#include "config.h"
#include "cpu.h"
#include "crt0-common.S"

/* implement vectors, specify section and reset address */
RBIMPL_VECTORS .vectors, start

/* When starting, we will be running at 0x40000000 most probably
 * but the code is expected to be loaded at 0x4xxxxxxx (uncached) and to be
 * running at virtual address 0xyyyyyyyy (cached). So we first
 * need to move everything to the right locationn then we setup the mmu and
 * jump to the final virtual address. */
.text
.global start
    /** The code below must be able to run at any address **/
start:
    /* Copy running address */
    sub     r7, pc, #8
    /* Save r0 */
    mov     r6, r0

    RBIMPL_START_BEGIN
    /* Disable MMU, disable caching and buffering;
     * use low exception range address (the core uses high range by default) */
    mrc     p15, 0, r0, c1, c0, 0
    ldr     r1, =0x3005
    bic     r0, r1
    mcr     p15, 0, r0, c1, c0, 0

    /* To call the C code we need a stack, since the stack is in virtual memory
     * use the stack's physical address */
    ldr     sp, =stackend_phys

    /* Enable MMU */
    bl      memory_init

    /* Copy the DRAM
     * Assume the dram binary blob is located at the loading address (r5) */
    mov    r2, r7
    ldr    r3, =_dramcopystart
    ldr    r4, =_dramcopyend
1:
    cmp     r4, r3
    ldrhi   r5, [r2], #4
    strhi   r5, [r3], #4
    bhi     1b

    mov     r2, #0
    mcr     p15, 0, r2, c7, c5, 0   @ Invalidate ICache
    
    /* Jump to real location */
    ldr     pc, =remap
remap:
    RBIMPL_DO_STD_COPY _iedata, _iend, _iramcopy, _iramstart, _iramend, _initcopy, _initstart, _initend, _edata, _end

    RBIMPL_INIT_STACKS stackend, stackbegin, iriq_stack, fiq_stack
    RBIMPL_START_END

    /* Jump to main */
    mov     r0, r6
    mov     r1, r7
    bl      main
1:
    b       1b

RBIMPL_STD_HANDLERS

RBIMPL_IRQ_FIQ_STACKS irq_stack, fiq_stack

end: