Revert "AMDGPU/SI: Use flat for global load/store when targeting HSA"

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

diff --git a/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index b59683f2c938803a951e1365d5b9e302f3e4041a..ea7c6429b7dfa06fc1f504653f4e00fb7535ae19 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -11,6 +11,8 @@
 /// \brief Defines an instruction selector for the AMDGPU target.
 //
 //===----------------------------------------------------------------------===//
+
+#include "AMDGPUDiagnosticInfoUnsupported.h"
 #include "AMDGPUInstrInfo.h"
 #include "AMDGPUISelLowering.h" // For AMDGPUISD
 #include "AMDGPURegisterInfo.h"
@@ -20,9 +22,9 @@
 #include "SIISelLowering.h"
 #include "SIMachineFunctionInfo.h"
 #include "llvm/CodeGen/FunctionLoweringInfo.h"
-#include "llvm/CodeGen/PseudoSourceValue.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
 #include "llvm/CodeGen/SelectionDAG.h"
 #include "llvm/CodeGen/SelectionDAGISel.h"
 #include "llvm/IR/Function.h"
@@ -285,22 +287,7 @@ SDNode *AMDGPUDAGToDAGISel::glueCopyToM0(SDNode *N) const {
   return N;
 }
 
-static unsigned selectVectorRegClassID(unsigned NumVectorElts, bool UseVGPR) {
-  if (UseVGPR) {
-    switch (NumVectorElts) {
-    case 1:
-      return AMDGPU::VGPR_32RegClassID;
-    case 2:
-      return AMDGPU::VReg_64RegClassID;
-    case 4:
-      return AMDGPU::VReg_128RegClassID;
-    case 8:
-      return AMDGPU::VReg_256RegClassID;
-    case 16:
-      return AMDGPU::VReg_512RegClassID;
-    }
-  }
-
+static unsigned selectSGPRVectorRegClassID(unsigned NumVectorElts) {
   switch (NumVectorElts) {
   case 1:
     return AMDGPU::SReg_32RegClassID;
@@ -350,23 +337,7 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     EVT EltVT = VT.getVectorElementType();
     assert(EltVT.bitsEq(MVT::i32));
     if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
-      bool UseVReg = false;
-
-      for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
-                                                    U != E; ++U) {
-        if (!U->isMachineOpcode()) {
-          continue;
-        }
-        const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
-        if (!RC) {
-          continue;
-        }
-        if (static_cast<const SIRegisterInfo *>(TRI)->isSGPRClass(RC)) {
-          UseVReg = false;
-        }
-      }
-
-      RegClassID = selectVectorRegClassID(NumVectorElts, UseVReg);
+      RegClassID = selectSGPRVectorRegClassID(NumVectorElts);
     } else {
       // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
       // that adds a 128 bits reg copy when going through TwoAddressInstructions
@@ -489,41 +460,6 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     N = glueCopyToM0(N);
     break;
   }
-  case AMDGPUISD::REGISTER_LOAD: {
-    if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
-      break;
-    SDValue Addr, Offset;
-
-    SDLoc DL(N);
-    SelectADDRIndirect(N->getOperand(1), Addr, Offset);
-    const SDValue Ops[] = {
-      Addr,
-      Offset,
-      CurDAG->getTargetConstant(0, DL, MVT::i32),
-      N->getOperand(0),
-    };
-    return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, DL,
-                                  CurDAG->getVTList(MVT::i32, MVT::i64,
-                                                    MVT::Other),
-                                  Ops);
-  }
-  case AMDGPUISD::REGISTER_STORE: {
-    if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
-      break;
-    SDValue Addr, Offset;
-    SelectADDRIndirect(N->getOperand(2), Addr, Offset);
-    SDLoc DL(N);
-    const SDValue Ops[] = {
-      N->getOperand(1),
-      Addr,
-      Offset,
-      CurDAG->getTargetConstant(0, DL, MVT::i32),
-      N->getOperand(0),
-    };
-    return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, DL,
-                                        CurDAG->getVTList(MVT::Other),
-                                        Ops);
-  }
 
   case AMDGPUISD::BFE_I32:
   case AMDGPUISD::BFE_U32: {
@@ -1093,36 +1029,10 @@ bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc,
 
   SDLoc DL(Addr);
   MachineFunction &MF = CurDAG->getMachineFunction();
-  const SIRegisterInfo *TRI =
-      static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
-  MachineRegisterInfo &MRI = MF.getRegInfo();
-  const SITargetLowering& Lowering =
-    *static_cast<const SITargetLowering*>(getTargetLowering());
+  const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
 
-  unsigned ScratchOffsetReg =
-      TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
-  Lowering.CreateLiveInRegister(*CurDAG, &AMDGPU::SReg_32RegClass,
-                                ScratchOffsetReg, MVT::i32);
-  SDValue Sym0 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD0", MVT::i32);
-  SDValue ScratchRsrcDword0 =
-      SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym0), 0);
-
-  SDValue Sym1 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD1", MVT::i32);
-  SDValue ScratchRsrcDword1 =
-      SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym1), 0);
-
-  const SDValue RsrcOps[] = {
-      CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
-      ScratchRsrcDword0,
-      CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
-      ScratchRsrcDword1,
-      CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32),
-  };
-  SDValue ScratchPtr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
-                                              MVT::v2i32, RsrcOps), 0);
-  Rsrc = SDValue(Lowering.buildScratchRSRC(*CurDAG, DL, ScratchPtr), 0);
-  SOffset = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
-      MRI.getLiveInVirtReg(ScratchOffsetReg), MVT::i32);
+  Rsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
+  SOffset = CurDAG->getRegister(Info->getScratchWaveOffsetReg(), MVT::i32);
 
   // (add n0, c1)
   if (CurDAG->isBaseWithConstantOffset(Addr)) {
@@ -1300,13 +1210,14 @@ SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
   AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
   SDLoc DL(N);
 
+  const MachineFunction &MF = CurDAG->getMachineFunction();
+  DiagnosticInfoUnsupported NotImplemented(*MF.getFunction(),
+                                           "addrspacecast not implemented");
+  CurDAG->getContext()->diagnose(NotImplemented);
+
   assert(Subtarget->hasFlatAddressSpace() &&
          "addrspacecast only supported with flat address space!");
 
-  assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
-          ASC->getDestAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) &&
-         "Cannot cast address space to / from constant address!");
-
   assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
           ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
          "Can only cast to / from flat address space!");