leak the affects of closures on the free-region-map, like we used to

This restores the behavior of regionck with respect to the
free-region-map: that is, it collects all the relations from the fn
and its closures. This feels a bit fishy but it's the behavior we've
had for some time, and it will go away with NLL, so seems best to just
keep it.

src/librustc/infer/outlives/env.rs

@@ -89,6 +89,52 @@ impl<'a, 'gcx: 'tcx, 'tcx: 'a> OutlivesEnvironment<'tcx> {
         self.free_region_map
     }
 
+    /// This is a hack to support the old-skool regionck, which
+    /// processes region constraints from the main function and the
+    /// closure together. In that context, when we enter a closure, we
+    /// want to be able to "save" the state of the surrounding a
+    /// function. We can then add implied bounds and the like from the
+    /// closure arguments into the environment -- these should only
+    /// apply in the closure body, so once we exit, we invoke
+    /// `pop_snapshot_post_closure` to remove them.
+    ///
+    /// Example:
+    ///
+    /// ```
+    /// fn foo<T>() {
+    ///    callback(for<'a> |x: &'a T| {
+    ///         // ^^^^^^^ not legal syntax, but probably should be
+    ///         // within this closure body, `T: 'a` holds
+    ///    })
+    /// }
+    /// ```
+    ///
+    /// This "containment" of closure's effects only works so well. In
+    /// particular, we (intentionally) leak relationships between free
+    /// regions that are created by the closure's bounds. The case
+    /// where this is useful is when you have (e.g.) a closure with a
+    /// signature like `for<'a, 'b> fn(x: &'a &'b u32)` -- in this
+    /// case, we want to keep the relationship `'b: 'a` in the
+    /// free-region-map, so that later if we have to take `LUB('b,
+    /// 'a)` we can get the result `'b`.
+    ///
+    /// I have opted to keep **all modifications** to the
+    /// free-region-map, however, and not just those that concern free
+    /// variables bound in the closure. The latter seems more correct,
+    /// but it is not the existing behavior, and I could not find a
+    /// case where the existing behavior went wrong. In any case, it
+    /// seems like it'd be readily fixed if we wanted. There are
+    /// similar leaks around givens that seem equally suspicious, to
+    /// be honest. --nmatsakis
+    pub fn push_snapshot_pre_closure(&self) -> usize {
+        self.region_bound_pairs.len()
+    }
+
+    /// See `push_snapshot_pre_closure`.
+    pub fn pop_snapshot_post_closure(&mut self, len: usize) {
+        self.region_bound_pairs.truncate(len);
+    }
+
     /// This method adds "implied bounds" into the outlives environment.
     /// Implied bounds are outlives relationships that we can deduce
     /// on the basis that certain types must be well-formed -- these are

src/librustc_typeck/check/regionck.rs

@@ -428,17 +428,17 @@ impl<'a, 'gcx, 'tcx> Visitor<'gcx> for RegionCtxt<'a, 'gcx, 'tcx> {
 
         // Save state of current function before invoking
         // `visit_fn_body`.  We will restore afterwards.
-        let outlives_environment = self.outlives_environment.clone();
         let old_body_id = self.body_id;
         let old_call_site_scope = self.call_site_scope;
+        let env_snapshot = self.outlives_environment.push_snapshot_pre_closure();
 
         let body = self.tcx.hir.body(body_id);
         self.visit_fn_body(id, body, span);
 
         // Restore state from previous function.
+        self.outlives_environment.pop_snapshot_post_closure(env_snapshot);
         self.call_site_scope = old_call_site_scope;
         self.body_id = old_body_id;
-        self.outlives_environment = outlives_environment;
     }
 
     //visit_pat: visit_pat, // (..) see above

src/test/run-pass/implied-bounds-closure-arg-outlives.rs

@@ -0,0 +1,44 @@
+// Copyright 2016 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+// Test that we are able to handle the relationships between free
+// regions bound in a closure callback.
+
+#[derive(Copy, Clone)]
+struct MyCx<'short, 'long: 'short> {
+    short: &'short u32,
+    long: &'long u32,
+}
+
+impl<'short, 'long> MyCx<'short, 'long> {
+    fn short(self) -> &'short u32 { self.short }
+    fn long(self) -> &'long u32 { self.long }
+    fn set_short(&mut self, v: &'short u32) { self.short = v; }
+}
+
+fn with<F, R>(op: F) -> R
+where
+    F: for<'short, 'long> FnOnce(MyCx<'short, 'long>) -> R,
+{
+    op(MyCx {
+        short: &22,
+        long: &22,
+    })
+}
+
+fn main() {
+    with(|mut cx| {
+        // For this to type-check, we need to be able to deduce that
+        // the lifetime of `l` can be `'short`, even though it has
+        // input from `'long`.
+        let l = if true { cx.long() } else { cx.short() };
+        cx.set_short(l);
+    });
+}