diff mbox series

aarch64: Prefer full-vector SVE main loops for hip12

Message ID 717fa9845be54931a606016b1aba3227@huawei.com
State New
Headers show
Series aarch64: Prefer full-vector SVE main loops for hip12 | expand

Commit Message

xiezhiheng Nov. 4, 2026, 8:49 a.m. UTC
For the assembly of fully-masked WHILELO loops, such as the following:

.L2:
  ld1h    z31.h, p7/z, [x0, x2, lsl 1]
  ld1h    z30.h, p7/z, [x1, x2, lsl 1]
  ...
  whilelo p7.h, w2, w4
  b.any   .L2

When WHILELO computes the predicate register p7, register dependencies
prevent the load instruction of the next loop iteration from being issued.
This behavior performs poorly on the hip12 core, and performance
regressions have been observed on hip12 for general fully-masked WHILELO
loop cases. Therefore, on the hip12 core, param_vect_partial_vector_usage
is set to 1 to generate full-vector main loops.

Bootstrapped and tested on aarch64-linux-gnu.

gcc/ChangeLog:

	* config/aarch64/aarch64-tuning-flags.def (AARCH64_EXTRA_TUNING_OPTION):
	Add PREFER_SVE_FULL_VECTORS tune parameter.
	* config/aarch64/aarch64.cc (aarch64_override_options_internal):
	Set param_vect_partial_vector_usage to 1.
	* config/aarch64/tuning_models/hip12.h: Set tune parameter
	PREFER_SVE_FULL_VECTORS.

Signed-off-by: Zhiheng Xie <xiezhiheng@huawei.com>
---
 gcc/config/aarch64/aarch64-tuning-flags.def | 4 ++++
 gcc/config/aarch64/aarch64.cc               | 8 ++++++++
 gcc/config/aarch64/tuning_models/hip12.h    | 3 ++-
 3 files changed, 14 insertions(+), 1 deletion(-)

Comments

Tamar Christina Nov. 10, 2026, 10:11 a.m. UTC | #1
> -----Original Message-----
> From: xiezhiheng <xiezhiheng@huawei.com>
> Sent: 04 November 2026 08:50
> To: gcc-patches@gcc.gnu.org
> Cc: Kyrylo Tkachov <ktkachov@nvidia.com>; Tamar Christina
> <Tamar.Christina@arm.com>
> Subject: aarch64: Prefer full-vector SVE main loops for hip12
> 
> For the assembly of fully-masked WHILELO loops, such as the following:
> 
> .L2:
>   ld1h    z31.h, p7/z, [x0, x2, lsl 1]
>   ld1h    z30.h, p7/z, [x1, x2, lsl 1]
>   ...
>   whilelo p7.h, w2, w4
>   b.any   .L2
> 
> When WHILELO computes the predicate register p7, register dependencies
> prevent the load instruction of the next loop iteration from being issued.
> This behavior performs poorly on the hip12 core, and performance
> regressions have been observed on hip12 for general fully-masked WHILELO
> loop cases. Therefore, on the hip12 core, param_vect_partial_vector_usage
> is set to 1 to generate full-vector main loops.
> 
> Bootstrapped and tested on aarch64-linux-gnu.
> 
> gcc/ChangeLog:
> 
> 	* config/aarch64/aarch64-tuning-flags.def
> (AARCH64_EXTRA_TUNING_OPTION):
> 	Add PREFER_SVE_FULL_VECTORS tune parameter.
> 	* config/aarch64/aarch64.cc (aarch64_override_options_internal):
> 	Set param_vect_partial_vector_usage to 1.
> 	* config/aarch64/tuning_models/hip12.h: Set tune parameter
> 	PREFER_SVE_FULL_VECTORS.
> 

OK.

At some point I'll look into making this driveable from costing instead.

You have a few patches now waiting on the sched2 pressure patch.
Do you want me to apply the chunks that are OK now?

Thanks,
Tamar

> Signed-off-by: Zhiheng Xie <xiezhiheng@huawei.com>
> ---
>  gcc/config/aarch64/aarch64-tuning-flags.def | 4 ++++
>  gcc/config/aarch64/aarch64.cc               | 8 ++++++++
>  gcc/config/aarch64/tuning_models/hip12.h    | 3 ++-
>  3 files changed, 14 insertions(+), 1 deletion(-)
> 
> diff --git a/gcc/config/aarch64/aarch64-tuning-flags.def
> b/gcc/config/aarch64/aarch64-tuning-flags.def
> index 64a773be34c..b29189bac72 100644
> --- a/gcc/config/aarch64/aarch64-tuning-flags.def
> +++ b/gcc/config/aarch64/aarch64-tuning-flags.def
> @@ -84,4 +84,8 @@ AARCH64_EXTRA_TUNING_OPTION
> ("narrow_gp_writes", NARROW_GP_WRITES)
>  /* Enable when the target prefers SVE merging movimm over zeroing.  */
>  AARCH64_EXTRA_TUNING_OPTION ("avoid_zeroing_movimm",
> AVOID_MOVIMM_Z)
> 
> +/* Enable if the target prefers full-vector main loops rather than
> +   fully-masked whilelo loops (i.e. -param=vect-partial-vector-usage=1).  */
> +AARCH64_EXTRA_TUNING_OPTION ("prefer_sve_full_vectors",
> PREFER_SVE_FULL_VECTORS)
> +
>  #undef AARCH64_EXTRA_TUNING_OPTION
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index 6c1229d0a95..cee5f6e449c 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -20436,6 +20436,14 @@ aarch64_override_options_internal (struct
> gcc_options *opts,
>        & AARCH64_EXTRA_TUNE_DISPATCH_SCHED)
>      gcc_assert (aarch64_tune_params.dispatch_constraints != NULL);
> 
> +  /* Prefer SVE main loops that operate on full vectors, handling the
> +     tail in a separate epilogue rather than with fully-masked WHILELO
> +     loops.  */
> +  if (aarch64_tune_params.extra_tuning_flags
> +      & AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS)
> +    SET_OPTION_IF_UNSET (opts, opts_set, param_vect_partial_vector_usage,
> +			 1);
> +
>    /* Enable possible unprofitable vectorization.  */
>    if (opts->x_flag_aarch64_max_vectorization)
>      SET_OPTION_IF_UNSET (opts, opts_set,
> diff --git a/gcc/config/aarch64/tuning_models/hip12.h
> b/gcc/config/aarch64/tuning_models/hip12.h
> index cfb204b2e53..30d07d990a0 100644
> --- a/gcc/config/aarch64/tuning_models/hip12.h
> +++ b/gcc/config/aarch64/tuning_models/hip12.h
> @@ -217,7 +217,8 @@ static const struct tune_params hip12_tunings =
>    0,    /* max_case_values.  */
>    tune_params::AUTOPREFETCHER_WEAK,    /* autoprefetcher_model.  */
>    (AARCH64_EXTRA_TUNE_BASE
> -   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT),    /*
> tune_flags.  */
> +   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT
> +   | AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS),    /* tune_flags.
> */
>    &generic_armv8_a_prefetch_tune,
>    AARCH64_LDP_STP_POLICY_ALWAYS,    /* ldp_policy_model.  */
>    AARCH64_LDP_STP_POLICY_ALWAYS,    /* stp_policy_model.  */
> --
> 2.43.0
xiezhiheng Nov. 10, 2026, 11:25 a.m. UTC | #2
> -----Original Message-----
> From: Tamar Christina <Tamar.Christina@arm.com>
> Sent: Tuesday, November 10, 2026 6:12 PM
> To: xiezhiheng <xiezhiheng@huawei.com>; gcc-patches@gcc.gnu.org
> Cc: Kyrylo Tkachov <ktkachov@nvidia.com>
> Subject: RE: aarch64: Prefer full-vector SVE main loops for hip12
> 
> > -----Original Message-----
> > From: xiezhiheng <xiezhiheng@huawei.com>
> > Sent: 04 November 2026 08:50
> > To: gcc-patches@gcc.gnu.org
> > Cc: Kyrylo Tkachov <ktkachov@nvidia.com>; Tamar Christina
> > <Tamar.Christina@arm.com>
> > Subject: aarch64: Prefer full-vector SVE main loops for hip12
> >
> > For the assembly of fully-masked WHILELO loops, such as the following:
> >
> > .L2:
> >   ld1h    z31.h, p7/z, [x0, x2, lsl 1]
> >   ld1h    z30.h, p7/z, [x1, x2, lsl 1]
> >   ...
> >   whilelo p7.h, w2, w4
> >   b.any   .L2
> >
> > When WHILELO computes the predicate register p7, register dependencies
> > prevent the load instruction of the next loop iteration from being issued.
> > This behavior performs poorly on the hip12 core, and performance
> > regressions have been observed on hip12 for general fully-masked WHILELO
> > loop cases. Therefore, on the hip12 core, param_vect_partial_vector_usage
> > is set to 1 to generate full-vector main loops.
> >
> > Bootstrapped and tested on aarch64-linux-gnu.
> >
> > gcc/ChangeLog:
> >
> > 	* config/aarch64/aarch64-tuning-flags.def
> > (AARCH64_EXTRA_TUNING_OPTION):
> > 	Add PREFER_SVE_FULL_VECTORS tune parameter.
> > 	* config/aarch64/aarch64.cc (aarch64_override_options_internal):
> > 	Set param_vect_partial_vector_usage to 1.
> > 	* config/aarch64/tuning_models/hip12.h: Set tune parameter
> > 	PREFER_SVE_FULL_VECTORS.
> >
> 
> OK.
> 
> At some point I'll look into making this driveable from costing instead.
> 
> You have a few patches now waiting on the sched2 pressure patch.
> Do you want me to apply the chunks that are OK now?

Sure, please help push this patch and the hip12 scheduling model patch.
https://gcc.gnu.org/pipermail/gcc-patches/2026-October/736869.html

Thanks a lot,
Zhiheng


> 
> Thanks,
> Tamar
> 
> > Signed-off-by: Zhiheng Xie <xiezhiheng@huawei.com>
> > ---
> >  gcc/config/aarch64/aarch64-tuning-flags.def | 4 ++++
> >  gcc/config/aarch64/aarch64.cc               | 8 ++++++++
> >  gcc/config/aarch64/tuning_models/hip12.h    | 3 ++-
> >  3 files changed, 14 insertions(+), 1 deletion(-)
> >
> > diff --git a/gcc/config/aarch64/aarch64-tuning-flags.def
> > b/gcc/config/aarch64/aarch64-tuning-flags.def
> > index 64a773be34c..b29189bac72 100644
> > --- a/gcc/config/aarch64/aarch64-tuning-flags.def
> > +++ b/gcc/config/aarch64/aarch64-tuning-flags.def
> > @@ -84,4 +84,8 @@ AARCH64_EXTRA_TUNING_OPTION
> > ("narrow_gp_writes", NARROW_GP_WRITES)
> >  /* Enable when the target prefers SVE merging movimm over zeroing.  */
> >  AARCH64_EXTRA_TUNING_OPTION ("avoid_zeroing_movimm",
> > AVOID_MOVIMM_Z)
> >
> > +/* Enable if the target prefers full-vector main loops rather than
> > +   fully-masked whilelo loops (i.e. -param=vect-partial-vector-usage=1).  */
> > +AARCH64_EXTRA_TUNING_OPTION ("prefer_sve_full_vectors",
> > PREFER_SVE_FULL_VECTORS)
> > +
> >  #undef AARCH64_EXTRA_TUNING_OPTION
> > diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> > index 6c1229d0a95..cee5f6e449c 100644
> > --- a/gcc/config/aarch64/aarch64.cc
> > +++ b/gcc/config/aarch64/aarch64.cc
> > @@ -20436,6 +20436,14 @@ aarch64_override_options_internal (struct
> > gcc_options *opts,
> >        & AARCH64_EXTRA_TUNE_DISPATCH_SCHED)
> >      gcc_assert (aarch64_tune_params.dispatch_constraints != NULL);
> >
> > +  /* Prefer SVE main loops that operate on full vectors, handling the
> > +     tail in a separate epilogue rather than with fully-masked WHILELO
> > +     loops.  */
> > +  if (aarch64_tune_params.extra_tuning_flags
> > +      & AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS)
> > +    SET_OPTION_IF_UNSET (opts, opts_set, param_vect_partial_vector_usage,
> > +			 1);
> > +
> >    /* Enable possible unprofitable vectorization.  */
> >    if (opts->x_flag_aarch64_max_vectorization)
> >      SET_OPTION_IF_UNSET (opts, opts_set,
> > diff --git a/gcc/config/aarch64/tuning_models/hip12.h
> > b/gcc/config/aarch64/tuning_models/hip12.h
> > index cfb204b2e53..30d07d990a0 100644
> > --- a/gcc/config/aarch64/tuning_models/hip12.h
> > +++ b/gcc/config/aarch64/tuning_models/hip12.h
> > @@ -217,7 +217,8 @@ static const struct tune_params hip12_tunings =
> >    0,    /* max_case_values.  */
> >    tune_params::AUTOPREFETCHER_WEAK,    /* autoprefetcher_model.  */
> >    (AARCH64_EXTRA_TUNE_BASE
> > -   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT),    /*
> > tune_flags.  */
> > +   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT
> > +   | AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS),    /* tune_flags.
> > */
> >    &generic_armv8_a_prefetch_tune,
> >    AARCH64_LDP_STP_POLICY_ALWAYS,    /* ldp_policy_model.  */
> >    AARCH64_LDP_STP_POLICY_ALWAYS,    /* stp_policy_model.  */
> > --
> > 2.43.0
>
Tamar Christina Nov. 10, 2026, 11:45 a.m. UTC | #3
> -----Original Message-----
> From: xiezhiheng <xiezhiheng@huawei.com>
> Sent: 10 November 2026 11:25
> To: Tamar Christina <Tamar.Christina@arm.com>; gcc-patches@gcc.gnu.org
> Cc: Kyrylo Tkachov <ktkachov@nvidia.com>
> Subject: RE: aarch64: Prefer full-vector SVE main loops for hip12
> 
> > -----Original Message-----
> > From: Tamar Christina <Tamar.Christina@arm.com>
> > Sent: Tuesday, November 10, 2026 6:12 PM
> > To: xiezhiheng <xiezhiheng@huawei.com>; gcc-patches@gcc.gnu.org
> > Cc: Kyrylo Tkachov <ktkachov@nvidia.com>
> > Subject: RE: aarch64: Prefer full-vector SVE main loops for hip12
> >
> > > -----Original Message-----
> > > From: xiezhiheng <xiezhiheng@huawei.com>
> > > Sent: 04 November 2026 08:50
> > > To: gcc-patches@gcc.gnu.org
> > > Cc: Kyrylo Tkachov <ktkachov@nvidia.com>; Tamar Christina
> > > <Tamar.Christina@arm.com>
> > > Subject: aarch64: Prefer full-vector SVE main loops for hip12
> > >
> > > For the assembly of fully-masked WHILELO loops, such as the following:
> > >
> > > .L2:
> > >   ld1h    z31.h, p7/z, [x0, x2, lsl 1]
> > >   ld1h    z30.h, p7/z, [x1, x2, lsl 1]
> > >   ...
> > >   whilelo p7.h, w2, w4
> > >   b.any   .L2
> > >
> > > When WHILELO computes the predicate register p7, register dependencies
> > > prevent the load instruction of the next loop iteration from being issued.
> > > This behavior performs poorly on the hip12 core, and performance
> > > regressions have been observed on hip12 for general fully-masked
> WHILELO
> > > loop cases. Therefore, on the hip12 core, param_vect_partial_vector_usage
> > > is set to 1 to generate full-vector main loops.
> > >
> > > Bootstrapped and tested on aarch64-linux-gnu.
> > >
> > > gcc/ChangeLog:
> > >
> > > 	* config/aarch64/aarch64-tuning-flags.def
> > > (AARCH64_EXTRA_TUNING_OPTION):
> > > 	Add PREFER_SVE_FULL_VECTORS tune parameter.
> > > 	* config/aarch64/aarch64.cc (aarch64_override_options_internal):
> > > 	Set param_vect_partial_vector_usage to 1.
> > > 	* config/aarch64/tuning_models/hip12.h: Set tune parameter
> > > 	PREFER_SVE_FULL_VECTORS.
> > >
> >
> > OK.
> >
> > At some point I'll look into making this driveable from costing instead.
> >
> > You have a few patches now waiting on the sched2 pressure patch.
> > Do you want me to apply the chunks that are OK now?
> 
> Sure, please help push this patch and the hip12 scheduling model patch.
> https://gcc.gnu.org/pipermail/gcc-patches/2026-October/736869.html

Done, pushed as

3a08e8b95a4 aarch64: Prefer full-vector SVE main loops for hip12
746a369911a aarch64: Add hip12 pipeline scheduling model


Thanks,
Tamar

> 
> Thanks a lot,
> Zhiheng
> 
> 
> >
> > Thanks,
> > Tamar
> >
> > > Signed-off-by: Zhiheng Xie <xiezhiheng@huawei.com>
> > > ---
> > >  gcc/config/aarch64/aarch64-tuning-flags.def | 4 ++++
> > >  gcc/config/aarch64/aarch64.cc               | 8 ++++++++
> > >  gcc/config/aarch64/tuning_models/hip12.h    | 3 ++-
> > >  3 files changed, 14 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/gcc/config/aarch64/aarch64-tuning-flags.def
> > > b/gcc/config/aarch64/aarch64-tuning-flags.def
> > > index 64a773be34c..b29189bac72 100644
> > > --- a/gcc/config/aarch64/aarch64-tuning-flags.def
> > > +++ b/gcc/config/aarch64/aarch64-tuning-flags.def
> > > @@ -84,4 +84,8 @@ AARCH64_EXTRA_TUNING_OPTION
> > > ("narrow_gp_writes", NARROW_GP_WRITES)
> > >  /* Enable when the target prefers SVE merging movimm over zeroing.  */
> > >  AARCH64_EXTRA_TUNING_OPTION ("avoid_zeroing_movimm",
> > > AVOID_MOVIMM_Z)
> > >
> > > +/* Enable if the target prefers full-vector main loops rather than
> > > +   fully-masked whilelo loops (i.e. -param=vect-partial-vector-usage=1).  */
> > > +AARCH64_EXTRA_TUNING_OPTION ("prefer_sve_full_vectors",
> > > PREFER_SVE_FULL_VECTORS)
> > > +
> > >  #undef AARCH64_EXTRA_TUNING_OPTION
> > > diff --git a/gcc/config/aarch64/aarch64.cc
> b/gcc/config/aarch64/aarch64.cc
> > > index 6c1229d0a95..cee5f6e449c 100644
> > > --- a/gcc/config/aarch64/aarch64.cc
> > > +++ b/gcc/config/aarch64/aarch64.cc
> > > @@ -20436,6 +20436,14 @@ aarch64_override_options_internal (struct
> > > gcc_options *opts,
> > >        & AARCH64_EXTRA_TUNE_DISPATCH_SCHED)
> > >      gcc_assert (aarch64_tune_params.dispatch_constraints != NULL);
> > >
> > > +  /* Prefer SVE main loops that operate on full vectors, handling the
> > > +     tail in a separate epilogue rather than with fully-masked WHILELO
> > > +     loops.  */
> > > +  if (aarch64_tune_params.extra_tuning_flags
> > > +      & AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS)
> > > +    SET_OPTION_IF_UNSET (opts, opts_set,
> param_vect_partial_vector_usage,
> > > +			 1);
> > > +
> > >    /* Enable possible unprofitable vectorization.  */
> > >    if (opts->x_flag_aarch64_max_vectorization)
> > >      SET_OPTION_IF_UNSET (opts, opts_set,
> > > diff --git a/gcc/config/aarch64/tuning_models/hip12.h
> > > b/gcc/config/aarch64/tuning_models/hip12.h
> > > index cfb204b2e53..30d07d990a0 100644
> > > --- a/gcc/config/aarch64/tuning_models/hip12.h
> > > +++ b/gcc/config/aarch64/tuning_models/hip12.h
> > > @@ -217,7 +217,8 @@ static const struct tune_params hip12_tunings =
> > >    0,    /* max_case_values.  */
> > >    tune_params::AUTOPREFETCHER_WEAK,    /* autoprefetcher_model.  */
> > >    (AARCH64_EXTRA_TUNE_BASE
> > > -   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT),    /*
> > > tune_flags.  */
> > > +   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT
> > > +   | AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS),    /*
> tune_flags.
> > > */
> > >    &generic_armv8_a_prefetch_tune,
> > >    AARCH64_LDP_STP_POLICY_ALWAYS,    /* ldp_policy_model.  */
> > >    AARCH64_LDP_STP_POLICY_ALWAYS,    /* stp_policy_model.  */
> > > --
> > > 2.43.0
> >
diff mbox series

Patch

diff --git a/gcc/config/aarch64/aarch64-tuning-flags.def b/gcc/config/aarch64/aarch64-tuning-flags.def
index 64a773be34c..b29189bac72 100644
--- a/gcc/config/aarch64/aarch64-tuning-flags.def
+++ b/gcc/config/aarch64/aarch64-tuning-flags.def
@@ -84,4 +84,8 @@  AARCH64_EXTRA_TUNING_OPTION ("narrow_gp_writes", NARROW_GP_WRITES)
 /* Enable when the target prefers SVE merging movimm over zeroing.  */
 AARCH64_EXTRA_TUNING_OPTION ("avoid_zeroing_movimm", AVOID_MOVIMM_Z)
 
+/* Enable if the target prefers full-vector main loops rather than
+   fully-masked whilelo loops (i.e. -param=vect-partial-vector-usage=1).  */
+AARCH64_EXTRA_TUNING_OPTION ("prefer_sve_full_vectors", PREFER_SVE_FULL_VECTORS)
+
 #undef AARCH64_EXTRA_TUNING_OPTION
diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
index 6c1229d0a95..cee5f6e449c 100644
--- a/gcc/config/aarch64/aarch64.cc
+++ b/gcc/config/aarch64/aarch64.cc
@@ -20436,6 +20436,14 @@  aarch64_override_options_internal (struct gcc_options *opts,
       & AARCH64_EXTRA_TUNE_DISPATCH_SCHED)
     gcc_assert (aarch64_tune_params.dispatch_constraints != NULL);
 
+  /* Prefer SVE main loops that operate on full vectors, handling the
+     tail in a separate epilogue rather than with fully-masked WHILELO
+     loops.  */
+  if (aarch64_tune_params.extra_tuning_flags
+      & AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS)
+    SET_OPTION_IF_UNSET (opts, opts_set, param_vect_partial_vector_usage,
+			 1);
+
   /* Enable possible unprofitable vectorization.  */
   if (opts->x_flag_aarch64_max_vectorization)
     SET_OPTION_IF_UNSET (opts, opts_set,
diff --git a/gcc/config/aarch64/tuning_models/hip12.h b/gcc/config/aarch64/tuning_models/hip12.h
index cfb204b2e53..30d07d990a0 100644
--- a/gcc/config/aarch64/tuning_models/hip12.h
+++ b/gcc/config/aarch64/tuning_models/hip12.h
@@ -217,7 +217,8 @@  static const struct tune_params hip12_tunings =
   0,    /* max_case_values.  */
   tune_params::AUTOPREFETCHER_WEAK,    /* autoprefetcher_model.  */
   (AARCH64_EXTRA_TUNE_BASE
-   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT),    /* tune_flags.  */
+   | AARCH64_EXTRA_TUNE_MATCHED_VECTOR_THROUGHPUT
+   | AARCH64_EXTRA_TUNE_PREFER_SVE_FULL_VECTORS),    /* tune_flags.  */
   &generic_armv8_a_prefetch_tune,
   AARCH64_LDP_STP_POLICY_ALWAYS,    /* ldp_policy_model.  */
   AARCH64_LDP_STP_POLICY_ALWAYS,    /* stp_policy_model.  */