auto const p = std::unique_ptr(new unsigned char[1024 * 50] { 0 });
link.exe
зависает, кушая при этом процессор.
auto const p = std::unique_ptr(new unsigned char[1024 * 50] { 0 });
Результат: fatal error C1063: compiler limit : compiler stack overflow
auto const p = std::unique_ptr(new unsigned char[1024] { 0 });
Сгенерированные инструкции:
00007FF6D1CF1A3A call operator new[] (07FF6D1CF1E30h)
00007FF6D1CF1A3F xor ecx,ecx
00007FF6D1CF1A41 test rax,rax
00007FF6D1CF1A44 je wmain+383h (07FF6D1CF1D9Bh)
00007FF6D1CF1A4A mov qword ptr [rax],rcx
00007FF6D1CF1A4D mov qword ptr [rax+8],rcx
00007FF6D1CF1A51 mov qword ptr [rax+10h],rcx
00007FF6D1CF1A55 mov qword ptr [rax+18h],rcx
00007FF6D1CF1A59 mov qword ptr [rax+20h],rcx
00007FF6D1CF1A5D mov qword ptr [rax+28h],rcx
00007FF6D1CF1A61 mov qword ptr [rax+30h],rcx
00007FF6D1CF1A65 mov qword ptr [rax+38h],rcx
00007FF6D1CF1A69 mov qword ptr [rax+40h],rcx
00007FF6D1CF1A6D mov qword ptr [rax+48h],rcx
00007FF6D1CF1A71 mov qword ptr [rax+50h],rcx
00007FF6D1CF1A75 mov qword ptr [rax+58h],rcx
00007FF6D1CF1A79 mov qword ptr [rax+60h],rcx
00007FF6D1CF1A7D mov qword ptr [rax+68h],rcx
00007FF6D1CF1A81 mov qword ptr [rax+70h],rcx
00007FF6D1CF1A85 mov qword ptr [rax+78h],rcx
00007FF6D1CF1A89 mov qword ptr [rax+80h],rcx
...
Заполнение пасяти нулями шмомпилятор от Быдлософт развернул в 128 инструкций mov
.
То же, но компилятором Intel C++:
00007FF755B1102B call operator new[] (07FF755B13C50h)
...
00007FF755B11053 call _intel_fast_memcpy (07FF755B11B50h)
_intel_fast_memcpy
для моего Core i5 2XXX выбрала реализацию с циклом следующего вида:
00007FF755B132A0 movdqa xmm0,xmmword ptr [rdx]
00007FF755B132A4 movdqa xmm1,xmmword ptr [rdx+10h]
00007FF755B132A9 movdqa xmmword ptr [rcx],xmm0
00007FF755B132AD movdqa xmmword ptr [rcx+10h],xmm1
00007FF755B132B2 lea r8,[r8-80h]
00007FF755B132B6 movdqa xmm2,xmmword ptr [rdx+20h]
00007FF755B132BB movdqa xmm3,xmmword ptr [rdx+30h]
00007FF755B132C0 movdqa xmmword ptr [rcx+20h],xmm2
00007FF755B132C5 movdqa xmmword ptr [rcx+30h],xmm3
00007FF755B132CA movdqa xmm0,xmmword ptr [rdx+40h]
00007FF755B132CF movdqa xmm1,xmmword ptr [rdx+50h]
00007FF755B132D4 cmp r8,0A8h
00007FF755B132DB movdqa xmmword ptr [rcx+40h],xmm0
00007FF755B132E0 movdqa xmmword ptr [rcx+50h],xmm1
00007FF755B132E5 movdqa xmm2,xmmword ptr [rdx+60h]
00007FF755B132EA movdqa xmm3,xmmword ptr [rdx+70h]
00007FF755B132EF lea rdx,[rdx+80h]
00007FF755B132F6 movdqa xmmword ptr [rcx+60h],xmm2
00007FF755B132FB movdqa xmmword ptr [rcx+70h],xmm3
00007FF755B13300 lea rcx,[rcx+80h]
00007FF755B13307 jge __intel_memcpy+0E90h (07FF755B132A0h)
А если выбрать Favor Small Code в настройках компилятора, то заполнение нулями превращается в ожидаемый и привычный rep movs
:
...
00007FF615041053 rep movs qword ptr [rdi],qword ptr [rsi]