push! will always have some performance burden if the reserved memory is not large enough. This is to be expected. Here, I am focusing on the case where the new values will fit into the already reserved memory. With Julia 1.10 and earlier a ccall is necessary, which takes a bit of time, so push! ha…

[image] sgaure: Looking at the native code, the bounds check in length_separate is hoisted out of the loop. The compiler can prove that it will not surprisingly throw, it’s sufficient to check once prior to the loop. It does not assume that a.ref can change concurrently. My point is that it s…

[image] sgaure: Looking at the native code, the bounds check in length_separate is hoisted out of the loop. The compiler can prove that it will not surprisingly throw, it’s sufficient to check once prior to the loop. It does not assume that a.ref can change concurrently. Talking to myself her…

[image] sgaure: Actually, the length_separate is split into two loops. It checks the bounds before the loop. If it’s ok, it runs a vectorized loop. But if it determines that there will be a bound fail, it runs a non-vectorized loop with bounds checking instead, to fill the vector up until it thr…

[image] Sukera: That’s just the regular non-vectorized path that’s kept around for any cleanup of regions that don’t fit the vector width. No, I don’t think so. There’s a cleanup loop right after the vector loop. It ends with a jump to the return code. In between there’s a serial loop which c…

I think we’re talking past each other - can you link to the lines from the gist I posted above you’re referring to?

[image] PatrickHaecker: However, we learn that we should mostly avoid simple mutable structs in performance critical code, so although having an immutable growing vector feels unnatural it’s worth a try I’ve occasionally been working on certain RNG operations . I recall having a similar experi…

[image] Sukera: I think we’re talking past each other - can you link to the lines from the gist I posted above you’re referring to? On line 22 it does a conditional jump to main.pseudo.exit, bypassing the vector loop. It continues to L6.postloop on line 120, then to L12.postloop on line 133 …

[image] sgaure: On line 22 it does a conditional jump to main.pseudo.exit, bypassing the vector loop. I might be reading something wrong here, but isn’t that branch only taken when the array is empty? %isnotneg.inv is true when %.size.sroa.0.0.copyload is less than 0, and that can’t happen. S…

[image] Sukera: I might be reading something wrong here, but isn’t that branch only taken when the array is empty? %isnotneg.inv is true when %.size.sroa.0.0.copyload is less than 0, and that can’t happen. So the branch in line 21 should always return whatever %.not30 is, and that’s only true wh…

[image] sgaure: So, why doesn’t it vectorize with the mutable struct? The mutable struct loop vectorizes if the error branch is removed, so it can’t be the mutable struct as such. Because the possibility of throwing an error is an observable sideeffect of the computation. It’s pretty difficul…

Why is push! still so slow?

Internals & Design

mikmoore March 27, 2024, 3:45pm 17

I’ve occasionally been working on certain RNG operations. I recall having a similar experience with mutable RNG states, which had me tinkering with immutable ones instead.

Topic		Replies	Views
Improving my mental model: Why is `sizehint!()` so much slower than `zeros()`? Performance	9	1646	March 31, 2022
Thread-safe array building General Usage multithreading	21	7820	October 24, 2017
Push! and insert! very slow? (julia 1.11) (EDIT: No; just one-time allocations.) Performance	21	529	November 12, 2024
Documenting performance model of `empty!`, `sizehint!`, `push!` & friends Internals & Design proposal	14	3787	May 2, 2019
How to create a vector of variable length in julia? General Usage	7	11832	January 25, 2019

Why is push! still so slow?

Related topics