P3216 R0 views::slice

[wg21bot]

P3216R0 views::slice (Hewill Kang)

[dhollman]

SG9 St. Louis

SG9 saw this paper in St. Louis. The following topics were discussed, with the following lists of arguments made on each side.

Length vs. End as Second Argument

Arguments in favor of second argument being size

• strided_slice in submdspan uses an size
• slice in valarray uses size and not end
• Ranges interaction with SIMD (length can eventually be static even if the first argument is dynamic)
• Separate SIMD thing should have similar semantics

Arguments in favor of second argument being end

• Consistent with Python and Ruby
• Ranges does not care for static extents
• SIMD thing should be separate
• range_v3 does this
• This is a different domain than mdspan and valarray

Not an issue:

• Safety (past-the-end truncates in either case)

Two only versus also three arguments

Two

• Avoid extra complexity to start with
• Avoid confusion with strided_slice
• It's really hard to read code with the third argument
• Colons in other languages are very different from argument lists
• range_v3 does this

Three

• People are going to come with papers anyway
• It's really easy to read for people used to Python/Matlab/Rust
• It's expected for people coming from Python/Matlab/Rust

Other considerations

• Negative offsets and/or strides
  • Extra branch
  • What if you have an infinite range
• New view versus combination of views
  • Most of the room was in favor of a new view here because this is so fundamental and common in other languages
  • range_v3 does this

A couple of polls were taken:

---

"SG9 prefers a new, dedicated view type over expression equivalence to a combination of take_view and drop_view"

Unanimous consent (weak quorum: 8 people) Note that the percentage of mdspan users was higher than usual

---

"SG9 approves these comments for posting on GitHub as a representation of the conversation in the room."

Unanimous consent