inst_sel.rs

use ast::ir::*;
use ast::ptr::*;
use ast::inst::*;
use ast::op;
use ast::op::OpCode;
use ast::types;
use ast::types::*;
use vm::VM;
use runtime::mm;
use runtime::ValueLocation;
use runtime::thread;
use runtime::entrypoints;
use runtime::entrypoints::RuntimeEntrypoint;

use compiler::CompilerPass;
use compiler::backend;
use compiler::backend::PROLOGUE_BLOCK_NAME;
use compiler::backend::x86_64;
use compiler::backend::x86_64::CodeGenerator;
use compiler::backend::x86_64::ASMCodeGen;
use compiler::machine_code::CompiledFunction;
use compiler::frame::Frame;

use std::collections::HashMap;
use std::any::Any;

pub struct InstructionSelection {
    name: &'static str,
    backend: Box<CodeGenerator>,
    
    current_callsite_id: usize,
    current_frame: Option<Frame>,
    current_block: Option<MuName>,
    current_func_start: Option<ValueLocation>,
    // key: block id, val: callsite that names the block as exception block
    current_exn_callsites: HashMap<MuID, Vec<ValueLocation>>,
    // key: block id, val: block location
    current_exn_blocks: HashMap<MuID, ValueLocation>     
}

impl <'a> InstructionSelection {
    #[cfg(feature = "aot")]
    pub fn new() -> InstructionSelection {
        InstructionSelection{
            name: "Instruction Selection (x64)",
            backend: Box::new(ASMCodeGen::new()),
            
            current_callsite_id: 0,
            current_frame: None,
            current_block: None,
            current_func_start: None,
            // key: block id, val: callsite that names the block as exception block
            current_exn_callsites: HashMap::new(), 
            current_exn_blocks: HashMap::new()
        }
    }

    #[cfg(feature = "jit")]
    pub fn new() -> InstructionSelection {
        unimplemented!()
    }
    
    // in this pass, we assume that
    // 1. all temporaries will use 64bit registers
    // 2. we do not need to backup/restore caller-saved registers
    // 3. we need to backup/restore all the callee-saved registers
    // if any of these assumption breaks, we will need to re-emit the code
    #[allow(unused_variables)]
    fn instruction_select(&mut self, node: &'a TreeNode, f_content: &FunctionContent, f_context: &mut FunctionContext, vm: &VM) {
        trace!("instsel on node {}", node);
        
        match node.v {
            TreeNode_::Instruction(ref inst) => {
                match inst.v {
                    Instruction_::Branch2{cond, ref true_dest, ref false_dest, true_prob} => {
                        trace!("instsel on BRANCH2");
                        // 'branch_if_true' == true, we emit cjmp the same as CmpOp  (je  for EQ, jne for NE)
                        // 'branch_if_true' == false, we emit opposite cjmp as CmpOp (jne for EQ, je  for NE)
                        let (fallthrough_dest, branch_dest, branch_if_true) = {
                            if true_prob > 0.5f32 {
                                (true_dest, false_dest, false)
                            } else {
                                (false_dest, true_dest, true)
                            }
                        };
                        
                        let ops = inst.ops.read().unwrap();
                        
                        self.process_dest(&ops, fallthrough_dest, f_content, f_context, vm);
                        self.process_dest(&ops, branch_dest, f_content, f_context, vm);
                        
                        let branch_target = f_content.get_block(branch_dest.target).name().unwrap();
    
                        let ref cond = ops[cond];
                        
                        if self.match_cmp_res(cond) {
                            trace!("emit cmp_res-branch2");
                            match self.emit_cmp_res(cond, f_content, f_context, vm) {
                                op::CmpOp::EQ => {
                                    if branch_if_true {
                                        self.backend.emit_je(branch_target);
                                    } else {
                                        self.backend.emit_jne(branch_target);
                                    }
                                },
                                op::CmpOp::NE => {
                                    if branch_if_true {
                                        self.backend.emit_jne(branch_target);
                                    } else {
                                        self.backend.emit_je(branch_target);
                                    }
                                },
                                op::CmpOp::UGE => {
                                    if branch_if_true {
                                        self.backend.emit_jae(branch_target);
                                    } else {
                                        self.backend.emit_jb(branch_target);
                                    }
                                },
                                op::CmpOp::UGT => {
                                    if branch_if_true {
                                        self.backend.emit_ja(branch_target);
                                    } else {
                                        self.backend.emit_jbe(branch_target);
                                    }
                                },
                                op::CmpOp::ULE => {
                                    if branch_if_true {
                                        self.backend.emit_jbe(branch_target);
                                    } else {
                                        self.backend.emit_ja(branch_target);
                                    }
                                },
                                op::CmpOp::ULT => {
                                    if branch_if_true {
                                        self.backend.emit_jb(branch_target);
                                    } else {
                                        self.backend.emit_jae(branch_target);
                                    }
                                },
                                op::CmpOp::SGE => {
                                    if branch_if_true {
                                        self.backend.emit_jge(branch_target);
                                    } else {
                                        self.backend.emit_jl(branch_target);
                                    }
                                },
                                op::CmpOp::SGT => {
                                    if branch_if_true {
                                        self.backend.emit_jg(branch_target);
                                    } else {
                                        self.backend.emit_jle(branch_target);
                                    }
                                },
                                op::CmpOp::SLE => {
                                    if branch_if_true {
                                        self.backend.emit_jle(branch_target);
                                    } else {
                                        self.backend.emit_jg(branch_target);
                                    }
                                },
                                op::CmpOp::SLT => {
                                    if branch_if_true {
                                        self.backend.emit_jl(branch_target);
                                    } else {
                                        self.backend.emit_jge(branch_target);
                                    }
                                },

                                // floating point

                                op::CmpOp::FOEQ | op::CmpOp::FUEQ => {
                                    if branch_if_true {
                                        self.backend.emit_je(branch_target);
                                    } else {
                                        self.backend.emit_jne(branch_target);
                                    }
                                },
                                op::CmpOp::FONE | op::CmpOp::FUNE => {
                                    if branch_if_true {
                                        self.backend.emit_jne(branch_target);
                                    } else {
                                        self.backend.emit_je(branch_target);
                                    }
                                },
                                op::CmpOp::FOGT | op::CmpOp::FUGT => {
                                    if branch_if_true {
                                        self.backend.emit_ja(branch_target);
                                    } else {
                                        self.backend.emit_jbe(branch_target);
                                    }
                                },
                                op::CmpOp::FOGE | op::CmpOp::FUGE => {
                                    if branch_if_true {
                                        self.backend.emit_jae(branch_target);
                                    } else {
                                        self.backend.emit_jb(branch_target);
                                    }
                                },
                                op::CmpOp::FOLT | op::CmpOp::FULT => {
                                    if branch_if_true {
                                        self.backend.emit_jb(branch_target);
                                    } else {
                                        self.backend.emit_jae(branch_target);
                                    }
                                },
                                op::CmpOp::FOLE | op::CmpOp::FULE => {
                                    if branch_if_true {
                                        self.backend.emit_jbe(branch_target);
                                    } else {
                                        self.backend.emit_ja(branch_target);
                                    }
                                },

                                _ => unimplemented!()
                            }
                        } else if self.match_ireg(cond) {
                            trace!("emit ireg-branch2");
                            
                            let cond_reg = self.emit_ireg(cond, f_content, f_context, vm);
                            
                            // emit: cmp cond_reg 1
                            self.backend.emit_cmp_imm_r(1, &cond_reg);
                            // emit: je #branch_dest
                            self.backend.emit_je(branch_target);
                        } else {
                            unimplemented!();
                        }
                    },

                    Instruction_::Select{cond, true_val, false_val} => {
                        use ast::op::CmpOp::*;

                        trace!("instsel on SELECT");
                        let ops = inst.ops.read().unwrap();

                        let ref cond = ops[cond];
                        let ref true_val = ops[true_val];
                        let ref false_val = ops[false_val];

                        if self.match_ireg(true_val) {
                            // moving integers/pointers
                            let tmp_res   = self.get_result_value(node);

                            // generate compare
                            let cmpop = if self.match_cmp_res(cond) {
                                self.emit_cmp_res(cond, f_content, f_context, vm)
                            } else if self.match_ireg(cond) {
                                let tmp_cond = self.emit_ireg(cond, f_content, f_context, vm);
                                // emit: cmp cond_reg 1
                                self.backend.emit_cmp_imm_r(1, &tmp_cond);

                                EQ
                            } else {
                                panic!("expected ireg, found {}", cond)
                            };

                            // use cmov for 16/32/64bit integeer
                            // use jcc  for 8 bit
                            match tmp_res.ty.get_int_length() {
                                // cmov
                                Some(len) if len > 8 => {
                                    let tmp_true  = self.emit_ireg(true_val, f_content, f_context, vm);
                                    let tmp_false = self.emit_ireg(false_val, f_content, f_context, vm);

                                    // mov tmp_false -> tmp_res
                                    self.backend.emit_mov_r_r(&tmp_res, &tmp_false);

                                    match cmpop {
                                        EQ  => self.backend.emit_cmove_r_r (&tmp_res, &tmp_true),
                                        NE  => self.backend.emit_cmovne_r_r(&tmp_res, &tmp_true),
                                        SGE => self.backend.emit_cmovge_r_r(&tmp_res, &tmp_true),
                                        SGT => self.backend.emit_cmovg_r_r (&tmp_res, &tmp_true),
                                        SLE => self.backend.emit_cmovle_r_r(&tmp_res, &tmp_true),
                                        SLT => self.backend.emit_cmovl_r_r (&tmp_res, &tmp_true),
                                        UGE => self.backend.emit_cmovae_r_r(&tmp_res, &tmp_true),
                                        UGT => self.backend.emit_cmova_r_r (&tmp_res, &tmp_true),
                                        ULE => self.backend.emit_cmovbe_r_r(&tmp_res, &tmp_true),
                                        ULT => self.backend.emit_cmovb_r_r (&tmp_res, &tmp_true),

                                        FOEQ | FUEQ => self.backend.emit_cmove_r_r (&tmp_res, &tmp_true),
                                        FONE | FUNE => self.backend.emit_cmovne_r_r(&tmp_res, &tmp_true),
                                        FOGT | FUGT => self.backend.emit_cmova_r_r (&tmp_res, &tmp_true),
                                        FOGE | FUGE => self.backend.emit_cmovae_r_r(&tmp_res, &tmp_true),
                                        FOLT | FULT => self.backend.emit_cmovb_r_r (&tmp_res, &tmp_true),
                                        FOLE | FULE => self.backend.emit_cmovbe_r_r(&tmp_res, &tmp_true),

                                        _ => unimplemented!()
                                    }
                                }
                                // jcc
                                _ => {
                                    let blk_true = format!("{}_select_true", node.id());
                                    let blk_end   = format!("{}_select_end", node.id());

                                    // jump to blk_true if true
                                    match cmpop {
                                        EQ  => self.backend.emit_je (blk_true.clone()),
                                        NE  => self.backend.emit_jne(blk_true.clone()),
                                        SGE => self.backend.emit_jge(blk_true.clone()),
                                        SGT => self.backend.emit_jg (blk_true.clone()),
                                        SLE => self.backend.emit_jle(blk_true.clone()),
                                        SLT => self.backend.emit_jl (blk_true.clone()),
                                        UGE => self.backend.emit_jae(blk_true.clone()),
                                        UGT => self.backend.emit_ja (blk_true.clone()),
                                        ULE => self.backend.emit_jbe(blk_true.clone()),
                                        ULT => self.backend.emit_jb (blk_true.clone()),

                                        FOEQ | FUEQ => self.backend.emit_je (blk_true.clone()),
                                        FONE | FUNE => self.backend.emit_jne(blk_true.clone()),
                                        FOGT | FUGT => self.backend.emit_ja (blk_true.clone()),
                                        FOGE | FUGE => self.backend.emit_jae(blk_true.clone()),
                                        FOLT | FULT => self.backend.emit_jb (blk_true.clone()),
                                        FOLE | FULE => self.backend.emit_jbe(blk_true.clone()),

                                        _ => unimplemented!()
                                    }

                                    // mov false result here
                                    self.emit_move_node_to_value(&tmp_res, &false_val, f_content, f_context, vm);

                                    // jmp to end
                                    self.backend.emit_jmp(blk_end.clone());

                                    // finishing current block
                                    let cur_block = self.current_block.as_ref().unwrap().clone();
                                    self.backend.end_block(cur_block.clone());

                                    // blk_true:
                                    self.current_block = Some(blk_true.clone());
                                    self.backend.start_block(blk_true.clone());
                                    // mov true value -> result
                                    self.emit_move_node_to_value(&tmp_res, &true_val, f_content, f_context, vm);

                                    self.backend.end_block(blk_true.clone());

                                    // blk_end:
                                    self.backend.start_block(blk_end.clone());
                                    self.current_block = Some(blk_end.clone());
                                }
                            }
                        } else {
                            // moving vectors, floatingpoints
                            unimplemented!()
                        }
                    },

                    Instruction_::CmpOp(op, op1, op2) => {
                        use ast::op::CmpOp::*;

                        trace!("instsel on CMPOP");
                        let ops = inst.ops.read().unwrap();
                        let ref op1 = ops[op1];
                        let ref op2 = ops[op2];

                        let tmp_res = self.get_result_value(node);

                        // cmov only take (16/32/64bits registers)
                        // make res64, and set to zero
                        let tmp_res64 = self.make_temporary(f_context, UINT64_TYPE.clone(), vm);
                        self.backend.emit_xor_r_r(&tmp_res64, &tmp_res64);

                        // set tmp1 as 1 (cmov doesnt allow immediate or reg8 as operand)
                        let tmp_1 = self.make_temporary(f_context, UINT64_TYPE.clone(), vm);
                        self.backend.emit_mov_r_imm(&tmp_1, 1);

                        // cmov 1 to result
                        match self.emit_cmp_res(node, f_content, f_context, vm) {
                            EQ  => self.backend.emit_cmove_r_r (&tmp_res64, &tmp_1),
                            NE  => self.backend.emit_cmovne_r_r(&tmp_res64, &tmp_1),
                            SGE => self.backend.emit_cmovge_r_r(&tmp_res64, &tmp_1),
                            SGT => self.backend.emit_cmovg_r_r (&tmp_res64, &tmp_1),
                            SLE => self.backend.emit_cmovle_r_r(&tmp_res64, &tmp_1),
                            SLT => self.backend.emit_cmovl_r_r (&tmp_res64, &tmp_1),
                            UGE => self.backend.emit_cmovae_r_r(&tmp_res64, &tmp_1),
                            UGT => self.backend.emit_cmova_r_r (&tmp_res64, &tmp_1),
                            ULE => self.backend.emit_cmovbe_r_r(&tmp_res64, &tmp_1),
                            ULT => self.backend.emit_cmovb_r_r (&tmp_res64, &tmp_1),

                            FOEQ | FUEQ => self.backend.emit_cmove_r_r (&tmp_res64, &tmp_1),
                            FONE | FUNE => self.backend.emit_cmovne_r_r(&tmp_res64, &tmp_1),
                            FOGT | FUGT => self.backend.emit_cmova_r_r (&tmp_res64, &tmp_1),
                            FOGE | FUGE => self.backend.emit_cmovae_r_r(&tmp_res64, &tmp_1),
                            FOLT | FULT => self.backend.emit_cmovb_r_r (&tmp_res64, &tmp_1),
                            FOLE | FULE => self.backend.emit_cmovbe_r_r(&tmp_res64, &tmp_1),

                            _ => unimplemented!()
                        }

                        // truncate tmp_res64 to tmp_res (probably u8)
                        self.backend.emit_mov_r_r(&tmp_res, &tmp_res64);
                    }

                    Instruction_::Branch1(ref dest) => {
                        trace!("instsel on BRANCH1");
                        let ops = inst.ops.read().unwrap();
                                            
                        self.process_dest(&ops, dest, f_content, f_context, vm);
                        
                        let target = f_content.get_block(dest.target).name().unwrap();
                        
                        trace!("emit branch1");
                        // jmp
                        self.backend.emit_jmp(target);
                    },

                    Instruction_::Switch{cond, ref default, ref branches} => {
                        trace!("instsel on SWITCH");
                        let ops = inst.ops.read().unwrap();

                        let ref cond = ops[cond];

                        if self.match_ireg(cond) {
                            let tmp_cond = self.emit_ireg(cond, f_content, f_context, vm);

                            // emit each branch
                            for &(case_op_index, ref case_dest) in branches {
                                let ref case_op = ops[case_op_index];

                                // process dest
                                self.process_dest(&ops, case_dest, f_content, f_context, vm);

                                let target = f_content.get_block(case_dest.target).name().unwrap();

                                if self.match_iimm(case_op) {
                                    let imm = self.node_iimm_to_i32(case_op);

                                    // cmp case cond
                                    self.backend.emit_cmp_imm_r(imm, &tmp_cond);
                                    // je dest
                                    self.backend.emit_je(target);
                                } else if self.match_ireg(case_op) {
                                    let tmp_case_op = self.emit_ireg(case_op, f_content, f_context, vm);

                                    // cmp case cond
                                    self.backend.emit_cmp_r_r(&tmp_case_op, &tmp_cond);
                                    // je dest
                                    self.backend.emit_je(target);
                                } else {
                                    panic!("expecting ireg cond to be either iimm or ireg: {}", cond);
                                }
                            }

                            // emit default
                            self.process_dest(&ops, default, f_content, f_context, vm);
                            
                            let default_target = f_content.get_block(default.target).name().unwrap();
                            self.backend.emit_jmp(default_target);
                        } else {
                            panic!("expecting cond in switch to be ireg: {}", cond);
                        }
                    }
                    
                    Instruction_::ExprCall{ref data, is_abort} => {
                        trace!("instsel on EXPRCALL");

                        if is_abort {
                            unimplemented!()
                        }
                        
                        self.emit_mu_call(
                            inst, // inst: &Instruction,
                            data, // calldata: &CallData,
                            None, // resumption: Option<&ResumptionData>,
                            node, // cur_node: &TreeNode, 
                            f_content, f_context, vm);
                    },
                    
                    Instruction_::Call{ref data, ref resume} => {
                        trace!("instsel on CALL");

                        self.emit_mu_call(
                            inst, 
                            data, 
                            Some(resume), 
                            node, 
                            f_content, f_context, vm);
                    },

                    Instruction_::ExprCCall{ref data, is_abort} => {
                        trace!("instsel on EXPRCCALL");

                        if is_abort {
                            unimplemented!()
                        }

                        self.emit_c_call_ir(inst, data, None, node, f_content, f_context, vm);
                    }

                    Instruction_::CCall{ref data, ref resume} => {
                        trace!("instsel on CCALL");

                        self.emit_c_call_ir(inst, data, Some(resume), node, f_content, f_context, vm);
                    }
                    
                    Instruction_::Return(_) => {
                        trace!("instsel on RETURN");

                        self.emit_common_epilogue(inst, f_content, f_context, vm);
                        
                        self.backend.emit_ret();
                    },
                    
                    Instruction_::BinOp(op, op1, op2) => {
                        trace!("instsel on BINOP");

                        let ops = inst.ops.read().unwrap();

                        let res_tmp = self.get_result_value(node);
                        
                        match op {
                            op::BinOp::Add => {
                                if self.match_ireg(&ops[op1]) && self.match_iimm(&ops[op2]) {
                                    trace!("emit add-ireg-imm");
                                    
                                    let reg_op1 = self.emit_ireg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.node_iimm_to_i32(&ops[op2]);
                                    
                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &reg_op1);
                                    // add op2, res
                                    self.backend.emit_add_r_imm(&res_tmp, reg_op2);
                                } else if self.match_ireg(&ops[op1]) && self.match_mem(&ops[op2]) {
                                    trace!("emit add-ireg-mem");
                                    
                                    let reg_op1 = self.emit_ireg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_mem(&ops[op2], vm);
                                    
                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &reg_op1);
                                    // add op2 res
                                    self.backend.emit_add_r_mem(&res_tmp, &reg_op2);
                                } else if self.match_ireg(&ops[op1]) && self.match_ireg(&ops[op2]) {
                                    trace!("emit add-ireg-ireg");

                                    let reg_op1 = self.emit_ireg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_ireg(&ops[op2], f_content, f_context, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &reg_op1);
                                    // add op2 res
                                    self.backend.emit_add_r_r(&res_tmp, &reg_op2);
                                } else {
                                    unimplemented!()
                                }
                            },
                            op::BinOp::Sub => {
                                if self.match_ireg(&ops[op1]) && self.match_iimm(&ops[op2]) {
                                    trace!("emit sub-ireg-imm");

                                    let reg_op1 = self.emit_ireg(&ops[op1], f_content, f_context, vm);
                                    let imm_op2 = self.node_iimm_to_i32(&ops[op2]);
                                    
                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &reg_op1);
                                    // add op2, res
                                    self.backend.emit_sub_r_imm(&res_tmp, imm_op2);
                                } else if self.match_ireg(&ops[op1]) && self.match_mem(&ops[op2]) {
                                    trace!("emit sub-ireg-mem");
                                    
                                    let reg_op1 = self.emit_ireg(&ops[op1], f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(&ops[op2], vm);
                                    
                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &reg_op1);
                                    // sub op2 res
                                    self.backend.emit_sub_r_mem(&res_tmp, &mem_op2);
                                } else if self.match_ireg(&ops[op1]) && self.match_ireg(&ops[op2]) {
                                    trace!("emit sub-ireg-ireg");

                                    let reg_op1 = self.emit_ireg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_ireg(&ops[op2], f_content, f_context, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &reg_op1);
                                    // add op2 res
                                    self.backend.emit_sub_r_r(&res_tmp, &reg_op2);
                                } else {
                                    unimplemented!()
                                }
                            },
                            op::BinOp::And => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                if self.match_ireg(op1) && self.match_iimm(op2) {
                                    trace!("emit and-ireg-iimm");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let imm_op2 = self.node_iimm_to_i32(op2);

                                    // mov op1 -> res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // and op2, res -> res
                                    self.backend.emit_and_r_imm(&res_tmp, imm_op2);
                                } else if self.match_ireg(op1) && self.match_mem(op2) {
                                    trace!("emit and-ireg-mem");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(op2, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // and op2, res -> res
                                    self.backend.emit_and_r_mem(&res_tmp, &mem_op2);
                                } else if self.match_ireg(op1) && self.match_ireg(op2) {
                                    trace!("emit and-ireg-ireg");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let tmp_op2 = self.emit_ireg(op2, f_content, f_context, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // and op2, res -> res
                                    self.backend.emit_and_r_r(&res_tmp, &tmp_op2);
                                } else {
                                    unimplemented!()
                                }
                            },
                            op::BinOp::Or => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                if self.match_ireg(op1) && self.match_iimm(op2) {
                                    trace!("emit or-ireg-iimm");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let imm_op2 = self.node_iimm_to_i32(op2);

                                    // mov op1 -> res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // Or op2, res -> res
                                    self.backend.emit_or_r_imm(&res_tmp, imm_op2);
                                } else if self.match_ireg(op1) && self.match_mem(op2) {
                                    trace!("emit or-ireg-mem");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(op2, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // Or op2, res -> res
                                    self.backend.emit_or_r_mem(&res_tmp, &mem_op2);
                                } else if self.match_ireg(op1) && self.match_ireg(op2) {
                                    trace!("emit or-ireg-ireg");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let tmp_op2 = self.emit_ireg(op2, f_content, f_context, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // Or op2, res -> res
                                    self.backend.emit_or_r_r(&res_tmp, &tmp_op2);
                                } else {
                                    unimplemented!()
                                }
                            },
                            op::BinOp::Xor => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                if self.match_ireg(op1) && self.match_iimm(op2) {
                                    trace!("emit xor-ireg-iimm");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let imm_op2 = self.node_iimm_to_i32(op2);

                                    // mov op1 -> res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // xor op2, res -> res
                                    self.backend.emit_xor_r_imm(&res_tmp, imm_op2);
                                } else if self.match_ireg(op1) && self.match_mem(op2) {
                                    trace!("emit xor-ireg-mem");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(op2, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // xor op2, res -> res
                                    self.backend.emit_xor_r_mem(&res_tmp, &mem_op2);
                                } else if self.match_ireg(op1) && self.match_ireg(op2) {
                                    trace!("emit xor-ireg-ireg");

                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);
                                    let tmp_op2 = self.emit_ireg(op2, f_content, f_context, vm);

                                    // mov op1, res
                                    self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    // xor op2, res -> res
                                    self.backend.emit_xor_r_r(&res_tmp, &tmp_op2);
                                } else {
                                    unimplemented!()
                                }
                            }
                            op::BinOp::Mul => {
                                // mov op1 -> rax
                                let op1 = &ops[op1];

                                let mreg_op1 = match op1.clone_value().ty.get_int_length() {
                                    Some(64) => x86_64::RAX.clone(),
                                    Some(32) => x86_64::EAX.clone(),
                                    Some(16) => x86_64::AX.clone(),
                                    Some(8)  => x86_64::AL.clone(),
                                    _ => unimplemented!()
                                };

                                if self.match_iimm(op1) {
                                    let imm_op1 = self.node_iimm_to_i32(op1);
                                    
                                    self.backend.emit_mov_r_imm(&mreg_op1, imm_op1);
                                } else if self.match_mem(op1) {
                                    let mem_op1 = self.emit_mem(op1, vm);
                                    
                                    self.backend.emit_mov_r_mem(&mreg_op1, &mem_op1);
                                } else if self.match_ireg(op1) {
                                    let reg_op1 = self.emit_ireg(op1, f_content, f_context, vm);

                                    self.backend.emit_mov_r_r(&mreg_op1, &reg_op1);
                                } else {
                                    unimplemented!();
                                }
                                
                                // mul op2
                                let op2 = &ops[op2];
                                if self.match_iimm(op2) {
                                    let imm_op2 = self.node_iimm_to_i32(op2);
                                    
                                    // put imm in a temporary
                                    // here we use result reg as temporary
                                    self.backend.emit_mov_r_imm(&res_tmp, imm_op2);
                                    
                                    self.backend.emit_mul_r(&res_tmp);
                                } else if self.match_mem(op2) {
                                    let mem_op2 = self.emit_mem(op2, vm);
                                    
                                    self.backend.emit_mul_mem(&mem_op2);
                                } else if self.match_ireg(op2) {
                                    let reg_op2 = self.emit_ireg(op2, f_content, f_context, vm);

                                    self.backend.emit_mul_r(&reg_op2);
                                } else {
                                    unimplemented!();
                                }
                                
                                // mov rax -> result
                                match res_tmp.ty.get_int_length() {
                                    Some(64) => self.backend.emit_mov_r_r(&res_tmp, &x86_64::RAX),
                                    Some(32) => self.backend.emit_mov_r_r(&res_tmp, &x86_64::EAX),
                                    Some(16) => self.backend.emit_mov_r_r(&res_tmp, &x86_64::AX),
                                    Some(8)  => self.backend.emit_mov_r_r(&res_tmp, &x86_64::AL),
                                    _ => unimplemented!()
                                }

                            },
                            op::BinOp::Udiv => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                self.emit_udiv(op1, op2, f_content, f_context, vm);

                                // mov rax -> result
                                match res_tmp.ty.get_int_length() {
                                    Some(64) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::RAX);
                                    }
                                    Some(32) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::EAX);
                                    }
                                    Some(16) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::AX);
                                    }
                                    Some(8)  => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::AL);
                                    }
                                    _ => unimplemented!()
                                }
                            },
                            op::BinOp::Sdiv => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                self.emit_idiv(op1, op2, f_content, f_context, vm);

                                // mov rax -> result
                                match res_tmp.ty.get_int_length() {
                                    Some(64) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::RAX);
                                    }
                                    Some(32) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::EAX);
                                    }
                                    Some(16) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::AX);
                                    }
                                    Some(8)  => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::AL);
                                    }
                                    _ => unimplemented!()
                                }
                            },
                            op::BinOp::Urem => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                self.emit_udiv(op1, op2, f_content, f_context, vm);

                                // mov rdx -> result
                                match res_tmp.ty.get_int_length() {
                                    Some(64) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::RDX);
                                    }
                                    Some(32) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::EDX);
                                    }
                                    Some(16) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::DX);
                                    }
                                    Some(8)  => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::AH);
                                    }
                                    _ => unimplemented!()
                                }
                            },
                            op::BinOp::Srem => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                self.emit_idiv(op1, op2, f_content, f_context, vm);

                                // mov rdx -> result
                                match res_tmp.ty.get_int_length() {
                                    Some(64) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::RDX);
                                    }
                                    Some(32) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::EDX);
                                    }
                                    Some(16) => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::DX);
                                    }
                                    Some(8)  => {
                                        self.backend.emit_mov_r_r(&res_tmp, &x86_64::AH);
                                    }
                                    _ => unimplemented!()
                                }
                            },

                            op::BinOp::Shl => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                if self.match_mem(op1) {
                                    unimplemented!()
                                } else if self.match_ireg(op1) {
                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);

                                    if self.match_iimm(op2) {
                                        let imm_op2 = self.node_iimm_to_i32(op2) as i8;

                                        // shl op1, op2 -> op1
                                        self.backend.emit_shl_r_imm8(&tmp_op1, imm_op2);

                                        // mov op1 -> result
                                        self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    } else if self.match_ireg(op2) {
                                        let tmp_op2 = self.emit_ireg(op2, f_content, f_context, vm);

                                        // mov op2 -> cl
                                        self.backend.emit_mov_r_r(&x86_64::CL, &tmp_op2);

                                        // shl op1, cl -> op1
                                        self.backend.emit_shl_r_cl(&tmp_op1);

                                        // mov op1 -> result
                                        self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    } else {
                                        panic!("unexpected op2 (not ireg not iimm): {}", op2);
                                    }
                                } else {
                                    panic!("unexpected op1 (not ireg not mem): {}", op1);
                                }
                            },
                            op::BinOp::Lshr => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                if self.match_mem(op1) {
                                    unimplemented!()
                                } else if self.match_ireg(op1) {
                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);

                                    if self.match_iimm(op2) {
                                        let imm_op2 = self.node_iimm_to_i32(op2) as i8;

                                        // shr op1, op2 -> op1
                                        self.backend.emit_shr_r_imm8(&tmp_op1, imm_op2);

                                        // mov op1 -> result
                                        self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    } else if self.match_ireg(op2) {
                                        let tmp_op2 = self.emit_ireg(op2, f_content, f_context, vm);

                                        // mov op2 -> cl
                                        self.backend.emit_mov_r_r(&x86_64::CL, &tmp_op2);

                                        // shr op1, cl -> op1
                                        self.backend.emit_shr_r_cl(&tmp_op1);

                                        // mov op1 -> result
                                        self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    } else {
                                        panic!("unexpected op2 (not ireg not iimm): {}", op2);
                                    }
                                } else {
                                    panic!("unexpected op1 (not ireg not mem): {}", op1);
                                }
                            },
                            op::BinOp::Ashr => {
                                let op1 = &ops[op1];
                                let op2 = &ops[op2];

                                if self.match_mem(op1) {
                                    unimplemented!()
                                } else if self.match_ireg(op1) {
                                    let tmp_op1 = self.emit_ireg(op1, f_content, f_context, vm);

                                    if self.match_iimm(op2) {
                                        let imm_op2 = self.node_iimm_to_i32(op2) as i8;

                                        // sar op1, op2 -> op1
                                        self.backend.emit_sar_r_imm8(&tmp_op1, imm_op2);

                                        // mov op1 -> result
                                        self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    } else if self.match_ireg(op2) {
                                        let tmp_op2 = self.emit_ireg(op2, f_content, f_context, vm);

                                        // mov op2 -> cl
                                        self.backend.emit_mov_r_r(&x86_64::CL, &tmp_op2);

                                        // sar op1, cl -> op1
                                        self.backend.emit_sar_r_cl(&tmp_op1);

                                        // mov op1 -> result
                                        self.backend.emit_mov_r_r(&res_tmp, &tmp_op1);
                                    } else  {
                                        panic!("unexpected op2 (not ireg not iimm): {}", op2);
                                    }
                                } else {
                                    panic!("unexpected op1 (not ireg not mem): {}", op1);
                                }
                            },


                            // floating point
                            op::BinOp::FAdd => {
                                if self.match_fpreg(&ops[op1]) && self.match_mem(&ops[op2]) {
                                    trace!("emit add-fpreg-mem");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(&ops[op2], vm);

                                    // mov op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // sub op2 res
                                    self.backend.emit_addsd_f64_mem64(&res_tmp, &mem_op2);
                                } else if self.match_fpreg(&ops[op1]) && self.match_fpreg(&ops[op2]) {
                                    trace!("emit add-fpreg-fpreg");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_fpreg(&ops[op2], f_content, f_context, vm);

                                    // movsd op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // add op2 res
                                    self.backend.emit_addsd_f64_f64(&res_tmp, &reg_op2);
                                } else {
                                    panic!("unexpected fadd: {}", node)
                                }
                            }

                            op::BinOp::FSub => {
                                if self.match_fpreg(&ops[op1]) && self.match_mem(&ops[op2]) {
                                    trace!("emit sub-fpreg-mem");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(&ops[op2], vm);

                                    // mov op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // sub op2 res
                                    self.backend.emit_subsd_f64_mem64(&res_tmp, &mem_op2);
                                } else if self.match_fpreg(&ops[op1]) && self.match_fpreg(&ops[op2]) {
                                    trace!("emit sub-fpreg-fpreg");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_fpreg(&ops[op2], f_content, f_context, vm);

                                    // movsd op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // sub op2 res
                                    self.backend.emit_subsd_f64_f64(&res_tmp, &reg_op2);
                                } else {
                                    panic!("unexpected fsub: {}", node)
                                }
                            }

                            op::BinOp::FMul => {
                                if self.match_fpreg(&ops[op1]) && self.match_mem(&ops[op2]) {
                                    trace!("emit mul-fpreg-mem");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(&ops[op2], vm);

                                    // mov op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // mul op2 res
                                    self.backend.emit_mulsd_f64_mem64(&res_tmp, &mem_op2);
                                } else if self.match_fpreg(&ops[op1]) && self.match_fpreg(&ops[op2]) {
                                    trace!("emit mul-fpreg-fpreg");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_fpreg(&ops[op2], f_content, f_context, vm);

                                    // movsd op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // mul op2 res
                                    self.backend.emit_mulsd_f64_f64(&res_tmp, &reg_op2);
                                } else {
                                    panic!("unexpected fmul: {}", node)
                                }
                            }

                            op::BinOp::FDiv => {
                                if self.match_fpreg(&ops[op1]) && self.match_mem(&ops[op2]) {
                                    trace!("emit div-fpreg-mem");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let mem_op2 = self.emit_mem(&ops[op2], vm);

                                    // mov op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // div op2 res
                                    self.backend.emit_divsd_f64_mem64(&res_tmp, &mem_op2);
                                } else if self.match_fpreg(&ops[op1]) && self.match_fpreg(&ops[op2]) {
                                    trace!("emit div-fpreg-fpreg");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_fpreg(&ops[op2], f_content, f_context, vm);

                                    // movsd op1, res
                                    self.backend.emit_movsd_f64_f64(&res_tmp, &reg_op1);
                                    // div op2 res
                                    self.backend.emit_divsd_f64_f64(&res_tmp, &reg_op2);
                                } else {
                                    panic!("unexpected fdiv: {}", node)
                                }
                            }

                            op::BinOp::FRem => {
                                if self.match_fpreg(&ops[op1]) && self.match_fpreg(&ops[op2]) {
                                    trace!("emit frem-fpreg-fpreg");

                                    let reg_op1 = self.emit_fpreg(&ops[op1], f_content, f_context, vm);
                                    let reg_op2 = self.emit_fpreg(&ops[op2], f_content, f_context, vm);

                                    let reg_tmp = self.get_result_value(node);

                                    self.emit_runtime_entry(&entrypoints::FREM, vec![reg_op1.clone(), reg_op2.clone()], Some(vec![reg_tmp.clone()]), Some(node), f_content, f_context, vm);
                                } else {
                                    panic!("unexpected fdiv: {}", node)
                                }
                            }
                        }
                    }

                    Instruction_::ConvOp{operation, ref from_ty, ref to_ty, operand} => {
                        trace!("instsel on CONVOP");

                        let ops = inst.ops.read().unwrap();

                        let ref op = ops[operand];

                        match operation {
                            op::ConvOp::TRUNC => {
                                if self.match_ireg(op) {
                                    let tmp_op = self.emit_ireg(op, f_content, f_context, vm);
                                    let tmp_res = self.get_result_value(node);

                                    // mov op -> result
                                    self.backend.emit_mov_r_r(&tmp_res, &tmp_op);
                                } else {
                                    panic!("unexpected op (expect ireg): {}", op);
                                }
                            }
                            op::ConvOp::ZEXT => {
                                if self.match_ireg(op) {
                                    let tmp_op = self.emit_ireg(op, f_content, f_context, vm);
                                    let tmp_res = self.get_result_value(node);

                                    // movz op -> result
                                    let from_ty_size = vm.get_backend_type_info(from_ty.id()).size;
                                    let to_ty_size   = vm.get_backend_type_info(to_ty.id()).size;

                                    if from_ty_size != to_ty_size {
                                        if from_ty_size == 4 && to_ty_size == 8 {
                                            // zero extend from 32 bits to 64 bits is a mov instruction
                                            // x86 does not have movzlq (32 to 64)

                                            // tmp_op is int32, but tmp_res is int64
                                            // we want to force a 32-to-32 mov, so high bits of the destination will be zeroed

                                            let tmp_res32 = unsafe {tmp_res.as_type(UINT32_TYPE.clone())};

                                            self.backend.emit_mov_r_r(&tmp_res32, &tmp_op);
                                        } else {
                                            self.backend.emit_movz_r_r(&tmp_res, &tmp_op);
                                        }
                                    } else {
                                        self.backend.emit_mov_r_r(&tmp_res, &tmp_op);
                                    }
                                } else {
                                    panic!("unexpected op (expect ireg): {}", op);
                                }
                            },
                            op::ConvOp::SEXT => {
                                if self.match_ireg(op) {
                                    let tmp_op = self.emit_ireg(op, f_content, f_context, vm);
                                    let tmp_res = self.get_result_value(node);

                                    // movs op -> result
                                    let from_ty_size = vm.get_backend_type_info(from_ty.id()).size;
                                    let to_ty_size   = vm.get_backend_type_info(to_ty.id()).size;

                                    if from_ty_size != to_ty_size {
                                        self.backend.emit_movs_r_r(&tmp_res, &tmp_op);
                                    } else {