[Bug tree-optimization/108821] [11/12/13 Regression] LIM reissuing a violatile store when it cannot/should not

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108821

Franz Sirl <sirl at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenther at suse dot de

--- Comment #2 from Franz Sirl <sirl at gcc dot gnu.org> ---
Started with commit b6ff3ddecfa93d53867afaaa078f85fc848abbbd:

tree-optimization/94988 - enhance SM some more

This enhances store-order preserving store motion to handle the case
of non-invariant dependent stores in the sequence of unconditionally
executed stores on exit by re-issueing them as part of the sequence
of stores on the exit.  This fixes the observed regression of
gcc.target/i386/pr64110.c which relies on store-motion of 'b'
for a loop like

  for (int i = 0; i < j; ++i)
    *b++ = x;

where for correctness we now no longer apply store-motion.  With
the patch we emit the correct

  tem = b;
  for (int i = 0; i < j; ++i)
    {
      tem = tem + 1;
      *tem = x;
    }
  b = tem;
  *tem = x;

preserving the original order of stores.  A testcase reflecting
the miscompilation done by earlier GCC is added as well.

This also fixes the reported ICE in PR95025 and adds checking code
to catch it earlier - the issue was not-supported refs propagation
leaving stray refs in the sequence.

2020-05-11  Richard Biener  <rguenther@suse.de>

PR tree-optimization/94988
PR tree-optimization/95025
* tree-ssa-loop-im.c (seq_entry): Make a struct, add from.
(sm_seq_push_down): Take extra parameter denoting where we
moved the ref to.
(execute_sm_exit): Re-issue sm_other stores in the correct
order.
(sm_seq_valid_bb): When always executed, allow sm_other to
prevail inbetween sm_ord and record their stored value.
(hoist_memory_references): Adjust refs_not_supported propagation
and prune sm_other from the end of the ordered sequences.

* gcc.dg/torture/pr94988.c: New testcase.
* gcc.dg/torture/pr95025.c: Likewise.
* gcc.dg/torture/pr95045.c: Likewise.
* g++.dg/asan/pr95025.C: New testcase.