-
Notifications
You must be signed in to change notification settings - Fork 13k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Optimize insertion sort #40807
Optimize insertion sort #40807
Conversation
This change slightly changes the main iteration loop so that LLVM can optimize it more efficiently. Benchmark: name before ns/iter after ns/iter diff ns/iter diff % slice::sort_unstable_small_ascending 39 (2051 MB/s) 38 (2105 MB/s) -1 -2.56% slice::sort_unstable_small_big_random 579 (2210 MB/s) 575 (2226 MB/s) -4 -0.69% slice::sort_unstable_small_descending 80 (1000 MB/s) 70 (1142 MB/s) -10 -12.50% slice::sort_unstable_small_random 396 (202 MB/s) 386 -10 -2.53%
r? @aturon (rust_highfive has picked a reviewer for you, use r? to override) |
@bors: r+ |
📌 Commit 2c816f7 has been approved by |
…=alexcrichton Optimize insertion sort This change slightly changes the main iteration loop so that LLVM can optimize it more efficiently. Benchmark: ``` name before ns/iter after ns/iter diff ns/iter diff % slice::sort_unstable_small_ascending 39 (2051 MB/s) 38 (2105 MB/s) -1 -2.56% slice::sort_unstable_small_big_random 579 (2210 MB/s) 575 (2226 MB/s) -4 -0.69% slice::sort_unstable_small_descending 80 (1000 MB/s) 70 (1142 MB/s) -10 -12.50% slice::sort_unstable_small_random 396 (202 MB/s) 386 -10 -2.53% ``` The benchmark is not a fluke. I can see that performance on `small_descending` is consistently better after this change. I'm not 100% sure why this makes things faster, but my guess would be that `v.len()+1` to the compiler looks like it could in theory overflow.
Since this is internal to libstd/core, could you check whether the inclusive range syntax makes things better as well? i.e. |
@nagisa Inclusive range syntax makes performance slightly worse, actually... With the old insertion sort and with inclusive range syntax there's a bound check at the beginning of insertion sort. This PR removes the bound check. If you want to play with this, here's a playpen link. |
⌛ Testing commit 2c816f7 with merge 04e47d7... |
…=alexcrichton Optimize insertion sort This change slightly changes the main iteration loop so that LLVM can optimize it more efficiently. Benchmark: ``` name before ns/iter after ns/iter diff ns/iter diff % slice::sort_unstable_small_ascending 39 (2051 MB/s) 38 (2105 MB/s) -1 -2.56% slice::sort_unstable_small_big_random 579 (2210 MB/s) 575 (2226 MB/s) -4 -0.69% slice::sort_unstable_small_descending 80 (1000 MB/s) 70 (1142 MB/s) -10 -12.50% slice::sort_unstable_small_random 396 (202 MB/s) 386 -10 -2.53% ``` The benchmark is not a fluke. I can see that performance on `small_descending` is consistently better after this change. I'm not 100% sure why this makes things faster, but my guess would be that `v.len()+1` to the compiler looks like it could in theory overflow.
…=alexcrichton Optimize insertion sort This change slightly changes the main iteration loop so that LLVM can optimize it more efficiently. Benchmark: ``` name before ns/iter after ns/iter diff ns/iter diff % slice::sort_unstable_small_ascending 39 (2051 MB/s) 38 (2105 MB/s) -1 -2.56% slice::sort_unstable_small_big_random 579 (2210 MB/s) 575 (2226 MB/s) -4 -0.69% slice::sort_unstable_small_descending 80 (1000 MB/s) 70 (1142 MB/s) -10 -12.50% slice::sort_unstable_small_random 396 (202 MB/s) 386 -10 -2.53% ``` The benchmark is not a fluke. I can see that performance on `small_descending` is consistently better after this change. I'm not 100% sure why this makes things faster, but my guess would be that `v.len()+1` to the compiler looks like it could in theory overflow.
This change slightly changes the main iteration loop so that LLVM can optimize it more efficiently.
Benchmark:
The benchmark is not a fluke. I can see that performance on
small_descending
is consistently better after this change. I'm not 100% sure why this makes things faster, but my guess would be thatv.len()+1
to the compiler looks like it could in theory overflow.