windows xp avx enabler

use32
org                   $47023C

Continue_46C54A   equ $46C54A
Continue_46C48F   equ $46C48F
AVXEnable         equ $4760E5
KiIpiGenericCall  equ $42B034

macro LOAD_AX arg1
{
    call    GetPIC_EAX
    add     eax, (arg1 - base1)
}


macro avxsave reg0 {
    VEXTRACTF128 [reg0+$120], ymm0,1
    VEXTRACTF128 [reg0+$130], ymm1,1
    VEXTRACTF128 [reg0+$140], ymm2,1
    VEXTRACTF128 [reg0+$150], ymm3,1
    VEXTRACTF128 [reg0+$160], ymm4,1
    VEXTRACTF128 [reg0+$170], ymm5,1
    VEXTRACTF128 [reg0+$180], ymm6,1
    VEXTRACTF128 [reg0+$190], ymm7,1
}

macro avxload reg0 {
    VBROADCASTF128 ymm0, [reg0+$120]
    VBROADCASTF128 ymm1, [reg0+$130]
    VBROADCASTF128 ymm2, [reg0+$140]
    VBROADCASTF128 ymm3, [reg0+$150]
    VBROADCASTF128 ymm4, [reg0+$160]
    VBROADCASTF128 ymm5, [reg0+$170]
    VBROADCASTF128 ymm6, [reg0+$180]
    VBROADCASTF128 ymm7, [reg0+$190]
}

Hook_427DA4:      ;47023C
        jmp short Save_eax_427DA4
Hook_5CE746:      ;47023E
        jmp short Save_eax_5CE746
Hook_46C542:      ;470240
        jmp short Save_ecx_46C542
Hook_46DDB7:      ;470242
        jmp short Save_ecx_46DDB7
Hook_46DF41:      ;470244
        jmp short Save_ecx_46DF41
Hook_46DF8D:      ;470246
        jmp short Save_ecx_46DF8D
Hook_46ED62:      ;470248
        jmp short Save_ecx_46ED62
Empty_Hook:       ;47024A
        jmp short $
Hook_42DBCE:      ;47024C
        jmp short Load_eax_42DBCE
Hook_46C488:      ;47024E
        jmp short Load_ecx_46C488
Hook_46DF51:      ;470250
        jmp short Load_ecx_46DF51
Init_Hook:        ;470252
        jmp short Init

Save_eax_427DA4:
Save_eax_427D61:
    mov     eax, [ebp-$8]
    jmp     Save_Eax

Save_eax_5CE746:
    mov     eax, [ebp-$18]
    jmp     Save_Eax

Save_ecx_46C542:
    call    Save_Ecx
    jmp     Continue_46C54A

Save_ecx_46DDB7:
    call    Save_Ecx
    or      ebx, $0A
    ret

Save_ecx_46DF41:
    call    Save_Ecx
    mov     byte [eax+$31], $0A
    ret

Save_ecx_46DF8D:
    call    Save_Ecx
    mov     ecx, [esp+$14+$4] ; +4 - orig stack
    ret

Save_ecx_46ED62:
    call    Save_Ecx
    mov     byte [edi+$31], $0A
    ret

Load_eax_42DBCE:
    mov     cr0, ebx
    call    Load_Eax
    ret

Load_ecx_46C488:
    call    Load_Ecx
    jmp     Continue_46C48F

Load_ecx_46DF51:
    call    Load_Ecx
    mov     edx, [ecx+$20C]
    ret




Save_Eax:
    call    check_avx
    jnz     Save_Eax_part2
    fxsave  [eax]
    ret
;Save_Eax part1

Init:
    ;eax = NumberProcessors
    mov     [ebp-4], eax ; stolen bytes, ebp-4 = &NumberProcessors

    push    ebx
    push    edi

    call    GetPIC_AVXEnable
    mov     edi,eax
    xor     eax,eax
    mov     byte [edi],al

    ; cpuid
    mov     al, 1                 ; fxn: get features
    cpuid
    bt      ecx, 28               ; AVX support by CPU?
    jnc     L1

    xor     ecx, ecx
    xor     eax, eax
    mov     al, $0d
    cpuid
    bt      eax, 2               ; XCR0.AVX can be set to 1 ?
    jnc     L1

    lea     eax, [ebp-4]
    push    eax
    LOAD_AX EnableAvxPerCPU
    push    eax
    call    KiIpiGenericCall
    mov     byte [edi], $1

L1:
    pop     edi
    pop     ebx

    ;stolen bytes
    lea     eax, [ebp-4]
    ret
;Init

Save_Eax_part2:
    fxsave  [eax]
    avxsave eax
    ret
;Save_Eax_part2



GetPIC_EAX:
    call    $+5
base1:
    pop     eax
    ret
;GetPIC_EAX


GetPIC_AVXEnable:
    LOAD_AX AVXEnable
    ret
;GetPIC_AVXEnable


Save_Ecx:
    push    eax
    mov     eax, ecx
    call    Save_Eax
    pop     eax
    ret
;Save_Ecx


Load_Ecx:
    push    eax
    mov     eax, ecx
    call    Load_Eax
    pop     eax
    ret
;Load_Ecx


Load_Eax:
    call check_avx
    jz      L3
    avxload eax
L3:
    fxrstor [eax]
    ret
;Load_Eax


check_avx:
    push    eax
    call    GetPIC_AVXEnable
    test    byte [eax],$01
    pop     eax
    ret
;check_avx


EnableAvxPerCPU:
    ; cr4.osxsave
    mov     eax, cr4              ; load cr4
    or      eax, $00040000        ; set bit 18 OSXSAVE
    mov     cr4, eax              ; write new content
    ; XCR0.FPU+SSE+AVX
    xor     ecx, ecx              ; XFEATURE_ENABLED_MASK register
    xor     edx, edx
    xor     eax, eax
    mov     al , $7               ; mask in edx:eax, fpu+sse+avx
    xsetbv
    ret     4
;EnableAvxPerCPU



EndAdress:
db  ($470400 - $) dup 0