Yacc may not be dead, but parser generation should be, in my opinion. If one is designing a language today one should, and indeed most new lanvuages in the last 5-10 years have, avoided the pathological ambiguities of C++. Hence implementing parsing via recursive descent is sufficient, and i say trivial as opposed to learning and using any particular parser generator (error handling being the lrimary culprit in my view). Even for C++, i don't know of a major project that does not use a handcrafted rec descent parser.
Can anyone convince me of the value of parser genaration, othr than an interesting academic exercise?
I'm making a language that is equivalent to (and compatible with) C++, but has a regular grammar, bison is quite pleasant to work with, and the lalr grammar is potentially reusable elsewhere (in javascript, with jison, in haskell, with happy, in go, with go-yacc and so on).
It also allows me to change the language structure very quickly which is very important when designing the language.
In the future I'll write a recdes parser as well, because I also want pretty error messages, but the parser generator saved me a ton of time so far and will continue to for a while.
It has its value for "quick and dirty" parser generation for experimentation, or for use in DSLs/ small tools where throwing something together is sufficient. But I agree with you overall - much more sophisticated parser generators than Yacc have still not gotten us to a point where using a parser generator for a full scale compiler for example is particularly viable as a long term strategy.
Parser generators allow you to work on a higher level. The reason why you may want to use one is the same as the reason why you'd want to use a higher level language rather than assembly. Instead of repeating the same pattern to implement every rule in the grammar you simply write down the grammar and the parser generator expands that so that you don't have to. Unfortunately most parser generators are limited in the class of languages they parse, or they are limited in the languages that they can express conveniently. Even parser generators that support full context free grammars are not enough. You need some method to abstract common patterns. For example if you want to express operator parsing in a context free grammar you end up with a separate rule for each level of precedence. Parser combinator libraries do allow you to use the full abstraction facilities of the programming language, but they usually are weak in terms of which grammars they allow you to parse in polynomial time (usually LL), whereas ideally you would be able to parse regular languages in O(n), deterministic languages in O(n) and context free languages in O(n^3). Also parser combinators often do not support streaming & incrementality because their reliance on backtracking forces them to keep the entire input in memory.
I don't know of any parser generator or parser combinator library that simultaneously supports abstraction and supports efficient streaming parsing. Does anybody know one?
PEG parsers are not streaming. They have to keep the entire input in memory in case backtracking happens. I'll look into Pratt parsing. I've actually implemented a Pratt parser in the past but I thought it was just for parsing operators with precedence?
PEG is a superset of recursive descent. You can structure you grammar in a way that backtracking is not required at all or minimised.
> They have to keep the entire input in memory in case backtracking happens.
Not necessarily. You only need to keep something like around a current statement (or other small syntax entity), discarding everything you've already streamed.
> I've actually implemented a Pratt parser in the past but I thought it was just for parsing operators with precedence?
Exactly. And it's really easy to mix it into an otherwise PEG-based parser, eliminating the need to backtrack for the worst backtracking case (binary expressions).
For example, there is a very efficient implementation of such an approach (pure PEG+Pratt, no memoisation and no backtracking) in Nemerle.
Parser generation saves development time when it comes to small or maybe even middle sized languages. Also, it is easier to see bigger picture on how language works when you have single file with rules.
Major project will necessary end up handcrafted, because major project usually need to be fast and optimized. Most projects are not that sensitive - they need to be developed reasonably fast and then need to be easy to be maintained.
If the target is security sensitive, there's whole classes of errors that are avoided by parser generators.
They are also useful to use while shaping or discovering a syntax. Parser generators will quickly show you where the irregularities are because they will become harder (or impossible) to implement.
> Hence implementing parsing via recursive descent is sufficient
And why exactly do you want an overbloated parser implementation full of leaky abstractions, while you can nicely generate it out of a BNF-like dense description?
> Can anyone convince me of the value of parser genaration
The biggest advantage of Earley Parsing is its accessibility.
Most other tools such as parser generators, parsing
expression grammars, or combinator libraries feature
restrictions that often make them hard to use. Use the wrong
kind of grammar, and your PEG will enter an infinite loop.
Use another wrong kind of grammar, and most parser
generators will fail. To a beginner, these restrictions feel
most arbitrary: it looks like it should work, but it doesn't.
There are workarounds of course, but they make these tools
more complex.
Earley parsing Just Works™.
On the flip side, to get this generality we must sacrifice
some speed. Earley parsing cannot compete with speed demons
such as Flex/Bison in terms of raw speed. It's not that bad,
however:
• Earley parsing is cubic in the worst cases, which is the
state of the art (and possibly the best we can do). The speed
demons often don't work at all for those worst cases. Other
parsers are prone to exponential combinatorial explosion.
• Most simple grammars can be parsed in linear time.
• Even the worst unambiguous grammars can be parsed in
quadratic time.
My advice would be to use Earley parsing by default, and only
revert to more specific methods if performance is an issue…
In 2014, we now have Earley parsers in C, JavaScript, Lua, Perl and Python.
Earley parsers give no guarantee that the grammar is not ambiguous. This is very important, since you don't want your parser return some arbitrary parse tree, but you want guarantees that the parse tree that you intended is returned. yacc does this IMHO in the correct way: give warnings (shift/reduce or reduce/reduce conflict in case of a possible ambiguity) and let the user resolve these manually. Since these conflicts are a strong sign of a bad specification of the language, the parser generator should be very cautious in resolving these ambiguities - doing this automatically probably leads to strange consequences the user did not intend.
> you don't want your parser return some arbitrary parse tree
You have a mistake in your assessment of Earley parsers. For ambiguous grammars and input, they do not return an arbitrary tree, but all possible trees. In this regard they are no worse than yacc: "the user resolve[s] these manually" by picking the correct result(s). To a human, it's visible at a glance which is the correct result(s), entirely without needing to learn how to decipher these bizarre warning messages you mentioned.
> conflicts are a strong sign of a bad specification of the language
So what? That's not pragmatic thinking. We cannot go back in time and influence the design of ambiguous computer languages so they are not ambiguous. Natural languages never are without ambiguities!
The parsing job needs to be done, no matter the complexity of the language. Instead of wishing it weren't so, just use a tool that can deal with it. Earley parsers can, yacc cannot.
> You have a mistake in your assessment of Earley parsers. For ambiguous grammars and input, they do not return an arbitrary tree, but all possible trees.
There can easily be an exponential number of parse trees for an ambiguous grammar, which is a contradiction to the runtime guarantee of the Earley algorithm.
> To a human, it's visible at a glance which is the correct result(s), entirely without needing to learn how to decipher these bizarre warning messages you mentioned.
The bad warning messages are a problem of yacc and not of the LALR(1) algorithm that yacc uses by default.
> So what? That's not pragmatic thinking. We cannot go back in time and influence the design of ambiguous computer languages so they are not ambiguous.
Please explain how you arrive at the conclusion that an Earley parser optimises entirely for the wrong thing. What is the wrong thing?
Competing with a manually written parser is easy: Earley parsers are derived from a CFG in standard format, such as ISO EBNF or IETF ABNF. Such a grammar is much easier to reason about than parser in a general programming language such as C.
"Earley parsers, which Cox apparently didn't know about in 2010"
How do you draw that conclusion? I see nothing in the article which says that he did or didn't know about Earley parsers. A quick search finds this posting by Cox from 17 Apr 2006 at http://compilers.iecc.com/comparch/article/06-04-111 :
> Although few people do use Earley and Tomita parsers in practice now, I think general approaches, especially GLR, are gaining ground.
Furthermore, the Wikipedia page for GLR says:
> Recognition using the GLR algorithm has the same worst-case time complexity as the CYK algorithm and Earley algorithm: O(n^3). However, GLR carries two additional advantages:
> - The time required to run the algorithm is proportional to the degree of nondeterminism in the grammar: on deterministic grammars the GLR algorithm runs in O(n) time (this is not true of the Earley[citation needed] and CYK algorithms, but the original Earley algorithms can be modified to ensure it)
> - The GLR algorithm is "online" – that is, it consumes the input tokens in a specific order and performs as much work as possible after consuming each token.
> Compared to other algorithms capable of handling the full class of context-free grammars (such as Earley or CYK), the GLR algorithm gives better performance on these "nearly deterministic" grammars, because only a single stack will be active during the majority of the parsing process.
Perhaps Cox knew about and rejected bringing up Earley in favor of GLR, for several sound reasons that you didn't know about in 2014?
No need to get so agitated. Your reply comes across unnecessarily hostile for no good reason.
> How do you draw that conclusion? I see nothing in the article which says that he did or didn't know about Earley parsers.
Simple inference from it not being mentioned, even though I thought it deserved to be. Since that does not prove anything, I wrote "apparently" – I anticipated my assessment could be wrong, and indeed it was.
> the Wikipedia page for GLR says
I'm not happy with that article. It gives people the wrong ideas, it's not realistically useful to make comparisons with the decades-old original algorithm. Modern Earley parsers do contain optimisations that makes those distinctions mentioned there moot. And unless I completely misunderstand what the WP contributor aimed to express, the Earley algorithm is "online", too, and that is the case even for unmodified/unoptimised Earley parsing. See http://web.stanford.edu/class/archive/cs/cs143/cs143.1128/le... or just step through an implementation with a debugger. I think the reasons are not as "sound" as you concluded them to be.
To me it appears after all that GLR and Earley are equal in power, so Cox shouldn't simply reject, and implementations compete in areas other than the algorithm, e.g. sensible error reporting, simple interface for simple use cases, ability to consume grammars in standard formats, coverage by number of programming languages and such like.
Unnecessarily hostile? I even used "Perhaps" where you used "apparently", and quoted a block of third-party text like you did.
Cox wrote "These tools and many others all have the guarantee that if they tell you the grammar is unambiguous, they'll give you a linear-time parser, and if not, they'll give you at worst a cubic-time parser. Computer science theory doesn't know a better way. But any of these is better than an exponential time parser."
It's more generous to believe that Earley is simply one of the "many others" that were unenumerated, but equal in power to GLR.
You can certainly argue that there are pluses and minus to all of them, but they are irrelevant for the context of the essay. That section is very short and can't be seen as being a complete summary of alternatives, but rather observation that "newer tools that provide compelling alternatives still embody [the spirit of yacc]", including bison.
The lack of a reference to Earley is not indicative that the author does not know it. Consider that ANLR uses adaptive LL( * ) because:
> The biggest problem for the average practitioner is that most parser generators do not produce code you can load into a debugger and step through. This immediately removes bottom-up parser generators and the really powerful GLR parser generators from consideration by the average programmer. There are a few other tools that generate source code like ANTLR does, but they don't have v4's adaptive LL( * ) parsers. You will be stuck with contorting your grammar to fit the needs of the tool's weaker, say, LL(k) parsing strategy. PEG-based tools have a number of weaknesses, but to mention one, they have essentially no error recovery because they cannot report an error and until they have parsed the entire input.
FWIW, I was using an Earley-based parser for Python as part of the SPARK package back in 2000, and I'm by far an expert in the field, so I think it's unreasonable to assume, as you did, that a practitioner in the field wouldn't know about it and have other reasons for not enumerating it specifically.
"Reject" is my word, not Cox's.
Nor did I mean to imply that the reasons on Wikipedia were the same as the ones the Cox used when deciding to not mention Earley, only that there could be reasons. Quoting Parr at http://blog.athico.com/2007/06/interview-with-antlr-30-autho... " GLR and Earley and CYK can deal with the same class of grammars (all context-free grammars), but GLR is more efficient." That one reason alone might be enough for Cox to have decided to mention GLR and leave Earley in the category "and many other[ tools]".
From the article: "...though Bison still retains yacc's infuriating lack of detail in error messages. (I use an awk script to parse the bison.output file and tell me what really went wrong.)"
Oh dear. Now I'm going to have to get a new Irony meter, because mine just blew...
> Elkhound is a parser generator, similar to Bison. The parsers it generates use the Generalized LR (GLR) parsing algorithm. GLR works with any context-free grammar, whereas LR parsers (such as Bison) require grammars to be LALR(1).
According to this swtch.com essay, "GNU Bison can optionally generate a GLR parser instead of an LALR(1) parsers" and checking history shows that GLR was available in Bison 1.75 in 2002 (See http://lists.gnu.org/archive/html/info-gnu/2002-10/msg00008.... .)
Regarding "so language specific"; proper language support for a given language is hard, and that's where most of the development time goes.
Adding support for two languages is more than twice as hard as support for one language. Take a look at the comments for Java support in Bison, at http://www.gnu.org/software/bison/manual/bison.html#Java-Par... to see some of the difficulties and incomplete aspects of that port. Now consider a port to Python, which doesn't have a switch statement so likely needs very different code generation style.
I used Elsa a lot for static C++ code analysis, it was long before Clang was available. Cannot compare to antlr, since I've never seen a usable antlr C++ grammar implementation.
My favorite utility for this task is spitbol. I do not know of any software that is more naturally suited to working with BNF. Nothing I have seen is as flexible, either. I'm currently learning an additional, interpreted language and testing its limits; it is quite fast, so my opinion could change. But I doubt it.
What killed my understanding of Yacc is the ad-hoc nature of semantic actions, I could never grasp what was in scope when it happened. You could access some state. Well I was never imperative oriented. I feel it could be enhanced with better integrated constructs like closures. C++ have them, I've seen people adding lambdas to C too, so maybe ...
I've been working on a derivative parser for PEGs; it's not quite working yet, but the inherent lack of ambiguity in PEGs is helpful to the time bounds there (I think I can make it worst case cubic, and linear in a lot of common cases). I've got some ideas how to modify the algorithm to a better derivative parser for CFGs; I should be able to recognize arbitrary CFGs in linear time, and I think parse them in cubic (carrying around the set of current parse tree options is expensive, but I think if you store them as a DAG of parsing paths rather than a parse tree you can make it tolerable).
31 comments
[ 4.1 ms ] story [ 66.8 ms ] threadCan anyone convince me of the value of parser genaration, othr than an interesting academic exercise?
We certainly don't want languages with hidden ambiguities, which are resolved in different ways in different implementations.
I don't know of any parser generator or parser combinator library that simultaneously supports abstraction and supports efficient streaming parsing. Does anybody know one?
PEG is a superset of recursive descent. You can structure you grammar in a way that backtracking is not required at all or minimised.
> They have to keep the entire input in memory in case backtracking happens.
Not necessarily. You only need to keep something like around a current statement (or other small syntax entity), discarding everything you've already streamed.
> I've actually implemented a Pratt parser in the past but I thought it was just for parsing operators with precedence?
Exactly. And it's really easy to mix it into an otherwise PEG-based parser, eliminating the need to backtrack for the worst backtracking case (binary expressions).
For example, there is a very efficient implementation of such an approach (pure PEG+Pratt, no memoisation and no backtracking) in Nemerle.
There is also a Packrat+Pratt parser used in https://github.com/combinatorylogic/mbase
Major project will necessary end up handcrafted, because major project usually need to be fast and optimized. Most projects are not that sensitive - they need to be developed reasonably fast and then need to be easy to be maintained.
They are also useful to use while shaping or discovering a syntax. Parser generators will quickly show you where the irregularities are because they will become harder (or impossible) to implement.
And why exactly do you want an overbloated parser implementation full of leaky abstractions, while you can nicely generate it out of a BNF-like dense description?
> Can anyone convince me of the value of parser genaration
https://en.wikipedia.org/wiki/Parsing_expression_grammar
Further discussion on killing yacc:
http://jeffreykegler.github.io/Ocean-of-Awareness-blog/indiv... http://jeffreykegler.github.io/Ocean-of-Awareness-blog/indiv... http://jeffreykegler.github.io/Ocean-of-Awareness-blog/indiv... http://jeffreykegler.github.io/Ocean-of-Awareness-blog/indiv...
You have a mistake in your assessment of Earley parsers. For ambiguous grammars and input, they do not return an arbitrary tree, but all possible trees. In this regard they are no worse than yacc: "the user resolve[s] these manually" by picking the correct result(s). To a human, it's visible at a glance which is the correct result(s), entirely without needing to learn how to decipher these bizarre warning messages you mentioned.
> conflicts are a strong sign of a bad specification of the language
So what? That's not pragmatic thinking. We cannot go back in time and influence the design of ambiguous computer languages so they are not ambiguous. Natural languages never are without ambiguities!
The parsing job needs to be done, no matter the complexity of the language. Instead of wishing it weren't so, just use a tool that can deal with it. Earley parsers can, yacc cannot.
There can easily be an exponential number of parse trees for an ambiguous grammar, which is a contradiction to the runtime guarantee of the Earley algorithm.
> To a human, it's visible at a glance which is the correct result(s), entirely without needing to learn how to decipher these bizarre warning messages you mentioned.
The bad warning messages are a problem of yacc and not of the LALR(1) algorithm that yacc uses by default.
> So what? That's not pragmatic thinking. We cannot go back in time and influence the design of ambiguous computer languages so they are not ambiguous.
The languages are typically not ambiguous, but their reference grammar is (most famous problem is the dangling else (http://en.wikipedia.org/wiki/Dangling_else)).
Its very useful to be able to identify ambiguities statically, during the language design stage.
And "killing Yacc" is a pointless exercise if the point is to get traction. What it needs to compete with is hand-written recursive descent parsers.
Competing with a manually written parser is easy: Earley parsers are derived from a CFG in standard format, such as ISO EBNF or IETF ABNF. Such a grammar is much easier to reason about than parser in a general programming language such as C.
How do you draw that conclusion? I see nothing in the article which says that he did or didn't know about Earley parsers. A quick search finds this posting by Cox from 17 Apr 2006 at http://compilers.iecc.com/comparch/article/06-04-111 :
> Although few people do use Earley and Tomita parsers in practice now, I think general approaches, especially GLR, are gaining ground.
Furthermore, the Wikipedia page for GLR says:
> Recognition using the GLR algorithm has the same worst-case time complexity as the CYK algorithm and Earley algorithm: O(n^3). However, GLR carries two additional advantages:
> - The time required to run the algorithm is proportional to the degree of nondeterminism in the grammar: on deterministic grammars the GLR algorithm runs in O(n) time (this is not true of the Earley[citation needed] and CYK algorithms, but the original Earley algorithms can be modified to ensure it)
> - The GLR algorithm is "online" – that is, it consumes the input tokens in a specific order and performs as much work as possible after consuming each token.
> Compared to other algorithms capable of handling the full class of context-free grammars (such as Earley or CYK), the GLR algorithm gives better performance on these "nearly deterministic" grammars, because only a single stack will be active during the majority of the parsing process.
Perhaps Cox knew about and rejected bringing up Earley in favor of GLR, for several sound reasons that you didn't know about in 2014?
> How do you draw that conclusion? I see nothing in the article which says that he did or didn't know about Earley parsers.
Simple inference from it not being mentioned, even though I thought it deserved to be. Since that does not prove anything, I wrote "apparently" – I anticipated my assessment could be wrong, and indeed it was.
> the Wikipedia page for GLR says
I'm not happy with that article. It gives people the wrong ideas, it's not realistically useful to make comparisons with the decades-old original algorithm. Modern Earley parsers do contain optimisations that makes those distinctions mentioned there moot. And unless I completely misunderstand what the WP contributor aimed to express, the Earley algorithm is "online", too, and that is the case even for unmodified/unoptimised Earley parsing. See http://web.stanford.edu/class/archive/cs/cs143/cs143.1128/le... or just step through an implementation with a debugger. I think the reasons are not as "sound" as you concluded them to be.
To me it appears after all that GLR and Earley are equal in power, so Cox shouldn't simply reject, and implementations compete in areas other than the algorithm, e.g. sensible error reporting, simple interface for simple use cases, ability to consume grammars in standard formats, coverage by number of programming languages and such like.
Cox wrote "These tools and many others all have the guarantee that if they tell you the grammar is unambiguous, they'll give you a linear-time parser, and if not, they'll give you at worst a cubic-time parser. Computer science theory doesn't know a better way. But any of these is better than an exponential time parser."
It's more generous to believe that Earley is simply one of the "many others" that were unenumerated, but equal in power to GLR.
You can certainly argue that there are pluses and minus to all of them, but they are irrelevant for the context of the essay. That section is very short and can't be seen as being a complete summary of alternatives, but rather observation that "newer tools that provide compelling alternatives still embody [the spirit of yacc]", including bison.
The lack of a reference to Earley is not indicative that the author does not know it. Consider that ANLR uses adaptive LL( * ) because:
> The biggest problem for the average practitioner is that most parser generators do not produce code you can load into a debugger and step through. This immediately removes bottom-up parser generators and the really powerful GLR parser generators from consideration by the average programmer. There are a few other tools that generate source code like ANTLR does, but they don't have v4's adaptive LL( * ) parsers. You will be stuck with contorting your grammar to fit the needs of the tool's weaker, say, LL(k) parsing strategy. PEG-based tools have a number of weaknesses, but to mention one, they have essentially no error recovery because they cannot report an error and until they have parsed the entire input.
That's from https://theantlrguy.atlassian.net/wiki/pages/viewpage.action... . The page doesn't mention Earley parsers either. I don't think that Terence Parr, author of ANTL and that quote, is ignorant of Earley parsers in 2013. (Especially as Parr mentions Earley in 2007 in http://blog.athico.com/2007/06/interview-with-antlr-30-autho... . Note also the issues with GLR in https://qconsf.com/system/files/presentation-slides/quest-fo... and compare to the lone reference in that presentation to Earley).
FWIW, I was using an Earley-based parser for Python as part of the SPARK package back in 2000, and I'm by far an expert in the field, so I think it's unreasonable to assume, as you did, that a practitioner in the field wouldn't know about it and have other reasons for not enumerating it specifically.
"Reject" is my word, not Cox's.
Nor did I mean to imply that the reasons on Wikipedia were the same as the ones the Cox used when deciding to not mention Earley, only that there could be reasons. Quoting Parr at http://blog.athico.com/2007/06/interview-with-antlr-30-autho... " GLR and Earley and CYK can deal with the same class of grammars (all context-free grammars), but GLR is more efficient." That one reason alone might be enough for Cox to have decided to mention GLR and leave Earley in the category "and many other[ tools]".
Oh dear. Now I'm going to have to get a new Irony meter, because mine just blew...
Also, why do parser generators always have to be so language specific?
[1] http://scottmcpeak.com/elkhound/
> Elkhound is a parser generator, similar to Bison. The parsers it generates use the Generalized LR (GLR) parsing algorithm. GLR works with any context-free grammar, whereas LR parsers (such as Bison) require grammars to be LALR(1).
Regarding "so language specific"; proper language support for a given language is hard, and that's where most of the development time goes.
Adding support for two languages is more than twice as hard as support for one language. Take a look at the comments for Java support in Bison, at http://www.gnu.org/software/bison/manual/bison.html#Java-Par... to see some of the difficulties and incomplete aspects of that port. Now consider a port to Python, which doesn't have a switch statement so likely needs very different code generation style.
PS: Also, see that article thread about limitations of Parsing (composability) and other ideas. https://news.ycombinator.com/item?id=2327313