views:

696

answers:

10

Memory access through pointers is said to be more efficient than memory access through an array. I am learning C and the above is stated in K&R. Specifically they say

Any operation that can be achieved by array subscripting can also be done with pointers. The pointer version will in general be faster

I dis-assembled the following code using visual C++.(Mine is a 686 processor. I have disabled all optimizations.)

int a[10], *p = a, temp;

void foo()
{
    temp = a[0];
    temp = *p;
}

To my surprise I see that memory access through a pointer takes 3 instructions to the two taken by memory access through an array. Below is the corresponding code.

; 5    : temp = a[0];

    mov eax, DWORD PTR _a
    mov DWORD PTR _temp, eax

; 6    : temp = *p;

    mov eax, DWORD PTR _p
    mov ecx, DWORD PTR [eax]
    mov DWORD PTR _temp, ecx

Please help me understand. What am I missing here??


As pointed out by many answers and comments I had used a compile time constant as the array index thus making it arguably easier for the access through an array. Below is the assembly code with a variable as the index. I now have equal number of instructions for access through pointer and arrays. My broader questions still holds good. The memory access through a pointer is not lending itself as being more efficient.

; 7    :        temp = a[i];

    mov eax, DWORD PTR _i
    mov ecx, DWORD PTR _a[eax*4]
    mov DWORD PTR _temp, ecx

; 8    : 
; 9    :    
; 10   :        temp = *p;

    mov eax, DWORD PTR _p
    mov ecx, DWORD PTR [eax]
    mov DWORD PTR _temp, ecx
+4  A: 

In the first case, the compiler directly knows the address of the array (which is also the address of the first element) and accesses it. In the second case, he knows the address of the pointer and reads the pointer's value, which points to that memory location. That's actually one additional indirection, so it's presumably slower here.

Alexander Gessler
This is true *if* this is not "optimized out" by the compiler as per Paxdiablo's answer.
dangerstat
Or, if the processor doesn't have an array access addressing mode, because if it does it can do the math in the same time in either case. At which point you're just down to additional register pressure from having to store the base pointer and the offset.
Andrew McGregor
Johannes Schaub - litb
+12  A: 

Memory access through pointers is said to be more efficient than memory access through an array.

That may have been true in the past when compilers were relatively stupid beasts. You only need to look at some of the code output by gcc in high optimization modes to know that it is no longer true. Some of that code is very hard to understand but, once you do, its brilliance is evident.

A decent compiler will generate the same code for pointer accesses and array accesses and you should probably not be worrying about that level of performance. The people that write compilers know far more about their target architectures than we mere mortals. Concentrate more on the macro level when optimizing your code (algorithm selection and so on) and trust in your tool-makers to do their job.


In fact, I'm surprised the compiler didn't optimize the entire

temp = a[0];

line out of existence since temp is over-written in the very next line with a different value and a is in no way marked volatile.

I remember an urban myth from long ago about a benchmark for the latest VAX Fortran compiler (showing my age here) that outperformed its competitors by several orders of magnitude.

Turns out the compiler figured out that the result from the benchmark calculation wasn't used anyhwere so it optimized the entire calculation loop into oblivion. Hence the substantial improvement in run speed.


Update: The reason that unoptimized code is more efficient in your particular case is because of the way you find the location. a will be at a fixed location decided at link/load time and the reference to it will be fixed up at the same time. So a[0] or indeed a[any constant] will be at a fixed location.

And p itself will also be at a fixed location for the same reason. But *p (the contents of p) is variable and therefore will have an extra lookup involved to find the correct memory location.

You'll probably find that having yet another variable x set to 0 (not const) and using a[x] would also introduce extra calculations.


In one of your comments, you state:

Doing as you suggested resulted in 3 instructions for memory access through arrays too (fetch index, fetch value of array element, store in temp). But I am still unable to see the efficiency. :-(

My response to that is that you very likely won't see an efficiency in using pointers. Modern compilers are more than up to the task of figuring out that array operations and pointer operations can be turned into the same underlying machine code.

In fact, without optimisation turned on, pointer code can be less efficient. Consider the following translations:

int *pa, i, a[10];

for (i = 0; i < 10; i++)
    a[i] = 100;
/*
    movl    $0, -16(%ebp)              ; this is i, init to 0
L2:
    cmpl    $9, -16(%ebp)              ; from 0 to 9
    jg      L3
    movl    -16(%ebp), %eax            ; load i into register
    movl    $100, -72(%ebp,%eax,4)     ; store 100 based on array/i
    leal    -16(%ebp), %eax            ; get address of i
    incl    (%eax)                     ; increment
    jmp     L2                         ; and loop
L3:
*/

for (pa = a; pa < a + 10; pa++)
    *pa = 100;
/*
    leal    -72(%ebp), %eax
    movl    %eax, -12(%ebp)            ; this is pa, init to &a[0]
L5:
    leal    -72(%ebp), %eax
    addl    $40, %eax
    cmpl    -12(%ebp), %eax            ; is pa at &(a[10])
    jbe     L6                         ; yes, stop
    movl    -12(%ebp), %eax            ; get pa
    movl    $100, (%eax)               ; store 100
    leal    -12(%ebp), %eax            ; get pa
    addl    $4, (%eax)                 ; add 4 (sizeof int)
    jmp     L5                         ; loop around
L6:
*/

From that example, you can actually see that the pointer example is longer, and unnecessarily so. It loads pa into %eax multiple times without it changing and indeed alternates %eax between pa and &(a[10]). The default optimisation here is basically none at all.

When you switch up to optimisation level 2, the code you get is:

    xorl    %eax, %eax
L5:
    movl    $100, %edx
    movl    %edx, -56(%ebp,%eax,4)
    incl    %eax
    cmpl    $9, %eax
    jle     L5

for the array version, and:

    leal    -56(%ebp), %eax
    leal    -16(%ebp), %edx
    jmp     L14
L16:
    movl    $100, (%eax)
    addl    $4, %eax
L14:
    cmpl    %eax, %edx
    ja      L16

for the pointer version.

I'm not going to do an analysis on clock cycles here (since it's too much work and I'm basically lazy) but I will point out one thing. There's not a huge difference in the code for both versions in terms of assembler instructions and, given the speeds that modern CPUs actually run at, you won't notice a difference unless you're doing billions of these operations. I always tend to prefer writing code for readability and only worrying about performance if it becomes an issue.

As an aside, that statement you reference:

5.3 Pointers and Arrays: The pointer version will in general be faster but, at least to the uninitiated, somewhat harder to grasp immediately.

dates back to the earliest versions of K&R, including my ancient 1978 one where functions are still written:

getint(pn)
int *pn;
{
    ...
}

Compilers have come an awfully long way since back then.

paxdiablo
Not an urban myth, this happened to me: we had a test for platform ports, part of a published conformance test set, intended to demonstrate that there was a certain "depth" of call stack available, by making a certain number of recursive calls. GCC made a tail-call optimisation. Oops. Writing optimisation-proof benchmarks is never as easy as you think...
Steve Jessop
I had specifically disabled all optimizations to try to see in assembly as to how pointers are more efficient.I am adding this info to my question just in case.
Abhijith Madhav
Johannes Schaub - litb
"You'll probably find that having yet another variable x set to 0 (not const) and using a[x] would also introduce extra calculations.". You are right. As @Droksha had commented earlier this makes the number of assembly instructions equal for access through array and pointer. I have added this info in the question.
Abhijith Madhav
Porculus
@Porculus, ah i see now. I thought it's just about `a[i]` vs `*(p+i)`.
Johannes Schaub - litb
Abhijith Madhav
I understand that, @Abhijith. My contention is that you *won't* see it, simply because it's no longer true. It would only be true for those compilers that calculated base+offset for every index access within an array and those compilers should be thrown immediately on the scrap heap :-)
paxdiablo
+1  A: 

"The pointer version will in general be faster" means that in most cases it's easier for the compiler to generate more efficient code having a pointer (which just needs to be dereferenced) than having an array and subscript (which means that the compiler needs to shift the address from the start of the array). With the modern processors and optimizing compilers, however, array access in the typical case is not slower than pointer access.

Specifically in your case, you would need to switch on the optimization, in order to get the same result.

Vlad
+1  A: 

Since 0 is defined as a constant, a[0] is a constant too, and the compiler knows where it is at compile time. In the "normal" case, the compiler would have to compute the element address from a base + offset (with offset being scaled according to the element size).

OTOH, p is a variable, and the indirection requires an extra move.

On a general basis, array index is internally handled as pointer arithmetic anyway, so I'm not sure to see the point that the K&R was trying to make.

filofel
+2  A: 

The speed is gained in loops, most of all. When you use an array, you would use a counter which you increment. To calculate the position, the system multiplies this counter with the size of the array element, then adds the address of the first element to get the address. With pointers, all you need to do to go to the next element is to increase the current pointer with the size of the element to get the next one, assuming all elements are next to each other in-memory.

Pointer arithmetic thus takes a bit less calculations when doing loops. Also, having pointers to the right element is faster than using an index within an array.

Modern development is slowly getting rid of many pointer operations, though. Processors are getting faster and faster and arrays are easier to manage than pointers. Also, arrays tend to reduce the amount of bugs in code. Array will allow index checks, making sure you're not accessing data outside the array.

Workshop Alex
"To calculate the position, the system multiplies this counter with the size of the array element, then adds the address of the first element to get the address." This entire sequence of steps are attained by a single assembly instruction( mov ecx, DWORD PTR _a[eax*4]) atleast in x86 thus granting no advantage to access through pointers as far as I can see.
Abhijith Madhav
On other platforms (non-x86), you'll have to multiply the index by the size of the array elements as a separate op. You might be able to simply left shift (if the size is a power of 2), but you'll have to multiply in other cases.
tomlogic
@workshop alex .. can u explain how this happens with Example ?
mr_eclair
+2  A: 

As paxdiablo said, Any new compiler will make them very similar.

Even more, I saw situations where array was faster then pointers. This was on a DSP processor which uses vector operations.

In this case, using arrays was similar to using restrict pointers. Because by using two arrays the compiler -implicitly- knows that they don't point to the same location. But if you deal with 2 pointer, the compiler may think that they point to same location and will skip pipe lining.

for example:

int a[10],b[10],c[10];
int *pa=a, *pb=b, *pc=c;
int i;

// fill a and b.
fill_arrays(a,b);

// set c[i] = a[i]+b[i];
for (i = 0; i<10; i++)
{
   c[i] = a[i] + b[i];
}

// set *pc++ = *pa++ + *pb++;
for (i = 0; i<10; i++)
{
   *pc++ = *pa++ + *pb++;
}

In case 1, the compiler will easily do pipe-lining of adding a and b, and storing value to c.

In case 2, the compiler will not pipe-line, because he might be overwriting a or b while saving to C.

Yousf
Good point about the use of restrict - pointer aliasing is pretty much the 'elephant in the room'. I think restrict is now a keyword in C++0x which is an excellent decision IMO
zebrabox
+1  A: 

If you're programming embedded platforms, you quickly learn that the pointer method is a lot faster than using an index.

struct bar a[10], *p;

void foo()
{
    int i;

    // slow loop
    for (i = 0; i < 10; ++i)
        printf( a[i].value);

    // faster loop
    for (p = a; p < &a[10]; ++p)
        printf( p->value);
}

The slow loop has to calculate a + (i * sizeof(struct bar)) each time through, whereas the second just has to add sizeof(struct bar) to p each time through. The multiply operation uses more clock cycles than the add on many processors.

You really start to see improvements if you reference a[i] multiple times inside the loop. Some compilers don't cache that address, so it may be recalculated multiple times inside the loop.

Try updating your sample to use a struct and reference multiple elements.

tomlogic
That would actually be down to the compiler, really, rather than the target environment. Any good compiler would turn that first loop into the same code as the second. Of course, if you're using an old 8051 compiler (for example) that *doesn't* do the insane levels of optimization that gcc is capable of then, yes, the first may well be slower.
paxdiablo
+2  A: 

You're getting good answers to your question here, but since you are learning, it is worth pointing out that efficiencies at that level are seldom noticeable.

When you are tuning a program for maximum performance, you should give at least as much attention to finding and fixing larger issues in the structure of the program. After those have been fixed, low-level optimizations can make a further difference.

Here's an example of how this can be done.

Mike Dunlavey
+1  A: 

Pointers used to be faster than arrays. Certainly back when the C language was designed pointers were quite a bit faster. But these days, optimizers can usually do a better job optimizing arrays than it can with pointers because arrays are more restricted.

Instruction sets of modern processors have also been designed to help optimize array access.

So the bottom line is that arrays are often faster these days, especially when used in loops with index variables.

Of course you would still want to use pointers for things like linked lists, but the old time optimization of walking a pointer through an array rather than using an index variable is now likely to be a dis-optimization.

John Knoeller
+2  A: 

Pointers naturally express simple induction variables while subscripts somewhat intrinsically require more sophisticated compiler optimizations


In many cases just using a subscripted expression requires that an extra layer be added to the problem. A loop that increments a subscript i can be though of as a state machine, and the expression a[i] technically requires, each time it is used, that i be multiplied times the size of each element and added to the base address.

In order to transform that access pattern to use pointers the compiler must analyze the entire loop and determine that, say, each element is being accessed. Then the compiler can replace the multiple instances of multiplying the subscript by the element size with a simple increment of the previous loop value. This process combines optimizations called common subexpression elimination and induction variable strength reduction.

When writing with pointers, the entire optimization process is not necessary because the programmer will typically just step through the array to start with.

Sometimes the compiler can do the optimization and sometimes it can't. It's more common in recent years to have a sophisticated compiler at hand, so pointer-based code is not always faster.

Because arrrays must usually be contiguous, another advantage for pointers is in creating incrementally allocated composite structures.

DigitalRoss