Assembly 编译程序集与手写程序集的性能差异

我一直在玩围棋的汇编语言，我写了一个函数作为练习

我已经基于本机Go版本，而程序集版本基于。在对这两个函数进行基准测试后，我发现原生Go版本比汇编版本快1.5倍-2倍左右，尽管手写汇编版本与

Go tool 6g-S popcount.Go的输出几乎相同

从
go test-bench=。

PASS
BenchmarkPopCount       100000000               19.4 ns/op 
BenchmarkPopCount_g     200000000                8.97 ns/op
ok      popcount   4.777s

"".popCount_g t=1 size=64 value=0 args=0x10 locals=0
        000000 00000 (popcount.go:5)    TEXT    "".popCount_g+0(SB),4,$0-16
        000000 00000 (popcount.go:5)    NOP     ,
        000000 00000 (popcount.go:5)    NOP     ,
        000000 00000 (popcount.go:5)    MOVL    "".i+8(FP),BP
        0x0004 00004 (popcount.go:5)    FUNCDATA        $2,gclocals┬À9308e7ef08d2cc2f72ae1228688dacf9+0(SB)
        0x0004 00004 (popcount.go:5)    FUNCDATA        $3,gclocals┬À3280bececceccd33cb74587feedb1f9f+0(SB)
        0x0004 00004 (popcount.go:6)    MOVL    BP,BX
        0x0006 00006 (popcount.go:6)    SHRL    $1,BX
        0x0008 00008 (popcount.go:6)    ANDL    $1431655765,BX
        0x000e 00014 (popcount.go:6)    SUBL    BX,BP
        0x0010 00016 (popcount.go:7)    MOVL    BP,AX
        0x0012 00018 (popcount.go:7)    ANDL    $858993459,AX
        0x0017 00023 (popcount.go:7)    SHRL    $2,BP
        0x001a 00026 (popcount.go:7)    ANDL    $858993459,BP
        0x0020 00032 (popcount.go:7)    ADDL    BP,AX
        0x0022 00034 (popcount.go:8)    MOVL    AX,BX
        0x0024 00036 (popcount.go:8)    SHRL    $4,BX
        0x0027 00039 (popcount.go:8)    ADDL    AX,BX
        0x0029 00041 (popcount.go:8)    ANDL    $252645135,BX
        0x002f 00047 (popcount.go:8)    IMULL   $16843009,BX
        0x0035 00053 (popcount.go:8)    SHRL    $24,BX
        0x0038 00056 (popcount.go:8)    MOVL    BX,"".~r1+16(FP)
        0x003c 00060 (popcount.go:8)    RET     ,

popcount.go

package popcount

func popCount(i uint32) uint32 // Defined in popcount_amd64.s

func popCount_g(i uint32) uint32 {
    i = i - ((i >> 1) & 0x55555555)
    i = (i & 0x33333333) + ((i >> 2) & 0x33333333)
    return (((i + (i >> 4)) & 0x0F0F0F0F) * 0x01010101) >> 24
}

package popcount

import "testing"

func TestPopcount(t *testing.T) {
    for i := uint32(0); i < uint32(100); i++ {
        if popCount(i) != popCount_g(i) {
            t.Fatalf("failed on input = %v", i)
        }
    }
}

func BenchmarkPopCount(b *testing.B) {
    for i := 0; i < b.N; i++ {
        popCount(uint32(i))
    }
}

func BenchmarkPopCount_g(b *testing.B) {
    for i := 0; i < b.N; i++ {
        popCount_g(uint32(i))
    }
}

popcount\u测试开始

package popcount

func popCount(i uint32) uint32 // Defined in popcount_amd64.s

func popCount_g(i uint32) uint32 {
    i = i - ((i >> 1) & 0x55555555)
    i = (i & 0x33333333) + ((i >> 2) & 0x33333333)
    return (((i + (i >> 4)) & 0x0F0F0F0F) * 0x01010101) >> 24
}

package popcount

import "testing"

func TestPopcount(t *testing.T) {
    for i := uint32(0); i < uint32(100); i++ {
        if popCount(i) != popCount_g(i) {
            t.Fatalf("failed on input = %v", i)
        }
    }
}

func BenchmarkPopCount(b *testing.B) {
    for i := 0; i < b.N; i++ {
        popCount(uint32(i))
    }
}

func BenchmarkPopCount_g(b *testing.B) {
    for i := 0; i < b.N; i++ {
        popCount_g(uint32(i))
    }
}

go工具6g-S popcount.go的输出

PASS
BenchmarkPopCount       100000000               19.4 ns/op 
BenchmarkPopCount_g     200000000                8.97 ns/op
ok      popcount   4.777s

"".popCount_g t=1 size=64 value=0 args=0x10 locals=0
        000000 00000 (popcount.go:5)    TEXT    "".popCount_g+0(SB),4,$0-16
        000000 00000 (popcount.go:5)    NOP     ,
        000000 00000 (popcount.go:5)    NOP     ,
        000000 00000 (popcount.go:5)    MOVL    "".i+8(FP),BP
        0x0004 00004 (popcount.go:5)    FUNCDATA        $2,gclocals┬À9308e7ef08d2cc2f72ae1228688dacf9+0(SB)
        0x0004 00004 (popcount.go:5)    FUNCDATA        $3,gclocals┬À3280bececceccd33cb74587feedb1f9f+0(SB)
        0x0004 00004 (popcount.go:6)    MOVL    BP,BX
        0x0006 00006 (popcount.go:6)    SHRL    $1,BX
        0x0008 00008 (popcount.go:6)    ANDL    $1431655765,BX
        0x000e 00014 (popcount.go:6)    SUBL    BX,BP
        0x0010 00016 (popcount.go:7)    MOVL    BP,AX
        0x0012 00018 (popcount.go:7)    ANDL    $858993459,AX
        0x0017 00023 (popcount.go:7)    SHRL    $2,BP
        0x001a 00026 (popcount.go:7)    ANDL    $858993459,BP
        0x0020 00032 (popcount.go:7)    ADDL    BP,AX
        0x0022 00034 (popcount.go:8)    MOVL    AX,BX
        0x0024 00036 (popcount.go:8)    SHRL    $4,BX
        0x0027 00039 (popcount.go:8)    ADDL    AX,BX
        0x0029 00041 (popcount.go:8)    ANDL    $252645135,BX
        0x002f 00047 (popcount.go:8)    IMULL   $16843009,BX
        0x0035 00053 (popcount.go:8)    SHRL    $24,BX
        0x0038 00056 (popcount.go:8)    MOVL    BX,"".~r1+16(FP)
        0x003c 00060 (popcount.go:8)    RET     ,

我知道
FUNCDATA
行包含垃圾收集器的信息，但除此之外，我没有看到任何明显的区别

是什么导致这两个函数在速度上存在如此大的差异？
如果您查看函数调用的伪汇编程序，您将看到调用了
.s
（汇编程序）版本，并且调用开销很大，
.go
版本是内联的

func S() {
    pc := popCount(uint32(0))
    _ = pc
}

"".S t=1 size=48 value=0 args=0x0 locals=0x10
    0x0000 00000 (popcount.go:11)   TEXT    "".S+0(SB),$16-0
    0x0000 00000 (popcount.go:11)   MOVQ    (TLS),CX
    0x0009 00009 (popcount.go:11)   CMPQ    SP,(CX)
    0x000c 00012 (popcount.go:11)   JHI ,21
    0x000e 00014 (popcount.go:11)   CALL    ,runtime.morestack00_noctxt(SB)
    0x0013 00019 (popcount.go:11)   JMP ,0
    0x0015 00021 (popcount.go:11)   SUBQ    $16,SP
    0x0019 00025 (popcount.go:11)   FUNCDATA    $2,gclocals·3280bececceccd33cb74587feedb1f9f+0(SB)
    0x0019 00025 (popcount.go:11)   FUNCDATA    $3,gclocals·3280bececceccd33cb74587feedb1f9f+0(SB)
    0x0019 00025 (popcount.go:12)   MOVL    $0,(SP)
    0x0020 00032 (popcount.go:12)   PCDATA  $1,$0
    0x0020 00032 (popcount.go:12)   CALL    ,"".popCount(SB)
    0x0025 00037 (popcount.go:12)   MOVL    8(SP),BX
    0x0029 00041 (popcount.go:12)   NOP ,
    0x0029 00041 (popcount.go:14)   ADDQ    $16,SP
    0x002d 00045 (popcount.go:14)   RET ,

func S_G() {
    pc := popCount_g(uint32(0))
    _ = pc
}

"".S_G t=1 size=64 value=0 args=0x0 locals=0x8
    0x0000 00000 (popcount.go:16)   TEXT    "".S_G+0(SB),4,$8-0
    0x0000 00000 (popcount.go:16)   SUBQ    $8,SP
    0x0004 00004 (popcount.go:16)   FUNCDATA    $2,gclocals·3280bececceccd33cb74587feedb1f9f+0(SB)
    0x0004 00004 (popcount.go:16)   FUNCDATA    $3,gclocals·3280bececceccd33cb74587feedb1f9f+0(SB)
    0x0004 00004 (popcount.go:17)   MOVL    $0,BP
    0x0006 00006 (popcount.go:17)   MOVL    BP,BX
    0x0008 00008 (popcount.go:17)   SHRL    $1,BX
    0x000a 00010 (popcount.go:17)   ANDL    $1431655765,BX
    0x0010 00016 (popcount.go:17)   SUBL    BX,BP
    0x0012 00018 (popcount.go:17)   MOVL    BP,AX
    0x0014 00020 (popcount.go:17)   ANDL    $858993459,AX
    0x0019 00025 (popcount.go:17)   SHRL    $2,BP
    0x001c 00028 (popcount.go:17)   ANDL    $858993459,BP
    0x0022 00034 (popcount.go:17)   ADDL    BP,AX
    0x0024 00036 (popcount.go:17)   MOVL    AX,BX
    0x0026 00038 (popcount.go:17)   SHRL    $4,BX
    0x0029 00041 (popcount.go:17)   ADDL    AX,BX
    0x002b 00043 (popcount.go:17)   ANDL    $252645135,BX
    0x0031 00049 (popcount.go:17)   IMULL   $16843009,BX
    0x0037 00055 (popcount.go:17)   SHRL    $24,BX
    0x003a 00058 (popcount.go:19)   ADDQ    $8,SP
    0x003e 00062 (popcount.go:19)   RET ,

如果您真的想要一个popcnt，那么它在X86处理器上已经是本机指令一段时间了。您的编译器可能有它的内在特性，例如Microsoft的uuuPopCnt（）（也有uuPopCnt16（）和uuuPopCnt64（）），或GCC的uuuuBuiltin_popcount（）（或64位的uuuuBuiltin_popcountll（））
 需要记住的一点是，go中的链接器比传统的C链接器聪明得多。特别是它可以转换指令、重新排序代码等，因此我建议您使用objdump查看最终的二进制文件，而不是比较go汇编程序的输出。@NickCraig-Wood非常好的建议。我会检查一下。@NickCraig Wood objdump显示程序集仍然非常相似，但是手工编写的程序集版本在调用函数之前显示
callq 420610
。本机go版本在
main.main中有整个程序集，因此我猜区别在于分配更多内存所需的时间，因为额外的跳转不会花费那么长的时间。有什么想法吗？@NickCraig-Wood根据peterSO的回答，看起来每次操作的呼叫开销可能会花费几纳秒。啊，好吧，现在我需要尝试学习如何内联汇编代码，如果可以的话。。。我去兔子洞！AFAIK：内联汇编代码是不可能的：-（在过去的开发列表中肯定已经讨论过了。谢谢，你是对的（我使用@NickCraig-Wood的建议使用objdump检查最终的二进制文件发现了同样的事情）是的，我知道最近英特尔和AMD处理器上的
popcnt
指令。这实际上只是学习Go使用的基于Plan9的汇编语言的一个练习，我在测试时遇到了这种性能差异。我感觉某些指令（如
popcnt
）Plan9/Go汇编语言不支持，并且似乎没有可用的
popcnt
或类似函数。我上次检查时不支持它，但您可以自己对指令进行编码，并在汇编中包含十六进制格式的文本字节，例如字节$0xf3；字节$0x48；字节$0x0f；字节$0xb8；字节$0x16//popcntQ（SI），DX