fixedstring: a @safe, @nogc string type

[H. S. Teoh]

On Tue, Jan 11, 2022 at 11:16:13AM +0000, Moth via Digitalmars-d-announce wrote:
> On Tuesday, 11 January 2022 at 03:20:22 UTC, Salih Dincer wrote:
> > [snip]
> 
> glad to hear you're finding it useful! =]

One minor usability issue I found just glancing over the code: many of
your methods take char[] as argument. Generally, you want const(char)[]
instead, so that it will work with both char[] and immutable(char)[].
No reason why you can't copy some immutable chars into a FixedString,
for example.

Another potential issue is with the range interface. Your .popFront is
implemented by copying the entire buffer 1 char forwards, which can
easily become a hidden performance bottleneck. Iteration over a
FixedString currently is O(N^2), which is a problem if performance is
your concern.

Generally, I'd advise not conflating your containers with ranges over
your containers: I'd make .opSlice return a traditional D slice (i.e.,
const(char)[]) instead of a FixedString, and just require writing `[]`
when you need to iterate over the string as a range:

	FixedString!64 mystr;
	foreach (ch; mystr[]) { // <-- iterates over const(char)[]
		...
	}

This way, no redundant copying of data is done during iteration.

Another issue is the way concatenation is implemented. Since
FixedStrings have compile-time size, this potentially means every time
you concatenate a string in your code you get another instantiation of
FixedString. This can lead to a LOT of template bloat if you're not
careful, which may quickly outweigh any benefits you may have gained
from not using the built-in strings.


> hm, i'm not sure how i would go about fixing that double character
> issue. i know there's currently some wierdness with wchars / dchars
> equality that needs to be fixed [shouldn't be too much trouble, just
> need to set aside the time for it], but i think being able to tell how
> many chars there are in a glyph requires unicode awareness? i'll look
> into it.
[...]

Yes, you will require Unicode-awareness, and no, it will NOT be as
simple as you imagine.

First of all, you have the wide-character issue: if you're dealing with
anything outside of the ASCII range, you will need to deal with code
points (potentially wchar, dchar).  You can either take the lazy way out
(FixedString!(n, wchar), FixedString!(n, dchar)), but that will
exacerbate your template bloat very quickly. Plus, it wastes a lot of
memory, esp. if you start using dchar[] -- 4 bytes per character
potentially makes ASCII strings use up 4x more memory. (And even if you
decide using dchar[] isn't a concern, there's still the issue of
graphemes -- see below, which requires non-trivial decoding anyway.)

Or you can handle UTF-8, which is a better solution in terms of memory
usage. But then you will immediately run into the encoding/decoding
problem. Your .opSlice, for example, will not work correctly unless you
auto-decode. But that will be a performance hit -- this is one of the
design mistakes in hindsight that's still plaguing Phobos today. IMO the
better approach is to iterate over the string *without* decoding, but
just detecting codepoint boundaries.  Regardless, you will need *some*
way of iterating over code points instead of code units in order to deal
with this properly.

But that's only the beginning of the story. In Unicode, a "code point"
is NOT what most people imagine a "character" is. For most European
languages this is the case, but once you go outside of that, you'll
start finding things like accented characters that are composed of
multiple code points.  In Unicode, that's called a Grapheme, and here's
the bad news: the length of a Grapheme is technically unbounded (even
though in practice it's usually 2 or occasionally 3 -- but you *will*
find more on rare occasions). And worst of all, determining the length
of a grapheme requires an expensive, non-trivial algorithm that will
KILL your performance if you blindly do it every time you traverse your
string.

And generally, you don't *want* to do grapheme segmentation anyway --
most code doesn't even care what the graphemes are, it just wants to
treat strings as opaque data that you may occasionally want to segment
into substrings (and substrings don't necessarily require grapheme
segmentation to compute, depending on what the final goal is). But
occasionally you *will* need grapheme segmentation (e.g., if you need to
know how many visual "characters" there are in a string); for that, you
will need std.uni. And no, it's not something you can implement
overnight.  It requires some heavy-duty lookup tables and a (very
careful!) implementation of TR14.

Because of the foregoing, you have at least 4 different definitions of
the length of the string:

1. The number of code units it occupies, i.e., the number of chars /
wchars / dchars.

2. The number of code points it contains, which, in UTF-8, is a
non-trivial quantity that requires iterating over the entire string to
compute. Or you can just use wchar[] or dchar[], but then your memory
footprint will increase, potentially up to 4x.

3. The number of graphemes it contains, i.e., how many "visual
characters" (the way most people understand the word "character") it
contains. This requires grapheme segmentation, is expensive to compute,
and generally shouldn't be done unless you have some concrete reason why
you want to do this.

4. The rendered width of the string, i.e., how much space it occupies if
displayed on the screen. Even on a monospace-font text terminal, this is
a non-trivial quantity because some Unicode codepoints are double-width
(e.g., East Asian block), and some are *zero*-width (e.g., shy hyphens,
zero-width breaking spaces). And it depends on how your terminal
emulator renders these characters (what Unicode defines as a
double-width may not necessarily be rendered that way).  And of course,
on a GUI application measuring the length of a string requires font
details.

Welcome to the *cough* wonderful world of Unicode, where everything is
possible but nothing is simple. :-D


T

-- 
This sentence is false.