BTree: remove Ord bound from new

[nbdd0121]

K: Ord bound is unnecessary on BTree{Map,Set}::new and their Default impl. No elements exist so there are nothing to compare anyway, so I don't think "future proof" would be a blocker here. This is analogous to HashMap::new not having a K: Eq + Hash bound.

#79245 originally does this and for some reason drops the change to new and Default. I can see why changes to other methods like entry or symmetric_difference need to be careful but I couldn't find out any reason not to do it on new.

Removing the bound also makes the stabilisation of const fn new not depending on const trait bounds.

cc @steffahn who suggests me to make this PR.

r? @dtolnay

[dtolnay]

No elements exist so there are nothing to compare anyway

I don't think "future proof" would be a blocker here

I couldn't find out any reason not to do it

This is not as clear cut as the PR description implies.

I am hesitant about this because this API change makes it impossible to selectively tackle some of the code bloat caused by monomorphizations inside the internals of these collections. According to https://www.stroustrup.com/SCARY.pdf such optimizations deliver 1.2x to 2.1x faster performance and 1x to 25x smaller code size in real applications.

I pasted below a sketch of the kind of restructure that this change hinders.

Clearly we still have bounds available via the insert and contains methods etc, so a comparator could be passed down from there as a function argument to every places that needs it, but the extra argument passing could negate a chunk of the performance benefit relative to stashing a comparator at suitable places inside of the data structure up front.

@rust-lang/libs-api @rust-lang/libs

useself::private::{Element,TypeErasedSet};use std::cmp::Ordering;use std::marker::PhantomData;use std::mem;pubstructSet<T>{imp:TypeErasedSet,marker:PhantomData<T>,}impl<T>Set<T>{pubfnnew() -> SelfwhereT:Ord,{unsafefnerased_cmp<T:Ord>(a:&dynElement,b:&dynElement) -> Ordering{T::cmp(&*(a as*const_as*const_),&*(b as*const_as*const_))}Set{imp:unsafe{TypeErasedSet::new(erased_cmp::<T>)},marker:PhantomData,}}pubfninsert(&mutself,value:T)whereT:Ord,{let boxed = Box::new(value);unsafe{let erased = mem::transmute::<Box<dynElement>,Box<dynElement + 'static>>(boxed);self.imp.insert(erased)}}pubfncontains(&self,value:&T) -> boolwhereT:Ord,{unsafe{self.imp.contains(value)}}}mod private {use std::cmp::Ordering;pub(super)traitElement{}impl<Sized>ElementforSized{}typeComparator = unsafefn(&dynElement,&dynElement) -> Ordering;pub(super)structTypeErasedSet{// Obviously you could build a more efficient representation than this.// Just illustrating that the concept works without a monomorphized T.repr:Vec<Box<dynElement>>,cmp:Comparator,}implTypeErasedSet{pub(super)unsafefnnew(cmp:Comparator) -> Self{TypeErasedSet{repr:Vec::new(),
cmp,}}pub(super)unsafefninsert(&mutself,value:Box<dynElement>){matchself.find(&*value){Ok(present) => self.repr[present] = value,Err(absent) => self.repr.insert(absent, value),}}pub(super)unsafefncontains(&self,value:&dynElement) -> bool{self.find(value).is_ok()}unsafefnfind(&self,value:&dynElement) -> Result<usize,usize>{self.repr.binary_search_by(|elem| unsafe{(self.cmp)(&**elem, value)})}}}fnmain(){letmut set = Set::new();
set.insert(5);
set.insert(2);
set.insert(9);
set.insert(4);
set.insert(9);println!("{}", set.contains(&4));println!("{}", set.contains(&3));}

[nbdd0121]

Clearly we still have bounds available via the insert and contains methods etc, so a comparator could be passed down from there as a function argument to every places that needs it, but the extra argument passing could negate a chunk of the performance benefit relative to stashing a comparator at suitable places inside of the data structure up front.

I suppose this wouldn't be a big deal for BTreeMap, since when searching the tree we currently only take NodeRef without having the references to the BTreeMap itself. So:

• If the comparator is to be stored in BTreeMap but not within the nodes, then extra argument is needed anyway. This would mean that it doesn't make sense to store comparator in the BTreeMap.
• If the comparator is to be stored within nodes, then the comparator is not needed to create an empty BTreeMap because it starts with no nodes (root = None).

[steffahn]

Storing a comparator in the data structure might kill covariance and ignoring that issue might be unsound. I will investigate more on that later.

Passing a comparator with every call (to insert, etc) seems more reasonable to me. Especially since you'd need to "pass down" the function anyway if it's stored at top level. Whether it originates from a read operation or from a compile-time constant passed down from one layer above in the call hierarchy doesn't seem like the most significant difference. But I don't even know where the proclaimed performance benefit in that "SCARY" paper comes from (it's a bit lengthy, havent read through much of it yet) so I can't be sure to have the full picture. (The code size reduction is obvious on the other hand.)

Passing the comparator every call also allows for more flexible API, where you could offer a version of BTreeMap (using the same underlying erased data structure) that supports a comparator dependent on short-term borrows of another data structure. E. g. a BTreeSet of indices into another Vec would be possible. (Sorted by the Order of the Vec elements).

[steffahn]

@dtolnay

[…] ignoring that issue might be unsound. I will investigate more on that later.

Here’s a demonstration using your code example.

use std::any::Any;use once_cell::sync::OnceCell;fnmain(){let s = Set::<Wrapper<for<'a> fn(&'astr)>>::new();letmut s:Set<Wrapper<fn(&'staticstr)>> = s;staticS:OnceCell<&'staticstr> = OnceCell::new();let f = |s| {let _ = S.try_insert(s);};
s.insert(Wrapper(f));
s.contains(&Wrapper(f));println!("{}",S.get().unwrap());}// for the record, BTreeSet is indeed covariant, toouse std::collections::BTreeSet;fn_b_tree_set_is_covariant<'a:'b,'b,T>(x:BTreeSet<&'aT>) -> BTreeSet<&'bT>{
x
}structWrapper<T>(T);impl<T:'static>OrdforWrapper<T>{fncmp(&self,_other:&Self) -> Ordering{ifletSome(f) = (selfas&dynAny).downcast_ref::<Wrapper<for<'a> fn(&'astr)>>(){
f.0(&String::from("Hello world!"));}Ordering::Equal}}impl<T:'static>PartialOrd<Self>forWrapper<T>{fnpartial_cmp(&self,other:&Self) -> Option<Ordering>{Some(self.cmp(other))}}impl<T:'static>EqforWrapper<T>{}impl<T:'static>PartialEq<Self>forWrapper<T>{fneq(&self,other:&Self) -> bool{self.cmp(other).is_eq()}}

(in the playground)
Example output:

���)

---

This unsoundness is not a small bug in your code, but an inherent problem of storing a comparison function in a covariant collection. Since BTreeSet will never lose its covariance, I don’t think your argument against removing K: Ord bounds from BTreeMap::new can be salvaged.

[Amanieu]

Incidentally hashbrown already does something similar internally (passing the comparator as an argument) but for different reasons:

• It makes RawTable independent of the details of how a value is hashed or compared.
• It makes the RawTable API more flexible since the comparator closure can contain borrowed data that isn't tied to the lifetime of the table.

[steffahn]

Incidentally hashbrown already does something similar internally (passing the comparator as an argument)

It’s a very slight similarity only when considering the motivation discussed here (i.e. less monomorphization), since the hasher arguments are generic (impl Fn(&T) -> u64), so at run-time there’s no “passing” but instead still full monomorphization if the calls come from a wrapping HashMap (the passed function items are zero-sized types, not function pointers).

[dtolnay]

@rfcbot fcp merge

#88040 (comment) is persuasive as far as the Ord impl which exists during new being totally useless at any point after new has returned, so I am on board with this.

[rfcbot]

Team member @dtolnay has proposed to merge this. The next step is review by the rest of the tagged team members:

• @Amanieu
• @BurntSushi
• @dtolnay
• @joshtriplett
• @m-ou-se
• @yaahc

No concerns currently listed.

Once a majority of reviewers approve (and at most 2 approvals are outstanding), this will enter its final comment period. If you spot a major issue that hasn't been raised at any point in this process, please speak up!

See this document for info about what commands tagged team members can give me.

[steffahn]

@ anyone reviewing this, FYI:

this suggestion

cc @steffahn who suggests me to make this PR.

happened in a (short) topic on Zulip which can be found here (or in the archive).

[nbdd0121]

Surprised to see that covariance are not asserted, filed #88058

EDIT: It is asserted just using a different name from the other containers.

[bors]

☔ The latest upstream changes (presumably #88094) made this pull request unmergeable. Please resolve the merge conflicts.

[bors]

☔ The latest upstream changes (presumably #86808) made this pull request unmergeable. Please resolve the merge conflicts.

[rfcbot]

🔔 This is now entering its final comment period, as per the review above. 🔔

[stlankes]

Naive question... Would this a solution for issue #88244 ?

[nbdd0121]

Naive question... Would this a solution for issue #88244 ?

Yes

[steffahn]

I don’t think that this PR has anything to do with resolving or not resolving #88244. While it makes the particular code example compile again, there’s no reason why other trait bounds on other const functions shouldn’t have the same problem (w.r.t. usability in associated consts for trait impls). Especially the fact that it’s a change/regression in compiler behavior means that either #88244 would be closed as “working as intended” (breaking changes to unstable features are okay) or it needs a bugfix.

[rfcbot]

The final comment period, with a disposition to merge, as per the review above, is now complete.

As the automated representative of the governance process, I would like to thank the author for their work and everyone else who contributed.

The RFC will be merged soon.

[m-ou-se]

@bors r+

[bors]

📌 Commit f33f266 has been approved by m-ou-se