On 9/8/19 12:02 PM, Steve Downey wrote:
Character repertoire sounds good, and I will eventually learn to spell it. Character set is definitely terminology from the pre-unicode times, and unfortunately tends to merge the repertoire and encoding, https://www.iana.org/assignments/character-sets/character-sets.xhtml

I think I was a little over zealous earlier in stating that Unicode uses "character repertoire" as I described.  I looked again and don't find that term formally defined in the standard.  However, "repertoire" is used throughout the standard in ways that I believe are consistent with my description.  I wasn't able to find an alternative formal term.

The way I've been thinking about it is that a "character repertoire" describes a set of abstract characters (a formal Unicode term) and a "character set" describes a set of encoded characters (a formal Unicode term) that associate each abstract character member of a "character repertoire" with a code point (a formal Unicode term) within a codespace (A formal Unicode term).  See sections 2.4 and 3.4 of Unicode 12 and uses of the word "repertoire" within those chapters.  The Unicode standard does use the term "character set", but I didn't find a formal definition.

Basic source character set is defined in [lex.charset] http://eel.is/c++draft/lex.charset#def:character_set,basic_source
Yes, and it defines a character repertoire.  "Physical source file characters" is the closest I've found to a term that describes the actual implementation defined source character set.

I'd like to get away from "execution encoding" because it conflates the presumed encoding and the one selected by the current locale. Now, admittedly, everyone conflates these and it's a source of error and mojibake, but perhaps with better words it would be easier to teach.
I agree.  I like "dynamic encoding" because it accurately reflects the reality that the encoding can be changed dynamically (by calls to std::setlocale).

As to UB. I'd like, if possible, to avoid creating new UB classes. Some things should probably be ill-formed, like unencodable characters. Others fall into existing UB, like specifying an inline string literal with two different encodings. Reading a string with the wrong encoding, I think, should be at worst unspecified, unless for some reason your decoder has UB, in which case it's the decoders problem, not the incorrect or mixed encoding isssue. That said, I'd defer to Core on this.
Wherever Core says we can get away with unspecified, I'm all for it.

Internal encoding is required to preserve distinct universal character names and treat all representations of the same universal character the same. So, the standard effectively requires unicode, but in terms of observables.

Agreed, I don't think anything is accomplished by trying to prescribe implementation details.


On Sun, Sep 8, 2019 at 5:39 AM Corentin Jabot <corentinjabot@gmail.com> wrote:

On Sun, 8 Sep 2019 at 05:46, Tom Honermann <tom@honermann.net> wrote:
On 9/5/19 9:41 PM, Steve Downey wrote:
Because I needed to circulate what I'm doing for Belfast, I've thrown together an abstract for the paper we've peripherally discussed about modernizing and tightening the specification around encodings of characters generally, and the source and execution character sets. 

This document proposes new standard terms for the various encodings for character and string literals, and the encodings associated with some character types. It also proposes that the wording used for [lex.charset], [lex.ccon], [lex.string], and [basic.fundamental] 8 be modified to reflect the new terminology. This paper does not intend to propose any changes that would require changes in any currently conforming implementation.

I'm hoping to have some preliminary work by the next telecon. The direction I'm thinking is that both Source and Execution Character Set are descriptions of the abstract characters, selected from 10646, that must be present to support C++. Encodings, both source and execution, are implementation defined. I would like to introduce terminology to describe the encoding used when translating narrow and wide character and string literals. I'd also like to make it explicit somewhere up front that there are associated encodings for some, but not all, character types. This is mentioned now in filesystem, but should be moved to a section with wider scope. The encoding for `char` and `wchar_t` is controlled by `locale`. The encoding for the unicode character types is fixed. The encoding used for literals was chosen at compile time, and is implementation defined. If locale and that endcoding conflict, behavior is unspecified. Combining TU with different encodings is in general unspecified, unless it results in an ODR violation.
This all sounds great.  My only question is behavior being unspecified vs undefined.  It seems challenging to get away with making it only unspecified.

Specifically, I'd like something along the line of:
If a character literal contains a c-char that do not have the same representation in the character literal encoding (aka *presumed" execution encoding) and the execution encoding, the behavior is undefined.


Some possible terms:
{"",Narrow,Wide} Literal Encoding - encoding on char and string literals
Dynamic Encoding - encoding implied by locale
*Character Set - A set of abstract characters ( Latin Capital letter A, Digit Zero, Left Parenthesis ...)
Unicode uses "character repertoire" for abstract sets of characters.  I favor following suit there.

+1 to sticking to Unicode terms 
*Basic Character Set - minimum required to be encoded
*Extended Character Set - what can be encoded
*Source Character Set - must be encodable in C++ source
I don't think "source character set" is defined today.  The closest we get is "Physical source file characters" in [lex.phases]p1.
*Execution Character Set - Source + control characters

More seriously i think it would be beneficial (necessary even) to have a source character encoding / character repertoire.

I wonder if we could specified that the internal character repertoire is Unicode. It kinda has to be already make that clearer.

I would also propose

Universal Character Name -> Unicode Code point
(character name should be reserved to the \N proposal)

* Current terms, with what I think the actual meanings are today.

I think these are good.  With these, there is no need for a term like "execution encoding", correct?  At compile-time, "literal encoding" encodes "execution character set" characters, and at run-time, "dynamic encoding" encodes "extended character set" characters, yes?

I prefer "execution" to dynamic

I like that this doesn't stray far from the existing terms.


SG16 Unicode mailing list