Dynamic recompilation

From Wikipedia, the free encyclopedia

In computer science, dynamic recompilation (sometimes abbreviated to dynarec) is a feature of some emulators and virtual machines, where the system may recompile some part of a program during execution. By compiling during execution, the system can tailor the generated code to reflect the program's run-time environment, and perhaps produce more efficient code by exploiting information that is not available to a traditional static compiler.

In other cases, a system may employ dynamic recompilation as part of an adaptive optimization strategy to execute a portable program representation such as Java or .NET Common Language Runtime bytecodes. Full-speed debuggers could also utilize it to reduce the space overhead incurred in most deoptimization techniques, and many other features such as dynamic thread migration.

[edit] Example

Suppose a program is being run in an emulator and needs to copy a null-terminated string. The program is compiled originally for a very simple processor. This processor can only copy a byte at a time, and must do so by first reading it from the source string into a register, then writing it from that register into the destination string. The original program might look something like this:

beginning:
    mov A,[first string pointer]    ; Put location of first character of source string
                                    ; in register A
    mov B,[second string pointer]   ; Put location of first character of destination string
                                    ; in register B
loop:
    mov C,[A]            ; Copy byte at address in register A to register C
    mov [B],C            ; Copy byte in register C to the address in register B
    cmp C,#0             ; Compare the data we just copied to 0 (string end marker)
    inc A                ; Increment the address in register A to point to
                         ; the next byte
    inc B                ; Increment the address in register B to point to
                         ; the next byte
    jnz loop             ; If it wasn't 0 then we have more to copy, so go back
                         ; and copy the next byte
end:                     ; If we didn't loop then we must have finished,
                         ; so carry on with something else.

The emulator might be running on a processor which is similar, but extremely good at copying strings, and the emulator knows it can take advantage of this. It might recognize the string copy sequence of instructions and decide to rewrite them more efficiently just before execution, to speed up the emulation.

Say there is an instruction on our new processor called movs, specifically designed to copy strings efficiently. Our theoretical movs instruction copies 16 bytes at a time, without having to load them into register C in between, but will stop if it copies a 0 byte (which marks the end of a string) and set the zero flag. It also knows that the addresses of the strings will be in registers A and B, so it increments A and B by 16 every time it executes, ready for the next copy.

Our new recompiled code might look something like this:

beginning:
    mov A,[first string pointer]    ; Put location of first character of source string
                                    ; in register A
    mov B,[second string pointer]   ; Put location of first character of destination string
                                    ; in register B
loop:
    movs [B],[A]            ; Copy 16 bytes at address in register A to address
                            ; in register B, then increment A and B by 16
    jnz loop                ; If the zero flag isn't set then we haven't reached
                            ; the end of the string, so go back and copy some more.
end:                        ; If we didn't loop then we must have finished,
                            ; so carry on with something else.

There is an immediate speed benefit simply because the processor doesn't have to load so many instructions to do the same task, but also because the movs instruction is likely to be optimized by the processor designer to be more efficient than the sequence used in the first example (for example it may make better use of parallel execution in the processor to increment A and B while it is still copying bytes).

In an ironic twist in real world usage, the first sequence of instructions (RISC) is generally preferred over the next (CISC). The reasons provided are slow CISC processor execution, prevention of pipeline stalls, and lower hardware overheads.

[edit] Products using dynamic recompilation

Many Java virtual machines feature dynamic recompilation.
Apple's new Rosetta for Mac OS X on x86, allows PowerPC code to be run on the x86 architecture.
The emulator used in Mac OS to run 680x0 code on the PowerPC hardware.
Psyco, a specializing compiler for Python.
The HP Dynamo project, an example of a transparent binary dynamic optimizer.
The VX32 virtual machine employs dynamic recompilation to create OS-independent x86-architecture sandboxes for safe application plug-ins.
Virtual PC for Mac, used to run x86 code on PowerPC.
QEMU, an open-source full system emulator.
The backwards compatibility functionality of the Xbox 360 (i.e. running games written for the original Xbox) is widely assumed to use dynamic recompilation.
DGD, an LPC server for muds and beyond, supports dynamic recompilation of objects, thereby allowing fully persistent systems to evolve over time.