(closes #2684) Fix the failure case when we have a WHERE construct with no array index notation.

[LonelyCat124]

No description provided.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.86%. Comparing base (70f9793) to head (9340fd6).
Report is 35 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2687      +/-   ##
==========================================
- Coverage   99.86%   99.86%   -0.01%     
==========================================
  Files         354      354              
  Lines       49082    49046      -36     
==========================================
- Hits        49018    48982      -36     
  Misses         64       64              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[LonelyCat124]

@sergisiso ready for a first look, I'm not super sure on the implementation details so its possible there's some things we're not testing for that i've missed.

[sergisiso]

Thanks @LonelyCat124 I think it's a step in the right directions. See comments inline but essentially I would like to see some more testing about what happens with "references" where we can not figure out its type ow when we have different kinds of intrinsics.

[LonelyCat124]

I think this is ready for another look @sergisiso . The outstanding thing I've not addressed is intrinsics (which were already "forbidden" in fparser2 before this change) - I'm not sure if/when it makes sense for them to exist.

[sergisiso]

@LonelyCat124 wrt the failing integration test, as we discussed last week, we need to do the following change:

         # For performance in lib_fortran, mark serial routines as GPU-enabled
         if psyir.name == "lib_fortran.f90":
-            if not subroutine.walk(Loop):
-                try:
-                    # We need the 'force' option.
-                    # SIGN_ARRAY_1D has a CodeBlock because of a WHERE without
-                    # array notation. (TODO #717)
-                    OMPDeclareTargetTrans().apply(subroutine,
-                                                  options={"force": True})
-                except TransformationError as err:
-                    print(err)
+            if subroutine.name.lower().startswith("sign_"):
+                OMPDeclareTargetTrans().apply(subroutine)

In omp_gpu_trans.py, acc_loops_trans.py, and acc_kernels_trans.py. I checked that it works and doesn't degrade the performance. The changes are:

• We don't identify the routine by having no loops anymore, we identify them by name.
• The force is gone because there is no Codeblock anymore as the WHERE is expanded. Well done.
• I removed the except because if this fails it will produce code that doesn't compile, so we better fail early

[LonelyCat124]

I've set the integration tests running again now with those changes.

[sergisiso]

@LonelyCat124 I made changes in some comments and pointed out some additional issues in #2722. But since this already fixes some of the reported issues and it passes the integration tests, it is a good amount of progress to merge this and defer remaining issues to a future PR. This is ready to merge.