Re: [PATCH] Allow inserting non-BMP characters

[Philipp Stephani <p.stephani2@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 3856 bytes --]

Eli Zaretskii <eliz@gnu.org> schrieb am Di., 26. Dez. 2017 um 17:11 Uhr:

> > From: Philipp Stephani <p.stephani2@gmail.com>
> > Date: Tue, 26 Dec 2017 10:35:42 +0000
> > Cc: emacs-devel@gnu.org, phst@google.com
> >
> >  Suggest to move surrogates_to_codepoint to coding.c, and then use the
> >  macros UTF_16_HIGH_SURROGATE_P and UTF_16_LOW_SURROGATE_P defined
> >  there.
> >
> > Hmm, I'd rather go the other way round and remove these macros later.
> They are macros, thus worse than
> > functions,
>
> I don't think we have a policy to prefer inline functions to macros,
> and I don't think we should have such a policy.  We use inline
> functions when that's necessary, but we don't in general prefer them.
> They have their own problems, see the comments in lisp.h for some of
> that.
>

Thanks, the only discussion I saw there was about some performance issues:

   Some operations are so commonly executed that they are implemented
   as macros, not functions, because otherwise runtime performance would
   suffer too much when compiling with GCC without optimization.

   FIXME: Remove the lisp_h_OP macros, and define just the inline OP
   functions, once "gcc -Og" (new to GCC 4.8) works well enough for
   Emacs developers.  Maybe in the year 2020.  See Bug#11935.

That bug says that GCC 4.8 should be available in Debian stable to support
-Og. Debian stable now already has 6.3 (
https://wiki.debian.org/DebianStretch#Packages_.26_versions). So maybe it's
time to compile development versions with -Og and try to reapply the
original patch. 5 years have passed, and compilers have gotten a lot better.

In any case, the new functions are certainly not commonly executed. They
are currently only executed when converting non-BMP keystrokes on macOS,
which is rare enough.



>
> > and don't seem to be correct either (what about a value such as
> 0x11DC00?).
>
> ??? They care correct for UTF-16 sequences, which are 16-bit numbers.
> If you need to augment them by testing the high-order bits to be zero
> in your case, that's okay, but I don't see any need for introducing
> similar but different functionality.
>

I'd be OK with using the macros since they already exist, but I wouldn't
want to touch them without converting them to functions first, and for
using them in nsterm.m I'd have to move them around.


>
> > No new macros please if we can avoid it. Functions are strictly better.
>
> Sorry, I disagree.  Each has its advantages, and on balance I find
> macros to be slightly better, certainly not worse.  There's no need to
> avoid them in C.
>

I disagree, see e.g. https://gcc.gnu.org/onlinedocs/cpp/Macro-Pitfalls.html and
many other sources.
Sometimes macros are unavoidable, but not here.


>
> > I don't care much whether they are in character.h or coding.h, but
> char_surrogate_p is already in character.h.
>
> char_surrogate_p should have used the coding.h macros as well.
>
> >  > +  USE_SAFE_ALLOCA;
> >  > +  unichar *utf16_buffer;
> >  > +  SAFE_NALLOCA (utf16_buffer, 1, len);
> >
> >  Maximum length of a UTF-16 sequence is known in advance, so why do you
> >  need SAFE_NALLOCA here?  Couldn't you use a buffer of fixed length
> >  instead?
> >
> > The text being inserted can be arbitrarily long. Even single characters
> (i.e. extended grapheme clusters) can
> > be arbitrarily long.
>
> Yes, but why do you first copy the input into a separate buffer?  Why
> not convert each UTF-16 sequence separately, as you go through the
> loop?
>

Message (method) invocations in Objective-C have high overhead because they
are late-bound. Therefore it is advisable to minimize the number of
messages sent.
https://developer.apple.com/documentation/foundation/nsstring/1408720-getcharacters?language=objc
also
indicates that a (properly implemented) getCharacters call is faster than
calling characterAtIndex in a loop.

[-- Attachment #2: Type: text/html, Size: 5500 bytes --]