Running Lua Scripts on the CLR through Bytecode Translation
Departamento de Informatica, PUC-Rio
Rua Marques de Sao Vicente, 225 - 22453-900
Rio de Janeiro, RJ, Brasil
Departamento de Informatica, PUC-Rio
Rua Marques de Sao Vicente, 225 - 22453-900
Rio de Janeiro, RJ, Brasil
Abstract: The .NET Common Language Runtime (CLR) aims to provide
interoperability among code written in several different languages, but
porting scripting languages to it, so that scripts can run natively, has
been hard. This paper presents our approach for running scripts written
in Lua, a scripting language, on the .NET CLR.
Previous approaches for running scripting languages on the CLR have
focused on extending the CLR, statically generating CLR classes from
user-defined types in the source languages. They required either
language extensions or restrictions on the languages' dynamic
Our approach, on the other hand, focused on keeping the syntax and semantics
of the original language intact, while giving the ability to manipulate
CLR objects. We implemented a translator of Lua virtual machine bytecodes
to CLR bytecodes. Benchmarks show that the code our translator generates
performs better than the code generated by compilers that use the previous
Key Words: virtual machines, compilers, common language runtime,
The Microsoft .NET Framework provides interoperability among several
different languages, through a Common Language Runtime [Meijer
and Gough, 2002]. The .NET CLR specification is an ISO and ECMA standard
[Microsoft, 2002]. Microsoft has a commercial implementation
of the CLR for its Windows platform, and non-commercial implementations
already exist for several other platforms [Stutz, 2002,
Ximian, 2005]. Some languages already have compilers
for the CLR, and compilers for other languages are in several stages of
development [Bock, 2005].
Lua [Ierusalimschy, 2003, Ierusalimschy
et al., 1996] is a scripting language that is easy to embed, small,
fast, and flexible. It is interpreted and dynamically typed, has a simple
syntax, and has several reflexive facilities. Lua also has first-class
functions, lexical scoping, and coroutines. It is widely used in the development
of computer games.
Scripting languages are often used for connecting components
written in other languages ("glue" code). They are also used
for building prototypes, and as languages for configuration files. The
dynamic nature of these languages lets them use components without
previous type declarations and without the need for a compilation
phase. Although they lack static type checks, they perform extensive
type checking at runtime and provide detailed information in case of
errors. Ousterhout argues that the combination of these features can
increase developer productivity by a factor of two or more [Ousterhout, 1998].
This paper presents an approach for running Lua scripts natively on
the CLR, by translating bytecodes of the Lua Virtual Machine to bytecodes
of the Common Intermediate Language. The Common Intermediate Language,
or CIL, is the underlying language of the CLR. Our approach leaves the
syntax and semantics of the Lua scripts intact, while achieving adequate
performance. The bytecode translator is called Lua2IL.
Porting scripting languages to the CLR has been hard. ActiveState has
tried to build Perl and Python compilers, but abandoned both projects years
ago [ActiveState, 2000, Hammond,
2000]. Smallscript Inc. has been working on a Smalltalk compiler for
the CLR since 1999 [Smallscript Inc., 2000], but
there was no version of it available on February 2005.
A common trend among those projects is their emphasis on extending the
CLR. They map user-defined types in the source languages to new types in
the CLR, and then generate these types during compilation. The scripting
languages mentioned in the previous paragraph are object-oriented, and
their compilers attempt to map classes in those languages to CLR classes.
The emphasis on extending the CLR makes porting harder, as dynamic creation
and modification of types is a common feature of scripting languages, Lua
included. To provide a mapping from user-defined types in the source language
to CLR types, the compilers have to either restrict dynamic features of
the source languages or extend the syntax and semantics of the languages
to allow the definition of user-defined types that cannot be extended at
Our approach, on the other hand, emphasizes the full implementation
of the features of the original language, without impairing its ability
as a consumer. We create a runtime system for the scripting language (Lua,
in our case) on top of the CLR, adding the features of the language that
the CLR does not support. This isolates the language from the rest of the
CLR, however, so we also implement an interface layer (a bridge) that gives
full access to CLR types to the scripts. This layer has the capabilities
of a full CLS consumer.
The Common Language Specification (CLS) is a subset of the CLR specification
that establishes a set of rules for language interoperability [Microsoft,
2002, CLI Partition I Section 7.2.2].
Compilers that generate code capable of using CLS-compliant libraries
are called CLS consumers. Compilers that can produce new libraries
or extend existing ones are called CLS extenders (any language that
can define new types is an extender, in essence). A full CLS consumer must
be able to call any CLS-compliant method or delegate, even methods with
names that are keywords of the language; to call distinct methods of a
type with the same signature but from different interfaces; to instantiate
any CLS-compliant type, including nested types; to read and write any CLS-compliant
property; and access any CLS-compliant event. All of these features are
supported by the interface layer of Lua2IL, and are available to Lua scripts.
The rest of this paper is structured as follows: Section
2 describes the byte-code translator and the interface layer. Section
3 presents some related work and performance evaluations, and Section
4 presents some conclusions and future developments.
2 Translating Lua scripts to the CLR
Translating a Lua script to the Common Language Runtime involves several
issues, the actual translation of the bytecodes being just one of them.
First there should be a way to represent Lua types using the types in the
CLR; we cover this on Section 2.1. Then there is the
implementation of the VM instructions (the translation itself), covered
in Section 2.2. Then there are the features of the
Lua language that the Lua runtime environment implements: coroutines and
weak tables. We cover coroutines in Section 2.3 and
weak tables in Section 2.4. Finally, Lua scripts need
to manipulate other CLR objects (instantiate them, access their fields,
call their methods, and so on). Section 2.5 details
the implementation of Lua wrappers for other CLR objects.
2.1 Representing Lua types in the CLR
A naive approach to represent Lua types in the CLR would be to map
strings, booleans and nil directly to their respective CLR
types (double, string, bool, and
null). Two new CLR types would represent tables (associative
arrays) and functions. The advantage is that CLR code written in other
languages would work with Lua types directly, and vice-versa.
There is a severe disadvantage, though: Lua is dynamically typed,
so the code that Lua2IL produces would have to use the lowest common
denominator among CLR types, the object type. Most operations
would require a type check (with the isinst instruction of
the CLR) and a type cast. The code would have to box and unbox all
numbers and booleans operate on them, wasting memory and worsening
numbers are floating point numbers with double precision.
Lua2IL does not use this naive representation. Internally, the code
that it generates deals with instances of the LuaValue structure.
This structure has two fields, O, of type LuaReference,
and N, of type double. LuaValue instances with
the O field set to null represent Lua numbers. Subclasses
of LuaReference represent all the other Lua types.
Instances of LuaString, which use a CLR string internally,
represent Lua strings. Instances of LuaTable represent tables,
implementing a C# version of the algorithm that the Lua interpreter
uses. This algorithm breaks a table in an array part and a hash part, to
optimize the use of tables as arrays.
Each Lua function gets its own class, subclassed from a common ancestor,
the LuaClosure class. The classes have a Call virtual
method that executes the function (the translated bytecodes). Lua functions
are first class values, so the actual functions are instances of their
respective classes. The main body of the script is also a function, represented
by a class named MainFunction. Instantiating this class and then
calling its Call method runs the script.
Booleans and nil are a special case. There is a singleton object for
each boolean value (the singleton instances of TrueClass and FalseClass).
The same happens with nil (the singleton instance of NilClass).
Userdata are a Lua type that represents data from a host application
or a library. The LuaWrapper class represents userdata, and instances
of this class are proxies to CLR objects. We cover this class in more detail
in Section 2.5.
The representation we use does not need type casts, as there are common
denominators among all types (LuaValue and LuaReference).
To check if a value is a number, for example, the code just checks whether
its O field is null. If the O field is null,
the number is stored in the N field. As another example, the code
to index a value just checks if its O field is not null,
then calls the get_Item method of the O field. If the
value does not support this operation, the implementation of this operation
in the value's class throws an exception (the Lua interpreter would flag
an error in this case).
2.2 Translation of the bytecodes
When Lua2IL translates a Lua script (previously compiled to Lua Virtual
Machine bytecodes), it first reads the script and builds an in-memory tree
structure of it. Each function the script defines is a node of this tree,
and the body of the script is the root. Lua2IL walks this tree, in preorder,
compiling each node to a subclass of LuaClosure. The end result
is a library containing all those subclasses. For example, take the following
local nested = function()
When read, this will generate a tree with the body as root, functions
foo and bar as its children, and function nested
as child of function foo. Each of the nodes of this tree will
be compiled to a subclass of LuaClosure.
Calls to Lua functions do not use the standard CLR parameter passing
mechanism. When Lua2IL compiles a call to a Lua function, it has no way
of knowing how many parameters there are in the function being called,
nor how many values it will return (Lua functions can return multiple values).
One possible way to pass parameters to Lua functions would be using an
array, with return values collected using another array. The downside is
that two arrays must be instantiated and filled in every function call,
so we used an alternative way.
This alternative way is to have a Lua stack, an auxiliary stack
parallel to the CLR execution stack. Lua2IL uses the Lua stack for
parameter passing and collecting return values (the CLR stack still
keeps track of funcion calls and returns). Each function receives the
Lua stack when it is called, along with how many arguments it is
receiving, and returns how many return values it pushed on the
stack. For example, a function foo calling another function
bar with 10 and 3 as arguments would push both
arguments to the stack, then call bar passing the stack and
the number 2 (for two arguments). If bar wants to
return the values 3 and 1, it would push them into the
stack and return the number 2 (for two return values).
Lua functions also use the Lua stack to store their local variables,
instead of using CLR locals. This is required by our implementation of
lexical closures. The code does not use strict stack discipline when operating
on locals, however. For example, an addition operation of two locals gets
their values directly from their stack positions, storing it in another
stack position. There is no need to push the values to the top of the stack
before operating on them.
The LuaClosure class also defines a helper method that receives
an array of arguments, builds a new Lua stack, calls the function with
this stack, pops the return values, and then returns them in another array.
Other CLR programs can use this helper method as a more natural interface
to Lua functions.
The Lua stack is implemented as an array of LuaValue instances.
The stack starts small and automatically grows as needed, doubling in size
each time it is filled. The stack never shrinks, although object references
are cleared as the stack unwinds.
Using an auxiliary stack mimics the way that the Lua interpreter implements
the Lua Virtual Machine, and this is part of our approach of implementing
a runtime system on top of the CLR. The Lua VM is register-based, but its
registers are actually virtual, mapped to positions in the Lua execution
stack [Ierusalimschy, 2002]. Parameter passing and
return in the Lua interpreter works just as described earlier in this section.
The Lua stack also lets Lua2IL reuse the interpreter's implementation of
Due to the similarity between the execution models of the Lua interpreter
and of Lua2IL, we could do, for most of the Lua VM instructions, a straightforward
translation from the original ANSI C implementation of the Lua interpreter
to the Common Intermediate Language of the CLR. The translation of some
instructions is not as straightforward, though. The Lua interpreter implements
function calls, tail calls and function returns by creating and maintaining
its own activation records for each function call. Lua2IL uses the CLR
stack to do this, letting the CLR keep track of activation records for
each Lua function call, as each Lua function call is also a CLR method
The implementation of the function call instruction invokes the Call
method of the callee, passing the stack and number of arguments (pushed
into the stack by previous instructions). A preamble in the Call
method adjusts the arguments in the stack to the number of arguments that
the function expects, then clears the stack space that the function will
use (possibly growing the stack). The implementation of tail calls is slightly
different: Lua2IL first copies the arguments to the beginning of the stack
of the caller, then invokes the Call method using the tail call
instruction of the CIL. The implementation of function return copies the
return values to the end of the caller's stack space, then unwinds the
Lua stack and does a CIL method return.
The first prototype of Lua2IL translated each instruction as a call
to a helper method, like a threaded interpreter. The helper methods were
implemented in C#. This approach was slower, but easier to debug.
After implementations of all the VM instructions were done and debugged,
we changed Lua2IL to directly emit CIL code instead of just calling helper
methods, effectively inlining those methods. This inlining allowed a few
more optimizations. Many of the Lua VM instructions can operate on either
literal values or registers. In the threaded translator, the helper method
that implemented the instruction did the tests to see whether the operands
were literals or registers. The inlined translator does these tests at
translation time, and the CIL code that it generates is specialized to
operate either on a literal or a register.
All instructions are inlined, but a few of them are partially implemented
by C# helper methods. In these cases, the inlined portion deals with the
common case, and delegates other cases to a helper.
For example, the inlined implementation of arithmetic instructions does
the arithmetic operation itself when both operands are numbers, delegating
to a helper method when dealing with operands of other types.
Lua supports full asymmetric coroutines [Moura et al.,
2004]. A Lua coroutine is a first-class value. During its execution,
the coroutine can yield control back to its caller at any time, including
deep inside nested function calls. When a coroutine yields, its execution
is suspended. It can be later resumed from any point in the script, even
inside other coroutines. Returning from the main function of a coroutine
also yields control back to the caller, but the coroutine is marked as
dead and can no longer be resumed. If an error occurs during the execution
of a coroutine, this error is captured and returned to the caller, and
then the coroutine is marked as dead.
Lua2IL implements coroutines on top of CLR threads, using semaphores
for synchronization. Each coroutine has its own Lua stack, plus a CLR thread
and two binary semaphores. The semaphores are called resume and
yield, and are initially closed. When the script creates a coroutine,
the thread of the coroutine is started. The first action of this thread
is to try to decrement its resume semaphore, making CLR suspend
When another thread resumes a coroutine, it increments the
resume semaphore of the coroutine, restarting the
execution of the coroutine's thread. Then the caller thread decrements
the yield semaphore of the coroutine, suspending itself.
When a coroutine yields back to its caller, it increments its yield
semaphore, restarting the execution of the caller thread. Then the coroutine
decrements its resume semaphore, suspending itself. When the coroutine
returns (finishes executing), it increments its yield semaphore,
again restarting the caller thread, then the coroutine is flagged as terminated.
Any exception occurring during execution of a coroutine is trapped and
terminates the coroutine.
The downside of this implementation is the overhead caused by context
switches and synchronization, as each CLR thread is an OS thread, and swapping
among them involves a full context switch. This overhead is not present
in the coroutine implementation of the Lua interpreter. However, this is
the only way of implementing coroutines on the CLR using managed code (that
is, in a portable way). A native code implementation exists that uses Windows
fibers (cooperative threads), but it is not portable, it has problems interacting
with the CLR garbage collector and exception handling subsystems, and it
uses undocumented API calls [Shankar, 2003].
2.4 Weak Tables
The Lua VM supports weak references through weak tables. A weak
table may have weak keys, weak values, or both. If a weak key or value
is collected then its pair is removed from the table. The garbage collector
of the Lua interpreter puts weak tables in a list during the mark phase;
in the end of this phase the collector traverses the tables, removes all
pairs with unmarked weak references, and then proceeds with the sweep phase
of garbage collection.
The Lua2IL runtime implements weak tables by storing a CLR weak reference
to the key (or value) instead of the key itself. A CLR weak reference is
an instance of System.WeakReference; the Lua2IL runtime wraps
weak keys and values with instances of this type.
This implementation introduces overhead in every table access, unlike
the implementation the Lua interpreter uses. Besides this added overhead,
the current implementation does not remove a weak reference from the table
after the object it references is collected. The only event the CLR associates
with garbage collection is object finalization, through a Finalize
method. This method adds overhead to garbage collection (objects with this
method are collected differently). Implementing a notification system on
top of Finalize is possible: each object can keep a list of the
tables that have weak references to them, and the Finalize method
of each object can go through this list, removing the pairs that contain
the object. Besides the lack of elegance of this solution, it also implies
a performance hit over the whole Lua2IL runtime, as every Lua object would
have a Finalize method, even if the object is never put inside
a weak table.
A better way would be if the CLR notified the weak reference when it
became invalid, or if it let applications register methods that would be
executed after each garbage collection cycle. Another possible mechanism
would be the one present in the Java Virtual Machine: associate a queue
with each weak reference, and when the reference becomes invalid it is
added to this queue.
2.5 Working with CLR objects
Our approach manages to keep the syntax and semantics of the Lua language
intact. This comes with a price, though, as the scripts are isolated from
the rest of the CLR; they have no direct notion of external CLR types.
But we can give them access to these types through a layer that sits between
the Lua environment and the rest of the CLR, automatically translating
Lua types to CLR types and vice-versa, all at runtime. This integration
layer is a full CLS consumer. It lets Lua scripts manipulate CLR objects.
The scripts can get references to CLR types and use these references to
instantiate objects, then access fields of those objects and call their
methods. The scripts do all these operations with the standard Lua syntax.
They can even pass Lua functions to methods that expect delegates, to handle
events with Lua code, for example.
Lua2IL represents types and objects from the CLR with the LuaWrapper
class, which has two other subclasses; one of them represents types, and
is responsible for object instantiation and access to static members, while
the other represents instances, and is responsible for access to instance
members. The LuaWrapper class and its subclasses have methods
that implement indexing (both to read and write values) and function invocation.
For example, an expression like obj:foo(arg1, arg2) is translated
by the Lua parser to the equivalent expression obj["foo"](obj,
arg1, arg2). The obj["foo"] subexpression emits
a bytecode for an indexing operation, and Lua2IL translates this bytecode
to a call to an indexing method on obj. If obj is a Lua
table, this method returns the value stored in the table under the "foo"
key. If obj is an instance of LuaWrapper, its indexing
method searches for a method foo in the CLR object represented
by obj, using the CLR reflection API. If the search finds the
method then the indexing method returns a proxy to it, otherwise it returns
Continuing the previous example, the compilation of the call to the
value returned by obj["foo"] emits a call (or tail call)
bytecode, which Lua2IL translates as a call to the method Call
of the proxy. The proxy's Call method pops the arguments from
the Lua stack, converts them to the types that the CLR method requires,
and then calls it. If the method is overloaded, the proxy tries to call
each of the methods, in the order they are defined, and throws an excep
tion if all the calls fail because of incompatible arguments.
The cost of searching for a method with the reflection API is high,
so the instances of LuaWrapper cache proxies. This cache is shared
by all instances of a same type. Proxies to overloaded methods cache the
last successful method that was called; on the next call the proxy tries
this method first.
Going back to the obj["foo"] example, if foo
is a field, the indexing method of obj finds foo, using
reflection, and returns the value of foo in the CLR object represented
by obj. The proxy caches the field (not its value), so the next
access does not need a new reflexive search. Properties are treated in
a similar manner. Writing to a field, like in obj.foo = bar, emits
an indexing bytecode that sets the value at the index. This is translated
by Lua2IL to a call to an indexing method that sets the value. This method
finds the foo field, using reflection, converts bar to
its type, and assigns to the field. The proxy caches the foo field,
as mentioned in the previous paragraph. Writing to properties is again
treated in a similar manner.
The Lua2IL runtime automatically converts Lua functions to delegates,
if a method expects a delegate as a parameter. A script can use this to
register Lua functions as event handlers, for example. The runtime dynamically
generates a new class that implements a method with the delegate's signature.
This method dispatches to a Lua function. The runtime instantiates this
class with the function being converted, and creates a delegate from this
instance. The dynamic classes are generated with the Reflection.Emit API
and kept in a temporary, memory-only library.
3 Related Work
During the years 1999 and 2000, Microsoft sponsored the development
of a Python compiler for the CLR, called Python for .NET [Hammond,
2000]. Python for .NET traverses the abstract syntax tree generated
by the CPython interpreter, emitting CIL code through the Reflection.Emit
API. The implementation has some similarities to the implementation of
Lua2IL: Python for .NET defines a PyObject structure for its values,
and a IPyType interface that define what operations can be done
on those values (the Lua2IL equivalents are the LuaValue structure
and LuaReference class, respectively).
Python for .NET is different from Lua2IL in the sense that it
generates CLR classes from Python classes, aided by special
annotations (comments) in the source code. Primitive types of the
Python language are mapped to primitive CLR types. Around 95% of the
Python core is implemented, according the author. Missing features are
primitive types without a direct mapping to CLR primitive types
(arbitrary size integers, complex numbers and ellipses), and built-in
methods of Python classes, used for dynamic extension of classes and
objects. The language syntax was not modified. The development of this
compiler halted about three years ago. The last available prototype is
dated April 2002, with parts of it dated April 2000.
Perl for .NET is a Perl compiler for the CLR, and was developed
by ActiveState between 1999 and 2000 [ActiveState, 2000].
The compiler works as a back-end to the Perl interpreter, generating C#
code (not CIL) that calls a Perl runtime for its operations. The compiler
also generates CLR classes from Perl classes, but there is no information
about how much of the Perl language is covered by the compiler, and the
source code for it is not available. The last available prototype is dated
June 2000, and does not work with release versions of the CLR, only with
JScript.NET is an extension of the JScript language (or EcmaScript)
with a compiler for the CLR, and is part of the Microsoft .NET Software
Development Kit. The language was extended with classes and optional type
declarations. The dynamic features of JScript are still available, although
interoperation with other CLR code is compromised if type declarations
are not used: delegates must be declared with the correct signature (including
type declarations), and declared inside a class. The code generated by
the compiler uses CLR types natively, requiring type checking and casts
in every operation with dynamic typing.
||Ackermann function, arguments 3 and 8
||Fibonacci numbers, the 30th number
||Random number generation, generate 1,000,000 numbers between 0 and
||Sieve of Eratosthenes, from 2 to 8,192, 100 runs
||30 × 30 matrix multiplication, 100 runs
||Heap sort on an array of 100,000 random numbers
Table 1: Scripts for the first compiler performance test
S# is a dialect of Smalltalk developed by SmallScript Corporation,
and S#.NET is a S# compiler for the CLR. According to its
author, the compiler and the language runtime are ready, but still need
to be integrated with the Visual Studio.NET development environment before
being released to the public. The compiler has been under development since
IronPython is a new Python compiler for the CLR, and is being developed
by Jim Hugunin [Hugunin, 2004]. IronPython is the
most similar to Lua2IL in its approach: although it uses its own parser,
written in C#, it maintains the Python syntax, supports all of the
Python core, including all of its dynamic features, and does not generate
CLR classes from Python classes. IronPython does some aggressive optimizations
on its generated code, specially if some of the more dynamic features of
Python are not used. Like the JScript.NET compiler, it uses native CLR
types whenever possible, but does not use type annotations, and any Python
function can be a delegate.
3.1 Performance Evaluation
Our first performance test is the execution of six scripts from
The Great Win32 Computer Language Shootout [Bagley, 2005], mainly involving arithmetic
operations, recursion, and array accesses. The goal is to evaluate
the performance of the code generated by the compilers when running
the primitive operations of the languages. A description of each test
script and the arguments of its execution is on Table
We tested the Lua2IL, JScript.NET, Python for .NET and IronPython 0.6
compilers. The same scripts compiled by Lua2IL were also executed by the
Lua 5.0.2 interpreter, and the same scripts compiled by Python for .NET
and Iron-Python were executed by the CPython 2.4.1 interpreter. We did
not test the Perl for .NET compiler, as it did not work with the version
of the CLR we used.
Figure 1: Results for the first performance test
The results are shown on Figure 1. The times are
in seconds, and all the scripts were run on the same machine, under the
The CPython interpreter is the binary distribution, available from www.python.org.
Python for .NET did not compile the matrix and heapsort
scripts, even though these scripts were syntactically correct. IronPython
successfully compiled the ack script, but it ran out of stack space
during execution and crashed. CPython also ran out of stack space during
execution of this script.
The Python for .NET compiler lags behind the others, as its authors
halted the development of the compiler before writing an optimizer. Next
comes the JScript.NET compiler, penalized by the inefficient code it generates
for numerical operations. Binary operations, except addition, are computed
by a generic evaluator object that receives a numeric code for the operation
and both operands. These evaluator objects are created in the heap, at
each execution of the function, so heavily recursive numerical code is
memory-intensive and very demanding on the garbage-collector.
2.8GHz HT, with 512Mb memory, running Windows XP Professional with version
1.1 of the .NET Common Language Runtime. The Lua interpreter was compiled
by the Microsoft 32-bit C/C++ optimizing compiler, version 13.10.3077 for
80x86, with the /O2 switch.
Both Lua2IL and IronPython show close results, with an advantage for
Lua2IL in numerical code, probably due to the type checks present in the
code IronPython generates. IronPython is at a slight advantage in code
that uses arrays. Arrays are an optimization of tables in Lua2IL, and the
Lua2IL run-time must check, at each array access, if the index is an integer
and if it is in the bounds of the array part of the table, defaulting to
use the hash part if each of these tests fail. Both Lua2IL and IronPython
are ahead of the CPython interpreter, but the Lua interpreter is the fastest
The Lua, Python, and JScript languages have similar semantics for the
scope of these tests, so it is fair to assume that differences in performance
are due to how the compilers and interpreters were implemented, and not
due to intrinsic differences among the languages.
The second performance test is a measuring of the time it takes to complete
a method call to a CLR object. The test was done with code generated by
the Lua2IL, JScript.NET (using late binding, with no type declarations),
and IronPython compilers. The Python for .NET compiler could not instantiate
the types in the assembly used in this test. We evaluated times for calls
to six distinct methods. They vary by the number and types of their parameters.
Three of the methods have all parameter and return values of type Int32,
and are called with zero, one, and two parameters. The other three methods
have parameters and return values of type object.
The results of the test are show on Figure 2, and
are in microseconds. They were collected on the same machine and under
the same conditions of the first performance test. The Lua columns
show the times for calls from the Lua interpreter, using the LuaInterface
[Mascarenhas and Ierusalimschy, 2004] library, a Lua
to CLR bridge. The other columns show the times for calls from code generated
by the respective compilers.
For this test, the code generated by JScript.NET and Lua2IL are very
close, within 10% of each other. This shows that any overhead introduced
by the peculiarities of the code generated by each compiler is dwarfed
by the time for the actual reflexive invocation of the method. IronPython,
on the other hand, clearly does not optimize method calls as well as it
optimizes the execution of Python code.
The higher times for the calls from the Lua interpreter are a result
of the overhead involved in passing values from the environment of the
Lua interpreter to the managed environment of the CLR. This shows the performance
advantage of code running directly under the CLR, which needs much less
This paper presented an approach for running scripts from Lua, a dynamically
typed language, on the Common Language Runtime. The approach works by translating
the bytecodes of the Lua virtual machine to bytecodes of the CLR.
Figure 2: Times for method calls
The goal was to keep the syntax and semantics of the language unchanged;
any script that the Lua interpreter executes, as long as it does not use
library code, should be translatable to CIL code that with the same behavior.
There is also an integration layer that lets scripts freely manipulate
Previous attempts at creating CLR compilers for scripting languages
have focused on static generation of classes, either by extending the language,
in the case of JScript, or by restricting dynamic features, in the case
of the Python for .NET compiler. Our approach focuses on reproducing the
semantics of the language, even if it requires extensive runtime support,
while offering access to CLR objects. We think the role of a consumer,
instead of a creator, of CLR types is more suited to scripting languages.
A recent Python compiler for the CLR, IronPython, uses a similar approach
to ours, and matches some of our results.
The goal of keeping the semantics of the language was almost fulfilled,
with only weak tables having a different semantic, due to the absence of
any mechanism in the CLR that notifies when a weak reference becomes invalid.
Lua2IL does some optimizations in the generated code, like generating
specialized implementations of Lua bytecodes. The integration layer also
optimizes calls to methods of CLR objects, caching the methods that are
discovered through reflection. We compared the performance of the code
generated by Lua2IL with code generated by three other compilers for dynamically
typed languages: a commercial compiler of the JScript language (developed
by Microsoft), and two open source prototype implementations of Python
compilers. We also compared the performance with that of the same code
executed by the latest release of the Lua and Python interpreters. The
results are mixed, with the code generated by Lua2IL performing better
than the others in tests that are not dominated by array accesses. Lua2IL,
like the Lua interpreter, implements arrays as an optimization of tables,
not as a dedicated array type. Lua2IL performs well even in tests dominated
by array accesses, though, coming close to the fastest compiler, IronPython,
an ahead of the rest.
Performance evaluation of the time taken by calls to the methods of
other CLR objects shows that the code generated by Lua2IL performs similarly
to code generated by JScript.NET, and better than the code generated both
by IronPython and by a Lua to CLR bridge. The overall time is dominated
by reflexive invocation, on code generated by both Lua2IL and JScript.NET.
For the future, we are working on adding an implementation of coroutines
on the CLR that does not depend on threads. This will enable a more effcient
implementation of Lua coroutines. We also plan on making the CLR garbage
collector more flexible, so it can better adapt to languages with finalization
semantics different from the one used by C#, such as the weak tables
of the Lua language. Another plan is to investigate how to enable faster
execution of script- ing languages by the CLR, to bring the performance
nearer to the performance of statically-typed languages.
[ActiveState, 2000] ActiveState (2000). Release
Information for the ActiveState Perl for .NET compiler. Available at http://www.activestate.com/Corporate/
[Bagley, 2005] Bagley, D. (2005). The Great Computer
Language Shootout. Available at http://dada.perl.it/shootout/.
[Bock, 2005] Bock, J. (2005). .NET Languages. Available
[Hammond, 2000] Hammond, M. (2000). Python for .NET:
Lessons Learned. Avail able at http://www.activestate.com/Corporate/Initiatives/NET/Python
for . NET whitepaper.pdf.
[Hugunin, 2004] Hugunin, J. (2004). IronPython:
A fast Python implementation for .NET and Mono. In PyCON 2004 International
[Ierusalimschy, 2002] Ierusalimschy, R. (2002).
The Virtual Machine of Lua 5.0. In Lightweight Languages 2003 Workshop.
Available at http://www.inf.puc-rio.br/~roberto/talks/lua-ll3.pdf.
[Ierusalimschy, 2003] Ierusalimschy, R. (2003).
Programming in Lua. Lua.org.
[Ierusalimschy et al., 1996] Ierusalimschy, R.,
de Figueiredo, L. H., and Celes, W. (1996). Lua — An Extensible Extension
Language. Software: Practice and Experience, 26(6):635-652.
[Mascarenhas and Ierusalimschy, 2004] Mascarenhas,
F. and Ierusalimschy, R. (2004). LuaInterface: Scripting the .NET CLR with
Lua. Journal of Universal Computer Science, 10(7):892-908.
[Meijer and Gough, 2002] Meijer, E. and Gough, J.
(2002). Technical Overview of the Common Language Runtime. Technical report,
Microsoft Research. Available at http://research.microsoft.com/~emeijer/Papers/CLR.pdf.
[Microsoft, 2002] Microsoft (2002). ECMA C# and
Common Language Infrastructure Standards. Available at http://msdn.microsoft.com/net/ecma/.
[Moura et al., 2004] Moura, A. L. d.,
Rodriguez, N., and Ierusalimschy, R. (2004). Coroutines in Lua. Journal of Universal Computer
[Ousterhout, 1998] Ousterhout,
J. (1998). Scripting: Higher Level Programming for the 21st
Century. IEEE Computer, 31(3):23-30.
[Shankar, 2003] Shankar,
A. (2003). Implementing Coroutines for .NET by Wrapping the Unmanaged
Fiber API. MSDN Magazine, 18(9). Available at http://msdn.microsoft.com/msdnmag/issues/03/09/CoroutinesinNET/default.aspx.
[Smallscript Inc., 2000] Smallscript Inc. (2000).
S#.NET Techpreview Software Release. Available at http://www.smallscript.com/Community/calendar_home.asp.
[Stutz, 2002] Stutz, D. (2002). The Microsoft
Shared Source CLI Implementation. Available at http://msdn.microsoft.com/library/en-us/Dndotnet/html/
[Ximian, 2005] Ximian (2005). The Mono Project.
Available at http://www.go-mono.com/.
All of the links in these references have been verified and are working
as of February 2005.