[lug] Various Arch/Compiler Binaries living together

[Jeffrey Siegal]

Timothy C. Klein wrote:
> While I would not be overly surprised if you two are correct, I am not
> going to take your word for it :-)  If it is true, though, it would seem
> to imply the following:  the only difference between the i386 and
> Athlon/Pentium III/Pentium IV really boils down to clock speed (and I
> guess maybe memory bandwidth / cache size).  I am not sure I buy that.

There was more of a difference on earlier architectures, the original 
Pentium being probably the biggest because of the need to do pairing 
correctly, and avoid pipeline stalls.  Since the Pentium-II, precise 
instruction selection has become much less important for two reasons: 1) 
out of order execution gives the processor more freedom to execute 
suboptimal instruction sequences closer to optimally, and 2) performance 
is more limited by memory and cache performance.

That's not to say processor-aware optimziation makes *no* difference, 
but it makes very little difference for most applications.

> There *are* new, and more effiecent, instructions in the newer chips
> (3DNow, SSE, SSE2, MMX, etc..., to name those that have catchy names).
> So for the average application such things make no difference?

Pretty much true.  Those make a difference for certain data-intensive 
operations (compression, etc.), but for most apps, they don't.  Also, 
GCC won't generate them anyway.

> Have any links to scientific analysis?

I don't, but I have over a decade experience with performance analysis, 
and my expert opinion is that the original poster was right.  If that's 
not good enough, someone else might have links.