[Bug c/64459] __attribute__((returns_nonnull)) creates wrong code on amd64

[Bug c/64459] __attribute__((returns_nonnull)) creates wrong code on amd64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64459

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2014-12-31
     Ever confirmed|0                           |1

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
> Since we use 992 returns_nonnull annotated functions in parrot, the bisect
> which function and asm exactly is the problem will need some time...

Without this, it is hard to tell if this is a bug in GCC or parrot.  So waiting
on a testcase that shows the problem.

[Bug c/64459] __attribute__((returns_nonnull)) creates wrong code on amd64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64459

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Maybe parrot has some code it where returns_nonnull is on the function but it
actually can return null and clang/MSVC does not optimize as good as GCC does.

[Bug c/64459] __attribute__((returns_nonnull)) creates wrong code on amd64

[rurban@x-ray.at]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64459

Reini Urban <rurban@x-ray.at> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
         Resolution|---                         |INVALID

--- Comment #3 from Reini Urban <rurban@x-ray.at> ---
This ticket can be closed, not a bug.

My test script finished after some hours of grinding, 
which tested all of the 992 returns_nonnull annotated functions in parrot.
Disabling returns_nonnull on one function made it work. 
I've checked the assembly.

It was Parrot_cx_check_scheduler(), which had a wrong 
if (Parrot_cx_check_scheduler(...) == 0)
check, and consequently this branch was removed by gcc.

So it's really a missed clang and msvc dead-code removal opportunity.

[Bug c/64459] __attribute__((returns_nonnull)) creates wrong code on amd64

[rurban@x-ray.at]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64459

--- Comment #4 from Reini Urban <rurban@x-ray.at> ---
maybe the script is of value to anyone:
https://gist.github.com/rurban/b8b951d03ff22734f05b

[Bug c/64459] __attribute__((returns_nonnull)) creates wrong code on amd64

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64459

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC5 supports in -fsanitize=undefined runtime warning or error when
returns_nonnull function does return NULL, so in theory it should be able to
tell you without such script.