-*-text-*-
C2HS TODO
=========
Next: -= phase re-ordering & parser clean up =-
- Pending suggestions from Axel Simon:
+ succ/pred definitions for enum hooks
+ Compilation of multiple files in one c2hs invocation
- Pending suggestions from Armin Sander:
+ MarshalFlags.hs
- We need to handle the time stamps of .chi files more carefully;
otherwise, we often get lots of unnecessary re-runs of c2hs.
- Conceptual problem with import hooks: Currently, everything imported
from a .chi file by a binding module is also dumped into its own .chi
file. Is is not generally correct, but it would also not be correct
to omit everything that has been imported. The conceptually correct
solution would be to dump an entry for a type into a .chi file if the
corresponding Haskell type is exported by the binding module. To
achieve this, we would also need an export hook.
- the idea of the hipar/mk/*.mk file to identify which projects are
present doesn't really cut it as `cvs update' will always check these
files out; fix this!
- when we first read the .chs file and pre-process it to generate some
more C header code, it might be nice to move the reading of .chi files
into the pre-processing phase, too
- GenBind.mergePtrMap could be improved (see the comment)
- Idea from Alastair: To make life easier for people who want to
distribute tools including c2hs-generated code, it would be good to
have an option to delay the inclusion of system-dependent information
until the generated files are compiled (possibly generating files that
have to be run through cpp or sed or so). With the New FFI Libraries
this shouldn't be too difficult, because most of the autoconf
information is to build c2hs and doesn't have an impact on the
generated code. The obvious exception is the use of the sizes and
alignments of primitive C types bei c2hs. So, here we should Haskell
expression computing these values from Storable rather than constants
computed by c2hs.
- {#enum ...#} should create an instance of Storable as follows:
instance Storable MyEnum where
sizeOf _ = sizeOf (undefined :: CInt)
alignment _ = alignment (undefined :: CInt)
peek p = liftM cToEnum $ peek (castPtr p :: Ptr CInt)
poke p v = poke (castPtr p :: Ptr CInt) (cFromEnum v)
- Vorschlag von Axel Krauth <krauth@infosun.fmi.uni-passau.de>:
An option with which c2hs prints a list of all functions that have not
been bound in a header file. The Position info schould be sufficient
to determine this. Maybe also add {#ignore foo#}, which makes it not
list a given identifier (for private functions listed in the header).
- Simon Marlow's bug (+inbox/37148)
- add lndir to hipar/ (start utils dir)
- In `GtkCList.cListGetNoOfRows' wird der Offset falsch berechnet
(sollte 84 sein)
- make clean and make cleanhi don't clean the lib/ subdir
- In `GdkGL.chs', the prefix is not properly removed from the
constructors of `GdkGL.Configs'; same problem with
`GdkEvents.EventType'
- maybe we should replace the keyword `fun' by `pure' (but retain
`fun' as a deprecated keyword for a while to avoid breaking code;
maybe emit a warning)
- install the executable (c2hs and c2hs-config) with a version number
suffix and make a symbolic link for the name without version number
(makes installing multiple versions easier)
- what about the bug re cyclic structures in nhc with the new version
1.0? supposedly, fixed
- the import path (/usr/lib/c2hs-0.7.5/ghc4/import) has to be more
specific and has to also include the minor version number of ghc
(as interface file formats change with minor versions :-(
- #define enum's (see below) - also requires reordering of phases
- Add an optional header="..." option in context hooks (can be used
instead of giving the .h file on the command line)?
STARTED: `header' tag is recognised in context hooks, BUT not yet
used - usage requires reordering the phases
- intro a safe flag (as opposite to unsafe, but make safe the default)
- overload stdAddr etc for ForeignObj (stable ptr & cast)
- All the C2HS binding hook keyword, such as `type', cannot be used in
access paths of get/set hooks! (See `gdk/GdkVisual.chs'.)
- C2HSMarsh: from qrczak@knm.org.pl (Marcin 'Qrczak' Kowalczyk)
mallocForeignObj:: Int -> IO ForeignObj -- malloc + freeing finalizer
- C2HSMarsh:
(!!!) :: FromAddr a => Ptr a -> Int -> IO a
p !!! i = do
v <- fixIO $ deref (p `plusAddr` sizeof v * i)
return v
(Use in GdkColor, then)
Part of New FFI, so use that in GdkColor!
- C2HSMarhs:
for compound structures it is inconvenient that in a pointer-based
out, an initial value is needed; thus
out :: (Storable a, FromAddr a) => a -> Marsh a Addr
out x = malloc x :> addrStd
and this is also convenient
inp :: ToAddr a => a -> Marsh () Addr
inp x = stdAddr x :> free
Axel also proposes this
byValue :: b -> Marsh () b
byValue x = use x :> forget
- C2HSMarsh:
- should re-export unsafePerformIO
- define marsh4 and marsh5
- Axel suggests,
2b. Automatisches Marshalling von Strukturen.
Eigentlich müßte es doch möglich sein, automatisch Instanzen von
Storable für Strukturen zu erstellen. Falls man mal einen generische
Pointer hat, kann man ja mit der ... as ... Methode eigene
Marshalling Funktionen zur Verfügung stellen. Oder habe ich da etwas
grundsätzliches übersehen?
- there are still optimisation related !!! in Parsers.hs
- ghttp auskoppeln; extra C->HS library page
- what about supporting producing bindings for nhc? An added problem
here is that the library `C2HS' has to be compiled with the compiler
that is target with the bindings. To possible options: (1) Distribute
the source of `C2HS' even in C->HS binary releases and support
compiling the library source with other Haskell systems. (2) Add a
`nhc' library-only compilation mode that requires to install the C->HS
binary for GHC first, then get the complete C->HS source, and the
special library-only compile and installation.
- maybe compilers based on the CTK should add a README.1ST to the root
directory when they are tarified
- kill the --old-ffi option when the old FFI disappeared
Marshalling templates
~~~~~~~~~~~~~~~~~~~~~
* the generic version for arguments is
{#call bar {inExpr >>> outExpr}#}
=>
do {
arg <- inExpr;
res <- bar arg;
argres <- outExpr arg;
return (res, argres)
}
we may optionally allow the tag "inout" before the argument description
* special versions:
in{e} => {e >>> free}
pure{e} => {return e >>> return . const ()}
out{e} => {malloc >>>
\arg -> do{res <- liftM e $ peek arg; free arg; return res}}
* example:
{#call foo in{newCString str} pure{cFromEnum kind} out{cToEnum}#}
=>
do {
arg1 <- newCString str;
let arg2 = cFromEnum kind;
arg3 <- malloc :: Ptr CInt;
res <- foo arg1 arg2 arg3;
free arg1;
arg3res <- liftM cToEnum $ peek arg3;
free arg3;
return (res, arg3res)
}
Short term
~~~~~~~~~~
* Explain the grabbing of cpp -I options from -cppopts= (aka -C) values better
in the docu and add something like:
I prefer that to giving an option to c2hs and passing it to
cpp, because - as you say - you usually already have a
variable with the cpp options in your makefile and this way
you can easily reuse it.
* #define enums in C: siehe unten - scheinen aber wichtig
* `GdkMarsh.GFlag' should be in the C2HS library (as `Flag') and we might want
Something like {#enum flag ...#} for marshalling masks like `GdkEventMask'.
The latter would not generate an `Enum' instance, but a `Flag' instance.
(See also `GdkEvents.EventMask'.)
* line directives for the Haskell compiler (emit only after chunks of
C->HS generated code); there should be an option to disable this
* In pointer hooks after a `->', we currently allow only type identifiers;
other forms of Haskell types would be nice, too (especially, `()').
* A function prototype that uses a defined type on its left hand side may
declare a function, while that is not obvious from the declaration itself
(without also considering the `typedef'). This is not understood by
`GenBind' so far.
* How about using `ForeignObj's instead of `Addr's whenever passing an address
to C? This is not desirable, for example, in case of an `Addr' obtained by
a foreign export dynamic. Maybe allow to specify that in C types pointer to
certain types mean `ForeignObj's (or default `ForeignObj's and some mean
Addr).
Actually, the argument in
file:/usr/doc/ghc-4.02/docs/libraries/libs-16.html#ss16.6
for not using an `Addr' shouldn't apply to FFI calls. If passing the `Addr'
to C land is the last action done with a `ForeignObj' in Haskell, then the
finaliser will still be called - so, it doesn't make any difference if we
had a function to get the `Addr' out of the `ForeignObj' in Haskell and
would pass this to C. ** Last statement isn't true anymore since the
special rule concerning the lifetime of ForeignObjs **
We definitely want Haskell side ForeignObj, MutableArrays, etc to Addr casts.
* Why is a stable name an `unsigned long' in C land? How about a `void *' and
being able to get the `Addr' also in Haskell land?
* if in a binding file erroneously `t->m' is used instead of `t.m' and `t' is
the tag of a struct, the error message just complains that there is no type
object for called `t'; it would be more user friendly to report in addition
that the tag `t' exists, but cannot be used in this expression
* improve C parser in c/
The space leak got smaller, but there is still too much heap used during
parsing.
* Should it be possible to specify a calling convention in context hook?
Isn't it a mitake to have explicit calling convention in the FFI? Shouldn't
that be adapted automatically depending on the target architecture?
* Tutorial with ideas about conventions for binding libraries (naming
conventions etc)
* Sven: How about exporting Haskell functions (also dynamic export)?
- callback registration function that explicitly have the type of the
callback function (from which we might want to generate a foreign export
dynamic).
* `mapM raise errs' in `lexC' increases the heap usage by a factor of _8_ when
running the lexer alone on `gtkext.h' - probably because the whole analysis
has to be completed before we can be sure to have no error (and this kills
an interlocked produced/consumer scheme for the lexer and whatever function
is consuming the tokens).
* sizeof and type hooks do only allow defined, but not basic types as
arguments. This is not a real problem as C2HS provides all the needed
information to circumvent such usage, but it would still be more elegant to
support basic types as arguments, too.
Side issues:
------------
* FFI: How can structs as function results be realised?
* FFI: If the FFI for C functions returning a structure would return a pointer
to that structure, we could handle them without extra impedance code in C.
The problem is of course, where to store the structure (and how to free it).
Tip from Alastair:
If you compile a C file (with gcc) with debugging on
and use
objdump --debugging main.o
to examine it, it prints out the size and offset of the fields of a
struct.
Middle term
~~~~~~~~~~~
* A hook {#const <C expr>#} would be nice (as in hsc2hs). Unfortunately, it
is not that easy to realise. We need to parse the <C expr> in the binding
module. Moreover, the main value of this would be when the <C expr> is
put into the C field together with `enum define' stuff, so that all
pre-processor symbols in it are resolved. However, this can easily give us
C compiler errors. A cheap way out would be {#const "C expr"#} to avoid
parsing the C expression, but then we can get even more errors during
compiling the C.
* Directories contained in `C_INCLUDE_PATH' should be searched for header
files after directories extracted from the -I option in cpp options, but
before the standard header file directories are searched.
* C->HS might get a lot easier to use by providing as optional marshalling
libraries modules that handle often occurring standard stuff like converting
`time_t' to `CalendarTime' or handle sockets etc. We would, then, probably
like to have a matching Posix or so library.
------------ Pre-1.0 rewrite ---------------
Problems:
* There are some implicit requirements on the position of binding hooks, which
the tool doesn't really enforce: context must be first, context and enum may
only occur where a toplevel definition is allowed. (This should be checked
before `GenBind' is used.) We can not really check that a hook is in a
position where a toplevel definition is allowed (without analysing
significant parts of the Haskell code), but we can at least guarantee that
these hooks occur in column position 0.
* Sven: #define enums in C: Introduce (#enum define SomeEnum {...}#) hooks
that collect `#define' symbols into an enumeration type; see also
+haskell/4025. One probem: If identifiers with the same lexeme as
`SomeEnum' or the enum members are already defined in the C header, we might
get conflicts.
The problem of this approach is that if the macro expands to something that
is not a constant expression C, we will get error messages from the
preprocessor, which are strange to the user.
Further idea, Michael's: extensible enums
There are, however, a number of interesting options supported by gcc -E that
might make alternative solutions to the problem feasible.
Approach:
* Split `CHS' into a more conventional `CHSAST', `CHSSyntax', and `CHSAnalysis'
(or similar) structure and with `CHSAnalysis' add a pass that goes over the
binding file before `GenBind' is used. Tasks: Static semantic checks
(context hook is first etc), collect all enum-define hooks etc
* .chs has to be read before .h (due to the enum-define hooks)
* Generating a header file that includes the bound to header and running this
through cpp has the added advantage that we don't have to give a path for
headers that are in any of the standard paths
END of ----- Pre-1.0 rewrite ---------------
* We could specify which header files matches a Haskell binding module in a
hook (context hook?) in the binding file. We could even include some
verbatim C code, instead of having extra C files.
* Do we like direct support for mapping complete structs (if they are
sufficiently wellbehaved?) into Haskell data structures - both by generating
the Haskell data type definitions and by generating a `cFrom<Struct>' and a
`cTo<Struct>' routine. The latter would be generated as a cascade of field
hooks. H/Direct's formal definition of structure marshalling might be
helpful here.
* Support for evaluating constants is not complete yet. In this context, it
should probably also be checked, when there are two overlapping tags in an
enum (this is allowed in C, but is problematic for marshaling).
* How about Hugs support? Is it already possible with current Hugs?
Release Checklist
~~~~~~~~~~~~~~~~~
(1) In root of working directory,
% make tar-c2hs
(2) Compile the resulting source distribution with latest ghc stable release:
% tar xzf c2hs-x.y.z.tar.gz
% cd c2hs-x.y.z
% ./configure
% make
(3) Install and regression testing:
- Installation procedure
% make install
- Tests in build/ghc?/c2hs/tests/ directory
- Build libghttp example
- Build Gtk+HS
(4) Check documentation and add release notes
(5) Extended build test: build with older stable release of ghc and with the
cvs version
(6) Register CVS tags for CTK and C->HS (syntax: Release-c2hs_x_y_z)
(7) Make newest `tar' and build rpm
(8) Put tar.gz and rpm sources and binaries up on Web page
(9) Optionally also release the current version of CTK
(10) Update the C2HS library files under the Web page's lib/ directory
(11) Announce: haskell@haskell.org, freshmeat.net
