#if defined(lint) || defined(__lint)

#include "arcfour.h"

/* ARGSUSED */
void
arcfour_crypt_asm(ARCFour_key *key, uchar_t *in, uchar_t *out, size_t len)
{}

/* ARGSUSED */
void
arcfour_key_init(ARCFour_key *key, uchar_t *keyval, int keyvallen)
{}

#else
#include <sys/asm_linkage.h>

ENTRY_NP(arcfour_crypt_asm)
	or	%rcx,%rcx # If (len == 0) return
	jne	.Lentry
	ret
.Lentry:
	push	%r12
	push	%r13

	/ Set %rdi to beginning of array, key->arr[0]
	add	$8,%rdi
	/ Get key->j
	movl	-8(%rdi),%r8d
	/ Get key->i
	movl	-4(%rdi),%r12d

	/
	/ Use a 4-byte key schedule element array
	/
	inc	%r8b
	movl	(%rdi,%r8,4),%r9d
	test	$-8,%rcx
	jz	.Lloop1
	jmp	.Lloop8

.align	16
.Lloop8:
	add	%r9b,%r12b
	mov	%r8,%r10
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 0=0
	inc	%r10b
	movl	(%rdi,%r10,4),%r11d
	cmp	%r10,%r12
	movl	%r9d,(%rdi,%r12,4)
	cmove	%r9,%r11
	movl	%r13d,(%rdi,%r8,4)
	add	%r9b,%r13b
	movb	(%rdi,%r13,4),%al
	add	%r11b,%r12b
	mov	%r10,%r8
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 1=0
	inc	%r8b
	movl	(%rdi,%r8,4),%r9d
	cmp	%r8,%r12
	movl	%r11d,(%rdi,%r12,4)
	cmove	%r11,%r9
	movl	%r13d,(%rdi,%r10,4)
	add	%r11b,%r13b
	movb	(%rdi,%r13,4),%al
	add	%r9b,%r12b
	mov	%r8,%r10
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 2=0
	inc	%r10b
	movl	(%rdi,%r10,4),%r11d
	cmp	%r10,%r12
	movl	%r9d,(%rdi,%r12,4)
	cmove	%r9,%r11
	movl	%r13d,(%rdi,%r8,4)
	add	%r9b,%r13b
	movb	(%rdi,%r13,4),%al
	add	%r11b,%r12b
	mov	%r10,%r8
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 3=0
	inc	%r8b
	movl	(%rdi,%r8,4),%r9d
	cmp	%r8,%r12
	movl	%r11d,(%rdi,%r12,4)
	cmove	%r11,%r9
	movl	%r13d,(%rdi,%r10,4)
	add	%r11b,%r13b
	movb	(%rdi,%r13,4),%al
	add	%r9b,%r12b
	mov	%r8,%r10
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 4=0
	inc	%r10b
	movl	(%rdi,%r10,4),%r11d
	cmp	%r10,%r12
	movl	%r9d,(%rdi,%r12,4)
	cmove	%r9,%r11
	movl	%r13d,(%rdi,%r8,4)
	add	%r9b,%r13b
	movb	(%rdi,%r13,4),%al
	add	%r11b,%r12b
	mov	%r10,%r8
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 5=0
	inc	%r8b
	movl	(%rdi,%r8,4),%r9d
	cmp	%r8,%r12
	movl	%r11d,(%rdi,%r12,4)
	cmove	%r11,%r9
	movl	%r13d,(%rdi,%r10,4)
	add	%r11b,%r13b
	movb	(%rdi,%r13,4),%al
	add	%r9b,%r12b
	mov	%r8,%r10
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 6=0
	inc	%r10b
	movl	(%rdi,%r10,4),%r11d
	cmp	%r10,%r12
	movl	%r9d,(%rdi,%r12,4)
	cmove	%r9,%r11
	movl	%r13d,(%rdi,%r8,4)
	add	%r9b,%r13b
	movb	(%rdi,%r13,4),%al
	add	%r11b,%r12b
	mov	%r10,%r8
	movl	(%rdi,%r12,4),%r13d
	ror	$8,%rax			# ror is redundant when 7=0
	inc	%r8b
	movl	(%rdi,%r8,4),%r9d
	cmp	%r8,%r12
	movl	%r11d,(%rdi,%r12,4)
	cmove	%r11,%r9
	movl	%r13d,(%rdi,%r10,4)
	add	%r11b,%r13b
	movb	(%rdi,%r13,4),%al
	ror	$8,%rax
	sub	$8,%rcx

	xor	(%rsi),%rax
	add	$8,%rsi
	mov	%rax,(%rdx)
	add	$8,%rdx

	test	$-8,%rcx
	jnz	.Lloop8
	cmp	$0,%rcx
	jne	.Lloop1

.Lexit:
	/
	/ Cleanup and exit code
	/
	/ --i to undo ++i done at entry
	sub	$1,%r8b
	/ set key->i
	movl	%r8d,-8(%rdi)
	/ set key->j
	movl	%r12d,-4(%rdi)

	pop	%r13
	pop	%r12
	ret

.align	16
.Lloop1:
	add	%r9b,%r12b
	movl	(%rdi,%r12,4),%r13d
	movl	%r9d,(%rdi,%r12,4)
	movl	%r13d,(%rdi,%r8,4)
	add	%r13b,%r9b
	inc	%r8b
	movl	(%rdi,%r9,4),%r13d
	movl	(%rdi,%r8,4),%r9d
	xorb	(%rsi),%r13b
	inc	%rsi
	movb	%r13b,(%rdx)
	inc	%rdx
	dec	%rcx
	jnz	.Lloop1
	jmp	.Lexit

	ret
SET_SIZE(arcfour_crypt_asm)
	/ int arcfour_crypt_on_intel(void);
.extern	arcfour_crypt_on_intel

ENTRY_NP(arcfour_key_init)
	/ Find out if we're running on Intel or something else (e.g., AMD64).
	/ This sets %eax to 1 for Intel, otherwise 0.
	push	%rdi		/ Save arg1
	push	%rsi		/ Save arg2
	push	%rdx		/ Save arg3
	call	arcfour_crypt_on_intel
	pop	%rdx		/ Restore arg3
	pop	%rsi		/ Restore arg2
	pop	%rdi		/ Restore arg1
	/ Save return value in key->flag (1=Intel, 0=AMD)
	movl	%eax,1032(%rdi)

	/ Set %rdi to beginning of array, key->arr[0]
	lea	8(%rdi),%rdi
	lea	(%rsi,%rdx),%rsi
	neg	%rdx
	mov	%rdx,%rcx

	xor	%eax,%eax
	xor	%r9,%r9
	xor	%r10,%r10
	xor	%r11,%r11

	/ Use a 4-byte data array
	jmp	.Lw1stloop

.align	16
.Lw1stloop:
	/ AMD64 (4-byte array)
	mov	%eax,(%rdi,%rax,4)
	add	$1,%al
	jnc	.Lw1stloop

	xor	%r9,%r9
	xor	%r8,%r8

.align	16
.Lw2ndloop:
	mov	(%rdi,%r9,4),%r10d
	add	(%rsi,%rdx,1),%r8b
	add	%r10b,%r8b
	add	$1,%rdx
	mov	(%rdi,%r8,4),%r11d
	cmovz	%rcx,%rdx
	mov	%r10d,(%rdi,%r8,4)
	mov	%r11d,(%rdi,%r9,4)
	add	$1,%r9b
	jnc	.Lw2ndloop

	/ Exit code
	xor	%eax,%eax
	mov	%eax,-8(%rdi)
	mov	%eax,-4(%rdi)

	ret
SET_SIZE(arcfour_key_init)
.asciz	"RC4 for x86_64, CRYPTOGAMS by <appro@openssl.org>"
#endif /* !lint && !__lint */