; $Id: HMR0Mixed.mac 48965 2013-10-08 09:06:05Z vboxsync $
;; @file
; HM - Ring-0 Host 32/64, Guest 32/64 world-switch routines
;
; Darwin uses this to build two versions in the hybrid case.
; Included by HMR0A.asm with RT_ARCH_AMD64 defined or undefined.
;

;
; Copyright (C) 2006-2013 Oracle Corporation
;
; This file is part of VirtualBox Open Source Edition (OSE), as
; available from http://www.virtualbox.org. This file is free software;
; you can redistribute it and/or modify it under the terms of the GNU
; General Public License (GPL) as published by the Free Software
; Foundation, in version 2 as it comes in the "COPYING" file of the
; VirtualBox OSE distribution. VirtualBox OSE is distributed in the
; hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
;

%ifdef RT_ARCH_AMD64
 ;;
 ; Keep these macro definitions in this file as it gets included and compiled
 ; with RT_ARCH_AMD64 once and RT_ARCH_X86 once.
 %define VMX_SKIP_GDTR
 %ifdef RT_OS_DARWIN
 ; Darwin (Mavericks) uses IDTR limit to store the CPUID so we need to restore it always.
 ; See @bugref{6875}.
 %elifdef RT_OS_WINDOWS
 ; Windows 8.1 RTM also seems to be using the IDTR limit for something. See @bugref{6956}.
 ;; @todo figure out what exactly it does and try and restrict it more.
 %else
  %define VMX_SKIP_IDTR
 %endif
 %define VMX_SKIP_TR
%endif

;; @def RESTORESTATEVM32
; Macro restoring essential host state and updating guest state
; for common host, 32-bit guest for VT-x.
;
; @param 1  Jump label suffix 1.
; @param 2  Jump label suffix 2.
; @param 3  Jump label suffix 3.
%macro RESTORESTATEVM32 3
    ; Restore base and limit of the IDTR & GDTR.
 %ifndef VMX_SKIP_IDTR
    lidt    [xSP]
    add     xSP, xCB * 2
 %endif
 %ifndef VMX_SKIP_GDTR
    lgdt    [xSP]
    add     xSP, xCB * 2
 %endif

    push    xDI
 %ifndef VMX_SKIP_TR
    mov     xDI, [xSP + xCB * 3]         ; pCtx (*3 to skip the saved xDI, TR, LDTR).
 %else
    mov     xDI, [xSP + xCB * 2]         ; pCtx (*2 to skip the saved xDI, LDTR).
 %endif

    mov     [ss:xDI + CPUMCTX.eax], eax
    mov     [ss:xDI + CPUMCTX.ebx], ebx
    mov     [ss:xDI + CPUMCTX.ecx], ecx
    mov     [ss:xDI + CPUMCTX.edx], edx
    mov     [ss:xDI + CPUMCTX.esi], esi
    mov     [ss:xDI + CPUMCTX.ebp], ebp
    mov     xAX, cr2
    mov     [ss:xDI + CPUMCTX.cr2], xAX

 %ifdef RT_ARCH_AMD64
    pop     xAX                                 ; The guest edi we pushed above.
    mov     dword [ss:xDI + CPUMCTX.edi], eax
 %else
    pop     dword [ss:xDI + CPUMCTX.edi]        ; The guest edi we pushed above.
 %endif

 %ifndef VMX_SKIP_TR
    ; Restore TSS selector; must mark it as not busy before using ltr (!)
    ; ASSUME that this is supposed to be 'BUSY'. (saves 20-30 ticks on the T42p)
    ; @todo get rid of sgdt
    pop     xBX         ; Saved TR
    sub     xSP, xCB * 2
    sgdt    [xSP]
    mov     xAX, xBX
    and     eax, X86_SEL_MASK_OFF_RPL               ; Mask away TI and RPL bits leaving only the descriptor offset.
    add     xAX, [xSP + 2]                          ; eax <- GDTR.address + descriptor offset.
    and     dword [ss:xAX + 4], ~RT_BIT(9)          ; Clear the busy flag in TSS desc (bits 0-7=base, bit 9=busy bit).
    ltr     bx
    add     xSP, xCB * 2
 %endif

    pop     xAX         ; Saved LDTR
 %ifdef RT_ARCH_AMD64
    cmp     eax, 0
    je      .skipldtwrite32%1
 %endif
    lldt    ax

.skipldtwrite32%1:
    add     xSP, xCB     ; pCtx

 %ifdef VMX_USE_CACHED_VMCS_ACCESSES
    pop     xDX         ; Saved pCache

    ; Note! If we get here as a result of invalid VMCS pointer, all the following
    ; vmread's will fail (only eflags.cf=1 will be set) but that shouldn't cause any
    ; trouble only just less efficient.
    mov     ecx, [ss:xDX + VMCSCACHE.Read.cValidEntries]
    cmp     ecx, 0      ; Can't happen
    je      .no_cached_read32%2
    jmp     .cached_read32%3

ALIGN(16)
.cached_read32%3:
    dec     xCX
    mov     eax, [ss:xDX + VMCSCACHE.Read.aField + xCX * 4]
    vmread  [ss:xDX + VMCSCACHE.Read.aFieldVal + xCX * 8], xAX
    cmp     xCX, 0
    jnz     .cached_read32%3
.no_cached_read32%2:
 %endif

    ; Restore segment registers.
    MYPOPSEGS xAX, ax

    ; Restore general purpose registers.
    MYPOPAD
%endmacro


;/**
; * Prepares for and executes VMLAUNCH/VMRESUME (32 bits guest mode)
; *
; * @returns VBox status code
; * @param   fResume    x86:[ebp+8], msc:rcx,gcc:rdi     Whether to use vmlauch/vmresume.
; * @param   pCtx       x86:[ebp+c], msc:rdx,gcc:rsi     Pointer to the guest-CPU context.
; * @param   pCache     x86:[esp+10],msc:r8, gcc:rdx     Pointer to the VMCS cache.
; */
ALIGNCODE(16)
BEGINPROC MY_NAME(VMXR0StartVM32)
    push    xBP
    mov     xBP, xSP

    pushf
    cli

    ; Save all general purpose host registers.
    MYPUSHAD

    ; First we have to save some final CPU context registers.
    mov     eax, VMX_VMCS_HOST_RIP
%ifdef RT_ARCH_AMD64
    lea     r10, [.vmlaunch_done wrt rip]
    vmwrite rax, r10
%else
    mov     ecx, .vmlaunch_done
    vmwrite eax, ecx
%endif
    ; Note: assumes success!

    ; Save guest-CPU context pointer.
%ifdef RT_ARCH_AMD64
 %ifdef ASM_CALL64_GCC
    ; fResume already in rdi
    ; pCtx    already in rsi
    mov     rbx, rdx        ; pCache
 %else
    mov     rdi, rcx        ; fResume
    mov     rsi, rdx        ; pCtx
    mov     rbx, r8         ; pCache
 %endif
%else
    mov     edi, [ebp + 8]  ; fResume
    mov     esi, [ebp + 12] ; pCtx
    mov     ebx, [ebp + 16] ; pCache
%endif

    ; Save segment registers.
    ; Note: MYPUSHSEGS trashes rdx & rcx, so we moved it here (msvc amd64 case).
    MYPUSHSEGS xAX, ax

%ifdef VMX_USE_CACHED_VMCS_ACCESSES
    mov     ecx, [xBX + VMCSCACHE.Write.cValidEntries]
    cmp     ecx, 0
    je      .no_cached_writes
    mov     edx, ecx
    mov     ecx, 0
    jmp     .cached_write

ALIGN(16)
.cached_write:
    mov     eax, [xBX + VMCSCACHE.Write.aField + xCX * 4]
    vmwrite xAX, [xBX + VMCSCACHE.Write.aFieldVal + xCX * 8]
    inc     xCX
    cmp     xCX, xDX
    jl     .cached_write

    mov     dword [xBX + VMCSCACHE.Write.cValidEntries], 0
.no_cached_writes:

    ; Save the pCache pointer.
    push    xBX
%endif

    ; Save the pCtx pointer.
    push    xSI

    ; Save host LDTR.
    xor     eax, eax
    sldt    ax
    push    xAX

%ifndef VMX_SKIP_TR
    ; The host TR limit is reset to 0x67; save & restore it manually.
    str     eax
    push    xAX
%endif

%ifndef VMX_SKIP_GDTR
    ; VT-x only saves the base of the GDTR & IDTR and resets the limit to 0xffff; we must restore the limit correctly!
    sub     xSP, xCB * 2
    sgdt    [xSP]
%endif
%ifndef VMX_SKIP_IDTR
    sub     xSP, xCB * 2
    sidt    [xSP]
%endif

    ; Load CR2 if necessary (may be expensive as writing CR2 is a synchronizing instruction).
    mov     xBX, [xSI + CPUMCTX.cr2]
    mov     xDX, cr2
    cmp     xBX, xDX
    je      .skipcr2write32
    mov     cr2, xBX

.skipcr2write32:
    mov     eax, VMX_VMCS_HOST_RSP
    vmwrite xAX, xSP
    ; Note: assumes success!
    ; Don't mess with ESP anymore!!!

    ; Load guest general purpose registers.
    mov     eax, [xSI + CPUMCTX.eax]
    mov     ebx, [xSI + CPUMCTX.ebx]
    mov     ecx, [xSI + CPUMCTX.ecx]
    mov     edx, [xSI + CPUMCTX.edx]
    mov     ebp, [xSI + CPUMCTX.ebp]

    ; Resume or start VM?
    cmp     xDI, 0                  ; fResume
    je      .vmlaunch_launch

    ; Load guest edi & esi.
    mov     edi, [xSI + CPUMCTX.edi]
    mov     esi, [xSI + CPUMCTX.esi]

    vmresume
    jmp     .vmlaunch_done;      ; Here if vmresume detected a failure.

.vmlaunch_launch:
    ; Save guest edi & esi.
    mov     edi, [xSI + CPUMCTX.edi]
    mov     esi, [xSI + CPUMCTX.esi]

    vmlaunch
    jmp     .vmlaunch_done;      ; Here if vmlaunch detected a failure.

ALIGNCODE(16) ;; @todo YASM BUG - this alignment is wrong on darwin, it's 1 byte off.
.vmlaunch_done:
    jc      near .vmxstart_invalid_vmcs_ptr
    jz      near .vmxstart_start_failed

    RESTORESTATEVM32 A, B, C
    mov     eax, VINF_SUCCESS

.vmstart_end:
    popf
    pop     xBP
    ret

.vmxstart_invalid_vmcs_ptr:
    RESTORESTATEVM32 D, E, F
    mov     eax, VERR_VMX_INVALID_VMCS_PTR_TO_START_VM
    jmp     .vmstart_end

.vmxstart_start_failed:
    RESTORESTATEVM32 G, H, I
    mov     eax, VERR_VMX_UNABLE_TO_START_VM
    jmp     .vmstart_end

ENDPROC MY_NAME(VMXR0StartVM32)


%ifdef RT_ARCH_AMD64
;; @def RESTORESTATEVM64
; Macro restoring essential host state and updating guest state
; for 64-bit host, 64-bit guest for VT-x.
;
; @param 1  Jump label suffix 1.
; @param 2  Jump label suffix 2.
; @param 3  Jump label suffix 3.
%macro RESTORESTATEVM64 3
    ; Restore base and limit of the IDTR & GDTR
 %ifndef VMX_SKIP_IDTR
    lidt    [xSP]
    add     xSP, xCB * 2
 %endif
 %ifndef VMX_SKIP_GDTR
    lgdt    [xSP]
    add     xSP, xCB * 2
 %endif

    push    xDI
 %ifndef VMX_SKIP_TR
    mov     xDI, [xSP + xCB * 3]        ; pCtx (*3 to skip the saved xDI, TR, LDTR)
 %else
    mov     xDI, [xSP + xCB * 2]        ; pCtx (*2 to skip the saved xDI, LDTR)
 %endif

    mov     qword [xDI + CPUMCTX.eax], rax
    mov     qword [xDI + CPUMCTX.ebx], rbx
    mov     qword [xDI + CPUMCTX.ecx], rcx
    mov     qword [xDI + CPUMCTX.edx], rdx
    mov     qword [xDI + CPUMCTX.esi], rsi
    mov     qword [xDI + CPUMCTX.ebp], rbp
    mov     qword [xDI + CPUMCTX.r8],  r8
    mov     qword [xDI + CPUMCTX.r9],  r9
    mov     qword [xDI + CPUMCTX.r10], r10
    mov     qword [xDI + CPUMCTX.r11], r11
    mov     qword [xDI + CPUMCTX.r12], r12
    mov     qword [xDI + CPUMCTX.r13], r13
    mov     qword [xDI + CPUMCTX.r14], r14
    mov     qword [xDI + CPUMCTX.r15], r15
    mov     rax, cr2
    mov     qword [xDI + CPUMCTX.cr2], rax

    pop     xAX                                 ; The guest rdi we pushed above
    mov     qword [xDI + CPUMCTX.edi], rax

 %ifndef VMX_SKIP_TR
    ; Restore TSS selector; must mark it as not busy before using ltr (!)
    ; ASSUME that this is supposed to be 'BUSY'. (saves 20-30 ticks on the T42p).
    ; @todo get rid of sgdt
    pop     xBX         ; Saved TR
    sub     xSP, xCB * 2
    sgdt    [xSP]
    mov     xAX, xBX
    and     eax, X86_SEL_MASK_OFF_RPL           ; Mask away TI and RPL bits leaving only the descriptor offset.
    add     xAX, [xSP + 2]                      ; eax <- GDTR.address + descriptor offset.
    and     dword [xAX + 4], ~RT_BIT(9)         ; Clear the busy flag in TSS desc (bits 0-7=base, bit 9=busy bit).
    ltr     bx
    add     xSP, xCB * 2
 %endif

    pop     xAX         ; Saved LDTR
    cmp     eax, 0
    je      .skipldtwrite64%1
    lldt    ax

.skipldtwrite64%1:
    pop     xSI         ; pCtx (needed in rsi by the macros below)

 %ifndef VBOX_WITH_AUTO_MSR_LOAD_RESTORE
    ; Save the guest MSRs and load the host MSRs.
    LOADHOSTMSREX MSR_K8_KERNEL_GS_BASE, CPUMCTX.msrKERNELGSBASE
    LOADHOSTMSREX MSR_K8_SF_MASK,        CPUMCTX.msrSFMASK
    LOADHOSTMSREX MSR_K6_STAR,           CPUMCTX.msrSTAR
    LOADHOSTMSREX MSR_K8_LSTAR,          CPUMCTX.msrLSTAR
 %endif

 %ifdef VMX_USE_CACHED_VMCS_ACCESSES
    pop     xDX         ; Saved pCache

    ; Note! If we get here as a result of invalid VMCS pointer, all the following
    ; vmread's will fail (only eflags.cf=1 will be set) but that shouldn't cause any
    ; trouble only just less efficient.
    mov     ecx, [xDX + VMCSCACHE.Read.cValidEntries]
    cmp     ecx, 0      ; Can't happen
    je      .no_cached_read64%2
    jmp     .cached_read64%3

ALIGN(16)
.cached_read64%3:
    dec     xCX
    mov     eax, [xDX + VMCSCACHE.Read.aField + xCX * 4]
    vmread  [xDX + VMCSCACHE.Read.aFieldVal + xCX * 8], xAX
    cmp     xCX, 0
    jnz     .cached_read64%3
.no_cached_read64%2:
 %endif

    ; Restore segment registers.
    MYPOPSEGS xAX, ax

    ; Restore general purpose registers.
    MYPOPAD
%endmacro


;/**
; * Prepares for and executes VMLAUNCH/VMRESUME (64 bits guest mode)
; *
; * @returns VBox status code
; * @param   fResume    msc:rcx, gcc:rdi     Whether to use vmlauch/vmresume.
; * @param   pCtx       msc:rdx, gcc:rsi     Pointer to the guest-CPU context.
; * @param   pCache     msc:r8,  gcc:rdx     Pointer to the VMCS cache.
; */
ALIGNCODE(16)
BEGINPROC MY_NAME(VMXR0StartVM64)
    push    xBP
    mov     xBP, xSP

    pushf
    cli

    ; Save all general purpose host registers.
    MYPUSHAD

    ; First we have to save some final CPU context registers.
    lea     r10, [.vmlaunch64_done wrt rip]
    mov     rax, VMX_VMCS_HOST_RIP      ; Return address (too difficult to continue after VMLAUNCH?).
    vmwrite rax, r10
    ; Note: assumes success!

    ; Save guest-CPU context pointer.
%ifdef ASM_CALL64_GCC
    ; fResume already in rdi
    ; pCtx    already in rsi
    mov     rbx, rdx        ; pCache
%else
    mov     rdi, rcx        ; fResume
    mov     rsi, rdx        ; pCtx
    mov     rbx, r8         ; pCache
%endif

    ; Save segment registers.
    ; Note: MYPUSHSEGS trashes rdx & rcx, so we moved it here (msvc amd64 case).
    MYPUSHSEGS xAX, ax

%ifdef VMX_USE_CACHED_VMCS_ACCESSES
    mov     ecx, [xBX + VMCSCACHE.Write.cValidEntries]
    cmp     ecx, 0
    je      .no_cached_writes
    mov     edx, ecx
    mov     ecx, 0
    jmp     .cached_write

ALIGN(16)
.cached_write:
    mov     eax, [xBX + VMCSCACHE.Write.aField + xCX * 4]
    vmwrite xAX, [xBX + VMCSCACHE.Write.aFieldVal + xCX * 8]
    inc     xCX
    cmp     xCX, xDX
    jl     .cached_write

    mov     dword [xBX + VMCSCACHE.Write.cValidEntries], 0
.no_cached_writes:

    ; Save the pCache pointer.
    push    xBX
%endif

%ifndef VBOX_WITH_AUTO_MSR_LOAD_RESTORE
    ; Save the host MSRs and load the guest MSRs.
    LOADGUESTMSR MSR_K8_LSTAR,          CPUMCTX.msrLSTAR
    LOADGUESTMSR MSR_K6_STAR,           CPUMCTX.msrSTAR
    LOADGUESTMSR MSR_K8_SF_MASK,        CPUMCTX.msrSFMASK
    LOADGUESTMSR MSR_K8_KERNEL_GS_BASE, CPUMCTX.msrKERNELGSBASE
%endif

    ; Save the pCtx pointer.
    push    xSI

    ; Save host LDTR.
    xor     eax, eax
    sldt    ax
    push    xAX

%ifndef VMX_SKIP_TR
    ; The host TR limit is reset to 0x67; save & restore it manually.
    str     eax
    push    xAX
%endif

%ifndef VMX_SKIP_GDTR
    ; VT-x only saves the base of the GDTR & IDTR and resets the limit to 0xffff; we must restore the limit correctly!
    sub     xSP, xCB * 2
    sgdt    [xSP]
%endif
%ifndef VMX_SKIP_IDTR
    sub     xSP, xCB * 2
    sidt    [xSP]
%endif

    ; Load CR2 if necessary (may be expensive as writing CR2 is a synchronizing instruction).
    mov     rbx, qword [xSI + CPUMCTX.cr2]
    mov     rdx, cr2
    cmp     rbx, rdx
    je      .skipcr2write
    mov     cr2, rbx

.skipcr2write:
    mov     eax, VMX_VMCS_HOST_RSP
    vmwrite xAX, xSP
    ; Note: assumes success!
    ; Don't mess with ESP anymore!!!

    ; Load guest general purpose registers.
    mov     rax, qword [xSI + CPUMCTX.eax]
    mov     rbx, qword [xSI + CPUMCTX.ebx]
    mov     rcx, qword [xSI + CPUMCTX.ecx]
    mov     rdx, qword [xSI + CPUMCTX.edx]
    mov     rbp, qword [xSI + CPUMCTX.ebp]
    mov     r8,  qword [xSI + CPUMCTX.r8]
    mov     r9,  qword [xSI + CPUMCTX.r9]
    mov     r10, qword [xSI + CPUMCTX.r10]
    mov     r11, qword [xSI + CPUMCTX.r11]
    mov     r12, qword [xSI + CPUMCTX.r12]
    mov     r13, qword [xSI + CPUMCTX.r13]
    mov     r14, qword [xSI + CPUMCTX.r14]
    mov     r15, qword [xSI + CPUMCTX.r15]

    ; Resume or start VM?
    cmp     xDI, 0                  ; fResume
    je      .vmlaunch64_launch

    ; Load guest rdi & rsi.
    mov     rdi, qword [xSI + CPUMCTX.edi]
    mov     rsi, qword [xSI + CPUMCTX.esi]

    vmresume
    jmp     .vmlaunch64_done;      ; Here if vmresume detected a failure.

.vmlaunch64_launch:
    ; Save guest rdi & rsi.
    mov     rdi, qword [xSI + CPUMCTX.edi]
    mov     rsi, qword [xSI + CPUMCTX.esi]

    vmlaunch
    jmp     .vmlaunch64_done;      ; Here if vmlaunch detected a failure.

ALIGNCODE(16)
.vmlaunch64_done:
    jc      near .vmxstart64_invalid_vmcs_ptr
    jz      near .vmxstart64_start_failed

    RESTORESTATEVM64 a, b, c
    mov     eax, VINF_SUCCESS

.vmstart64_end:
    popf
    pop     xBP
    ret

.vmxstart64_invalid_vmcs_ptr:
    RESTORESTATEVM64 d, e, f
    mov     eax, VERR_VMX_INVALID_VMCS_PTR_TO_START_VM
    jmp     .vmstart64_end

.vmxstart64_start_failed:
    RESTORESTATEVM64 g, h, i
    mov     eax, VERR_VMX_UNABLE_TO_START_VM
    jmp     .vmstart64_end
ENDPROC MY_NAME(VMXR0StartVM64)
%endif ; RT_ARCH_AMD64


;/**
; * Prepares for and executes VMRUN (32 bits guests)
; *
; * @returns VBox status code
; * @param   HCPhysVMCB     Physical address of host VMCB.
; * @param   HCPhysVMCB     Physical address of guest VMCB.
; * @param   pCtx           Pointer to the guest CPU-context.
; */
ALIGNCODE(16)
BEGINPROC MY_NAME(SVMR0VMRun)
%ifdef RT_ARCH_AMD64 ; fake a cdecl stack frame
 %ifdef ASM_CALL64_GCC
    push    rdx
    push    rsi
    push    rdi
 %else
    push    r8
    push    rdx
    push    rcx
 %endif
    push    0
%endif
    push    xBP
    mov     xBP, xSP
    pushf

    ; Save all general purpose host registers.
    MYPUSHAD

    ; Save guest CPU-context pointer.
    mov     xSI, [xBP + xCB * 2 + RTHCPHYS_CB * 2]  ; pCtx
    push    xSI                                     ; push for saving the state at the end

    ; Save host fs, gs, sysenter msr etc.
    mov     xAX, [xBP + xCB * 2]                    ; pVMCBHostPhys (64 bits physical address; x86: take low dword only)
    push    xAX                                     ; save for the vmload after vmrun
    vmsave

    ; Setup eax for VMLOAD.
    mov     xAX, [xBP + xCB * 2 + RTHCPHYS_CB]      ; pVMCBPhys (64 bits physical address; take low dword only)

    ; Load guest general purpose registers.
    ; eax is loaded from the VMCB by VMRUN.
    mov     ebx, [xSI + CPUMCTX.ebx]
    mov     ecx, [xSI + CPUMCTX.ecx]
    mov     edx, [xSI + CPUMCTX.edx]
    mov     edi, [xSI + CPUMCTX.edi]
    mov     ebp, [xSI + CPUMCTX.ebp]
    mov     esi, [xSI + CPUMCTX.esi]

    ; Clear the global interrupt flag & execute sti to make sure external interrupts cause a world switch.
    clgi
    sti

    ; Load guest fs, gs, sysenter msr etc.
    vmload
    ; Run the VM.
    vmrun

    ; eax is in the VMCB already; we can use it here.

    ; Save guest fs, gs, sysenter msr etc.
    vmsave

    ; Load host fs, gs, sysenter msr etc.
    pop     xAX                     ; Pushed above
    vmload

    ; Set the global interrupt flag again, but execute cli to make sure IF=0.
    cli
    stgi

    pop     xAX                     ; pCtx

    mov     [ss:xAX + CPUMCTX.ebx], ebx
    mov     [ss:xAX + CPUMCTX.ecx], ecx
    mov     [ss:xAX + CPUMCTX.edx], edx
    mov     [ss:xAX + CPUMCTX.esi], esi
    mov     [ss:xAX + CPUMCTX.edi], edi
    mov     [ss:xAX + CPUMCTX.ebp], ebp

    ; Restore host general purpose registers.
    MYPOPAD

    mov     eax, VINF_SUCCESS

    popf
    pop     xBP
%ifdef RT_ARCH_AMD64
    add     xSP, 4*xCB
%endif
    ret
ENDPROC MY_NAME(SVMR0VMRun)

%ifdef RT_ARCH_AMD64
;/**
; * Prepares for and executes VMRUN (64 bits guests)
; *
; * @returns VBox status code
; * @param   HCPhysVMCB     Physical address of host VMCB.
; * @param   HCPhysVMCB     Physical address of guest VMCB.
; * @param   pCtx           Pointer to the guest-CPU context.
; */
ALIGNCODE(16)
BEGINPROC MY_NAME(SVMR0VMRun64)
    ; Fake a cdecl stack frame
 %ifdef ASM_CALL64_GCC
    push    rdx
    push    rsi
    push    rdi
 %else
    push    r8
    push    rdx
    push    rcx
 %endif
    push    0
    push    rbp
    mov     rbp, rsp
    pushf

    ; Manual save and restore:
    ;  - General purpose registers except RIP, RSP, RAX
    ;
    ; Trashed:
    ;  - CR2 (we don't care)
    ;  - LDTR (reset to 0)
    ;  - DRx (presumably not changed at all)
    ;  - DR7 (reset to 0x400)
    ;

    ; Save all general purpose host registers.
    MYPUSHAD

    ; Save guest CPU-context pointer.
    mov     rsi, [rbp + xCB * 2 + RTHCPHYS_CB * 2]  ; pCtx
    push    rsi                                     ; push for saving the state at the end

    ; Save host fs, gs, sysenter msr etc.
    mov     rax, [rbp + xCB * 2]                    ; pVMCBHostPhys (64 bits physical address; x86: take low dword only)
    push    rax                                     ; Save for the vmload after vmrun
    vmsave

    ; Setup eax for VMLOAD.
    mov     rax, [rbp + xCB * 2 + RTHCPHYS_CB]      ; pVMCBPhys (64 bits physical address; take low dword only)

    ; Load guest general purpose registers.
    ; rax is loaded from the VMCB by VMRUN.
    mov     rbx, qword [xSI + CPUMCTX.ebx]
    mov     rcx, qword [xSI + CPUMCTX.ecx]
    mov     rdx, qword [xSI + CPUMCTX.edx]
    mov     rdi, qword [xSI + CPUMCTX.edi]
    mov     rbp, qword [xSI + CPUMCTX.ebp]
    mov     r8,  qword [xSI + CPUMCTX.r8]
    mov     r9,  qword [xSI + CPUMCTX.r9]
    mov     r10, qword [xSI + CPUMCTX.r10]
    mov     r11, qword [xSI + CPUMCTX.r11]
    mov     r12, qword [xSI + CPUMCTX.r12]
    mov     r13, qword [xSI + CPUMCTX.r13]
    mov     r14, qword [xSI + CPUMCTX.r14]
    mov     r15, qword [xSI + CPUMCTX.r15]
    mov     rsi, qword [xSI + CPUMCTX.esi]

    ; Clear the global interrupt flag & execute sti to make sure external interrupts cause a world switch.
    clgi
    sti

    ; Load guest fs, gs, sysenter msr etc.
    vmload
    ; Run the VM.
    vmrun

    ; rax is in the VMCB already; we can use it here.

    ; Save guest fs, gs, sysenter msr etc.
    vmsave

    ; Load host fs, gs, sysenter msr etc.
    pop     rax                     ; pushed above
    vmload

    ; Set the global interrupt flag again, but execute cli to make sure IF=0.
    cli
    stgi

    pop     rax                     ; pCtx

    mov     qword [rax + CPUMCTX.ebx], rbx
    mov     qword [rax + CPUMCTX.ecx], rcx
    mov     qword [rax + CPUMCTX.edx], rdx
    mov     qword [rax + CPUMCTX.esi], rsi
    mov     qword [rax + CPUMCTX.edi], rdi
    mov     qword [rax + CPUMCTX.ebp], rbp
    mov     qword [rax + CPUMCTX.r8],  r8
    mov     qword [rax + CPUMCTX.r9],  r9
    mov     qword [rax + CPUMCTX.r10], r10
    mov     qword [rax + CPUMCTX.r11], r11
    mov     qword [rax + CPUMCTX.r12], r12
    mov     qword [rax + CPUMCTX.r13], r13
    mov     qword [rax + CPUMCTX.r14], r14
    mov     qword [rax + CPUMCTX.r15], r15

    ; Restore host general purpose registers.
    MYPOPAD

    mov     eax, VINF_SUCCESS

    popf
    pop     rbp
    add     rsp, 4 * xCB
    ret
ENDPROC MY_NAME(SVMR0VMRun64)
%endif ; RT_ARCH_AMD64