On Fri, Aug 05, 2016 at 04:25:26PM -0400, Michael Meissner wrote:
> Ummm, this patch looks wrong. Because the insn uses the SFDF iterator, the
> mode of operands[1] could be either SFmode or DFmode. I think it should be:
>
> - rtx tmp, stack;
> + rtx src = force_reg (mode, operands[1])
On Wed, Aug 03, 2016 at 12:02:54AM +0930, Alan Modra wrote:
> This is a hack I tried to avoid having to poke at lra code for
> pr71680..
>
> The idea of adding force_reg here was that it removes subregs from
> fix_trunc, emitting the subreg mode conversion in a separate set insn.
> That avoids the
On Wed, Aug 03, 2016 at 12:02:54AM +0930, Alan Modra wrote:
> So this patch isn't a particularly good solution to pr71680, but
> a) force_reg for an operand that must be a reg is 100% safe, and
> b) it's good to expose more combine opportunities.
I'm not sure it actually does fix the PR. But the
This is a hack I tried to avoid having to poke at lra code for
pr71680..
The idea of adding force_reg here was that it removes subregs from
fix_trunc, emitting the subreg mode conversion in a separate set insn.
That avoids the underlying lra issue, by virtue of combine merging the
SF subreg with t
