Norihiro Tanaka wrote:
> Now I wrote the additional patch.

Sorry, I don't understand that patch.  I applied your earlier patch 
(attachment 1) and then merged that patch (attachment 2), but the 
resulting 'grep' failed 'make check'; I got a "FAIL: fedora" and the 
test after "file" never finished.  This was on Fedora 20 x86-64 with my 
own installation of GCC 4.9.0.

Instead of trying to understand that patch, I wrote a different patch 
(attachment 3) that simplifies the code and removes the part that is 
x86-specific.  This improves the performance quite a bit for the test 
case given in the ChangeLog entry.  I also tested performance on Sparc 
Solaris 10, and memchr was a big win there.  Perhaps there are platforms 
where memchr is a big loss, but we can cross that bridge if we come to it.

Perhaps I merged that patch incorrectly, so I'll leave this bug open for 
now.  To be honest I hope that patch doesn't improve performance 
significantly if we get it to work correctly, as it looks tricky.