XCoreISelLowering.cpp revision 1d8f975890bba8a328b1e940c9b87675d8d4949c
1//===-- XCoreISelLowering.cpp - XCore DAG Lowering Implementation   ------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements the XCoreTargetLowering class.
11//
12//===----------------------------------------------------------------------===//
13
14#define DEBUG_TYPE "xcore-lower"
15
16#include "XCoreISelLowering.h"
17#include "XCoreMachineFunctionInfo.h"
18#include "XCore.h"
19#include "XCoreTargetMachine.h"
20#include "XCoreSubtarget.h"
21#include "llvm/DerivedTypes.h"
22#include "llvm/Function.h"
23#include "llvm/Intrinsics.h"
24#include "llvm/CallingConv.h"
25#include "llvm/GlobalVariable.h"
26#include "llvm/GlobalAlias.h"
27#include "llvm/CodeGen/CallingConvLower.h"
28#include "llvm/CodeGen/MachineFrameInfo.h"
29#include "llvm/CodeGen/MachineFunction.h"
30#include "llvm/CodeGen/MachineInstrBuilder.h"
31#include "llvm/CodeGen/MachineRegisterInfo.h"
32#include "llvm/CodeGen/SelectionDAGISel.h"
33#include "llvm/CodeGen/ValueTypes.h"
34#include "llvm/Support/Debug.h"
35#include "llvm/ADT/VectorExtras.h"
36#include <queue>
37#include <set>
38using namespace llvm;
39
40const char *XCoreTargetLowering::
41getTargetNodeName(unsigned Opcode) const
42{
43  switch (Opcode)
44  {
45    case XCoreISD::BL                : return "XCoreISD::BL";
46    case XCoreISD::PCRelativeWrapper : return "XCoreISD::PCRelativeWrapper";
47    case XCoreISD::DPRelativeWrapper : return "XCoreISD::DPRelativeWrapper";
48    case XCoreISD::CPRelativeWrapper : return "XCoreISD::CPRelativeWrapper";
49    case XCoreISD::STWSP             : return "XCoreISD::STWSP";
50    case XCoreISD::RETSP             : return "XCoreISD::RETSP";
51    default                           : return NULL;
52  }
53}
54
55XCoreTargetLowering::XCoreTargetLowering(XCoreTargetMachine &XTM)
56  : TargetLowering(XTM),
57    TM(XTM),
58    Subtarget(*XTM.getSubtargetImpl()) {
59
60  // Set up the register classes.
61  addRegisterClass(MVT::i32, XCore::GRRegsRegisterClass);
62
63  // Compute derived properties from the register classes
64  computeRegisterProperties();
65
66  // Division is expensive
67  setIntDivIsCheap(false);
68
69  setShiftAmountType(MVT::i32);
70  // shl X, 32 == 0
71  setShiftAmountFlavor(Extend);
72  setStackPointerRegisterToSaveRestore(XCore::SP);
73
74  setSchedulingPreference(SchedulingForRegPressure);
75
76  // Use i32 for setcc operations results (slt, sgt, ...).
77  setSetCCResultContents(ZeroOrOneSetCCResult);
78
79  // XCore does not have the NodeTypes below.
80  setOperationAction(ISD::BR_CC,     MVT::Other, Expand);
81  setOperationAction(ISD::SELECT_CC, MVT::i32,   Custom);
82  setOperationAction(ISD::ADDC, MVT::i32, Expand);
83  setOperationAction(ISD::ADDE, MVT::i32, Expand);
84  setOperationAction(ISD::SUBC, MVT::i32, Expand);
85  setOperationAction(ISD::SUBE, MVT::i32, Expand);
86
87  // Stop the combiner recombining select and set_cc
88  setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
89
90  // 64bit
91  setOperationAction(ISD::ADD, MVT::i64, Custom);
92  setOperationAction(ISD::SUB, MVT::i64, Custom);
93
94  if (Subtarget.isXS1A()) {
95    setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
96  }
97  setOperationAction(ISD::MULHS, MVT::i32, Expand);
98  setOperationAction(ISD::MULHU, MVT::i32, Expand);
99  setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
100  setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
101  setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
102
103  // Bit Manipulation
104  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
105  setOperationAction(ISD::ROTL , MVT::i32, Expand);
106  setOperationAction(ISD::ROTR , MVT::i32, Expand);
107
108  // Expand jump tables for now
109  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
110  setOperationAction(ISD::JumpTable, MVT::i32, Custom);
111
112  // RET must be custom lowered, to meet ABI requirements
113  setOperationAction(ISD::RET,           MVT::Other, Custom);
114
115  setOperationAction(ISD::GlobalAddress, MVT::i32,   Custom);
116
117  // Thread Local Storage
118  setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
119
120  // Conversion of i64 -> double produces constantpool nodes
121  setOperationAction(ISD::ConstantPool, MVT::i32,   Custom);
122
123  // Loads
124  setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
125  setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
126  setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
127
128  setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
129  setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Expand);
130
131  // Varargs
132  setOperationAction(ISD::VAEND, MVT::Other, Expand);
133  setOperationAction(ISD::VACOPY, MVT::Other, Expand);
134  setOperationAction(ISD::VAARG, MVT::Other, Custom);
135  setOperationAction(ISD::VASTART, MVT::Other, Custom);
136
137  // Dynamic stack
138  setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
139  setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
140  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
141
142  // Debug
143  setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
144  setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
145}
146
147SDValue XCoreTargetLowering::
148LowerOperation(SDValue Op, SelectionDAG &DAG) {
149  switch (Op.getOpcode())
150  {
151  case ISD::CALL:             return LowerCALL(Op, DAG);
152  case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
153  case ISD::RET:              return LowerRET(Op, DAG);
154  case ISD::GlobalAddress:    return LowerGlobalAddress(Op, DAG);
155  case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
156  case ISD::ConstantPool:     return LowerConstantPool(Op, DAG);
157  case ISD::JumpTable:        return LowerJumpTable(Op, DAG);
158  case ISD::SELECT_CC:        return LowerSELECT_CC(Op, DAG);
159  case ISD::VAARG:            return LowerVAARG(Op, DAG);
160  case ISD::VASTART:          return LowerVASTART(Op, DAG);
161  // FIXME: Remove these when LegalizeDAGTypes lands.
162  case ISD::ADD:
163  case ISD::SUB:              return SDValue(ExpandADDSUB(Op.getNode(), DAG),0);
164
165  case ISD::FRAMEADDR:        return LowerFRAMEADDR(Op, DAG);
166  default:
167    assert(0 && "unimplemented operand");
168    return SDValue();
169  }
170}
171
172SDNode *XCoreTargetLowering::
173ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
174  switch (N->getOpcode()) {
175  case ISD::SUB:
176  case ISD::ADD:
177    return ExpandADDSUB(N, DAG);
178  default:
179    assert(0 && "Wasn't expecting to be able to lower this!");
180    return NULL;
181  }
182}
183
184//===----------------------------------------------------------------------===//
185//  Misc Lower Operation implementation
186//===----------------------------------------------------------------------===//
187
188SDValue XCoreTargetLowering::
189LowerSELECT_CC(SDValue Op, SelectionDAG &DAG)
190{
191  SDValue Cond = DAG.getNode(ISD::SETCC, MVT::i32, Op.getOperand(2),
192                             Op.getOperand(3), Op.getOperand(4));
193  return DAG.getNode(ISD::SELECT, MVT::i32, Cond, Op.getOperand(0),
194                     Op.getOperand(1));
195}
196
197SDValue XCoreTargetLowering::
198getGlobalAddressWrapper(SDValue GA, GlobalValue *GV, SelectionDAG &DAG)
199{
200  if (isa<Function>(GV)) {
201    return DAG.getNode(XCoreISD::PCRelativeWrapper, MVT::i32, GA);
202  } else if (!Subtarget.isXS1A()) {
203    const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
204    if (!GVar) {
205      // If GV is an alias then use the aliasee to determine constness
206      if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
207        GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal());
208    }
209    bool isConst = GVar && GVar->isConstant();
210    if (isConst) {
211      return DAG.getNode(XCoreISD::CPRelativeWrapper, MVT::i32, GA);
212    }
213  }
214  return DAG.getNode(XCoreISD::DPRelativeWrapper, MVT::i32, GA);
215}
216
217SDValue XCoreTargetLowering::
218LowerGlobalAddress(SDValue Op, SelectionDAG &DAG)
219{
220  GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
221  SDValue GA = DAG.getTargetGlobalAddress(GV, MVT::i32);
222  // If it's a debug information descriptor, don't mess with it.
223  if (DAG.isVerifiedDebugInfoDesc(Op))
224    return GA;
225  return getGlobalAddressWrapper(GA, GV, DAG);
226}
227
228static inline SDValue BuildGetId(SelectionDAG &DAG) {
229  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, MVT::i32,
230                     DAG.getConstant(Intrinsic::xcore_getid, MVT::i32));
231}
232
233static inline bool isZeroLengthArray(const Type *Ty) {
234  const ArrayType *AT = dyn_cast_or_null<ArrayType>(Ty);
235  return AT && (AT->getNumElements() == 0);
236}
237
238SDValue XCoreTargetLowering::
239LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG)
240{
241  // transform to label + getid() * size
242  GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
243  SDValue GA = DAG.getTargetGlobalAddress(GV, MVT::i32);
244  const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
245  if (!GVar) {
246    // If GV is an alias then use the aliasee to determine size
247    if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
248      GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal());
249  }
250  if (! GVar) {
251    assert(0 && "Thread local object not a GlobalVariable?");
252    return SDValue();
253  }
254  const Type *Ty = cast<PointerType>(GV->getType())->getElementType();
255  if (!Ty->isSized() || isZeroLengthArray(Ty)) {
256    cerr << "Size of thread local object " << GVar->getName()
257         << " is unknown\n";
258    abort();
259  }
260  SDValue base = getGlobalAddressWrapper(GA, GV, DAG);
261  const TargetData *TD = TM.getTargetData();
262  unsigned Size = TD->getABITypeSize(Ty);
263  SDValue offset = DAG.getNode(ISD::MUL, MVT::i32, BuildGetId(DAG),
264                       DAG.getConstant(Size, MVT::i32));
265  return DAG.getNode(ISD::ADD, MVT::i32, base, offset);
266}
267
268SDValue XCoreTargetLowering::
269LowerConstantPool(SDValue Op, SelectionDAG &DAG)
270{
271  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
272  if (Subtarget.isXS1A()) {
273    assert(0 && "Lowering of constant pool unimplemented");
274    return SDValue();
275  } else {
276    MVT PtrVT = Op.getValueType();
277    SDValue Res;
278    if (CP->isMachineConstantPoolEntry()) {
279      Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
280                                      CP->getAlignment());
281    } else {
282      Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
283                                      CP->getAlignment());
284    }
285    return DAG.getNode(XCoreISD::CPRelativeWrapper, MVT::i32, Res);
286  }
287}
288
289SDValue XCoreTargetLowering::
290LowerJumpTable(SDValue Op, SelectionDAG &DAG)
291{
292  MVT PtrVT = Op.getValueType();
293  JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
294  SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
295  return DAG.getNode(XCoreISD::DPRelativeWrapper, MVT::i32, JTI);
296}
297
298SDNode *XCoreTargetLowering::
299ExpandADDSUB(SDNode *N, SelectionDAG &DAG)
300{
301  assert(N->getValueType(0) == MVT::i64 &&
302         (N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) &&
303        "Unknown operand to lower!");
304
305  // Extract components
306  SDValue LHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(0),
307                             DAG.getConstant(0, MVT::i32));
308  SDValue LHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(0),
309                             DAG.getConstant(1, MVT::i32));
310  SDValue RHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(1),
311                             DAG.getConstant(0, MVT::i32));
312  SDValue RHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, N->getOperand(1),
313                             DAG.getConstant(1, MVT::i32));
314
315  // Expand
316  if (Subtarget.isXS1A()) {
317    SDValue Lo = DAG.getNode(N->getOpcode(), MVT::i32, LHSL, RHSL);
318
319    ISD::CondCode CarryCC = (N->getOpcode() == ISD::ADD) ? ISD::SETULT :
320                                                           ISD::SETUGT;
321    SDValue Carry = DAG.getSetCC(MVT::i32, Lo, LHSL, CarryCC);
322
323    SDValue Hi = DAG.getNode(N->getOpcode(), MVT::i32, LHSH, Carry);
324    Hi = DAG.getNode(N->getOpcode(), MVT::i32, Hi, RHSH);
325    // Merge the pieces
326    return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi).getNode();
327  }
328  unsigned Opcode = (N->getOpcode() == ISD::ADD) ? XCoreISD::LADD :
329                                                   XCoreISD::LSUB;
330  SDValue Zero = DAG.getConstant(0, MVT::i32);
331  SDValue Carry = DAG.getNode(Opcode, DAG.getVTList(MVT::i32, MVT::i32),
332                                  LHSL, RHSL, Zero);
333  SDValue Lo(Carry.getNode(), 1);
334
335  SDValue Ignored = DAG.getNode(Opcode, DAG.getVTList(MVT::i32, MVT::i32),
336                                  LHSH, RHSH, Carry);
337  SDValue Hi(Ignored.getNode(), 1);
338  // Merge the pieces
339  return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi).getNode();
340}
341
342SDValue XCoreTargetLowering::
343LowerVAARG(SDValue Op, SelectionDAG &DAG)
344{
345  assert(0 && "unimplemented");
346  // FIX Arguments passed by reference need a extra dereference.
347  SDNode *Node = Op.getNode();
348  const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
349  MVT VT = Node->getValueType(0);
350  SDValue VAList = DAG.getLoad(getPointerTy(), Node->getOperand(0),
351                               Node->getOperand(1), V, 0);
352  // Increment the pointer, VAList, to the next vararg
353  SDValue Tmp3 = DAG.getNode(ISD::ADD, getPointerTy(), VAList,
354                     DAG.getConstant(VT.getSizeInBits(),
355                                     getPointerTy()));
356  // Store the incremented VAList to the legalized pointer
357  Tmp3 = DAG.getStore(VAList.getValue(1), Tmp3, Node->getOperand(1), V, 0);
358  // Load the actual argument out of the pointer VAList
359  return DAG.getLoad(VT, Tmp3, VAList, NULL, 0);
360}
361
362SDValue XCoreTargetLowering::
363LowerVASTART(SDValue Op, SelectionDAG &DAG)
364{
365  // vastart stores the address of the VarArgsFrameIndex slot into the
366  // memory location argument
367  MachineFunction &MF = DAG.getMachineFunction();
368  XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
369  SDValue Addr = DAG.getFrameIndex(XFI->getVarArgsFrameIndex(), MVT::i32);
370  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
371  return DAG.getStore(Op.getOperand(0), Addr, Op.getOperand(1), SV, 0);
372}
373
374SDValue XCoreTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
375  // Depths > 0 not supported yet!
376  if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
377    return SDValue();
378
379  MachineFunction &MF = DAG.getMachineFunction();
380  const TargetRegisterInfo *RegInfo = getTargetMachine().getRegisterInfo();
381  return DAG.getCopyFromReg(DAG.getEntryNode(), RegInfo->getFrameRegister(MF),
382    MVT::i32);
383}
384
385//===----------------------------------------------------------------------===//
386//                      Calling Convention Implementation
387//
388//  The lower operations present on calling convention works on this order:
389//      LowerCALL (virt regs --> phys regs, virt regs --> stack)
390//      LowerFORMAL_ARGUMENTS (phys --> virt regs, stack --> virt regs)
391//      LowerRET (virt regs --> phys regs)
392//      LowerCALL (phys regs --> virt regs)
393//
394//===----------------------------------------------------------------------===//
395
396#include "XCoreGenCallingConv.inc"
397
398//===----------------------------------------------------------------------===//
399//                  CALL Calling Convention Implementation
400//===----------------------------------------------------------------------===//
401
402/// XCore custom CALL implementation
403SDValue XCoreTargetLowering::
404LowerCALL(SDValue Op, SelectionDAG &DAG)
405{
406  CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
407  unsigned CallingConv = TheCall->getCallingConv();
408  // For now, only CallingConv::C implemented
409  switch (CallingConv)
410  {
411    default:
412      assert(0 && "Unsupported calling convention");
413    case CallingConv::Fast:
414    case CallingConv::C:
415      return LowerCCCCallTo(Op, DAG, CallingConv);
416  }
417}
418
419/// LowerCCCCallTo - functions arguments are copied from virtual
420/// regs to (physical regs)/(stack frame), CALLSEQ_START and
421/// CALLSEQ_END are emitted.
422/// TODO: isTailCall, sret.
423SDValue XCoreTargetLowering::
424LowerCCCCallTo(SDValue Op, SelectionDAG &DAG, unsigned CC)
425{
426  CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
427  SDValue Chain  = TheCall->getChain();
428  SDValue Callee = TheCall->getCallee();
429  bool isVarArg  = TheCall->isVarArg();
430
431  // Analyze operands of the call, assigning locations to each operand.
432  SmallVector<CCValAssign, 16> ArgLocs;
433  CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
434
435  // The ABI dictates there should be one stack slot available to the callee
436  // on function entry (for saving lr).
437  CCInfo.AllocateStack(4, 4);
438
439  CCInfo.AnalyzeCallOperands(TheCall, CC_XCore);
440
441  // Get a count of how many bytes are to be pushed on the stack.
442  unsigned NumBytes = CCInfo.getNextStackOffset();
443
444  Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes,
445                                 getPointerTy(), true));
446
447  SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass;
448  SmallVector<SDValue, 12> MemOpChains;
449
450  // Walk the register/memloc assignments, inserting copies/loads.
451  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
452    CCValAssign &VA = ArgLocs[i];
453
454    // Arguments start after the 5 first operands of ISD::CALL
455    SDValue Arg = TheCall->getArg(i);
456
457    // Promote the value if needed.
458    switch (VA.getLocInfo()) {
459      default: assert(0 && "Unknown loc info!");
460      case CCValAssign::Full: break;
461      case CCValAssign::SExt:
462        Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
463        break;
464      case CCValAssign::ZExt:
465        Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
466        break;
467      case CCValAssign::AExt:
468        Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
469        break;
470    }
471
472    // Arguments that can be passed on register must be kept at
473    // RegsToPass vector
474    if (VA.isRegLoc()) {
475      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
476    } else {
477      assert(VA.isMemLoc());
478
479      int Offset = VA.getLocMemOffset();
480
481      MemOpChains.push_back(DAG.getNode(XCoreISD::STWSP, MVT::Other, Chain, Arg,
482                                        DAG.getConstant(Offset/4, MVT::i32)));
483    }
484  }
485
486  // Transform all store nodes into one single node because
487  // all store nodes are independent of each other.
488  if (!MemOpChains.empty())
489    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
490                        &MemOpChains[0], MemOpChains.size());
491
492  // Build a sequence of copy-to-reg nodes chained together with token
493  // chain and flag operands which copy the outgoing args into registers.
494  // The InFlag in necessary since all emited instructions must be
495  // stuck together.
496  SDValue InFlag;
497  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
498    Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first,
499                             RegsToPass[i].second, InFlag);
500    InFlag = Chain.getValue(1);
501  }
502
503  // If the callee is a GlobalAddress node (quite common, every direct call is)
504  // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
505  // Likewise ExternalSymbol -> TargetExternalSymbol.
506  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
507    Callee = DAG.getTargetGlobalAddress(G->getGlobal(), MVT::i32);
508  else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee))
509    Callee = DAG.getTargetExternalSymbol(E->getSymbol(), MVT::i32);
510
511  // XCoreBranchLink = #chain, #target_address, #opt_in_flags...
512  //             = Chain, Callee, Reg#1, Reg#2, ...
513  //
514  // Returns a chain & a flag for retval copy to use.
515  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
516  SmallVector<SDValue, 8> Ops;
517  Ops.push_back(Chain);
518  Ops.push_back(Callee);
519
520  // Add argument registers to the end of the list so that they are
521  // known live into the call.
522  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
523    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
524                                  RegsToPass[i].second.getValueType()));
525
526  if (InFlag.getNode())
527    Ops.push_back(InFlag);
528
529  Chain  = DAG.getNode(XCoreISD::BL, NodeTys, &Ops[0], Ops.size());
530  InFlag = Chain.getValue(1);
531
532  // Create the CALLSEQ_END node.
533  Chain = DAG.getCALLSEQ_END(Chain,
534                             DAG.getConstant(NumBytes, getPointerTy(), true),
535                             DAG.getConstant(0, getPointerTy(), true),
536                             InFlag);
537  InFlag = Chain.getValue(1);
538
539  // Handle result values, copying them out of physregs into vregs that we
540  // return.
541  return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
542                 Op.getResNo());
543}
544
545/// LowerCallResult - Lower the result values of an ISD::CALL into the
546/// appropriate copies out of appropriate physical registers.  This assumes that
547/// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
548/// being lowered. Returns a SDNode with the same number of values as the
549/// ISD::CALL.
550SDNode *XCoreTargetLowering::
551LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
552        unsigned CallingConv, SelectionDAG &DAG) {
553  bool isVarArg = TheCall->isVarArg();
554
555  // Assign locations to each value returned by this call.
556  SmallVector<CCValAssign, 16> RVLocs;
557  CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs);
558
559  CCInfo.AnalyzeCallResult(TheCall, RetCC_XCore);
560  SmallVector<SDValue, 8> ResultVals;
561
562  // Copy all of the result registers out of their specified physreg.
563  for (unsigned i = 0; i != RVLocs.size(); ++i) {
564    Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
565                                 RVLocs[i].getValVT(), InFlag).getValue(1);
566    InFlag = Chain.getValue(2);
567    ResultVals.push_back(Chain.getValue(0));
568  }
569
570  ResultVals.push_back(Chain);
571
572  // Merge everything together with a MERGE_VALUES node.
573  return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
574                     &ResultVals[0], ResultVals.size()).getNode();
575}
576
577//===----------------------------------------------------------------------===//
578//             FORMAL_ARGUMENTS Calling Convention Implementation
579//===----------------------------------------------------------------------===//
580
581/// XCore custom FORMAL_ARGUMENTS implementation
582SDValue XCoreTargetLowering::
583LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG)
584{
585  unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
586  switch(CC)
587  {
588    default:
589      assert(0 && "Unsupported calling convention");
590    case CallingConv::C:
591    case CallingConv::Fast:
592      return LowerCCCArguments(Op, DAG);
593  }
594}
595
596/// LowerCCCArguments - transform physical registers into
597/// virtual registers and generate load operations for
598/// arguments places on the stack.
599/// TODO: sret
600SDValue XCoreTargetLowering::
601LowerCCCArguments(SDValue Op, SelectionDAG &DAG)
602{
603  MachineFunction &MF = DAG.getMachineFunction();
604  MachineFrameInfo *MFI = MF.getFrameInfo();
605  MachineRegisterInfo &RegInfo = MF.getRegInfo();
606  SDValue Root = Op.getOperand(0);
607  bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
608  unsigned CC = MF.getFunction()->getCallingConv();
609
610  // Assign locations to all of the incoming arguments.
611  SmallVector<CCValAssign, 16> ArgLocs;
612  CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
613
614  CCInfo.AnalyzeFormalArguments(Op.getNode(), CC_XCore);
615
616  unsigned StackSlotSize = XCoreFrameInfo::stackSlotSize();
617
618  SmallVector<SDValue, 16> ArgValues;
619
620  unsigned LRSaveSize = StackSlotSize;
621
622  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
623
624    CCValAssign &VA = ArgLocs[i];
625
626    if (VA.isRegLoc()) {
627      // Arguments passed in registers
628      MVT RegVT = VA.getLocVT();
629      switch (RegVT.getSimpleVT()) {
630      default:
631        cerr << "LowerFORMAL_ARGUMENTS Unhandled argument type: "
632             << RegVT.getSimpleVT()
633             << "\n";
634        abort();
635      case MVT::i32:
636        unsigned VReg = RegInfo.createVirtualRegister(
637                          XCore::GRRegsRegisterClass);
638        RegInfo.addLiveIn(VA.getLocReg(), VReg);
639        ArgValues.push_back(DAG.getCopyFromReg(Root, VReg, RegVT));
640      }
641    } else {
642      // sanity check
643      assert(VA.isMemLoc());
644      // Load the argument to a virtual register
645      unsigned ObjSize = VA.getLocVT().getSizeInBits()/8;
646      if (ObjSize > StackSlotSize) {
647        cerr << "LowerFORMAL_ARGUMENTS Unhandled argument type: "
648             << VA.getLocVT().getSimpleVT()
649             << "\n";
650      }
651      // Create the frame index object for this incoming parameter...
652      int FI = MFI->CreateFixedObject(ObjSize,
653                                      LRSaveSize + VA.getLocMemOffset());
654
655      // Create the SelectionDAG nodes corresponding to a load
656      //from this parameter
657      SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
658      ArgValues.push_back(DAG.getLoad(VA.getLocVT(), Root, FIN, NULL, 0));
659    }
660  }
661
662  if (isVarArg) {
663    /* Argument registers */
664    static const unsigned ArgRegs[] = {
665      XCore::R0, XCore::R1, XCore::R2, XCore::R3
666    };
667    XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
668    unsigned FirstVAReg = CCInfo.getFirstUnallocated(ArgRegs,
669                                                     array_lengthof(ArgRegs));
670    if (FirstVAReg < array_lengthof(ArgRegs)) {
671      SmallVector<SDValue, 4> MemOps;
672      int offset = 0;
673      // Save remaining registers, storing higher register numbers at a higher
674      // address
675      for (unsigned i = array_lengthof(ArgRegs) - 1; i >= FirstVAReg; --i) {
676        // Create a stack slot
677        int FI = MFI->CreateFixedObject(4, offset);
678        if (i == FirstVAReg) {
679          XFI->setVarArgsFrameIndex(FI);
680        }
681        offset -= StackSlotSize;
682        SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
683        // Move argument from phys reg -> virt reg
684        unsigned VReg = RegInfo.createVirtualRegister(
685                          XCore::GRRegsRegisterClass);
686        RegInfo.addLiveIn(ArgRegs[i], VReg);
687        SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::i32);
688        // Move argument from virt reg -> stack
689        SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
690        MemOps.push_back(Store);
691      }
692      if (!MemOps.empty())
693        Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
694                           &MemOps[0], MemOps.size());
695    } else {
696      // This will point to the next argument passed via stack.
697      XFI->setVarArgsFrameIndex(
698          MFI->CreateFixedObject(4, LRSaveSize + CCInfo.getNextStackOffset()));
699    }
700  }
701
702  ArgValues.push_back(Root);
703
704  // Return the new list of results.
705  std::vector<MVT> RetVT(Op.getNode()->value_begin(),
706                                    Op.getNode()->value_end());
707  return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
708}
709
710//===----------------------------------------------------------------------===//
711//               Return Value Calling Convention Implementation
712//===----------------------------------------------------------------------===//
713
714SDValue XCoreTargetLowering::
715LowerRET(SDValue Op, SelectionDAG &DAG)
716{
717  // CCValAssign - represent the assignment of
718  // the return value to a location
719  SmallVector<CCValAssign, 16> RVLocs;
720  unsigned CC   = DAG.getMachineFunction().getFunction()->getCallingConv();
721  bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
722
723  // CCState - Info about the registers and stack slot.
724  CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs);
725
726  // Analize return values of ISD::RET
727  CCInfo.AnalyzeReturn(Op.getNode(), RetCC_XCore);
728
729  // If this is the first return lowered for this function, add
730  // the regs to the liveout set for the function.
731  if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
732    for (unsigned i = 0; i != RVLocs.size(); ++i)
733      if (RVLocs[i].isRegLoc())
734        DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
735  }
736
737  // The chain is always operand #0
738  SDValue Chain = Op.getOperand(0);
739  SDValue Flag;
740
741  // Copy the result values into the output registers.
742  for (unsigned i = 0; i != RVLocs.size(); ++i) {
743    CCValAssign &VA = RVLocs[i];
744    assert(VA.isRegLoc() && "Can only return in registers!");
745
746    // ISD::RET => ret chain, (regnum1,val1), ...
747    // So i*2+1 index only the regnums
748    Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
749
750    // guarantee that all emitted copies are
751    // stuck together, avoiding something bad
752    Flag = Chain.getValue(1);
753  }
754
755  // Return on XCore is always a "retsp 0"
756  if (Flag.getNode())
757    return DAG.getNode(XCoreISD::RETSP, MVT::Other,
758                       Chain, DAG.getConstant(0, MVT::i32), Flag);
759  else // Return Void
760    return DAG.getNode(XCoreISD::RETSP, MVT::Other,
761                       Chain, DAG.getConstant(0, MVT::i32));
762}
763
764//===----------------------------------------------------------------------===//
765//  Other Lowering Code
766//===----------------------------------------------------------------------===//
767
768MachineBasicBlock *
769XCoreTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
770                                               MachineBasicBlock *BB) {
771  const TargetInstrInfo &TII = *getTargetMachine().getInstrInfo();
772  assert((MI->getOpcode() == XCore::SELECT_CC) &&
773         "Unexpected instr type to insert");
774
775  // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
776  // control-flow pattern.  The incoming instruction knows the destination vreg
777  // to set, the condition code register to branch on, the true/false values to
778  // select between, and a branch opcode to use.
779  const BasicBlock *LLVM_BB = BB->getBasicBlock();
780  MachineFunction::iterator It = BB;
781  ++It;
782
783  //  thisMBB:
784  //  ...
785  //   TrueVal = ...
786  //   cmpTY ccX, r1, r2
787  //   bCC copy1MBB
788  //   fallthrough --> copy0MBB
789  MachineBasicBlock *thisMBB = BB;
790  MachineFunction *F = BB->getParent();
791  MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
792  MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
793  BuildMI(BB, TII.get(XCore::BRFT_lru6))
794    .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
795  F->insert(It, copy0MBB);
796  F->insert(It, sinkMBB);
797  // Update machine-CFG edges by transferring all successors of the current
798  // block to the new block which will contain the Phi node for the select.
799  sinkMBB->transferSuccessors(BB);
800  // Next, add the true and fallthrough blocks as its successors.
801  BB->addSuccessor(copy0MBB);
802  BB->addSuccessor(sinkMBB);
803
804  //  copy0MBB:
805  //   %FalseValue = ...
806  //   # fallthrough to sinkMBB
807  BB = copy0MBB;
808
809  // Update machine-CFG edges
810  BB->addSuccessor(sinkMBB);
811
812  //  sinkMBB:
813  //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
814  //  ...
815  BB = sinkMBB;
816  BuildMI(BB, TII.get(XCore::PHI), MI->getOperand(0).getReg())
817    .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
818    .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
819
820  F->DeleteMachineInstr(MI);   // The pseudo instruction is gone now.
821  return BB;
822}
823
824//===----------------------------------------------------------------------===//
825//  Addressing mode description hooks
826//===----------------------------------------------------------------------===//
827
828static inline bool isImmUs(int64_t val)
829{
830  return (val >= 0 && val <= 11);
831}
832
833static inline bool isImmUs2(int64_t val)
834{
835  return (val%2 == 0 && isImmUs(val/2));
836}
837
838static inline bool isImmUs4(int64_t val)
839{
840  return (val%4 == 0 && isImmUs(val/4));
841}
842
843/// isLegalAddressingMode - Return true if the addressing mode represented
844/// by AM is legal for this target, for a load/store of the specified type.
845bool
846XCoreTargetLowering::isLegalAddressingMode(const AddrMode &AM,
847                                              const Type *Ty) const {
848  MVT VT = getValueType(Ty, true);
849  // Get expected value type after legalization
850  switch (VT.getSimpleVT()) {
851  // Legal load / stores
852  case MVT::i8:
853  case MVT::i16:
854  case MVT::i32:
855    break;
856  // Expand i1 -> i8
857  case MVT::i1:
858    VT = MVT::i8;
859    break;
860  // Everything else is lowered to words
861  default:
862    VT = MVT::i32;
863    break;
864  }
865  if (AM.BaseGV) {
866    return VT == MVT::i32 && !AM.HasBaseReg && AM.Scale == 0 &&
867                 AM.BaseOffs%4 == 0;
868  }
869
870  switch (VT.getSimpleVT()) {
871  default:
872    return false;
873  case MVT::i8:
874    // reg + imm
875    if (AM.Scale == 0) {
876      return isImmUs(AM.BaseOffs);
877    }
878    return AM.Scale == 1 && AM.BaseOffs == 0;
879  case MVT::i16:
880    // reg + imm
881    if (AM.Scale == 0) {
882      return isImmUs2(AM.BaseOffs);
883    }
884    return AM.Scale == 2 && AM.BaseOffs == 0;
885  case MVT::i32:
886    // reg + imm
887    if (AM.Scale == 0) {
888      return isImmUs4(AM.BaseOffs);
889    }
890    // reg + reg<<2
891    return AM.Scale == 4 && AM.BaseOffs == 0;
892  }
893
894  return false;
895}
896
897//===----------------------------------------------------------------------===//
898//                           XCore Inline Assembly Support
899//===----------------------------------------------------------------------===//
900
901std::vector<unsigned> XCoreTargetLowering::
902getRegClassForInlineAsmConstraint(const std::string &Constraint,
903                                  MVT VT) const
904{
905  if (Constraint.size() != 1)
906    return std::vector<unsigned>();
907
908  switch (Constraint[0]) {
909    default : break;
910    case 'r':
911      return make_vector<unsigned>(XCore::R0, XCore::R1,  XCore::R2,
912                                   XCore::R3, XCore::R4,  XCore::R5,
913                                   XCore::R6, XCore::R7,  XCore::R8,
914                                   XCore::R9, XCore::R10, XCore::R11, 0);
915      break;
916  }
917  return std::vector<unsigned>();
918}
919