(j3.2006) (SC22WG5.4909) [ukfortran] Comment on a comment on the WG5 letterballot on N1947

[Malcolm Cohen]

Keith Bierman wrote:
>OpenVMS, as far as I know, has only run on processors which physical support 
>for overflow detection (Alpha, VAX). So this is not a very useful proof point 
>for the significant number of processors that do not have such hardware 
>support.

The vast majority of processors, including the x86 family, have such hardware 
support.

>>        2.  Detection of integer overflow conditions was part of John Reid's 
>> ENABLE block
>>proposal in 1994.
>
>John was proposing a language feature;
> he didn't have the burden of actually implementing it,
> so again this example
>proves nothing germane to Bill's objections.

John's ENABLE block was not shot down by people raising objections to integer 
overflow detection.

>Look at current GPU architecture manuals and hardware specifications for 
>examples of compute engines for which such handling could prove quite costly 
>(indeed, making moving computation off the CPU possibly pointless). GPUs are 
>similar (in many ways) to the old FPS style array processors and Cray style 
>vector machines (just not nearly as polished or easy to program).
>
>I'm not saying that these aren't useful or interesting language features. But 
>those without any "skin the game" w.r.t. >implementing either the hardware or 
>software should not blithely ignore the costs proposed.

(1) No-one is blithely ignoring anything.
(2) I do in fact know what it costs in hardware and software; in hardware the 
costs are not zero but they are miniscule compared to many other cute features 
that are being implemented.
(3) I have in fact implemented it in software without using the hardware 
support.

Or the snarkier but more amusing:

"Losing a rocket on the launch pad: a few billion dollars.
Being able to brag about your GPU: priceless."

Cheers,
-- 
................................Malcolm Cohen, Nihon NAG, Tokyo.