Re: [OT] GCC vs Intel C++ compiler benchmark

[Tim Prince <tprinceusa at mindspring dot com>]

On Sunday 27 January 2002 04:45, Florian Weimer wrote:
> Claus Fischer <claus.fischer@clausfischer.com> writes:
> > I'm not concerned with that but with the notable OS difference
> > between Linux (SuSE 7.3) and Windows (XP Pro). IMHO a CPU bound
> > benchmark should see less than 1 % influence from OS and C library.
> > The graphics shows roughly 7% better performance on Windows for the
> > same (Intel) compiler.
> >
> > Are there any ABI differences which would justify such a difference?
>
> A few years ago, I noticed the following difference: The Microsoft C
> library on Win32 switches the FPU to 64 bit precision.  GNU/Linux uses
> 80 bit precision, which used to be slower. 
The Intel compilers follow the usual Windows vs linux practice, as you point 
out.  Thus, x87 divide and square root are significantly slower and more 
accurate on linux.  Choosing the sse code generation options eliminates this 
difference. 
> Depending on the working
> set, VM performance can affect benchmark results, too.  For example,
> Windows 95 taints the first-level data cache of the processor
> regularly, measurably decreasing performance.
>
> However, I haven't got any current data.
In my experience, Win2K takes better advantage of L2 cache than linux, 
provided that the machine is dedicated to a single task, the time slice is 
set long, and nothing special is done in linux to take advantage of that.  At 
most a 2% differential there.