trapflag-trace-vs-ptrace-learn

exploring the use of ptrace vs the x86 trapflag to step through every instruction of a program.

there are three small programs executing a certain number of instructions:

• count-ptrace.c uses ptrace to step through every instruction, executing ~1e6 instructions
• count-trapflag.c uses the x86 trap flag and signal handlers to execute ~1e7 instructions
• loop.c executes ~1e9 instructions without any instrumentation

These are the execution times on my netbook (32-bit Intel Atom N450, 1.66GHz):

• 1e6 instructinos in 61.7s in count-ptrace.c (16.2KHz)
• 1e7 instructions in 38.1s in count-trapflag.c (262.5KHz)
• 1e9 instructions in 0.61s in loop.c (1,639MHz)

This implies the following slow-downs of the different instrumentation schemes:

• 101,172x slower with ptrace
• 6,244x slower with trapflag and x86

This implies using the trapflag+signal handlers is about 16x faster than using ptrace.