rustc: simpler ParameterEnvironment and free regions. #41914

eddyb · 2017-05-11T15:36:40Z

The commits describe the slow transformation but the highlights are:

ReEarlyBound is considered free, with a scope based on the item that defined the lifetime parameter, and the root body of the RegionMaps in use, removing the need for free_substs
liberate_late_bound_regions and implicit_region_bound moved to typeck
CodeExtent not interned at all now - ideally it would be 2 u32 but it's small anyway

Future work building up on this could include:

ParameterEnvironment becoming just the result of predicates_of
- interning makes my "parent chain" scheme unnecessary
implicit_region_bound could be retrieved from RegionMaps
renaming CodeExtent to Scope
- generalizing "call site" to "use site" or something better to include constants
renaming RegionMaps to ScopeTree and its API to talk about "parents" explicitly

rust-highfive · 2017-05-11T15:36:51Z

r? @pnkfelix

(rust_highfive has picked a reviewer for you, use r? to override)

nikomatsakis · 2017-05-11T16:55:27Z

r? @nikomatsakis

nikomatsakis

OK, this looks pretty nice. I left two small nits.

nikomatsakis · 2017-05-11T17:34:03Z

src/librustc/middle/region.rs

+            body
+        })
+        .unwrap_or_else(|| {
+            let root = tcx.hir.as_local_node_id(fr.scope).unwrap();


I can't really follow what's going on here; maybe add a comment or two for what kinds of rust code falls into which case here?

This is explained on root_parent but do note this is not the final version of that function.
Let me know on the final functions where it seems unclear.

nikomatsakis · 2017-05-11T18:13:36Z

src/librustc/infer/region_inference/mod.rs

@@ -213,6 +215,7 @@ pub struct RegionVarBindings<'a, 'gcx: 'a+'tcx, 'tcx: 'a> {
    // record the fact that `'a <= 'b` is implied by the fn signature,
    // and then ignore the constraint when solving equations. This is
    // a bit of a hack but seems to work.
+    early_bound_givens: RefCell<FxHashSet<(ty::EarlyBoundRegion, ty::RegionVid)>>,


maybe combine these by changing the type to RefCell<FxHashSet<(ty::Region<'tcx>, ty::RegionVid)>>, where the region is either an early-bound or free?

I was wondering if that would be more expensive but I doubt it now. Will do.

nikomatsakis · 2017-05-11T18:23:44Z

I see that the librustc_driver unit tests are failing. I do hate those. I've been wanting to refactor them in such a way that we can encode the same scenarios in .rs files, but it's kind of difficult. I'm tempted to just tell you to delete them -- I'm not sure to what extent we have complete coverage of those scenarios elsewhere though (specifically some of the LUB tests).

arielb1 · 2017-05-11T19:28:37Z

CodeExtent was interned as a u32 once upon a time. Not sure why that changed.

arielb1 · 2017-05-11T19:36:31Z

src/librustc/middle/region.rs

+            }
+            _ => fr.scope
+        };
+        assert_eq!(param_owner, fr.scope);


Why can't you use fr.scope directly?

I wanted to have no regressions from lifetime parameters not defined on the scope leaking into e.g. the signature of a closure.

nikomatsakis · 2017-05-11T20:01:19Z

@arielb1

CodeExtent was interned as a u32 once upon a time. Not sure why that changed.

I changed that to a pointer because there were various contexts that needed to be able to access the data without having enough context to get a region-maps (note that there is now more than one region-maps). But passing by value is probably even better. (Also, in the meantime, we've started to intern Regions, so I don't think that the size matters so much anymore anyway.)

nikomatsakis · 2017-05-12T13:47:24Z

src/librustc/middle/region.rs

+
+        let param_owner_id = tcx.hir.as_local_node_id(param_owner).unwrap();
+        let body_id = tcx.hir.maybe_body_owned_by(param_owner_id).unwrap_or_else(|| {
+            // The lifetime was defined on node that doesn't own a body,


👍 nice comment

nikomatsakis · 2017-05-12T13:52:45Z

@bors r+

bors · 2017-05-12T13:52:46Z

📌 Commit 788e70e has been approved by nikomatsakis

Rollup of 15 pull requests - Successful merges: #41820, #41843, #41847, #41860, #41896, #41912, #41916, #41918, #41919, #41921, #41923, #41934, #41940, #41942, #41943 - Failed merges: #41914

bors · 2017-05-13T14:03:03Z

☔ The latest upstream changes (presumably #41965) made this pull request unmergeable. Please resolve the merge conflicts.

eddyb · 2017-05-13T14:43:51Z

@bors r=nikomatsakis

bors · 2017-05-13T14:43:52Z

📌 Commit 6da4123 has been approved by nikomatsakis

eddyb · 2017-05-13T17:04:40Z

@bors p=1000 (separate for perf. I wish we had a guaranteed to keep PRs out of rollups)

bors · 2017-05-13T17:56:16Z

⌛ Testing commit 6da4123 with merge 826d8f3...

rustc: simpler ParameterEnvironment and free regions. The commits describe the slow transformation but the highlights are: * `ReEarlyBound` is considered free, with a scope based on the item that defined the lifetime parameter, and the root body of the `RegionMaps` in use, removing the need for `free_substs` * `liberate_late_bound_regions` and `implicit_region_bound` moved to typeck * `CodeExtent` not interned at all now - ideally it would be 2 `u32` but it's small anyway Future work building up on this could include: * `ParameterEnvironment` becoming just the result of `predicates_of` * interning makes my "parent chain" scheme unnecessary * `implicit_region_bound` could be retrieved from `RegionMaps` * renaming `CodeExtent` to `Scope` * generalizing "call site" to "use site" or something better to include constants * renaming `RegionMaps` to `ScopeTree` and its API to talk about "parents" explicitly

bors · 2017-05-13T20:19:20Z

☀️ Test successful - status-appveyor, status-travis
Approved by: nikomatsakis
Pushing 826d8f3 to master...

rust-highfive assigned pnkfelix May 11, 2017

rust-highfive assigned nikomatsakis and unassigned pnkfelix May 11, 2017

nikomatsakis requested changes May 11, 2017

View reviewed changes

arielb1 reviewed May 11, 2017

View reviewed changes

eddyb force-pushed the region-refactor branch from 457e481 to bb37801 Compare May 11, 2017 21:57

alexcrichton added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label May 11, 2017

eddyb force-pushed the region-refactor branch from bb37801 to 33c653e Compare May 12, 2017 08:57

eddyb mentioned this pull request May 12, 2017

Lifetime scope note messages have off-by-one column numbers. #41938

Closed

eddyb force-pushed the region-refactor branch from 33c653e to 788e70e Compare May 12, 2017 11:05

nikomatsakis reviewed May 12, 2017

View reviewed changes

nikomatsakis approved these changes May 12, 2017

View reviewed changes

Mark-Simulacrum mentioned this pull request May 12, 2017

Rollup of 15 pull requests #41954

Closed

eddyb added 7 commits May 13, 2017 17:42

rustc: uniformly compute ParameterEnvironment's "free outlive scope".

bd8246b

rustc: use call_site_extent for closure environment free regions.

9809fe0

rustc_mir: remove unnecessary drop scope for constants.

72aab7a

rustc: use DefId instead of CodeExtent for FreeRegion's scope.

6d4c214

rustc: move liberate_late_bound_regions to rustc_typeck.

74c6788

rustc_typeck: do not leak late-bound lifetimes from bounds to closures.

dbae169

rustc: treat ReEarlyBound as free without replacing it with ReFree.

2da080e

eddyb added 3 commits May 13, 2017 17:42

rustc: remove redundant fn_id's from CodeExtentData.

da8b6e2

rustc: stop interning CodeExtent, it's small enough.

ea1c6df

rustc: don't keep a second reference to the HIR map in middle::region.

6da4123

eddyb force-pushed the region-refactor branch from 788e70e to 6da4123 Compare May 13, 2017 14:43

bors merged commit 6da4123 into rust-lang:master May 13, 2017

eddyb deleted the region-refactor branch May 14, 2017 02:58

Mark-Simulacrum mentioned this pull request Jun 6, 2017

Rust 1.19 regression, treebitmap 0.2.2 #42467

Closed

dtolnay mentioned this pull request Jan 28, 2018

Rust compiler panics when trying to reference impl-ed constants with Self #47814

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

rustc: simpler ParameterEnvironment and free regions. #41914

rustc: simpler ParameterEnvironment and free regions. #41914

eddyb commented May 11, 2017

rust-highfive commented May 11, 2017

nikomatsakis commented May 11, 2017

nikomatsakis left a comment

nikomatsakis May 11, 2017

eddyb May 11, 2017

nikomatsakis May 11, 2017

eddyb May 11, 2017

nikomatsakis commented May 11, 2017

arielb1 commented May 11, 2017

arielb1 May 11, 2017

eddyb May 11, 2017

nikomatsakis commented May 11, 2017

nikomatsakis May 12, 2017

nikomatsakis commented May 12, 2017

bors commented May 12, 2017

bors commented May 13, 2017

eddyb commented May 13, 2017

bors commented May 13, 2017

eddyb commented May 13, 2017

bors commented May 13, 2017

bors commented May 13, 2017

rustc: simpler ParameterEnvironment and free regions. #41914

rustc: simpler ParameterEnvironment and free regions. #41914

Conversation

eddyb commented May 11, 2017

rust-highfive commented May 11, 2017

nikomatsakis commented May 11, 2017

nikomatsakis left a comment

Choose a reason for hiding this comment

nikomatsakis May 11, 2017

Choose a reason for hiding this comment

eddyb May 11, 2017

Choose a reason for hiding this comment

nikomatsakis May 11, 2017

Choose a reason for hiding this comment

eddyb May 11, 2017

Choose a reason for hiding this comment

nikomatsakis commented May 11, 2017

arielb1 commented May 11, 2017

arielb1 May 11, 2017

Choose a reason for hiding this comment

eddyb May 11, 2017

Choose a reason for hiding this comment

nikomatsakis commented May 11, 2017

nikomatsakis May 12, 2017

Choose a reason for hiding this comment

nikomatsakis commented May 12, 2017

bors commented May 12, 2017

bors commented May 13, 2017

eddyb commented May 13, 2017

bors commented May 13, 2017

eddyb commented May 13, 2017

bors commented May 13, 2017

bors commented May 13, 2017