Pass by reference or by value?

[sturlamolden]

On Jul 13, 9:10 pm, Robert Dailey <rcdai... at gmail.com> wrote:

> I noticed in Python all function parameters seem to be passed by
> reference. This means that when I modify the value of a variable of a
> function, the value of the variable externally from the function is
> also modified.


Pass-by-value and pass-by-reference usually refer to the semantics
used in C and Fortran, respectively.


Here is an example showing the distinction between the 'pass-by-value'
semantics of C and the 'pass-by-reference' semantics of Fortran:


In C you can pass pointers, but it is still a pure 'pass-by-value'
semantics. The pointer itself is passed by value (which is an
address). What that means, is that the function receives a copy of the
argument used in the call:

int a = 1;
foo(&a);
bar(&a);

void foo(int *a)
{
  a = (int*) 0; /* no side effect */
}

void bar(int *a)
{
  *a = 0; /* side effect, but a is still passed by value */
}

Now take a look at how the functions foo and bar are called. They both
have the same signature,
void ()(* int). Thus they must be called equivalently. But still one
produces a side-effect and the other does not. Often in C litterature
you can see what happens in the function bar called 'pass-by-
reference'. But that is very incorrect, C never passes anything by
reference.

'Pass-by-value' is thus equivalent to 'pass-a-copy'.


In Fortran you can only pass references.

integer(4) :: a
a = 1
call bar(a)

subroutine bar(a)
integer(4) :: a
a = 0             ! side-effect
end subroutine

That means, when a variable is used to call a function, the function
receives a pointer to the actual argument, not a local copy. That is
very different from C's copy-passing behaviour.

C++ can pass-by-value and pass-by-reference:

int a = 1;
foo(&a);  // no side-effect
bar(&a);  // side-effect
foo2(a);  // no side-effect
bar2(a);  // side-effect!!!

void foo(int *a)  // pass-by-value
{
  a = (int*) 0; // no side effect
}

void bar(int *a)  // pass-by-value
{
  *a = 0; // side effect, but a is still passed by value
}


void foo2(int a) // pass-by-value
{
   a = 0;  // no side effect !!!
}


void bar2(int &a) // pass-by-reference
{
   a = 0;  // side effect !!!
}

The C++ example clearly shows what we mean by pass-by-reference. It
seems when we call the function bar2 that we give it the value of a,
when we in fact give it a reference to a. That is the same thing that
happens in Fortran.

What it all boils down to is this:


Local copies are made in the 'pass-by-value' semantics.
Local copies are not made in the 'pass-by-reference' semantics.



Now that we have defined 'pass-by-value' and 'pass-by-reference'
precisely, it is time to examine what Python does.

It turns out that Python neither works like C nor like Fortran.
Rather, Python works like LISP. That is:


Python names are pointers bound to values.
Python always pass pointers to values.
Python never pass local copies.
Function arguments are referenced by names in the function's local
namespace.
Names referencing function arguments can be rebound in the local
scope.


>From this you might think that Python does pass-by-reference. Afterall
it does not create local copies. But as names can be rebound in the
local scope, function calls in Python and Fortran behave very
differently.

That is:

a = 1 # a points to an int(1)
foobar(a)
# a still points to an int(1)

def foobar(a):
   a = 0 # rebinds a locally, produces no side-effect
         # in Fortran this would have produced a side-effect


Now, examine this code sniplet:

a = 1 # a points to an int(1)
foobar(a)
# a still points to an int(1) as an int is immutable
a = [1, 2, 3] # rebinds a, a points to a mutable type
foobar(a)
# a points to [1,2,3,1,2,3]

def foobar(a):
   a *= 2 # rebinds a only if a points to an immutable type,
          # otherwise the value pointed to by a is changed


If you take a careful look at these examples, you should be able to
figure out what Python does.

Python does 'pass-by-reference' in the sense that it always pass
pointers.

Python does 'pass-by-value' in the sense that names has a local scope
and can be rebound in the local scope.

The correct statement would be to say that Python does neither pass-by-
reference nor pass-by-value in the conventional meaning of the terms.

Rather Python does 'pass-as-lisp'.



Sturla Molden



P.S. You could think of Python as a Java only having reference types.
What makes the semantics of Java and C# 'pass-by-value' like C, is the
fact that they make copies of value-types when passed to a function. A
reference type is also passed by value, in this case the value of the
reference is passed. Python does not have value types. But Python does
have immutable types. An int is an immutable type in Python and a
value type in Java. What happens when Python and Java calls a function
with an int as argument is very different. Java makes a copy of the
int and passes that. Python passes a pointer to the int. But as an int
is immutable in Python, it may look like Python passes ints 'by value'
when in fact it never creates a local copy.