Re: [PATCH v2 0/5] Speed up uNN_chr and uNN_strchr with Boyer-Moore algorithm

[Pádraig Brady]

On 29/07/10 10:26, Paolo Bonzini wrote:
> On 07/29/2010 11:05 AM, Pádraig Brady wrote:
>>> I would better recommend to use
>>> the u8_strstr function.
>>
>> I wonder could we speed that up for UTF-8
>> by just deferring to strstr() ?
>> I've not tested this so feel free to bin it.
> 
> An alternative idea: check if there is more than 1 character using
> u8_mbtouc.  If there is one, use the faster u8_strchr algorithm (you can
> just use u8_strchr, even though that does a useless conversion back to
> UTF-8).  If there is more than one, use strstr directly.

I thought of that but thought it would be too much overhead for the
general case since it seemed like it would need strlen, u8_mbsnlen & u8_mbtouc,
with conversions back and forth.

> And this gives another opportunity: we can now define u8_strstr and
> u16_strstr to always return a pointer to a valid UTF-8 sequence:
> 
> /* Find the first occurrence of NEEDLE in HAYSTACK.  For UTF-8 and
>    UTF-16, the returned value will never point in the middle of a
>    multibyte sequence or surrogate pair.  If this cannot be satisfied,
>    for example due to an invalid NEEDLE, NULL is returned.  */
> 
> Why?  Because the only cases when you can get a false positive are: 1)
> when you have illegal sequences at the beginning of the needle, which
> now we could easily check; 2) when both the needle and haystack have the
> same illegal UTF-8 in it, which I think could be ignored.
> 
> For u16_strstr you need to look for an invalid surrogate character at
> the beginning of the needle.
> 
> Thoughts?  Would anybody give a shot to the above?

But with those benefits it might be worth it?

cheers,
Pádraig.