MipsSEISelLowering.cpp revision 3d60241c3e86973be281660bc5971c3a46cfdc47 
1//===-- MipsSEISelLowering.cpp - MipsSE DAG Lowering Interface --*- C++ -*-===//
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// Subclass of MipsTargetLowering specialized for mips32/64.
11//
12//===----------------------------------------------------------------------===//
13#include "MipsSEISelLowering.h"
14#include "MipsRegisterInfo.h"
15#include "MipsTargetMachine.h"
16#include "llvm/CodeGen/MachineInstrBuilder.h"
17#include "llvm/CodeGen/MachineRegisterInfo.h"
18#include "llvm/Support/CommandLine.h"
19#include "llvm/Target/TargetInstrInfo.h"
20
21using namespace llvm;
22
23static cl::opt<bool>
24EnableMipsTailCalls("enable-mips-tail-calls", cl::Hidden,
25                    cl::desc("MIPS: Enable tail calls."), cl::init(false));
26
27MipsSETargetLowering::MipsSETargetLowering(MipsTargetMachine &TM)
28  : MipsTargetLowering(TM) {
29  // Set up the register classes
30
31  clearRegisterClasses();
32
33  addRegisterClass(MVT::i32, &Mips::CPURegsRegClass);
34
35  if (HasMips64)
36    addRegisterClass(MVT::i64, &Mips::CPU64RegsRegClass);
37
38  if (Subtarget->hasDSP()) {
39    MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
40
41    for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
42      addRegisterClass(VecTys[i], &Mips::DSPRegsRegClass);
43
44      // Expand all builtin opcodes.
45      for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
46        setOperationAction(Opc, VecTys[i], Expand);
47
48      setOperationAction(ISD::ADD, VecTys[i], Legal);
49      setOperationAction(ISD::SUB, VecTys[i], Legal);
50      setOperationAction(ISD::LOAD, VecTys[i], Legal);
51      setOperationAction(ISD::STORE, VecTys[i], Legal);
52      setOperationAction(ISD::BITCAST, VecTys[i], Legal);
53    }
54  }
55
56  if (Subtarget->hasDSPR2())
57    setOperationAction(ISD::MUL, MVT::v2i16, Legal);
58
59  if (!TM.Options.UseSoftFloat) {
60    addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
61
62    // When dealing with single precision only, use libcalls
63    if (!Subtarget->isSingleFloat()) {
64      if (HasMips64)
65        addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
66      else
67        addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
68    }
69  }
70
71  setOperationAction(ISD::SMUL_LOHI,          MVT::i32, Custom);
72  setOperationAction(ISD::UMUL_LOHI,          MVT::i32, Custom);
73  setOperationAction(ISD::MULHS,              MVT::i32, Custom);
74  setOperationAction(ISD::MULHU,              MVT::i32, Custom);
75
76  if (HasMips64) {
77    setOperationAction(ISD::MULHS,            MVT::i64, Custom);
78    setOperationAction(ISD::MULHU,            MVT::i64, Custom);
79    setOperationAction(ISD::MUL,              MVT::i64, Custom);
80  }
81
82  setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
83  setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
84  setOperationAction(ISD::SDIVREM, MVT::i64, Custom);
85  setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
86  setOperationAction(ISD::MEMBARRIER,         MVT::Other, Custom);
87  setOperationAction(ISD::ATOMIC_FENCE,       MVT::Other, Custom);
88  setOperationAction(ISD::LOAD,               MVT::i32, Custom);
89  setOperationAction(ISD::STORE,              MVT::i32, Custom);
90
91  setTargetDAGCombine(ISD::ADDE);
92  setTargetDAGCombine(ISD::SUBE);
93
94  computeRegisterProperties();
95}
96
97const MipsTargetLowering *
98llvm::createMipsSETargetLowering(MipsTargetMachine &TM) {
99  return new MipsSETargetLowering(TM);
100}
101
102
103bool
104MipsSETargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
105  MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
106
107  switch (SVT) {
108  case MVT::i64:
109  case MVT::i32:
110    if (Fast)
111      *Fast = true;
112    return true;
113  default:
114    return false;
115  }
116}
117
118SDValue MipsSETargetLowering::LowerOperation(SDValue Op,
119                                             SelectionDAG &DAG) const {
120  switch(Op.getOpcode()) {
121  case ISD::SMUL_LOHI: return lowerMulDiv(Op, MipsISD::Mult, true, true, DAG);
122  case ISD::UMUL_LOHI: return lowerMulDiv(Op, MipsISD::Multu, true, true, DAG);
123  case ISD::MULHS:     return lowerMulDiv(Op, MipsISD::Mult, false, true, DAG);
124  case ISD::MULHU:     return lowerMulDiv(Op, MipsISD::Multu, false, true, DAG);
125  case ISD::MUL:       return lowerMulDiv(Op, MipsISD::Mult, true, false, DAG);
126  case ISD::SDIVREM:   return lowerMulDiv(Op, MipsISD::DivRem, true, true, DAG);
127  case ISD::UDIVREM:   return lowerMulDiv(Op, MipsISD::DivRemU, true, true, DAG);
128  }
129
130  return MipsTargetLowering::LowerOperation(Op, DAG);
131}
132
133// selectMADD -
134// Transforms a subgraph in CurDAG if the following pattern is found:
135//  (addc multLo, Lo0), (adde multHi, Hi0),
136// where,
137//  multHi/Lo: product of multiplication
138//  Lo0: initial value of Lo register
139//  Hi0: initial value of Hi register
140// Return true if pattern matching was successful.
141static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
142  // ADDENode's second operand must be a flag output of an ADDC node in order
143  // for the matching to be successful.
144  SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
145
146  if (ADDCNode->getOpcode() != ISD::ADDC)
147    return false;
148
149  SDValue MultHi = ADDENode->getOperand(0);
150  SDValue MultLo = ADDCNode->getOperand(0);
151  SDNode *MultNode = MultHi.getNode();
152  unsigned MultOpc = MultHi.getOpcode();
153
154  // MultHi and MultLo must be generated by the same node,
155  if (MultLo.getNode() != MultNode)
156    return false;
157
158  // and it must be a multiplication.
159  if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
160    return false;
161
162  // MultLo amd MultHi must be the first and second output of MultNode
163  // respectively.
164  if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
165    return false;
166
167  // Transform this to a MADD only if ADDENode and ADDCNode are the only users
168  // of the values of MultNode, in which case MultNode will be removed in later
169  // phases.
170  // If there exist users other than ADDENode or ADDCNode, this function returns
171  // here, which will result in MultNode being mapped to a single MULT
172  // instruction node rather than a pair of MULT and MADD instructions being
173  // produced.
174  if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
175    return false;
176
177  DebugLoc DL = ADDENode->getDebugLoc();
178
179  // Initialize accumulator.
180  SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
181                                  ADDCNode->getOperand(1),
182                                  ADDENode->getOperand(1));
183
184  // create MipsMAdd(u) node
185  MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
186
187  SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
188                                 MultNode->getOperand(0),// Factor 0
189                                 MultNode->getOperand(1),// Factor 1
190                                 ACCIn);
191
192  // replace uses of adde and addc here
193  if (!SDValue(ADDCNode, 0).use_empty()) {
194    SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
195    SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
196                                    LoIdx);
197    CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
198  }
199  if (!SDValue(ADDENode, 0).use_empty()) {
200    SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
201    SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
202                                    HiIdx);
203    CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
204  }
205
206  return true;
207}
208
209// selectMSUB -
210// Transforms a subgraph in CurDAG if the following pattern is found:
211//  (addc Lo0, multLo), (sube Hi0, multHi),
212// where,
213//  multHi/Lo: product of multiplication
214//  Lo0: initial value of Lo register
215//  Hi0: initial value of Hi register
216// Return true if pattern matching was successful.
217static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
218  // SUBENode's second operand must be a flag output of an SUBC node in order
219  // for the matching to be successful.
220  SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
221
222  if (SUBCNode->getOpcode() != ISD::SUBC)
223    return false;
224
225  SDValue MultHi = SUBENode->getOperand(1);
226  SDValue MultLo = SUBCNode->getOperand(1);
227  SDNode *MultNode = MultHi.getNode();
228  unsigned MultOpc = MultHi.getOpcode();
229
230  // MultHi and MultLo must be generated by the same node,
231  if (MultLo.getNode() != MultNode)
232    return false;
233
234  // and it must be a multiplication.
235  if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
236    return false;
237
238  // MultLo amd MultHi must be the first and second output of MultNode
239  // respectively.
240  if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
241    return false;
242
243  // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
244  // of the values of MultNode, in which case MultNode will be removed in later
245  // phases.
246  // If there exist users other than SUBENode or SUBCNode, this function returns
247  // here, which will result in MultNode being mapped to a single MULT
248  // instruction node rather than a pair of MULT and MSUB instructions being
249  // produced.
250  if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
251    return false;
252
253  DebugLoc DL = SUBENode->getDebugLoc();
254
255  // Initialize accumulator.
256  SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
257                                  SUBCNode->getOperand(0),
258                                  SUBENode->getOperand(0));
259
260  // create MipsSub(u) node
261  MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
262
263  SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
264                                 MultNode->getOperand(0),// Factor 0
265                                 MultNode->getOperand(1),// Factor 1
266                                 ACCIn);
267
268  // replace uses of sube and subc here
269  if (!SDValue(SUBCNode, 0).use_empty()) {
270    SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
271    SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
272                                    LoIdx);
273    CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
274  }
275  if (!SDValue(SUBENode, 0).use_empty()) {
276    SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
277    SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
278                                    HiIdx);
279    CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
280  }
281
282  return true;
283}
284
285static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
286                                  TargetLowering::DAGCombinerInfo &DCI,
287                                  const MipsSubtarget *Subtarget) {
288  if (DCI.isBeforeLegalize())
289    return SDValue();
290
291  if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
292      selectMADD(N, &DAG))
293    return SDValue(N, 0);
294
295  return SDValue();
296}
297
298static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
299                                  TargetLowering::DAGCombinerInfo &DCI,
300                                  const MipsSubtarget *Subtarget) {
301  if (DCI.isBeforeLegalize())
302    return SDValue();
303
304  if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
305      selectMSUB(N, &DAG))
306    return SDValue(N, 0);
307
308  return SDValue();
309}
310
311SDValue
312MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
313  SelectionDAG &DAG = DCI.DAG;
314
315  switch (N->getOpcode()) {
316  case ISD::ADDE:
317    return performADDECombine(N, DAG, DCI, Subtarget);
318  case ISD::SUBE:
319    return performSUBECombine(N, DAG, DCI, Subtarget);
320  default:
321    return MipsTargetLowering::PerformDAGCombine(N, DCI);
322  }
323}
324
325MachineBasicBlock *
326MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
327                                                  MachineBasicBlock *BB) const {
328  switch (MI->getOpcode()) {
329  default:
330    return MipsTargetLowering::EmitInstrWithCustomInserter(MI, BB);
331  case Mips::BPOSGE32_PSEUDO:
332    return emitBPOSGE32(MI, BB);
333  }
334}
335
336bool MipsSETargetLowering::
337isEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
338                                  unsigned NextStackOffset,
339                                  const MipsFunctionInfo& FI) const {
340  if (!EnableMipsTailCalls)
341    return false;
342
343  // Return false if either the callee or caller has a byval argument.
344  if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
345    return false;
346
347  // Return true if the callee's argument area is no larger than the
348  // caller's.
349  return NextStackOffset <= FI.getIncomingArgSize();
350}
351
352void MipsSETargetLowering::
353getOpndList(SmallVectorImpl<SDValue> &Ops,
354            std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
355            bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
356            CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
357  // T9 should contain the address of the callee function if
358  // -reloction-model=pic or it is an indirect call.
359  if (IsPICCall || !GlobalOrExternal) {
360    unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
361    RegsToPass.push_front(std::make_pair(T9Reg, Callee));
362  } else
363    Ops.push_back(Callee);
364
365  MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal,
366                                  InternalLinkage, CLI, Callee, Chain);
367}
368
369SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
370                                          bool HasLo, bool HasHi,
371                                          SelectionDAG &DAG) const {
372  EVT Ty = Op.getOperand(0).getValueType();
373  DebugLoc DL = Op.getDebugLoc();
374  SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
375                             Op.getOperand(0), Op.getOperand(1));
376  SDValue Lo, Hi;
377
378  if (HasLo)
379    Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
380                     DAG.getConstant(Mips::sub_lo, MVT::i32));
381  if (HasHi)
382    Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, Ty, Mult,
383                     DAG.getConstant(Mips::sub_hi, MVT::i32));
384
385  if (!HasLo || !HasHi)
386    return HasLo ? Lo : Hi;
387
388  SDValue Vals[] = { Lo, Hi };
389  return DAG.getMergeValues(Vals, 2, DL);
390}
391
392MachineBasicBlock * MipsSETargetLowering::
393emitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
394  // $bb:
395  //  bposge32_pseudo $vr0
396  //  =>
397  // $bb:
398  //  bposge32 $tbb
399  // $fbb:
400  //  li $vr2, 0
401  //  b $sink
402  // $tbb:
403  //  li $vr1, 1
404  // $sink:
405  //  $vr0 = phi($vr2, $fbb, $vr1, $tbb)
406
407  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
408  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
409  const TargetRegisterClass *RC = &Mips::CPURegsRegClass;
410  DebugLoc DL = MI->getDebugLoc();
411  const BasicBlock *LLVM_BB = BB->getBasicBlock();
412  MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
413  MachineFunction *F = BB->getParent();
414  MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
415  MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
416  MachineBasicBlock *Sink  = F->CreateMachineBasicBlock(LLVM_BB);
417  F->insert(It, FBB);
418  F->insert(It, TBB);
419  F->insert(It, Sink);
420
421  // Transfer the remainder of BB and its successor edges to Sink.
422  Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
423               BB->end());
424  Sink->transferSuccessorsAndUpdatePHIs(BB);
425
426  // Add successors.
427  BB->addSuccessor(FBB);
428  BB->addSuccessor(TBB);
429  FBB->addSuccessor(Sink);
430  TBB->addSuccessor(Sink);
431
432  // Insert the real bposge32 instruction to $BB.
433  BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
434
435  // Fill $FBB.
436  unsigned VR2 = RegInfo.createVirtualRegister(RC);
437  BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
438    .addReg(Mips::ZERO).addImm(0);
439  BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
440
441  // Fill $TBB.
442  unsigned VR1 = RegInfo.createVirtualRegister(RC);
443  BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
444    .addReg(Mips::ZERO).addImm(1);
445
446  // Insert phi function to $Sink.
447  BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
448          MI->getOperand(0).getReg())
449    .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
450
451  MI->eraseFromParent();   // The pseudo instruction is gone now.
452  return Sink;
453}
454