I've taken a deep dive into the world of parsers recently, wanting to create my own programming language.

However, I found out that there exist two somewhat different approaches of writing parsers: Parser Generators and Parser Combinators.

Interestingly, I have been unable to find any resource that explained in what cases which approach is better; Rather, many resources (and persons) I queried about the subject did not know of the other approach, only explaining their approach as the approach and not mentioning the other at all:

Simple Overview:

Parser Generator

A Parser Generator takes a file written in a DSL that is some dialect of Extended Backus-Naur form, and turns it into source code that can then (when compiled) become a parser for the input language that was described in this DSL.

This means that the compilation process is done in two separate steps. Interestingly, Parser Generators themselves are also compilers (and many of them are indeed self-hosting).

Parser Combinator

A Parser Combinator describes simple functions called parsers that all take an input as parameter, and try to pluck off the first character(s) of this input if they match. They return a tuple (result, rest_of_input), where result might be empty (e.g. nil or Nothing) if the parser was unable to parse anything from this input. An example would be an digit parser. Other parsers can of course take parsers as first arguments (the final argument still remaining the input string) to combine them: e.g. many1 attempts to match another parser as many times as possible (but at least once, or it itself fails).

You can now of course combine (compose) digit and many1, to create a new parser, say integer.

Also, a higher-level choice parser can be written that takes a list of parsers, trying each of them in turn.

In this way, very complex lexers/parsers can be built. In languages supporting operator overloading, this also looks very much like EBNF, even though it is still written directly in the target language (and you can use all features of the target language you desire).

Simple Differences


  • Parser Generators are written in a combination of the EBNF-ish DSL and the code that these statements should generate to when they match.
  • Parser Combinators are written in the target language directly.


  • Parser Generators have a very distinct difference between the 'lexer' (which splits a string into tokens that might be tagged to show what kind of value we are dealing with) and the 'parser' (which takes the output list of tokens from the lexer and attempts to combine them, forming an Abstract Syntax Tree).
  • Parser Combinators do not have/need this distinction; usually, simple parsers perform the work of the 'lexer' and the more high-level parsers call these simpler ones to decide which kind of AST-node to create.


However, even given these differences (and this list of differences is probably far from complete!), I cannot make an educated choice on when to use which one. I fail to see what the implications/consequences are of these differences.

What problem properties would indicate that a problem would better be solved using a Parser Generator? What problem properties would indicate that a problem would better be solved using a Parser Combinator?

  • 10
    There are at least two more ways of implementing parsers that you didn't mention: parser interpreters (similar to parser generators, except instead of compiling the parser language to e.g. C or Java, the parser language is executed directly), and simply writing the parser by hand. Writing the parser by hand is the preferred form of implementation for many modern production-ready industrial-strength language implementations (e.g. GCC, Clang, javac, Scala). It gives you the most control over the internal parser state, which helps with generating good error messages (which in recent years … Dec 22, 2016 at 12:14
  • 7
    … has become a very high priority for language implementors). Also, many existing parser generators / interpreters / combinators are not really designed to cope with the great variety of demands that modern language implementations must fulfill. E.g. many modern language implementations use the same piece of code for batch compilation, IDE background compilation, syntax highlighting, automated refactoring, intelligent code completion, automatic documentation generation, automatic diagramming, etc. Scala even uses the compiler for runtime reflection and its macro system. Many existing parser … Dec 22, 2016 at 12:17
  • 5
    … frameworks are not flexible enough to deal with that. Note also that there are parser frameworks that are not based on EBNF. E.g. packrat parsers for Parsing Expression Grammars. Dec 22, 2016 at 12:20
  • 2
    I think it heavily depends on the language you are trying to compile. What type is it (LR, ...)?
    – user188153
    Dec 22, 2016 at 12:26
  • 3
    … somehow a natural choice, but rather a constraint forced upon us by inferior tooling. Look at e.g. Perl6 or Fortress for some languages with powerful syntactic capabilities that cannot cleanly expressed in terms of lex/YACC derivatives. Dec 22, 2016 at 23:42

4 Answers 4


I have done a lot of research these past few days, to understand better why these separate technologies exist, and what their strengths and weaknesses are.

Some of the already-existing answers hinted at some of their differences, but they did not give the complete picture, and seemed to be somewhat opinionated, which is why this answer was written.

This exposition is long, but important. bear with me (Or if you're impatient, scroll to the end to see a flowchart).

To understand the differences between Parser Combinators and Parser Generators, one first needs to understand the difference between the various kinds of parsing that exist.


Parsing is the process of analyzing of a string of symbols according to a formal grammar. (In Computing Science,) parsing is used to be able to let a computer understand text written in a language, usually creating a parse tree that represents the written text, storing the meaning of the different written parts in each node of the tree. This parse tree can then be used for a variety of different purposes, such as translating it to another language (used in many compilers), interpreting the written instructions directly in some way (SQL, HTML), allowing tools like Linters to do their work, etc. Sometimes, a parse tree is not explicitly generated, but rather the action that should be performed at each type of node in the tree is executed directly. This increases efficiency, but underwater still an implicit parse tree exists.

Parsing is a problem that is computationally difficult. There has been over fifty years of research on this subject, but there is still much to learn.

Roughly speaking, there are four general algorithms to let a computer parse input:

  • LL parsing. (Context-free, top-down parsing.)
  • LR parsing. (Context-free, bottom-up parsing.)
  • PEG + Packrat parsing.
  • Earley Parsing.

Note that these types of parsing are very general, theoretical descriptions. There are multiple ways to implement each of these algorithms on physical machines, with different tradeoffs.

LL and LR can only look at Context-Free grammars (that is; the context around the tokens that are written is not important to understand how they are used).

PEG/Packrat parsing and Earley parsing are used a lot less: Earley-parsing is nice in that it can handle a whole lot more grammars (including those that are not necessarily Context-Free) but it is less efficient (as claimed by the dragon book (section 4.1.1); I am not sure if these claims are still accurate). Parsing Expression Grammar + Packrat-parsing is a method that is relatively efficient and can also handle more grammars than both LL and LR, but hides ambiguities, as will quickly be touched on below.

LL (Left-to-right, Leftmost derivation)

This is possibly the most natural way to think about parsing. The idea is to look at the next token in the input string and then decide which one of maybe multiple possible recursive calls should be taken to generate a tree structure.

This tree is built 'top-down', meaning that we start at the root of the tree, and travel the grammar rules in the same way as we travel through the input string. It can also be seen as constructing a 'postfix' equivalent for the 'infix' token stream that is being read.

Parsers performing LL-style parsing can be written to look very much like the original grammar that was specified. This makes it relatively easy to understand, debug and enhance them. Classical Parser Combinators are nothing more than 'lego pieces' that can be put together to build an LL-style parser.

LR (Left-to-right, Rightmost derivation)

LR parsing travels the other way, bottom-up: At each step, the top element(s) on the stack are compared to the list of grammar, to see if they could be reduced to a higher-level rule in the grammar. If not, the next token from the input stream is shift ed and placed on top of the stack.

A program is correct if at the end we end up with a single node on the stack which represents the starting rule from our grammar.


In either of these two systems, it sometimes is necessary to peek at more tokens from the input before being able to decide which choice to make. This is the (0), (1), (k) or (*)-syntax you see after the names of these two general algorithms, such as LR(1) or LL(k). k usually stands for 'as much as your grammar needs', while * usually stands for 'this parser performs backtracking', which is more powerful/easy to implement, but has a much higher memory and time usage than a parser that can just keep on parsing linearly.

Note that LR-style parsers already have many tokens on the stack when they might decide to 'look ahead', so they already have more information to dispatch on. This means that they often need less 'lookahead' than an LL-style parser for the same grammar.

LL vs. LR: Ambiguity

When reading the two descriptions above, one might wonder why LR-style parsing exists, as LL-style parsing seems a lot more natural.

However, LL-style parsing has a problem: Left Recursion.

It is very natural to write a grammar like:

expr ::= expr '+' expr | term
term ::= integer | float

But, a LL-style parser will get stuck in an infinite recursive loop when parsing this grammar: When trying out the left-most possibility of the expr rule, it recurses to this rule again without consuming any input.

There are ways to resolve this problem. The simplest is to rewrite your grammar so that this kind of recursion does not happen any more:

expr ::= term expr_rest
expr_rest ::= '+' expr | ϵ
term ::= integer | float

(Here, ϵ stands for the 'empty string')

This grammar now is right recursive. Note that it immediately is a lot more difficult to read.

In practice, left-recursion might happen indirectly with many other steps in-between. This makes it a hard problem to look out for. But trying to solve it makes your grammar harder to read.

As Section 2.5 of the Dragon Book states:

We appear to have a conflict: on the one hand we need a grammar that facilitates translation, on the other hand we need a significantly different grammar that facilitates parsing. The solution is to begin with the grammar for easy translation and carefully transform it to facilitate parsing. By eliminating the left recursion we can obtain a grammar suitable for use in a predictive recursive-descent translator.

LR-style parsers do not have the problem of this left-recursion, as they build the tree from the bottom-up. However, the mental translation of a grammar like above to an LR-style parser (which is often implemented as a Finite-State Automaton)
is very hard (and error-prone) to do, as often there are hundreds or thousands of states + state transitions to consider. This is why LR-style parsers are usually generated by a Parser Generator, which is also known as a 'compiler compiler'.

How to resolve Ambiguities

We saw two methods to resolve Left-recursion ambiguities above: 1) rewrite the syntax 2) use an LR-parser.

But there are other kinds of ambiguities which are harder to solve: What if two different rules are equally applicable at the same time?

Some common examples are:

Both LL-style and LR-style parsers have problems with these. Problems with parsing arithmetic expressions can be solved by introducing operator precedence. In a similar way, other problems like the Dangling Else can be solved, by picking one precedence behaviour and sticking with it. (In C/C++, for instance, the dangling else always belongs to the closest 'if').

Another 'solution' to this is to use Parser Expression Grammar (PEG): This is similar to the BNF-grammar used above, but in the case of an ambiguity, always 'pick the first'. Of course, this does not really 'solve' the problem, but rather hide that an ambiguity actually exists: The end users might not know which choice the parser makes, and this might lead to unexpected results.

More information that is a whole lot more in-depth than this post, including why it is impossible in general to know if your grammar does not have any ambiguities and the implications of this is the wonderful blog article LL and LR in context: Why parsing tools are hard. I can highly recommend it; it helped me a lot to understand all the things I am talking about right now.

50 years of research

But life goes on. It turned out that 'normal' LR-style parsers implemented as finite state automatons often needed thousands of states + transitions, which was a problem in program size. So, variants such as Simple LR (SLR) and LALR (Look-ahead LR) were written that combine other techniques to make the automaton smaller, reducing the disk and memory footprint of the parser programs.

Also, another way to resolve the ambiguities listed above is to use generalized techniques in which, in the case of an ambiguity, both possibilities are kept and parsed: Either one might fail to parse down the line (in which case the other possibility is the 'correct' one), as well as returning both (and in this way showing that an ambiguity exists) in the case they both are correct.

Interestingly, after the Generalized LR algorithm was described, it turned out that a similar approach could be used to implement Generalized LL parsers, which is similarly fast ( $O(n^3)$ time complexity for ambiguous grammars, $ O(n) $ for completely unambiguous grammars, albeit with more bookkeeping than a simple (LA)LR parser, which means a higher constant-factor) but again allow a parser to be written in recursive descent (top-down) style that is a lot more natural to write and debug.

Parser Combinators, Parser Generators

So, with this long exposition, we are now arriving at the core of the question:

What is the difference of Parser Combinators and Parser Generators, and when should one be used over the other?

They are really different kinds of beasts:

Parser Combinators were created because people were writing top-down parsers and realized that many of these had a lot in common.

Parser Generators were created because people were looking to build parsers that did not have the problems that LL-style parsers had (i.e. LR-style parsers), which proved very hard to do by hand. Common ones include Yacc/Bison, that implement (LA)LR).

Interestingly, nowadays the landscape is muddled somewhat:

  • It is possible to write Parser Combinators that work with the GLL algorithm, resolving the ambiguity-issues that classical LL-style parsers had, while being just as readable/understandable as all kinds of top-down parsing.

  • Parser Generators can also be written for LL-style parsers. ANTLR does exactly that, and uses other heuristics (Adaptive LL(*)) to resolve the ambiguities that classical LL-style parsers had.

In general, creating an LR parser generator and and debugging the output of an (LA)LR-style parser generator running on your grammar is difficult, because of the translation of your original grammar to the 'inside-out' LR form. On the other hand, tools like Yacc/Bison have had many years of optimisations, and seen a lot of use in the wild, which means that many people now consider it as the way to do parsing and are sceptical towards new approaches.

Which one you should use, depends on how hard your grammar is, and how fast the parser needs to be. Depending on the grammar, one of these techniques (/implementations of the different techniques) might be faster, have a smaller memory footprint, have a smaller disk footprint, or be more extensible or easier to debug than the others. Your Mileage May Vary.

Side note: On the subject of Lexical Analysis.

Lexical Analysis can be used both for Parser Combinators and Parser Generators. The idea is to have a 'dumb' parser that is very easy to implement (and therefore fast) that performs a first pass over your source code, removing for instance repeating white spaces, comments, etc, and possibly 'tokenizing' in a very coarse way the different elements that make up your language.

The main advantage is that this first step makes the real parser a lot simpler (and because of that possibly faster). The main disadvantage is that you have a separate translation step, and e.g. error reporting with line- and column numbers becomes harder because of the removal of white-space.

A lexer in the end is 'just' another parser and can be implemented using any of the techniques above. Because of its simplicity, often other techniques are used than for the main parser, and for instance extra 'lexer generators' exist.


Here is a flowchart that is applicable to most cases: enter image description here

  • 1
    @Sjoerd It indeed is a lot of text, as it turned out to be a very hard problem. If you know of a way that I can make the final paragraph clearer, I'm all ears: "Which one you should use, depends on how hard your grammar is, and how fast the parser needs to be. Depending on the grammar, one of these techniques (/implementations of the different techniques) might be faster, have a smaller memory footprint, have a smaller disk footprint, or be more extensible or easier to debug than the others. Your Mileage May Vary."
    – Qqwy
    Dec 26, 2016 at 14:12
  • 2
    The other answers are both much shorter and much clearer, and do a much better job at answering.
    – Sjoerd
    Dec 26, 2016 at 14:40
  • 6
    @Sjoerd the reason I wrote this answer was because the other answers were either oversimplifying the problem, presenting a partial answer as full answer and/or falling into the anecdotal fallacy trap. The answer above is the embodiement of the discussion Jörg W Mittag, Thomas Killian and I had in the comments of the question after understanding what they were talking about and presenting it without assuming prior knowledge.
    – Qqwy
    Dec 26, 2016 at 15:35
  • 2
    Parser Combinators do indeed fail to solve the problem when you don't actually use them. There are more combinators than just |, that's the whole point. The correct re-write for expr is the even more succinct expr = term 'sepBy' "+" (where single quotes here are substituting for backticks to turn a function infix, because mini-markdown does not have character escaping). In the more general case there is the chainBy combinator as well. I realize it is difficult to find a simple parsing task as an example that is not well suited to PCs, but thats a strong argument in their favour really. Dec 30, 2018 at 8:38
  • 1
    @Qqwy In my recent experience, LL parser combinator libraries tend to be extremely slow and produce annoying error messages, and should be avoided. There are more robust parser combinator libraries that not only handle lexing but tokenizing as well, and they look at the input stream one token at a time. In practice, you can build a parser combinator that's reasonably fast if: (1) You tokenize the input, which results in less frequent backtracking, (2) combined with the avoidance of allocations (3) indirect dispatch. The last one, indirect dispatch, depends if you're using a tracing JIT. Nov 7, 2019 at 12:50

For input that is guaranteed to be free of syntax errors, or where an overall pass/fail on syntactic correctness is okay, parser combinators are much simpler to work with, especially in functional programming languages. These are situations like programming puzzles, reading data files, etc.

The feature that makes you want to add the complexity of parser generators is error messages. You want error messages that point the user to a line and column, and hopefully are also understandable by a human. It takes a lot of code to do that properly, and the better parser generators like antlr can help you out with that.

However, automatic generation can only get you so far, and most commercial and long-lived open source compilers end up manually writing their parsers. I presume if you felt comfortable doing this, you wouldn't have asked this question, so I'd recommend going with the parser generator.

  • 4
    Thank you for your answer! Why would it be easier to build readable error-messages using a Parser Generator than a Parser Combinator? (Regardless of what implementation we are talking about, specifically) For instance, I know that both Parsec and Spirit contain functionality to print error messages including line+column information, so it seems definitely possible to do this in Parser Combinators as well.
    – Qqwy
    Dec 22, 2016 at 18:59
  • It's not that you can't print error messages with parser combinators, it's that their advantages are less evident when you throw error messages into the mix. Do a relatively complex grammar using both methods and you'll see what I mean. Dec 22, 2016 at 19:50
  • 1
    With a Parser Combinator, by definition all you can get in an error condition is "Starting at this point, no legal input was found". This doesn't really tell you what was wrong. In theory, the individual parsers called at that point could tell you what it expected and DIDN'T find, but all you could do is print all that out, making for a looooooong error message. Dec 25, 2016 at 12:00
  • 4
    Parser generators aren't exactly known for their good error messages either, to be honest.
    – mrr
    Dec 27, 2016 at 21:13
  • 1
    Not by default, no, but they have more convenient hooks for adding good error messages. Dec 27, 2016 at 22:25

Sam Harwell, one of the maintainers of the ANTLR parser generator, recently wrote:

I found [combinators] do not meet my needs:

  1. ANTLR provides me with tools for managing things like ambiguities. During development there are tools that can show me ambiguous parse results so I can eliminate those ambiguities in the grammar. At runtime I can leverage ambiguity resulting from incomplete input in the IDE to produce more accurate results in features like code completion.
  2. In practice I've found that parser combinators were not suitable for meeting my performance goals. Part of this goes back
  3. When parse results are used for features like outlining, code completion, and smart indent, it's easy for subtle changes to the grammar to impact the accuracy of those results. ANTLR provides tools that can turn these mismatches into compile errors, even in cases where the types would otherwise compile. I can with confidence prototype a new language feature that affects the grammar knowing that all the additional code that forms the IDE will provide a complete experience for the new feature from the start. My fork of ANTLR 4 (which the C# target is based on) is the only tool I know of that even attempts to provide this feature.

Essentially, parser combinators are a cool toy to play with, but they're simply not cut out for doing serious work.

  • 2
    Parsers are often needed for much simpler things than a fully fledged programming language + IDE. How often did you ever code something like this? Once? Never? 90% of the time one just has to parse that weird custom file format your company's system is supposed to import from a client. That's were parser combinators shine: when the task is too simple to justify the immense overhead of parser generators, but writing a recursive-decent parser by hand is too annoying. Feb 9, 2020 at 10:00
  • @GoodNightNerdPride "How often did you ever code something like this?" Quite frequently, in fact. ANTLR pops up all over the place for parsing custom formats that are "simpler things than a fully fledged programming language." Feb 10, 2020 at 13:28
  • 1
    that's exactly my point! Parser generators are being used "all over the place" when they shouldn't (one exemplary exception being a "fully fledged programming language + IDE"). I claim that parser combinators are quite sufficient 90% of the time when some kind of parser is needed. Yet people waste time and energy with parser generators. Feb 10, 2020 at 14:27
  • This answer is nothing more than appealing to the authority of someone who is clearly biased. Oct 7, 2022 at 16:52
  • Remember, folks, conflating conclusions with biases says more about you than it does about the person who did the research and reached the conclusion. Oct 7, 2022 at 17:16

As Karl mentions, parser generators tend to have better error reporting. In addition:

  • they tend to be faster, since the generated code can be specialized for the syntax and generate jump tables for the lookahead.
  • they tend to have better tooling, to identify ambiguous syntax, remove left recursion, fill in error branches, etc.
  • they tend to handle recursive definitions better.
  • they tend to be more robust, since generators have been along longer and do more of the boilerplate for you, reducing the chance of you screwing it up.

On the other hand, combinators' have their own advantages:

  • they are in code, so if your syntax varies at runtime, you can more easily mutate things.
  • they tend to be easier to tie into and actually consume (the output of parser generators tends to be very generic and awkward to use).
  • they are in code, so tend to be a little easier to debug when your grammar doesn't do what you expect.
  • they tend to have a shallower learning curve since they work like any other code. Parser generators tend to have their own quirks to learn to get stuff working.
  • Parser generators tend to have horrible error reporting relative to hand-written LL recursive-descent parsers used in the real world. Parser generators rarely offer state table transition hooks necessary to add great diagnostics. This is why nearly every real compiler doesn't use parser combinators or parser generators. LL recursive-decent parsers are trivial to construct, although not as "clean" PCs/PGs, they're more useful.
    – dhchdhd
    Jan 6, 2019 at 8:42

Not the answer you're looking for? Browse other questions tagged or ask your own question.