Make all glue functions take values by alias to remove the need for
bitcasts at the top of every glue function. Use static type
information to produce the correct type for glue functions so that
LLVM can enforce the type system at call sites.
(specifically in the method_trait case) -- if you wrote x.f()
and x has type @T for a trait T, x wasn't getting auto-deref'ed.
This was bad.
Closes#2935
we now detect inconsistent modes, binding names, and various other errors.
typeck/trans integration is mostly done.
borrowck not so much.
more tests needed.
Region pointers were considered a scalar type, so compare_scalar_type would
get called to compare region pointers in trans. This would fail, since
compare_scalar_type has no case for region pointers.
Changed type_is_scalar to return false for region pointers. This had the side
effect of breaking casts to types of the form &T. To ameliorate that, I added
library functions ptr::assimilate (taking a &T to a *T) and ptr::to_uint
(taking a &T to a uint), both of which use reinterpret_cast.
While I was at it, I removed ty::type_has_resources, which is dead code.
If the expected type is none (due to a type error), we shouldn't fail with an ICE, but
rather, just print out another type error. Changed the do-expr type error message to
make sense in this context (see the test case for how it works).
Closes#3044.
API is (for now) mostly by value, there are options to use it by
reference if you like. Hash and equality functions must be pure
and by reference (forward looking to the day when something
like send_map becomes the standard map).
What we now do is to create a region variable for each &
expression (and also each borrow). The lifetime of this
variable will be checked by borrowck to ensure it is not greater
than the lifetime of the underlying data. This both leads to
shorter lifetimes in some cases but also longer in others,
such as taking the address to the interior of unique boxes
tht are rooted in region pointers (e.g., returning a pointer
to the interior of a sendable map).
This may lead to issue #2977 if the rvalue is not POD, because
we may drop the data in trans sooner than borrowck expects us
to. Need to work out precisely where that fix ought to occur.
Updating types std::sha1::sha1 -> hash::streaming was a relatively
small change. Renaming the variables to reflect that things aren't
sha1s any more touched far more lines.
Rewrote bitv as a class that uses a 32-bit int as its representation
for bit vectors of 32 bits or less, and a vector (the old representation)
otherwise. I didn't benchmark very much, but a bit of informal benchmarking
suggested this is a win.
Closes#2341
Previously, resolve was allowing impls, traits or classes that were
nested within a fn to refer to upvars, as well as referring to type
parameters bound by the fn. Fixing this required adding a new kind of
def: def_typaram_binder, which can refer to any of an impl, trait or
class that has bound ty params. resolve uses this to enforce that
methods can refer to their parent item's type parameters, but not to
outer items' type parameters; other stages ignore it. I also made
sure that impl, trait and class methods get checked inside a
MethodRibKind thing so as to forbid upvars, and changed the definition
of MethodRibKind so that its second argument is an optional node_id
(so that required trait method signatures can be checked with a
MethodRibKind as well).
New style is allow(foo), warn(foo), deny(foo) and forbid(foo),
mirrored by -A foo, -W foo, -D foo and -F foo on command line.
These replace -W no-foo, -W foo, -W err-foo, respectively.
Forbid is new, and means "deny, and you can't override it".
It will warn you if you use the default mode for something that
is expensive to copy, and it will warn you if you use any explicit
mode other than copy. So you should migrate over to using the
default mode for most things (and borrowed pointers when you don't
want to copy) and copy mode for things you really wanted to copy.
