[PATCH] D22900: Revert r244207 - Mark calls in thunk functions as tail-call optimization

[David Majnemer via cfe-commits]

On Sun, Jul 31, 2016 at 10:32 PM, Gerolf Hoflehner <ghoflehner at apple.com>
wrote:

>
> > On Jul 31, 2016, at 1:46 AM, Amjad Aboud <amjad.aboud at intel.com> wrote:
> >
> > aaboud added a comment.
> >
> >> ISTM that the DWARF spec intended such thunks to be encoded as
> `DW_AT_trampoline`.  That seems more appropriate than relying on codegen
> emitting a tailcall.  This way the debugger can make the policy decision of
> whether or not thunks should show up in the backtrace.
> >
> >>
> >
> >> In any case, correctness must always trump all else.  Reverting to
> green should take precedence over a QoI bug like PR24235.
> >
> >
> > I agree to the revert, though I am not sure about the new test, it looks
> too complected, especially the command line.
>
> An app crashed somewhere because it loaded a garbage value from the stack.
> To show that problem the test is.a little longer than it would be if we
> only wanted to check " no tail + byval “.
>

It is frowned upon to have tests which rely on LLVM's optimizers, it makes
the resulting test susceptible to changes in code unrelated to clang.  I'd
recommend we change the test to check for no tail + byval.


> > I will let David decide on accepting that test or ask for improvement.
> >
> > Regards,
> > Amjad
> >
> >
> > https://reviews.llvm.org/D22900
> >
> >
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20160801/5a17eea2/attachment.html>