Implement Fuse with Option #70366

cuviper · 2020-03-24T19:21:28Z

The former done flag was roughly similar to an Option tag, but left
the possibity of misuse. By using a real Option, we can set None
when the iterator is exhausted, removing any way to call it again. We
also allow niche layout this way, so the Fuse may be smaller.

The FusedIterator specialization does want to ignore the possibility
of exhaustion though, so it uses unsafe { intrinsics::unreachable() }
to optimize that branch away. The entire Fuse implementation is now
isolated in its own module to contain that unsafety.

r? @scottmcm

The former `done` flag was roughly similar to an `Option` tag, but left the possibity of misuse. By using a real `Option`, we can set `None` when the iterator is exhausted, removing any way to call it again. We also allow niche layout this way, so the `Fuse` may be smaller. The `FusedIterator` specialization does want to ignore the possibility of exhaustion though, so it uses `unsafe { intrinsics::unreachable() }` to optimize that branch away. The entire `Fuse` implementation is now isolated in its own module to contain that unsafety.

cuviper · 2020-03-24T19:22:11Z

This was extracted from #70332 -- let's go ahead and get a perf run:

@bors try @rust-timer queue

rust-timer · 2020-03-24T19:22:12Z

Awaiting bors try build completion

bors · 2020-03-24T19:22:22Z

⌛ Trying commit 212e6ce with merge bad75affe428dbcd6e24d9ae6a248ac2e9364c7f...

cuviper · 2020-03-24T19:33:35Z

If this one also slows down codegen, I think it might help to use direct match instead of all those uses of as_ref(), as_mut(), and ?.

Mark-Simulacrum · 2020-03-24T19:50:20Z

src/libcore/iter/adapters/fuse.rs

+        match self.iter {
+            Some(ref iter) => iter,
+            // SAFETY: the specialized iterator never sets `None`
+            None => unsafe { intrinsics::unreachable() },


Ideally all of these should be hint::unreachable_unchecked() today I think

Why is that better? For core that's free to use intrinsics, going through hint just adds more inline indirection...

Clearer meaning, IMO. But not that important, certainly no need to change things in this PR.

Why is that better?

For example, if we make the hint method panic in debug-builds at some point (to detect incorrect use), then using the intrinsic would not benefit from that check.

One should note that this code triggers UB if an iterator implements FusedIterator by error. The implementor don't even have to use unsafe in order to achieve that.

I don't think so. Regardless of I's FusedIterator implementation or anything in I, it can't set the iter field of Fuse to None. This modules never sets iter to None if I is FusedIterator.

bors · 2020-03-24T21:50:59Z

☀️ Try build successful - checks-azure
Build commit: bad75affe428dbcd6e24d9ae6a248ac2e9364c7f (bad75affe428dbcd6e24d9ae6a248ac2e9364c7f)

rust-timer · 2020-03-24T21:51:01Z

Queued bad75affe428dbcd6e24d9ae6a248ac2e9364c7f with parent 2dcf54f, future comparison URL.

timvermeulen · 2020-03-25T05:18:22Z

Does this affect when drop handlers are called? (I don't know how important this is)

scottmcm · 2020-03-25T08:07:36Z

@timvermeulen Yes it does.

@rust-lang/libs I was going to r+ this, as it's a size win and seems fine in perf. But now that #70374 showed up too, I wanted to ask how you feel about different approaches here. (And to double-check that where Drop happens isn't something that needs to be preserved.)

dtolnay

This PR drops the impl TrustedRandomAccess for Fuse<I> if I is !FusedIterator, which seems unintentional.

  unsafe impl<I> TrustedRandomAccess for Fuse<I>
  where
-     I: TrustedRandomAccess,
+     I: TrustedRandomAccess + FusedIterator,

Otherwise looks good to me.

cuviper · 2020-03-25T16:12:36Z

@dtolnay that was semi-intentional, because what is that supposed to do if we've already set None? The old code just ignored done and called the iterator regardless.

But TrustedRandomAccess is an internal trait anyway, and all who implement that are already FusedIterator too. Maybe we should just formalize that?

dtolnay · 2020-03-25T16:33:44Z

The caller of get_unchecked guarantees that the fuse isn't already None, because they have a specific index at which they know an element exists. So you should be able to use the same approach as in as_inner_mut.

cuviper · 2020-03-25T18:11:54Z

@dtolnay OK, done!

dtolnay · 2020-03-25T18:13:52Z

@bors r+

bors · 2020-03-25T18:13:53Z

📌 Commit 4f429c0 has been approved by dtolnay

@scottmcm

Implement Fuse with Option The former `done` flag was roughly similar to an `Option` tag, but left the possibity of misuse. By using a real `Option`, we can set `None` when the iterator is exhausted, removing any way to call it again. We also allow niche layout this way, so the `Fuse` may be smaller. The `FusedIterator` specialization does want to ignore the possibility of exhaustion though, so it uses `unsafe { intrinsics::unreachable() }` to optimize that branch away. The entire `Fuse` implementation is now isolated in its own module to contain that unsafety. r? @scottmcm

@scottmcm

Implement Fuse with Option The former `done` flag was roughly similar to an `Option` tag, but left the possibity of misuse. By using a real `Option`, we can set `None` when the iterator is exhausted, removing any way to call it again. We also allow niche layout this way, so the `Fuse` may be smaller. The `FusedIterator` specialization does want to ignore the possibility of exhaustion though, so it uses `unsafe { intrinsics::unreachable() }` to optimize that branch away. The entire `Fuse` implementation is now isolated in its own module to contain that unsafety. r? @scottmcm

AnthonyMikh · 2020-03-25T19:12:05Z

@dtolnay since #70374 modifes Fuse too, may you take a look at that PR as well?

scottmcm · 2020-03-25T19:13:23Z

Thanks for taking this on, BTW, @cuviper!

@ghost

Rollup of 5 pull requests Successful merges: - rust-lang#70226 (use checked casts and arithmetic in Miri engine) - rust-lang#70319 (correctly normalize constants) - rust-lang#70352 (Add long error explanation for E0710 ) - rust-lang#70366 (Implement Fuse with Option) - rust-lang#70379 (fix incorrect type name in doc comments) Failed merges: - rust-lang#70375 (avoid catching InterpError) r? @ghost

Match options directly in the Fuse implementation Rather than using `as_ref()`, `as_mut()`, and `?`, we can use `match` directly to save a lot of generated code. This was mentioned as a possibility in rust-lang#70366 (comment), and I found that it had a very large impact on rust-lang#70332 using `Fuse` within `Chain`. Let's evaluate this change on its own first.

@scottmcm

Implement Chain with Option fuses The iterators are now "fused" with `Option` so we don't need separate state to track which part is already exhausted, and we may also get niche layout for `None`. We don't use the real `Fuse` adapter because its specialization for `FusedIterator` unconditionally descends into the iterator, and that could be expensive to keep revisiting stuff like nested chains. It also hurts compiler performance to add more iterator layers to `Chain`. This change was inspired by the [proposal](https://internals.rust-lang.org/t/proposal-implement-iter-chain-using-fuse/12006) on the internals forum. This is an alternate to rust-lang#70332, directly employing some of the same `Fuse` optimizations as rust-lang#70366 and rust-lang#70750. r? @scottmcm

rust-highfive assigned scottmcm Mar 24, 2020

rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label Mar 24, 2020

Mark-Simulacrum reviewed Mar 24, 2020

View reviewed changes

WaffleLapkin mentioned this pull request Mar 25, 2020

Make iter::Fuse truly zero cost #70374

Closed

dtolnay requested changes Mar 25, 2020

View reviewed changes

impl TrustedRandomAccess for Fuse without FusedIterator

4f429c0

dtolnay approved these changes Mar 25, 2020

View reviewed changes

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Mar 25, 2020

Dylan-DPC-zz mentioned this pull request Mar 25, 2020

Rollup of 4 pull requests #70402

Closed

This was referenced Mar 25, 2020

Rollup of 5 pull requests #70403

Closed

Rollup of 5 pull requests #70404

Merged

bors merged commit 530c320 into rust-lang:master Mar 25, 2020

AnthonyMikh mentioned this pull request Mar 28, 2020

Specialize internal state of Fuse depending on underlying iterator #70502

Closed

cuviper mentioned this pull request Apr 3, 2020

Implement Chain with Fuses #70332

Closed

cuviper deleted the option-fuse branch April 3, 2020 18:32

cuviper restored the option-fuse branch April 3, 2020 18:32

cuviper deleted the option-fuse branch April 3, 2020 18:37

cuviper mentioned this pull request Apr 3, 2020

Match options directly in the Fuse implementation #70750

Merged

cuviper mentioned this pull request Apr 7, 2020

Implement Chain with Option fuses #70896

Merged

cuviper mentioned this pull request Jun 30, 2021

iter::Fuse is unsound with how specialization currently behaves around HRTB fn pointers #85863

Closed

dtolnay assigned dtolnay and unassigned scottmcm Mar 24, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement Fuse with Option #70366

Implement Fuse with Option #70366

cuviper commented Mar 24, 2020

cuviper commented Mar 24, 2020

rust-timer commented Mar 24, 2020

bors commented Mar 24, 2020

cuviper commented Mar 24, 2020

Mark-Simulacrum Mar 24, 2020

cuviper Mar 24, 2020

Mark-Simulacrum Mar 24, 2020

RalfJung Mar 25, 2020 •

edited

Loading

dtolnay Mar 26, 2020

bors commented Mar 24, 2020

rust-timer commented Mar 24, 2020

timvermeulen commented Mar 25, 2020

scottmcm commented Mar 25, 2020 •

edited

Loading

dtolnay left a comment

cuviper commented Mar 25, 2020

dtolnay commented Mar 25, 2020

cuviper commented Mar 25, 2020

dtolnay commented Mar 25, 2020

bors commented Mar 25, 2020

AnthonyMikh commented Mar 25, 2020

scottmcm commented Mar 25, 2020

Implement Fuse with Option #70366

Implement Fuse with Option #70366

Conversation

cuviper commented Mar 24, 2020

cuviper commented Mar 24, 2020

rust-timer commented Mar 24, 2020

bors commented Mar 24, 2020

cuviper commented Mar 24, 2020

Mark-Simulacrum Mar 24, 2020

Choose a reason for hiding this comment

cuviper Mar 24, 2020

Choose a reason for hiding this comment

Mark-Simulacrum Mar 24, 2020

Choose a reason for hiding this comment

RalfJung Mar 25, 2020 • edited Loading

Choose a reason for hiding this comment

dtolnay Mar 26, 2020

Choose a reason for hiding this comment

bors commented Mar 24, 2020

rust-timer commented Mar 24, 2020

timvermeulen commented Mar 25, 2020

scottmcm commented Mar 25, 2020 • edited Loading

dtolnay left a comment

Choose a reason for hiding this comment

cuviper commented Mar 25, 2020

dtolnay commented Mar 25, 2020

cuviper commented Mar 25, 2020

dtolnay commented Mar 25, 2020

bors commented Mar 25, 2020

AnthonyMikh commented Mar 25, 2020

scottmcm commented Mar 25, 2020

RalfJung Mar 25, 2020 •

edited

Loading

scottmcm commented Mar 25, 2020 •

edited

Loading