gistfile1.txt

% pprof --list=thread_scan a.out a.out.prof 
Using local file a.out.
Using local file a.out.prof.
Removing _L_unlock_16 from all stack traces.
ROUTINE ====================== thread_scan in /afs/andrew.cmu.edu/usr9/areece/15-418/profiling/try_perf.c
   199   1087 Total samples (flat / cumulative)
     .      .   55:   size_t size;
     .      .   56:   size_t* num_perfect_squares;
     .      .   57:   mutex_t * num_perfect_squares_lock;
     .      .   58: } thread_arg_t;
     .      .   59: 
---
     .      .   60: void* thread_scan(void* void_arg) {
     .      .   61:     // TODO(awreece) Copy locally so dont interfere with each other.
     .      .   62: 	thread_arg_t* args = (thread_arg_t*) void_arg;
     .      .   63: 	size_t i;
     .      .   64: 
    83     99   65: 	for (i = 0; i < args->size; i++) {
    41     80   66: 	  	uint32_t val = args->input[i];
    26    744   67: 		uint32_t sqrt = integer_square_root(val);
    39     74   68: 		if (sqrt * sqrt == val) {
    10     90   69: 		  __sync_fetch_and_add(args->num_perfect_squares, 1);
     .      .   70: 		}
     .      .   71: 	}
     .      .   72: }
---
     .      .   73: 
     .      .   74: int main(int argc, char** argv) {
     .      .   75: 	size_t size = 1<<27;
     .      .   76: 	uint32_t* inarray = (uint32_t*) malloc(size * sizeof(uint32_t));
     .      .   77: 	size_t i;