incr.comp.: Introduce the concept of anonymous DepNodes and use them for trait selection #42769
Conversation
arielb1
added
the
S-waiting-on-author
| if self.current_node().is_some() { | ||
| let source = self.make_node(v); | ||
| self.add_edge_from_current_node(|current| (source, current)) | ||
| match self.task_stack.last_mut() { |
There was a problem hiding this comment.
Wait, this can't be right: We are ignoring reads from anonymous tasks :/
There was a problem hiding this comment.
The OpenTask::Anon variant is missing from the match, so reads during anonymous tasks are just dropped. I should have made the match exhaustive indeed :)
I'm in the process of fixing this but there are some complications (especially around
rust/src/librustc/traits/fulfill.rs
Line 628 in bd32b1b
There was a problem hiding this comment.
Oh, I see what you mean. Indeed, this is broken. Ha, score one for my persnicketiness about being exhaustive. =)
src/librustc/dep_graph/shadow.rs
Outdated
| let forbidden_edge = if !ENABLED { | ||
| None | ||
| } else { | ||
| match env::var("RUST_FORBID_DEP_GRAPH_EDGE") { |
There was a problem hiding this comment.
Is there no value to this check anymore? Does it not make sense in the new environment?
There was a problem hiding this comment.
(It still seems fairly useful to me =)
| reads.push(v); | ||
| } | ||
| } | ||
| _ => { |
There was a problem hiding this comment.
Nit: exhaustive plz =)
|
@michaelwoerister Looks like this is waiting on some changes from you. |
|
@Mark-Simulacrum Yes, this PR is currently on hold because of an issue with caching |
|
I'm closing this in favor of #43028 and a subsequent PR that will introduce anonymous nodes without using them for trait selection yet. |
incr.comp.: Deduplicate some DepNodes and introduce anonymous DepNodes This is a parallel PR to the pending #42769. It implements most of what is possible in terms of DepNode re-opening without having anonymous DepNodes yet (#42298). r? @nikomatsakis
This PR implements "anonymous"
DepNodes, which areDepNodesthe identity of which is not known beforehand and is automatically computed from its set of input edges. These nodes allow for fine-grained dependency tracking in cases like trait selection where it is hard to come up with a goodDepNodeID.This is an important step towards implementing #42298.
r? @nikomatsakis