lock-s.diff

9c9
< 	subq	$16, %rsp
---
> 	subq	$32, %rsp
14c14,19
< 	movl	$0, (%rax)
---
> 	movq	%rax, -16(%rbp)
> 	movl	$0, -20(%rbp)
> 	movl	-20(%rbp), %eax
> 	movl	%eax, %edx
> 	movq	-16(%rbp), %rax
> 	movl	%edx, (%rax)
28,33c33,34
< 	movl	$0, -4(%rbp)
< L6:
< 	movq	-24(%rbp), %rax
< 	movl	(%rax), %eax
< 	cmpl	%eax, -4(%rbp)
< 	jne	L6
---
> 	movl	$0, -12(%rbp)
> L7:
35c36,54
< 	movl	$1, (%rax)
---
> 	movq	%rax, -8(%rbp)
> 	movl	$1, -16(%rbp)
> 	movl	-16(%rbp), %eax
> 	movl	%eax, %ecx
> 	movq	-8(%rbp), %rsi
> 	leaq	-12(%rbp), %rdx
> 	movl	(%rdx), %eax
> 	lock; cmpxchgl	%ecx, (%rsi)
> 	movl	%eax, %ecx
> 	sete	%al
> 	testb	%al, %al
> 	jne	L4
> 	movl	%ecx, (%rdx)
> L4:
> 	testb	%al, %al
> 	jne	L10
> 	movl	$0, -12(%rbp)
> 	jmp	L7
> L10:
48c67,72
< 	movq	%rdi, -8(%rbp)
---
> 	movq	%rdi, -24(%rbp)
> 	movq	-24(%rbp), %rax
> 	movq	%rax, -8(%rbp)
> 	movl	$0, -12(%rbp)
> 	movl	-12(%rbp), %eax
> 	movl	%eax, %edx
50c74,76
< 	movl	$0, (%rax)
---
> 	xchgl	(%rax), %edx
> 	movl	%edx, %eax
> 	movl	%eax, -16(%rbp)