Ravi is a derivative/dialect of Lua 5.3 with limited optional static typing and an LLVM powered JIT compiler. The name Ravi comes from the Sanskrit word for the Sun. Interestingly a precursor to Lua was Sol which had support for static types; Sol means the Sun in Portugese.

Lua is perfect as a small embeddable dynamic language so why a derivative? Ravi extends Lua with static typing for greater performance under JIT compilation. However, the static typing is optional and therefore Lua programs are also valid Ravi programs.

There are other attempts to add static typing to Lua - e.g. Typed Lua but these efforts are mostly about adding static type checks in the language while leaving the VM unmodified. The Typed Lua effort is very similar to the approach taken by Typescript in the JavaScript world. The static typing is to aid programming in the large - the code is eventually translated to standard Lua and executed in the unmodified Lua VM.

My motivation is somewhat different - I want to enhance the VM to support more efficient operations when types are known. Type information can be exploited by JIT compilation technology to improve performance. At the same time, I want to keep the language safe and therefore usable by non-expert programmers.

• Optional static typing for Lua
• Type specific bytecodes to improve performance
• Compatibility with Lua 5.3 (see Compatibility section below)
• LLVM powered JIT compiler
• Additionally a libgccjit based alternative JIT compiler is also available

See Ravi Documentation. As more stuff is built I will keep updating the documentation so please revisit for latest information.

Also see the slides I presented at the Lua 2015 Workshop.

The project was kicked off in January 2015.

The LLVM JIT compiler is functional. The Lua and Ravi bytecodes currently implemented in LLVM are described in JIT Status page.

Ravi also provides an LLVM binding; this is still work in progress so please check documentation for the latest status.

As of July 2015 the libgccjit based JIT implementation is also functional but some byte codes are not yet compiled, and featurewise this implementation is somewhat lagging behind the LLVM based implementation.

For performance benchmarks please visit the Ravi Performance Benchmarks page.

Ravi allows you to annotate local variables and function parameters with static types. The supported types and the resulting behaviour are as follows:

integer

denotes an integral value of 64-bits.

number

denotes a double (floating point) value of 8 bytes.

integer[]

denotes an array of integers

number[]

denotes an array of numbers

table

a Lua table

Declaring the types of local variables and function parameters has following advantages.

• integer and number types are automatically initialized to zero
• Arithmetic operations on numeric types make use of type specific bytecodes which leads to more efficient JIT compilation
• Specialised operators to get/set from array types are implemented; this makes array access more efficient in JIT mode as the access can be inlined
• Declared tables allow specialized opcodes for table gets involving integer and short literal string keys; these opcodes result in more efficient JIT code
• Values assigned to typed variables are checked statically when possible; if the values are results from a function call then runtime type checking is performed
• The standard table operations on arrays are checked to ensure that the type is not subverted
• Even if a typed variable is captured in a closure its type must be respected
• When function parameters are decorated with types, Ravi performs an implicit coersion of those parameters to the required types. If the coersion fails a runtime error occurs.

The array types (number[] and integer[]) are specializations of Lua table with some additional special behaviour:

• Array types are not compatible with declared table variables, i.e. following is not allowed:

  local t: table = {}
  local t2: number[] = t  -- error!
  
  local t3: number[] = {}
  local t4: table = t3    -- error!

  But following is okay:

  local t5: number[] = {}
  local t6 = t5           -- t6 treated as table

  The reason for these restrictions is that declared table types generate optimized JIT code which assumes that the keys are integers or literal short strings. Similarly declared array types result in specialized JIT code that assume integer keys and numeric values. The generated JIT code would be incorrect if the types were not as expected.

• Indices >= 1 should be used when accessing array elements. Ravi arrays (and slices) have a hidden slot at index 0 for performance reasons, but this is not visible in pairs() or ipairs(), or when initializing an array using a literal initializer; only direct access via the [] operator can see this slot.

• Arrays must always be initialized:

  local t: number[] = {} -- okay
  local t2: number[]     -- error!

  This restriction is placed as otherwise the JIT code would need to insert tests to validate that the variable is not nil.

• An array will grow automatically if user sets the element just past the array length:

  local t: number[] = {}
  t[1] = 4.2             -- okay, array grows by 1
  t[5] = 2.4             -- error! as attempt to set value 

• It is an error to attempt to set an element that is beyond len+1
• The current used length of the array is recorded and returned by len operations
• The array only permits the right type of value to be assigned (this is also checked at runtime to allow compatibility with Lua)
• Accessing out of bounds elements will cause an error, except for setting the len+1 element

• It is possible to pass arrays to functions and return arrays from functions. Arrays passed to functions appear as Lua tables inside those functions if the parameters are untyped - however the tables will still be subject to restrictions as above. If the parameters are typed then the arrays will be recognized at compile time:

  local function f(a, b: integer[], c)
    -- Here a is dynamic type
    -- b is declared as integer[]
    -- c is also a dynamic type
    b[1] = a[1] -- Okay only if a is actually also integer[]
    b[1] = c[1] -- Will fail if c[1] cannot be converted to an integer
  end
  
  local a : integer[] = {1}
  local b : integer[] = {}
  local c = {1}
  
  f(a,b,c)        -- ok as c[1] is integer
  f(a,b, {'hi'})  -- error!

• Arrays returned from functions can be stored into appropriately typed local variables - there is validation that the types match:

  local t: number[] = f() -- type will be checked at runtime

• Operations on array types can be optimised to special bytecode and JIT only when the array type is statically known. Otherwise regular table access will be used subject to runtime checks.
• Array types ignore __index, __newindex and __len metamethods.
• Array types cannot be set as metatables for other values.
• pairs() and ipairs() work on arrays as normal
• There is no way to delete an array element.
• The array data is stored in contiguous memory just like native C arrays; morever the garbage collector does not scan the array data

A declared table (as shown below) has some additional nuances:

local t: table = {}

• Like array types, a variable of table type must be initialized

• Array types are not compatible with declared table variables, i.e. following is not allowed:

  local t: table = {}
  local t2: number[] = t -- error!

• When short string literals are used to access a table element, specialized bytecodes are generated that are more efficiently JIT compiled:

  local t: table = { name='dibyendu'}
  print(t.name) -- The GETTABLE opcode is specialized in this case

• As with array types, specialized bytecodes are generated when integer keys are used

Following library functions allow creation of array types of defined length.

table.intarray(num_elements, initial_value)

creates an integer array of specified size, and initializes with initial value. The return type is integer[]. The size of the array cannot be changed dynamically, i.e. it is fixed to the initial specified size. This allows slices to be created on such arrays.

table.numarray(num_elements, initial_value)

creates an number array of specified size, and initializes with initial value. The return type is number[]. The size of the array cannot be changed dynamically, i.e. it is fixed to the initial specified size. This allows slices to be created on such arrays.

Ravi does not support defining new types, or structured types based on tables. This creates some practical issues when dynamic types are mixed with static types. For example:

local t = { 1,2,3 }
local i: integer = t[1] -- generates an error

Above code generates an error as the compiler does not know that the value in t[1] is an integer. However often we as programmers know the type that is expected and it would be nice to be able to tell the compiler what the expected type of t[1] is above. To enable this Ravi supports type assertion operators. A type assertion is introduced by the '@' symbol, which must be followed by the type name. So we can rewrite the above example as:

local t = { 1,2,3 }
local i: integer = @integer( t[1] )

The type assertion operator is a unary operator and binds to the expression following the operator. We use the parenthesis above to enure that the type assertion is applied to t[1] rather than t. More examples are shown below:

local a: number[] = @number[] { 1,2,3 }
local t = { @number[] { 4,5,6 }, @integer[] { 6,7,8 } }
local a1: number[] = @number[]( t[1] )
local a2: integer[] = @integer[]( t[2] )

For a real example of how type assertions can be used, please have a look at the test program gaussian2.lua

Since release 0.6 Ravi supports array slices. An array slice allows a portion of a Ravi array to be treated as if it is an array - this allows efficient access to the underlying array elements. Following new functions are available:

table.slice(array, start_index, num_elements)

creates a slice from an existing fixed size array - allowing efficient access to the underlying array elements.

Slices access the memory of the underlying array; hence a slice can only be created on fixed size arrays (constructed by table.numarray() or table.intarray()). This ensures that the array memory cannot be reallocated while a slice is referring to it. Ravi does not track the slices that refer to arrays - slices get garbage collected as normal.

Slices cannot extend the array size for the same reasons above.

The type of a slice is the same as that of the underlying array - hence slices get the same optimized JIT operations for array access.

Each slice holds an internal reference to the underlying array to ensure that the garbage collector does not reclaim the array while there are slices pointing to it.

For an example use of slices please see the matmul1.ravi benchmark program in the repository. Note that this feature is highly experimental and not very well tested.

Example of code that works - you can copy this to the command line input:

function tryme()
  local i,j = 5,6
  return i,j
end
local i:integer, j:integer = tryme(); print(i+j)

When values from a function call are assigned to a typed variable, an implicit type coersion takes place. In above example an error would occur if the function returned values that could not converted to integers.

In the following example, the parameter j is defined as a number, hence it is an error to pass a value that cannot be converted to a number:

function tryme(j: number)
  for i=1,1000000000 do
    j = j+1
  end
  return j
end
print(tryme(0.0))

An example with arrays:

function tryme()
  local a : number[], j:number = {}
  for i=1,10 do
    a[i] = i
    j = j + a[i]
  end
  return j
end
print(tryme())

Another example using arrays. Here the function receives a parameter arr of type number[] - it would be an error to pass any other type to the function because only number[] types can be converted to number[] types:

function sum(arr: number[]) 
  local n: number = 0.0
  for i = 1,#arr do
    n = n + arr[i]
  end
  return n
end

print(sum(table.numarray(10, 2.0)))

The table.numarray(n, initial_value) creates a number[] of specified size and initializes the array with the given initial value.

To keep with Lua's dynamic nature Ravi uses a mix of compile type checking and runtime type checks. However due to the dynamic nature of Lua, compilation happens at runtime anyway so effectually all checks are at runtime.

The LLVM based JIT compiler is functional. Most bytecodes other than bit-wise operators are JIT compiled when using LLVM, but there are restrictions as described in compatibility section below. Everything described below relates to using LLVM as the JIT compiler.

There are two modes of JIT compilation.

auto mode

in this mode the compiler decides when to compile a Lua function. The current implementation is very simple - any Lua function call is checked to see if the bytecodes contained in it can be compiled. If this is true then the function is compiled provided either a) function has a fornum loop, or b) it is largish (greater than 150 bytecodes) or c) it is being executed many times (> 50). Because of the simplistic behaviour performance the benefit of JIT compilation is only available if the JIT compiled functions will be executed many times so that the cost of JIT compilation can be amortized.

manual mode

in this mode user must explicitly request compilation. This is the default mode. This mode is suitable for library developers who can pre compile the functions in library module table.

A JIT api is available with following functions:

ravi.jit([b])

returns enabled setting of JIT compiler; also enables/disables the JIT compiler; defaults to true

ravi.auto([b [, min_size [, min_executions]]])

returns setting of auto compilation and compilation thresholds; also sets the new settings if values are supplied; defaults are false, 150, 50.

ravi.compile(func_or_table[, options])

compiles a Lua function (or functions if a table is supplied) if possible, returns true if compilation was successful for at least one function. options is an optional table with compilation options - in particular omitArrayGetRangeCheck - which disables range checks in array get operations to improve performance in some cases. Note that at present if the first argument is a table of functions and has more than 100 functions then only the first 100 will be compiled. You can invoke compile() repeatedly on the table until it returns false. Each invocation leads to a new module being created; any functions already compiled are skipped.

ravi.iscompiled(func)

returns the JIT status of a function

ravi.dumplua(func)

dumps the Lua bytecode of the function

ravi.dumpir(func)

dumps the IR of the compiled function (only if function was compiled; only LLVM version)

ravi.dumpasm(func)

dumps the machine code using the currently set optimization level (only if function was compiled; only LLVM)

ravi.optlevel([n])

sets LLVM optimization level (0, 1, 2, 3); defaults to 2

ravi.sizelevel([n])

sets LLVM size level (0, 1, 2); defaults to 0

ravi.tracehook([b])

Enables support for line hooks via the debug api. Note that enabling this option will result in inefficient JIT as a call to a C function will be inserted at beginning of every Lua bytecode boundary; use this option only when you want to use the debug api to step through code line by line

Ravi should be able to run all Lua 5.3 programs in interpreted mode, but there are some differences:

• Ravi supports optional typing and enhanced types such as arrays (described above). Programs using these features cannot be run by standard Lua. However all types in Ravi can be passed to Lua functions; operations on Ravi arrays within Lua code will be subject to restrictions as described in the section above on arrays.
• Values crossing from Lua to Ravi will be subjected to typechecks should these values be assigned to typed variables.
• Upvalues cannot subvert the static typing of local variables (issue #26)
• Certain Lua limits are reduced due to changed byte code structure. These are described below.

When JIT compilation is enabled some things will not work:

• You cannot yield from a compiled function as compiled code does not support coroutines (issue 14); as a workaround Ravi will only execute JITed code from the main Lua thread; any secondary threads (coroutines) execute in interpreter mode.
• In JITed code tailcalls are implemented as regular calls so unlike Lua VM which supports infinite tail recursion JIT compiled code only supports tail recursion to a depth of about 110 (issue #17)

• CMake
• LLVM 3.7

The build is CMake based.

I built LLVM 3.7 from source. I used the following sequence from the VS2015 command window:

cd \github\llvm
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=c:\LLVM37 -DLLVM_TARGETS_TO_BUILD="X86" -G "Visual Studio 14 Win64" ..  

I then opened the generated solution in VS2015 and performed a INSTALL build from there. Note that if you perform a Release build of LLVM then you will also need to do a Release build of Ravi otherwise you will get link errors.

On Ubuntu I found that the official LLVM distributions don't work with CMake. The CMake config files appear to be broken. So I ended up downloading and building LLVM 3.7 from source and that worked. The approach is similar to that described for MAC OS X below.

I am using Max OSX Yosemite. Pre-requisites are XCode 6.1 and CMake. Ensure cmake is on the path. Assuming that LLVM source has been extracted to $HOME/llvm-3.7.0.src I follow these steps:

cd llvm-3.7.0.src
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$HOME/LLVM -DLLVM_TARGETS_TO_BUILD="X86" ..
make install

I am developing Ravi using Visual Studio 2015 Community Edition on Windows 8.1 64bit, gcc on Unbuntu 64-bit, and clang/Xcode on MAC OS X.

Assuming that LLVM has been installed as described above, then on Windows I invoke the cmake config as follows:

cd build
cmake -DLLVM_JIT=ON -DCMAKE_INSTALL_PREFIX=c:\ravi -DLLVM_DIR=c:\LLVM37\share\llvm\cmake -G "Visual Studio 14 Win64" ..

I then open the solution in VS2015 and do a build from there.

On Ubuntu I use:

cd build
cmake -DLLVM_JIT=ON -DCMAKE_INSTALL_PREFIX=$HOME/ravi -DLLVM_DIR=$HOME/LLVM/share/llvm/cmake -DCMAKE_BUILD_TYPE=Release -G "Unix Makefiles" ..
make

Note that on a clean install of Ubuntu 15.10 I had to install following packages:

• cmake
• git
• libreadline-dev

On MAC OS X I use:

cd build
cmake -DLLVM_JIT=ON -DCMAKE_INSTALL_PREFIX=$HOME/ravi -DLLVM_DIR=$HOME/LLVM/share/llvm/cmake -DCMAKE_BUILD_TYPE=Release -G "Xcode" ..

I open the generated project in Xcode and do a build from there.

You can omit -DLLVM_JIT=ON option above to build Ravi with a null JIT implementation.

The Ravi build creates a shared library, the Lua executable and some test programs.

The lua command recognizes following environment variables. Note that these are only for internal debugging purposes.

RAVI_DEBUG_EXPR

if set to a value this triggers debug output of expression parsing

RAVI_DEBUG_CODEGEN

if set to a value this triggers a dump of the code being generated

RAVI_DEBUG_VARS

if set this triggers a dump of local variables construction and destruction

Also see section above on available API for dumping either Lua bytecode or LLVM IR for compiled code.

• Feb-Jun 2015 - implement JIT compilation using LLVM
• Jun-Jul 2015 - libgccjit based alternative JIT
• 2016 priorties
  • IDE support (Visual Studio Code)
  • BLAS and LAPACK
  • GNU Scientific library
  • symengine

MIT License for LLVM version.

Since the reason for introducing optional static typing is to enhance performance primarily - not all types benefit from this capability. In fact it is quite hard to extend this to generic recursive structures such as tables without encurring significant overhead. For instance - even to represent a recursive type in the parser will require dynamic memory allocation and add great overhead to the parser.

From a performance point of view the only types that seem worth specializing are:

• integer (64-bit int)
• number (double)
• array of integers
• array of numbers

I want to build on existing Lua types rather than introducing completely new types to the Lua system. I quite like the minimalist nature of Lua. However, to make the execution efficient I am adding new type specific opcodes and enhancing the Lua parser/code generator to encode these opcodes only when types are known. The new opcodes will execute more efficiently as they will not need to perform type checks. Morever, type specific instructions will lend themselves to more efficient JIT compilation.

I am adding new opcodes that cover arithmetic operations, array operations, variable assignments, etc..

An immediate issue is that the Lua bytecode structure has a 6-bit opcode which is insufficient to hold the various opcodes that I will need. Simply extending the size of this is problematic as then it reduces the space available to the operands A B and C. Furthermore the way Lua bytecodes work means that B and C operands must be 1-bit larger than A - as the extra bit is used to flag whether the operand refers to a constant or a register. (Thanks to Dirk Laurie for pointing this out).

I am amending the bit mapping in the 32-bit instruction to allow 9-bits for the byte-code, 7-bits for operand A, and 8-bits for operands B and C. This means that some of the Lua limits (maximum number of variables in a function, etc.) have to be revised to be lower than the default.

The new instructions are specialised for types, and also for register/versus constant. So for example OP_RAVI_ADDFI means add number and integer. And OP_RAVI_ADDFF means add number and number. The existing Lua opcodes that these are based on define which operands are used.

Example:

local i=0; i=i+1

Above standard Lua code compiles to:

[0] LOADK A=0 Bx=-1
[1] ADD A=0 B=0 C=-2
[2] RETURN A=0 B=1

We add type info using Ravi extensions:

local i:integer=0; i=i+1

Now the code compiles to:

[0] LOADK A=0 Bx=-1
[1] ADDII A=0 B=0 C=-2
[2] RETURN A=0 B=1

Above uses type specialised opcode OP_RAVI_ADDII.