[Bug tree-optimization/44485] [4.6 Regression] ICE in get_expr_operands, at tree-ssa-operands.c:1020

[rguenth at gcc dot gnu dot org]

------- Comment #7 from rguenth at gcc dot gnu dot org  2010-08-25 17:26 -------
(In reply to comment #6)
> The problem is that func_21 is first marked as
> TREE_READONLY/DECL_LOOPING_CONST_OR_PURE_P by local_pure_const, thus there are
> no vops on it, but afterwards it is by another local_pure_const made
> also TREE_THIS_VOLATILE (i.e. noreturn).
> flags_from_decl_or_type doesn't set ECF_CONST for const noreturn functions
> though for some reason:
> 603      /* Process the pure and const attributes.  */
> 604      if (TREE_READONLY (exp) && ! TREE_THIS_VOLATILE (exp))
> 605        flags |= ECF_CONST;
> 606      if (DECL_PURE_P (exp))
> 607        flags |= ECF_PURE;
> (strangely it does set ECF_PURE for pure noreturn calls), which means when
> fixup_noreturn_call calls update_stmt to drop lhs from the call, it adds vops
> that were missing before and .MEM needs renaming, but the pass in which
> fixup_noreturn_call is called doesn't expect that ssa needs updating.
> I wonder what is the reason for that " && ! TREE_THIS_VOLATILE (exp)" above.

I'd say remove it - it probably was a measure to avoid DCEing the side-effect
or so, before we had proper looping-pure-or-const.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44485