Prefer Buffer.Memmove over Array.Copy in some scenarios

[GrabYourPitchforks]

Minor perf optimizations to Array.Resize<T> and RuntimeHelpers.GetSubArray. In both of these methods it's safe to use Buffer.Memmove rather than bouncing through the more heavyweight Array.Copy. Due to the nature of how we're constructing the arrays we know these operations to be valid.

[jkotas]

Any measurable improvements?

[GrabYourPitchforks]

Quick benchmark, showing an approx. 10% savings in RuntimeHelpers.GetSubArray:

public class ArrayRunner
{
    private object[] _objects;

    [Params(0, 1, 4, 12, 24, 128)]
    public int ArrayLength { get; set; }

    [GlobalSetup]
    public void Setup()
    {
        _objects = Enumerable.Range(0, 128).Select(i => i.ToString()).ToArray();
    }

    [Benchmark]
    public object[] ArrSlice()
    {
        return _objects[..ArrayLength];
    }
}

Method	Job	Toolchain	ArrayLength	Mean	Error	StdDev	Ratio	RatioSD
ArrSlice	Job-XTQBTF	master	0	91.96 ns	1.456 ns	1.362 ns	1.00	0.00
ArrSlice	Job-ECZSQW	memmove	0	85.51 ns	1.713 ns	1.602 ns	0.93	0.02
ArrSlice	Job-XTQBTF	master	1	100.25 ns	1.599 ns	1.417 ns	1.00	0.00
ArrSlice	Job-ECZSQW	memmove	1	89.24 ns	1.682 ns	1.574 ns	0.89	0.02
ArrSlice	Job-XTQBTF	master	4	107.82 ns	1.226 ns	1.147 ns	1.00	0.00
ArrSlice	Job-ECZSQW	memmove	4	97.51 ns	1.557 ns	1.456 ns	0.90	0.02
ArrSlice	Job-XTQBTF	master	12	113.85 ns	1.978 ns	2.641 ns	1.00	0.00
ArrSlice	Job-ECZSQW	memmove	12	104.73 ns	1.160 ns	0.969 ns	0.91	0.03
ArrSlice	Job-XTQBTF	master	24	128.17 ns	1.740 ns	1.627 ns	1.00	0.00
ArrSlice	Job-ECZSQW	memmove	24	115.81 ns	2.310 ns	2.472 ns	0.90	0.02
ArrSlice	Job-XTQBTF	master	128	242.94 ns	4.920 ns	13.134 ns	1.00	0.00
ArrSlice	Job-ECZSQW	memmove	128	222.49 ns	4.485 ns	10.123 ns	0.91	0.06

[stephentoub]

Interesting. I'd have expected a fixed overhead reduction. Is the savings scaling with length?

[GrabYourPitchforks]

@stephentoub I would've expected a fixed-size reduction as well. Maybe something's off on my test box? Here's the benchmarks for Array.Resize<T>, which clearly shows the fixed-size overhead reduction across all test cases.

public class ArrayRunner
{
    private int[] _ints;

    [Params(0, 1, 4, 12, 24, 128)]
    public int ArrayLength { get; set; }

    [GlobalSetup]
    public void Setup()
    {
        _ints = Enumerable.Range(0, 32).ToArray();
    }

    [Benchmark]
    public int[] ArrResize()
    {
        int[] ints = _ints;
        Array.Resize(ref ints, ArrayLength);
        return ints;
    }

    [Benchmark]
    public uint[] ArrResizeCompatibleValueType()
    {
        uint[] uints = Unsafe.As<uint[]>(_ints); // safe, but avoid 'castclass' for benchmark
        Array.Resize(ref uints, ArrayLength);
        return uints;
    }
}

Method	Job	Toolchain	ArrayLength	Mean	Error	StdDev	Median	Ratio	RatioSD
ArrResize	Job-OGUJUP	master	0	16.21 ns	0.398 ns	0.620 ns	16.08 ns	1.00	0.00
ArrResize	Job-PCCOVY	memmove	0	13.10 ns	0.326 ns	0.799 ns	12.94 ns	0.82	0.05
ArrResizeCompatibleValueType	Job-OGUJUP	master	0	48.71 ns	1.044 ns	2.927 ns	48.01 ns	1.00	0.00
ArrResizeCompatibleValueType	Job-PCCOVY	memmove	0	12.98 ns	0.336 ns	0.654 ns	12.81 ns	0.27	0.02
ArrResize	Job-OGUJUP	master	1	17.13 ns	0.477 ns	1.376 ns	16.70 ns	1.00	0.00
ArrResize	Job-PCCOVY	memmove	1	12.39 ns	0.318 ns	0.282 ns	12.37 ns	0.68	0.04
ArrResizeCompatibleValueType	Job-OGUJUP	master	1	44.54 ns	0.778 ns	0.727 ns	44.68 ns	1.00	0.00
ArrResizeCompatibleValueType	Job-PCCOVY	memmove	1	11.93 ns	0.272 ns	0.241 ns	11.90 ns	0.27	0.01
ArrResize	Job-OGUJUP	master	4	16.08 ns	0.247 ns	0.219 ns	16.06 ns	1.00	0.00
ArrResize	Job-PCCOVY	memmove	4	13.12 ns	0.337 ns	0.795 ns	12.84 ns	0.87	0.06
ArrResizeCompatibleValueType	Job-OGUJUP	master	4	44.88 ns	0.872 ns	0.773 ns	44.77 ns	1.00	0.00
ArrResizeCompatibleValueType	Job-PCCOVY	memmove	4	12.32 ns	0.297 ns	0.330 ns	12.24 ns	0.27	0.01
ArrResize	Job-OGUJUP	master	12	19.24 ns	0.432 ns	0.660 ns	19.15 ns	1.00	0.00
ArrResize	Job-PCCOVY	memmove	12	16.05 ns	0.396 ns	0.941 ns	15.72 ns	0.85	0.06
ArrResizeCompatibleValueType	Job-OGUJUP	master	12	49.75 ns	1.066 ns	1.459 ns	49.40 ns	1.00	0.00
ArrResizeCompatibleValueType	Job-PCCOVY	memmove	12	14.27 ns	0.312 ns	0.261 ns	14.26 ns	0.29	0.01
ArrResize	Job-OGUJUP	master	24	22.87 ns	0.260 ns	0.203 ns	22.83 ns	1.00	0.00
ArrResize	Job-PCCOVY	memmove	24	19.64 ns	0.358 ns	0.352 ns	19.57 ns	0.86	0.02
ArrResizeCompatibleValueType	Job-OGUJUP	master	24	53.02 ns	1.126 ns	1.205 ns	52.62 ns	1.00	0.00
ArrResizeCompatibleValueType	Job-PCCOVY	memmove	24	18.80 ns	0.423 ns	0.396 ns	18.86 ns	0.36	0.01
ArrResize	Job-OGUJUP	master	128	52.46 ns	0.915 ns	0.811 ns	52.69 ns	1.00	0.00
ArrResize	Job-PCCOVY	memmove	128	49.79 ns	0.918 ns	0.814 ns	49.62 ns	0.95	0.02
ArrResizeCompatibleValueType	Job-OGUJUP	master	128	83.35 ns	1.715 ns	2.229 ns	83.42 ns	1.00	0.00
ArrResizeCompatibleValueType	Job-PCCOVY	memmove	128	48.83 ns	0.905 ns	0.803 ns	48.76 ns	0.59	0.02

[jkotas]

I guess there is noise from GC or alignment.

[stephentoub]

Seems worthwhile even if just for small array sizes. Thanks.

[GrabYourPitchforks]

Looking into why CI's on the floor. I don't think it's related to this PR but want to double-check.

[GrabYourPitchforks]

Interesting that it's only the Mono debug builds failing. I don't see these tests failing in other PRs. Might be some weirdness in introducing GetArrayDataReference in these code paths. Still investigating.

[GrabYourPitchforks]

/azp run runtime

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[GrabYourPitchforks]

@marek-safar we have a few unit tests that are failing only in mono debug builds. Would you be able to assign somebody to take a look? There appears to be some weirdness in calling MemoryMarshal.GetArrayDataReference from mono debug builds.

[marek-safar]

@vargaz could you please help them with Mono side

[vargaz]

This looks like a problem in Mono's implementation of BulkMoveWithWriteBarrier. Our old Memmove implementation had a type argument because it used different different copy/write barrier functions for different types, but the .netcore version doesn't have this argument.
Perhaps we should move the Memmove method into runtime specific code so Mono can have its own implementation ?

[GrabYourPitchforks]

@vargaz If you point me to the mono-preferred implementation I can send a PR to split these back out. Thanks!

[vargaz]

#37920

[GrabYourPitchforks]

/azp run runtime

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[GrabYourPitchforks]

@vargaz Looks like mono debug tests are still failing, even after merging in your latest changes. See https://github.com/dotnet/runtime/pull/35733/checks.

Ideas?

[GrabYourPitchforks]

/azp run runtime

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[GrabYourPitchforks]

Failing leg seems to have been pulled from current CI runs, and on rerun today everything that ran was green. Going ahead with the merge.