Hacker News Reader

Comment by wiz21c | Hacker News Reader

Long time ago I coded much of my professional stuff in assembly, I was hired specifically to optimize stuff. But it was 20 years ago, compilers were not very smart.

How smart are compilers these days ? Say, to optimize small function, for example computing a scalar product or applying a 3D matrix transformation to a set of points.

View on HN

IME: "it depends". You'll have to check the generated assembly code and then tweak your "high-level" C code to appease the compiler to generate code that's acceptable. Sometimes a small change in the high-level code is enough to break the "pattern matching" in optimizer passes.

Here's an example that looks like magic at first glance where the compiler converts manual bit twiddling code to a popcnt instruction (with the right compiler setting), but do the bit counting any other way, and the whole thing falls apart:

https://www.godbolt.org/z/KaM6jWjdx

Pretty smart. The examples you gave are math-heavy, so to get the best performance you need to do use some kind of SIMD instructions. For these you need to drop down a level, although not really assembly - there are compiler intrinsics that you can use. And for simple functions, compilers are getting fairly good at autovectorization, meaning to introduce SIMD instructions automatically. But it's not something you can rely on.

Generally, they do lots of inlining, and then once you inline you can get some more optimizations in, rinse and repeat. Ends up pretty optimal. (This is C++, can't speak for other languages.)

We work on very perf-sensitive code and we never drop down to assembly. For hot loops, we usually inspect the generated assembly and if it's not great, it's fairly easy to "nudge" the compiler towards the better-performing solutions by tweaking the source code. Also some manual unrolling might be needed to better saturate the vector processing cores of modern CPUs.

And when you're working with signed integers, you still have to do stuff like a >> 1 instead of a / 2 :)

If it was on StackOverflow, I would give it a "correct answer" flag :-)

But if it's now up to some nudging, it's vastly better to me.

I provided math stuff and vectorisable stuff on purpose :-) Happy to see that vectorisation is somewhat automatic. I remember the MMX days and there were not that funny :-)

Someone well-versed in assembly can still do much better than the compiler in many cases over reasonably small function. You could of course do better for large functions too, but at some point the cost becomes prohibitive: if you're going to do this, stick to a few hotspots which are as compact as possible.

GCC and ICC typically do a reasonable job, with the right flags: https://godbolt.org/z/6a8no4n5a.

Reply on news.ycombinator.com