vm.rs

// Copyright 2017 The Australian National University
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

use std::collections::HashMap;

use rodal;
use ast::ptr::*;
use ast::ir::*;
use ast::inst::*;
use ast::types;
use ast::types::*;
use compiler::{Compiler, CompilerPolicy};
use compiler::backend;
use compiler::backend::BackendType;
use compiler::machine_code::{CompiledFunction, CompiledCallsite};

use runtime::thread::*;
use runtime::*;
use utils::ByteSize;
use utils::BitSize;
use utils::Address;
use runtime::mm as gc;
use vm::handle::*;
use vm::vm_options::VMOptions;
use vm::vm_options::MuLogLevel;

use log::LogLevel;
use std::sync::Arc;
use std::sync::RwLock;
use std::sync::Mutex;
use std::sync::RwLockWriteGuard;
use std::sync::atomic::{AtomicUsize, ATOMIC_USIZE_INIT, Ordering};
use std::thread::JoinHandle;
use std::collections::LinkedList;
use std;
use utils::bit_utils::{bits_ones, u64_asr};

/// The VM struct. This stores metadata for the currently running Zebu instance.
/// This struct gets persisted in the boot image, and when the boot image is loaded,
/// everything should be back to the same status as before persisting.
///
/// This struct is usually used as Arc<VM> so it can be shared among threads. The
/// Arc<VM> is stored in every thread local of a Mu thread, so that they can refer
/// to the VM easily.
///
/// We are using fine-grained lock on VM to allow mutability on different fields in VM.
/// Also we use two-level locks for some data structures such as MuFunction/
/// MuFunctionVersion/CompiledFunction so that we can mutate on two
/// different functions/funcvers/etc at the same time.

//  FIXME: However, there are problems with fine-grained lock design,
//  and we will need to rethink. See Issue #2.
//  TODO: besides fields in VM, there are some 'globals' we need to persist
//  such as STRUCT_TAG_MAP, INTERNAL_ID and internal types from ir crate. The point is
//  ir crate should be independent and self-contained. But when persisting the 'world',
//  besides persisting VM struct (containing most of the 'world'), we also need to
//  specifically persist those globals.
pub struct VM {
    // The comments are the offset into the struct
    // ---serialize---
    /// next MuID to assign
    next_id: AtomicUsize, // +0
    /// a map from MuID to MuName (for client to query)
    id_name_map: RwLock<HashMap<MuID, MuName>>, // +8
    /// a map from MuName to ID (for client to query)
    name_id_map: RwLock<HashMap<MuName, MuID>>, // +64
    /// types declared to the VM
    types: RwLock<HashMap<MuID, P<MuType>>>, // +120
    /// Ref types declared by 'make_strong_type', the key is the ID of the Referant
    ref_types: RwLock<HashMap<MuID, P<MuType>>>,
    /// types that are resolved as BackendType
    backend_type_info: RwLock<HashMap<MuID, Box<BackendType>>>,
    /// constants declared to the VM
    constants: RwLock<HashMap<MuID, P<Value>>>,
    /// globals declared to the VM
    globals: RwLock<HashMap<MuID, P<Value>>>,
    /// function signatures declared
    func_sigs: RwLock<HashMap<MuID, P<MuFuncSig>>>,
    /// functions declared to the VM
    funcs: RwLock<HashMap<MuID, RwLock<MuFunction>>>,
    /// primordial function that is set to make boot image
    primordial: RwLock<Option<PrimordialThreadInfo>>,

    /// current options for this VM
    pub vm_options: VMOptions, // +624

    // ---partially serialize---
    /// compiled functions
    /// (we are not persisting generated code with compiled function)
    compiled_funcs: RwLock<HashMap<MuID, RwLock<CompiledFunction>>>, // +728

    /// match each functions version to a map, mapping each of it's containing callsites
    /// to the name of the catch block
    callsite_table: RwLock<HashMap<MuID, Vec<Callsite>>>, // +784

    // ---do not serialize---
    /// global cell locations. We use this map to create handles for global cells,
    /// or dump globals into boot image. (this map does not get persisted because
    /// the location is changed in different runs)
    global_locations: RwLock<HashMap<MuID, ValueLocation>>,
    func_vers: RwLock<HashMap<MuID, RwLock<MuFunctionVersion>>>,

    /// all the funcref that clients want to store for AOT which are pending stores
    /// For AOT scenario, when client tries to store funcref to the heap, the store
    /// happens before we have an actual address for the function so we store a fake
    /// funcref and when generating boot image, we fix the funcref with a relocatable symbol
    aot_pending_funcref_store: RwLock<HashMap<Address, ValueLocation>>,

    /// runtime callsite table for exception handling
    /// a map from callsite address to CompiledCallsite
    compiled_callsite_table: RwLock<HashMap<Address, CompiledCallsite>>, // 896

    /// Nnmber of callsites in the callsite tables
    callsite_count: AtomicUsize,
    pub pending_joins: Mutex<LinkedList<JoinHandle<()>>> // A list of all threads currently waiting to be joined
}

unsafe impl rodal::Dump for VM {
    fn dump<D: ?Sized + rodal::Dumper>(&self, dumper: &mut D) {
        dumper.debug_record("VM", "dump");

        dumper.dump_object(&self.next_id);
        dumper.dump_object(&self.id_name_map);
        dumper.dump_object(&self.name_id_map);
        dumper.dump_object(&self.types);
        dumper.dump_object(&self.ref_types);
        dumper.dump_object(&self.backend_type_info);
        dumper.dump_object(&self.constants);
        dumper.dump_object(&self.globals);
        dumper.dump_object(&self.func_sigs);
        dumper.dump_object(&self.funcs);
        dumper.dump_object(&self.primordial);
        dumper.dump_object(&self.vm_options);
        dumper.dump_object(&self.compiled_funcs);
        dumper.dump_object(&self.callsite_table);

        // Dump empty maps so that we can safely read and modify them once loaded
        dumper.dump_padding(&self.global_locations);
        dumper.dump_object_here(&RwLock::new(
            rodal::EmptyHashMap::<MuID, ValueLocation>::new()
        ));

        dumper.dump_padding(&self.func_vers);
        dumper.dump_object_here(&RwLock::new(
            rodal::EmptyHashMap::<MuID, RwLock<MuFunctionVersion>>::new()
        ));

        dumper.dump_padding(&self.aot_pending_funcref_store);
        dumper.dump_object_here(&RwLock::new(
            rodal::EmptyHashMap::<Address, ValueLocation>::new()
        ));

        // Dump an emepty hashmap for the other hashmaps
        dumper.dump_padding(&self.compiled_callsite_table);
        dumper.dump_object_here(&RwLock::new(
            rodal::EmptyHashMap::<Address, CompiledCallsite>::new()
        ));
        dumper.dump_object(&self.callsite_count);

        dumper.dump_padding(&self.pending_joins);
        dumper.dump_object_here(&Mutex::new(
            rodal::EmptyLinkedList::<JoinHandle<()>>::new()
        ));
    }
}

/// a fake funcref to store for AOT when client tries to store a funcref via API
//  For AOT scenario, when client tries to store funcref to the heap, the store
//  happens before we have an actual address for the function so we store a fake
//  funcref and when generating boot image, we fix the funcref with a relocatable symbol
const PENDING_FUNCREF: u64 = {
    use std::u64;
    u64::MAX
};

/// a macro to generate int8/16/32/64 from/to API calls
macro_rules! gen_handle_int {
    ($fn_from: ident, $fn_to: ident, $int_ty: ty) => {
        pub fn $fn_from (&self, num: $int_ty, len: BitSize) -> APIHandleResult {
            let handle_id = self.next_id();
            self.new_handle (APIHandle {
                id: handle_id,
                v: APIHandleValue::Int(num as u64, len)
            })
        }

        pub fn $fn_to (&self, handle: APIHandleArg) -> $int_ty {
            handle.v.as_int() as $int_ty
        }
    }
}

impl<'a> VM {
    /// creates a VM with default options
    pub fn new() -> VM {
        VM::new_internal(VMOptions::default())
    }

    /// creates a VM with specified options
    pub fn new_with_opts(str: &str) -> VM {
        VM::new_internal(VMOptions::init(str))
    }

    /// internal function to create a VM with options
    #[cfg(not(feature = "sel4-rumprun"))]
    fn new_internal(options: VMOptions) -> VM {
        VM::start_logging(options.flag_log_level);

        let ret = VM {
            next_id: ATOMIC_USIZE_INIT,
            vm_options: options,
            id_name_map: RwLock::new(HashMap::new()),
            name_id_map: RwLock::new(HashMap::new()),
            constants: RwLock::new(HashMap::new()),
            types: RwLock::new(HashMap::new()),
            ref_types: RwLock::new(HashMap::new()),
            backend_type_info: RwLock::new(HashMap::new()),
            globals: RwLock::new(HashMap::new()),
            global_locations: RwLock::new(hashmap!{}),
            func_sigs: RwLock::new(HashMap::new()),
            func_vers: RwLock::new(HashMap::new()),
            funcs: RwLock::new(HashMap::new()),
            compiled_funcs: RwLock::new(HashMap::new()),
            callsite_table: RwLock::new(HashMap::new()),
            primordial: RwLock::new(None),
            aot_pending_funcref_store: RwLock::new(HashMap::new()),
            compiled_callsite_table: RwLock::new(HashMap::new()),
            callsite_count: ATOMIC_USIZE_INIT,
            pending_joins: Mutex::new(LinkedList::new()),
        };

        // insert all internal types
        {
            let mut types = ret.types.write().unwrap();
            for ty in INTERNAL_TYPES.iter() {
                types.insert(ty.id(), ty.clone());
            }
        }

        // starts allocating ID from USER_ID_START
        ret.next_id.store(USER_ID_START, Ordering::Relaxed);

        // init types
        types::init_types();

        // init runtime
        ret.init_runtime();

        ret
    }

    /// internal function to create a VM with options for sel4-rumprun
    /// default memory sizes are different from other platforms
    #[cfg(feature = "sel4-rumprun")]
    fn new_internal(options: VMOptions) -> VM {
        VM::start_logging(options.flag_log_level);

        let mut ret = VM {
            next_id: ATOMIC_USIZE_INIT,
            vm_options: options,
            id_name_map: RwLock::new(HashMap::new()),
            name_id_map: RwLock::new(HashMap::new()),
            constants: RwLock::new(HashMap::new()),
            types: RwLock::new(HashMap::new()),
            backend_type_info: RwLock::new(HashMap::new()),
            globals: RwLock::new(HashMap::new()),
            global_locations: RwLock::new(hashmap!{}),
            func_sigs: RwLock::new(HashMap::new()),
            func_vers: RwLock::new(HashMap::new()),
            funcs: RwLock::new(HashMap::new()),
            compiled_funcs: RwLock::new(HashMap::new()),
            callsite_table: RwLock::new(HashMap::new()),
            primordial: RwLock::new(None),
            aot_pending_funcref_store: RwLock::new(HashMap::new()),
            compiled_callsite_table: RwLock::new(HashMap::new()),
            callsite_count: ATOMIC_USIZE_INIT
        };

        // currently, the default sizes don't work on sel4-rumprun platform
        // this is due to memory allocation size limitations
        ret.vm_options.flag_gc_immixspace_size = 1 << 19;
        ret.vm_options.flag_gc_lospace_size = 1 << 19;

        // insert all internal types
        {
            let mut types = ret.types.write().unwrap();
            for ty in INTERNAL_TYPES.iter() {
                types.insert(ty.id(), ty.clone());
            }
        }

        // starts allocating ID from USER_ID_START
        ret.next_id.store(USER_ID_START, Ordering::Relaxed);

        // init types
        types::init_types();

        // init runtime
        ret.init_runtime();

        ret
    }

    /// initializes runtime
    fn init_runtime(&self) {
        // init gc
        {
            let ref options = self.vm_options;
            gc::gc_init(
                options.flag_gc_immixspace_size,
                options.flag_gc_lospace_size,
                options.flag_gc_nthreads,
                !options.flag_gc_disable_collection
            );
        }
    }

    /// starts logging based on MuLogLevel flag
    fn start_logging(level: MuLogLevel) {
        use std::env;
        match level {
            MuLogLevel::None => {}
            MuLogLevel::Error => VM::start_logging_internal(LogLevel::Error),
            MuLogLevel::Warn => VM::start_logging_internal(LogLevel::Warn),
            MuLogLevel::Info => VM::start_logging_internal(LogLevel::Info),
            MuLogLevel::Debug => VM::start_logging_internal(LogLevel::Debug),
            MuLogLevel::Trace => VM::start_logging_internal(LogLevel::Trace),
            MuLogLevel::Env => {
                match env::var("MU_LOG_LEVEL") {
                    Ok(s) => VM::start_logging(MuLogLevel::from_string(s)),
                    _ => {} // Don't log
                }
            }
        }
    }

    /// starts trace-level logging
    pub fn start_logging_trace() {
        VM::start_logging_internal(LogLevel::Trace)
    }

    /// starts logging based on MU_LOG_LEVEL environment variable
    pub fn start_logging_env() {
        VM::start_logging(MuLogLevel::Env)
    }

    /// starts logging based on Rust's LogLevel
    /// (this function actually initializes logger and deals with error)
    fn start_logging_internal(level: LogLevel) {
        use stderrlog;

        let verbose = match level {
            LogLevel::Error => 0,
            LogLevel::Warn => 1,
            LogLevel::Info => 2,
            LogLevel::Debug => 3,
            LogLevel::Trace => 4
        };

        match stderrlog::new().verbosity(verbose).init() {
            Ok(()) => { info!("logger initialized") }
            Err(e) => {
                error!(
                    "failed to init logger, probably already initialized: {:?}",
                    e
                )
            }
        }
    }

    /// adds an exception callsite and catch block
    /// (later we will use this info to build an exception table for unwinding use)
    pub fn add_exception_callsite(&self, callsite: Callsite, fv: MuID) {
        let mut table = self.callsite_table.write().unwrap();

        if table.contains_key(&fv) {
            table.get_mut(&fv).unwrap().push(callsite);
        } else {
            table.insert(fv, vec![callsite]);
        };
        // TODO: do wee need a stronger ordering??
        self.callsite_count.fetch_add(1, Ordering::Relaxed);
    }

    /// resumes persisted VM. Ideally the VM should be back to the status when we start
    /// persisting it except a few fields that we do not want to persist.
    pub fn resume_vm(dumped_vm: *mut Arc<VM>) -> Arc<VM> {
        // load the vm back
        let vm = unsafe { rodal::load_asm_pointer_move(dumped_vm) };

        // initialize runtime
        vm.init_runtime();

        // construct exception table
        vm.build_callsite_table();

        // restore gc types
        {
            let type_info_guard = vm.backend_type_info.read().unwrap();
            let mut type_info_vec: Vec<Box<BackendType>> =
                type_info_guard.values().map(|x| x.clone()).collect();
            type_info_vec.sort_by(|a, b| a.gc_type.id.cmp(&b.gc_type.id));

            let mut expect_id = 0;
            for ty_info in type_info_vec.iter() {
                use runtime::mm;

                let ref gc_type = ty_info.gc_type;

                if gc_type.id != expect_id {
                    debug_assert!(expect_id < gc_type.id);

                    while expect_id < gc_type.id {
                        use runtime::mm::common::gctype::GCType;

                        mm::add_gc_type(GCType::new_noreftype(0, 8));
                        expect_id += 1;
                    }
                }

                // now expect_id == gc_type.id
                debug_assert!(expect_id == gc_type.id);

                mm::add_gc_type(gc_type.as_ref().clone());
                expect_id += 1;
            }
        }
        // construct exception table
        vm.build_callsite_table();
        vm
    }

    /// builds a succinct exception table for fast query during exception unwinding
    /// We need this step because for AOT compilation, we do not know symbol address at compile,
    /// and resolving symbol address during exception handling is expensive. Thus when boot image
    /// gets executed, we first resolve symbols and store the results in another table for fast
    /// query.
    pub fn build_callsite_table(&self) {
        let callsite_table = self.callsite_table.read().unwrap();
        let compiled_funcs = self.compiled_funcs.read().unwrap();
        let mut compiled_callsite_table = self.compiled_callsite_table.write().unwrap();
        // TODO: Use a different ordering?
        compiled_callsite_table.reserve(self.callsite_count.load(Ordering::Relaxed));
        for (fv, callsite_list) in callsite_table.iter() {
            let compiled_func = compiled_funcs.get(fv).unwrap().read().unwrap();
            let callee_saved_table = Arc::new(compiled_func.frame.callee_saved.clone());
            for callsite in callsite_list.iter() {
                compiled_callsite_table.insert(
                    resolve_symbol(callsite.name.clone()),
                    CompiledCallsite::new(
                        &callsite,
                        compiled_func.func_ver_id,
                        callee_saved_table.clone()
                    )
                );
            }
        }
    }

    /// returns a valid ID for use next
    pub fn next_id(&self) -> MuID {
        // This only needs to be atomic, and does not need to be a synchronisation operation. The
        // only requirement for IDs is that all IDs obtained from `next_id()` are different. So
        // `Ordering::Relaxed` is sufficient.
        self.next_id.fetch_add(1, Ordering::Relaxed)
    }

    /// are we doing AOT compilation? (feature = aot when building Zebu)
    pub fn is_doing_aot(&self) -> bool {
        return cfg!(feature = "aot");
    }

    /// are we doing JIT compilation? (feature = jit when building Zebu)
    pub fn is_doing_jit(&self) -> bool {
        return cfg!(feature = "jit");
    }

    /// informs VM about a client-supplied name
    pub fn set_name(&self, entity: &MuEntity) {
        let id = entity.id();
        let name = entity.name();

        let mut map = self.id_name_map.write().unwrap();
        map.insert(id, name.clone());

        let mut map2 = self.name_id_map.write().unwrap();
        map2.insert(name, id);
    }

    /// returns Mu ID for a client-supplied name
    /// This function should only used by client, 'name' used internally may be slightly different
    /// due to removal of some special symbols in the MuName. See name_check() in ir.rs
    pub fn id_of(&self, name: &str) -> MuID {
        let map = self.name_id_map.read().unwrap();
        match map.get(name) {
            Some(id) => *id,
            None => panic!("cannot find id for name: {}", name)
        }
    }

    /// returns the client-supplied Mu name for Mu ID
    /// This function should only used by client, 'name' used internally may be slightly different
    /// due to removal of some special symbols in the MuName. See name_check() in ir.rs
    pub fn name_of(&self, id: MuID) -> MuName {
        let map = self.id_name_map.read().unwrap();
        map.get(&id).unwrap().clone()
    }

    /// declares a constant
    pub fn declare_const(&self, entity: MuEntityHeader, ty: P<MuType>, val: Constant) -> P<Value> {
        let ret = P(Value {
            hdr: entity,
            ty: ty,
            v: Value_::Constant(val)
        });

        let mut constants = self.constants.write().unwrap();
        self.declare_const_internal(&mut constants, ret.id(), ret.clone());

        ret
    }

    /// adds a constant to the map (already acquired lock)
    fn declare_const_internal(
        &self,
        map: &mut RwLockWriteGuard<HashMap<MuID, P<Value>>>,
        id: MuID,
        val: P<Value>
    ) {
        debug_assert!(!map.contains_key(&id));

        info!("declare const #{} = {}", id, val);
        map.insert(id, val);
    }

    /// gets the constant P<Value> for a given Mu ID, panics if there is no type with the ID
    pub fn get_const(&self, id: MuID) -> P<Value> {
        let const_lock = self.constants.read().unwrap();
        match const_lock.get(&id) {
            Some(ret) => ret.clone(),
            None => panic!("cannot find const #{}", id)
        }
    }

    /// allocates memory for a constant that needs to be put in memory
    /// For AOT, we simply create a label for it, and let code emitter allocate the memory
    #[cfg(feature = "aot")]
    pub fn allocate_const(&self, val: P<Value>) -> ValueLocation {
        let id = val.id();
        let name = format!("CONST_{}_{}", id, val.name());

        ValueLocation::Relocatable(backend::RegGroup::GPR, name)
    }

    /// declares a global
    pub fn declare_global(&self, entity: MuEntityHeader, ty: P<MuType>) -> P<Value> {
        // create iref value for the global
        let global = P(Value {
            hdr: entity,
            ty: self.declare_type(
                MuEntityHeader::unnamed(self.next_id()),
                MuType_::iref(ty.clone())
            ),
            v: Value_::Global(ty)
        });

        let mut globals = self.globals.write().unwrap();
        let mut global_locs = self.global_locations.write().unwrap();
        self.declare_global_internal(&mut globals, &mut global_locs, global.id(), global.clone());

        global
    }

    /// adds the global to the map (already acquired lock), and allocates memory for it
    fn declare_global_internal(
        &self,
        globals: &mut RwLockWriteGuard<HashMap<MuID, P<Value>>>,
        global_locs: &mut RwLockWriteGuard<HashMap<MuID, ValueLocation>>,
        id: MuID,
        val: P<Value>
    ) {
        self.declare_global_internal_no_alloc(globals, id, val.clone());
        self.alloc_global(global_locs, id, val);
    }

    /// adds the global to the map (already acquired lock)
    /// when bulk declaring, we hold locks for everything, we cannot resolve backend type
    /// and do alloc so we add globals to the map, and then allocate them later
    fn declare_global_internal_no_alloc(
        &self,
        globals: &mut RwLockWriteGuard<HashMap<MuID, P<Value>>>,
        id: MuID,
        val: P<Value>
    ) {
        debug_assert!(!globals.contains_key(&id));

        info!("declare global #{} = {}", id, val);
        globals.insert(id, val.clone());
    }

    /// allocates memory for a global cell
    fn alloc_global(
        &self,
        global_locs: &mut RwLockWriteGuard<HashMap<MuID, ValueLocation>>,
        id: MuID,
        val: P<Value>
    ) {
        let backend_ty = self.get_backend_type_info(val.ty.get_referent_ty().unwrap().id());
        let loc = gc::allocate_global(val, backend_ty);
        trace!("allocate global #{} as {}", id, loc);
        global_locs.insert(id, loc);
    }

    /// declares a type
    pub fn declare_type(&self, entity: MuEntityHeader, ty: MuType_) -> P<MuType> {
        let ty = P(MuType { hdr: entity, v: ty });

        let mut types = self.types.write().unwrap();
        self.declare_type_internal(&mut types, ty.id(), ty.clone());

        ty
    }

    /// adds the type to the map (already acquired lock)
    fn declare_type_internal(
        &self,
        types: &mut RwLockWriteGuard<HashMap<MuID, P<MuType>>>,
        id: MuID,
        ty: P<MuType>
    ) {
        debug_assert!(!types.contains_key(&id));

        types.insert(id, ty.clone());
        info!("declare type #{} = {}", id, ty);

        // for struct/hybrid, also adds to struct/hybrid tag map
        if ty.is_struct() {
            let tag = ty.get_struct_hybrid_tag().unwrap();
            let struct_map_guard = STRUCT_TAG_MAP.read().unwrap();
            let struct_inner = struct_map_guard.get(&tag).unwrap();
            trace!("  {}", struct_inner);
        } else if ty.is_hybrid() {
            let tag = ty.get_struct_hybrid_tag().unwrap();
            let hybrid_map_guard = HYBRID_TAG_MAP.read().unwrap();
            let hybrid_inner = hybrid_map_guard.get(&tag).unwrap();
            trace!("  {}", hybrid_inner);
        }
    }

    /// gets the type for a given Mu ID, panics if there is no type with the ID
    pub fn get_type(&self, id: MuID) -> P<MuType> {
        let type_lock = self.types.read().unwrap();
        match type_lock.get(&id) {
            Some(ret) => ret.clone(),
            None => panic!("cannot find type #{}", id)
        }
    }

    /// declares a function signature
    pub fn declare_func_sig(
        &self,
        entity: MuEntityHeader,
        ret_tys: Vec<P<MuType>>,
        arg_tys: Vec<P<MuType>>
    ) -> P<MuFuncSig> {
        let ret = P(MuFuncSig {
            hdr: entity,
            ret_tys: ret_tys,
            arg_tys: arg_tys
        });

        let mut func_sigs = self.func_sigs.write().unwrap();
        self.declare_func_sig_internal(&mut func_sigs, ret.id(), ret.clone());

        ret
    }

    /// adds a function signature to the map (already acquired lock)
    fn declare_func_sig_internal(
        &self,
        sigs: &mut RwLockWriteGuard<HashMap<MuID, P<MuFuncSig>>>,
        id: MuID,
        sig: P<MuFuncSig>
    ) {
        debug_assert!(!sigs.contains_key(&id));

        info!("declare func sig #{} = {}", id, sig);
        sigs.insert(id, sig);
    }

    /// gets the function signature for a given ID, panics if there is no func sig with the ID
    pub fn get_func_sig(&self, id: MuID) -> P<MuFuncSig> {
        let func_sig_lock = self.func_sigs.read().unwrap();
        match func_sig_lock.get(&id) {
            Some(ret) => ret.clone(),
            None => panic!("cannot find func sig #{}", id)
        }
    }

    /// declares a Mu function
    pub fn declare_func(&self, func: MuFunction) {
        let mut funcs = self.funcs.write().unwrap();

        self.declare_func_internal(&mut funcs, func.id(), func);
    }

    /// adds a Mu function to the map (already acquired lock)
    fn declare_func_internal(
        &self,
        funcs: &mut RwLockWriteGuard<HashMap<MuID, RwLock<MuFunction>>>,
        id: MuID,
        func: MuFunction
    ) {
        debug_assert!(!funcs.contains_key(&id));

        info!("declare func #{} = {}", id, func);
        funcs.insert(id, RwLock::new(func));
    }

    /// gets the function name for a function (by ID), panics if there is no function with the ID
    /// Note this name is the internal name, which is different than
    /// the client-supplied name from vm.name_of()
    pub fn get_name_for_func(&self, id: MuID) -> MuName {
        let funcs_lock = self.funcs.read().unwrap();
        match funcs_lock.get(&id) {
            Some(func) => func.read().unwrap().name(),
            None => panic!("cannot find name for Mu function #{}")
        }
    }

    /// gets the function signature for a function (by ID),
    /// panics if there is no function with the ID
    pub fn get_sig_for_func(&self, id: MuID) -> P<MuFuncSig> {
        let funcs_lock = self.funcs.read().unwrap();
        match funcs_lock.get(&id) {
            Some(func) => func.read().unwrap().sig.clone(),
            None => panic!("cannot find Mu function #{}", id)
        }
    }

    /// gets the current function version for a Mu function (by ID)
    /// returns None if the function does not exist, or no version is defined for the function
    pub fn get_cur_version_for_func(&self, fid: MuID) -> Option<MuID> {
        let funcs_guard = self.funcs.read().unwrap();
        match funcs_guard.get(&fid) {
            Some(rwlock_func) => {
                let func_guard = rwlock_func.read().unwrap();
                func_guard.cur_ver
            }
            None => None
        }
    }

    /// gets the address as ValueLocation of a Mu function (by ID)
    pub fn get_address_for_func(&self, func_id: MuID) -> ValueLocation {
        let funcs = self.funcs.read().unwrap();
        let func: &MuFunction = &funcs.get(&func_id).unwrap().read().unwrap();

        if self.is_doing_jit() {
            unimplemented!()
        } else {
            ValueLocation::Relocatable(backend::RegGroup::GPR, func.name())
        }
    }

    /// defines a function version
    pub fn define_func_version(&self, func_ver: MuFunctionVersion) {
        info!("define function version {}", func_ver);
        // add this funcver to map
        let func_ver_id = func_ver.id();
        {
            let mut func_vers = self.func_vers.write().unwrap();
            func_vers.insert(func_ver_id, RwLock::new(func_ver));
        }

        // acquire a reference to the func_ver
        let func_vers = self.func_vers.read().unwrap();
        let func_ver = func_vers.get(&func_ver_id).unwrap().write().unwrap();

        // change current version of the function to new version (obsolete old versions)
        let funcs = self.funcs.read().unwrap();
        // it should be declared before defining
        debug_assert!(funcs.contains_key(&func_ver.func_id));
        let mut func = funcs.get(&func_ver.func_id).unwrap().write().unwrap();

        func.new_version(func_ver.id());

        if self.is_doing_jit() {
            // redefinition may happen, we need to check
            unimplemented!()
        }
    }

    pub fn make_strong_type(&self, ty: P<MuType>) -> P<MuType> {
        match &ty.v {
            &MuType_::WeakRef(ref t) => {
                let res = self.ref_types
                    .read()
                    .unwrap()
                    .get(&t.id())
                    .map(|x| x.clone());
                match res {
                    Some(ty) => ty,
                    None => {
                        let ty = P(MuType::new(self.next_id(), MuType_::muref(t.clone())));
                        self.ref_types.write().unwrap().insert(t.id(), ty.clone());
                        ty
                    }
                }
            }
            _ => ty.clone()
        }
    }

    /// adds a new bundle into VM.
    /// This function will drain the contents of all arguments. Ideally, this function should
    /// happen atomically. e.g. The client should not see a new type added without also seeing
    /// a new function added.
    pub fn declare_many(
        &self,
        new_id_name_map: &mut HashMap<MuID, MuName>,
        new_types: &mut HashMap<MuID, P<MuType>>,
        new_func_sigs: &mut HashMap<MuID, P<MuFuncSig>>,
        new_constants: &mut HashMap<MuID, P<Value>>,
        new_globals: &mut HashMap<MuID, P<Value>>,
        new_funcs: &mut HashMap<MuID, Box<MuFunction>>,
        new_func_vers: &mut HashMap<MuID, Box<MuFunctionVersion>>,
        arc_vm: Arc<VM>
    ) {
        // Make sure other components, if ever acquiring multiple locks at the same time, acquire
        // them in this order, to prevent deadlock.
        {
            let mut id_name_map = self.id_name_map.write().unwrap();
            let mut name_id_map = self.name_id_map.write().unwrap();
            let mut types = self.types.write().unwrap();
            let mut constants = self.constants.write().unwrap();
            let mut globals = self.globals.write().unwrap();
            let mut func_sigs = self.func_sigs.write().unwrap();
            let mut funcs = self.funcs.write().unwrap();
            let mut func_vers = self.func_vers.write().unwrap();

            for (id, name) in new_id_name_map.drain() {
                id_name_map.insert(id, name.clone());
                name_id_map.insert(name, id);
            }

            for (id, obj) in new_types.drain() {
                self.declare_type_internal(&mut types, id, obj);
            }

            for (id, obj) in new_constants.drain() {
                self.declare_const_internal(&mut constants, id, obj);
            }

            for (id, obj) in new_globals.drain() {
                // we bulk allocate later
                // (since we are holding all the locks, we cannot find ty info)
                self.declare_global_internal_no_alloc(&mut globals, id, obj);
            }

            for (id, obj) in new_func_sigs.drain() {
                self.declare_func_sig_internal(&mut func_sigs, id, obj);
            }

            for (id, obj) in new_funcs.drain() {
                self.declare_func_internal(&mut funcs, id, *obj);
            }

            for (id, obj) in new_func_vers.drain() {
                let func_id = obj.func_id;
                func_vers.insert(id, RwLock::new(*obj));

                {
                    trace!("Adding funcver {} as a version of {}...", id, func_id);
                    let func = funcs.get_mut(&func_id).unwrap();
                    func.write().unwrap().new_version(id);
                    trace!(
                        "Added funcver {} as a version of {} {:?}.",
                        id,
                        func_id,
                        func
                    );
                }
            }
        }
        // Locks released here

        // allocate all the globals defined
        {
            let globals = self.globals.read().unwrap();
            let mut global_locs = self.global_locations.write().unwrap();

            // make sure current thread has allocator
            let created =
                unsafe { MuThread::current_thread_as_mu_thread(Address::zero(), arc_vm.clone()) };

            for (id, global) in globals.iter() {
                self.alloc_global(&mut global_locs, *id, global.clone());
            }

            if created {
                unsafe { MuThread::cleanup_current_mu_thread() };
            }