This isn't a production-grade compiler. It is merely a personal, educational haskell/llvm/compiler/static-vs-dynamic learning exercise. As such, I apologize in advance for atrocities such as mostly missing comments and/or lack of unit tests (only integration tests were written).
Do not try running this at home on your 32-bit machine! The compiler can be targeted towards 64-bit machines only. Moreover, it was only tested on a(n Arch) Linux environment.
A compiler for Crisp (Lisp/Scheme subset) in Haskell, with an LLVM backend.
I initially wanted Crisp to be inspired by Scheme, and branch off it's own path. However, due to recent time constraints, and how bringing the compiler to it's current state took more time than I've anticipated, I settled on a subset of Scheme for now.
Here's an example of implementing merge-sort in Crisp.
(define empty? (lambda(x) (= x '())))
(define list-length
(lambda (lst)
(if (empty? lst)
0
(+ 1 (list-length (cdr lst))))))
(define take
(lambda (n lst)
(if (= n 0)
'()
(cons (car lst)
(take (- n 1) (cdr lst))))))
(define drop
(lambda (n lst)
(if (= n 0)
lst
(drop (- n 1) (cdr lst)))))
(define split
(lambda (n lst)
(cons (take n lst) (drop n lst))))
(define merge-sorted
(lambda (lstA lstB)
(cond ((empty? lstA)
lstB)
((empty? lstB)
lstA)
((< (car lstA) (car lstB))
(cons (car lstA) (merge-sorted (cdr lstA) lstB)))
(else
(cons (car lstB) (merge-sorted lstA (cdr lstB)))))))
(define merge-sort
(lambda (lst)
(let ((len (list-length lst)))
(if (< len 2)
lst
(let ((halves-pair (split (/ len 2) lst)))
(merge-sorted (merge-sort (car halves-pair))
(merge-sort (cdr halves-pair))))))))
(merge-sort '(34 87 24 90 74 10 47))
Compile and execute:
$ crc -i merge-sort.crs -o merge-sort
Compiled successfully.
$ ./merge-sort
(10 24 34 47 74 87 90)
Here is what portions of the LLVM assembly for this module looks like: This is what merge-sort compiled to, not including it's generated inner-lambdas.
define i64 @initGlobals-lambda6(i64 %__env, i64 %lst) {
entry:
%0 = alloca i64
store i64 %__env, i64* %0
%1 = alloca i64
store i64 %lst, i64* %1
%2 = inttoptr i64 %__env to <{}>*
%3 = call i64 @memalign(i64 1, i64 16)
%4 = inttoptr i64 %3 to <{ i64, i64 }>*
%5 = call i64 @memalign(i64 1, i64 8)
%6 = inttoptr i64 %5 to <{ i64 }>*
%7 = load i64* %1
%8 = call i64 @memalign(i64 1, i64 8)
%9 = inttoptr i64 %8 to i64*
store i64 %7, i64* %9
%10 = getelementptr inbounds <{ i64 }>* %6, i32 0, i32 0
store i64 %8, i64* %10
%11 = getelementptr inbounds <{ i64, i64 }>* %4, i32 0, i32 0
store i64 %5, i64* %11
%12 = getelementptr inbounds <{ i64, i64 }>* %4, i32 0, i32 1
store i64 ptrtoint (i64 (i64, i64)* @initGlobals-lambda6-lambda to i64), i64* %12
%13 = inttoptr i64 %3 to <{ i64, i64 }>*
%14 = getelementptr inbounds <{ i64, i64 }>* %13, i32 0, i32 0
%15 = load i64* %14
%16 = getelementptr inbounds <{ i64, i64 }>* %13, i32 0, i32 1
%17 = load i64* %16
%18 = inttoptr i64 %17 to i64 (i64, i64)*
%19 = load i64* @list-length
%20 = inttoptr i64 %19 to <{ i64, i64 }>*
%21 = getelementptr inbounds <{ i64, i64 }>* %20, i32 0, i32 0
%22 = load i64* %21
%23 = getelementptr inbounds <{ i64, i64 }>* %20, i32 0, i32 1
%24 = load i64* %23
%25 = inttoptr i64 %24 to i64 (i64, i64)*
%26 = load i64* %9
%27 = call i64 %25(i64 %22, i64 %26)
%28 = call i64 %18(i64 %15, i64 %27)
ret i64 %28
}
Implemented features, as of this writing:
-
Crisp
- Data types
- Simple types
- numbers, chars, bools
- Functions such as: number?, char?, boolean?, +, -, *, /, =, <, <=, >, >=
- Composite types
- strings, lists, vectors
- Example literals: "A string", '(1, #t, "a list item"), #(1, #t, "a vector item")
- Functions such as: cons, car, cdr, set-car!, make-vector, vector-ref, vector-set!
- Special types
- lambdas, closures
- Example:
(let ((x 5)) (lambda (y) (+ y x)))
- Simple types
- Keywords such as: define, lambda, if, cond(else), and, or, not, let, set!
- Data types
-
Compiler
- Normal compilation mode
- REPL mode
-
Missing features
- Negative numbers, floating point numbers.
- They can be implemented without too much difficulty and wouldn't benefit me much educationally. It would be mainly fiddling with how to do this in LLVM.
- A standard library
- Helper functions such as map or accumulate, which could easily be implemented in crisp itself and linked as a standard library of sorts, are missing, as I didn't find implementing them, nor attempting to be able to compile every conceivable scheme file, conducive to this learning exercise.
- Negative numbers, floating point numbers.
The crc (crisp-compiler) is a command-line application. The options are:
-i FILEPATH --input-filepath=FILEPATH Path of input file. REQUIRED if not in repl-mode
-o FILEPATH --output-filepath=FILEPATH Path of output file. Default: ./a.out
-r --repl-mode Interactive REPL mode. Default: false
-p --print-llvm Output resulting LLVM IR. Default: false
-h --help Show help
The compiler has 2 modes of operation:
- File compilation mode (-i, -o)
- A Crisp file is passed as an input, and an output my be passed as well. (otherwise ./a.out is used)
- The file is compiled and a binary file is written in output.
- The target machine is the one running the compiler.
- REPL mode (-r)
- This works like your standard REPL (read-eval-print-loop) session.
- This mode utilizes LLVM's JIT capabilities.
- -i and -o flags will be ignored in this mode.
- Both modes allow for the resulting LLVM IR, that is a natural side product
of the compilation process, to be sent to stdout.
- Out of convenience, the LLVM IR that is printed is before optimization passes. Having said that, during a repl session the resulting LLVM IR will consist of all definitions up to that point, thus definitions from prior evaluations will be displayed optimized. However, code resulting from the latest invocation or definition will not.
- The Haskell compiler GHC is required, as some GHC extensions are used in the compiler source files.
- LLVM bindings to version 3.4.1 were used, as that is what was (still is?) available at the time of writing the bulk of crc. Make sure it is installed on your system: http://llvm.org/releases/download.html#3.4.2 It should be available in your favorite package manager. If you're using Arch Linux, do note that already at the time of this writing, pacman is offering a newer version, which isn't compatible with the llvm-haskell bindings. You can use the Arch Rollback Machine. http://seblu.net/a/arm/packages/l/llvm/llvm-3.4.1-2-x86_64.pkg.tar.xz
- (Might be superfluous but) gcc (GNU's C Compiler), as Compiling the LLVM modules results in object files which will be linked by the gcc linker (it will be invoked implicitly by the Crisp compiler).
- Although this repository already contains the neccessary primitive functions compiled as llvm-modules, if you wish to recompile them, you'd need clang as they are .c files which should reside as LLVM modules come link-time. Your best bet is getting a version which matches the LLVM's, through the link given above. And for Arch linux users: http://seblu.net/a/arm/packages/c/clang/clang-3.4.1-2-x86_64.pkg.tar.xz
- The test-suite runs a test per scheme (.scm) file in the resources/tests folder (which should reside in a different folder after you cabal installed),
- The tests compiles x.scm, runs a gcc linker over it, executes it, and asserts the result equals to the text in x.ans.
- test-cond.scm and test-cond.ans files for example:
(define empty? (lambda(x) (= x '())))
(define merge-sort (lambda (lstA lstB)
(cond ((empty? lstA) lstB)
((empty? lstB) lstA)
((< (car lstA) (car lstB)) (cons (car lstA) (merge-sort (cdr lstA) lstB)))
(else (cons (car lstB) (merge-sort lstA (cdr lstB))))
)
))
(merge-sort '(2 6 9) '(3 5 10))
(2 3 5 6 9 10)
- The HUnit tests are integration tests. And require that the binary be installed (doesn't matter if globally or 'sandbox'ically, it uses the binDir that is cabal-autogenerated).
###Option 1: Install package from hackage### Package not yet uploaded to Hackage.
###Option 2: clone from Github###
To install:
$ git clone https://github.com/talw/crisp-compiler.git
$ cd crisp-compiler
$ cabal sandbox init
$ cabal install
After which you should have a crc in .cabal-sandbox/bin/ where you 'cabal install'ed.