Optimize: core slice binary_search_by

[RoDmitry]

I have found a more optimal algorithm for the binary search. I have ran benchmarks and it works faster than the previous one. Also I have looked at the compiled assembly, with and without an early return, and it looks good either way. And benchmarks show that it's faster with an early return. Looks strange for me that "CPU can reliably predict the loop count" based on the mathematical operation as it was before (substraction). But I don't know much about branch prediction, so feel free to improve it if possible.

Also I would like to know, is there any attitude towards optimizations using SIMD? Can I write platform specific SIMD intrinsics, or what should I use if I would like to contribute SIMD code optimizations? Is there already any SIMD code optimizations in core/std, I could look at as an example?

P.S. It's my first contribution here.

[rustbot]

r? @thomcc

rustbot has assigned @thomcc.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[the8472]

There have been many changes and attempts to binary_search.

• Rewrite binary search implementation #128254
• Improve slice::binary_search_by #127007
• Tell LLVM that partition_point returns a valid fencepost #102535
• replace if with match in binary_search #106969
• Instruct LLVM that binary_search returns a valid index #81354
• Improve slice.binary_search_by()'s best-case performance to O(1) #74024
• Improve SliceExt::binary_search performance #45333
• and probably some I have missed

Some versions of this look quite like yours, so you'll have to justify this change by a more detailed analysis and benchmark results. std already has some that you can try to run.
Which implementation performs best also depends on the exact types, the comparison function and the distribution of the values being used for benchmarking.

Also I would like to know, is there any attitude towards optimizations using SIMD? Can I write platform specific SIMD intrinsics, or what should I use if I would like to contribute SIMD code optimizations? Is there already any SIMD code optimizations in core/std, I could look at as an example?

That's possible, but only worthwhile in a few cases where good results can be achieved with the baseline SIMD instructions available on tier1 platforms, e.g. SSE2 on x86-64-linux, since runtime detection isn't available in core.
So we only have a few of those and mostly rely on autovectorization. But you can look at #103779 for an example.

[hanna-kruppe]

I have ran benchmarks and it works faster than the previous one.

Which benchmarks? Do they happen to look up the same value in the same slice repeatedly? Or otherwise have predictable patterns in the outcome of the match cmp that this change introduces? I wouldn’t be surprised if this change will look like a performance improvement for such benchmarks, but that’s not representative for many real world uses of binary search.

[RoDmitry]

I agree that you need a more detailed analysis and benchmark results. Have found benchmarks in library/coretests/benches/slice.rs. Will try them later.

[rust-log-analyzer]

The job mingw-check failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)