[AArch64] Optimized memcpy.

This is an optimized memcpy for AArch64.  Copies are split into 3 main
cases: small copies of up to 16 bytes, medium copies of 17..96 bytes
which are fully unrolled.  Large copies of more than 96 bytes align
the destination and use an unrolled loop processing 64 bytes per
iteration.  In order to share code with memmove, small and medium
copies read all data before writing, allowing any kind of overlap.  On
a random copy test memcpy is 40.8% faster on A57 and 28.4% on A53.

newlib/ChangeLog

@@ -1,3 +1,8 @@
+2015-07-13  Wilco Dijkstra  <wdijkstr@arm.com>
+
+	* newlib/libc/machine/aarch64/memcpy.S (memcpy):
+	Rewrite of optimized memcpy.
+
 2015-07-13  Wilco Dijkstra  <wdijkstr@arm.com>
 
 	* newlib/libc/machine/aarch64/memove.S (memmove):

newlib/libc/machine/aarch64/memcpy.S

@@ -24,10 +24,37 @@
    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */
 
+/*
+ * Copyright (c) 2015 ARM Ltd
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. The name of the company may not be used to endorse or promote
+ *    products derived from this software without specific prior written
+ *    permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+ * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+ * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
 /* Assumptions:
  *
- * ARMv8-a, AArch64
+ * ARMv8-a, AArch64, unaligned accesses.
- * Unaligned accesses
  *
  */
 
@@ -38,22 +65,26 @@
 #define dstin	x0
 #define src	x1
 #define count	x2
-#define tmp1	x3
+#define dst	x3
-#define tmp1w	w3
+#define srcend	x4
-#define tmp2	x4
+#define dstend	x5
-#define tmp2w	w4
+#define A_l	x6
-#define tmp3	x5
+#define A_lw	w6
-#define tmp3w	w5
+#define A_h	x7
-#define dst	x6
+#define A_hw	w7
+#define B_l	x8
+#define B_h	x9
+#define C_l	x10
+#define C_h	x11
+#define D_l	x12
+#define D_h	x13
+#define E_l	src
+#define E_h	count
+#define F_l	srcend
+#define F_h	dst
+#define tmp1	x9
 
-#define A_l	x7
+#define L(l) .L ## l
-#define A_h	x8
-#define B_l	x9
-#define B_h	x10
-#define C_l	x11
-#define C_h	x12
-#define D_l	x13
-#define D_h	x14
 
 	.macro def_fn f p2align=0
 	.text
@@ -63,133 +94,129 @@
 \f:
 	.endm
 
+/* Copies are split into 3 main cases: small copies of up to 16 bytes,
+   medium copies of 17..96 bytes which are fully unrolled. Large copies
+   of more than 96 bytes align the destination and use an unrolled loop
+   processing 64 bytes per iteration.
+   Small and medium copies read all data before writing, allowing any
+   kind of overlap, and memmove tailcalls memcpy for these cases as
+   well as non-overlapping copies.
+*/
+
 def_fn memcpy p2align=6
+	add	srcend, src, count
+	add	dstend, dstin, count
+	cmp	count, 96
+	b.hi	L(copy_long)
+	cmp	count, 16
+	b.hs	L(copy_medium)
 
-	mov	dst, dstin
+	/* Small copies: 0..16 bytes.  */
-	cmp	count, #64
+L(copy16):
-	b.ge	.Lcpy_not_short
+	tbz	count, 3, 1f
-	cmp	count, #15
+	ldr	A_l, [src]
-	b.le	.Ltail15tiny
+	ldr	A_h, [srcend, -8]
-
+	str	A_l, [dstin]
-	/* Deal with small copies quickly by dropping straight into the
+	str	A_h, [dstend, -8]
-	 * exit block.  */
-.Ltail63:
-	/* Copy up to 48 bytes of data.  At this point we only need the
-	 * bottom 6 bits of count to be accurate.  */
-	ands	tmp1, count, #0x30
-	b.eq	.Ltail15
-	add	dst, dst, tmp1
-	add	src, src, tmp1
-	cmp	tmp1w, #0x20
-	b.eq	1f
-	b.lt	2f
-	ldp	A_l, A_h, [src, #-48]
-	stp	A_l, A_h, [dst, #-48]
-1:
-	ldp	A_l, A_h, [src, #-32]
-	stp	A_l, A_h, [dst, #-32]
-2:
-	ldp	A_l, A_h, [src, #-16]
-	stp	A_l, A_h, [dst, #-16]
-
-.Ltail15:
-	ands	count, count, #15
-	beq	1f
-	add	src, src, count
-	ldp	A_l, A_h, [src, #-16]
-	add	dst, dst, count
-	stp	A_l, A_h, [dst, #-16]
-1:
 	ret
-
-.Ltail15tiny:
-	/* Copy up to 15 bytes of data.  Does not assume additional data
-	   being copied.  */
-	tbz	count, #3, 1f
-	ldr	tmp1, [src], #8
-	str	tmp1, [dst], #8
-1:
-	tbz	count, #2, 1f
-	ldr	tmp1w, [src], #4
-	str	tmp1w, [dst], #4
-1:
-	tbz	count, #1, 1f
-	ldrh	tmp1w, [src], #2
-	strh	tmp1w, [dst], #2
-1:
-	tbz	count, #0, 1f
-	ldrb	tmp1w, [src]
-	strb	tmp1w, [dst]
 1:
+	tbz	count, 2, 1f
+	ldr	A_lw, [src]
+	ldr	A_hw, [srcend, -4]
+	str	A_lw, [dstin]
+	str	A_hw, [dstend, -4]
 	ret
+	.p2align 4
+1:
+	cbz	count, 2f
+	ldrb	A_lw, [src]
+	tbz	count, 1, 1f
+	ldrh	A_hw, [srcend, -2]
+	strh	A_hw, [dstend, -2]
+1:	strb	A_lw, [dstin]
+2:	ret
 
-.Lcpy_not_short:
+	.p2align 4
-	/* We don't much care about the alignment of DST, but we want SRC
+	/* Medium copies: 17..96 bytes.	 */
-	 * to be 128-bit (16 byte) aligned so that we don't cross cache line
+L(copy_medium):
-	 * boundaries on both loads and stores.  */
-	neg	tmp2, src
-	ands	tmp2, tmp2, #15		/* Bytes to reach alignment.  */
-	b.eq	2f
-	sub	count, count, tmp2
-	/* Copy more data than needed; it's faster than jumping
-	 * around copying sub-Quadword quantities.  We know that
-	 * it can't overrun.  */
 	ldp	A_l, A_h, [src]
-	add	src, src, tmp2
+	tbnz	count, 6, L(copy96)
-	stp	A_l, A_h, [dst]
+	ldp	D_l, D_h, [srcend, -16]
-	add	dst, dst, tmp2
+	tbz	count, 5, 1f
-	/* There may be less than 63 bytes to go now.  */
+	ldp	B_l, B_h, [src, 16]
-	cmp	count, #63
+	ldp	C_l, C_h, [srcend, -32]
-	b.le	.Ltail63
+	stp	B_l, B_h, [dstin, 16]
-2:
+	stp	C_l, C_h, [dstend, -32]
-	subs	count, count, #128
+1:
-	b.ge	.Lcpy_body_large
+	stp	A_l, A_h, [dstin]
-	/* Less than 128 bytes to copy, so handle 64 here and then jump
+	stp	D_l, D_h, [dstend, -16]
-	 * to the tail.  */
-	ldp	A_l, A_h, [src]
-	ldp	B_l, B_h, [src, #16]
-	ldp	C_l, C_h, [src, #32]
-	ldp	D_l, D_h, [src, #48]
-	stp	A_l, A_h, [dst]
-	stp	B_l, B_h, [dst, #16]
-	stp	C_l, C_h, [dst, #32]
-	stp	D_l, D_h, [dst, #48]
-	tst	count, #0x3f
-	add	src, src, #64
-	add	dst, dst, #64
-	b.ne	.Ltail63
 	ret
 
-	/* Critical loop.  Start at a new cache line boundary.  Assuming
+	.p2align 4
-	 * 64 bytes per line this ensures the entire loop is in one line.  */
+	/* Copy 64..96 bytes.  Copy 64 bytes from the start and
-	.p2align 6
+	   32 bytes from the end.  */
-.Lcpy_body_large:
+L(copy96):
-	/* There are at least 128 bytes to copy.  */
+	ldp	B_l, B_h, [src, 16]
-	ldp	A_l, A_h, [src, #0]
+	ldp	C_l, C_h, [src, 32]
-	sub	dst, dst, #16		/* Pre-bias.  */
+	ldp	D_l, D_h, [src, 48]
-	ldp	B_l, B_h, [src, #16]
+	ldp	E_l, E_h, [srcend, -32]
-	ldp	C_l, C_h, [src, #32]
+	ldp	F_l, F_h, [srcend, -16]
-	ldp	D_l, D_h, [src, #48]!	/* src += 64 - Pre-bias.  */
+	stp	A_l, A_h, [dstin]
-1:
+	stp	B_l, B_h, [dstin, 16]
-	stp	A_l, A_h, [dst, #16]
+	stp	C_l, C_h, [dstin, 32]
-	ldp	A_l, A_h, [src, #16]
+	stp	D_l, D_h, [dstin, 48]
-	stp	B_l, B_h, [dst, #32]
+	stp	E_l, E_h, [dstend, -32]
-	ldp	B_l, B_h, [src, #32]
+	stp	F_l, F_h, [dstend, -16]
-	stp	C_l, C_h, [dst, #48]
-	ldp	C_l, C_h, [src, #48]
-	stp	D_l, D_h, [dst, #64]!
-	ldp	D_l, D_h, [src, #64]!
-	subs	count, count, #64
-	b.ge	1b
-	stp	A_l, A_h, [dst, #16]
-	stp	B_l, B_h, [dst, #32]
-	stp	C_l, C_h, [dst, #48]
-	stp	D_l, D_h, [dst, #64]
-	add	src, src, #16
-	add	dst, dst, #64 + 16
-	tst	count, #0x3f
-	b.ne	.Ltail63
 	ret
+
+	/* Align DST to 16 byte alignment so that we don't cross cache line
+	   boundaries on both loads and stores.	 There are at least 96 bytes
+	   to copy, so copy 16 bytes unaligned and then align.	The loop
+	   copies 64 bytes per iteration and prefetches one iteration ahead.  */
+
+	.p2align 4
+L(copy_long):
+	and	tmp1, dstin, 15
+	bic	dst, dstin, 15
+	ldp	D_l, D_h, [src]
+	sub	src, src, tmp1
+	add	count, count, tmp1	/* Count is now 16 too large.  */
+	ldp	A_l, A_h, [src, 16]
+	stp	D_l, D_h, [dstin]
+	ldp	B_l, B_h, [src, 32]
+	ldp	C_l, C_h, [src, 48]
+	ldp	D_l, D_h, [src, 64]!
+	subs	count, count, 128 + 16	/* Test and readjust count.  */
+	b.ls	2f
+1:
+	stp	A_l, A_h, [dst, 16]
+	ldp	A_l, A_h, [src, 16]
+	stp	B_l, B_h, [dst, 32]
+	ldp	B_l, B_h, [src, 32]
+	stp	C_l, C_h, [dst, 48]
+	ldp	C_l, C_h, [src, 48]
+	stp	D_l, D_h, [dst, 64]!
+	ldp	D_l, D_h, [src, 64]!
+	subs	count, count, 64
+	b.hi	1b
+
+	/* Write the last full set of 64 bytes.	 The remainder is at most 64
+	   bytes, so it is safe to always copy 64 bytes from the end even if
+	   there is just 1 byte left.  */
+2:
+	ldp	E_l, E_h, [srcend, -64]
+	stp	A_l, A_h, [dst, 16]
+	ldp	A_l, A_h, [srcend, -48]
+	stp	B_l, B_h, [dst, 32]
+	ldp	B_l, B_h, [srcend, -32]
+	stp	C_l, C_h, [dst, 48]
+	ldp	C_l, C_h, [srcend, -16]
+	stp	D_l, D_h, [dst, 64]
+	stp	E_l, E_h, [dstend, -64]
+	stp	A_l, A_h, [dstend, -48]
+	stp	B_l, B_h, [dstend, -32]
+	stp	C_l, C_h, [dstend, -16]
+	ret
+
 	.size	memcpy, . - memcpy
-
 #endif