mlua/src/lua.rs

1245 lines
44 KiB
Rust
Raw Normal View History

A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
use std::{cmp, mem, ptr, str};
use std::sync::{Arc, Mutex};
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
use std::cell::{Cell, RefCell};
use std::ffi::CString;
use std::any::TypeId;
2017-05-21 18:50:59 -05:00
use std::marker::PhantomData;
use std::collections::HashMap;
2017-05-21 18:50:59 -05:00
use std::os::raw::{c_char, c_int, c_void};
use libc;
2017-05-21 18:50:59 -05:00
use ffi;
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
use error::{Error, Result};
use util::{callback_error, check_stack, check_stack_err, gc_guard, get_userdata,
get_wrapped_error, init_error_metatables, pop_error, protect_lua, protect_lua_closure,
push_string, push_userdata, push_wrapped_error, safe_pcall, safe_xpcall,
userdata_destructor, StackGuard};
use value::{FromLua, FromLuaMulti, MultiValue, Nil, ToLua, ToLuaMulti, Value};
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
use types::{Callback, Integer, LightUserData, LuaRef, Number, RefType, RegistryKey};
use string::String;
use table::Table;
use function::Function;
use thread::Thread;
2017-10-23 15:42:20 -05:00
use userdata::{AnyUserData, MetaMethod, UserData, UserDataMethods};
use scope::Scope;
2017-05-21 18:50:59 -05:00
2017-06-17 20:23:17 -05:00
/// Top level Lua struct which holds the Lua state itself.
2017-05-21 18:50:59 -05:00
pub struct Lua {
pub(crate) state: *mut ffi::lua_State,
2017-05-21 18:50:59 -05:00
ephemeral: bool,
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ref_stack_slots: [Cell<usize>; REF_STACK_SIZE as usize],
2017-05-21 18:50:59 -05:00
}
// Data associated with the main lua_State via lua_getextraspace.
struct ExtraData {
registered_userdata: HashMap<TypeId, c_int>,
registry_unref_list: Arc<Mutex<Option<Vec<c_int>>>>,
}
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
const REF_STACK_SIZE: c_int = 16;
static FUNCTION_METATABLE_REGISTRY_KEY: u8 = 0;
unsafe impl Send for Lua {}
2017-05-21 18:50:59 -05:00
impl Drop for Lua {
fn drop(&mut self) {
unsafe {
if cfg!(debug_assertions) {
for use_count in &self.ref_stack_slots {
rlua_assert!(use_count.get() == 0, "live stack reference detected");
}
if !self.ephemeral {
let top = ffi::lua_gettop(self.state);
rlua_assert!(
top == REF_STACK_SIZE,
"stack problem detected, stack top is {}",
top - REF_STACK_SIZE
);
}
}
if !self.ephemeral {
let extra_data = *(ffi::lua_getextraspace(self.state) as *mut *mut ExtraData);
*(*extra_data).registry_unref_list.lock().unwrap() = None;
Box::from_raw(extra_data);
ffi::lua_close(self.state);
2017-05-21 18:50:59 -05:00
}
}
}
}
impl Lua {
/// Creates a new Lua state and loads standard library without the `debug` library.
2017-05-21 18:50:59 -05:00
pub fn new() -> Lua {
unsafe { Lua::create_lua(false) }
2017-05-21 18:50:59 -05:00
}
/// Creates a new Lua state and loads the standard library including the `debug` library.
///
/// The debug library is very unsound, loading it and using it breaks all the guarantees of
/// rlua.
pub unsafe fn new_with_debug() -> Lua {
Lua::create_lua(true)
}
/// Loads a chunk of Lua code and returns it as a function.
2017-06-17 20:23:17 -05:00
///
/// The source can be named by setting the `name` parameter. This is generally recommended as it
/// results in better error traces.
2017-06-17 20:23:17 -05:00
///
/// Equivalent to Lua's `load` function.
pub fn load(&self, source: &str, name: Option<&str>) -> Result<Function> {
2017-05-21 18:50:59 -05:00
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 1);
match if let Some(name) = name {
let name =
CString::new(name.to_owned()).map_err(|e| Error::ToLuaConversionError {
from: "&str",
to: "string",
message: Some(e.to_string()),
})?;
ffi::luaL_loadbuffer(
self.state,
source.as_ptr() as *const c_char,
source.len(),
name.as_ptr(),
)
} else {
ffi::luaL_loadbuffer(
self.state,
source.as_ptr() as *const c_char,
source.len(),
ptr::null(),
)
} {
ffi::LUA_OK => Ok(Function(self.pop_ref())),
err => Err(pop_error(self.state, err)),
}
2017-05-21 18:50:59 -05:00
}
}
/// Execute a chunk of Lua code.
///
/// This is equivalent to simply loading the source with `load` and then calling the resulting
/// function with no arguments.
///
/// Returns the values returned by the chunk.
pub fn exec<'lua, R: FromLuaMulti<'lua>>(
&'lua self,
source: &str,
name: Option<&str>,
) -> Result<R> {
self.load(source, name)?.call(())
}
2017-06-17 20:23:17 -05:00
/// Evaluate the given expression or chunk inside this Lua state.
///
/// If `source` is an expression, returns the value it evaluates to. Otherwise, returns the
/// values returned by the chunk (if any).
pub fn eval<'lua, R: FromLuaMulti<'lua>>(
&'lua self,
source: &str,
name: Option<&str>,
) -> Result<R> {
// First, try interpreting the lua as an expression by adding
// "return", then as a statement. This is the same thing the
// actual lua repl does.
self.load(&format!("return {}", source), name)
.or_else(|_| self.load(source, name))?
.call(())
}
2017-07-24 19:13:11 -05:00
/// Pass a `&str` slice to Lua, creating and returning an interned Lua string.
pub fn create_string(&self, s: &str) -> Result<String> {
2017-05-21 18:50:59 -05:00
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 4);
push_string(self.state, s)?;
Ok(String(self.pop_ref()))
2017-05-21 18:50:59 -05:00
}
}
2017-06-17 20:23:17 -05:00
/// Creates and returns a new table.
pub fn create_table(&self) -> Result<Table> {
2017-05-21 18:50:59 -05:00
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 3);
unsafe extern "C" fn new_table(state: *mut ffi::lua_State) -> c_int {
ffi::lua_newtable(state);
1
}
protect_lua(self.state, 0, new_table)?;
Ok(Table(self.pop_ref()))
2017-05-21 18:50:59 -05:00
}
}
2017-06-17 20:23:17 -05:00
/// Creates a table and fills it with values from an iterator.
pub fn create_table_from<'lua, K, V, I>(&'lua self, cont: I) -> Result<Table<'lua>>
2017-06-15 09:26:39 -05:00
where
K: ToLua<'lua>,
V: ToLua<'lua>,
I: IntoIterator<Item = (K, V)>,
2017-05-21 18:50:59 -05:00
{
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 5);
unsafe extern "C" fn new_table(state: *mut ffi::lua_State) -> c_int {
ffi::lua_newtable(state);
1
}
protect_lua(self.state, 0, new_table)?;
for (k, v) in cont {
self.push_value(k.to_lua(self)?);
self.push_value(v.to_lua(self)?);
unsafe extern "C" fn raw_set(state: *mut ffi::lua_State) -> c_int {
ffi::lua_rawset(state, -3);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
1
}
protect_lua(self.state, 3, raw_set)?;
}
Ok(Table(self.pop_ref()))
2017-05-21 18:50:59 -05:00
}
}
2017-06-17 20:23:17 -05:00
/// Creates a table from an iterator of values, using `1..` as the keys.
pub fn create_sequence_from<'lua, T, I>(&'lua self, cont: I) -> Result<Table<'lua>>
2017-06-15 09:26:39 -05:00
where
T: ToLua<'lua>,
I: IntoIterator<Item = T>,
2017-05-21 18:50:59 -05:00
{
self.create_table_from(cont.into_iter().enumerate().map(|(k, v)| (k + 1, v)))
2017-05-21 18:50:59 -05:00
}
2017-06-17 20:23:17 -05:00
/// Wraps a Rust function or closure, creating a callable Lua function handle to it.
2017-07-25 18:54:40 -05:00
///
/// The function's return value is always a `Result`: If the function returns `Err`, the error
/// is raised as a Lua error, which can be caught using `(x)pcall` or bubble up to the Rust code
/// that invoked the Lua code. This allows using the `?` operator to propagate errors through
/// intermediate Lua code.
///
/// If the function returns `Ok`, the contained value will be converted to one or more Lua
/// values. For details on Rust-to-Lua conversions, refer to the [`ToLua`] and [`ToLuaMulti`]
/// traits.
///
2017-07-25 18:54:40 -05:00
/// # Examples
///
/// Create a function which prints its argument:
///
/// ```
/// # extern crate rlua;
/// # use rlua::{Lua, Result};
/// # fn try_main() -> Result<()> {
/// let lua = Lua::new();
///
/// let greet = lua.create_function(|_, name: String| {
2017-07-25 18:54:40 -05:00
/// println!("Hello, {}!", name);
/// Ok(())
2017-07-25 18:54:40 -05:00
/// });
/// # let _ = greet; // used
/// # Ok(())
/// # }
/// # fn main() {
/// # try_main().unwrap();
/// # }
/// ```
///
/// Use tuples to accept multiple arguments:
2017-07-25 18:54:40 -05:00
///
/// ```
/// # extern crate rlua;
/// # use rlua::{Lua, Result};
/// # fn try_main() -> Result<()> {
/// let lua = Lua::new();
///
/// let print_person = lua.create_function(|_, (name, age): (String, u8)| {
2017-07-25 18:54:40 -05:00
/// println!("{} is {} years old!", name, age);
/// Ok(())
2017-07-25 18:54:40 -05:00
/// });
/// # let _ = print_person; // used
/// # Ok(())
/// # }
/// # fn main() {
/// # try_main().unwrap();
/// # }
/// ```
///
/// [`ToLua`]: trait.ToLua.html
/// [`ToLuaMulti`]: trait.ToLuaMulti.html
pub fn create_function<'lua, 'callback, A, R, F>(&'lua self, func: F) -> Result<Function<'lua>>
2017-06-15 09:26:39 -05:00
where
A: FromLuaMulti<'callback>,
R: ToLuaMulti<'callback>,
F: 'static + Send + Fn(&'callback Lua, A) -> Result<R>,
2017-05-21 18:50:59 -05:00
{
self.create_callback(Box::new(move |lua, args| {
func(lua, A::from_lua_multi(args, lua)?)?.to_lua_multi(lua)
}))
2017-05-21 18:50:59 -05:00
}
2018-02-11 16:54:17 -06:00
/// Wraps a Rust mutable closure, creating a callable Lua function handle to it.
///
/// This is a version of [`create_function`] that accepts a FnMut argument. Refer to
/// [`create_function`] for more information about the implementation.
///
/// [`create_function`]: #method.create_function
pub fn create_function_mut<'lua, 'callback, A, R, F>(
&'lua self,
func: F,
) -> Result<Function<'lua>>
where
A: FromLuaMulti<'callback>,
R: ToLuaMulti<'callback>,
F: 'static + Send + FnMut(&'callback Lua, A) -> Result<R>,
{
let func = RefCell::new(func);
self.create_function(move |lua, args| {
(&mut *func.try_borrow_mut()
.map_err(|_| Error::RecursiveMutCallback)?)(lua, args)
})
}
2017-06-17 20:23:17 -05:00
/// Wraps a Lua function into a new thread (or coroutine).
///
/// Equivalent to `coroutine.create`.
pub fn create_thread<'lua>(&'lua self, func: Function<'lua>) -> Result<Thread<'lua>> {
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 2);
let thread_state =
protect_lua_closure(self.state, 0, 1, |state| ffi::lua_newthread(state))?;
self.push_ref(&func.0);
ffi::lua_xmove(self.state, thread_state, 1);
Ok(Thread(self.pop_ref()))
}
}
2017-06-17 20:23:17 -05:00
/// Create a Lua userdata object from a custom userdata type.
pub fn create_userdata<T>(&self, data: T) -> Result<AnyUserData>
2017-06-15 09:26:39 -05:00
where
T: Send + UserData,
2017-05-21 18:50:59 -05:00
{
unsafe { self.make_userdata(data) }
2017-05-21 18:50:59 -05:00
}
2017-06-17 20:23:17 -05:00
/// Returns a handle to the global environment.
pub fn globals(&self) -> Table {
2017-05-21 18:50:59 -05:00
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 2);
ffi::lua_rawgeti(self.state, ffi::LUA_REGISTRYINDEX, ffi::LUA_RIDX_GLOBALS);
Table(self.pop_ref())
2017-05-21 18:50:59 -05:00
}
}
/// Calls the given function with a `Scope` parameter, giving the function the ability to create
/// userdata from rust types that are !Send, and rust callbacks that are !Send and not 'static.
///
/// The lifetime of any function or userdata created through `Scope` lasts only until the
/// completion of this method call, on completion all such created values are automatically
/// dropped and Lua references to them are invalidated. If a script accesses a value created
/// through `Scope` outside of this method, a Lua error will result. Since we can ensure the
/// lifetime of values created through `Scope`, and we know that `Lua` cannot be sent to another
/// thread while `Scope` is live, it is safe to allow !Send datatypes and functions whose
/// lifetimes only outlive the scope lifetime.
///
/// Handles that `Lua::scope` produces have a `'lua` lifetime of the scope parameter, to prevent
/// the handles from escaping the callback. However, this is not the only way for values to
/// escape the callback, as they can be smuggled through Lua itself. This is safe to do, but
/// not very useful, because after the scope is dropped, all references to scoped values,
/// whether in Lua or in rust, are invalidated. `Function` types will error when called, and
/// `AnyUserData` types will be typeless.
pub fn scope<'scope, 'lua: 'scope, F, R>(&'lua self, f: F) -> R
where
F: FnOnce(&Scope<'scope>) -> R,
{
let scope = Scope::new(self);
let r = f(&scope);
drop(scope);
r
}
2017-06-17 20:23:17 -05:00
/// Coerces a Lua value to a string.
///
/// The value must be a string (in which case this is a no-op) or a number.
pub fn coerce_string<'lua>(&'lua self, v: Value<'lua>) -> Result<String<'lua>> {
2017-05-21 18:50:59 -05:00
match v {
Value::String(s) => Ok(s),
2017-05-21 18:50:59 -05:00
v => unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 4);
let ty = v.type_name();
self.push_value(v);
let s =
protect_lua_closure(self.state, 1, 1, |state| ffi::lua_tostring(state, -1))?;
if s.is_null() {
Err(Error::FromLuaConversionError {
from: ty,
to: "String",
message: Some("expected string or number".to_string()),
})
} else {
Ok(String(self.pop_ref()))
}
2017-05-21 18:50:59 -05:00
},
}
}
2017-06-17 20:23:17 -05:00
/// Coerces a Lua value to an integer.
///
/// The value must be an integer, or a floating point number or a string that can be converted
/// to an integer. Refer to the Lua manual for details.
pub fn coerce_integer(&self, v: Value) -> Result<Integer> {
2017-05-21 18:50:59 -05:00
match v {
Value::Integer(i) => Ok(i),
2017-05-21 18:50:59 -05:00
v => unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 2);
let ty = v.type_name();
self.push_value(v);
let mut isint = 0;
let i = ffi::lua_tointegerx(self.state, -1, &mut isint);
if isint == 0 {
Err(Error::FromLuaConversionError {
from: ty,
to: "integer",
message: None,
})
} else {
Ok(i)
}
2017-05-21 18:50:59 -05:00
},
}
}
2017-06-17 20:23:17 -05:00
/// Coerce a Lua value to a number.
///
/// The value must be a number or a string that can be converted to a number. Refer to the Lua
/// manual for details.
pub fn coerce_number(&self, v: Value) -> Result<Number> {
2017-05-21 18:50:59 -05:00
match v {
Value::Number(n) => Ok(n),
2017-05-21 18:50:59 -05:00
v => unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 2);
let ty = v.type_name();
self.push_value(v);
let mut isnum = 0;
let n = ffi::lua_tonumberx(self.state, -1, &mut isnum);
if isnum == 0 {
Err(Error::FromLuaConversionError {
from: ty,
to: "number",
message: Some("number or string coercible to number".to_string()),
})
} else {
Ok(n)
}
2017-05-21 18:50:59 -05:00
},
}
}
/// Converts a value that implements `ToLua` into a `Value` instance.
pub fn pack<'lua, T: ToLua<'lua>>(&'lua self, t: T) -> Result<Value<'lua>> {
2017-05-21 18:50:59 -05:00
t.to_lua(self)
}
/// Converts a `Value` instance into a value that implements `FromLua`.
pub fn unpack<'lua, T: FromLua<'lua>>(&'lua self, value: Value<'lua>) -> Result<T> {
2017-05-21 18:50:59 -05:00
T::from_lua(value, self)
}
/// Converts a value that implements `ToLuaMulti` into a `MultiValue` instance.
pub fn pack_multi<'lua, T: ToLuaMulti<'lua>>(&'lua self, t: T) -> Result<MultiValue<'lua>> {
2017-05-21 18:50:59 -05:00
t.to_lua_multi(self)
}
/// Converts a `MultiValue` instance into a value that implements `FromLuaMulti`.
pub fn unpack_multi<'lua, T: FromLuaMulti<'lua>>(
&'lua self,
value: MultiValue<'lua>,
) -> Result<T> {
2017-05-21 18:50:59 -05:00
T::from_lua_multi(value, self)
}
2017-12-16 17:05:53 -06:00
/// Set a value in the Lua registry based on a string name.
///
/// This value will be available to rust from all `Lua` instances which share the same main
/// state.
2017-12-16 17:05:53 -06:00
pub fn set_named_registry_value<'lua, T: ToLua<'lua>>(
&'lua self,
name: &str,
t: T,
) -> Result<()> {
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 5);
push_string(self.state, name)?;
self.push_value(t.to_lua(self)?);
unsafe extern "C" fn set_registry(state: *mut ffi::lua_State) -> c_int {
ffi::lua_rawset(state, ffi::LUA_REGISTRYINDEX);
0
}
protect_lua(self.state, 2, set_registry)
}
}
2017-12-16 17:05:53 -06:00
/// Get a value from the Lua registry based on a string name.
///
2018-02-11 16:54:17 -06:00
/// Any Lua instance which shares the underlying main state may call this method to
/// get a value previously set by [`set_named_registry_value`].
///
/// [`set_named_registry_value`]: #method.set_named_registry_value
2017-12-16 17:05:53 -06:00
pub fn named_registry_value<'lua, T: FromLua<'lua>>(&'lua self, name: &str) -> Result<T> {
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 4);
push_string(self.state, name)?;
unsafe extern "C" fn get_registry(state: *mut ffi::lua_State) -> c_int {
ffi::lua_rawget(state, ffi::LUA_REGISTRYINDEX);
1
}
protect_lua(self.state, 1, get_registry)?;
T::from_lua(self.pop_value(), self)
}
}
/// Removes a named value in the Lua registry.
2017-12-16 17:05:53 -06:00
///
2018-02-11 17:17:15 -06:00
/// Equivalent to calling [`set_named_registry_value`] with a value of Nil.
2018-02-11 16:54:17 -06:00
///
/// [`set_named_registry_value`]: #method.set_named_registry_value
2017-12-16 17:05:53 -06:00
pub fn unset_named_registry_value<'lua>(&'lua self, name: &str) -> Result<()> {
self.set_named_registry_value(name, Nil)
}
/// Place a value in the Lua registry with an auto-generated key.
///
/// This value will be available to rust from all `Lua` instances which share the same main
/// state.
pub fn create_registry_value<'lua, T: ToLua<'lua>>(&'lua self, t: T) -> Result<RegistryKey> {
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 2);
self.push_value(t.to_lua(self)?);
let registry_id = gc_guard(self.state, || {
ffi::luaL_ref(self.state, ffi::LUA_REGISTRYINDEX)
});
Ok(RegistryKey {
registry_id,
unref_list: (*self.extra()).registry_unref_list.clone(),
})
}
}
/// Get a value from the Lua registry by its `RegistryKey`
///
2018-02-11 16:54:17 -06:00
/// Any Lua instance which shares the underlying main state may call this method to get a value
/// previously placed by [`create_registry_value`].
///
/// [`create_registry_value`]: #method.create_registry_value
pub fn registry_value<'lua, T: FromLua<'lua>>(&'lua self, key: &RegistryKey) -> Result<T> {
unsafe {
if !self.owns_registry_value(key) {
return Err(Error::MismatchedRegistryKey);
}
let _sg = StackGuard::new(self.state);
check_stack(self.state, 2);
ffi::lua_rawgeti(
self.state,
ffi::LUA_REGISTRYINDEX,
key.registry_id as ffi::lua_Integer,
);
T::from_lua(self.pop_value(), self)
}
}
/// Removes a value from the Lua registry.
///
/// You may call this function to manually remove a value placed in the registry with
2018-02-11 16:54:17 -06:00
/// [`create_registry_value`]. In addition to manual `RegistryKey` removal, you can also call
/// [`expire_registry_values`] to automatically remove values from the registry whose
/// `RegistryKey`s have been dropped.
2018-02-11 16:54:17 -06:00
///
/// [`create_registry_value`]: #method.create_registry_value
/// [`expire_registry_values`]: #method.expire_registry_values
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
pub fn remove_registry_value(&self, key: RegistryKey) -> Result<()> {
unsafe {
if !self.owns_registry_value(&key) {
return Err(Error::MismatchedRegistryKey);
}
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::luaL_unref(self.state, ffi::LUA_REGISTRYINDEX, key.take());
Ok(())
}
}
/// Returns true if the given `RegistryKey` was created by a `Lua` which shares the underlying
/// main state with this `Lua` instance.
///
/// Other than this, methods that accept a `RegistryKey` will return
/// `Error::MismatchedRegistryKey` if passed a `RegistryKey` that was not created with a
/// matching `Lua` state.
pub fn owns_registry_value(&self, key: &RegistryKey) -> bool {
unsafe { Arc::ptr_eq(&key.unref_list, &(*self.extra()).registry_unref_list) }
}
2018-02-11 16:54:17 -06:00
/// Remove any registry values whose `RegistryKey`s have all been dropped.
///
/// Unlike normal handle values, `RegistryKey`s do not automatically remove themselves on Drop,
/// but you can call this method to remove any unreachable registry values not manually removed
/// by `Lua::remove_registry_value`.
pub fn expire_registry_values(&self) {
unsafe {
let unref_list = mem::replace(
&mut *(*self.extra()).registry_unref_list.lock().unwrap(),
Some(Vec::new()),
);
for id in unref_list.unwrap() {
ffi::luaL_unref(self.state, ffi::LUA_REGISTRYINDEX, id);
}
}
}
// Uses 2 stack spaces, does not call checkstack
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
pub(crate) unsafe fn push_value(&self, value: Value) {
match value {
Value::Nil => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_pushnil(self.state);
}
Value::Boolean(b) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_pushboolean(self.state, if b { 1 } else { 0 });
}
Value::LightUserData(ud) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_pushlightuserdata(self.state, ud.0);
}
Value::Integer(i) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_pushinteger(self.state, i);
}
Value::Number(n) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_pushnumber(self.state, n);
}
Value::String(s) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
self.push_ref(&s.0);
}
Value::Table(t) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
self.push_ref(&t.0);
}
Value::Function(f) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
self.push_ref(&f.0);
}
Value::Thread(t) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
self.push_ref(&t.0);
}
Value::UserData(ud) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
self.push_ref(&ud.0);
}
Value::Error(e) => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
push_wrapped_error(self.state, e);
}
}
}
// Uses 2 stack spaces, does not call checkstack
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
pub(crate) unsafe fn pop_value(&self) -> Value {
match ffi::lua_type(self.state, -1) {
ffi::LUA_TNIL => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_pop(self.state, 1);
Nil
}
ffi::LUA_TBOOLEAN => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
let b = Value::Boolean(ffi::lua_toboolean(self.state, -1) != 0);
ffi::lua_pop(self.state, 1);
b
}
ffi::LUA_TLIGHTUSERDATA => {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
let ud = Value::LightUserData(LightUserData(ffi::lua_touserdata(self.state, -1)));
ffi::lua_pop(self.state, 1);
ud
}
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::LUA_TNUMBER => if ffi::lua_isinteger(self.state, -1) != 0 {
let i = Value::Integer(ffi::lua_tointeger(self.state, -1));
ffi::lua_pop(self.state, 1);
2017-10-23 15:42:20 -05:00
i
} else {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
let n = Value::Number(ffi::lua_tonumber(self.state, -1));
ffi::lua_pop(self.state, 1);
2017-10-23 15:42:20 -05:00
n
},
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::LUA_TSTRING => Value::String(String(self.pop_ref())),
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::LUA_TTABLE => Value::Table(Table(self.pop_ref())),
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::LUA_TFUNCTION => Value::Function(Function(self.pop_ref())),
ffi::LUA_TUSERDATA => {
// It should not be possible to interact with userdata types other than custom
// UserData types OR a WrappedError. WrappedPanic should never be able to be caught
// in lua, so it should never be here.
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
if let Some(err) = get_wrapped_error(self.state, -1).as_ref() {
let err = err.clone();
ffi::lua_pop(self.state, 1);
Value::Error(err)
} else {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
Value::UserData(AnyUserData(self.pop_ref()))
}
}
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::LUA_TTHREAD => Value::Thread(Thread(self.pop_ref())),
_ => rlua_panic!("LUA_TNONE in pop_value"),
}
}
2017-10-23 14:17:46 -05:00
// Used 1 stack space, does not call checkstack
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
pub(crate) unsafe fn push_ref(&self, lref: &LuaRef) {
assert!(
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
lref.lua as *const Lua == self as *const Lua,
"Lua instance passed Value created from a different Lua"
2017-06-15 09:26:39 -05:00
);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
match lref.ref_type {
RefType::Nil => ffi::lua_pushnil(self.state),
RefType::Stack { stack_slot } => {
ffi::lua_pushvalue(self.state, stack_slot);
}
RefType::Registry { registry_id } => {
ffi::lua_rawgeti(
self.state,
ffi::LUA_REGISTRYINDEX,
registry_id as ffi::lua_Integer,
);
}
}
}
// Pops the topmost element of the stack and stores a reference to it in the
// registry.
//
// This pins the object, preventing garbage collection until the returned
// `LuaRef` is dropped.
//
// pop_ref uses 1 extra stack space and does not call checkstack
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
pub(crate) unsafe fn pop_ref(&self) -> LuaRef {
for i in 0..REF_STACK_SIZE {
let ref_slot = &self.ref_stack_slots[i as usize];
if ref_slot.get() == 0 {
ref_slot.set(1);
ffi::lua_replace(self.state, i + 1);
return LuaRef {
lua: self,
ref_type: RefType::Stack { stack_slot: i + 1 },
};
}
}
let registry_id = gc_guard(self.state, || {
ffi::luaL_ref(self.state, ffi::LUA_REGISTRYINDEX)
});
if registry_id == ffi::LUA_REFNIL {
LuaRef {
lua: self,
ref_type: RefType::Nil,
}
} else {
LuaRef {
lua: self,
ref_type: RefType::Registry {
registry_id: registry_id,
},
}
}
}
pub(crate) fn clone_ref(&self, lref: &LuaRef) -> LuaRef {
unsafe {
match lref.ref_type {
RefType::Nil => LuaRef {
lua: self,
ref_type: RefType::Nil,
},
RefType::Stack { stack_slot } => {
let ref_slot = &self.ref_stack_slots[(stack_slot - 1) as usize];
ref_slot.set(ref_slot.get() + 1);
LuaRef {
lua: self,
ref_type: RefType::Stack { stack_slot },
}
}
RefType::Registry { registry_id } => {
check_stack(self.state, 2);
ffi::lua_rawgeti(
self.state,
ffi::LUA_REGISTRYINDEX,
registry_id as ffi::lua_Integer,
);
let registry_id = gc_guard(self.state, || {
ffi::luaL_ref(self.state, ffi::LUA_REGISTRYINDEX)
});
LuaRef {
lua: self,
ref_type: RefType::Registry {
registry_id: registry_id,
},
}
}
}
}
}
pub(crate) fn drop_ref(&self, lref: &mut LuaRef) {
unsafe {
match lref.ref_type {
RefType::Nil => {}
RefType::Stack { stack_slot } => {
let ref_slot = &self.ref_stack_slots[(stack_slot - 1) as usize];
let ref_count = ref_slot.get();
rlua_debug_assert!(ref_count > 0, "ref slot use count has gone below zero");
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ref_slot.set(ref_count - 1);
if ref_count == 1 {
ffi::lua_pushnil(self.state);
ffi::lua_replace(self.state, stack_slot);
}
}
RefType::Registry { registry_id } => {
ffi::luaL_unref(self.state, ffi::LUA_REGISTRYINDEX, registry_id);
}
}
}
}
pub(crate) unsafe fn userdata_metatable<T: UserData>(&self) -> Result<c_int> {
2017-05-21 18:50:59 -05:00
// Used if both an __index metamethod is set and regular methods, checks methods table
// first, then __index metamethod.
unsafe extern "C" fn meta_index_impl(state: *mut ffi::lua_State) -> c_int {
ffi::luaL_checkstack(state, 2, ptr::null());
2017-10-23 14:17:46 -05:00
2017-05-21 18:50:59 -05:00
ffi::lua_pushvalue(state, -1);
ffi::lua_gettable(state, ffi::lua_upvalueindex(1));
if ffi::lua_isnil(state, -1) == 0 {
ffi::lua_insert(state, -3);
ffi::lua_pop(state, 2);
1
} else {
ffi::lua_pop(state, 1);
ffi::lua_pushvalue(state, ffi::lua_upvalueindex(2));
ffi::lua_insert(state, -3);
ffi::lua_call(state, 2, 1);
1
}
}
if let Some(table_id) = (*self.extra()).registered_userdata.get(&TypeId::of::<T>()) {
return Ok(*table_id);
}
let _sg = StackGuard::new(self.state);
check_stack(self.state, 6);
2017-05-21 18:50:59 -05:00
let mut methods = UserDataMethods {
methods: HashMap::new(),
meta_methods: HashMap::new(),
_type: PhantomData,
};
T::add_methods(&mut methods);
protect_lua_closure(self.state, 0, 1, |state| {
ffi::lua_newtable(state);
})?;
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
let has_methods = !methods.methods.is_empty();
2017-05-21 18:50:59 -05:00
if has_methods {
push_string(self.state, "__index")?;
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
protect_lua_closure(self.state, 0, 1, |state| {
ffi::lua_newtable(state);
})?;
2017-05-21 18:50:59 -05:00
for (k, m) in methods.methods {
push_string(self.state, &k)?;
self.push_value(Value::Function(self.create_callback(m)?));
protect_lua_closure(self.state, 3, 1, |state| {
ffi::lua_rawset(state, -3);
})?;
}
protect_lua_closure(self.state, 3, 1, |state| {
ffi::lua_rawset(state, -3);
})?;
}
2017-05-21 18:50:59 -05:00
for (k, m) in methods.meta_methods {
if k == MetaMethod::Index && has_methods {
push_string(self.state, "__index")?;
ffi::lua_pushvalue(self.state, -1);
ffi::lua_gettable(self.state, -3);
self.push_value(Value::Function(self.create_callback(m)?));
protect_lua_closure(self.state, 2, 1, |state| {
ffi::lua_pushcclosure(state, meta_index_impl, 2);
})?;
2017-05-21 18:50:59 -05:00
protect_lua_closure(self.state, 3, 1, |state| {
ffi::lua_rawset(state, -3);
})?;
} else {
let name = match k {
MetaMethod::Add => "__add",
MetaMethod::Sub => "__sub",
MetaMethod::Mul => "__mul",
MetaMethod::Div => "__div",
MetaMethod::Mod => "__mod",
MetaMethod::Pow => "__pow",
MetaMethod::Unm => "__unm",
MetaMethod::IDiv => "__idiv",
MetaMethod::BAnd => "__band",
MetaMethod::BOr => "__bor",
MetaMethod::BXor => "__bxor",
MetaMethod::BNot => "__bnot",
MetaMethod::Shl => "__shl",
MetaMethod::Shr => "__shr",
MetaMethod::Concat => "__concat",
MetaMethod::Len => "__len",
MetaMethod::Eq => "__eq",
MetaMethod::Lt => "__lt",
MetaMethod::Le => "__le",
MetaMethod::Index => "__index",
MetaMethod::NewIndex => "__newindex",
MetaMethod::Call => "__call",
MetaMethod::ToString => "__tostring",
};
push_string(self.state, name)?;
self.push_value(Value::Function(self.create_callback(m)?));
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
protect_lua_closure(self.state, 3, 1, |state| {
ffi::lua_rawset(state, -3);
})?;
}
}
2017-05-21 18:50:59 -05:00
push_string(self.state, "__gc")?;
ffi::lua_pushcfunction(self.state, userdata_destructor::<RefCell<T>>);
protect_lua_closure(self.state, 3, 1, |state| {
ffi::lua_rawset(state, -3);
})?;
push_string(self.state, "__metatable")?;
ffi::lua_pushboolean(self.state, 0);
protect_lua_closure(self.state, 3, 1, |state| {
ffi::lua_rawset(state, -3);
})?;
let id = gc_guard(self.state, || {
ffi::luaL_ref(self.state, ffi::LUA_REGISTRYINDEX)
});
(*self.extra())
.registered_userdata
.insert(TypeId::of::<T>(), id);
Ok(id)
2017-05-21 18:50:59 -05:00
}
pub(crate) fn create_callback<'lua, 'callback>(
&'lua self,
func: Callback<'callback, 'static>,
) -> Result<Function<'lua>> {
unsafe extern "C" fn callback_call_impl(state: *mut ffi::lua_State) -> c_int {
callback_error(state, || {
if ffi::lua_type(state, ffi::lua_upvalueindex(1)) == ffi::LUA_TNIL {
return Err(Error::CallbackDestructed);
}
let lua = Lua {
state: state,
ephemeral: true,
ref_stack_slots: Default::default(),
};
let args = lua.setup_callback_stack_slots();
let func = get_userdata::<Callback>(state, ffi::lua_upvalueindex(1));
let results = (*func)(&lua, args)?;
let nresults = results.len() as c_int;
check_stack_err(state, nresults)?;
for r in results {
lua.push_value(r);
}
Ok(nresults)
})
}
unsafe {
let _sg = StackGuard::new(self.state);
check_stack(self.state, 4);
push_userdata::<Callback>(self.state, func)?;
ffi::lua_pushlightuserdata(
self.state,
&FUNCTION_METATABLE_REGISTRY_KEY as *const u8 as *mut c_void,
);
ffi::lua_rawget(self.state, ffi::LUA_REGISTRYINDEX);
ffi::lua_setmetatable(self.state, -2);
protect_lua_closure(self.state, 1, 1, |state| {
ffi::lua_pushcclosure(state, callback_call_impl, 1);
})?;
Ok(Function(self.pop_ref()))
}
}
// Does not require Send bounds, which can lead to unsafety.
pub(crate) unsafe fn make_userdata<T>(&self, data: T) -> Result<AnyUserData>
where
T: UserData,
{
let _sg = StackGuard::new(self.state);
check_stack(self.state, 4);
push_userdata::<RefCell<T>>(self.state, RefCell::new(data))?;
ffi::lua_rawgeti(
self.state,
ffi::LUA_REGISTRYINDEX,
self.userdata_metatable::<T>()? as ffi::lua_Integer,
);
ffi::lua_setmetatable(self.state, -2);
Ok(AnyUserData(self.pop_ref()))
}
unsafe fn create_lua(load_debug: bool) -> Lua {
unsafe extern "C" fn allocator(
_: *mut c_void,
ptr: *mut c_void,
_: usize,
nsize: usize,
) -> *mut c_void {
if nsize == 0 {
libc::free(ptr as *mut libc::c_void);
ptr::null_mut()
} else {
let p = libc::realloc(ptr as *mut libc::c_void, nsize);
if p.is_null() {
// We require that OOM results in an abort, and that the lua allocator function
// never errors. Since this is what rust itself normally does on OOM, this is
// not really a huge loss. Importantly, this allows us to turn off the gc, and
// then know that calling Lua API functions marked as 'm' will not result in a
// 'longjmp' error while the gc is off.
abort!("out of memory in Lua allocation, aborting!");
} else {
p as *mut c_void
}
}
}
let state = ffi::lua_newstate(allocator, ptr::null_mut());
// Ignores or `unwrap()`s 'm' errors, because this is assuming that nothing in the lua
// standard library will have a `__gc` metamethod error.
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
// Do not open the debug library, it can be used to cause unsafety.
ffi::luaL_requiref(state, cstr!("_G"), ffi::luaopen_base, 1);
ffi::luaL_requiref(state, cstr!("coroutine"), ffi::luaopen_coroutine, 1);
ffi::luaL_requiref(state, cstr!("table"), ffi::luaopen_table, 1);
ffi::luaL_requiref(state, cstr!("io"), ffi::luaopen_io, 1);
ffi::luaL_requiref(state, cstr!("os"), ffi::luaopen_os, 1);
ffi::luaL_requiref(state, cstr!("string"), ffi::luaopen_string, 1);
ffi::luaL_requiref(state, cstr!("utf8"), ffi::luaopen_utf8, 1);
ffi::luaL_requiref(state, cstr!("math"), ffi::luaopen_math, 1);
ffi::luaL_requiref(state, cstr!("package"), ffi::luaopen_package, 1);
ffi::lua_pop(state, 9);
init_error_metatables(state);
if load_debug {
ffi::luaL_requiref(state, cstr!("debug"), ffi::luaopen_debug, 1);
ffi::lua_pop(state, 1);
}
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
// Create the function metatable
ffi::lua_pushlightuserdata(
state,
&FUNCTION_METATABLE_REGISTRY_KEY as *const u8 as *mut c_void,
);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_newtable(state);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
push_string(state, "__gc").unwrap();
ffi::lua_pushcfunction(state, userdata_destructor::<Callback>);
ffi::lua_rawset(state, -3);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
push_string(state, "__metatable").unwrap();
ffi::lua_pushboolean(state, 0);
ffi::lua_rawset(state, -3);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_rawset(state, ffi::LUA_REGISTRYINDEX);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
// Override pcall and xpcall with versions that cannot be used to catch rust panics.
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_rawgeti(state, ffi::LUA_REGISTRYINDEX, ffi::LUA_RIDX_GLOBALS);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
push_string(state, "pcall").unwrap();
ffi::lua_pushcfunction(state, safe_pcall);
ffi::lua_rawset(state, -3);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
push_string(state, "xpcall").unwrap();
ffi::lua_pushcfunction(state, safe_xpcall);
ffi::lua_rawset(state, -3);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ffi::lua_pop(state, 1);
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
// Create ExtraData, and place it in the lua_State "extra space"
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
let extra_data = Box::into_raw(Box::new(ExtraData {
registered_userdata: HashMap::new(),
registry_unref_list: Arc::new(Mutex::new(Some(Vec::new()))),
}));
*(ffi::lua_getextraspace(state) as *mut *mut ExtraData) = extra_data;
rlua_debug_assert!(ffi::lua_gettop(state) == 0, "stack leak during creation");
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
check_stack(state, REF_STACK_SIZE);
ffi::lua_settop(state, REF_STACK_SIZE);
Lua {
state,
ephemeral: false,
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
ref_stack_slots: Default::default(),
}
}
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
// Set up the stack slot area in a callback, returning all arguments on the stack as a
// MultiValue
fn setup_callback_stack_slots<'lua>(&'lua self) -> MultiValue<'lua> {
unsafe {
check_stack(self.state, 2);
let nargs = ffi::lua_gettop(self.state);
let stack_nargs = cmp::min(REF_STACK_SIZE, nargs);
let mut args = MultiValue::new();
args.reserve(stack_nargs as usize);
// Convert all of the reference types in the ref stack area into LuaRef types in-place.
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
for i in 0..stack_nargs {
let n = stack_nargs - i;
let make_ref = || {
self.ref_stack_slots[(n - 1) as usize].set(1);
LuaRef {
lua: self,
ref_type: RefType::Stack { stack_slot: n },
}
};
match ffi::lua_type(self.state, n) {
ffi::LUA_TNIL => {
args.push_front(Value::Nil);
}
ffi::LUA_TBOOLEAN => {
args.push_front(Value::Boolean(ffi::lua_toboolean(self.state, n) != 0));
}
ffi::LUA_TLIGHTUSERDATA => {
args.push_front(Value::LightUserData(LightUserData(
ffi::lua_touserdata(self.state, n),
)));
}
ffi::LUA_TNUMBER => if ffi::lua_isinteger(self.state, n) != 0 {
args.push_front(Value::Integer(ffi::lua_tointeger(self.state, n)));
} else {
args.push_front(Value::Number(ffi::lua_tonumber(self.state, n)));
},
ffi::LUA_TSTRING => {
args.push_front(Value::String(String(make_ref())));
}
ffi::LUA_TTABLE => {
args.push_front(Value::Table(Table(make_ref())));
}
ffi::LUA_TFUNCTION => {
args.push_front(Value::Function(Function(make_ref())));
}
ffi::LUA_TUSERDATA => {
if let Some(err) = get_wrapped_error(self.state, n).as_ref() {
args.push_front(Value::Error(err.clone()));
} else {
args.push_front(Value::UserData(AnyUserData(make_ref())));
}
}
ffi::LUA_TTHREAD => {
args.push_front(Value::Thread(Thread(make_ref())));
}
_ => rlua_panic!("LUA_TNONE in setup_callback_stack_slots"),
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
}
}
if nargs < REF_STACK_SIZE {
check_stack(self.state, REF_STACK_SIZE - nargs);
ffi::lua_settop(self.state, REF_STACK_SIZE);
args
} else if nargs > REF_STACK_SIZE {
// If the total number of arguments exceeds the ref stack area, pop off the rest of
// the arguments as normal.
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
let mut extra_args = Vec::new();
extra_args.reserve((nargs - REF_STACK_SIZE) as usize);
for _ in REF_STACK_SIZE..nargs {
extra_args.push(self.pop_value());
}
extra_args.extend(args.into_vec_rev());
MultiValue::from_vec_rev(extra_args)
} else {
args
}
}
}
unsafe fn extra(&self) -> *mut ExtraData {
A lot of performance changes. Okay, so this is kind of a mega-commit of a lot of performance related changes to rlua, some of which are pretty complicated. There are some small improvements here and there, but most of the benefits of this change are from a few big changes. The simplest big change is that there is now `protect_lua` as well as `protect_lua_call`, which allows skipping a lightuserdata parameter and some stack manipulation in some cases. Second simplest is the change to use Vec instead of VecDeque for MultiValue, and to have MultiValue be used as a sort of "backwards-only" Vec so that ToLuaMulti / FromLuaMulti still work correctly. The most complex change, though, is a change to the way LuaRef works, so that LuaRef can optionally point into the Lua stack instead of only registry values. At state creation a set number of stack slots is reserved for the first N LuaRef types (currently 16), and space for these are also allocated separately allocated at callback time. There is a huge breaking change here, which is that now any LuaRef types MUST only be used with the Lua on which they were created, and CANNOT be used with any other Lua callback instance. This mostly will affect people using LuaRef types from inside a scope callback, but hopefully in those cases `Function::bind` will be a suitable replacement. On the plus side, the rules for LuaRef types are easier to state now. There is probably more easy-ish perf on the table here, but here's the preliminary results, based on my very limited benchmarks: create table time: [314.13 ns 315.71 ns 317.44 ns] change: [-36.154% -35.670% -35.205%] (p = 0.00 < 0.05) create array 10 time: [2.9731 us 2.9816 us 2.9901 us] change: [-16.996% -16.600% -16.196%] (p = 0.00 < 0.05) Performance has improved. create string table 10 time: [5.6904 us 5.7164 us 5.7411 us] change: [-53.536% -53.309% -53.079%] (p = 0.00 < 0.05) Performance has improved. call add function 3 10 time: [5.1134 us 5.1222 us 5.1320 us] change: [-4.1095% -3.6910% -3.1781%] (p = 0.00 < 0.05) Performance has improved. call callback add 2 10 time: [5.4408 us 5.4480 us 5.4560 us] change: [-6.4203% -5.7780% -5.0013%] (p = 0.00 < 0.05) Performance has improved. call callback append 10 time: [9.8243 us 9.8410 us 9.8586 us] change: [-26.937% -26.702% -26.469%] (p = 0.00 < 0.05) Performance has improved. create registry 10 time: [3.7005 us 3.7089 us 3.7174 us] change: [-8.4965% -8.1042% -7.6926%] (p = 0.00 < 0.05) Performance has improved. I think that a lot of these benchmarks are too "easy", and most API usage is going to be more like the 'create string table 10' benchmark, where there are a lot of handles and tables and strings, so I think that 25%-50% improvement is a good guess for most use cases.
2018-03-11 22:20:10 -05:00
*(ffi::lua_getextraspace(self.state) as *mut *mut ExtraData)
}
2017-05-21 18:50:59 -05:00
}