[Bug tree-optimization/61757] [4.10 Regression] genmodes failure with enable-checking

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61757

--- Comment #17 from Michael Matz <matz at gcc dot gnu.org> ---
The testsuite failure on m68k is indeed caused by r212352.  In particular
this hunk:

@@ -1641,8 +1609,7 @@ record_equality (tree x, tree y)
      long as we canonicalize on one value.  */
   if (is_gimple_min_invariant (y))
     ;
-  else if (is_gimple_min_invariant (x)
-          || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
+  else if (is_gimple_min_invariant (x))
     prev_x = x, x = y, y = prev_x, prev_x = prev_y;
   else if (prev_x && is_gimple_min_invariant (prev_x))
     x = y, y = prev_x, prev_x = prev_y;

The test is <=, so without loops (both SSA names being at 'depth' 0) the check
succeeds always and hence will canonicalize on x (i.e. record x as value of y).
The testcase is written so that this is the expected order for the third
if to be found trivial.

The problem doesn't happen on x86-64 because with -O1 the && condition is
rewritten into arithmetic on '&', so even though the wanted transformation
doesn't happen there are only two 'if ' statements, which is what is searched
for.

Unfortunately this is all harmless and not causing the miscompilations, it's
only a missed optimization.

FWIW, I think the canonicalization that DOM wants here should be retained
in some form (canonicalize to the value that covers a larger area).  Possibly
by checking if one definition dominates the other