Does the restrict keyword provide significant benefits in gcc/g++?

Does the restrict keyword provide significant benefits in gcc / g++ ?

It can reduce the number of instructions as shown on the example below, so use it whenever possible.

GCC 4.8 Linux x86-64 exmample

Input:

void f(int *a, int *b, int *x) {
  *a += *x;
  *b += *x;
}

void fr(int *restrict a, int *restrict b, int *restrict x) {
  *a += *x;
  *b += *x;
}

Compile and decompile:

gcc -g -std=c99 -O0 -c main.c
objdump -S main.o

With -O0, they are the same.

With -O3:

void f(int *a, int *b, int *x) {
    *a += *x;
   0:   8b 02                   mov    (%rdx),%eax
   2:   01 07                   add    %eax,(%rdi)
    *b += *x;
   4:   8b 02                   mov    (%rdx),%eax
   6:   01 06                   add    %eax,(%rsi)  

void fr(int *restrict a, int *restrict b, int *restrict x) {
    *a += *x;
  10:   8b 02                   mov    (%rdx),%eax
  12:   01 07                   add    %eax,(%rdi)
    *b += *x;
  14:   01 06                   add    %eax,(%rsi) 

For the uninitiated, the calling convention is:

• rdi = first parameter
• rsi = second parameter
• rdx = third parameter

Conclusion: 3 instructions instead of 4.

Of course, instructions can have different latencies, but this gives a good idea.

Why GCC was able to optimize that?

The code above was taken from the Wikipedia example which is very illuminating.

Pseudo assembly for f:

load R1 ← *x    ; Load the value of x pointer
load R2 ← *a    ; Load the value of a pointer
add R2 += R1    ; Perform Addition
set R2 → *a     ; Update the value of a pointer
; Similarly for b, note that x is loaded twice,
; because x may point to a (a aliased by x) thus 
; the value of x will change when the value of a
; changes.
load R1 ← *x
load R2 ← *b
add R2 += R1
set R2 → *b

For fr:

load R1 ← *x
load R2 ← *a
add R2 += R1
set R2 → *a
; Note that x is not reloaded,
; because the compiler knows it is unchanged
; "load R1 ← *x" is no longer needed.
load R2 ← *b
add R2 += R1
set R2 → *b

Is it really any faster?

Ermmm... not for this simple test:

.text
    .global _start
    _start:
        mov $0x10000000, %rbx
        mov $x, %rdx
        mov $x, %rdi
        mov $x, %rsi
    loop:
        # START of interesting block
        mov (%rdx),%eax
        add %eax,(%rdi)
        mov (%rdx),%eax # Comment out this line.
        add %eax,(%rsi)
        # END ------------------------
        dec %rbx
        cmp $0, %rbx
        jnz loop
        mov $60, %rax
        mov $0, %rdi
        syscall
.data
    x:
        .int 0

And then:

as -o a.o a.S && ld a.o && time ./a.out

on Ubuntu 14.04 AMD64 CPU Intel i5-3210M.

I confess that I still don't understand modern CPUs. Let me know if you:

• found a flaw in my method
• found an assembler test case where it becomes much faster
• understand why there wasn't a difference

The restrict keyword does a difference.

I've seen improvements of factor 2 and more in some situations (image processing). Most of the time the difference is not that large though. About 10%.

Here is a little example that illustrate the difference. I've written a very basic 4x4 vector * matrix transform as a test. Note that I have to force the function not to be inlined. Otherwise GCC detects that there aren't any aliasing pointers in my benchmark code and restrict wouldn't make a difference due to inlining.

I could have moved the transform function to a different file as well.

#include <math.h>

#ifdef USE_RESTRICT
#else
#define __restrict
#endif


void transform (float * __restrict dest, float * __restrict src, 
                float * __restrict matrix, int n) __attribute__ ((noinline));

void transform (float * __restrict dest, float * __restrict src, 
                float * __restrict matrix, int n)
{
  int i;

  // simple transform loop.

  // written with aliasing in mind. dest, src and matrix 
  // are potentially aliasing, so the compiler is forced to reload
  // the values of matrix and src for each iteration.

  for (i=0; i<n; i++)
  {
    dest[0] = src[0] * matrix[0] + src[1] * matrix[1] + 
              src[2] * matrix[2] + src[3] * matrix[3];

    dest[1] = src[0] * matrix[4] + src[1] * matrix[5] + 
              src[2] * matrix[6] + src[3] * matrix[7];

    dest[2] = src[0] * matrix[8] + src[1] * matrix[9] + 
              src[2] * matrix[10] + src[3] * matrix[11];

    dest[3] = src[0] * matrix[12] + src[1] * matrix[13] + 
              src[2] * matrix[14] + src[3] * matrix[15];

    src  += 4;
    dest += 4;
  }
}

float srcdata[4*10000];
float dstdata[4*10000];

int main (int argc, char**args)
{
  int i,j;
  float matrix[16];

  // init all source-data, so we don't get NANs  
  for (i=0; i<16; i++)   matrix[i] = 1;
  for (i=0; i<4*10000; i++) srcdata[i] = i;

  // do a bunch of tests for benchmarking. 
  for (j=0; j<10000; j++)
    transform (dstdata, srcdata, matrix, 10000);
}

---

Results: (on my 2 Ghz Core Duo)

nils@doofnase:~$ gcc -O3 test.c
nils@doofnase:~$ time ./a.out

real    0m2.517s
user    0m2.516s
sys     0m0.004s

nils@doofnase:~$ gcc -O3 -DUSE_RESTRICT test.c
nils@doofnase:~$ time ./a.out

real    0m2.034s
user    0m2.028s
sys     0m0.000s

Over the thumb 20% faster execution, on that system.

To show how much it depends on the architecture I've let the same code run on a Cortex-A8 embedded CPU (adjusted the loop count a bit cause I don't want to wait that long):

root@beagleboard:~# gcc -O3 -mcpu=cortex-a8 -mfpu=neon -mfloat-abi=softfp test.c
root@beagleboard:~# time ./a.out

real    0m 7.64s
user    0m 7.62s
sys     0m 0.00s

root@beagleboard:~# gcc -O3 -mcpu=cortex-a8 -mfpu=neon -mfloat-abi=softfp -DUSE_RESTRICT test.c 
root@beagleboard:~# time ./a.out

real    0m 7.00s
user    0m 6.98s
sys     0m 0.00s

Here the difference is just 9% (same compiler btw.)