[OpenMP] Add regression test: this verifies that only one call to the runtime for reduction is done in a combined construct with distribute, without simd, and with at least one pragma accepting reduction as a clause.

That's correct, debug printfs were reporting zero threads to be used in SPMD region instead of the actual number used.

Committed revision 321759.

Committed revision 319696.

Committed revision 318692.

Committed revision 317893.

Committed revision 314905.

Committed revision 302031.

Re-Committed revision 301340.

Committed revision 301223.

[OpenMP] Refactor after split of original patch in two patches (sema and cg). Transform lambdas to static functions, where possible, define default constructors, format, and update against trunk.

Committed revision 300802.

Also reported by bbot slave:

Committed revision r295497

Address comments on initial patch after splitting. Added a dependency with sema patch.

About this:

I will split this patch into smaller patches, but I will keep this open in case I need to refer back to it: some of the changes that can go into independent patches require a justification and this patch shows it.
Thanks for your comments!

[OpenMP] Updating the patch to show full context

Committed revision r294195

Committed to r294026

Commited r290898.

Committed revision 275282.

Committed revision 275271.

[OpenMP] Add regression test for ast, remove 'brief's, and rebase.

[OpenMP] Add regression test to check correctness of ast building, remove 'brief's, and rebase.

[OpenMP] Remove generation of capture (move to codegen) and rebase.

[OpenMP] Apply comments: remove creation of private variable and update regression tests to include references to pointers.

No positive tests for the construct

[OpenMP] Remove 'brief' comment tags and change error report to standard 'expected' form.

[OpenMP] Remove private variable creation from this patch as requested by comments and rebase.

I think that Kelvin is right. This is what that sentence refers to:
int * bla = ..;
#pragma omp target data use_device_ptr(bla)
{

.. bla .. // <-- this is the 'reference' that sentence is about

}

[OpenMP] Delete trailing whitespaces in regression test file and rebase.

Yes, the purpose of these patches was to check correctness of code gen for the related clauses when targeting nvptx exclusively.
Since I wrote these tests, the base support changed to generate same code on host and any device type, and it seemed to me that the need for a differentiated regression test was not there any more (happy to be told otherwise).

Resubmitted at revision 273884 after fixes.

Committed revision 273705.

Thanks for the hint - I have updated the diff to use Context.getSizeType(). Please let me know if this is what you meant.

[OpenMP] Change type of distribute lower and upper bound fields to portable size_t (Context.getSizeType()).

Thanks for these further comments. I have just updated the diff applying the suggestions.

[OpenMP] Address comments: emit statements in code gen as an inlined directive (no actual code gen, just a placeholder); change type of parallel outlined function parameters to pointer to enable casting in code gen to the right type, instead of inferring value from loop analysis.

[OpenMP] Update patch to trunk and to reflect comments. Remove fields from OMPDistributeParallelFor class and will add them to the related code gen patch.

Thanks very much for the quick review. I am about to update the patch reflecting your comments and to the latest trunk.

[OpenMP] Add regression test for distribute used as if-clause modifier and check that an error is generated

Codegen of region bodies should not be specialized based on target.

Committed revision 267972.

[OPENMP] Remove unnecessary templating of function.

Great, thanks for the hint. I do much prefer your version of the function. I am updating the diff.

[OPENMP] Remove template for intervening statement and only allow CompoundStmt.

Thanks for the review. I addressed your comment in a new version of the revision.
Samuel already reviewed the patch, as you can see in the message exchange. I am very happy to apply any design change you can think of.

[OPENMP] Rebase patch on new support for codegen for teams in nvptx target. private and firstprivate clauses are now dealt with in the teams patch and this one only becomes regression tests.

Committed revision 265304.

Hi Alexey

[OPENMP] Even though this patch was already accepted in its previous form, comments on depending patch D18286 (http://reviews.llvm.org/D18286) revealed that a new approach for this patch was necessary. Instead of committing something that I have to change anyway later on, I decided to provide a new version of this base patch. Please review it again and let me know about any comments you may have.

If I understand correctly the problem, I would like to add something on top of Samuel's comment.
My understanding is that Alexey is suggesting that we pass a reference type to kernels for every pointer input parameter, regardless of the actual type and data sharing attribute of the related variable. Ignore the remained of this comment if this is not the case.

Simplify the search function and generalize by using a template.

Thinking more carefully, I believe you are right - no, it is not. I guess we would not call at all shouldEmitInlinedTeamsDirective for the case of paralle.

Hi Jonas

To make clear my comment I an updating the patch following Alexey's comment and modifying it to make it work.
You can see that now we have the following check:
CGM.getTarget().getTriple().getArch() == llvm::Triple::nvptx ||
CGM.getTarget().getTriple().getArch() == llvm::Triple::nvptx64

Hi Alexey

Committed revision 263837.

Update against trunk as dependence was committed to trunk.

Committed revision 263654.

Modify diff to reflect comments.

[OPENMP] Add missing check macros in regression test and remove reference to firstprivate (will come in next patch).

[OPENMP] Add testing for 32-bit arch

I can actually add a regression test that checks that the compiler does not break when using a global in a teams region, even without declare target, if that is what is required here.