arm64: lib: Implement optimized memset routine

This patch, based on Linaro's Cortex Strings library, improves
the performance of the assembly optimized memset() function.

Signed-off-by: Zhichang Yuan <zhichang.yuan@linaro.org>
Signed-off-by: Deepak Saxena <dsaxena@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>

arch/arm64/lib/memset.S

@@ -1,5 +1,13 @@
 /*
  * Copyright (C) 2013 ARM Ltd.
+ * Copyright (C) 2013 Linaro.
+ *
+ * This code is based on glibc cortex strings work originally authored by Linaro
+ * and re-licensed under GPLv2 for the Linux kernel. The original code can
+ * be found @
+ *
+ * http://bazaar.launchpad.net/~linaro-toolchain-dev/cortex-strings/trunk/
+ * files/head:/src/aarch64/
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 as
@@ -16,6 +24,7 @@
 
 #include <linux/linkage.h>
 #include <asm/assembler.h>
+#include <asm/cache.h>
 
 /*
  * Fill in the buffer with character c (alignment handled by the hardware)
@@ -27,27 +36,181 @@
  * Returns:
  *	x0 - buf
  */
+
+dstin		.req	x0
+val		.req	w1
+count		.req	x2
+tmp1		.req	x3
+tmp1w		.req	w3
+tmp2		.req	x4
+tmp2w		.req	w4
+zva_len_x	.req	x5
+zva_len		.req	w5
+zva_bits_x	.req	x6
+
+A_l		.req	x7
+A_lw		.req	w7
+dst		.req	x8
+tmp3w		.req	w9
+tmp3		.req	x9
+
 ENTRY(memset)
-	mov	x4, x0
-	and	w1, w1, #0xff
-	orr	w1, w1, w1, lsl #8
-	orr	w1, w1, w1, lsl #16
-	orr	x1, x1, x1, lsl #32
-	subs	x2, x2, #8
-	b.mi	2f
-1:	str	x1, [x4], #8
-	subs	x2, x2, #8
-	b.pl	1b
-2:	adds	x2, x2, #4
-	b.mi	3f
-	sub	x2, x2, #4
-	str	w1, [x4], #4
-3:	adds	x2, x2, #2
-	b.mi	4f
-	sub	x2, x2, #2
-	strh	w1, [x4], #2
-4:	adds	x2, x2, #1
-	b.mi	5f
-	strb	w1, [x4]
-5:	ret
+	mov	dst, dstin	/* Preserve return value.  */
+	and	A_lw, val, #255
+	orr	A_lw, A_lw, A_lw, lsl #8
+	orr	A_lw, A_lw, A_lw, lsl #16
+	orr	A_l, A_l, A_l, lsl #32
+
+	cmp	count, #15
+	b.hi	.Lover16_proc
+	/*All store maybe are non-aligned..*/
+	tbz	count, #3, 1f
+	str	A_l, [dst], #8
+1:
+	tbz	count, #2, 2f
+	str	A_lw, [dst], #4
+2:
+	tbz	count, #1, 3f
+	strh	A_lw, [dst], #2
+3:
+	tbz	count, #0, 4f
+	strb	A_lw, [dst]
+4:
+	ret
+
+.Lover16_proc:
+	/*Whether  the start address is aligned with 16.*/
+	neg	tmp2, dst
+	ands	tmp2, tmp2, #15
+	b.eq	.Laligned
+/*
+* The count is not less than 16, we can use stp to store the start 16 bytes,
+* then adjust the dst aligned with 16.This process will make the current
+* memory address at alignment boundary.
+*/
+	stp	A_l, A_l, [dst] /*non-aligned store..*/
+	/*make the dst aligned..*/
+	sub	count, count, tmp2
+	add	dst, dst, tmp2
+
+.Laligned:
+	cbz	A_l, .Lzero_mem
+
+.Ltail_maybe_long:
+	cmp	count, #64
+	b.ge	.Lnot_short
+.Ltail63:
+	ands	tmp1, count, #0x30
+	b.eq	3f
+	cmp	tmp1w, #0x20
+	b.eq	1f
+	b.lt	2f
+	stp	A_l, A_l, [dst], #16
+1:
+	stp	A_l, A_l, [dst], #16
+2:
+	stp	A_l, A_l, [dst], #16
+/*
+* The last store length is less than 16,use stp to write last 16 bytes.
+* It will lead some bytes written twice and the access is non-aligned.
+*/
+3:
+	ands	count, count, #15
+	cbz	count, 4f
+	add	dst, dst, count
+	stp	A_l, A_l, [dst, #-16]	/* Repeat some/all of last store. */
+4:
+	ret
+
+	/*
+	* Critical loop. Start at a new cache line boundary. Assuming
+	* 64 bytes per line, this ensures the entire loop is in one line.
+	*/
+	.p2align	L1_CACHE_SHIFT
+.Lnot_short:
+	sub	dst, dst, #16/* Pre-bias.  */
+	sub	count, count, #64
+1:
+	stp	A_l, A_l, [dst, #16]
+	stp	A_l, A_l, [dst, #32]
+	stp	A_l, A_l, [dst, #48]
+	stp	A_l, A_l, [dst, #64]!
+	subs	count, count, #64
+	b.ge	1b
+	tst	count, #0x3f
+	add	dst, dst, #16
+	b.ne	.Ltail63
+.Lexitfunc:
+	ret
+
+	/*
+	* For zeroing memory, check to see if we can use the ZVA feature to
+	* zero entire 'cache' lines.
+	*/
+.Lzero_mem:
+	cmp	count, #63
+	b.le	.Ltail63
+	/*
+	* For zeroing small amounts of memory, it's not worth setting up
+	* the line-clear code.
+	*/
+	cmp	count, #128
+	b.lt	.Lnot_short /*count is at least  128 bytes*/
+
+	mrs	tmp1, dczid_el0
+	tbnz	tmp1, #4, .Lnot_short
+	mov	tmp3w, #4
+	and	zva_len, tmp1w, #15	/* Safety: other bits reserved.  */
+	lsl	zva_len, tmp3w, zva_len
+
+	ands	tmp3w, zva_len, #63
+	/*
+	* ensure the zva_len is not less than 64.
+	* It is not meaningful to use ZVA if the block size is less than 64.
+	*/
+	b.ne	.Lnot_short
+.Lzero_by_line:
+	/*
+	* Compute how far we need to go to become suitably aligned. We're
+	* already at quad-word alignment.
+	*/
+	cmp	count, zva_len_x
+	b.lt	.Lnot_short		/* Not enough to reach alignment.  */
+	sub	zva_bits_x, zva_len_x, #1
+	neg	tmp2, dst
+	ands	tmp2, tmp2, zva_bits_x
+	b.eq	2f			/* Already aligned.  */
+	/* Not aligned, check that there's enough to copy after alignment.*/
+	sub	tmp1, count, tmp2
+	/*
+	* grantee the remain length to be ZVA is bigger than 64,
+	* avoid to make the 2f's process over mem range.*/
+	cmp	tmp1, #64
+	ccmp	tmp1, zva_len_x, #8, ge	/* NZCV=0b1000 */
+	b.lt	.Lnot_short
+	/*
+	* We know that there's at least 64 bytes to zero and that it's safe
+	* to overrun by 64 bytes.
+	*/
+	mov	count, tmp1
+1:
+	stp	A_l, A_l, [dst]
+	stp	A_l, A_l, [dst, #16]
+	stp	A_l, A_l, [dst, #32]
+	subs	tmp2, tmp2, #64
+	stp	A_l, A_l, [dst, #48]
+	add	dst, dst, #64
+	b.ge	1b
+	/* We've overrun a bit, so adjust dst downwards.*/
+	add	dst, dst, tmp2
+2:
+	sub	count, count, zva_len_x
+3:
+	dc	zva, dst
+	add	dst, dst, zva_len_x
+	subs	count, count, zva_len_x
+	b.ge	3b
+	ands	count, count, zva_bits_x
+	b.ne	.Ltail_maybe_long
+	ret
 ENDPROC(memset)