I used to use LC_ALL=C, but, as I vaguely recall, it got in the way of dealing with UNICODE. I tried a couple LC values aimed at UNICODE and the US, but something always went pear-shaped. I finally give up. I am perfectly happy to suffer a tiny bit of performance, to have most things work without thinking. A factor of 6, or 35, is not tiny, since I use grep and friends intensely. That's how I discovered the performance problem to begin with. Anyway, thank you for fixing my problem. I suspect that many of us pioneers (using UNIX since 1973) have '[0-9]' wired into our fingers.

On Wed, Mar 22, 2017 at 2:01 PM, Paul Eggert <eggert@cs.ucla.edu> wrote:

On 03/22/2017 05:44 AM, John P. Linderman wrote:

That puts the runtimes on equal footing:

In my measurements, P[0-9] is still a tiny bit slower if one is using glibc regex, due to a performance problem in glibc. You can work around it by configuring --with-included-regex. It's probably not worth worrying about, though.

By the way, using LC_ALL=C should help avoid performance problems like these in the future, if all you're doing is something where single-byte pattern matching suffices.