#include <variant/core.h>
#include <asm/regs.h>
#include <asm/asmmacro.h>
#include <asm/cacheasm.h>
        /*
         * RB-Data: RedBoot data/bss
         * P:       Boot-Parameters
         * L:       Kernel-Loader
         *
         * The Linux-Kernel image including the loader must be loaded
         * to a position so that the kernel and the boot parameters
         * can fit in the space before the load address.
         *  ______________________________________________________
         * |_RB-Data_|_P_|__________|_L_|___Linux-Kernel___|______|
         *                          ^
         *                          ^ Load address
         *  ______________________________________________________
         * |___Linux-Kernel___|_P_|_L_|___________________________|
         *
         * The loader copies the parameter to the position that will
         * be the end of the kernel and itself to the end of the
         * parameter list.
         */

/* Make sure we have enough space for the 'uncompressor' */

#define STACK_SIZE 32768
#define HEAP_SIZE (131072*4)

        # a2: Parameter list
        # a3: Size of parameter list

        .section .start, "ax"

        .globl __start
        /* this must be the first byte of the loader! */
__start:
        entry   sp, 32          # we do not intend to return
        _call0  _start
__start_a0:
        .align 4

        .section .text, "ax"
        .begin literal_prefix .text

        /* put literals in here! */

        .globl _start
_start:

        /* 'reset' window registers */

        movi    a4, 1
        wsr     a4, ps
        rsync

        rsr     a5, windowbase
        ssl     a5
        sll     a4, a4
        wsr     a4, windowstart
        rsync

        movi    a4, 0x00040000
        wsr     a4, ps
        rsync

        /* copy the loader to its address
         * Note: The loader itself is a very small piece, so we assume we
         *       don't partially overlap. We also assume (even more important)
         *       that the kernel image is out of the way. Usually, when the
         *       load address of this image is not at an arbitrary address,
         *       but aligned to some 10K's we shouldn't overlap.
         */

        /* Note: The assembler cannot relax "addi a0, a0, ..." to an
           l32r, so we load to a4 first. */

        # addi  a4, a0, __start - __start_a0
        # mov   a0, a4

        movi    a4, __start
        movi    a5, __start_a0
        add     a4, a0, a4
        sub     a0, a4, a5

        movi    a4, __start
        movi    a5, __reloc_end

        # a0: address where this code has been loaded
        # a4: compiled address of __start
        # a5: compiled end address

        mov.n   a7, a0
        mov.n   a8, a4

1:
        l32i    a10, a7, 0
        l32i    a11, a7, 4
        s32i    a10, a8, 0
        s32i    a11, a8, 4
        l32i    a10, a7, 8
        l32i    a11, a7, 12
        s32i    a10, a8, 8
        s32i    a11, a8, 12
        addi    a8, a8, 16
        addi    a7, a7, 16
        blt     a8, a5, 1b


        /* We have to flush and invalidate the caches here before we jump. */

#if XCHAL_DCACHE_IS_WRITEBACK

        ___flush_dcache_all a5 a6

#endif

        ___invalidate_icache_all a5 a6
        isync

        movi    a11, _reloc
        jx      a11

        .globl _reloc
_reloc:

        /* RedBoot is now at the end of the memory, so we don't have
         * to copy the parameter list. Keep the code around; in case
         * we need it again. */
#if 0
        # a0: load address
        # a2: start address of parameter list
        # a3: length of parameter list
        # a4: __start

        /* copy the parameter list out of the way */

        movi    a6, _param_start
        add     a3, a2, a3
2:
        l32i    a8, a2, 0
        s32i    a8, a6, 0
        addi    a2, a2, 4
        addi    a6, a6, 4
        blt     a2, a3, 2b
#endif

        /* clear BSS section */
        movi    a6, __bss_start
        movi    a7, __bss_end
        movi.n  a5, 0
3:
        s32i    a5, a6, 0
        addi    a6, a6, 4
        blt     a6, a7, 3b

        movi    a5, -16
        movi    a1, _stack + STACK_SIZE
        and     a1, a1, a5

        /* Uncompress the kernel */

        # a0: load address
        # a2: boot parameter
        # a4: __start

        movi    a3, __image_load
        sub     a4, a3, a4
        add     a8, a0, a4

        # a1  Stack
        # a8(a4)  Load address of the image

        movi    a6, _image_start
        movi    a10, _image_end
        movi    a7, 0x1000000
        sub     a11, a10, a6
        movi    a9, complen
        s32i    a11, a9, 0

        movi    a0, 0

        # a6 destination
        # a7 maximum size of destination
        # a8 source
        # a9 ptr to length

        .extern gunzip
        movi    a4, gunzip
        beqz    a4, 1f

        callx4  a4

        j       2f


        # a6 destination start
        # a7 maximum size of destination
        # a8 source start
        # a9 ptr to length
        # a10 destination end

1:
        l32i    a9, a8, 0
        l32i    a11, a8, 4
        s32i    a9, a6, 0
        s32i    a11, a6, 4
        l32i    a9, a8, 8
        l32i    a11, a8, 12
        s32i    a9, a6, 8
        s32i    a11, a6, 12
        addi    a6, a6, 16
        addi    a8, a8, 16
        blt     a6, a10, 1b


        /* jump to the kernel */
2:
#if XCHAL_DCACHE_IS_WRITEBACK

        ___flush_dcache_all a5 a6

#endif

        ___invalidate_icache_all a5 a6

        isync

        # a2  Boot parameter list

        movi    a0, _image_start
        jx      a0

        .align 16
        .data
        .globl avail_ram
avail_ram:
        .long   _heap
        .globl end_avail
end_avail:
        .long   _heap + HEAP_SIZE

        .comm _stack, STACK_SIZE
        .comm _heap, HEAP_SIZE

        .globl end_avail
        .comm complen, 4

        .end    literal_prefix