BBVectorize, when using VTTI, should not form types that will be split.

This is needed so that perl's SHA can be compiled (otherwise
BBVectorize takes far too long to find its fixed point).

I'll try to come up with a reduced test case.

llvm-svn: 166738

1 files changed, 19 insertions, 0 deletions

diff --git a/llvm/lib/Transforms/Vectorize/BBVectorize.cpp b/llvm/lib/Transforms/Vectorize/BBVectorize.cpp
index 61e8d735e41..95b78322caa 100644
--- a/llvm/lib/Transforms/Vectorize/BBVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/BBVectorize.cpp
@@ -817,6 +817,15 @@ namespace {
                                                  IAddressSpace);
           if (VCost > ICost + JCost)
             return false;
+
+          // FIXME: We don't want to fuse to a type that will be split, even
+          // if the two input types will also be split and there is no other
+          // associated cost. This check depends on the fact
+          // that the current implementation of getMemoryOpCost returns only
+          // the type-splitting cost.
+          if (VCost > 1)
+            return false;
+
           CostSavings = ICost + JCost - VCost;
         }
       } else {
@@ -831,6 +840,16 @@ namespace {
 
       if (VCost > ICost + JCost)
         return false;
+
+      // FIXME: We don't want to fuse to a type that will be split, even
+      // if the two input types will also be split and there is no other
+      // associated cost. This check depends on the fact
+      // that the current implementation of getMemoryOpCost returns only
+      // the type-splitting cost (and does nothing else).
+      unsigned VTypeCost = VTTI->getMemoryOpCost(I->getOpcode(), VT1, 0, 0);
+      if (VTypeCost > 1)
+        return false;
+
       CostSavings = ICost + JCost - VCost;
     }