This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH, rs6000] power8 patches, patch #4 (revised), new power8 builtins
- From: Michael Meissner <meissner at linux dot vnet dot ibm dot com>
- To: David Edelsohn <dje dot gcc at gmail dot com>
- Cc: Michael Meissner <meissner at linux dot vnet dot ibm dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>, Pat Haugen <pthaugen at us dot ibm dot com>, Peter Bergner <bergner at vnet dot ibm dot com>
- Date: Wed, 5 Jun 2013 12:13:32 -0400
- Subject: Re: [PATCH, rs6000] power8 patches, patch #4 (revised), new power8 builtins
- References: <20130520204053 dot GA21090 at ibm-tiger dot the-meissners dot org> <20130521234717 dot GA27879 at ibm-tiger dot the-meissners dot org> <20130604184853 dot GA12768 at ibm-tiger dot the-meissners dot org> <CAGWvnyni0U+749Y-_aS1mB9vMQ1pKHhiikGhxJmV26KVP2Rbtw at mail dot gmail dot com>
On Wed, Jun 05, 2013 at 10:28:02AM -0400, David Edelsohn wrote:
> +;; The canonical form is to have the negated elment first, so we need to
> +;; reverse arguments.
>
> Please fix the typo in the comment: "element".
Ok. I need to proof-read the patches before sending them out.
> +;; Like VSX_L, but don't support TImode for doing logical instructions in
> +;; 32-bit
> +(define_mode_iterator VSX_L2 [V16QI V8HI V4SI V2DI V4SF V2DF])
> +
> ;; Iterator for memory move. Handle TImode specially to allow
> ;; it to use gprs as well as vsx registers.
> (define_mode_iterator VSX_M [V16QI V8HI V4SI V2DI V4SF V2DF])
>
> +(define_mode_iterator VSX_M2 [V16QI
> + V8HI
> + V4SI
> + V2DI
> + V4SF
> + V2DF
> + (TI "TARGET_VSX_TIMODE")])
>
> The patch adds new iterators VSX_L2 and VSX_M2. The original
> ChangeLog only mentioned M2 and the new ChangeLog only mentions L2.
> What's going on?
I thought I had deleted VSX_M2 from this patch. It will be needed in patch #8
for the fusion peephole. The difference is VSX_L2 avoids TImode altogether,
and was used by the logical ops to prevent TImode operations in VSX registers
in 32-bit.
The problem is unless we have expanders/splitters for logical DImode, the
compiler when it wants to do a logical DImode operation says, aha I have a
TImode operation, and then it converts the DImode value to TImode, does the
operation (which in turn may mean transfer between GPR and VSX registers).
I can add splitters and such for 32-bit DImode to prevent this, but I don't
know if you want me to do it in the context of this patch, or do it as a later
patch.
> * config/rs6000/rs6000.md (eqv<mode>3): Add support for powerp eqv instruction.
>
> Why isn't this covered by boolean_operator and %q output operand? And
> why can't that predicate and output operand handle vsx as well, e.g.,
> *vsx_eqv? Why don't we simply have vsx_bool<mode>3, etc.
See my reply to Segher that boolc doesn't use the canonical form, so eqv is not
currently generated.
--
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460, USA
email: meissner@linux.vnet.ibm.com, phone: +1 (978) 899-4797