radeon/llvm: Lower loads from USE_SGPR adddress space during DAG lowering

5 files changed, 50 insertions, 66 deletions

diff --git a/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp b/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp
index 2e67828..e22df8e 100644
--- a/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp
+++ b/src/gallium/drivers/radeon/AMDGPUISelLowering.cpp
@@ -311,13 +311,6 @@ bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const
   return false;
 }
 
-void AMDGPUTargetLowering::addLiveIn(MachineInstr * MI,
-    MachineFunction * MF, MachineRegisterInfo & MRI,
-    const TargetInstrInfo * TII, unsigned reg) const
-{
-  AMDGPU::utilAddLiveIn(MF, MRI, TII, reg, MI->getOperand(0).getReg());
-}
-
 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
                                                   const TargetRegisterClass *RC,
                                                    unsigned Reg, EVT VT) const {
diff --git a/src/gallium/drivers/radeon/AMDGPUISelLowering.h b/src/gallium/drivers/radeon/AMDGPUISelLowering.h
index b05de38..cd2f0c0 100644
--- a/src/gallium/drivers/radeon/AMDGPUISelLowering.h
+++ b/src/gallium/drivers/radeon/AMDGPUISelLowering.h
@@ -29,16 +29,6 @@ private:
 
 protected:
 
-  /// addLiveIn - This functions adds reg to the live in list of the entry block
-  /// and emits a copy from reg to MI.getOperand(0).
-  ///
-  //  Some registers are loaded with values before the program
-  /// begins to execute.  The loading of these values is modeled with pseudo
-  /// instructions which are lowered using this function. 
-  void addLiveIn(MachineInstr * MI, MachineFunction * MF,
-                 MachineRegisterInfo & MRI, const TargetInstrInfo * TII,
-		 unsigned reg) const;
-
   /// CreateLiveInRegister - Helper function that adds Reg to the LiveIn list
   /// of the DAG's MachineFunction.  This returns a Register SDNode representing
   /// Reg. 
diff --git a/src/gallium/drivers/radeon/SIISelLowering.cpp b/src/gallium/drivers/radeon/SIISelLowering.cpp
index 270e4a1..092c2fa 100644
--- a/src/gallium/drivers/radeon/SIISelLowering.cpp
+++ b/src/gallium/drivers/radeon/SIISelLowering.cpp
@@ -13,6 +13,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "SIISelLowering.h"
+#include "AMDIL.h"
 #include "AMDILIntrinsicInfo.h"
 #include "SIInstrInfo.h"
 #include "SIRegisterInfo.h"
@@ -47,6 +48,11 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
 
   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
 
+  // We need to custom lower loads from the USER_SGPR address space, so we can
+  // add the SGPRs as livein registers.
+  setOperationAction(ISD::LOAD, MVT::i32, Custom);
+  setOperationAction(ISD::LOAD, MVT::i64, Custom);
+
   setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
   setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
 
@@ -126,11 +132,6 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
   case AMDGPU::SI_V_CNDLT:
     LowerSI_V_CNDLT(MI, *BB, I, MRI);
     break;
-  case AMDGPU::USE_SGPR_32:
-  case AMDGPU::USE_SGPR_64:
-    lowerUSE_SGPR(MI, BB->getParent(), MRI);
-    MI->eraseFromParent();
-    break;
   }
   return BB;
 }
@@ -209,21 +210,6 @@ void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
   MI->eraseFromParent();
 }
 
-void SITargetLowering::lowerUSE_SGPR(MachineInstr *MI,
-    MachineFunction * MF, MachineRegisterInfo & MRI) const
-{
-  const TargetInstrInfo * TII = getTargetMachine().getInstrInfo();
-  unsigned dstReg = MI->getOperand(0).getReg();
-  int64_t newIndex = MI->getOperand(1).getImm();
-  const TargetRegisterClass * dstClass = MRI.getRegClass(dstReg);
-  unsigned DwordWidth = dstClass->getSize() / 4;
-  assert(newIndex % DwordWidth == 0 && "USER_SGPR not properly aligned");
-  newIndex = newIndex / DwordWidth;
-
-  unsigned newReg = dstClass->getRegister(newIndex);
-  addLiveIn(MI, MF, MRI, TII, newReg); 
-}
-
 EVT SITargetLowering::getSetCCResultType(EVT VT) const
 {
   return MVT::i1;
@@ -238,6 +224,7 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const
   switch (Op.getOpcode()) {
   default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
   case ISD::BR_CC: return LowerBR_CC(Op, DAG);
+  case ISD::LOAD: return LowerLOAD(Op, DAG);
   case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
   case ISD::AND: return Loweri1ContextSwitch(Op, DAG, ISD::AND);
   case ISD::INTRINSIC_WO_CHAIN: {
@@ -303,6 +290,48 @@ SDValue SITargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const
   return Result;
 }
 
+SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
+{
+  EVT VT = Op.getValueType();
+  LoadSDNode *Ptr = dyn_cast<LoadSDNode>(Op);
+
+  assert(Ptr);
+
+  unsigned AddrSpace = Ptr->getPointerInfo().getAddrSpace();
+
+  // We only need to lower USER_SGPR address space loads
+  if (AddrSpace != AMDGPUAS::USER_SGPR_ADDRESS) {
+    return SDValue();
+  }
+
+  // Loads from the USER_SGPR address space can only have constant value
+  // pointers.
+  ConstantSDNode *BasePtr = dyn_cast<ConstantSDNode>(Ptr->getBasePtr());
+  assert(BasePtr);
+
+  unsigned TypeDwordWidth = VT.getSizeInBits() / 32;
+  const TargetRegisterClass * dstClass;
+  switch (TypeDwordWidth) {
+    default:
+      assert(!"USER_SGPR value size not implemented");
+      return SDValue();
+    case 1:
+      dstClass = &AMDGPU::SReg_32RegClass;
+      break;
+    case 2:
+      dstClass = &AMDGPU::SReg_64RegClass;
+      break;
+  }
+  uint64_t Index = BasePtr->getZExtValue();
+  assert(Index % TypeDwordWidth == 0 && "USER_SGPR not properly aligned");
+  unsigned SGPRIndex = Index / TypeDwordWidth;
+  unsigned Reg = dstClass->getRegister(SGPRIndex);
+
+  DAG.ReplaceAllUsesOfValueWith(Op, CreateLiveInRegister(DAG, dstClass, Reg,
+                                                         VT));
+  return SDValue();
+}
+
 SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
 {
   SDValue LHS = Op.getOperand(0);
diff --git a/src/gallium/drivers/radeon/SIISelLowering.h b/src/gallium/drivers/radeon/SIISelLowering.h
index d545452..cf655a1 100644
--- a/src/gallium/drivers/radeon/SIISelLowering.h
+++ b/src/gallium/drivers/radeon/SIISelLowering.h
@@ -35,12 +35,11 @@ class SITargetLowering : public AMDGPUTargetLowering
               MachineBasicBlock::iterator I) const;
   void LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
               MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
-  void lowerUSE_SGPR(MachineInstr *MI, MachineFunction * MF,
-                     MachineRegisterInfo & MRI) const;
 
   SDValue Loweri1ContextSwitch(SDValue Op, SelectionDAG &DAG,
                                            unsigned VCCNode) const;
   SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
 
 public:
diff --git a/src/gallium/drivers/radeon/SIInstructions.td b/src/gallium/drivers/radeon/SIInstructions.td
index aad2ade..152d735 100644
--- a/src/gallium/drivers/radeon/SIInstructions.td
+++ b/src/gallium/drivers/radeon/SIInstructions.td
@@ -7,19 +7,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-def load_user_sgpr : PatFrag<(ops node:$ptr),
-  (load node:$ptr),
-  [{
-    const Value *Src = cast<LoadSDNode>(N)->getSrcValue();
-    if (Src) {
-      PointerType * PT = dyn_cast<PointerType>(Src->getType());
-      return PT && PT->getAddressSpace() == AMDGPUAS::USER_SGPR_ADDRESS;
-    }
-    return false;
-  }]
->;
-
-
 def isSI : Predicate<"Subtarget.device()"
                             "->getGeneration() == AMDGPUDeviceInfo::HD7XXX">;
 
@@ -956,20 +943,6 @@ def SI_INTERP_CONST : InstSI <
                                                  imm:$attr, SReg_32:$params))]
 >;
 
-def USE_SGPR_32 : InstSI <
-  (outs SReg_32:$dst),
-  (ins i32imm:$src0),
-  "USE_SGPR_32",
-  [(set (i32 SReg_32:$dst), (load_user_sgpr imm:$src0))]
->;
-
-def USE_SGPR_64 : InstSI <
-  (outs SReg_64:$dst),
-  (ins i32imm:$src0),
-  "USE_SGPR_64",
-  [(set (i64 SReg_64:$dst), (load_user_sgpr imm:$src0))]
->;
-
 } // end usesCustomInserter 
 
 // SI Psuedo branch instructions.  These are used by the CFG structurizer pass