Re: MPI code error raises a question

Richard Maine wrote:
> glen herrmannsfeldt <gah@[EMAIL PROTECTED]
> wrote:
> 
>> Richard Maine wrote:
>> (snip)
>>
>>> What you are looking for is the VOLATILE attribute. Memory addresses
>>> that get updated via means independent of the Fortran program are
>>> exactly the kind of thing VOLATILE is about.
>> That is the way it looks to me, for VOLATILE:
>>
>>    "An allocatable object with the VOLATILE attribute may
>>     additionally have its allocation status, dynamic type
>>     and type parameters, and array bounds changed by means
>>     not specified by the program."
>>
>> I thought for MPI that only the data would change,
> 
> Most likely, but that seems irrelevant. Something doesn't actually
> *HAVE* to do every single thing that VOLATILE allows in order to be
> appropriate for VOLATILE. Many things can use VOLATILE MPI is one of
> them.
> 
>> and that
>> (even though intended for Fortran ASYNCHRONOUS I/O) that
>> ASYNCHRONOUS might be more appropriate.

I think people, especially Glen?, are confused about what
ASYNCHRONOUS means.  It basically provides a set of restrictions
on the user and hardly any restrictions on the compiler.  If
a variable is in a pending input/output sequence (meaning it
has the ASYNCHRONOUS attribute and is in an unWAITed for
I/O transfer) then the user is not allowed to reference or define
it if it's in an input transfer and not allowed to define it if
it's in an output transfer.  (F2003 page 195, around line 10,
gives the gory details.)

This means that in a code snippet like
       asynchronous  ::  x,y
          ...
       x = y + y
the compiler doesn't have to do anything special.  It is the
users responsibility to guarantee that x and y aren't in an
active i/o statement of the wrong kind.

This is totally different from VOLATILE where the compiler is
encouraged to fetch variables from memory and store them back
as soon as it can.

About the only compiler restrictions for ASYNCHRONOUS is
prevention of code motion across WAIT statements (and this
applies to any call statement where the compiler can't
analyze the callee well enough) and an effective prohibition
of copy-in/copy-out argument passing of asynchronous variables.



Dick Hendrickson

> 
> So MPI doesn't fit VOLATILE because it doesn't use every possible arcane
> part of VOLATILE, but it does fit ASYNCHRONOUS even though it doesn't
> have anything to do with any of the statements that relate to
> ASYNCRONOUS? The logic here escapes me.
>