Re: A patch for PPro

[Jim Wilson <wilson@cygnus.com>]

I believe this patch is the wrong approach, because back-end files should
not be trying to use private data which is used in varasm.c internals.
That destroys modularity, and makes it harder to maintain gcc.  If we
really want to take this route, then we need to make this data public rather
than private, instead of adding hooks to allow other code to use varasm.c's
private data.  This situation is a little confusing because this private
data is not stored in static variables inside varasm.c, but rather in
the global RTL, but it is still private data.

In any case, I don't believe that this really fixes the problem.
There is a general problem that we can't support input reloads with an
insn using cc0, and there is no safe way to prevent generation of input
reloads.

I think the only safe fix is to modify the conditional move patterns to
emit both the compare and the conditional move, and split them after reload.

Jim