This destabilises all the implementation details of `thread_local!`,
since they do not *need* to be stable with the new attribute.

Unstable items used in a macro expansion will now always trigger
stability warnings, *unless* the unstable items are directly inside a
macro marked with `#[allow_internal_unstable]`. IOW, the compiler warns
unless the span of the unstable item is a subspan of the definition of a
macro marked with that attribute.

E.g.

    #[allow_internal_unstable]
    macro_rules! foo {
        ($e: expr) => {{
            $e;
            unstable(); // no warning
            only_called_by_foo!();
        }}
    }

    macro_rules! only_called_by_foo {
        () => { unstable() } // warning
    }

    foo!(unstable()) // warning

The unstable inside `foo` is fine, due to the attribute. But the
`unstable` inside `only_called_by_foo` is not, since that macro doesn't
have the attribute, and the `unstable` passed into `foo` is also not
fine since it isn't contained in the macro itself (that is, even though
it is only used directly in the macro).

In the process this makes the stability tracking much more precise,
e.g. previously `println!("{}", unstable())` got no warning, but now it
does. As such, this is a bug fix that may cause [breaking-change]s.

The attribute is definitely feature gated, since it explicitly allows
side-stepping the feature gating system.

This commit deprecates the majority of std::old_io::fs in favor of std::fs and
its new functionality. Some functions remain non-deprecated but are now behind a
feature gate called `old_fs`. These functions will be deprecated once
suitable replacements have been implemented.

The compiler has been migrated to new `std::fs` and `std::path` APIs where
appropriate as part of this change.

This allows to create proper debuginfo line information for items inlined from other crates (e.g. instantiations of generics).
Only the codemap's 'metadata' is stored in a crate's metadata. That is, just filename, line-beginnings, etc. but not the actual source code itself. We are thus missing the opportunity of making Rust the first "open-source-only" programming language out there. Pity.

Many of the modifications putting in `Box::new` calls also include a
pointer to Issue 22405, which tracks going back to `box <expr>` if
possible in the future.

(Still tried to use `Box<_>` where it sufficed; thus some tests still
have `box_syntax` enabled, as they use a mix of `box` and `Box::new`.)

Precursor for overloaded-`box` and placement-`in`; see Issue 22181.

(You shouldn't use it, but it is a semi-reasonable way to annotate
types when necessary.)

This is the kind of change that one is expected to need to make to
accommodate overloaded-`box`.

----

Note that this is not *all* of the changes necessary to accommodate
Issue 22181.  It is merely the subset of those cases where there was
already a let-binding in place that made it easy to add the necesasry
type ascription.

(For unnamed intermediate `Box` values, one must go down a different
route; `Box::new` is the option that maximizes portability, but has
potential inefficiency depending on whether the call is inlined.)

----

There is one place worth note, `run-pass/coerce-match.rs`, where I
used an ugly form of `Box<_>` type ascription where I would have
preferred to use `Box::new` to accommodate overloaded-`box`.  I
deliberately did not use `Box::new` here, because that is already done
in coerce-match-calls.rs.

----

Precursor for overloaded-`box` and placement-`in`; see Issue 22181.

See buildlog here for evidence of such occurring:
  http://buildbot.rust-lang.org/builders/auto-linux-32-opt/builds/3910/steps/test/logs/stdio

Conflicts:
	src/libstd/thread.rs

Suggestion was here:

  https://github.com/rust-lang/rust/pull/22532#discussion_r25505609

Added pair of regression tests that I should have made when I thought
of doing this in the first place.

However, I left the ICEs rather than switching to nicer error
reporting there; these cases *should* be detected prior to hitting
trans, and thus ICE'ing here is appropriate.

(So why switch to `eval_const_expr_partial`? Because I want to try to
eliminate all uses of `eval_const_expr` entirely.)

 1. Detect and report arithmetic overflow during const-expr eval.

 2. Instead `eval_const_expr_partial` returning `Err(String)`, it now
    has a dedicated enum of different cases. The main benefit of this
    is the ability to pass along an interpretable payload, namely the
    two inputs that caused an overlfow.

I attempted to minimize fallout to error output in tests, but some was
unavoidable. Those changes are in a follow-on commit.

This allows computation to proceed and find further errors.

(However, this is also annoying at times when the subsequent errors
are just reporting that a ty_err occurred. I have thoughts on ways to
fix this that I will experiment with separately.)

This only replaces the conditional arith-overflow asserts with
unconditional errors from the guts of const-eval; it does *not*
attempt to sanely handle such errors e.g. with a nice error message
from `rustc`.  So the same test that led me to add this commit are
still failing, and must be addressed.

(The bug was in `impl RandomAccessIterator for Rev`; it may or may not
have been innocuous, depending on what guarantees one has about the
behavior of `idx` for the underlying iterator.)

Regarding the `rand` changes: It is unfortunate that Wrapping(T) does
not support the `+=` operator.  We may want to try to fix that before
1.0 to make porting code like this palatable.

Regarding `std::rand`, just arith-overflow in first example from
`std::rand::random()` doc.

 * `core::num`: adjust `UnsignedInt::is_power_of_two`,
   `UnsignedInt::next_power_of_two`, `Int::pow`.

   In particular for `Int::pow`: (1.) do not panic when `base`
   overflows if `acc` never observes the overflowed `base`, and (2.)
   if `acc` does observe the overflowed `base`, make sure we only
   panic if we would have otherwise (e.g. during a computation of
   `base * base`).

 * also in `core::num`: avoid underflow during computation of `uint::MAX`.

 * `std::num`: adjust tests `uint::test_uint_from_str_overflow`,
   `uint::test_uint_to_str_overflow`, `strconv`

 * `coretest::num`: adjust `test::test_int_from_str_overflow`.

* The error patterns had a typo.

* Our current constant evaluation would silently allow the overflow
  (filed as Issue 22531).

* The overflowing-mul test was accidentally doing addition instead of
  multiplication.

* `collections::btree::node`: accommodate (transient) underflow.

* `collections::btree::map`: avoid underflow during `fn next`
  for `BTreeMap::range` methods.

* `collections::slice`: note that pnkfelix deliberately used
  `new_pos_wrapping` only once; the other cases of arithmetic do not
  over- nor underflow, which is a useful property to leave implicitly
  checked/documented via the remaining calls to `fn new_pos(..)`.

* `collections::vec_deque` applied wrapping ops (somewhat blindly)
  to two implementation methods, and many tests.

* `std::collections::hash::table` : Use `OverflowingOps` trait to
  track overflow during `calculate_offsets` and `calculate_allocation`
  functions.

These return the result of the operation *plus* an overflow/underflow bit.

This can make it easier to write operations where you want to chain
some arithmetic together, but also want to return a flag signalling if
overflow every occurred.

During my clean-up of rebase errors, I took the opportunity to implement
parse_opt_bool so that it isn't identical to parse_bool wrapped in
`Some`.

parse_opt_bool considers no value to be true, a value of 'y', 'yes' or
'on' to be true and 'n', 'no' or 'off' to be false. All other values are
an error.

Many of the core rust libraries have places that rely on integer
wrapping behaviour. These places have been altered to use the wrapping_*
methods:

 * core::hash::sip - A number of macros
 * core::str - The `maximal_suffix` method in `TwoWaySearcher`
 * rustc::util::nodemap - Implementation of FnvHash
 * rustc_back::sha2 - A number of macros and other places
 * rand::isaac - Isaac64Rng, changed to use the Wrapping helper type

Some places had "benign" underflow. This is when underflow or overflow
occurs, but the unspecified value is not used due to other conditions.

 * collections::bit::Bitv - underflow when `self.nbits` is zero.
 * collections::hash::{map,table} - Underflow when searching an empty
   table. Did cause undefined behaviour in this case due to an
   out-of-bounds ptr::offset based on the underflowed index. However the
   resulting pointers would never be read from.
 * syntax::ext::deriving::encodable - Underflow when calculating the
   index of the last field in a variant with no fields.

These cases were altered to avoid the underflow, often by moving the
underflowing operation to a place where underflow could not happen.

There was one case that relied on the fact that unsigned arithmetic and
two's complement arithmetic are identical with wrapping semantics. This
was changed to use the wrapping_* methods.

Finally, the calculation of variant discriminants could overflow if the
preceeding discriminant was `U64_MAX`. The logic in `rustc::middle::ty`
for this was altered to avoid the overflow completely, while the
remaining places were changed to use wrapping methods. This is because
`rustc::middle::ty::enum_variants` now throws an error when the
calculated discriminant value overflows a `u64`.

This behaviour can be triggered by the following code:

```
enum Foo {
  A = U64_MAX,
  B
}
```

This commit also implements the remaining integer operators for
Wrapped<T>.

Adds overflow checking to integer addition, multiplication, and subtraction
when `-Z force-overflow-checks` is true, or if `--cfg ndebug` is not passed to
the compiler. On overflow, it panics with `arithmetic operation overflowed`.
Also adds `overflowing_add`, `overflowing_sub`, and `overflowing_mul`
intrinsics for doing unchecked arithmetic.

[breaking-change]

Namely, I have been annoyed in the past when I have done
`RUST_BACKTRACE=1 make check` only to discover (again) that such a
trick causes this test to fail, because it assumes that the
`RUST_BACKTRACE` environment variable is not set.

Also clarified the mysterious `_next_int` method.