RT，今天尝试手写 uint128_t，在测试除移位以外的运算符都与系统自带的汇编代码一模一样，但在写左移和右移时发现无论什么我可以想到的写法都不能避免分支预测，无法使用 cmovne 指令。

struct _U {
	unsigned long long r, l;
	_U operator=(unsigned long long v) {
		r = v, l = 0;
		return *this;
	}
	_U operator<<(unsigned int amount) {
		const auto lo = r, hi = l;
		return
			amount < 64	? _U{lo << amount, (hi << amount) | (lo >> (64 - amount))}
							: _U{0, lo << amount - 64};
	}
};
_U x1, x2;
unsigned long long v;
int main() {
	__builtin_scanf("%llu", &v); x1 = v;
	__builtin_scanf("%llu", &v);
	x2 = x1 << v;
}

#define _U __int128 unsigned
_U x1, x2;
unsigned long long v;
int main() {
	__builtin_scanf("%llu", &v); x1 = v;
	__builtin_scanf("%llu", &v);
	x2 = x1 << v;
}

这两份代码在计算左移时的输出是

# 自定义 uint128_t
	mov	rdx, QWORD PTR v[rip]
	mov	rax, QWORD PTR x1[rip]
	mov	r8, QWORD PTR x1[rip+8]
	cmp	edx, 63
	jbe	.L5
	lea	ecx, -64[rdx]
	xor	r9d, r9d
	sal	rax, cl
.L3:
	mov	QWORD PTR x2[rip+8], rax
	xor	eax, eax
	mov	QWORD PTR x2[rip], r9
	add	rsp, 40
	ret
.L5:
	mov	r9, rax
	mov	ecx, edx
	sal	r9, cl
	mov	ecx, 64
	sub	ecx, edx
	shr	rax, cl
	mov	ecx, edx
	sal	r8, cl
	or	rax, r8
	jmp	.L3

# 自带的 uint128_t
	movzx	ecx, BYTE PTR v[rip]
	mov	rax, QWORD PTR x1[rip]
	xor	r8d, r8d
	mov	rdx, QWORD PTR x1[rip+8]
	shld	rdx, rax, cl
	sal	rax, cl
	test	cl, 64
	cmovne	rdx, rax
	cmovne	rax, r8
	mov	QWORD PTR x2[rip], rax
	xor	eax, eax
	mov	QWORD PTR x2[rip+8], rdx
	add	rsp, 40
	ret

求问怎样的写法才可以与系统自带的 uint128_t 输出大致相同？（开 -O2）