#include "../executor.h"
#include "../../include/std/alloc.h"
#include "../../include/std/buffer.h"
#include "../../include/std/log.h"
#include "asm.h"
#include <assert.h>

/*
    TODO: Currently almost all operations are performed on memory. This should be optimized
    to take advantage of registers.

    TODO: Operations with memory registers could access outside the stack. Should this be checked?
*/

/*
    TODO: Allow this to dynamically change up to 1<<16, to match jump instruction.
    This is a sane limit for now
*/
#define MAX_LABELS 128

typedef struct {
    u32 asm_index;
    u16 target_label;
    bool condition;
} JumpDefer;

typedef struct {
    Asm asm;
    u16 func_counter;
    Buffer/*JumpDefer*/ jump_defer;
    u32 label_asm_index[MAX_LABELS];
    int label_counter;
    int num_args;
    int num_pushed_values;
} amal_executor_impl;

#define ASM_ENSURE_CAPACITY return_if_error(asm_ensure_capacity(&impl->asm, 256));

#define IMPL_START \
    amal_executor_impl *impl = (amal_executor_impl*)self; \
    ASM_ENSURE_CAPACITY

/*
    @reg will be a positive value when accessing local variables, in which case the first
    local variable is located at -sizeof(usize) and the next one is at -(2 * sizeof(usize)).
    @reg will be a negative value starting at -1 when accessing parameters.
    The first parameter is located at 3*sizeof(usize) and the next one is at 4*sizeof(usize).
    Parameter starts at 3*sizeof(usize) because offset 0 is the return address, offset 1*sizeof(usize) is the
    saved RBP and 2*sizeof(usize) is saved RBX.
    TODO: Use different offset when saving more registers, for example on Microsoft Windows.
*/
#define get_register_stack_offset(reg) \
    (reg >= 0 ? (i32)(-reg * (int)sizeof(usize) - sizeof(usize)) : (i32)(-reg * (int)sizeof(usize) + 2*sizeof(usize)))

static i64 abs_i64(i64 value) {
    return value >= 0 ? value : -value;
}

const Reg64 SYS_V_REG_PARAMS[] = { RDI, RSI, RDX, RCX, R8, R9, R10, R11 };
const int NUM_REG_PARAMS = 8;

int amal_executor_init(amal_executor **self) {
    amal_executor_impl **impl;
    impl = (amal_executor_impl**)self;
    *impl = NULL;
    return_if_error(am_malloc(sizeof(amal_executor_impl), (void**)impl));
    (*impl)->func_counter = 0;
    (*impl)->label_counter = 0;
    (*impl)->num_args = 0;
    (*impl)->num_pushed_values = 0;
    ignore_result_int(buffer_init(&(*impl)->jump_defer, NULL));
    return asm_init(&(*impl)->asm);
}

void amal_executor_deinit(amal_executor *self) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    buffer_deinit(&impl->jump_defer);
    asm_deinit(&impl->asm);
    am_free(impl);
}

int amal_executor_run(amal_executor *self, u32 offset) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    return asm_execute(&impl->asm, offset);
}

u32 amal_exec_get_code_offset(amal_executor *self) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    return asm_get_size(&impl->asm);
}

int amal_executor_instructions_start(amal_executor *self, u16 num_functions) {
    (void)self;
    (void)num_functions;
    return 0;
}

int amal_executor_instructions_end(amal_executor *self) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    impl->func_counter = 0;
    return 0;
}

int amal_exec_nop(amal_executor *self) {
    IMPL_START
    asm_nop(&impl->asm);
    return 0;
}

int amal_exec_setz(amal_executor *self, i8 dst_reg) {
    AsmPtr dst;
    IMPL_START
    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_mov_mi(&impl->asm, &dst, 0);
    return 0;
}

int amal_exec_mov(amal_executor *self, i8 dst_reg, i8 src_reg) {
    AsmPtr ptr;
    IMPL_START

    asm_ptr_init_disp(&ptr, RBP, get_register_stack_offset(src_reg));
    asm_mov_rm(&impl->asm, RAX, &ptr);

    asm_ptr_init_disp(&ptr, RBP, get_register_stack_offset(dst_reg));
    asm_mov_mr(&impl->asm, &ptr, RAX);
    return 0;
}

int amal_exec_movi(amal_executor *self, i8 dst_reg, i64 imm) {
    IMPL_START
    /* TODO: if @number is a float then use float instructions */
    if(abs_i64(imm) <= INT32_MAX) {
        AsmPtr dst;
        asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
        asm_mov_mi(&impl->asm, &dst, imm);
    } else {
        AsmPtr dst;
        asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
        asm_mov_ri(&impl->asm, RAX, imm);
        asm_mov_mr(&impl->asm, &dst, RAX);
    }
    return 0;
}

int amal_exec_movd(amal_executor *self, i8 dst_reg, BufferView data) {
    AsmPtr dst;
    IMPL_START

    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_mov_ri(&impl->asm, RAX, (uintptr_t)data.data);
    asm_mov_mr(&impl->asm, &dst, RAX);
    return 0;
}

int amal_exec_add(amal_executor *self, i8 dst_reg, i8 src_reg1, i8 src_reg2) {
    AsmPtr dst;
    AsmPtr reg1;
    AsmPtr reg2;
    IMPL_START

    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_ptr_init_disp(&reg1, RBP, get_register_stack_offset(src_reg1));
    asm_ptr_init_disp(&reg2, RBP, get_register_stack_offset(src_reg2));

    asm_mov_rm(&impl->asm, RAX, &reg1);
    asm_mov_rm(&impl->asm, RCX, &reg2);
    asm_add_rr(&impl->asm, RAX, RCX);
    asm_mov_mr(&impl->asm, &dst, RAX);
    return 0;
}

int amal_exec_sub(amal_executor *self, i8 dst_reg, i8 src_reg1, i8 src_reg2) {
    AsmPtr dst;
    AsmPtr reg1;
    AsmPtr reg2;
    IMPL_START

    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_ptr_init_disp(&reg1, RBP, get_register_stack_offset(src_reg1));
    asm_ptr_init_disp(&reg2, RBP, get_register_stack_offset(src_reg2));

    asm_mov_rm(&impl->asm, RAX, &reg1);
    asm_mov_rm(&impl->asm, RCX, &reg2);
    asm_sub_rr(&impl->asm, RAX, RCX);
    asm_mov_mr(&impl->asm, &dst, RAX);
    return 0;
}

int amal_exec_imul(amal_executor *self, i8 dst_reg, i8 src_reg1, i8 src_reg2) {
    AsmPtr dst;
    AsmPtr reg1;
    AsmPtr reg2;
    IMPL_START

    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_ptr_init_disp(&reg1, RBP, get_register_stack_offset(src_reg1));
    asm_ptr_init_disp(&reg2, RBP, get_register_stack_offset(src_reg2));

    asm_mov_rm(&impl->asm, RAX, &reg1);
    asm_mov_rm(&impl->asm, RCX, &reg2);
    asm_imul_rr(&impl->asm, RAX, RCX);
    asm_mov_mr(&impl->asm, &dst, RAX);
    return 0;
}

int amal_exec_mul(amal_executor *self, i8 dst_reg, i8 src_reg1, i8 src_reg2) {
    (void)self;
    (void)dst_reg;
    (void)src_reg1;
    (void)src_reg2;
    /* TODO: Implement! */
    #if 0
    AsmPtr dst;
    AsmPtr reg1;
    AsmPtr reg2;

    asm_ptr_init_disp(&dst, RBP, -(i32)get_register_at_offset(0));
    asm_ptr_init_disp(&reg1, RBP, -(i32)get_register_at_offset(1));
    asm_ptr_init_disp(&reg2, RBP, -(i32)get_register_at_offset(2));

    return_if_error(asm_mov_rm(&self->asm, RAX, &reg1));
    return_if_error(asm_mov_rm(&self->asm, RCX, &reg2));
    return_if_error(asm_mul_rr(&self->asm, RAX, RCX));
    return_if_error(asm_mov_mr(&self->asm, &dst, RAX));
    #endif
    assert(bool_false && "TODO: Implement!");
    return 0;
}

int amal_exec_idiv(amal_executor *self, i8 dst_reg, i8 src_reg1, i8 src_reg2) {
    AsmPtr dst;
    AsmPtr reg1;
    AsmPtr reg2;
    IMPL_START

    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_ptr_init_disp(&reg1, RBP, get_register_stack_offset(src_reg1));
    asm_ptr_init_disp(&reg2, RBP, get_register_stack_offset(src_reg2));

    asm_mov_rm(&impl->asm, RAX, &reg1);
    asm_mov_rm(&impl->asm, RCX, &reg2);
    asm_cqo(&impl->asm);
    asm_idiv_rr(&impl->asm, RCX);
    asm_mov_mr(&impl->asm, &dst, RAX);
    return 0;
}

int amal_exec_div(amal_executor *self, i8 dst_reg, i8 src_reg1, i8 src_reg2) {
    (void)self;
    (void)dst_reg;
    (void)src_reg1;
    (void)src_reg2;
    /* TODO: Implement! */
    assert(bool_false && "TODO: Implement!");
    return 0;
}

int amal_exec_push(amal_executor *self, i8 reg) {
    AsmPtr reg_ptr;
    IMPL_START

    asm_ptr_init_disp(&reg_ptr, RBP, get_register_stack_offset(reg));
    if(impl->num_pushed_values < NUM_REG_PARAMS) {
        asm_mov_rm(&impl->asm, SYS_V_REG_PARAMS[impl->num_pushed_values], &reg_ptr);
    } else {
        asm_mov_rm(&impl->asm, RAX, &reg_ptr);
        asm_pushr(&impl->asm, RAX);
    }
    ++impl->num_pushed_values;
    return 0;
}

int amal_exec_pushi(amal_executor *self, i64 imm) {
    (void)self;
    (void)imm;
    /* TODO: Implement! */
    assert(bool_false && "TODO: Implement!");
    return 0;
}

int amal_exec_pushd(amal_executor *self, BufferView data) {
    (void)self;
    (void)data;
    /* TODO: Implement! */
    assert(bool_false && "TODO: Implement!");
    return 0;
}

int amal_exec_call_start(amal_executor *self, u8 num_args) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    impl->num_args = num_args;
    return 0;
}

int amal_exec_call(amal_executor *self, u32 code_offset, i8 dst_reg) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    /* TODO: Preserve necessary registers before call? */
    /* TODO: This assumes all arguments are isize */
    /* Do the function call */
    isize asm_offset = asm_get_size(&impl->asm);
    int num_pushed_stack = impl->num_pushed_values - (int)NUM_REG_PARAMS;
    ASM_ENSURE_CAPACITY

    assert((num_pushed_stack <= 0 || num_pushed_stack % 2 == 0) && "TODO: Align stack to 16-bytes before calling functions");
    assert(code_offset < asm_offset);
    asm_call_rel32(&impl->asm, (isize)code_offset - asm_offset);

    /* Handle function result and cleanup */
    {
        AsmPtr dst;
        asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
        /* TODO: Make this work when result is not stored in RAX (multiple return results) */
        asm_mov_mr(&impl->asm, &dst, RAX);
    }
    if(num_pushed_stack > 0)
        asm_add_rm64_imm(&impl->asm, RSP, num_pushed_stack * sizeof(isize));
    impl->num_pushed_values = 0;
    return 0;
}

void amal_exec_call_overwrite(amal_executor *self, u32 call_code_offset, i32 new_target_rel32) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    asm_overwrite_call_rel32(&impl->asm, call_code_offset, new_target_rel32);
}

/*
    TODO: Make argument passing work for different calling conventions and different ABI.
    This currently assumes x86_64 system v abi.
    System-V ABI parameters:
     	RDI, RSI, RDX, RCX, R8, R9, XMM0–7.
        The rest are passed in the stack.
*/
/* TODO: Make this work when function returns something else than a POD */
int amal_exec_calle(amal_executor *self, void *func, i8 dst_reg) {
    AsmPtr dst;
    amal_executor_impl *impl = (amal_executor_impl*)self;
    int num_pushed_stack = impl->num_pushed_values - (int)NUM_REG_PARAMS;
    ASM_ENSURE_CAPACITY

    assert((num_pushed_stack <= 0 || num_pushed_stack % 2 == 0) && "TODO: Align stack to 16-bytes before calling functions");

    /* TODO: Preserve necessary registers before call? */
    /* TODO: This assumes all arguments are isize */
    asm_mov_ri(&impl->asm, RAX, (intptr_t)func);
    asm_callr(&impl->asm, RAX);
    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_mov_mr(&impl->asm, &dst, RAX);
    if(num_pushed_stack > 0)
        asm_add_rm64_imm(&impl->asm, RSP, num_pushed_stack * sizeof(isize));
    impl->num_pushed_values = 0;
    return 0;
}

/*
int amal_exec_callr(i8 dst_reg, BufferView data) {

}
*/

int amal_exec_cmp(amal_executor *self, i8 dst_reg, i8 src_reg1, i8 src_reg2) {
    AsmPtr dst, src1, src2;
    IMPL_START

    asm_ptr_init_disp(&dst, RBP, get_register_stack_offset(dst_reg));
    asm_ptr_init_disp(&src1, RBP, get_register_stack_offset(src_reg1));
    asm_ptr_init_disp(&src2, RBP, get_register_stack_offset(src_reg2));

    asm_mov_rm(&impl->asm, RCX, &dst);
    asm_xor_rm64(&impl->asm, RCX, RCX);

    asm_mov_rm(&impl->asm, RAX, &src1);
    asm_cmp_rm(&impl->asm, RAX, &src2);
    asm_sete_r(&impl->asm, RCX);
    asm_mov_mr(&impl->asm, &dst, RCX);
    return 0;
}

int amal_exec_jz(amal_executor *self, i8 reg, u16 target_label) {
    AsmPtr ptr;
    u32 asm_offset;
    IMPL_START

    asm_ptr_init_disp(&ptr, RBP, get_register_stack_offset(reg));
    asm_mov_rm(&impl->asm, RAX, &ptr);
    asm_cmp_rm64_imm(&impl->asm, RAX, 0);

    asm_offset = asm_get_size(&impl->asm);
    if(target_label < impl->label_counter) {
        asm_jz(&impl->asm, (i32)impl->label_asm_index[target_label] - (i32)asm_offset);
        return 0;
    } else {
        JumpDefer jump_defer;
        jump_defer.asm_index = asm_offset;
        jump_defer.target_label = target_label;
        jump_defer.condition = bool_true;
        /* 
            Insert dummy target, but it has to be above INT16_MAX, so the target can be replaced
            no matter how large the jump will be
        */
        asm_jz(&impl->asm, INT32_MAX);
        return buffer_append(&impl->jump_defer, &jump_defer, sizeof(jump_defer));
    }
}

int amal_exec_jmp(amal_executor *self, u16 target_label) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    u32 asm_offset = asm_get_size(&impl->asm);
    ASM_ENSURE_CAPACITY
    if(target_label < impl->label_counter) {
        asm_jmp(&impl->asm, (i32)impl->label_asm_index[target_label] - (i32)asm_offset);
        return 0;
    } else {
        JumpDefer jump_defer;
        jump_defer.asm_index = asm_offset;
        jump_defer.target_label = target_label;
        jump_defer.condition = bool_false;
        /* 
            Insert dummy target, but it has to be above INT16_MAX, so the target can be replaced
            no matter how large the jump will be
        */
        asm_jmp(&impl->asm, INT32_MAX);
        return buffer_append(&impl->jump_defer, &jump_defer, sizeof(jump_defer));
    }
}

int amal_exec_ret(amal_executor *self, i8 reg) {
    AsmPtr ret_reg;
    IMPL_START

    asm_ptr_init_disp(&ret_reg, RBP, get_register_stack_offset(reg));
    /* Result is returned in RAX register. TODO: Make this work when returning more than one result */
    asm_mov_rm(&impl->asm, RAX, &ret_reg);
    return amal_exec_func_end(self);
}

static u32 get_next_uneven_number(u32 value) {
    return value + !(value & 1);
}

int amal_exec_func_start(amal_executor *self, u16 num_regs) {
    /*
        TODO: Validate stack size, or maybe remove all validation? do we really need validation?
        If we need security, we could fork the process instead.
    */

    /*
        Some registers need to be preserved before entering a function scope and these registers are different on different platforms.
        32-bit: EBX, ESI, EDI, EBP
        64-bit Windows: RBX, RSI, RDI, RBP, R12-R15, XMM6-XMM15
        64-bit Linux,BSD,Mac: RBX, RBP, R12-R15
    */
    IMPL_START
    asm_pushr(&impl->asm, RBX);
    asm_pushr(&impl->asm, RBP);
    asm_mov_rr(&impl->asm, RBP, RSP);
    /*
        Functions are entered with a stack alignment of 8 (because of call return address is pushed to stack).
        Make sure to align to to next 16-byte even if the extra bytes are not used.
    */
    asm_sub_rm64_imm(&impl->asm, RSP, get_next_uneven_number(num_regs) * sizeof(isize));
    return 0;
}

int amal_exec_func_end(amal_executor *self) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    JumpDefer *jump_defer = buffer_begin(&impl->jump_defer);
    JumpDefer *jump_defer_end = buffer_end(&impl->jump_defer);
    ASM_ENSURE_CAPACITY
    for(; jump_defer != jump_defer_end; ++jump_defer) {
        i32 jump_offset;
        if(jump_defer->target_label >= impl->label_counter) {
            amal_log_error("Program attempted to jump to a label that doesn't exist (label %u, while there are only %u labels)", jump_defer->target_label, impl->label_counter);
            return -1;
        }
        jump_offset = (isize)impl->label_asm_index[jump_defer->target_label] - (isize)jump_defer->asm_index;
        if(jump_defer->condition)
            asm_overwrite_jcc_rel32(&impl->asm, jump_defer->asm_index, jump_offset);
        else
            asm_overwrite_jmp_rel32(&impl->asm, jump_defer->asm_index, jump_offset);
    }
    buffer_clear(&impl->jump_defer);
    impl->label_counter = 0;

    asm_mov_rr(&impl->asm, RSP, RBP);
    asm_popr(&impl->asm, RBP);
    asm_popr(&impl->asm, RBX);
    asm_ret(&impl->asm, 0);
    return 0;
}

int amal_exec_label(amal_executor *self) {
    amal_executor_impl *impl = (amal_executor_impl*)self;
    assert(impl->label_counter < MAX_LABELS);
    impl->label_asm_index[impl->label_counter++] = asm_get_size(&impl->asm);
    return 0;
}