stepanov on i686-pc-linux-gnu

[Joe Buck]

> |> Sigh.  I'm so used to looking at Sparc assembly with the delayed
> |> branch.
> |> 
> |> Nevertheless, there are two memory accesses in the loop
> 
> No, there is only one.  'leal' only calculates the effective address,
> without any memory access.

Sigh.

Rather than embarrassing myself further, I ask someone who is familiar
with x86 assembly to compile the Stepanov benchmark with -O2 -S,
and compare the slow accumulate loops with the fast accumulate loops
and post an analysis of what's going wrong in the slow loops.