AArch64: Add optimized Q-register memcpy

Add a new memcpy using 128-bit Q registers - this is faster on modern
cores and reduces codesize.  Similar to the generic memcpy, small cases
include copies up to 32 bytes.  64-128 byte copies are split into two
cases to improve performance of 64-96 byte copies.  Large copies align
the source rather than the destination.

bench-memcpy-random is ~9% faster than memcpy_falkor on Neoverse N1,
so make this memcpy the default on N1 (on Centriq it is 15% faster than
memcpy_falkor).

Passes GLIBC regression tests.

Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
(cherry picked from commit 4a733bf375)

sysdeps/aarch64/multiarch/Makefile

@@ -1,5 +1,5 @@
 ifeq ($(subdir),string)
-sysdep_routines += memcpy_generic memcpy_thunderx memcpy_thunderx2 \
+sysdep_routines += memcpy_generic memcpy_advsimd memcpy_thunderx memcpy_thunderx2 \
 		   memcpy_falkor memmove_falkor \
 		   memset_generic memset_falkor memset_emag \
 		   memchr_generic memchr_nosimd \

sysdeps/aarch64/multiarch/ifunc-impl-list.c

@@ -42,11 +42,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_thunderx)
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_thunderx2)
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_falkor)
+	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_simd)
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_generic))
   IFUNC_IMPL (i, name, memmove,
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_thunderx)
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_thunderx2)
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_falkor)
+	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_simd)
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_generic))
   IFUNC_IMPL (i, name, memset,
 	      /* Enable this on non-falkor processors too so that other cores

sysdeps/aarch64/multiarch/memcpy.c

@@ -29,6 +29,7 @@
 extern __typeof (__redirect_memcpy) __libc_memcpy;
 
 extern __typeof (__redirect_memcpy) __memcpy_generic attribute_hidden;
+extern __typeof (__redirect_memcpy) __memcpy_simd attribute_hidden;
 extern __typeof (__redirect_memcpy) __memcpy_thunderx attribute_hidden;
 extern __typeof (__redirect_memcpy) __memcpy_thunderx2 attribute_hidden;
 extern __typeof (__redirect_memcpy) __memcpy_falkor attribute_hidden;
@@ -36,11 +37,11 @@ extern __typeof (__redirect_memcpy) __memcpy_falkor attribute_hidden;
 libc_ifunc (__libc_memcpy,
             (IS_THUNDERX (midr)
 	     ? __memcpy_thunderx
-	     : (IS_FALKOR (midr) || IS_PHECDA (midr) || IS_ARES (midr)
+	     : (IS_FALKOR (midr) || IS_PHECDA (midr)
 		? __memcpy_falkor
 		: (IS_THUNDERX2 (midr) || IS_THUNDERX2PA (midr)
 		  ? __memcpy_thunderx2
-		  : __memcpy_generic))));
+		  : (IS_ARES (midr) ? __memcpy_simd : __memcpy_generic)))));
 
 # undef memcpy
 strong_alias (__libc_memcpy, memcpy);

sysdeps/aarch64/multiarch/memcpy_advsimd.S

@@ -0,0 +1,247 @@
+/* Generic optimized memcpy using SIMD.
+   Copyright (C) 2020 Free Software Foundation, Inc.
+
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library.  If not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64, Advanced SIMD, unaligned accesses.
+ *
+ */
+
+#define dstin	x0
+#define src	x1
+#define count	x2
+#define dst	x3
+#define srcend	x4
+#define dstend	x5
+#define A_l	x6
+#define A_lw	w6
+#define A_h	x7
+#define B_l	x8
+#define B_lw	w8
+#define B_h	x9
+#define C_lw	w10
+#define tmp1	x14
+
+#define A_q	q0
+#define B_q	q1
+#define C_q	q2
+#define D_q	q3
+#define E_q	q4
+#define F_q	q5
+#define G_q	q6
+#define H_q	q7
+
+
+/* This implementation supports both memcpy and memmove and shares most code.
+   It uses unaligned accesses and branchless sequences to keep the code small,
+   simple and improve performance.
+
+   Copies are split into 3 main cases: small copies of up to 32 bytes, medium
+   copies of up to 128 bytes, and large copies.  The overhead of the overlap
+   check in memmove is negligible since it is only required for large copies.
+
+   Large copies use a software pipelined loop processing 64 bytes per
+   iteration.  The destination pointer is 16-byte aligned to minimize
+   unaligned accesses.  The loop tail is handled by always copying 64 bytes
+   from the end.  */
+
+ENTRY (__memcpy_simd)
+	DELOUSE (0)
+	DELOUSE (1)
+	DELOUSE (2)
+
+	add	srcend, src, count
+	add	dstend, dstin, count
+	cmp	count, 128
+	b.hi	L(copy_long)
+	cmp	count, 32
+	b.hi	L(copy32_128)
+
+	/* Small copies: 0..32 bytes.  */
+	cmp	count, 16
+	b.lo	L(copy16)
+	ldr	A_q, [src]
+	ldr	B_q, [srcend, -16]
+	str	A_q, [dstin]
+	str	B_q, [dstend, -16]
+	ret
+
+	/* Copy 8-15 bytes.  */
+L(copy16):
+	tbz	count, 3, L(copy8)
+	ldr	A_l, [src]
+	ldr	A_h, [srcend, -8]
+	str	A_l, [dstin]
+	str	A_h, [dstend, -8]
+	ret
+
+	/* Copy 4-7 bytes.  */
+L(copy8):
+	tbz	count, 2, L(copy4)
+	ldr	A_lw, [src]
+	ldr	B_lw, [srcend, -4]
+	str	A_lw, [dstin]
+	str	B_lw, [dstend, -4]
+	ret
+
+	/* Copy 0..3 bytes using a branchless sequence.  */
+L(copy4):
+	cbz	count, L(copy0)
+	lsr	tmp1, count, 1
+	ldrb	A_lw, [src]
+	ldrb	C_lw, [srcend, -1]
+	ldrb	B_lw, [src, tmp1]
+	strb	A_lw, [dstin]
+	strb	B_lw, [dstin, tmp1]
+	strb	C_lw, [dstend, -1]
+L(copy0):
+	ret
+
+	.p2align 4
+	/* Medium copies: 33..128 bytes.  */
+L(copy32_128):
+	ldp	A_q, B_q, [src]
+	ldp	C_q, D_q, [srcend, -32]
+	cmp	count, 64
+	b.hi	L(copy128)
+	stp	A_q, B_q, [dstin]
+	stp	C_q, D_q, [dstend, -32]
+	ret
+
+	.p2align 4
+	/* Copy 65..128 bytes.  */
+L(copy128):
+	ldp	E_q, F_q, [src, 32]
+	cmp	count, 96
+	b.ls	L(copy96)
+	ldp	G_q, H_q, [srcend, -64]
+	stp	G_q, H_q, [dstend, -64]
+L(copy96):
+	stp	A_q, B_q, [dstin]
+	stp	E_q, F_q, [dstin, 32]
+	stp	C_q, D_q, [dstend, -32]
+	ret
+
+	/* Align loop64 below to 16 bytes.  */
+	nop
+
+	/* Copy more than 128 bytes.  */
+L(copy_long):
+	/* Copy 16 bytes and then align src to 16-byte alignment.  */
+	ldr	D_q, [src]
+	and	tmp1, src, 15
+	bic	src, src, 15
+	sub	dst, dstin, tmp1
+	add	count, count, tmp1	/* Count is now 16 too large.  */
+	ldp	A_q, B_q, [src, 16]
+	str	D_q, [dstin]
+	ldp	C_q, D_q, [src, 48]
+	subs	count, count, 128 + 16	/* Test and readjust count.  */
+	b.ls	L(copy64_from_end)
+L(loop64):
+	stp	A_q, B_q, [dst, 16]
+	ldp	A_q, B_q, [src, 80]
+	stp	C_q, D_q, [dst, 48]
+	ldp	C_q, D_q, [src, 112]
+	add	src, src, 64
+	add	dst, dst, 64
+	subs	count, count, 64
+	b.hi	L(loop64)
+
+	/* Write the last iteration and copy 64 bytes from the end.  */
+L(copy64_from_end):
+	ldp	E_q, F_q, [srcend, -64]
+	stp	A_q, B_q, [dst, 16]
+	ldp	A_q, B_q, [srcend, -32]
+	stp	C_q, D_q, [dst, 48]
+	stp	E_q, F_q, [dstend, -64]
+	stp	A_q, B_q, [dstend, -32]
+	ret
+
+END (__memcpy_simd)
+libc_hidden_builtin_def (__memcpy_simd)
+
+
+ENTRY (__memmove_simd)
+	DELOUSE (0)
+	DELOUSE (1)
+	DELOUSE (2)
+
+	add	srcend, src, count
+	add	dstend, dstin, count
+	cmp	count, 128
+	b.hi	L(move_long)
+	cmp	count, 32
+	b.hi	L(copy32_128)
+
+	/* Small moves: 0..32 bytes.  */
+	cmp	count, 16
+	b.lo	L(copy16)
+	ldr	A_q, [src]
+	ldr	B_q, [srcend, -16]
+	str	A_q, [dstin]
+	str	B_q, [dstend, -16]
+	ret
+
+L(move_long):
+	/* Only use backward copy if there is an overlap.  */
+	sub	tmp1, dstin, src
+	cbz	tmp1, L(move0)
+	cmp	tmp1, count
+	b.hs	L(copy_long)
+
+	/* Large backwards copy for overlapping copies.
+	   Copy 16 bytes and then align srcend to 16-byte alignment.  */
+L(copy_long_backwards):
+	ldr	D_q, [srcend, -16]
+	and	tmp1, srcend, 15
+	bic	srcend, srcend, 15
+	sub	count, count, tmp1
+	ldp	A_q, B_q, [srcend, -32]
+	str	D_q, [dstend, -16]
+	ldp	C_q, D_q, [srcend, -64]
+	sub	dstend, dstend, tmp1
+	subs	count, count, 128
+	b.ls	L(copy64_from_start)
+
+L(loop64_backwards):
+	stp	A_q, B_q, [dstend, -32]
+	ldp	A_q, B_q, [srcend, -96]
+	stp	C_q, D_q, [dstend, -64]
+	ldp	C_q, D_q, [srcend, -128]
+	sub	srcend, srcend, 64
+	sub	dstend, dstend, 64
+	subs	count, count, 64
+	b.hi	L(loop64_backwards)
+
+	/* Write the last iteration and copy 64 bytes from the start.  */
+L(copy64_from_start):
+	ldp	E_q, F_q, [src, 32]
+	stp	A_q, B_q, [dstend, -32]
+	ldp	A_q, B_q, [src]
+	stp	C_q, D_q, [dstend, -64]
+	stp	E_q, F_q, [dstin, 32]
+	stp	A_q, B_q, [dstin]
+L(move0):
+	ret
+
+END (__memmove_simd)
+libc_hidden_builtin_def (__memmove_simd)

sysdeps/aarch64/multiarch/memmove.c

@@ -29,6 +29,7 @@
 extern __typeof (__redirect_memmove) __libc_memmove;
 
 extern __typeof (__redirect_memmove) __memmove_generic attribute_hidden;
+extern __typeof (__redirect_memmove) __memmove_simd attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_thunderx attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_thunderx2 attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_falkor attribute_hidden;
@@ -40,7 +41,7 @@ libc_ifunc (__libc_memmove,
 		? __memmove_falkor
 		: (IS_THUNDERX2 (midr) || IS_THUNDERX2PA (midr)
 		  ? __memmove_thunderx2
-		  : __memmove_generic))));
+		  : (IS_ARES (midr) ? __memmove_simd : __memmove_generic)))));
 
 # undef memmove
 strong_alias (__libc_memmove, memmove);