xref: /freebsd/contrib/arm-optimized-routines/string/aarch64/strcpy-sve.S (revision e6bfd18d21b225af6a0ed67ceeaf1293b7b9eba5)
1/*
2 * strcpy/stpcpy - copy a string returning pointer to start/end.
3 *
4 * Copyright (c) 2018-2021, Arm Limited.
5 * SPDX-License-Identifier: MIT
6 */
7
8#include "../asmdefs.h"
9
10#if __ARM_FEATURE_SVE
11/* Assumptions:
12 *
13 * ARMv8-a, AArch64
14 * SVE Available.
15 */
16
17/* To build as stpcpy, define BUILD_STPCPY before compiling this file.  */
18#ifdef BUILD_STPCPY
19#define FUNC  __stpcpy_aarch64_sve
20#else
21#define FUNC  __strcpy_aarch64_sve
22#endif
23
24ENTRY (FUNC)
25	PTR_ARG (0)
26	PTR_ARG (1)
27	setffr				/* initialize FFR */
28	ptrue	p2.b, all		/* all ones; loop invariant */
29	mov	x2, 0			/* initialize offset */
30
31	.p2align 4
32	/* Read a vector's worth of bytes, stopping on first fault.  */
330:	ldff1b	z0.b, p2/z, [x1, x2]
34	rdffrs	p0.b, p2/z
35	b.nlast	1f
36
37	/* First fault did not fail: the whole vector is valid.
38	   Avoid depending on the contexts of FFR beyond the branch.  */
39	cmpeq	p1.b, p2/z, z0.b, 0	/* search for zeros */
40	b.any	2f
41
42	/* No zero found.  Store the whole vector and loop.  */
43	st1b	z0.b, p2, [x0, x2]
44	incb	x2, all
45	b	0b
46
47	/* First fault failed: only some of the vector is valid.
48	   Perform the comparison only on the valid bytes.  */
491:	cmpeq	p1.b, p0/z, z0.b, 0	/* search for zeros */
50	b.any	2f
51
52	/* No zero found.  Store the valid portion of the vector and loop.  */
53	setffr				/* re-init FFR */
54	st1b	z0.b, p0, [x0, x2]
55	incp	x2, p0.b
56	b	0b
57
58	/* Zero found.  Crop the vector to the found zero and finish.  */
592:	brka	p0.b, p2/z, p1.b
60	st1b	z0.b, p0, [x0, x2]
61#ifdef BUILD_STPCPY
62	add	x0, x0, x2
63	sub	x0, x0, 1
64	incp	x0, p0.b
65#endif
66	ret
67
68END (FUNC)
69
70#endif
71
72