[Bug tree-optimization/56778] [4.9 Regression] ICE on several benchmarks after r196775.

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56778

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-02 11:19:35 UTC ---
The issue is that gimple_seq_add_seq updates stmts added.  But we update SSA
form _before_ adding the sequence into the IL stream:

  /* End loop-exit-fixes after versioning.  */

  update_ssa (TODO_update_ssa);
  if (cond_expr_stmt_list)
    {
      cond_exp_gsi = gsi_last_bb (condition_bb);
      gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list,
                             GSI_SAME_STMT);
    }

gimplify.c has a function that does gimple_seq_add_seq_without_update.

There are two ways to fix this, one is to insert the sequence before
updating SSA form, another is to consistently use
gimple_seq_add_seq_without_update.
Or to make gimple_seq_add_seq not update stmts (sequences are inserted
later anyway, which can do the update).  Not sure why it would make sense
at all to update stmt operands for sth not in the IL ...

Well.  I'm going for the first.

Which doesn't fix it because we create an alias-check for a load of a
load appearantly where the latter is not loop invariant.  I presume
this is GATHER support, and eventually the DDR mangling in data-ref
analysis I removed did avoid this to happen.