We do want to allow LoadPRE to perform LICM-like transformations: we already consider PHI nodes to be negligible for

code size (making this transform code size neutral), and it allows us to hoist values out of loops, which is always
a good thing.

llvm-svn: 115205

1 files changed, 0 insertions, 13 deletions

diff --git a/llvm/lib/Transforms/Scalar/GVN.cpp b/llvm/lib/Transforms/Scalar/GVN.cpp
index de99c3f8b69..f1272d4956d 100644
--- a/llvm/lib/Transforms/Scalar/GVN.cpp
+++ b/llvm/lib/Transforms/Scalar/GVN.cpp
@@ -1545,19 +1545,6 @@ bool GVN::processNonLocalLoad(LoadInst *LI,
   assert(TmpBB);
   LoadBB = TmpBB;
 
-  // If we have a repl set with LI itself in it, this means we have a loop where
-  // at least one of the values is LI.  Since this means that we won't be able
-  // to eliminate LI even if we insert uses in the other predecessors, we will
-  // end up increasing code size.  Reject this by scanning for LI.
-  for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
-    if (ValuesPerBlock[i].isSimpleValue() &&
-        ValuesPerBlock[i].getSimpleValue() == LI) {
-      // Skip cases where LI is the only definition, even for EnableFullLoadPRE.
-      if (!EnableFullLoadPRE || e == 1)
-        return false;
-    }
-  }
-
   // FIXME: It is extremely unclear what this loop is doing, other than
   // artificially restricting loadpre.
   if (isSinglePred) {