Re: Regimplification enhancements 1/3

[Richard Biener <richard dot guenther at gmail dot com>]

On Mon, Jun 16, 2014 at 12:56 PM, Bernd Schmidt <bernds@codesourcery.com> wrote:
> For the ptx port, I've needed to write a new pass which ensures all objects
> go into address spaces as required by the machine. This uses the
> regimplification code in gimplify-me.c, and that requires some fixes and
> upgrades.

Can you explain how and why you need re-gimplification here?  IMHO
you should apply the necessary steps yourself, not put in GENERIC
into GIMPLE and rely on the gimplifier.

> Here's the first. When address spaces change, an ADDR_EXPR may have to be
> changed to ADDR_SPACE_CONVERT_EXPR, and these two have different
> representations in gimple. This patch stops the regimplification code from
> creating invalid gimple in such a case.
>
> Bootstrapped and tested on x86_64-linux, ok?

+      if (is_gimple_assign (stmt)
+         && num_ops == 2
+         && (get_gimple_rhs_class (gimple_expr_code (stmt))
+             == GIMPLE_SINGLE_RHS)
+         && (get_gimple_rhs_class (TREE_CODE (gimple_op (stmt, 1)))
+             != GIMPLE_SINGLE_RHS))

err ... that's a very crappy gimple condition that isn't going to reliably work.

You can't turn a x = &foo; into x = (convert)&foo; this way abusing
the re-gimplification routines.

The re-gimplification routines do _not_ accept random GENERIC
operands.  They were written to handle registers becoming non-registers
and nothing more.

Richard.

>
> Bernd