Re: Forbid __builtin_return_address when inlining

[Andrew Haley <aph at redhat dot com>]

Mark Mitchell writes:
 > Andrew Haley wrote:
 > > Mark Mitchell writes:
 > >  > Andrew Haley wrote:
 > >  > > __builtin_return_address(0) returns the wrong address if its caller
 > >  > > has been inlined.  Weirdly, we never seem to have detected this
 > >  > > before, but it surely must be wrong.
 > >  > 
 > >  > The manual says:
 > >  > 
 > >  > >  When
 > >  > >      inlining the expected behavior is that the function will return
 > >  > >      the address of the function that will be returned to.  To work
 > >  > >      around this behavior use the `noinline' function attribute.
 > >  > 
 > >  > which seems to suggest that the behavior you are seeing as is
 > >  > documented, peculiar though it seems.
 > > 
 > > Hmm, OK.  But if I fixed the peculiar behavour then I could also
 > > remove the peculiar documentation, couldn't I?  That would surely be
 > > an improvement.,,
 > 
 > The problem is that we don't know how people may already be using
 > this "feature".  Since someone bothered to document it, I wouldn't
 > be surprised if something (the Linux kernel?  a scripting language
 > interpreter?) is using it in some way that depends on the current
 > behavior.

No kidding.  Yes, that's surely possible.

 > So, I fully agree that the semantics you're proposing seem more
 > sensible (i.e., tell me where *this function* returns, and, if
 > necessary, make sure that it has somewhere to return to by not
 > inlining it), but I'd be nervous about making the change.  Is there
 > a compelling reason to change this, other than just that it makes
 > sense?

No, not really.  Java methods sometimes need to find their caller, and
I was testing that when I discovered this anomalous behaviour when a
method was being inlined.  The most obvious solution is for the Java
FE to mark the methods that need to find their caller noinline, and
that's what I'll do.

Andrew.