Before this patch, instructions such as div, mul, shifts with count
in CL, cmpxchg are mishandled.

This patch adds vex prefix handling. In particular, it avoids colliding
with register operand encoded in vex.vvvv field.

Since we need to avoid two possible register operands, the selection of
scratch register needs to be from at least three registers.

After looking through a lot of CPU docs, it looks like the safest choice
is SI,DI,BX. Selecting BX needs care to not collide with implicit use of
BX by cmpxchg8b.

Test-case:

	#include <stdio.h>

	static const char *const pass[] = { "FAIL", "pass" };

	long two = 2;
	void test1(void)
	{
		long ax = 0, dx = 0;
		asm volatile("\n"
	"			xor	%%edx,%%edx\n"
	"			lea	2(%%edx),%%eax\n"
	// We divide 2 by 2. Result (in eax) should be 1:
	"	probe1:		.globl	probe1\n"
	"			divl	two(%%rip)\n"
	// If we have a bug (eax mangled on entry) the result will be 2,
	// because eax gets restored by probe machinery.
		: "=a" (ax), "=d" (dx) /*out*/
		: "0" (ax), "1" (dx) /*in*/
		: "memory" /*clobber*/
		);
		dprintf(2, "%s: %s\n", __func__,
			pass[ax == 1]
		);
	}

	long val2 = 0;
	void test2(void)
	{
		long old_val = val2;
		long ax = 0, dx = 0;
		asm volatile("\n"
	"			mov	val2,%%eax\n"     // eax := val2
	"			lea	1(%%eax),%%edx\n" // edx := eax+1
	// eax is equal to val2. cmpxchg should store edx to val2:
	"	probe2:		.globl  probe2\n"
	"			cmpxchg %%edx,val2(%%rip)\n"
	// If we have a bug (eax mangled on entry), val2 will stay unchanged
		: "=a" (ax), "=d" (dx) /*out*/
		: "0" (ax), "1" (dx) /*in*/
		: "memory" /*clobber*/
		);
		dprintf(2, "%s: %s\n", __func__,
			pass[val2 == old_val + 1]
		);
	}

	long val3[2] = {0,0};
	void test3(void)
	{
		long old_val = val3[0];
		long ax = 0, dx = 0;
		asm volatile("\n"
	"			mov	val3,%%eax\n"  // edx:eax := val3
	"			mov	val3+4,%%edx\n"
	"			mov	%%eax,%%ebx\n" // ecx:ebx := edx:eax + 1
	"			mov	%%edx,%%ecx\n"
	"			add	$1,%%ebx\n"
	"			adc	$0,%%ecx\n"
	// edx:eax is equal to val3. cmpxchg8b should store ecx:ebx to val3:
	"	probe3:		.globl  probe3\n"
	"			cmpxchg8b val3(%%rip)\n"
	// If we have a bug (edx:eax mangled on entry), val3 will stay unchanged.
	// If ecx:edx in mangled, val3 will get wrong value.
		: "=a" (ax), "=d" (dx) /*out*/
		: "0" (ax), "1" (dx) /*in*/
		: "cx", "bx", "memory" /*clobber*/
		);
		dprintf(2, "%s: %s\n", __func__,
			pass[val3[0] == old_val + 1 && val3[1] == 0]
		);
	}

	int main(int argc, char **argv)
	{
		test1();
		test2();
		test3();
		return 0;
	}

Before this change all tests fail if probe{1,2,3} are probed.
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Reviewed-by: Jim Keniston <jkenisto@us.ibm.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>