@AlexGuteniev, do I need to update benchmarks, or simply put benchmark numbers as part of the commit message?

@AlexGuteniev, do I need to update benchmarks, or simply put benchmark numbers as part of the commit message?

I think the current benchmarks are fine, just post before/after number

Ideally, we'd need ARM64 numbers as well. But I think we can do without them.

I think the current benchmarks are fine, just post before/after number

You also need to pass somehow /arch:AVX or greater to make it different, perhaps by temporarily alterint CmakeFIle.txt locally, or by any other working method.

You also need to pass somehow /arch:AVX or greater to make it different, perhaps by temporarily alterint CmakeFIle.txt locally, or by any other working method.

shouldn't unit tests automatically run with/without these switches?

shouldn't unit tests automatically run with/without these switches?

Unit tests should.
I mean benchmarks, which currently don't

Also please edit the PR description to link it with the issue using one of Github keywords.

I'm getting conflicting results for the test: popcount on my pc is somehow slower by around 30%
it's a tiger-lake-u core i7-11390H, shouldn't have slow popcount :)

I'm getting conflicting results for the test: popcount on my pc is somehow slower by around 30% it's a tiger-lake-u core i7-11390H, shouldn't have slow popcount :)

no idea how is that it doesn't get inlined :)

wow, accidentially, i had /Only __inline (/Ob1) option. Apparently, this destroys most of std code that doesn't have inlines. Imo, all these functions should be marked inline.

popcount is on the right. So, it does improve, but will be 30% slower for projects using /Ob1

Imo, all these functions should be marked inline.

No; I asked a while ago, see #619

provides 5-10% speed up

✋ we don't do that here.

See https://github.com/microsoft/STL/wiki/Benchmarking-the-STL#calculating-speedups

So you probably mean 1.05 to 1.1 times speed up

So you probably mean 1.05 to 1.1 times speed up

updated. I based it on my previous commit message: #5367

(actual git history doesn't have that anyways)

Mixed results for _if version, but consistent improvement of the branchless one

i5-1235U (Alder Lake)

P Cores

E Cores

Thanks! 😻 Results on my 5950X, comparing main to this PR with set _CL_=/arch:AVX2 injected non-intrusively:

I'm not worried about the 0.96, that looks like noise (I reran and it was slightly better), and the 1.41 speedups are very healthy indeed.

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

no changes planned.

0️⃣ 1️⃣ 0️⃣