[benchmark] Add benchmark for UnsafeMutableRawBufferPointer firstIndex(of:) and lastIndex(of:)

[valeriyvan]

No description provided.

[glessard]

@swift-ci please benchmark

[glessard]

Benchmark looks reasonable. Do we really need to run it with 4 different orders of magnitude?

[glessard]

Benchmarking issue:

✅	Benchmark Check Report
⚠️⏱	RawBuffer.1000.findFirst execution took 3 μs. Increase the workload of RawBuffer.1000.findFirst to be more than 20 μs.

Removing the versions that take fewer than 10k elements should fix this.

[glessard]

The benchmark harness needs each iteration to run for longer than some minimum amount of time. It appears that only the largest of these benchmark invocations meets the bar.

[valeriyvan]

The benchmark harness needs each iteration to run for longer than some minimum amount of time. It appears that only the largest of these benchmark invocations meets the bar.

Search in 100K byte buffer looks to happen quite rare. Contrary, search in buffer of few hundred bytes (or shorter) looks more realistic. E.g. search for nil terminator in string buffers. Are we interested to keep an eye on how future code change might slow down search in short buffer? If yes, then we are interested to keep benchmarks of search in short buffers. Should we just increase multiplicator from 10 to 100 or 1000 and keep all benchmarks?

[glessard]

If you can increase the multiplier, then that would take care of the sanitizer's complaint. I'm not sure we need all those lengths, though. It matters to have a short length (less than 1 stride), a short length that vectorizes (more than 1 stride and less than 2 strides), and a large length (many strides.) Probably 7, 15 and 1000 are fine. We don't want to have an unnecessarily large number of benchmarks because of the load on the resources.

[glessard]

Created new issue to describe this work: #63200

[valeriyvan]

If you can increase the multiplier, then that would take care of the sanitizer's complaint. I'm not sure we need all those lengths, though. It matters to have a short length (less than 1 stride), a short length that vectorizes (more than 1 stride and less than 2 strides), and a large length (many strides.) Probably 7, 15 and 1000 are fine. We don't want to have an unnecessarily large number of benchmarks because of the load on the resources.

Done. Let's see how it works.

[valeriyvan]

I don't know how review request from @lorentey got removed. It was unintentional. Sorry.

[glessard]

@swift-ci please benchmark

[glessard]

Looks like we can make the multipliers much larger:

✅	Benchmark Check Report
⚠️⏱	RawBuffer.7.findLast execution took 3 μs. Increase the workload of RawBuffer.7.findLast to be more than 20 μs. unable to compute memory footprint of RawBuffer.7.findLast
⚠️⏱	RawBuffer.7.findFirst execution took 4 μs. Increase the workload of RawBuffer.7.findFirst to be more than 20 μs. unable to compute memory footprint of RawBuffer.7.findFirst unable to compute memory footprint of RawBuffer.1000.findFirst unable to compute memory footprint of RawBuffer.1000.findLast
⚠️⏱	RawBuffer.15.findLast execution took 5 μs. Increase the workload of RawBuffer.15.findLast to be more than 20 μs. unable to compute memory footprint of RawBuffer.15.findLast
⚠️⏱	RawBuffer.15.findFirst execution took 6 μs. Increase the workload of RawBuffer.15.findFirst to be more than 20 μs. unable to compute memory footprint of RawBuffer.15.findFirst

Since we expect to speed up the search a lot, we can probably shoot for the millisecond range, so that the benchmark still executes for long enough after the vectorized implementations are in place.

[valeriyvan]

Since we expect to speed up the search a lot, we can probably shoot for the millisecond range, so that the benchmark still executes for long enough after the vectorized implementations are in place.

Increased multiplier to 1000_000

[glessard]

@swift-ci please benchmark

[glessard]

@swift-ci please benchmark

[glessard]

@swift-ci please benchmark

[glessard]

Thanks for this benchmark! We can tune the remaining details once we have an implementation.

The benchmark checker still warns that the 1000-length iterations are slightly too long, but if the improvements are reasonable, then the result will be under the warning threshold.

[glessard]

@swift-ci please smoke test and merge