AngelScript
 
Loading...
Searching...
No Matches
How to build a JIT compiler

AngelScript doesn't provide a built-in JIT compiler, instead it permits an external JIT compiler to be implemented through a public interface.

To use JIT compilation, the scripts must be compiled with a few extra instructions that provide hints to the JIT compiler and also entry points so that the VM will know when to pass control to the JIT compiled function. By default this is turned off, and must thus be turned on by setting the engine property asEP_INCLUDE_JIT_INSTRUCTIONS.

If the application sets the JIT compiler with SetJITCompiler AngelScript will automatically invoke it to provide the JIT functions with each compilation or loading of pre-compiled bytecode.

The structure of the JIT function

The JIT compiled function must follow certain rules in order to behave well with the virtual machine. The intention is that the VM will pass the control to the JIT function, and when the execution is to be suspended the JIT function returns the control to the VM, updating the internal state of the VM so that the VM can resume the execution when requested. Each time the JIT function returns control to the VM it must make sure that the VM registers and stack values have been updated according to the code that was executed.

The byte code will have a special instruction, JitEntry, which defines the positions where the VM can pass the control to the JIT function. These are usually placed for every script statement, and after each instruction that calls another function. This implies that the JIT compiled function needs to be able to start the execution at different points based on the argument in the JitEntry instruction. The value of the argument is defined by the JIT compiler and how it is interpreted is also up to the JIT compiler, with the exception of 0 that means that the control should not be passed to the JIT function.

Some byte code instructions are not meant to be converted into native code. These are usually the ones that have a more global effect on the VM, e.g. the instructions that setup a call to a new script function, or that return from a previous instruction. When these functions are encountered, the JIT function should return the control to the VM, and then the VM will execute the instruction.

Other byte code instructions may be partially implemented by the JIT function, for example those that can throw an exception based on specific conditions. One such example is the instructions for divisions, if the divider is 0 the VM will set an exception and abort the execution. For these instructions the JIT compiler should preferrably implement the condition that doesn't throw an exception, and if an exception is to be thrown the JIT function will instead break out to the VM.

The following shows a possible structure of a JIT compiled function:

  void jitCompiledFunc(asSVMRegisters *regs, asPWORD jitArg)
  {
    Read desired VM registers into CPU registers.
    Jump to the current position of the function based on the 'jitArg' argument.
  1:
    Execute code in block 1.
    Jump to exit if an illegal operation is done, e.g. divide by zero. 
    Jump to exit if block ends with an instruction that should not be executed by JIT function. 
  2:
    ...
  3:
    ...
  exit:
    Update the VM registers before returning control to VM.
    If necessary the function can invoke the methods of the context informed 
    in the regs, e.g. to suspend the execution, or to set a script exception.
  }

Traversing the byte code

int CJITCompiler::CompileFunction(asIScriptFunction *func, asJITFunction *output)
{
bool success = StartNewCompilation();
// Get the script byte code
asUINT length;
asDWORD *byteCode = func->GetByteCode(&length);
asDWORD *end = byteCode + length;
while( byteCode < end )
{
// Determine the instruction
asEBCInstr op = asEBCInstr(*(asBYTE*)byteCode);
switch( op )
{
// Translate each byte code instruction into native code.
// The codes that cannot be translated should return the control
// to the VM, so that it can continue the processing. When
// the VM encounters the next JitEntry instruction it will
// transfer the control back to the JIT function.
...
// Update the argument for the JitEntry instruction with
// the argument that should be sent to the jit function.
// Remember that 0 means that the VM should not pass
// control to the JIT function.
asBC_PTRARG(byteCode) = DetermineJitEntryArg();
break;
}
// Move to next instruction
byteCode += asBCTypeSize[asBCInfo[op].type];
}
if( success )
{
*output = GetCompiledFunction();
return 0;
}
return -1;
}
unsigned char asBYTE
8 bit unsigned integer
Definition: angelscript.h:608
unsigned long asDWORD
32 bit unsigned integer
Definition: angelscript.h:624
const int asBCTypeSize[21]
Lookup table for determining the size of each type of bytecode instruction.
Definition: angelscript.h:4995
#define asBC_PTRARG(x)
Macro to access the first pointer argument in the bytecode instruction.
Definition: angelscript.h:5336
asEBCInstr
The bytecode instructions used by the VM.
Definition: angelscript.h:4531
@ asBC_JitEntry
If a JIT function is available and the argument is not 0 then call the JIT function.
Definition: angelscript.h:4883
const asSBCInfo asBCInfo[256]
Information on each bytecode instruction.
Definition: angelscript.h:5064
unsigned int asUINT
32 bit unsigned integer
Definition: angelscript.h:610
void(* asJITFunction)(asSVMRegisters *registers, asPWORD jitArg)
The function signature of a JIT compiled function.
Definition: angelscript.h:4496
The interface for a script function description.
Definition: angelscript.h:4000
virtual asDWORD * GetByteCode(asUINT *length=0)=0
Returns the byte code buffer and length.
asEBCType type
Instruction argument layout.
Definition: angelscript.h:5035

The following macros should be used to read the arguments from the bytecode instruction. The layout of the arguments is determined from the asBCInfo array.

What each byte code instruction does is described in Byte code instructions, but the exact implementation of each byte code instruction is best determined from the implementation in the VM, i.e. the asCScriptContext::ExecuteNext method.