Counting machine instructions of a process using PTRACE_SINGLESTEP

Question

Hi,

on a Linux machine, I am using ptrace with the PTRACE_SINGLESTEP parameter to count the number of machine instructions of a program. I followed this article: http://www.ncsu.edu/it/mirror/ldp/LDP/LGNET/81/sandeep.html.

However, the result seems odd to me. For a very simple program, over 95000 machine instructions are counted. The test program is

int main(void) { return 23; }

What's going on here? Is the code from the article wrong? (I can't see what's wrong with it.) If not, what causes such a simple program to require >95000 instructions?

Answer 1

It's due to something called "software bloat". You have to initialize and finalize stdio, and maybe even some threading code that bled into the standard C runtime. If you read a little further and profile it you may find out exactly what. Or you could just read the source.

Update: Actually, I realized later that you have probably been tracing through the operation of the dynamic linker, which has a lot of work to do. I see that someone left such a comment, so I upvoted the comment. If you didn't link the program statically, then both of our original answers were basically wrong.

Answer 2

The C program you're compiling is linked to C library. It contains the _start symbol which the program execution starts from. At that point, C library initializes itself and eventually calls main. After main returns, the control flows back to _start and there are a bunch of other instructions to execute and return the program return value. Note that using PTRACE_SINGLESTEP successively doesn't count the number of compiled instructions. It counts the number of executed instructions. That means 95k instructions are executed before entering main, when executing main and after exiting main.