[Bug ipa/111157] [14 Regression] 416.gamess fails with a run-time abort when compiled with -O2 -flto after r14-3226-gd073e2d75d9ed4

["jamborm at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111157

--- Comment #6 from Martin Jambor <jamborm at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #5)
> I think if IPA modref declares the argument dead at the call site then IPA
> CP/SRA cannot declare it known constant.

It is declared "killed" by the function.  I still need to figure out
whether that is all I need or whether the fact that it is not read
either is the combination I am after.  But I agree that IPA-CP should
refrain from propagating clearly unneeded info in that case.

> 
> Now, I wonder why IPA CP/SRA does not replace the known constant parameter
> with an automatic var like
> 
> point.constprop.isra (double ISRA.1740, int & restrict ipoint, double &
> restrict x, double & restrict y, double & restrict z, int & restrict istat)
> {
> ...
>   const int istat.local = 0;
>   istat = &istat.local;
> 
> ?  So if not all uses of 'istat' get resolved we avoid generating wrong
> code.  The expense is a constant pool entry (if not all uses are removed),
> but I think that's OK.  It would also work for aggregates.  It would also
> relieve IPA-CP modification phase from doing anything but trival value
> replacement (in case the arg isn't apointer).

I'm afraid I don't understand.  Even in this particular case, istat is
checked by the caller and the callee can assign to it also other
values, not just the one which happens to be what it it initialized to
by the caller - and in the original code it does when there is an
error - those writes cannot be redirected to a local variable.