R600/SI: Add a MUBUF store pattern for Reg+Imm offsets

[oota-llvm.git] / lib / Target / R600 / AMDGPUISelDAGToDAG.cpp

diff --git a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
index 95037ba6a9e6db76a0e68c03ba8eb17b9b0323f3..a9891350e570f422151b7b90e65ee67dfea8f14f 100644 (file)
--- a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
@@ -195,6 +195,7 @@ bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) {
 SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
   unsigned int Opc = N->getOpcode();
   if (N->isMachineOpcode()) {
+    N->setNodeId(-1);
     return NULL;   // Already selected.
   }
   switch (Opc) {
@@ -308,40 +309,42 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
                                   SDLoc(N), N->getValueType(0), Ops);
   }
+  case AMDGPUISD::REGISTER_LOAD: {
+    const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
+    if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
+      break;
+    SDValue Addr, Offset;
+
+    SelectADDRIndirect(N->getOperand(1), Addr, Offset);
+    const SDValue Ops[] = {
+      Addr,
+      Offset,
+      CurDAG->getTargetConstant(0, MVT::i32),
+      N->getOperand(0),
+    };
+    return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, SDLoc(N),
+                                  CurDAG->getVTList(MVT::i32, MVT::i64, MVT::Other),
+                                  Ops);
   }
-  SDNode *Result = SelectCode(N);
-
-  // Fold operands of selected node
-
-  const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
-  if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
-    const R600InstrInfo *TII =
-        static_cast<const R600InstrInfo*>(TM.getInstrInfo());
-    if (Result && Result->isMachineOpcode() &&
-        !(TII->get(Result->getMachineOpcode()).TSFlags & R600_InstFlag::VECTOR)
-        && TII->hasInstrModifiers(Result->getMachineOpcode())) {
-      // If node has a single use which is CLAMP_R600, folds it
-      if (Result->hasOneUse() && Result->isMachineOpcode()) {
-        SDNode *PotentialClamp = *Result->use_begin();
-        if (PotentialClamp->isMachineOpcode() &&
-            PotentialClamp->getMachineOpcode() == AMDGPU::CLAMP_R600) {
-          unsigned ClampIdx =
-            TII->getOperandIdx(Result->getMachineOpcode(), AMDGPU::OpName::clamp);
-          std::vector<SDValue> Ops;
-          unsigned NumOp = Result->getNumOperands();
-          for (unsigned i = 0; i < NumOp; ++i) {
-            Ops.push_back(Result->getOperand(i));
-          }
-          Ops[ClampIdx - 1] = CurDAG->getTargetConstant(1, MVT::i32);
-          Result = CurDAG->SelectNodeTo(PotentialClamp,
-              Result->getMachineOpcode(), PotentialClamp->getVTList(),
-              Ops.data(), NumOp);
-        }
-      }
-    }
+  case AMDGPUISD::REGISTER_STORE: {
+    const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
+    if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
+      break;
+    SDValue Addr, Offset;
+    SelectADDRIndirect(N->getOperand(2), Addr, Offset);
+    const SDValue Ops[] = {
+      N->getOperand(1),
+      Addr,
+      Offset,
+      CurDAG->getTargetConstant(0, MVT::i32),
+      N->getOperand(0),
+    };
+    return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, SDLoc(N),
+                                        CurDAG->getVTList(MVT::Other),
+                                        Ops);
   }
-
-  return Result;
+  }
+  return SelectCode(N);
 }