gprof

[Ken Raeburn]

On Sep 20, 2005, at 06:31, Michael Trimarchi wrote:
> >But also, if you care about the accuracy of the results, you  may  
> need to modify the C runtime support code for >profiling, which   
> typically updates the per-function data in a manner that is not   
> thread-safe.
> may you explain more precisely this point?
> Regards
> Michael

With basic profiling, the runtime support code keeps track of how  
often the CPU program counter is in a given range of values, with  
fairly fine granularity.  Later this table is dumped out, and (g)prof  
interprets it in combination with symbol information from the  
executable.  For example, addresses XXX through YYY correspond to  
function A, and so many ticks at a certain frequency were counted  
with the PC in that range, so here's the amount of time the program  
spent in that function.  But that may be inaccurate in multithreaded  
programs if the counter is implemented as read counter value N,  
increment value, another thread runs for a bit and changes the  
counter, store counter value N+1; you've just lost the change made by  
the other thread.

For graph profiling, you also need data recorded on entry to a  
function indicating where the function was called from, and how many  
times it was called from each call site.  Since you could have  
arbitrarily many such call sites, this is likely to involve dynamic  
memory allocation, walking through some data structures, etc.  If  
it's not done just right, it might even result in crashes in  
multiprocessor, multithreaded situations, if you're really unlucky  
with the timing.

I've never gotten around to modifying the support code to try to make  
it thread-safe.  Using mutex locks would be the obvious approach, but  
probably kind of expensive compared to some of the atomic operations  
a few processors have available, or "store if another cpu or thread  
hasn't stored here, and set condition codes to tell me", etc.

You might just get kind of lucky with the existing support code,  
though, your program might not crash, and the numbers might even be  
vaguely accurate...

Ken