A mythical beast called
I have recently bumped into one of those articles which title sounds rather like “one weird trick” ad. Guiding just from it, you can imagine how small amount of information is contained there, and by actually going through that link there you will definitely make sure that it has some knowledge, but no real explanations are in place.
Of course, a final explanation is indeed really simple -
Function.prototype.toString() returns everything even the comments inside the function body, size of that string is a feature that allows the optimizing compiler inside V8 to make a decision to inline that particular function or not. There is even a command line switch which to use to modify a limit and a default value. Simple enough? Not really.
That article left a lot of unanswered questions. Probably because of that JIT compilers and its optimization techniques are kind of black magic.
Why is it hard?
Returning to the main topic, one of many tricks that they are using, which is used also in aforementioned case, is called compilation cache.
V8, compilation cache and inlining
It is a really simple trick (do not confuse it with inline caches - it is a different technique!), where compiler textually compares the implementation of the function and if it is the same as the other one, merges it together and uses since then one representation. Why? Because if it is a hard process, why compiler should do it many times for the same code sample? If we can avoid it, we can speed-up the execution even more.
Second thing which is using compilation cache is inlining, where instead of function call (which can be expensive) we are putting directly small piece of code. The more similarities we detect, the better cache utilization will be. And of course we can free program related memory (or rather reuse it inside V8 to other things).
But there are a culprit related to those techniques - it is related with the textual representation of function body. As in the original example, if the code differs by even a small piece that is inside a comment, it will be treated as a different thing - and in consequence compilation cache and inlining cannot be used properly. As a side note - regarding those two tricks compilers often are not working directly on the source code level, but at first they are transforming it to an AST or other more friendlier data structure for them. But not in case of V8 - probably because of additional checks and computation that needs to be done, so authors relied on simple string comparison.
There is one additional trick that we can use, that is really well described (and related to both optimization techniques) by one of my favorite presenters and compiler evangelists Vyacheslav Egorov - he is using UUIDs hard-coded inside a comment in the function body to distinguish logically different functions, that could be potentially compiled and inlined from a single cache entry.
Why he would like to avoid that? In order to minimize the cache rebuilding and cache misses after recompilation and further passes of JIT compilation, when we’re generating code or do other fancy stuff (if you do that fancy stuff, you are probably aware of how that things work in the first place).
Yeah, JIT and compilation techniques are hard.