bug#33793: sed bug with regular expressions

[Eric Blake]

tag 33793 notabug
thanks

On 12/18/18 6:50 AM, Uladzimir Panasiuk wrote:
> Hi. I've found the bug using sed. There is how to reproduce:
> 1) Run bash
> 2) Exec command \
> echo weather -5.0 | sed
> 's/[^0-9\-\.]//g'

You used two range expressions in this regex, but the result is the same 
as if you had used this regex with only one range expression::

's/[^0-9\.]//g'

Either way, you requested all characters except for the 10 digits, a 
literal backslash, or a literal dot.  Remember, a range expression [\-\] 
selects a single character of the backslash.  Since '-' is not excluded 
from the [] expression, sed correctly strips it.

> 3) You will get "5.0". Expected output is "-5.0"

You might be remembering the behavior of perl regex, where \ inside [] 
is an escape character.  But that's not how POSIX regex behaves - inside 
[], \ is literal, and there are no escape characters.

> BUT
> If you exec
> echo weather -5.0 | sed 's/[^0-9\.\-]//g'

Here, your regex only has one range expression, but lists \ twice.  The 
repetition is harmless, but means that your expression is the same as 
this shorter:

's/[^0-9\.-]//g'

It is not obvious from your input whether you intended to be filtering 
out literal backslash or not, but if not, you probably meant to write:

's/[^0-9.-]//g'

with no backslash, and with the - last (as that is one of the few places 
that you can write - to be matched as itself rather than treated as a 
range operator between neighboring characters).

I'm closing this as not a bug, but feel free to reply with further 
questions or comments.

--
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org