Re: [Gomp-discuss] The tree nodes ...

[Diego Novillo]

On Mon, 10 Feb 2003, Lars Segerlund wrote:

>  ( also want to vent my opinion that we should make the distinction 
> between section running on the same machine and sections slated to 
> target another cpu ).
> 
At compile time you can either tell if a section is sequential or
concurrent.  How will this distinction even help the compiler?

>  As for the discussion about HPC gnu initiative, I think that we can 
> postpone that and rather work with the existing standards, such as 
> Threading libraries and open MPI or PVM implementations, ( the last two 
> in the future ), but what it actualizes is the need not to restrict 
> ourself to much.
> 
You have to have a way of getting hints from the front end that a
section needs to run in parallel.  I would suggest that we first
solve the problem of thread concurrency.  Message passing
concurrency is fundamentally different both at runtime and
compile time.  The memory model plays an important role.  Even
with thread concurrency you have to stick by a particular memory
model.

My suggestion is to keep implementing the FE changes to accept
OMP #pragmas.  We can make our lives simpler by initially
accepting a subset of all the OMP pragmas.  For instance, we
could start with just the pragmas that mark parallel loops,
parallel sections and private/shared variables.

With that subset you can start implementing a runtime and the
minimal compiler support to allow people to write concurrent code
in GCC.  Once you get that going, you can start refining the
pragmas and getting more elaborate with the runtime and
optimizations.

Also, let's not get caught up in a debate about the OpenMP
standard and whether or not it will be accepted in mainline GCC.
Let's just implement something that has value, once we have
something concrete to show we can cross the next bridge.  No
point having vacuous flamewars.


Diego.