[LCP]register question

[David Spencer]

Erik Mouw wrote:
> 
> To wake you up from a pretty dream: Pentium class CPUs (even the P-4)
> are not at all considered modern. It's a design that dates back from
> the early eighties (8086), and it has legacy things like segmentation,
> a real mode, and an overly extensive instruction set. Even worse is
> that it has only four general purpose registers (eax - edx), of which
> three are also used for things like indexing and shift counts, which
> leaves you with only one real general purpose register. Compare that
> with other CPUs which usually have a RISC architecture with plenty of
> general purpose registers (at least 16).

I thought the segmentation, real mode and the rest of the legacy stuff
were "emulated" and that the "normal" operation of the Pentium was in
fact relatively modern (e.g. the memory model of the Penty is actually
flat, but you can do segmented if you want).  Of course just having the
legacy stuff on chip will slow the thing down by virtue of the extra
decoding.

> The drawback is that you can get less optimal code because you force
> the compiler to put the wrong variable in a register

No - the register keyword is only a hint, which the compiler does not
have to use at all.  These days though most fairly frequently accessed
memory will be cached anyway so moving it a few mm closer to the ALU by
putting it into registers will be pretty futile anyway.  There are other
ways of speeding stuff up: use a better algorithm, rewrite the tight
stuff in assembler etc, and if you are already using optimal code, buy a
faster processor, or buy a multiprocessor system so that your tight loop
isn't interrupted by keyboard scans.  Or, ultimately, reimplement the
software in hardware.

But yes the original point is still valid: you have to profile an
application to find out the bottlenecks before you can streamline it.

Dave.