Re: [Valgrind-users] Regarding nulgrind and on how to find out number of x86 instructions inserted?

[Nicholas N. <nj...@cs...>]

On Tue, 17 Oct 2006, Mohit Tiwari wrote:

> 1. I compared the runtimes of nulgrind with that of reduced memcheck, that
> performs no instrumentation and has its MC_(instrument) function simply
> return bb_in,
> and noticed that there is a significant difference in their running times.
> For gcc, with a lgred data input, the timing is 134s for nulgrind and 170s
> for reduced emcheck. Why could that be? Shouldn't their performance be
> similar?

Memcheck also replaces malloc, free and related functions with its own 
version.  That will make a difference, especially on an allocation-heavy 
program like gcc.  I expect the difference would be smaller on Fortran 
benchmarks, for example.

Also, Memcheck's event handlers (for new_mem_*, die_mem_*, etc) will still 
be running, eg. for when system calls execute.  You'll need to disable them 
too.

> 2. I wished to find out how the peformance of an instrumented program
> degrades depending upon the number of instructions that get inserted by the
> *_(instrument) function. Is there a way I can find out how many IRs and x86
> instructions got added in?

Try the --help-debug flag, and look at the output.  --trace-flags shows you 
the IR at every stage.  It doesn't count it though, you'll have to add code 
to do that yourself.

Nick