; x64 POSIX vfork continuation. The parent enters the resume/wait syscall without
; a nested userspace return frame while the child owns the caller's stack.

INCLUDE syscall_nr.inc

VF_ReturnAddress       EQU 0
VF_ParentRsp           EQU 8
VF_Rbx                 EQU 16
VF_Rbp                 EQU 24
VF_Rsi                 EQU 32
VF_Rdi                 EQU 40
VF_R12                 EQU 48
VF_R13                 EQU 56
VF_R14                 EQU 64
VF_R15                 EQU 72
VF_Xmm6                 EQU 80
VF_Xmm7                 EQU 96
VF_Xmm8                 EQU 112
VF_Xmm9                 EQU 128
VF_Xmm10                EQU 144
VF_Xmm11                EQU 160
VF_Xmm12                EQU 176
VF_Xmm13                EQU 192
VF_Xmm14                EQU 208
VF_Xmm15                EQU 224
VF_Mxcsr                EQU 240
VF_X87Control           EQU 244
VF_ThreadHandle         EQU 256
VF_PreviousSuspendCount EQU 264
VF_ContextSize          EQU 280
VF_LocalSize            EQU 312

EXTERN PosixVforkPrepare:PROC
EXTERN PosixVforkParentFinish:PROC

.CODE

PUBLIC vfork
vfork PROC EXPORT
    sub     rsp, VF_LocalSize
    lea     r10, [rsp + 32]

    mov     rax, [rsp + VF_LocalSize]
    mov     [r10 + VF_ReturnAddress], rax
    lea     rax, [rsp + VF_LocalSize + 8]
    mov     [r10 + VF_ParentRsp], rax
    mov     [r10 + VF_Rbx], rbx
    mov     [r10 + VF_Rbp], rbp
    mov     [r10 + VF_Rsi], rsi
    mov     [r10 + VF_Rdi], rdi
    mov     [r10 + VF_R12], r12
    mov     [r10 + VF_R13], r13
    mov     [r10 + VF_R14], r14
    mov     [r10 + VF_R15], r15
    movdqu  xmmword ptr [r10 + VF_Xmm6], xmm6
    movdqu  xmmword ptr [r10 + VF_Xmm7], xmm7
    movdqu  xmmword ptr [r10 + VF_Xmm8], xmm8
    movdqu  xmmword ptr [r10 + VF_Xmm9], xmm9
    movdqu  xmmword ptr [r10 + VF_Xmm10], xmm10
    movdqu  xmmword ptr [r10 + VF_Xmm11], xmm11
    movdqu  xmmword ptr [r10 + VF_Xmm12], xmm12
    movdqu  xmmword ptr [r10 + VF_Xmm13], xmm13
    movdqu  xmmword ptr [r10 + VF_Xmm14], xmm14
    movdqu  xmmword ptr [r10 + VF_Xmm15], xmm15
    stmxcsr dword ptr [r10 + VF_Mxcsr]
    fnstcw  word ptr [r10 + VF_X87Control]

    mov     rcx, r10
    call    PosixVforkPrepare
    test    rax, rax
    jnz     vfork_wait
    add     rsp, VF_LocalSize
    mov     eax, -1
    ret

vfork_wait:
    ; R12 deliberately holds the heap context across the kernel wait. The child
    ; restores the caller's original R12 in its independent native context.
    mov     r12, rax
    mov     ecx, dword ptr [r12 + VF_ThreadHandle]
    lea     rdx, [r12 + VF_PreviousSuspendCount]
    mov     eax, SYS_RESUME_THREAD_AND_WAIT
    int     40h
    mov     r9d, eax

    ; The child may have overwritten every byte of the old vfork/helper frames.
    ; Rebuild only a synthetic return frame from the heap continuation.
    mov     r10, r12
    mov     r11, [r10 + VF_ParentRsp]
    mov     rax, [r10 + VF_ReturnAddress]
    mov     rsp, r11
    sub     rsp, 8
    mov     [rsp], rax
    sub     rsp, 40

    ldmxcsr dword ptr [r10 + VF_Mxcsr]
    fldcw   word ptr [r10 + VF_X87Control]
    movdqu  xmm6, xmmword ptr [r10 + VF_Xmm6]
    movdqu  xmm7, xmmword ptr [r10 + VF_Xmm7]
    movdqu  xmm8, xmmword ptr [r10 + VF_Xmm8]
    movdqu  xmm9, xmmword ptr [r10 + VF_Xmm9]
    movdqu  xmm10, xmmword ptr [r10 + VF_Xmm10]
    movdqu  xmm11, xmmword ptr [r10 + VF_Xmm11]
    movdqu  xmm12, xmmword ptr [r10 + VF_Xmm12]
    movdqu  xmm13, xmmword ptr [r10 + VF_Xmm13]
    movdqu  xmm14, xmmword ptr [r10 + VF_Xmm14]
    movdqu  xmm15, xmmword ptr [r10 + VF_Xmm15]
    mov     rbx, [r10 + VF_Rbx]
    mov     rbp, [r10 + VF_Rbp]
    mov     rsi, [r10 + VF_Rsi]
    mov     rdi, [r10 + VF_Rdi]
    mov     r12, [r10 + VF_R12]
    mov     r13, [r10 + VF_R13]
    mov     r14, [r10 + VF_R14]
    mov     r15, [r10 + VF_R15]

    mov     rcx, r10
    mov     edx, r9d
    call    PosixVforkParentFinish
    add     rsp, 40
    ret
vfork ENDP

PUBLIC PosixVforkEnterChild
PosixVforkEnterChild PROC
    mov     r10, rcx
    mov     r11, [r10 + VF_ParentRsp]
    mov     rax, [r10 + VF_ReturnAddress]
    mov     rsp, r11
    sub     rsp, 8
    mov     [rsp], rax

    ldmxcsr dword ptr [r10 + VF_Mxcsr]
    fldcw   word ptr [r10 + VF_X87Control]
    movdqu  xmm6, xmmword ptr [r10 + VF_Xmm6]
    movdqu  xmm7, xmmword ptr [r10 + VF_Xmm7]
    movdqu  xmm8, xmmword ptr [r10 + VF_Xmm8]
    movdqu  xmm9, xmmword ptr [r10 + VF_Xmm9]
    movdqu  xmm10, xmmword ptr [r10 + VF_Xmm10]
    movdqu  xmm11, xmmword ptr [r10 + VF_Xmm11]
    movdqu  xmm12, xmmword ptr [r10 + VF_Xmm12]
    movdqu  xmm13, xmmword ptr [r10 + VF_Xmm13]
    movdqu  xmm14, xmmword ptr [r10 + VF_Xmm14]
    movdqu  xmm15, xmmword ptr [r10 + VF_Xmm15]
    mov     rbx, [r10 + VF_Rbx]
    mov     rbp, [r10 + VF_Rbp]
    mov     rsi, [r10 + VF_Rsi]
    mov     rdi, [r10 + VF_Rdi]
    mov     r12, [r10 + VF_R12]
    mov     r13, [r10 + VF_R13]
    mov     r14, [r10 + VF_R14]
    mov     r15, [r10 + VF_R15]
    xor     eax, eax
    ret
PosixVforkEnterChild ENDP

END