Re: [aarch64] Code-gen for vector initialization involving constants

[Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>]

[-- Attachment #1: Type: text/plain, Size: 1613 bytes --]

On Thu, 25 May 2023 at 15:26, Prathamesh Kulkarni
<prathamesh.kulkarni@linaro.org> wrote:
>
> On Thu, 25 May 2023 at 13:04, Richard Sandiford
> <richard.sandiford@arm.com> wrote:
> >
> > LGTM, just a couple of comment tweaks:
> >
> > Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
> > > diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> > > index d6fc94015fa..db7ca4c28c3 100644
> > > --- a/gcc/config/aarch64/aarch64.cc
> > > +++ b/gcc/config/aarch64/aarch64.cc
> > > @@ -22332,6 +22332,46 @@ aarch64_unzip_vector_init (machine_mode mode, rtx vals, bool even_p)
> > >    return gen_rtx_PARALLEL (new_mode, vec);
> > >  }
> > >
> > > +/* Return true if INSN is a scalar move.  */
> >
> > s/INSN/SET/
> >
> > > +
> > > +static bool
> > > +scalar_move_insn_p (rtx set)
> > > +{
> > > +  rtx src = SET_SRC (set);
> > > +  rtx dest = SET_DEST (set);
> > > +  return (is_a<scalar_mode> (GET_MODE (dest))
> > > +       && aarch64_mov_operand (src, GET_MODE (dest)));
> > > +}
> > > +
> > > +/* Similar to seq_cost, but ignore cost for scalar moves.  This function
> > > +   is called from aarch64_expand_vector_init.  */
> >
> > Probably best to drop the second sentence.
> >
> > OK with those changes, thanks (no need to retest).
> Thanks, committed as ea9154dbc8fc86d4c617503ca5e6f02fed3a6a56.
Hi Richard,
The s32 case for single constant patch doesn't regress now after the
above commit.
Bootstrapped+tested on aarch64-linux-gnu, and verified that the new
tests pass for aarch64_be-linux-gnu.
Is it OK to commit ?

Thanks,
Prathamesh
>
> Thanks,
> Prathamesh
> >
> > Richard

[-- Attachment #2: gnu-780-9.txt --]
[-- Type: text/plain, Size: 6503 bytes --]

[aarch64] Improve code-gen for vector initialization with single constant element.

gcc/ChangeLog:
	* config/aarch64/aarc64.cc (aarch64_expand_vector_init): Tweak condition
	if (n_var == n_elts && n_elts <= 16) to allow a single constant,
	and if maxv == 1, use constant element for duplicating into register.

gcc/testsuite/ChangeLog:
	* gcc.target/aarch64/vec-init-single-const.c: New test.
	* gcc.target/aarch64/vec-init-single-const-be.c: Likewise.
	* gcc.target/aarch64/vec-init-single-const-2.c: Likewise.

diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
index 5b046d32b37..30d6e3e8d83 100644
--- a/gcc/config/aarch64/aarch64.cc
+++ b/gcc/config/aarch64/aarch64.cc
@@ -22192,7 +22192,7 @@ aarch64_expand_vector_init_fallback (rtx target, rtx vals)
      and matches[X][1] with the count of duplicate elements (if X is the
      earliest element which has duplicates).  */
 
-  if (n_var == n_elts && n_elts <= 16)
+  if (n_var >= n_elts - 1 && n_elts <= 16)
     {
       int matches[16][2] = {0};
       for (int i = 0; i < n_elts; i++)
@@ -22209,12 +22209,23 @@ aarch64_expand_vector_init_fallback (rtx target, rtx vals)
 	}
       int maxelement = 0;
       int maxv = 0;
+      rtx const_elem = NULL_RTX;
+      int const_elem_pos = 0;
+
       for (int i = 0; i < n_elts; i++)
-	if (matches[i][1] > maxv)
-	  {
-	    maxelement = i;
-	    maxv = matches[i][1];
-	  }
+	{
+	  if (matches[i][1] > maxv)
+	    {
+	      maxelement = i;
+	      maxv = matches[i][1];
+	    }
+	  if (CONST_INT_P (XVECEXP (vals, 0, i))
+	      || CONST_DOUBLE_P (XVECEXP (vals, 0, i)))
+	    {
+	      const_elem_pos = i;
+	      const_elem = XVECEXP (vals, 0, i);
+	    }
+	}
 
       /* Create a duplicate of the most common element, unless all elements
 	 are equally useless to us, in which case just immediately set the
@@ -22252,8 +22263,19 @@ aarch64_expand_vector_init_fallback (rtx target, rtx vals)
 	     vector register.  For big-endian we want that position to hold
 	     the last element of VALS.  */
 	  maxelement = BYTES_BIG_ENDIAN ? n_elts - 1 : 0;
-	  rtx x = force_reg (inner_mode, XVECEXP (vals, 0, maxelement));
-	  aarch64_emit_move (target, lowpart_subreg (mode, x, inner_mode));
+
+	  /* If we have a single constant element, use that for duplicating
+	     instead.  */
+	  if (const_elem)
+	    {
+	      maxelement = const_elem_pos;
+	      aarch64_emit_move (target, gen_vec_duplicate (mode, const_elem));
+	    }
+	  else
+	    {
+	      rtx x = force_reg (inner_mode, XVECEXP (vals, 0, maxelement));
+	      aarch64_emit_move (target, lowpart_subreg (mode, x, inner_mode));
+	    }
 	}
       else
 	{
diff --git a/gcc/testsuite/gcc.target/aarch64/vec-init-single-const-2.c b/gcc/testsuite/gcc.target/aarch64/vec-init-single-const-2.c
new file mode 100644
index 00000000000..f4dcab429c1
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/vec-init-single-const-2.c
@@ -0,0 +1,30 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+#include <arm_neon.h>
+
+/* In case where there are no duplicate elements in vector initializer,
+   check that the constant is used for duplication.  */
+
+int8x16_t f_s8(int8_t a0, int8_t a1, int8_t a2, int8_t a3, int8_t a4,
+                int8_t a5, int8_t a6, int8_t a7, int8_t a8, int8_t a9,
+                int8_t a10, int8_t a11, int8_t a12, int8_t a13, int8_t a14)
+{
+  return (int8x16_t) { a0, a1, a2, a3, a4, a5, a6, a7,
+                       a8, a9, a10, a11, a12, a13, a14, 1 };
+}
+
+int16x8_t f_s16(int16_t a0, int16_t a1, int16_t a2, int16_t a3, int16_t a4,
+		int16_t a5, int16_t a6)
+{
+  return (int16x8_t) { a0, a1, a2, a3, a4, a5, a6, 1 };
+}
+
+int32x4_t f_s32(int32_t a0, int32_t a1, int32_t a2)
+{
+  return (int32x4_t) { a0, a1, a2, 1 };
+}
+
+/* { dg-final { scan-assembler {\tmovi\tv[0-9]+\.8b, 0x1} } } */ 
+/* { dg-final { scan-assembler {\tmovi\tv[0-9]+\.4h, 0x1} } } */ 
+/* { dg-final { scan-assembler {\tmovi\tv[0-9]+\.2s, 0x1} } } */ 
diff --git a/gcc/testsuite/gcc.target/aarch64/vec-init-single-const-be.c b/gcc/testsuite/gcc.target/aarch64/vec-init-single-const-be.c
new file mode 100644
index 00000000000..3140e007b5d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/vec-init-single-const-be.c
@@ -0,0 +1,58 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-schedule-insns -fno-schedule-insns2" } */
+/* { dg-final { check-function-bodies "**" "" "" { target { be } } } } */
+
+#include <arm_neon.h>
+
+/*
+** f_s8:
+**	dup	v0.16b, w0
+**	movi	(v[0-9]+)\.8b, 0x1
+**	ins	v0.b\[0\], \1\.b\[0\]
+**	ret
+*/
+
+int8x16_t f_s8(int8_t x)
+{
+  return (int8x16_t) { x, x, x, x, x, x, x, x,
+                       x, x, x, x, x, x, x, 1 };
+}
+
+/*
+** f_s16:
+**	dup	v0.8h, w0
+**	movi	(v[0-9]+)\.4h, 0x1
+**	ins	v0.h\[0\], \1\.h\[0\]
+**	ret
+*/
+
+int16x8_t f_s16(int16_t x)
+{
+  return (int16x8_t) { x, x, x, x, x, x, x, 1 };
+}
+
+/*
+** f_s32:
+**	dup	v0.4s, w0
+**	movi	(v[0-9]+)\.2s, 0x1
+**	ins	v0.s\[0\], \1\.s\[0\]
+**	ret
+*/
+
+int32x4_t f_s32(int32_t x)
+{
+  return (int32x4_t) { x, x, x, 1 };
+}
+
+/*
+** f_s64:
+**	adrp	x[0-9]+, .LC[0-9]+
+**	ldr	q0, \[x[0-9]+, #:lo12:.LC[0-9]+\]
+**	ins	v0\.d\[1\], x0
+**	ret
+*/
+
+int64x2_t f_s64(int64_t x)
+{
+  return (int64x2_t) { x, 1 };
+}
diff --git a/gcc/testsuite/gcc.target/aarch64/vec-init-single-const.c b/gcc/testsuite/gcc.target/aarch64/vec-init-single-const.c
new file mode 100644
index 00000000000..274b0b39ac4
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/vec-init-single-const.c
@@ -0,0 +1,58 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-schedule-insns -fno-schedule-insns2" } */
+/* { dg-final { check-function-bodies "**" "" "" { target { le } } } } */
+
+#include <arm_neon.h>
+
+/*
+** f_s8:
+**	dup	v0.16b, w0
+**	movi	(v[0-9]+)\.8b, 0x1
+**	ins	v0.b\[15\], \1\.b\[0\]
+**	ret
+*/
+
+int8x16_t f_s8(int8_t x)
+{
+  return (int8x16_t) { x, x, x, x, x, x, x, x,
+                       x, x, x, x, x, x, x, 1 };
+}
+
+/*
+** f_s16:
+**	dup	v0.8h, w0
+**	movi	(v[0-9]+)\.4h, 0x1
+**	ins	v0.h\[7\], \1\.h\[0\]
+**	ret
+*/
+
+int16x8_t f_s16(int16_t x)
+{
+  return (int16x8_t) { x, x, x, x, x, x, x, 1 };
+}
+
+/*
+** f_s32:
+**	dup	v0.4s, w0
+**	movi	(v[0-9]+)\.2s, 0x1
+**	ins	v0.s\[3\], \1\.s\[0\]
+**	ret
+*/
+
+int32x4_t f_s32(int32_t x)
+{
+  return (int32x4_t) { x, x, x, 1 };
+}
+
+/*
+** f_s64:
+**	adrp	x[0-9]+, .LC[0-9]+
+**	ldr	q0, \[x[0-9]+, #:lo12:.LC[0-9]+\]
+**	ins	v0\.d\[0\], x0
+**	ret
+*/
+
+int64x2_t f_s64(int64_t x)
+{
+  return (int64x2_t) { x, 1 };
+}