#pragma once

#include "codegen.hpp"

class StackFasmX86_64Generator : public CodeGenerator
{
public:
    ~StackFasmX86_64Generator() override = default;
private:
    int GetStackSize(const IR::OpView ops)
    {
        int stackSize = 0;
        for (auto &op : ops)
        {
            switch (op->GetType())
            {
            case IR::OpType::STORE:
            case IR::OpType::LOAD:
            case IR::OpType::LOAD_CONST:
            case IR::OpType::ADD:
            case IR::OpType::CALL:
                stackSize += 4;
            }
        }
        return stackSize;
    }

    StringView GetTempAddr(IR::Reg reg)
    {
        return std::move((StringBuilder() << 't' << reg).view());
    }
private:
    void GenerateOp(const IR::Op *op)
    {
        switch (op->GetType())
        {
        case IR::OpType::EXTERN:
        {
            auto extrn = reinterpret_cast<const IR::ExternOp *>(op);
            auto symbol = extrn->symbol();
            printf("extrn '%s' as __%s\n", symbol.c_str(), symbol.c_str());
            printf("%s = PLT __%s\n", symbol.c_str(), symbol.c_str());
        }
        break;
        case IR::OpType::FN:
        {
            auto fn = reinterpret_cast<const IR::FnOp *>(op);
            auto name = fn->name();
            printf("public %s\n", name.c_str());
            printf("%s:\n", name.c_str());
            printf("push rbp\n");
            printf("mov rbp, rsp\n");
            m_stack = new StackAllocator();
            int stackSize = GetStackSize(fn->ops());
            printf("sub rsp, %d\n", stackSize);
            if (fn->params().size > 0)
            {
                auto param_slot = m_stack->Allocate(fn->params().data[0]);
                printf("mov [rbp-%d], edi\n", param_slot.offset);
            }
            for(auto &fOp : fn->ops())
            {
                GenerateOp(fOp);
            }
            operator delete(m_stack);
            m_stack = nullptr;
            printf("leave\nret\n");
        }
        break;
        case IR::OpType::CALL:
        {
            auto call = reinterpret_cast<const IR::CallOp *>(op);
            // TODO: support several arguments
            if (call->args().size == 1)
            {
                auto slot = m_stack->Resolve(GetTempAddr(call->args().data[0]));
                printf("mov edi, [rbp-%d]\n", slot.offset);
            }
            printf("call %s\n", call->callee().c_str());
            auto result_slot = m_stack->Allocate(GetTempAddr(call->result()));
            printf("mov dword [rbp-%d], eax\n", result_slot.offset);
        }
        break;
        case IR::OpType::LOAD_CONST:
        {
            auto lc = reinterpret_cast<const IR::LoadConstOp *>(op);
            auto addr = GetTempAddr(lc->result());
            auto slot = m_stack->Allocate(addr);
            printf("mov dword [rbp-%d], %ld\n", slot.offset, lc->value());
        }
        break;
        case IR::OpType::STORE:
        {
            auto s = reinterpret_cast<const IR::StoreOp *>(op);
            printf("; DEBUG: resolving stack slot at %s\n", s->addr().c_str());
            auto slot = m_stack->Allocate(s->addr());
            auto value_slot = m_stack->Resolve(GetTempAddr(s->src()));
            printf("mov eax, [rbp-%d]\n", value_slot.offset);
            printf("mov dword [rbp-%d], eax\n", slot.offset);
        }
        break;
        case IR::OpType::LOAD:
        {
            auto l = reinterpret_cast<const IR::LoadOp *>(op);
            auto value_slot = m_stack->Allocate(GetTempAddr(l->result()));
            auto variable_slot = m_stack->Resolve(l->addr());
            printf("mov eax, [rbp-%d]\n", variable_slot.offset);
            printf("mov dword [rbp-%d], eax\n", value_slot.offset);
        }
        break;
        case IR::OpType::ADD:
        {
            auto expr = reinterpret_cast<const IR::AddOp *>(op);
            auto lhs_slot = m_stack->Resolve(GetTempAddr(expr->lhs()));
            printf("mov eax, [rbp-%d]\n", lhs_slot.offset);
            auto rhs_slot = m_stack->Resolve(GetTempAddr(expr->rhs()));
            printf("add eax, [rbp-%d]\n", rhs_slot.offset);
            auto result_slot = m_stack->Allocate(GetTempAddr(expr->result()));
            printf("mov dword [rbp-%d], eax\n", result_slot.offset);
        }
        break;
        default: printf("; NOT HANDLED\n; %s\n", op->Format(0).c_str()); break;
        }
    }
public:
    void Generate(const char* filename, View<IR::Op*> ops) override
    {
        printf("; fasm x86_64 linux generated assembly using pl\n");
        printf("format ELF64\n");
        printf("section '.text' executable\n");

        for (auto& op : ops)
        {
            GenerateOp(op);
        }
    }
public:
    // TODO: handle sub-blocks
    StackAllocator* m_stack = nullptr;
};