1 files changed, 28 insertions, 50 deletions

diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 708d36f6f978..091ad8b18516 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -8420,17 +8420,25 @@ SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
                             Intrinsic::ppc_altivec_lvsl);
       SDValue PermCntl = BuildIntrinsicOp(Intr, Ptr, DAG, dl, MVT::v16i8);
 
-      // Refine the alignment of the original load (a "new" load created here
-      // which was identical to the first except for the alignment would be
-      // merged with the existing node regardless).
+      // Create the new MMO for the new base load. It is like the original MMO,
+      // but represents an area in memory almost twice the vector size centered
+      // on the original address. If the address is unaligned, we might start
+      // reading up to (sizeof(vector)-1) bytes below the address of the
+      // original unaligned load.
       MachineFunction &MF = DAG.getMachineFunction();
-      MachineMemOperand *MMO =
-        MF.getMachineMemOperand(LD->getPointerInfo(),
-                                LD->getMemOperand()->getFlags(),
-                                LD->getMemoryVT().getStoreSize(),
-                                ABIAlignment);
-      LD->refineAlignment(MMO);
-      SDValue BaseLoad = SDValue(LD, 0);
+      MachineMemOperand *BaseMMO =
+        MF.getMachineMemOperand(LD->getMemOperand(),
+                                -LD->getMemoryVT().getStoreSize()+1,
+                                2*LD->getMemoryVT().getStoreSize()-1);
+
+      // Create the new base load.
+      SDValue LDXIntID = DAG.getTargetConstant(Intrinsic::ppc_altivec_lvx,
+                                               getPointerTy());
+      SDValue BaseLoadOps[] = { Chain, LDXIntID, Ptr };
+      SDValue BaseLoad =
+        DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
+                                DAG.getVTList(MVT::v4i32, MVT::Other),
+                                BaseLoadOps, MVT::v4i32, BaseMMO);
 
       // Note that the value of IncOffset (which is provided to the next
       // load's pointer info offset value, and thus used to calculate the
@@ -8452,21 +8460,18 @@ SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
       SDValue Increment = DAG.getConstant(IncValue, getPointerTy());
       Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
 
+      MachineMemOperand *ExtraMMO =
+        MF.getMachineMemOperand(LD->getMemOperand(),
+                                1, 2*LD->getMemoryVT().getStoreSize()-1);
+      SDValue ExtraLoadOps[] = { Chain, LDXIntID, Ptr };
       SDValue ExtraLoad =
-        DAG.getLoad(VT, dl, Chain, Ptr,
-                    LD->getPointerInfo().getWithOffset(IncOffset),
-                    LD->isVolatile(), LD->isNonTemporal(),
-                    LD->isInvariant(), ABIAlignment);
+        DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
+                                DAG.getVTList(MVT::v4i32, MVT::Other),
+                                ExtraLoadOps, MVT::v4i32, ExtraMMO);
 
       SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
         BaseLoad.getValue(1), ExtraLoad.getValue(1));
 
-      if (BaseLoad.getValueType() != MVT::v4i32)
-        BaseLoad = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, BaseLoad);
-
-      if (ExtraLoad.getValueType() != MVT::v4i32)
-        ExtraLoad = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, ExtraLoad);
-
       // Because vperm has a big-endian bias, we must reverse the order
       // of the input vectors and complement the permute control vector
       // when generating little endian code.  We have already handled the
@@ -8483,36 +8488,9 @@ SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
       if (VT != MVT::v4i32)
         Perm = DAG.getNode(ISD::BITCAST, dl, VT, Perm);
 
-      // Now we need to be really careful about how we update the users of the
-      // original load. We cannot just call DCI.CombineTo (or
-      // DAG.ReplaceAllUsesWith for that matter), because the load still has
-      // uses created here (the permutation for example) that need to stay.
-      SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
-      while (UI != UE) {
-        SDUse &Use = UI.getUse();
-        SDNode *User = *UI;
-        // Note: BaseLoad is checked here because it might not be N, but a
-        // bitcast of N.
-        if (User == Perm.getNode() || User == BaseLoad.getNode() ||
-            User == TF.getNode() || Use.getResNo() > 1) {
-          ++UI;
-          continue;
-        }
-
-        SDValue To = Use.getResNo() ? TF : Perm;
-        ++UI;
-
-        SmallVector<SDValue, 8> Ops;
-        for (const SDUse &O : User->ops()) {
-          if (O == Use)
-            Ops.push_back(To);
-          else
-            Ops.push_back(O);
-        }
-
-        DAG.UpdateNodeOperands(User, Ops);
-      }
-
+      // The output of the permutation is our loaded result, the TokenFactor is
+      // our new chain.
+      DCI.CombineTo(N, Perm, TF);
       return SDValue(N, 0);
     }
     }