For small types, just swap directly #43351
Conversation
Don't ptr::swap_nonoverlapping for things smaller than its block size. That simpler for LLVM -- particularly in debug -- and means the load/store instructions are generated with full alignment information.
| r? @BurntSushi (rust_highfive has picked a reviewer for you, use r? to override) |
| pub fn swap<T>(x: &mut T, y: &mut T) { | ||
| unsafe { | ||
| ptr::swap_nonoverlapping(x, y, 1); | ||
| if size_of::<T>() < 32 { |
There was a problem hiding this comment.
Isn't 32 a bit magical? Can we get the "correct" size from some platform info?
There was a problem hiding this comment.
32 is the block size used in ptr::swap_nonoverlapping_bytes, so is the point where the simd loop there starts to potentially do something interesting.
That choice is from PR #40454, which has history and a bunch of perf numbers.
There was a problem hiding this comment.
Is it possible to make it so this function and swap_nonoverlapping_bytes are always in sync? e.g., If the latter is updated, can we also guarantee that the former also must be updated?
alexcrichton added the S-waiting-on-review 
| Are there benchmarks for this? |
alexcrichton added S-waiting-on-author 
| ping @scottmcm, thoughts on the benchmarks @BurntSushi mentioned? |
| I'm going to close this to keep the queue clean, but we'd love to take this so if you manage to come back to it @scottmcm just let us know and we'll reopen! |
6 participants
In
mem::swap, don'tptr::swap_nonoverlappingfor things smaller than its block size. That lets debug mode not run the simd loop and means the load/store instructions in release are generated with full alignment information. (Swap on u64 in stable has align 8, but in nightly it has align 1.)Inspired by @oyvindln in #43299 (comment).