A notable amount of code I’ve written in my professional life has made its way onto GitHub. Here’s a collection of some of my commits that I think are neat. (Disclaimer: I wrote these while employed at Meta, where I am still employed at the time of writing. However, this is my personal space and my employer does not necessarily endorse the things I stay in my personal spaces. In particular, I am merely pointing to and summarizing public information below.)
Back in 2021, I wrote about unions in my professional capacity (disclaimer: this is my personal space and my employer does not necessarily endorse what I say here) on the PyTorch Dev Discussions forum. Stick with (or skip) the C++ 101 “What is a union?” section and you’ll be rewarded with the advanced material in the second half. Please enjoy!
Copy elision is a C++ compiler optimization that, as its name suggests, eliminates extra copy and move operations. It is similar to the classical copy propagation optimization, but specifically performed on C++ objects that may have non-trivial copy and move constructors. In this post, I’ll walk through an example where an obvious optimization you might expect from your compiler doesn’t actually happen in practice.
ARM64 is a computer architecture that competes with the popular Intel x86-64 architecture used for the CPUs in desktops, laptops, and so on. ARM64 is common in mobile phones1, as well as Graviton-based Amazon EC2 instances, the Raspberry Pi 3 and 4, and the much ballyhooed Apple M1 chips, so knowing about it might be useful! In fact, I have almost certainly spent more time with ARM64 than x86-64 because of the iPhone.
Now that we know how to read assembly language, we can talk about how parameter passing works at the machine level and its consequences for writing faster code. We will focus on x86_64, but ARM64 works in a roughly similar way.
UPDATE: This article now has an ARM64 port.
Why, in 2021, does anyone need to learn about assembly language? First, reading assembly language is the way to know exactly what your program is doing. Why, exactly, is that C++ program 1 MiB (say) instead of 100 KiB? Is it possible to squeeze some more performance out of that function that gets called all the time?
Why is inlining so important in C++? Clearly, it reduces function call overhead: if a function is inlined, there is no need to spend time setting up its arguments, jumping to it, creating a stack frame, and then undoing all that upon returning. More interestingly, though, inlining enables other compiler optimizations. In this article, I will show examples of constant propagation and loop-invariant code motion (LICM). Then, I will explain how inlining enables these optimizations to apply more widely and show an example of the consequences when that doesn’t happen.
std::function does what it says on the tin: it holds any callable object. For example, you can use lambdas with or without captures. It has a couple key drawbacks:
The C++ standard library has many classes and functions that are easy to use and relatively safe. However, in situations where performance and efficiency really matter, straightforward use of the standard library may not always be the best choice. This post is the first in a series that will catalogue some of those opportunities for improvement. It is meant to be accessible to people who are not C++ gurus or language lawyers.
Putting writing on a website in 2021 is surprisingly pleasant, despite a few glitches.
I wanted to have a plain old static site. I can write raw HTML well
enough, and I knew about Bootstrap,
but that still seemed like too much work. Instead, I looked at a
couple
overviews of static site generators, and
I ended up picking Hugo as the one most fitting
my sensibilities. You write pages and posts in Markdown, run Hugo, and
get static HTML/CSS/JS out. It has a built-in web server with hot
reloading so that you can run hugo server, edit your site, and see
it in your browser right away. The builds are really fast; I am using
a potato cheap Chromebook from
2018
and the site still builds in less than 100 ms.