Why would it be hard to generate the code for a direct-threaded interpreter?

haberman · on Oct 8, 2012

That's exactly the question that Mike's whole post is answering. It turns out that the common optimizations like hoisting, CSE, tail-merging, etc. perform badly on direct-threaded interpreter code. Which is one of the primary reasons Mike wrote the interpreter for LuaJIT2 in hand-written assembly instead of C.

tptacek · on Oct 8, 2012

I think it is easier for the compiler for a high-level language to predict usage patterns --- which branches will be taken, which variables should be in regs or L1, which functions are best served by slower code with a small memory footprint vs. faster code with a larger footprint, what's best handled with a jump table versus with conditional jumps, &c &c --- that can a C compiler, and that over the long run languages with better, more consistent abstractions than C are going to have better runtime characteristics than the outputs of a C compiler.

Put differently: C compilers are hamstrung by all the dumb C programs a developer could express with C.

Javascript compilers have to deal with bad Javascript, don't get me wrong, but they don't have to deal so much with bad list and table implementations.

I get what the LuaJIT2 post is saying (and let's be clear: I'm not putting myself on a level with the Lua JIT people).

For what it's worth: C is my first language, and my favorite language. My background is systems programming. I would not enjoy figuring out how to reliably write a memory-mapped CSR while masking the right set of interrupts in Javascript. :)

haberman · on Oct 8, 2012

> I think it is easier for the compiler for a high-level language to predict usage patterns

Profile-guided optimizations for C can give a C compiler a lot of the same information. It's true that these patterns could vary over the run of the program, though at that point you have to weigh the speed gained from being adaptive vs. the runtime overhead of monitoring the program's execution and doing multiple passes of code generation.

> For what it's worth: C is my first language, and my favorite language.

Yeah I know. :) A lot of what I wrote was directed more generally at the "faster than C" sentiment and not at you specifically.

> I'm not putting myself on a level with the Lua JIT people

I think you mean the LuaJIT person. I got to finally meet him this summer and verify that he is human, though I guess I can't guarantee that he doesn't have cybernetic implants or work together with others under a single pseudonym. :)