Function Call Optimization

Function Call Optimization

• 1.
  Function Call Optimization Thisnote describes the observations on the following Function Calls: 1. Constructor call 2. Constructor call of Base Class 3. Get / Set Methods. To Test for the above a sample Base Class has been developed. class Base { public: size_t nameLen; char *name; protected: Base(char *nameStr): nameLen((nameStr)? strlen(nameStr): 0), name(strcpy(new char[nameLen+1], nameStr)) {} … }; This generates the following assembly in debug build using g++. The function calls have been marked in red. _ZN4BaseC2EPc: => Base::Base() .LFB1443: .loc 3 17 0 pushq %rbp # .LCFI22: movq %rsp, %rbp #, .LCFI23: pushq %rbx # .LCFI24: subq $40, %rsp #, .LCFI25: movq %rdi, -16(%rbp) # this, this movq %rsi, -24(%rbp) # nameStr, nameStr .LBB13: .loc 3 20 0 movq -16(%rbp), %rax # this, movq %rax, -32(%rbp) #, cmpq $0, -24(%rbp) #, nameStr je .L39 #, movq -24(%rbp), %rdi # nameStr, nameStr call strlen # => strlen() movq %rax, -40(%rbp) # tmp61, jmp .L40 # .L39: movq $0, -40(%rbp) #,
• 2.
  .L40: movq -40(%rbp), %rax #, movq -32(%rbp), %rdx #, movq %rax, (%rdx) #, <variable>.nameLen movq -16(%rbp), %rbx # this, this movq -16(%rbp), %rax # this, this movq (%rax), %rdi # <variable>.nameLen, tmp63 incq %rdi # tmp63 call _Znam # => operator new() movq %rax, %rdi #, tmp65 movq -24(%rbp), %rsi # nameStr, nameStr call strcpy # => strcpy() movq %rax, 8(%rbx) #, <variable>.name .LBE13: addq $40, %rsp #, popq %rbx # leave ret The Derived Class, class Derived: public Base { Base *myBase; public: Derived(char *name): Base(name), myBase((Base*)this) {} … }; generates the following assembly: _ZN7DerivedC1EPc: => Derived::Derived() .LFB1452: .loc 3 53 0 pushq %rbp # .LCFI19: movq %rsp, %rbp #, .LCFI20: subq $16, %rsp #, .LCFI21: movq %rdi, -8(%rbp) # this, this movq %rsi, -16(%rbp) # name, name .LBB12: .loc 3 56 0 movq -16(%rbp), %rsi # name, name movq -8(%rbp), %rdi # this, this call _ZN4BaseC2EPc # => Base::Base() movq -8(%rbp), %rdx # this, this movq -8(%rbp), %rax # this, this movq %rax, 16(%rdx) # this, <variable>.myBase .LBE12: leave ret Finally, the instantiation of a derived Class Object as in char *s; Derived d(s);
• 3.
  generates: .loc 2 8 0 movq -24(%rbp), %rsi # s, s leaq -64(%rbp), %rdi #, tmp59 .LEHB0: call _ZN7DerivedC1EPc # => Derived::Derived() This means that the functions are called respectively as we had expected. Building the instantiation in release mode, we see the following: .LCFI2: testq %rdi, %rdi # s movq %rsp, %rbp #, tmp114 je .L3 #, call strlen # => strlen() .L3: .L5: leaq 1(%rax), %rdi #, tmp67 movq %rax, (%rbp) # tmp63, <variable>.nameLen .LEHB0: call _Znam # => operator new() .LEHE0: movq %rbx, %rsi # s, nameStr movq %rax, %rdi #, <anonymous> call strcpy # => strcpy() movq %rax, 8(%rbp) # tmp70, <variable>.name movq 8(%rsp), %rdi # <variable>.name, <variable>.name movq %rbp, 16(%rbp) # tmp114, <variable>.myBase testq %rdi, %rdi # <variable>.name jne .L44 #, Thus we see that the Derived (and Base) Constructor calls have been totally optimized. The above is illustrated in terms of the constructors. Similar optimizations do (mostly) take place for (non-virtual) destructors, copy constructor, copy assignment operator, all non-virtual inline member functions and all inline global functions. Only virtual member functions are not inlined as their call sequence is runtime dependent. The same behavior is observed in case of Get / Set Methods.
• 4.
  class Base { … char *GetName() const { return name; } void SetName(char *nameStr) { if (nameStr) { if (name) { delete [] name; } name = strcpy(new char[strlen(nameStr)+1], nameStr); } } … }; The assemblies for the Get / Set Methods are shown below:
• 5.
  _ZNK4Base7GetNameEv: => Base::GetName() .LFB1448: .loc 3 29 0 pushq %rbp # .LCFI17: movq %rsp, %rbp #, .LCFI18: movq %rdi, -8(%rbp) # this, this .LBB11: .loc 3 29 0 movq -8(%rbp), %rax # this, this movq 8(%rax), %rax # <variable>.name, <variable>.name .LBE11: leave ret _ZN4Base7SetNameEPc: => Base::SetName() .LFB1449: .loc 3 31 0 pushq %rbp # .LCFI13: movq %rsp, %rbp #, .LCFI14: pushq %rbx # .LCFI15: subq $24, %rsp #, .LCFI16: movq %rdi, -16(%rbp) # this, this movq %rsi, -24(%rbp) # nameStr, nameStr .LBB10: .loc 3 32 0 cmpq $0, -24(%rbp) #, nameStr je .L29 #, .loc 3 34 0 movq -16(%rbp), %rax # this, this cmpq $0, 8(%rax) #, <variable>.name je .L31 #, .loc 3 36 0 movq -16(%rbp), %rax # this, this cmpq $0, 8(%rax) #, <variable>.name je .L31 #, movq -16(%rbp), %rax # this, this movq 8(%rax), %rdi # <variable>.name, <variable>.name call _ZdaPv # => operator delete() .L31: .loc 3 39 0 movq -16(%rbp), %rbx # this, this movq -24(%rbp), %rdi # nameStr, nameStr call strlen # => strlen() movq %rax, %rdi # tmp65, tmp63 incq %rdi # tmp63 call _Znam # => operatot new() movq %rax, %rdi #, tmp66 movq -24(%rbp), %rsi # nameStr, nameStr call strcpy # => strcpy() movq %rax, 8(%rbx) #, <variable>.name .L29: .LBE10: .loc 3 41 0 addq $24, %rsp #, popq %rbx # leave ret
• 6.
  The calls, char *oldName = d.GetName(); char *newName = "My Gang"; d.SetName(newName); generate the following assembly in debug build: .LEHE0: .loc 2 11 0 leaq -64(%rbp), %rdi #, tmp60 call _ZNK4Base7GetNameEv # => Base::GetName() movq %rax, -72(%rbp) # tmp61, oldName .loc 2 12 0 movq $.LC0, -80(%rbp) #, newName .loc 2 13 0 movq -80(%rbp), %rsi # newName, newName leaq -64(%rbp), %rdi #, tmp63 .LEHB1: call _ZN4Base7SetNameEPc # => Base::SetName() .loc 2 16 0 movq $.LC1, -80(%rbp) #, newName The same in release build is: .L10: movl $.LC0, %edi #, nameStr call strlen # => strlen() leaq 1(%rax), %rdi #, tmp80 .LEHB1: call _Znam # => operator new() movq %rax, %rdi #, tmp85 movl $.LC0, %esi #, nameStr call strcpy # => strcpy() testq %rax, %rax # tmp86 movq %rax, 8(%rsp) # tmp86, <variable>.name je .L14 #, movq %rax, %rdi # tmp86, <variable>.name call _ZdaPv # => operator delete() movl $.LC1, %edi #, newName call strlen # leaq 1(%rax), %rdi #, tmp90 call _Znam # => operator new() movq %rax, %rdi #, tmp95 movl $.LC1, %esi #, newName call strcpy # => strcpy() movq %rax, 8(%rsp) # tmp96, <variable>.name