> On Sep 22, 2016, at 6:11 PM, Xiaodi Wu <xiaodi...@gmail.com> wrote:
> 
> On Thu, Sep 22, 2016 at 7:44 PM, Michael Gottesman <mgottes...@apple.com 
> <mailto:mgottes...@apple.com>> wrote:
> 
>> On Sep 22, 2016, at 5:09 PM, Xiaodi Wu <xiaodi...@gmail.com 
>> <mailto:xiaodi...@gmail.com>> wrote:
>> 
>> On Thu, Sep 22, 2016 at 6:54 PM, Michael Gottesman <mgottes...@apple.com 
>> <mailto:mgottes...@apple.com>> wrote:
>> 
>>> On Sep 22, 2016, at 4:19 PM, Xiaodi Wu <xiaodi...@gmail.com 
>>> <mailto:xiaodi...@gmail.com>> wrote:
>>> 
>>> You mean values of type String?
>> 
>> I was speaking solely of constant strings.
>> 
>>> I would want those to be exactly what I say they are; NFC normalization is 
>>> available, if I recall, as part of Foundation, but by no means should my 
>>> String values be silently changed!
>> 
>> Why.
>> 
>> For one, I don't want to pay the computational cost of normalization at 
>> runtime unless necessary.
> 
> This would only happen with strings that are known to be constant at compile 
> time (and as such the transformation would occur at compile time). There 
> would be no runtime cost.
> 
> Yes, for constant strings only there would be no runtime cost.
>  
> 
>> For another, I expect to be able to round-trip user input. 
> 
> String checks for canonical equivalence, IIRC.
> 
> Sure, but I'm not talking about using comparison operators here. I mean that 
> if we have `let str = "[some non-NFC string]"`, I should be able to write 
> that out to a file with all the non-canonical glyphs intact.

I would argue that most people that is not an interesting distinction. 
Naturally there would be a way to escape such canonicalization to get the 
non-canonicalized String.

> 
> There are known issues with NFC that are acceptable for normalizing Swift 
> identifiers but make it unsuitable for general use. For example, the 
> normalized form of Greek ano teleia is middle dot, but these two glyphs are 
> rendered differently in many fonts, and substituting a middle dot in place of 
> the Greek punctuation mark is actually quite inadequate for Greek text (ano 
> teleia is supposed to be around x-height; middle dot is not). Even for 
> constant strings, it is essential that one can output ano teleia when it is 
> specified rather than middle dot. However, Unicode normalization algorithms 
> guarantee stability and will forever require swapping the former for the 
> latter. I understand that other such problematic characters exist.

I would argue that that is a problem with the unicode standard and with the 
fonts. This is not a problem for Swift to solve.

> 
>> Normalization is not lossless and cannot be reversed. Finally, if I want to 
>> use normalization form D (NFD), your proposal
>> would make it impossible, because (IIUC) serial NFC + NFD normalization can 
>> produce different output than NFD normalization alone.
> 
> Why would you want to do this/care about this? I.e. what is the use case?
> 
> Use cases for NFD include searching, where you'd find substrings considered 
> "compatible." For instance, the fi ligature is considered compatible with the 
> letters f and i, but they are not equal. If you've ever successfully searched 
> for a word like "finance" in a PDF document that's been typeset with 
> ligatures, you've benefited from NFD. Roughly speaking (IIUC), the difference 
> between searching NFC-normalized strings and NFD-normalized strings is 
> analogous to the difference between a case-sensitive and a case-insensitive 
> search. Therefore, given a string x, it's sometimes important to be able to 
> obtain NFD(x). If every string x is now automatically NFC(x), then the best 
> one can do is NFD(NFC(x)), which is not guaranteed equal to NFD(x) even with 
> canonical comparison (i.e. NFC(NFD(NFC(x))) != NFC(NFD(x)) for all x).

There are issues here related to String design. For instance, one could make an 
argument that such searching is really only interesting for a "Text" use case 
which is different from a String use case. That being said, I don't want to 
argue about this here since we are hijacking this thread ; ).

>  
> As an aside, I am not formally proposing this. I am just discussing potential 
> opportunities for optimization given that we would need (as apart of this 
> proposal) to add knowledge of unicode to the compiler which would allow for 
> compile time transformations.
> 
> I'd be interested to know what performance gains you're envisioning with such 
> an optimization of constant strings at compile time.

I would have to measure such wins to say anything concrete. Algorithmically one 
would be able to avoid normalization during common unicode operations when you 
know you are using constant strings. Even though this may provide a runtime 
win, the major win from teaching the compiler about unicode would be in terms 
of applying unicode operations such as encoding/decoding to constant strings.

That being said, this is not the proposal that is being discussed here or even 
being proposed here. [i.e. lets stop hijacking this thread ; )]

> 
>>> On Thu, Sep 22, 2016 at 6:10 PM, Michael Gottesman <mgottes...@apple.com 
>>> <mailto:mgottes...@apple.com>> wrote:
>>> 
>>> > On Sep 22, 2016, at 10:50 AM, Joe Groff via swift-evolution 
>>> > <swift-evolution@swift.org <mailto:swift-evolution@swift.org>> wrote:
>>> >
>>> >
>>> >> On Jul 26, 2016, at 12:26 PM, Xiaodi Wu via swift-evolution 
>>> >> <swift-evolution@swift.org <mailto:swift-evolution@swift.org>> wrote:
>>> >>
>>> >> +1. Even if it's too late for Swift 3, though, I'd argue that it's 
>>> >> highly unlikely to be code-breaking in practice. Any existing code that 
>>> >> would get tripped up by this normalization is arguably broken already.
>>> >
>>> > I'm inclined to agree. To be paranoid about perfect compatibility, we 
>>> > could conceivably allow existing code with differently-normalized 
>>> > identifiers with a warning based on Swift version, but it's probably not 
>>> > worth it. It'd be interesting to data-mine Github or the iOS Swift 
>>> > Playgrounds app and see if this breaks any Swift 3 code in practice.
>>> 
>>> As an additional interesting point here, we could in general normalize 
>>> unicode strings. This could potentially reduce the size of unicode 
>>> characters or allow us to constant propagate certain unicode algorithms in 
>>> the optimizer.
>>> 
>>> >
>>> > -Joe
>>> > _______________________________________________
>>> > swift-evolution mailing list
>>> > swift-evolution@swift.org <mailto:swift-evolution@swift.org>
>>> > https://lists.swift.org/mailman/listinfo/swift-evolution 
>>> > <https://lists.swift.org/mailman/listinfo/swift-evolution>
_______________________________________________
swift-evolution mailing list
swift-evolution@swift.org
https://lists.swift.org/mailman/listinfo/swift-evolution

Reply via email to