This repository contains an implementation of an extended subset of Haskell. It uses combinators for the runtime execution.
The runtime system has minimal dependencies, and can be compiled even for micro-controllers.
The boards/
directory contains some samples, e.g., some sample code for an STM32F407G-DISC1 board.
The compiler can compile itself.
You can find my presentation from the Haskell Symposium 2024, video.
There is also a short paper in doc/hs2024.pdf
.
There are two different ways to compile MicroHs:
- Using GHC.
Makefile
targetbin/gmhs
- Using the included combinator file and runtime.
Makefile
targetbin/mhs
These different ways of compiling need slightly different imports etc.
This happens by GHC looking in the ghc/
subdirectory first for any extras/overrides.
Compiling MicroHs is really best done using make
, but there is also a MicroHs.cabal
file
for use with cabal
/mcabal
. This only builds what corresponds to the first target.
Doing cabal install
will install the compiler.
Note that mhs
built with ghc does not have all the functionality.
Also note that there is no need to have a Haskell compiler to run MicroHs. All you need is a C compiler, and MicroHs can bootstrap, given the included combinator file.
To install mhs
use make install
. This will install mhs
in ~/.mcabal
in the same
way as mcabal
(MicroCabal) would have. It will install a compiler binary and a compiled base package.
You will have to add ~/.mcabal/bin
to your PATH
.
Alternatively, to install mhs
use make oldinstall
. By default this copies the files to /usr/local
,
but this can be overridden by make PREFIX=dir oldinstall
.
You also need to set the environment variable MHSDIR
.
To compile on Windows make sure cl
is in the path, and then use nmake
with Makefile.windows
.
The compiler can also be used with emscripten to produce JavaScript/WASM, see Makefile.emscripten
.
The language is an extended subset of Haskell-2010.
Differences:
- There is only deriving for
Bounded
,Enum
,Eq
,Ord
,Show
, andTypeable
. - Kind variables need an explicit
forall
. - Always enabled extension:
- BangPatterns
- ConstraintKinds
- DefaultSignatures
- DoAndIfThenElse
- DuplicateRecordFields
- EmptyDataDecls
- ExistentialQuantification
- ExtendedDefaultRules
- FlexibleContexts
- FlexibleInstance
- ForeignFunctionInterface
- FunctionalDependencies
- GADTs
- GADTsyntax
- ImportShadowing
- ImportQualifiedPost
- IncoherentInstances
- KindSignatures
- LambdaCase
- MonoLocalBinds
- MultilineStrings
- MultiParamTypeClasses
- MultiWayIf
- NamedDefaults
- NamedFieldPuns
- NegativeLiterals
- NoMonomorphismRestriction
- NoStarIsType
- OrPatterns
- OverlappingInstances
- OverloadedRecordDot
- OverloadedRecordUpdate
- OverloadedStrings
- PolyKinds
- RankNTypes
- RecordWildCards
- QualifiedDo
- ScopedTypeVariables
- StandaloneKindSignatures
- TupleSections (only pairs right now)
- TypeLits
- TypeSynonymInstances
- UndecidableInstances
- UndecidableSuperClasses
- ViewPatterns
main
in the top module given tomhs
serves at the program entry point.- Many things that should be an error (but which are mostly harmless) are not reported.
- Text file I/O uses UTF8, but the source code does not allow Unicode.
- The
BangPatterns
extension is parsed, but only effective at the a top levellet
/where
. - More differences that I don't remember right now.
Mutually recursive modules are allowed the same way as with GHC, using .hs-boot
files.
The file Example.hs
contains the following:
module Example(main) where
fac :: Int -> Int
fac 0 = 1
fac n = n * fac(n-1)
main :: IO ()
main = do
let rs = map fac [1,2,3,10]
putStrLn "Some factorials"
print rs
First, make sure the compiler is built by doing make
.
Then compile the file by bin/mhs Example -oEx
which produces Ex
.
Finally, run the binary file by ./Ex
.
This should produce
Some factorials
[1,2,6,3628800]
The Prelude
contains the functions from the Haskell Report and a few extensions,
with the notable exception that Foldable
and Traversable
are not part of the Prelude
.
They can be imported separately, though.
There are some primitive data types, e.g Int
, IO
, Ptr
, and Double
.
These are known by the runtime system and various primitive operations work on them.
The function type, ->
, is (of course) also built in.
All other types are defined with the language. They are converted to lambda terms using an encoding. For types with few constructors (< 5) it uses Scott encoding, otherwise it is a pair with an integer tag and a tuple (Scott encoded) with all arguments. The runtime system knows how lists and booleans are encoded.
The compiler is written in Micro Haskell.
It takes a name of a module (or a file name) and compiles to a target (see below).
This module should contain the function main
of type IO ()
and
it will be the entry point to the program.
--version
show version number-i
set module search path to empty-iDIR
appendDIR
to module search path-oFILE
output file. If theFILE
ends in.comb
it will produce a textual combinator file. IfFILE
ends in.c
it will produce a C file with the combinators. For all otherFILE
it will compile the combinators together with the runtime system to produce a regular executable.-r
run directly-v
be more verbose, flag can be repeated-CW
write compilation cache to.mhscache
at the end of compilation-CR
read compilation cache from.mhscache
at the start of compilation-C
short for-CW
and-CR
-T
generate dynamic function usage statistics-z
compress combinator code generated in the.c
file-l
show every time a module is loaded-XCPP
runcpphs
on source files-Dxxx
passed tocpphs
-Ixxx
passed tocpphs
-tTARGET
select target-a
set package search path to empty-aDIR
prependDIR
to package search path-PPKG
create packagePKG
-LFILE
list all modules in a package-Q FILE [DIR]
install package--
marks end of compiler arguments
With the -v
flag the processing time for each module is reported.
E.g.
importing done MicroHs.Exp, 284ms (91 + 193)
which means that processing the module MicroHs.Exp
took 284ms,
with parsing taking 91ms and typecheck&desugar taking 193ms.
With the -C
flag the compiler writes out its internal cache of compiled modules to the file .mhscache
at the end of compilation. At startup it reads this file if it exists, and then validates the contents
by an MD5 checksum for all the files in the cache.
This can make compilation much faster since the compiler will not parse and typecheck a module if it is in
the cache.
Do NOT use -C
when you are changing the compiler itself; if the cached data types change the compiler will probably just crash.
MHSDIR
the directory wherelib/
andsrc/
are expected to be. Defaults to./
.MHSCC
command use to compile C file to produce binaries. Look at the source for more information.MHSCPPHS
command to use with-XCPP
flag. Defaults tocpphs
.MHSCONF
which runtime to use, defaults tounix-32/64
depending on your host's word size
Abstract
, combinator bracket abstraction and optimization.Compile
, top level compiler. Maintains a cache of already compiled modules.CompileCache
, cache for compiled modules.Deriving
, do deriving for various type classes.Desugar
, desugar full expressions to simple expressions.EncodeData
, data type encoding.Exp
, simple expression type.ExpPrint
, serializeExp
for the runtime system.Expr
, parsed expression type.FFI
, generate C wrappers for FFI.Fixity
, resolve operator fixities.Flags
, compiler flags.Graph
, strongly connected component algorithm.Ident
, identifiers and related types.IdentMap
, map from identifiers to something.Interactive
, top level for the interactive REPL.Lex
, lexical analysis and indentation processing.Main
, the main module. Decodes flags, compiles, and writes result.MakeCArray
, generate a C version of the combinator file.Parse
, parse and build and abstract syntax tree.StateIO
, state + IO monad.SymTab
, symbol table manipulation.TCMonad
, type checking monad.Translate
, convert an expression tree to its value.TypeCheck
, type checker.
If no module name is given the compiler enters interactive mode.
You can enter expressions to be evaluated, or top level definitions (including import
).
Simple line editing is available.
All definitions are saved in the file Interactive.hs
and all input
lines as saved in .mhsi
. The latter file is read on startup so
the command history is persisted.
Available commands:
:quit
Quit the interactive system:clear
Get back to start state:del STR
Delete all definitions that begin withSTR
:reload
Reload all modules:type EXPR
Show type ofEXPR
:kind TYPE
Show kind ofTYPE
expr
Evaluate expression.defn
Add definition (can also be animport
)
When mhs
is built, targets.conf is generated. It will look something like this:
[default]
cc = "cc"
conf = "unix-64"
You can add other targets to this file, changing which compiler command is used and which runtime is
selected and then use the -t
argument to select which target you would like.
To avoid compiling everything from source all the time there is a notion of a precompiled package.
A package is simply a set of modules that are compiled together and that can then be installed in
a known place (typically ~/.mcabal/mhs-VERSION/packages/
).
Packages can depend on already installed packages.
There is a search path for installed packages, controlled by the -a
flag.
To compile a package use the command mhs -Ppackage-name.pkg modules...
where modules...
are all the modules you wish to expose from the package. If other modules are needed they will
automatically be included in the package.
You typically also want to use the -o
flag to give the package a sensible name.
To install a package use the command mhs -Q package-name.pkg [install-dir]
.
If the install-dir
is left out the package is installed in the default place.
There is no need for any extra flags to mhs
to use installed packages, they are all visible at all times.
When compiling to a binary only the used parts of a package will be included in the binary.
A (maybe) short-coming of the package system is that there can only be one version of a
package installed at a time. If you need multiple version, you have to use different directories for them
and use -a
to control it. There is absolutely no checks for consistency among packages.
There is also no compatibility between packages compiled with different versions of the compiler.
There is a number of subdirectories:
Tools/
a few useful tools for compressions etc.bin/
executables are put heregenerated/
this contains the (machine generated) combinator file for the compiler.lib/
this contains thePrelude
and other base library file.src/MicroHs/
the compiler sourcesrc/runtime/
the runtime sourcetests/
some tests
The runtime system is written in C and is in src/runtime/eval.c
.
It uses combinators for handling variables, and has primitive operations
for built in types and for executing IO operations.
There is a also a simple mark-scan garbage collector.
The runtime system is written in a reasonably portable C code.
Runtime flags are given between the flags +RTS
and -RTS
.
Between those the runtime decodes the flags, everything else is available to
the running program.
-HSIZE
set heap size toSIZE
cells, can be suffixed byk
,M
, orG
, default is50M
-KSIZE
set stack size toSIZE
entries, can be suffixed byk
,M
, orG
, default is100k
-rFILE
read combinators fromFILE
, instead ofout.comb
-v
be more verbose, flag can be repeated
For example, bin/mhseval +RTS -H1M -v -RTS hello
runs out.comb
and the program gets the argument hello
,
whereas the runtime system sets the heap to 1M cells and is verbose.
MicroHs supports calling C functions.
When running the program directly (using -r
) or when generating a .comb
file only the functions in the table built
into src/runtime/eval.c
can be used. When generating a .c
file or an executable any C function can be called.
MicroHs implements the record dot extensions.
So accessing a field a
in record r
is written r.a
, as well as the usual a r
.
The former is overloaded and can access any a
field, whereas the latter is the usual monomorphic field selector.
Updating a field has the usual Haskell syntax r{ a = e }
, but the type is overloaded so this can update the a
field in any record.
The typeclasses HasField
and SetField
capture this.
HasField "name" rec ty
expresses that the record type rec
has a field name
with type ty
that can be extracted with getField
.
SetField "name" rec ty
expresses that the record type rec
has a field name
with type ty
that can be set setField
.
Record updates can also update nested fields, e.g., r{ a.b.c = e }
. Note that this will not easily work in GHC, since GHC does not
fully implement OverloadedRecordUpdate
. When GHC decides how to do it, MicroHs will follow suit.
Note that record updates cannot change the type of polymorphic fields.
The runtime system can serialize and deserialize any expression
and keep its graph structure (sharing and cycles).
The only exceptions to this are C pointers (e.g., file handles), which cannot be serialized (except for stdin
, stdout
, and stderr
).
Memory allocation is based on cells. Each cell has room for two pointers (i.e., two words, typically 16 bytes), so it can represent an application node. One bit is used to indicate if the cell is an application or something else. If it is something else one word is a tag indicating what it is, e.g., a combinator or an integer. The second word is then used to store any payload, e.g., the number itself for an integer node.
Memory allocation has a bitmap with one bit per cell. Allocating a cell consists of finding the next free cell using the bitmap, and then marking it as used. The garbage collector first clears the bitmap and then (recursively) marks every used cell in the bitmap. There is no explicit scan phase since that is baked into the allocation. Allocation is fast assuming the CPU has some kind of FindFirstSet instruction.
It is possible to use smaller cells by using 32 bit "pointers" instead of 64 bit pointers. This has a performance penalty, though.
The C code for the evaluator does not use any special features, and should be portable to many platforms. It has mostly been tested with MacOS and Linux, and somewhat with Windows.
The code has mostly been tested on 64 bit platforms, so again, there are lurking problems with other word sizes, but they should be easy to fix.
The src/runtime/
directory contains configuration files for different platform.
Use the appropriate src/runtime/eval-
platform.c
.
The compiler can compile itself. To replace bin/mhs
with a new version,
do make bootstrap
. This will recompile the compiler twice and compare
the outputs to make sure the new compiler still works.
Sadly, compiling a lot of Haskell packages needs the C preprocessor.
To this end, the distribution contains the combinator code for cpphs
.
Doing make bin/cpphs
will create the binary for the preprocessor.
To bootstrap cpphs
you can do make bootstrapcpphs
.
This assumes that you have git
to download the needed packages.
At the moment, the downloaded packages are forks of the original to
make it compile with mhs
.
-
- Q: When will it get insert feature?
- A: Maybe some time, maybe never.
-
- Q: Why are the error messages so bad?
- A: Error messages are boring.
-
- Q: Why is the so much source code?
- A: I wonder this myself. 7000+ lines of Haskell seems excessive. 2500+ lines of C is also more than I'd like for such a simple system.
-
- Q: Why are the binaries so big?
- A: The combinator file is rather verbose. The combinator file for the compiler shrinks from 350kB to 75kB when compressed with upx. The evaluator alone is about 70kB (26kB compressed with upx).