[LLVMdev] [RFC] Progress towards OpenMP support
hfinkel at anl.gov
Wed Sep 12 20:13:09 PDT 2012
On Tue, 11 Sep 2012 01:21:59 +0530
Sanjoy Das <sanjoy at playingwithpointers.com> wrote:
Thanks for working on this! Comments below...
> Hi all,
> I made some progress on implementing Hal's proposal  for
> implementing OpenMP support in LLVM. The patch I've attached just
> barely compiles, but I'd like to get some input on the design early on
> to prevent trouble later. I'd especially like some input on the
> following points:
> * Metadata is never mutated or dropped
> I think it is better to have an analysis pass that simply provides a
> consistent "view" of the current !parallel nodes instead of one that
> mutates the IR in order to make it consistent. By addRequired<> ing
> it (called ParallizationMetadata currently) in the lowering pass and
> only indirectly accessing the metadata through the analysis pass, we
> are assured that we don't parallelize regions with inconsistent
My rationale for proposing the self-consistent metadata solution was
that it seemed to be the safest option. If we simply insist that all
relevant passes use the ParallizationMetadata pass, without any
verification, then we could end up with parallelization-unaware passes
silently miscompiling code when parallelization is enabled. Do you have
a way of avoiding that?
> * No information is optional
> It simplifies the implementation greatly if we change the spec to
> assume this. I don't think this restriction shifts any significant
> complexity to the frontends.
> * Loops don't have lists of special handling regions
> Since loops are always nested inside parallel regions, can't we
> maintain this list in the parent region itself?
I'm not sure. We do need to make sure, for example, that an 'ordered'
region stays properly nested within its parent loop.
> * Tripping assertions vs. silently ignoring metadata
> I'm not very clear on when it is "right" to silently ignore metadata
> and when to assert. Right now I assert only when a !parallel MDNode
> has an incorrect number of children (this is one thing simplified by
> making all fields compulsory). The verifier probably needs to know
> about this constraint, something I haven't addressed yet. Should our
> assertions be stricter or must we not assert at all?
We should assert when an individual metadata entry is malformed; we
should drop non-self-consistent sets of metadata entries (as these
might have resulted from the actions of non-parallelization-aware
> (The code is also up on Github ).
>  http://lists.cs.uiuc.edu/pipermail/llvmdev/2012-August/052472.html
>  https://github.com/sanjoy/llvm/tree/parallel-md
> Sanjoy Das
Leadership Computing Facility
Argonne National Laboratory
More information about the llvm-dev