https://issues.dlang.org/show_bug.cgi?id=17650

Jon Degenhardt <jrdemail2000-dlang@yahoo.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|enhancement                 |regression

--

https://issues.dlang.org/show_bug.cgi?id=17650

Martin Nowak <code@dawg.eu> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |code@dawg.eu

--- Comment #1 from Martin Nowak <code@dawg.eu> ---
We'll address this with 2.075.1 to not delay the release even further.

--

https://issues.dlang.org/show_bug.cgi?id=17650

--- Comment #2 from Vladimir Panteleev <dlang-bugzilla@thecybershadow.net> ---
Introduced in https://github.com/dlang/phobos/pull/5351

--

https://issues.dlang.org/show_bug.cgi?id=17650

--- Comment #3 from Vladimir Panteleev <dlang-bugzilla@thecybershadow.net> ---
Interesting - looking at the PR, this doesn't really seem like a regression, rather that the addition of the @safe attribute exposed an out-of-bounds array access that was always there. Feel free to reclassify, Jon.

--

https://issues.dlang.org/show_bug.cgi?id=17650

--- Comment #4 from Vladimir Panteleev <dlang-bugzilla@thecybershadow.net> ---
Actually, I believe we do count a breakage as a regression if something breaks for the end-user, regardless of what is going on under the hood. E.g. it's possible that for supported architectures, the range violation was completely benign.

--

https://issues.dlang.org/show_bug.cgi?id=17650

--- Comment #5 from Jon Degenhardt <jrdemail2000-dlang@yahoo.com> ---
Wow. So I introduced this, initiated by trying add new unit tests. And my own unit tests caught it, but unfortunately, not the unit tests I added to Phobos. How ironic. Thanks for investigating Vladimir.

--

https://issues.dlang.org/show_bug.cgi?id=17650

ZombineDev <petar.p.kirov@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |petar.p.kirov@gmail.com

--- Comment #6 from ZombineDev <petar.p.kirov@gmail.com> ---
The error doesn't make sense to me. @safe adds bounds checks where there are none only when building with -release. I.e. if we're not building with -release, there shouldn't be any change in behavior - bounds checking should already be present (-boundscheck defaults to 'on' in that case). Am I missing something?

--

https://issues.dlang.org/show_bug.cgi?id=17650

--- Comment #7 from Jon Degenhardt <jrdemail2000-dlang@yahoo.com> ---
(In reply to ZombineDev from comment #6)
> The error doesn't make sense to me. @safe adds bounds checks where there are none only when building with -release. I.e. if we're not building with -release, there shouldn't be any change in behavior - bounds checking should already be present (-boundscheck defaults to 'on' in that case). Am I missing something?

There was already an existing range violation, one that didn't have a unit test in the std.getopt module. Adding @safe to the function turned on bounds checking in debug mode. I had a unit test in my own code (not std.geopt) that triggered the out-of-bounds case. This unit test didn't get run until I actually ran my own unit tests with the new code.

The out-of-bounds case occurs when there is an command line argument that is a single dash. e.g.

   $ myprogram -x a  # okay
   $ myprogram -x -  # out-of-bounds

In the single dash case there's an attempt to access one character past it. Kind of understandable there wasn't a unit test for this originally. If you want to see the sequence that led to adding @safe read https://github.com/dlang/phobos/pull/5347.

--

https://issues.dlang.org/show_bug.cgi?id=17650

--- Comment #8 from Jon Degenhardt <jrdemail2000-dlang@yahoo.com> ---
I'm preparing a PR to fix the underlying issue. A concern for the general release is that a single hyphen is often used to represent standard input in command line args. e.g.

$ cat file.txt | my_d_program --file - > out.txt

This will trigger the bounds check error if my_d_program is compiled in safe mode.

--

https://issues.dlang.org/show_bug.cgi?id=17650

--- Comment #9 from Jon Degenhardt <jrdemail2000-dlang@yahoo.com> ---
https://github.com/dlang/phobos/pull/5612

--