The code is somewhat invasive, but it seems hard to do this in a
clean way, since the design itself involves a bunch of 'action
at a distance'.
Issue #1819
This change uses the same code for handling the "self" reference for
classes as is already used for impls/ifaces. This allows removing the
extra maybe_self_id argument (which was just for classes) to trans_closure
that I added before. I also rewrote the translation for class ctors so
that it doesn't generate new AST nodes (instead translating directly).
Also changed visit so that it visits class ctors correctly with visit_fn,
and changed typestate to not do return-checking when visiting a class ctor.
Cross-crate method calls don't work yet. Added
run-pass/class-method-cross-crate to test that, but it's xfailed
References to fields within methods don't work yet. Added
run-pass/class-methods to test that, but it's also xfailed
In particular, use the ast::method type to represent a class method,
and try to reuse as much iface code as possible. (This makes sense now
since I'll be allowing polymorphic class methods.)
Cross-crate metadata for classes works well enough that programs with
classes in other crates compile successfully, but output wrong results.
Checking in work so far to avoid merge hassles. (Tests are xfailed.)
We used to generate a module T with a serialize() and deserialize() fn,
but this was suboptimal for a number of reasons:
- it required moving serialization into core so that uint etc worked
- it was harder to override the serialization behavior locally
(this is now trivial)
- Move io, run and rand to core.
- Remove incorrect ctypes module (use libc).
- Remove os-specific modules for os and fs.
- Split fs between core::path and core::os.
As per discussion on IRC. I am about to file an RFC for further
discussion about the more general issue of whether to enforce
invariants through types, typestate, or dynamic checks, but for now,
removing the misleading name "last_unsafe".
This required changing almost all users of hashmaps to import the hashmap interface first.
The `size` member in the hashmap structure was renamed to `count` to work around a name conflict.
This required changing almost all users of hashmaps to import the hashmap interface first.
The `size` member in the hashmap structure was renamed to `count` to work around a name conflict.
Programs using classes with fields only (no methods) compile and run,
as long as nothing refers to a class in a different crate (todo).
Also changed the AST representation of classes to have a separate
record for constructor info (instead of inlining the fields in the
item_class node), and fixed up spans and pretty-printing for
classes.
The prestate for calls was getting set incorrectly to the poststate for the operator
in the call. This worked before since most of the time, operator expressions are
pure. Issue 1895 shows how this breaks when the operator is a closure that has a
move-in capture clause.
(I had a several-day, multi-file patch for this that didn't work... and then it
turned out to be a one-line fix. The joys of programming.)
Closes#1895
Class tests aren't working yet, but they fail a little later :-)
Also, make the parser correctly set a constructor's result type to
its enclosing class type.
Issue #352Closes#1720
The old checker would happily accept things like 'alt x { @some(a) { a } }'.
It now properly descends into patterns, checks exhaustiveness of booleans,
and complains when number/string patterns aren't exhaustive.
This allows you to take the value of, for example, `[1].len`, or bind
it with `bind x.map(_)` syntax.
I'm holding off on implementing this for dynamic methods (those on
bounded type parameters or iface types) until it's clearer what we
will do with monomorphization.
Issue #435
It would fail to start out with a fresh scope when going over
a loop or block function for the second time, and thus not
recognize last uses of locals defined inside the block.
Closes#1818
This function creates a new scheduler with a specified number of threads and
immediately executes a task on it. The scheduler is configured to terminate
when the task dies. This is the minimum API necessary to enable blocking C
calls.
Added class support to the parser, prettyprinter, fold, and visit.
(See Issue 1726.)
This is WIP -- the test case is xfailed, and attempting to compile
it will error out in resolve.
Now that core exports "option" as a synonym for option::t, search-and-
replace option::t with option.
The only place that still refers to option::t are the modules in libcore
that use option, because fixing this requires a new snapshot
(forthcoming).
Since item_consts can't refer to or modify local variables, they
don't participate in typestate and thus get empty pre and
postconditions by default.
Closes#1660
3 tests, pretty/block-disambig.rs, run-pass/operator-overloading.rs,
and run-pass/weird-exprs.rs, all included the ternary operator. These
were changed to use the if-then-else construct instead.
2 tests, run-pass/block-arg-in-ternary.rs and run-pass/ternary.rs, were
only there because of the ternary operator, and were removed.
The methods used to implement operators now simply use
the name of the operator itself, except for unary -, which is called
min to not clash with binary -. Index is called [].
Closes#1520
When no built-in interpretation is found for one of the operators
mentioned below, the typechecker will try to turn it into a method
call with the name written next to it. For binary operators, the
method will be called on the LHS with the RHS as only parameter.
Binary:
+ op_add
- op_sub
* op_mul
/ op_div
% op_rem
& op_and
| op_or
^ op_xor
<< op_shift_left
>> op_shift_right
>>> op_ashift_right
Unary:
- op_neg
! op_not
Overloading of the indexing ([]) operator isn't finished yet.
Issue #1520
See issue 1426 for details. Now, the semantics of "export t;" where t is a tag are
to export all of t's variants as well. "export t{};" exports t but not its
variants, while "export t{a, b, c};" exports only variants a, b, c of t.
To do:
- documentation
- there's currently no checking that a, b, c are actually variants of t in the
above example
- there's also no checking that t is an enum type, in the second two examples above
- change the modules listed in issue 1426 that should have the old export
semantics to use the t{} syntax
I deleted the test export-no-tag-variants since we're doing the opposite now,
and other tests cover the same behavior.
Support Lenny222's proposed syntax for exporting a tag without
its variants, or selected tags from a variant, in the AST and parser.
No support further down the line yet. Tests are xfailed.
typestate was using the enclosing function ID for the "this function
returns" constraint, which meant confusion and panic in the case
where a predicate p includes "check p()". Fixed it to use a fresh
ID.
Closes#933
This is not my ideal way of going about things. I'd prefer not
to have expressions typed as fn*(), for example, but I couldn't
get that to work together with inferring the modes of arguments
and other corner cases.
Although the old version of GEP_tup_like was incorrect in some
cases, I do not believe we ever used it in an incorrect fashion.
In particular, it could go wrong with extended index sequences
like [0, 1, 3], but as near as I can tell we only ever use it
with short sequences like [0, i].
This commit allows patterns like:
alt x { some(_) { ... } none { } }
without the '.' after none. The parser suspends judgment about
whether a bare ident is a tag or a new bound variable; instead,
the resolver disambiguates.
This means that any code after resolution that pattern-matches on
patterns needs to call pat_util::normalize_pat, which consults
an environment to do this disambiguation.
In addition, local variables are no longer allowed to shadow
tag names, so this required changing some code (e.g. renaming
variables named "mut", and renaming ast::sub to subtract).
The parser currently accepts patterns with and without the '.'.
Once the compiler and libraries are changed, it will no longer
accept the '.'.
