[r11-3641 Regression] FAIL: gcc.dg/torture/pta-ptrarith-1.c -Os scan-tree-dump alias "ESCAPED = {[^\n}]* i f [^\n}]*}" on Linux/x86_64 (-m32 -march=cascadelake)

[Jan Hubicka]

> > >
> > > A number of people routinely send emails similar to these to this
> > > list to point out regressions on their targets.  I find both kinds
> > > of emails very useful and don't mind the additional traffic.
> > >
> > > What would be an improvement is sending just one email for all
> > > the testsuite regressions rather than one for each test or run
> > > as seems to be happening.
> >
> > Sunil, can you update your script to send out a single email
> > for all regressions caused by one commit?  Thanks.
> >
> Sure, I'll work on it.

In any case this mail was useful (thanks for running the script).  The
bug in mod-ref is by using poly_int64 to remember offset from MEM_REF:

     tree offset = TREE_CODE (base) == MEM_REF
                  ? TREE_OPERAND (base, 1) : NULL_TREE;

a.parm_offset_known = offset && wi::to_poly_offset (offset).to_shwi
				    (&a.parm_offset).to_shwi (&a.parm_offset)

For 64bit targets this is OK, but for 32bit targets we mistakely
consider negative offsets to be large positive integers here:

                  ao_ref_init_from_ptr_and_range
                         (&ref2, arg, true,
                          access_node->offset
                          + (access_node->parm_offset
                             << LOG2_BITS_PER_UNIT), access_node->size,
                          access_node->max_size);

I think I am intended to use mem_ref_offset and poly_offset_int type.
This will need new streaming support that is not hard to add, but I am
not sure how to do the computation above (that probably needs overflow
check) and how to print the offset to dump file.

Richard, can you help me? :))
I had code in ao_ref_init_from_ptr_and_range that took parm_code
separately and re-inserted it to the MEM_REF built. However this does
not happen in all cases: when base is a decl we omit constructing
MEM_REF and then offset adjustment is necessary.
Without going through this folding we are unable to disambiguate base
pointers: MEM_REF containing ADDR_EXPR of DECL does not go through the
usual decl compares.

Honza
> 
> >
> > > I'm not sure that automatically opening bugs instead would be
> > > better, certainly not one per test, and not if the code author
> > > wasn't also CC'd if not automatically assigned to it.
> > >
> > > Martin
> >
> >
> >
> > --
> > H.J.
> >