为什么 "MOVQ 0x30(SP), DX" 慢？

Question

请参阅以下 pprof session。在 treesort.add 的第 42 行中，有一个 int 比较。我认为它占所有 cpu 时间的 64%。在 disasm 中，操作是 "MOVQ 0x30(SP), DX"。为什么这么慢？

File: treesort_bench.test.exe
Type: cpu
Time: Sep 7, 2018 at 3:15pm (EDT)
Duration: 2.60s, Total samples = 2.43s (93.44%)
Entering interactive mode (type "help" for commands, "o" for options)
(pprof) top 10
Showing nodes accounting for 2.41s, 99.18% of 2.43s total
Dropped 2 nodes (cum <= 0.01s)
      flat  flat%   sum%        cum   cum%
     2.40s 98.77% 98.77%      2.42s 99.59%  gopl.io/ch4/treesort.add
     0.01s  0.41% 99.18%      0.02s  0.82%  runtime.mallocgc
         0     0% 99.18%      0.26s 10.70%  gopl.io/ch4/treesort.Sort
         0     0% 99.18%      0.25s 10.29%  gopl.io/ch4/treesort_bench.BenchmarkSort
         0     0% 99.18%      0.26s 10.70%  gopl.io/ch4/treesort_bench.run
         0     0% 99.18%      0.02s  0.82%  runtime.newobject
         0     0% 99.18%      0.22s  9.05%  testing.(*B).launch
         0     0% 99.18%      0.02s  0.82%  testing.(*B).run1.func1
         0     0% 99.18%      0.25s 10.29%  testing.(*B).runN
(pprof) list add
Total: 2.43s
ROUTINE ======================== gopl.io/ch4/treesort.add in go\src\gopl.io\ch4\treesort\sort.go
     2.40s      4.45s (flat, cum) 183.13% of Total
         .          .     30:           values = appendValues(values, t.right)
         .          .     31:   }
         .          .     32:   return values
         .          .     33:}
         .          .     34:
      90ms       90ms     35:func add(t *tree, value int) *tree {
         .          .     36:   if t == nil {
         .          .     37:           // Equivalent to return &tree{value: value}.
         .       20ms     38:           t = new(tree)
         .          .     39:           t.value = value
         .          .     40:           return t
         .          .     41:   }
     1.55s      1.55s     42:   flag := value < t.value
         .          .     43:   if flag {
         .      240ms     44:           t.left = add(t.left, value)
         .          .     45:   } else {
     630ms      2.42s     46:           t.right = add(t.right, value)
         .          .     47:   }
     130ms      130ms     48:   return t
         .          .     49:}
         .          .     50:
         .          .     51://!-
(pprof) disasm add
Total: 2.43s
ROUTINE ======================== gopl.io/ch4/treesort.add
     2.40s      5.08s (flat, cum) 209.05% of Total
      50ms       50ms     4fcb66: MOVQ 0(AX), CX                          ;gopl.io/ch4/treesort.add sort.go:42
     1.48s      1.48s     4fcb69: MOVQ 0x30(SP), DX
      20ms       20ms     4fcb6e: CMPQ CX, DX
         .          .     4fcb71: JGE 0x4fcbbb                            ;sort.go:43

Answer 1

Why is “MOVQ 0x30(SP), DX” slow?

您提供的证据不足以证明指令速度慢。

MOVQ — 移动四字 — 是来自 Intel 64 和 IA-32 架构指令集的指令。参见 Intel® 64 and IA-32 Architectures Software Developer Manuals

MOVQ 0x30(SP), DX 指令将 tree.value 变量的 8 个字节从内存移动到 DX 寄存器。

与任何其他科学努力一样，性能测量依赖于可重现的结果。您提供的信息不足以重现您的结果。比如treesort_bench.test.exe的代码在哪里，什么处理器，什么内存，什么操作系统。

我试过了，但无法重现您的结果。将您的代码和重现结果的步骤添加到您的问题中。

为什么 "MOVQ 0x30(SP), DX" 慢？

Why is "MOVQ 0x30(SP), DX" slow?

comparison

performance

go

disassembly

pprof