diff --git a/libc/arch-arm64/cortex-a53/bionic/memmove.S b/libc/arch-arm64/cortex-a53/bionic/memmove.S
new file mode 100644
index 0000000..c50112d
--- /dev/null
+++ b/libc/arch-arm64/cortex-a53/bionic/memmove.S
@@ -0,0 +1,153 @@
+/* Copyright (c) 2013, Linaro Limited
+   All rights reserved.
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are met:
+       * Redistributions of source code must retain the above copyright
+         notice, this list of conditions and the following disclaimer.
+       * 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.
+       * Neither the name of the Linaro nor the
+         names of its contributors may be used to endorse or promote products
+         derived from this software without specific prior written permission.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   "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 THE COPYRIGHT
+   HOLDER OR CONTRIBUTORS 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. */
+
+/*
+ * 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, unaligned accesses, wchar_t is 4 bytes
+ */
+
+#include <private/bionic_asm.h>
+
+/* Parameters and result.  */
+#define dstin	x0
+#define src	x1
+#define count	x2
+#define srcend	x3
+#define dstend	x4
+#define tmp1	x5
+#define A_l	x6
+#define A_h	x7
+#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	count
+#define E_h	tmp1
+
+/* All memmoves up to 96 bytes are done by memcpy as it supports overlaps.
+   Larger backwards copies are also handled by memcpy. The only remaining
+   case is forward large copies.  The destination is aligned, and an
+   unrolled loop processes 64 bytes per iteration.
+*/
+
+#if defined(WMEMMOVE)
+ENTRY(wmemmove)
+	lsl	count, count, #2
+#else
+ENTRY(memmove)
+#endif
+	sub	tmp1, dstin, src
+	cmp	count, 96
+	ccmp	tmp1, count, 2, hi
+	b.hs	memcpy
+
+	cbz	tmp1, 3f
+	add	dstend, dstin, count
+	add	srcend, src, count
+
+	/* Align dstend 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.  */
+
+	and	tmp1, dstend, 15
+	ldp	D_l, D_h, [srcend, -16]
+	sub	srcend, srcend, tmp1
+	sub	count, count, tmp1
+	ldp	A_l, A_h, [srcend, -16]
+	stp	D_l, D_h, [dstend, -16]
+	ldp	B_l, B_h, [srcend, -32]
+	ldp	C_l, C_h, [srcend, -48]
+	ldp	D_l, D_h, [srcend, -64]!
+	sub	dstend, dstend, tmp1
+	subs	count, count, 128
+	b.ls	2f
+	nop
+1:
+	stp	A_l, A_h, [dstend, -16]
+	ldp	A_l, A_h, [srcend, -16]
+	stp	B_l, B_h, [dstend, -32]
+	ldp	B_l, B_h, [srcend, -32]
+	stp	C_l, C_h, [dstend, -48]
+	ldp	C_l, C_h, [srcend, -48]
+	stp	D_l, D_h, [dstend, -64]!
+	ldp	D_l, D_h, [srcend, -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 start even if
+	   there is just 1 byte left.  */
+2:
+	ldp	E_l, E_h, [src, 48]
+	stp	A_l, A_h, [dstend, -16]
+	ldp	A_l, A_h, [src, 32]
+	stp	B_l, B_h, [dstend, -32]
+	ldp	B_l, B_h, [src, 16]
+	stp	C_l, C_h, [dstend, -48]
+	ldp	C_l, C_h, [src]
+	stp	D_l, D_h, [dstend, -64]
+	stp	E_l, E_h, [dstin, 48]
+	stp	A_l, A_h, [dstin, 32]
+	stp	B_l, B_h, [dstin, 16]
+	stp	C_l, C_h, [dstin]
+3:	ret
+
+#if defined(WMEMMOVE)
+END(wmemmove)
+#else
+END(memmove)
+#endif
diff --git a/libc/arch-arm64/generic/bionic/memcpy_base.S b/libc/arch-arm64/generic/bionic/memcpy_base.S
index c5d42ce..f850624 100644
--- a/libc/arch-arm64/generic/bionic/memcpy_base.S
+++ b/libc/arch-arm64/generic/bionic/memcpy_base.S
@@ -1,4 +1,4 @@
-/* Copyright (c) 2012, Linaro Limited
+/* Copyright (c) 2012-2013, Linaro Limited
    All rights reserved.
 
    Redistribution and use in source and binary forms, with or without
@@ -22,158 +22,196 @@
    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.
-*/
+   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
- * Unaligned accesses
+ * ARMv8-a, AArch64, unaligned accesses.
  *
  */
 
+#include <private/bionic_asm.h>
+
 #define dstin	x0
 #define src	x1
 #define count	x2
-#define tmp1	x3
-#define tmp1w	w3
-#define tmp2	x4
-#define tmp2w	w4
-#define tmp3	x5
-#define tmp3w	w5
-#define dst	x6
+#define dst	x3
+#define srcend	x4
+#define dstend	x5
+#define A_l	x6
+#define A_lw	w6
+#define A_h	x7
+#define A_hw	w7
+#define B_l	x8
+#define B_lw   w8
+#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 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
+#define L(l) .L ## l
 
-	mov	dst, dstin
-	cmp	count, #64
-	b.ge	.Lcpy_not_short
-	cmp	count, #15
-	b.le	.Ltail15tiny
+/* 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.
+*/
 
-	/* Deal with small copies quickly by dropping straight into the
-	 * 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]
+	prfm    PLDL1KEEP, [src]
+	add	srcend, src, count
+	add	dstend, dstin, count
+        cmp     count, 16
+        b.ls    L(copy16)
+	cmp	count, 96
+	b.hi	L(copy_long)
 
-.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:
-	ret
-
-.Lcpy_not_short:
-	/* We don't much care about the alignment of DST, but we want SRC
-	 * to be 128-bit (16 byte) aligned so that we don't cross cache line
-	 * 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.  */
+	/* Medium copies: 17..96 bytes.  */
+	sub	tmp1, count, 1
 	ldp	A_l, A_h, [src]
-	add	src, src, tmp2
-	stp	A_l, A_h, [dst]
-	add	dst, dst, tmp2
-	/* There may be less than 63 bytes to go now.  */
-	cmp	count, #63
-	b.le	.Ltail63
-2:
-	subs	count, count, #128
-	b.ge	.Lcpy_body_large
-	/* Less than 128 bytes to copy, so handle 64 here and then jump
-	 * 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
+	tbnz	tmp1, 6, L(copy96)
+	ldp	D_l, D_h, [srcend, -16]
+	tbz	tmp1, 5, 1f
+	ldp	B_l, B_h, [src, 16]
+	ldp	C_l, C_h, [srcend, -32]
+	stp	B_l, B_h, [dstin, 16]
+	stp	C_l, C_h, [dstend, -32]
+1:
+	stp	A_l, A_h, [dstin]
+	stp	D_l, D_h, [dstend, -16]
 	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 6
-.Lcpy_body_large:
-	/* There are at least 128 bytes to copy.  */
-	ldp	A_l, A_h, [src, #0]
-	sub	dst, dst, #16		/* Pre-bias.  */
-	ldp	B_l, B_h, [src, #16]
-	ldp	C_l, C_h, [src, #32]
-	ldp	D_l, D_h, [src, #48]!	/* src += 64 - Pre-bias.  */
+	.p2align 4
+
+	/* Small copies: 0..16 bytes.  */
+L(copy16):
+	cmp	count, 8
+	b.lo	1f
+	ldr	A_l, [src]
+	ldr	A_h, [srcend, -8]
+	str	A_l, [dstin]
+	str	A_h, [dstend, -8]
+	ret
+	.p2align 4
 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.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
+	tbz	count, 2, 1f
+	ldr	A_lw, [src]
+	ldr	A_hw, [srcend, -4]
+	str	A_lw, [dstin]
+	str	A_hw, [dstend, -4]
+	ret
+
+	/* Copy 0..3 bytes.  Use a branchless sequence that copies the same
+	   byte 3 times if count==1, or the 2nd byte twice if count==2.  */
+1:
+	cbz	count, 2f
+	lsr	tmp1, count, 1
+	ldrb	A_lw, [src]
+	ldrb	A_hw, [srcend, -1]
+	ldrb	B_lw, [src, tmp1]
+	strb	A_lw, [dstin]
+	strb	B_lw, [dstin, tmp1]
+	strb	A_hw, [dstend, -1]
+2:	ret
+
+	.p2align 4
+	/* Copy 64..96 bytes.  Copy 64 bytes from the start and
+	   32 bytes from the end.  */
+L(copy96):
+	ldp	B_l, B_h, [src, 16]
+	ldp	C_l, C_h, [src, 32]
+	ldp	D_l, D_h, [src, 48]
+	ldp	E_l, E_h, [srcend, -32]
+	ldp	F_l, F_h, [srcend, -16]
+	stp	A_l, A_h, [dstin]
+	stp	B_l, B_h, [dstin, 16]
+	stp	C_l, C_h, [dstin, 32]
+	stp	D_l, D_h, [dstin, 48]
+	stp	E_l, E_h, [dstend, -32]
+	stp	F_l, F_h, [dstend, -16]
+	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
diff --git a/libc/arch-arm64/generic/bionic/memset.S b/libc/arch-arm64/generic/bionic/memset.S
index 9626c0b..3416cfb 100644
--- a/libc/arch-arm64/generic/bionic/memset.S
+++ b/libc/arch-arm64/generic/bionic/memset.S
@@ -1,4 +1,4 @@
-/* Copyright (c) 2012, Linaro Limited
+/* Copyright (c) 2012-2013, Linaro Limited
    All rights reserved.
 
    Redistribution and use in source and binary forms, with or without
@@ -22,13 +22,39 @@
    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.
-*/
+   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
- * Unaligned accesses
+ * ARMv8-a, AArch64, unaligned accesses
  *
  */
 
@@ -45,24 +71,22 @@
    values rather than re-reading them each call.  */
 
 #define dstin		x0
-#define val		w1
+#define val		x1
+#define valw		w1
 #define count		x2
-#define dst_count x3 /* for __memset_chk */
-#define tmp1		x3
-#define tmp1w		w3
-#define tmp2		x4
-#define tmp2w		w4
-#define zva_len_x	x5
-#define zva_len		w5
-#define zva_bits_x	x6
+#define dst 		x3
+#define dstend		x4
+#define tmp1		x5
+#define tmp1w		w5
+#define tmp2		x6
+#define tmp2w		w6
+#define zva_len		x5
+#define zva_lenw	w7
 
-#define A_l		x7
-#define A_lw		w7
-#define dst		x8
-#define tmp3w		w9
+#define L(l) .L ## l
 
 ENTRY(__memset_chk)
-  cmp count, dst_count
+  cmp count, dst
   bls memset
 
   // Preserve for accurate backtrace.
@@ -76,171 +100,152 @@
 
 ENTRY(memset)
 
-	mov	dst, dstin		/* Preserve return value.  */
-	ands	A_lw, val, #255
-	b.eq	.Lzero_mem
-	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
-.Ltail_maybe_long:
-	cmp	count, #64
-	b.ge	.Lnot_short
-.Ltail_maybe_tiny:
-	cmp	count, #15
-	b.le	.Ltail15tiny
-.Ltail63:
-	ands	tmp1, count, #0x30
-	b.eq	.Ltail15
-	add	dst, dst, tmp1
-	cmp	tmp1w, #0x20
-	b.eq	1f
-	b.lt	2f
-	stp	A_l, A_l, [dst, #-48]
-1:
-	stp	A_l, A_l, [dst, #-32]
-2:
-	stp	A_l, A_l, [dst, #-16]
+	dup	v0.16B, valw
+	add	dstend, dstin, count
 
-.Ltail15:
-	and	count, count, #15
-	add	dst, dst, count
-	stp	A_l, A_l, [dst, #-16]	/* Repeat some/all of last store. */
+	cmp	count, 96
+	b.hi	L(set_long)
+	cmp	count, 16
+	b.hs	L(set_medium)
+	mov	val, v0.D[0]
+
+	/* Set 0..15 bytes.  */
+	tbz	count, 3, 1f
+	str	val, [dstin]
+	str	val, [dstend, -8]
+	ret
+	nop
+1:	tbz	count, 2, 2f
+	str	valw, [dstin]
+	str	valw, [dstend, -4]
+	ret
+2:	cbz	count, 3f
+	strb	valw, [dstin]
+	tbz	count, 1, 3f
+	strh	valw, [dstend, -2]
+3:	ret
+
+	/* Set 17..96 bytes.  */
+L(set_medium):
+	str	q0, [dstin]
+	tbnz	count, 6, L(set96)
+	str	q0, [dstend, -16]
+	tbz	count, 5, 1f
+	str	q0, [dstin, 16]
+	str	q0, [dstend, -32]
+1:	ret
+
+	.p2align 4
+	/* Set 64..96 bytes.  Write 64 bytes from the start and
+	   32 bytes from the end.  */
+L(set96):
+	str	q0, [dstin, 16]
+	stp	q0, q0, [dstin, 32]
+	stp	q0, q0, [dstend, -32]
 	ret
 
-.Ltail15tiny:
-	/* Set up to 15 bytes.  Does not assume earlier memory
-	   being set.  */
-	tbz	count, #3, 1f
-	str	A_l, [dst], #8
-1:
-	tbz	count, #2, 1f
-	str	A_lw, [dst], #4
-1:
-	tbz	count, #1, 1f
-	strh	A_lw, [dst], #2
-1:
-	tbz	count, #0, 1f
-	strb	A_lw, [dst]
-1:
+	.p2align 3
+	nop
+L(set_long):
+	and	valw, valw, 255
+	bic	dst, dstin, 15
+	str	q0, [dstin]
+	cmp	count, 256
+	ccmp	valw, 0, 0, cs
+	b.eq	L(try_zva)
+L(no_zva):
+	sub	count, dstend, dst	/* Count is 16 too large.  */
+	add	dst, dst, 16
+	sub	count, count, 64 + 16	/* Adjust count and bias for loop.  */
+1:	stp	q0, q0, [dst], 64
+	stp	q0, q0, [dst, -32]
+L(tail64):
+	subs	count, count, 64
+	b.hi	1b
+2:	stp	q0, q0, [dstend, -64]
+	stp	q0, q0, [dstend, -32]
 	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 6
-.Lnot_short:
-	neg	tmp2, dst
-	ands	tmp2, tmp2, #15
-	b.eq	2f
-	/* Bring DST to 128-bit (16-byte) alignment.  We know that there's
-	 * more than that to set, so we simply store 16 bytes and advance by
-	 * the amount required to reach alignment.  */
-	sub	count, count, tmp2
-	stp	A_l, A_l, [dst]
-	add	dst, dst, tmp2
-	/* There may be less than 63 bytes to go now.  */
-	cmp	count, #63
-	b.le	.Ltail63
-2:
-	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
-	ret
-
-	/* For zeroing memory, check to see if we can use the ZVA feature to
-	 * zero entire 'cache' lines.  */
-.Lzero_mem:
-	mov	A_l, #0
-	cmp	count, #63
-	b.le	.Ltail_maybe_tiny
-	neg	tmp2, dst
-	ands	tmp2, tmp2, #15
-	b.eq	1f
-	sub	count, count, tmp2
-	stp	A_l, A_l, [dst]
-	add	dst, dst, tmp2
-	cmp	count, #63
-	b.le	.Ltail63
-1:
-	/* For zeroing small amounts of memory, it's not worth setting up
-	 * the line-clear code.  */
-	cmp	count, #128
-	b.lt	.Lnot_short
-#ifdef MAYBE_VIRT
-	/* For efficiency when virtualized, we cache the ZVA capability.  */
-	adrp	tmp2, .Lcache_clear
-	ldr	zva_len, [tmp2, #:lo12:.Lcache_clear]
-	tbnz	zva_len, #31, .Lnot_short
-	cbnz	zva_len, .Lzero_by_line
+	.p2align 3
+L(try_zva):
 	mrs	tmp1, dczid_el0
-	tbz	tmp1, #4, 1f
-	/* ZVA not available.  Remember this for next time.  */
-	mov	zva_len, #~0
-	str	zva_len, [tmp2, #:lo12:.Lcache_clear]
-	b	.Lnot_short
-1:
-	mov	tmp3w, #4
-	and	zva_len, tmp1w, #15	/* Safety: other bits reserved.  */
-	lsl	zva_len, tmp3w, zva_len
-	str	zva_len, [tmp2, #:lo12:.Lcache_clear]
-#else
-	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
-#endif
+	tbnz	tmp1w, 4, L(no_zva)
+	and	tmp1w, tmp1w, 15
+	cmp	tmp1w, 4	/* ZVA size is 64 bytes.  */
+	b.ne	 L(zva_128)
 
-.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	1f			/* Already aligned.  */
-	/* Not aligned, check that there's enough to copy after alignment.  */
-	sub	tmp1, count, tmp2
-	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
-2:
-	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	2b
-	/* We've overrun a bit, so adjust dst downwards.  */
-	add	dst, dst, tmp2
-1:
-	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
+	/* Write the first and last 64 byte aligned block using stp rather
+	   than using DC ZVA.  This is faster on some cores.
+	 */
+L(zva_64):
+	str	q0, [dst, 16]
+	stp	q0, q0, [dst, 32]
+	bic	dst, dst, 63
+	stp	q0, q0, [dst, 64]
+	stp	q0, q0, [dst, 96]
+	sub	count, dstend, dst	/* Count is now 128 too large.	*/
+	sub	count, count, 128+64+64	/* Adjust count and bias for loop.  */
+	add	dst, dst, 128
+	nop
+1:	dc	zva, dst
+	add	dst, dst, 64
+	subs	count, count, 64
+	b.hi	1b
+	stp	q0, q0, [dst, 0]
+	stp	q0, q0, [dst, 32]
+	stp	q0, q0, [dstend, -64]
+	stp	q0, q0, [dstend, -32]
 	ret
+
+	.p2align 3
+L(zva_128):
+	cmp	tmp1w, 5	/* ZVA size is 128 bytes.  */
+	b.ne	L(zva_other)
+
+	str	q0, [dst, 16]
+	stp	q0, q0, [dst, 32]
+	stp	q0, q0, [dst, 64]
+	stp	q0, q0, [dst, 96]
+	bic	dst, dst, 127
+	sub	count, dstend, dst	/* Count is now 128 too large.	*/
+	sub	count, count, 128+128	/* Adjust count and bias for loop.  */
+	add	dst, dst, 128
+1:	dc	zva, dst
+	add	dst, dst, 128
+	subs	count, count, 128
+	b.hi	1b
+	stp	q0, q0, [dstend, -128]
+	stp	q0, q0, [dstend, -96]
+	stp	q0, q0, [dstend, -64]
+	stp	q0, q0, [dstend, -32]
+	ret
+
+L(zva_other):
+	mov	tmp2w, 4
+	lsl	zva_lenw, tmp2w, tmp1w
+	add	tmp1, zva_len, 64	/* Max alignment bytes written.	 */
+	cmp	count, tmp1
+	blo	L(no_zva)
+
+	sub	tmp2, zva_len, 1
+	add	tmp1, dst, zva_len
+	add	dst, dst, 16
+	subs	count, tmp1, dst	/* Actual alignment bytes to write.  */
+	bic	tmp1, tmp1, tmp2	/* Aligned dc zva start address.  */
+	beq	2f
+1:	stp	q0, q0, [dst], 64
+	stp	q0, q0, [dst, -32]
+	subs	count, count, 64
+	b.hi	1b
+2:	mov	dst, tmp1
+	sub	count, dstend, tmp1	/* Remaining bytes to write.  */
+	subs	count, count, zva_len
+	b.lo	4f
+3:	dc	zva, dst
+	add	dst, dst, zva_len
+	subs	count, count, zva_len
+	b.hs	3b
+4:	add	count, count, zva_len
+	b	L(tail64)
+
 END(memset)
-
-#ifdef MAYBE_VIRT
-	.bss
-	.p2align 2
-.Lcache_clear:
-	.space 4
-#endif
diff --git a/libc/arch-arm64/generic/bionic/strlen.S b/libc/arch-arm64/generic/bionic/strlen.S
index 3bd9809..6e540fc 100644
--- a/libc/arch-arm64/generic/bionic/strlen.S
+++ b/libc/arch-arm64/generic/bionic/strlen.S
@@ -1,16 +1,16 @@
-/* Copyright (c) 2014, Linaro Limited
+/* Copyright (c) 2013-2015, Linaro Limited
    All rights reserved.
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are met:
        * Redistributions of source code must retain the above copyright
-         notice, this list of conditions and the following disclaimer.
+	 notice, this list of conditions and the following disclaimer.
        * 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.
+	 notice, this list of conditions and the following disclaimer in the
+	 documentation and/or other materials provided with the distribution.
        * Neither the name of the Linaro nor the
-         names of its contributors may be used to endorse or promote products
-         derived from this software without specific prior written permission.
+	 names of its contributors may be used to endorse or promote products
+	 derived from this software without specific prior written permission.
 
    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
@@ -22,16 +22,19 @@
    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.
-*/
+   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */
 
 /* Assumptions:
  *
- * ARMv8-a, AArch64
+ * ARMv8-a, AArch64, unaligned accesses, min page size 4k.
  */
 
 #include <private/bionic_asm.h>
 
+/* To test the page crossing code path more thoroughly, compile with
+   -DTEST_PAGE_CROSS - this will force all calls through the slower
+   entry path.  This option is not intended for production use.	 */
+
 /* Arguments and results.  */
 #define srcin		x0
 #define len		x0
@@ -40,87 +43,185 @@
 #define src		x1
 #define data1		x2
 #define data2		x3
-#define data2a		x4
-#define has_nul1	x5
-#define has_nul2	x6
-#define tmp1		x7
-#define tmp2		x8
-#define tmp3		x9
-#define tmp4		x10
-#define zeroones	x11
-#define pos		x12
+#define has_nul1	x4
+#define has_nul2	x5
+#define tmp1		x4
+#define tmp2		x5
+#define tmp3		x6
+#define tmp4		x7
+#define zeroones	x8
+
+#define L(l) .L ## l
+
+	/* NUL detection works on the principle that (X - 1) & (~X) & 0x80
+	   (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
+	   can be done in parallel across the entire word. A faster check
+	   (X - 1) & 0x80 is zero for non-NUL ASCII characters, but gives
+	   false hits for characters 129..255.	*/
 
 #define REP8_01 0x0101010101010101
 #define REP8_7f 0x7f7f7f7f7f7f7f7f
 #define REP8_80 0x8080808080808080
 
-	/* Start of critial section -- keep to one 64Byte cache line.  */
-ENTRY(strlen)
-	mov	zeroones, #REP8_01
-	bic	src, srcin, #15
-	ands	tmp1, srcin, #15
-	b.ne	.Lmisaligned
-	/* NUL detection works on the principle that (X - 1) & (~X) & 0x80
-	   (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
-	   can be done in parallel across the entire word.  */
-	/* The inner loop deals with two Dwords at a time.  This has a
-	   slightly higher start-up cost, but we should win quite quickly,
-	   especially on cores with a high number of issue slots per
-	   cycle, as we get much better parallelism out of the operations.  */
-.Lloop:
-	ldp	data1, data2, [src], #16
-.Lrealigned:
-	sub	tmp1, data1, zeroones
-	orr	tmp2, data1, #REP8_7f
-	sub	tmp3, data2, zeroones
-	orr	tmp4, data2, #REP8_7f
-	bic	has_nul1, tmp1, tmp2
-	bics	has_nul2, tmp3, tmp4
-	ccmp	has_nul1, #0, #0, eq	/* NZCV = 0000  */
-	b.eq	.Lloop
-	/* End of critical section -- keep to one 64Byte cache line.  */
-
-	sub	len, src, srcin
-	cbz	has_nul1, .Lnul_in_data2
-#ifdef __AARCH64EB__
-	mov	data2, data1
+#ifdef TEST_PAGE_CROSS
+# define MIN_PAGE_SIZE 15
+#else
+# define MIN_PAGE_SIZE 4096
 #endif
-	sub	len, len, #8
-	mov	has_nul2, has_nul1
-.Lnul_in_data2:
+
+	/* Since strings are short on average, we check the first 16 bytes
+	   of the string for a NUL character.  In order to do an unaligned ldp
+	   safely we have to do a page cross check first.  If there is a NUL
+	   byte we calculate the length from the 2 8-byte words using
+	   conditional select to reduce branch mispredictions (it is unlikely
+	   strlen will be repeatedly called on strings with the same length).
+
+	   If the string is longer than 16 bytes, we align src so don't need
+	   further page cross checks, and process 32 bytes per iteration
+	   using the fast NUL check.  If we encounter non-ASCII characters,
+	   fallback to a second loop using the full NUL check.
+
+	   If the page cross check fails, we read 16 bytes from an aligned
+	   address, remove any characters before the string, and continue
+	   in the main loop using aligned loads.  Since strings crossing a
+	   page in the first 16 bytes are rare (probability of
+	   16/MIN_PAGE_SIZE ~= 0.4%), this case does not need to be optimized.
+
+	   AArch64 systems have a minimum page size of 4k.  We don't bother
+	   checking for larger page sizes - the cost of setting up the correct
+	   page size is just not worth the extra gain from a small reduction in
+	   the cases taking the slow path.  Note that we only care about
+	   whether the first fetch, which may be misaligned, crosses a page
+	   boundary.  */
+
+ENTRY(strlen)
+	and	tmp1, srcin, MIN_PAGE_SIZE - 1
+	mov	zeroones, REP8_01
+	cmp	tmp1, MIN_PAGE_SIZE - 16
+	b.gt	L(page_cross)
+	ldp	data1, data2, [srcin]
 #ifdef __AARCH64EB__
 	/* For big-endian, carry propagation (if the final byte in the
-	   string is 0x01) means we cannot use has_nul directly.  The
-	   easiest way to get the correct byte is to byte-swap the data
-	   and calculate the syndrome a second time.  */
+	   string is 0x01) means we cannot use has_nul1/2 directly.
+	   Since we expect strings to be small and early-exit,
+	   byte-swap the data now so has_null1/2 will be correct.  */
+	rev	data1, data1
 	rev	data2, data2
-	sub	tmp1, data2, zeroones
-	orr	tmp2, data2, #REP8_7f
-	bic	has_nul2, tmp1, tmp2
 #endif
-	sub	len, len, #8
-	rev	has_nul2, has_nul2
-	clz	pos, has_nul2
-	add	len, len, pos, lsr #3		/* Bits to bytes.  */
+	sub	tmp1, data1, zeroones
+	orr	tmp2, data1, REP8_7f
+	sub	tmp3, data2, zeroones
+	orr	tmp4, data2, REP8_7f
+	bics	has_nul1, tmp1, tmp2
+	bic	has_nul2, tmp3, tmp4
+	ccmp	has_nul2, 0, 0, eq
+	beq	L(main_loop_entry)
+
+	/* Enter with C = has_nul1 == 0.  */
+	csel	has_nul1, has_nul1, has_nul2, cc
+	mov	len, 8
+	rev	has_nul1, has_nul1
+	clz	tmp1, has_nul1
+	csel	len, xzr, len, cc
+	add	len, len, tmp1, lsr 3
 	ret
 
-.Lmisaligned:
-	cmp	tmp1, #8
-	neg	tmp1, tmp1
-	ldp	data1, data2, [src], #16
-	lsl	tmp1, tmp1, #3		/* Bytes beyond alignment -> bits.  */
-	mov	tmp2, #~0
+	/* The inner loop processes 32 bytes per iteration and uses the fast
+	   NUL check.  If we encounter non-ASCII characters, use a second
+	   loop with the accurate NUL check.  */
+	.p2align 4
+L(main_loop_entry):
+	bic	src, srcin, 15
+	sub	src, src, 16
+L(main_loop):
+	ldp	data1, data2, [src, 32]!
+.Lpage_cross_entry:
+	sub	tmp1, data1, zeroones
+	sub	tmp3, data2, zeroones
+	orr	tmp2, tmp1, tmp3
+	tst	tmp2, zeroones, lsl 7
+	bne	1f
+	ldp	data1, data2, [src, 16]
+	sub	tmp1, data1, zeroones
+	sub	tmp3, data2, zeroones
+	orr	tmp2, tmp1, tmp3
+	tst	tmp2, zeroones, lsl 7
+	beq	L(main_loop)
+	add	src, src, 16
+1:
+	/* The fast check failed, so do the slower, accurate NUL check.	 */
+	orr	tmp2, data1, REP8_7f
+	orr	tmp4, data2, REP8_7f
+	bics	has_nul1, tmp1, tmp2
+	bic	has_nul2, tmp3, tmp4
+	ccmp	has_nul2, 0, 0, eq
+	beq	L(nonascii_loop)
+
+	/* Enter with C = has_nul1 == 0.  */
+L(tail):
 #ifdef __AARCH64EB__
-	/* Big-endian.  Early bytes are at MSB.  */
-	lsl	tmp2, tmp2, tmp1	/* Shift (tmp1 & 63).  */
+	/* For big-endian, carry propagation (if the final byte in the
+	   string is 0x01) means we cannot use has_nul1/2 directly.  The
+	   easiest way to get the correct byte is to byte-swap the data
+	   and calculate the syndrome a second time.  */
+	csel	data1, data1, data2, cc
+	rev	data1, data1
+	sub	tmp1, data1, zeroones
+	orr	tmp2, data1, REP8_7f
+	bic	has_nul1, tmp1, tmp2
+#else
+	csel	has_nul1, has_nul1, has_nul2, cc
+#endif
+	sub	len, src, srcin
+	rev	has_nul1, has_nul1
+	add	tmp2, len, 8
+	clz	tmp1, has_nul1
+	csel	len, len, tmp2, cc
+	add	len, len, tmp1, lsr 3
+	ret
+
+L(nonascii_loop):
+	ldp	data1, data2, [src, 16]!
+	sub	tmp1, data1, zeroones
+	orr	tmp2, data1, REP8_7f
+	sub	tmp3, data2, zeroones
+	orr	tmp4, data2, REP8_7f
+	bics	has_nul1, tmp1, tmp2
+	bic	has_nul2, tmp3, tmp4
+	ccmp	has_nul2, 0, 0, eq
+	bne	L(tail)
+	ldp	data1, data2, [src, 16]!
+	sub	tmp1, data1, zeroones
+	orr	tmp2, data1, REP8_7f
+	sub	tmp3, data2, zeroones
+	orr	tmp4, data2, REP8_7f
+	bics	has_nul1, tmp1, tmp2
+	bic	has_nul2, tmp3, tmp4
+	ccmp	has_nul2, 0, 0, eq
+	beq	L(nonascii_loop)
+	b	L(tail)
+
+	/* Load 16 bytes from [srcin & ~15] and force the bytes that precede
+	   srcin to 0x7f, so we ignore any NUL bytes before the string.
+	   Then continue in the aligned loop.  */
+L(page_cross):
+	bic	src, srcin, 15
+	ldp	data1, data2, [src]
+	lsl	tmp1, srcin, 3
+	mov	tmp4, -1
+#ifdef __AARCH64EB__
+	/* Big-endian.	Early bytes are at MSB.	 */
+	lsr	tmp1, tmp4, tmp1	/* Shift (tmp1 & 63).  */
 #else
 	/* Little-endian.  Early bytes are at LSB.  */
-	lsr	tmp2, tmp2, tmp1	/* Shift (tmp1 & 63).  */
+	lsl	tmp1, tmp4, tmp1	/* Shift (tmp1 & 63).  */
 #endif
-	orr	data1, data1, tmp2
-	orr	data2a, data2, tmp2
-	csinv	data1, data1, xzr, le
-	csel	data2, data2, data2a, le
-	b	.Lrealigned
+	orr	tmp1, tmp1, REP8_80
+	orn	data1, data1, tmp1
+	orn	tmp2, data2, tmp1
+	tst	srcin, 8
+	csel	data1, data1, tmp4, eq
+	csel	data2, data2, tmp2, eq
+	b	L(page_cross_entry)
 
 END(strlen)