Many instructions going into the CPU don’t have to be executed. Take ADD B, which takes one cycle, and does A += B. Compare the cost in python vs C.
With optimizations off, this c program turns into:
int __declspec(noinline) add(int A, int B) {
return A + B;
}
#pragma optimize("", off)
int main(int ArgCount, char **Args) {
return add(1234, 5678);
}This asm:
lea eax, [rcx + rdx]In python it actually takes 181 instructions:
00007FFDFD04F40F mov qword ptr [r11+38h],r14
00007FFDFD04F413 add r14,2
00007FFDFD04F417 jmp _PyEval_EvalFrameDefault+5872h (07FFDFD04F462h)
00007FFDFD04F462 mov rdi,qword ptr [r12-8]
00007FFDFD04F467 lea rax,[__ImageBase (07FFDFCE20000h)]
00007FFDFD04F46E mov rsi,qword ptr [r12-10h]
00007FFDFD04F473 sub r12,8
00007FFDFD04F477 mov rdx,rdi
00007FFDFD04F47A mov rcx,rsi
00007FFDFD04F47D call qword ptr [rax+r13*8+3E7A70h]
00007FFDFCF07380 mov qword ptr [rsp+8],rbx
00007FFDFCF07385 push rdi
00007FFDFCF07386 sub rsp,30h
00007FFDFCF0738A xor r8d,r8d
00007FFDFCF0738D mov rdi,rdx
00007FFDFCF07390 mov rbx,rcx
00007FFDFCF07393 call binary_op1 (07FFDFCF06C70h)
00007FFDFCF06C70 mov qword ptr [rsp+10h],rbp
00007FFDFCF06C75 mov qword ptr [rsp+18h],rsi
00007FFDFCF06C7A mov qword ptr [rsp+20h],rdi
00007FFDFCF06C7F push r14
00007FFDFCF06C81 sub rsp,20h
00007FFDFCF06C85 mov rsi,rdx
00007FFDFCF06C88 movsxd r9,r8d
00007FFDFCF06C8B mov rdx,qword ptr [rcx+8]
00007FFDFCF06C8F mov rbp,rcx
00007FFDFCF06C92 mov rdi,qword ptr [rdx+60h]
00007FFDFCF06C96 test rdi,rdi
00007FFDFCF06C99 je binary_op1+2Fh (07FFDFCF06C9Fh)
00007FFDFCF06C9B mov rdi,qword ptr [rdi+r9]
00007FFDFCF06C9F mov rcx,qword ptr [rsi+8]
00007FFDFCF06CA3 mov qword ptr [rsp+30h],rbx
00007FFDFCF06CA8 cmp rcx,rdx
00007FFDFCF06CAB je binary_op1+55h (07FFDFCF06CC5h)
00007FFDFCF06CC5 xor ebx,ebx
00007FFDFCF06CC7 lea r14,[_Py_NotImplementedStruct (07FFDFD2E6AF0h)]
00007FFDFCF06CCE test rdi,rdi
00007FFDFCF06CD1 je binary_op1+0ADh (07FFDFCF06D1Dh)
00007FFDFCF06CD3 test rbx,rbx
00007FFDFCF06CD6 je binary_op1+90h (07FFDFCF06D00h)
00007FFDFCF06D00 mov rdx,rsi
00007FFDFCF06D03 mov rcx,rbp
00007FFDFCF06D06 call rdi
00007FFDFCF60410 mov rax,qword ptr [rcx+8]
00007FFDFCF60414 test dword ptr [rax+0A8h],1000000h
00007FFDFCF6041E je long_add+24h (07FFDFCF60434h)
00007FFDFCF60420 mov rax,qword ptr [rdx+8]
00007FFDFCF60424 test dword ptr [rax+0A8h],1000000h
00007FFDFCF6042E jne _PyLong_Add (07FFDFCF60380h)
00007FFDFCF60380 sub rsp,28h
00007FFDFCF60384 mov r8,rdx
00007FFDFCF60387 mov rdx,qword ptr [rcx+10h]
00007FFDFCF6038B lea rax,[rdx+1]
00007FFDFCF6038F cmp rax,3
00007FFDFCF60393 jae _PyLong_Add+3Eh (07FFDFCF603BEh)
00007FFDFCF60395 mov r9,qword ptr [r8+10h]
00007FFDFCF60399 lea rax,[r9+1]
00007FFDFCF6039D cmp rax,3
00007FFDFCF603A1 jae _PyLong_Add+3Eh (07FFDFCF603BEh)
00007FFDFCF603A3 mov ecx,dword ptr [rcx+18h]
00007FFDFCF603A6 mov eax,dword ptr [r8+18h]
00007FFDFCF603AA imul rcx,rdx
00007FFDFCF603AE imul rax,r9
00007FFDFCF603B2 add rcx,rax
00007FFDFCF603B5 add rsp,28h
00007FFDFCF603B9 jmp _PyLong_FromSTwoDigits (07FFDFCF5AF10h)
00007FFDFCF5AF10 push rdi
00007FFDFCF5AF12 sub rsp,20h
00007FFDFCF5AF16 lea rax,[rcx+5]
00007FFDFCF5AF1A mov rdi,rcx
00007FFDFCF5AF1D cmp rax,105h
00007FFDFCF5AF23 ja _PyLong_FromSTwoDigits+2Fh (07FFDFCF5AF3Fh)
00007FFDFCF5AF3F lea rax,[rcx+3FFFFFFFh]
00007FFDFCF5AF46 cmp rax,7FFFFFFFh
00007FFDFCF5AF4C jae _PyLong_FromSTwoDigits+48h (07FFDFCF5AF58h)
00007FFDFCF5AF4E add rsp,20h
00007FFDFCF5AF52 pop rdi
00007FFDFCF5AF53 jmp _PyLong_FromMedium (07FFDFCF5AE50h)
00007FFDFCF5AE50 mov qword ptr [rsp+10h],rbx
00007FFDFCF5AE55 push rsi
00007FFDFCF5AE56 sub rsp,20h
00007FFDFCF5AE5A movsxd rsi,ecx
00007FFDFCF5AE5D mov edx,20h
00007FFDFCF5AE62 mov rcx,qword ptr [_PyObject (07FFDFD3B8CA8h)]
00007FFDFCF5AE69 call qword ptr [_PyObject+8h (07FFDFD3B8CB0h)]
00007FFDFCF73E60 push rbx
00007FFDFCF73E62 sub rsp,20h
00007FFDFCF73E66 mov rbx,rdx
00007FFDFCF73E69 call pymalloc_alloc (07FFDFCF73C80h)
00007FFDFCF73C80 sub rsp,28h
00007FFDFCF73C84 lea rax,[rdx-1]
00007FFDFCF73C88 cmp rax,1FFh
00007FFDFCF73C8E jbe pymalloc_alloc+17h (07FFDFCF73C97h)
00007FFDFCF73C97 mov qword ptr [rsp+30h],rbx
00007FFDFCF73C9C mov qword ptr [rsp+38h],rsi
00007FFDFCF73CA1 mov qword ptr [rsp+40h],rdi
00007FFDFCF73CA6 lea edi,[rdx-1]
00007FFDFCF73CA9 shr edi,4
00007FFDFCF73CAC mov qword ptr [rsp+20h],r14
00007FFDFCF73CB1 lea r14,[__ImageBase (07FFDFCE20000h)]
00007FFDFCF73CB8 lea eax,[rdi+rdi]
00007FFDFCF73CBB lea rsi,[rax*8+4C7120h]
00007FFDFCF73CC3 mov rdx,qword ptr [rsi+r14]
00007FFDFCF73CC7 mov rcx,qword ptr [rdx+10h]
00007FFDFCF73CCB cmp rdx,rcx
00007FFDFCF73CCE je pymalloc_alloc+9Fh (07FFDFCF73D1Fh)
00007FFDFCF73CD0 mov r9,qword ptr [rdx+8]
00007FFDFCF73CD4 inc dword ptr [rdx]
00007FFDFCF73CD6 mov rax,qword ptr [r9]
00007FFDFCF73CD9 mov qword ptr [rdx+8],rax
00007FFDFCF73CDD test rax,rax
00007FFDFCF73CE0 jne pymalloc_alloc+1C1h (07FFDFCF73E41h)
00007FFDFCF73E41 mov r14,qword ptr [rsp+20h]
00007FFDFCF73E46 mov rax,r9
00007FFDFCF73E49 mov rdi,qword ptr [rsp+40h]
00007FFDFCF73E4E mov rsi,qword ptr [rsp+38h]
00007FFDFCF73E53 mov rbx,qword ptr [rsp+30h]
00007FFDFCF73E58 add rsp,28h
00007FFDFCF73E5C ret
00007FFDFCF73E6E test rax,rax
00007FFDFCF73E71 jne _PyObject_Malloc+46h (07FFDFCF73EA6h)
00007FFDFCF73EA6 add rsp,20h
00007FFDFCF73EAA pop rbx
00007FFDFCF73EAB ret
00007FFDFCF5AE6F mov rbx,rax
00007FFDFCF5AE72 test rax,rax
00007FFDFCF5AE75 jne _PyLong_FromMedium+4Ah (07FFDFCF5AE9Ah)
00007FFDFCF5AE9A mov qword ptr [rsp+30h],rdi
00007FFDFCF5AE9F mov eax,esi
00007FFDFCF5AEA1 cdq
00007FFDFCF5AEA2 mov rcx,rsi
00007FFDFCF5AEA5 sar rcx,3Fh
00007FFDFCF5AEA9 mov edi,eax
00007FFDFCF5AEAB and rcx,0FFFFFFFFFFFFFFFEh
00007FFDFCF5AEAF lea rax,[PyLong_Type (07FFDFD2E52E0h)]
00007FFDFCF5AEB6 xor edi,edx
00007FFDFCF5AEB8 mov qword ptr [rbx+8],rax
00007FFDFCF5AEBC inc rcx
00007FFDFCF5AEBF sub edi,edx
00007FFDFCF5AEC1 mov qword ptr [rbx+10h],rcx
00007FFDFCF5AEC5 test dword ptr [PyLong_Type+0A8h (07FFDFD2E5388h)],200h
00007FFDFCF5AECF je _PyLong_FromMedium+88h (07FFDFCF5AED8h)
00007FFDFCF5AED8 cmp dword ptr [_Py_tracemalloc_config+4h (07FFDFD2E7324h)],0
00007FFDFCF5AEDF je _PyLong_FromMedium+99h (07FFDFCF5AEE9h)
00007FFDFCF5AEE9 mov dword ptr [rbx+18h],edi
00007FFDFCF5AEEC mov rax,rbx
00007FFDFCF5AEEF mov rdi,qword ptr [rsp+30h]
00007FFDFCF5AEF4 mov qword ptr [rbx],1
00007FFDFCF5AEFB mov rbx,qword ptr [rsp+38h]
00007FFDFCF5AF00 add rsp,20h
00007FFDFCF5AF04 pop rsi
00007FFDFCF5AF05 ret
00007FFDFCF06D08 mov rcx,rax
00007FFDFCF06D0B cmp rax,r14
00007FFDFCF06D0E jne binary_op1+0D9h (07FFDFCF06D49h)
00007FFDFCF06D49 mov rbx,qword ptr [rsp+30h]
00007FFDFCF06D4E mov rbp,qword ptr [rsp+38h]
00007FFDFCF06D53 mov rsi,qword ptr [rsp+40h]
00007FFDFCF06D58 mov rdi,qword ptr [rsp+48h]
00007FFDFCF06D5D add rsp,20h
00007FFDFCF06D61 pop r14
00007FFDFCF06D63 ret
00007FFDFCF07398 mov rcx,rax
00007FFDFCF0739B lea rax,[_Py_NotImplementedStruct (07FFDFD2E6AF0h)]
00007FFDFCF073A2 cmp rcx,rax
00007FFDFCF073A5 je PyNumber_Add+35h (07FFDFCF073B5h)
00007FFDFCF073A7 mov rax,rcx
00007FFDFCF073AA mov rbx,qword ptr [rsp+40h]
00007FFDFCF073AF add rsp,30h
00007FFDFCF073B3 pop rdi
00007FFDFCF073B4 ret
00007FFDFD04F485 sub qword ptr [rsi],1
00007FFDFD04F489 mov r15,rax
00007FFDFD04F48C jne _PyEval_EvalFrameDefault+58AFh (07FFDFD04F49Fh)
00007FFDFD04F49F sub qword ptr [rdi],1
00007FFDFD04F4A3 jne _PyEval_EvalFrameDefault+58C6h (07FFDFD04F4B6h)
00007FFDFD04F4B6 mov qword ptr [r12-8],r15
00007FFDFD04F4BB test r15,r15
00007FFDFD04F4BE mov r15,qword ptr [rsp+30h]
00007FFDFD04F4C3 je _PyEval_EvalFrameDefault+161h (07FFDFD049D51h)
00007FFDFD04F4C9 add r14,2
00007FFDFD04F4CD jmp _PyEval_EvalFrameDefault+59FCh (07FFDFD04F5ECh) Similarly, running through a python program that sums an array of 4096 integers, a C program can do ~0.8 adds per cycle, and a python version can do 150x less, and this is without optimizations. An aggressive compiler would optimize more.