Re: DCE eliminating valid statement for ACATS c34007p

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

    You want to look at tree-ssa-operands.c:get_expr_operands() and see where
    things go wrong.  Also for D.861 not in SSA form, there might be a missing
    call to mark_vars_to_rename and/or update_ssa somewhere.  At which point
    in the pass flow does the above happen?  Is it ever "correct"?

A couple of hours after sending my email, I figured out what happened.
The reason why D.861 was not in SSA form was that it was TREE_ADDRESSABLE
and everything else was due to that.

This turned out to be the "well known" problem that the Ada front end is
making an ADDR_EXPR of odd things, in this case a COMPOUND_EXPR.

If I put code in the language-specific gimplifier to do:

      /* Otherwise, if we are taking the address of something that is neither
	 a refence, declaration, or constant, make a variable for the operand
	 here and then take its address.  If we don't do it this way, we may
	 confuse the gimplifier because it needs to know the variable is
	 addressable at this point.  This duplicates code in
	 internal_get_tmp_var, which is unfortunate.  */
      else if (TREE_CODE_CLASS (TREE_CODE (op)) != tcc_reference
	       && TREE_CODE_CLASS (TREE_CODE (op)) != tcc_declaration
	       && TREE_CODE_CLASS (TREE_CODE (op)) != tcc_constant)
	{
	  tree new_var = create_tmp_var (TREE_TYPE (op), "A");
	  tree mod = build (MODIFY_EXPR, TREE_TYPE (op), new_var, op);

	  TREE_ADDRESSABLE (new_var) = 1;
	  if (TREE_CODE (TREE_TYPE (op)) == COMPLEX_TYPE)
	    DECL_COMPLEX_GIMPLE_REG_P (new_var) = 1;

	  if (EXPR_HAS_LOCATION (op))
	    SET_EXPR_LOCUS (mod, EXPR_LOCUS (op));

	  gimplify_and_add (mod, pre_p);
	  TREE_OPERAND (expr, 0) = new_var;
	  recompute_tree_invarant_for_addr_expr (expr);
	  return GS_ALL_DONE;
	}

it all works.  Whether this should be here or in gimplify.c is a matter
mostly of philosphy.  I don't care for the duplication of code in
internal_get_tmp_var from a maintainability point of view and that argues for
adding an "addressable" parameter to the routine and calling it that way from
gimplify_addr_expr.  The counter argument is that this tree is only made by
Ada, so it should be done within the Ada front end.  On the other hand, we
define "valid GENERIC" to be those trees that were generated by the compiler
immediately before the tree-ssa conversion and trees of the above form were
generated then.  So they are valid GENERIC and ought to be handled by the
gimplifier.

It would also be good to put an assert into the gimplifier that detects the
case where one of its variable is suddenly marked addressable since that
meant it produced invalid gimple earlier, but since we have no flag to denote
such, I don't think we can do it now.

Are there are feelings about where this code should go?  My preference
is gimplify.c, but if there are strong feelings the other way, I'm OK
with leaving it where it is now.