This involved adding 'copy' to more generics than I hoped, but an
experiment with making it implicit showed that that way lies madness --
unless enforced, you will not remember to mark functions that don't
copy as not requiring copyable kind.
Issue #1177
This goes before a snapshot, so that subsequenct patches can make the
transition without breaking the build. Disables kind checking pass, makes
parser accept both new and old-style kind annotation.
Issue #1177
This patch changes how to specify ABI and link name of a native module.
Before:
native "cdecl" mod llvm = "rustllvm" {...}
After:
#[abi = "cdecl"]
#[link_name = "rustllvm"]
native mod llvm {...}
The old optional syntax for ABI and link name is no longer supported.
Fixes issue #547
It now threads information about invalidated aliases through the AST
properly. This makes it more permissive for conditionals (invalidating
an alias in one branch doesn't prevent you from using it in another),
and less permissive for loops (it now properly notices when a loop
invalidates an alias that it might still use in another iteration).
Closes#1144
Patch to error and fail instead of using all available memory
then crashing to detect the error condition of an unmatched
double quote before the end of a file.
I couldn't get it to show nice error messages, so this may not be
the ideal fix.
A test case for this situation has also been added.
So *resource, ~resource, [resource] are all pinned. This is counter to the
design of the kind system, but this way is a much clearer path to type safety.
Once we've established a good baseline with lots of tests, then we can try to
make raising pinned kinds work.
We were only using it in a single place, and there for no discernable reason
(probably as part of the bare-fn-vals-are-not-copyable plan). It seems more
surprising than useful.
It is now 1-based, rather than 0 based. (Seems more natural, and allows 0 to
be used to refer to self and maybe to closure.)
Also allows non-referenced args to be implicitly copied again.
Issue #918
Blocks (or statements involving blocks) that end in a semicolon are no
longer considered the block-expression of their outer block. This used
to be an expression block, but now is a statement block:
{ if foo { ret 1; } else { ret 10; } }
This helps clear up some ambiguities in our grammar.
Upvars are now marked with def_upvar throughout, not just when going
through freevars::lookup_def. This makes things less error-prone. One
thing to watch out for is that def_upvar is used in `for each` bodies
too, when they refer to a local outside the body.
Having it in the alias pass was slightly more efficient (finding
expression roots has to be done in both passes), but further muddled
up the already complex alias checker.
Also factors out some duplication in the mutability-checking code.
Closes#868. Unfortunately, this causes certain invalid programs to
fail type-checking instead of failing type-state when a type-state
error message would probably be more intuitive. (Although, by any
reasonable interpretation of the static semantics, it technically
ought to be a type error.)
Autoderef on binops is basically unused, kind of silly, and
complicates typechecking. There were only three instances of it in the
compiler and the test drivers, two of which were of the form "*foo =
foo + 1", which should be written as "*foo += 1" anyways.
I tried to pay attention to what was actually being tested so, e.g. when I
test was just using a vec as a boxed thing, I converted to boxed ints, etc.
Haven't converted the macro tests yet. Not sure what to do there.
Previously, typestate was initializing the init constraint for
a declared-but-not-initialized variable (like x in "let x;") to False,
but other constraints to Don't-know. This led to over-lenient results
when a variable was used before declaration (see the included test
case). Now, everything gets initialized to False in the prestate/poststate-
finding phase, and Don't-know should only be used in pre/postconditions.
This aspect of the algorithm really needs formalization (just on paper),
but for now, this closes#700
In the writeback phase, the typechecker now checks that it isn't
replacing a type variable T with a type that contains T. It
also does an occurs check in do_autoderef in order to avoid
getting into an infinite chain of derefs.
I'm a bit worried that there are more places where the occurs
check needs to happen where I'm not doing it now, though.
Closes#768
While it is still technically possible to test stage 0, it is not part of any
of the main testing rules and maintaining xfail-stage0 is a chore. Nobody
should worry about how tests fare in stage0.
The logic for how the "returns" constraint was handled was always
dodgy, for reasons explained in the comments I added to
auxiliary::fn_info in this commit. Fixed it by adding distinct
"returns" and "diverges" constraints for each function, which
are both handled positively (that is: for a ! function, the
"diverges" constraint must be true on every exit path; for
any other function, the "returns" constraint must be true
on every exit path).
Closes#779
This was previously disallowed by the typechecker and not properly handled
in trans. I removed the typechecker check (replacing it with a simpler
check that spawned functions don't have type params) and fixed trans.
Closes#756.
This replaces the make-based test runner with a set of Rust-based test
runners. I believe that all existing functionality has been
preserved. The primary objective is to dogfood the Rust test
framework.
A few main things happen here:
1) The run-pass/lib-* tests are all moved into src/test/stdtest. This
is a standalone test crate intended for all standard library tests. It
compiles to build/test/stdtest.stageN.
2) rustc now compiles into yet another build artifact, this one a test
runner that runs any tests contained directly in the rustc crate. This
allows much more fine-grained unit testing of the compiler. It
compiles to build/test/rustctest.stageN.
3) There is a new custom test runner crate at src/test/compiletest
that reproduces all the functionality for running the compile-fail,
run-fail, run-pass and bench tests while integrating with Rust's test
framework. It compiles to build/test/compiletest.stageN.
4) The build rules have been completely changed to use the new test
runners, while also being less redundant, following the example of the
recent stageN.mk rewrite.
It adds two new features to the cfail/rfail/rpass/bench tests:
1) Tests can specify multiple 'error-pattern' directives which must be
satisfied in order.
2) Tests can specify a 'compile-flags' directive which will make the
test runner provide additional command line arguments to rustc.
There are some downsides, the primary being that Rust has to be
functioning pretty well just to run _any_ tests, which I imagine will
be the source of some frustration when the entire test suite
breaks. Will also cause some headaches during porting.
Not having individual make rules, each rpass, etc test no longer
remembers between runs whether it completed successfully. As a result,
it's not possible to incrementally fix multiple tests by just running
'make check', fixing a test, and repeating without re-running all the
tests contained in the test runner. Instead you can filter just the
tests you want to run by using the TESTNAME environment variable.
This also dispenses with the ability to run stage0 tests, but they
tended to be broken more often than not anyway.
Programs with constrained types now parse and typecheck, but
typestate doesn't check them specially, so the one relevant test
case so far is XFAILed.
Also rewrote all of the constraint-related data structures in the
process (again), for some reason. I got rid of a superfluous
data structure in the context that was mapping front-end constraints
to resolved constraints, instead handling constraints in the same
way in which everything else gets resolved.
Capturing a type argument in the enclosing scope should be an error --
this commit implements that check in resolve, avoiding a potential
assertion failure in trans.
Closes#648.
Typestate was failing to check some code because if it saw an item,
it would quit immediately. This was to avoid checking nested items
in the same context as the lexically enclosing item, but it was
having the wrong effect: not checking the code after the item at all.
Fixed by switching to visit and skipping over items in a proper
nested fashion. Closes#668.
A check in trans didn't have a corresponding check in typeck, causing
some programs (to wit, compile-fail/chan-parameterized-args.rs - part of this
commit) to fail with an assertion failure in trans instead of a type error.
Fixed it. In short, arguments that are future thunk arguments (any spawn
arguments, and _ arguments in bind) need to either not contain type params
or type vars, or be by-reference.
Closes#665.
You can now say
expr_move(?dst, ?src) | expr_assign(?dst, ?src) { ... }
to match both expr_move and expr_assign. The names, types, and number
of bound names have to match in all the patterns.
Closes#449.
This is important since we are going to be making functions noncopyable
soon, which means we'll be seeing a lot of boxed functions.
(*f)(...) is really just too heavyweight.
Doing the autodereferencing was a very little bit tricky since
trans_call works with an *lval* of the function whereas existing
autoderef code was not for lvals.
Resources are now defined like...
resource fd(int n) { close(n); }
Calling fd with an int will then produce a non-copyable value
that, when dropped, will call close on the given int.
Wrote some small test cases that use while loops and moves, to
make sure the poststate for the loop body gets propagated into the
new prestate and deinitialization gets reflected.
Along with that, rewrite the code for intersecting states. I still
find it dodgy, but I guess I'll continue trying to add more tests.
Also, I'll probably feel better about it once I start formalizing
the algorithm.
Includes assignment operations. Add regression tests for lots of less useful,
less used or unexpected combinations, as well as a selection of compile-fail
tests. Closes#500 (again!)
Modified typestate to throw away any constraints mentioning a
variable on the LHS of an assignment, recv, assign_op, or on
either side of a swap.
Some code cleanup as well.
If you use a function expecting an alias argument in a context that
expects a function expecting a value argument, or vice versa, the
previous error message complained that the number of arguments was
wrong. Fixed the error message to be accurate.
typestate now drops constraints correctly in the post-state of
a move expression or a declaration whose op is a move. It doesn't
yet drop constraints mentioning variables that get updated.
To do this, I had to change typestate to use trit-vectors instead
of bit-vectors, because for every constraint, there are three
possible values: known-to-be-false (e.g. after x <- y, init(y) is
known-to-be-false), known-to-be-true, and unknown. Before, we
conflated known-to-be-false with unknown. But move requires them
to be treated differently. Consider:
(program a)
(a1) x = 1;
(a2) y <- x;
(a3) log x;
(program b)
(b1) x = 1;
(b2) y <- z;
(b3) log x;
With only two values, the postcondition of statement a2 for
constraint init(x) is the same as that of b2: 0. But in (a2)'s
postcondition, init(x) *must* be false, but in (b2)'s condition,
it's just whatever it was in the postcondition of the preceding statement.
This code was causing a bounds check failure:
fn hd[U](&vec[U] v) -> U {
fn hd1(&vec[U] w) -> U {
ret w.(0);
}
ret hd1(v);
}
because in hd1, U was being treated as if it referred to a type
parameter of hd1, rather than referring to the lexically enclosing binding
for U that's part of hd.
I'm actually not sure whether this is a legit program or not. But I wanted
to get rid of the bounds check error, so I assumed that program shouldn't
compile and made it a proper error message.
This involved, in part, changing the ast::def type so that a def_fn
has a "purity" field. This lets the typechecker determine whether
functions defined in other crates are pure.
It also required updating some error messages in tests. As a test
for cross-crate constrained functions, I added a safe_slice function
to std::str (slice(), with one of the asserts replaced with a
function precondition) and some test cases (various versions of
fn-constraint.rs) that call it. Also, I changed "fn" to "pred" for
some of the boolean functions in std::uint.
I noticed that typestate was being lazier than it should be,
because it was only checking typestate for statements and
top-level expression (that is, the expression in a stmt_expr, but
not any subexpressions). So I rewrote the checks in tstate/ck.rs
to use walk, which exposed a few bugs in typestate that I fixed.
Also added some more test cases for if-check.
I don't currently know how to deal with syntax extensions that appear betweeen
an attribute and an item, so this test captures the error that occurs.
Issue #487
I added a "resolved" version of the ast::constr type -- ty::constr_def
-- that has a def_id field instead of an ann_field. This is more
consistent with other types and eliminates some checking.
Incidentally, I removed the def_map argument to the top-level function
in middle::alias, since the ty::ctxt already has a def_map field.
This is a somewhat odd place to put these checks, but the data tracked
by that pass, and the available functions, make it trivial to do such
a check there.
Before, all aliases were implicitly mutable, and writing
&mutable was the same as writing &. Now, the two are
distinguished, and assignments to regular aliases are
no longer allowed.
Improve error message in the case where a use of a polymorphic tag has
insufficient type arguments given. Before, the typechecker was
just crashing with a bounds check error.
The old system tried to ensure that the location an alias pointed at
would retain its type. That turned out to not be strong enough in the
face of aliases to the inside of tags.
The new system instead proves that values pointed to by aliases are
not replaced (or invalidated in some other way) at all. It knows of
two sufficient conditions for this, and tries to prove at least of
them:
A) The alias is 'immutably rooted' in a local, and this local is not
reassigned for the lifetime of the alias. Immutably rooted means
the alias refers to the local itself, or to something reachable
from the local through immutable dereferencing.
B) No value whose type might include the type of the 'inner mutable
element' of the thing the alias refers to (for example, the box in
rec(mutable x = @mutable int)) is from the outer scope is accessed
for the lifetime of the alias. This means for functions, no other
argument types may include the alias's inner mutable type. For alt,
for each, and for, it means the body does not refer to any locals
originating from outside their scope that include this type.
The lifetime of an alias in an alt, for each, or for body is defined
as the range from its definition to its last use, not to the point
where it goes out of scope. This makes working around these
restrictions somewhat less annoying. For example, you can assign to
your alt-ed value you don't refer to any bindings afterwards.
The alias checker works by ensuring that any value to which an alias
is created is rooted in some way that ensures it outlives the alias.
It is now disallowed to create an alias to the content of a mutable
box, or to a box hanging off a mutable field. There is also machinery
in place to prevent assignment to local variables whenever they are
the root of a live alias.
