enters_scope can reach a CodeBlock node

[sergisiso]

Doing:
psyclone -s ukca_aero_ctl.py $GLOMAP/ukca_aero_ctl.F90 -o ukca_aero_ctl_trans.f90

Fails while lowering the OpenMP with:
AttributeError: 'CodeBlock' object has no attribute 'enters_scope'

[JonElsey]

Hi Sergi,

I've tried running this on ukca_aero_ctl

https://code.metoffice.gov.uk/trac/ukca/browser/main/trunk/src/science/core/aerosols/glomap/ukca_aero_ctl.F90

using

https://code.metoffice.gov.uk/trac/lfric_apps/browser/main/branches/dev/alanjhewitt/r14198_ukca_glomap_cpu_openmp/applications/lfric_atm/optimisation/minimum/transmute/science/ukca/src/science/core/aerosols/glomap/ukca_aero_ctl.py )

This worked at PSyclone 3.1.

PSyclone 3.2 at master previously gave the error mentioned above. With this branch, I get:

  File ".../psyclone/src/psyclone/psyir/symbols/symbol_table.py", line 1077, in lookup
    raise KeyError(f"Could not find '{name}' in the Symbol "
KeyError: "Could not find 'item' in the Symbol Table."

I believe that the error is due to a derived type (ukcaD1codes(n)%item) inside a CodeBlock.

(Pdb) access.description
"'! PSyclone CodeBlock (unsupported code) reason:\n!  - Unsupported statement: Write_Stmt\nWRITE(umMessage, '(A16,3i4,3e14.4)') 'UKCA_MODE diag: ', k, n, ukcaD1codes(n) % item, SUM(mode_diags(:, :, :, k)), MAXVAL(mode_diags(:, :, :, k)), MINVAL(mode_diags(:, :, :, k))\n'"

This variable is fine to be PUBLIC, if there is an override/workaround.

[sergisiso]

Ok, the issue is in the conversion from signature name to a symbol in:

  File "psyir/nodes/data_sharing_attribute_mixin.py", line 225, in infer_sharing_attributes
    symbol = access.node.scope.symbol_table.lookup(name)

However, in addition this there is a clear performance problem. The fparser2_reader (not fparser) takes more than a minute to complete. The profiler showed that the problem was in get_callees, and I thought it is probably my change to references_accesses that is now looking at the subroutine declarations, so I shortcut that change and it went down to seconds.

But then I brought it to master (it was 2 PRs behind) and it when back up to minutes to complete (maybe due to #3067 @arporter @hiker ?), so I also shortcuted the get_callees implementation, and it when back down to seconds.

The patch to make it fast is:

+++ b/src/psyclone/psyir/nodes/call.py
@@ -308,6 +308,7 @@ class Call(Statement, DataNode):

         '''
         var_accesses = VariablesAccessMap()
+        return var_accesses

         # The RoutineSymbol has a CALL access.
         sig, indices_list = self.routine.get_signature_and_indices()
@@ -492,6 +493,7 @@ class Call(Statement, DataNode):
             limitation prevents definite determination of the target routine.

         '''
+        return []
         rsym = self.routine.symbol
         if rsym.is_unresolved:
             # Search for the Routine in the current file. This search is

Obviouly this loses functionallity, so we need to think of a proper fix

[sergisiso]

I need to try merging #3163 maybe it is all due to attempting to finding external files on-the-fly?

[sergisiso]

Another idea is to differentiate between get references and get reference_access.

In several places we just need to know what symbols are touched, but doing it through reference_accesses means we need to follow get_callees (to know the intents), but also do the map and map updates to calculate the proper access.

A simple get references would only use a more efficient set, and not bother with expensive updates or finding declarations.

[arporter]

I'm surprised #3067 caused (possibly) such a large jump - I wouldn't have expected it to be much worse than what we had before. There may be some unintended behaviour in there.

[sergisiso]

I don't discard an uninteded behaviour, the slowdown was not in the reader anymore, but the profiler was not very informative. I need some time to understand this.

[sergisiso]

I will tackle the performance issue separately in #3183