Reducing the cost of autodecoding
Patrick Schluter via Digitalmars-d
digitalmars-d at puremagic.com
Sat Oct 15 09:42:24 PDT 2016
On Saturday, 15 October 2016 at 00:50:08 UTC, Stefan Koch wrote:
> On Friday, 14 October 2016 at 20:47:39 UTC, Stefan Koch wrote:
>> On Thursday, 13 October 2016 at 21:49:22 UTC, safety0ff wrote:
>>>> Bad benchmark! Bad! -- Andrei
>>>
>>> Also, I suspect a benchmark with a larger loop body might not
>>> benefit as significantly from branch hints as this one.
>>
>> I disagree in longer loops code compactness is as important as
>> in small ones.
>>
>> This is about the smallest inline version of decode I could
>> come up with :
>>
>> __gshared static immutable ubyte[] charWidthTab = [
>> 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
>> 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
>> 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
>> 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 6, 6, 1, 1
>> ];
>>
>> dchar myFront(ref char[] str) pure nothrow
>> {
>> dchar c = cast(dchar) str[0];
>> if ((c & 128))
>> {
>> if (c & 64)
>> final switch(charWidthTab[c - 192])
>> {
>> case 2 :
>> c |= ((str[1] & 0x80) >> 5);
>> break;
>> case 3 :
>> c |= ((str[1] & 0x80) >> 4);
>> c |= ((str[2] & 0x80) >> 10);
>> break;
>> case 4 :
>> c |= ((str[1] & 0x80) >> 3);
>> c |= ((str[2] & 0x80) >> 9);
>> c |= ((str[3] & 0x80) >> 15);
>> break;
>> case 5,6,1 :
>> goto Linvalid;
>> }
>> else
>> Linvalid :
>> c = dchar.init;
>>
>> }
>> return c;
>> }
>
> Disregard all that code.
> It is horribly wrong!
>
> This is more correct : (Tough for some reason it does not pass
> the unittests)
>
> dchar myFront(ref char[] str) pure
> {
> dchar c = cast(dchar) str.ptr[0];
> if (c & 128)
> {
> if (c & 64)
> {
> auto l = charWidthTab.ptr[c - 192];
> if (str.length < l)
> goto Linvalid;
>
> final switch (l)
> {
> case 2:
> c = ((c & ~(64 | 128)) << 6);
> c |= (str.ptr[1] & ~0x80);
> break;
> case 3:
> c = ((c & ~(32 | 64 | 128)) << 12);
> c |= ((str.ptr[1] & ~0x80) << 6);
> c |= ((str.ptr[2] & ~0x80));
> break;
> case 4:
> c = ((c & ~(16 | 32 | 64 | 128)) << 18);
> c |= ((str.ptr[1] & ~0x80) << 12);
> c |= ((str.ptr[2] & ~0x80) << 6);
> c |= ((str.ptr[3] & ~0x80));
> break;
> case 5, 6, 1:
> goto Linvalid;
> }
> }
> else
> Linvalid : throw new Exception("yadayada");
>
> }
> return c;
> }
Looks very verbose to me. I had found in the BSD codebase a very
clever utf-8 conversion function in C, maybe it can be used here.
Sorry if I do not participate on the testing as I don't have a
proper compilation environment here at home. Here the routine I
use at work (it's in C), put that here for inspiration.
DEFINE_INLINE uint_t xctomb(char *r, wchar_t wc)
{
uint_t u8l = utf8len(wc);
switch(u8l) {
/* Note: code falls through cases! */
case 4: r[3] = 0x80 | (wc & 0x3f); wc >>= 6; wc |= 0x10000;
case 3: r[2] = 0x80 | (wc & 0x3f); wc >>= 6; wc |= 0x800;
case 2: r[1] = 0x80 | (wc & 0x3f); wc >>= 6; wc |= 0xc0;
case 1: r[0] = wc;
}
return u8l;
}
utf8len being
DEFINE_INLINE uint_t utf8len(wchar_t wc)
{
if(wc < 0x80)
return 1;
else if(wc < 0x800)
return 2;
else
if(wc < 0x10000)
return 3;
else
return 4;
}
The code generated on SPARC with gcc 3.4.6 was really good. On
x86_64 with gcc 5.1 was also not bad. I have not tried a lot of
alternatives as UTF-8 coding is not a bottle neck on our project.
There's also no check for length 5 and 6 as they are not possible
on our system, but for here it has to be added. (the
DEFINE_INLINE macro is either extern inline or inline depending
on some macro magic that is not of importance here).
More information about the Digitalmars-d
mailing list