Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Make view of range with range be a range #26872

Merged
merged 5 commits into from
Apr 30, 2020
Merged

Make view of range with range be a range #26872

merged 5 commits into from
Apr 30, 2020

Conversation

andyferris
Copy link
Member

I had brought this one up on Slack a few months ago.

Typical AbstractRange types are compact and immutable, and we can remove the overhead of SubArray in this case while the cost of view remains O(1).

New behavior:

julia> view(1:10, 1:10)
1:10

@andyferris andyferris added arrays [a, r, r, a, y, s] performance Must go faster labels Apr 21, 2018
@StefanKarpinski StefanKarpinski requested a review from mbauman April 21, 2018 17:18
@andyferris
Copy link
Member Author

Bump

@mbauman
Copy link
Member

mbauman commented May 1, 2018

Yeah, I think I'm in favor here. We should probably also document this constraint for AbstractRange — that indexing a range by a range must return a range — but unfortunately we don't have any existing documentation there, so that's a bigger ask.

@andyferris
Copy link
Member Author

Before merging - I'm wondering now about some of the floating-point unit ranges. Is it really even possible for all of these getindex operations to return an AbstractRange? Some algorithms go through hoops to hit their end points but I wonder if a subset of a LinSpace is necessarily a LinSpace (to full floating point precision), for example.

One option is to limit this signature to integer ranges... thoughts?

@mbauman
Copy link
Member

mbauman commented May 2, 2018

Yes, that's a good consideration. I'm not sure, either. We could simply duplicate the getindex signatures we have. That's probably the best option.

julia> methods(getindex, Tuple{AbstractRange, AbstractRange})
# 8 methods for generic function "getindex":
[1] getindex(r::Base.OneTo{T}, s::Base.OneTo) where T in Base at range.jl:537
[2] getindex(r::AbstractUnitRange, s::AbstractUnitRange{#s57} where #s57<:Integer) in Base at range.jl:529
[3] getindex(r::AbstractUnitRange, s::StepRange{#s57,S} where S where #s57<:Integer) in Base at range.jl:543
[4] getindex(r::StepRange, s::AbstractRange{#s57} where #s57<:Integer) in Base at range.jl:550
[5] getindex(r::StepRangeLen{T,#s57,#s56} where #s56<:Base.TwicePrecision where #s57<:Base.TwicePrecision, s::OrdinalRange{#s55,S} where S where #s55<:Integer) where T in Base at twiceprecision.jl:460
[6] getindex(r::StepRangeLen{T,R,S} where S where R, s::OrdinalRange{#s57,S} where S where #s57<:Integer) where T in Base at range.jl:557
[7] getindex(r::LinRange, s::OrdinalRange{#s57,S} where S where #s57<:Integer) in Base at range.jl:567

@andyferris andyferris force-pushed the ajf/view-of-range branch from d9006dc to afd6d10 Compare May 17, 2018 04:41
@andyferris
Copy link
Member Author

andyferris commented May 17, 2018

I've now specialized on the first three four - the typical integer cases. Unlike my earlier approach, this should at least be perfectly safe.

Typical `AbstractRange` types are compact and immutable, and we can
remove the overhead of `SubArray` in this case while the cost of `view`
remains O(1).

Here we specialize only on valid combinations of integer ranges,
otherwise still relying on `SubArray`.
@andyferris andyferris force-pushed the ajf/view-of-range branch from afd6d10 to f49cee2 Compare May 17, 2018 06:17
@JeffBezanson
Copy link
Member

Bump. Seems useful.

@mbauman
Copy link
Member

mbauman commented Aug 16, 2018

Agreed! Let's do this. Needs a rebase. Also, do we now need NEWS? How conservatively are we going to be with regards to annotating changes like this that change types but not behavior in 1.x?

@fredrikekre fredrikekre added the needs news A NEWS entry is required for this change label Aug 16, 2018
@timholy
Copy link
Member

timholy commented Aug 16, 2018

I like it now that it's safe. Like @mbauman, my head spins with questions about whether changing return type is a breaking change.

@JeffBezanson
Copy link
Member

I think it's too breaking for a 1.0.x release, but might be ok for 1.1. In particular it's ok if the new type supports the same interface as the old one.

@andyferris
Copy link
Member Author

Cool.

Do we have a branch for 1.1 for this to target?

@andyferris
Copy link
Member Author

andyferris commented Aug 16, 2018

I think the problem of answering the question “how breaking it is this?” is that we don’t have a rigorous software engineering answer to “what is an interface?”

Ideally the type system or something would check that you haven’t changed the interface of the return type that you have documented as that which should be assumed by the user of your function. Or something - it’s a tricky thing, it seems.

In this case what is the interface assumed by a view? That it is cheap (O(1) and no large memory allocations)? That it tracks mutations to the parent? Both? Something else? How is this documented and how could this possibly be verified by tests or the type system or the compiler?

Tricky, tricky.

@StefanKarpinski
Copy link
Member

StefanKarpinski commented Aug 17, 2018

Since this is technically breaking I think it should definitely not be done in a point release—so not in a 1.0.x release. However, if we run PkgEval on the entire registered ecosystem and nothing breaks, I think we should go ahead with it in a 1.x release. I feel that we should take this approach in general: anything that is mildly technically breaking but doesn't actually break any packages can be changed in a minor release.

@fredrikekre
Copy link
Member

It could still break things in the future.

@StefanKarpinski
Copy link
Member

StefanKarpinski commented Aug 17, 2018

Yes, of course, that's what "technically breaking" means—that something could potentially break. The point of running PkgEval on all registered packages is to see if people seem to actually be relying on the behavior that is changing. If not a single registered package is, that's strong evidence that in practice no one cares about the behavioral change and therefore it's ok to make it.

Keep in mind that all changes are potentially breaking, including so-called "bug fixes", which are just breaking changes where we have made a judgement call that no one is likely to have been relying on the old behavior. Making no changes to the language that could potentially break anyone's code would mean not changing anything at all, which is clearly not desirable or practical.

@mbauman
Copy link
Member

mbauman commented Oct 31, 2019

Triage generally liked this — it just needs a rebase and NEWS.

While technically breaking, this is going from a more complicated (immutable) AbstractArray to a much simpler (immutable) one. If folks are dispatching on a particular AbstractArray type in the first place, they'll probably like this even better.

@JeffBezanson
Copy link
Member

Bump

@JeffBezanson JeffBezanson removed the triage This should be discussed on a triage call label Nov 14, 2019
@mbauman
Copy link
Member

mbauman commented Nov 14, 2019

Looks like we're missing a few methods here:

julia> for m in methods(getindex, Tuple{AbstractRange, AbstractRange})
           tt = Base.tuple_type_tail(m.sig)
           tt == Tuple{AbstractArray,Vararg{Any,N}} where N && continue
           vmt = collect(methods(view, Tuple{AbstractRange, AbstractRange}))
           vm = findfirst(sig->tt <: Base.tuple_type_tail(sig.sig), vmt)
           if vmt[vm].sig == Tuple{typeof(view),AbstractArray,Vararg{Any,N}} where N
               println(tt)
           end
       end
Tuple{StepRangeLen{T,#s75,#s74} where #s74<:Base.TwicePrecision where #s75<:Base.TwicePrecision,OrdinalRange{#s73,S} where S where #s73<:Integer} where T
Tuple{StepRangeLen{T,R,S} where S where R,OrdinalRange{#s75,S} where S where #s75<:Integer} where T
Tuple{LinRange,OrdinalRange{#s75,S} where S where #s75<:Integer}

The above should probably be converted into a unit test.

@StefanKarpinski
Copy link
Member

Bump. This seems like it would be good to go with Matt's comment addressed.

mbauman added 3 commits April 28, 2020 15:18
* origin/master: (833 commits)
  Improve typesubtract for tuples (#35600)
  Make searchsorted*/findnext/findprev return values of keytype (#32978)
  fix buggy rand(RandomDevice(), Bool) (#35590)
  remove `Ref` allocation on task switch (#35606)
  Revert "partr: fix multiqueue resorting stability" (#35589)
  exclude types with free variables from `type_morespecific` (#35555)
  fix small typo in NEWS.md (#35611)
  enable inline allocation of structs with pointers (#34126)
  SparseArrays: Speed up right-division by diagonal matrices (#35533)
  Allow hashing 1D OffsetArrays
  NEWS item for introspection macros (#35594)
  Special case empty covec-diagonal-vec product (#35557)
  [GCChecker] fix a few tests by looking through casts
  Use norm instead of abs in generic lu factorization (#34575)
  [GCChecker,NFC] run clang-format -style=llvm
  [GCChecker] fix tests and add Makefile
  Add introspection macros support for dot syntax (#35522)
  Specialize `union` for `OneTo` (#35577)
  add pop!(vector, idx, [default]) (#35513)
  bump Pkg version (#35584)
  ...
@mbauman mbauman removed the needs news A NEWS entry is required for this change label Apr 28, 2020
@mbauman
Copy link
Member

mbauman commented Apr 29, 2020

I added the missing methods, a test, and NEWS. The win32 failure looks unrelated — this is now good to go!

@andyferris
Copy link
Member Author

Awesome - thanks Matt!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
arrays [a, r, r, a, y, s] performance Must go faster
Projects
None yet
Development

Successfully merging this pull request may close these issues.

6 participants