deprecate API's with string/seq + start/last in favor of ones with openArray

[timotheecour]

proposal

• for all new APIs, instead of passing a string/seq with optional start/end params, use a single openArray[T] param (openArray[char] for string)
• for existing APIs, deprecate existing such routines and add an overload taking a openArray[T] param, while ensuring that the openArray overload is picked when no start/end param is passed (see https://github.com/nim-lang/Nim/pull/18173#issuecomment-855012363 for how to do that)
• if the string/seq is a var param, a VM+js limitation (https://github.com/nim-lang/Nim/issues/15952) currently prevents this, but that bug is fixable (refs https://github.com/nim-lang/Nim/issues/15952#issuecomment-786312417)
• the only exception is when a lookbehind/lookahead behavior is needed, as in regex, refs https://github.com/nim-lang/Nim/issues/14284#issuecomment-626562406

benefits

• simpler APIs (less overloads needed) and separation of concern (callee doesn't need to care where the slice came from), less code bloat
• no broken edge cases (see example 2 below), no design by convention
• more flexible

example 1

• findBounds has 6 overloads (https://github.com/nim-lang/Nim/pull/18028#issue-645255843), and even then it's still not flexible enough, instead we should have a single proc:

  proc findBounds*(input: openArray[char], pattern: Regex, matches: var openArray[Slice[int]],
  start = 0): tuple[Slice[int]]

  which would subsume all others and also be strictly more flexible, and less error prone

refs https://github.com/nim-lang/Nim/pull/18028#issuecomment-841790728

example 2

(refs https://github.com/nim-lang/Nim/pull/18173) optional start/last parameters lead to bad API's where a valid value ends up hijacked with a different meaning, eg:

func find*(a: SkipTable, s, sub: string, start: Natural = 0, last = 0): int

here, when user specifies last = 0, it ends up meaning last = sub.high, so there's no way for user to specify the edge condition of last = 0, which should be a valid choice

changing the signature to func find*(a: SkipTable, s, sub: string, start: Natural = 0, last = -1): int would be an improvement but is still unsatisfactory, as this could equally be interpreted as meaning s.high or empty slice of s (EDIT: func find*(a: SkipTable, s, sub: string, start: Natural = 0, last = sub.high): int would be possible too)

instead, func find*(a: SkipTable, s: openArray[char], sub: string): int would be self-documenting and not have any edge cases, the user then would call simply toOpenArray(s, start, end) if they want a slice

[Araq]

here, when user specifies last = 0, it ends up meaning last = sub.high, so there's no way for user to specify the edge condition of last = 0, which should be a valid choice

Well but Nim allows for

func find*(a: SkipTable, s, sub: string, start: Natural = 0, last = sub.high): int

it's just that the stdlib was written before this feature existed (and I personally don't like the feature).

It's a minor point, but RFCs should be correct. :-)

Regarding the RFC itself, IMO we can simply write new code with openArray and leave the old code untouched, I don't enjoy deprecating stuff that works well. Yes really, it does work well -- you can use Nim as a better Basic thanks to strutils and openArray soon enough will bite back with the required ownership rules.

[timotheecour]

It's a minor point, but RFCs should be correct. :-)

yes, I've updated RFC and included that sub.high as alternative

leave the old code untouched, I don't enjoy deprecating stuff that works well.

except that old code tends to not use sub.high and falls into "example 2" category, causing edge conditions for otherwise valid params (eg https://github.com/nim-lang/Nim/pull/18173, there are/were other similar cases). Also uniformity in APIs (when applied to best practices) is still a good thing.

So we can still introduce the openArray overload for old APIs (with the special care described in https://github.com/nim-lang/Nim/pull/18173#issuecomment-855012363 to ensure the openArray overload is selected when stard/end is unspecified), and whether to deprecate/soft-deprecate (with just a doc comment)/no-deprecate is secondary, at least new code can use old APIs using new style.

soon enough will bite back with the required ownership rules

do you have an example?

[Araq]

do you have an example?

You cannot write proc echo(x: varargs[openArray[char]]). Now you can argue that this should be allowed, as you can for all the other cases I could come up with, but this means we need to refine first-class openArray support first, make it non-experimental etc -- before we can build new APIs on top of it.