[PATCH] D17381: [Polly] [RFC] Memory reference level dependency analysis

[Tobias Grosser via llvm-commits]

grosser added a comment.

Comments inline.


================
Comment at: lib/Analysis/DependenceInfo.cpp:305-309
@@ -266,4 +304,7 @@
+
+  if (isl_union_map_is_empty(AccessSchedule) &&
+      OptAnalysisLevel == STATEMENT_LEVEL_ANALYSIS) {
     isl_union_map_free(AccessSchedule);
     Schedule = S.getScheduleTree();
   } else {
     auto *ScheduleMap =
----------------
grosser wrote:
> etherzhhb wrote:
> > Can we improve this if-else?
> Yes, see https://llvm.org/bugs/show_bug.cgi?id=26320 for a first description why the second code path is there. Johannes, who wrote this code, might be able to tell you better what would be needed to move reduction dependences to schedule trees.
> 
> 
@ether. I just realized you are using the second code path for reference level dependence analysis. I would prefer to use the first code path. Dependence analysis on schedule trees can be significantly faster (we observed this) as the analysis can exploit the tree structure. This is why we moved to schedule trees in the first place.


Repository:
  rL LLVM

http://reviews.llvm.org/D17381