xref: /linux/arch/arm64/lib/strncmp.S (revision 8137a49e1567726eb10fcf55ad141ac19804ca6b)
1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright (c) 2013-2021, Arm Limited.
4 *
5 * Adapted from the original at:
6 * https://github.com/ARM-software/optimized-routines/blob/e823e3abf5f89ecb/string/aarch64/strncmp.S
7 */
8
9#include <linux/linkage.h>
10#include <asm/assembler.h>
11
12/* Assumptions:
13 *
14 * ARMv8-a, AArch64
15 */
16
17#define L(label) .L ## label
18
19#define REP8_01 0x0101010101010101
20#define REP8_7f 0x7f7f7f7f7f7f7f7f
21#define REP8_80 0x8080808080808080
22
23/* Parameters and result.  */
24#define src1		x0
25#define src2		x1
26#define limit		x2
27#define result		x0
28
29/* Internal variables.  */
30#define data1		x3
31#define data1w		w3
32#define data2		x4
33#define data2w		w4
34#define has_nul		x5
35#define diff		x6
36#define syndrome	x7
37#define tmp1		x8
38#define tmp2		x9
39#define tmp3		x10
40#define zeroones	x11
41#define pos		x12
42#define limit_wd	x13
43#define mask		x14
44#define endloop		x15
45#define count		mask
46
47SYM_FUNC_START_WEAK_PI(strncmp)
48	cbz	limit, L(ret0)
49	eor	tmp1, src1, src2
50	mov	zeroones, #REP8_01
51	tst	tmp1, #7
52	and	count, src1, #7
53	b.ne	L(misaligned8)
54	cbnz	count, L(mutual_align)
55	/* Calculate the number of full and partial words -1.  */
56	sub	limit_wd, limit, #1	/* limit != 0, so no underflow.  */
57	lsr	limit_wd, limit_wd, #3	/* Convert to Dwords.  */
58
59	/* NUL detection works on the principle that (X - 1) & (~X) & 0x80
60	   (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
61	   can be done in parallel across the entire word.  */
62	.p2align 4
63L(loop_aligned):
64	ldr	data1, [src1], #8
65	ldr	data2, [src2], #8
66L(start_realigned):
67	subs	limit_wd, limit_wd, #1
68	sub	tmp1, data1, zeroones
69	orr	tmp2, data1, #REP8_7f
70	eor	diff, data1, data2	/* Non-zero if differences found.  */
71	csinv	endloop, diff, xzr, pl	/* Last Dword or differences.  */
72	bics	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
73	ccmp	endloop, #0, #0, eq
74	b.eq	L(loop_aligned)
75	/* End of main loop */
76
77	/* Not reached the limit, must have found the end or a diff.  */
78	tbz	limit_wd, #63, L(not_limit)
79
80	/* Limit % 8 == 0 => all bytes significant.  */
81	ands	limit, limit, #7
82	b.eq	L(not_limit)
83
84	lsl	limit, limit, #3	/* Bits -> bytes.  */
85	mov	mask, #~0
86#ifdef __AARCH64EB__
87	lsr	mask, mask, limit
88#else
89	lsl	mask, mask, limit
90#endif
91	bic	data1, data1, mask
92	bic	data2, data2, mask
93
94	/* Make sure that the NUL byte is marked in the syndrome.  */
95	orr	has_nul, has_nul, mask
96
97L(not_limit):
98	orr	syndrome, diff, has_nul
99
100#ifndef	__AARCH64EB__
101	rev	syndrome, syndrome
102	rev	data1, data1
103	/* The MS-non-zero bit of the syndrome marks either the first bit
104	   that is different, or the top bit of the first zero byte.
105	   Shifting left now will bring the critical information into the
106	   top bits.  */
107	clz	pos, syndrome
108	rev	data2, data2
109	lsl	data1, data1, pos
110	lsl	data2, data2, pos
111	/* But we need to zero-extend (char is unsigned) the value and then
112	   perform a signed 32-bit subtraction.  */
113	lsr	data1, data1, #56
114	sub	result, data1, data2, lsr #56
115	ret
116#else
117	/* For big-endian we cannot use the trick with the syndrome value
118	   as carry-propagation can corrupt the upper bits if the trailing
119	   bytes in the string contain 0x01.  */
120	/* However, if there is no NUL byte in the dword, we can generate
121	   the result directly.  We can't just subtract the bytes as the
122	   MSB might be significant.  */
123	cbnz	has_nul, 1f
124	cmp	data1, data2
125	cset	result, ne
126	cneg	result, result, lo
127	ret
1281:
129	/* Re-compute the NUL-byte detection, using a byte-reversed value.  */
130	rev	tmp3, data1
131	sub	tmp1, tmp3, zeroones
132	orr	tmp2, tmp3, #REP8_7f
133	bic	has_nul, tmp1, tmp2
134	rev	has_nul, has_nul
135	orr	syndrome, diff, has_nul
136	clz	pos, syndrome
137	/* The MS-non-zero bit of the syndrome marks either the first bit
138	   that is different, or the top bit of the first zero byte.
139	   Shifting left now will bring the critical information into the
140	   top bits.  */
141	lsl	data1, data1, pos
142	lsl	data2, data2, pos
143	/* But we need to zero-extend (char is unsigned) the value and then
144	   perform a signed 32-bit subtraction.  */
145	lsr	data1, data1, #56
146	sub	result, data1, data2, lsr #56
147	ret
148#endif
149
150L(mutual_align):
151	/* Sources are mutually aligned, but are not currently at an
152	   alignment boundary.  Round down the addresses and then mask off
153	   the bytes that precede the start point.
154	   We also need to adjust the limit calculations, but without
155	   overflowing if the limit is near ULONG_MAX.  */
156	bic	src1, src1, #7
157	bic	src2, src2, #7
158	ldr	data1, [src1], #8
159	neg	tmp3, count, lsl #3	/* 64 - bits(bytes beyond align). */
160	ldr	data2, [src2], #8
161	mov	tmp2, #~0
162	sub	limit_wd, limit, #1	/* limit != 0, so no underflow.  */
163#ifdef __AARCH64EB__
164	/* Big-endian.  Early bytes are at MSB.  */
165	lsl	tmp2, tmp2, tmp3	/* Shift (count & 63).  */
166#else
167	/* Little-endian.  Early bytes are at LSB.  */
168	lsr	tmp2, tmp2, tmp3	/* Shift (count & 63).  */
169#endif
170	and	tmp3, limit_wd, #7
171	lsr	limit_wd, limit_wd, #3
172	/* Adjust the limit. Only low 3 bits used, so overflow irrelevant.  */
173	add	limit, limit, count
174	add	tmp3, tmp3, count
175	orr	data1, data1, tmp2
176	orr	data2, data2, tmp2
177	add	limit_wd, limit_wd, tmp3, lsr #3
178	b	L(start_realigned)
179
180	.p2align 4
181	/* Don't bother with dwords for up to 16 bytes.  */
182L(misaligned8):
183	cmp	limit, #16
184	b.hs	L(try_misaligned_words)
185
186L(byte_loop):
187	/* Perhaps we can do better than this.  */
188	ldrb	data1w, [src1], #1
189	ldrb	data2w, [src2], #1
190	subs	limit, limit, #1
191	ccmp	data1w, #1, #0, hi	/* NZCV = 0b0000.  */
192	ccmp	data1w, data2w, #0, cs	/* NZCV = 0b0000.  */
193	b.eq	L(byte_loop)
194L(done):
195	sub	result, data1, data2
196	ret
197	/* Align the SRC1 to a dword by doing a bytewise compare and then do
198	   the dword loop.  */
199L(try_misaligned_words):
200	lsr	limit_wd, limit, #3
201	cbz	count, L(do_misaligned)
202
203	neg	count, count
204	and	count, count, #7
205	sub	limit, limit, count
206	lsr	limit_wd, limit, #3
207
208L(page_end_loop):
209	ldrb	data1w, [src1], #1
210	ldrb	data2w, [src2], #1
211	cmp	data1w, #1
212	ccmp	data1w, data2w, #0, cs	/* NZCV = 0b0000.  */
213	b.ne	L(done)
214	subs	count, count, #1
215	b.hi	L(page_end_loop)
216
217L(do_misaligned):
218	/* Prepare ourselves for the next page crossing.  Unlike the aligned
219	   loop, we fetch 1 less dword because we risk crossing bounds on
220	   SRC2.  */
221	mov	count, #8
222	subs	limit_wd, limit_wd, #1
223	b.lo	L(done_loop)
224L(loop_misaligned):
225	and	tmp2, src2, #0xff8
226	eor	tmp2, tmp2, #0xff8
227	cbz	tmp2, L(page_end_loop)
228
229	ldr	data1, [src1], #8
230	ldr	data2, [src2], #8
231	sub	tmp1, data1, zeroones
232	orr	tmp2, data1, #REP8_7f
233	eor	diff, data1, data2	/* Non-zero if differences found.  */
234	bics	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
235	ccmp	diff, #0, #0, eq
236	b.ne	L(not_limit)
237	subs	limit_wd, limit_wd, #1
238	b.pl	L(loop_misaligned)
239
240L(done_loop):
241	/* We found a difference or a NULL before the limit was reached.  */
242	and	limit, limit, #7
243	cbz	limit, L(not_limit)
244	/* Read the last word.  */
245	sub	src1, src1, 8
246	sub	src2, src2, 8
247	ldr	data1, [src1, limit]
248	ldr	data2, [src2, limit]
249	sub	tmp1, data1, zeroones
250	orr	tmp2, data1, #REP8_7f
251	eor	diff, data1, data2	/* Non-zero if differences found.  */
252	bics	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
253	ccmp	diff, #0, #0, eq
254	b.ne	L(not_limit)
255
256L(ret0):
257	mov	result, #0
258	ret
259
260SYM_FUNC_END_PI(strncmp)
261EXPORT_SYMBOL_NOKASAN(strncmp)
262