Two separate changes that got intertwined (sorry):
Make to_str pure. Closes#3691
In str, change functions like push_char to take an &mut str instead of
an &str. Closes#3710
The parser now warns about use of mutbl-ref mode, though it's kind
of a lie since this commit doesn't remove support for the mode.
Changed move_val_init to have stage0 and stage1/2 versions, the latter of
which is demoded.
Changed the type that the typechecker expects the move_val_init
intrinsic to have. After this is pushed, I can make a new snapshot,
which will remove the need for the stage0 versions.
This removes most explicit uses of the + argument mode. Pending a
snapshot, I had to remove the forbid(deprecated_modes) pragma from
a bunch of files. I'll put it back!
+ mode still has to be used in a few places for functions that get
moved (see task.rs)
The changes outside core and std are due to the to_bytes trait and
making the compiler (with legacy modes on) agree with the libraries
(with legacy modes off) about modes.
Everything should now call ptr::addr_of instead of
ptr::p2::addr_of. Only the pipes macro code when compiled
by stage0 will call ptr::p2::addr_of. Needs a snapshot to get
rid of that.
two closures in the same function body can end up having
the same mangled name. This small patch uses gensym to
uniquely name each closure body so there won't be conflicts.
Destructors were internally declared with an extra (hidden) nil-typed
argument that was passed in by-ref mode. This was causing spurious
mode warnings. Deleted it. Also some misc. cleanup because I
couldn't help myself.
This will help with the auto_serialize2 migration. We have to change
ident from a type alias to uint into a unique type. We need to use
a struct instead of a "enum ident = token::str_num" because structs
support constants, but newtypes do not.
It's not clear that impls should have to be public to be used
from outside the crate; in particular, I don't know whether
they have to be public to be used from other modules within
the same crate.
