xref: /freebsd/contrib/arm-optimized-routines/string/aarch64/memchr-sve.S (revision 072a4ba82a01476eaee33781ccd241033eefcf0b)
131914882SAlex Richardson/*
231914882SAlex Richardson * memchr - find a character in a memory zone
331914882SAlex Richardson *
4*072a4ba8SAndrew Turner * Copyright (c) 2018-2022, Arm Limited.
5*072a4ba8SAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
631914882SAlex Richardson */
731914882SAlex Richardson
8*072a4ba8SAndrew Turner#include "asmdefs.h"
931914882SAlex Richardson
1031914882SAlex Richardson#if __ARM_FEATURE_SVE
1131914882SAlex Richardson/* Assumptions:
1231914882SAlex Richardson *
1331914882SAlex Richardson * ARMv8-a, AArch64
1431914882SAlex Richardson * SVE Available.
1531914882SAlex Richardson */
1631914882SAlex Richardson
1731914882SAlex RichardsonENTRY (__memchr_aarch64_sve)
1831914882SAlex Richardson	PTR_ARG (0)
1931914882SAlex Richardson	SIZE_ARG (2)
2031914882SAlex Richardson	dup	z1.b, w1			/* duplicate c to a vector */
2131914882SAlex Richardson	setffr					/* initialize FFR */
2231914882SAlex Richardson	mov	x3, 0				/* initialize off */
2331914882SAlex Richardson
2431914882SAlex Richardson	.p2align 4
2531914882SAlex Richardson0:	whilelo	p1.b, x3, x2			/* make sure off < max */
2631914882SAlex Richardson	b.none	9f
2731914882SAlex Richardson
2831914882SAlex Richardson	/* Read a vector's worth of bytes, bounded by max,
2931914882SAlex Richardson	   stopping on first fault.  */
3031914882SAlex Richardson	ldff1b	z0.b, p1/z, [x0, x3]
3131914882SAlex Richardson	rdffrs	p0.b, p1/z
3231914882SAlex Richardson	b.nlast	2f
3331914882SAlex Richardson
3431914882SAlex Richardson	/* First fault did not fail: the vector bounded by max is valid.
3531914882SAlex Richardson	   Avoid depending on the contents of FFR beyond the branch.  */
3631914882SAlex Richardson	incb	x3				/* speculate increment */
3731914882SAlex Richardson	cmpeq	p2.b, p1/z, z0.b, z1.b		/* search for c */
3831914882SAlex Richardson	b.none	0b
3931914882SAlex Richardson	decb	x3				/* undo speculate */
4031914882SAlex Richardson
4131914882SAlex Richardson	/* Found C.  */
4231914882SAlex Richardson1:	brkb	p2.b, p1/z, p2.b	/* find the first c */
4331914882SAlex Richardson	add	x0, x0, x3		/* form partial pointer */
4431914882SAlex Richardson	incp	x0, p2.b		/* form final pointer to c */
4531914882SAlex Richardson	ret
4631914882SAlex Richardson
4731914882SAlex Richardson	/* First fault failed: only some of the vector is valid.
4831914882SAlex Richardson	   Perform the comparision only on the valid bytes.  */
4931914882SAlex Richardson2:	cmpeq	p2.b, p0/z, z0.b, z1.b
5031914882SAlex Richardson	b.any	1b
5131914882SAlex Richardson
5231914882SAlex Richardson	/* No C found.  Re-init FFR, increment, and loop.  */
5331914882SAlex Richardson	setffr
5431914882SAlex Richardson	incp	x3, p0.b
5531914882SAlex Richardson	b	0b
5631914882SAlex Richardson
5731914882SAlex Richardson	/* Found end of count.  */
5831914882SAlex Richardson9:	mov	x0, 0			/* return null */
5931914882SAlex Richardson	ret
6031914882SAlex Richardson
6131914882SAlex RichardsonEND (__memchr_aarch64_sve)
6231914882SAlex Richardson
6331914882SAlex Richardson#endif
6431914882SAlex Richardson
65