Skip to content(if available)orjump to list(if available)

Revisiting Loop Recognition in C++ in Rust

pjmlp

I was expecting that the C++ code would have been updated to C++23 best practices, and what are all those std::list doing in modern CPU cache lines?

npalli

Not only is it not C++23, the C++ code is from the 2011 paper!, so pre-C++11. Rust maybe better than code written before C++11 is not a strong take.

pjmlp

Yes it has a pretty much C++98/C++ARM feeling to the code style.

quietbritishjim

In fairness, updating to a range-based for loop would give the same effective code as using iterators manually in the loop as they have done. I'm not convinced any new features in C++23 (after C++11) would have given a performance improvement - was there something you had in mind?

The choice of std::list is a bit more dubious. Looking at the code (I found a mirror [1]), it seems like it's mostly for appending and iterating but there is pop_front in one place. Maybe std::deque would be better? Or std::vector with an iterator noting the effective start? Or maybe they were deliberately choosing a poor data structure to stress the compiler (and if so the OP should've used a linked list in Rust for fairness). The article doesn't comment so we can only guess their motivation.

[1] https://github.com/hundt98847/multi-language-bench/blob/mast...

pjmlp

> was there something you had in mind?

More like on how to write the code in more ergonomic way than the performance by itself, maybe some of that stuff could even be constepxr/eval.

npalli

std::list to std::vector should be the big one.

std::map to std::unordered_map could be next.

then, really ranges/constexpr/std::move could make a difference, hard to say definitely.

Beyond these, Modern C++ would have most definitely led to much shorter code as that was a metric for comparison.

quietbritishjim

> std::list to std::vector should be the big one

That is not a "C++23 best practice", which is what I was replying to. It doesn't even need C++11! And one use of this type uses pop_front() so std::vector is not obviously a good choice here.

> std::map to std::unordered_map could be next.

Again, I called out C++11 as possibly making a difference - sure this could help but it doesn't need C++23.

> then, really ranges/constexpr/std::move could make a difference, hard to say definitely.

How? Ranges are a nice syntax but what would they speed up here? There doesn't seem to be anything evaluated at compile time so what's the benefit of constexpr? (Even std::move doesn't have an obvious use in the benchmark code but that's C++11 anyway.)

> Beyond these, Modern C++ would have most definitely led to much shorter code as that was a metric for comparison.

I agree that would be interesting, but I'd be surprised if the code was much shorter. I'd guess something like 10%.

kvemkon

> Raspberry Pi OS Bookworm

> g++ (Debian 12.2.0-14) 12.2.0

> cargo 1.87.0 (99624be96 2025-05-06)

Looks like cargo is not from OS repo. So either use both outdated gcc and rust from RaspiOS or install the recent gcc-15 from experimental.

SkiFire13

Moreover they also differ in the optimizer being used. A more fair comparison would use Clang for C++ or rustc_codegen_gcc for Rust, possibly with the same versions of LLVM and/or GCC/libgccjit.

Joker_vD

The whole blogpost reads surprisingly angry, for some reason. Is it just my impression?

> And Google can go to hell.

Oh, apparently the tone is intended. Why though?..

krona

By my reading, the C++ FindSet() implementation of the union/find algorithm builds a temporary list! Incredible scenes.