moved store output

# BB#24:                                # %nujB
                                        #   in Loop: Header=BB257_1 Depth=1
	addl	84(%esi), %edi
	movl	20(%esi), %ebx
	movsd	32(%ebp), %xmm0
	addl	80(%esi), %edx
	movsd	8(%ebp), %xmm1
	addl	76(%esi), %ecx
	addl	72(%esi), %eax
	subsd	8(%ebx,%edi,8), %xmm0
	movl	16(%esi), %edi
	subsd	8(%edi,%edx,8), %xmm1
	movl	12(%esi), %edx
	mulsd	%xmm0, %xmm1
	movsd	56(%ebp), %xmm0
	subsd	8(%edx,%ecx,8), %xmm0
	movl	4(%esi), %ecx
	movl	%ecx, 4(%esp)           # 4-byte Spill
	movl	8(%esi), %edx
	movsd	88(%ebp), %xmm2
	movl	(%ebp), %ebx
	movl	48(%ebp), %ecx
	movl	76(%ebp), %edi
	incl	%edi
	incl	%ecx
	subsd	8(%edx,%eax,8), %xmm2
	movl	24(%ebp), %eax
	movl	72(%ebp), %edx
	movl	%eax, 20(%esp)          # 4-byte Spill
	movl	4(%esp), %eax           # 4-byte Reload
	mulsd	%xmm0, %xmm2
	incl	%edx
	incl	20(%esp)                # 4-byte Folded Spill
	subsd	%xmm1, %xmm2
	movsd	%xmm2, 8(%eax,%ebx,8)
	movl	32(%esp), %eax          # 4-byte Reload
	incl	%ebx
	decl	%eax
	movl	%eax, 96(%ebp)
	movl	%edi, 76(%ebp)
	movl	%edx, 72(%ebp)
	movl	28(%esp), %edx          # 4-byte Reload
	decl	%edx
	movl	%edx, 52(%ebp)
	movl	%ecx, 48(%ebp)
	movl	8(%esp), %ecx           # 4-byte Reload
	movl	20(%esp), %edx          # 4-byte Reload
	decl	%ecx
	movl	%ecx, 28(%ebp)
	movl	24(%esp), %ecx          # 4-byte Reload
	movl	%edx, 24(%ebp)
	decl	%ecx
	movl	%ecx, 4(%ebp)
	movl	%ebx, (%ebp)
	jmp	.LBB257_1