Re: xdelta speedups (was Re: [Monotone-devel] updates to net.venge.monotone.experiment.performance)

[Graydon Hoare]

Eric Anderson wrote:

> My best guess was that widen was there to make sure that when widening
> a signed to an unsigned the right thing happened, but I wasn't sure it
> seemed unnecessary for an unsigned to unsigned widen.

Yup.

The main part of the function actually is the "implicitly convert to the 
explicit widen<>() argument type, to use in the explicit conversion" 
part. If you just do static_cast<u32>(somestring[i]) you are doing a 
sign extension on a platform with signed chars, because somestring[i] 
has char type.

You could write static_cast<u32>(static_cast<u8>(somestring[i])) 
instead, but I felt like widen<u32,u8>(somestring[i]) was more pleasant 
on the eyes.

(I don't actually know that static_cast<u32>(u8) is guaranteed to not 
sign-extend, because I can't find a precise definition of the conversion 
rules; but I'd be pretty surprised if there was a platform anywhere 
where it did. It's probably OK to assume unless it's in netsync code, 
where "being wrong" means "buffer overrun")

-graydon