[Bug rtl-optimization/57425] New: RTL alias analysis unprepared to handle stack slot sharing

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57425

            Bug ID: 57425
           Summary: RTL alias analysis unprepared to handle stack slot
                    sharing
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: major
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: amylaar at gcc dot gnu.org

Created attachment 30198
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30198&action=edit
testcase variant using a int / long union

I see a failures for gcc.dg/torture/pr25654.c for avr for -Os, -O2 and higher
optimization levels.  The .172r.cse1 debugging dump still looks OK, but the
.173r.fwprop1 dump shows that the second memory read to compute sum
has been eliminated by reusing the value of the first one, despite the store
that happened in the mean time.

AFAICT, this patch:
http://gcc.gnu.org/viewcvs/gcc?view=revision&revision=188667
has removed the infrastructure that was put in place to fix PR 25654, without
putting any in any replacement.

Now, what makes the avr different from the majority of tier1 targets is that
short and int are the same size.

If I change the testcase a bit to use int/long instead of short/int, it starts
failing on i686-pc-linux-gnu.  This is a 4.8 regression relative to 4.7 .
A variant of the testcase without unions also fails.  This is again at -Os /
-O2
and higher optimization levels.