xref: /freebsd/contrib/arm-optimized-routines/string/arm/memchr.S (revision 3c4ba5f55438f7afd4f4b0b56f88f2bb505fd6a6)
1/*
2 * memchr - scan memory for a character
3 *
4 * Copyright (c) 2010-2021, Arm Limited.
5 * SPDX-License-Identifier: MIT
6 */
7
8/*
9   Written by Dave Gilbert <david.gilbert@linaro.org>
10
11   This __memchr_arm routine is optimised on a Cortex-A9 and should work on
12   all ARMv7 processors.   It has a fast past for short sizes, and has
13   an optimised path for large data sets; the worst case is finding the
14   match early in a large data set.
15
16 */
17
18@ 2011-02-07 david.gilbert@linaro.org
19@    Extracted from local git a5b438d861
20@ 2011-07-14 david.gilbert@linaro.org
21@    Import endianness fix from local git ea786f1b
22@ 2011-12-07 david.gilbert@linaro.org
23@    Removed unneeded cbz from align loop
24
25	.syntax unified
26	.arch armv7-a
27
28@ this lets us check a flag in a 00/ff byte easily in either endianness
29#ifdef __ARMEB__
30#define CHARTSTMASK(c) 1<<(31-(c*8))
31#else
32#define CHARTSTMASK(c) 1<<(c*8)
33#endif
34	.thumb
35
36@ ---------------------------------------------------------------------------
37	.thumb_func
38	.align 2
39	.p2align 4,,15
40	.global __memchr_arm
41	.type __memchr_arm,%function
42__memchr_arm:
43	@ r0 = start of memory to scan
44	@ r1 = character to look for
45	@ r2 = length
46	@ returns r0 = pointer to character or NULL if not found
47	and	r1,r1,#0xff	@ Don't think we can trust the caller to actually pass a char
48
49	cmp	r2,#16		@ If it's short don't bother with anything clever
50	blt	20f
51
52	tst	r0, #7		@ If it's already aligned skip the next bit
53	beq	10f
54
55	@ Work up to an aligned point
565:
57	ldrb	r3, [r0],#1
58	subs	r2, r2, #1
59	cmp	r3, r1
60	beq	50f		@ If it matches exit found
61	tst	r0, #7
62	bne	5b		@ If not aligned yet then do next byte
63
6410:
65	@ At this point, we are aligned, we know we have at least 8 bytes to work with
66	push	{r4,r5,r6,r7}
67	orr	r1, r1, r1, lsl #8	@ expand the match word across to all bytes
68	orr	r1, r1, r1, lsl #16
69	bic	r4, r2, #7	@ Number of double words to work with
70	mvns	r7, #0		@ all F's
71	movs	r3, #0
72
7315:
74	ldmia	r0!,{r5,r6}
75	subs	r4, r4, #8
76	eor	r5,r5, r1	@ Get it so that r5,r6 have 00's where the bytes match the target
77	eor	r6,r6, r1
78	uadd8	r5, r5, r7	@ Parallel add 0xff - sets the GE bits for anything that wasn't 0
79	sel	r5, r3, r7	@ bytes are 00 for none-00 bytes, or ff for 00 bytes - NOTE INVERSION
80	uadd8	r6, r6, r7	@ Parallel add 0xff - sets the GE bits for anything that wasn't 0
81	sel	r6, r5, r7	@ chained....bytes are 00 for none-00 bytes, or ff for 00 bytes - NOTE INVERSION
82	cbnz	r6, 60f
83	bne	15b		@ (Flags from the subs above) If not run out of bytes then go around again
84
85	pop	{r4,r5,r6,r7}
86	and	r1,r1,#0xff	@ Get r1 back to a single character from the expansion above
87	and	r2,r2,#7	@ Leave the count remaining as the number after the double words have been done
88
8920:
90	cbz	r2, 40f		@ 0 length or hit the end already then not found
91
9221:  @ Post aligned section, or just a short call
93	ldrb	r3,[r0],#1
94	subs	r2,r2,#1
95	eor	r3,r3,r1	@ r3 = 0 if match - doesn't break flags from sub
96	cbz	r3, 50f
97	bne	21b		@ on r2 flags
98
9940:
100	movs	r0,#0		@ not found
101	bx	lr
102
10350:
104	subs	r0,r0,#1	@ found
105	bx	lr
106
10760:  @ We're here because the fast path found a hit - now we have to track down exactly which word it was
108	@ r0 points to the start of the double word after the one that was tested
109	@ r5 has the 00/ff pattern for the first word, r6 has the chained value
110	cmp	r5, #0
111	itte	eq
112	moveq	r5, r6		@ the end is in the 2nd word
113	subeq	r0,r0,#3	@ Points to 2nd byte of 2nd word
114	subne	r0,r0,#7	@ or 2nd byte of 1st word
115
116	@ r0 currently points to the 3rd byte of the word containing the hit
117	tst	r5, # CHARTSTMASK(0)	@ 1st character
118	bne	61f
119	adds	r0,r0,#1
120	tst	r5, # CHARTSTMASK(1)	@ 2nd character
121	ittt	eq
122	addeq	r0,r0,#1
123	tsteq	r5, # (3<<15)		@ 2nd & 3rd character
124	@ If not the 3rd must be the last one
125	addeq	r0,r0,#1
126
12761:
128	pop	{r4,r5,r6,r7}
129	subs	r0,r0,#1
130	bx	lr
131
132	.size	__memchr_arm, . - __memchr_arm
133