[Bug tree-optimization/120043] [15/16 Regression] wrong code at -O{s,2,3} with -fallow-store-data-races on x86_64-linux-gnu since r15-1391-g4b75ed33fa5fd6

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120043

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Version|unknown                     |15.1.0
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot 
gnu.org
             Status|NEW                         |ASSIGNED

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
I'll have a look.  Yes, tree_could_trap_p does not know whether it sees a
load or a store.

OTOH another issue is that we turn *b = 1 into a = 1 with conditional
propagation:

  <bb 2> [local count: 1073741824]:
  b.1_1 = b;
  if (b.1_1 != &a)
    goto <bb 4>; [70.00%]
  else
    goto <bb 3>; [30.00%]

  <bb 3> [local count: 322122544]:
  a = 1;

  <bb 4> [local count: 1073741824]:
  return 0;

handling this in tree_could_trap_p conservatively would be quite prohibiting.
There is ifcvt_memrefs_wont_trap from if-conversion that handles this case
correctly (but it uses flow analysis to improve results).

I'll see to fix cselim and enhance tree_could_trap_p with a lhs_p flag.