[Bug tree-optimization/107087] [13 Regression] bits/stl_algobase.h:431: warning: 'void* __builtin_memcpy(void*, const void*, unsigned int)' reading between 8 and 2147483644 bytes from a region of size 4 [-Wstringop-overread]

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107087

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|NEW                         |ASSIGNED

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Huh.

<bb 14> [local count: 226234232]:
_634 = _76 + _546;
_618 = -12;
if (_618 > 4)
  goto <bb 15>; [90.00%]
else
  goto <bb 16>; [10.00%]

left around by t.ii.204t.forwprop4

   _634 = _76 + _546;
-  _618 = 8 - prephitmp_260;
+  _618 = -12;
   if (_618 > 4)

and this constant is produced by cleanup_cfg via replace_uses_by which
replaces prephitmp_260 with 20 and folds the stmt when merging blocks
30 and 36.

If we'd realize the condition would be false we wouldn't emit the bad
diagnostics.  Note it's the very last forwprop and there's no copy
propagation done afterwards.  IMHO even CFG cleanup propagating to
uses instead of replacing the PHI with a copy is premature, so asking
for it to propagate even further there is wrong I think.

forwprop could eventually be taught to track known not executable edges
in its non-iterating RPO walk and so eventually optimize the copy itself.
The testcase has no loops at least.

Let me try.