compile speed on lfs 5.0

John Gay johngay at eircom.net
Tue May 11 08:24:12 PDT 2004


On Sat 08 May 2004 20:40, Ken Moffat wrote:
> On Sat, 8 May 2004, John Gay wrote:
> > Actually, as I understand it, gcc-3.x makes major improvements in
> > optimizing for MMX ans SSE instructions in the Pentium family. As these
> > are non-trivial to optimize for, I imagine it takes much larger memory
> > structures to search for optimizing. I would consider this a required
> > feature that was missing from the 2.9x family of gcc, not feature bloat.
>
> Gcc-3, particularly the 3.3 series, is much more standards-compliant for
> c++ [ if I've understood the issues correctly ].  gcc-3.3.3 is much
> better at coping with combinations of CFLAGS which can sometimes create
> internal errors in previous 3.3-series versions.  The resulting code may
> even run faster.  The downside is that compilation is much slower.
>
I know that the progression from gcc-3.0 through gcc-3.3 has been 
concentrating on improving c++ compliance, but hte major revision change from 
gcc-2.x to gcc-3.x was much improved optimisation for MMX/SSE instructions. 
Prior to the gcc-3.x series, MXX/SSE optimization had to be purposely coded 
in assembly before it could be optimized. This would account for the much 
longer compilation times, as gcc now has to check for matrix-type 
calculations and see if MMX/SSE instructions can be used. This would require 
very large memory structures to verify the calculations can be optimized, 
which is what slows gcc down so much in the first place.

Cheers,

	John Gay




More information about the lfs-support mailing list