Declare variables at top of function or in separate scopes? - c++

Question

Which is preferred, method 1 or method 2?
Method 1:
LRESULT CALLBACK wpMainWindow(HWND hwnd, UINT msg, WPARAM wparam, LPARAM lparam)
{
switch (msg)
{
case WM_PAINT:
{
HDC hdc;
PAINTSTRUCT ps;
RECT rc;
GetClientRect(hwnd, &rc);
hdc = BeginPaint(hwnd, &ps);
// drawing here
EndPaint(hwnd, &ps);
break;
}
default:
return DefWindowProc(hwnd, msg, wparam, lparam);
}
return 0;
}
Method 2:
LRESULT CALLBACK wpMainWindow(HWND hwnd, UINT msg, WPARAM wparam, LPARAM lparam)
{
HDC hdc;
PAINTSTRUCT ps;
RECT rc;
switch (msg)
{
case WM_PAINT:
GetClientRect(hwnd, &rc);
hdc = BeginPaint(hwnd, &ps);
// drawing here
EndPaint(hwnd, &ps);
break;
default:
return DefWindowProc(hwnd, msg, wparam, lparam);
}
return 0;
}
In method 1, if msg = WM_PAINT when wpMainWindow function is called, does it allocate memory for all the variables on the stack at the beginning? or only when it enters the WM_PAINT scope?
Would method 1 only use the memory when the message is WM_PAINT, and method 2 would use the memory no matter what msg equaled?

Answer 1

Variables should be declared as locally as possible.
Declaring variables "at the top of the function" is always a disastrously bad practice. Even in C89/90 language, where variables can only be declared at the beginning of the block, it is better to declare them as locally as possible, i.e. at the beginning of smallest local block that covers the desired lifetime of the variable. Sometimes it might even make sense to introduce a "redundant" local block with the only purpose of "localizing" the variable declaration.
In C++ and C99, where it is possible to declare variable anywhere in the code, the answer is pretty straightforward: again, declare each variable as locally as possible, and as close as possible to the point where you use it the very first time. The primary rationale for that is that in most cases this will allow you to supply a meaningful initializer to the variable at the point of declaration (instead of declaring it without initializer or with a dummy initializer).
As for the memory usage, in general a typical implementation will immediately (as you enter the function) allocate the maximum space required for all variables that exist at the same time. However, your declaration habits might affect the exact size of that space. For example, in this code
void foo() {
int a, b, c;
if (...) {
}
if (...) {
}
}
all three variables exist at the same time and generally the space for all three has to be allocated. But in this code
void foo() {
int a;
if (...) {
int b;
}
if (...) {
int c;
}
}
only two variables exist at any given moment, meaning that space for only two variables will be allocated by a typical implementation (b and c will share the same space). This is another reason to declare variables as locally as possible.

Answer 2

Whether something's allocated on the stack in case 1 is implementation defined. Implementations aren't even required to have a stack.
It's usually no slower to do so since the operation tends to be a simple subtraction (for a downward growing stack) of one value from the stack pointer for the entire local variable area.
The thing that's important here is that the scope should be as local as possible. In other words, declare your variables as late as possible and only keep them around as long as needed.
Note that declaring here is at a different abstraction level to allocating space for them. The actual space may be allocated at the start of the function (implementation level) but you can only use those variables while they're scoped (C level).
Locality of information is important, just like its cousin, encapsulation.

Answer 3

I like Method 3:
LRESULT wpMainWindowPaint(HWND hwnd)
{
HDC hdc;
PAINTSTRUCT ps;
RECT rc;
GetClientRect(hwnd, &rc);
hdc = BeginPaint(hwnd, &ps);
// drawing here
EndPaint(hwnd, &ps);
return 0;
}
LRESULT CALLBACK wpMainWindow(HWND hwnd, UINT msg, WPARAM wparam, LPARAM lparam)
{
switch (msg)
{
case WM_PAINT: return wpMainWindowPaint(hwnd);
default: return DefWindowProc(hwnd, msg, wparam, lparam);
}
}
If it deserves its own scope for organization purposes, it deserves its own function. If you're worried about function call overhead, make it inline.

Answer 4

Since it's the compiler's job to optimize my code, and an hour of compiler-time is way cheaper than an hour of my time, and my time gets wasted if I need to scroll up and down the code to see where a variable was declared, I think my company wants me to keep everything as local as possible.
Not even am I talking about 'the smallest block', but 'as near to the place where it is used'!
LRESULT CALLBACK wpMainWindow(HWND hwnd, UINT msg, WPARAM wparam, LPARAM lparam)
{
switch (msg)
{
case WM_PAINT:
{
RECT rc;
GetClientRect(hwnd, &rc);
{ // sometimes I even create an arbitrary block
// to show correlated statements.
// as a side-effect, the compiler may not need to allocate space for
// variables declared here...
PAINTSTRUCT ps;
HDC hdc = BeginPaint(hwnd, &ps);
// drawing here
EndPaint(hwnd, &ps);
}
break;
}
default:
return DefWindowProc(hwnd, msg, wparam, lparam);
}
return 0;
}

Answer 5

Define the variables in the narrowest scope where they are relevant. There's no reason to use Method 2 above in my opinion.
Stack space is only likely to be used when the variables are in scope. As #paxdiablo points out, your locals may wind up in registers rather than on the stack, if the compiler can find the space for them.

Answer 6

Memory allocation is not specified in the Standard to this detail, so for a real answer you'll have to specify compiler and platform. It's not going to matter for performance.
What you want is readability, and in general that's done by declaring variables in the smallest usable scope, and preferably when you can immediately initialize them with reasonable values. The smaller a variable's scope, the less it can potentially interact with the rest of the program in unpredictable ways. The closer the declaration to initialization, the less opportunity for anything bad to happen.
What would probably be better is something like
RECT rc;
GetClientRect(hwnd, &rc);
PAINTSTRUCT ps;
HDC hdc = BeginPaint(hwnd, &ps);
This is for C++. For C, the rule is similar, except that earlier versions of C required all the variables to be declared at the top of a block.

Answer 7

You can't know at what point the stack reservation is done.
For readability I would go with C99 (or C++). That allows you the declaration of a variable really there where first use it.
HDC hdc = BeginPaint(hwnd, &ps);

Answer 8

For Java programming language, the common practice is to declare local variables only when needed in a method.
void foo(int i) {
if (i == 1)
return;
Map map1 = new HashMap();
if (i == 2)
return;
Map map2 = new HashMap();
}
For C++ programming language, I also suggest the same practice since declaring variables with a non-trivial constructor involves execution cost. Putting all these declarations at the method beginning causes unneeded cost if some of these variables will be used.
void foo(int i)
{
if (i == 1)
return;
std::map<int, int> map1; // constructor is executed here
if (i == 2)
return;
std::map<int, int> map2; // constructor is executed here
}
For C, the story is different. It depends on architecture and compiler. For x86 and GCC, putting all the declarations at function beginning and declaring variables only when needed have the same performance. The reason is that C variables don't have a constructor. And the effect on stack memory allocation by these two approaches is the same. Here is an example:
void foo(int i)
{
int m[50];
int n[50];
switch (i) {
case 0:
break;
case 1:
break;
default:
break;
}
}
void bar(int i)
{
int m[50];
switch (i) {
case 0:
break;
case 1:
break;
default:
break;
}
int n[50];
}
For both functions, the assembly code for stack manipulation is:
pushl %ebp
movl %esp, %ebp
subl $400, %esp
Putting all the declarations at function beginning is common in Linux kernel code.

Answer 9

There's no need to pollute the stack with variables that are possibly never used. Allocate your vars right before they are used. Overlooking the RECT rc and the subsequent call to GetClientRect, Ben Voight's method is the way to go.