[Bug tree-optimization/82446] [8 Regression] Missed equalities in dr_group_sort_cmp

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82446

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 42311
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42311&action=edit
failed patch attempt

I tried to address this in split_constant_offset, better canonicalizing trees
there.  While it handles the case in the original PR it regresses some others,
for example gcc.dg/vect/pr69783.c because it canonicalizes before folding,
possibly messing up canonicalization that happened to be ok before.

So the solution is probably to add a canonicalize_tree function that does just
this (without folding, just swapping operands in-place -- requiring tree
unsharing before if it is possibly shared).

Ideally we'd have a expr hash on each tree_exp that swap_tree_operands ()
could use so we'd properly canonicalize already during folding (conveniently
tree_exp _does_ have 4 bytes of unused padding...).  buildN could compute
this for us as well.

Anyway, a separate function looks more feasible at the moment, we can call that
from dr_innermost analysis (there's already some kind of weird canonicalization
function).