Unrelated to this specific post I’m such a fan of Zed. It’s the first feature complete text editor in recent memory that I’ve truly enjoyed using (i.e. it stays out of the way, is really fast, feels well engineered). I’m coming to Zed after years of Emacs which I still have love for but no longer feels like a competitive piece of software (it does not take full advantage of how good computers are today, e.g. gpu rendering or multicore). I really hope Zed stays a fast and lightweight text editor instead of becoming some bloated growth-at-all-cost VC ware (not that they’ve exhibited any signs of that happening). I’d also happily pay for Zed without a subscription based thing for access to LLM features (which I do not use).
Why does Emacs need that though? I hear people say this all the time and I don't get it. Multicore kind of works against the structure that Emacs touts as a feature. And GPU rendering? In many applications, I totally agree with these complaints. But it's a text editor.
I tried Zed myself, and it's good. But it doesn't dethrone Emacs (for me personally).
Long time emacs user here (+20 years, yikes). I've used it on all kinds of computers during this time. Even a relatively modest computer from 2024 is an absolute beast compared to something from the year 2000.
With that said, there are text editing operations that can cause it to grind to a complete halt, especially working with large files (or very long lines). And it shouldn't, you know? I think emacs users sort of internalize which operations they should avoid. It's kind of ridiculous to have to do that in a text editor with the massive amounts of compute that our computers have today.
I have consistent issues with emacs locking up when executing network requests. I'm sure there's a specific bug that could be hunted down and addressed, but this sort of thing shouldn't happen much in an editor that's multicore by default.
I'm not trying to dismiss emacs' reasoning, of course, but I can understand being disgruntled with it.
The actual rendering I've been quite please by, though!
Most new TTY projects use GPUs for that reason
The are serious about local vs. remote vs. shared. They are serious about hardware acceleration because they care about users who type fast and edit big files. They care about real computer science on how to push the current hardware to serve the user, rather than treating the hardware as a crutch to give the user a slightly worse experience at a fraction of the development cost. They care about code highlighting the snippets on their blog like very similar to the default Zed theme.
These are cool, serious people. If they give me emacs key bindings with full paredit-everywhere, I might switch my daily driver.
And this is about using modern SIMD-style stuff, branch less stuff, Lemire stuff in a modern and relevant context. Other commenters have pointed out that you can do the mask or popcount better with intrinsically, and yeah, but they probably know that, and B. They got the massive win, the remaining factor of K on the pop count is coming.
Long Zed.
You don't need information about the position of newlines in all the chunks located before the one your offset lands on
> // Parallel bit count intermediates
> let a = v - ((v >> 1) & (u64::MAX / 3));
> let b = (a & (u64::MAX / 5)) + ((a >> 2) & (u64::MAX / 5));
> let c = (b + (b >> 4)) & (u64::MAX / 0x11);
> let d = (c + (c >> 8)) & (u64::MAX / 0x101);
On processors with BMI2 the whole algorithm reduces to a PDEP as mentioned in another comment, but if you don't have that this is pretty much the best you can do (unless you use lookup tables but those have pros and cons).
The above code is completely source- and binary-portable and reasonably fast -- certainly faster than naively looping through the bits, and within a small constant factor of a CPU POPCOUNT instruction.
And the compiler is not required to lower it to a single instruction. It will if the target architecture is specified appropriately, but there's nothing that says it has to explode if it can't. In fact, by doing it this way, the compiler is actually more free to generate code in a way that's optimal for the architecture in all cases, because all the implementation details are hidden e.g. loads of large constants may be avoided if the compiler is allowed to choose the exact implementation, while using the portable version may tie its hands more depending on how it's feeling on that day. Here's __builtin_popcount working just fine while targeting a ~20yo architecture without native support for SSE4.2; it can generate this code knowing what the proper instructions and schedules are: https://godbolt.org/z/ra7n5T5f3
The moral here is that the primitives are there for you to use. Just use them and save yourself and your would-be code reviewer's time.
Also, pretty much every compiler for a very long time has supported __builtin_popcount or equivalent.
Clang supports __builtin_popcount() too. And MSVC has __popcnt().
However, as someone already mentioned, this is not equivalent to a pair of popcounts.
Bitstring rank/select is a well-known problem, and the BMI and non-BMI (Hacker's Delight) versions are available as a reference.
I've been saying this a lot recently, that CPU's are powerful 1-bit vector machines!
I know that they have some recordings from their coding session, but I am looking for something more overall.
> filter: blur(1px);
Wonderful styling!
aa -> -> bb -> -> bb
It only takes up two spaces, after all