xref: /netbsd-src/lib/libnvmm/libnvmm_x86.c (revision e6b71c148711fd8b0d6ef07729e79923e9aa810f)

/*	$NetBSD: libnvmm_x86.c,v 1.43 2020/12/27 20:56:14 reinoud Exp $	*/

/*
 * Copyright (c) 2018-2020 Maxime Villard, m00nbsd.net
 * All rights reserved.
 *
 * This code is part of the NVMM hypervisor.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 */

#include <sys/cdefs.h>

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <errno.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <machine/vmparam.h>
#include <machine/pte.h>
#include <machine/psl.h>

#define MIN(X, Y) (((X) < (Y)) ? (X) : (Y))
#define __cacheline_aligned __attribute__((__aligned__(64)))

#include <x86/specialreg.h>

/* -------------------------------------------------------------------------- */

/*
 * Undocumented debugging function. Helpful.
 */
int
nvmm_vcpu_dump(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu)
{
	struct nvmm_x64_state *state = vcpu->state;
	uint16_t *attr;
	size_t i;
	int ret;

	const char *segnames[] = {
		"ES", "CS", "SS", "DS", "FS", "GS", "GDT", "IDT", "LDT", "TR"
	};

	ret = nvmm_vcpu_getstate(mach, vcpu, NVMM_X64_STATE_ALL);
	if (ret == -1)
		return -1;

	printf("+ VCPU id=%d\n", (int)vcpu->cpuid);
	printf("| -> RAX=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RAX]);
	printf("| -> RCX=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RCX]);
	printf("| -> RDX=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RDX]);
	printf("| -> RBX=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RBX]);
	printf("| -> RSP=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RSP]);
	printf("| -> RBP=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RBP]);
	printf("| -> RSI=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RSI]);
	printf("| -> RDI=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RDI]);
	printf("| -> RIP=%"PRIx64"\n", state->gprs[NVMM_X64_GPR_RIP]);
	printf("| -> RFLAGS=%p\n", (void *)state->gprs[NVMM_X64_GPR_RFLAGS]);
	for (i = 0; i < NVMM_X64_NSEG; i++) {
		attr = (uint16_t *)&state->segs[i].attrib;
		printf("| -> %s: sel=0x%x base=%"PRIx64", limit=%x, "
		    "attrib=%x [type=%d,l=%d,def=%d]\n",
		    segnames[i],
		    state->segs[i].selector,
		    state->segs[i].base,
		    state->segs[i].limit,
		    *attr,
		    state->segs[i].attrib.type,
		    state->segs[i].attrib.l,
		    state->segs[i].attrib.def);
	}
	printf("| -> MSR_EFER=%"PRIx64"\n", state->msrs[NVMM_X64_MSR_EFER]);
	printf("| -> CR0=%"PRIx64"\n", state->crs[NVMM_X64_CR_CR0]);
	printf("| -> CR3=%"PRIx64"\n", state->crs[NVMM_X64_CR_CR3]);
	printf("| -> CR4=%"PRIx64"\n", state->crs[NVMM_X64_CR_CR4]);
	printf("| -> CR8=%"PRIx64"\n", state->crs[NVMM_X64_CR_CR8]);

	return 0;
}

/* -------------------------------------------------------------------------- */

#define PTE32_L1_SHIFT	12
#define PTE32_L2_SHIFT	22

#define PTE32_L2_MASK	0xffc00000
#define PTE32_L1_MASK	0x003ff000

#define PTE32_L2_FRAME	(PTE32_L2_MASK)
#define PTE32_L1_FRAME	(PTE32_L2_FRAME|PTE32_L1_MASK)

#define pte32_l1idx(va)	(((va) & PTE32_L1_MASK) >> PTE32_L1_SHIFT)
#define pte32_l2idx(va)	(((va) & PTE32_L2_MASK) >> PTE32_L2_SHIFT)

#define CR3_FRAME_32BIT	__BITS(31, 12)

typedef uint32_t pte_32bit_t;

static int
x86_gva_to_gpa_32bit(struct nvmm_machine *mach, uint64_t cr3,
    gvaddr_t gva, gpaddr_t *gpa, bool has_pse, nvmm_prot_t *prot)
{
	gpaddr_t L2gpa, L1gpa;
	uintptr_t L2hva, L1hva;
	pte_32bit_t *pdir, pte;
	nvmm_prot_t pageprot;

	/* We begin with an RWXU access. */
	*prot = NVMM_PROT_ALL;

	/* Parse L2. */
	L2gpa = (cr3 & CR3_FRAME_32BIT);
	if (nvmm_gpa_to_hva(mach, L2gpa, &L2hva, &pageprot) == -1)
		return -1;
	pdir = (pte_32bit_t *)L2hva;
	pte = pdir[pte32_l2idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if ((pte & PTE_PS) && !has_pse)
		return -1;
	if (pte & PTE_PS) {
		*gpa = (pte & PTE32_L2_FRAME);
		*gpa = *gpa + (gva & PTE32_L1_MASK);
		return 0;
	}

	/* Parse L1. */
	L1gpa = (pte & PTE_FRAME);
	if (nvmm_gpa_to_hva(mach, L1gpa, &L1hva, &pageprot) == -1)
		return -1;
	pdir = (pte_32bit_t *)L1hva;
	pte = pdir[pte32_l1idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if (pte & PTE_PS)
		return -1;

	*gpa = (pte & PTE_FRAME);
	return 0;
}

/* -------------------------------------------------------------------------- */

#define	PTE32_PAE_L1_SHIFT	12
#define	PTE32_PAE_L2_SHIFT	21
#define	PTE32_PAE_L3_SHIFT	30

#define	PTE32_PAE_L3_MASK	0xc0000000
#define	PTE32_PAE_L2_MASK	0x3fe00000
#define	PTE32_PAE_L1_MASK	0x001ff000

#define	PTE32_PAE_L3_FRAME	(PTE32_PAE_L3_MASK)
#define	PTE32_PAE_L2_FRAME	(PTE32_PAE_L3_FRAME|PTE32_PAE_L2_MASK)
#define	PTE32_PAE_L1_FRAME	(PTE32_PAE_L2_FRAME|PTE32_PAE_L1_MASK)

#define pte32_pae_l1idx(va)	(((va) & PTE32_PAE_L1_MASK) >> PTE32_PAE_L1_SHIFT)
#define pte32_pae_l2idx(va)	(((va) & PTE32_PAE_L2_MASK) >> PTE32_PAE_L2_SHIFT)
#define pte32_pae_l3idx(va)	(((va) & PTE32_PAE_L3_MASK) >> PTE32_PAE_L3_SHIFT)

#define CR3_FRAME_32BIT_PAE	__BITS(31, 5)

typedef uint64_t pte_32bit_pae_t;

static int
x86_gva_to_gpa_32bit_pae(struct nvmm_machine *mach, uint64_t cr3,
    gvaddr_t gva, gpaddr_t *gpa, nvmm_prot_t *prot)
{
	gpaddr_t L3gpa, L2gpa, L1gpa;
	uintptr_t L3hva, L2hva, L1hva;
	pte_32bit_pae_t *pdir, pte;
	nvmm_prot_t pageprot;

	/* We begin with an RWXU access. */
	*prot = NVMM_PROT_ALL;

	/* Parse L3. */
	L3gpa = (cr3 & CR3_FRAME_32BIT_PAE);
	if (nvmm_gpa_to_hva(mach, L3gpa, &L3hva, &pageprot) == -1)
		return -1;
	pdir = (pte_32bit_pae_t *)L3hva;
	pte = pdir[pte32_pae_l3idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if (pte & PTE_NX)
		*prot &= ~NVMM_PROT_EXEC;
	if (pte & PTE_PS)
		return -1;

	/* Parse L2. */
	L2gpa = (pte & PTE_FRAME);
	if (nvmm_gpa_to_hva(mach, L2gpa, &L2hva, &pageprot) == -1)
		return -1;
	pdir = (pte_32bit_pae_t *)L2hva;
	pte = pdir[pte32_pae_l2idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if (pte & PTE_NX)
		*prot &= ~NVMM_PROT_EXEC;
	if (pte & PTE_PS) {
		*gpa = (pte & PTE32_PAE_L2_FRAME);
		*gpa = *gpa + (gva & PTE32_PAE_L1_MASK);
		return 0;
	}

	/* Parse L1. */
	L1gpa = (pte & PTE_FRAME);
	if (nvmm_gpa_to_hva(mach, L1gpa, &L1hva, &pageprot) == -1)
		return -1;
	pdir = (pte_32bit_pae_t *)L1hva;
	pte = pdir[pte32_pae_l1idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if (pte & PTE_NX)
		*prot &= ~NVMM_PROT_EXEC;
	if (pte & PTE_PS)
		return -1;

	*gpa = (pte & PTE_FRAME);
	return 0;
}

/* -------------------------------------------------------------------------- */

#define PTE64_L1_SHIFT	12
#define PTE64_L2_SHIFT	21
#define PTE64_L3_SHIFT	30
#define PTE64_L4_SHIFT	39

#define PTE64_L4_MASK	0x0000ff8000000000
#define PTE64_L3_MASK	0x0000007fc0000000
#define PTE64_L2_MASK	0x000000003fe00000
#define PTE64_L1_MASK	0x00000000001ff000

#define PTE64_L4_FRAME	PTE64_L4_MASK
#define PTE64_L3_FRAME	(PTE64_L4_FRAME|PTE64_L3_MASK)
#define PTE64_L2_FRAME	(PTE64_L3_FRAME|PTE64_L2_MASK)
#define PTE64_L1_FRAME	(PTE64_L2_FRAME|PTE64_L1_MASK)

#define pte64_l1idx(va)	(((va) & PTE64_L1_MASK) >> PTE64_L1_SHIFT)
#define pte64_l2idx(va)	(((va) & PTE64_L2_MASK) >> PTE64_L2_SHIFT)
#define pte64_l3idx(va)	(((va) & PTE64_L3_MASK) >> PTE64_L3_SHIFT)
#define pte64_l4idx(va)	(((va) & PTE64_L4_MASK) >> PTE64_L4_SHIFT)

#define CR3_FRAME_64BIT	__BITS(51, 12)

typedef uint64_t pte_64bit_t;

static inline bool
x86_gva_64bit_canonical(gvaddr_t gva)
{
	/* Bits 63:47 must have the same value. */
#define SIGN_EXTEND	0xffff800000000000ULL
	return (gva & SIGN_EXTEND) == 0 || (gva & SIGN_EXTEND) == SIGN_EXTEND;
}

static int
x86_gva_to_gpa_64bit(struct nvmm_machine *mach, uint64_t cr3,
    gvaddr_t gva, gpaddr_t *gpa, nvmm_prot_t *prot)
{
	gpaddr_t L4gpa, L3gpa, L2gpa, L1gpa;
	uintptr_t L4hva, L3hva, L2hva, L1hva;
	pte_64bit_t *pdir, pte;
	nvmm_prot_t pageprot;

	/* We begin with an RWXU access. */
	*prot = NVMM_PROT_ALL;

	if (!x86_gva_64bit_canonical(gva))
		return -1;

	/* Parse L4. */
	L4gpa = (cr3 & CR3_FRAME_64BIT);
	if (nvmm_gpa_to_hva(mach, L4gpa, &L4hva, &pageprot) == -1)
		return -1;
	pdir = (pte_64bit_t *)L4hva;
	pte = pdir[pte64_l4idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if (pte & PTE_NX)
		*prot &= ~NVMM_PROT_EXEC;
	if (pte & PTE_PS)
		return -1;

	/* Parse L3. */
	L3gpa = (pte & PTE_FRAME);
	if (nvmm_gpa_to_hva(mach, L3gpa, &L3hva, &pageprot) == -1)
		return -1;
	pdir = (pte_64bit_t *)L3hva;
	pte = pdir[pte64_l3idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if (pte & PTE_NX)
		*prot &= ~NVMM_PROT_EXEC;
	if (pte & PTE_PS) {
		*gpa = (pte & PTE64_L3_FRAME);
		*gpa = *gpa + (gva & (PTE64_L2_MASK|PTE64_L1_MASK));
		return 0;
	}

	/* Parse L2. */
	L2gpa = (pte & PTE_FRAME);
	if (nvmm_gpa_to_hva(mach, L2gpa, &L2hva, &pageprot) == -1)
		return -1;
	pdir = (pte_64bit_t *)L2hva;
	pte = pdir[pte64_l2idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if (pte & PTE_NX)
		*prot &= ~NVMM_PROT_EXEC;
	if (pte & PTE_PS) {
		*gpa = (pte & PTE64_L2_FRAME);
		*gpa = *gpa + (gva & PTE64_L1_MASK);
		return 0;
	}

	/* Parse L1. */
	L1gpa = (pte & PTE_FRAME);
	if (nvmm_gpa_to_hva(mach, L1gpa, &L1hva, &pageprot) == -1)
		return -1;
	pdir = (pte_64bit_t *)L1hva;
	pte = pdir[pte64_l1idx(gva)];
	if ((pte & PTE_P) == 0)
		return -1;
	if ((pte & PTE_U) == 0)
		*prot &= ~NVMM_PROT_USER;
	if ((pte & PTE_W) == 0)
		*prot &= ~NVMM_PROT_WRITE;
	if (pte & PTE_NX)
		*prot &= ~NVMM_PROT_EXEC;
	if (pte & PTE_PS)
		return -1;

	*gpa = (pte & PTE_FRAME);
	return 0;
}

static inline int
x86_gva_to_gpa(struct nvmm_machine *mach, struct nvmm_x64_state *state,
    gvaddr_t gva, gpaddr_t *gpa, nvmm_prot_t *prot)
{
	bool is_pae, is_lng, has_pse;
	uint64_t cr3;
	size_t off;
	int ret;

	if ((state->crs[NVMM_X64_CR_CR0] & CR0_PG) == 0) {
		/* No paging. */
		*prot = NVMM_PROT_ALL;
		*gpa = gva;
		return 0;
	}

	off = (gva & PAGE_MASK);
	gva &= ~PAGE_MASK;

	is_pae = (state->crs[NVMM_X64_CR_CR4] & CR4_PAE) != 0;
	is_lng = (state->msrs[NVMM_X64_MSR_EFER] & EFER_LMA) != 0;
	has_pse = (state->crs[NVMM_X64_CR_CR4] & CR4_PSE) != 0;
	cr3 = state->crs[NVMM_X64_CR_CR3];

	if (is_pae && is_lng) {
		/* 64bit */
		ret = x86_gva_to_gpa_64bit(mach, cr3, gva, gpa, prot);
	} else if (is_pae && !is_lng) {
		/* 32bit PAE */
		ret = x86_gva_to_gpa_32bit_pae(mach, cr3, gva, gpa, prot);
	} else if (!is_pae && !is_lng) {
		/* 32bit */
		ret = x86_gva_to_gpa_32bit(mach, cr3, gva, gpa, has_pse, prot);
	} else {
		ret = -1;
	}

	if (ret == -1) {
		errno = EFAULT;
	}

	*gpa = *gpa + off;

	return ret;
}

int
nvmm_gva_to_gpa(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    gvaddr_t gva, gpaddr_t *gpa, nvmm_prot_t *prot)
{
	struct nvmm_x64_state *state = vcpu->state;
	int ret;

	ret = nvmm_vcpu_getstate(mach, vcpu,
	    NVMM_X64_STATE_CRS | NVMM_X64_STATE_MSRS);
	if (ret == -1)
		return -1;

	return x86_gva_to_gpa(mach, state, gva, gpa, prot);
}

/* -------------------------------------------------------------------------- */

#define DISASSEMBLER_BUG()	\
	do {			\
		errno = EINVAL;	\
		return -1;	\
	} while (0);

static inline bool
is_long_mode(struct nvmm_x64_state *state)
{
	return (state->msrs[NVMM_X64_MSR_EFER] & EFER_LMA) != 0;
}

static inline bool
is_64bit(struct nvmm_x64_state *state)
{
	return (state->segs[NVMM_X64_SEG_CS].attrib.l != 0);
}

static inline bool
is_32bit(struct nvmm_x64_state *state)
{
	return (state->segs[NVMM_X64_SEG_CS].attrib.l == 0) &&
	    (state->segs[NVMM_X64_SEG_CS].attrib.def == 1);
}

static inline bool
is_16bit(struct nvmm_x64_state *state)
{
	return (state->segs[NVMM_X64_SEG_CS].attrib.l == 0) &&
	    (state->segs[NVMM_X64_SEG_CS].attrib.def == 0);
}

static int
segment_check(struct nvmm_x64_state_seg *seg, gvaddr_t gva, size_t size)
{
	uint64_t limit;

	/*
	 * This is incomplete. We should check topdown, etc, really that's
	 * tiring.
	 */
	if (__predict_false(!seg->attrib.p)) {
		goto error;
	}

	limit = (uint64_t)seg->limit + 1;
	if (__predict_true(seg->attrib.g)) {
		limit *= PAGE_SIZE;
	}

	if (__predict_false(gva + size > limit)) {
		goto error;
	}

	return 0;

error:
	errno = EFAULT;
	return -1;
}

static inline void
segment_apply(struct nvmm_x64_state_seg *seg, gvaddr_t *gva)
{
	*gva += seg->base;
}

static inline uint64_t
size_to_mask(size_t size)
{
	switch (size) {
	case 1:
		return 0x00000000000000FF;
	case 2:
		return 0x000000000000FFFF;
	case 4:
		return 0x00000000FFFFFFFF;
	case 8:
	default:
		return 0xFFFFFFFFFFFFFFFF;
	}
}

static uint64_t
rep_get_cnt(struct nvmm_x64_state *state, size_t adsize)
{
	uint64_t mask, cnt;

	mask = size_to_mask(adsize);
	cnt = state->gprs[NVMM_X64_GPR_RCX] & mask;

	return cnt;
}

static void
rep_set_cnt(struct nvmm_x64_state *state, size_t adsize, uint64_t cnt)
{
	uint64_t mask;

	/* XXX: should we zero-extend? */
	mask = size_to_mask(adsize);
	state->gprs[NVMM_X64_GPR_RCX] &= ~mask;
	state->gprs[NVMM_X64_GPR_RCX] |= cnt;
}

static int
read_guest_memory(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    gvaddr_t gva, uint8_t *data, size_t size)
{
	struct nvmm_x64_state *state = vcpu->state;
	struct nvmm_mem mem;
	nvmm_prot_t prot;
	gpaddr_t gpa;
	uintptr_t hva;
	bool is_mmio;
	int ret, remain;

	ret = x86_gva_to_gpa(mach, state, gva, &gpa, &prot);
	if (__predict_false(ret == -1)) {
		return -1;
	}
	if (__predict_false(!(prot & NVMM_PROT_READ))) {
		errno = EFAULT;
		return -1;
	}

	if ((gva & PAGE_MASK) + size > PAGE_SIZE) {
		remain = ((gva & PAGE_MASK) + size - PAGE_SIZE);
	} else {
		remain = 0;
	}
	size -= remain;

	ret = nvmm_gpa_to_hva(mach, gpa, &hva, &prot);
	is_mmio = (ret == -1);

	if (is_mmio) {
		mem.mach = mach;
		mem.vcpu = vcpu;
		mem.data = data;
		mem.gpa = gpa;
		mem.write = false;
		mem.size = size;
		(*vcpu->cbs.mem)(&mem);
	} else {
		if (__predict_false(!(prot & NVMM_PROT_READ))) {
			errno = EFAULT;
			return -1;
		}
		memcpy(data, (uint8_t *)hva, size);
	}

	if (remain > 0) {
		ret = read_guest_memory(mach, vcpu, gva + size,
		    data + size, remain);
	} else {
		ret = 0;
	}

	return ret;
}

static int
write_guest_memory(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    gvaddr_t gva, uint8_t *data, size_t size)
{
	struct nvmm_x64_state *state = vcpu->state;
	struct nvmm_mem mem;
	nvmm_prot_t prot;
	gpaddr_t gpa;
	uintptr_t hva;
	bool is_mmio;
	int ret, remain;

	ret = x86_gva_to_gpa(mach, state, gva, &gpa, &prot);
	if (__predict_false(ret == -1)) {
		return -1;
	}
	if (__predict_false(!(prot & NVMM_PROT_WRITE))) {
		errno = EFAULT;
		return -1;
	}

	if ((gva & PAGE_MASK) + size > PAGE_SIZE) {
		remain = ((gva & PAGE_MASK) + size - PAGE_SIZE);
	} else {
		remain = 0;
	}
	size -= remain;

	ret = nvmm_gpa_to_hva(mach, gpa, &hva, &prot);
	is_mmio = (ret == -1);

	if (is_mmio) {
		mem.mach = mach;
		mem.vcpu = vcpu;
		mem.data = data;
		mem.gpa = gpa;
		mem.write = true;
		mem.size = size;
		(*vcpu->cbs.mem)(&mem);
	} else {
		if (__predict_false(!(prot & NVMM_PROT_WRITE))) {
			errno = EFAULT;
			return -1;
		}
		memcpy((uint8_t *)hva, data, size);
	}

	if (remain > 0) {
		ret = write_guest_memory(mach, vcpu, gva + size,
		    data + size, remain);
	} else {
		ret = 0;
	}

	return ret;
}

/* -------------------------------------------------------------------------- */

static int fetch_segment(struct nvmm_machine *, struct nvmm_vcpu *);

#define NVMM_IO_BATCH_SIZE	32

static int
assist_io_batch(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    struct nvmm_io *io, gvaddr_t gva, uint64_t cnt)
{
	uint8_t iobuf[NVMM_IO_BATCH_SIZE];
	size_t i, iosize, iocnt;
	int ret;

	cnt = MIN(cnt, NVMM_IO_BATCH_SIZE);
	iosize = MIN(io->size * cnt, NVMM_IO_BATCH_SIZE);
	iocnt = iosize / io->size;

	io->data = iobuf;

	if (!io->in) {
		ret = read_guest_memory(mach, vcpu, gva, iobuf, iosize);
		if (ret == -1)
			return -1;
	}

	for (i = 0; i < iocnt; i++) {
		(*vcpu->cbs.io)(io);
		io->data += io->size;
	}

	if (io->in) {
		ret = write_guest_memory(mach, vcpu, gva, iobuf, iosize);
		if (ret == -1)
			return -1;
	}

	return iocnt;
}

int
nvmm_assist_io(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu)
{
	struct nvmm_x64_state *state = vcpu->state;
	struct nvmm_vcpu_exit *exit = vcpu->exit;
	struct nvmm_io io;
	uint64_t cnt = 0; /* GCC */
	uint8_t iobuf[8];
	int iocnt = 1;
	gvaddr_t gva = 0; /* GCC */
	int reg = 0; /* GCC */
	int ret, seg;
	bool psld = false;

	if (__predict_false(exit->reason != NVMM_VCPU_EXIT_IO)) {
		errno = EINVAL;
		return -1;
	}

	io.mach = mach;
	io.vcpu = vcpu;
	io.port = exit->u.io.port;
	io.in = exit->u.io.in;
	io.size = exit->u.io.operand_size;
	io.data = iobuf;

	ret = nvmm_vcpu_getstate(mach, vcpu,
	    NVMM_X64_STATE_GPRS | NVMM_X64_STATE_SEGS |
	    NVMM_X64_STATE_CRS | NVMM_X64_STATE_MSRS);
	if (ret == -1)
		return -1;

	if (exit->u.io.rep) {
		cnt = rep_get_cnt(state, exit->u.io.address_size);
		if (__predict_false(cnt == 0)) {
			state->gprs[NVMM_X64_GPR_RIP] = exit->u.io.npc;
			goto out;
		}
	}

	if (__predict_false(state->gprs[NVMM_X64_GPR_RFLAGS] & PSL_D)) {
		psld = true;
	}

	/*
	 * Determine GVA.
	 */
	if (exit->u.io.str) {
		if (io.in) {
			reg = NVMM_X64_GPR_RDI;
		} else {
			reg = NVMM_X64_GPR_RSI;
		}

		gva = state->gprs[reg];
		gva &= size_to_mask(exit->u.io.address_size);

		if (exit->u.io.seg != -1) {
			seg = exit->u.io.seg;
		} else {
			if (io.in) {
				seg = NVMM_X64_SEG_ES;
			} else {
				seg = fetch_segment(mach, vcpu);
				if (seg == -1)
					return -1;
			}
		}

		if (__predict_true(is_long_mode(state))) {
			if (seg == NVMM_X64_SEG_GS || seg == NVMM_X64_SEG_FS) {
				segment_apply(&state->segs[seg], &gva);
			}
		} else {
			ret = segment_check(&state->segs[seg], gva, io.size);
			if (ret == -1)
				return -1;
			segment_apply(&state->segs[seg], &gva);
		}

		if (exit->u.io.rep && !psld) {
			iocnt = assist_io_batch(mach, vcpu, &io, gva, cnt);
			if (iocnt == -1)
				return -1;
			goto done;
		}
	}

	if (!io.in) {
		if (!exit->u.io.str) {
			memcpy(io.data, &state->gprs[NVMM_X64_GPR_RAX], io.size);
		} else {
			ret = read_guest_memory(mach, vcpu, gva, io.data,
			    io.size);
			if (ret == -1)
				return -1;
		}
	}

	(*vcpu->cbs.io)(&io);

	if (io.in) {
		if (!exit->u.io.str) {
			memcpy(&state->gprs[NVMM_X64_GPR_RAX], io.data, io.size);
			if (io.size == 4) {
				/* Zero-extend to 64 bits. */
				state->gprs[NVMM_X64_GPR_RAX] &= size_to_mask(4);
			}
		} else {
			ret = write_guest_memory(mach, vcpu, gva, io.data,
			    io.size);
			if (ret == -1)
				return -1;
		}
	}

done:
	if (exit->u.io.str) {
		if (__predict_false(psld)) {
			state->gprs[reg] -= iocnt * io.size;
		} else {
			state->gprs[reg] += iocnt * io.size;
		}
	}

	if (exit->u.io.rep) {
		cnt -= iocnt;
		rep_set_cnt(state, exit->u.io.address_size, cnt);
		if (cnt == 0) {
			state->gprs[NVMM_X64_GPR_RIP] = exit->u.io.npc;
		}
	} else {
		state->gprs[NVMM_X64_GPR_RIP] = exit->u.io.npc;
	}

out:
	ret = nvmm_vcpu_setstate(mach, vcpu, NVMM_X64_STATE_GPRS);
	if (ret == -1)
		return -1;

	return 0;
}

/* -------------------------------------------------------------------------- */

struct x86_emul {
	bool readreg;
	bool backprop;
	bool notouch;
	void (*func)(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
};

static void x86_func_or(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_and(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_xchg(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_sub(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_xor(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_cmp(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_test(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_mov(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_stos(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);
static void x86_func_lods(struct nvmm_vcpu *, struct nvmm_mem *, uint64_t *);

static const struct x86_emul x86_emul_or = {
	.readreg = true,
	.func = x86_func_or
};

static const struct x86_emul x86_emul_and = {
	.readreg = true,
	.func = x86_func_and
};

static const struct x86_emul x86_emul_xchg = {
	.readreg = true,
	.backprop = true,
	.func = x86_func_xchg
};

static const struct x86_emul x86_emul_sub = {
	.readreg = true,
	.func = x86_func_sub
};

static const struct x86_emul x86_emul_xor = {
	.readreg = true,
	.func = x86_func_xor
};

static const struct x86_emul x86_emul_cmp = {
	.notouch = true,
	.func = x86_func_cmp
};

static const struct x86_emul x86_emul_test = {
	.notouch = true,
	.func = x86_func_test
};

static const struct x86_emul x86_emul_mov = {
	.func = x86_func_mov
};

static const struct x86_emul x86_emul_stos = {
	.func = x86_func_stos
};

static const struct x86_emul x86_emul_lods = {
	.func = x86_func_lods
};

/* Legacy prefixes. */
#define LEG_LOCK	0xF0
#define LEG_REPN	0xF2
#define LEG_REP		0xF3
#define LEG_OVR_CS	0x2E
#define LEG_OVR_SS	0x36
#define LEG_OVR_DS	0x3E
#define LEG_OVR_ES	0x26
#define LEG_OVR_FS	0x64
#define LEG_OVR_GS	0x65
#define LEG_OPR_OVR	0x66
#define LEG_ADR_OVR	0x67

struct x86_legpref {
	bool opr_ovr:1;
	bool adr_ovr:1;
	bool rep:1;
	bool repn:1;
	bool repe:1;
	int8_t seg;
};

struct x86_rexpref {
	bool b:1;
	bool x:1;
	bool r:1;
	bool w:1;
	bool present:1;
};

struct x86_reg {
	int num;	/* NVMM GPR state index */
	uint64_t mask;
};

struct x86_dualreg {
	int reg1;
	int reg2;
};

enum x86_disp_type {
	DISP_NONE,
	DISP_0,
	DISP_1,
	DISP_2,
	DISP_4
};

struct x86_disp {
	enum x86_disp_type type;
	uint64_t data; /* 4 bytes, but can be sign-extended */
};

struct x86_regmodrm {
	uint8_t mod:2;
	uint8_t reg:3;
	uint8_t rm:3;
};

struct x86_immediate {
	uint64_t data;
};

struct x86_sib {
	uint8_t scale;
	const struct x86_reg *idx;
	const struct x86_reg *bas;
};

enum x86_store_type {
	STORE_NONE,
	STORE_REG,
	STORE_DUALREG,
	STORE_IMM,
	STORE_SIB,
	STORE_DMO
};

struct x86_store {
	enum x86_store_type type;
	union {
		const struct x86_reg *reg;
		struct x86_dualreg dualreg;
		struct x86_immediate imm;
		struct x86_sib sib;
		uint64_t dmo;
	} u;
	struct x86_disp disp;
	int hardseg;
};

struct x86_instr {
	uint8_t len;
	struct x86_legpref legpref;
	struct x86_rexpref rexpref;
	struct x86_regmodrm regmodrm;
	uint8_t operand_size;
	uint8_t address_size;
	uint64_t zeroextend_mask;

	const struct x86_opcode *opcode;
	const struct x86_emul *emul;

	struct x86_store src;
	struct x86_store dst;
	struct x86_store *strm;
};

struct x86_decode_fsm {
	/* vcpu */
	bool is64bit;
	bool is32bit;
	bool is16bit;

	/* fsm */
	int (*fn)(struct x86_decode_fsm *, struct x86_instr *);
	uint8_t *buf;
	uint8_t *end;
};

struct x86_opcode {
	bool valid:1;
	bool regmodrm:1;
	bool regtorm:1;
	bool dmo:1;
	bool todmo:1;
	bool movs:1;
	bool cmps:1;
	bool stos:1;
	bool lods:1;
	bool szoverride:1;
	bool group1:1;
	bool group3:1;
	bool group11:1;
	bool immediate:1;
	uint8_t defsize;
	uint8_t flags;
	const struct x86_emul *emul;
};

struct x86_group_entry {
	const struct x86_emul *emul;
};

#define OPSIZE_BYTE 0x01
#define OPSIZE_WORD 0x02 /* 2 bytes */
#define OPSIZE_DOUB 0x04 /* 4 bytes */
#define OPSIZE_QUAD 0x08 /* 8 bytes */

#define FLAG_imm8	0x01
#define FLAG_immz	0x02
#define FLAG_ze		0x04

static const struct x86_group_entry group1[8] __cacheline_aligned = {
	[1] = { .emul = &x86_emul_or },
	[4] = { .emul = &x86_emul_and },
	[6] = { .emul = &x86_emul_xor },
	[7] = { .emul = &x86_emul_cmp }
};

static const struct x86_group_entry group3[8] __cacheline_aligned = {
	[0] = { .emul = &x86_emul_test },
	[1] = { .emul = &x86_emul_test }
};

static const struct x86_group_entry group11[8] __cacheline_aligned = {
	[0] = { .emul = &x86_emul_mov }
};

static const struct x86_opcode primary_opcode_table[256] __cacheline_aligned = {
	/*
	 * Group1
	 */
	[0x80] = {
		/* Eb, Ib */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.group1 = true,
		.immediate = true,
		.emul = NULL /* group1 */
	},
	[0x81] = {
		/* Ev, Iz */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.group1 = true,
		.immediate = true,
		.flags = FLAG_immz,
		.emul = NULL /* group1 */
	},
	[0x83] = {
		/* Ev, Ib */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.group1 = true,
		.immediate = true,
		.flags = FLAG_imm8,
		.emul = NULL /* group1 */
	},

	/*
	 * Group3
	 */
	[0xF6] = {
		/* Eb, Ib */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.group3 = true,
		.immediate = true,
		.emul = NULL /* group3 */
	},
	[0xF7] = {
		/* Ev, Iz */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.group3 = true,
		.immediate = true,
		.flags = FLAG_immz,
		.emul = NULL /* group3 */
	},

	/*
	 * Group11
	 */
	[0xC6] = {
		/* Eb, Ib */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.group11 = true,
		.immediate = true,
		.emul = NULL /* group11 */
	},
	[0xC7] = {
		/* Ev, Iz */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.group11 = true,
		.immediate = true,
		.flags = FLAG_immz,
		.emul = NULL /* group11 */
	},

	/*
	 * OR
	 */
	[0x08] = {
		/* Eb, Gb */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_or
	},
	[0x09] = {
		/* Ev, Gv */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_or
	},
	[0x0A] = {
		/* Gb, Eb */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_or
	},
	[0x0B] = {
		/* Gv, Ev */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_or
	},

	/*
	 * AND
	 */
	[0x20] = {
		/* Eb, Gb */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_and
	},
	[0x21] = {
		/* Ev, Gv */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_and
	},
	[0x22] = {
		/* Gb, Eb */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_and
	},
	[0x23] = {
		/* Gv, Ev */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_and
	},

	/*
	 * SUB
	 */
	[0x28] = {
		/* Eb, Gb */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_sub
	},
	[0x29] = {
		/* Ev, Gv */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_sub
	},
	[0x2A] = {
		/* Gb, Eb */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_sub
	},
	[0x2B] = {
		/* Gv, Ev */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_sub
	},

	/*
	 * XOR
	 */
	[0x30] = {
		/* Eb, Gb */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_xor
	},
	[0x31] = {
		/* Ev, Gv */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_xor
	},
	[0x32] = {
		/* Gb, Eb */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_xor
	},
	[0x33] = {
		/* Gv, Ev */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_xor
	},

	/*
	 * XCHG
	 */
	[0x86] = {
		/* Eb, Gb */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_xchg
	},
	[0x87] = {
		/* Ev, Gv */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_xchg
	},

	/*
	 * MOV
	 */
	[0x88] = {
		/* Eb, Gb */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_mov
	},
	[0x89] = {
		/* Ev, Gv */
		.valid = true,
		.regmodrm = true,
		.regtorm = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_mov
	},
	[0x8A] = {
		/* Gb, Eb */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_mov
	},
	[0x8B] = {
		/* Gv, Ev */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_mov
	},
	[0xA0] = {
		/* AL, Ob */
		.valid = true,
		.dmo = true,
		.todmo = false,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_mov
	},
	[0xA1] = {
		/* rAX, Ov */
		.valid = true,
		.dmo = true,
		.todmo = false,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_mov
	},
	[0xA2] = {
		/* Ob, AL */
		.valid = true,
		.dmo = true,
		.todmo = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_mov
	},
	[0xA3] = {
		/* Ov, rAX */
		.valid = true,
		.dmo = true,
		.todmo = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_mov
	},

	/*
	 * MOVS
	 */
	[0xA4] = {
		/* Yb, Xb */
		.valid = true,
		.movs = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = NULL
	},
	[0xA5] = {
		/* Yv, Xv */
		.valid = true,
		.movs = true,
		.szoverride = true,
		.defsize = -1,
		.emul = NULL
	},

	/*
	 * CMPS
	 */
	[0xA6] = {
		/* Yb, Xb */
		.valid = true,
		.cmps = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = NULL
	},
	[0xA7] = {
		/* Yv, Xv */
		.valid = true,
		.cmps = true,
		.szoverride = true,
		.defsize = -1,
		.emul = NULL
	},

	/*
	 * STOS
	 */
	[0xAA] = {
		/* Yb, AL */
		.valid = true,
		.stos = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_stos
	},
	[0xAB] = {
		/* Yv, rAX */
		.valid = true,
		.stos = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_stos
	},

	/*
	 * LODS
	 */
	[0xAC] = {
		/* AL, Xb */
		.valid = true,
		.lods = true,
		.szoverride = false,
		.defsize = OPSIZE_BYTE,
		.emul = &x86_emul_lods
	},
	[0xAD] = {
		/* rAX, Xv */
		.valid = true,
		.lods = true,
		.szoverride = true,
		.defsize = -1,
		.emul = &x86_emul_lods
	},
};

static const struct x86_opcode secondary_opcode_table[256] __cacheline_aligned = {
	/*
	 * MOVZX
	 */
	[0xB6] = {
		/* Gv, Eb */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = true,
		.defsize = OPSIZE_BYTE,
		.flags = FLAG_ze,
		.emul = &x86_emul_mov
	},
	[0xB7] = {
		/* Gv, Ew */
		.valid = true,
		.regmodrm = true,
		.regtorm = false,
		.szoverride = true,
		.defsize = OPSIZE_WORD,
		.flags = FLAG_ze,
		.emul = &x86_emul_mov
	},
};

static const struct x86_reg gpr_map__rip = { NVMM_X64_GPR_RIP, 0xFFFFFFFFFFFFFFFF };

/* [REX-present][enc][opsize] */
static const struct x86_reg gpr_map__special[2][4][8] __cacheline_aligned = {
	[false] = {
		/* No REX prefix. */
		[0b00] = {
			[0] = { NVMM_X64_GPR_RAX, 0x000000000000FF00 }, /* AH */
			[1] = { NVMM_X64_GPR_RSP, 0x000000000000FFFF }, /* SP */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RSP, 0x00000000FFFFFFFF }, /* ESP */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 },
		},
		[0b01] = {
			[0] = { NVMM_X64_GPR_RCX, 0x000000000000FF00 }, /* CH */
			[1] = { NVMM_X64_GPR_RBP, 0x000000000000FFFF }, /* BP */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RBP, 0x00000000FFFFFFFF },	/* EBP */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 },
		},
		[0b10] = {
			[0] = { NVMM_X64_GPR_RDX, 0x000000000000FF00 }, /* DH */
			[1] = { NVMM_X64_GPR_RSI, 0x000000000000FFFF }, /* SI */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RSI, 0x00000000FFFFFFFF }, /* ESI */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 },
		},
		[0b11] = {
			[0] = { NVMM_X64_GPR_RBX, 0x000000000000FF00 }, /* BH */
			[1] = { NVMM_X64_GPR_RDI, 0x000000000000FFFF }, /* DI */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RDI, 0x00000000FFFFFFFF }, /* EDI */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 },
		}
	},
	[true] = {
		/* Has REX prefix. */
		[0b00] = {
			[0] = { NVMM_X64_GPR_RSP, 0x00000000000000FF }, /* SPL */
			[1] = { NVMM_X64_GPR_RSP, 0x000000000000FFFF }, /* SP */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RSP, 0x00000000FFFFFFFF }, /* ESP */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RSP, 0xFFFFFFFFFFFFFFFF }, /* RSP */
		},
		[0b01] = {
			[0] = { NVMM_X64_GPR_RBP, 0x00000000000000FF }, /* BPL */
			[1] = { NVMM_X64_GPR_RBP, 0x000000000000FFFF }, /* BP */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RBP, 0x00000000FFFFFFFF }, /* EBP */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RBP, 0xFFFFFFFFFFFFFFFF }, /* RBP */
		},
		[0b10] = {
			[0] = { NVMM_X64_GPR_RSI, 0x00000000000000FF }, /* SIL */
			[1] = { NVMM_X64_GPR_RSI, 0x000000000000FFFF }, /* SI */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RSI, 0x00000000FFFFFFFF }, /* ESI */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RSI, 0xFFFFFFFFFFFFFFFF }, /* RSI */
		},
		[0b11] = {
			[0] = { NVMM_X64_GPR_RDI, 0x00000000000000FF }, /* DIL */
			[1] = { NVMM_X64_GPR_RDI, 0x000000000000FFFF }, /* DI */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RDI, 0x00000000FFFFFFFF }, /* EDI */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RDI, 0xFFFFFFFFFFFFFFFF }, /* RDI */
		}
	}
};

/* [depends][enc][size] */
static const struct x86_reg gpr_map[2][8][8] __cacheline_aligned = {
	[false] = {
		/* Not extended. */
		[0b000] = {
			[0] = { NVMM_X64_GPR_RAX, 0x00000000000000FF }, /* AL */
			[1] = { NVMM_X64_GPR_RAX, 0x000000000000FFFF }, /* AX */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RAX, 0x00000000FFFFFFFF }, /* EAX */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RAX, 0xFFFFFFFFFFFFFFFF }, /* RAX */
		},
		[0b001] = {
			[0] = { NVMM_X64_GPR_RCX, 0x00000000000000FF }, /* CL */
			[1] = { NVMM_X64_GPR_RCX, 0x000000000000FFFF }, /* CX */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RCX, 0x00000000FFFFFFFF }, /* ECX */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RCX, 0xFFFFFFFFFFFFFFFF }, /* RCX */
		},
		[0b010] = {
			[0] = { NVMM_X64_GPR_RDX, 0x00000000000000FF }, /* DL */
			[1] = { NVMM_X64_GPR_RDX, 0x000000000000FFFF }, /* DX */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RDX, 0x00000000FFFFFFFF }, /* EDX */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RDX, 0xFFFFFFFFFFFFFFFF }, /* RDX */
		},
		[0b011] = {
			[0] = { NVMM_X64_GPR_RBX, 0x00000000000000FF }, /* BL */
			[1] = { NVMM_X64_GPR_RBX, 0x000000000000FFFF }, /* BX */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_RBX, 0x00000000FFFFFFFF }, /* EBX */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_RBX, 0xFFFFFFFFFFFFFFFF }, /* RBX */
		},
		[0b100] = {
			[0] = { -1, 0 }, /* SPECIAL */
			[1] = { -1, 0 }, /* SPECIAL */
			[2] = { -1, 0 },
			[3] = { -1, 0 }, /* SPECIAL */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 }, /* SPECIAL */
		},
		[0b101] = {
			[0] = { -1, 0 }, /* SPECIAL */
			[1] = { -1, 0 }, /* SPECIAL */
			[2] = { -1, 0 },
			[3] = { -1, 0 }, /* SPECIAL */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 }, /* SPECIAL */
		},
		[0b110] = {
			[0] = { -1, 0 }, /* SPECIAL */
			[1] = { -1, 0 }, /* SPECIAL */
			[2] = { -1, 0 },
			[3] = { -1, 0 }, /* SPECIAL */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 }, /* SPECIAL */
		},
		[0b111] = {
			[0] = { -1, 0 }, /* SPECIAL */
			[1] = { -1, 0 }, /* SPECIAL */
			[2] = { -1, 0 },
			[3] = { -1, 0 }, /* SPECIAL */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { -1, 0 }, /* SPECIAL */
		},
	},
	[true] = {
		/* Extended. */
		[0b000] = {
			[0] = { NVMM_X64_GPR_R8, 0x00000000000000FF }, /* R8B */
			[1] = { NVMM_X64_GPR_R8, 0x000000000000FFFF }, /* R8W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R8, 0x00000000FFFFFFFF }, /* R8D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R8, 0xFFFFFFFFFFFFFFFF }, /* R8 */
		},
		[0b001] = {
			[0] = { NVMM_X64_GPR_R9, 0x00000000000000FF }, /* R9B */
			[1] = { NVMM_X64_GPR_R9, 0x000000000000FFFF }, /* R9W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R9, 0x00000000FFFFFFFF }, /* R9D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R9, 0xFFFFFFFFFFFFFFFF }, /* R9 */
		},
		[0b010] = {
			[0] = { NVMM_X64_GPR_R10, 0x00000000000000FF }, /* R10B */
			[1] = { NVMM_X64_GPR_R10, 0x000000000000FFFF }, /* R10W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R10, 0x00000000FFFFFFFF }, /* R10D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R10, 0xFFFFFFFFFFFFFFFF }, /* R10 */
		},
		[0b011] = {
			[0] = { NVMM_X64_GPR_R11, 0x00000000000000FF }, /* R11B */
			[1] = { NVMM_X64_GPR_R11, 0x000000000000FFFF }, /* R11W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R11, 0x00000000FFFFFFFF }, /* R11D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R11, 0xFFFFFFFFFFFFFFFF }, /* R11 */
		},
		[0b100] = {
			[0] = { NVMM_X64_GPR_R12, 0x00000000000000FF }, /* R12B */
			[1] = { NVMM_X64_GPR_R12, 0x000000000000FFFF }, /* R12W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R12, 0x00000000FFFFFFFF }, /* R12D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R12, 0xFFFFFFFFFFFFFFFF }, /* R12 */
		},
		[0b101] = {
			[0] = { NVMM_X64_GPR_R13, 0x00000000000000FF }, /* R13B */
			[1] = { NVMM_X64_GPR_R13, 0x000000000000FFFF }, /* R13W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R13, 0x00000000FFFFFFFF }, /* R13D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R13, 0xFFFFFFFFFFFFFFFF }, /* R13 */
		},
		[0b110] = {
			[0] = { NVMM_X64_GPR_R14, 0x00000000000000FF }, /* R14B */
			[1] = { NVMM_X64_GPR_R14, 0x000000000000FFFF }, /* R14W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R14, 0x00000000FFFFFFFF }, /* R14D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R14, 0xFFFFFFFFFFFFFFFF }, /* R14 */
		},
		[0b111] = {
			[0] = { NVMM_X64_GPR_R15, 0x00000000000000FF }, /* R15B */
			[1] = { NVMM_X64_GPR_R15, 0x000000000000FFFF }, /* R15W */
			[2] = { -1, 0 },
			[3] = { NVMM_X64_GPR_R15, 0x00000000FFFFFFFF }, /* R15D */
			[4] = { -1, 0 },
			[5] = { -1, 0 },
			[6] = { -1, 0 },
			[7] = { NVMM_X64_GPR_R15, 0xFFFFFFFFFFFFFFFF }, /* R15 */
		},
	}
};

/* [enc] */
static const int gpr_dual_reg1_rm[8] __cacheline_aligned = {
	[0b000] = NVMM_X64_GPR_RBX, /* BX (+SI) */
	[0b001] = NVMM_X64_GPR_RBX, /* BX (+DI) */
	[0b010] = NVMM_X64_GPR_RBP, /* BP (+SI) */
	[0b011] = NVMM_X64_GPR_RBP, /* BP (+DI) */
	[0b100] = NVMM_X64_GPR_RSI, /* SI */
	[0b101] = NVMM_X64_GPR_RDI, /* DI */
	[0b110] = NVMM_X64_GPR_RBP, /* BP */
	[0b111] = NVMM_X64_GPR_RBX, /* BX */
};

static int
node_overflow(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	fsm->fn = NULL;
	return -1;
}

static int
fsm_read(struct x86_decode_fsm *fsm, uint8_t *bytes, size_t n)
{
	if (fsm->buf + n > fsm->end) {
		return -1;
	}
	memcpy(bytes, fsm->buf, n);
	return 0;
}

static inline void
fsm_advance(struct x86_decode_fsm *fsm, size_t n,
    int (*fn)(struct x86_decode_fsm *, struct x86_instr *))
{
	fsm->buf += n;
	if (fsm->buf > fsm->end) {
		fsm->fn = node_overflow;
	} else {
		fsm->fn = fn;
	}
}

static const struct x86_reg *
resolve_special_register(struct x86_instr *instr, uint8_t enc, size_t regsize)
{
	enc &= 0b11;
	if (regsize == 8) {
		/* May be 64bit without REX */
		return &gpr_map__special[1][enc][regsize-1];
	}
	return &gpr_map__special[instr->rexpref.present][enc][regsize-1];
}

/*
 * Special node, for MOVS. Fake two displacements of zero on the source and
 * destination registers.
 */
static int
node_movs(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	size_t adrsize;

	adrsize = instr->address_size;

	/* DS:RSI */
	instr->src.type = STORE_REG;
	instr->src.u.reg = &gpr_map__special[1][2][adrsize-1];
	instr->src.disp.type = DISP_0;

	/* ES:RDI, force ES */
	instr->dst.type = STORE_REG;
	instr->dst.u.reg = &gpr_map__special[1][3][adrsize-1];
	instr->dst.disp.type = DISP_0;
	instr->dst.hardseg = NVMM_X64_SEG_ES;

	fsm_advance(fsm, 0, NULL);

	return 0;
}

/*
 * Special node, for CMPS. Fake two displacements of zero on the source and
 * destination registers.
 * XXX coded as clone of movs as its similar in register usage
 * XXX might be merged with node_movs()
 */
static int
node_cmps(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	size_t adrsize;

	adrsize = instr->address_size;

	/* DS:RSI */
	instr->src.type = STORE_REG;
	instr->src.u.reg = &gpr_map__special[1][2][adrsize-1];
	instr->src.disp.type = DISP_0;

	/* ES:RDI, force ES */
	instr->dst.type = STORE_REG;
	instr->dst.u.reg = &gpr_map__special[1][3][adrsize-1];
	instr->dst.disp.type = DISP_0;
	instr->dst.hardseg = NVMM_X64_SEG_ES;

	fsm_advance(fsm, 0, NULL);

	return 0;
}

/*
 * Special node, for STOS and LODS. Fake a displacement of zero on the
 * destination register.
 */
static int
node_stlo(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode = instr->opcode;
	struct x86_store *stlo, *streg;
	size_t adrsize, regsize;

	adrsize = instr->address_size;
	regsize = instr->operand_size;

	if (opcode->stos) {
		streg = &instr->src;
		stlo = &instr->dst;
	} else {
		streg = &instr->dst;
		stlo = &instr->src;
	}

	streg->type = STORE_REG;
	streg->u.reg = &gpr_map[0][0][regsize-1]; /* ?AX */

	stlo->type = STORE_REG;
	if (opcode->stos) {
		/* ES:RDI, force ES */
		stlo->u.reg = &gpr_map__special[1][3][adrsize-1];
		stlo->hardseg = NVMM_X64_SEG_ES;
	} else {
		/* DS:RSI */
		stlo->u.reg = &gpr_map__special[1][2][adrsize-1];
	}
	stlo->disp.type = DISP_0;

	fsm_advance(fsm, 0, NULL);

	return 0;
}

static int
node_dmo(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode = instr->opcode;
	struct x86_store *stdmo, *streg;
	size_t adrsize, regsize;

	adrsize = instr->address_size;
	regsize = instr->operand_size;

	if (opcode->todmo) {
		streg = &instr->src;
		stdmo = &instr->dst;
	} else {
		streg = &instr->dst;
		stdmo = &instr->src;
	}

	streg->type = STORE_REG;
	streg->u.reg = &gpr_map[0][0][regsize-1]; /* ?AX */

	stdmo->type = STORE_DMO;
	if (fsm_read(fsm, (uint8_t *)&stdmo->u.dmo, adrsize) == -1) {
		return -1;
	}
	fsm_advance(fsm, adrsize, NULL);

	return 0;
}

static inline uint64_t
sign_extend(uint64_t val, int size)
{
	if (size == 1) {
		if (val & __BIT(7))
			val |= 0xFFFFFFFFFFFFFF00;
	} else if (size == 2) {
		if (val & __BIT(15))
			val |= 0xFFFFFFFFFFFF0000;
	} else if (size == 4) {
		if (val & __BIT(31))
			val |= 0xFFFFFFFF00000000;
	}
	return val;
}

static int
node_immediate(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode = instr->opcode;
	struct x86_store *store;
	uint8_t immsize;
	size_t sesize = 0;

	/* The immediate is the source */
	store = &instr->src;
	immsize = instr->operand_size;

	if (opcode->flags & FLAG_imm8) {
		sesize = immsize;
		immsize = 1;
	} else if ((opcode->flags & FLAG_immz) && (immsize == 8)) {
		sesize = immsize;
		immsize = 4;
	}

	store->type = STORE_IMM;
	if (fsm_read(fsm, (uint8_t *)&store->u.imm.data, immsize) == -1) {
		return -1;
	}
	fsm_advance(fsm, immsize, NULL);

	if (sesize != 0) {
		store->u.imm.data = sign_extend(store->u.imm.data, sesize);
	}

	return 0;
}

static int
node_disp(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode = instr->opcode;
	uint64_t data = 0;
	size_t n;

	if (instr->strm->disp.type == DISP_1) {
		n = 1;
	} else if (instr->strm->disp.type == DISP_2) {
		n = 2;
	} else if (instr->strm->disp.type == DISP_4) {
		n = 4;
	} else {
		DISASSEMBLER_BUG();
	}

	if (fsm_read(fsm, (uint8_t *)&data, n) == -1) {
		return -1;
	}

	if (__predict_true(fsm->is64bit)) {
		data = sign_extend(data, n);
	}

	instr->strm->disp.data = data;

	if (opcode->immediate) {
		fsm_advance(fsm, n, node_immediate);
	} else {
		fsm_advance(fsm, n, NULL);
	}

	return 0;
}

/*
 * Special node to handle 16bit addressing encoding, which can reference two
 * registers at once.
 */
static int
node_dual(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	int reg1, reg2;

	reg1 = gpr_dual_reg1_rm[instr->regmodrm.rm];

	if (instr->regmodrm.rm == 0b000 ||
	    instr->regmodrm.rm == 0b010) {
		reg2 = NVMM_X64_GPR_RSI;
	} else if (instr->regmodrm.rm == 0b001 ||
	    instr->regmodrm.rm == 0b011) {
		reg2 = NVMM_X64_GPR_RDI;
	} else {
		DISASSEMBLER_BUG();
	}

	instr->strm->type = STORE_DUALREG;
	instr->strm->u.dualreg.reg1 = reg1;
	instr->strm->u.dualreg.reg2 = reg2;

	if (instr->strm->disp.type == DISP_NONE) {
		DISASSEMBLER_BUG();
	} else if (instr->strm->disp.type == DISP_0) {
		/* Indirect register addressing mode */
		if (instr->opcode->immediate) {
			fsm_advance(fsm, 1, node_immediate);
		} else {
			fsm_advance(fsm, 1, NULL);
		}
	} else {
		fsm_advance(fsm, 1, node_disp);
	}

	return 0;
}

static const struct x86_reg *
get_register_idx(struct x86_instr *instr, uint8_t index)
{
	uint8_t enc = index;
	const struct x86_reg *reg;
	size_t regsize;

	regsize = instr->address_size;
	reg = &gpr_map[instr->rexpref.x][enc][regsize-1];

	if (reg->num == -1) {
		reg = resolve_special_register(instr, enc, regsize);
	}

	return reg;
}

static const struct x86_reg *
get_register_bas(struct x86_instr *instr, uint8_t base)
{
	uint8_t enc = base;
	const struct x86_reg *reg;
	size_t regsize;

	regsize = instr->address_size;
	reg = &gpr_map[instr->rexpref.b][enc][regsize-1];
	if (reg->num == -1) {
		reg = resolve_special_register(instr, enc, regsize);
	}

	return reg;
}

static int
node_sib(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode;
	uint8_t scale, index, base;
	bool noindex, nobase;
	uint8_t byte;

	if (fsm_read(fsm, &byte, sizeof(byte)) == -1) {
		return -1;
	}

	scale = ((byte & 0b11000000) >> 6);
	index = ((byte & 0b00111000) >> 3);
	base  = ((byte & 0b00000111) >> 0);

	opcode = instr->opcode;

	noindex = false;
	nobase = false;

	if (index == 0b100 && !instr->rexpref.x) {
		/* Special case: the index is null */
		noindex = true;
	}

	if (instr->regmodrm.mod == 0b00 && base == 0b101) {
		/* Special case: the base is null + disp32 */
		instr->strm->disp.type = DISP_4;
		nobase = true;
	}

	instr->strm->type = STORE_SIB;
	instr->strm->u.sib.scale = (1 << scale);
	if (!noindex)
		instr->strm->u.sib.idx = get_register_idx(instr, index);
	if (!nobase)
		instr->strm->u.sib.bas = get_register_bas(instr, base);

	/* May have a displacement, or an immediate */
	if (instr->strm->disp.type == DISP_1 ||
	    instr->strm->disp.type == DISP_2 ||
	    instr->strm->disp.type == DISP_4) {
		fsm_advance(fsm, 1, node_disp);
	} else if (opcode->immediate) {
		fsm_advance(fsm, 1, node_immediate);
	} else {
		fsm_advance(fsm, 1, NULL);
	}

	return 0;
}

static const struct x86_reg *
get_register_reg(struct x86_instr *instr, const struct x86_opcode *opcode)
{
	uint8_t enc = instr->regmodrm.reg;
	const struct x86_reg *reg;
	size_t regsize;

	regsize = instr->operand_size;

	reg = &gpr_map[instr->rexpref.r][enc][regsize-1];
	if (reg->num == -1) {
		reg = resolve_special_register(instr, enc, regsize);
	}

	return reg;
}

static const struct x86_reg *
get_register_rm(struct x86_instr *instr, const struct x86_opcode *opcode)
{
	uint8_t enc = instr->regmodrm.rm;
	const struct x86_reg *reg;
	size_t regsize;

	if (instr->strm->disp.type == DISP_NONE) {
		regsize = instr->operand_size;
	} else {
		/* Indirect access, the size is that of the address. */
		regsize = instr->address_size;
	}

	reg = &gpr_map[instr->rexpref.b][enc][regsize-1];
	if (reg->num == -1) {
		reg = resolve_special_register(instr, enc, regsize);
	}

	return reg;
}

static inline bool
has_sib(struct x86_instr *instr)
{
	return (instr->address_size != 2 && /* no SIB in 16bit addressing */
	    instr->regmodrm.mod != 0b11 &&
	    instr->regmodrm.rm == 0b100);
}

static inline bool
is_rip_relative(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	return (fsm->is64bit && /* RIP-relative only in 64bit mode */
	    instr->regmodrm.mod == 0b00 &&
	    instr->regmodrm.rm == 0b101);
}

static inline bool
is_disp32_only(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	return (!fsm->is64bit && /* no disp32-only in 64bit mode */
	    instr->address_size != 2 && /* no disp32-only in 16bit addressing */
	    instr->regmodrm.mod == 0b00 &&
	    instr->regmodrm.rm == 0b101);
}

static inline bool
is_disp16_only(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	return (instr->address_size == 2 && /* disp16-only only in 16bit addr */
	    instr->regmodrm.mod == 0b00 &&
	    instr->regmodrm.rm == 0b110);
}

static inline bool
is_dual(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	return (instr->address_size == 2 &&
	    instr->regmodrm.mod != 0b11 &&
	    instr->regmodrm.rm <= 0b011);
}

static enum x86_disp_type
get_disp_type(struct x86_instr *instr)
{
	switch (instr->regmodrm.mod) {
	case 0b00:	/* indirect */
		return DISP_0;
	case 0b01:	/* indirect+1 */
		return DISP_1;
	case 0b10:	/* indirect+{2,4} */
		if (__predict_false(instr->address_size == 2)) {
			return DISP_2;
		}
		return DISP_4;
	case 0b11:	/* direct */
	default:	/* llvm */
		return DISP_NONE;
	}
	__unreachable();
}

static int
node_regmodrm(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	struct x86_store *strg, *strm;
	const struct x86_opcode *opcode;
	const struct x86_reg *reg;
	uint8_t byte;

	if (fsm_read(fsm, &byte, sizeof(byte)) == -1) {
		return -1;
	}

	opcode = instr->opcode;

	instr->regmodrm.rm  = ((byte & 0b00000111) >> 0);
	instr->regmodrm.reg = ((byte & 0b00111000) >> 3);
	instr->regmodrm.mod = ((byte & 0b11000000) >> 6);

	if (opcode->regtorm) {
		strg = &instr->src;
		strm = &instr->dst;
	} else { /* RM to REG */
		strm = &instr->src;
		strg = &instr->dst;
	}

	/* Save for later use. */
	instr->strm = strm;

	/*
	 * Special cases: Groups. The REG field of REGMODRM is the index in
	 * the group. op1 gets overwritten in the Immediate node, if any.
	 */
	if (opcode->group1) {
		if (group1[instr->regmodrm.reg].emul == NULL) {
			return -1;
		}
		instr->emul = group1[instr->regmodrm.reg].emul;
	} else if (opcode->group3) {
		if (group3[instr->regmodrm.reg].emul == NULL) {
			return -1;
		}
		instr->emul = group3[instr->regmodrm.reg].emul;
	} else if (opcode->group11) {
		if (group11[instr->regmodrm.reg].emul == NULL) {
			return -1;
		}
		instr->emul = group11[instr->regmodrm.reg].emul;
	}

	if (!opcode->immediate) {
		reg = get_register_reg(instr, opcode);
		if (reg == NULL) {
			return -1;
		}
		strg->type = STORE_REG;
		strg->u.reg = reg;
	}

	/* The displacement applies to RM. */
	strm->disp.type = get_disp_type(instr);

	if (has_sib(instr)) {
		/* Overwrites RM */
		fsm_advance(fsm, 1, node_sib);
		return 0;
	}

	if (is_rip_relative(fsm, instr)) {
		/* Overwrites RM */
		strm->type = STORE_REG;
		strm->u.reg = &gpr_map__rip;
		strm->disp.type = DISP_4;
		fsm_advance(fsm, 1, node_disp);
		return 0;
	}

	if (is_disp32_only(fsm, instr)) {
		/* Overwrites RM */
		strm->type = STORE_REG;
		strm->u.reg = NULL;
		strm->disp.type = DISP_4;
		fsm_advance(fsm, 1, node_disp);
		return 0;
	}

	if (__predict_false(is_disp16_only(fsm, instr))) {
		/* Overwrites RM */
		strm->type = STORE_REG;
		strm->u.reg = NULL;
		strm->disp.type = DISP_2;
		fsm_advance(fsm, 1, node_disp);
		return 0;
	}

	if (__predict_false(is_dual(fsm, instr))) {
		/* Overwrites RM */
		fsm_advance(fsm, 0, node_dual);
		return 0;
	}

	reg = get_register_rm(instr, opcode);
	if (reg == NULL) {
		return -1;
	}
	strm->type = STORE_REG;
	strm->u.reg = reg;

	if (strm->disp.type == DISP_NONE) {
		/* Direct register addressing mode */
		if (opcode->immediate) {
			fsm_advance(fsm, 1, node_immediate);
		} else {
			fsm_advance(fsm, 1, NULL);
		}
	} else if (strm->disp.type == DISP_0) {
		/* Indirect register addressing mode */
		if (opcode->immediate) {
			fsm_advance(fsm, 1, node_immediate);
		} else {
			fsm_advance(fsm, 1, NULL);
		}
	} else {
		fsm_advance(fsm, 1, node_disp);
	}

	return 0;
}

static size_t
get_operand_size(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode = instr->opcode;
	int opsize;

	/* Get the opsize */
	if (!opcode->szoverride) {
		opsize = opcode->defsize;
	} else if (instr->rexpref.present && instr->rexpref.w) {
		opsize = 8;
	} else {
		if (!fsm->is16bit) {
			if (instr->legpref.opr_ovr) {
				opsize = 2;
			} else {
				opsize = 4;
			}
		} else { /* 16bit */
			if (instr->legpref.opr_ovr) {
				opsize = 4;
			} else {
				opsize = 2;
			}
		}
	}

	return opsize;
}

static size_t
get_address_size(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	if (fsm->is64bit) {
		if (__predict_false(instr->legpref.adr_ovr)) {
			return 4;
		}
		return 8;
	}

	if (fsm->is32bit) {
		if (__predict_false(instr->legpref.adr_ovr)) {
			return 2;
		}
		return 4;
	}

	/* 16bit. */
	if (__predict_false(instr->legpref.adr_ovr)) {
		return 4;
	}
	return 2;
}

static int
node_primary_opcode(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode;
	uint8_t byte;

	if (fsm_read(fsm, &byte, sizeof(byte)) == -1) {
		return -1;
	}

	opcode = &primary_opcode_table[byte];
	if (__predict_false(!opcode->valid)) {
		return -1;
	}

	instr->opcode = opcode;
	instr->emul = opcode->emul;
	instr->operand_size = get_operand_size(fsm, instr);
	instr->address_size = get_address_size(fsm, instr);

	if (fsm->is64bit && (instr->operand_size == 4)) {
		/* Zero-extend to 64 bits. */
		instr->zeroextend_mask = ~size_to_mask(4);
	}

	if (opcode->regmodrm) {
		fsm_advance(fsm, 1, node_regmodrm);
	} else if (opcode->dmo) {
		/* Direct-Memory Offsets */
		fsm_advance(fsm, 1, node_dmo);
	} else if (opcode->stos || opcode->lods) {
		fsm_advance(fsm, 1, node_stlo);
	} else if (opcode->movs) {
		fsm_advance(fsm, 1, node_movs);
	} else if (opcode->cmps) {
		fsm_advance(fsm, 1, node_cmps);
	} else {
		return -1;
	}

	return 0;
}

static int
node_secondary_opcode(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	const struct x86_opcode *opcode;
	uint8_t byte;

	if (fsm_read(fsm, &byte, sizeof(byte)) == -1) {
		return -1;
	}

	opcode = &secondary_opcode_table[byte];
	if (__predict_false(!opcode->valid)) {
		return -1;
	}

	instr->opcode = opcode;
	instr->emul = opcode->emul;
	instr->operand_size = get_operand_size(fsm, instr);
	instr->address_size = get_address_size(fsm, instr);

	if (fsm->is64bit && (instr->operand_size == 4)) {
		/* Zero-extend to 64 bits. */
		instr->zeroextend_mask = ~size_to_mask(4);
	}

	if (opcode->flags & FLAG_ze) {
		/*
		 * Compute the mask for zero-extend. Update the operand size,
		 * we move fewer bytes.
		 */
		instr->zeroextend_mask |= size_to_mask(instr->operand_size);
		instr->zeroextend_mask &= ~size_to_mask(opcode->defsize);
		instr->operand_size = opcode->defsize;
	}

	if (opcode->regmodrm) {
		fsm_advance(fsm, 1, node_regmodrm);
	} else {
		return -1;
	}

	return 0;
}

static int
node_main(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	uint8_t byte;

#define ESCAPE	0x0F
#define VEX_1	0xC5
#define VEX_2	0xC4
#define XOP	0x8F

	if (fsm_read(fsm, &byte, sizeof(byte)) == -1) {
		return -1;
	}

	/*
	 * We don't take XOP. It is AMD-specific, and it was removed shortly
	 * after being introduced.
	 */
	if (byte == ESCAPE) {
		fsm_advance(fsm, 1, node_secondary_opcode);
	} else if (!instr->rexpref.present) {
		if (byte == VEX_1) {
			return -1;
		} else if (byte == VEX_2) {
			return -1;
		} else {
			fsm->fn = node_primary_opcode;
		}
	} else {
		fsm->fn = node_primary_opcode;
	}

	return 0;
}

static int
node_rex_prefix(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	struct x86_rexpref *rexpref = &instr->rexpref;
	uint8_t byte;
	size_t n = 0;

	if (fsm_read(fsm, &byte, sizeof(byte)) == -1) {
		return -1;
	}

	if (byte >= 0x40 && byte <= 0x4F) {
		if (__predict_false(!fsm->is64bit)) {
			return -1;
		}
		rexpref->b = ((byte & 0x1) != 0);
		rexpref->x = ((byte & 0x2) != 0);
		rexpref->r = ((byte & 0x4) != 0);
		rexpref->w = ((byte & 0x8) != 0);
		rexpref->present = true;
		n = 1;
	}

	fsm_advance(fsm, n, node_main);
	return 0;
}

static int
node_legacy_prefix(struct x86_decode_fsm *fsm, struct x86_instr *instr)
{
	uint8_t byte;

	if (fsm_read(fsm, &byte, sizeof(byte)) == -1) {
		return -1;
	}

	if (byte == LEG_OPR_OVR) {
		instr->legpref.opr_ovr = 1;
	} else if (byte == LEG_OVR_DS) {
		instr->legpref.seg = NVMM_X64_SEG_DS;
	} else if (byte == LEG_OVR_ES) {
		instr->legpref.seg = NVMM_X64_SEG_ES;
	} else if (byte == LEG_REP) {
		instr->legpref.rep = 1;
	} else if (byte == LEG_OVR_GS) {
		instr->legpref.seg = NVMM_X64_SEG_GS;
	} else if (byte == LEG_OVR_FS) {
		instr->legpref.seg = NVMM_X64_SEG_FS;
	} else if (byte == LEG_ADR_OVR) {
		instr->legpref.adr_ovr = 1;
	} else if (byte == LEG_OVR_CS) {
		instr->legpref.seg = NVMM_X64_SEG_CS;
	} else if (byte == LEG_OVR_SS) {
		instr->legpref.seg = NVMM_X64_SEG_SS;
	} else if (byte == LEG_REPN) {
		instr->legpref.repn = 1;
	} else if (byte == LEG_LOCK) {
		/* ignore */
	} else {
		/* not a legacy prefix */
		fsm_advance(fsm, 0, node_rex_prefix);
		return 0;
	}

	fsm_advance(fsm, 1, node_legacy_prefix);
	return 0;
}

static int
x86_decode(uint8_t *inst_bytes, size_t inst_len, struct x86_instr *instr,
    struct nvmm_x64_state *state)
{
	struct x86_decode_fsm fsm;
	int ret;

	memset(instr, 0, sizeof(*instr));
	instr->legpref.seg = -1;
	instr->src.hardseg = -1;
	instr->dst.hardseg = -1;

	fsm.is64bit = is_64bit(state);
	fsm.is32bit = is_32bit(state);
	fsm.is16bit = is_16bit(state);

	fsm.fn = node_legacy_prefix;
	fsm.buf = inst_bytes;
	fsm.end = inst_bytes + inst_len;

	while (fsm.fn != NULL) {
		ret = (*fsm.fn)(&fsm, instr);
		if (ret == -1) {
#ifdef NVMM_DEBUG
			printf("\n%s debug: unrecognized instruction found " \
			       "with max length %ld : [ ", __func__, inst_len);
			for (uint i = 0; i < inst_len; i++)
			      printf("%02x ", inst_bytes[i]);
			printf("]\n");
			fflush(stdout);
#endif
			return -1;
		}
	}

	instr->len = fsm.buf - inst_bytes;

	return 0;
}

/* -------------------------------------------------------------------------- */

#define EXEC_INSTR(sz, instr)						\
static uint##sz##_t							\
exec_##instr##sz(uint##sz##_t op1, uint##sz##_t op2, uint64_t *rflags)	\
{									\
	uint##sz##_t res;						\
	__asm __volatile (						\
		#instr"	%2, %3;"					\
		"mov	%3, %1;"					\
		"pushfq;"						\
		"popq	%0"						\
	    : "=r" (*rflags), "=r" (res)				\
	    : "r" (op1), "r" (op2));					\
	return res;							\
}

#define EXEC_DISPATCHER(instr)						\
static uint64_t								\
exec_##instr(uint64_t op1, uint64_t op2, uint64_t *rflags, size_t opsize) \
{									\
	switch (opsize) {						\
	case 1:								\
		return exec_##instr##8(op1, op2, rflags);		\
	case 2:								\
		return exec_##instr##16(op1, op2, rflags);		\
	case 4:								\
		return exec_##instr##32(op1, op2, rflags);		\
	default:							\
		return exec_##instr##64(op1, op2, rflags);		\
	}								\
}

/* SUB: ret = op1 - op2 */
#define PSL_SUB_MASK	(PSL_V|PSL_C|PSL_Z|PSL_N|PSL_PF|PSL_AF)
EXEC_INSTR(8, sub)
EXEC_INSTR(16, sub)
EXEC_INSTR(32, sub)
EXEC_INSTR(64, sub)
EXEC_DISPATCHER(sub)

/* OR:  ret = op1 | op2 */
#define PSL_OR_MASK	(PSL_V|PSL_C|PSL_Z|PSL_N|PSL_PF)
EXEC_INSTR(8, or)
EXEC_INSTR(16, or)
EXEC_INSTR(32, or)
EXEC_INSTR(64, or)
EXEC_DISPATCHER(or)

/* AND: ret = op1 & op2 */
#define PSL_AND_MASK	(PSL_V|PSL_C|PSL_Z|PSL_N|PSL_PF)
EXEC_INSTR(8, and)
EXEC_INSTR(16, and)
EXEC_INSTR(32, and)
EXEC_INSTR(64, and)
EXEC_DISPATCHER(and)

/* XOR: ret = op1 ^ op2 */
#define PSL_XOR_MASK	(PSL_V|PSL_C|PSL_Z|PSL_N|PSL_PF)
EXEC_INSTR(8, xor)
EXEC_INSTR(16, xor)
EXEC_INSTR(32, xor)
EXEC_INSTR(64, xor)
EXEC_DISPATCHER(xor)

/* -------------------------------------------------------------------------- */

/*
 * Emulation functions. We don't care about the order of the operands, except
 * for SUB, CMP and TEST. For these ones we look at mem->write to determine who
 * is op1 and who is op2.
 */

static void
x86_func_or(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	uint64_t *retval = (uint64_t *)mem->data;
	const bool write = mem->write;
	uint64_t *op1, op2, fl, ret;

	op1 = (uint64_t *)mem->data;
	op2 = 0;

	/* Fetch the value to be OR'ed (op2). */
	mem->data = (uint8_t *)&op2;
	mem->write = false;
	(*vcpu->cbs.mem)(mem);

	/* Perform the OR. */
	ret = exec_or(*op1, op2, &fl, mem->size);

	if (write) {
		/* Write back the result. */
		mem->data = (uint8_t *)&ret;
		mem->write = true;
		(*vcpu->cbs.mem)(mem);
	} else {
		/* Return data to the caller. */
		*retval = ret;
	}

	gprs[NVMM_X64_GPR_RFLAGS] &= ~PSL_OR_MASK;
	gprs[NVMM_X64_GPR_RFLAGS] |= (fl & PSL_OR_MASK);
}

static void
x86_func_and(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	uint64_t *retval = (uint64_t *)mem->data;
	const bool write = mem->write;
	uint64_t *op1, op2, fl, ret;

	op1 = (uint64_t *)mem->data;
	op2 = 0;

	/* Fetch the value to be AND'ed (op2). */
	mem->data = (uint8_t *)&op2;
	mem->write = false;
	(*vcpu->cbs.mem)(mem);

	/* Perform the AND. */
	ret = exec_and(*op1, op2, &fl, mem->size);

	if (write) {
		/* Write back the result. */
		mem->data = (uint8_t *)&ret;
		mem->write = true;
		(*vcpu->cbs.mem)(mem);
	} else {
		/* Return data to the caller. */
		*retval = ret;
	}

	gprs[NVMM_X64_GPR_RFLAGS] &= ~PSL_AND_MASK;
	gprs[NVMM_X64_GPR_RFLAGS] |= (fl & PSL_AND_MASK);
}

static void
x86_func_xchg(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	uint64_t *op1, op2;

	op1 = (uint64_t *)mem->data;
	op2 = 0;

	/* Fetch op2. */
	mem->data = (uint8_t *)&op2;
	mem->write = false;
	(*vcpu->cbs.mem)(mem);

	/* Write op1 in op2. */
	mem->data = (uint8_t *)op1;
	mem->write = true;
	(*vcpu->cbs.mem)(mem);

	/* Write op2 in op1. */
	*op1 = op2;
}

static void
x86_func_sub(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	uint64_t *retval = (uint64_t *)mem->data;
	const bool write = mem->write;
	uint64_t *op1, *op2, fl, ret;
	uint64_t tmp;
	bool memop1;

	memop1 = !mem->write;
	op1 = memop1 ? &tmp : (uint64_t *)mem->data;
	op2 = memop1 ? (uint64_t *)mem->data : &tmp;

	/* Fetch the value to be SUB'ed (op1 or op2). */
	mem->data = (uint8_t *)&tmp;
	mem->write = false;
	(*vcpu->cbs.mem)(mem);

	/* Perform the SUB. */
	ret = exec_sub(*op1, *op2, &fl, mem->size);

	if (write) {
		/* Write back the result. */
		mem->data = (uint8_t *)&ret;
		mem->write = true;
		(*vcpu->cbs.mem)(mem);
	} else {
		/* Return data to the caller. */
		*retval = ret;
	}

	gprs[NVMM_X64_GPR_RFLAGS] &= ~PSL_SUB_MASK;
	gprs[NVMM_X64_GPR_RFLAGS] |= (fl & PSL_SUB_MASK);
}

static void
x86_func_xor(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	uint64_t *retval = (uint64_t *)mem->data;
	const bool write = mem->write;
	uint64_t *op1, op2, fl, ret;

	op1 = (uint64_t *)mem->data;
	op2 = 0;

	/* Fetch the value to be XOR'ed (op2). */
	mem->data = (uint8_t *)&op2;
	mem->write = false;
	(*vcpu->cbs.mem)(mem);

	/* Perform the XOR. */
	ret = exec_xor(*op1, op2, &fl, mem->size);

	if (write) {
		/* Write back the result. */
		mem->data = (uint8_t *)&ret;
		mem->write = true;
		(*vcpu->cbs.mem)(mem);
	} else {
		/* Return data to the caller. */
		*retval = ret;
	}

	gprs[NVMM_X64_GPR_RFLAGS] &= ~PSL_XOR_MASK;
	gprs[NVMM_X64_GPR_RFLAGS] |= (fl & PSL_XOR_MASK);
}

static void
x86_func_cmp(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	uint64_t *op1, *op2, fl;
	uint64_t tmp;
	bool memop1;

	memop1 = !mem->write;
	op1 = memop1 ? &tmp : (uint64_t *)mem->data;
	op2 = memop1 ? (uint64_t *)mem->data : &tmp;

	/* Fetch the value to be CMP'ed (op1 or op2). */
	mem->data = (uint8_t *)&tmp;
	mem->write = false;
	(*vcpu->cbs.mem)(mem);

	/* Perform the CMP. */
	exec_sub(*op1, *op2, &fl, mem->size);

	gprs[NVMM_X64_GPR_RFLAGS] &= ~PSL_SUB_MASK;
	gprs[NVMM_X64_GPR_RFLAGS] |= (fl & PSL_SUB_MASK);
}

static void
x86_func_test(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	uint64_t *op1, *op2, fl;
	uint64_t tmp;
	bool memop1;

	memop1 = !mem->write;
	op1 = memop1 ? &tmp : (uint64_t *)mem->data;
	op2 = memop1 ? (uint64_t *)mem->data : &tmp;

	/* Fetch the value to be TEST'ed (op1 or op2). */
	mem->data = (uint8_t *)&tmp;
	mem->write = false;
	(*vcpu->cbs.mem)(mem);

	/* Perform the TEST. */
	exec_and(*op1, *op2, &fl, mem->size);

	gprs[NVMM_X64_GPR_RFLAGS] &= ~PSL_AND_MASK;
	gprs[NVMM_X64_GPR_RFLAGS] |= (fl & PSL_AND_MASK);
}

static void
x86_func_mov(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	/*
	 * Nothing special, just move without emulation.
	 */
	(*vcpu->cbs.mem)(mem);
}

static void
x86_func_stos(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	/*
	 * Just move, and update RDI.
	 */
	(*vcpu->cbs.mem)(mem);

	if (gprs[NVMM_X64_GPR_RFLAGS] & PSL_D) {
		gprs[NVMM_X64_GPR_RDI] -= mem->size;
	} else {
		gprs[NVMM_X64_GPR_RDI] += mem->size;
	}
}

static void
x86_func_lods(struct nvmm_vcpu *vcpu, struct nvmm_mem *mem, uint64_t *gprs)
{
	/*
	 * Just move, and update RSI.
	 */
	(*vcpu->cbs.mem)(mem);

	if (gprs[NVMM_X64_GPR_RFLAGS] & PSL_D) {
		gprs[NVMM_X64_GPR_RSI] -= mem->size;
	} else {
		gprs[NVMM_X64_GPR_RSI] += mem->size;
	}
}

/* -------------------------------------------------------------------------- */

static inline uint64_t
gpr_read_address(struct x86_instr *instr, struct nvmm_x64_state *state, int gpr)
{
	uint64_t val;

	val = state->gprs[gpr];
	val &= size_to_mask(instr->address_size);

	return val;
}

static int
store_to_gva(struct nvmm_x64_state *state, struct x86_instr *instr,
    struct x86_store *store, gvaddr_t *gvap, size_t size)
{
	struct x86_sib *sib;
	gvaddr_t gva = 0;
	uint64_t reg;
	int ret, seg;

	if (store->type == STORE_SIB) {
		sib = &store->u.sib;
		if (sib->bas != NULL)
			gva += gpr_read_address(instr, state, sib->bas->num);
		if (sib->idx != NULL) {
			reg = gpr_read_address(instr, state, sib->idx->num);
			gva += sib->scale * reg;
		}
	} else if (store->type == STORE_REG) {
		if (store->u.reg == NULL) {
			/* The base is null. Happens with disp32-only and
			 * disp16-only. */
		} else {
			gva = gpr_read_address(instr, state, store->u.reg->num);
		}
	} else if (store->type == STORE_DUALREG) {
		gva = gpr_read_address(instr, state, store->u.dualreg.reg1) +
		    gpr_read_address(instr, state, store->u.dualreg.reg2);
	} else {
		gva = store->u.dmo;
	}

	if (store->disp.type != DISP_NONE) {
		gva += store->disp.data;
	}

	if (store->hardseg != -1) {
		seg = store->hardseg;
	} else {
		if (__predict_false(instr->legpref.seg != -1)) {
			seg = instr->legpref.seg;
		} else {
			seg = NVMM_X64_SEG_DS;
		}
	}

	if (__predict_true(is_long_mode(state))) {
		if (seg == NVMM_X64_SEG_GS || seg == NVMM_X64_SEG_FS) {
			segment_apply(&state->segs[seg], &gva);
		}
	} else {
		ret = segment_check(&state->segs[seg], gva, size);
		if (ret == -1)
			return -1;
		segment_apply(&state->segs[seg], &gva);
	}

	*gvap = gva;
	return 0;
}

static int
fetch_segment(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu)
{
	struct nvmm_x64_state *state = vcpu->state;
	uint8_t inst_bytes[5], byte;
	size_t i, fetchsize;
	gvaddr_t gva;
	int ret, seg;

	fetchsize = sizeof(inst_bytes);

	gva = state->gprs[NVMM_X64_GPR_RIP];
	if (__predict_false(!is_long_mode(state))) {
		ret = segment_check(&state->segs[NVMM_X64_SEG_CS], gva,
		    fetchsize);
		if (ret == -1)
			return -1;
		segment_apply(&state->segs[NVMM_X64_SEG_CS], &gva);
	}

	ret = read_guest_memory(mach, vcpu, gva, inst_bytes, fetchsize);
	if (ret == -1)
		return -1;

	seg = NVMM_X64_SEG_DS;
	for (i = 0; i < fetchsize; i++) {
		byte = inst_bytes[i];

		if (byte == LEG_OVR_DS) {
			seg = NVMM_X64_SEG_DS;
		} else if (byte == LEG_OVR_ES) {
			seg = NVMM_X64_SEG_ES;
		} else if (byte == LEG_OVR_GS) {
			seg = NVMM_X64_SEG_GS;
		} else if (byte == LEG_OVR_FS) {
			seg = NVMM_X64_SEG_FS;
		} else if (byte == LEG_OVR_CS) {
			seg = NVMM_X64_SEG_CS;
		} else if (byte == LEG_OVR_SS) {
			seg = NVMM_X64_SEG_SS;
		} else if (byte == LEG_OPR_OVR) {
			/* nothing */
		} else if (byte == LEG_ADR_OVR) {
			/* nothing */
		} else if (byte == LEG_REP) {
			/* nothing */
		} else if (byte == LEG_REPN) {
			/* nothing */
		} else if (byte == LEG_LOCK) {
			/* nothing */
		} else {
			return seg;
		}
	}

	return seg;
}

static int
fetch_instruction(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    struct nvmm_vcpu_exit *exit)
{
	struct nvmm_x64_state *state = vcpu->state;
	size_t fetchsize;
	gvaddr_t gva;
	int ret;

	fetchsize = sizeof(exit->u.mem.inst_bytes);

	gva = state->gprs[NVMM_X64_GPR_RIP];
	if (__predict_false(!is_long_mode(state))) {
		ret = segment_check(&state->segs[NVMM_X64_SEG_CS], gva,
		    fetchsize);
		if (ret == -1)
			return -1;
		segment_apply(&state->segs[NVMM_X64_SEG_CS], &gva);
	}

	ret = read_guest_memory(mach, vcpu, gva, exit->u.mem.inst_bytes,
	    fetchsize);
	if (ret == -1)
		return -1;

	exit->u.mem.inst_len = fetchsize;

	return 0;
}

static int
assist_mem_movs(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    struct x86_instr *instr)
{
	struct nvmm_x64_state *state = vcpu->state;
	uint64_t *gprs;
	uint8_t data[8];
	gvaddr_t gva;
	size_t size;
	int ret;

	size = instr->operand_size;
	gprs = state->gprs;

	/* Source. */
	ret = store_to_gva(state, instr, &instr->src, &gva, size);
	if (ret == -1)
		return -1;
	ret = read_guest_memory(mach, vcpu, gva, data, size);
	if (ret == -1)
		return -1;

	/* Destination. */
	ret = store_to_gva(state, instr, &instr->dst, &gva, size);
	if (ret == -1)
		return -1;
	ret = write_guest_memory(mach, vcpu, gva, data, size);
	if (ret == -1)
		return -1;

	/*
	 * Inlined x86_func_movs() call
	 * 	(*instr->emul->func)(vcpu, &mem, state->gprs);
	 */

	if (gprs[NVMM_X64_GPR_RFLAGS] & PSL_D) {
		gprs[NVMM_X64_GPR_RSI] -= size;
		gprs[NVMM_X64_GPR_RDI] -= size;
	} else {
		gprs[NVMM_X64_GPR_RSI] += size;
		gprs[NVMM_X64_GPR_RDI] += size;
	}

	return 0;
}

static int
assist_mem_cmps(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    struct x86_instr *instr)
{
	struct nvmm_x64_state *state = vcpu->state;
	uint64_t *gprs, op1, op2, fl;
	uint8_t data1[8], data2[8];
	gvaddr_t gva;
	size_t size;
	int ret;

	size = instr->operand_size;
	gprs = state->gprs;

	/* Source 1. */
	ret = store_to_gva(state, instr, &instr->src, &gva, size);
	if (ret == -1)
		return -1;
	ret = read_guest_memory(mach, vcpu, gva, data1, size);
	if (ret == -1)
		return -1;

	/* Source 2. */
	ret = store_to_gva(state, instr, &instr->dst, &gva, size);
	if (ret == -1)
		return -1;
	ret = read_guest_memory(mach, vcpu, gva, data2, size);
	if (ret == -1)
		return -1;

	/*
	 * Inlined x86_func_cmps() call
	 * 	(*instr->emul->func)(vcpu, &mem, state->gprs);
	 */

	/* Perform the CMP. */
	op1 = *((uint64_t *) data1);
	op2 = *((uint64_t *) data2);
	exec_sub(op1, op2, &fl, size);

	gprs[NVMM_X64_GPR_RFLAGS] &= ~PSL_SUB_MASK;
	gprs[NVMM_X64_GPR_RFLAGS] |= (fl & PSL_SUB_MASK);

	if (gprs[NVMM_X64_GPR_RFLAGS] & PSL_D) {
		gprs[NVMM_X64_GPR_RSI] -= size;
		gprs[NVMM_X64_GPR_RDI] -= size;
	} else {
		gprs[NVMM_X64_GPR_RSI] += size;
		gprs[NVMM_X64_GPR_RDI] += size;
	}

	return 0;
}

static int
assist_mem_single(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu,
    struct x86_instr *instr)
{
	struct nvmm_x64_state *state = vcpu->state;
	struct nvmm_vcpu_exit *exit = vcpu->exit;
	struct nvmm_mem mem;
	uint8_t membuf[8];
	uint64_t val;

	memset(membuf, 0, sizeof(membuf));

	mem.mach = mach;
	mem.vcpu = vcpu;
	mem.gpa = exit->u.mem.gpa;
	mem.size = instr->operand_size;
	mem.data = membuf;

	/* Determine the direction. */
	switch (instr->src.type) {
	case STORE_REG:
		if (instr->src.disp.type != DISP_NONE) {
			/* Indirect access. */
			mem.write = false;
		} else {
			/* Direct access. */
			mem.write = true;
		}
		break;
	case STORE_DUALREG:
		if (instr->src.disp.type == DISP_NONE) {
			DISASSEMBLER_BUG();
		}
		mem.write = false;
		break;
	case STORE_IMM:
		mem.write = true;
		break;
	case STORE_SIB:
		mem.write = false;
		break;
	case STORE_DMO:
		mem.write = false;
		break;
	default:
		DISASSEMBLER_BUG();
	}

	if (mem.write) {
		switch (instr->src.type) {
		case STORE_REG:
			/* The instruction was "reg -> mem". Fetch the register
			 * in membuf. */
			if (__predict_false(instr->src.disp.type != DISP_NONE)) {
				DISASSEMBLER_BUG();
			}
			val = state->gprs[instr->src.u.reg->num];
			val = __SHIFTOUT(val, instr->src.u.reg->mask);
			memcpy(mem.data, &val, mem.size);
			break;
		case STORE_IMM:
			/* The instruction was "imm -> mem". Fetch the immediate
			 * in membuf. */
			memcpy(mem.data, &instr->src.u.imm.data, mem.size);
			break;
		default:
			DISASSEMBLER_BUG();
		}
	} else if (instr->emul->readreg) {
		/* The instruction was "mem -> reg", but the value of the
		 * register matters for the emul func. Fetch it in membuf. */
		if (__predict_false(instr->dst.type != STORE_REG)) {
			DISASSEMBLER_BUG();
		}
		if (__predict_false(instr->dst.disp.type != DISP_NONE)) {
			DISASSEMBLER_BUG();
		}
		val = state->gprs[instr->dst.u.reg->num];
		val = __SHIFTOUT(val, instr->dst.u.reg->mask);
		memcpy(mem.data, &val, mem.size);
	}

	(*instr->emul->func)(vcpu, &mem, state->gprs);

	if (instr->emul->notouch) {
		/* We're done. */
		return 0;
	}

	if (!mem.write) {
		/* The instruction was "mem -> reg". The emul func has filled
		 * membuf with the memory content. Install membuf in the
		 * register. */
		if (__predict_false(instr->dst.type != STORE_REG)) {
			DISASSEMBLER_BUG();
		}
		if (__predict_false(instr->dst.disp.type != DISP_NONE)) {
			DISASSEMBLER_BUG();
		}
		memcpy(&val, membuf, sizeof(uint64_t));
		val = __SHIFTIN(val, instr->dst.u.reg->mask);
		state->gprs[instr->dst.u.reg->num] &= ~instr->dst.u.reg->mask;
		state->gprs[instr->dst.u.reg->num] |= val;
		state->gprs[instr->dst.u.reg->num] &= ~instr->zeroextend_mask;
	} else if (instr->emul->backprop) {
		/* The instruction was "reg -> mem", but the memory must be
		 * back-propagated to the register. Install membuf in the
		 * register. */
		if (__predict_false(instr->src.type != STORE_REG)) {
			DISASSEMBLER_BUG();
		}
		if (__predict_false(instr->src.disp.type != DISP_NONE)) {
			DISASSEMBLER_BUG();
		}
		memcpy(&val, membuf, sizeof(uint64_t));
		val = __SHIFTIN(val, instr->src.u.reg->mask);
		state->gprs[instr->src.u.reg->num] &= ~instr->src.u.reg->mask;
		state->gprs[instr->src.u.reg->num] |= val;
		state->gprs[instr->src.u.reg->num] &= ~instr->zeroextend_mask;
	}

	return 0;
}

int
nvmm_assist_mem(struct nvmm_machine *mach, struct nvmm_vcpu *vcpu)
{
	struct nvmm_x64_state *state = vcpu->state;
	struct nvmm_vcpu_exit *exit = vcpu->exit;
	struct x86_instr instr;
	uint64_t cnt = 0; /* GCC */
	int ret;

	if (__predict_false(exit->reason != NVMM_VCPU_EXIT_MEMORY)) {
		errno = EINVAL;
		return -1;
	}

	ret = nvmm_vcpu_getstate(mach, vcpu,
	    NVMM_X64_STATE_GPRS | NVMM_X64_STATE_SEGS |
	    NVMM_X64_STATE_CRS | NVMM_X64_STATE_MSRS);
	if (ret == -1)
		return -1;

	if (exit->u.mem.inst_len == 0) {
		/*
		 * The instruction was not fetched from the kernel. Fetch
		 * it ourselves.
		 */
		ret = fetch_instruction(mach, vcpu, exit);
		if (ret == -1)
			return -1;
	}

	ret = x86_decode(exit->u.mem.inst_bytes, exit->u.mem.inst_len,
	    &instr, state);
	if (ret == -1) {
		errno = ENODEV;
		return -1;
	}

	if (instr.legpref.rep || instr.legpref.repn) {
		cnt = rep_get_cnt(state, instr.address_size);
		if (__predict_false(cnt == 0)) {
			state->gprs[NVMM_X64_GPR_RIP] += instr.len;
			goto out;
		}
	}

	if (instr.opcode->movs) {
		ret = assist_mem_movs(mach, vcpu, &instr);
	} else if (instr.opcode->cmps) {
		instr.legpref.repe = !instr.legpref.repn;
		ret = assist_mem_cmps(mach, vcpu, &instr);
	} else {
		ret = assist_mem_single(mach, vcpu, &instr);
	}
	if (ret == -1) {
		errno = ENODEV;
		return -1;
	}

	if (instr.legpref.rep || instr.legpref.repn) {
		cnt -= 1;
		rep_set_cnt(state, instr.address_size, cnt);
		if (cnt == 0) {
			state->gprs[NVMM_X64_GPR_RIP] += instr.len;
		} else if (__predict_false(instr.legpref.repn)) {
			/* repn */
			if (state->gprs[NVMM_X64_GPR_RFLAGS] & PSL_Z) {
				state->gprs[NVMM_X64_GPR_RIP] += instr.len;
			}
		} else if (__predict_false(instr.legpref.repe)) {
			/* repe */
			if ((state->gprs[NVMM_X64_GPR_RFLAGS] & PSL_Z) == 0) {
				state->gprs[NVMM_X64_GPR_RIP] += instr.len;
			}
		}
	} else {
		state->gprs[NVMM_X64_GPR_RIP] += instr.len;
	}

out:
	ret = nvmm_vcpu_setstate(mach, vcpu, NVMM_X64_STATE_GPRS);
	if (ret == -1)
		return -1;

	return 0;
}