perf/sarp.c

   1 // This artificial program allocates and deallocates a lot of large objects
   2 // on the stack.  It is a stress test for Memcheck's set_address_range_perms
   3 // (sarp) function.  Pretty much all Valgrind versions up to 3.1.X do very
   4 // badly on it, ie. a slowdown of at least 100x.
   5 //
   6 // It is representative of tsim_arch, the simulator for the University of
   7 // Texas's TRIPS processor, whose performance under Valgrind is dominated by
   8 // the handling of one frequently-called function that allocates 8348 bytes
   9 // on the stack.
  10
  11 #include <assert.h>
  12 #include <time.h>
  13
  14 #define REPS   1000*1000*10
  15
  16 __attribute__((noinline))
  17 int f(int i)
  18 {
  19    // This nonsense is just to ensure that the compiler does not optimise
  20    // away the stack allocation.
  21    char big_array[500];
  22    big_array[  0] = 12;
  23    big_array[ 23] = 34;
  24    big_array[256] = 56;
  25    big_array[434] = 78;
  26    assert( 480 == (&big_array[490] - &big_array[10]) );
  27    return big_array[i];
  28 }
  29
  30 int main(void)
  31 {
  32    int i, sum = 0;
  33
  34    struct timespec req __attribute__((unused));
  35    req.tv_sec  = 0;
  36    req.tv_nsec = 100*1000*1000;   // 0.1s
  37
  38    // Pause for a bit so that the native run-time is not 0.00, which leads
  39    // to ridiculous slow-down figures.
  40    //nanosleep(&req, NULL);
  41
  42    for (i = 0; i < REPS; i++) {
  43       sum += f(i & 0xff);
  44    }
  45    return ( sum == 0xdeadbeef ? 1 : 0 );
  46 }
  47