gistfile1.txt

Percent |      Source code & Disassembly of a.out
------------------------------------------------
         :
         :
         :
         :      Disassembly of section .text:
         :
         :      0000000000400869 <thread_scan>:
         :        size_t size;
         :        size_t* num_perfect_squares;
         :        mutex_t * num_perfect_squares_lock;
         :      } thread_arg_t;
         :
         :      void* thread_scan(void* void_arg) {
    0.00 :        400869:       push   %rbp
    0.00 :        40086a:       mov    %rsp,%rbp
    0.00 :        40086d:       sub    $0x28,%rsp
    0.00 :        400871:       mov    %rdi,-0x28(%rbp)
         :              thread_arg_t* args = (thread_arg_t*) void_arg;
    0.00 :        400875:       mov    -0x28(%rbp),%rax
    0.00 :        400879:       mov    %rax,-0x10(%rbp)
          :              size_t i;
          :
          :              for (i = 0; i < args->size; i++) {
     0.00 :        40087d:       movq   $0x0,-0x8(%rbp)
     0.00 :        400885:       jmp    4008c9 <thread_scan+0x60>
          :                      uint32_t val = args->input[i];
     3.73 :        400887:       mov    -0x10(%rbp),%rax
     0.00 :        40088b:       mov    (%rax),%rax
     1.58 :        40088e:       mov    -0x8(%rbp),%rdx
     0.00 :        400892:       shl    $0x2,%rdx
     1.58 :        400896:       add    %rdx,%rax
     0.53 :        400899:       mov    (%rax),%eax
    14.17 :        40089b:       mov    %eax,-0x14(%rbp)
          :                      uint32_t sqrt = integer_square_root(val);
     1.82 :        40089e:       mov    -0x14(%rbp),%eax
     8.09 :        4008a1:       mov    %eax,%edi
     2.15 :        4008a3:       callq  40080c <integer_square_root>
     2.68 :        4008a8:       mov    %eax,-0x18(%rbp)
          :                      if (sqrt * sqrt == val) {
     1.96 :        4008ab:       mov    -0x18(%rbp),%eax
     5.60 :        4008ae:       imul   -0x18(%rbp),%eax
     6.65 :        4008b2:       cmp    -0x14(%rbp),%eax
     0.00 :        4008b5:       jne    4008c4 <thread_scan+0x5b>
          :                        __sync_fetch_and_add(args->num_perfect_squares, 1)
     3.59 :        4008b7:       mov    -0x10(%rbp),%rax
     1.77 :        4008bb:       mov    0x10(%rax),%rax
     0.91 :        4008bf:       lock addq $0x1,(%rax)
          :      void* thread_scan(void* void_arg) {
          :              // TODO(awreece) Copy locally so dont interfere with each ot
          :              thread_arg_t* args = (thread_arg_t*) void_arg;
          :              size_t i;
          :
          :              for (i = 0; i < args->size; i++) {
    38.10 :        4008c4:       addq   $0x1,-0x8(%rbp)
     0.72 :        4008c9:       mov    -0x10(%rbp),%rax
     0.29 :        4008cd:       mov    0x8(%rax),%rax
     4.07 :        4008d1:       cmp    -0x8(%rbp),%rax
     0.00 :        4008d5:       ja     400887 <thread_scan+0x1e>
          :                      uint32_t sqrt = integer_square_root(val);
          :                      if (sqrt * sqrt == val) {
          :                        __sync_fetch_and_add(args->num_perfect_squares, 1)
          :                      }
          :              }
          :      }
     0.00 :        4008d7:       leaveq 
     0.00 :        4008d8:       retq