Re: [GCC 12 backport] Disable generating load/store vector pairs for block copies.

[Michael Meissner <meissner@linux.ibm.com>]

On Thu, Jul 14, 2022 at 04:12:14PM -0500, Segher Boessenkool wrote:
> On Thu, Jul 14, 2022 at 11:20:56AM -0400, Michael Meissner wrote:
> > I have applied the patch to GCC 12.
> > 
> > | From 22736f3d0d4fb8ce4afb3230023f8accdb03a623 Mon Sep 17 00:00:00 2001
> > | From: Michael Meissner <meissner@linux.ibm.com>
> > | Date: Thu, 14 Jul 2022 11:16:08 -0400
> > | Subject: [PATCH] [BACKPORT] Disable generating load/store vector pairs for block copies.
> > 
> > Testing has found that using load and store vector pair for block copies
> > can result in a slow down on power10.  This patch disables using the
> > vector pair instructions for block copies if we are tuning for power10.
> > 
> > 2022-06-11   Michael Meissner  <meissner@linux.ibm.com>
> > 
> > gcc/
> > 
> > 	* config/rs6000/rs6000.cc (rs6000_option_override_internal): Do
> > 	not generate block copies with vector pair instructions if we are
> > 	tuning for power10.  Back port from master branch.
> 
> You never posted the trunk version of this, so that never was approved
> either.

I did post the trunk version on June 10th, and your only comment was fix the
commit message, which I thought I did in the commit.

> > +++ b/gcc/config/rs6000/rs6000.cc
> > @@ -4151,7 +4151,10 @@ rs6000_option_override_internal (bool global_init_p)
> >  
> >    if (!(rs6000_isa_flags_explicit & OPTION_MASK_BLOCK_OPS_VECTOR_PAIR))
> >      {
> > -      if (TARGET_MMA && TARGET_EFFICIENT_UNALIGNED_VSX)
> > +      /* Do not generate lxvp and stxvp on power10 since there are some
> > +	 performance issues.  */
> > +      if (TARGET_MMA && TARGET_EFFICIENT_UNALIGNED_VSX
> > +	  && rs6000_tune != PROCESSOR_POWER10)
> >  	rs6000_isa_flags |= OPTION_MASK_BLOCK_OPS_VECTOR_PAIR;
> >        else
> >  	rs6000_isa_flags &= ~OPTION_MASK_BLOCK_OPS_VECTOR_PAIR;
> 
> The TARGET_MMA in that should not be there.  Please fix that (that
> probably needs more changes).


All of the movoo and movxo support require TARGET_MMA as does the code in
rs6000-string.cc that could possibly generate load/store vector pair.  To
remove the check here would mean also fixing all of the vector load and store
pairs in mma.md.

> This statement does the opposite of what the comment says.
> 
> Please fix this.  On trunk, first.

-- 
Michael Meissner, IBM
PO Box 98, Ayer, Massachusetts, USA, 01432
email: meissner@linux.ibm.com