fail delinearization when the size of subscripts differs

Because the delinearization is not a global analysis pass, it will compute the
delinearization independently of knowledge about the way the delinearization
happened for other data accesses to the same array: the dependence analysis will
only trigger the delinearization on a tuple of access functions, and thus
delinearization may compute different subscripts sizes for a same array.  When
that happens the safest is to discard the delinearized information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@201866 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 14 insertions, 1 deletions

diff --git a/lib/Analysis/DependenceAnalysis.cpp b/lib/Analysis/DependenceAnalysis.cpp
index b74140db2c..6f9effa127 100644
--- a/lib/Analysis/DependenceAnalysis.cpp
+++ b/lib/Analysis/DependenceAnalysis.cpp
@@ -3193,10 +3193,23 @@ DependenceAnalysis::tryDelinearize(const SCEV *SrcSCEV, const SCEV *DstSCEV,
   DstAR->delinearize(*SE, DstSubscripts, DstSizes);
 
   int size = SrcSubscripts.size();
+  // Fail when there is only a subscript: that's a linearized access function.
+  if (size < 2)
+    return false;
+
   int dstSize = DstSubscripts.size();
-  if (size != dstSize || size < 2)
+  // Fail when the number of subscripts in Src and Dst differ.
+  if (size != dstSize)
     return false;
 
+  // Fail when the size of any of the subscripts in Src and Dst differs: the
+  // dependence analysis assumes that elements in the same array have same size.
+  // SCEV delinearization does not have a context based on which it would decide
+  // globally the size of subscripts that would best fit all the array accesses.
+  for (int i = 0; i < size; ++i)
+    if (SrcSizes[i] != DstSizes[i])
+      return false;
+
 #ifndef NDEBUG
   DEBUG(errs() << "\nSrcSubscripts: ");
   for (int i = 0; i < size; i++)