System.arraycopy vs Arrays.copyOf

648 views

Skip to first unread message

Carl Mastrangelo

unread,

Feb 9, 2017, 11:33:12 AM2/9/17

to mechanical-sympathy

Hi,

I'm trying to copy an array of bytes and return the copy. I noticed that there are two ways of doing this: System.arraycopy and Arrays.copyOf, both of which have intrinsic versions in OpenJDK. I was hoping that the Arrays.copyOf version would skip the initial zeroing of the destination array, and thus be slightly faster. However, benchmarking them seems to show that neither one is consistently faster than the other. I'm not that familiar with assembly, so I may be reading it wrong. That said, reading the "Hottest Region 2" below, it looks like copyOf is still zeroing? Can that be right?

Benchmark:

@State(Scope.Benchmark)

public class Benchmark {

@Param({"1048576"})

public int size;

public long[] src;

@Setup

public void setUp() throws Exception {

src = new long[size];

for (int i = 0; i < src.length; i++) {

src[i] = i;

}

/**

* Javadoc comment.

@Benchmark

@BenchmarkMode(Mode.SampleTime)

@OutputTimeUnit(TimeUnit.NANOSECONDS)

public long[] syscopy() {

long[] dest = new long[src.length];

System.arraycopy(src, 0, dest, 0, src.length);

return dest;

}

/**

* Javadoc comment.

@Benchmark

@BenchmarkMode(Mode.SampleTime)

@OutputTimeUnit(TimeUnit.NANOSECONDS)

public long[] arrcopy() {

return Arrays.copyOf(src, src.length);

}

And the Output for arrcopy:

# JMH 1.17.3 (released 56 days ago)

# VM version: JDK 1.8.0_92, VM 25.92-b14

# VM invoker: ~/Downloads/jdk1.8.0_92/jre/bin/java

# VM options: -server -Xms2g -Xmx2g -XX:+UnlockDiagnosticVMOptions -XX:LogFile=/tmp/thelogs -XX:+LogCompilation -XX:+PrintNMethods -XX:+PrintNativeNMethods -XX:+PrintAssembly -XX:+PrintInlining -XX:PrintAssemblyOptions=syntax -XX:+PrintCompilation

# Warmup: 10 iterations, 1 s each

# Measurement: 1 iterations, 1 s each

# Timeout: 10 min per iteration

# Threads: 1 thread, will synchronize iterations

# Benchmark mode: Sampling time

# Benchmark: Benchmark.arrcopy

# Parameters: (size = 1048576)

# Run progress: 0.00% complete, ETA 00:00:22

# Fork: 1 of 1

# Preparing profilers: LinuxPerfAsmProfiler

# Profilers consume stdout and stderr from target VM, use -v EXTRA to copy to console

# Warmup Iteration 1: 3064247.755 ±(99.9%) 1101881.397 ns/op

# Warmup Iteration 2: 2104675.705 ±(99.9%) 116863.885 ns/op

# Warmup Iteration 3: 1933284.271 ±(99.9%) 73166.251 ns/op

# Warmup Iteration 4: 2022234.170 ±(99.9%) 31234.827 ns/op

# Warmup Iteration 5: 2054921.988 ±(99.9%) 171828.509 ns/op

# Warmup Iteration 6: 1984621.714 ±(99.9%) 70876.189 ns/op

# Warmup Iteration 7: 2029098.580 ±(99.9%) 32632.162 ns/op

# Warmup Iteration 8: 1888186.445 ±(99.9%) 27772.059 ns/op

# Warmup Iteration 9: 1956954.176 ±(99.9%) 110294.977 ns/op

# Warmup Iteration 10: 1961879.592 ±(99.9%) 75665.475 ns/op

Iteration 1: 2082120.533 ±(99.9%) 31688.951 ns/op

arrcopy·p0.00: 1851392.000 ns/op

arrcopy·p0.50: 2059264.000 ns/op

arrcopy·p0.90: 2231910.400 ns/op

arrcopy·p0.95: 2273280.000 ns/op

arrcopy·p0.99: 3109150.720 ns/op

arrcopy·p0.999: 4194304.000 ns/op

arrcopy·p0.9999: 4194304.000 ns/op

arrcopy·p1.00: 4194304.000 ns/op

# Processing profiler results: LinuxPerfAsmProfiler

Result "arrcopy":

N = 480

mean = 2082120.533 ±(99.9%) 31688.951 ns/op

Histogram, ns/op:

[1000000.000, 1250000.000) = 0

[1250000.000, 1500000.000) = 0

[1500000.000, 1750000.000) = 0

[1750000.000, 2000000.000) = 166

[2000000.000, 2250000.000) = 276

[2250000.000, 2500000.000) = 26

[2500000.000, 2750000.000) = 6

[2750000.000, 3000000.000) = 1

[3000000.000, 3250000.000) = 1

[3250000.000, 3500000.000) = 1

[3500000.000, 3750000.000) = 1

[3750000.000, 4000000.000) = 1

[4000000.000, 4250000.000) = 1

[4250000.000, 4500000.000) = 0

[4500000.000, 4750000.000) = 0

Percentiles, ns/op:

p(0.0000) = 1851392.000 ns/op

p(50.0000) = 2059264.000 ns/op

p(90.0000) = 2231910.400 ns/op

p(95.0000) = 2273280.000 ns/op

p(99.0000) = 3109150.720 ns/op

p(99.9000) = 4194304.000 ns/op

p(99.9900) = 4194304.000 ns/op

p(99.9990) = 4194304.000 ns/op

p(99.9999) = 4194304.000 ns/op

p(100.0000) = 4194304.000 ns/op

Secondary result "·asm":

PrintAssembly processed: 156748 total address lines.

Perf output processed (skipped 10.568 seconds):

Column 1: cycles (1148 events)

Column 2: instructions (1051 events)

Hottest code regions (>10.00% "cycles" events):

....[Hottest Region 1]..............................................................................

runtime stub, StubRoutines::jlong_disjoint_arraycopy (24 bytes)

0x00007fd8a505278e: neg %rdx

╭ 0x00007fd8a5052791: jmpq Stub::jlong_disjoint_arraycopy+72 0x0x7fd8a50527c8

│↗ 0x00007fd8a5052796: mov 0x8(%rdi,%rdx,8),%rax

││ 0x00007fd8a505279b: mov %rax,0x8(%rcx,%rdx,8)

││ 0x00007fd8a50527a0: inc %rdx

│╰ 0x00007fd8a50527a3: jne Stub::jlong_disjoint_arraycopy+22 0x0x7fd8a5052796

│ 0x00007fd8a50527a5: xor %rax,%rax

│ 0x00007fd8a50527a8: leaveq

│ 0x00007fd8a50527a9: retq

│ 0x00007fd8a50527aa: nopw 0x0(%rax,%rax,1)

0.17% 0.19% │ ↗ 0x00007fd8a50527b0: vmovdqu -0x38(%rdi,%rdx,8),%ymm0

37.11% 12.94% │ │ 0x00007fd8a50527b6: vmovdqu %ymm0,-0x38(%rcx,%rdx,8)

5.66% 1.52% │ │ 0x00007fd8a50527bc: vmovdqu -0x18(%rdi,%rdx,8),%ymm1

0.87% 0.38% │ │ 0x00007fd8a50527c2: vmovdqu %ymm1,-0x18(%rcx,%rdx,8)

5.14% 0.48% ↘ │ 0x00007fd8a50527c8: add $0x8,%rdx

╰ 0x00007fd8a50527cc: jle Stub::jlong_disjoint_arraycopy+48 0x0x7fd8a50527b0

0x00007fd8a50527ce: sub $0x4,%rdx

╭ 0x00007fd8a50527d2: jg Stub::jlong_disjoint_arraycopy+100 0x0x7fd8a50527e4

│ 0x00007fd8a50527d4: vmovdqu -0x18(%rdi,%rdx,8),%ymm0

│ 0x00007fd8a50527da: vmovdqu %ymm0,-0x18(%rcx,%rdx,8)

│ 0x00007fd8a50527e0: add $0x4,%rdx

↘ 0x00007fd8a50527e4: (bad)

0x00007fd8a50527e7: rol $0xf5,%ch

0x00007fd8a50527ea: out %eax,(%dx)

0x00007fd8a50527eb: leaveq

....................................................................................................

48.95% 15.51% <total for region 1>

....[Hottest Region 2]..............................................................................

C1, level 3, java.lang.String::hashCode, version 1 (8 bytes)

0x00007fd8a5106814: mov %ecx,0x8(%rax)

0x00007fd8a5106817: mov %ebx,0xc(%rax)

0x00007fd8a510681a: mov 0xa(%rdx),%cl

0x00007fd8a510681d: and $0xff,%rcx

0x00007fd8a5106824: sub %rcx,%rsi

0x00007fd8a5106827: add %rax,%rcx

0x00007fd8a510682a: sub $0x0,%rsi

╭ 0x00007fd8a510682e: je 0x00007fd8a5106845

│ 0x00007fd8a5106834: xor %rdi,%rdi

│ 0x00007fd8a5106837: shr $0x3,%rsi

4.36% 19.70% │↗ 0x00007fd8a510683b: mov %rdi,-0x8(%rcx,%rsi,8)

32.49% 49.57% ││ 0x00007fd8a5106840: dec %rsi

0.17% │╰ 0x00007fd8a5106843: jne 0x00007fd8a510683b

↘ 0x00007fd8a5106845: retq

0x00007fd8a5106846: push %rbp

0x00007fd8a5106847: mov %rsp,%rbp

0x00007fd8a510684a: mov %rsp,-0x28(%rsp)

0x00007fd8a510684f: sub $0x80,%rsp

0x00007fd8a5106856: mov %rax,0x78(%rsp)

0x00007fd8a510685b: mov %rcx,0x70(%rsp)

0x00007fd8a5106860: mov %rdx,0x68(%rsp)

0x00007fd8a5106865: mov %rbx,0x60(%rsp)

0x00007fd8a510686a: mov %rbp,0x50(%rsp)

....................................................................................................

37.02% 69.27% <total for region 2>

....[Hottest Regions]...............................................................................

48.95% 15.51% runtime stub StubRoutines::jlong_disjoint_arraycopy (24 bytes)

37.02% 69.27% C1, level 3 java.lang.String::hashCode, version 1 (8 bytes)

8.62% 9.32% [kernel.kallsyms] [unknown] (0 bytes)

1.31% 1.05% C1, level 3 java.util.zip.ZipCoder::getBytes, version 582 (9 bytes)

0.61% 0.76% C1, level 3 java.util.zip.ZipCoder::getBytes, version 582 (17 bytes)

0.52% 0.95% C1, level 3 java.lang.String::hashCode, version 1 (5 bytes)

0.44% 0.95% C1, level 3 java.util.zip.ZipCoder::getBytes, version 582 (24 bytes)

0.17% 0.29% C1, level 3 java.util.zip.ZipCoder::getBytes, version 582 (10 bytes)

0.09% [kernel.kallsyms] [unknown] (0 bytes)

0.09% 0.10% [kernel.kallsyms] [unknown] (0 bytes)

0.09% [kernel.kallsyms] [unknown] (0 bytes)

0.09% 0.10% [kernel.kallsyms] [unknown] (0 bytes)

1.13% 1.71% <...other 27 warm regions...>

....................................................................................................

99.83% 100.00% <totals>

....[Hottest Methods (after inlining)]..............................................................

48.95% 15.51% runtime stub StubRoutines::jlong_disjoint_arraycopy

37.54% 70.22% C1, level 3 java.lang.String::hashCode, version 1

9.76% 10.18% [kernel.kallsyms] [unknown]

3.22% 3.71% C1, level 3 java.util.zip.ZipCoder::getBytes, version 582

0.09% interpreter iconst_0 3 iconst_0

0.09% interpreter if_icmpge 162 if_icmpge

0.09% interpreter fast_aputfield 211 fast_aputfield

0.09% 0.19% C1, level 3 Benchmark::arrcopy, version 455

....................................................................................................

99.83% 99.81% <totals>

....[Distribution by Source]........................................................................

48.95% 15.51% runtime stub

40.85% 74.22% C1, level 3

9.76% 10.18% [kernel.kallsyms]

0.26% interpreter

....................................................................................................

99.83% 100.00% <totals>

Norman Maurer

unread,

Feb 9, 2017, 11:41:28 AM2/9/17

to mechanica...@googlegroups.com

As far as I know both do zero out.

--
You received this message because you are subscribed to the Google Groups "mechanical-sympathy" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mechanical-symp...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Jean-Philippe BEMPEL

unread,

Feb 10, 2017, 5:33:41 AM2/10/17

to mechanical-sympathy

Hi Carl,

Did you read this: https://shipilev.net/blog/2016/arrays-wisdom-ancients/ ? :)

Cheers

Nikolay Tsankov

unread,

Feb 15, 2017, 4:27:09 AM2/15/17

to mechanica...@googlegroups.com

I strongly suspect that you are not giving it enough time to warm up.

The zero-ing happens when you allocate the destination arrays and hotspot can often optimize it away if you copy into it right after allocation. As Mr. Shipilev pointed there are cases when it doesn't yet, e.g. when you use a field for the length parameter.

To test this assumption, I've added another benchmark to your test and run it with a longer warmup:

@Benchmark
public long[] arraycopy_field() {
    long[] dst = new long[size];
    System.arraycopy(src, 0, dst, 0, size);
    return dst;
}
Results:
Benchmark                    (size)  Mode  Cnt        Score       Error  Units
ArrayBench.arraycopy_field  1048576  avgt   30  1907102.105 ± 30966.659  ns/op       <- zeroing slows it down
ArrayBench.arrcopy          1048576  avgt   30  1402851.015 ±  4220.862  ns/op       <- no zeroing
ArrayBench.syscopy          1048576  avgt   30  1422812.870 ± 26030.386  ns/op       <- no zeroing

I have not looked at the assembly though, so I might be wrong

--
You received this message because you are subscribed to the Google Groups "mechanical-sympathy" group.

To unsubscribe from this group and stop receiving emails from it, send an email to mechanical-sympathy+unsub...@googlegroups.com.

Reply all

Reply to author

Forward

0 new messages