diff mbox series

tree-optimization: [PR101540] Simplify CONSTRUCTOR for vector(1) to be VCE

Message ID 1638122190-9379-1-git-send-email-apinski@marvell.com
State New
Headers
Series tree-optimization: [PR101540] Simplify CONSTRUCTOR for vector(1) to be VCE |

Commit Message

Li, Pan2 via Gcc-patches Nov. 28, 2021, 5:56 p.m. UTC 
  From: Andrew Pinski <apinski@marvell.com>

This just adds a simplification to simplify_vector_constructor for
vector of 1 element to be VCE which should reduce memory usage in
the compiler and maybe allow for some more optimizations.

OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.

	PR tree-optimization/101540

gcc/ChangeLog:

	* tree-ssa-forwprop.c (simplify_vector_constructor):
	Simplify constructor of vector of 1 element to just
	be a VIEW_CONVERT_EXPR.

gcc/testsuite/ChangeLog:

	* gcc.dg/tree-ssa/pr101540-1.c: New test.
---
 gcc/testsuite/gcc.dg/tree-ssa/pr101540-1.c | 13 +++++++++++++
 gcc/tree-ssa-forwprop.c                    | 13 +++++++++++++
 2 files changed, 26 insertions(+)
 create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr101540-1.c

Comments

Jeff Law Nov. 28, 2021, 8:25 p.m. UTC | #1 
On 11/28/2021 10:56 AM, apinski--- via Gcc-patches wrote:
> From: Andrew Pinski <apinski@marvell.com>
>
> This just adds a simplification to simplify_vector_constructor for
> vector of 1 element to be VCE which should reduce memory usage in
> the compiler and maybe allow for some more optimizations.
>
> OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
>
> 	PR tree-optimization/101540
>
> gcc/ChangeLog:
>
> 	* tree-ssa-forwprop.c (simplify_vector_constructor):
> 	Simplify constructor of vector of 1 element to just
> 	be a VIEW_CONVERT_EXPR.
>
> gcc/testsuite/ChangeLog:
>
> 	* gcc.dg/tree-ssa/pr101540-1.c: New test.
So why generate a VCE here if the type conversion is useless?  Why not 
just a NOP_EXPR?  Is there something special about converting between 
the element type and the outer vector type that requires VCE rather than 
NOP_EXR?  Neither an ACK or NAK, just trying to understand it a bit better.

Jeff
Andrew Pinski Nov. 29, 2021, 12:56 a.m. UTC | #2 
On Sun, Nov 28, 2021 at 12:25 PM Jeff Law via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
>
>
> On 11/28/2021 10:56 AM, apinski--- via Gcc-patches wrote:
> > From: Andrew Pinski <apinski@marvell.com>
> >
> > This just adds a simplification to simplify_vector_constructor for
> > vector of 1 element to be VCE which should reduce memory usage in
> > the compiler and maybe allow for some more optimizations.
> >
> > OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
> >
> >       PR tree-optimization/101540
> >
> > gcc/ChangeLog:
> >
> >       * tree-ssa-forwprop.c (simplify_vector_constructor):
> >       Simplify constructor of vector of 1 element to just
> >       be a VIEW_CONVERT_EXPR.
> >
> > gcc/testsuite/ChangeLog:
> >
> >       * gcc.dg/tree-ssa/pr101540-1.c: New test.
> So why generate a VCE here if the type conversion is useless?  Why not
> just a NOP_EXPR?  Is there something special about converting between
> the element type and the outer vector type that requires VCE rather than
> NOP_EXR?  Neither an ACK or NAK, just trying to understand it a bit better.


Because right now tree-cfg.c has this check for vector types for NOP_EXPR:
        /* Allow conversions between vectors with the same number of elements,
           provided that the conversion is OK for the element types too.  */
        if (VECTOR_TYPE_P (lhs_type)
            && VECTOR_TYPE_P (rhs1_type)
            && known_eq (TYPE_VECTOR_SUBPARTS (lhs_type),
                         TYPE_VECTOR_SUBPARTS (rhs1_type)))
          {
            lhs_type = TREE_TYPE (lhs_type);
            rhs1_type = TREE_TYPE (rhs1_type);
          }
        else if (VECTOR_TYPE_P (lhs_type) || VECTOR_TYPE_P (rhs1_type))
          {
            error ("invalid vector types in nop conversion");
            debug_generic_expr (lhs_type);
            debug_generic_expr (rhs1_type);
            return true;
          }

We can change this check here for NOP_EXPR and vector types but VCE is
still a nop in most cases and handled as such really. But I wonder if
the rest of the compiler is ready for it though.

Thanks,
Andrew Pinski

>
> Jeff
>
>
Richard Biener Nov. 29, 2021, 8:59 a.m. UTC | #3 
On Mon, Nov 29, 2021 at 1:57 AM Andrew Pinski via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Sun, Nov 28, 2021 at 12:25 PM Jeff Law via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> >
> >
> > On 11/28/2021 10:56 AM, apinski--- via Gcc-patches wrote:
> > > From: Andrew Pinski <apinski@marvell.com>
> > >
> > > This just adds a simplification to simplify_vector_constructor for
> > > vector of 1 element to be VCE which should reduce memory usage in
> > > the compiler and maybe allow for some more optimizations.
> > >
> > > OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
> > >
> > >       PR tree-optimization/101540
> > >
> > > gcc/ChangeLog:
> > >
> > >       * tree-ssa-forwprop.c (simplify_vector_constructor):
> > >       Simplify constructor of vector of 1 element to just
> > >       be a VIEW_CONVERT_EXPR.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > >       * gcc.dg/tree-ssa/pr101540-1.c: New test.
> > So why generate a VCE here if the type conversion is useless?  Why not
> > just a NOP_EXPR?  Is there something special about converting between
> > the element type and the outer vector type that requires VCE rather than
> > NOP_EXR?  Neither an ACK or NAK, just trying to understand it a bit better.
>
>
> Because right now tree-cfg.c has this check for vector types for NOP_EXPR:
>         /* Allow conversions between vectors with the same number of elements,
>            provided that the conversion is OK for the element types too.  */
>         if (VECTOR_TYPE_P (lhs_type)
>             && VECTOR_TYPE_P (rhs1_type)
>             && known_eq (TYPE_VECTOR_SUBPARTS (lhs_type),
>                          TYPE_VECTOR_SUBPARTS (rhs1_type)))
>           {
>             lhs_type = TREE_TYPE (lhs_type);
>             rhs1_type = TREE_TYPE (rhs1_type);
>           }
>         else if (VECTOR_TYPE_P (lhs_type) || VECTOR_TYPE_P (rhs1_type))
>           {
>             error ("invalid vector types in nop conversion");
>             debug_generic_expr (lhs_type);
>             debug_generic_expr (rhs1_type);
>             return true;
>           }
>
> We can change this check here for NOP_EXPR and vector types but VCE is
> still a nop in most cases and handled as such really. But I wonder if
> the rest of the compiler is ready for it though.

It's definitely not a NOP, I think the original patch is OK.

Thanks,
Richard.

>
> Thanks,
> Andrew Pinski
>
> >
> > Jeff
> >
> >
diff mbox series

Patch

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr101540-1.c b/gcc/testsuite/gcc.dg/tree-ssa/pr101540-1.c
new file mode 100644
index 00000000000..73fb342e029
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr101540-1.c
@@ -0,0 +1,13 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-forwprop1" } */
+/* PR tree-optimization/101540 */
+typedef unsigned char __attribute__((__vector_size__ (1))) W;
+
+W foo (unsigned char uc)
+{
+  return (W){uc};
+}
+/* The constructor in the above function should be converted into a VCE.  */
+/* { dg-final { scan-tree-dump-times "VIEW_CONVERT_EXPR" 1 "forwprop1"} } */
+// {uc_1(D)}
+/* { dg-final { scan-tree-dump-times "{uc_\[0-9\]+.D.}" 0 "forwprop1"} } */
diff --git a/gcc/tree-ssa-forwprop.c b/gcc/tree-ssa-forwprop.c
index a830bab78ba..94b92d3d0af 100644
--- a/gcc/tree-ssa-forwprop.c
+++ b/gcc/tree-ssa-forwprop.c
@@ -2392,6 +2392,19 @@  simplify_vector_constructor (gimple_stmt_iterator *gsi)
   elem_type = TREE_TYPE (type);
   elem_size = TREE_INT_CST_LOW (TYPE_SIZE (elem_type));
 
+  /* Special case V1 constructor with the same type to being a VCE.  */
+  if (nelts == 1 && CONSTRUCTOR_NELTS (op) == 1)
+    {
+      tree op1 = CONSTRUCTOR_ELT (op, 0)->value;
+      if (useless_type_conversion_p (elem_type, TREE_TYPE (op1)))
+	{
+	  op1 = build1 (VIEW_CONVERT_EXPR, type, op1);
+	  gimple_assign_set_rhs_from_tree (gsi, op1);
+	  update_stmt (gsi_stmt (*gsi));
+	  return true;
+	}
+    }
+
   orig[0] = NULL;
   orig[1] = NULL;
   conv_code = ERROR_MARK;