[Polly] Make ScopInfo handle PHI nodes in exit block

[Meinersbur]

Summary

after iterating over all blocks of the region, also iterate over the exit block which normally does not belong to the region. The exit block may contain PHI nodes that will later be moved into the region.

Such PHI nodes cannot yet occur without changes upstream (TempScopInfo, ScopDetection) which will follow.

[jdoerfert]

I am unsure if I missed something but this looks to me like two code extractions into helper functions and hidden in line 1693ff some code that does create an extra statement. I do not mind the first part (extraction) but I have some comments regarding the second:

1. Please commit them separately
2. For the extra statement I think we need at least one test case as well as some motivation how this actually helps us.

[Meinersbur]

In D11869#219805, @jdoerfert wrote:

2. For the extra statement I think we need at least one test case as well as some motivation how this actually helps us.

This patch alone has no effect. See summary.

[Meinersbur]

Rebase to http://reviews.llvm.org/D11878

[grosser]

LGTM. Please merge with other changes to ensure we get test coverage at the first commit.

[Meinersbur]

Merged into D11869

[Meinersbur]

Merged into D11870, not D11869

lib/Analysis/ScopInfo.cpp

   if (!Schedule)
     Schedule = isl_schedule_empty(getParamSpace());
 
+  // Build the statement for PHIs in the exit block (never nested in a
+  // scopped-loop)
+  isl_schedule *StmtSchedule =
+      buildBBScopStmt(getRegion().getExit(), TempScop, getRegion(), NestLoops);
+  Schedule = ScheduleCombineSequence(Schedule, StmtSchedule);
+
   realignParams();
   addParameterBounds();
   simplifyAssumedContext();