression of a production as a new non-terminal. Also, I suppose we need to use the tokens defined in spec (although bison allows a more readable syntax).
So to yacc-ify this EBNF production, we first introducing the new non-terminal and replace the token names:
FieldDecl: T_VAR IdentifierList Type T_SEMICOLON
Which leaves the definition of IdentifierList. Repetition in BNF is always produced with recursion, following a very simple model which uses two productions:
the base, which is the simplest possible repetition (usually either nothing or a single list item), and
the recursion, which describes a longer possibility by extending a shorter one.
In this case, the list must have at least one item, and we extend by adding a comma and another item:
IdentifierList
: T_ID /* base case */
| IdentifierList T_COMMA T_ID /* Recursive extension */
The point of this exercise is to develop your skills in thinking grammatically: that is, factoring out the syntax and semantics of the language. So you should try to understand the grammars presented, both for Decaf and for the author's version of EBNF, and avoid blindly copying code (including grammars). Good luck!

Counterpart of regular expressions for parsing nested strucures

Regular expressions are a standard tool used for parsing strings across many languages. However their scope of applicability is limited. Regular expressions can only match a list. There is no way to describe arbitrary deep nested structures using regular expressions. Question: what is a technology/framework as widely used/spread and as standatd as regular expessions are that can match tree structures (produce AST).
Regular expressions describe a finite-state automaton.
Since the late 1960's, the "bread and butter" of parsing (though not necessarily the "state of the art") has been push-down automata generated by parser generators according to "LR" algorithms like LALR(1).
The connection to regular expressions is this: the parsing machine does in fact use rules very similar to regular expressions in order to recognize viable prefixes. The "shift" state transitions among the "core LR(0) items" constitute a finite automaton, and can be described by a regular expression. The recursion is is handled thanks to the semantic action of pushing symbols onto a stack when doing the "shifts", and removing them ("reducing"). Reductions rewrite a portion of the stack, and perform a "goto" to another state. This type of goto, together with the stack, is absent in the regular expression automaton.
Parse Expression Grammars are also related to regular expressions. Regular expressions themselves can be endowed with recursion. Firstly, we can take pieces of regular expressions and give them names, and then construct bigger regular expressions by writing expressions which invoke these names. (Such as feature is found in the lex tool where you can define a named expressions like letters [A-Za-z]+ and refer to it later as {letters}. Now suppose you allow circular references, like letters [A-Za-z]{letters}?. You now have recursion; the only problem is to adjust the model in order to implement it.
Implementations of so-called "regular expressions" in various modern languages and environments in fact support recursion. Perl-compatible regular expressions (PCRE) support it, for instance.
Expressions that feature recursion or backreferencing are not handled by the classic NFA compilation route (possibly converted to a DFA); they cannot be.
How the above letters recursion can be handled is with actual recursion. The ? operator can be implemented as a function which tries to match its respective argument object. If it succeeds, then it consumes whatever it has matched, otherwise it consumes nothing. That is to say, the regular expression can be converted to a syntax tree, and interpreted "as is" rather than compiled to a state machine (or trivially compiled to functions corresponding to the nodes of the tree), and such interpretation can naturally handle recursion. The interpretation then constitutes, effectively, a syntax-directed recursive-descent parser. (Note how I avoided left recursion in defining letters to make that example compatible with this approach).
Example: parenthesis-matching regex:
par-match := ({par-match})|
This gets compiled to a tree:
branch-op <-- "par-match" name points at this node
/ \
catenate-op <empty>
/ \
"(" catenate-op
/ \
{par-match} ")"
This can then converted to a recursive descent parser, or interpreted directly.
Pattern matching starts by invoking the top-level "branch-op". This operator simply tries all of the alternatives. Suppose the input is empty. Then the left alternative will fail: it demands an open parenthesis. So then the right alternative will succeed: empty matches empty. (The operators either "fail" or indicate "success" and consume input.)
But suppose your input is (()). The left catenate-op will in turn invoke its left subtree, which matches and consumes the left parenthesis, leaving ()). It will then invoke its right subtree, another catenate-op. This catenate-op matches its left subtree, which triggers recursion into the top level via the named par-match references. That recursion will match and consume (), leaving ). The catenate-op then invokes its right subtree which matches ). Control returns up to branch-op. (Though the left side of branch-op matched something, branch-op must still try the other alternative; more than one branch can match, and some can match longer than others.)
This is closely related to Parsing Expression Grammars work.
Practically speaking, the recursive definition could be encoded into the regex syntax somehow. Say we invent some new operator like (?name:definition) which means "match definition which is allowed to contain invocations of itself via name. The invocation syntax could be (*name), so that we can write the par-match example as (?par-match:\((*par-match)\)|). The combinations (? and (* are invalid under "classic" regex syntax and so we can use them for extension.
As a final note, regexes correspond to grammars. That is the fundamental connection btween regexes and parsing. That is to say, regexes correspond to a particular subset of grammars describe only regular languages. An example of a grammar which describes a regular language:
S -> A | B
B -> b
A -> A a | c
Although there is A -> A ... recursion, this is still regular, and corresponds to the regex ac*|b, which is just a more compact way to denote the same language. The grammar lets us notate languages that aren't regular and for which we can't write a regex, but as we have seen, we can extend the regex notation and semantics to express some of these things. Regular expressions aren't separate from grammars. The two aren't counterparts, but rather one is a special case or subset of the other.
Parser generators like Yacc, Bison, and derivatives are what you're after. They aren't as widespread as regular expressions because they generate actual C code. There are translations like Jison for example which implement the Yacc/Bison syntax using javascript. I know there are similar tools for other languages.
I get the impression Parsing expression grammar systems are up and coming though.

Is D's grammar really context-free?

I've posted this on the D newsgroup some months ago, but for some reason, the answer never really convinced me, so I thought I'd ask it here.
The grammar of D is apparently context-free.
The grammar of C++, however, isn't (even without macros). (Please read this carefully!)
Now granted, I know nothing (officially) about compilers, lexers, and parsers. All I know is from what I've learned on the web.
And here is what (I believe) I have understood regarding context, in not-so-technical lingo:
The grammar of a language is context-free if and only if you can always understand the meaning (though not necessarily the exact behavior) of a given piece of its code without needing to "look" anywhere else.
Or, in even less rigor:
The grammar cannot be context-free if I need I can't tell the type of an expression just by looking at it.
So, for example, C++ fails the context-free test because the meaning of confusing<sizeof(x)>::q < 3 > (2) depends on the value of q.
So far, so good.
Now my question is: Can the same thing be said of D?
In D, hashtables can be created through a Value[Key] declaration, for example
int[string] peoplesAges; // Maps names to ages
Static arrays can be defined in a similar syntax:
int[3] ages; // Array of 3 elements
And templates can be used to make them confusing:
template Test1(T...)
{
alias int[T[0]] Test;
}
template Test2(U...)
{
alias int[U] Test2; // LGTM
}
Test1!(5) foo;
Test1!(int) bar;
Test2!(int) baz; // Guess what? It's invalid code.
This means that I cannot tell the meaning of T[0] or U just by looking at it (i.e. it could be a number, it could be a data type, or it could be a tuple of God-knows-what). I can't even tell if the expression is grammatically valid (since int[U] certainly isn't -- you can't have a hashtable with tuples as keys or values).
Any parsing tree that I attempt to make for Test would fail to make any sense (since it would need to know whether the node contains a data type versus a literal or an identifier) unless it delays the result until the value of T is known (making it context-dependent).
Given this, is D actually context-free, or am I misunderstanding the concept?
Why/why not?
Update:
I just thought I'd comment: It's really interesting to see the answers, since:
Some answers claim that C++ and D can't be context-free
Some answers claim that C++ and D are both context-free
Some answers support the claim that C++ is context-sensitive while D isn't
No one has yet claimed that C++ is context-free while D is context-sensitive :-)
I can't tell if I'm learning or getting more confused, but either way, I'm kind of glad I asked this... thanks for taking the time to answer, everyone!
Being context free is first a property of generative grammars. It means that what a non-terminal can generate will not depend on the context in which the non-terminal appears (in non context-free generative grammar, the very notion of "string generated by a given non-terminal" is in general difficult to define). This doesn't prevent the same string of symbols to be generated by two non-terminals (so for the same strings of symbols to appear in two different contexts with a different meaning) and has nothing to do with type checking.
It is common to extend the context-free definition from grammars to language by stating that a language is context-free if there is at least one context free grammar describing it.
In practice, no programming language is context-free because things like "a variable must be declared before it is used" can't be checked by a context-free grammar (they can be checked by some other kinds of grammars). This isn't bad, in practice the rules to be checked are divided in two: those you want to check with the grammar and those you check in a semantic pass (and this division also allows for better error reporting and recovery, so you sometimes want to accept more in the grammar than what would be possible in order to give your users better diagnostics).
What people mean by stating that C++ isn't context-free is that doing this division isn't possible in a convenient way (with convenient including as criteria "follows nearly the official language description" and "my parser generator tool support that kind of division"; allowing the grammar to be ambiguous and the ambiguity to be resolved by the semantic check is an relatively easy way to do the cut for C++ and follow quite will the C++ standard, but it is inconvenient when you are relying on tools which don't allow ambiguous grammars, when you have such tools, it is convenient).
I don't know enough about D to know if there is or not a convenient cut of the language rules in a context-free grammar with semantic checks, but what you show is far from proving the case there isn't.
The property of being context free is a very formal concept; you can find a definition here. Note that it applies to grammars: a language is said to be context free if there is at least one context free grammar that recognizes it. Note that there may be other grammars, possibly non context free, that recognize the same language.
Basically what it means is that the definition of a language element cannot change according to which elements surround it. By language elements I mean concepts like expressions and identifiers and not specific instances of these concepts inside programs, like a + b or count.
Let's try and build a concrete example. Consider this simple COBOL statement:
01 my-field PICTURE 9.9 VALUE 9.9.
Here I'm defining a field, i.e. a variable, which is dimensioned to hold one integral digit, the decimal point, and one decimal digit, with initial value 9.9 . A very incomplete grammar for this could be:
field-declaration ::= level-number identifier 'PICTURE' expression 'VALUE' expression '.'
expression ::= digit+ ( '.' digit+ )
Unfortunately the valid expressions that can follow PICTURE are not the same valid expressions that can follow VALUE. I could rewrite the second production in my grammar as follows:
'PICTURE' expression ::= digit+ ( '.' digit+ ) | 'A'+ | 'X'+
'VALUE' expression ::= digit+ ( '.' digit+ )
This would make my grammar context-sensitive, because expression would be a different thing according to whether it was found after 'PICTURE' or after 'VALUE'. However, as it has been pointed out, this doesn't say anything about the underlying language. A better alternative would be:
field-declaration ::= level-number identifier 'PICTURE' format 'VALUE' expression '.'
format ::= digit+ ( '.' digit+ ) | 'A'+ | 'X'+
expression ::= digit+ ( '.' digit+ )
which is context-free.
As you can see this is very different from your understanding. Consider:
a = b + c;
There is very little you can say about this statement without looking up the declarations of a,b and c, in any of the languages for which this is a valid statement, however this by itself doesn't imply that any of those languages is not context free. Probably what is confusing you is the fact that context freedom is different from ambiguity. This a simplified version of your C++ example:
a < b > (c)
This is ambiguous in that by looking at it alone you cannot tell whether this is a function template call or a boolean expression. The previous example on the other hand is not ambiguous; From the point of view of grammars it can only be interpreted as:
identifier assignment identifier binary-operator identifier semi-colon
In some cases you can resolve ambiguities by introducing context sensitivity at the grammar level. I don't think this is the case with the ambiguous example above: in this case you cannot eliminate the ambiguity without knowing whether a is a template or not. Note that when such information is not available, for instance when it depends on a specific template specialization, the language provides ways to resolve ambiguities: that is why you sometimes have to use typename to refer to certain types within templates or to use template when you call member function templates.
There are already a lot of good answers, but since you are uninformed about grammars, parsers and compilers etc, let me demonstrate this by an example.
First, the concept of grammars are quite intuitive. Imagine a set of rules:
S -> a T
T -> b G t
T -> Y d
b G -> a Y b
Y -> c
Y -> lambda (nothing)
And imagine you start with S. The capital letters are non-terminals and the small letters are terminals. This means that if you get a sentence of all terminals, you can say the grammar generated that sentence as a "word" in the language. Imagine such substitutions with the above grammar (The phrase between *phrase* is the one being replaced):
*S* -> a *T* -> a *b G* t -> a a *Y* b t -> a a b t
So, I could create aabt with this grammar.
Ok, back to main line.
Let us assume a simple language. You have numbers, two types (int and string) and variables. You can do multiplication on integers and addition on strings but not the other way around.
First thing you need, is a lexer. That is usually a regular grammar (or equal to it, a DFA, or equally a regular expression) that matches the program tokens. It is common to express them in regular expressions. In our example:
(I'm making these syntaxes up)
number: [1-9][0-9]* // One digit from 1 to 9, followed by any number
// of digits from 0-9
variable: [a-zA-Z_][a-zA-Z_0-9]* // You get the idea. First a-z or A-Z or _
// then as many a-z or A-Z or _ or 0-9
// this is similar to C
int: 'i' 'n' 't'
string: 's' 't' 'r' 'i' 'n' 'g'
equal: '='
plus: '+'
multiply: '*'
whitespace: (' ' or '\n' or '\t' or '\r')* // to ignore this type of token
So, now you got a regular grammar, tokenizing your input, but it understands nothing of the structure.
Then you need a parser. The parser, is usually a context free grammar. A context free grammar means, in the grammar you only have single nonterminals on the left side of grammar rules. In the example in the beginning of this answer, the rule
b G -> a Y b
makes the grammar context-sensitive because on the left you have b G and not just G. What does this mean?
Well, when you write a grammar, each of the nonterminals have a meaning. Let's write a context-free grammar for our example (| means or. As if writing many rules in the same line):
program -> statement program | lambda
statement -> declaration | executable
declaration -> int variable | string variable
executable -> variable equal expression
expression -> integer_type | string_type
integer_type -> variable multiply variable |
variable multiply number |
number multiply variable |
number multiply number
string_type -> variable plus variable
Now this grammar can accept this code:
x = 1*y
int x
string y
z = x+y
Grammatically, this code is correct. So, let's get back to what context-free means. As you can see in the example above, when you expand executable, you generate one statement of the form variable = operand operator operand without any consideration which part of code you are at. Whether the very beginning or middle, whether the variables are defined or not, or whether the types match, you don't know and you don't care.
Next, you need semantics. This is were context-sensitive grammars come into play. First, let me tell you that in reality, no one actually writes a context sensitive grammar (because parsing it is too difficult), but rather bit pieces of code that the parser calls when parsing the input (called action routines. Although this is not the only way). Formally, however, you can define all you need. For example, to make sure you define a variable before using it, instead of this
executable -> variable equal expression
you have to have something like:
declaration some_code executable -> declaration some_code variable equal expression
more complex though, to make sure the variable in declaration matches the one being calculated.
Anyway, I just wanted to give you the idea. So, all these things are context-sensitive:
Type checking
Number of arguments to function
default value to function
if member exists in obj in code: obj.member
Almost anything that's not like: missing ; or }
I hope you got an idea what are the differences (If you didn't, I'd be more than happy to explain).
So in summary:
Lexer uses a regular grammar to tokenize input
Parser uses a context-free grammar to make sure the program is in correct structure
Semantic analyzer uses a context-sensitive grammar to do type-checking, parameter matching etc etc
It is not necessarily always like that though. This just shows you how each level needs to get more powerful to be able to do more stuff. However, each of the mentioned compiler levels could in fact be more powerful.
For example, one language that I don't remember, used array subscription and function call both with parentheses and therefore it required the parser to go look up the type (context-sensitive related stuff) of the variable and determine which rule (function_call or array_substitution) to take.
If you design a language with lexer that has regular expressions that overlap, then you would need to also look up the context to determine which type of token you are matching.
To get to your question! With the example you mentioned, it is clear that the c++ grammar is not context-free. The language D, I have absolutely no idea, but you should be able to reason about it now. Think of it this way: In a context free grammar, a nonterminal can expand without taking into consideration anything, BUT the structure of the language. Similar to what you said, it expands, without "looking" anywhere else.
A familiar example would be natural languages. For example in English, you say:
sentence -> subject verb object clause
clause -> .... | lambda
Well, sentence and clause are nonterminals here. With this grammar you can create these sentences:
I go there because I want to
or
I jump you that I is air
As you can see, the second one has the correct structure, but is meaningless. As long as a context free grammar is concerned, the meaning doesn't matter. It just expands verb to whatever verb without "looking" at the rest of the sentence.
So if you think D has to at some point check how something was defined elsewhere, just to say the program is structurally correct, then its grammar is not context-free. If you isolate any part of the code and it still can say that it is structurally correct, then it is context-free.
There is a construct in D's lexer:
string ::= q" Delim1 Chars newline Delim2 "
where Delim1 and Delim2 are matching identifiers, and Chars does not contain newline Delim2.
This construct is context sensitive, therefore D's lexer grammar is context sensitive.
It's been a few years since I've worked with D's grammar much, so I can't remember all the trouble spots off the top of my head, or even if any of them make D's parser grammar context sensitive, but I believe they do not. From recall, I would say D's grammar is context free, not LL(k) for any k, and it has an obnoxious amount of ambiguity.
The grammar cannot be context-free if I need I can't tell the type of
an expression just by looking at it.
No, that's flat out wrong. The grammar cannot be context-free if you can't tell if it is an expression just by looking at it and the parser's current state (am I in a function, in a namespace, etc).
The type of an expression, however, is a semantic meaning, not syntactic, and the parser and the grammar do not give a penny about types or semantic validity or whether or not you can have tuples as values or keys in hashmaps, or if you defined that identifier before using it.
The grammar doesn't care what it means, or if that makes sense. It only cares about what it is.
To answer the question of if a programming language is context free you must first decide where to draw the line between syntax and semantics. As an extreme example, it is illegal in C for a program to use the value of some kinds of integers after they have been allowed to overflow. Clearly this can't be checked at compile time, let alone parse time:
void Fn() {
int i = INT_MAX;
FnThatMightNotReturn(); // halting problem?
i++;
if(Test(i)) printf("Weeee!\n");
}
As a less extreme example that others have pointed out, deceleration before use rules can't be enforced in a context free syntax so if you wish to keep your syntax pass context free, then that must be deferred to the next pass.
As a practical definition, I would start with the question of: Can you correctly and unambiguously determine the parse tree of all correct programs using a context free grammar and, for all incorrect programs (that the language requires be rejected), either reject them as syntactically invalid or produce a parse tree that the later passes can identify as invalid and reject?
Given that the most correct spec for the D syntax is a parser (IIRC an LL parser) I strongly suspect that it is in fact context free by the definition I suggested.
Note: the above says nothing about what grammar the language documentation or a given parser uses, only if a context free grammar exists. Also, the only full documentation on the D language is the source code of the compiler DMD.
These answers are making my head hurt.
First of all, the complications with low level languages and figuring out whether they are context-free or not, is that the language you write in is often processed in many steps.
In C++ (order may be off, but that shouldn't invalidate my point):
it has to process macros and other preprocessor stuffs
it has to interpret templates
it finally interprets your code.
Because the first step can change the context of the second step and the second step can change the context of the third step, the language YOU write in (including all of these steps) is context sensitive.
The reason people will try and defend a language (stating it is context-free) is, because the only exceptions that adds context are the traceable preprocessor statements and template calls. You only have to follow two restricted exceptions to the rules to pretend the language is context-free.
Most languages are context-sensitive overall, but most languages only have these minor exceptions to being context-free.

Finite State Machine parser

I would like to parse a self-designed file format with a FSM-like parser in C++ (this is a teach-myself-c++-the-hard-way-by-doing-something-big-and-difficult kind of project :)). I have a tokenized string with newlines signifying the end of a euh... line. See here for an input example. All the comments will and junk is filtered out, so I have a std::string like this:
global \n { \n SOURCE_DIRS src \n HEADER_DIRS include \n SOURCES bitwise.c framing.c \n HEADERS ogg/os_types.h ogg/ogg.h \n } \n ...
Syntax explanation:
{ } are scopes, and capitalized words signify that a list of options/files is to follow.
\n are only important in a list of options/files, signifying the end of the list.
So I thought that a FSM would be simple/extensible enough for my needs/knowledge. As far as I can tell (and want my file design to be), I don't need concurrent states or anything fancy like that. Some design/implementation questions:
Should I use an enum or an abstract class + derivatives for my states? The first is probably better for small syntax, but could get ugly later, and the second is the exact opposite. I'm leaning to the first, for its simplicity. enum example and class example. EDIT: what about this suggestion for goto, I thought they were evil in C++?
When reading a list, I need to NOT ignore \n. My preferred way of using the string via stringstream, will ignore \n by default. So I need simple way of telling (the same!) stringstream to not ignore newlines when a certain state is enabled.
Will the simple enum states suffice for multi-level parsing (scopes within scopes {...{...}...}) or would that need hacky implementations?
Here's the draft states I have in mind:
upper: reads global, exe, lib+ target names...
normal: inside a scope, can read SOURCES..., create user variables...
list: adds items to a list until a newline is encountered.
Each scope will have a kind of conditional (e.g. win32:global { gcc:CFLAGS = ... }) and will need to be handled in the exact same fashion eveywhere (even in the list state, per item).
Thanks for any input.
If you have nesting scopes, then a Finite State Machine is not the right way to go, and you should look at a Context Free Grammar parser. An LL(1) parser can be written as a set of recursive funcitons, or an LALR(1) parser can be written using a parser generator such as Bison.
If you add a stack to an FSM, then you're getting into pushdown automaton territory. A nondeterministic pushdown automaton is equivalent to a context free grammar (though a deterministic pushdown automaton is strictly less powerful.) LALR(1) parser generators actually generate a deterministic pushdown automaton internally. A good compiler design textbook will cover the exact algorithm by which the pushdown automaton is constructed from the grammar. (In this way, adding a stack isn't "hacky".) This Wikipedia article also describes how to construct the LR(1) pushdown automaton from your grammar, but IMO, the article is not as clear as it could be.
If your scopes nest only finitely deep (i.e. you have the upper, normal and list levels but you don't have nested lists or nested normals), then you can use a FSM without a stack.
There are two stages to analyzing a text input stream for parsing:
Lexical Analysis: This is where your input stream is broken into lexical units. It looks at a sequence of characters and generates tokens (analagous to word in spoken or written languages). Finite state machines are very good at lexical analysis provided you've made good design decision about the lexical structure. From your data above, individal lexemes would be things like your keywords (e.g. "global"), identifiers (e.g. "bitwise", "SOURCES"), symbolic tokesn (e.g. "{" "}", ".", "/"), numeric values, escape values (e.g. "\n"), etc.
Syntactic / Grammatic Analysis: Upon generating a sequence of tokens (or perhaps while you're doing so) you need to be able to analyze the structure to determine if the sequence of tokens is consistent with your language design. You generally need some sort of parser for this, though if the language structure is not very complicated, you may be able to do it with a finite state machine instead. In general (and since you want nesting structures in your case in particular) you will need to use one of the techniques Ken Bloom describes.
So in response to your questions:
Should I use an enum or an abstract class + derivatives for my states?
I found that for small tokenizers, a matrix of state / transition values is suitable, something like next_state = state_transitions[current_state][current_input_char]. In this case, the next_state and current_state are some integer types (including possibly an enumerated type). Input errors are detected when you transition to an invalid state. The end of an token is identified based on the state identification of valid endstates with no valid transition available to another state given the next input character. If you're concerned about space, you could use a vector of maps instead. Making the states classes is possible, but I think that's probably making thing more difficult than you need.
When reading a list, I need to NOT ignore \n.
You can either create a token called "\n", or a more generalize escape token (an identifier preceded by a backslash. If you're talking about identifying line breaks in the source, then those are simply characters you need to create transitions for in your state transition matrix (be aware of the differnce between Unix and Windows line breaks, however; you could create a FSM that operates on either).
Will the simple enum states suffice for multi-level parsing (scopes within scopes {...{...}...}) or would that need hacky implementations?
This is where you will need a grammar or pushdown automaton unless you can guarantee that the nesting will not exceed a certain level. Even then, it will likely make your FSM very complex.
Here's the draft states I have in mind: ...
See my commments on lexical and grammatical analysis above.
For parsing I always try to use something already proven to work: ANTLR with ANTLRWorks which is of great help for designing and testing a grammar. You can generate code for C/C++ (and other languages) but you need to build the ANTLR runtime for those languages.
Of course if you find flex or bison easier to use you can use them too (I know that they generate only C and C++ but I may be wrong since I didn't use them for some time).

Complexity of Regex substitution

I didn't get the answer to this anywhere. What is the runtime complexity of a Regex match and substitution?
Edit: I work in python. But would like to know in general about most popular languages/tools (java, perl, sed).
From a purely theoretical stance:
The implementation I am familiar with would be to build a Deterministic Finite Automaton to recognize the regex. This is done in O(2^m), m being the size of the regex, using a standard algorithm. Once this is built, running a string through it is linear in the length of the string - O(n), n being string length. A replacement on a match found in the string should be constant time.
So overall, I suppose O(2^m + n).
Other theoretical info of possible interest.
For clarity, assume the standard definition for a regular expression
http://en.wikipedia.org/wiki/Regular_language
from the formal language theory. Practically, this means that the only building
material are alphabet symbols, operators of concatenation, alternation and
Kleene closure, along with the unit and zero constants (which appear for
group-theoretic reasons). Generally it's a good idea not to overload this term
despite the everyday practice in scripting languages which leads to
ambiguities.
There is an NFA construction that solves the matching problem for a regular
expression r and an input text t in O(|r| |t|) time and O(|r|) space, where
|-| is the length function. This algorithm was further improved by Myers
http://doi.acm.org/10.1145/128749.128755
to the time and space complexity O(|r| |t| / log |t|) by using automaton node listings and the Four Russians paradigm. This paradigm seems to be named after four Russian guys who wrote a groundbreaking paper which is not
online. However, the paradigm is illustrated in these computational biology
lecture notes
http://lyle.smu.edu/~saad/courses/cse8354/lectures/lecture5.pdf
I find it hilarious to name a paradigm by the number and
the nationality of authors instead of their last names.
The matching problem for regular expressions with added backreferences is
NP-complete, which was proven by Aho
http://portal.acm.org/citation.cfm?id=114877
by a reduction from the vertex-cover problem which is a classical NP-complete problem.
To match regular expressions with backreferences deterministically we could
employ backtracking (not unlike the Perl regex engine) to keep track of the
possible subwords of the input text t that can be assigned to the variables in
r. There are only O(|t|^2) subwords that can be assigned to any one variable
in r. If there are n variables in r, then there are O(|t|^2n) possible
assignments. Once an assignment of substrings to variables is fixed, the
problem reduces to the plain regular expression matching. Therefore the
worst-case complexity for matching regular expressions with backreferences is
O(|t|^2n).
Note however, regular expressions with backreferences are not yet
full-featured regexen.
Take, for example, the "don't care" symbol apart from any other
operators. There are several polynomial algorithms deciding whether a set of
patterns matches an input text. For example, Kucherov and Rusinowitch
http://dx.doi.org/10.1007/3-540-60044-2_46
define a pattern as a word w_1#w_2#...#w_n where each w_i is a word (not a regular expression) and "#" is a variable length "don't care" symbol not contained in either of w_i. They derive an O((|t| + |P|) log |P|) algorithm for matching a set of patterns P against an input text t, where |t| is the length of the text, and |P| is the length of all the words in P.
It would be interesting to know how these complexity measures combine and what
is the complexity measure of the matching problem for regular expressions with
backreferences, "don't care" and other interesting features of practical
regular expressions.
Alas, I haven't said a word about Python... :)
Depends on what you define by regex. If you allow operators of concatenation, alternative and Kleene-star, the time can actually be O(m*n+m), where m is size of a regex and n is length of the string. You do it by constructing a NFA (that is linear in m), and then simulating it by maintaining the set of states you're in and updating that (in O(m)) for every letter of input.
Things that make regex parsing difficult:
parentheses and backreferences: capturing is still OK with the aforementioned algorithm, although it would get the complexity higher, so it might be infeasable. Backreferences raise the recognition power of the regex, and its difficulty is well
positive look-ahead: is just another name for intersection, which raises the complexity of the aforementioned algorithm to O(m^2+n)
negative look-ahead: a disaster for constructing the automaton (O(2^m), possibly PSPACE-complete). But should still be possible to tackle with the dynamic algorithm in something like O(n^2*m)
Note that with a concrete implementation, things might get better or worse. As a rule of thumb, simple features should be fast enough, and unambiguous (eg. not like a*a*) regexes are better.
To delve into theprise's answer, for the construction of the automaton, O(2^m) is the worst case, though it really depends on the form of the regular expression (for a very simple one that matches a word, it's in O(m), using for example the Knuth-Morris-Pratt algorithm).
Depends on the implementation. What language/library/class? There may be a best case, but it would be very specific to the number of features in the implementation.
You can trade space for speed by building a nondeterministic finite automaton instead of a DFA. This can be traversed in linear time. Of course, in the worst case this could need O(2^m) space. I'd expect the tradeoff to be worth it.
If you're after matching and substitution, that implies grouping and backreferences.
Here is a perl example where grouping and backreferences can be used to solve an NP complete problem:
http://perl.plover.com/NPC/NPC-3SAT.html
This (coupled with a few other theoretical tidbits) means that using regular expressions for matching and substitution is NP-complete.
Note that this is different from the formal definition of a regular expression - which don't have the notion of grouping - and match in polynomial time as described by the other answers.
In python's re library, even if a regex is compiled, the complexity can still be exponential (in string length) in some cases, as it is not built on DFA. Some references here, here or here.