理解 JIT 对 for 循环的重写

Question

我有以下Java代码（所有数组在我们调用“arrays”之前都已初始化，并且所有数组的大小均为“arraySize”）

int arraySize = 64;

float[] a;
float[] b;
float[] result;

public void arrays() {
    for (int i = 0; i < arraySize; i++) {
        result[i] = ((a[i] * b[i] + b[i] - b[i]) / b[i]) +
                     a[i] + a[i] + a[i] + a[i];
    }
}

JIT 的输出是：

# {method} {0x00000001034751a8} 'arrays' '()V' in 'main/ComplexExpression'
#           [sp+0x30]  (sp of caller)
[Entry Point]
0x000000010c4c55a0: mov 0x8(%rsi),%r10d
0x000000010c4c55a4: movabs $0x800000000,%r11
0x000000010c4c55ae: add %r11,%r10
0x000000010c4c55b1: cmp %r10,%rax
0x000000010c4c55b4: jne 0x000000010c44b780  ;   {runtime_call ic_miss_stub}
0x000000010c4c55ba: xchg %ax,%ax
0x000000010c4c55bc: nopl 0x0(%rax)
[Verified Entry Point]
0x000000010c4c55c0: mov %eax,-0x14000(%rsp)
0x000000010c4c55c7: push %rbp
0x000000010c4c55c8: sub $0x20,%rsp  ;*synchronization entry
                                    ; - main.ComplexExpression::arrays@-1 (line 51)
0x000000010c4c55cc: mov %rsi,%rcx
0x000000010c4c55cf: mov 0xc(%rsi),%ebp  ;*getfield arraySize {reexecute=0 rethrow=0 return_oop=0}
                                        ; - main.ComplexExpression::arrays@4 (line 51)
0x000000010c4c55d2: test %ebp,%ebp
0x000000010c4c55d4: jle L0006  ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0}
                               ; - main.ComplexExpression::arrays@7 (line 51)
0x000000010c4c55da: mov 0x10(%rsi),%r11d  ;*getfield a {reexecute=0 rethrow=0 return_oop=0}
                                          ; - main.ComplexExpression::arrays@16 (line 52)
0x000000010c4c55de: xchg %ax,%ax
0x000000010c4c55e0: mov 0xc(%r12,%r11,8),%r10d  ; implicit exception: dispatches to 0x000000010c4c58a3
                                                ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c55e5: test %r10d,%r10d
0x000000010c4c55e8: jbe L0007
0x000000010c4c55ee: movslq %ebp,%r9
0x000000010c4c55f1: movslq %r10d,%r10
0x000000010c4c55f4: dec %r9
0x000000010c4c55f7: cmp %r10,%r9
0x000000010c4c55fa: nopw 0x0(%rax,%rax,1)
0x000000010c4c5600: jae L0007
0x000000010c4c5606: mov 0x14(%rsi),%ebx  ;*getfield b {reexecute=0 rethrow=0 return_oop=0}
                                         ; - main.ComplexExpression::arrays@22 (line 52)
0x000000010c4c5609: mov 0xc(%r12,%rbx,8),%r10d  ; implicit exception: dispatches to 0x000000010c4c58a3
                                                ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c560e: test %r10d,%r10d
0x000000010c4c5611: jbe L0007
0x000000010c4c5617: movslq %r10d,%r10
0x000000010c4c561a: nopw 0x0(%rax,%rax,1)
0x000000010c4c5620: cmp %r10,%r9
0x000000010c4c5623: jae L0007
0x000000010c4c5629: mov 0x18(%rsi),%r8d  ;*getfield result {reexecute=0 rethrow=0 return_oop=0}
                                         ; - main.ComplexExpression::arrays@11 (line 52)
0x000000010c4c562d: mov 0xc(%r12,%r8,8),%r10d  ; implicit exception: dispatches to 0x000000010c4c58a3
                                               ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                               ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c5632: test %r10d,%r10d
0x000000010c4c5635: jbe L0007
0x000000010c4c563b: movslq %r10d,%r10
0x000000010c4c563e: xchg %ax,%ax
0x000000010c4c5640: cmp %r10,%r9
0x000000010c4c5643: jae L0007
0x000000010c4c5649: lea (%r12,%r8,8),%rdx
0x000000010c4c564d: lea (%r12,%r11,8),%rdi
0x000000010c4c5651: mov %edx,%r11d
0x000000010c4c5654: lea (%r12,%rbx,8),%rax
0x000000010c4c5658: shr $0x2,%r11d
0x000000010c4c565c: and $0x7,%r11d
0x000000010c4c5660: mov $0x3,%r9d
0x000000010c4c5666: sub %r11d,%r9d
0x000000010c4c5669: and $0x7,%r9d
0x000000010c4c566d: inc %r9d
0x000000010c4c5670: cmp %ebp,%r9d
0x000000010c4c5673: cmovg %ebp,%r9d
0x000000010c4c5677: xor %r10d,%r10d
0x000000010c4c567a: xor %r11d,%r11d  ;*aload_0 {reexecute=0 rethrow=0 return_oop=0}
                                     ; - main.ComplexExpression::arrays@10 (line 52)
             L0000: vmovss 0x10(%rax,%r11,4),%xmm1  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                    ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c5684: vmovss 0x10(%rdi,%r11,4),%xmm0  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                    ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c568b: vmulss %xmm1,%xmm0,%xmm3
0x000000010c4c568f: vaddss %xmm1,%xmm3,%xmm2
0x000000010c4c5693: vsubss %xmm1,%xmm2,%xmm3
0x000000010c4c5697: vdivss %xmm1,%xmm3,%xmm1
0x000000010c4c569b: vaddss %xmm0,%xmm1,%xmm2
0x000000010c4c569f: vaddss %xmm0,%xmm2,%xmm1
0x000000010c4c56a3: vaddss %xmm0,%xmm1,%xmm2
0x000000010c4c56a7: vaddss %xmm0,%xmm2,%xmm0
0x000000010c4c56ab: vmovss %xmm0,0x10(%rdx,%r11,4)  ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                                    ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c56b2: inc %r11d  ;*iinc {reexecute=0 rethrow=0 return_oop=0}
                               ; - main.ComplexExpression::arrays@78 (line 51)
0x000000010c4c56b5: cmp %r9d,%r11d
0x000000010c4c56b8: jl L0000  ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0}
                              ; - main.ComplexExpression::arrays@7 (line 51)
0x000000010c4c56ba: mov %ebp,%r9d
0x000000010c4c56bd: add $0xffffffe1,%r9d
0x000000010c4c56c1: mov $0x80000000,%r8d
0x000000010c4c56c7: cmp %r9d,%ebp
0x000000010c4c56ca: cmovl %r8d,%r9d
0x000000010c4c56ce: cmp %r9d,%r11d
0x000000010c4c56d1: jge L0004
0x000000010c4c56d7: mov $0x7d00,%ebx
             L0001: mov %r9d,%esi
0x000000010c4c56df: sub %r11d,%esi
0x000000010c4c56e2: cmp %r11d,%r9d
0x000000010c4c56e5: cmovl %r10d,%esi
0x000000010c4c56e9: cmp $0x7d00,%esi
0x000000010c4c56ef: cmova %ebx,%esi
0x000000010c4c56f2: add %r11d,%esi
0x000000010c4c56f5: data16 data16 nopw 0x0(%rax,%rax,1)  ;*aload_0 {reexecute=0 rethrow=0 return_oop=0}
                                                         ; - main.ComplexExpression::arrays@10 (line 52)
             L0002: vmovdqu 0x10(%rax,%r11,4),%ymm0  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c5707: vmovdqu 0x10(%rdi,%r11,4),%ymm1  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c570e: vmulps %ymm0,%ymm1,%ymm2
0x000000010c4c5712: vaddps %ymm0,%ymm2,%ymm2
0x000000010c4c5716: vsubps %ymm0,%ymm2,%ymm2
0x000000010c4c571a: vdivps %ymm0,%ymm2,%ymm0
0x000000010c4c571e: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c5722: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c5726: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c572a: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c572e: vmovdqu %ymm0,0x10(%rdx,%r11,4)  ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c5735: vmovdqu 0x30(%rdi,%r11,4),%ymm0  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c573c: vmovdqu 0x30(%rax,%r11,4),%ymm1  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c5743: vmulps %ymm1,%ymm0,%ymm2
0x000000010c4c5747: vaddps %ymm1,%ymm2,%ymm2
0x000000010c4c574b: vsubps %ymm1,%ymm2,%ymm2
0x000000010c4c574f: vdivps %ymm1,%ymm2,%ymm1
0x000000010c4c5753: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c5757: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c575b: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c575f: vaddps %ymm0,%ymm1,%ymm0
0x000000010c4c5763: vmovdqu %ymm0,0x30(%rdx,%r11,4)  ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c576a: vmovdqu 0x50(%rdi,%r11,4),%ymm0  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c5771: vmovdqu 0x50(%rax,%r11,4),%ymm1  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c5778: vmulps %ymm1,%ymm0,%ymm2
0x000000010c4c577c: vaddps %ymm1,%ymm2,%ymm2
0x000000010c4c5780: vsubps %ymm1,%ymm2,%ymm2
0x000000010c4c5784: vdivps %ymm1,%ymm2,%ymm1
0x000000010c4c5788: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c578c: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c5790: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c5794: vaddps %ymm0,%ymm1,%ymm0
0x000000010c4c5798: vmovdqu %ymm0,0x50(%rdx,%r11,4)  ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c579f: vmovdqu 0x70(%rdi,%r11,4),%ymm0  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c57a6: vmovdqu 0x70(%rax,%r11,4),%ymm1  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c57ad: vmulps %ymm1,%ymm0,%ymm2
0x000000010c4c57b1: vaddps %ymm1,%ymm2,%ymm2
0x000000010c4c57b5: vsubps %ymm1,%ymm2,%ymm2
0x000000010c4c57b9: vdivps %ymm1,%ymm2,%ymm1
0x000000010c4c57bd: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c57c1: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c57c5: vaddps %ymm0,%ymm1,%ymm1
0x000000010c4c57c9: vaddps %ymm0,%ymm1,%ymm0
0x000000010c4c57cd: vmovdqu %ymm0,0x70(%rdx,%r11,4)  ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c57d4: add $0x20,%r11d  ;*iinc {reexecute=0 rethrow=0 return_oop=0}
                                     ; - main.ComplexExpression::arrays@78 (line 51)
0x000000010c4c57d8: cmp %esi,%r11d
0x000000010c4c57db: nopl 0x0(%rax,%rax,1)
0x000000010c4c57e0: jl L0002  ;*goto {reexecute=0 rethrow=0 return_oop=0}
                              ; - main.ComplexExpression::arrays@81 (line 51)
0x000000010c4c57e6: mov 0x348(%r15),%rsi  ; ImmutableOopMap {rcx=Oop rdi=Oop rdx=Oop rax=Oop }
                                          ;*goto {reexecute=1 rethrow=0 return_oop=0}
                                          ; - (reexecute) main.ComplexExpression::arrays@81 (line 51)
0x000000010c4c57ed: test %eax,(%rsi)  ;*goto {reexecute=0 rethrow=0 return_oop=0}
                                      ; - main.ComplexExpression::arrays@81 (line 51)
                                      ;   {poll} *** SAFEPOINT POLL ***
0x000000010c4c57ef: cmp %r9d,%r11d
0x000000010c4c57f2: jl L0001
0x000000010c4c57f8: mov %ebp,%r10d
0x000000010c4c57fb: add $0xfffffff9,%r10d
0x000000010c4c57ff: cmp %r10d,%ebp
0x000000010c4c5802: cmovl %r8d,%r10d
0x000000010c4c5806: cmp %r10d,%r11d
0x000000010c4c5809: jge L0004
0x000000010c4c580b: nop  ;*aload_0 {reexecute=0 rethrow=0 return_oop=0}
                         ; - main.ComplexExpression::arrays@10 (line 52)
             L0003: vmovdqu 0x10(%rax,%r11,4),%ymm0  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c5813: vmovdqu 0x10(%rdi,%r11,4),%ymm1  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c581a: vmulps %ymm0,%ymm1,%ymm2
0x000000010c4c581e: vaddps %ymm0,%ymm2,%ymm2
0x000000010c4c5822: vsubps %ymm0,%ymm2,%ymm2
0x000000010c4c5826: vdivps %ymm0,%ymm2,%ymm0
0x000000010c4c582a: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c582e: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c5832: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c5836: vaddps %ymm1,%ymm0,%ymm0
0x000000010c4c583a: vmovdqu %ymm0,0x10(%rdx,%r11,4)  ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                                     ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c5841: add $0x8,%r11d  ;*iinc {reexecute=0 rethrow=0 return_oop=0}
                                    ; - main.ComplexExpression::arrays@78 (line 51)
0x000000010c4c5845: cmp %r10d,%r11d
0x000000010c4c5848: jl L0003  ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0}
                              ; - main.ComplexExpression::arrays@7 (line 51)
             L0004: cmp %ebp,%r11d
0x000000010c4c584d: jge L0006
0x000000010c4c584f: nop  ;*aload_0 {reexecute=0 rethrow=0 return_oop=0}
                         ; - main.ComplexExpression::arrays@10 (line 52)
             L0005: vmovss 0x10(%rax,%r11,4),%xmm1  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                    ; - main.ComplexExpression::arrays@26 (line 52)
0x000000010c4c5857: vmovss 0x10(%rdi,%r11,4),%xmm0  ;*faload {reexecute=0 rethrow=0 return_oop=0}
                                                    ; - main.ComplexExpression::arrays@20 (line 52)
0x000000010c4c585e: vmulss %xmm1,%xmm0,%xmm3
0x000000010c4c5862: vaddss %xmm1,%xmm3,%xmm2
0x000000010c4c5866: vsubss %xmm1,%xmm2,%xmm3
0x000000010c4c586a: vdivss %xmm1,%xmm3,%xmm1
0x000000010c4c586e: vaddss %xmm0,%xmm1,%xmm2
0x000000010c4c5872: vaddss %xmm0,%xmm2,%xmm1
0x000000010c4c5876: vaddss %xmm0,%xmm1,%xmm2
0x000000010c4c587a: vaddss %xmm0,%xmm2,%xmm0
0x000000010c4c587e: vmovss %xmm0,0x10(%rdx,%r11,4)  ;*fastore {reexecute=0 rethrow=0 return_oop=0}
                                                    ; - main.ComplexExpression::arrays@77 (line 52)
0x000000010c4c5885: inc %r11d  ;*iinc {reexecute=0 rethrow=0 return_oop=0}
                               ; - main.ComplexExpression::arrays@78 (line 51)
0x000000010c4c5888: cmp %ebp,%r11d
0x000000010c4c588b: jl L0005
             L0006: vzeroupper
0x000000010c4c5890: add $0x20,%rsp
0x000000010c4c5894: pop %rbp
0x000000010c4c5895: cmp 0x340(%r15),%rsp  ;   {poll_return} *** SAFEPOINT POLL ***
0x000000010c4c589c: ja L0008
0x000000010c4c58a2: ret  ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0}
                         ; - main.ComplexExpression::arrays@7 (line 51)
             L0007: mov $0xffffff76,%esi
0x000000010c4c58a8: mov %rcx,(%rsp)
0x000000010c4c58ac: vzeroupper
0x000000010c4c58af: call 0x000000010c451000  ; ImmutableOopMap {[0]=Oop }
                                             ;*if_icmpge {reexecute=1 rethrow=0 return_oop=0}
                                             ; - (reexecute) main.ComplexExpression::arrays@7 (line 51)
                                             ;   {runtime_call UncommonTrapBlob}
             L0008: movabs $0x10c4c5895,%r10  ;   {internal_word}
0x000000010c4c58be: mov %r10,0x358(%r15)
0x000000010c4c58c5: jmp 0x000000010c452100  ;   {runtime_call SafepointBlob}
0x000000010c4c58ca: hlt
0x000000010c4c58cb: hlt
0x000000010c4c58cc: hlt
0x000000010c4c58cd: hlt
0x000000010c4c58ce: hlt
0x000000010c4c58cf: hlt
0x000000010c4c58d0: hlt
0x000000010c4c58d1: hlt
0x000000010c4c58d2: hlt
0x000000010c4c58d3: hlt
0x000000010c4c58d4: hlt
0x000000010c4c58d5: hlt
0x000000010c4c58d6: hlt
0x000000010c4c58d7: hlt
0x000000010c4c58d8: hlt
0x000000010c4c58d9: hlt
0x000000010c4c58da: hlt
0x000000010c4c58db: hlt
0x000000010c4c58dc: hlt
0x000000010c4c58dd: hlt
0x000000010c4c58de: hlt
0x000000010c4c58df: hlt
[Exception Handler]
0x000000010c4c58e0: jmp 0x000000010c464a00  ;   {no_reloc}
[Deopt Handler Code]
0x000000010c4c58e5: call 0x000000010c4c58ea
0x000000010c4c58ea: subq $0x5,(%rsp)
0x000000010c4c58ef: jmp 0x000000010c4513a0  ;   {runtime_call DeoptimizationBlob}
0x000000010c4c58f4: hlt
0x000000010c4c58f5: hlt
0x000000010c4c58f6: hlt
0x000000010c4c58f7: hlt

看起来 for 循环被翻译成多个“汇编循环”。我猜主要部分在 L0002 中，您可以在其中看到循环展开，但还有 L0000、L0003 和 L0005，它们似乎也是 for 循环的一部分，但我很难理解它们是如何融入的。

有人可以解释所有标签中的所有部分如何组成实际的循环吗？我想象它是 JIT 中的一些“已知”模式，但我只是不知道。

Answer 1

仅讨论跳转标签所示的主循环。

\n

L0000循环：与 8 个元素（32 字节）的倍数对齐。这and $0x7是一个致命的赠品。当涉及到向量指令时，向量对齐几乎到处都有讨论。我没有一个黄金子弹参考链接来解释该主题，但您会在 SO 和其他地方找到大量资源。Intel \xc2\xae 64 和 IA-32 架构优化参考手册也应该涵盖它。

\n

即使没有，最有效的内存访问应该是不言而喻的，即数据不跨缓存线或更糟糕的内存页进行分割。与自然向量大小对齐可以确保这一点。

\n

AVX 和 AVX2 硬件通常可以很好地处理未对齐情况，但较旧的 SSE和较新的 AVX512 硬件需要对齐才能获得最佳性能。我假设代码生成器只是以相同的方式处理所有这些情况。

\n

由于我们有两个输入和一个输出，因此我们只能真正与其中之一对齐，然后希望其他输入也同样对齐。如果我正确地读取了代码，则对齐是在输出上执行的，而不是在输入上执行的，但我可能读错了。

\n

L0002循环：主循环展开 4 次，每次迭代 4 x 8 = 32 个元素。

\n

L0003循环：与主循环相同，但未展开以处理最后 0-3 个完整向量。

\n

L0005循环：将最后 0-7 个元素作为标量处理。有趣的是，使用 16 字节 XMM 寄存器进行一次迭代可以节省 4 次迭代，但编译器选择不这样做。我猜他们认为这不值得。可以在循环中完成相同的操作L0000。

\n

编译器知道输出不会与输入重叠。否则，您会期望另一个单独的循环来处理两者重叠得如此紧密以致于无法进行矢量化的情况。这可能只是一个单独的条件跳转到L0005循环。

\n

对齐循环计数器

\n

让我们尝试解开对齐环L0000。它的循环计数器%r11d（在循环之前立即设置为 0）作为%r9d其循环限制。那么让我们回溯%r9d一下。

\n

0x000000010c4c5670: cmp %ebp,%r9d\n0x000000010c4c5673: cmovg %ebp,%r9d\n

\n

我们之前成立的%ebp，arraySize所以这基本上是r9d = min(r9d, arraySize)。说得通。在一个小的、未对齐的数组中，L0000循环处理全部内容。

\n

在那之前：

\n

0x000000010c4c5649: lea (%r12,%r8,8),%rdx\n0x000000010c4c5651: mov %edx,%r11d\n0x000000010c4c5658: shr $0x2,%r11d\n0x000000010c4c565c: and $0x7,%r11d\n0x000000010c4c5660: mov $0x3,%r9d\n0x000000010c4c5666: sub %r11d,%r9d\n0x000000010c4c5669: and $0x7,%r9d\n0x000000010c4c566d: inc %r9d\n

\n

r12从未在该代码片段中初始化。这个问题表明它持有堆基数。再往上走，我们找到了这条线0x000000010c4c5629: mov 0x18(%rsi),%r8d ;*getfield result。因此，我们假设lea (%r12,%r8,8),%rdx给出了结果数组的起始地址。

\n

但是，该地址似乎不是数据内容的开始。稍后在循环中您会发现所有内存访问都遵循该模式vmovss %xmm0,0x10(%rdx,%r11,4)。注意额外的偏移量0x10 = 16。我们可以猜测数组的前 16 个字节保存元数据，例如其大小。该偏移量被合并到后续的每个计算中。啊。

\n

除以shr $0x24，从字节到浮点索引。and $0x7给出距最后一个 8 的倍数的偏移量。在 C 代码中，这将是

\n

0x000000010c4c5670: cmp %ebp,%r9d\n0x000000010c4c5673: cmovg %ebp,%r9d\n

\n

然后我们计算%r9d：

\n

0x000000010c4c5649: lea (%r12,%r8,8),%rdx\n0x000000010c4c5651: mov %edx,%r11d\n0x000000010c4c5658: shr $0x2,%r11d\n0x000000010c4c565c: and $0x7,%r11d\n0x000000010c4c5660: mov $0x3,%r9d\n0x000000010c4c5666: sub %r11d,%r9d\n0x000000010c4c5669: and $0x7,%r9d\n0x000000010c4c566d: inc %r9d\n

\n

我们已经确定该%r11d值在 0 到 7 之间，因此我们可以只查看数字。

\n

| r11d |  r9d |\n|   0  |   4  |\n|   1  |   3  |\n|   2  |   2  |\n|   3  |   1  |\n|   4  |   8  |\n|   5  |   7  |\n|   6  |   6  |\n|   7  |   5  |\n

\n

只有当我们记住所有内存访问都被 4 个额外浮点偏移时，这才有意义。假设%r11d为 1，表示数组头在最后 32 字节对齐之后从 1 float = 4 字节开始。对齐后，数组内容从 1 + 4 = 5 float = 20 字节开始。然后我们进行 3 次标量迭代以获得下一个 32 字节对齐。

\n

我对这段代码遇到的一个问题是，对于r11d = 4，我们不应该进行任何标量迭代。我们可以直接进入矢量化代码。额外的and $0x7,r9d可以解决这个问题。但另一方面，这节省了条件分支指令，因为循环总是至少执行一次。

\n

后续问题

\n

\n

所以“循环对齐”不是用于循环控制的“i”的对齐，它与缓存行大小的对齐？

\n

\n

不是缓存行大小，而是自然向量大小。此代码使用 AVX 向量（YMM 寄存器）。它们的大小是 32 字节，因此我们尝试获得 32 字节对齐。缓存行为 64 字节。但请注意 32 字节向量的 32 字节对齐访问永远不会跨越缓存行边界。这是主要目标。

\n

\n

如果是这样，您知道确定迭代次数的伪算法吗？我会假设只是（地址％8），但似乎L0000之前代码中唯一的“lea”没有找到进入r9d或r11d的方式，它们用于循环控制

\n

\n

我想我上面已经说得很清楚了。然而，一般来说，用 C 语言编写时，您可以这样写：

\n

struct Array {\n   int64_t meta1, meta2;\n   float content[];\n};\nArray* result = heap_base + offset_of_result;\nunsigned r11d = (unsigned) result / sizeof(float) % 8;\n

\n

有更短的版本，但这应该是最容易理解的。需要注意的关键点是，我们必须将数组指针重新解释为整数才能使其工作，这就是我使用 C 而不是 Java 的原因。

\n

更短、更优雅的版本可能是

\n

unsigned r9d = (3 - r11d) % 8 + 1;\n

\n