[Bug tree-optimization/57359] store motion causes wrong code for union access at -O3

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57359

--- Comment #23 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #22)
> Created attachment 48311 [details]
> patch
> 
> Note that apart from the possible bad impact on optimization when fixing this
> bug an actual fix is complicated by the custom "optimized" dependence
> analysis
> code in the loop invariant motion pass.
> 
> A conservative "simple" patch would be the attached but that doesn't preserve
> store-motion for the following (because the LIM data dependence code doesn't
> care about stmt order):
> 
> typedef int A;
> typedef float B;
> 
> void __attribute__((noinline,noclone))
> foo(A *p, B *q, long unk)
> {
>   for (long i = 0; i < unk; ++i) {
>       q[i] = 42;
>       *p = 1;
>   }
> }
> 
> usually this bug doesn't manifest itself but of course the fix will be
> experienced everywhere.  Benchmarking the simple patch might reveal
> it's not an issue (but I doubt that...).

Which means a better approach might be to, in addition to the existing
dependence testing, verify we can sink the stores through all exits
(IIRC there is/was a similar bug involving ordering of stores sunk where we'd
have to preserve the order).  There's again the difficult cases like

 for ()
  {
    *p = 1;
    q[i] = 42;
    if (test)
      *p = 2; 
  }

which we currently sink as

 for ()
   {
     p_1 = 1;
     q[i] = 42;
     if (test)
       p_1 = 2;
   }
 *p = p_1;

if we want to preserve all sinking we'd have to replay all [possibly
aliased] stores - again more difficult when the store we sink is
in the latch (or when multiple exits are involved).  For the above
example do

  for ()
   {
     p_1 = 1;
     q[i] = 42;
     if (test)
       p_1 = 2;
   }
  *p = p_1;
  q[i] = 42;
  if (test)
    *p = p_1;

with scanning for valid re-orderings starting from all exits we want
to consider sinking two stores that cannot be reordered.

Or we want to re-think the whole store-motion data dependence code.