Re: GNAT (GCC) Profile Guided Compilation

[Keean Schupke <keean.schupke@googlemail.com>]

On Friday, 29 June 2012 10:34:19 UTC+1, Dmitry A. Kazakov  wrote:
> On Fri, 29 Jun 2012 02:17:19 -0700 (PDT), Keean Schupke wrote:
> 
> > Anyone have any ideas why profile guided compilation performs so poorly
> > with Ada compared to C++?
> 
> Probably because you programmed it using low level C-ish stuff like
> addresses instead of accessing data directly. In general, if you write a C
> program in Ada (or any other language), you cannot expect it do better than
> in C. To beat C, the precondition is to design it in Ada way.
> 
> -- 
> Regards,
> Dmitry A. Kazakov
> http://www.dmitry-kazakov.de

I don't think you are right here, but I would be happy to be proved wrong. 

In general I find the choice of algorithm dominates performance. In this case most of the performance is dominated by a disjoint-set (union-find) algorithm operating on an array of data. The data nominally represents a 2D space (although I am using a 1D array to avoid the cost of multiplication on every access).

Most operations access the neighbouring cells, hence require a +/- one cell, or +/- one row operations.

Maintaining the address of the current cell is faster than using indexed addressing (mainly because of the extra parameter that needs to be passed around increasing the pressure on register allocation).

I would be interested to understand how an algorithm can be implemented in a more Ada way?


Cheers,
Keean.