gistfile1.txt

% pprof --disas=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
   199   1087 samples (flat, cumulative) 103.4% of total
-------------------- .../andrew.cmu.edu/usr9/areece/15-418/profiling/try_perf.c
     .      .    60: void* thread_scan(void* void_arg) {
     .      .      400869: push   %rbp
     .      .      40086a: mov    %rsp,%rbp
     .      .      40086d: sub    $0x28,%rsp
     .      .      400871: mov    %rdi,-0x28(%rbp)
     .      .    62: thread_arg_t* args = (thread_arg_t*) void_arg;
     .      .      400875: mov    -0x28(%rbp),%rax
     .      .      400879: mov    %rax,-0x10(%rbp)
     .      7    65: for (i = 0; i < args->size; i++) {
     .      .      40087d: movq   $0x0,-0x8(%rbp)
     .      7      400885: jmp    4008c9 <thread_scan+0x60>
    41     80    66: uint32_t val = args->input[i];
     7      7      400887: mov    -0x10(%rbp),%rax
     .      2      40088b: mov    (%rax),%rax
     2      2      40088e: mov    -0x8(%rbp),%rdx
     .      3      400892: shl    $0x2,%rdx
     3      4      400896: add    %rdx,%rax
     1     29      400899: mov    (%rax),%eax
    28     33      40089b: mov    %eax,-0x14(%rbp)
    26    744    67: uint32_t sqrt = integer_square_root(val);
     5     17      40089e: mov    -0x14(%rbp),%eax
    12     14      4008a1: mov    %eax,%edi
     2    698      4008a3: callq  40080c <integer_square_root>
     7     15      4008a8: mov    %eax,-0x18(%rbp)
    39     74    68: if (sqrt * sqrt == val) {
     8     21      4008ab: mov    -0x18(%rbp),%eax
    13     31      4008ae: imul   -0x18(%rbp),%eax
    18     18      4008b2: cmp    -0x14(%rbp),%eax
     .      4      4008b5: jne    4008c4 <thread_scan+0x5b>
    10     90    69: __sync_fetch_and_add(args->num_perfect_squares, 1);
     4      7      4008b7: mov    -0x10(%rbp),%rax
     3      6      4008bb: mov    0x10(%rax),%rax
     3     77      4008bf: lock addq $0x1,(%rax)
    83     92    65: for (i = 0; i < args->size; i++) {
    74     76      4008c4: addq   $0x1,-0x8(%rbp)
     2      2      4008c9: mov    -0x10(%rbp),%rax
     .      7      4008cd: mov    0x8(%rax),%rax
     7      7      4008d1: cmp    -0x8(%rbp),%rax
     .      .      4008d5: ja     400887 <thread_scan+0x1e>
     .      .    72: }
     .      .      4008d7: leaveq 
     .      .      4008d8: retq