Possible explaination why large_real_kind_form_io_2.f90 is failing on PPC/OSX

[Tim Prince]

Dominique Dhumieres wrote:
> To make things clearer:
>
> (1) I am aware that using softaware emulation is slower tha using hardware,
> so I do expect that using real(16) will lead to much slower code tha using
> real(8), though a significant part of the increased time may come from the
> memory bandwith (the same may apply to the use of real(10) especially with
> inappropriate padding).
>
>   
The reason for padding real(10) out to 16 bytes is to limit memory 
bandwidth consumption to what it would be with real(16) (or real(8) 
parallel instructions).  This assumes that you don't use options which 
reduce alignments (-Os, -mpreferred-stack-alignment=8).  When you permit 
it to do so, the compiler will assure that real(10) doesn't straddle 
cache line boundaries, which would be very expensive in memory access 
related operations.