VirtualBox

Changeset 87344 in vbox


Ignore:
Timestamp:
Jan 21, 2021 11:36:03 AM (4 years ago)
Author:
vboxsync
Message:

VMM/HMR0A.asm: Moved hmR0SVMRunWrapXMM next to the function it is supposed to wrap for better mojo.

File:
1 edited

Legend:

Unmodified
Added
Removed

  • trunk/src/VBox/VMM/VMMR0/HMR0A.asm

    r87337 r87344  
    447447
    448448%ifdef VBOX_WITH_KERNEL_USING_XMM
    449 
    450449;;
    451450; Wrapper around vmx.pfnStartVM that preserves host XMM registers and
    … …  
    628627        jmp     .restore_non_volatile_host_xmm_regs
    629628ENDPROC   hmR0VMXStartVMWrapXMM
    630 
    631 ;;
    632 ; Wrapper around svm.pfnVMRun that preserves host XMM registers and
    633 ; load the guest ones when necessary.
    634 ;
    635 ; @cproto       DECLASM(int) hmR0SVMRunWrapXMM(PVM pVM, PVMCPU pVCpu, RTHCPHYS HCPhysVmcb, PFNHMSVMVMRUN pfnVMRun);
    636 ;
    637 ; @returns      eax
    638 ;
    639 ; @param        pVM             msc:rcx
    640 ; @param        pVCpu           msc:rdx        The cross context virtual CPU structure of the calling EMT.
    641 ; @param        HCPhysVmcb      msc:r8
    642 ; @param        pfnVMRun        msc:r9
    643 ;
    644 ; @remarks      This is essentially the same code as hmR0VMXStartVMWrapXMM, only the parameters differ a little bit.
    645 ;
    646 ; @remarks      Drivers shouldn't use AVX registers without saving+loading:
    647 ;                   https://msdn.microsoft.com/en-us/library/windows/hardware/ff545910%28v=vs.85%29.aspx?f=255&MSPPError=-2147217396
    648 ;               However the compiler docs have different idea:
    649 ;                   https://msdn.microsoft.com/en-us/library/9z1stfyw.aspx
    650 ;               We'll go with the former for now.
    651 ;
    652 ; ASSUMING 64-bit and windows for now.
    653 ALIGNCODE(64)
    654 BEGINPROC hmR0SVMRunWrapXMM
    655         SEH64_END_PROLOGUE
    656         push    xBP
    657         mov     xBP, xSP
    658         sub     xSP, 0b0h + 040h        ; don't bother optimizing the frame size
    659 
    660 %ifndef ASM_CALL64_MSC
    661  %error "MSC only"
    662 %endif
    663         ; Spill input parameters.
    664         mov     [xBP + 010h], rcx       ; pVM
    665         mov     [xBP + 018h], rdx       ; pVCpu
    666         mov     [xBP + 020h], r8        ; HCPhysVmcb
    667         mov     [xBP + 028h], r9        ; pfnVMRun
    668 
    669         ; Ask CPUM whether we've started using the FPU yet.
    670 ;; @todo implement this in assembly, it's just checking a couple of things. Or have the C code do it.
    671         mov     rcx, rdx                ; pVCpu
    672         call    NAME(CPUMIsGuestFPUStateActive)
    673         test    al, al
    674 
    675         mov     rcx, [xBP + 010h]       ; pVM
    676         mov     rdx, [xBP + 018h]       ; pVCpu
    677         mov     r8,  [xBP + 020h]       ; HCPhysVmcb
    678         mov     r9,  [xBP + 028h]       ; pfnVMRun
    679 
    680         jnz     .guest_fpu_state_active
    681 
    682         ; No need to mess with XMM registers just call the start routine and return.
    683         call    r9
    684 
    685         leave
    686         ret
    687 
    688 ALIGNCODE(8)
    689 .guest_fpu_state_active:
    690         ; Save the non-volatile host XMM registers.
    691 ;; @todo change to rbp relative addressing as that saves a byte per instruction!
    692         movdqa  [rsp + 040h + 000h], xmm6
    693         movdqa  [rsp + 040h + 010h], xmm7
    694         movdqa  [rsp + 040h + 020h], xmm8
    695         movdqa  [rsp + 040h + 030h], xmm9
    696         movdqa  [rsp + 040h + 040h], xmm10
    697         movdqa  [rsp + 040h + 050h], xmm11
    698         movdqa  [rsp + 040h + 060h], xmm12
    699         movdqa  [rsp + 040h + 070h], xmm13
    700         movdqa  [rsp + 040h + 080h], xmm14
    701         movdqa  [rsp + 040h + 090h], xmm15
    702         stmxcsr [rsp + 040h + 0a0h]
    703 
    704         mov     r11, rdx                ; r11 = pVCpu (rdx may get trashed)
    705         mov     eax, [rdx + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask]
    706         test    eax, eax
    707         jz      .guest_fpu_state_manually
    708 
    709         ;
    710         ; Using XSAVE.
    711         ;
    712         and     eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
    713         xor     edx, edx
    714         mov     r10, [r11 + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
    715         xrstor  [r10]
    716 
    717         ; Make the call (same as in the other case).
    718         mov     rdx, r11                ; restore pVCpu to rdx
    719         call    r9
    720 
    721         mov     r10d, eax               ; save return value (xsave below uses eax)
    722 
    723         ; Save the guest XMM registers.
    724         mov     rcx, [xBP + 018h]       ; pVCpu
    725         mov     eax, [rcx + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask]
    726         and     eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
    727         mov     rcx, [rcx + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
    728         xor     edx, edx
    729         xsave   [rcx]
    730 
    731         mov     eax, r10d               ; restore return value
    732 
    733 .restore_non_volatile_host_xmm_regs:
    734         ; Load the non-volatile host XMM registers.
    735 ;; @todo change to rbp relative addressing as that saves a byte per instruction!
    736         movdqa  xmm6,  [rsp + 040h + 000h]
    737         movdqa  xmm7,  [rsp + 040h + 010h]
    738         movdqa  xmm8,  [rsp + 040h + 020h]
    739         movdqa  xmm9,  [rsp + 040h + 030h]
    740         movdqa  xmm10, [rsp + 040h + 040h]
    741         movdqa  xmm11, [rsp + 040h + 050h]
    742         movdqa  xmm12, [rsp + 040h + 060h]
    743         movdqa  xmm13, [rsp + 040h + 070h]
    744         movdqa  xmm14, [rsp + 040h + 080h]
    745         movdqa  xmm15, [rsp + 040h + 090h]
    746         ldmxcsr [rsp + 040h + 0a0h]
    747         leave
    748         ret
    749 
    750         ;
    751         ; No XSAVE, load and save the guest XMM registers manually.
    752         ;
    753 ALIGNCODE(8)
    754 .guest_fpu_state_manually:
    755         ; Load the full guest XMM register state.
    756         mov     rdx, [r11 + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
    757         movdqa  xmm0,  [rdx + XMM_OFF_IN_X86FXSTATE + 000h]
    758         movdqa  xmm1,  [rdx + XMM_OFF_IN_X86FXSTATE + 010h]
    759         movdqa  xmm2,  [rdx + XMM_OFF_IN_X86FXSTATE + 020h]
    760         movdqa  xmm3,  [rdx + XMM_OFF_IN_X86FXSTATE + 030h]
    761         movdqa  xmm4,  [rdx + XMM_OFF_IN_X86FXSTATE + 040h]
    762         movdqa  xmm5,  [rdx + XMM_OFF_IN_X86FXSTATE + 050h]
    763         movdqa  xmm6,  [rdx + XMM_OFF_IN_X86FXSTATE + 060h]
    764         movdqa  xmm7,  [rdx + XMM_OFF_IN_X86FXSTATE + 070h]
    765         movdqa  xmm8,  [rdx + XMM_OFF_IN_X86FXSTATE + 080h]
    766         movdqa  xmm9,  [rdx + XMM_OFF_IN_X86FXSTATE + 090h]
    767         movdqa  xmm10, [rdx + XMM_OFF_IN_X86FXSTATE + 0a0h]
    768         movdqa  xmm11, [rdx + XMM_OFF_IN_X86FXSTATE + 0b0h]
    769         movdqa  xmm12, [rdx + XMM_OFF_IN_X86FXSTATE + 0c0h]
    770         movdqa  xmm13, [rdx + XMM_OFF_IN_X86FXSTATE + 0d0h]
    771         movdqa  xmm14, [rdx + XMM_OFF_IN_X86FXSTATE + 0e0h]
    772         movdqa  xmm15, [rdx + XMM_OFF_IN_X86FXSTATE + 0f0h]
    773         ldmxcsr        [rdx + X86FXSTATE.MXCSR]
    774 
    775         ; Make the call (same as in the other case).
    776         mov     rdx, r11                ; restore pVCpu to rdx
    777         call    r9
    778 
    779         ; Save the guest XMM registers.
    780         mov     rdx, [xBP + 018h]       ; pVCpu
    781         mov     rdx, [rdx + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
    782         stmxcsr [rdx + X86FXSTATE.MXCSR]
    783         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 000h], xmm0
    784         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 010h], xmm1
    785         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 020h], xmm2
    786         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 030h], xmm3
    787         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 040h], xmm4
    788         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 050h], xmm5
    789         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 060h], xmm6
    790         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 070h], xmm7
    791         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 080h], xmm8
    792         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 090h], xmm9
    793         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0a0h], xmm10
    794         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0b0h], xmm11
    795         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0c0h], xmm12
    796         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0d0h], xmm13
    797         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0e0h], xmm14
    798         movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0f0h], xmm15
    799         jmp     .restore_non_volatile_host_xmm_regs
    800 ENDPROC   hmR0SVMRunWrapXMM
    801 
    802629%endif ; VBOX_WITH_KERNEL_USING_XMM
    803630
    … …  
    1089916
    1090917
     918%ifdef VBOX_WITH_KERNEL_USING_XMM
     919;;
     920; Wrapper around svm.pfnVMRun that preserves host XMM registers and
     921; load the guest ones when necessary.
     922;
     923; @cproto       DECLASM(int) hmR0SVMRunWrapXMM(PVM pVM, PVMCPU pVCpu, RTHCPHYS HCPhysVmcb, PFNHMSVMVMRUN pfnVMRun);
     924;
     925; @returns      eax
     926;
     927; @param        pVM             msc:rcx
     928; @param        pVCpu           msc:rdx        The cross context virtual CPU structure of the calling EMT.
     929; @param        HCPhysVmcb      msc:r8
     930; @param        pfnVMRun        msc:r9
     931;
     932; @remarks      This is essentially the same code as hmR0VMXStartVMWrapXMM, only the parameters differ a little bit.
     933;
     934; @remarks      Drivers shouldn't use AVX registers without saving+loading:
     935;                   https://msdn.microsoft.com/en-us/library/windows/hardware/ff545910%28v=vs.85%29.aspx?f=255&MSPPError=-2147217396
     936;               However the compiler docs have different idea:
     937;                   https://msdn.microsoft.com/en-us/library/9z1stfyw.aspx
     938;               We'll go with the former for now.
     939;
     940; ASSUMING 64-bit and windows for now.
     941ALIGNCODE(64)
     942BEGINPROC hmR0SVMRunWrapXMM
     943        SEH64_END_PROLOGUE
     944        push    xBP
     945        mov     xBP, xSP
     946        sub     xSP, 0b0h + 040h        ; don't bother optimizing the frame size
     947
     948%ifndef ASM_CALL64_MSC
     949 %error "MSC only"
     950%endif
     951        ; Spill input parameters.
     952        mov     [xBP + 010h], rcx       ; pVM
     953        mov     [xBP + 018h], rdx       ; pVCpu
     954        mov     [xBP + 020h], r8        ; HCPhysVmcb
     955        mov     [xBP + 028h], r9        ; pfnVMRun
     956
     957        ; Ask CPUM whether we've started using the FPU yet.
     958;; @todo implement this in assembly, it's just checking a couple of things. Or have the C code do it.
     959        mov     rcx, rdx                ; pVCpu
     960        call    NAME(CPUMIsGuestFPUStateActive)
     961        test    al, al
     962
     963        mov     rcx, [xBP + 010h]       ; pVM
     964        mov     rdx, [xBP + 018h]       ; pVCpu
     965        mov     r8,  [xBP + 020h]       ; HCPhysVmcb
     966        mov     r9,  [xBP + 028h]       ; pfnVMRun
     967
     968        jnz     .guest_fpu_state_active
     969
     970        ; No need to mess with XMM registers just call the start routine and return.
     971        call    r9
     972
     973        leave
     974        ret
     975
     976ALIGNCODE(8)
     977.guest_fpu_state_active:
     978        ; Save the non-volatile host XMM registers.
     979;; @todo change to rbp relative addressing as that saves a byte per instruction!
     980        movdqa  [rsp + 040h + 000h], xmm6
     981        movdqa  [rsp + 040h + 010h], xmm7
     982        movdqa  [rsp + 040h + 020h], xmm8
     983        movdqa  [rsp + 040h + 030h], xmm9
     984        movdqa  [rsp + 040h + 040h], xmm10
     985        movdqa  [rsp + 040h + 050h], xmm11
     986        movdqa  [rsp + 040h + 060h], xmm12
     987        movdqa  [rsp + 040h + 070h], xmm13
     988        movdqa  [rsp + 040h + 080h], xmm14
     989        movdqa  [rsp + 040h + 090h], xmm15
     990        stmxcsr [rsp + 040h + 0a0h]
     991
     992        mov     r11, rdx                ; r11 = pVCpu (rdx may get trashed)
     993        mov     eax, [rdx + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask]
     994        test    eax, eax
     995        jz      .guest_fpu_state_manually
     996
     997        ;
     998        ; Using XSAVE.
     999        ;
     1000        and     eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
     1001        xor     edx, edx
     1002        mov     r10, [r11 + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
     1003        xrstor  [r10]
     1004
     1005        ; Make the call (same as in the other case).
     1006        mov     rdx, r11                ; restore pVCpu to rdx
     1007        call    r9
     1008
     1009        mov     r10d, eax               ; save return value (xsave below uses eax)
     1010
     1011        ; Save the guest XMM registers.
     1012        mov     rcx, [xBP + 018h]       ; pVCpu
     1013        mov     eax, [rcx + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask]
     1014        and     eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
     1015        mov     rcx, [rcx + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
     1016        xor     edx, edx
     1017        xsave   [rcx]
     1018
     1019        mov     eax, r10d               ; restore return value
     1020
     1021.restore_non_volatile_host_xmm_regs:
     1022        ; Load the non-volatile host XMM registers.
     1023;; @todo change to rbp relative addressing as that saves a byte per instruction!
     1024        movdqa  xmm6,  [rsp + 040h + 000h]
     1025        movdqa  xmm7,  [rsp + 040h + 010h]
     1026        movdqa  xmm8,  [rsp + 040h + 020h]
     1027        movdqa  xmm9,  [rsp + 040h + 030h]
     1028        movdqa  xmm10, [rsp + 040h + 040h]
     1029        movdqa  xmm11, [rsp + 040h + 050h]
     1030        movdqa  xmm12, [rsp + 040h + 060h]
     1031        movdqa  xmm13, [rsp + 040h + 070h]
     1032        movdqa  xmm14, [rsp + 040h + 080h]
     1033        movdqa  xmm15, [rsp + 040h + 090h]
     1034        ldmxcsr [rsp + 040h + 0a0h]
     1035        leave
     1036        ret
     1037
     1038        ;
     1039        ; No XSAVE, load and save the guest XMM registers manually.
     1040        ;
     1041ALIGNCODE(8)
     1042.guest_fpu_state_manually:
     1043        ; Load the full guest XMM register state.
     1044        mov     rdx, [r11 + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
     1045        movdqa  xmm0,  [rdx + XMM_OFF_IN_X86FXSTATE + 000h]
     1046        movdqa  xmm1,  [rdx + XMM_OFF_IN_X86FXSTATE + 010h]
     1047        movdqa  xmm2,  [rdx + XMM_OFF_IN_X86FXSTATE + 020h]
     1048        movdqa  xmm3,  [rdx + XMM_OFF_IN_X86FXSTATE + 030h]
     1049        movdqa  xmm4,  [rdx + XMM_OFF_IN_X86FXSTATE + 040h]
     1050        movdqa  xmm5,  [rdx + XMM_OFF_IN_X86FXSTATE + 050h]
     1051        movdqa  xmm6,  [rdx + XMM_OFF_IN_X86FXSTATE + 060h]
     1052        movdqa  xmm7,  [rdx + XMM_OFF_IN_X86FXSTATE + 070h]
     1053        movdqa  xmm8,  [rdx + XMM_OFF_IN_X86FXSTATE + 080h]
     1054        movdqa  xmm9,  [rdx + XMM_OFF_IN_X86FXSTATE + 090h]
     1055        movdqa  xmm10, [rdx + XMM_OFF_IN_X86FXSTATE + 0a0h]
     1056        movdqa  xmm11, [rdx + XMM_OFF_IN_X86FXSTATE + 0b0h]
     1057        movdqa  xmm12, [rdx + XMM_OFF_IN_X86FXSTATE + 0c0h]
     1058        movdqa  xmm13, [rdx + XMM_OFF_IN_X86FXSTATE + 0d0h]
     1059        movdqa  xmm14, [rdx + XMM_OFF_IN_X86FXSTATE + 0e0h]
     1060        movdqa  xmm15, [rdx + XMM_OFF_IN_X86FXSTATE + 0f0h]
     1061        ldmxcsr        [rdx + X86FXSTATE.MXCSR]
     1062
     1063        ; Make the call (same as in the other case).
     1064        mov     rdx, r11                ; restore pVCpu to rdx
     1065        call    r9
     1066
     1067        ; Save the guest XMM registers.
     1068        mov     rdx, [xBP + 018h]       ; pVCpu
     1069        mov     rdx, [rdx + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
     1070        stmxcsr [rdx + X86FXSTATE.MXCSR]
     1071        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 000h], xmm0
     1072        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 010h], xmm1
     1073        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 020h], xmm2
     1074        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 030h], xmm3
     1075        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 040h], xmm4
     1076        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 050h], xmm5
     1077        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 060h], xmm6
     1078        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 070h], xmm7
     1079        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 080h], xmm8
     1080        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 090h], xmm9
     1081        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0a0h], xmm10
     1082        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0b0h], xmm11
     1083        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0c0h], xmm12
     1084        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0d0h], xmm13
     1085        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0e0h], xmm14
     1086        movdqa  [rdx + XMM_OFF_IN_X86FXSTATE + 0f0h], xmm15
     1087        jmp     .restore_non_volatile_host_xmm_regs
     1088ENDPROC   hmR0SVMRunWrapXMM
     1089
     1090%endif ; VBOX_WITH_KERNEL_USING_XMM
     1091
    10911092;;
    10921093; Prepares for and executes VMRUN (32-bit and 64-bit guests).
Note: See TracChangeset for help on using the changeset viewer.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette