.text
	.align	4, 0x90
	.globl	___gmpn_sqr_basecase
	
	
___gmpn_sqr_basecase:

	

	cmp	$2, %rdx
	jae	Lgt1

	mov	(%rsi), %rdx
	.byte	0xc4,226,251,0xf6,210
	mov	%rax, (%rdi)
	mov	%rdx, 8(%rdi)
	
	ret

Lgt1:	jne	Lgt2

	mov	(%rsi), %rdx
	mov	8(%rsi), %rcx
	.byte	0xc4,98,179,0xf6,209	
	.byte	0xc4,98,251,0xf6,194	
	mov	%rcx, %rdx
	.byte	0xc4,226,163,0xf6,210	
	add	%r9, %r9		
	adc	%r10, %r10		
	adc	$0, %rdx		
	add	%r9, %r8		
	adc	%r11, %r10		
	adc	$0, %rdx		
	mov	%rax, (%rdi)
	mov	%r8, 8(%rdi)
	mov	%r10, 16(%rdi)
	mov	%rdx, 24(%rdi)
	
	ret

Lgt2:	cmp	$4, %rdx
	jae	Lgt3

	push	%rbx
	mov	(%rsi), %rdx
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xc4,98,187,0xf6,78,16
	add	%r11, %r8
	mov	8(%rsi), %rdx
	.byte	0xc4,98,251,0xf6,94,16
	adc	%rax, %r9
	adc	$0, %r11
	test	%ebx, %ebx
	mov	(%rsi), %rdx
	.byte	0xc4,226,227,0xf6,202
	mov	%rbx, (%rdi)
	mov	8(%rsi), %rdx
	.byte	0xc4,226,251,0xf6,218
	mov	16(%rsi), %rdx
	.byte	0xc4,226,203,0xf6,210
	.byte	0x66,77,0x0f,0x38,0xf6,210
	.byte	0x66,77,0x0f,0x38,0xf6,192
	.byte	0x66,77,0x0f,0x38,0xf6,201
	.byte	0x66,77,0x0f,0x38,0xf6,219
	.byte	0xf3,73,0x0f,0x38,0xf6,202
	.byte	0xf3,73,0x0f,0x38,0xf6,192
	.byte	0xf3,73,0x0f,0x38,0xf6,217
	.byte	0xf3,73,0x0f,0x38,0xf6,243
	mov	$0, %r8d
	.byte	0xf3,73,0x0f,0x38,0xf6,208
	.byte	0x66,73,0x0f,0x38,0xf6,208
	mov	%rcx, 8(%rdi)
	mov	%rax, 16(%rdi)
	mov	%rbx, 24(%rdi)
	mov	%rsi, 32(%rdi)
	mov	%rdx, 40(%rdi)
	pop	%rbx
	
	ret

Lgt3:	push	%rbx

	lea	-3(%rdx), %ebx
	lea	5(%rdx), %ecx
	mov	%edx, %eax
	and	$-8, %ebx
	shr	$3, %ecx		
	neg	%rbx			
	and	$7, %eax		

	mov	(%rsi), %rdx

	lea	Lmtab(%rip), %r10
	movslq	(%r10,%rax,4), %r8
	lea	(%r8, %r10), %r10
	jmp	*%r10


Lmf0:	.byte	0xc4,98,187,0xf6,202		
	add	%rdx, %rdx
	.byte	0xc4,98,171,0xf6,94,8
	lea	64(%rsi), %rsi
	add	%r9, %r10
	jmp	Lmb0

Lmf3:	.byte	0xc4,98,171,0xf6,218		
	add	%rdx, %rdx
	mov	%r10, (%rdi)
	.byte	0xc4,98,187,0xf6,78,8
	lea	24(%rsi), %rsi
	lea	24(%rdi), %rdi
	add	%r11, %r8
	jmp	Lmb3

Lmf4:	.byte	0xc4,98,187,0xf6,202		
	add	%rdx, %rdx
	.byte	0xc4,98,171,0xf6,94,8
	mov	%r8, (%rdi)
	lea	32(%rsi), %rsi
	lea	32(%rdi), %rdi
	add	%r9, %r10
	jmp	Lmb4

Lmf5:	.byte	0xc4,98,171,0xf6,218		
	add	%rdx, %rdx
	.byte	0xc4,98,187,0xf6,78,8
	mov	%r10, (%rdi)
	lea	40(%rsi), %rsi
	lea	40(%rdi), %rdi
	add	%r11, %r8
	jmp	Lmb5

Lmf6:	.byte	0xc4,98,187,0xf6,202		
	add	%rdx, %rdx
	.byte	0xc4,98,171,0xf6,94,8
	mov	%r8, (%rdi)
	lea	48(%rsi), %rsi
	lea	48(%rdi), %rdi
	add	%r9, %r10
	jmp	Lmb6

Lmf7:	.byte	0xc4,98,171,0xf6,218		
	add	%rdx, %rdx
	.byte	0xc4,98,187,0xf6,78,8
	mov	%r10, (%rdi)
	lea	56(%rsi), %rsi
	lea	56(%rdi), %rdi
	add	%r11, %r8
	jmp	Lmb7

Lmf1:	.byte	0xc4,98,171,0xf6,218		
	add	%rdx, %rdx
	.byte	0xc4,98,187,0xf6,78,8
	mov	%r10, (%rdi)
	lea	8(%rsi), %rsi
	lea	8(%rdi), %rdi
	add	%r11, %r8
	jmp	Lmb1

Lmf2:	.byte	0xc4,98,187,0xf6,202		
	add	%rdx, %rdx
	.byte	0xc4,98,171,0xf6,94,8
	mov	%r8, (%rdi)
	lea	16(%rsi), %rsi
	lea	16(%rdi), %rdi
	dec	%ecx
	add	%r9, %r10
	.byte	0xc4,98,187,0xf6,14

	.align	4, 0x90
Ltop:	mov	%r10, -8(%rdi)
	adc	%r11, %r8
Lmb1:	.byte	0xc4,98,171,0xf6,94,8
	adc	%r9, %r10
	lea	64(%rsi), %rsi
Lmb0:	mov	%r8, (%rdi)
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,187,0xf6,78,208
	lea	64(%rdi), %rdi
	adc	%r11, %r8
Lmb7:	.byte	0xc4,98,171,0xf6,94,216
	mov	%r8, -48(%rdi)
	adc	%r9, %r10
Lmb6:	mov	%r10, -40(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	adc	%r11, %r8
Lmb5:	.byte	0xc4,98,171,0xf6,94,232
	mov	%r8, -32(%rdi)
	adc	%r9, %r10
Lmb4:	.byte	0xc4,98,187,0xf6,78,240
	mov	%r10, -24(%rdi)
	adc	%r11, %r8
Lmb3:	.byte	0xc4,98,171,0xf6,94,248
	adc	%r9, %r10
	mov	%r8, -16(%rdi)
	dec	%ecx
	.byte	0xc4,98,187,0xf6,14
	jnz	Ltop

Lend:	mov	%r10, -8(%rdi)
	adc	%r11, %r8




	lea	Latab(%rip), %r10
	movslq	(%r10,%rax,4), %r11
	lea	(%r11, %r10), %r11

	mov	$63, %eax
	jmp	*%r11

Led0:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf7:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	-64(%rsi,%rbx,8), %rsi
	mov	%ebx, %ecx
	lea	-56(%rdi,%rbx,8), %rdi
	mov	(%rsi), %r9		
	mov	8(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,193
	.byte	0xc4,66,250,0xf7,201
	and	%rdx, %r9			
	.byte	0xc4,98,171,0xf6,218		
	lea	(%r8,%rdx,2), %rdx		
	jmp	Lb7

	.align	4, 0x90
Ltp0:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led0
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
Lb0:	mov	%r8, (%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp0

Led1:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf0:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	-64(%rsi,%rbx,8), %rsi
	mov	%ebx, %ecx
	lea	-56(%rdi,%rbx,8), %rdi
	mov	-8(%rsi), %r11		
	mov	(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,211
	.byte	0xc4,66,250,0xf7,219
	and	%rdx, %r11			
	.byte	0xc4,98,187,0xf6,202		
	lea	(%r10,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	jmp	Lb0

	.align	4, 0x90
Ltp1:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led1
Lb1:	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
	mov	%r8, (%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp1

Led2:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf1:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	(%rsi,%rbx,8), %rsi
	mov	%ebx, %ecx
	lea	8(%rbx), %rbx
	lea	-56(%rdi,%rbx,8), %rdi
	mov	-16(%rsi), %r9		
	mov	-8(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,193
	.byte	0xc4,66,250,0xf7,201
	and	%rdx, %r9			
	.byte	0xc4,98,171,0xf6,218		
	lea	(%r8,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,209			
	.byte	0xc4,98,187,0xf6,14
	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jmp	Lb1

	.align	4, 0x90
Ltp2:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led2
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
	mov	%r8, (%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
Lb2:	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp2

Led3:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf2:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	(%rsi,%rbx,8), %rsi
	or	%ebx, %ecx
	jz	Lcor3
	lea	-56(%rdi,%rbx,8), %rdi
	mov	-24(%rsi), %r11		
	mov	-16(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,211
	.byte	0xc4,66,250,0xf7,219
	and	%rdx, %r11			
	.byte	0xc4,98,187,0xf6,202		
	lea	(%r10,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,195
	jmp	Lb2

	.align	4, 0x90
Ltp3:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led3
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
	mov	%r8, (%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
Lb3:	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp3

Led4:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf3:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	(%rsi,%rbx,8), %rsi
	mov	%ebx, %ecx
	lea	-56(%rdi,%rbx,8), %rdi
	mov	-32(%rsi), %r9		
	mov	-24(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,193
	.byte	0xc4,66,250,0xf7,201
	and	%rdx, %r9			
	.byte	0xc4,98,171,0xf6,218		
	lea	(%r8,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,209
	jmp	Lb3

	.align	4, 0x90
Ltp4:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led4
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
	mov	%r8, (%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
Lb4:	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp4

Led5:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf4:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	(%rsi,%rbx,8), %rsi
	mov	%ebx, %ecx
	lea	-56(%rdi,%rbx,8), %rdi
	mov	-40(%rsi), %r11		
	mov	-32(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,211
	.byte	0xc4,66,250,0xf7,219
	and	%rdx, %r11			
	.byte	0xc4,98,187,0xf6,202		
	lea	(%r10,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,195
	jmp	Lb4

	.align	4, 0x90
Ltp5:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led5
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
	mov	%r8, (%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
Lb5:	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp5

Led6:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf5:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	(%rsi,%rbx,8), %rsi
	mov	%ebx, %ecx
	lea	-56(%rdi,%rbx,8), %rdi
	mov	-48(%rsi), %r9		
	mov	-40(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,193
	.byte	0xc4,66,250,0xf7,201
	and	%rdx, %r9			
	.byte	0xc4,98,171,0xf6,218		
	lea	(%r8,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,209
	jmp	Lb5

	.align	4, 0x90
Ltp6:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led6
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
	mov	%r8, (%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
Lb6:	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp6

Led7:	.byte	0xf3,76,0x0f,0x38,0xf6,7
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
Lf6:	mov	%r8, (%rdi)
	adc	%rcx, %r9		
	mov	%r9, 8(%rdi)
	lea	(%rsi,%rbx,8), %rsi
	mov	%ebx, %ecx
	lea	-56(%rdi,%rbx,8), %rdi
	mov	-56(%rsi), %r11		
	mov	-48(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,211
	.byte	0xc4,66,250,0xf7,219
	and	%rdx, %r11			
	.byte	0xc4,98,187,0xf6,202		
	lea	(%r10,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xc4,98,171,0xf6,94,216
	jmp	Lb6

	.align	4, 0x90
Ltp7:	.byte	0xf3,76,0x0f,0x38,0xf6,87,248
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, -8(%rdi)
	jrcxz	Led7
	.byte	0xc4,98,171,0xf6,94,8
	.byte	0xf3,76,0x0f,0x38,0xf6,7
	lea	8(%rcx), %ecx
	mov	%r8, (%rdi)
Lb7:	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,78,16
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,87,8
	mov	%r10, 8(%rdi)
	.byte	0xc4,98,171,0xf6,94,24
	lea	64(%rsi), %rsi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,16
	mov	%r8, 16(%rdi)
	.byte	0xc4,98,187,0xf6,78,224
	.byte	0xf3,76,0x0f,0x38,0xf6,87,24
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 24(%rdi)
	.byte	0xc4,98,171,0xf6,94,232
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xf3,76,0x0f,0x38,0xf6,71,32
	mov	%r8, 32(%rdi)
	.byte	0xc4,98,187,0xf6,78,240
	.byte	0xf3,76,0x0f,0x38,0xf6,87,40
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 40(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,48
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 48(%rdi)
	lea	64(%rdi), %rdi
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,187,0xf6,14
	jmp	Ltp7

Lcor3:lea	-64(%rdi), %rdi
	mov	-24(%rsi), %r11		
	mov	-16(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,211
	.byte	0xc4,66,250,0xf7,219
	and	%rdx, %r11			
	.byte	0xc4,98,187,0xf6,202		
	lea	(%r10,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,71,56
	.byte	0xc4,98,171,0xf6,94,248
	mov	%r8, 56(%rdi)
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,98,227,0xf6,14
	.byte	0xf3,76,0x0f,0x38,0xf6,87,64
	.byte	0x66,73,0x0f,0x38,0xf6,219
	mov	%r10, 64(%rdi)
	.byte	0xf3,72,0x0f,0x38,0xf6,95,72
	.byte	0xf3,76,0x0f,0x38,0xf6,201		
	adc	%rcx, %r9		
	mov	%r9, 80(%rdi)	

	mov	-16(%rsi), %r9		
	mov	-8(%rsi), %rdx		
	.byte	0xc4,66,251,0xf7,193
	.byte	0xc4,66,250,0xf7,201
	and	%rdx, %r9			
	.byte	0xc4,98,171,0xf6,218		
	lea	(%r8,%rdx,2), %rdx		
	.byte	0x66,77,0x0f,0x38,0xf6,209
	.byte	0xc4,226,187,0xf6,6
	.byte	0xf3,76,0x0f,0x38,0xf6,211
	.byte	0x66,77,0x0f,0x38,0xf6,195
	mov	%r10, 72(%rdi)
	.byte	0xf3,76,0x0f,0x38,0xf6,71,80
	.byte	0xf3,72,0x0f,0x38,0xf6,193		
	mov	%r8, 80(%rdi)
	adc	%rcx, %rax		

	mov	-8(%rsi), %r11		
	mov	(%rsi), %rdx		
	sar	$63, %r11
	and	%rdx, %r11			
	.byte	0xc4,98,187,0xf6,202		
	.byte	0x66,77,0x0f,0x38,0xf6,195
	.byte	0xf3,76,0x0f,0x38,0xf6,192
	mov	%r8, 88(%rdi)
	.byte	0x66,76,0x0f,0x38,0xf6,201
	.byte	0xf3,76,0x0f,0x38,0xf6,201
	mov	%r9, 96(%rdi)

	pop	%rbx
	
	ret

	.text
	.align	3, 0x90
Lmtab:.set	Lmf7_tmp, Lmf7-Lmtab
	.long	Lmf7_tmp

	.set	Lmf0_tmp, Lmf0-Lmtab
	.long	Lmf0_tmp

	.set	Lmf1_tmp, Lmf1-Lmtab
	.long	Lmf1_tmp

	.set	Lmf2_tmp, Lmf2-Lmtab
	.long	Lmf2_tmp

	.set	Lmf3_tmp, Lmf3-Lmtab
	.long	Lmf3_tmp

	.set	Lmf4_tmp, Lmf4-Lmtab
	.long	Lmf4_tmp

	.set	Lmf5_tmp, Lmf5-Lmtab
	.long	Lmf5_tmp

	.set	Lmf6_tmp, Lmf6-Lmtab
	.long	Lmf6_tmp

Latab:.set	Lf6_tmp, Lf6-Latab
	.long	Lf6_tmp

	.set	Lf7_tmp, Lf7-Latab
	.long	Lf7_tmp

	.set	Lf0_tmp, Lf0-Latab
	.long	Lf0_tmp

	.set	Lf1_tmp, Lf1-Latab
	.long	Lf1_tmp

	.set	Lf2_tmp, Lf2-Latab
	.long	Lf2_tmp

	.set	Lf3_tmp, Lf3-Latab
	.long	Lf3_tmp

	.set	Lf4_tmp, Lf4-Latab
	.long	Lf4_tmp

	.set	Lf5_tmp, Lf5-Latab
	.long	Lf5_tmp

	.text