Merge branch 'for-next/strings' into for-next/core

* for-next/strings:
  Revert "arm64: Mitigate MTE issues with str{n}cmp()"
  arm64: lib: Import latest version of Arm Optimized Routines' strncmp
  arm64: lib: Import latest version of Arm Optimized Routines' strcmp

4 files changed, 269 insertions, 213 deletions

diff --git a/arch/arm64/include/asm/assembler.h b/arch/arm64/include/asm/assembler.h
index e8bd0af0141c..8df412178efb 100644
--- a/arch/arm64/include/asm/assembler.h
+++ b/arch/arm64/include/asm/assembler.h
@@ -535,11 +535,6 @@ alternative_endif
 #define EXPORT_SYMBOL_NOKASAN(name)	EXPORT_SYMBOL(name)
 #endif
 
-#ifdef CONFIG_KASAN_HW_TAGS
-#define EXPORT_SYMBOL_NOHWKASAN(name)
-#else
-#define EXPORT_SYMBOL_NOHWKASAN(name)	EXPORT_SYMBOL_NOKASAN(name)
-#endif
 	/*
 	 * Emit a 64-bit absolute little endian symbol reference in a way that
 	 * ensures that it will be resolved at build time, even when building a
diff --git a/arch/arm64/include/asm/string.h b/arch/arm64/include/asm/string.h
index 95f7686b728d..3a3264ff47b9 100644
--- a/arch/arm64/include/asm/string.h
+++ b/arch/arm64/include/asm/string.h
@@ -12,13 +12,11 @@ extern char *strrchr(const char *, int c);
 #define __HAVE_ARCH_STRCHR
 extern char *strchr(const char *, int c);
 
-#ifndef CONFIG_KASAN_HW_TAGS
 #define __HAVE_ARCH_STRCMP
 extern int strcmp(const char *, const char *);
 
 #define __HAVE_ARCH_STRNCMP
 extern int strncmp(const char *, const char *, __kernel_size_t);
-#endif
 
 #define __HAVE_ARCH_STRLEN
 extern __kernel_size_t strlen(const char *);
diff --git a/arch/arm64/lib/strcmp.S b/arch/arm64/lib/strcmp.S
index cda7de747efc..9b89b4533607 100644
--- a/arch/arm64/lib/strcmp.S
+++ b/arch/arm64/lib/strcmp.S
@@ -1,9 +1,9 @@
 /* SPDX-License-Identifier: GPL-2.0-only */
 /*
- * Copyright (c) 2012-2021, Arm Limited.
+ * Copyright (c) 2012-2022, Arm Limited.
  *
  * Adapted from the original at:
- * https://github.com/ARM-software/optimized-routines/blob/afd6244a1f8d9229/string/aarch64/strcmp.S
+ * https://github.com/ARM-software/optimized-routines/blob/189dfefe37d54c5b/string/aarch64/strcmp.S
  */
 
 #include <linux/linkage.h>
@@ -11,161 +11,175 @@
 
 /* Assumptions:
  *
- * ARMv8-a, AArch64
+ * ARMv8-a, AArch64.
+ * MTE compatible.
  */
 
 #define L(label) .L ## label
 
 #define REP8_01 0x0101010101010101
 #define REP8_7f 0x7f7f7f7f7f7f7f7f
-#define REP8_80 0x8080808080808080
 
-/* Parameters and result.  */
 #define src1		x0
 #define src2		x1
 #define result		x0
 
-/* Internal variables.  */
 #define data1		x2
 #define data1w		w2
 #define data2		x3
 #define data2w		w3
 #define has_nul		x4
 #define diff		x5
+#define off1		x5
 #define syndrome	x6
-#define tmp1		x7
-#define tmp2		x8
-#define tmp3		x9
-#define zeroones	x10
-#define pos		x11
-
-	/* Start of performance-critical section  -- one 64B cache line.  */
-	.align 6
+#define tmp		x6
+#define data3		x7
+#define zeroones	x8
+#define shift		x9
+#define off2		x10
+
+/* On big-endian early bytes are at MSB and on little-endian LSB.
+   LS_FW means shifting towards early bytes.  */
+#ifdef __AARCH64EB__
+# define LS_FW lsl
+#else
+# define LS_FW lsr
+#endif
+
+/* 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.
+   Since carry propagation makes 0x1 bytes before a NUL byte appear
+   NUL too in big-endian, byte-reverse the data before the NUL check.  */
+
+
 SYM_FUNC_START(__pi_strcmp)
-	eor	tmp1, src1, src2
-	mov	zeroones, #REP8_01
-	tst	tmp1, #7
+	sub	off2, src2, src1
+	mov	zeroones, REP8_01
+	and	tmp, src1, 7
+	tst	off2, 7
 	b.ne	L(misaligned8)
-	ands	tmp1, src1, #7
-	b.ne	L(mutual_align)
-	/* 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.  */
+	cbnz	tmp, L(mutual_align)
+
+	.p2align 4
+
 L(loop_aligned):
-	ldr	data1, [src1], #8
-	ldr	data2, [src2], #8
+	ldr	data2, [src1, off2]
+	ldr	data1, [src1], 8
 L(start_realigned):
-	sub	tmp1, data1, zeroones
-	orr	tmp2, data1, #REP8_7f
-	eor	diff, data1, data2	/* Non-zero if differences found.  */
-	bic	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
+#ifdef __AARCH64EB__
+	rev	tmp, data1
+	sub	has_nul, tmp, zeroones
+	orr	tmp, tmp, REP8_7f
+#else
+	sub	has_nul, data1, zeroones
+	orr	tmp, data1, REP8_7f
+#endif
+	bics	has_nul, has_nul, tmp	/* Non-zero if NUL terminator.  */
+	ccmp	data1, data2, 0, eq
+	b.eq	L(loop_aligned)
+#ifdef __AARCH64EB__
+	rev	has_nul, has_nul
+#endif
+	eor	diff, data1, data2
 	orr	syndrome, diff, has_nul
-	cbz	syndrome, L(loop_aligned)
-	/* End of performance-critical section  -- one 64B cache line.  */
-
 L(end):
-#ifndef	__AARCH64EB__
+#ifndef __AARCH64EB__
 	rev	syndrome, syndrome
 	rev	data1, data1
-	/* The MS-non-zero bit of the syndrome marks either the first bit
-	   that is different, or the top bit of the first zero byte.
-	   Shifting left now will bring the critical information into the
-	   top bits.  */
-	clz	pos, syndrome
 	rev	data2, data2
-	lsl	data1, data1, pos
-	lsl	data2, data2, pos
-	/* But we need to zero-extend (char is unsigned) the value and then
-	   perform a signed 32-bit subtraction.  */
-	lsr	data1, data1, #56
-	sub	result, data1, data2, lsr #56
-	ret
-#else
-	/* For big-endian we cannot use the trick with the syndrome value
-	   as carry-propagation can corrupt the upper bits if the trailing
-	   bytes in the string contain 0x01.  */
-	/* However, if there is no NUL byte in the dword, we can generate
-	   the result directly.  We can't just subtract the bytes as the
-	   MSB might be significant.  */
-	cbnz	has_nul, 1f
-	cmp	data1, data2
-	cset	result, ne
-	cneg	result, result, lo
-	ret
-1:
-	/* Re-compute the NUL-byte detection, using a byte-reversed value.  */
-	rev	tmp3, data1
-	sub	tmp1, tmp3, zeroones
-	orr	tmp2, tmp3, #REP8_7f
-	bic	has_nul, tmp1, tmp2
-	rev	has_nul, has_nul
-	orr	syndrome, diff, has_nul
-	clz	pos, syndrome
-	/* The MS-non-zero bit of the syndrome marks either the first bit
-	   that is different, or the top bit of the first zero byte.
+#endif
+	clz	shift, syndrome
+	/* The most-significant-non-zero bit of the syndrome marks either the
+	   first bit that is different, or the top bit of the first zero byte.
 	   Shifting left now will bring the critical information into the
 	   top bits.  */
-	lsl	data1, data1, pos
-	lsl	data2, data2, pos
+	lsl	data1, data1, shift
+	lsl	data2, data2, shift
 	/* But we need to zero-extend (char is unsigned) the value and then
 	   perform a signed 32-bit subtraction.  */
-	lsr	data1, data1, #56
-	sub	result, data1, data2, lsr #56
+	lsr	data1, data1, 56
+	sub	result, data1, data2, lsr 56
 	ret
-#endif
+
+	.p2align 4
 
 L(mutual_align):
 	/* Sources are mutually aligned, but are not currently at an
 	   alignment boundary.  Round down the addresses and then mask off
-	   the bytes that preceed the start point.  */
-	bic	src1, src1, #7
-	bic	src2, src2, #7
-	lsl	tmp1, tmp1, #3		/* Bytes beyond alignment -> bits.  */
-	ldr	data1, [src1], #8
-	neg	tmp1, tmp1		/* Bits to alignment -64.  */
-	ldr	data2, [src2], #8
-	mov	tmp2, #~0
-#ifdef __AARCH64EB__
-	/* Big-endian.  Early bytes are at MSB.  */
-	lsl	tmp2, tmp2, tmp1	/* Shift (tmp1 & 63).  */
-#else
-	/* Little-endian.  Early bytes are at LSB.  */
-	lsr	tmp2, tmp2, tmp1	/* Shift (tmp1 & 63).  */
-#endif
-	orr	data1, data1, tmp2
-	orr	data2, data2, tmp2
+	   the bytes that precede the start point.  */
+	bic	src1, src1, 7
+	ldr	data2, [src1, off2]
+	ldr	data1, [src1], 8
+	neg	shift, src2, lsl 3	/* Bits to alignment -64.  */
+	mov	tmp, -1
+	LS_FW	tmp, tmp, shift
+	orr	data1, data1, tmp
+	orr	data2, data2, tmp
 	b	L(start_realigned)
 
 L(misaligned8):
 	/* Align SRC1 to 8 bytes and then compare 8 bytes at a time, always
-	   checking to make sure that we don't access beyond page boundary in
-	   SRC2.  */
-	tst	src1, #7
-	b.eq	L(loop_misaligned)
+	   checking to make sure that we don't access beyond the end of SRC2.  */
+	cbz	tmp, L(src1_aligned)
 L(do_misaligned):
-	ldrb	data1w, [src1], #1
-	ldrb	data2w, [src2], #1
-	cmp	data1w, #1
-	ccmp	data1w, data2w, #0, cs	/* NZCV = 0b0000.  */
+	ldrb	data1w, [src1], 1
+	ldrb	data2w, [src2], 1
+	cmp	data1w, 0
+	ccmp	data1w, data2w, 0, ne	/* NZCV = 0b0000.  */
 	b.ne	L(done)
-	tst	src1, #7
+	tst	src1, 7
 	b.ne	L(do_misaligned)
 
-L(loop_misaligned):
-	/* Test if we are within the last dword of the end of a 4K page.  If
-	   yes then jump back to the misaligned loop to copy a byte at a time.  */
-	and	tmp1, src2, #0xff8
-	eor	tmp1, tmp1, #0xff8
-	cbz	tmp1, L(do_misaligned)
-	ldr	data1, [src1], #8
-	ldr	data2, [src2], #8
-
-	sub	tmp1, data1, zeroones
-	orr	tmp2, data1, #REP8_7f
-	eor	diff, data1, data2	/* Non-zero if differences found.  */
-	bic	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
+L(src1_aligned):
+	neg	shift, src2, lsl 3
+	bic	src2, src2, 7
+	ldr	data3, [src2], 8
+#ifdef __AARCH64EB__
+	rev	data3, data3
+#endif
+	lsr	tmp, zeroones, shift
+	orr	data3, data3, tmp
+	sub	has_nul, data3, zeroones
+	orr	tmp, data3, REP8_7f
+	bics	has_nul, has_nul, tmp
+	b.ne	L(tail)
+
+	sub	off1, src2, src1
+
+	.p2align 4
+
+L(loop_unaligned):
+	ldr	data3, [src1, off1]
+	ldr	data2, [src1, off2]
+#ifdef __AARCH64EB__
+	rev	data3, data3
+#endif
+	sub	has_nul, data3, zeroones
+	orr	tmp, data3, REP8_7f
+	ldr	data1, [src1], 8
+	bics	has_nul, has_nul, tmp
+	ccmp	data1, data2, 0, eq
+	b.eq	L(loop_unaligned)
+
+	lsl	tmp, has_nul, shift
+#ifdef __AARCH64EB__
+	rev	tmp, tmp
+#endif
+	eor	diff, data1, data2
+	orr	syndrome, diff, tmp
+	cbnz	syndrome, L(end)
+L(tail):
+	ldr	data1, [src1]
+	neg	shift, shift
+	lsr	data2, data3, shift
+	lsr	has_nul, has_nul, shift
+#ifdef __AARCH64EB__
+	rev     data2, data2
+	rev	has_nul, has_nul
+#endif
+	eor	diff, data1, data2
 	orr	syndrome, diff, has_nul
-	cbz	syndrome, L(loop_misaligned)
 	b	L(end)
 
 L(done):
@@ -173,4 +187,4 @@ L(done):
 	ret
 SYM_FUNC_END(__pi_strcmp)
 SYM_FUNC_ALIAS_WEAK(strcmp, __pi_strcmp)
-EXPORT_SYMBOL_NOHWKASAN(strcmp)
+EXPORT_SYMBOL_NOKASAN(strcmp)
diff --git a/arch/arm64/lib/strncmp.S b/arch/arm64/lib/strncmp.S
index a848abcec975..fe7bbc0b42a7 100644
--- a/arch/arm64/lib/strncmp.S
+++ b/arch/arm64/lib/strncmp.S
@@ -1,9 +1,9 @@
 /* SPDX-License-Identifier: GPL-2.0-only */
 /*
- * Copyright (c) 2013-2021, Arm Limited.
+ * Copyright (c) 2013-2022, Arm Limited.
  *
  * Adapted from the original at:
- * https://github.com/ARM-software/optimized-routines/blob/e823e3abf5f89ecb/string/aarch64/strncmp.S
+ * https://github.com/ARM-software/optimized-routines/blob/189dfefe37d54c5b/string/aarch64/strncmp.S
  */
 
 #include <linux/linkage.h>
@@ -11,14 +11,14 @@
 
 /* Assumptions:
  *
- * ARMv8-a, AArch64
+ * ARMv8-a, AArch64.
+ * MTE compatible.
  */
 
 #define L(label) .L ## label
 
 #define REP8_01 0x0101010101010101
 #define REP8_7f 0x7f7f7f7f7f7f7f7f
-#define REP8_80 0x8080808080808080
 
 /* Parameters and result.  */
 #define src1		x0
@@ -39,10 +39,24 @@
 #define tmp3		x10
 #define zeroones	x11
 #define pos		x12
-#define limit_wd	x13
-#define mask		x14
-#define endloop		x15
+#define mask		x13
+#define endloop		x14
 #define count		mask
+#define offset		pos
+#define neg_offset	x15
+
+/* Define endian dependent shift operations.
+   On big-endian early bytes are at MSB and on little-endian LSB.
+   LS_FW means shifting towards early bytes.
+   LS_BK means shifting towards later bytes.
+   */
+#ifdef __AARCH64EB__
+#define LS_FW lsl
+#define LS_BK lsr
+#else
+#define LS_FW lsr
+#define LS_BK lsl
+#endif
 
 SYM_FUNC_START(__pi_strncmp)
 	cbz	limit, L(ret0)
@@ -52,9 +66,6 @@ SYM_FUNC_START(__pi_strncmp)
 	and	count, src1, #7
 	b.ne	L(misaligned8)
 	cbnz	count, L(mutual_align)
-	/* Calculate the number of full and partial words -1.  */
-	sub	limit_wd, limit, #1	/* limit != 0, so no underflow.  */
-	lsr	limit_wd, limit_wd, #3	/* Convert to Dwords.  */
 
 	/* NUL detection works on the principle that (X - 1) & (~X) & 0x80
 	   (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
@@ -64,56 +75,52 @@ L(loop_aligned):
 	ldr	data1, [src1], #8
 	ldr	data2, [src2], #8
 L(start_realigned):
-	subs	limit_wd, limit_wd, #1
+	subs	limit, limit, #8
 	sub	tmp1, data1, zeroones
 	orr	tmp2, data1, #REP8_7f
 	eor	diff, data1, data2	/* Non-zero if differences found.  */
-	csinv	endloop, diff, xzr, pl	/* Last Dword or differences.  */
+	csinv	endloop, diff, xzr, hi	/* Last Dword or differences.  */
 	bics	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
 	ccmp	endloop, #0, #0, eq
 	b.eq	L(loop_aligned)
 	/* End of main loop */
 
-	/* Not reached the limit, must have found the end or a diff.  */
-	tbz	limit_wd, #63, L(not_limit)
-
-	/* Limit % 8 == 0 => all bytes significant.  */
-	ands	limit, limit, #7
-	b.eq	L(not_limit)
-
-	lsl	limit, limit, #3	/* Bits -> bytes.  */
-	mov	mask, #~0
-#ifdef __AARCH64EB__
-	lsr	mask, mask, limit
-#else
-	lsl	mask, mask, limit
-#endif
-	bic	data1, data1, mask
-	bic	data2, data2, mask
-
-	/* Make sure that the NUL byte is marked in the syndrome.  */
-	orr	has_nul, has_nul, mask
-
-L(not_limit):
+L(full_check):
+#ifndef __AARCH64EB__
 	orr	syndrome, diff, has_nul
-
-#ifndef	__AARCH64EB__
+	add	limit, limit, 8	/* Rewind limit to before last subs. */
+L(syndrome_check):
+	/* Limit was reached. Check if the NUL byte or the difference
+	   is before the limit. */
 	rev	syndrome, syndrome
 	rev	data1, data1
-	/* The MS-non-zero bit of the syndrome marks either the first bit
-	   that is different, or the top bit of the first zero byte.
-	   Shifting left now will bring the critical information into the
-	   top bits.  */
 	clz	pos, syndrome
 	rev	data2, data2
 	lsl	data1, data1, pos
+	cmp	limit, pos, lsr #3
 	lsl	data2, data2, pos
 	/* But we need to zero-extend (char is unsigned) the value and then
 	   perform a signed 32-bit subtraction.  */
 	lsr	data1, data1, #56
 	sub	result, data1, data2, lsr #56
+	csel result, result, xzr, hi
 	ret
 #else
+	/* Not reached the limit, must have found the end or a diff.  */
+	tbz	limit, #63, L(not_limit)
+	add	tmp1, limit, 8
+	cbz	limit, L(not_limit)
+
+	lsl	limit, tmp1, #3	/* Bits -> bytes.  */
+	mov	mask, #~0
+	lsr	mask, mask, limit
+	bic	data1, data1, mask
+	bic	data2, data2, mask
+
+	/* Make sure that the NUL byte is marked in the syndrome.  */
+	orr	has_nul, has_nul, mask
+
+L(not_limit):
 	/* For big-endian we cannot use the trick with the syndrome value
 	   as carry-propagation can corrupt the upper bits if the trailing
 	   bytes in the string contain 0x01.  */
@@ -134,10 +141,11 @@ L(not_limit):
 	rev	has_nul, has_nul
 	orr	syndrome, diff, has_nul
 	clz	pos, syndrome
-	/* The MS-non-zero bit of the syndrome marks either the first bit
-	   that is different, or the top bit of the first zero byte.
+	/* The most-significant-non-zero bit of the syndrome marks either the
+	   first bit that is different, or the top bit of the first zero byte.
 	   Shifting left now will bring the critical information into the
 	   top bits.  */
+L(end_quick):
 	lsl	data1, data1, pos
 	lsl	data2, data2, pos
 	/* But we need to zero-extend (char is unsigned) the value and then
@@ -159,22 +167,12 @@ L(mutual_align):
 	neg	tmp3, count, lsl #3	/* 64 - bits(bytes beyond align). */
 	ldr	data2, [src2], #8
 	mov	tmp2, #~0
-	sub	limit_wd, limit, #1	/* limit != 0, so no underflow.  */
-#ifdef __AARCH64EB__
-	/* Big-endian.  Early bytes are at MSB.  */
-	lsl	tmp2, tmp2, tmp3	/* Shift (count & 63).  */
-#else
-	/* Little-endian.  Early bytes are at LSB.  */
-	lsr	tmp2, tmp2, tmp3	/* Shift (count & 63).  */
-#endif
-	and	tmp3, limit_wd, #7
-	lsr	limit_wd, limit_wd, #3
-	/* Adjust the limit. Only low 3 bits used, so overflow irrelevant.  */
-	add	limit, limit, count
-	add	tmp3, tmp3, count
+	LS_FW	tmp2, tmp2, tmp3	/* Shift (count & 63).  */
+	/* Adjust the limit and ensure it doesn't overflow.  */
+	adds	limit, limit, count
+	csinv	limit, limit, xzr, lo
 	orr	data1, data1, tmp2
 	orr	data2, data2, tmp2
-	add	limit_wd, limit_wd, tmp3, lsr #3
 	b	L(start_realigned)
 
 	.p2align 4
@@ -197,13 +195,11 @@ L(done):
 	/* Align the SRC1 to a dword by doing a bytewise compare and then do
 	   the dword loop.  */
 L(try_misaligned_words):
-	lsr	limit_wd, limit, #3
-	cbz	count, L(do_misaligned)
+	cbz	count, L(src1_aligned)
 
 	neg	count, count
 	and	count, count, #7
 	sub	limit, limit, count
-	lsr	limit_wd, limit, #3
 
 L(page_end_loop):
 	ldrb	data1w, [src1], #1
@@ -214,48 +210,101 @@ L(page_end_loop):
 	subs	count, count, #1
 	b.hi	L(page_end_loop)
 
-L(do_misaligned):
-	/* Prepare ourselves for the next page crossing.  Unlike the aligned
-	   loop, we fetch 1 less dword because we risk crossing bounds on
-	   SRC2.  */
-	mov	count, #8
-	subs	limit_wd, limit_wd, #1
-	b.lo	L(done_loop)
-L(loop_misaligned):
-	and	tmp2, src2, #0xff8
-	eor	tmp2, tmp2, #0xff8
-	cbz	tmp2, L(page_end_loop)
+	/* The following diagram explains the comparison of misaligned strings.
+	   The bytes are shown in natural order. For little-endian, it is
+	   reversed in the registers. The "x" bytes are before the string.
+	   The "|" separates data that is loaded at one time.
+	   src1     | a a a a a a a a | b b b c c c c c | . . .
+	   src2     | x x x x x a a a   a a a a a b b b | c c c c c . . .
+
+	   After shifting in each step, the data looks like this:
+	                STEP_A              STEP_B              STEP_C
+	   data1    a a a a a a a a     b b b c c c c c     b b b c c c c c
+	   data2    a a a a a a a a     b b b 0 0 0 0 0     0 0 0 c c c c c
 
+	   The bytes with "0" are eliminated from the syndrome via mask.
+
+	   Align SRC2 down to 16 bytes. This way we can read 16 bytes at a
+	   time from SRC2. The comparison happens in 3 steps. After each step
+	   the loop can exit, or read from SRC1 or SRC2. */
+L(src1_aligned):
+	/* Calculate offset from 8 byte alignment to string start in bits. No
+	   need to mask offset since shifts are ignoring upper bits. */
+	lsl	offset, src2, #3
+	bic	src2, src2, #0xf
+	mov	mask, -1
+	neg	neg_offset, offset
 	ldr	data1, [src1], #8
-	ldr	data2, [src2], #8
-	sub	tmp1, data1, zeroones
-	orr	tmp2, data1, #REP8_7f
-	eor	diff, data1, data2	/* Non-zero if differences found.  */
-	bics	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
-	ccmp	diff, #0, #0, eq
-	b.ne	L(not_limit)
-	subs	limit_wd, limit_wd, #1
-	b.pl	L(loop_misaligned)
+	ldp	tmp1, tmp2, [src2], #16
+	LS_BK	mask, mask, neg_offset
+	and	neg_offset, neg_offset, #63	/* Need actual value for cmp later. */
+	/* Skip the first compare if data in tmp1 is irrelevant. */
+	tbnz	offset, 6, L(misaligned_mid_loop)
 
-L(done_loop):
-	/* We found a difference or a NULL before the limit was reached.  */
-	and	limit, limit, #7
-	cbz	limit, L(not_limit)
-	/* Read the last word.  */
-	sub	src1, src1, 8
-	sub	src2, src2, 8
-	ldr	data1, [src1, limit]
-	ldr	data2, [src2, limit]
-	sub	tmp1, data1, zeroones
-	orr	tmp2, data1, #REP8_7f
+L(loop_misaligned):
+	/* STEP_A: Compare full 8 bytes when there is enough data from SRC2.*/
+	LS_FW	data2, tmp1, offset
+	LS_BK	tmp1, tmp2, neg_offset
+	subs	limit, limit, #8
+	orr	data2, data2, tmp1	/* 8 bytes from SRC2 combined from two regs.*/
+	sub	has_nul, data1, zeroones
 	eor	diff, data1, data2	/* Non-zero if differences found.  */
-	bics	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
-	ccmp	diff, #0, #0, eq
-	b.ne	L(not_limit)
+	orr	tmp3, data1, #REP8_7f
+	csinv	endloop, diff, xzr, hi	/* If limit, set to all ones. */
+	bic	has_nul, has_nul, tmp3	/* Non-zero if NUL byte found in SRC1. */
+	orr	tmp3, endloop, has_nul
+	cbnz	tmp3, L(full_check)
+
+	ldr	data1, [src1], #8
+L(misaligned_mid_loop):
+	/* STEP_B: Compare first part of data1 to second part of tmp2. */
+	LS_FW	data2, tmp2, offset
+#ifdef __AARCH64EB__
+	/* For big-endian we do a byte reverse to avoid carry-propagation
+	problem described above. This way we can reuse the has_nul in the
+	next step and also use syndrome value trick at the end. */
+	rev	tmp3, data1
+	#define data1_fixed tmp3
+#else
+	#define data1_fixed data1
+#endif
+	sub	has_nul, data1_fixed, zeroones
+	orr	tmp3, data1_fixed, #REP8_7f
+	eor	diff, data2, data1	/* Non-zero if differences found.  */
+	bic	has_nul, has_nul, tmp3	/* Non-zero if NUL terminator.  */
+#ifdef __AARCH64EB__
+	rev	has_nul, has_nul
+#endif
+	cmp	limit, neg_offset, lsr #3
+	orr	syndrome, diff, has_nul
+	bic	syndrome, syndrome, mask	/* Ignore later bytes. */
+	csinv	tmp3, syndrome, xzr, hi	/* If limit, set to all ones. */
+	cbnz	tmp3, L(syndrome_check)
+
+	/* STEP_C: Compare second part of data1 to first part of tmp1. */
+	ldp	tmp1, tmp2, [src2], #16
+	cmp	limit, #8
+	LS_BK	data2, tmp1, neg_offset
+	eor	diff, data2, data1	/* Non-zero if differences found.  */
+	orr	syndrome, diff, has_nul
+	and	syndrome, syndrome, mask	/* Ignore earlier bytes. */
+	csinv	tmp3, syndrome, xzr, hi	/* If limit, set to all ones. */
+	cbnz	tmp3, L(syndrome_check)
+
+	ldr	data1, [src1], #8
+	sub	limit, limit, #8
+	b	L(loop_misaligned)
+
+#ifdef	__AARCH64EB__
+L(syndrome_check):
+	clz	pos, syndrome
+	cmp	pos, limit, lsl #3
+	b.lo	L(end_quick)
+#endif
 
 L(ret0):
 	mov	result, #0
 	ret
 SYM_FUNC_END(__pi_strncmp)
 SYM_FUNC_ALIAS_WEAK(strncmp, __pi_strncmp)
-EXPORT_SYMBOL_NOHWKASAN(strncmp)
+EXPORT_SYMBOL_NOKASAN(strncmp)