tree-optimization/110221 - SLP and loop mask/len

Message ID 20231110131658.09A5D13398@imap2.suse-dmz.suse.de
State Accepted
Headers
Series tree-optimization/110221 - SLP and loop mask/len |

Checks

Context Check Description
snail/gcc-patch-check success Github commit url

Commit Message

Richard Biener Nov. 10, 2023, 1:16 p.m. UTC
  The following fixes the issue that when SLP stmts are internal defs
but appear invariant because they end up only using invariant defs
then they get scheduled outside of the loop.  This nice optimization
breaks down when loop masks or lens are applied since those are not
explicitly tracked as dependences.  The following makes sure to never
schedule internal defs outside of the vectorized loop when the
loop uses masks/lens.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

	PR tree-optimization/110221
	* tree-vect-slp.cc (vect_schedule_slp_node): When loop
	masking / len is applied make sure to not schedule
	intenal defs outside of the loop.

	* gfortran.dg/pr110221.f: New testcase.
---
 gcc/testsuite/gfortran.dg/pr110221.f | 17 +++++++++++++++++
 gcc/tree-vect-slp.cc                 | 10 ++++++++++
 2 files changed, 27 insertions(+)
 create mode 100644 gcc/testsuite/gfortran.dg/pr110221.f
  

Comments

Andre Vieira (lists) March 1, 2024, 9:24 a.m. UTC | #1
Hi,

Bootstrapped and tested the gcc-13 backport of this on gcc-12 for 
aarch64-unknown-linux-gnu and x86_64-pc-linux-gnu and no regressions.

OK to push to gcc-12 branch?

Kind regards,
Andre Vieira

On 10/11/2023 13:16, Richard Biener wrote:
> The following fixes the issue that when SLP stmts are internal defs
> but appear invariant because they end up only using invariant defs
> then they get scheduled outside of the loop.  This nice optimization
> breaks down when loop masks or lens are applied since those are not
> explicitly tracked as dependences.  The following makes sure to never
> schedule internal defs outside of the vectorized loop when the
> loop uses masks/lens.
> 
> Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
> 
> 	PR tree-optimization/110221
> 	* tree-vect-slp.cc (vect_schedule_slp_node): When loop
> 	masking / len is applied make sure to not schedule
> 	intenal defs outside of the loop.
> 
> 	* gfortran.dg/pr110221.f: New testcase.
> ---
>   gcc/testsuite/gfortran.dg/pr110221.f | 17 +++++++++++++++++
>   gcc/tree-vect-slp.cc                 | 10 ++++++++++
>   2 files changed, 27 insertions(+)
>   create mode 100644 gcc/testsuite/gfortran.dg/pr110221.f
> 
> diff --git a/gcc/testsuite/gfortran.dg/pr110221.f b/gcc/testsuite/gfortran.dg/pr110221.f
> new file mode 100644
> index 00000000000..8b57384313a
> --- /dev/null
> +++ b/gcc/testsuite/gfortran.dg/pr110221.f
> @@ -0,0 +1,17 @@
> +C PR middle-end/68146
> +C { dg-do compile }
> +C { dg-options "-O2 -w" }
> +C { dg-additional-options "-mavx512f --param vect-partial-vector-usage=2" { target avx512f } }
> +      SUBROUTINE CJYVB(V,Z,V0,CBJ,CDJ,CBY,CYY)
> +      IMPLICIT DOUBLE PRECISION (A,B,G,O-Y)
> +      IMPLICIT COMPLEX*16 (C,Z)
> +      DIMENSION CBJ(0:*),CDJ(0:*),CBY(0:*)
> +      N=INT(V)
> +      CALL GAMMA2(VG,GA)
> +      DO 65 K=1,N
> +        CBY(K)=CYY
> +65    CONTINUE
> +      CDJ(0)=V0/Z*CBJ(0)-CBJ(1)
> +      DO 70 K=1,N
> +70      CDJ(K)=-(K+V0)/Z*CBJ(K)+CBJ(K-1)
> +      END
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index 3e5814c3a31..80e279d8f50 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -9081,6 +9081,16 @@ vect_schedule_slp_node (vec_info *vinfo,
>         /* Emit other stmts after the children vectorized defs which is
>   	 earliest possible.  */
>         gimple *last_stmt = NULL;
> +      if (auto loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
> +	if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
> +	    || LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo))
> +	  {
> +	    /* But avoid scheduling internal defs outside of the loop when
> +	       we might have only implicitly tracked loop mask/len defs.  */
> +	    gimple_stmt_iterator si
> +	      = gsi_after_labels (LOOP_VINFO_LOOP (loop_vinfo)->header);
> +	    last_stmt = *si;
> +	  }
>         bool seen_vector_def = false;
>         FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
>   	if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
  
Richard Biener March 1, 2024, 9:48 a.m. UTC | #2
On Fri, 1 Mar 2024, Andre Vieira (lists) wrote:

> Hi,
> 
> Bootstrapped and tested the gcc-13 backport of this on gcc-12 for
> aarch64-unknown-linux-gnu and x86_64-pc-linux-gnu and no regressions.
> 
> OK to push to gcc-12 branch?

OK.

Thanks,
Richard.

> Kind regards,
> Andre Vieira
> 
> On 10/11/2023 13:16, Richard Biener wrote:
> > The following fixes the issue that when SLP stmts are internal defs
> > but appear invariant because they end up only using invariant defs
> > then they get scheduled outside of the loop.  This nice optimization
> > breaks down when loop masks or lens are applied since those are not
> > explicitly tracked as dependences.  The following makes sure to never
> > schedule internal defs outside of the vectorized loop when the
> > loop uses masks/lens.
> > 
> > Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
> > 
> >  PR tree-optimization/110221
> >  * tree-vect-slp.cc (vect_schedule_slp_node): When loop
> >  masking / len is applied make sure to not schedule
> >  intenal defs outside of the loop.
> > 
> > 	* gfortran.dg/pr110221.f: New testcase.
> > ---
> >   gcc/testsuite/gfortran.dg/pr110221.f | 17 +++++++++++++++++
> >   gcc/tree-vect-slp.cc                 | 10 ++++++++++
> >   2 files changed, 27 insertions(+)
> >   create mode 100644 gcc/testsuite/gfortran.dg/pr110221.f
> > 
> > diff --git a/gcc/testsuite/gfortran.dg/pr110221.f
> > b/gcc/testsuite/gfortran.dg/pr110221.f
> > new file mode 100644
> > index 00000000000..8b57384313a
> > --- /dev/null
> > +++ b/gcc/testsuite/gfortran.dg/pr110221.f
> > @@ -0,0 +1,17 @@
> > +C PR middle-end/68146
> > +C { dg-do compile }
> > +C { dg-options "-O2 -w" }
> > +C { dg-additional-options "-mavx512f --param vect-partial-vector-usage=2" {
> > target avx512f } }
> > +      SUBROUTINE CJYVB(V,Z,V0,CBJ,CDJ,CBY,CYY)
> > +      IMPLICIT DOUBLE PRECISION (A,B,G,O-Y)
> > +      IMPLICIT COMPLEX*16 (C,Z)
> > +      DIMENSION CBJ(0:*),CDJ(0:*),CBY(0:*)
> > +      N=INT(V)
> > +      CALL GAMMA2(VG,GA)
> > +      DO 65 K=1,N
> > +        CBY(K)=CYY
> > +65    CONTINUE
> > +      CDJ(0)=V0/Z*CBJ(0)-CBJ(1)
> > +      DO 70 K=1,N
> > +70      CDJ(K)=-(K+V0)/Z*CBJ(K)+CBJ(K-1)
> > +      END
> > diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> > index 3e5814c3a31..80e279d8f50 100644
> > --- a/gcc/tree-vect-slp.cc
> > +++ b/gcc/tree-vect-slp.cc
> > @@ -9081,6 +9081,16 @@ vect_schedule_slp_node (vec_info *vinfo,
> >           /* Emit other stmts after the children vectorized defs which is
> >     earliest possible.  */
> >         gimple *last_stmt = NULL;
> > +      if (auto loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
> > +	if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
> > +	    || LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo))
> > +	  {
> > +	    /* But avoid scheduling internal defs outside of the loop when
> > +	       we might have only implicitly tracked loop mask/len defs.  */
> > +	    gimple_stmt_iterator si
> > +	      = gsi_after_labels (LOOP_VINFO_LOOP (loop_vinfo)->header);
> > +	    last_stmt = *si;
> > +	  }
> >         bool seen_vector_def = false;
> >         FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
> >    if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
>
  

Patch

diff --git a/gcc/testsuite/gfortran.dg/pr110221.f b/gcc/testsuite/gfortran.dg/pr110221.f
new file mode 100644
index 00000000000..8b57384313a
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/pr110221.f
@@ -0,0 +1,17 @@ 
+C PR middle-end/68146
+C { dg-do compile }
+C { dg-options "-O2 -w" }
+C { dg-additional-options "-mavx512f --param vect-partial-vector-usage=2" { target avx512f } }
+      SUBROUTINE CJYVB(V,Z,V0,CBJ,CDJ,CBY,CYY)
+      IMPLICIT DOUBLE PRECISION (A,B,G,O-Y)
+      IMPLICIT COMPLEX*16 (C,Z)
+      DIMENSION CBJ(0:*),CDJ(0:*),CBY(0:*)
+      N=INT(V)
+      CALL GAMMA2(VG,GA)
+      DO 65 K=1,N
+        CBY(K)=CYY
+65    CONTINUE
+      CDJ(0)=V0/Z*CBJ(0)-CBJ(1)
+      DO 70 K=1,N
+70      CDJ(K)=-(K+V0)/Z*CBJ(K)+CBJ(K-1)
+      END
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 3e5814c3a31..80e279d8f50 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -9081,6 +9081,16 @@  vect_schedule_slp_node (vec_info *vinfo,
       /* Emit other stmts after the children vectorized defs which is
 	 earliest possible.  */
       gimple *last_stmt = NULL;
+      if (auto loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
+	if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
+	    || LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo))
+	  {
+	    /* But avoid scheduling internal defs outside of the loop when
+	       we might have only implicitly tracked loop mask/len defs.  */
+	    gimple_stmt_iterator si
+	      = gsi_after_labels (LOOP_VINFO_LOOP (loop_vinfo)->header);
+	    last_stmt = *si;
+	  }
       bool seen_vector_def = false;
       FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
 	if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)