Re: [PATCH], PR target/64205, Fix -mcpu=G5 breakage on PowerPC

[David Edelsohn <dje dot gcc at gmail dot com>]

On Fri, Feb 6, 2015 at 9:41 AM, Michael Meissner
<meissner@linux.vnet.ibm.com> wrote:
> This patch fixes PR target/64205, which was caused due to my previous patch to
> add the -mupper-regs support for scalar floating point to occupy the VSX
> floating point registers that are overlaid on top of the traditional Altivec
> registers, starting with ISA 2.06 (power7).
>
> The problem was, I added secondary reload handlers for SDmode unconditionally.
> On systems before ISA 2.06, there are no stores and loads for 32-bit SDmode in
> a floating point register, and so you can't do normal secondary reload
> operations on this (you have to do a 64-bit store and a 32-bit load on the
> bottom bits).
>
> I have tested this with a full bootstrap on PowerPC 64-bit systems with no
> regressions.  In addition, I built a bootstrap compiler using the --with-cpu=G5
> option and it succeeded.  Is it ok to install the patches in trunk, and in any
> branches that the upper regs patches were applied to?
>
> [gcc]
> 2015-02-06  Michael Meissner  <meissner@linux.vnet.ibm.com>
>
>         PR target/64205
>         * config/rs6000/rs6000.c (rs6000_init_hard_regno_mode_ok): Do not
>         add a general secondary reload handler for SDmode, unless we have
>         both read/write support for SDmode.
>
> [gcc/testsuite]
> 2015-02-06  Michael Meissner  <meissner@linux.vnet.ibm.com>
>
>         PR target/64205
>         * gcc.target/powerpc/pr64205.c: New file.

Please clarify the comments in rs6000.c.  Too many double negatives
and ambiguous terms.

For example, "Use SDmode reload patterns if float load/store integer
word instructions are available."

Okay with an improved comment.

Thanks, David