1//===-- MipsISelLowering.cpp - Mips 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 defines the interfaces that Mips uses to lower LLVM code into a
11// selection DAG.
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "mips-lower"
16#include "MipsISelLowering.h"
17#include "MipsMachineFunction.h"
18#include "MipsTargetMachine.h"
19#include "MipsTargetObjectFile.h"
20#include "MipsSubtarget.h"
21#include "llvm/DerivedTypes.h"
22#include "llvm/Function.h"
23#include "llvm/GlobalVariable.h"
24#include "llvm/Intrinsics.h"
25#include "llvm/CallingConv.h"
26#include "InstPrinter/MipsInstPrinter.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/Support/ErrorHandling.h"
36using namespace llvm;
37
38// If I is a shifted mask, set the size (Size) and the first bit of the
39// mask (Pos), and return true.
40// For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
41static bool IsShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
42  if (!isUInt<32>(I) || !isShiftedMask_32(I))
43     return false;
44
45  Size = CountPopulation_32(I);
46  Pos = CountTrailingZeros_32(I);
47  return true;
48}
49
50const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
51  switch (Opcode) {
52  case MipsISD::JmpLink:           return "MipsISD::JmpLink";
53  case MipsISD::Hi:                return "MipsISD::Hi";
54  case MipsISD::Lo:                return "MipsISD::Lo";
55  case MipsISD::GPRel:             return "MipsISD::GPRel";
56  case MipsISD::TlsGd:             return "MipsISD::TlsGd";
57  case MipsISD::TprelHi:           return "MipsISD::TprelHi";
58  case MipsISD::TprelLo:           return "MipsISD::TprelLo";
59  case MipsISD::ThreadPointer:     return "MipsISD::ThreadPointer";
60  case MipsISD::Ret:               return "MipsISD::Ret";
61  case MipsISD::FPBrcond:          return "MipsISD::FPBrcond";
62  case MipsISD::FPCmp:             return "MipsISD::FPCmp";
63  case MipsISD::CMovFP_T:          return "MipsISD::CMovFP_T";
64  case MipsISD::CMovFP_F:          return "MipsISD::CMovFP_F";
65  case MipsISD::FPRound:           return "MipsISD::FPRound";
66  case MipsISD::MAdd:              return "MipsISD::MAdd";
67  case MipsISD::MAddu:             return "MipsISD::MAddu";
68  case MipsISD::MSub:              return "MipsISD::MSub";
69  case MipsISD::MSubu:             return "MipsISD::MSubu";
70  case MipsISD::DivRem:            return "MipsISD::DivRem";
71  case MipsISD::DivRemU:           return "MipsISD::DivRemU";
72  case MipsISD::BuildPairF64:      return "MipsISD::BuildPairF64";
73  case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
74  case MipsISD::WrapperPIC:        return "MipsISD::WrapperPIC";
75  case MipsISD::DynAlloc:          return "MipsISD::DynAlloc";
76  case MipsISD::Sync:              return "MipsISD::Sync";
77  case MipsISD::Ext:               return "MipsISD::Ext";
78  case MipsISD::Ins:               return "MipsISD::Ins";
79  default:                         return NULL;
80  }
81}
82
83MipsTargetLowering::
84MipsTargetLowering(MipsTargetMachine &TM)
85  : TargetLowering(TM, new MipsTargetObjectFile()),
86    Subtarget(&TM.getSubtarget<MipsSubtarget>()),
87    HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()) {
88
89  // Mips does not have i1 type, so use i32 for
90  // setcc operations results (slt, sgt, ...).
91  setBooleanContents(ZeroOrOneBooleanContent);
92  setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
93
94  // Set up the register classes
95  addRegisterClass(MVT::i32, Mips::CPURegsRegisterClass);
96  addRegisterClass(MVT::f32, Mips::FGR32RegisterClass);
97
98  if (HasMips64)
99    addRegisterClass(MVT::i64, Mips::CPU64RegsRegisterClass);
100
101  // When dealing with single precision only, use libcalls
102  if (!Subtarget->isSingleFloat()) {
103    if (HasMips64)
104      addRegisterClass(MVT::f64, Mips::FGR64RegisterClass);
105    else
106      addRegisterClass(MVT::f64, Mips::AFGR64RegisterClass);
107  }
108
109  // Load extented operations for i1 types must be promoted
110  setLoadExtAction(ISD::EXTLOAD,  MVT::i1,  Promote);
111  setLoadExtAction(ISD::ZEXTLOAD, MVT::i1,  Promote);
112  setLoadExtAction(ISD::SEXTLOAD, MVT::i1,  Promote);
113
114  // MIPS doesn't have extending float->double load/store
115  setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
116  setTruncStoreAction(MVT::f64, MVT::f32, Expand);
117
118  // Used by legalize types to correctly generate the setcc result.
119  // Without this, every float setcc comes with a AND/OR with the result,
120  // we don't want this, since the fpcmp result goes to a flag register,
121  // which is used implicitly by brcond and select operations.
122  AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
123
124  // Mips Custom Operations
125  setOperationAction(ISD::GlobalAddress,      MVT::i32,   Custom);
126  setOperationAction(ISD::GlobalAddress,      MVT::i64,   Custom);
127  setOperationAction(ISD::BlockAddress,       MVT::i32,   Custom);
128  setOperationAction(ISD::GlobalTLSAddress,   MVT::i32,   Custom);
129  setOperationAction(ISD::JumpTable,          MVT::i32,   Custom);
130  setOperationAction(ISD::ConstantPool,       MVT::i32,   Custom);
131  setOperationAction(ISD::SELECT,             MVT::f32,   Custom);
132  setOperationAction(ISD::SELECT,             MVT::f64,   Custom);
133  setOperationAction(ISD::SELECT,             MVT::i32,   Custom);
134  setOperationAction(ISD::BRCOND,             MVT::Other, Custom);
135  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32,   Custom);
136  setOperationAction(ISD::VASTART,            MVT::Other, Custom);
137
138  setOperationAction(ISD::SDIV, MVT::i32, Expand);
139  setOperationAction(ISD::SREM, MVT::i32, Expand);
140  setOperationAction(ISD::UDIV, MVT::i32, Expand);
141  setOperationAction(ISD::UREM, MVT::i32, Expand);
142  setOperationAction(ISD::SDIV, MVT::i64, Expand);
143  setOperationAction(ISD::SREM, MVT::i64, Expand);
144  setOperationAction(ISD::UDIV, MVT::i64, Expand);
145  setOperationAction(ISD::UREM, MVT::i64, Expand);
146
147  // Operations not directly supported by Mips.
148  setOperationAction(ISD::BR_JT,             MVT::Other, Expand);
149  setOperationAction(ISD::BR_CC,             MVT::Other, Expand);
150  setOperationAction(ISD::SELECT_CC,         MVT::Other, Expand);
151  setOperationAction(ISD::UINT_TO_FP,        MVT::i32,   Expand);
152  setOperationAction(ISD::FP_TO_UINT,        MVT::i32,   Expand);
153  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1,    Expand);
154  setOperationAction(ISD::CTPOP,             MVT::i32,   Expand);
155  setOperationAction(ISD::CTTZ,              MVT::i32,   Expand);
156  setOperationAction(ISD::ROTL,              MVT::i32,   Expand);
157  setOperationAction(ISD::ROTL,              MVT::i64,   Expand);
158
159  if (!Subtarget->hasMips32r2())
160    setOperationAction(ISD::ROTR, MVT::i32,   Expand);
161
162  if (!Subtarget->hasMips64r2())
163    setOperationAction(ISD::ROTR, MVT::i64,   Expand);
164
165  setOperationAction(ISD::SHL_PARTS,         MVT::i32,   Expand);
166  setOperationAction(ISD::SRA_PARTS,         MVT::i32,   Expand);
167  setOperationAction(ISD::SRL_PARTS,         MVT::i32,   Expand);
168  setOperationAction(ISD::FCOPYSIGN,         MVT::f32,   Custom);
169  setOperationAction(ISD::FCOPYSIGN,         MVT::f64,   Custom);
170  setOperationAction(ISD::FSIN,              MVT::f32,   Expand);
171  setOperationAction(ISD::FSIN,              MVT::f64,   Expand);
172  setOperationAction(ISD::FCOS,              MVT::f32,   Expand);
173  setOperationAction(ISD::FCOS,              MVT::f64,   Expand);
174  setOperationAction(ISD::FPOWI,             MVT::f32,   Expand);
175  setOperationAction(ISD::FPOW,              MVT::f32,   Expand);
176  setOperationAction(ISD::FPOW,              MVT::f64,   Expand);
177  setOperationAction(ISD::FLOG,              MVT::f32,   Expand);
178  setOperationAction(ISD::FLOG2,             MVT::f32,   Expand);
179  setOperationAction(ISD::FLOG10,            MVT::f32,   Expand);
180  setOperationAction(ISD::FEXP,              MVT::f32,   Expand);
181  setOperationAction(ISD::FMA,               MVT::f32,   Expand);
182  setOperationAction(ISD::FMA,               MVT::f64,   Expand);
183
184  setOperationAction(ISD::EXCEPTIONADDR,     MVT::i32, Expand);
185  setOperationAction(ISD::EHSELECTION,       MVT::i32, Expand);
186
187  setOperationAction(ISD::VAARG,             MVT::Other, Expand);
188  setOperationAction(ISD::VACOPY,            MVT::Other, Expand);
189  setOperationAction(ISD::VAEND,             MVT::Other, Expand);
190
191  // Use the default for now
192  setOperationAction(ISD::STACKSAVE,         MVT::Other, Expand);
193  setOperationAction(ISD::STACKRESTORE,      MVT::Other, Expand);
194
195  setOperationAction(ISD::MEMBARRIER,        MVT::Other, Custom);
196  setOperationAction(ISD::ATOMIC_FENCE,      MVT::Other, Custom);
197
198  setOperationAction(ISD::ATOMIC_LOAD,       MVT::i32,    Expand);
199  setOperationAction(ISD::ATOMIC_STORE,      MVT::i32,    Expand);
200
201  setInsertFencesForAtomic(true);
202
203  if (Subtarget->isSingleFloat())
204    setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
205
206  if (!Subtarget->hasSEInReg()) {
207    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
208    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
209  }
210
211  if (!Subtarget->hasBitCount())
212    setOperationAction(ISD::CTLZ, MVT::i32, Expand);
213
214  if (!Subtarget->hasSwap())
215    setOperationAction(ISD::BSWAP, MVT::i32, Expand);
216
217  setTargetDAGCombine(ISD::ADDE);
218  setTargetDAGCombine(ISD::SUBE);
219  setTargetDAGCombine(ISD::SDIVREM);
220  setTargetDAGCombine(ISD::UDIVREM);
221  setTargetDAGCombine(ISD::SETCC);
222  setTargetDAGCombine(ISD::AND);
223  setTargetDAGCombine(ISD::OR);
224
225  setMinFunctionAlignment(2);
226
227  setStackPointerRegisterToSaveRestore(Mips::SP);
228  computeRegisterProperties();
229
230  setExceptionPointerRegister(Mips::A0);
231  setExceptionSelectorRegister(Mips::A1);
232}
233
234bool MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
235  MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
236  return SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16;
237}
238
239EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
240  return MVT::i32;
241}
242
243// SelectMadd -
244// Transforms a subgraph in CurDAG if the following pattern is found:
245//  (addc multLo, Lo0), (adde multHi, Hi0),
246// where,
247//  multHi/Lo: product of multiplication
248//  Lo0: initial value of Lo register
249//  Hi0: initial value of Hi register
250// Return true if pattern matching was successful.
251static bool SelectMadd(SDNode* ADDENode, SelectionDAG* CurDAG) {
252  // ADDENode's second operand must be a flag output of an ADDC node in order
253  // for the matching to be successful.
254  SDNode* ADDCNode = ADDENode->getOperand(2).getNode();
255
256  if (ADDCNode->getOpcode() != ISD::ADDC)
257    return false;
258
259  SDValue MultHi = ADDENode->getOperand(0);
260  SDValue MultLo = ADDCNode->getOperand(0);
261  SDNode* MultNode = MultHi.getNode();
262  unsigned MultOpc = MultHi.getOpcode();
263
264  // MultHi and MultLo must be generated by the same node,
265  if (MultLo.getNode() != MultNode)
266    return false;
267
268  // and it must be a multiplication.
269  if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
270    return false;
271
272  // MultLo amd MultHi must be the first and second output of MultNode
273  // respectively.
274  if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
275    return false;
276
277  // Transform this to a MADD only if ADDENode and ADDCNode are the only users
278  // of the values of MultNode, in which case MultNode will be removed in later
279  // phases.
280  // If there exist users other than ADDENode or ADDCNode, this function returns
281  // here, which will result in MultNode being mapped to a single MULT
282  // instruction node rather than a pair of MULT and MADD instructions being
283  // produced.
284  if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
285    return false;
286
287  SDValue Chain = CurDAG->getEntryNode();
288  DebugLoc dl = ADDENode->getDebugLoc();
289
290  // create MipsMAdd(u) node
291  MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
292
293  SDValue MAdd = CurDAG->getNode(MultOpc, dl,
294                                 MVT::Glue,
295                                 MultNode->getOperand(0),// Factor 0
296                                 MultNode->getOperand(1),// Factor 1
297                                 ADDCNode->getOperand(1),// Lo0
298                                 ADDENode->getOperand(1));// Hi0
299
300  // create CopyFromReg nodes
301  SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
302                                              MAdd);
303  SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
304                                              Mips::HI, MVT::i32,
305                                              CopyFromLo.getValue(2));
306
307  // replace uses of adde and addc here
308  if (!SDValue(ADDCNode, 0).use_empty())
309    CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
310
311  if (!SDValue(ADDENode, 0).use_empty())
312    CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
313
314  return true;
315}
316
317// SelectMsub -
318// Transforms a subgraph in CurDAG if the following pattern is found:
319//  (addc Lo0, multLo), (sube Hi0, multHi),
320// where,
321//  multHi/Lo: product of multiplication
322//  Lo0: initial value of Lo register
323//  Hi0: initial value of Hi register
324// Return true if pattern matching was successful.
325static bool SelectMsub(SDNode* SUBENode, SelectionDAG* CurDAG) {
326  // SUBENode's second operand must be a flag output of an SUBC node in order
327  // for the matching to be successful.
328  SDNode* SUBCNode = SUBENode->getOperand(2).getNode();
329
330  if (SUBCNode->getOpcode() != ISD::SUBC)
331    return false;
332
333  SDValue MultHi = SUBENode->getOperand(1);
334  SDValue MultLo = SUBCNode->getOperand(1);
335  SDNode* MultNode = MultHi.getNode();
336  unsigned MultOpc = MultHi.getOpcode();
337
338  // MultHi and MultLo must be generated by the same node,
339  if (MultLo.getNode() != MultNode)
340    return false;
341
342  // and it must be a multiplication.
343  if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
344    return false;
345
346  // MultLo amd MultHi must be the first and second output of MultNode
347  // respectively.
348  if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
349    return false;
350
351  // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
352  // of the values of MultNode, in which case MultNode will be removed in later
353  // phases.
354  // If there exist users other than SUBENode or SUBCNode, this function returns
355  // here, which will result in MultNode being mapped to a single MULT
356  // instruction node rather than a pair of MULT and MSUB instructions being
357  // produced.
358  if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
359    return false;
360
361  SDValue Chain = CurDAG->getEntryNode();
362  DebugLoc dl = SUBENode->getDebugLoc();
363
364  // create MipsSub(u) node
365  MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
366
367  SDValue MSub = CurDAG->getNode(MultOpc, dl,
368                                 MVT::Glue,
369                                 MultNode->getOperand(0),// Factor 0
370                                 MultNode->getOperand(1),// Factor 1
371                                 SUBCNode->getOperand(0),// Lo0
372                                 SUBENode->getOperand(0));// Hi0
373
374  // create CopyFromReg nodes
375  SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
376                                              MSub);
377  SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
378                                              Mips::HI, MVT::i32,
379                                              CopyFromLo.getValue(2));
380
381  // replace uses of sube and subc here
382  if (!SDValue(SUBCNode, 0).use_empty())
383    CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
384
385  if (!SDValue(SUBENode, 0).use_empty())
386    CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
387
388  return true;
389}
390
391static SDValue PerformADDECombine(SDNode *N, SelectionDAG& DAG,
392                                  TargetLowering::DAGCombinerInfo &DCI,
393                                  const MipsSubtarget* Subtarget) {
394  if (DCI.isBeforeLegalize())
395    return SDValue();
396
397  if (Subtarget->hasMips32() && SelectMadd(N, &DAG))
398    return SDValue(N, 0);
399
400  return SDValue();
401}
402
403static SDValue PerformSUBECombine(SDNode *N, SelectionDAG& DAG,
404                                  TargetLowering::DAGCombinerInfo &DCI,
405                                  const MipsSubtarget* Subtarget) {
406  if (DCI.isBeforeLegalize())
407    return SDValue();
408
409  if (Subtarget->hasMips32() && SelectMsub(N, &DAG))
410    return SDValue(N, 0);
411
412  return SDValue();
413}
414
415static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG& DAG,
416                                    TargetLowering::DAGCombinerInfo &DCI,
417                                    const MipsSubtarget* Subtarget) {
418  if (DCI.isBeforeLegalizeOps())
419    return SDValue();
420
421  EVT Ty = N->getValueType(0);
422  unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
423  unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
424  unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
425                                                  MipsISD::DivRemU;
426  DebugLoc dl = N->getDebugLoc();
427
428  SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
429                               N->getOperand(0), N->getOperand(1));
430  SDValue InChain = DAG.getEntryNode();
431  SDValue InGlue = DivRem;
432
433  // insert MFLO
434  if (N->hasAnyUseOfValue(0)) {
435    SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty,
436                                            InGlue);
437    DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
438    InChain = CopyFromLo.getValue(1);
439    InGlue = CopyFromLo.getValue(2);
440  }
441
442  // insert MFHI
443  if (N->hasAnyUseOfValue(1)) {
444    SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
445                                            HI, Ty, InGlue);
446    DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
447  }
448
449  return SDValue();
450}
451
452static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
453  switch (CC) {
454  default: llvm_unreachable("Unknown fp condition code!");
455  case ISD::SETEQ:
456  case ISD::SETOEQ: return Mips::FCOND_OEQ;
457  case ISD::SETUNE: return Mips::FCOND_UNE;
458  case ISD::SETLT:
459  case ISD::SETOLT: return Mips::FCOND_OLT;
460  case ISD::SETGT:
461  case ISD::SETOGT: return Mips::FCOND_OGT;
462  case ISD::SETLE:
463  case ISD::SETOLE: return Mips::FCOND_OLE;
464  case ISD::SETGE:
465  case ISD::SETOGE: return Mips::FCOND_OGE;
466  case ISD::SETULT: return Mips::FCOND_ULT;
467  case ISD::SETULE: return Mips::FCOND_ULE;
468  case ISD::SETUGT: return Mips::FCOND_UGT;
469  case ISD::SETUGE: return Mips::FCOND_UGE;
470  case ISD::SETUO:  return Mips::FCOND_UN;
471  case ISD::SETO:   return Mips::FCOND_OR;
472  case ISD::SETNE:
473  case ISD::SETONE: return Mips::FCOND_ONE;
474  case ISD::SETUEQ: return Mips::FCOND_UEQ;
475  }
476}
477
478
479// Returns true if condition code has to be inverted.
480static bool InvertFPCondCode(Mips::CondCode CC) {
481  if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
482    return false;
483
484  if (CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT)
485    return true;
486
487  assert(false && "Illegal Condition Code");
488  return false;
489}
490
491// Creates and returns an FPCmp node from a setcc node.
492// Returns Op if setcc is not a floating point comparison.
493static SDValue CreateFPCmp(SelectionDAG& DAG, const SDValue& Op) {
494  // must be a SETCC node
495  if (Op.getOpcode() != ISD::SETCC)
496    return Op;
497
498  SDValue LHS = Op.getOperand(0);
499
500  if (!LHS.getValueType().isFloatingPoint())
501    return Op;
502
503  SDValue RHS = Op.getOperand(1);
504  DebugLoc dl = Op.getDebugLoc();
505
506  // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
507  // node if necessary.
508  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
509
510  return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
511                     DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
512}
513
514// Creates and returns a CMovFPT/F node.
515static SDValue CreateCMovFP(SelectionDAG& DAG, SDValue Cond, SDValue True,
516                            SDValue False, DebugLoc DL) {
517  bool invert = InvertFPCondCode((Mips::CondCode)
518                                 cast<ConstantSDNode>(Cond.getOperand(2))
519                                 ->getSExtValue());
520
521  return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
522                     True.getValueType(), True, False, Cond);
523}
524
525static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG& DAG,
526                                   TargetLowering::DAGCombinerInfo &DCI,
527                                   const MipsSubtarget* Subtarget) {
528  if (DCI.isBeforeLegalizeOps())
529    return SDValue();
530
531  SDValue Cond = CreateFPCmp(DAG, SDValue(N, 0));
532
533  if (Cond.getOpcode() != MipsISD::FPCmp)
534    return SDValue();
535
536  SDValue True  = DAG.getConstant(1, MVT::i32);
537  SDValue False = DAG.getConstant(0, MVT::i32);
538
539  return CreateCMovFP(DAG, Cond, True, False, N->getDebugLoc());
540}
541
542static SDValue PerformANDCombine(SDNode *N, SelectionDAG& DAG,
543                                 TargetLowering::DAGCombinerInfo &DCI,
544                                 const MipsSubtarget* Subtarget) {
545  // Pattern match EXT.
546  //  $dst = and ((sra or srl) $src , pos), (2**size - 1)
547  //  => ext $dst, $src, size, pos
548  if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
549    return SDValue();
550
551  SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
552
553  // Op's first operand must be a shift right.
554  if (ShiftRight.getOpcode() != ISD::SRA && ShiftRight.getOpcode() != ISD::SRL)
555    return SDValue();
556
557  // The second operand of the shift must be an immediate.
558  uint64_t Pos;
559  ConstantSDNode *CN;
560  if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
561    return SDValue();
562
563  Pos = CN->getZExtValue();
564
565  uint64_t SMPos, SMSize;
566  // Op's second operand must be a shifted mask.
567  if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
568      !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
569    return SDValue();
570
571  // Return if the shifted mask does not start at bit 0 or the sum of its size
572  // and Pos exceeds the word's size.
573  if (SMPos != 0 || Pos + SMSize > 32)
574    return SDValue();
575
576  return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), MVT::i32,
577                     ShiftRight.getOperand(0),
578                     DAG.getConstant(Pos, MVT::i32),
579                     DAG.getConstant(SMSize, MVT::i32));
580}
581
582static SDValue PerformORCombine(SDNode *N, SelectionDAG& DAG,
583                                TargetLowering::DAGCombinerInfo &DCI,
584                                const MipsSubtarget* Subtarget) {
585  // Pattern match INS.
586  //  $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
587  //  where mask1 = (2**size - 1) << pos, mask0 = ~mask1
588  //  => ins $dst, $src, size, pos, $src1
589  if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
590    return SDValue();
591
592  SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
593  uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
594  ConstantSDNode *CN;
595
596  // See if Op's first operand matches (and $src1 , mask0).
597  if (And0.getOpcode() != ISD::AND)
598    return SDValue();
599
600  if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
601      !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
602    return SDValue();
603
604  // See if Op's second operand matches (and (shl $src, pos), mask1).
605  if (And1.getOpcode() != ISD::AND)
606    return SDValue();
607
608  if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
609      !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
610    return SDValue();
611
612  // The shift masks must have the same position and size.
613  if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
614    return SDValue();
615
616  SDValue Shl = And1.getOperand(0);
617  if (Shl.getOpcode() != ISD::SHL)
618    return SDValue();
619
620  if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
621    return SDValue();
622
623  unsigned Shamt = CN->getZExtValue();
624
625  // Return if the shift amount and the first bit position of mask are not the
626  // same.
627  if (Shamt != SMPos0)
628    return SDValue();
629
630  return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), MVT::i32,
631                     Shl.getOperand(0),
632                     DAG.getConstant(SMPos0, MVT::i32),
633                     DAG.getConstant(SMSize0, MVT::i32),
634                     And0.getOperand(0));
635}
636
637SDValue  MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
638  const {
639  SelectionDAG &DAG = DCI.DAG;
640  unsigned opc = N->getOpcode();
641
642  switch (opc) {
643  default: break;
644  case ISD::ADDE:
645    return PerformADDECombine(N, DAG, DCI, Subtarget);
646  case ISD::SUBE:
647    return PerformSUBECombine(N, DAG, DCI, Subtarget);
648  case ISD::SDIVREM:
649  case ISD::UDIVREM:
650    return PerformDivRemCombine(N, DAG, DCI, Subtarget);
651  case ISD::SETCC:
652    return PerformSETCCCombine(N, DAG, DCI, Subtarget);
653  case ISD::AND:
654    return PerformANDCombine(N, DAG, DCI, Subtarget);
655  case ISD::OR:
656    return PerformORCombine(N, DAG, DCI, Subtarget);
657  }
658
659  return SDValue();
660}
661
662SDValue MipsTargetLowering::
663LowerOperation(SDValue Op, SelectionDAG &DAG) const
664{
665  switch (Op.getOpcode())
666  {
667    case ISD::BRCOND:             return LowerBRCOND(Op, DAG);
668    case ISD::ConstantPool:       return LowerConstantPool(Op, DAG);
669    case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
670    case ISD::GlobalAddress:      return LowerGlobalAddress(Op, DAG);
671    case ISD::BlockAddress:       return LowerBlockAddress(Op, DAG);
672    case ISD::GlobalTLSAddress:   return LowerGlobalTLSAddress(Op, DAG);
673    case ISD::JumpTable:          return LowerJumpTable(Op, DAG);
674    case ISD::SELECT:             return LowerSELECT(Op, DAG);
675    case ISD::VASTART:            return LowerVASTART(Op, DAG);
676    case ISD::FCOPYSIGN:          return LowerFCOPYSIGN(Op, DAG);
677    case ISD::FRAMEADDR:          return LowerFRAMEADDR(Op, DAG);
678    case ISD::MEMBARRIER:         return LowerMEMBARRIER(Op, DAG);
679    case ISD::ATOMIC_FENCE:       return LowerATOMIC_FENCE(Op, DAG);
680  }
681  return SDValue();
682}
683
684//===----------------------------------------------------------------------===//
685//  Lower helper functions
686//===----------------------------------------------------------------------===//
687
688// AddLiveIn - This helper function adds the specified physical register to the
689// MachineFunction as a live in value.  It also creates a corresponding
690// virtual register for it.
691static unsigned
692AddLiveIn(MachineFunction &MF, unsigned PReg, TargetRegisterClass *RC)
693{
694  assert(RC->contains(PReg) && "Not the correct regclass!");
695  unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
696  MF.getRegInfo().addLiveIn(PReg, VReg);
697  return VReg;
698}
699
700// Get fp branch code (not opcode) from condition code.
701static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
702  if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
703    return Mips::BRANCH_T;
704
705  if (CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT)
706    return Mips::BRANCH_F;
707
708  return Mips::BRANCH_INVALID;
709}
710
711static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
712                                        DebugLoc dl,
713                                        const MipsSubtarget* Subtarget,
714                                        const TargetInstrInfo *TII,
715                                        bool isFPCmp, unsigned Opc) {
716  // There is no need to expand CMov instructions if target has
717  // conditional moves.
718  if (Subtarget->hasCondMov())
719    return BB;
720
721  // To "insert" a SELECT_CC instruction, we actually have to insert the
722  // diamond control-flow pattern.  The incoming instruction knows the
723  // destination vreg to set, the condition code register to branch on, the
724  // true/false values to select between, and a branch opcode to use.
725  const BasicBlock *LLVM_BB = BB->getBasicBlock();
726  MachineFunction::iterator It = BB;
727  ++It;
728
729  //  thisMBB:
730  //  ...
731  //   TrueVal = ...
732  //   setcc r1, r2, r3
733  //   bNE   r1, r0, copy1MBB
734  //   fallthrough --> copy0MBB
735  MachineBasicBlock *thisMBB  = BB;
736  MachineFunction *F = BB->getParent();
737  MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
738  MachineBasicBlock *sinkMBB  = F->CreateMachineBasicBlock(LLVM_BB);
739  F->insert(It, copy0MBB);
740  F->insert(It, sinkMBB);
741
742  // Transfer the remainder of BB and its successor edges to sinkMBB.
743  sinkMBB->splice(sinkMBB->begin(), BB,
744                  llvm::next(MachineBasicBlock::iterator(MI)),
745                  BB->end());
746  sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
747
748  // Next, add the true and fallthrough blocks as its successors.
749  BB->addSuccessor(copy0MBB);
750  BB->addSuccessor(sinkMBB);
751
752  // Emit the right instruction according to the type of the operands compared
753  if (isFPCmp)
754    BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
755  else
756    BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
757      .addReg(Mips::ZERO).addMBB(sinkMBB);
758
759  //  copy0MBB:
760  //   %FalseValue = ...
761  //   # fallthrough to sinkMBB
762  BB = copy0MBB;
763
764  // Update machine-CFG edges
765  BB->addSuccessor(sinkMBB);
766
767  //  sinkMBB:
768  //   %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
769  //  ...
770  BB = sinkMBB;
771
772  if (isFPCmp)
773    BuildMI(*BB, BB->begin(), dl,
774            TII->get(Mips::PHI), MI->getOperand(0).getReg())
775      .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
776      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
777  else
778    BuildMI(*BB, BB->begin(), dl,
779            TII->get(Mips::PHI), MI->getOperand(0).getReg())
780      .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
781      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
782
783  MI->eraseFromParent();   // The pseudo instruction is gone now.
784  return BB;
785}
786
787MachineBasicBlock *
788MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
789                                                MachineBasicBlock *BB) const {
790  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
791  DebugLoc dl = MI->getDebugLoc();
792
793  switch (MI->getOpcode()) {
794  default:
795    assert(false && "Unexpected instr type to insert");
796    return NULL;
797  case Mips::MOVT:
798  case Mips::MOVT_S:
799  case Mips::MOVT_D:
800    return ExpandCondMov(MI, BB, dl, Subtarget, TII, true, Mips::BC1F);
801  case Mips::MOVF:
802  case Mips::MOVF_S:
803  case Mips::MOVF_D:
804    return ExpandCondMov(MI, BB, dl, Subtarget, TII, true, Mips::BC1T);
805  case Mips::MOVZ_I:
806  case Mips::MOVZ_S:
807  case Mips::MOVZ_D:
808    return ExpandCondMov(MI, BB, dl, Subtarget, TII, false, Mips::BNE);
809  case Mips::MOVN_I:
810  case Mips::MOVN_S:
811  case Mips::MOVN_D:
812    return ExpandCondMov(MI, BB, dl, Subtarget, TII, false, Mips::BEQ);
813
814  case Mips::ATOMIC_LOAD_ADD_I8:
815    return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
816  case Mips::ATOMIC_LOAD_ADD_I16:
817    return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
818  case Mips::ATOMIC_LOAD_ADD_I32:
819    return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
820
821  case Mips::ATOMIC_LOAD_AND_I8:
822    return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
823  case Mips::ATOMIC_LOAD_AND_I16:
824    return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
825  case Mips::ATOMIC_LOAD_AND_I32:
826    return EmitAtomicBinary(MI, BB, 4, Mips::AND);
827
828  case Mips::ATOMIC_LOAD_OR_I8:
829    return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
830  case Mips::ATOMIC_LOAD_OR_I16:
831    return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
832  case Mips::ATOMIC_LOAD_OR_I32:
833    return EmitAtomicBinary(MI, BB, 4, Mips::OR);
834
835  case Mips::ATOMIC_LOAD_XOR_I8:
836    return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
837  case Mips::ATOMIC_LOAD_XOR_I16:
838    return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
839  case Mips::ATOMIC_LOAD_XOR_I32:
840    return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
841
842  case Mips::ATOMIC_LOAD_NAND_I8:
843    return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
844  case Mips::ATOMIC_LOAD_NAND_I16:
845    return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
846  case Mips::ATOMIC_LOAD_NAND_I32:
847    return EmitAtomicBinary(MI, BB, 4, 0, true);
848
849  case Mips::ATOMIC_LOAD_SUB_I8:
850    return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
851  case Mips::ATOMIC_LOAD_SUB_I16:
852    return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
853  case Mips::ATOMIC_LOAD_SUB_I32:
854    return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
855
856  case Mips::ATOMIC_SWAP_I8:
857    return EmitAtomicBinaryPartword(MI, BB, 1, 0);
858  case Mips::ATOMIC_SWAP_I16:
859    return EmitAtomicBinaryPartword(MI, BB, 2, 0);
860  case Mips::ATOMIC_SWAP_I32:
861    return EmitAtomicBinary(MI, BB, 4, 0);
862
863  case Mips::ATOMIC_CMP_SWAP_I8:
864    return EmitAtomicCmpSwapPartword(MI, BB, 1);
865  case Mips::ATOMIC_CMP_SWAP_I16:
866    return EmitAtomicCmpSwapPartword(MI, BB, 2);
867  case Mips::ATOMIC_CMP_SWAP_I32:
868    return EmitAtomicCmpSwap(MI, BB, 4);
869  }
870}
871
872// This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
873// Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
874MachineBasicBlock *
875MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
876                                     unsigned Size, unsigned BinOpcode,
877                                     bool Nand) const {
878  assert(Size == 4 && "Unsupported size for EmitAtomicBinary.");
879
880  MachineFunction *MF = BB->getParent();
881  MachineRegisterInfo &RegInfo = MF->getRegInfo();
882  const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
883  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
884  DebugLoc dl = MI->getDebugLoc();
885
886  unsigned OldVal = MI->getOperand(0).getReg();
887  unsigned Ptr = MI->getOperand(1).getReg();
888  unsigned Incr = MI->getOperand(2).getReg();
889
890  unsigned StoreVal = RegInfo.createVirtualRegister(RC);
891  unsigned AndRes = RegInfo.createVirtualRegister(RC);
892  unsigned Success = RegInfo.createVirtualRegister(RC);
893
894  // insert new blocks after the current block
895  const BasicBlock *LLVM_BB = BB->getBasicBlock();
896  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
897  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
898  MachineFunction::iterator It = BB;
899  ++It;
900  MF->insert(It, loopMBB);
901  MF->insert(It, exitMBB);
902
903  // Transfer the remainder of BB and its successor edges to exitMBB.
904  exitMBB->splice(exitMBB->begin(), BB,
905                  llvm::next(MachineBasicBlock::iterator(MI)),
906                  BB->end());
907  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
908
909  //  thisMBB:
910  //    ...
911  //    fallthrough --> loopMBB
912  BB->addSuccessor(loopMBB);
913  loopMBB->addSuccessor(loopMBB);
914  loopMBB->addSuccessor(exitMBB);
915
916  //  loopMBB:
917  //    ll oldval, 0(ptr)
918  //    <binop> storeval, oldval, incr
919  //    sc success, storeval, 0(ptr)
920  //    beq success, $0, loopMBB
921  BB = loopMBB;
922  BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(Ptr).addImm(0);
923  if (Nand) {
924    //  and andres, oldval, incr
925    //  nor storeval, $0, andres
926    BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr);
927    BuildMI(BB, dl, TII->get(Mips::NOR), StoreVal)
928      .addReg(Mips::ZERO).addReg(AndRes);
929  } else if (BinOpcode) {
930    //  <binop> storeval, oldval, incr
931    BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
932  } else {
933    StoreVal = Incr;
934  }
935  BuildMI(BB, dl, TII->get(Mips::SC), Success)
936    .addReg(StoreVal).addReg(Ptr).addImm(0);
937  BuildMI(BB, dl, TII->get(Mips::BEQ))
938    .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
939
940  MI->eraseFromParent();   // The instruction is gone now.
941
942  return exitMBB;
943}
944
945MachineBasicBlock *
946MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
947                                             MachineBasicBlock *BB,
948                                             unsigned Size, unsigned BinOpcode,
949                                             bool Nand) const {
950  assert((Size == 1 || Size == 2) &&
951      "Unsupported size for EmitAtomicBinaryPartial.");
952
953  MachineFunction *MF = BB->getParent();
954  MachineRegisterInfo &RegInfo = MF->getRegInfo();
955  const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
956  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
957  DebugLoc dl = MI->getDebugLoc();
958
959  unsigned Dest = MI->getOperand(0).getReg();
960  unsigned Ptr = MI->getOperand(1).getReg();
961  unsigned Incr = MI->getOperand(2).getReg();
962
963  unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
964  unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
965  unsigned Mask = RegInfo.createVirtualRegister(RC);
966  unsigned Mask2 = RegInfo.createVirtualRegister(RC);
967  unsigned NewVal = RegInfo.createVirtualRegister(RC);
968  unsigned OldVal = RegInfo.createVirtualRegister(RC);
969  unsigned Incr2 = RegInfo.createVirtualRegister(RC);
970  unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
971  unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
972  unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
973  unsigned AndRes = RegInfo.createVirtualRegister(RC);
974  unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
975  unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
976  unsigned StoreVal = RegInfo.createVirtualRegister(RC);
977  unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
978  unsigned SrlRes = RegInfo.createVirtualRegister(RC);
979  unsigned SllRes = RegInfo.createVirtualRegister(RC);
980  unsigned Success = RegInfo.createVirtualRegister(RC);
981
982  // insert new blocks after the current block
983  const BasicBlock *LLVM_BB = BB->getBasicBlock();
984  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
985  MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
986  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
987  MachineFunction::iterator It = BB;
988  ++It;
989  MF->insert(It, loopMBB);
990  MF->insert(It, sinkMBB);
991  MF->insert(It, exitMBB);
992
993  // Transfer the remainder of BB and its successor edges to exitMBB.
994  exitMBB->splice(exitMBB->begin(), BB,
995                  llvm::next(MachineBasicBlock::iterator(MI)),
996                  BB->end());
997  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
998
999  BB->addSuccessor(loopMBB);
1000  loopMBB->addSuccessor(loopMBB);
1001  loopMBB->addSuccessor(sinkMBB);
1002  sinkMBB->addSuccessor(exitMBB);
1003
1004  //  thisMBB:
1005  //    addiu   masklsb2,$0,-4                # 0xfffffffc
1006  //    and     alignedaddr,ptr,masklsb2
1007  //    andi    ptrlsb2,ptr,3
1008  //    sll     shiftamt,ptrlsb2,3
1009  //    ori     maskupper,$0,255               # 0xff
1010  //    sll     mask,maskupper,shiftamt
1011  //    nor     mask2,$0,mask
1012  //    sll     incr2,incr,shiftamt
1013
1014  int64_t MaskImm = (Size == 1) ? 255 : 65535;
1015  BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1016    .addReg(Mips::ZERO).addImm(-4);
1017  BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1018    .addReg(Ptr).addReg(MaskLSB2);
1019  BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1020  BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1021  BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1022    .addReg(Mips::ZERO).addImm(MaskImm);
1023  BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1024    .addReg(ShiftAmt).addReg(MaskUpper);
1025  BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1026  BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1027
1028
1029  // atomic.load.binop
1030  // loopMBB:
1031  //   ll      oldval,0(alignedaddr)
1032  //   binop   binopres,oldval,incr2
1033  //   and     newval,binopres,mask
1034  //   and     maskedoldval0,oldval,mask2
1035  //   or      storeval,maskedoldval0,newval
1036  //   sc      success,storeval,0(alignedaddr)
1037  //   beq     success,$0,loopMBB
1038
1039  // atomic.swap
1040  // loopMBB:
1041  //   ll      oldval,0(alignedaddr)
1042  //   and     newval,incr2,mask
1043  //   and     maskedoldval0,oldval,mask2
1044  //   or      storeval,maskedoldval0,newval
1045  //   sc      success,storeval,0(alignedaddr)
1046  //   beq     success,$0,loopMBB
1047
1048  BB = loopMBB;
1049  BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1050  if (Nand) {
1051    //  and andres, oldval, incr2
1052    //  nor binopres, $0, andres
1053    //  and newval, binopres, mask
1054    BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1055    BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1056      .addReg(Mips::ZERO).addReg(AndRes);
1057    BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1058  } else if (BinOpcode) {
1059    //  <binop> binopres, oldval, incr2
1060    //  and newval, binopres, mask
1061    BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1062    BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1063  } else {// atomic.swap
1064    //  and newval, incr2, mask
1065    BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1066  }
1067
1068  BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1069    .addReg(OldVal).addReg(Mask2);
1070  BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1071    .addReg(MaskedOldVal0).addReg(NewVal);
1072  BuildMI(BB, dl, TII->get(Mips::SC), Success)
1073    .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1074  BuildMI(BB, dl, TII->get(Mips::BEQ))
1075    .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1076
1077  //  sinkMBB:
1078  //    and     maskedoldval1,oldval,mask
1079  //    srl     srlres,maskedoldval1,shiftamt
1080  //    sll     sllres,srlres,24
1081  //    sra     dest,sllres,24
1082  BB = sinkMBB;
1083  int64_t ShiftImm = (Size == 1) ? 24 : 16;
1084
1085  BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1086    .addReg(OldVal).addReg(Mask);
1087  BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1088      .addReg(ShiftAmt).addReg(MaskedOldVal1);
1089  BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1090      .addReg(SrlRes).addImm(ShiftImm);
1091  BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1092      .addReg(SllRes).addImm(ShiftImm);
1093
1094  MI->eraseFromParent();   // The instruction is gone now.
1095
1096  return exitMBB;
1097}
1098
1099MachineBasicBlock *
1100MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1101                                      MachineBasicBlock *BB,
1102                                      unsigned Size) const {
1103  assert(Size == 4 && "Unsupported size for EmitAtomicCmpSwap.");
1104
1105  MachineFunction *MF = BB->getParent();
1106  MachineRegisterInfo &RegInfo = MF->getRegInfo();
1107  const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1108  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1109  DebugLoc dl = MI->getDebugLoc();
1110
1111  unsigned Dest    = MI->getOperand(0).getReg();
1112  unsigned Ptr     = MI->getOperand(1).getReg();
1113  unsigned OldVal  = MI->getOperand(2).getReg();
1114  unsigned NewVal  = MI->getOperand(3).getReg();
1115
1116  unsigned Success = RegInfo.createVirtualRegister(RC);
1117
1118  // insert new blocks after the current block
1119  const BasicBlock *LLVM_BB = BB->getBasicBlock();
1120  MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1121  MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1122  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1123  MachineFunction::iterator It = BB;
1124  ++It;
1125  MF->insert(It, loop1MBB);
1126  MF->insert(It, loop2MBB);
1127  MF->insert(It, exitMBB);
1128
1129  // Transfer the remainder of BB and its successor edges to exitMBB.
1130  exitMBB->splice(exitMBB->begin(), BB,
1131                  llvm::next(MachineBasicBlock::iterator(MI)),
1132                  BB->end());
1133  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1134
1135  //  thisMBB:
1136  //    ...
1137  //    fallthrough --> loop1MBB
1138  BB->addSuccessor(loop1MBB);
1139  loop1MBB->addSuccessor(exitMBB);
1140  loop1MBB->addSuccessor(loop2MBB);
1141  loop2MBB->addSuccessor(loop1MBB);
1142  loop2MBB->addSuccessor(exitMBB);
1143
1144  // loop1MBB:
1145  //   ll dest, 0(ptr)
1146  //   bne dest, oldval, exitMBB
1147  BB = loop1MBB;
1148  BuildMI(BB, dl, TII->get(Mips::LL), Dest).addReg(Ptr).addImm(0);
1149  BuildMI(BB, dl, TII->get(Mips::BNE))
1150    .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1151
1152  // loop2MBB:
1153  //   sc success, newval, 0(ptr)
1154  //   beq success, $0, loop1MBB
1155  BB = loop2MBB;
1156  BuildMI(BB, dl, TII->get(Mips::SC), Success)
1157    .addReg(NewVal).addReg(Ptr).addImm(0);
1158  BuildMI(BB, dl, TII->get(Mips::BEQ))
1159    .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1160
1161  MI->eraseFromParent();   // The instruction is gone now.
1162
1163  return exitMBB;
1164}
1165
1166MachineBasicBlock *
1167MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1168                                              MachineBasicBlock *BB,
1169                                              unsigned Size) const {
1170  assert((Size == 1 || Size == 2) &&
1171      "Unsupported size for EmitAtomicCmpSwapPartial.");
1172
1173  MachineFunction *MF = BB->getParent();
1174  MachineRegisterInfo &RegInfo = MF->getRegInfo();
1175  const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1176  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1177  DebugLoc dl = MI->getDebugLoc();
1178
1179  unsigned Dest    = MI->getOperand(0).getReg();
1180  unsigned Ptr     = MI->getOperand(1).getReg();
1181  unsigned CmpVal  = MI->getOperand(2).getReg();
1182  unsigned NewVal  = MI->getOperand(3).getReg();
1183
1184  unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1185  unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1186  unsigned Mask = RegInfo.createVirtualRegister(RC);
1187  unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1188  unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1189  unsigned OldVal = RegInfo.createVirtualRegister(RC);
1190  unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1191  unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1192  unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1193  unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1194  unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1195  unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1196  unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1197  unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1198  unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1199  unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1200  unsigned SllRes = RegInfo.createVirtualRegister(RC);
1201  unsigned Success = RegInfo.createVirtualRegister(RC);
1202
1203  // insert new blocks after the current block
1204  const BasicBlock *LLVM_BB = BB->getBasicBlock();
1205  MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1206  MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1207  MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1208  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1209  MachineFunction::iterator It = BB;
1210  ++It;
1211  MF->insert(It, loop1MBB);
1212  MF->insert(It, loop2MBB);
1213  MF->insert(It, sinkMBB);
1214  MF->insert(It, exitMBB);
1215
1216  // Transfer the remainder of BB and its successor edges to exitMBB.
1217  exitMBB->splice(exitMBB->begin(), BB,
1218                  llvm::next(MachineBasicBlock::iterator(MI)),
1219                  BB->end());
1220  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1221
1222  BB->addSuccessor(loop1MBB);
1223  loop1MBB->addSuccessor(sinkMBB);
1224  loop1MBB->addSuccessor(loop2MBB);
1225  loop2MBB->addSuccessor(loop1MBB);
1226  loop2MBB->addSuccessor(sinkMBB);
1227  sinkMBB->addSuccessor(exitMBB);
1228
1229  // FIXME: computation of newval2 can be moved to loop2MBB.
1230  //  thisMBB:
1231  //    addiu   masklsb2,$0,-4                # 0xfffffffc
1232  //    and     alignedaddr,ptr,masklsb2
1233  //    andi    ptrlsb2,ptr,3
1234  //    sll     shiftamt,ptrlsb2,3
1235  //    ori     maskupper,$0,255               # 0xff
1236  //    sll     mask,maskupper,shiftamt
1237  //    nor     mask2,$0,mask
1238  //    andi    maskedcmpval,cmpval,255
1239  //    sll     shiftedcmpval,maskedcmpval,shiftamt
1240  //    andi    maskednewval,newval,255
1241  //    sll     shiftednewval,maskednewval,shiftamt
1242  int64_t MaskImm = (Size == 1) ? 255 : 65535;
1243  BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1244    .addReg(Mips::ZERO).addImm(-4);
1245  BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1246    .addReg(Ptr).addReg(MaskLSB2);
1247  BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1248  BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1249  BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1250    .addReg(Mips::ZERO).addImm(MaskImm);
1251  BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1252    .addReg(ShiftAmt).addReg(MaskUpper);
1253  BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1254  BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1255    .addReg(CmpVal).addImm(MaskImm);
1256  BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1257    .addReg(ShiftAmt).addReg(MaskedCmpVal);
1258  BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1259    .addReg(NewVal).addImm(MaskImm);
1260  BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1261    .addReg(ShiftAmt).addReg(MaskedNewVal);
1262
1263  //  loop1MBB:
1264  //    ll      oldval,0(alginedaddr)
1265  //    and     maskedoldval0,oldval,mask
1266  //    bne     maskedoldval0,shiftedcmpval,sinkMBB
1267  BB = loop1MBB;
1268  BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1269  BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1270    .addReg(OldVal).addReg(Mask);
1271  BuildMI(BB, dl, TII->get(Mips::BNE))
1272    .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1273
1274  //  loop2MBB:
1275  //    and     maskedoldval1,oldval,mask2
1276  //    or      storeval,maskedoldval1,shiftednewval
1277  //    sc      success,storeval,0(alignedaddr)
1278  //    beq     success,$0,loop1MBB
1279  BB = loop2MBB;
1280  BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1281    .addReg(OldVal).addReg(Mask2);
1282  BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1283    .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1284  BuildMI(BB, dl, TII->get(Mips::SC), Success)
1285      .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1286  BuildMI(BB, dl, TII->get(Mips::BEQ))
1287      .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1288
1289  //  sinkMBB:
1290  //    srl     srlres,maskedoldval0,shiftamt
1291  //    sll     sllres,srlres,24
1292  //    sra     dest,sllres,24
1293  BB = sinkMBB;
1294  int64_t ShiftImm = (Size == 1) ? 24 : 16;
1295
1296  BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1297      .addReg(ShiftAmt).addReg(MaskedOldVal0);
1298  BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1299      .addReg(SrlRes).addImm(ShiftImm);
1300  BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1301      .addReg(SllRes).addImm(ShiftImm);
1302
1303  MI->eraseFromParent();   // The instruction is gone now.
1304
1305  return exitMBB;
1306}
1307
1308//===----------------------------------------------------------------------===//
1309//  Misc Lower Operation implementation
1310//===----------------------------------------------------------------------===//
1311SDValue MipsTargetLowering::
1312LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const
1313{
1314  MachineFunction &MF = DAG.getMachineFunction();
1315  MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1316
1317  assert(getTargetMachine().getFrameLowering()->getStackAlignment() >=
1318         cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue() &&
1319         "Cannot lower if the alignment of the allocated space is larger than \
1320          that of the stack.");
1321
1322  SDValue Chain = Op.getOperand(0);
1323  SDValue Size = Op.getOperand(1);
1324  DebugLoc dl = Op.getDebugLoc();
1325
1326  // Get a reference from Mips stack pointer
1327  SDValue StackPointer = DAG.getCopyFromReg(Chain, dl, Mips::SP, MVT::i32);
1328
1329  // Subtract the dynamic size from the actual stack size to
1330  // obtain the new stack size.
1331  SDValue Sub = DAG.getNode(ISD::SUB, dl, MVT::i32, StackPointer, Size);
1332
1333  // The Sub result contains the new stack start address, so it
1334  // must be placed in the stack pointer register.
1335  Chain = DAG.getCopyToReg(StackPointer.getValue(1), dl, Mips::SP, Sub,
1336                           SDValue());
1337
1338  // This node always has two return values: a new stack pointer
1339  // value and a chain
1340  SDVTList VTLs = DAG.getVTList(MVT::i32, MVT::Other);
1341  SDValue Ptr = DAG.getFrameIndex(MipsFI->getDynAllocFI(), getPointerTy());
1342  SDValue Ops[] = { Chain, Ptr, Chain.getValue(1) };
1343
1344  return DAG.getNode(MipsISD::DynAlloc, dl, VTLs, Ops, 3);
1345}
1346
1347SDValue MipsTargetLowering::
1348LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1349{
1350  // The first operand is the chain, the second is the condition, the third is
1351  // the block to branch to if the condition is true.
1352  SDValue Chain = Op.getOperand(0);
1353  SDValue Dest = Op.getOperand(2);
1354  DebugLoc dl = Op.getDebugLoc();
1355
1356  SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1357
1358  // Return if flag is not set by a floating point comparison.
1359  if (CondRes.getOpcode() != MipsISD::FPCmp)
1360    return Op;
1361
1362  SDValue CCNode  = CondRes.getOperand(2);
1363  Mips::CondCode CC =
1364    (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1365  SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1366
1367  return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1368                     Dest, CondRes);
1369}
1370
1371SDValue MipsTargetLowering::
1372LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1373{
1374  SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1375
1376  // Return if flag is not set by a floating point comparison.
1377  if (Cond.getOpcode() != MipsISD::FPCmp)
1378    return Op;
1379
1380  return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1381                      Op.getDebugLoc());
1382}
1383
1384SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1385                                               SelectionDAG &DAG) const {
1386  // FIXME there isn't actually debug info here
1387  DebugLoc dl = Op.getDebugLoc();
1388  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1389
1390  if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1391    SDVTList VTs = DAG.getVTList(MVT::i32);
1392
1393    MipsTargetObjectFile &TLOF = (MipsTargetObjectFile&)getObjFileLowering();
1394
1395    // %gp_rel relocation
1396    if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1397      SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1398                                              MipsII::MO_GPREL);
1399      SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl, VTs, &GA, 1);
1400      SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1401      return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode);
1402    }
1403    // %hi/%lo relocation
1404    SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1405                                              MipsII::MO_ABS_HI);
1406    SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1407                                              MipsII::MO_ABS_LO);
1408    SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, VTs, &GAHi, 1);
1409    SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1410    return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1411  }
1412
1413  EVT ValTy = Op.getValueType();
1414  bool HasGotOfst = (GV->hasInternalLinkage() ||
1415                     (GV->hasLocalLinkage() && !isa<Function>(GV)));
1416  unsigned GotFlag = IsN64 ?
1417                     (HasGotOfst ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT_DISP) :
1418                     MipsII::MO_GOT;
1419  SDValue GA = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0, GotFlag);
1420  GA = DAG.getNode(MipsISD::WrapperPIC, dl, ValTy, GA);
1421  SDValue ResNode = DAG.getLoad(ValTy, dl,
1422                                DAG.getEntryNode(), GA, MachinePointerInfo(),
1423                                false, false, 0);
1424  // On functions and global targets not internal linked only
1425  // a load from got/GP is necessary for PIC to work.
1426  if (!HasGotOfst)
1427    return ResNode;
1428  SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0,
1429                                            IsN64 ? MipsII::MO_GOT_OFST :
1430                                                    MipsII::MO_ABS_LO);
1431  SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, GALo);
1432  return DAG.getNode(ISD::ADD, dl, ValTy, ResNode, Lo);
1433}
1434
1435SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1436                                              SelectionDAG &DAG) const {
1437  const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1438  // FIXME there isn't actually debug info here
1439  DebugLoc dl = Op.getDebugLoc();
1440
1441  if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1442    // %hi/%lo relocation
1443    SDValue BAHi = DAG.getBlockAddress(BA, MVT::i32, true,
1444                                       MipsII::MO_ABS_HI);
1445    SDValue BALo = DAG.getBlockAddress(BA, MVT::i32, true,
1446                                       MipsII::MO_ABS_LO);
1447    SDValue Hi = DAG.getNode(MipsISD::Hi, dl, MVT::i32, BAHi);
1448    SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALo);
1449    return DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1450  }
1451
1452  SDValue BAGOTOffset = DAG.getBlockAddress(BA, MVT::i32, true,
1453                                            MipsII::MO_GOT);
1454  BAGOTOffset = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, BAGOTOffset);
1455  SDValue BALOOffset = DAG.getBlockAddress(BA, MVT::i32, true,
1456                                           MipsII::MO_ABS_LO);
1457  SDValue Load = DAG.getLoad(MVT::i32, dl,
1458                             DAG.getEntryNode(), BAGOTOffset,
1459                             MachinePointerInfo(), false, false, 0);
1460  SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALOOffset);
1461  return DAG.getNode(ISD::ADD, dl, MVT::i32, Load, Lo);
1462}
1463
1464SDValue MipsTargetLowering::
1465LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1466{
1467  // If the relocation model is PIC, use the General Dynamic TLS Model,
1468  // otherwise use the Initial Exec or Local Exec TLS Model.
1469  // TODO: implement Local Dynamic TLS model
1470
1471  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1472  DebugLoc dl = GA->getDebugLoc();
1473  const GlobalValue *GV = GA->getGlobal();
1474  EVT PtrVT = getPointerTy();
1475
1476  if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1477    // General Dynamic TLS Model
1478    SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32,
1479                                             0, MipsII::MO_TLSGD);
1480    SDValue Tlsgd = DAG.getNode(MipsISD::TlsGd, dl, MVT::i32, TGA);
1481    SDValue GP = DAG.getRegister(Mips::GP, MVT::i32);
1482    SDValue Argument = DAG.getNode(ISD::ADD, dl, MVT::i32, GP, Tlsgd);
1483
1484    ArgListTy Args;
1485    ArgListEntry Entry;
1486    Entry.Node = Argument;
1487    Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
1488    Args.push_back(Entry);
1489    std::pair<SDValue, SDValue> CallResult =
1490        LowerCallTo(DAG.getEntryNode(),
1491                    (Type *) Type::getInt32Ty(*DAG.getContext()),
1492                    false, false, false, false, 0, CallingConv::C, false, true,
1493                    DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG,
1494                    dl);
1495
1496    return CallResult.first;
1497  }
1498
1499  SDValue Offset;
1500  if (GV->isDeclaration()) {
1501    // Initial Exec TLS Model
1502    SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1503                                             MipsII::MO_GOTTPREL);
1504    Offset = DAG.getLoad(MVT::i32, dl,
1505                         DAG.getEntryNode(), TGA, MachinePointerInfo(),
1506                         false, false, 0);
1507  } else {
1508    // Local Exec TLS Model
1509    SDVTList VTs = DAG.getVTList(MVT::i32);
1510    SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1511                                               MipsII::MO_TPREL_HI);
1512    SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1513                                               MipsII::MO_TPREL_LO);
1514    SDValue Hi = DAG.getNode(MipsISD::TprelHi, dl, VTs, &TGAHi, 1);
1515    SDValue Lo = DAG.getNode(MipsISD::TprelLo, dl, MVT::i32, TGALo);
1516    Offset = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1517  }
1518
1519  SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1520  return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1521}
1522
1523SDValue MipsTargetLowering::
1524LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1525{
1526  SDValue ResNode;
1527  SDValue HiPart;
1528  // FIXME there isn't actually debug info here
1529  DebugLoc dl = Op.getDebugLoc();
1530  bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1531  unsigned char OpFlag = IsPIC ? MipsII::MO_GOT : MipsII::MO_ABS_HI;
1532
1533  EVT PtrVT = Op.getValueType();
1534  JumpTableSDNode *JT  = cast<JumpTableSDNode>(Op);
1535
1536  SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
1537
1538  if (!IsPIC) {
1539    SDValue Ops[] = { JTI };
1540    HiPart = DAG.getNode(MipsISD::Hi, dl, DAG.getVTList(MVT::i32), Ops, 1);
1541  } else {// Emit Load from Global Pointer
1542    JTI = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, JTI);
1543    HiPart = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(), JTI,
1544                         MachinePointerInfo(),
1545                         false, false, 0);
1546  }
1547
1548  SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
1549                                         MipsII::MO_ABS_LO);
1550  SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, JTILo);
1551  ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1552
1553  return ResNode;
1554}
1555
1556SDValue MipsTargetLowering::
1557LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1558{
1559  SDValue ResNode;
1560  ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1561  const Constant *C = N->getConstVal();
1562  // FIXME there isn't actually debug info here
1563  DebugLoc dl = Op.getDebugLoc();
1564
1565  // gp_rel relocation
1566  // FIXME: we should reference the constant pool using small data sections,
1567  // but the asm printer currently doesn't support this feature without
1568  // hacking it. This feature should come soon so we can uncomment the
1569  // stuff below.
1570  //if (IsInSmallSection(C->getType())) {
1571  //  SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1572  //  SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1573  //  ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1574
1575  if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1576    SDValue CPHi = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1577                                             N->getOffset(), MipsII::MO_ABS_HI);
1578    SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1579                                             N->getOffset(), MipsII::MO_ABS_LO);
1580    SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, MVT::i32, CPHi);
1581    SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1582    ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1583  } else {
1584    SDValue CP = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1585                                           N->getOffset(), MipsII::MO_GOT);
1586    CP = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, CP);
1587    SDValue Load = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(),
1588                               CP, MachinePointerInfo::getConstantPool(),
1589                               false, false, 0);
1590    SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1591                                             N->getOffset(), MipsII::MO_ABS_LO);
1592    SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1593    ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, Load, Lo);
1594  }
1595
1596  return ResNode;
1597}
1598
1599SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1600  MachineFunction &MF = DAG.getMachineFunction();
1601  MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1602
1603  DebugLoc dl = Op.getDebugLoc();
1604  SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1605                                 getPointerTy());
1606
1607  // vastart just stores the address of the VarArgsFrameIndex slot into the
1608  // memory location argument.
1609  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1610  return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1611                      MachinePointerInfo(SV),
1612                      false, false, 0);
1613}
1614
1615static SDValue LowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG) {
1616  // FIXME: Use ext/ins instructions if target architecture is Mips32r2.
1617  DebugLoc dl = Op.getDebugLoc();
1618  SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op.getOperand(0));
1619  SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op.getOperand(1));
1620  SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op0,
1621                             DAG.getConstant(0x7fffffff, MVT::i32));
1622  SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op1,
1623                             DAG.getConstant(0x80000000, MVT::i32));
1624  SDValue Result = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1625  return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Result);
1626}
1627
1628static SDValue LowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool isLittle) {
1629  // FIXME:
1630  //  Use ext/ins instructions if target architecture is Mips32r2.
1631  //  Eliminate redundant mfc1 and mtc1 instructions.
1632  unsigned LoIdx = 0, HiIdx = 1;
1633
1634  if (!isLittle)
1635    std::swap(LoIdx, HiIdx);
1636
1637  DebugLoc dl = Op.getDebugLoc();
1638  SDValue Word0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1639                              Op.getOperand(0),
1640                              DAG.getConstant(LoIdx, MVT::i32));
1641  SDValue Hi0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1642                            Op.getOperand(0), DAG.getConstant(HiIdx, MVT::i32));
1643  SDValue Hi1 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1644                            Op.getOperand(1), DAG.getConstant(HiIdx, MVT::i32));
1645  SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi0,
1646                             DAG.getConstant(0x7fffffff, MVT::i32));
1647  SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi1,
1648                             DAG.getConstant(0x80000000, MVT::i32));
1649  SDValue Word1 = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1650
1651  if (!isLittle)
1652    std::swap(Word0, Word1);
1653
1654  return DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64, Word0, Word1);
1655}
1656
1657SDValue MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG)
1658  const {
1659  EVT Ty = Op.getValueType();
1660
1661  assert(Ty == MVT::f32 || Ty == MVT::f64);
1662
1663  if (Ty == MVT::f32)
1664    return LowerFCOPYSIGN32(Op, DAG);
1665  else
1666    return LowerFCOPYSIGN64(Op, DAG, Subtarget->isLittle());
1667}
1668
1669SDValue MipsTargetLowering::
1670LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1671  // check the depth
1672  assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1673         "Frame address can only be determined for current frame.");
1674
1675  MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1676  MFI->setFrameAddressIsTaken(true);
1677  EVT VT = Op.getValueType();
1678  DebugLoc dl = Op.getDebugLoc();
1679  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, Mips::FP, VT);
1680  return FrameAddr;
1681}
1682
1683// TODO: set SType according to the desired memory barrier behavior.
1684SDValue MipsTargetLowering::LowerMEMBARRIER(SDValue Op,
1685                                            SelectionDAG& DAG) const {
1686  unsigned SType = 0;
1687  DebugLoc dl = Op.getDebugLoc();
1688  return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1689                     DAG.getConstant(SType, MVT::i32));
1690}
1691
1692SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1693                                              SelectionDAG& DAG) const {
1694  // FIXME: Need pseudo-fence for 'singlethread' fences
1695  // FIXME: Set SType for weaker fences where supported/appropriate.
1696  unsigned SType = 0;
1697  DebugLoc dl = Op.getDebugLoc();
1698  return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1699                     DAG.getConstant(SType, MVT::i32));
1700}
1701
1702//===----------------------------------------------------------------------===//
1703//                      Calling Convention Implementation
1704//===----------------------------------------------------------------------===//
1705
1706#include "MipsGenCallingConv.inc"
1707
1708//===----------------------------------------------------------------------===//
1709// TODO: Implement a generic logic using tblgen that can support this.
1710// Mips O32 ABI rules:
1711// ---
1712// i32 - Passed in A0, A1, A2, A3 and stack
1713// f32 - Only passed in f32 registers if no int reg has been used yet to hold
1714//       an argument. Otherwise, passed in A1, A2, A3 and stack.
1715// f64 - Only passed in two aliased f32 registers if no int reg has been used
1716//       yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
1717//       not used, it must be shadowed. If only A3 is avaiable, shadow it and
1718//       go to stack.
1719//
1720//  For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
1721//===----------------------------------------------------------------------===//
1722
1723static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
1724                       MVT LocVT, CCValAssign::LocInfo LocInfo,
1725                       ISD::ArgFlagsTy ArgFlags, CCState &State) {
1726
1727  static const unsigned IntRegsSize=4, FloatRegsSize=2;
1728
1729  static const unsigned IntRegs[] = {
1730      Mips::A0, Mips::A1, Mips::A2, Mips::A3
1731  };
1732  static const unsigned F32Regs[] = {
1733      Mips::F12, Mips::F14
1734  };
1735  static const unsigned F64Regs[] = {
1736      Mips::D6, Mips::D7
1737  };
1738
1739  // ByVal Args
1740  if (ArgFlags.isByVal()) {
1741    State.HandleByVal(ValNo, ValVT, LocVT, LocInfo,
1742                      1 /*MinSize*/, 4 /*MinAlign*/, ArgFlags);
1743    unsigned NextReg = (State.getNextStackOffset() + 3) / 4;
1744    for (unsigned r = State.getFirstUnallocated(IntRegs, IntRegsSize);
1745         r < std::min(IntRegsSize, NextReg); ++r)
1746      State.AllocateReg(IntRegs[r]);
1747    return false;
1748  }
1749
1750  // Promote i8 and i16
1751  if (LocVT == MVT::i8 || LocVT == MVT::i16) {
1752    LocVT = MVT::i32;
1753    if (ArgFlags.isSExt())
1754      LocInfo = CCValAssign::SExt;
1755    else if (ArgFlags.isZExt())
1756      LocInfo = CCValAssign::ZExt;
1757    else
1758      LocInfo = CCValAssign::AExt;
1759  }
1760
1761  unsigned Reg;
1762
1763  // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
1764  // is true: function is vararg, argument is 3rd or higher, there is previous
1765  // argument which is not f32 or f64.
1766  bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
1767      || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
1768  unsigned OrigAlign = ArgFlags.getOrigAlign();
1769  bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
1770
1771  if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
1772    Reg = State.AllocateReg(IntRegs, IntRegsSize);
1773    // If this is the first part of an i64 arg,
1774    // the allocated register must be either A0 or A2.
1775    if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
1776      Reg = State.AllocateReg(IntRegs, IntRegsSize);
1777    LocVT = MVT::i32;
1778  } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
1779    // Allocate int register and shadow next int register. If first
1780    // available register is Mips::A1 or Mips::A3, shadow it too.
1781    Reg = State.AllocateReg(IntRegs, IntRegsSize);
1782    if (Reg == Mips::A1 || Reg == Mips::A3)
1783      Reg = State.AllocateReg(IntRegs, IntRegsSize);
1784    State.AllocateReg(IntRegs, IntRegsSize);
1785    LocVT = MVT::i32;
1786  } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
1787    // we are guaranteed to find an available float register
1788    if (ValVT == MVT::f32) {
1789      Reg = State.AllocateReg(F32Regs, FloatRegsSize);
1790      // Shadow int register
1791      State.AllocateReg(IntRegs, IntRegsSize);
1792    } else {
1793      Reg = State.AllocateReg(F64Regs, FloatRegsSize);
1794      // Shadow int registers
1795      unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
1796      if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
1797        State.AllocateReg(IntRegs, IntRegsSize);
1798      State.AllocateReg(IntRegs, IntRegsSize);
1799    }
1800  } else
1801    llvm_unreachable("Cannot handle this ValVT.");
1802
1803  unsigned SizeInBytes = ValVT.getSizeInBits() >> 3;
1804  unsigned Offset = State.AllocateStack(SizeInBytes, OrigAlign);
1805
1806  if (!Reg)
1807    State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1808  else
1809    State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1810
1811  return false; // CC must always match
1812}
1813
1814//===----------------------------------------------------------------------===//
1815//                  Call Calling Convention Implementation
1816//===----------------------------------------------------------------------===//
1817
1818static const unsigned O32IntRegsSize = 4;
1819
1820static const unsigned O32IntRegs[] = {
1821  Mips::A0, Mips::A1, Mips::A2, Mips::A3
1822};
1823
1824// Return next O32 integer argument register.
1825static unsigned getNextIntArgReg(unsigned Reg) {
1826  assert((Reg == Mips::A0) || (Reg == Mips::A2));
1827  return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
1828}
1829
1830// Write ByVal Arg to arg registers and stack.
1831static void
1832WriteByValArg(SDValue& ByValChain, SDValue Chain, DebugLoc dl,
1833              SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
1834              SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
1835              MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
1836              const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
1837              MVT PtrType, bool isLittle) {
1838  unsigned LocMemOffset = VA.getLocMemOffset();
1839  unsigned Offset = 0;
1840  uint32_t RemainingSize = Flags.getByValSize();
1841  unsigned ByValAlign = Flags.getByValAlign();
1842
1843  // Copy the first 4 words of byval arg to registers A0 - A3.
1844  // FIXME: Use a stricter alignment if it enables better optimization in passes
1845  //        run later.
1846  for (; RemainingSize >= 4 && LocMemOffset < 4 * 4;
1847       Offset += 4, RemainingSize -= 4, LocMemOffset += 4) {
1848    SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1849                                  DAG.getConstant(Offset, MVT::i32));
1850    SDValue LoadVal = DAG.getLoad(MVT::i32, dl, Chain, LoadPtr,
1851                                  MachinePointerInfo(),
1852                                  false, false, std::min(ByValAlign,
1853                                                         (unsigned )4));
1854    MemOpChains.push_back(LoadVal.getValue(1));
1855    unsigned DstReg = O32IntRegs[LocMemOffset / 4];
1856    RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
1857  }
1858
1859  if (RemainingSize == 0)
1860    return;
1861
1862  // If there still is a register available for argument passing, write the
1863  // remaining part of the structure to it using subword loads and shifts.
1864  if (LocMemOffset < 4 * 4) {
1865    assert(RemainingSize <= 3 && RemainingSize >= 1 &&
1866           "There must be one to three bytes remaining.");
1867    unsigned LoadSize = (RemainingSize == 3 ? 2 : RemainingSize);
1868    SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1869                                  DAG.getConstant(Offset, MVT::i32));
1870    unsigned Alignment = std::min(ByValAlign, (unsigned )4);
1871    SDValue LoadVal = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
1872                                     LoadPtr, MachinePointerInfo(),
1873                                     MVT::getIntegerVT(LoadSize * 8), false,
1874                                     false, Alignment);
1875    MemOpChains.push_back(LoadVal.getValue(1));
1876
1877    // If target is big endian, shift it to the most significant half-word or
1878    // byte.
1879    if (!isLittle)
1880      LoadVal = DAG.getNode(ISD::SHL, dl, MVT::i32, LoadVal,
1881                            DAG.getConstant(32 - LoadSize * 8, MVT::i32));
1882
1883    Offset += LoadSize;
1884    RemainingSize -= LoadSize;
1885
1886    // Read second subword if necessary.
1887    if (RemainingSize != 0)  {
1888      assert(RemainingSize == 1 && "There must be one byte remaining.");
1889      LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1890                            DAG.getConstant(Offset, MVT::i32));
1891      unsigned Alignment = std::min(ByValAlign, (unsigned )2);
1892      SDValue Subword = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
1893                                       LoadPtr, MachinePointerInfo(),
1894                                       MVT::i8, false, false, Alignment);
1895      MemOpChains.push_back(Subword.getValue(1));
1896      // Insert the loaded byte to LoadVal.
1897      // FIXME: Use INS if supported by target.
1898      unsigned ShiftAmt = isLittle ? 16 : 8;
1899      SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i32, Subword,
1900                                  DAG.getConstant(ShiftAmt, MVT::i32));
1901      LoadVal = DAG.getNode(ISD::OR, dl, MVT::i32, LoadVal, Shift);
1902    }
1903
1904    unsigned DstReg = O32IntRegs[LocMemOffset / 4];
1905    RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
1906    return;
1907  }
1908
1909  // Create a fixed object on stack at offset LocMemOffset and copy
1910  // remaining part of byval arg to it using memcpy.
1911  SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1912                            DAG.getConstant(Offset, MVT::i32));
1913  LastFI = MFI->CreateFixedObject(RemainingSize, LocMemOffset, true);
1914  SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
1915  ByValChain = DAG.getMemcpy(ByValChain, dl, Dst, Src,
1916                             DAG.getConstant(RemainingSize, MVT::i32),
1917                             std::min(ByValAlign, (unsigned)4),
1918                             /*isVolatile=*/false, /*AlwaysInline=*/false,
1919                             MachinePointerInfo(0), MachinePointerInfo(0));
1920}
1921
1922/// LowerCall - functions arguments are copied from virtual regs to
1923/// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
1924/// TODO: isTailCall.
1925SDValue
1926MipsTargetLowering::LowerCall(SDValue InChain, SDValue Callee,
1927                              CallingConv::ID CallConv, bool isVarArg,
1928                              bool &isTailCall,
1929                              const SmallVectorImpl<ISD::OutputArg> &Outs,
1930                              const SmallVectorImpl<SDValue> &OutVals,
1931                              const SmallVectorImpl<ISD::InputArg> &Ins,
1932                              DebugLoc dl, SelectionDAG &DAG,
1933                              SmallVectorImpl<SDValue> &InVals) const {
1934  // MIPs target does not yet support tail call optimization.
1935  isTailCall = false;
1936
1937  MachineFunction &MF = DAG.getMachineFunction();
1938  MachineFrameInfo *MFI = MF.getFrameInfo();
1939  const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
1940  bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1941  MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1942
1943  // Analyze operands of the call, assigning locations to each operand.
1944  SmallVector<CCValAssign, 16> ArgLocs;
1945  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1946		 getTargetMachine(), ArgLocs, *DAG.getContext());
1947
1948  if (Subtarget->isABI_O32())
1949    CCInfo.AnalyzeCallOperands(Outs, CC_MipsO32);
1950  else
1951    CCInfo.AnalyzeCallOperands(Outs, CC_Mips);
1952
1953  // Get a count of how many bytes are to be pushed on the stack.
1954  unsigned NextStackOffset = CCInfo.getNextStackOffset();
1955
1956  // Chain is the output chain of the last Load/Store or CopyToReg node.
1957  // ByValChain is the output chain of the last Memcpy node created for copying
1958  // byval arguments to the stack.
1959  SDValue Chain, CallSeqStart, ByValChain;
1960  SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
1961  Chain = CallSeqStart = DAG.getCALLSEQ_START(InChain, NextStackOffsetVal);
1962  ByValChain = InChain;
1963
1964  // If this is the first call, create a stack frame object that points to
1965  // a location to which .cprestore saves $gp.
1966  if (IsPIC && !MipsFI->getGPFI())
1967    MipsFI->setGPFI(MFI->CreateFixedObject(4, 0, true));
1968
1969  // Get the frame index of the stack frame object that points to the location
1970  // of dynamically allocated area on the stack.
1971  int DynAllocFI = MipsFI->getDynAllocFI();
1972
1973  // Update size of the maximum argument space.
1974  // For O32, a minimum of four words (16 bytes) of argument space is
1975  // allocated.
1976  if (Subtarget->isABI_O32())
1977    NextStackOffset = std::max(NextStackOffset, (unsigned)16);
1978
1979  unsigned MaxCallFrameSize = MipsFI->getMaxCallFrameSize();
1980
1981  if (MaxCallFrameSize < NextStackOffset) {
1982    MipsFI->setMaxCallFrameSize(NextStackOffset);
1983
1984    // Set the offsets relative to $sp of the $gp restore slot and dynamically
1985    // allocated stack space. These offsets must be aligned to a boundary
1986    // determined by the stack alignment of the ABI.
1987    unsigned StackAlignment = TFL->getStackAlignment();
1988    NextStackOffset = (NextStackOffset + StackAlignment - 1) /
1989                      StackAlignment * StackAlignment;
1990
1991    if (IsPIC)
1992      MFI->setObjectOffset(MipsFI->getGPFI(), NextStackOffset);
1993
1994    MFI->setObjectOffset(DynAllocFI, NextStackOffset);
1995  }
1996
1997  // With EABI is it possible to have 16 args on registers.
1998  SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
1999  SmallVector<SDValue, 8> MemOpChains;
2000
2001  int FirstFI = -MFI->getNumFixedObjects() - 1, LastFI = 0;
2002
2003  // Walk the register/memloc assignments, inserting copies/loads.
2004  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2005    SDValue Arg = OutVals[i];
2006    CCValAssign &VA = ArgLocs[i];
2007
2008    // Promote the value if needed.
2009    switch (VA.getLocInfo()) {
2010    default: llvm_unreachable("Unknown loc info!");
2011    case CCValAssign::Full:
2012      if (Subtarget->isABI_O32() && VA.isRegLoc()) {
2013        if (VA.getValVT() == MVT::f32 && VA.getLocVT() == MVT::i32)
2014          Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
2015        if (VA.getValVT() == MVT::f64 && VA.getLocVT() == MVT::i32) {
2016          SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2017                                   Arg, DAG.getConstant(0, MVT::i32));
2018          SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2019                                   Arg, DAG.getConstant(1, MVT::i32));
2020          if (!Subtarget->isLittle())
2021            std::swap(Lo, Hi);
2022          unsigned LocRegLo = VA.getLocReg();
2023          unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2024          RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2025          RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2026          continue;
2027        }
2028      }
2029      break;
2030    case CCValAssign::SExt:
2031      Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
2032      break;
2033    case CCValAssign::ZExt:
2034      Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
2035      break;
2036    case CCValAssign::AExt:
2037      Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
2038      break;
2039    }
2040
2041    // Arguments that can be passed on register must be kept at
2042    // RegsToPass vector
2043    if (VA.isRegLoc()) {
2044      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2045      continue;
2046    }
2047
2048    // Register can't get to this point...
2049    assert(VA.isMemLoc());
2050
2051    // ByVal Arg.
2052    ISD::ArgFlagsTy Flags = Outs[i].Flags;
2053    if (Flags.isByVal()) {
2054      assert(Subtarget->isABI_O32() &&
2055             "No support for ByVal args by ABIs other than O32 yet.");
2056      assert(Flags.getByValSize() &&
2057             "ByVal args of size 0 should have been ignored by front-end.");
2058      WriteByValArg(ByValChain, Chain, dl, RegsToPass, MemOpChains, LastFI, MFI,
2059                    DAG, Arg, VA, Flags, getPointerTy(), Subtarget->isLittle());
2060      continue;
2061    }
2062
2063    // Create the frame index object for this incoming parameter
2064    LastFI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
2065                                    VA.getLocMemOffset(), true);
2066    SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2067
2068    // emit ISD::STORE whichs stores the
2069    // parameter value to a stack Location
2070    MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2071                                       MachinePointerInfo(),
2072                                       false, false, 0));
2073  }
2074
2075  // Extend range of indices of frame objects for outgoing arguments that were
2076  // created during this function call. Skip this step if no such objects were
2077  // created.
2078  if (LastFI)
2079    MipsFI->extendOutArgFIRange(FirstFI, LastFI);
2080
2081  // If a memcpy has been created to copy a byval arg to a stack, replace the
2082  // chain input of CallSeqStart with ByValChain.
2083  if (InChain != ByValChain)
2084    DAG.UpdateNodeOperands(CallSeqStart.getNode(), ByValChain,
2085                           NextStackOffsetVal);
2086
2087  // Transform all store nodes into one single node because all store
2088  // nodes are independent of each other.
2089  if (!MemOpChains.empty())
2090    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2091                        &MemOpChains[0], MemOpChains.size());
2092
2093  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2094  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2095  // node so that legalize doesn't hack it.
2096  unsigned char OpFlag = IsPIC ? MipsII::MO_GOT_CALL : MipsII::MO_NO_FLAG;
2097  bool LoadSymAddr = false;
2098  SDValue CalleeLo;
2099
2100  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2101    if (IsPIC && G->getGlobal()->hasInternalLinkage()) {
2102      Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2103                                          getPointerTy(), 0,MipsII:: MO_GOT);
2104      CalleeLo = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(),
2105                                            0, MipsII::MO_ABS_LO);
2106    } else {
2107      Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2108                                          getPointerTy(), 0, OpFlag);
2109    }
2110
2111    LoadSymAddr = true;
2112  }
2113  else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2114    Callee = DAG.getTargetExternalSymbol(S->getSymbol(),
2115                                getPointerTy(), OpFlag);
2116    LoadSymAddr = true;
2117  }
2118
2119  SDValue InFlag;
2120
2121  // Create nodes that load address of callee and copy it to T9
2122  if (IsPIC) {
2123    if (LoadSymAddr) {
2124      // Load callee address
2125      Callee = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, Callee);
2126      SDValue LoadValue = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(), Callee,
2127                                      MachinePointerInfo::getGOT(),
2128                                      false, false, 0);
2129
2130      // Use GOT+LO if callee has internal linkage.
2131      if (CalleeLo.getNode()) {
2132        SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CalleeLo);
2133        Callee = DAG.getNode(ISD::ADD, dl, MVT::i32, LoadValue, Lo);
2134      } else
2135        Callee = LoadValue;
2136    }
2137
2138    // copy to T9
2139    Chain = DAG.getCopyToReg(Chain, dl, Mips::T9, Callee, SDValue(0, 0));
2140    InFlag = Chain.getValue(1);
2141    Callee = DAG.getRegister(Mips::T9, MVT::i32);
2142  }
2143
2144  // Build a sequence of copy-to-reg nodes chained together with token
2145  // chain and flag operands which copy the outgoing args into registers.
2146  // The InFlag in necessary since all emitted instructions must be
2147  // stuck together.
2148  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2149    Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2150                             RegsToPass[i].second, InFlag);
2151    InFlag = Chain.getValue(1);
2152  }
2153
2154  // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2155  //             = Chain, Callee, Reg#1, Reg#2, ...
2156  //
2157  // Returns a chain & a flag for retval copy to use.
2158  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2159  SmallVector<SDValue, 8> Ops;
2160  Ops.push_back(Chain);
2161  Ops.push_back(Callee);
2162
2163  // Add argument registers to the end of the list so that they are
2164  // known live into the call.
2165  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2166    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2167                                  RegsToPass[i].second.getValueType()));
2168
2169  if (InFlag.getNode())
2170    Ops.push_back(InFlag);
2171
2172  Chain  = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2173  InFlag = Chain.getValue(1);
2174
2175  // Create the CALLSEQ_END node.
2176  Chain = DAG.getCALLSEQ_END(Chain,
2177                             DAG.getIntPtrConstant(NextStackOffset, true),
2178                             DAG.getIntPtrConstant(0, true), InFlag);
2179  InFlag = Chain.getValue(1);
2180
2181  // Handle result values, copying them out of physregs into vregs that we
2182  // return.
2183  return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2184                         Ins, dl, DAG, InVals);
2185}
2186
2187/// LowerCallResult - Lower the result values of a call into the
2188/// appropriate copies out of appropriate physical registers.
2189SDValue
2190MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2191                                    CallingConv::ID CallConv, bool isVarArg,
2192                                    const SmallVectorImpl<ISD::InputArg> &Ins,
2193                                    DebugLoc dl, SelectionDAG &DAG,
2194                                    SmallVectorImpl<SDValue> &InVals) const {
2195  // Assign locations to each value returned by this call.
2196  SmallVector<CCValAssign, 16> RVLocs;
2197  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2198		 getTargetMachine(), RVLocs, *DAG.getContext());
2199
2200  CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2201
2202  // Copy all of the result registers out of their specified physreg.
2203  for (unsigned i = 0; i != RVLocs.size(); ++i) {
2204    Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2205                               RVLocs[i].getValVT(), InFlag).getValue(1);
2206    InFlag = Chain.getValue(2);
2207    InVals.push_back(Chain.getValue(0));
2208  }
2209
2210  return Chain;
2211}
2212
2213//===----------------------------------------------------------------------===//
2214//             Formal Arguments Calling Convention Implementation
2215//===----------------------------------------------------------------------===//
2216static void ReadByValArg(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2217                         std::vector<SDValue>& OutChains,
2218                         SelectionDAG &DAG, unsigned NumWords, SDValue FIN,
2219                         const CCValAssign &VA, const ISD::ArgFlagsTy& Flags) {
2220  unsigned LocMem = VA.getLocMemOffset();
2221  unsigned FirstWord = LocMem / 4;
2222
2223  // copy register A0 - A3 to frame object
2224  for (unsigned i = 0; i < NumWords; ++i) {
2225    unsigned CurWord = FirstWord + i;
2226    if (CurWord >= O32IntRegsSize)
2227      break;
2228
2229    unsigned SrcReg = O32IntRegs[CurWord];
2230    unsigned Reg = AddLiveIn(MF, SrcReg, Mips::CPURegsRegisterClass);
2231    SDValue StorePtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIN,
2232                                   DAG.getConstant(i * 4, MVT::i32));
2233    SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(Reg, MVT::i32),
2234                                 StorePtr, MachinePointerInfo(), false,
2235                                 false, 0);
2236    OutChains.push_back(Store);
2237  }
2238}
2239
2240/// LowerFormalArguments - transform physical registers into virtual registers
2241/// and generate load operations for arguments places on the stack.
2242SDValue
2243MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2244                                         CallingConv::ID CallConv,
2245                                         bool isVarArg,
2246                                         const SmallVectorImpl<ISD::InputArg>
2247                                         &Ins,
2248                                         DebugLoc dl, SelectionDAG &DAG,
2249                                         SmallVectorImpl<SDValue> &InVals)
2250                                          const {
2251  MachineFunction &MF = DAG.getMachineFunction();
2252  MachineFrameInfo *MFI = MF.getFrameInfo();
2253  MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2254
2255  MipsFI->setVarArgsFrameIndex(0);
2256
2257  // Used with vargs to acumulate store chains.
2258  std::vector<SDValue> OutChains;
2259
2260  // Assign locations to all of the incoming arguments.
2261  SmallVector<CCValAssign, 16> ArgLocs;
2262  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2263		 getTargetMachine(), ArgLocs, *DAG.getContext());
2264
2265  if (Subtarget->isABI_O32())
2266    CCInfo.AnalyzeFormalArguments(Ins, CC_MipsO32);
2267  else
2268    CCInfo.AnalyzeFormalArguments(Ins, CC_Mips);
2269
2270  int LastFI = 0;// MipsFI->LastInArgFI is 0 at the entry of this function.
2271
2272  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2273    CCValAssign &VA = ArgLocs[i];
2274
2275    // Arguments stored on registers
2276    if (VA.isRegLoc()) {
2277      EVT RegVT = VA.getLocVT();
2278      unsigned ArgReg = VA.getLocReg();
2279      TargetRegisterClass *RC = 0;
2280
2281      if (RegVT == MVT::i32)
2282        RC = Mips::CPURegsRegisterClass;
2283      else if (RegVT == MVT::i64)
2284        RC = Mips::CPU64RegsRegisterClass;
2285      else if (RegVT == MVT::f32)
2286        RC = Mips::FGR32RegisterClass;
2287      else if (RegVT == MVT::f64)
2288        RC = HasMips64 ? Mips::FGR64RegisterClass : Mips::AFGR64RegisterClass;
2289      else
2290        llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2291
2292      // Transform the arguments stored on
2293      // physical registers into virtual ones
2294      unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2295      SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2296
2297      // If this is an 8 or 16-bit value, it has been passed promoted
2298      // to 32 bits.  Insert an assert[sz]ext to capture this, then
2299      // truncate to the right size.
2300      if (VA.getLocInfo() != CCValAssign::Full) {
2301        unsigned Opcode = 0;
2302        if (VA.getLocInfo() == CCValAssign::SExt)
2303          Opcode = ISD::AssertSext;
2304        else if (VA.getLocInfo() == CCValAssign::ZExt)
2305          Opcode = ISD::AssertZext;
2306        if (Opcode)
2307          ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
2308                                 DAG.getValueType(VA.getValVT()));
2309        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2310      }
2311
2312      // Handle O32 ABI cases: i32->f32 and (i32,i32)->f64
2313      if (Subtarget->isABI_O32()) {
2314        if (RegVT == MVT::i32 && VA.getValVT() == MVT::f32)
2315          ArgValue = DAG.getNode(ISD::BITCAST, dl, MVT::f32, ArgValue);
2316        if (RegVT == MVT::i32 && VA.getValVT() == MVT::f64) {
2317          unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
2318                                    getNextIntArgReg(ArgReg), RC);
2319          SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
2320          if (!Subtarget->isLittle())
2321            std::swap(ArgValue, ArgValue2);
2322          ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
2323                                 ArgValue, ArgValue2);
2324        }
2325      }
2326
2327      InVals.push_back(ArgValue);
2328    } else { // VA.isRegLoc()
2329
2330      // sanity check
2331      assert(VA.isMemLoc());
2332
2333      ISD::ArgFlagsTy Flags = Ins[i].Flags;
2334
2335      if (Flags.isByVal()) {
2336        assert(Subtarget->isABI_O32() &&
2337               "No support for ByVal args by ABIs other than O32 yet.");
2338        assert(Flags.getByValSize() &&
2339               "ByVal args of size 0 should have been ignored by front-end.");
2340        unsigned NumWords = (Flags.getByValSize() + 3) / 4;
2341        LastFI = MFI->CreateFixedObject(NumWords * 4, VA.getLocMemOffset(),
2342                                        true);
2343        SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2344        InVals.push_back(FIN);
2345        ReadByValArg(MF, Chain, dl, OutChains, DAG, NumWords, FIN, VA, Flags);
2346
2347        continue;
2348      }
2349
2350      // The stack pointer offset is relative to the caller stack frame.
2351      LastFI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
2352                                      VA.getLocMemOffset(), true);
2353
2354      // Create load nodes to retrieve arguments from the stack
2355      SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2356      InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2357                                   MachinePointerInfo::getFixedStack(LastFI),
2358                                   false, false, 0));
2359    }
2360  }
2361
2362  // The mips ABIs for returning structs by value requires that we copy
2363  // the sret argument into $v0 for the return. Save the argument into
2364  // a virtual register so that we can access it from the return points.
2365  if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2366    unsigned Reg = MipsFI->getSRetReturnReg();
2367    if (!Reg) {
2368      Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i32));
2369      MipsFI->setSRetReturnReg(Reg);
2370    }
2371    SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2372    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2373  }
2374
2375  if (isVarArg && Subtarget->isABI_O32()) {
2376    // Record the frame index of the first variable argument
2377    // which is a value necessary to VASTART.
2378    unsigned NextStackOffset = CCInfo.getNextStackOffset();
2379    assert(NextStackOffset % 4 == 0 &&
2380           "NextStackOffset must be aligned to 4-byte boundaries.");
2381    LastFI = MFI->CreateFixedObject(4, NextStackOffset, true);
2382    MipsFI->setVarArgsFrameIndex(LastFI);
2383
2384    // If NextStackOffset is smaller than o32's 16-byte reserved argument area,
2385    // copy the integer registers that have not been used for argument passing
2386    // to the caller's stack frame.
2387    for (; NextStackOffset < 16; NextStackOffset += 4) {
2388      TargetRegisterClass *RC = Mips::CPURegsRegisterClass;
2389      unsigned Idx = NextStackOffset / 4;
2390      unsigned Reg = AddLiveIn(DAG.getMachineFunction(), O32IntRegs[Idx], RC);
2391      SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, MVT::i32);
2392      LastFI = MFI->CreateFixedObject(4, NextStackOffset, true);
2393      SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2394      OutChains.push_back(DAG.getStore(Chain, dl, ArgValue, PtrOff,
2395                                       MachinePointerInfo(),
2396                                       false, false, 0));
2397    }
2398  }
2399
2400  MipsFI->setLastInArgFI(LastFI);
2401
2402  // All stores are grouped in one node to allow the matching between
2403  // the size of Ins and InVals. This only happens when on varg functions
2404  if (!OutChains.empty()) {
2405    OutChains.push_back(Chain);
2406    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2407                        &OutChains[0], OutChains.size());
2408  }
2409
2410  return Chain;
2411}
2412
2413//===----------------------------------------------------------------------===//
2414//               Return Value Calling Convention Implementation
2415//===----------------------------------------------------------------------===//
2416
2417SDValue
2418MipsTargetLowering::LowerReturn(SDValue Chain,
2419                                CallingConv::ID CallConv, bool isVarArg,
2420                                const SmallVectorImpl<ISD::OutputArg> &Outs,
2421                                const SmallVectorImpl<SDValue> &OutVals,
2422                                DebugLoc dl, SelectionDAG &DAG) const {
2423
2424  // CCValAssign - represent the assignment of
2425  // the return value to a location
2426  SmallVector<CCValAssign, 16> RVLocs;
2427
2428  // CCState - Info about the registers and stack slot.
2429  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2430		 getTargetMachine(), RVLocs, *DAG.getContext());
2431
2432  // Analize return values.
2433  CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
2434
2435  // If this is the first return lowered for this function, add
2436  // the regs to the liveout set for the function.
2437  if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2438    for (unsigned i = 0; i != RVLocs.size(); ++i)
2439      if (RVLocs[i].isRegLoc())
2440        DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2441  }
2442
2443  SDValue Flag;
2444
2445  // Copy the result values into the output registers.
2446  for (unsigned i = 0; i != RVLocs.size(); ++i) {
2447    CCValAssign &VA = RVLocs[i];
2448    assert(VA.isRegLoc() && "Can only return in registers!");
2449
2450    Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2451                             OutVals[i], Flag);
2452
2453    // guarantee that all emitted copies are
2454    // stuck together, avoiding something bad
2455    Flag = Chain.getValue(1);
2456  }
2457
2458  // The mips ABIs for returning structs by value requires that we copy
2459  // the sret argument into $v0 for the return. We saved the argument into
2460  // a virtual register in the entry block, so now we copy the value out
2461  // and into $v0.
2462  if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2463    MachineFunction &MF      = DAG.getMachineFunction();
2464    MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2465    unsigned Reg = MipsFI->getSRetReturnReg();
2466
2467    if (!Reg)
2468      llvm_unreachable("sret virtual register not created in the entry block");
2469    SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2470
2471    Chain = DAG.getCopyToReg(Chain, dl, Mips::V0, Val, Flag);
2472    Flag = Chain.getValue(1);
2473  }
2474
2475  // Return on Mips is always a "jr $ra"
2476  if (Flag.getNode())
2477    return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2478                       Chain, DAG.getRegister(Mips::RA, MVT::i32), Flag);
2479  else // Return Void
2480    return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2481                       Chain, DAG.getRegister(Mips::RA, MVT::i32));
2482}
2483
2484//===----------------------------------------------------------------------===//
2485//                           Mips Inline Assembly Support
2486//===----------------------------------------------------------------------===//
2487
2488/// getConstraintType - Given a constraint letter, return the type of
2489/// constraint it is for this target.
2490MipsTargetLowering::ConstraintType MipsTargetLowering::
2491getConstraintType(const std::string &Constraint) const
2492{
2493  // Mips specific constrainy
2494  // GCC config/mips/constraints.md
2495  //
2496  // 'd' : An address register. Equivalent to r
2497  //       unless generating MIPS16 code.
2498  // 'y' : Equivalent to r; retained for
2499  //       backwards compatibility.
2500  // 'f' : Floating Point registers.
2501  if (Constraint.size() == 1) {
2502    switch (Constraint[0]) {
2503      default : break;
2504      case 'd':
2505      case 'y':
2506      case 'f':
2507        return C_RegisterClass;
2508        break;
2509    }
2510  }
2511  return TargetLowering::getConstraintType(Constraint);
2512}
2513
2514/// Examine constraint type and operand type and determine a weight value.
2515/// This object must already have been set up with the operand type
2516/// and the current alternative constraint selected.
2517TargetLowering::ConstraintWeight
2518MipsTargetLowering::getSingleConstraintMatchWeight(
2519    AsmOperandInfo &info, const char *constraint) const {
2520  ConstraintWeight weight = CW_Invalid;
2521  Value *CallOperandVal = info.CallOperandVal;
2522    // If we don't have a value, we can't do a match,
2523    // but allow it at the lowest weight.
2524  if (CallOperandVal == NULL)
2525    return CW_Default;
2526  Type *type = CallOperandVal->getType();
2527  // Look at the constraint type.
2528  switch (*constraint) {
2529  default:
2530    weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2531    break;
2532  case 'd':
2533  case 'y':
2534    if (type->isIntegerTy())
2535      weight = CW_Register;
2536    break;
2537  case 'f':
2538    if (type->isFloatTy())
2539      weight = CW_Register;
2540    break;
2541  }
2542  return weight;
2543}
2544
2545/// Given a register class constraint, like 'r', if this corresponds directly
2546/// to an LLVM register class, return a register of 0 and the register class
2547/// pointer.
2548std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2549getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
2550{
2551  if (Constraint.size() == 1) {
2552    switch (Constraint[0]) {
2553    case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2554    case 'y': // Same as 'r'. Exists for compatibility.
2555    case 'r':
2556      return std::make_pair(0U, Mips::CPURegsRegisterClass);
2557    case 'f':
2558      if (VT == MVT::f32)
2559        return std::make_pair(0U, Mips::FGR32RegisterClass);
2560      if (VT == MVT::f64)
2561        if ((!Subtarget->isSingleFloat()) && (!Subtarget->isFP64bit()))
2562          return std::make_pair(0U, Mips::AFGR64RegisterClass);
2563      break;
2564    }
2565  }
2566  return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2567}
2568
2569bool
2570MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2571  // The Mips target isn't yet aware of offsets.
2572  return false;
2573}
2574
2575bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2576  if (VT != MVT::f32 && VT != MVT::f64)
2577    return false;
2578  if (Imm.isNegZero())
2579    return false;
2580  return Imm.isZero();
2581}
2582