Improve strncmp for mutually misaligned inputs

The mutually misaligned inputs on aarch64 are compared with a simple
byte copy, which is not very efficient.  Enhance the comparison
similar to strcmp by loading a double-word at a time.  The peak
performance improvement (i.e. 4k maxlen comparisons) due to this on
the strncmp microbenchmark in glibc is as follows:

falkor: 3.5x (up to 72% time reduction)
cortex-a73: 3.5x (up to 71% time reduction)
cortex-a53: 3.5x (up to 71% time reduction)

All mutually misaligned inputs from 16 bytes maxlen onwards show
upwards of 15% improvement and there is no measurable effect on the
performance of aligned/mutually aligned inputs.

newlib/libc/machine/aarch64/strncmp.S

@@ -1,4 +1,4 @@
-/* Copyright (c) 2013, Linaro Limited
+/* Copyright (c) 2013, 2018, Linaro Limited
    All rights reserved.
 
    Redistribution and use in source and binary forms, with or without
@@ -67,6 +67,7 @@
 #define limit_wd	x13
 #define mask		x14
 #define endloop		x15
+#define count		mask
 
 	.text
 	.p2align 6
@@ -78,9 +79,9 @@ def_fn strncmp
 	eor	tmp1, src1, src2
 	mov	zeroones, #REP8_01
 	tst	tmp1, #7
+	and	count, src1, #7
 	b.ne	.Lmisaligned8
-	ands	tmp1, src1, #7
+	cbnz	count, .Lmutual_align
-	b.ne	.Lmutual_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.  */
@@ -185,44 +186,105 @@ def_fn strncmp
 	bic	src1, src1, #7
 	bic	src2, src2, #7
 	ldr	data1, [src1], #8
-	neg	tmp3, tmp1, lsl #3	/* 64 - bits(bytes beyond 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 (tmp1 & 63).  */
+	lsl	tmp2, tmp2, tmp3	/* Shift (count & 63).  */
 #else
 	/* Little-endian.  Early bytes are at LSB.  */
-	lsr	tmp2, tmp2, tmp3	/* Shift (tmp1 & 63).  */
+	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, tmp1
+	add	limit, limit, count
-	add	tmp3, tmp3, tmp1
+	add	tmp3, tmp3, count
 	orr	data1, data1, tmp2
 	orr	data2, data2, tmp2
 	add	limit_wd, limit_wd, tmp3, lsr #3
 	b	.Lstart_realigned
 
-.Lret0:
-	mov	result, #0
-	ret
-
 	.p2align 6
+	/* Don't bother with dwords for up to 16 bytes.  */
 .Lmisaligned8:
-	sub	limit, limit, #1
+	cmp	limit, #16
-1:
+	b.hs	.Ltry_misaligned_words
+
+.Lbyte_loop:
 	/* Perhaps we can do better than this.  */
 	ldrb	data1w, [src1], #1
 	ldrb	data2w, [src2], #1
 	subs	limit, limit, #1
-	ccmp	data1w, #1, #0, cs	/* NZCV = 0b0000.  */
+	ccmp	data1w, #1, #0, hi	/* NZCV = 0b0000.  */
 	ccmp	data1w, data2w, #0, cs	/* NZCV = 0b0000.  */
-	b.eq	1b
+	b.eq	.Lbyte_loop
+.Ldone:
 	sub	result, data1, data2
 	ret
-	.size strncmp, . - strncmp
+	/* Align the SRC1 to a dword by doing a bytewise compare and then do
+	   the dword loop.  */
+.Ltry_misaligned_words:
+	lsr	limit_wd, limit, #3
+	cbz	count, .Ldo_misaligned
 
+	neg	count, count
+	and	count, count, #7
+	sub	limit, limit, count
+	lsr	limit_wd, limit, #3
+
+.Lpage_end_loop:
+	ldrb	data1w, [src1], #1
+	ldrb	data2w, [src2], #1
+	cmp	data1w, #1
+	ccmp	data1w, data2w, #0, cs	/* NZCV = 0b0000.  */
+	b.ne	.Ldone
+	subs	count, count, #1
+	b.hi	.Lpage_end_loop
+
+.Ldo_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	.Ldone_loop
+.Lloop_misaligned:
+	and	tmp2, src2, #0xff8
+	eor	tmp2, tmp2, #0xff8
+	cbz	tmp2, .Lpage_end_loop
+
+	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	.Lnot_limit
+	subs	limit_wd, limit_wd, #1
+	b.pl	.Lloop_misaligned
+
+.Ldone_loop:
+	/* We found a difference or a NULL before the limit was reached.  */
+	and	limit, limit, #7
+	cbz	limit, .Lnot_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
+	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	.Lnot_limit
+
+.Lret0:
+	mov	result, #0
+	ret
+	.size strncmp, . - strncmp
 #endif