Say I have the following, in a toy DSL:
int foo(int bar = 0);
With a tool such as rust-peg, I could define some simple parser expression grammar (PEG) rules to match it (assume appropriate structs FnProto and 'Arg'):
function -> FnProto
= t:type " " n:name "(" v:arglist ");"
{ FnProto { return_type:t, name:n, args:v } }
arglist -> Vec<Arg>
= arg ** ","
arg -> Arg
= t:type " " n:name " = " z:integer { Arg { typename:t, name:n, value:z } }
type -> String
= "int" { match_str.to_string() }
name -> String
= [a-zA-Z_]+[a-zA-Z0-9_] { match_str.to_string() }
integer -> i64
= "-"? [0-9]+ { match_str.parse().unwrap() }
In practice such simple rules are insufficient, but they will serve to illustrate my point.
Now consider the following situation, where the default value of bar is a constant defined previously in the same file:
int BAZ = 0xDEADBEEF;
int foo(int bar = BAZ);
Now the rule for parsing functions needs to accept not only integer literals as default argument values, but also any previously declared constants.
I could do one pass to parse constants and substitute the appropriate values in a second pass, but do I really have to resort to two passes? Is there some way I can refer to previously parsed data from within a rule?
You are confusing "parsing" (the recognition of a valid program, perhaps including capture of a representation of it [e.g, as an AST]) and semantic analysis and/or execution.
Your parser should define what is legal to say, syntactically, in the language. Nothing less, and nothing more. You might be able to write some programs that are semantic nonsense that the parser will not complain about.
Having parsed the text, you now need "other passes" over the parsed data (not the source text) to build classic compiler structures such as symbol tables, and to check that all uses of symbols are valid. To do those other passes, you could arguably reparse the text but you've done that already once by assumption. The standard solution here is to have the first parse build an abstract syntax tree (AST) representing the essential details of the program. Those "other passes" operate by walking the AST rather than parsing the source text again.
This is all classic and taught in standard compiler classes and books. If you are serious about building a programming language, you will need this background.
Related
Any idea on how to implement macro syntax with Rascal and also how to implement the typing and expansion(translation) of the macro syntax in Rascal? Any link to projects or repositories on this problem would also be appreciated.
Macro's are definitions of code substitutions in syntax trees, which is definitely one of the main features of Rascal. Questions I would have before advicing specific techniques:
adding macro's to an existing languages, or to a new language?
macro's at refactoring time, at compile-time or a run-time?
which would inform the question whether or not to implement macro's on concrete syntaxt trees or abstract syntax trees.
I would not say macros are a "problem" per se. The raw substitutions in syntax trees are trivial with Rascal. However, "hygienic macros" are more involved. Here we have to consider the capturing of variables by the expanded macro bodies, and what we can do about this (renaming) to avoid it. The literature on how to make macros hygienic is plenty. The complexity of hygienic macros depends on the type and name analysis (scoping) system of the base language that macros are added to.
If you have a DSL that you want to translate in stages to the target code, that can also be called "macros", but you will not find that name in the documentation. Here is an example: https://github.com/usethesource/flybytes/blob/main/src/lang/flybytes/macros/ControlFlow.rsc where "macro" is used to rewrite an additional AST node to its semantics in the "core" language.
The basic mechanisms are:
pattern matching: detects what you want to expand, with macros this is often a single ADT constructor but it can also be a more complex special case like matching i+=1 to substitute it with i++ .
substitution: at the location where the match was found, we create a new AST value in a simpler language but with the same semantics. This is done with AST expressions in Rascal, the => operator in visit and insert statements, and return and = in functions.
traversal: guiding the pattern matching and substitution without having to write to much boilerplate recursive functions.
Small example:
data Bool(loc src=|unknown:///|)
= \and(Bool l, Bool r)
| \or(Bool r, Bool r)
| \true()
| \false()
| \not(Bool a)
;
I extend the language with a "macro":
data Bool = impl(Bool l, Bool r)
A first option is to rewrite the constructor immediately and always with an overloaded function:
Bool impl(Bool l, Bool r) = or(not(l), r);
However, we lose some information here for debugging purposes, so let's try to keep the information intact:
Bool impl(Bool l, Bool r, src=loc s) = or(not(l), r, src=s);
Sometimes we want to delay the expansion for a specific stage in the compiler. In particular with the above "rewrite rule" a type-checker will not see the different anymore between ==> and || which sometimes creates usability issues with error messages.
In that case we wrap the expansion in a visit and stage it as a function:
Bool macroExpansion(Bool input) = visit(input) {
case impl(Bool l, Bool r, src=loc s) => or(not(l), r, src=s)
// add more rules here
}
It is also possible to encapsulate rewrite rules as reusable functions:
Bool expand1(impl(Bool l, Bool r, src=loc s) = or(not(l), r, src=s);
Bool expand2(not(not(Bool b))) = b;
and then pass those or apply those: (expand1 + expand2)(myBool)
So to wrap this up:
pattern matching is the key to macro expansion, patterns can be wrapped in functions or visit cases or both, and functions can be passed around and combined.
watch out to do some "origin tracking" and forward src fields to the right-hand sides of rewrite rules, otherwise the generated code does not know where it comes from.
I am using Clang/libtooling (ASTComsumer with a Matcher) to visit ALL return statements (ReturnStmt). I need to extract the expression that comes after the keyword return in a string form so that I can put that in a macro that I am replacing return statement with.
For example, I want to replace the following line:
return somefunc() + 1;
with
FUNCTION_EXIT(somefunc() + 1); // FUNCTION_EXIT is a C macro
The macro will return from the function after doing some logging.
I am using ReturnStmt::getRetValue() that returns an Expr and tried to get it in string form (so that it can be passed to the macro), but I haven't found a way yet. Is there a way to stringify Expr?
Clang has a strict separation of concerns between the abstract syntax tree (AST) and the actual source code. The component that converts between these is the Lexer. To get the raw source for an Expr e:
const string text = Lexer::getSourceText(e.getSourceRange(), source_manager, opt);
Note that the SourceManager and LangOptions are available from the ASTContext. If the code you're parsing has macros then things get more complicated because you have to care about spelling location versus expansion location; SourceManager has a bunch of different functions to convert between these.
Good luck!
Is it possible to append terminals retrieved from a text file to a lexicon in Rascal? This would happen at run time, and I see no obvious way to achieve this. I would rather keep the data separate from the Rascal project. For example, if I had read in a list of countries from a text file, how would I add these to a lexicon (using the lexical keyword)?
In the data-dependent version of the Rascal parser this is even easier and faster but we haven't released this yet. For now I'd write a generic rule with a post-parse filter, like so:
rascal>set[str] lexicon = {"aap", "noot", "mies"};
set[str]: {"noot","mies","aap"}
rascal>lexical Word = [a-z]+;
ok
rascal>syntax LexiconWord = word: Word w;
ok
rascal>LexiconWord word(Word w) { // called when the LexiconWord.word rule is use to build a tree
>>>>>>> if ("<w>" notin lexicon)
>>>>>>> filter; // remove this parse tree
>>>>>>> else fail; // just build the tree
>>>>>>>}
rascal>[Sentence] "hello"
|prompt:///|(0,18,<1,0>,<1,18>): ParseError(|prompt:///|(0,18,<1,0>,<1,18>))
at $root$(|prompt:///|(0,64,<1,0>,<1,64>))
rascal>[Sentence] "aap"
Sentence: (Sentence) `aap`
rascal>
Because the filter function removed all possible derivations for hello, the parser eventually returns a parse error on hello. It does not do so for aap which is in the lexicon, so hurray. Of course you can make interestingly complex derivations with this kind of filtering. People sometimes write ambiguous grammars and use filters like so to make it unambiguous.
Parsing and filtering in this way is in cubic worst-case time in terms of the length of the input, if the filtering function is in amortized constant time. If the grammar is linear, then of course the entire process is also linear.
A completely different answer would be to dynamically update the grammar and generate a parser from this. This involves working against the internal grammar representation of Rascal like so:
set[str] lexicon = {"aap", "noot", "mies"};
syntax Word = ; // empty definition
typ = #Word;
grammar = typ.definitions;
grammar[sort("Word")] = { prod(sort("Word"), lit(x), {}) | x <- lexicon };
newTyp = type(sort("Word"), grammar);
This newType is a reified grammar + type for the definition of the lexicon, and which can now be used like so:
import ParseTree;
if (type[Word] staticGrammar := newType) {
parse(staticGrammar, "aap");
}
Now having written al this, two things:
I think this may trigger unknown bugs since we did not test dynamic parser generation, and
For a lexicon with a reasonable size, this will generate an utterly slow parser since the parser is optimized for keywords in programming languages and not large lexicons.
I have designed a new language for that I want to write a reasonable lexer and parser.
For the sake of brevity, I have reduced this language to a minimum so that my questions are still open.
The language has implicit and explicit strings, arrays and objects. An implicit string is just a sequence of characters that does not contain <, {, [ or ]. An explicit string looks like <salt<text>salt> where salt is an arbitrary identifier (i.e. [a-zA-Z][a-zA-Z0-9]*) and text is an arbitrary sequence of characters that does not contain the salt.
An array starts with [, followed by objects and/or strings and ends with ].
All characters within an array that don't belong to an array, object or explicit string do belong to an implicit string and the length of each implicit string is maximal and greater than 0.
An object starts with { and ends with } and consists of properties. A property starts with an identifier, followed by a colon, then optional whitespaces and then either an explicit string, array or object.
So the string [ name:test <xml<<html>[]</html>>xml> {name:<b<test>b>}<b<bla>b> ] represents an array with 6 items: " name:test ", "<html>[]</html>", " ", { name: "test" }, "bla" and " " (the object is notated in json).
As one can see, this language is not context free due to the explicit string (that I don't want to miss). However, the syntax tree is nonambiguous.
So my question is: Is a property a token that may be returned by a tokenizer? Or should the tokenizer return T_identifier, T_colon when he reads an object property?
The real language allows even prefixes in the identifier of a property, e.g. ns/name:<a<test>a> where ns is the prefix for a namespace.
Should the tokenizer return T_property_prefix("ns"), T_property_prefix_separator, T_property_name("name"), T_property_colon or just T_property("ns/name") or even T_identifier("ns"), T_slash, T_identifier("name"), T_colon?
If the tokenizer should recognize properties (which would be useful for syntax highlighters), he must have a stack, because name: is not a property if it is in an array. To decide whether bla: in [{foo:{bar:[test:baz]} bla:{}}] is a property or just an implicit string, the tokenizer must track when he enters and leave an object or array.
Thus, the tokenizer would not be a finite state machine any more.
Or does it make sense to have two tokenizers - the first, which separates whitespaces from alpha-numerical character sequences and special characters like : or [, the second, which uses the first to build more semantical tokens? The parser could then operate on top of the second tokenizer.
Anyways, the tokenizer must have an infinite lookahead to see when an explicit string ends. Or should the detection of the end of an explicit string happen inside the parser?
Or should I use a parser generator for my undertaking? Since my language is not context free, I don't think there is an appropriate parser generator.
Thanks in advance for your answers!
flex can be requested to provide a context stack, and many flex scanners use this feature. So, while it may not fit with a purist view of how a scanner scans, it is a perfectly acceptable and supported feature. See this chapter of the flex manual for details on how to have different lexical contexts (called "start conditions"); at the very end is a brief description of the context stack. (Don't miss the sentence which notes that you need %option stack to enable the stack.) [See Note 1]
Slightly trickier is the requirement to match strings with variable end markers. flex does not have any variable match feature, but it does allow you to read one character at a time from the scanner input, using the function input(). That's sufficient for your language (at least as described).
Here's a rough outline of a possible scanner:
%option stack
%x SC_OBJECT
%%
/* initial/array context only */
[^][{<]+ yylval = strdup(yytext); return STRING;
/* object context only */
<SC_OBJECT>{
[}] yy_pop_state(); return '}';
[[:alpha:]][[:alnum:]]* yylval = strdup(yytext); return ID;
[:/] return yytext[0];
[[:space:]]+ /* Ignore whitespace */
}
/* either context */
<*>{
[][] return yytext[0]; /* char class with [] */
[{] yy_push_state(SC_OBJECT); return '{';
"<"[[:alpha:]][[:alnum:]]*"<" {
/* We need to save a copy of the salt because yytext could
* be invalidated by input().
*/
char* salt = strdup(yytext);
char* saltend = salt + yyleng;
char* match = salt;
/* The string accumulator code is *not* intended
* to be a model for how to write string accumulators.
*/
yylval = NULL;
size_t length = 0;
/* change salt to what we're looking for */
*salt = *(saltend - 1) = '>';
while (match != saltend) {
int ch = input();
if (ch == EOF) {
yyerror("Unexpected EOF");
/* free the temps and do something */
}
if (ch == *match) ++match;
else if (ch == '>') match = salt + 1;
else match = salt;
/* Don't do this in real code */
yylval = realloc(yylval, ++length);
yylval[length - 1] = ch;
}
/* Get rid of the terminator */
yylval[length - yyleng] = 0;
free(salt);
return STRING;
}
. yyerror("Invalid character in object");
}
I didn't test that thoroughly, but here is what it looks like with your example input:
[ name:test <xml<<html>[]</html>>xml> {name:<b<test>b>}<b<bla>b> ]
Token: [
Token: STRING: -- name:test --
Token: STRING: --<html>[]</html>--
Token: STRING: -- --
Token: {
Token: ID: --name--
Token: :
Token: STRING: --test--
Token: }
Token: STRING: --bla--
Token: STRING: -- --
Token: ]
Notes
In your case, unless you wanted to avoid having a parser, you don't actually need a stack since the only thing that needs to be pushed onto the stack is an object context, and a stack with only one possible value can be replaced with a counter.
Consequently, you could just remove the %option stack and define a counter at the top of the scan. Instead of pushing the start condition, you increment the counter and set the start condition; instead of popping, you decrement the counter and reset the start condition if it drops to 0.
%%
/* Indented text before the first rule is inserted at the top of yylex */
int object_count = 0;
<*>[{] ++object_count; BEGIN(SC_OBJECT); return '{';
<SC_OBJECT[}] if (!--object_count) BEGIN(INITIAL); return '}'
Reading the input one character at a time is not the most efficient. Since in your case, a string terminate must start with >, it would probably be better to define a separate "explicit string" context, in which you recognized [^>]+ and [>]. The second of these would do the character-at-a-time match, as with the above code, but would terminate instead of looping if it found a non-matching character other than >. However, the simple code presented may turn out to be fast enough, and anyway it was just intended to be good enough to do a test run.
I think the traditional way to parse your language would be to have the tokenizer return T_identifier("ns"), T_slash, T_identifier("name"), T_colon for ns/name:
Anyway, I can see three reasonable ways you could implement support for your language:
Use lex/flex and yacc/bison. The tokenizers generated by lex/flex do not have stack so you should be using T_identifier and not T_context_specific_type. I didn't try the approach so I can't give a definite comment on whether your language could be parsed by lex/flex and yacc/bison. So, my comment is try it to see if it works. You may find information about the lexer hack useful: http://en.wikipedia.org/wiki/The_lexer_hack
Implement a hand-built recursive descent parser. Note that this can be easily built without separate lexer/parser stages. So, if the lexemes depend on context it is easy to handle when using this approach.
Implement your own parser generator which turns lexemes on and off based on the context of the parser. So, the lexer and the parser would be integrated together using this approach.
I once worked for a major network security vendor where deep packet inspection was performed by using approach (3), i.e. we had a custom parser generator. The reason for this is that approach (1) doesn't work for two reasons: firstly, data can't be fed to lex/flex and yacc/bison incrementally, and secondly, HTTP can't be parsed by using lex/flex and yacc/bison because the meaning of the string "HTTP" depends on its location, i.e. it could be a header value or the protocol specifier. The approach (2) didn't work because data can't be fed incrementally to recursive descent parsers.
I should add that if you want to have meaningful error messages, a recursive descent parser approach is heavily recommended. My understanding is that the current version of gcc uses a hand-built recursive descent parser.
I have read the GOLD Homepage ( http://www.devincook.com/goldparser/ ) docs, FAQ and Wikipedia to find out what practical application there could possibly be for GOLD. I was thinking along the lines of having a programming language (easily) available to my systems such as ABAP on SAP or X++ on Axapta - but it doesn't look feasible to me, at least not easily - even if you use GOLD.
The final use of the parsed result produced by GOLD escapes me - what do you do with the result of the parse?
EDIT: A practical example (description) would be great.
Parsing really consists of two phases. The first is "lexing", which convert the raw strings of character in to something that the program can more readily understand (commonly called tokens).
Simple example, lex would convert:
if (a + b > 2) then
In to:
IF_TOKEN LEFT_PAREN IDENTIFIER(a) PLUS_SIGN IDENTIFIER(b) GREATER_THAN NUMBER(2) RIGHT_PAREN THEN_TOKEN
The parse takes that stream of tokens, and attempts to make yet more sense out of them. In this case, it would try and match up those tokens to an IF_STATEMENT. To the parse, the IF _STATEMENT may well look like this:
IF ( BOOLEAN_EXPRESSION ) THEN
Where the result of the lexing phase is a token stream, the result of the parsing phase is a Parse Tree.
So, a parser could convert the above in to:
if_statement
|
v
boolean_expression.operator = GREATER_THAN
| |
| v
V numeric_constant.string="2"
expression.operator = PLUS_SIGN
| |
| v
v identifier.string = "b"
identifier.string = "a"
Here you see we have an IF_STATEMENT. An IF_STATEMENT has a single argument, which is a BOOLEAN_EXPRESSION. This was explained in some manner to the parser. When the parser is converting the token stream, it "knows" what a IF looks like, and know what a BOOLEAN_EXPRESSION looks like, so it can make the proper assignments when it sees the code.
For example, if you have just:
if (a + b) then
The parser could know that it's not a boolean expression (because the + is arithmetic, not a boolean operator) and the parse could throw an error at this point.
Next, we see that a BOOLEAN_EXPRESSION has 3 components, the operator (GREATER_THAN), and two sides, the left side and the right side.
On the left side, it points to yet another expression, the "a + b", while on the right is points to a NUMERIC_CONSTANT, in this case the string "2". Again, the parser "knows" this is a NUMERIC constant because we told it about strings of numbers. If it wasn't numbers, it would be an IDENTIFIER (like "a" and "b" are).
Note, that if we had something like:
if (a + b > "XYZ") then
That "parses" just fine (expression on the left, string constant on the right). We don't know from looking at this whether this is a valid expression or not. We don't know if "a" or "b" reference Strings or Numbers at this point. So, this is something the parser can't decided for us, can't flag as an error, as it simply doesn't know. That will happen when we evaluate (either execute or try to compile in to code) the IF statement.
If we did:
if [a > b ) then
The parser can readily see that syntax error as a problem, and will throw an error. That string of tokens doesn't look like anything it knows about.
So, the point being that when you get a complete parse tree, you have some assurance that at first cut the "code looks good". Now during execution, other errors may well come up.
To evaluate the parse tree, you just walk the tree. You'll have some code associated with the major nodes of the parse tree during the compile or evaluation part. Let's assuming that we have an interpreter.
public void execute_if_statment(ParseTreeNode node) {
// We already know we have a IF_STATEMENT node
Value value = evaluate_expression(node.getBooleanExpression());
if (value.getBooleanResult() == true) {
// we do the "then" part of the code
}
}
public Value evaluate_expression(ParseTreeNode node) {
Value result = null;
if (node.isConstant()) {
result = evaluate_constant(node);
return result;
}
if (node.isIdentifier()) {
result = lookupIdentifier(node);
return result;
}
Value leftSide = evaluate_expression(node.getLeftSide());
Value rightSide = evaluate_expression(node.getRightSide());
if (node.getOperator() == '+') {
if (!leftSide.isNumber() || !rightSide.isNumber()) {
throw new RuntimeError("Must have numbers for adding");
}
int l = leftSide.getIntValue();
int r = rightSide.getIntValue();
int sum = l + r;
return new Value(sum);
}
if (node.getOperator() == '>') {
if (leftSide.getType() != rightSide.getType()) {
throw new RuntimeError("You can only compare values of the same type");
}
if (leftSide.isNumber()) {
int l = leftSide.getIntValue();
int r = rightSide.getIntValue();
boolean greater = l > r;
return new Value(greater);
} else {
// do string compare instead
}
}
}
So, you can see that we have a recursive evaluator here. You see how we're checking the run time types, and performing the basic evaluations.
What will happen is the execute_if_statement will evaluate it's main expression. Even tho we wanted only BOOLEAN_EXPRESION in the parse, all expressions are mostly the same for our purposes. So, execute_if_statement calls evaluate_expression.
In our system, all expressions have an operator and a left and right side. Each side of an expression is ALSO an expression, so you can see how we immediately try and evaluate those as well to get their real value. The one note is that if the expression consists of a CONSTANT, then we simply return the constants value, if it's an identifier, we look it up as a variable (and that would be a good place to throw a "I can't find the variable 'a'" message), otherwise we're back to the left side/right side thing.
I hope you can see how a simple evaluator can work once you have a token stream from a parser. Note how during evaluation, the major elements of the language are in place, otherwise we'd have got a syntax error and never got to this phase. We can simply expect to "know" that when we have a, for example, PLUS operator, we're going to have 2 expressions, the left and right side. Or when we execute an IF statement, that we already have a boolean expression to evaluate. The parse is what does that heavy lifting for us.
Getting started with a new language can be a challenge, but you'll find once you get rolling, the rest become pretty straightforward and it's almost "magic" that it all works in the end.
Note, pardon the formatting, but underscores are messing things up -- I hope it's still clear.
I would recommend antlr.org for information and the 'free' tool I would use for any parser use.
GOLD can be used for any kind of application where you have to apply context-free grammars to input.
elaboration:
Essentially, CFGs apply to all programming languages. So if you wanted to develop a scripting language for your company, you'd need to write a parser- or get a parsing program. Alternatively, if you wanted to have a semi-natural language for input for non-programmers in the company, you could use a parser to read that input and spit out more "machine-readable" data. Essentially, a context-free grammar allows you to describe far more inputs than a regular expression. The GOLD system apparently makes the parsing problem somewhat easier than lex/yacc(the UNIX standard programs for parsing).