我一直在阅读有关由于英特尔 CPU 上地址位 6 到 11 的歧义导致加载/存储重叠导致的 4K 别名。所以我正在尝试编写各种简单的测试(在 i7-3770k、Win7、64 位、VS2017 上)来专门导致问题，以确保我在实践中理解它。
我一直在尝试但未能表现出行为的第一个测试是:
void Test4KAliasing1()
{
typedef float Value;// Also tried with double
const uint32_t ValueCount = 1024;
const uint32_t OffsetCount = 256;
const uint32_t TestCount = 512;
Value* a = (Value*)_aligned_malloc(ValueCount * sizeof(Value), 4096);
Value* b = (Value*)_aligned_malloc(ValueCount * sizeof(Value), 4096);
for (uint32_t i = 0; i < ValueCount; ++i)
a[i] = b[i] = (Value)rand();
for (uint32_t offset = 0; offset < OffsetCount; ++offset)
{
uint64_t startTime = StartCPUCycles();
for (uint32_t test = 0; test < TestCount; ++test)
{
for (uint32_t i = 0; i < ValueCount; ++i)
{
uint32_t j = (offset + i) % ValueCount;
a[i] += b[j] * 3.142f;
}
}
uint64_t duration = EndCPUCycles() - startTime;
printf("time: %llu\toffset: %u ", duration / TestCount, offset);
printf("\n", a, b);
}
_aligned_free(b);
_aligned_free(a);
}
灵感来自:http://richardstartin.uk/the-much-aligned-garbage-collector/
所以我不太确定为什么这没有从结果时间显示问题？正如我所认为的那样，由于乱序执行，存储然后在循环迭代中会发生到/从不明确的地址加载？
生成的程序集是:
000000013F2510E4 cpuid
000000013F2510E6 rdtsc
000000013F2510E8 shl rdx,20h
000000013F2510EC mov r9d,200h
000000013F2510F2 or rax,rdx
000000013F2510F5 mov r10,rax
000000013F2510F8 nop dword ptr [rax+rax]
000000013F251100 lea ebx,[rsi+1]
000000013F251103 mov r8d,80h
000000013F251109 lea rdx,[r14+8]
000000013F25110D nop dword ptr [rax]
000000013F251110 mov rax,rbx
000000013F251113 lea ecx,[rbx-1]
000000013F251116 and eax,3FFh
000000013F25111B lea rdx,[rdx+20h]
000000013F25111F and ecx,3FFh
000000013F251125 vmulss xmm1,xmm6,dword ptr [rdi+rcx*4]
000000013F25112A vaddss xmm2,xmm1,dword ptr [rdx-28h]
000000013F25112F vmovss dword ptr [rdx-28h],xmm2
000000013F251134 vmulss xmm1,xmm6,dword ptr [rdi+rax*4]
000000013F251139 vaddss xmm2,xmm1,dword ptr [rdx-24h]
000000013F25113E vmovss dword ptr [rdx-24h],xmm2
000000013F251143 lea eax,[rbx+1]
000000013F251146 and eax,3FFh
000000013F25114B vmulss xmm1,xmm6,dword ptr [rdi+rax*4]
000000013F251150 vaddss xmm2,xmm1,dword ptr [rdx-20h]
000000013F251155 vmovss dword ptr [rdx-20h],xmm2
000000013F25115A lea eax,[rbx+2]
000000013F25115D and eax,3FFh
000000013F251162 vmulss xmm1,xmm6,dword ptr [rdi+rax*4]
000000013F251167 vaddss xmm2,xmm1,dword ptr [rdx-1Ch]
000000013F25116C vmovss dword ptr [rdx-1Ch],xmm2
000000013F251171 lea eax,[rbx+3]
000000013F251174 and eax,3FFh
000000013F251179 vmulss xmm1,xmm6,dword ptr [rdi+rax*4]
000000013F25117E vaddss xmm2,xmm1,dword ptr [rdx-18h]
000000013F251183 vmovss dword ptr [rdx-18h],xmm2
000000013F251188 lea eax,[rbx+4]
000000013F25118B and eax,3FFh
000000013F251190 vmulss xmm1,xmm6,dword ptr [rdi+rax*4]
000000013F251195 vaddss xmm2,xmm1,dword ptr [rdx-14h]
000000013F25119A vmovss dword ptr [rdx-14h],xmm2
000000013F25119F lea eax,[rbx+5]
000000013F2511A2 and eax,3FFh
000000013F2511A7 vmulss xmm1,xmm6,dword ptr [rdi+rax*4]
000000013F2511AC vaddss xmm2,xmm1,dword ptr [rdx-10h]
000000013F2511B1 lea eax,[rbx+6]
000000013F2511B4 add ebx,8
000000013F2511B7 vmovss dword ptr [rdx-10h],xmm2
000000013F2511BC and eax,3FFh
000000013F2511C1 vmulss xmm1,xmm6,dword ptr [rdi+rax*4]
000000013F2511C6 vaddss xmm2,xmm1,dword ptr [rdx-0Ch]
000000013F2511CB vmovss dword ptr [rdx-0Ch],xmm2
000000013F2511D0 sub r8,1
000000013F2511D4 jne Test4KAliasing1+0B0h (013F251110h)
000000013F2511DA sub r9,1
000000013F2511DE jne Test4KAliasing1+0A0h (013F251100h)
000000013F2511E4 rdtsc
同样在网络上，我看到了各种描述，说底部 12 位必须匹配才能发生这种混叠，而在其他地方只有 6 到 11 位？由于最低 6 位是缓存行中的字节索引，而且一切都是基于缓存行的，那么我会认为它只需要第 6 位到第 11 位来匹配？
编辑:
同样根据彼得斯的回答，我尝试过:
a[i] *= 1.234f;
b[j] += 4.321f;
这似乎没有显示问题并产生:
000000013F6C10E8 cpuid
000000013F6C10EA rdtsc
000000013F6C10EC shl rdx,20h
000000013F6C10F0 mov ebx,200h
000000013F6C10F5 or rax,rdx
000000013F6C10F8 mov r9,rax
000000013F6C10FB nop dword ptr [rax+rax]
000000013F6C1100 lea edx,[rsi+1]
000000013F6C1103 mov r8d,80h
000000013F6C1109 lea rcx,[r14+8]
000000013F6C110D nop dword ptr [rax]
000000013F6C1110 vmulss xmm1,xmm6,dword ptr [rcx-8]
000000013F6C1115 vmovss dword ptr [rcx-8],xmm1
000000013F6C111A lea eax,[rdx-1]
000000013F6C111D and eax,3FFh
000000013F6C1122 lea rcx,[rcx+20h]
000000013F6C1126 vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C112B vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C1130 vmulss xmm1,xmm6,dword ptr [rcx-24h]
000000013F6C1135 vmovss dword ptr [rcx-24h],xmm1
000000013F6C113A mov rax,rdx
000000013F6C113D and eax,3FFh
000000013F6C1142 vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C1147 vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C114C vmulss xmm0,xmm6,dword ptr [rcx-20h]
000000013F6C1151 lea eax,[rdx+1]
000000013F6C1154 and eax,3FFh
000000013F6C1159 vmovss dword ptr [rcx-20h],xmm0
000000013F6C115E vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C1163 vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C1168 vmulss xmm1,xmm6,dword ptr [rcx-1Ch]
000000013F6C116D vmovss dword ptr [rcx-1Ch],xmm1
000000013F6C1172 lea eax,[rdx+2]
000000013F6C1175 and eax,3FFh
000000013F6C117A vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C117F vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C1184 vmulss xmm1,xmm6,dword ptr [rcx-18h]
000000013F6C1189 vmovss dword ptr [rcx-18h],xmm1
000000013F6C118E lea eax,[rdx+3]
000000013F6C1191 and eax,3FFh
000000013F6C1196 vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C119B vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C11A0 vmulss xmm1,xmm6,dword ptr [rcx-14h]
000000013F6C11A5 vmovss dword ptr [rcx-14h],xmm1
000000013F6C11AA lea eax,[rdx+4]
000000013F6C11AD and eax,3FFh
000000013F6C11B2 vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C11B7 vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C11BC vmulss xmm1,xmm6,dword ptr [rcx-10h]
000000013F6C11C1 lea eax,[rdx+5]
000000013F6C11C4 and eax,3FFh
000000013F6C11C9 vmovss dword ptr [rcx-10h],xmm1
000000013F6C11CE vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C11D3 vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C11D8 vmulss xmm1,xmm6,dword ptr [rcx-0Ch]
000000013F6C11DD lea eax,[rdx+6]
000000013F6C11E0 add edx,8
000000013F6C11E3 and eax,3FFh
000000013F6C11E8 vmovss dword ptr [rcx-0Ch],xmm1
000000013F6C11ED vaddss xmm1,xmm7,dword ptr [rdi+rax*4]
000000013F6C11F2 vmovss dword ptr [rdi+rax*4],xmm1
000000013F6C11F7 sub r8,1
000000013F6C11FB jne Test4KAliasing1+0B0h (013F6C1110h)
000000013F6C1201 sub rbx,1
000000013F6C1205 jne Test4KAliasing1+0A0h (013F6C1100h)
000000013F6C120B rdtsc
同样基于彼得提到的链接问题，我尝试了 3 个数组:
a[i] += b[j] + c[j];
这似乎也没有问题。生成的代码是:
000000013F5110F6 cpuid
000000013F5110F8 rdtsc
000000013F5110FA shl rdx,20h
000000013F5110FE mov r8d,200h
000000013F511104 or rax,rdx
000000013F511107 mov r10,rax
000000013F51110A nop word ptr [rax+rax]
000000013F511110 lea ebx,[rbp+1]
000000013F511113 mov r9d,100h
000000013F511119 lea rdx,[r13+8]
000000013F51111D nop dword ptr [rax]
000000013F511120 mov rax,rbx
000000013F511123 lea ecx,[rbx-1]
000000013F511126 and eax,7FFh
000000013F51112B lea rdx,[rdx+20h]
000000013F51112F and ecx,7FFh
000000013F511135 vmovss xmm0,dword ptr [rsi+rcx*4]
000000013F51113A vaddss xmm1,xmm0,dword ptr [rdi+rcx*4]
000000013F51113F vaddss xmm2,xmm1,dword ptr [rdx-28h]
000000013F511144 vmovss dword ptr [rdx-28h],xmm2
000000013F511149 vmovss xmm0,dword ptr [rsi+rax*4]
000000013F51114E vaddss xmm1,xmm0,dword ptr [rdi+rax*4]
000000013F511153 vaddss xmm2,xmm1,dword ptr [rdx-24h]
000000013F511158 vmovss dword ptr [rdx-24h],xmm2
000000013F51115D lea eax,[rbx+1]
000000013F511160 and eax,7FFh
000000013F511165 vmovss xmm0,dword ptr [rsi+rax*4]
000000013F51116A vaddss xmm1,xmm0,dword ptr [rdi+rax*4]
000000013F51116F vaddss xmm2,xmm1,dword ptr [rdx-20h]
000000013F511174 vmovss dword ptr [rdx-20h],xmm2
000000013F511179 lea eax,[rbx+2]
000000013F51117C and eax,7FFh
000000013F511181 vmovss xmm0,dword ptr [rsi+rax*4]
000000013F511186 vaddss xmm1,xmm0,dword ptr [rdi+rax*4]
000000013F51118B vaddss xmm2,xmm1,dword ptr [rdx-1Ch]
000000013F511190 vmovss dword ptr [rdx-1Ch],xmm2
000000013F511195 lea eax,[rbx+3]
000000013F511198 and eax,7FFh
000000013F51119D vmovss xmm0,dword ptr [rsi+rax*4]
000000013F5111A2 vaddss xmm1,xmm0,dword ptr [rdi+rax*4]
000000013F5111A7 vaddss xmm2,xmm1,dword ptr [rdx-18h]
000000013F5111AC vmovss dword ptr [rdx-18h],xmm2
000000013F5111B1 lea eax,[rbx+4]
000000013F5111B4 and eax,7FFh
000000013F5111B9 vmovss xmm0,dword ptr [rsi+rax*4]
000000013F5111BE vaddss xmm1,xmm0,dword ptr [rdi+rax*4]
000000013F5111C3 vaddss xmm2,xmm1,dword ptr [rdx-14h]
000000013F5111C8 vmovss dword ptr [rdx-14h],xmm2
000000013F5111CD lea eax,[rbx+5]
000000013F5111D0 and eax,7FFh
000000013F5111D5 vmovss xmm0,dword ptr [rsi+rax*4]
000000013F5111DA vaddss xmm1,xmm0,dword ptr [rdi+rax*4]
000000013F5111DF vaddss xmm2,xmm1,dword ptr [rdx-10h]
000000013F5111E4 lea eax,[rbx+6]
000000013F5111E7 add ebx,8
000000013F5111EA vmovss dword ptr [rdx-10h],xmm2
000000013F5111EF and eax,7FFh
000000013F5111F4 vmovss xmm0,dword ptr [rsi+rax*4]
000000013F5111F9 vaddss xmm1,xmm0,dword ptr [rdi+rax*4]
000000013F5111FE vaddss xmm2,xmm1,dword ptr [rdx-0Ch]
000000013F511203 vmovss dword ptr [rdx-0Ch],xmm2
000000013F511208 sub r9,1
000000013F51120C jne Test4KAliasing2+0C0h (013F511120h)
000000013F511212 sub r8,1
000000013F511216 jne Test4KAliasing2+0B0h (013F511110h)
000000013F51121C rdtsc
继彼得对他的回答的评论/更新之后，我尝试了:
a[i] *= 1.234f;
b[i] += 4.321f;
哪个没有显示问题。注意:我试图改变 i 的偏移量， j = i + 偏移量从零偏移量开始，对于大多数这些尝试，如果我能找到它，看看什么偏移量可以缓解问题。 (由于我的 x86 已经生锈了，我仍然在这里挖掘反汇编以了解地址生成)。
000000013F7D1104 cpuid
000000013F7D1106 rdtsc
000000013F7D1108 shl rdx,20h
000000013F7D110C or rax,rdx
000000013F7D110F mov edx,200h
000000013F7D1114 mov rbx,rax
000000013F7D1117 cmp rsi,r15
000000013F7D111A ja Test4KAliasing1+130h (013F7D1190h)
000000013F7D111C cmp rbp,r14
000000013F7D111F jb Test4KAliasing1+130h (013F7D1190h)
000000013F7D1121 lea rcx,[rsi+4]
000000013F7D1125 mov eax,100h
000000013F7D112A nop word ptr [rax+rax]
000000013F7D1130 vmulss xmm1,xmm6,dword ptr [rdi+rcx-4]
000000013F7D1136 vmovss dword ptr [rdi+rcx-4],xmm1
000000013F7D113C vaddss xmm1,xmm7,dword ptr [rcx-4]
000000013F7D1141 vmovss dword ptr [rcx-4],xmm1
000000013F7D1146 vmulss xmm1,xmm6,dword ptr [rcx+rdi]
000000013F7D114B vmovss dword ptr [rcx+rdi],xmm1
000000013F7D1150 vaddss xmm0,xmm7,dword ptr [rcx]
000000013F7D1154 vmovss dword ptr [rcx],xmm0
000000013F7D1158 vmulss xmm0,xmm6,dword ptr [rdi+rcx+4]
000000013F7D115E vmovss dword ptr [rdi+rcx+4],xmm0
000000013F7D1164 vaddss xmm0,xmm7,dword ptr [rcx+4]
000000013F7D1169 vmovss dword ptr [rcx+4],xmm0
000000013F7D116E vmulss xmm0,xmm6,dword ptr [rdi+rcx+8]
000000013F7D1174 vmovss dword ptr [rdi+rcx+8],xmm0
000000013F7D117A vaddss xmm0,xmm7,dword ptr [rcx+8]
000000013F7D117F vmovss dword ptr [rcx+8],xmm0
000000013F7D1184 lea rcx,[rcx+10h]
000000013F7D1188 sub rax,1
000000013F7D118C jne Test4KAliasing1+0D0h (013F7D1130h)
000000013F7D118E jmp Test4KAliasing1+1AEh (013F7D120Eh)
000000013F7D1190 vmovups xmm2,xmmword ptr [__xmm@3f9df3b63f9df3b63f9df3b63f9df3b6 (013F7EA0E0h)]
000000013F7D1198 vmovups xmm3,xmmword ptr [__xmm@408a45a2408a45a2408a45a2408a45a2 (013F7EA0F0h)]
000000013F7D11A0 lea rax,[rsi+10h]
000000013F7D11A4 mov ecx,40h
000000013F7D11A9 nop dword ptr [rax]
000000013F7D11B0 vmulps xmm1,xmm2,xmmword ptr [rdi+rax-10h]
000000013F7D11B6 vmovups xmmword ptr [rdi+rax-10h],xmm1
000000013F7D11BC vaddps xmm1,xmm3,xmmword ptr [rax-10h]
000000013F7D11C1 vmovups xmmword ptr [rax-10h],xmm1
000000013F7D11C6 vmulps xmm1,xmm2,xmmword ptr [rdi+rax]
000000013F7D11CB vmovups xmmword ptr [rdi+rax],xmm1
000000013F7D11D0 vaddps xmm1,xmm3,xmmword ptr [rax]
000000013F7D11D4 vmovups xmmword ptr [rax],xmm1
000000013F7D11D8 vmulps xmm1,xmm2,xmmword ptr [rdi+rax+10h]
000000013F7D11DE vmovups xmmword ptr [rdi+rax+10h],xmm1
000000013F7D11E4 vaddps xmm1,xmm3,xmmword ptr [rax+10h]
000000013F7D11E9 vmovups xmmword ptr [rax+10h],xmm1
000000013F7D11EE vmulps xmm1,xmm2,xmmword ptr [rdi+rax+20h]
000000013F7D11F4 vmovups xmmword ptr [rdi+rax+20h],xmm1
000000013F7D11FA vaddps xmm1,xmm3,xmmword ptr [rax+20h]
000000013F7D11FF vmovups xmmword ptr [rax+20h],xmm1
000000013F7D1204 lea rax,[rax+40h]
000000013F7D1208 sub rcx,1
000000013F7D120C jne Test4KAliasing1+150h (013F7D11B0h)
000000013F7D120E sub rdx,1
000000013F7D1212 jne Test4KAliasing1+0B7h (013F7D1117h)
000000013F7D1218 rdtsc
典型的计时运行:
a[i] *= 1.234f;
b[i] += 4.321f;
是:
time: 715 offset: 0
time: 647 offset: 1
time: 641 offset: 2
time: 703 offset: 3
time: 658 offset: 4
time: 657 offset: 5
time: 656 offset: 6
time: 657 offset: 7
time: 658 offset: 8
time: 657 offset: 9
time: 658 offset: 10
time: 653 offset: 11
time: 658 offset: 12
time: 652 offset: 13
time: 658 offset: 14
time: 657 offset: 15
time: 658 offset: 16
time: 656 offset: 17
time: 659 offset: 18
time: 656 offset: 19
time: 656 offset: 20
time: 656 offset: 21
time: 663 offset: 22
time: 657 offset: 23
time: 657 offset: 24
time: 704 offset: 25
time: 714 offset: 26
time: 657 offset: 27
time: 658 offset: 28
time: 658 offset: 29
time: 656 offset: 30
time: 656 offset: 31
time: 657 offset: 32
time: 658 offset: 33
time: 658 offset: 34
time: 656 offset: 35
time: 658 offset: 36
time: 658 offset: 37
time: 658 offset: 38
time: 658 offset: 39
time: 660 offset: 40
time: 660 offset: 41
time: 664 offset: 42
time: 656 offset: 43
time: 656 offset: 44
time: 658 offset: 45
time: 656 offset: 46
time: 656 offset: 47
time: 713 offset: 48
time: 658 offset: 49
time: 663 offset: 50
time: 662 offset: 51
time: 665 offset: 52
time: 663 offset: 53
time: 665 offset: 54
time: 658 offset: 55
time: 658 offset: 56
time: 658 offset: 57
time: 656 offset: 58
time: 657 offset: 59
time: 658 offset: 60
time: 658 offset: 61
time: 656 offset: 62
time: 666 offset: 63
time: 656 offset: 64
time: 658 offset: 65
time: 656 offset: 66
time: 657 offset: 67
time: 658 offset: 68
time: 658 offset: 69
time: 652 offset: 70
time: 658 offset: 71
time: 657 offset: 72
time: 658 offset: 73
time: 658 offset: 74
time: 656 offset: 75
time: 658 offset: 76
time: 665 offset: 77
time: 657 offset: 78
time: 656 offset: 79
time: 656 offset: 80
time: 666 offset: 81
time: 656 offset: 82
time: 702 offset: 83
time: 640 offset: 84
time: 640 offset: 85
time: 657 offset: 86
time: 657 offset: 87
time: 658 offset: 88
time: 658 offset: 89
time: 656 offset: 90
time: 657 offset: 91
time: 657 offset: 92
time: 657 offset: 93
time: 658 offset: 94
time: 662 offset: 95
time: 658 offset: 96
time: 656 offset: 97
time: 657 offset: 98
time: 663 offset: 99
time: 660 offset: 100
time: 663 offset: 101
time: 657 offset: 102
time: 656 offset: 103
time: 664 offset: 104
time: 659 offset: 105
time: 659 offset: 106
time: 658 offset: 107
time: 774 offset: 108
time: 707 offset: 109
time: 710 offset: 110
time: 658 offset: 111
time: 657 offset: 112
time: 661 offset: 113
time: 658 offset: 114
time: 656 offset: 115
time: 658 offset: 116
time: 657 offset: 117
time: 658 offset: 118
time: 660 offset: 119
time: 666 offset: 120
time: 657 offset: 121
time: 658 offset: 122
time: 651 offset: 123
time: 658 offset: 124
time: 657 offset: 125
time: 657 offset: 126
time: 658 offset: 127
time: 656 offset: 128
time: 658 offset: 129
time: 656 offset: 130
time: 658 offset: 131
time: 645 offset: 132
time: 640 offset: 133
time: 640 offset: 134
time: 659 offset: 135
time: 664 offset: 136
time: 658 offset: 137
time: 662 offset: 138
time: 656 offset: 139
time: 658 offset: 140
time: 656 offset: 141
time: 658 offset: 142
time: 660 offset: 143
time: 658 offset: 144
time: 658 offset: 145
time: 656 offset: 146
time: 657 offset: 147
time: 664 offset: 148
time: 656 offset: 149
time: 656 offset: 150
time: 658 offset: 151
time: 656 offset: 152
time: 668 offset: 153
time: 656 offset: 154
time: 656 offset: 155
time: 656 offset: 156
time: 658 offset: 157
time: 656 offset: 158
time: 658 offset: 159
time: 660 offset: 160
time: 658 offset: 161
time: 658 offset: 162
time: 658 offset: 163
time: 658 offset: 164
time: 656 offset: 165
time: 686 offset: 166
time: 656 offset: 167
time: 656 offset: 168
time: 658 offset: 169
time: 656 offset: 170
time: 658 offset: 171
time: 656 offset: 172
time: 656 offset: 173
time: 656 offset: 174
time: 658 offset: 175
time: 656 offset: 176
time: 658 offset: 177
time: 658 offset: 178
time: 654 offset: 179
time: 639 offset: 180
time: 639 offset: 181
time: 639 offset: 182
time: 657 offset: 183
time: 641 offset: 184
time: 640 offset: 185
time: 640 offset: 186
time: 640 offset: 187
time: 640 offset: 188
time: 640 offset: 189
time: 640 offset: 190
time: 700 offset: 191
time: 715 offset: 192
time: 657 offset: 193
time: 657 offset: 194
time: 662 offset: 195
time: 703 offset: 196
time: 640 offset: 197
time: 639 offset: 198
time: 638 offset: 199
time: 640 offset: 200
time: 640 offset: 201
time: 640 offset: 202
time: 704 offset: 203
time: 638 offset: 204
time: 640 offset: 205
time: 639 offset: 206
time: 657 offset: 207
time: 658 offset: 208
time: 657 offset: 209
time: 659 offset: 210
time: 663 offset: 211
time: 658 offset: 212
time: 658 offset: 213
time: 657 offset: 214
time: 667 offset: 215
time: 657 offset: 216
time: 657 offset: 217
time: 658 offset: 218
time: 657 offset: 219
time: 656 offset: 220
time: 661 offset: 221
time: 651 offset: 222
time: 658 offset: 223
time: 658 offset: 224
time: 656 offset: 225
time: 658 offset: 226
time: 658 offset: 227
time: 672 offset: 228
time: 658 offset: 229
time: 656 offset: 230
time: 649 offset: 231
time: 665 offset: 232
time: 657 offset: 233
time: 652 offset: 234
time: 664 offset: 235
time: 656 offset: 236
time: 662 offset: 237
time: 658 offset: 238
time: 665 offset: 239
time: 658 offset: 240
time: 657 offset: 241
time: 656 offset: 242
time: 658 offset: 243
time: 657 offset: 244
time: 658 offset: 245
time: 658 offset: 246
time: 656 offset: 247
time: 658 offset: 248
time: 656 offset: 249
time: 658 offset: 250
time: 656 offset: 251
time: 665 offset: 252
time: 658 offset: 253
time: 656 offset: 254
time: 658 offset: 255
但是:我想我犯了一个错误，现在发现它:
a[i] *= 1.234f;
b[j] += 4.321f;
作为一个典型的计时运行现在是:
time: 2794 offset: 0
time: 2737 offset: 1
time: 2655 offset: 2
time: 2748 offset: 3
time: 2605 offset: 4
time: 2730 offset: 5
time: 2665 offset: 6
time: 2703 offset: 7
time: 2571 offset: 8
time: 2558 offset: 9
time: 2213 offset: 10
time: 2200 offset: 11
time: 2325 offset: 12
time: 2200 offset: 13
time: 2200 offset: 14
time: 2264 offset: 15
time: 2264 offset: 16
time: 2355 offset: 17
time: 2348 offset: 18
time: 2262 offset: 19
time: 2260 offset: 20
time: 2262 offset: 21
time: 2260 offset: 22
time: 2490 offset: 23
time: 2261 offset: 24
time: 2260 offset: 25
time: 2255 offset: 26
time: 2261 offset: 27
time: 2263 offset: 28
time: 2260 offset: 29
time: 2260 offset: 30
time: 2262 offset: 31
time: 2264 offset: 32
time: 2355 offset: 33
time: 2266 offset: 34
time: 2270 offset: 35
time: 2260 offset: 36
time: 2268 offset: 37
time: 2260 offset: 38
time: 2260 offset: 39
time: 2262 offset: 40
time: 2260 offset: 41
time: 2259 offset: 42
time: 2260 offset: 43
time: 2260 offset: 44
time: 2255 offset: 45
time: 2260 offset: 46
time: 2265 offset: 47
time: 2263 offset: 48
time: 2355 offset: 49
time: 2293 offset: 50
time: 2204 offset: 51
time: 2323 offset: 52
time: 2200 offset: 53
time: 2200 offset: 54
time: 2460 offset: 55
time: 2200 offset: 56
随着偏移量变大，这可能是大约 20% 的差异吗？
最佳答案
底部 12 位是位 [11 : 0] 。第 11 位是第 12 位，因为我们从 0 开始计数。
CPU 以字节粒度检测加载/存储别名，而不仅仅是加载是否访问与旧存储相同的缓存行。存储到 array[1] 不会减慢 array[2] 的加载速度；这对性能来说真的很糟糕，因为循环遍历数组并一次对每个元素进行 RMW 是一种非常常见的模式。 (没有软件流水线来在我们存储的位置之前加载几个元素。)
所以我认为你在这里没有遇到问题，因为你只是在从 4k 页面内的相同偏移量加载后存储到一个位置。 如果你做了类似这个简单循环的事情(不需要额外的跨步或偏移到另一个额外的页面，不同页面中的两个数组就可以了。)
for (i = 0 ; i < limit ; i++) {
a[i] *= 1.234;
b[i] += 4.321; // load from the same offset we just wrote, but in another page
}
并且编译器在加载
a 之前制作了存储到 b 的 asm ，您会遇到问题，因为 a 和 b 相对于 4k 页面具有相同的对齐方式。(编译器可以在存储之前执行两次加载，如果它证明了
a != b ，或者在运行执行该操作的循环版本之前发出代码进行检查。或者使用自动矢量化和/或展开，如果它检查了矢量宽度的重叠乘以展开因子。)这不是一个完美的例子，但是使来自
b 的加载依赖于来自 a 的存储应该使乱序执行至少努力隐藏那么多延迟。创建 4k 别名的另一种简单方法是将 memcpy 从
src = srcpage 转换为 dst = dstpage + 16 或 ，当然 srcpage 和 dstpage 都是页面对齐的。存储到 dst[i] 就像 dstpage[i+16] (以字节为单位，而不是任何 C 元素大小)，因此存储 dst[i] 将在从 src[i+16] 加载之前(按程序顺序)发生。当循环到达该 i 值时，负载将被 4k 别名阻塞。参见 L1 memory bandwidth: 50% drop in efficiency using addresses which differ by 4096+64 bytes 示例，@HadiBrais 对包括 IvyBridge(如 i7-3770k)在内的 CPU 进行了性能分析。
关于performance - 了解英特尔 CPU 上的 4K 锯齿，我们在Stack Overflow上找到一个类似的问题：https://stackoverflow.com/questions/54415140/