Have Tracing JIT Compilers Won?

I'd say that for languages such as Python or Javascript tracing JIT compilers are the only viable approach. Because the languages allow anything you can't make any sensible predictions about types and call graphs at compile time (people tried; didn't work well). The incremental-JIT approach (LuaJIT) has less information to work with, and it's not aware which parts of the program are bottlenecks, so it can't intelligently decide which parts parts need optimization most.

The tracing JIT is such an obvious tactic that it should be part of any dynamic language. Measure which parts are slow, then optimize those parts using the parameter type information gathered at runtime. Fall back on original code on type mismatch. The implementation is hairy, but it's a no-brainer conceptually speaking.

I suspect that dynamic languages will evolve over the next 10 years to allow for really efficient compilers, but for the current languages such as Javascript I doubt we're ever going to see massive performance improvements.

Ahead of time compilers do work quite well for a subset of dynamic languages. Just like tracing JIT works well for a subset. This has been proven by projects like shedskin, cython, and tinypyC++. Given typing information, either explicitly or through type inference - massive speed-ups are possible.

Also, tracing JIT currently does not do multiple CPU optimizations - which static compilers have been doing. They also don't do SIMD or MIMD optimizations, which compilers like gcc/icc/etc are doing.

However, runtime assemblers - like 'Orc' 'corepy' and older ones like 'softwire' are proving to be lots faster in multimedia applications compared to traditional compilers or interpreters. These runtime assemblers give authors of programs access to the compilers/assemblers. Just like allowing people to specify typing, letting people construct programs at runtime can give massive speed-ups.

C has even had fast interpreters for a while - see tinycc as an example. It doesn't use a tracing compiler, but a very straight forward direct translation compiler which happens to be faster than many interpreters.

It's quite easy to argue 'No' to this question I think.

> C has even had fast interpreters for a while - see tinycc as an example. It doesn't use a tracing compiler, but a very straight forward direct translation compiler which happens to be faster than many interpreters.

How is tinycc an interpreter? It's a full-blown compiler, alas a very simple one with very little optimization. If it weren't faster than an interpreter, I'd be shocked. For well written code, it reaches 50%-70% of modern gcc speed from tests I did a couple of years ago. I think that's on par, or even slightly better, than gcc 1.x

It helps that C is designed to have an almost one to one mapping between its native operators and machine operations and addressing modes.

For example, the pre and post increment operators are expressions with side effects, can be confusing and can provoke undefined behaviour with ease. Having two variants also seems redundant. But the two variants may have different optimal translations to underlying CPU addressing modes, while leaving the pre or post modified value in a register is often a side-effect of the CPU translation. If the language only permitted ++ as a statement, the usefulness of side-effect would be lost.

It's a cliche that C is little more than portable assembler, but at a design level there's a lot of truth to it. The length of the assembly translation of C code is usually pretty directly proportional to the source code token length. In more expressive languages with more powerful abstractions, that's frequently not the case.

With modern IDEs, tracing JIT, and type inference, I'm surprised that someone hasn't made the same lateral thinking sidestep that Hennessey & Patterson made for RISC. They offloaded much of the decoding work from the chip to the compiler, yielding faster chips in a smaller die size.

Perhaps the Go language is the start of such a sidestep. We should be able to offload a lot of the work of type annotation to the IDE and compiler, yielding a compiled language with the dynamic feel of an interpreted one. In fact, for langs with clean runtime models like Lisp, we should be able to allow intermediate states of partial annotation, and allow some interpreted execution for fast prototyping. (But demand full annotation before deployment.)

(EDIT: Yes, I was aware of Haskell when I posted this. Haskell. There I said it!)

Is there a tool that uses a Bayesian technique for suggesting type annotation that cannot be inferred? This would be tremendously useful. (And very dangerous in the hands of the incompetent.)