/*
 * switch.S (aarch64) - Cooperative VM context switch primitive.
 *
 * FABRIC-3.md §XXVIII (Stage 2, preemptive-context-switch save/restore
 * primitive, 2026-09-13). Deliberately NOT isr.S's interrupt trap frame
 * (Stage 0) -- that saves everything because an interrupt has no call
 * site. sk_vm_switch_to() is an ordinary `bl`, so AAPCS64 already makes
 * every caller-saved register (x0-x18, v0-v7, v16-v31) the caller's own
 * problem across the call. Only the callee-saved set needs explicit
 * handling: x19-x28, x29 (fp), x30 (lr), and d8-d15 (only the low 64
 * bits of v8-v15 are callee-saved per AAPCS64).
 *
 * void sk_vm_switch_to(uint64_t *save_sp_out, uint64_t new_sp);
 *   x0 = save_sp_out, x1 = new_sp
 *
 * For a VM that has never run before, new_sp must point at a synthetic
 * 160-byte frame built by the C glue (switch.c) to this exact layout,
 * with x30 (lr) set to the trampoline entry address and everything else
 * zeroed, so the final `ret` below (which branches to lr) jumps there.
 */

    .section .text
    .balign 4

    .equ SW_X19, 0
    .equ SW_X21, 16
    .equ SW_X23, 32
    .equ SW_X25, 48
    .equ SW_X27, 64
    .equ SW_X29, 80   /* x29 (fp), x30 (lr) */
    .equ SW_D8,  96
    .equ SW_D10, 112
    .equ SW_D12, 128
    .equ SW_D14, 144
    .equ SW_SIZE, 160

/* .hidden is an ELF-only visibility directive with no COFF equivalent;
 * this file is also assembled for the aarch64 loader, which clang targets
 * at aarch64-pc-windows-msvc (PE/COFF) -- guard on _WIN32 (defined only by
 * that target), matching isr.S's own precedent exactly. */
#ifndef _WIN32
.hidden sk_vm_switch_to
#endif

    .global sk_vm_switch_to
sk_vm_switch_to:
    sub sp, sp, #SW_SIZE

    stp x19, x20, [sp, #SW_X19]
    stp x21, x22, [sp, #SW_X21]
    stp x23, x24, [sp, #SW_X23]
    stp x25, x26, [sp, #SW_X25]
    stp x27, x28, [sp, #SW_X27]
    stp x29, x30, [sp, #SW_X29]
    stp d8,  d9,  [sp, #SW_D8]
    stp d10, d11, [sp, #SW_D10]
    stp d12, d13, [sp, #SW_D12]
    stp d14, d15, [sp, #SW_D14]

    mov x2, sp
    str x2, [x0]         /* *save_sp_out = sp (this context's parked SP) */

    mov sp, x1            /* sp = new_sp (switch stacks) */

    ldp x19, x20, [sp, #SW_X19]
    ldp x21, x22, [sp, #SW_X21]
    ldp x23, x24, [sp, #SW_X23]
    ldp x25, x26, [sp, #SW_X25]
    ldp x27, x28, [sp, #SW_X27]
    ldp x29, x30, [sp, #SW_X29]
    ldp d8,  d9,  [sp, #SW_D8]
    ldp d10, d11, [sp, #SW_D10]
    ldp d12, d13, [sp, #SW_D12]
    ldp d14, d15, [sp, #SW_D14]

    add sp, sp, #SW_SIZE
    ret

/*
 * uint64_t sk_vm_switch_prime(uint64_t stack_top, void *entry);
 *
 * Builds the synthetic frame above at the top of a fresh native stack, so
 * the first ever sk_vm_switch_to() into it lands at `entry` (x30/lr) via
 * its final `ret`. Kept in this file (not C) so the layout can never
 * drift from sk_vm_switch_to() above -- one set of offsets, not two.
 */
#ifndef _WIN32
.hidden sk_vm_switch_prime
#endif

    .global sk_vm_switch_prime
sk_vm_switch_prime:
    sub x2, x0, #SW_SIZE
    stp xzr, xzr, [x2, #SW_X19]
    stp xzr, xzr, [x2, #SW_X21]
    stp xzr, xzr, [x2, #SW_X23]
    stp xzr, xzr, [x2, #SW_X25]
    stp xzr, xzr, [x2, #SW_X27]
    stp xzr, x1,  [x2, #SW_X29]   /* x29=0, x30(lr)=entry */
    stp xzr, xzr, [x2, #SW_D8]
    stp xzr, xzr, [x2, #SW_D10]
    stp xzr, xzr, [x2, #SW_D12]
    stp xzr, xzr, [x2, #SW_D14]
    mov x0, x2
    ret