R600/SI: Fix offset folding in some cases with shifted pointers.

Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
is only done if the add has one use. If the resulting constant
add can be folded into an addressing mode, force this to happen
for the pointer operand.

This ends up happening a lot because of how LDS objects are allocated.
Since the globals are allocated next to each other, acessing the first
element of the second object is directly indexed by a shifted pointer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215739 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/SIISelLowering.cpp

@@ -235,6 +235,26 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
 
   setTargetDAGCombine(ISD::UINT_TO_FP);
 
+  // All memory operations. Some folding on the pointer operand is done to help
+  // matching the constant offsets in the addressing modes.
+  setTargetDAGCombine(ISD::LOAD);
+  setTargetDAGCombine(ISD::STORE);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD);
+  setTargetDAGCombine(ISD::ATOMIC_STORE);
+  setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
+  setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
+  setTargetDAGCombine(ISD::ATOMIC_SWAP);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
+  setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
+
   setSchedulingPreference(Sched::RegPressure);
 }
 
@@ -1296,6 +1316,56 @@ SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
   return SDValue();
 }
 
+// (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2)
+
+// This is a variant of
+// (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2),
+//
+// The normal DAG combiner will do this, but only if the add has one use since
+// that would increase the number of instructions.
+//
+// This prevents us from seeing a constant offset that can be folded into a
+// memory instruction's addressing mode. If we know the resulting add offset of
+// a pointer can be folded into an addressing offset, we can replace the pointer
+// operand with the add of new constant offset. This eliminates one of the uses,
+// and may allow the remaining use to also be simplified.
+//
+SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
+                                               unsigned AddrSpace,
+                                               DAGCombinerInfo &DCI) const {
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+
+  if (N0.getOpcode() != ISD::ADD)
+    return SDValue();
+
+  const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
+  if (!CN1)
+    return SDValue();
+
+  const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+  if (!CAdd)
+    return SDValue();
+
+  const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
+      getTargetMachine().getSubtargetImpl()->getInstrInfo());
+
+  // If the resulting offset is too large, we can't fold it into the addressing
+  // mode offset.
+  APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue();
+  if (!TII->canFoldOffset(Offset.getZExtValue(), AddrSpace))
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc SL(N);
+  EVT VT = N->getValueType(0);
+
+  SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
+  SDValue COffset = DAG.getConstant(Offset, MVT::i32);
+
+  return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset);
+}
+
 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
                                             DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -1348,8 +1418,45 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::UINT_TO_FP: {
     return performUCharToFloatCombine(N, DCI);
   }
-  }
+  case ISD::LOAD:
+  case ISD::STORE:
+  case ISD::ATOMIC_LOAD:
+  case ISD::ATOMIC_STORE:
+  case ISD::ATOMIC_CMP_SWAP:
+  case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
+  case ISD::ATOMIC_SWAP:
+  case ISD::ATOMIC_LOAD_ADD:
+  case ISD::ATOMIC_LOAD_SUB:
+  case ISD::ATOMIC_LOAD_AND:
+  case ISD::ATOMIC_LOAD_OR:
+  case ISD::ATOMIC_LOAD_XOR:
+  case ISD::ATOMIC_LOAD_NAND:
+  case ISD::ATOMIC_LOAD_MIN:
+  case ISD::ATOMIC_LOAD_MAX:
+  case ISD::ATOMIC_LOAD_UMIN:
+  case ISD::ATOMIC_LOAD_UMAX: { // TODO: Target mem intrinsics.
+    if (DCI.isBeforeLegalize())
+      break;
 
+    MemSDNode *MemNode = cast<MemSDNode>(N);
+    SDValue Ptr = MemNode->getBasePtr();
+
+    // TODO: We could also do this for multiplies.
+    unsigned AS = MemNode->getAddressSpace();
+    if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) {
+      SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI);
+      if (NewPtr) {
+        SmallVector<SDValue, 8> NewOps;
+        for (unsigned I = 0, N = MemNode->getNumOperands(); I != N; ++I)
+          NewOps.push_back(MemNode->getOperand(I));
+
+        NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr;
+        return SDValue(DAG.UpdateNodeOperands(MemNode, NewOps), 0);
+      }
+    }
+    break;
+  }
+  }
   return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
 }