On Mon, May 14, 2007 at 10:47:13PM +0100, Mark Shinwell wrote:
> I'm fairly certain that this is the correct approach to fix this
> issue, but I'm less certain that I have correctly guarded the call
> to forget_old_reloads_1,
[snip]
> Index: reload1.c
> ===================================================================
> --- reload1.c (revision 170932)
> +++ reload1.c (working copy)
> @@ -7506,6 +7506,9 @@ emit_reload_insns (struct insn_chain *ch
> }
> }
>
> + if (i < 0 && rld[r].in != NULL_RTX && rld[r].reg_rtx != NULL_RTX)
> + forget_old_reloads_1 (rld[r].reg_rtx, NULL_RTX, NULL);
> +
> /* The following if-statement was #if 0'd in 1.34 (or before...).
> It's reenabled in 1.35 because supposedly nothing else
> deals with this problem. */
It seems to me that the only special thing happening in your testcase is
that reload.in(_reg) is a PLUS rather than a REG or MEM. Does you patch not
prevent reload inheritance in many cases where it would be OK?
Hmm, the immediately preceeding if() block begins:
/* I is nonneg if this reload used a register.
If rld[r].reg_rtx is 0, this is an optional reload
that we opted to ignore. */
if (i >= 0 && rld[r].reg_rtx != 0)
{
I'd like a similiar comment just before your code. It is the i < 0 in
your case which prevents the preceeding if() block from updating
reg_reloaded_valid and so on.
--
Rask Ingemann Lambertsen
