Re: Kernel checksum routime (>=5)x86 optimization

[Artur Skawina]

Kurt Garloff wrote:
> 
> By the way: Is it correct to use adcl in this piece of code???

yes

> Why do we want to add 1 instead of 1<<32 in case of overflow???

one's complement arithmetic (see rfc1071 for details on the checksum)

> > For instructions this simple the integer units are probably faster than
> > the instruction decoder anyway.
> 
> That might be the reason, yes.
> it might also be that only one load can be executed per cycle.

ppro+ breaks up instructions into something intel calls 'microops'.
There are three decoders, the first can do four microops per cycle,
the other two only one/cycle. so you can have up to three instructions
per cycle, but only one can be complex and the others must be simple.
For example that routine with "%adcl XX(%esi), %eax" replaced with
"addl  XX(%esi), %eax ; setc %ch; addb %ch, %bl" only does two
percent worse. With "addl  XX(%esi), %eax ; adcl $0, %ebx" it does
10% worse.

artur