[Toybox] pathological case in sed s///g

[Rich Felker]

On Mon, May 06, 2019 at 12:42:44PM -0500, Rob Landley wrote:
> Huh... I'll assume REG_STARTEND works in bionic since you're pointing me at it.
> It's not in the regex man page but it _is_ in the glibc headers...
> 
>   https://github.com/bminor/glibc/commit/6fefb4e0b16
> 
> Looks like it went into glibc in 2004, which is way past 7 years. I should poke
> Michael Kerrisk to update the man page.
> 
> And musl explicitly refused to do it, but Rich makes bad calls all the time:
> 
>   https://www.openwall.com/lists/musl/2013/01/15/26
> 
> I still haven't got an #ifdef __MUSL__ in portability because Rich insists his
> libc is the only perfect piece of software ever written, but I can probe for it

You can #ifdef REG_STARTEND and use it conditionally depending on
whether the functionality is offered. There is no reason to hard-code
an assumption that musl doesn't or does have this functionality; it's
been proposed and there's even a patch somewhere. (It's not costly to
support relative to the current bad regex implementation, but the
concern is that it might impose a nontrivial cost on a future good
implementation once we're locked into having it.)

> If this _does_ match NUL bytes it lets me remove regexec0() entirely from libc,
> which would be very nice. And since it started life as a BSD extension (they're
> the only man page _documenting_ it) I get support on FreeBSD and probably MacOS too.

I'm not sure if the proposed patch supports matching NUL or not (i.e.
whether it treats the start/end as *additional* constraints, to work
with a substring of a necessarily null-terminated string, or whether
they replace the null-termination criterion. If we do adopt it we
should ensure we do whatever other implementations do here, especially
if that's important to use cases you or others want.

Rich