There were a bunch of problems with consts where an enum was contained within some other type (vector, tuple, struct, etc.); some of these would cause LLVM assertion failures, and some would silently read from the wrong address. These changes should fix all of that.
It would be good if someone with access to a win32 host could do the equivalent of `make check-stageN-rpass TESTNAME=enum` on that platform before merging this.
This is wasted space if the const is just an enum, but optimizing that
case without breaking everything else is an issue that can be addressed
separately.
The first commit message has most of the comments, but this pull request basically fixes a lot of issues surrounding the `unused_imports` warning/deny attribute.
Before this patch there were these problems:
1. Unused imports from `prelude.rs` were warned about with dummy spans, leading to a large number of confusing warnings.
2. Unused imports from `intrinsic.rs` were warned about with the file `<intrinsic>` which couldn't be forced to go away
3. Methods used from imported traites (like `io::WriterUtil`) resulted in an unused warning of the import even though it was used.
4. If one `use` statement imported N modules, M of which weren't used, M warning statements were issued.
5. If a glob import statement was used, each public export of the target module which wasn't used had a warning issued.
This patch deals with all these cases by doing:
1. Ignore unused imports from `prelude.rs` (indicated by a dummy span of 0)
2. Ignore unused imports from `intrinsic.rs` (test on the imported module name, is there a better way?)
3. Track when imported modules are used as candidates for methods, and just assume they're used. This may not end up being the actual case, but in theory not warning about an unused thing is worse than warning about a used thing.
4. Only issue one warning statement
5. Only issue one warning statement.
This is the first time I've edited the compiler itself, and I tried to keep up with the style around, but I may have missed something here or there...
r?
It looks to me like the string_reader and tt_reader structs are
GC pointers only because they predate the modern borrow system.
This commit leaves the type names string_reader and tt_reader alone
(they still refer to GC-ed pointers), but internally the functions
now use borrowed pointers to refer to these structures. My guess
would be that it's possible to move this change outward and not
use the GCed pointers at all, but that change looks like it could be
a larger one. Actually, I'm delighted at how quick this change was.
5283a8b reworks the TreeMap lazy iterator to use `&mut` again, which closes#4763. It gets the performance of the set methods back in the same ballpark that it was pre-INHTWAMA which is nice. These can be turned back into methods eventually.
e5b6334 removes the transitional smallintmap attributes which closes#4737.
Also adds Rng::gen() for generically generating any type that implements the Rand trait. There's no way to generate things with a length (for e.g. strings or vectors), because I can't think of an elegant way to do that. Maybe have a RandLen trait that inherits Rand?
This can be used for a quickcheck mechanism I'm working on.
Each call to next() was doing a copy rather than a move. There's
currently no way for this to be a method that uses &mut self, so it has
to be a free function. Closes#4763.
