/*-
 * Copyright (c) 2023, The FreeBSD Foundation
 *
 * SPDX-License-Expression: BSD-2-Clause
 *
 * Portions of this software were developed by Robert Clausecker
 * <fuz@FreeBSD.org> under sponsorship from the FreeBSD Foundation.
 *
 * Adapted from NetBSD's common/lib/libc/arch/x86_64/string/strcat.S
 * written by J.T. Conklin <jtc@acorntoolworks.com>
 * that was originally dedicated to the public domain
 */

#include <machine/asm.h>
#if 0
	RCSID("$NetBSD: strcat.S,v 1.4 2004/07/26 18:51:21 drochner Exp $")
#endif

#include "amd64_archlevel.h"

ARCHFUNCS(strcat)
	ARCHFUNC(strcat, scalar)
	ARCHFUNC(strcat, baseline)
ENDARCHFUNCS(strcat)

ARCHENTRY(strcat, scalar)
	movq	%rdi,%rax
	movabsq	$0x0101010101010101,%r8
	movabsq	$0x8080808080808080,%r9

	/*
	 * Align destination to word boundary.
	 * Consider unrolling loop?
	 */
.Lscan:
.Lscan_align:
	testb	$7,%dil
	je	.Lscan_aligned
	cmpb	$0,(%rdi)
	je	.Lcopy
	incq	%rdi
	jmp	.Lscan_align

	.align	4
.Lscan_aligned:
.Lscan_loop:
	movq	(%rdi),%rdx
	addq	$8,%rdi
	subq	%r8,%rdx
	testq	%r9,%rdx
	je	.Lscan_loop

	/*
	 * In rare cases, the above loop may exit prematurely. We must
	 * return to the loop if none of the bytes in the word equal 0.
	 */

	cmpb	$0,-8(%rdi)	/* 1st byte == 0? */
	jne	1f
	subq	$8,%rdi
	jmp	.Lcopy

1:	cmpb	$0,-7(%rdi)	/* 2nd byte == 0? */
	jne	1f
	subq	$7,%rdi
	jmp	.Lcopy

1:	cmpb	$0,-6(%rdi)	/* 3rd byte == 0? */
	jne	1f
	subq	$6,%rdi
	jmp	.Lcopy

1:	cmpb	$0,-5(%rdi)	/* 4th byte == 0? */
	jne	1f
	subq	$5,%rdi
	jmp	.Lcopy

1:	cmpb	$0,-4(%rdi)	/* 5th byte == 0? */
	jne	1f
	subq	$4,%rdi
	jmp	.Lcopy

1:	cmpb	$0,-3(%rdi)	/* 6th byte == 0? */
	jne	1f
	subq	$3,%rdi
	jmp	.Lcopy

1:	cmpb	$0,-2(%rdi)	/* 7th byte == 0? */
	jne	1f
	subq	$2,%rdi
	jmp	.Lcopy

1:	cmpb	$0,-1(%rdi)	/* 8th byte == 0? */
	jne	.Lscan_loop
	subq	$1,%rdi

	/*
	 * Align source to a word boundary.
	 * Consider unrolling loop?
	 */
.Lcopy:
.Lcopy_align:
	testb	$7,%sil
	je	.Lcopy_aligned
	movb	(%rsi),%dl
	incq	%rsi
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl
	jne	.Lcopy_align
	ret

	.align	4
.Lcopy_loop:
	movq	%rdx,(%rdi)
	addq	$8,%rdi
.Lcopy_aligned:
	movq	(%rsi),%rdx
	movq	%rdx,%rcx
	addq	$8,%rsi
	subq	%r8,%rcx
	testq	%r9,%rcx
	je	.Lcopy_loop

	/*
	 * In rare cases, the above loop may exit prematurely. We must
	 * return to the loop if none of the bytes in the word equal 0.
	 */

	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 1st byte == 0? */
	je	.Ldone

	shrq	$8,%rdx
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 2nd byte == 0? */
	je	.Ldone

	shrq	$8,%rdx
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 3rd byte == 0? */
	je	.Ldone

	shrq	$8,%rdx
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 4th byte == 0? */
	je	.Ldone

	shrq	$8,%rdx
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 5th byte == 0? */
	je	.Ldone

	shrq	$8,%rdx
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 6th byte == 0? */
	je	.Ldone

	shrq	$8,%rdx
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 7th byte == 0? */
	je	.Ldone

	shrq	$8,%rdx
	movb	%dl,(%rdi)
	incq	%rdi
	testb	%dl,%dl		/* 8th byte == 0? */
	jne	.Lcopy_aligned

.Ldone:
	ret
ARCHEND(strcat, scalar)

/*
 * Call into strlen + strcpy if we have any SIMD at all.
 * The scalar implementation above is better for the scalar
 * case as it avoids the function call overhead, but pessimal
 * if we could call SIMD routines instead.
 */
ARCHENTRY(strcat, baseline)
	push	%rbp
	mov	%rsp, %rbp
	push	%rsi
	push	%rbx
	mov	%rdi, %rbx		# remember destination for later
	call	CNAME(strlen)		# strlen(dest)
	mov	-8(%rbp), %rsi
	lea	(%rbx, %rax, 1), %rdi	# dest + strlen(dest)
	call	CNAME(__stpcpy)		# stpcpy(dest + strlen(dest), src)
	mov	%rbx, %rax		# return dest
	pop	%rbx
	leave
	ret
ARCHEND(strcat, baseline)

	.section .note.GNU-stack,"",%progbits