Refactor std::path to use default methods #8909

lkuper · 2013-09-01T06:08:08Z

Parts I'm unsure about and would like a reviewer to look at are:

pub trait GenericPath : Clone + Eq + ToStr -- is this the done thing? I've never done trait inheritance before, let alone from multiple traits, but it seemed to be necessary to be able to call all the methods we have to be able to call on self.
changing the argument of components from self to &self, and having it return self.components.clone() instead of self.components; this was necessary to avoid move errors, but I'm not sure if it's the right thing. (The default methods impls now all have to call self.components() instead of just referencing the field self.components.)

alexcrichton · 2013-09-01T16:16:17Z

src/libstd/path.rs

@@ -57,54 +57,115 @@ pub fn PosixPath(s: &str) -> PosixPath {
    GenericPath::from_str(s)
 }

-pub trait GenericPath {
-    /// Converts a string to a Path
+pub trait GenericPath : Clone + Eq + ToStr {


It looks like this API is pretty old, and more recent apis have since been attempting to remove as many allocations as possible. It appears that this API itself is performing a lot of allocations all over the place when most of them aren't necessary. I'm not sure if you'd want to look into this in this pull request or possibly leave it to later, though. If you're modifying Path already it seems like a good time, but if you're busy, then no worries!

Regardless, this inheritance is fine (especially if the traits are usable in the default implementations of the methods). If you can remove them, however, feel free to. I'd love to drop the Clone bound, but that's coupled with changing these functions to perform fewer allocations anyway.

Re: "especially if the traits are usable in the default implementations of the methods" Yes, WindowsPath and PosixPath were already deriving from Clone, and now that some of the code from those has been moved up into default methods, I had to do this in order to use the traits in the default methods.

I'll see what I can do about having this code perform fewer allocations, but I probably will need some advice and examples.

I'm having a go at revising the Path API to have fewer allocations built in.

But its a bigger job than alexcrichton might have realized, because some clients (e.g. rustpkg) are taking advantage of the fact that the accessors return owned strings. Its not hard to fix, but just a bigger job and probably worthy of a separate ticket. It does not have to block landing lkuper's work here, if someone else wants to double check the concerns lkuper raised in the ticket description.

I believe @kballard is also experimenting with rewriting std::path.

I am. I'm pretty far along. I told @lkuper in IRC that refactoring to use default methods is fine and won't interfere with my work, but doing any more work on path is probably not a good idea as I'm replacing it entirely.

I'm currently on vacation but I will be resuming my work on the new path next week.

Hmm. I think I just finished the task of revising the current Path API to have fewer allocations built in. I'm going to go ahead and file the PR for it; if we end up switching to @kballard's API later, fine, but I figure I'll be happier if we fix this now.

(PR filed as #8978, just to tie off that task.)

@pnkfelix OK, now that #8978 has landed I'll rebase this. Thanks.

Rebased and updated the patch.

…py, r=huonw A [dialogue](#8909 (diff)) on PR #8909 inspired me to make this change. r? anyone (It is possible that `std::path` itself will soon be replaced with a new implementation that kballard's working on, as mentioned in the dialogue linked above, but this revision is simple enough that I figured I'd offer it up.)

alexcrichton · 2013-09-05T06:13:39Z

src/libstd/path.rs

+    fn filename<'a>(&'a self) -> Option<&'a str> {
+        match self.components().len() {
+            0 => None,
+            n => Some(self.components()[n - 1].clone()),


Does this actually compile? I'd be a little surprised if it did. Regardless I don't think the clone is necessary.

@alexcrichton what part of it did you think looked suspcious? The use of 0 in a match pattern?

The clone should create an ~ pointer which does not have the same lifetime as self, yet you're returning it as the same lifetime.

Ah! Scary!

Okay, I just tried out something similar in a little example, and the type checker yelled at me (expected &'a but found ~). So you are right: this code shouldn't compile, and probably does not.

@pnkfelix @alexcrichton Yes, you were right, this doesn't compile, and shouldn't. I messed up the merge, then was sloppy and didn't wait for LLVM to build. I just updated the patch to a version that actually compiles and passes make check-stage1-std.

alexcrichton · 2013-09-05T06:15:50Z

lgtm with removal of allocations

lkuper · 2013-09-05T18:07:05Z

Updated patch to incorporate #8978.

r? anyone

…chton (cc: #3227) Parts I'm unsure about and would like a reviewer to look at are: * `pub trait GenericPath : Clone + Eq + ToStr` -- is this the done thing? I've never done trait inheritance before, let alone from multiple traits, but it seemed to be necessary to be able to call all the methods we have to be able to call on `self`. * changing the argument of `components` from `self` to `&self`, and having it return `self.components.clone()` instead of `self.components`; this was necessary to avoid move errors, but I'm not sure if it's the right thing. (The default methods impls now all have to call `self.components()` instead of just referencing the field `self.components`.)

…py, r=huonw A [dialogue](rust-lang/rust#8909 (diff)) on PR #8909 inspired me to make this change. r? anyone (It is possible that `std::path` itself will soon be replaced with a new implementation that kballard's working on, as mentioned in the dialogue linked above, but this revision is simple enough that I figured I'd offer it up.)

alexcrichton reviewed Sep 1, 2013
View reviewed changes

pnkfelix mentioned this pull request Sep 4, 2013

Make path api less allocation happy #8978

Closed

alexcrichton reviewed Sep 5, 2013
View reviewed changes

lkuper added 2 commits September 5, 2013 06:25

Factor shared code out into default GenericPath methods.

8651011

Minor doc cleanup.

59291eb

bors closed this Sep 6, 2013

lkuper mentioned this pull request Sep 6, 2013

De-duplicate methods in core::path(2) when default trait methods work #3227

Closed

Refactor std::path to use default methods #8909

Refactor std::path to use default methods #8909

Uh oh!

Conversation

lkuper commented Sep 1, 2013

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alexcrichton commented Sep 5, 2013

Uh oh!

lkuper commented Sep 5, 2013

Uh oh!

Uh oh!