Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/class-{finalize,layout}-impl.lisp: Error checking on layout slots.
The layout slots which are autovivifying (`effective-methods',
`%ilayout', and `vtables') shouldn't be accessed prior to class
finalization, so gather their definitions together and check that this
is done properly.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/class-finalize-impl.lisp: Remove FIXME which was fixed ages ago.
The slots which were marked for autovivifying in fact have been for five
years. We can delete setting them explicitly without great loss.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/: Introduce a macro for defining on-demand slots.
This is a surprisingly major win.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
Major effort to plug slot-name leaks.
Arrange that all slot names, for structures and CLOS objects, are
internal symbols of the relevant package.
There used to be a number of bad words in slot names, including `class',
`method', `expr', `type', and `condition'. All of these have gone.
I've used two main approaches.
* Renaming the slots with a leading `%'. For structures, this
involves hacking the constructor function to initialize the slot
from a dummy argument with a less unpleasant name, and setting up
trivial reader and writer function wrappers, so there's a new macro
`define-access-wrapper' in utilities.lisp to do this. For CLOS
objects, the accessor functions are named explicitly so that's not
a problem, but there's a lot of work needed to track down direct
slot accesses through `slot-value' and `with-slots'.
* For classes defined through `definst', I've instead named the slots
with gensyms (at read time), because their names are used as part of
automagically defined methods.
I may not have fixed everything: this is a rather invasive change.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Report on generic function methods.
Show all methods on newly defined generic functions, and all methods
specialized on our own classes.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Strip duplicate exports.
They can happen legitimately because of automatic exporting.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Better pretty formatting for keywords.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Mark the start of the class tree dump.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/codegen-proto.lisp, doc/list-exports.lisp: Export `inst' readers.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/: More missing exports.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Sort sibling classes by name in the tree.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Search for exports inside `eval-when' blocks.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Don't get confused and thing `nil' isn't interned.
Check the correct value of `find-symbol', because `nil' is false.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Check for anomalies when preparing reports.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
doc/list-exports.lisp: Some sketchy code to report on exported symbols.
Currently produces a list with category indications, and a class
hierarchy. Useful for preparing documentation.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/: Fix up some wrong exports.
Various small bugs:
* Some symbols which should have been exported weren't.
* One export was mispslet.
* One export was duplicated.
* One was a failed attempt to export an instruction, but `definst' can
do that properly now.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/final.lisp, src/frontend.lisp: Compile methods before dumping.
Exercise the method machinery so we don't have to run the compiler as
much after dumping.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/frontend.lisp: Prepare the builtin module at load time.
This is better because load time happens before the image is dumped.
Also, remove some pointless `clear-the-decks' calls.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/final.lisp: Rename from `debug.lisp'.
It's going to get more stuff.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
vars.am, lib/: Add some manual pages for the library.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
lib/sod.h: Arrange that `SOD_XCHAIN' returns `void *'.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
lib/sod.h, src/class-output.lisp: Rename `SOD__CAR' to `SOD_CAR'.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/class-output.lisp: Make `__CONV_' macros more robust.
If the formal argument name is a simple identifier, then things come
unstuck when there's a chain head with that same nickname. So make the
argument name be something which isn't allowed as a class nickname.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/: Stuff vtable structures into a union.
For the same reason that ichains are collected into a union: to force
the compiler's hand into getting the various members to overlap in the
common prefixes.
This doesn't affect day-to-day use, but it does change the types of the
things which are actually exported.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
lib/sod.h: Make the `_vt' member of `sod_instance' be pointer-to-const.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/frontend.lisp: The command-line options aren't crazy any more.
Mark Wooding [Mon, 7 Sep 2015 14:20:11 +0000 (15:20 +0100)]
test/chimaera.sod: Reorder Serpent tickling decision.
It makes more sense to me to have the actions in the order they'll be
experienced.
Mark Wooding [Mon, 7 Sep 2015 14:13:12 +0000 (15:13 +0100)]
test/chimaera.sod: Use the macros for sending messages.
Rather than digging about in `_vt' directly.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
Makefile.am, debian/: Initial packaging; release 0.2.0.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
Makefile.am: Include a pkgconfig file.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
Include `confsubst' machinery.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
Makefile.am: Write release number into distribution packages.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
Makefile.am: Actually distribute the `auto-version' script.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/Makefile.am: Install `parser' files in the correct place.
We have to use the `nobase_...' variable, which is waaay too long. So
invent our own variable `LISP_SOURCES' and use that instead.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/Makefile.am: Distribute `src/parser/package.lisp'.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/Makefile.am, src/test-base.lisp: Check version numbers match.
We have version numbers in two places, and that's no good, but I can't
see a good way of generating the Lisp package version from the Git
version, so I'm just going to add a test that they match.
It's not a perfect test, but it ought to do.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/Makefile.am: Improve test running rule.
Catch errors and print them briefly rather than having a full stack
backtrace.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/test-base.lisp: Actually signal an error if the tests fail.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/parser/parser-test.lisp: Suppress maybe-unused variable warning.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/c-types-parse.lisp: Some better comment formatting.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/c-types-parse.lisp: Handle `...' in prefix-`(' disambiguation.
Now that we've actually got the machinery to parse ellipses in function
argument lists, it'd be nice to do it properly.
(I think -- but I might be wrong -- that the reason this wasn't done
before is that C doesn't let you put `...' as the only thing in an
argument list, because you need to have a named argument to initialize
the `va_list'. But Sod can, and does, because there's always at least a
receiver argument.)
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/module-parse.lisp: Reinstate `peek' around the main item parser.
This was removed for debugging. Now we can parse slot initializers
again.
The error handling for class items is still pretty awful and needs
fixing.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/module-parse.lisp: Don't allow a method body on a varargs message def.
It doesn't work because the method will have the wrong type.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/class-make-impl.lisp: Make `make-sod-MUMBLE' actually return a MUMBLE.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-{proto,impl}.lisp: Add `:valist' method-entry role.
For each varargs message, add a new entry whose name has a `__v' suffix,
and which takes a `va_list' argument in place of the variable-length
argument list of the `nil' entry. The `nil' entry now just sets up the
`va_list' pointer and invokes the corresponding `:valist' entry function.
This actually makes constructing the method-entry functions somewhat
cleaner, since the handling of variable argument lists is now
concentrated in the construction of a separate entry function.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/: Allow methods to have more than one entry each in a vtable.
The entries are assigned distinct `roles' to distinguish them. Each
role can have a different type. To accommodate this a number of changes
are made to the API. Note that no roles other than the standard `nil'
role are currently defined, so none of this change should have any
externally observable effect.
* The `make-method-entry' method is replaced by `make-method-entries',
which returns a list of entry objects. The standard method on
`compute-vtmsgs' collects these together into a big list.
* Slots in the `vtmsgs' structure are now given names by the method
entries directly, rather than being named after their messages.
There is a new generic function `method-entry-slot-name' to make
this work, and a little protocol `method-entry-slot-name-by-role' to
make extending this machinery easy.
* The `message-macro-name' function now takes a method-entry rather
than a message, because each entry needs its own macro.
* The `method-entry-function-name' function has grown an additional
`role' argument. The standard method inserts a non-nil role name in
an unimaginative manner.
* The standard method on `method-entry-function-type' now inspects the
entry role, but its behaviour is unchanged except to check that the
role is nil.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-impl.lisp (compute-method-entry-functions): Simplify `parm-n'.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-impl.lisp (compute-method-entry-functions): Reorder `let*'.
Currently the `method entry details' and `effective method function
details' blocks are independent, but they won't be soon.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/: Tweak handling of `va_list' pointers in generated functions.
In method entries, effective method bodies, and delegation trampolines,
the `va_list' argument is named `sod__ap', and a copy is made for
invoking direct methods into a temporary called `sod__tmp_ap'. The name
`sod__master_ap' is now only used in direct method argument lists and
the `CALL_NEXT_METHOD' machinery.
This reduces the amount of mangling of argument lists which needs to be
done. It will also make some future changes a little easier.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-proto.lisp: Fix varargs trampoline method construction.
The extra `ap' argument comes at the front, not the end. Trim the final
ellipsis instead of (wrongly) trying to convert it.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-impl.lisp: Don't mangle `ap' name in `basic-argument-names'.
It's already correct. If we change it then there'll be two `master_ap'
arguments when we invoke the method.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-impl.lisp: Argument name list should only contain names.
The whole argument object prints wrongly and makes a mess in the output.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-impl.lisp: Handle varargs methods in `next-method-type'.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/debug.lisp: Write output to a well-known file.
This way, Emacs can display it properly fontified in a buffer, and
refresh when it gets changed, rather than having an enormous dump of
text in the middle of a REPL session.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/c-types-parse.lisp: Plain types don't have `declspec' slots.
Type names count as declaration specifiers, so `scan-declspec' has to
cope with them. Fortunately, bare types don't come with tags, so this
is easy to work around.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/class-utilities.lisp: Compare C types with the right function.
There's something weird with the caching, I think. But this way leaves
a better taste anyway.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/class-output.lisp: Remove debugging hack.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/: Argument lists don't only contain `argument' objects.
They can also contain `:ellipsis' markers.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/c-types-parse.lisp: Parse `...' in argument lists.
Somehow this got missed out. There are a number of bugs elsewhere,
where things can't cope with the `:ellipsis' marker, to be fixed in
subsequent commits.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-aggregate.lisp: Expose number of methods to custom combination.
This permits pre-allocating an array for collecting results, for
example. Suggested by Tony Finch.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
lib/sod.h, src/class-{output,utilities}.lisp: Macros for messages.
Emit a C macro for each message defined, which just saves the ugly
messing duplication of the receiver and the messing about with vtables.
This feature isn't stable yet, so don't rely on it not changing. For
example, I've not yet decided on whether or not to uppercase the message
name.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-aggregate.lisp: Move return-type check into `define-' macro.
It seems a better fit this way.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-aggregate.lisp: Fix export for `aggregating-effective-method'.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/method-aggregate.lisp: Give aggregating combinations their own file.
The code has been reordered a little too, but there are no substantive
changes.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/classes.lisp, src/class-layout-proto.lisp: Docstring fixes.
Specifically, name literal symbols in lowercase, with quotes, to
distinguish from metasyntactic variables.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
lib/sod.h, test/chimaera.sod: Promote `SOD_DECL' to the library header.
It seems generally useful.
Mark Wooding [Sun, 30 Aug 2015 11:59:22 +0000 (12:59 +0100)]
test/: Make `nml.tickle' be a `progn' method.
I think this was always intended, but I didn't have the method
combination until now.
Mark Wooding [Sun, 30 Aug 2015 10:08:11 +0000 (11:08 +0100)]
test/chimaera.sod: Don't use `(void)' for message argument lists.
It's inaccurate because more arguments are added later.
Mark Wooding [Sun, 30 Aug 2015 10:07:21 +0000 (11:07 +0100)]
src/class-make-{proto,impl}.lisp: Better choice of default message class.
If a combination is specified, use the new `aggregating-message' by
default.
Mark Wooding [Sun, 30 Aug 2015 10:06:13 +0000 (11:06 +0100)]
src/method-impl.lisp: New protocol for aggregating method combinations.
Mark Wooding [Sat, 29 Aug 2015 13:36:34 +0000 (14:36 +0100)]
src/method-impl.lisp: Invoke `after' methods, even in `void' messages.
Previously, the code would have suppressed them for some bizarre
reason.
The `voidp' variable is no longer needed, so remove it.
Mark Wooding [Sat, 29 Aug 2015 13:33:40 +0000 (14:33 +0100)]
src/method-impl.lisp: Fix incorrect argument order.
Mark Wooding [Sat, 29 Aug 2015 13:33:02 +0000 (14:33 +0100)]
src/method-impl.lisp: There's a handy definition `c-type-void', so use it.
Mark Wooding [Sat, 29 Aug 2015 13:13:01 +0000 (14:13 +0100)]
src/method-impl.lisp: Remove unnecessary `with-slots' form.
Mark Wooding [Sat, 29 Aug 2015 13:12:07 +0000 (14:12 +0100)]
src/method-impl.lisp: Typo in docstring.
Mark Wooding [Sat, 29 Aug 2015 13:11:42 +0000 (14:11 +0100)]
src/method-impl.lisp: Default method for `primary-method-class'.
The `basic-direct-method' class seems like an appropriate choice.
Mark Wooding [Sun, 30 Aug 2015 10:03:22 +0000 (11:03 +0100)]
src/module-impl.lisp: Don't emit `#line' to non-position-aware streams.
Because we can't switch back afterwards, the compiler will blame the
wrong thing. Also, if the output stream isn't position-aware, it's
probably because we're in the midst of pretty-printing something, and
forcing `#line' directives will produce hideous results.
This is all rather unfortunate. It's clearly a bug that we can't
produce pretty output and have useful error location information, but I
think this is the best I can manage right now.
Mark Wooding [Sun, 30 Aug 2015 09:57:57 +0000 (10:57 +0100)]
src/classes.lisp: Add comment explaining why psets aren't captured.
This one had me wondering for a while, but it turns out there's a good
reason.
Mark Wooding [Sat, 29 Aug 2015 14:11:55 +0000 (15:11 +0100)]
src/: Rename the various `:lisp-class' properties.
Instead, there's now a separate property for each kind of metaobject.
The reason is that some declarations can produce two metaobjects, e.g.,
a message declaration might also include a direct method, and we can't
possibly use the same metaobject class for both. To resolve this, we
use different properties to name the different metaobject classes.
* For classes, we use `:lisp-metaclass' (to distinguish from a runtime
metaclass).
* For messages, `:message-class'.
* For methods, `:method-class'.
* For slots, `:slot-class'.
* For slot initializers, `:initializer-class'.
* For modules, `:module-class'.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/pset-parse.lisp: Allow brace-enclosed fragments as property values.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/codegen-{proto,impl}.lisp: Protocol for emitting custom declarations.
Mark Wooding [Sun, 30 Aug 2015 09:59:45 +0000 (10:59 +0100)]
src/codegen-proto.lisp: Include `;' in `var-inst' output.
Because other kinds of decls will have their own semicolons. Previously
the semicolons were added as part of the block output machinery.
Mark Wooding [Sun, 30 Aug 2015 10:04:18 +0000 (11:04 +0100)]
src/{codegen,method}-impl.lisp: Allow non-variables in `vars' lists.
It used to be assumed that everything in the `vars' list was a
`var-inst' object, but that's not going to be true any more.
Mark Wooding [Sat, 29 Aug 2015 13:09:54 +0000 (14:09 +0100)]
src/codegen-proto.lisp: A small number of new instructions.
Mark Wooding [Sun, 30 Aug 2015 09:58:38 +0000 (10:58 +0100)]
src/codegen-{proto,impl}.lisp: Have `definst' optionally export symbols.
The enormous export lists are annoying to maintain.
Mark Wooding [Sat, 29 Aug 2015 12:20:49 +0000 (13:20 +0100)]
src/sod.asd: Elide a transitively implied dependency.
Mark Wooding [Sat, 29 Aug 2015 12:20:28 +0000 (13:20 +0100)]
src/sod.asd: Fix another missing dependency.
Mark Wooding [Sat, 29 Aug 2015 12:19:33 +0000 (13:19 +0100)]
src/debug.lisp: New small utility function for interactive testing.
Testing at the command line is OK when things can be expected to work,
but it's annoying when they're likely to break. This simple function
makes things pleasant to debug from inside SLIME.
Mark Wooding [Tue, 25 Aug 2015 16:46:54 +0000 (17:46 +0100)]
pre-reorg/: Delete this old cruft.
Mark Wooding [Tue, 25 Aug 2015 16:44:54 +0000 (17:44 +0100)]
lib/sod.[ch]: The runtime library is LGPL.
Mark Wooding [Tue, 25 Aug 2015 09:44:20 +0000 (10:44 +0100)]
vars.am, */Makefile.am: Improve silent-rules building.
New silent-rules handling for Lisp dumping, testing, and for generating
Sod outputs.
Mark Wooding [Tue, 25 Aug 2015 09:27:36 +0000 (10:27 +0100)]
doc/: Work in progress.
Mark Wooding [Tue, 25 Aug 2015 09:25:30 +0000 (10:25 +0100)]
lib/sod.[ch]: Name instance arguments `obj', not `p'.
In some cases, the documentation already had this right.
Mark Wooding [Tue, 25 Aug 2015 09:23:50 +0000 (10:23 +0100)]
lib/sod.[ch] (sod_convert): Make the instance argument be `const void *'.
The documentation says that it should be, and I agree.
Mark Wooding [Tue, 25 Aug 2015 09:20:05 +0000 (10:20 +0100)]
src/module-output.lisp: Remove unused output tag `:classes'.
Mark Wooding [Wed, 19 Aug 2015 16:47:15 +0000 (17:47 +0100)]
Include test machinery in the new build system.
Mark Wooding [Wed, 19 Aug 2015 16:45:26 +0000 (17:45 +0100)]
configure.ac, src/Makefile.am: Rename the FASL extension variable.
Mark Wooding [Wed, 19 Aug 2015 16:37:17 +0000 (17:37 +0100)]
src/parser/scanner-charbuf-impl.lisp: Delay `charbuf-scanner-place'.
Similar to
1d087117..., we have mutually recursive references between a
class and structure, which we resolve by delaying the structure
definition.
Mark Wooding [Wed, 19 Aug 2015 16:34:12 +0000 (17:34 +0100)]
Factor common Makefile definitions into a `vars.am' file.
Mark Wooding [Wed, 19 Aug 2015 01:17:36 +0000 (02:17 +0100)]
Introduce a proper build system and clear away old cruft.