Re: Case mapping of sharp s

[Eli Zaretskii]

> Date: Sat, 21 Nov 2009 04:08:42 +0100
> From: grischka <address@hidden>
> Cc: address@hidden, address@hidden
> 
> Stefan Monnier wrote:
> >> Also how can it happen that a C function varies between 4 and 90
> >> seconds for the same action.
> > 
> > I have no explanation for that.
> 
> Turned out that the time of backwards simple_search depends mostly
> on the number of buffer markers in the buffer.
> 
> That's because of CHAR_TO_BYTE in the inner loop and then because
> that one doesn't mind checking hundreds of markers for each single
> char in the file.

CHAR_TO_BYTE could be expensive, yes.  But how else can you convert an
arbitrary character position to the corresponding byte position?  When
you scan forward, you know the byte length of a multi-byte UTF-8
sequence by the first byte, but what do you do when you scan backwards?

The markers CHAR_TO_BYTE considers are a kind of cache, and are
supposed to speed things up.  I don't know what measurements were done
at the time this caching was introduced, nor whether those
measurements were repeated when Emacs switched from Mule encoding to
the current extended UTF-8 encoding of characters.  Maybe nowadays
this caching no longer helps.  Maybe UTF-8 allows a simpler conversion
than just counting bytes since the beginning of the buffer.  Or maybe
this particular use-case does not benefit from the cache, and we
should have a no-cache method for doing the same.

IOW, more research is needed.