[Bug tree-optimization/69466] [6 Regression] ICE: Invalid PHI argument after vectorization (on -O2)

[aoliva at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69466

--- Comment #5 from Alexandre Oliva <aoliva at gcc dot gnu.org> ---
Created attachment 37486
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37486&action=edit
Patch I'm testing to fix the problem

The problem occurs because we call set_current_def for phi nodes after
duplicating some blocks, but it isn't always the case that the blocks passed to
slpeel_duplicate_current_defs_from_edges were copied from one another.  In the
given testcase, the exit blocks passed to this function don't seem to be
related at all, so taking information from corresponding phi nodes to call
set_current_def is most certainly incorrect.  Indeed, it's calling
set_current_def for a virtual operand from a "corresponding" non-virtual
operand that causes slpeel_duplicate_current_defs_from_edges's use of
get_current_def to set up an incorrect, non-virtual value for the
virtual-operand phi node.  With this patch, we refrain from setting current
defs when blocks don't have analogous phi nodes, and this is enough to address
the problem.

It would be nice, however, if someone more familiar with the loop vectorizer
would check whether our calling this function with mismatched blocks indicates
another latent problem, or whether we could check for the mismatch and bypass
the call more efficiently at the caller.  For this testcase, the mismatching
call is the single_exit (scalar_loop) one in
slpeel_tree_duplicate_loop_to_edge_cfg.