Self hosted C – breakdown
achacompilers.blogspot.com
achacompilers.blogspot.com
I notice you're using a recursive-descent parser, which I think is a good idea because it can actually reduce complexity over one made with parser-generators; in fact it's simple enough that even beginners who have just grasped recursion should be able to understand and write one. Once you have a basic parser that can evaluate arithmetic expressions, extending that to generating AST nodes and parse a (mostly) simple language like C is not so difficult. If I remember correctly, at least one college has an expression evaluator as a first-year CS assignment related to recursion. It's interesting that one of the sections of K&R, on reading and writing C declarations, even contains a simple parser for them --- I'm not aware of any introductory books for other languages which contain small pieces of their implementation.
That said, you can use the "precedence climbing" technique[1] to collapse many of the recursive functions for most of the expression operators into one function with a succinct loop and a table, making your parser even simpler and faster. C4[2] is an example of this technique but with a series of if-else in the body of the function instead of a table, and I think is worth some careful study just for its mindblowingly awesome ridiculous simplicity.
[1] https://www.engr.mun.ca/~theo/Misc/exp_parsing.htm#climbing
I feel like two of the very few contenders for this might be books on Forth or Lisp, though still maybe not an "introductory" book.
This is a great tutorial on expression parsing using Pratt Parsers by Bob Nystrom:
http://journal.stuffwithstuff.com/2011/03/19/pratt-parsers-e...
foo * bar;
(There's a really good article which I've been trying to find which goes into details about all the weird parsing edge cases due to the rules about exactly when a word goes into the symbol table which I've been completely unable to find...)
A language whose semantics you have to understand to understand its syntax is outside of the entire formal language realm inside which we formally recognize a "context-free" category. That formal language realm is purely syntactic: the terminal symbols are just "dumb" and stand for themselves. They are subject to rules which indicate how those symbols are combined.
C is "quasi context free" in that if we resolve what the raw tokens mean (which one is a type and which one isn't), then the resulting language is context-free. In fact stronger than context free: amenable not only to LALR(1) parsing approaches but to recursive descent, with a suitable refactoring of the left-recursive rules given in the standard.
> the ambiguity rests in not knowing what is the lexical category of foo in "foo * bar".
Right, so there are 2 possible parse trees. Being able to disambiguate them with more information doesn't mean it's not context-free. Given just a CFG, you were still able to determine that you have a syntactically valid C program.
> A language whose semantics you have to understand to understand its syntax is outside of the entire formal language realm inside which we formally recognize a "context-free" category.
A formal language is just the set of strings that belong to it. That's it. An absurd example might be "the set of strings corresponding to valid C programs that do not terminate". To recognize such a language, you would not only have to understand the semantics of C, you would even need to solve the halting problem, but that doesn't make it "outside of the entire formal language realm".
Not with the same symbols, though.
> A formal language is just the set of strings that belong to it. That's it.
That is correct, and in that framework, we cannot reason about the meaning of "foo"; that if it's a type apply this rule otherwise that rule.
If we do that reasoning, what we're really doing is replacing "foo" with one of two other symbols, and then parsing those; then we have different strings (so of course the parse trees cannot be the same, no matter what).
Statement -> Type Op Var ";"
Statement -> Var Op Var ";"
Op -> "*"
If C were parsed this way, nondeterministically, ultimately the ambiguity would be resolved by looking up the type info anyway. (The interesting possibility exists, though, that in some cases the type info could be inferred, based on how the declared identifier is used.)
That mapping requires semantic information (type declarations), but that information is something the compiler must maintain anyway so that it can type check the language.
Yes, if we simply treat "foo" as an IDENTIFIER token and feed that to the parser, thing swill be difficult. But we don't; what we return to the parser is a TYPESPEC token, based on peering into the current parse-time lexical environment where we resolve "foo" to a typedef name. The parser then quite easily goes through the state transitions corresponding to its declaration rule.
Sure, a parser which maintains parse-time lexical environments which are also visible to the lexer is slightly harder to write than one which doesn't have such stuff (and just returns an unadorned AST which you can walk again to build that info).
But if the year is 1973, you would likely do that anyway even if the language doesn't require it. Why? Because building the compile-time environment tree reduces the number of passes: from a single pass through the raw code, you already have the syntax tree with type info (and possibly type info that has even been checked).
not necessary. you can add a bit of information to each identifier token: is this the name of a type or not?
you have to split up one of the parser passes, so if you had a two pass "tokenizer" and "LR parser" compiler now you have a tokenizer pass followed by two LR parser passes. The first LR pass is "last mile tokenization". I guess another word would be a (very short) cascade tokenizer/parser.
The intermediate pass adds a piece of metadata to each token marking it as the name of a type or not (this makes the language actually context free instead of just nearly context free). If you do this, then the second pass doesn't see this as two identifiers at all because it knows which identifiers refer to types and which do not.
It's a different application (computing scope data instead of type data) but I've implemented the basic concept of annotating token data with the result of an intermediate parsing stage here...check out line 318 here:
There seem to be a lot of very good tutorials on writing a simple compiler, like http://scheme2006.cs.uchicago.edu/11-ghuloum.pdf http://compilers.iecc.com/crenshaw/
There are also readable code bases like tinyc or this one.
However these always use a simple template based approach to code generation and stop there.
Of course there are real world compiler (lvm, gcc, ...). However because of their size and complexity it's hard to learn from them.
What I am interested in would be going from: "I have an AST and for this node type I output this assembly." to "I translate the AST to an IR. Transform an the IR in multiple passes and finally output decent asm."
That is, something that talks about CFG analysis, register scheduling or maybe SSA form? I have read papers on SSA and Sea-of-Nodes IRs, but these of course always assume that you know how and why to use them. A more approachable text/github-repo that shows how to take the step from template base code gen to more advanced techniques would be great.
(I hope it is ok to ask this here. Did not want to hijack the thread.)
For LALR Parsers and DFA scanners and the like, I found the original Dragon Book http://www.amazon.co.uk/dp/B00EKYURHS is still best.
Both books are very accessible reading.
It is currently closed source but it shouldn't be hard for my compiler to target his IR in the future. His backend is currently around 6k lines of code so should be relatively easy to understand, though understanding some of the concepts is harder overall.
Corresponding lecture notes: https://www.clear.rice.edu/comp412/Lectures.
Another space you can look for interesting things (not necessarily optimising compilation) is the compilation of functional programming languages. They usually take a slightly different approach compared to imperative languages, though there's clearly a lot of overlap. I really enjoy reading in an old book called The Implementation of Functional Programming Languages from 1987, which is available online: http://research.microsoft.com/en-us/um/people/simonpj/papers...
I don't really know about that many other real, yet simple(ish) systems. Maybe Pharo Smalltalk, luajit, pypy, gnu guile (latest version)?
It's a small (few thousand lines), rather neat, weakly optimising compiler for a subset of OCaml, written in OCaml. It has more than one backend, a few optimisations, a register allocator. No SSA form or 'advanced' stuff though.
- Gcc is unreadable;
- Clang is "advanced", but not readable;
- yours and tcc are clean... But you know it's always a risk to compare yourself to Fabrice Bellard :)
edit:
Compiling my parse.c file, tcc is 34 times faster than gcc -O0 and my compiler is 3-4 times faster than gcc -O0. I should make a post with the actual stats.
(It won.)
pcc is simple, but very old school and a bit contorted.
libfirm is small and clean and generates excellent code. (Despite the name, it's a complete ANSI C compiler, as well as being a LLVM-light.)
The ACK is easy to port and is a complete standalone toolchain including compilers, assemblers and linkers, but (at least on modern processors) produces code so bad that it will make you want to stab yourself in the eyes to make the pain go away.
There's a look-but-don't-touch licensed compiler called vbcc which is by far the best C compiler I've ever seen; small, easy to understand, easy to port, and produces excellent code... but the author doesn't want to release it under an open source license.
Any others?
That said, in practice a large amount of C code out there is not pure standard C so if you want to be able to compile something "interesting" like e.g. a Linux kernel you'll need to implement some.
nwcc: small, slightly unfinished, looks like it's stalled.
sparse: technically this isn't a compiler, it's a code linter --- but it contains a complete ANSI C front end and a reasonable amount of a backend and can be made to generate code without much effort. I've used it to compile C to Java and Perl. There's no optimisation layer and it's a little buggy (it's not really really intended as a compiler).
I disagree; "brevity" and "readability" are unfortunately a tradeoff, and IMO the author's version optimizing the wrong metric.
n->For.init = i; n->For.step = st; ...
This is why it is clear. Though I could change it to be better.
Of course I know clang handles more, it is also designed as a library for other tools. The fact that clang mixes C and C++ parsing is not a merit, more like a necessary evil. This doesn't contradict my point on code clarity at all.
Probably not exactly what you're asking for, but hey, free book.
Only with difficulty, because your compiler is written in a language that is poorly suited for writing compilers.
How about using nice high level language whose implementation is partially written in C to write a C compiler in that nice high level language which then compiles that partially-written-in-C component of the implementation of that high level language.