ARM64 threads using Linux clone3 syscall

thread.S

.equ SYS_clone3, 435
.equ SYS_write, 64
.equ SYS_exit, 93
.equ CHILD_STACK_SIZE, 0x4000           // 16 KB per child
.equ CLONE_VM, 0x00000100
.equ CLONE_FS, 0x00000200
.equ CLONE_FILES, 0x00000400
.equ CLONE_SIGHAND, 0x00000800
.equ CLONE_PARENT, 0x00008000
.equ CLONE_THREAD, 0x00010000
.equ CLONE_IO, 0x80000000
.equ CLONE_FLAGS, 0x80018f00
.equ STDOUT_FILENO, 1
.equ NUM_THREADS, 32

.section .data
hello_msg: .ascii "Hello from \x1b[1;31mthread\x1b[0m!\n"
hello_msg_len = . - hello_msg
parent_msg: .ascii "Hello from \x1b[1;33mmain\x1b[0m!\n"
parent_msg_len = . - parent_msg

.section .bss
.align 16
// Stack space for child threads
child_stacks: .space (NUM_THREADS * CHILD_STACK_SIZE)
clone_args: .space 88               // clone_args structure

.text
.global _start

_start:
    // Create 3 threads
    mov x19, #0                         // Thread counter
    
create_thread_loop:
    cmp x19, #NUM_THREADS
    b.ge wait_loop
    
    // Create thread with current thread number
    mov x0, x19
    bl create_thread_clone3
    
    add x19, x19, #1
    b create_thread_loop

wait_loop:
    // Simple busy wait for threads to complete
    mov x0, #STDOUT_FILENO
    adr x1, parent_msg
    mov x2, #parent_msg_len
    mov x8, #SYS_write
    svc #0
    mov x0, #0x20000000
wait_inner:
    sub x0, x0, #1
    cbnz x0, wait_inner
    
    // Exit program
    mov x0, #0                          // Exit status
    mov x8, #SYS_exit
    svc #0

// Create thread using clone3 syscall
// x0 = thread ID
create_thread_clone3:
    // Save thread ID and registers
    sub sp, sp, #32
    // str x0, [sp]                        // Save thread ID
    str lr, [sp, #8]                   // Save return address
    
    // Calculate this thread's stack area
    mov x1, #CHILD_STACK_SIZE
    mul x1, x0, x1                      // thread_id * CHILD_STACK_SIZE
    adr x2, child_stacks
    add x20, x2, x1                     // x20 = base of this thread's stack
    
    // sub x20, x20, #32                   // Leave space for thread data
    
    // Store thread function and ID at the top of stack
    str x0, [x20]                   // Store thread ID
    //mov x0, x21
    // str x20, [x20, #16]                 // Store stack base for child
    
    // Set up clone_args structure
    adr x16, clone_args                 // Get address of clone_args
    ldr x2, =CLONE_FLAGS
    str x2, [x16]                       // flags (offset 0)
    str x20, [x16, #40]                 // stack (offset 40) - stack top
    mov x2, #CHILD_STACK_SIZE
    str x2, [x16, #48]                  // stack_size (offset 48)
    // Call clone3
    mov x0, x16                         // clone_args pointer
    mov x1, #88                         // size of clone_args structure
    mov x8, #SYS_clone3
    svc #0
    
    // Check if we're in the child process
    cmp x0, #0
    b.eq child_process
    
    // Parent process - restore and return
    ldr lr, [sp, #8]                   // Restore return address
    add sp, sp, #32
    ret

child_process:
    // We're in the child thread
    // Set up our stack pointer
    adr x2, thread_func
    // ldr x2, [x20]                       // Load function pointer
    ldr x19, [x20]                  // Load thread ID
    // Call the thread function
    blr x2                              // Branch to function pointer
    // Exit thread
    mov x0, x20
    mov x8, #SYS_exit
    svc #0

// Thread function - prints "Hello X" where X is thread ID
thread_func:
    // Thread ID is in x19 register
    add w4, w19, #'0'             // Add ASCII offset (48)
    // add x1, sp, #1
    // str x4, [sp, #32] 
    // add x1, sp, #-128
    strb w4, [sp, -16]
    //brk #0
    // strb w4, [sp, #16]
    mov x0, #STDOUT_FILENO
    add x1, sp, #-16
    mov x2, #1
    mov x8, #SYS_write
    svc #0
    // Print "Hello "
    mov x0, #STDOUT_FILENO
    adr x1, hello_msg
    mov x2, #hello_msg_len
    mov x8, #SYS_write
    svc #0
    ret

thread_mmap.S

.equ SYS_clone3, 435
.equ SYS_mmap, 222
.equ SYS_write, 64
.equ SYS_exit, 93
.equ PROT_READ, 1
.equ PROT_WRITE, 2
.equ MAP_PRIVATE, 2
.equ MAP_ANONYMOUS, 32
.equ MAP_GROWSDOWN, 256
.equ STACK_SIZE, (4096 * 1024)    // 4 MB mmap stack
.equ CLONE_VM, 0x00000100
.equ CLONE_FS, 0x00000200
.equ CLONE_FILES, 0x00000400
.equ CLONE_SIGHAND, 0x00000800
.equ CLONE_PARENT, 0x00008000
.equ CLONE_THREAD, 0x00010000
.equ CLONE_IO, 0x80000000
.equ CLONE_FLAGS, (CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | CLONE_PARENT | CLONE_THREAD | CLONE_IO) //  0x80018f00
.equ STDOUT_FILENO, 1
.equ NUM_THREADS, 32

.section .data
hello_msg: .ascii "Hello from \x1b[1;31mthread\x1b[0m!\n"
hello_msg_len = . - hello_msg
parent_msg: .ascii "Hello from \x1b[1;33mmain\x1b[0m!\n"
parent_msg_len = . - parent_msg

.section .bss
clone_args: .space 88               // clone_args structure

.text
.global _start

_start:
    mov x19, #0                         // Thread counter
    
create_thread_loop:
    cmp x19, #NUM_THREADS
    b.ge wait_loop
    
    // Create thread with current thread number
    mov x0, x19
    bl create_thread_clone3
    
    add x19, x19, #1
    b create_thread_loop

wait_loop:
    // Simple busy wait for threads to complete
    mov x0, #STDOUT_FILENO
    adr x1, parent_msg
    mov x2, #parent_msg_len
    mov x8, #SYS_write
    svc #0
    mov x0, #0x20000000
wait_inner:
    sub x0, x0, #1
    cbnz x0, wait_inner
    
    // Exit program
    mov x0, #0                          // Exit status
    mov x8, #SYS_exit
    svc #0

stack_create:
    // Set up arguments for mmap syscall
    mov x0, #0                          // addr = NULL
    mov x1, #STACK_SIZE                 // length = STACK_SIZE
    mov x2, #(PROT_READ | PROT_WRITE)   // prot = PROT_READ | PROT_WRITE
    mov x3, #(MAP_PRIVATE | MAP_ANONYMOUS | MAP_GROWSDOWN)  // flags
    mov x4, #-1                         // fd = -1 (for anonymous mapping)
    mov x5, #0                          // offset = 0
    
    // Make syscall
    mov x8, #SYS_mmap                   // syscall number in x8
    svc #0                              // supervisor call
    
    ret

// Create thread using clone3 syscall
// x0 = thread ID
create_thread_clone3:
    // Save thread ID and registers
    sub sp, sp, #16
    // str x0, [sp]                        // Save thread ID
    str lr, [sp, #8]                   // Save return address
    mov x22, x0 // Save threadId in x22 for now

    bl stack_create
    cmp x0, #-1
    b.eq thread_create_error

    mov x20, x0 // Save stack base in x20
    
    // Store thread function and ID at the top of stack
    str x22, [x20]                   // Store thread ID
    
    // Set up clone_args structure
    adr x16, clone_args                 // Get address of clone_args
    ldr x2, =CLONE_FLAGS
    str x2, [x16]                       // flags (offset 0)
    str x20, [x16, #40]                 // stack (offset 40)
    mov x2, #STACK_SIZE
    str x2, [x16, #48]                  // stack_size (offset 48)
    // Call clone3
    mov x0, x16                         // clone_args pointer
    mov x1, #88                         // size of clone_args structure
    mov x8, #SYS_clone3
    svc #0
    
    // Check if we're in the child process
    cmp x0, #0
    b.eq child_process
    
    // Parent process - restore and return
    ldr lr, [sp, #8]                   // Restore return address
    add sp, sp, #16
    ret
thread_create_error:
    add sp, sp, #16
    ret


child_process:
    // We're in the child thread
    // Set up our stack pointer
    adr x2, thread_func
    // ldr x2, [x20]                       // Load function pointer
    ldr x19, [x20]                  // Load thread ID
    // Call the thread function
    blr x2                              // Branch to function pointer
    // Exit thread
    mov x0, x20
    mov x8, #SYS_exit
    svc #0

// Thread function - prints "Hello X" where X is thread ID
thread_func:
    // Thread ID is in x19 register
    add w4, w19, #'0'             // Add ASCII offset (48)
    strb w4, [sp, #-16]
    //brk #0
    mov x0, #STDOUT_FILENO
    add x1, sp, #-16
    mov x2, #1
    mov x8, #SYS_write
    svc #0
    // Print "Hello "
    mov x0, #STDOUT_FILENO
    adr x1, hello_msg
    mov x2, #hello_msg_len
    mov x8, #SYS_write
    svc #0
    ret