1 //===- MipsISelLowering.cpp - Mips DAG Lowering Implementation ------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that Mips uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "MipsISelLowering.h"
15 #include "MCTargetDesc/MipsBaseInfo.h"
16 #include "MCTargetDesc/MipsInstPrinter.h"
17 #include "MCTargetDesc/MipsMCTargetDesc.h"
18 #include "MipsCCState.h"
19 #include "MipsInstrInfo.h"
20 #include "MipsMachineFunction.h"
21 #include "MipsRegisterInfo.h"
22 #include "MipsSubtarget.h"
23 #include "MipsTargetMachine.h"
24 #include "MipsTargetObjectFile.h"
25 #include "llvm/ADT/APFloat.h"
26 #include "llvm/ADT/ArrayRef.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/StringRef.h"
30 #include "llvm/ADT/StringSwitch.h"
31 #include "llvm/CodeGen/CallingConvLower.h"
32 #include "llvm/CodeGen/FunctionLoweringInfo.h"
33 #include "llvm/CodeGen/ISDOpcodes.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineInstr.h"
38 #include "llvm/CodeGen/MachineInstrBuilder.h"
39 #include "llvm/CodeGen/MachineJumpTableInfo.h"
40 #include "llvm/CodeGen/MachineMemOperand.h"
41 #include "llvm/CodeGen/MachineOperand.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/RuntimeLibcalls.h"
44 #include "llvm/CodeGen/SelectionDAG.h"
45 #include "llvm/CodeGen/SelectionDAGNodes.h"
46 #include "llvm/CodeGen/TargetFrameLowering.h"
47 #include "llvm/CodeGen/TargetInstrInfo.h"
48 #include "llvm/CodeGen/TargetRegisterInfo.h"
49 #include "llvm/CodeGen/ValueTypes.h"
50 #include "llvm/IR/CallingConv.h"
51 #include "llvm/IR/Constants.h"
52 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/DebugLoc.h"
54 #include "llvm/IR/DerivedTypes.h"
55 #include "llvm/IR/Function.h"
56 #include "llvm/IR/GlobalValue.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/Value.h"
59 #include "llvm/MC/MCContext.h"
60 #include "llvm/MC/MCRegisterInfo.h"
61 #include "llvm/Support/Casting.h"
62 #include "llvm/Support/CodeGen.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/Compiler.h"
65 #include "llvm/Support/ErrorHandling.h"
66 #include "llvm/Support/MachineValueType.h"
67 #include "llvm/Support/MathExtras.h"
68 #include "llvm/Target/TargetMachine.h"
69 #include "llvm/Target/TargetOptions.h"
70 #include <algorithm>
71 #include <cassert>
72 #include <cctype>
73 #include <cstdint>
74 #include <deque>
75 #include <iterator>
76 #include <utility>
77 #include <vector>
78 
79 using namespace llvm;
80 
81 #define DEBUG_TYPE "mips-lower"
82 
83 STATISTIC(NumTailCalls, "Number of tail calls");
84 
85 static cl::opt<bool>
86 NoZeroDivCheck("mno-check-zero-division", cl::Hidden,
87                cl::desc("MIPS: Don't trap on integer division by zero."),
88                cl::init(false));
89 
90 extern cl::opt<bool> EmitJalrReloc;
91 
92 static const MCPhysReg Mips64DPRegs[8] = {
93   Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
94   Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
95 };
96 
97 // If I is a shifted mask, set the size (Size) and the first bit of the
98 // mask (Pos), and return true.
99 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
100 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
101   if (!isShiftedMask_64(I))
102     return false;
103 
104   Size = countPopulation(I);
105   Pos = countTrailingZeros(I);
106   return true;
107 }
108 
109 // The MIPS MSA ABI passes vector arguments in the integer register set.
110 // The number of integer registers used is dependant on the ABI used.
111 MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
112                                                       CallingConv::ID CC,
113                                                       EVT VT) const {
114   if (!VT.isVector())
115     return getRegisterType(Context, VT);
116 
117   return Subtarget.isABI_O32() || VT.getSizeInBits() == 32 ? MVT::i32
118                                                            : MVT::i64;
119 }
120 
121 unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
122                                                            CallingConv::ID CC,
123                                                            EVT VT) const {
124   if (VT.isVector())
125     return std::max(((unsigned)VT.getSizeInBits() /
126                      (Subtarget.isABI_O32() ? 32 : 64)),
127                     1U);
128   return MipsTargetLowering::getNumRegisters(Context, VT);
129 }
130 
131 unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv(
132     LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
133     unsigned &NumIntermediates, MVT &RegisterVT) const {
134   // Break down vector types to either 2 i64s or 4 i32s.
135   RegisterVT = getRegisterTypeForCallingConv(Context, CC, VT);
136   IntermediateVT = RegisterVT;
137   NumIntermediates = VT.getSizeInBits() < RegisterVT.getSizeInBits()
138                          ? VT.getVectorNumElements()
139                          : VT.getSizeInBits() / RegisterVT.getSizeInBits();
140 
141   return NumIntermediates;
142 }
143 
144 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
145   MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
146   return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
147 }
148 
149 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
150                                           SelectionDAG &DAG,
151                                           unsigned Flag) const {
152   return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag);
153 }
154 
155 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty,
156                                           SelectionDAG &DAG,
157                                           unsigned Flag) const {
158   return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
159 }
160 
161 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty,
162                                           SelectionDAG &DAG,
163                                           unsigned Flag) const {
164   return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
165 }
166 
167 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty,
168                                           SelectionDAG &DAG,
169                                           unsigned Flag) const {
170   return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
171 }
172 
173 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty,
174                                           SelectionDAG &DAG,
175                                           unsigned Flag) const {
176   return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(),
177                                    N->getOffset(), Flag);
178 }
179 
180 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
181   switch ((MipsISD::NodeType)Opcode) {
182   case MipsISD::FIRST_NUMBER:      break;
183   case MipsISD::JmpLink:           return "MipsISD::JmpLink";
184   case MipsISD::TailCall:          return "MipsISD::TailCall";
185   case MipsISD::Highest:           return "MipsISD::Highest";
186   case MipsISD::Higher:            return "MipsISD::Higher";
187   case MipsISD::Hi:                return "MipsISD::Hi";
188   case MipsISD::Lo:                return "MipsISD::Lo";
189   case MipsISD::GotHi:             return "MipsISD::GotHi";
190   case MipsISD::TlsHi:             return "MipsISD::TlsHi";
191   case MipsISD::GPRel:             return "MipsISD::GPRel";
192   case MipsISD::ThreadPointer:     return "MipsISD::ThreadPointer";
193   case MipsISD::Ret:               return "MipsISD::Ret";
194   case MipsISD::ERet:              return "MipsISD::ERet";
195   case MipsISD::EH_RETURN:         return "MipsISD::EH_RETURN";
196   case MipsISD::FMS:               return "MipsISD::FMS";
197   case MipsISD::FPBrcond:          return "MipsISD::FPBrcond";
198   case MipsISD::FPCmp:             return "MipsISD::FPCmp";
199   case MipsISD::FSELECT:           return "MipsISD::FSELECT";
200   case MipsISD::MTC1_D64:          return "MipsISD::MTC1_D64";
201   case MipsISD::CMovFP_T:          return "MipsISD::CMovFP_T";
202   case MipsISD::CMovFP_F:          return "MipsISD::CMovFP_F";
203   case MipsISD::TruncIntFP:        return "MipsISD::TruncIntFP";
204   case MipsISD::MFHI:              return "MipsISD::MFHI";
205   case MipsISD::MFLO:              return "MipsISD::MFLO";
206   case MipsISD::MTLOHI:            return "MipsISD::MTLOHI";
207   case MipsISD::Mult:              return "MipsISD::Mult";
208   case MipsISD::Multu:             return "MipsISD::Multu";
209   case MipsISD::MAdd:              return "MipsISD::MAdd";
210   case MipsISD::MAddu:             return "MipsISD::MAddu";
211   case MipsISD::MSub:              return "MipsISD::MSub";
212   case MipsISD::MSubu:             return "MipsISD::MSubu";
213   case MipsISD::DivRem:            return "MipsISD::DivRem";
214   case MipsISD::DivRemU:           return "MipsISD::DivRemU";
215   case MipsISD::DivRem16:          return "MipsISD::DivRem16";
216   case MipsISD::DivRemU16:         return "MipsISD::DivRemU16";
217   case MipsISD::BuildPairF64:      return "MipsISD::BuildPairF64";
218   case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
219   case MipsISD::Wrapper:           return "MipsISD::Wrapper";
220   case MipsISD::DynAlloc:          return "MipsISD::DynAlloc";
221   case MipsISD::Sync:              return "MipsISD::Sync";
222   case MipsISD::Ext:               return "MipsISD::Ext";
223   case MipsISD::Ins:               return "MipsISD::Ins";
224   case MipsISD::CIns:              return "MipsISD::CIns";
225   case MipsISD::LWL:               return "MipsISD::LWL";
226   case MipsISD::LWR:               return "MipsISD::LWR";
227   case MipsISD::SWL:               return "MipsISD::SWL";
228   case MipsISD::SWR:               return "MipsISD::SWR";
229   case MipsISD::LDL:               return "MipsISD::LDL";
230   case MipsISD::LDR:               return "MipsISD::LDR";
231   case MipsISD::SDL:               return "MipsISD::SDL";
232   case MipsISD::SDR:               return "MipsISD::SDR";
233   case MipsISD::EXTP:              return "MipsISD::EXTP";
234   case MipsISD::EXTPDP:            return "MipsISD::EXTPDP";
235   case MipsISD::EXTR_S_H:          return "MipsISD::EXTR_S_H";
236   case MipsISD::EXTR_W:            return "MipsISD::EXTR_W";
237   case MipsISD::EXTR_R_W:          return "MipsISD::EXTR_R_W";
238   case MipsISD::EXTR_RS_W:         return "MipsISD::EXTR_RS_W";
239   case MipsISD::SHILO:             return "MipsISD::SHILO";
240   case MipsISD::MTHLIP:            return "MipsISD::MTHLIP";
241   case MipsISD::MULSAQ_S_W_PH:     return "MipsISD::MULSAQ_S_W_PH";
242   case MipsISD::MAQ_S_W_PHL:       return "MipsISD::MAQ_S_W_PHL";
243   case MipsISD::MAQ_S_W_PHR:       return "MipsISD::MAQ_S_W_PHR";
244   case MipsISD::MAQ_SA_W_PHL:      return "MipsISD::MAQ_SA_W_PHL";
245   case MipsISD::MAQ_SA_W_PHR:      return "MipsISD::MAQ_SA_W_PHR";
246   case MipsISD::DPAU_H_QBL:        return "MipsISD::DPAU_H_QBL";
247   case MipsISD::DPAU_H_QBR:        return "MipsISD::DPAU_H_QBR";
248   case MipsISD::DPSU_H_QBL:        return "MipsISD::DPSU_H_QBL";
249   case MipsISD::DPSU_H_QBR:        return "MipsISD::DPSU_H_QBR";
250   case MipsISD::DPAQ_S_W_PH:       return "MipsISD::DPAQ_S_W_PH";
251   case MipsISD::DPSQ_S_W_PH:       return "MipsISD::DPSQ_S_W_PH";
252   case MipsISD::DPAQ_SA_L_W:       return "MipsISD::DPAQ_SA_L_W";
253   case MipsISD::DPSQ_SA_L_W:       return "MipsISD::DPSQ_SA_L_W";
254   case MipsISD::DPA_W_PH:          return "MipsISD::DPA_W_PH";
255   case MipsISD::DPS_W_PH:          return "MipsISD::DPS_W_PH";
256   case MipsISD::DPAQX_S_W_PH:      return "MipsISD::DPAQX_S_W_PH";
257   case MipsISD::DPAQX_SA_W_PH:     return "MipsISD::DPAQX_SA_W_PH";
258   case MipsISD::DPAX_W_PH:         return "MipsISD::DPAX_W_PH";
259   case MipsISD::DPSX_W_PH:         return "MipsISD::DPSX_W_PH";
260   case MipsISD::DPSQX_S_W_PH:      return "MipsISD::DPSQX_S_W_PH";
261   case MipsISD::DPSQX_SA_W_PH:     return "MipsISD::DPSQX_SA_W_PH";
262   case MipsISD::MULSA_W_PH:        return "MipsISD::MULSA_W_PH";
263   case MipsISD::MULT:              return "MipsISD::MULT";
264   case MipsISD::MULTU:             return "MipsISD::MULTU";
265   case MipsISD::MADD_DSP:          return "MipsISD::MADD_DSP";
266   case MipsISD::MADDU_DSP:         return "MipsISD::MADDU_DSP";
267   case MipsISD::MSUB_DSP:          return "MipsISD::MSUB_DSP";
268   case MipsISD::MSUBU_DSP:         return "MipsISD::MSUBU_DSP";
269   case MipsISD::SHLL_DSP:          return "MipsISD::SHLL_DSP";
270   case MipsISD::SHRA_DSP:          return "MipsISD::SHRA_DSP";
271   case MipsISD::SHRL_DSP:          return "MipsISD::SHRL_DSP";
272   case MipsISD::SETCC_DSP:         return "MipsISD::SETCC_DSP";
273   case MipsISD::SELECT_CC_DSP:     return "MipsISD::SELECT_CC_DSP";
274   case MipsISD::VALL_ZERO:         return "MipsISD::VALL_ZERO";
275   case MipsISD::VANY_ZERO:         return "MipsISD::VANY_ZERO";
276   case MipsISD::VALL_NONZERO:      return "MipsISD::VALL_NONZERO";
277   case MipsISD::VANY_NONZERO:      return "MipsISD::VANY_NONZERO";
278   case MipsISD::VCEQ:              return "MipsISD::VCEQ";
279   case MipsISD::VCLE_S:            return "MipsISD::VCLE_S";
280   case MipsISD::VCLE_U:            return "MipsISD::VCLE_U";
281   case MipsISD::VCLT_S:            return "MipsISD::VCLT_S";
282   case MipsISD::VCLT_U:            return "MipsISD::VCLT_U";
283   case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
284   case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
285   case MipsISD::VNOR:              return "MipsISD::VNOR";
286   case MipsISD::VSHF:              return "MipsISD::VSHF";
287   case MipsISD::SHF:               return "MipsISD::SHF";
288   case MipsISD::ILVEV:             return "MipsISD::ILVEV";
289   case MipsISD::ILVOD:             return "MipsISD::ILVOD";
290   case MipsISD::ILVL:              return "MipsISD::ILVL";
291   case MipsISD::ILVR:              return "MipsISD::ILVR";
292   case MipsISD::PCKEV:             return "MipsISD::PCKEV";
293   case MipsISD::PCKOD:             return "MipsISD::PCKOD";
294   case MipsISD::INSVE:             return "MipsISD::INSVE";
295   }
296   return nullptr;
297 }
298 
299 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM,
300                                        const MipsSubtarget &STI)
301     : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) {
302   // Mips does not have i1 type, so use i32 for
303   // setcc operations results (slt, sgt, ...).
304   setBooleanContents(ZeroOrOneBooleanContent);
305   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
306   // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA
307   // does. Integer booleans still use 0 and 1.
308   if (Subtarget.hasMips32r6())
309     setBooleanContents(ZeroOrOneBooleanContent,
310                        ZeroOrNegativeOneBooleanContent);
311 
312   // Load extented operations for i1 types must be promoted
313   for (MVT VT : MVT::integer_valuetypes()) {
314     setLoadExtAction(ISD::EXTLOAD,  VT, MVT::i1,  Promote);
315     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1,  Promote);
316     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1,  Promote);
317   }
318 
319   // MIPS doesn't have extending float->double load/store.  Set LoadExtAction
320   // for f32, f16
321   for (MVT VT : MVT::fp_valuetypes()) {
322     setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
323     setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
324   }
325 
326   // Set LoadExtAction for f16 vectors to Expand
327   for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) {
328     MVT F16VT = MVT::getVectorVT(MVT::f16, VT.getVectorNumElements());
329     if (F16VT.isValid())
330       setLoadExtAction(ISD::EXTLOAD, VT, F16VT, Expand);
331   }
332 
333   setTruncStoreAction(MVT::f32, MVT::f16, Expand);
334   setTruncStoreAction(MVT::f64, MVT::f16, Expand);
335 
336   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
337 
338   // Used by legalize types to correctly generate the setcc result.
339   // Without this, every float setcc comes with a AND/OR with the result,
340   // we don't want this, since the fpcmp result goes to a flag register,
341   // which is used implicitly by brcond and select operations.
342   AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
343 
344   // Mips Custom Operations
345   setOperationAction(ISD::BR_JT,              MVT::Other, Expand);
346   setOperationAction(ISD::GlobalAddress,      MVT::i32,   Custom);
347   setOperationAction(ISD::BlockAddress,       MVT::i32,   Custom);
348   setOperationAction(ISD::GlobalTLSAddress,   MVT::i32,   Custom);
349   setOperationAction(ISD::JumpTable,          MVT::i32,   Custom);
350   setOperationAction(ISD::ConstantPool,       MVT::i32,   Custom);
351   setOperationAction(ISD::SELECT,             MVT::f32,   Custom);
352   setOperationAction(ISD::SELECT,             MVT::f64,   Custom);
353   setOperationAction(ISD::SELECT,             MVT::i32,   Custom);
354   setOperationAction(ISD::SETCC,              MVT::f32,   Custom);
355   setOperationAction(ISD::SETCC,              MVT::f64,   Custom);
356   setOperationAction(ISD::BRCOND,             MVT::Other, Custom);
357   setOperationAction(ISD::FCOPYSIGN,          MVT::f32,   Custom);
358   setOperationAction(ISD::FCOPYSIGN,          MVT::f64,   Custom);
359   setOperationAction(ISD::FP_TO_SINT,         MVT::i32,   Custom);
360 
361   if (!(TM.Options.NoNaNsFPMath || Subtarget.inAbs2008Mode())) {
362     setOperationAction(ISD::FABS, MVT::f32, Custom);
363     setOperationAction(ISD::FABS, MVT::f64, Custom);
364   }
365 
366   if (Subtarget.isGP64bit()) {
367     setOperationAction(ISD::GlobalAddress,      MVT::i64,   Custom);
368     setOperationAction(ISD::BlockAddress,       MVT::i64,   Custom);
369     setOperationAction(ISD::GlobalTLSAddress,   MVT::i64,   Custom);
370     setOperationAction(ISD::JumpTable,          MVT::i64,   Custom);
371     setOperationAction(ISD::ConstantPool,       MVT::i64,   Custom);
372     setOperationAction(ISD::SELECT,             MVT::i64,   Custom);
373     setOperationAction(ISD::LOAD,               MVT::i64,   Custom);
374     setOperationAction(ISD::STORE,              MVT::i64,   Custom);
375     setOperationAction(ISD::FP_TO_SINT,         MVT::i64,   Custom);
376     setOperationAction(ISD::SHL_PARTS,          MVT::i64,   Custom);
377     setOperationAction(ISD::SRA_PARTS,          MVT::i64,   Custom);
378     setOperationAction(ISD::SRL_PARTS,          MVT::i64,   Custom);
379   }
380 
381   if (!Subtarget.isGP64bit()) {
382     setOperationAction(ISD::SHL_PARTS,          MVT::i32,   Custom);
383     setOperationAction(ISD::SRA_PARTS,          MVT::i32,   Custom);
384     setOperationAction(ISD::SRL_PARTS,          MVT::i32,   Custom);
385   }
386 
387   setOperationAction(ISD::EH_DWARF_CFA,         MVT::i32,   Custom);
388   if (Subtarget.isGP64bit())
389     setOperationAction(ISD::EH_DWARF_CFA,       MVT::i64,   Custom);
390 
391   setOperationAction(ISD::SDIV, MVT::i32, Expand);
392   setOperationAction(ISD::SREM, MVT::i32, Expand);
393   setOperationAction(ISD::UDIV, MVT::i32, Expand);
394   setOperationAction(ISD::UREM, MVT::i32, Expand);
395   setOperationAction(ISD::SDIV, MVT::i64, Expand);
396   setOperationAction(ISD::SREM, MVT::i64, Expand);
397   setOperationAction(ISD::UDIV, MVT::i64, Expand);
398   setOperationAction(ISD::UREM, MVT::i64, Expand);
399 
400   // Operations not directly supported by Mips.
401   setOperationAction(ISD::BR_CC,             MVT::f32,   Expand);
402   setOperationAction(ISD::BR_CC,             MVT::f64,   Expand);
403   setOperationAction(ISD::BR_CC,             MVT::i32,   Expand);
404   setOperationAction(ISD::BR_CC,             MVT::i64,   Expand);
405   setOperationAction(ISD::SELECT_CC,         MVT::i32,   Expand);
406   setOperationAction(ISD::SELECT_CC,         MVT::i64,   Expand);
407   setOperationAction(ISD::SELECT_CC,         MVT::f32,   Expand);
408   setOperationAction(ISD::SELECT_CC,         MVT::f64,   Expand);
409   setOperationAction(ISD::UINT_TO_FP,        MVT::i32,   Expand);
410   setOperationAction(ISD::UINT_TO_FP,        MVT::i64,   Expand);
411   setOperationAction(ISD::FP_TO_UINT,        MVT::i32,   Expand);
412   setOperationAction(ISD::FP_TO_UINT,        MVT::i64,   Expand);
413   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1,    Expand);
414   if (Subtarget.hasCnMips()) {
415     setOperationAction(ISD::CTPOP,           MVT::i32,   Legal);
416     setOperationAction(ISD::CTPOP,           MVT::i64,   Legal);
417   } else {
418     setOperationAction(ISD::CTPOP,           MVT::i32,   Expand);
419     setOperationAction(ISD::CTPOP,           MVT::i64,   Expand);
420   }
421   setOperationAction(ISD::CTTZ,              MVT::i32,   Expand);
422   setOperationAction(ISD::CTTZ,              MVT::i64,   Expand);
423   setOperationAction(ISD::ROTL,              MVT::i32,   Expand);
424   setOperationAction(ISD::ROTL,              MVT::i64,   Expand);
425   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32,  Expand);
426   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64,  Expand);
427 
428   if (!Subtarget.hasMips32r2())
429     setOperationAction(ISD::ROTR, MVT::i32,   Expand);
430 
431   if (!Subtarget.hasMips64r2())
432     setOperationAction(ISD::ROTR, MVT::i64,   Expand);
433 
434   setOperationAction(ISD::FSIN,              MVT::f32,   Expand);
435   setOperationAction(ISD::FSIN,              MVT::f64,   Expand);
436   setOperationAction(ISD::FCOS,              MVT::f32,   Expand);
437   setOperationAction(ISD::FCOS,              MVT::f64,   Expand);
438   setOperationAction(ISD::FSINCOS,           MVT::f32,   Expand);
439   setOperationAction(ISD::FSINCOS,           MVT::f64,   Expand);
440   setOperationAction(ISD::FPOW,              MVT::f32,   Expand);
441   setOperationAction(ISD::FPOW,              MVT::f64,   Expand);
442   setOperationAction(ISD::FLOG,              MVT::f32,   Expand);
443   setOperationAction(ISD::FLOG2,             MVT::f32,   Expand);
444   setOperationAction(ISD::FLOG10,            MVT::f32,   Expand);
445   setOperationAction(ISD::FEXP,              MVT::f32,   Expand);
446   setOperationAction(ISD::FMA,               MVT::f32,   Expand);
447   setOperationAction(ISD::FMA,               MVT::f64,   Expand);
448   setOperationAction(ISD::FREM,              MVT::f32,   Expand);
449   setOperationAction(ISD::FREM,              MVT::f64,   Expand);
450 
451   // Lower f16 conversion operations into library calls
452   setOperationAction(ISD::FP16_TO_FP,        MVT::f32,   Expand);
453   setOperationAction(ISD::FP_TO_FP16,        MVT::f32,   Expand);
454   setOperationAction(ISD::FP16_TO_FP,        MVT::f64,   Expand);
455   setOperationAction(ISD::FP_TO_FP16,        MVT::f64,   Expand);
456 
457   setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
458 
459   setOperationAction(ISD::VASTART,           MVT::Other, Custom);
460   setOperationAction(ISD::VAARG,             MVT::Other, Custom);
461   setOperationAction(ISD::VACOPY,            MVT::Other, Expand);
462   setOperationAction(ISD::VAEND,             MVT::Other, Expand);
463 
464   // Use the default for now
465   setOperationAction(ISD::STACKSAVE,         MVT::Other, Expand);
466   setOperationAction(ISD::STACKRESTORE,      MVT::Other, Expand);
467 
468   if (!Subtarget.isGP64bit()) {
469     setOperationAction(ISD::ATOMIC_LOAD,     MVT::i64,   Expand);
470     setOperationAction(ISD::ATOMIC_STORE,    MVT::i64,   Expand);
471   }
472 
473   if (!Subtarget.hasMips32r2()) {
474     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
475     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
476   }
477 
478   // MIPS16 lacks MIPS32's clz and clo instructions.
479   if (!Subtarget.hasMips32() || Subtarget.inMips16Mode())
480     setOperationAction(ISD::CTLZ, MVT::i32, Expand);
481   if (!Subtarget.hasMips64())
482     setOperationAction(ISD::CTLZ, MVT::i64, Expand);
483 
484   if (!Subtarget.hasMips32r2())
485     setOperationAction(ISD::BSWAP, MVT::i32, Expand);
486   if (!Subtarget.hasMips64r2())
487     setOperationAction(ISD::BSWAP, MVT::i64, Expand);
488 
489   if (Subtarget.isGP64bit()) {
490     setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom);
491     setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom);
492     setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom);
493     setTruncStoreAction(MVT::i64, MVT::i32, Custom);
494   }
495 
496   setOperationAction(ISD::TRAP, MVT::Other, Legal);
497 
498   setTargetDAGCombine(ISD::SDIVREM);
499   setTargetDAGCombine(ISD::UDIVREM);
500   setTargetDAGCombine(ISD::SELECT);
501   setTargetDAGCombine(ISD::AND);
502   setTargetDAGCombine(ISD::OR);
503   setTargetDAGCombine(ISD::ADD);
504   setTargetDAGCombine(ISD::SUB);
505   setTargetDAGCombine(ISD::AssertZext);
506   setTargetDAGCombine(ISD::SHL);
507 
508   if (ABI.IsO32()) {
509     // These libcalls are not available in 32-bit.
510     setLibcallName(RTLIB::SHL_I128, nullptr);
511     setLibcallName(RTLIB::SRL_I128, nullptr);
512     setLibcallName(RTLIB::SRA_I128, nullptr);
513   }
514 
515   setMinFunctionAlignment(Subtarget.isGP64bit() ? Align(8) : Align(4));
516 
517   // The arguments on the stack are defined in terms of 4-byte slots on O32
518   // and 8-byte slots on N32/N64.
519   setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? Align(8)
520                                                             : Align(4));
521 
522   setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP);
523 
524   MaxStoresPerMemcpy = 16;
525 
526   isMicroMips = Subtarget.inMicroMipsMode();
527 }
528 
529 const MipsTargetLowering *
530 MipsTargetLowering::create(const MipsTargetMachine &TM,
531                            const MipsSubtarget &STI) {
532   if (STI.inMips16Mode())
533     return createMips16TargetLowering(TM, STI);
534 
535   return createMipsSETargetLowering(TM, STI);
536 }
537 
538 // Create a fast isel object.
539 FastISel *
540 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
541                                   const TargetLibraryInfo *libInfo) const {
542   const MipsTargetMachine &TM =
543       static_cast<const MipsTargetMachine &>(funcInfo.MF->getTarget());
544 
545   // We support only the standard encoding [MIPS32,MIPS32R5] ISAs.
546   bool UseFastISel = TM.Options.EnableFastISel && Subtarget.hasMips32() &&
547                      !Subtarget.hasMips32r6() && !Subtarget.inMips16Mode() &&
548                      !Subtarget.inMicroMipsMode();
549 
550   // Disable if either of the following is true:
551   // We do not generate PIC, the ABI is not O32, XGOT is being used.
552   if (!TM.isPositionIndependent() || !TM.getABI().IsO32() ||
553       Subtarget.useXGOT())
554     UseFastISel = false;
555 
556   return UseFastISel ? Mips::createFastISel(funcInfo, libInfo) : nullptr;
557 }
558 
559 EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
560                                            EVT VT) const {
561   if (!VT.isVector())
562     return MVT::i32;
563   return VT.changeVectorElementTypeToInteger();
564 }
565 
566 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
567                                     TargetLowering::DAGCombinerInfo &DCI,
568                                     const MipsSubtarget &Subtarget) {
569   if (DCI.isBeforeLegalizeOps())
570     return SDValue();
571 
572   EVT Ty = N->getValueType(0);
573   unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
574   unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
575   unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
576                                                   MipsISD::DivRemU16;
577   SDLoc DL(N);
578 
579   SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
580                                N->getOperand(0), N->getOperand(1));
581   SDValue InChain = DAG.getEntryNode();
582   SDValue InGlue = DivRem;
583 
584   // insert MFLO
585   if (N->hasAnyUseOfValue(0)) {
586     SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
587                                             InGlue);
588     DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
589     InChain = CopyFromLo.getValue(1);
590     InGlue = CopyFromLo.getValue(2);
591   }
592 
593   // insert MFHI
594   if (N->hasAnyUseOfValue(1)) {
595     SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
596                                             HI, Ty, InGlue);
597     DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
598   }
599 
600   return SDValue();
601 }
602 
603 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
604   switch (CC) {
605   default: llvm_unreachable("Unknown fp condition code!");
606   case ISD::SETEQ:
607   case ISD::SETOEQ: return Mips::FCOND_OEQ;
608   case ISD::SETUNE: return Mips::FCOND_UNE;
609   case ISD::SETLT:
610   case ISD::SETOLT: return Mips::FCOND_OLT;
611   case ISD::SETGT:
612   case ISD::SETOGT: return Mips::FCOND_OGT;
613   case ISD::SETLE:
614   case ISD::SETOLE: return Mips::FCOND_OLE;
615   case ISD::SETGE:
616   case ISD::SETOGE: return Mips::FCOND_OGE;
617   case ISD::SETULT: return Mips::FCOND_ULT;
618   case ISD::SETULE: return Mips::FCOND_ULE;
619   case ISD::SETUGT: return Mips::FCOND_UGT;
620   case ISD::SETUGE: return Mips::FCOND_UGE;
621   case ISD::SETUO:  return Mips::FCOND_UN;
622   case ISD::SETO:   return Mips::FCOND_OR;
623   case ISD::SETNE:
624   case ISD::SETONE: return Mips::FCOND_ONE;
625   case ISD::SETUEQ: return Mips::FCOND_UEQ;
626   }
627 }
628 
629 /// This function returns true if the floating point conditional branches and
630 /// conditional moves which use condition code CC should be inverted.
631 static bool invertFPCondCodeUser(Mips::CondCode CC) {
632   if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
633     return false;
634 
635   assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
636          "Illegal Condition Code");
637 
638   return true;
639 }
640 
641 // Creates and returns an FPCmp node from a setcc node.
642 // Returns Op if setcc is not a floating point comparison.
643 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
644   // must be a SETCC node
645   if (Op.getOpcode() != ISD::SETCC)
646     return Op;
647 
648   SDValue LHS = Op.getOperand(0);
649 
650   if (!LHS.getValueType().isFloatingPoint())
651     return Op;
652 
653   SDValue RHS = Op.getOperand(1);
654   SDLoc DL(Op);
655 
656   // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
657   // node if necessary.
658   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
659 
660   return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
661                      DAG.getConstant(condCodeToFCC(CC), DL, MVT::i32));
662 }
663 
664 // Creates and returns a CMovFPT/F node.
665 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
666                             SDValue False, const SDLoc &DL) {
667   ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
668   bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
669   SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
670 
671   return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
672                      True.getValueType(), True, FCC0, False, Cond);
673 }
674 
675 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
676                                     TargetLowering::DAGCombinerInfo &DCI,
677                                     const MipsSubtarget &Subtarget) {
678   if (DCI.isBeforeLegalizeOps())
679     return SDValue();
680 
681   SDValue SetCC = N->getOperand(0);
682 
683   if ((SetCC.getOpcode() != ISD::SETCC) ||
684       !SetCC.getOperand(0).getValueType().isInteger())
685     return SDValue();
686 
687   SDValue False = N->getOperand(2);
688   EVT FalseTy = False.getValueType();
689 
690   if (!FalseTy.isInteger())
691     return SDValue();
692 
693   ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False);
694 
695   // If the RHS (False) is 0, we swap the order of the operands
696   // of ISD::SELECT (obviously also inverting the condition) so that we can
697   // take advantage of conditional moves using the $0 register.
698   // Example:
699   //   return (a != 0) ? x : 0;
700   //     load $reg, x
701   //     movz $reg, $0, a
702   if (!FalseC)
703     return SDValue();
704 
705   const SDLoc DL(N);
706 
707   if (!FalseC->getZExtValue()) {
708     ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
709     SDValue True = N->getOperand(1);
710 
711     SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
712                          SetCC.getOperand(1),
713                          ISD::getSetCCInverse(CC, SetCC.getValueType()));
714 
715     return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
716   }
717 
718   // If both operands are integer constants there's a possibility that we
719   // can do some interesting optimizations.
720   SDValue True = N->getOperand(1);
721   ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True);
722 
723   if (!TrueC || !True.getValueType().isInteger())
724     return SDValue();
725 
726   // We'll also ignore MVT::i64 operands as this optimizations proves
727   // to be ineffective because of the required sign extensions as the result
728   // of a SETCC operator is always MVT::i32 for non-vector types.
729   if (True.getValueType() == MVT::i64)
730     return SDValue();
731 
732   int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue();
733 
734   // 1)  (a < x) ? y : y-1
735   //  slti $reg1, a, x
736   //  addiu $reg2, $reg1, y-1
737   if (Diff == 1)
738     return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False);
739 
740   // 2)  (a < x) ? y-1 : y
741   //  slti $reg1, a, x
742   //  xor $reg1, $reg1, 1
743   //  addiu $reg2, $reg1, y-1
744   if (Diff == -1) {
745     ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
746     SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
747                          SetCC.getOperand(1),
748                          ISD::getSetCCInverse(CC, SetCC.getValueType()));
749     return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True);
750   }
751 
752   // Could not optimize.
753   return SDValue();
754 }
755 
756 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG,
757                                     TargetLowering::DAGCombinerInfo &DCI,
758                                     const MipsSubtarget &Subtarget) {
759   if (DCI.isBeforeLegalizeOps())
760     return SDValue();
761 
762   SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2);
763 
764   ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse);
765   if (!FalseC || FalseC->getZExtValue())
766     return SDValue();
767 
768   // Since RHS (False) is 0, we swap the order of the True/False operands
769   // (obviously also inverting the condition) so that we can
770   // take advantage of conditional moves using the $0 register.
771   // Example:
772   //   return (a != 0) ? x : 0;
773   //     load $reg, x
774   //     movz $reg, $0, a
775   unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F :
776                                                          MipsISD::CMovFP_T;
777 
778   SDValue FCC = N->getOperand(1), Glue = N->getOperand(3);
779   return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(),
780                      ValueIfFalse, FCC, ValueIfTrue, Glue);
781 }
782 
783 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
784                                  TargetLowering::DAGCombinerInfo &DCI,
785                                  const MipsSubtarget &Subtarget) {
786   if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
787     return SDValue();
788 
789   SDValue FirstOperand = N->getOperand(0);
790   unsigned FirstOperandOpc = FirstOperand.getOpcode();
791   SDValue Mask = N->getOperand(1);
792   EVT ValTy = N->getValueType(0);
793   SDLoc DL(N);
794 
795   uint64_t Pos = 0, SMPos, SMSize;
796   ConstantSDNode *CN;
797   SDValue NewOperand;
798   unsigned Opc;
799 
800   // Op's second operand must be a shifted mask.
801   if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
802       !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
803     return SDValue();
804 
805   if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) {
806     // Pattern match EXT.
807     //  $dst = and ((sra or srl) $src , pos), (2**size - 1)
808     //  => ext $dst, $src, pos, size
809 
810     // The second operand of the shift must be an immediate.
811     if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))))
812       return SDValue();
813 
814     Pos = CN->getZExtValue();
815 
816     // Return if the shifted mask does not start at bit 0 or the sum of its size
817     // and Pos exceeds the word's size.
818     if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
819       return SDValue();
820 
821     Opc = MipsISD::Ext;
822     NewOperand = FirstOperand.getOperand(0);
823   } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) {
824     // Pattern match CINS.
825     //  $dst = and (shl $src , pos), mask
826     //  => cins $dst, $src, pos, size
827     // mask is a shifted mask with consecutive 1's, pos = shift amount,
828     // size = population count.
829 
830     // The second operand of the shift must be an immediate.
831     if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))))
832       return SDValue();
833 
834     Pos = CN->getZExtValue();
835 
836     if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 ||
837         Pos + SMSize > ValTy.getSizeInBits())
838       return SDValue();
839 
840     NewOperand = FirstOperand.getOperand(0);
841     // SMSize is 'location' (position) in this case, not size.
842     SMSize--;
843     Opc = MipsISD::CIns;
844   } else {
845     // Pattern match EXT.
846     //  $dst = and $src, (2**size - 1) , if size > 16
847     //  => ext $dst, $src, pos, size , pos = 0
848 
849     // If the mask is <= 0xffff, andi can be used instead.
850     if (CN->getZExtValue() <= 0xffff)
851       return SDValue();
852 
853     // Return if the mask doesn't start at position 0.
854     if (SMPos)
855       return SDValue();
856 
857     Opc = MipsISD::Ext;
858     NewOperand = FirstOperand;
859   }
860   return DAG.getNode(Opc, DL, ValTy, NewOperand,
861                      DAG.getConstant(Pos, DL, MVT::i32),
862                      DAG.getConstant(SMSize, DL, MVT::i32));
863 }
864 
865 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
866                                 TargetLowering::DAGCombinerInfo &DCI,
867                                 const MipsSubtarget &Subtarget) {
868   // Pattern match INS.
869   //  $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
870   //  where mask1 = (2**size - 1) << pos, mask0 = ~mask1
871   //  => ins $dst, $src, size, pos, $src1
872   if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
873     return SDValue();
874 
875   SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
876   uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
877   ConstantSDNode *CN, *CN1;
878 
879   // See if Op's first operand matches (and $src1 , mask0).
880   if (And0.getOpcode() != ISD::AND)
881     return SDValue();
882 
883   if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
884       !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
885     return SDValue();
886 
887   // See if Op's second operand matches (and (shl $src, pos), mask1).
888   if (And1.getOpcode() == ISD::AND &&
889       And1.getOperand(0).getOpcode() == ISD::SHL) {
890 
891     if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
892         !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
893       return SDValue();
894 
895     // The shift masks must have the same position and size.
896     if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
897       return SDValue();
898 
899     SDValue Shl = And1.getOperand(0);
900 
901     if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
902       return SDValue();
903 
904     unsigned Shamt = CN->getZExtValue();
905 
906     // Return if the shift amount and the first bit position of mask are not the
907     // same.
908     EVT ValTy = N->getValueType(0);
909     if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
910       return SDValue();
911 
912     SDLoc DL(N);
913     return DAG.getNode(MipsISD::Ins, DL, ValTy, Shl.getOperand(0),
914                        DAG.getConstant(SMPos0, DL, MVT::i32),
915                        DAG.getConstant(SMSize0, DL, MVT::i32),
916                        And0.getOperand(0));
917   } else {
918     // Pattern match DINS.
919     //  $dst = or (and $src, mask0), mask1
920     //  where mask0 = ((1 << SMSize0) -1) << SMPos0
921     //  => dins $dst, $src, pos, size
922     if (~CN->getSExtValue() == ((((int64_t)1 << SMSize0) - 1) << SMPos0) &&
923         ((SMSize0 + SMPos0 <= 64 && Subtarget.hasMips64r2()) ||
924          (SMSize0 + SMPos0 <= 32))) {
925       // Check if AND instruction has constant as argument
926       bool isConstCase = And1.getOpcode() != ISD::AND;
927       if (And1.getOpcode() == ISD::AND) {
928         if (!(CN1 = dyn_cast<ConstantSDNode>(And1->getOperand(1))))
929           return SDValue();
930       } else {
931         if (!(CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1))))
932           return SDValue();
933       }
934       // Don't generate INS if constant OR operand doesn't fit into bits
935       // cleared by constant AND operand.
936       if (CN->getSExtValue() & CN1->getSExtValue())
937         return SDValue();
938 
939       SDLoc DL(N);
940       EVT ValTy = N->getOperand(0)->getValueType(0);
941       SDValue Const1;
942       SDValue SrlX;
943       if (!isConstCase) {
944         Const1 = DAG.getConstant(SMPos0, DL, MVT::i32);
945         SrlX = DAG.getNode(ISD::SRL, DL, And1->getValueType(0), And1, Const1);
946       }
947       return DAG.getNode(
948           MipsISD::Ins, DL, N->getValueType(0),
949           isConstCase
950               ? DAG.getConstant(CN1->getSExtValue() >> SMPos0, DL, ValTy)
951               : SrlX,
952           DAG.getConstant(SMPos0, DL, MVT::i32),
953           DAG.getConstant(ValTy.getSizeInBits() / 8 < 8 ? SMSize0 & 31
954                                                         : SMSize0,
955                           DL, MVT::i32),
956           And0->getOperand(0));
957 
958     }
959     return SDValue();
960   }
961 }
962 
963 static SDValue performMADD_MSUBCombine(SDNode *ROOTNode, SelectionDAG &CurDAG,
964                                        const MipsSubtarget &Subtarget) {
965   // ROOTNode must have a multiplication as an operand for the match to be
966   // successful.
967   if (ROOTNode->getOperand(0).getOpcode() != ISD::MUL &&
968       ROOTNode->getOperand(1).getOpcode() != ISD::MUL)
969     return SDValue();
970 
971   // We don't handle vector types here.
972   if (ROOTNode->getValueType(0).isVector())
973     return SDValue();
974 
975   // For MIPS64, madd / msub instructions are inefficent to use with 64 bit
976   // arithmetic. E.g.
977   // (add (mul a b) c) =>
978   //   let res = (madd (mthi (drotr c 32))x(mtlo c) a b) in
979   //   MIPS64:   (or (dsll (mfhi res) 32) (dsrl (dsll (mflo res) 32) 32)
980   //   or
981   //   MIPS64R2: (dins (mflo res) (mfhi res) 32 32)
982   //
983   // The overhead of setting up the Hi/Lo registers and reassembling the
984   // result makes this a dubious optimzation for MIPS64. The core of the
985   // problem is that Hi/Lo contain the upper and lower 32 bits of the
986   // operand and result.
987   //
988   // It requires a chain of 4 add/mul for MIPS64R2 to get better code
989   // density than doing it naively, 5 for MIPS64. Additionally, using
990   // madd/msub on MIPS64 requires the operands actually be 32 bit sign
991   // extended operands, not true 64 bit values.
992   //
993   // FIXME: For the moment, disable this completely for MIPS64.
994   if (Subtarget.hasMips64())
995     return SDValue();
996 
997   SDValue Mult = ROOTNode->getOperand(0).getOpcode() == ISD::MUL
998                      ? ROOTNode->getOperand(0)
999                      : ROOTNode->getOperand(1);
1000 
1001   SDValue AddOperand = ROOTNode->getOperand(0).getOpcode() == ISD::MUL
1002                      ? ROOTNode->getOperand(1)
1003                      : ROOTNode->getOperand(0);
1004 
1005   // Transform this to a MADD only if the user of this node is the add.
1006   // If there are other users of the mul, this function returns here.
1007   if (!Mult.hasOneUse())
1008     return SDValue();
1009 
1010   // maddu and madd are unusual instructions in that on MIPS64 bits 63..31
1011   // must be in canonical form, i.e. sign extended. For MIPS32, the operands
1012   // of the multiply must have 32 or more sign bits, otherwise we cannot
1013   // perform this optimization. We have to check this here as we're performing
1014   // this optimization pre-legalization.
1015   SDValue MultLHS = Mult->getOperand(0);
1016   SDValue MultRHS = Mult->getOperand(1);
1017 
1018   bool IsSigned = MultLHS->getOpcode() == ISD::SIGN_EXTEND &&
1019                   MultRHS->getOpcode() == ISD::SIGN_EXTEND;
1020   bool IsUnsigned = MultLHS->getOpcode() == ISD::ZERO_EXTEND &&
1021                     MultRHS->getOpcode() == ISD::ZERO_EXTEND;
1022 
1023   if (!IsSigned && !IsUnsigned)
1024     return SDValue();
1025 
1026   // Initialize accumulator.
1027   SDLoc DL(ROOTNode);
1028   SDValue TopHalf;
1029   SDValue BottomHalf;
1030   BottomHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand,
1031                               CurDAG.getIntPtrConstant(0, DL));
1032 
1033   TopHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand,
1034                            CurDAG.getIntPtrConstant(1, DL));
1035   SDValue ACCIn = CurDAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
1036                                   BottomHalf,
1037                                   TopHalf);
1038 
1039   // Create MipsMAdd(u) / MipsMSub(u) node.
1040   bool IsAdd = ROOTNode->getOpcode() == ISD::ADD;
1041   unsigned Opcode = IsAdd ? (IsUnsigned ? MipsISD::MAddu : MipsISD::MAdd)
1042                           : (IsUnsigned ? MipsISD::MSubu : MipsISD::MSub);
1043   SDValue MAddOps[3] = {
1044       CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(0)),
1045       CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(1)), ACCIn};
1046   EVT VTs[2] = {MVT::i32, MVT::i32};
1047   SDValue MAdd = CurDAG.getNode(Opcode, DL, VTs, MAddOps);
1048 
1049   SDValue ResLo = CurDAG.getNode(MipsISD::MFLO, DL, MVT::i32, MAdd);
1050   SDValue ResHi = CurDAG.getNode(MipsISD::MFHI, DL, MVT::i32, MAdd);
1051   SDValue Combined =
1052       CurDAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResLo, ResHi);
1053   return Combined;
1054 }
1055 
1056 static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG,
1057                                  TargetLowering::DAGCombinerInfo &DCI,
1058                                  const MipsSubtarget &Subtarget) {
1059   // (sub v0 (mul v1, v2)) => (msub v1, v2, v0)
1060   if (DCI.isBeforeLegalizeOps()) {
1061     if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
1062         !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64)
1063       return performMADD_MSUBCombine(N, DAG, Subtarget);
1064 
1065     return SDValue();
1066   }
1067 
1068   return SDValue();
1069 }
1070 
1071 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
1072                                  TargetLowering::DAGCombinerInfo &DCI,
1073                                  const MipsSubtarget &Subtarget) {
1074   // (add v0 (mul v1, v2)) => (madd v1, v2, v0)
1075   if (DCI.isBeforeLegalizeOps()) {
1076     if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
1077         !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64)
1078       return performMADD_MSUBCombine(N, DAG, Subtarget);
1079 
1080     return SDValue();
1081   }
1082 
1083   // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
1084   SDValue Add = N->getOperand(1);
1085 
1086   if (Add.getOpcode() != ISD::ADD)
1087     return SDValue();
1088 
1089   SDValue Lo = Add.getOperand(1);
1090 
1091   if ((Lo.getOpcode() != MipsISD::Lo) ||
1092       (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
1093     return SDValue();
1094 
1095   EVT ValTy = N->getValueType(0);
1096   SDLoc DL(N);
1097 
1098   SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
1099                              Add.getOperand(0));
1100   return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
1101 }
1102 
1103 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
1104                                  TargetLowering::DAGCombinerInfo &DCI,
1105                                  const MipsSubtarget &Subtarget) {
1106   // Pattern match CINS.
1107   //  $dst = shl (and $src , imm), pos
1108   //  => cins $dst, $src, pos, size
1109 
1110   if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips())
1111     return SDValue();
1112 
1113   SDValue FirstOperand = N->getOperand(0);
1114   unsigned FirstOperandOpc = FirstOperand.getOpcode();
1115   SDValue SecondOperand = N->getOperand(1);
1116   EVT ValTy = N->getValueType(0);
1117   SDLoc DL(N);
1118 
1119   uint64_t Pos = 0, SMPos, SMSize;
1120   ConstantSDNode *CN;
1121   SDValue NewOperand;
1122 
1123   // The second operand of the shift must be an immediate.
1124   if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand)))
1125     return SDValue();
1126 
1127   Pos = CN->getZExtValue();
1128 
1129   if (Pos >= ValTy.getSizeInBits())
1130     return SDValue();
1131 
1132   if (FirstOperandOpc != ISD::AND)
1133     return SDValue();
1134 
1135   // AND's second operand must be a shifted mask.
1136   if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) ||
1137       !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
1138     return SDValue();
1139 
1140   // Return if the shifted mask does not start at bit 0 or the sum of its size
1141   // and Pos exceeds the word's size.
1142   if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits())
1143     return SDValue();
1144 
1145   NewOperand = FirstOperand.getOperand(0);
1146   // SMSize is 'location' (position) in this case, not size.
1147   SMSize--;
1148 
1149   return DAG.getNode(MipsISD::CIns, DL, ValTy, NewOperand,
1150                      DAG.getConstant(Pos, DL, MVT::i32),
1151                      DAG.getConstant(SMSize, DL, MVT::i32));
1152 }
1153 
1154 SDValue  MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
1155   const {
1156   SelectionDAG &DAG = DCI.DAG;
1157   unsigned Opc = N->getOpcode();
1158 
1159   switch (Opc) {
1160   default: break;
1161   case ISD::SDIVREM:
1162   case ISD::UDIVREM:
1163     return performDivRemCombine(N, DAG, DCI, Subtarget);
1164   case ISD::SELECT:
1165     return performSELECTCombine(N, DAG, DCI, Subtarget);
1166   case MipsISD::CMovFP_F:
1167   case MipsISD::CMovFP_T:
1168     return performCMovFPCombine(N, DAG, DCI, Subtarget);
1169   case ISD::AND:
1170     return performANDCombine(N, DAG, DCI, Subtarget);
1171   case ISD::OR:
1172     return performORCombine(N, DAG, DCI, Subtarget);
1173   case ISD::ADD:
1174     return performADDCombine(N, DAG, DCI, Subtarget);
1175   case ISD::SHL:
1176     return performSHLCombine(N, DAG, DCI, Subtarget);
1177   case ISD::SUB:
1178     return performSUBCombine(N, DAG, DCI, Subtarget);
1179   }
1180 
1181   return SDValue();
1182 }
1183 
1184 bool MipsTargetLowering::isCheapToSpeculateCttz() const {
1185   return Subtarget.hasMips32();
1186 }
1187 
1188 bool MipsTargetLowering::isCheapToSpeculateCtlz() const {
1189   return Subtarget.hasMips32();
1190 }
1191 
1192 bool MipsTargetLowering::shouldFoldConstantShiftPairToMask(
1193     const SDNode *N, CombineLevel Level) const {
1194   if (N->getOperand(0).getValueType().isVector())
1195     return false;
1196   return true;
1197 }
1198 
1199 void
1200 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
1201                                           SmallVectorImpl<SDValue> &Results,
1202                                           SelectionDAG &DAG) const {
1203   SDValue Res = LowerOperation(SDValue(N, 0), DAG);
1204 
1205   if (Res)
1206     for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
1207       Results.push_back(Res.getValue(I));
1208 }
1209 
1210 void
1211 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
1212                                        SmallVectorImpl<SDValue> &Results,
1213                                        SelectionDAG &DAG) const {
1214   return LowerOperationWrapper(N, Results, DAG);
1215 }
1216 
1217 SDValue MipsTargetLowering::
1218 LowerOperation(SDValue Op, SelectionDAG &DAG) const
1219 {
1220   switch (Op.getOpcode())
1221   {
1222   case ISD::BRCOND:             return lowerBRCOND(Op, DAG);
1223   case ISD::ConstantPool:       return lowerConstantPool(Op, DAG);
1224   case ISD::GlobalAddress:      return lowerGlobalAddress(Op, DAG);
1225   case ISD::BlockAddress:       return lowerBlockAddress(Op, DAG);
1226   case ISD::GlobalTLSAddress:   return lowerGlobalTLSAddress(Op, DAG);
1227   case ISD::JumpTable:          return lowerJumpTable(Op, DAG);
1228   case ISD::SELECT:             return lowerSELECT(Op, DAG);
1229   case ISD::SETCC:              return lowerSETCC(Op, DAG);
1230   case ISD::VASTART:            return lowerVASTART(Op, DAG);
1231   case ISD::VAARG:              return lowerVAARG(Op, DAG);
1232   case ISD::FCOPYSIGN:          return lowerFCOPYSIGN(Op, DAG);
1233   case ISD::FABS:               return lowerFABS(Op, DAG);
1234   case ISD::FRAMEADDR:          return lowerFRAMEADDR(Op, DAG);
1235   case ISD::RETURNADDR:         return lowerRETURNADDR(Op, DAG);
1236   case ISD::EH_RETURN:          return lowerEH_RETURN(Op, DAG);
1237   case ISD::ATOMIC_FENCE:       return lowerATOMIC_FENCE(Op, DAG);
1238   case ISD::SHL_PARTS:          return lowerShiftLeftParts(Op, DAG);
1239   case ISD::SRA_PARTS:          return lowerShiftRightParts(Op, DAG, true);
1240   case ISD::SRL_PARTS:          return lowerShiftRightParts(Op, DAG, false);
1241   case ISD::LOAD:               return lowerLOAD(Op, DAG);
1242   case ISD::STORE:              return lowerSTORE(Op, DAG);
1243   case ISD::EH_DWARF_CFA:       return lowerEH_DWARF_CFA(Op, DAG);
1244   case ISD::FP_TO_SINT:         return lowerFP_TO_SINT(Op, DAG);
1245   }
1246   return SDValue();
1247 }
1248 
1249 //===----------------------------------------------------------------------===//
1250 //  Lower helper functions
1251 //===----------------------------------------------------------------------===//
1252 
1253 // addLiveIn - This helper function adds the specified physical register to the
1254 // MachineFunction as a live in value.  It also creates a corresponding
1255 // virtual register for it.
1256 static unsigned
1257 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
1258 {
1259   Register VReg = MF.getRegInfo().createVirtualRegister(RC);
1260   MF.getRegInfo().addLiveIn(PReg, VReg);
1261   return VReg;
1262 }
1263 
1264 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI,
1265                                               MachineBasicBlock &MBB,
1266                                               const TargetInstrInfo &TII,
1267                                               bool Is64Bit, bool IsMicroMips) {
1268   if (NoZeroDivCheck)
1269     return &MBB;
1270 
1271   // Insert instruction "teq $divisor_reg, $zero, 7".
1272   MachineBasicBlock::iterator I(MI);
1273   MachineInstrBuilder MIB;
1274   MachineOperand &Divisor = MI.getOperand(2);
1275   MIB = BuildMI(MBB, std::next(I), MI.getDebugLoc(),
1276                 TII.get(IsMicroMips ? Mips::TEQ_MM : Mips::TEQ))
1277             .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill()))
1278             .addReg(Mips::ZERO)
1279             .addImm(7);
1280 
1281   // Use the 32-bit sub-register if this is a 64-bit division.
1282   if (Is64Bit)
1283     MIB->getOperand(0).setSubReg(Mips::sub_32);
1284 
1285   // Clear Divisor's kill flag.
1286   Divisor.setIsKill(false);
1287 
1288   // We would normally delete the original instruction here but in this case
1289   // we only needed to inject an additional instruction rather than replace it.
1290 
1291   return &MBB;
1292 }
1293 
1294 MachineBasicBlock *
1295 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
1296                                                 MachineBasicBlock *BB) const {
1297   switch (MI.getOpcode()) {
1298   default:
1299     llvm_unreachable("Unexpected instr type to insert");
1300   case Mips::ATOMIC_LOAD_ADD_I8:
1301     return emitAtomicBinaryPartword(MI, BB, 1);
1302   case Mips::ATOMIC_LOAD_ADD_I16:
1303     return emitAtomicBinaryPartword(MI, BB, 2);
1304   case Mips::ATOMIC_LOAD_ADD_I32:
1305     return emitAtomicBinary(MI, BB);
1306   case Mips::ATOMIC_LOAD_ADD_I64:
1307     return emitAtomicBinary(MI, BB);
1308 
1309   case Mips::ATOMIC_LOAD_AND_I8:
1310     return emitAtomicBinaryPartword(MI, BB, 1);
1311   case Mips::ATOMIC_LOAD_AND_I16:
1312     return emitAtomicBinaryPartword(MI, BB, 2);
1313   case Mips::ATOMIC_LOAD_AND_I32:
1314     return emitAtomicBinary(MI, BB);
1315   case Mips::ATOMIC_LOAD_AND_I64:
1316     return emitAtomicBinary(MI, BB);
1317 
1318   case Mips::ATOMIC_LOAD_OR_I8:
1319     return emitAtomicBinaryPartword(MI, BB, 1);
1320   case Mips::ATOMIC_LOAD_OR_I16:
1321     return emitAtomicBinaryPartword(MI, BB, 2);
1322   case Mips::ATOMIC_LOAD_OR_I32:
1323     return emitAtomicBinary(MI, BB);
1324   case Mips::ATOMIC_LOAD_OR_I64:
1325     return emitAtomicBinary(MI, BB);
1326 
1327   case Mips::ATOMIC_LOAD_XOR_I8:
1328     return emitAtomicBinaryPartword(MI, BB, 1);
1329   case Mips::ATOMIC_LOAD_XOR_I16:
1330     return emitAtomicBinaryPartword(MI, BB, 2);
1331   case Mips::ATOMIC_LOAD_XOR_I32:
1332     return emitAtomicBinary(MI, BB);
1333   case Mips::ATOMIC_LOAD_XOR_I64:
1334     return emitAtomicBinary(MI, BB);
1335 
1336   case Mips::ATOMIC_LOAD_NAND_I8:
1337     return emitAtomicBinaryPartword(MI, BB, 1);
1338   case Mips::ATOMIC_LOAD_NAND_I16:
1339     return emitAtomicBinaryPartword(MI, BB, 2);
1340   case Mips::ATOMIC_LOAD_NAND_I32:
1341     return emitAtomicBinary(MI, BB);
1342   case Mips::ATOMIC_LOAD_NAND_I64:
1343     return emitAtomicBinary(MI, BB);
1344 
1345   case Mips::ATOMIC_LOAD_SUB_I8:
1346     return emitAtomicBinaryPartword(MI, BB, 1);
1347   case Mips::ATOMIC_LOAD_SUB_I16:
1348     return emitAtomicBinaryPartword(MI, BB, 2);
1349   case Mips::ATOMIC_LOAD_SUB_I32:
1350     return emitAtomicBinary(MI, BB);
1351   case Mips::ATOMIC_LOAD_SUB_I64:
1352     return emitAtomicBinary(MI, BB);
1353 
1354   case Mips::ATOMIC_SWAP_I8:
1355     return emitAtomicBinaryPartword(MI, BB, 1);
1356   case Mips::ATOMIC_SWAP_I16:
1357     return emitAtomicBinaryPartword(MI, BB, 2);
1358   case Mips::ATOMIC_SWAP_I32:
1359     return emitAtomicBinary(MI, BB);
1360   case Mips::ATOMIC_SWAP_I64:
1361     return emitAtomicBinary(MI, BB);
1362 
1363   case Mips::ATOMIC_CMP_SWAP_I8:
1364     return emitAtomicCmpSwapPartword(MI, BB, 1);
1365   case Mips::ATOMIC_CMP_SWAP_I16:
1366     return emitAtomicCmpSwapPartword(MI, BB, 2);
1367   case Mips::ATOMIC_CMP_SWAP_I32:
1368     return emitAtomicCmpSwap(MI, BB);
1369   case Mips::ATOMIC_CMP_SWAP_I64:
1370     return emitAtomicCmpSwap(MI, BB);
1371 
1372   case Mips::ATOMIC_LOAD_MIN_I8:
1373     return emitAtomicBinaryPartword(MI, BB, 1);
1374   case Mips::ATOMIC_LOAD_MIN_I16:
1375     return emitAtomicBinaryPartword(MI, BB, 2);
1376   case Mips::ATOMIC_LOAD_MIN_I32:
1377     return emitAtomicBinary(MI, BB);
1378   case Mips::ATOMIC_LOAD_MIN_I64:
1379     return emitAtomicBinary(MI, BB);
1380 
1381   case Mips::ATOMIC_LOAD_MAX_I8:
1382     return emitAtomicBinaryPartword(MI, BB, 1);
1383   case Mips::ATOMIC_LOAD_MAX_I16:
1384     return emitAtomicBinaryPartword(MI, BB, 2);
1385   case Mips::ATOMIC_LOAD_MAX_I32:
1386     return emitAtomicBinary(MI, BB);
1387   case Mips::ATOMIC_LOAD_MAX_I64:
1388     return emitAtomicBinary(MI, BB);
1389 
1390   case Mips::ATOMIC_LOAD_UMIN_I8:
1391     return emitAtomicBinaryPartword(MI, BB, 1);
1392   case Mips::ATOMIC_LOAD_UMIN_I16:
1393     return emitAtomicBinaryPartword(MI, BB, 2);
1394   case Mips::ATOMIC_LOAD_UMIN_I32:
1395     return emitAtomicBinary(MI, BB);
1396   case Mips::ATOMIC_LOAD_UMIN_I64:
1397     return emitAtomicBinary(MI, BB);
1398 
1399   case Mips::ATOMIC_LOAD_UMAX_I8:
1400     return emitAtomicBinaryPartword(MI, BB, 1);
1401   case Mips::ATOMIC_LOAD_UMAX_I16:
1402     return emitAtomicBinaryPartword(MI, BB, 2);
1403   case Mips::ATOMIC_LOAD_UMAX_I32:
1404     return emitAtomicBinary(MI, BB);
1405   case Mips::ATOMIC_LOAD_UMAX_I64:
1406     return emitAtomicBinary(MI, BB);
1407 
1408   case Mips::PseudoSDIV:
1409   case Mips::PseudoUDIV:
1410   case Mips::DIV:
1411   case Mips::DIVU:
1412   case Mips::MOD:
1413   case Mips::MODU:
1414     return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false,
1415                                false);
1416   case Mips::SDIV_MM_Pseudo:
1417   case Mips::UDIV_MM_Pseudo:
1418   case Mips::SDIV_MM:
1419   case Mips::UDIV_MM:
1420   case Mips::DIV_MMR6:
1421   case Mips::DIVU_MMR6:
1422   case Mips::MOD_MMR6:
1423   case Mips::MODU_MMR6:
1424     return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, true);
1425   case Mips::PseudoDSDIV:
1426   case Mips::PseudoDUDIV:
1427   case Mips::DDIV:
1428   case Mips::DDIVU:
1429   case Mips::DMOD:
1430   case Mips::DMODU:
1431     return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, false);
1432 
1433   case Mips::PseudoSELECT_I:
1434   case Mips::PseudoSELECT_I64:
1435   case Mips::PseudoSELECT_S:
1436   case Mips::PseudoSELECT_D32:
1437   case Mips::PseudoSELECT_D64:
1438     return emitPseudoSELECT(MI, BB, false, Mips::BNE);
1439   case Mips::PseudoSELECTFP_F_I:
1440   case Mips::PseudoSELECTFP_F_I64:
1441   case Mips::PseudoSELECTFP_F_S:
1442   case Mips::PseudoSELECTFP_F_D32:
1443   case Mips::PseudoSELECTFP_F_D64:
1444     return emitPseudoSELECT(MI, BB, true, Mips::BC1F);
1445   case Mips::PseudoSELECTFP_T_I:
1446   case Mips::PseudoSELECTFP_T_I64:
1447   case Mips::PseudoSELECTFP_T_S:
1448   case Mips::PseudoSELECTFP_T_D32:
1449   case Mips::PseudoSELECTFP_T_D64:
1450     return emitPseudoSELECT(MI, BB, true, Mips::BC1T);
1451   case Mips::PseudoD_SELECT_I:
1452   case Mips::PseudoD_SELECT_I64:
1453     return emitPseudoD_SELECT(MI, BB);
1454   case Mips::LDR_W:
1455     return emitLDR_W(MI, BB);
1456   case Mips::LDR_D:
1457     return emitLDR_D(MI, BB);
1458   case Mips::STR_W:
1459     return emitSTR_W(MI, BB);
1460   case Mips::STR_D:
1461     return emitSTR_D(MI, BB);
1462   }
1463 }
1464 
1465 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1466 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1467 MachineBasicBlock *
1468 MipsTargetLowering::emitAtomicBinary(MachineInstr &MI,
1469                                      MachineBasicBlock *BB) const {
1470 
1471   MachineFunction *MF = BB->getParent();
1472   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1473   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1474   DebugLoc DL = MI.getDebugLoc();
1475 
1476   unsigned AtomicOp;
1477   bool NeedsAdditionalReg = false;
1478   switch (MI.getOpcode()) {
1479   case Mips::ATOMIC_LOAD_ADD_I32:
1480     AtomicOp = Mips::ATOMIC_LOAD_ADD_I32_POSTRA;
1481     break;
1482   case Mips::ATOMIC_LOAD_SUB_I32:
1483     AtomicOp = Mips::ATOMIC_LOAD_SUB_I32_POSTRA;
1484     break;
1485   case Mips::ATOMIC_LOAD_AND_I32:
1486     AtomicOp = Mips::ATOMIC_LOAD_AND_I32_POSTRA;
1487     break;
1488   case Mips::ATOMIC_LOAD_OR_I32:
1489     AtomicOp = Mips::ATOMIC_LOAD_OR_I32_POSTRA;
1490     break;
1491   case Mips::ATOMIC_LOAD_XOR_I32:
1492     AtomicOp = Mips::ATOMIC_LOAD_XOR_I32_POSTRA;
1493     break;
1494   case Mips::ATOMIC_LOAD_NAND_I32:
1495     AtomicOp = Mips::ATOMIC_LOAD_NAND_I32_POSTRA;
1496     break;
1497   case Mips::ATOMIC_SWAP_I32:
1498     AtomicOp = Mips::ATOMIC_SWAP_I32_POSTRA;
1499     break;
1500   case Mips::ATOMIC_LOAD_ADD_I64:
1501     AtomicOp = Mips::ATOMIC_LOAD_ADD_I64_POSTRA;
1502     break;
1503   case Mips::ATOMIC_LOAD_SUB_I64:
1504     AtomicOp = Mips::ATOMIC_LOAD_SUB_I64_POSTRA;
1505     break;
1506   case Mips::ATOMIC_LOAD_AND_I64:
1507     AtomicOp = Mips::ATOMIC_LOAD_AND_I64_POSTRA;
1508     break;
1509   case Mips::ATOMIC_LOAD_OR_I64:
1510     AtomicOp = Mips::ATOMIC_LOAD_OR_I64_POSTRA;
1511     break;
1512   case Mips::ATOMIC_LOAD_XOR_I64:
1513     AtomicOp = Mips::ATOMIC_LOAD_XOR_I64_POSTRA;
1514     break;
1515   case Mips::ATOMIC_LOAD_NAND_I64:
1516     AtomicOp = Mips::ATOMIC_LOAD_NAND_I64_POSTRA;
1517     break;
1518   case Mips::ATOMIC_SWAP_I64:
1519     AtomicOp = Mips::ATOMIC_SWAP_I64_POSTRA;
1520     break;
1521   case Mips::ATOMIC_LOAD_MIN_I32:
1522     AtomicOp = Mips::ATOMIC_LOAD_MIN_I32_POSTRA;
1523     NeedsAdditionalReg = true;
1524     break;
1525   case Mips::ATOMIC_LOAD_MAX_I32:
1526     AtomicOp = Mips::ATOMIC_LOAD_MAX_I32_POSTRA;
1527     NeedsAdditionalReg = true;
1528     break;
1529   case Mips::ATOMIC_LOAD_UMIN_I32:
1530     AtomicOp = Mips::ATOMIC_LOAD_UMIN_I32_POSTRA;
1531     NeedsAdditionalReg = true;
1532     break;
1533   case Mips::ATOMIC_LOAD_UMAX_I32:
1534     AtomicOp = Mips::ATOMIC_LOAD_UMAX_I32_POSTRA;
1535     NeedsAdditionalReg = true;
1536     break;
1537   case Mips::ATOMIC_LOAD_MIN_I64:
1538     AtomicOp = Mips::ATOMIC_LOAD_MIN_I64_POSTRA;
1539     NeedsAdditionalReg = true;
1540     break;
1541   case Mips::ATOMIC_LOAD_MAX_I64:
1542     AtomicOp = Mips::ATOMIC_LOAD_MAX_I64_POSTRA;
1543     NeedsAdditionalReg = true;
1544     break;
1545   case Mips::ATOMIC_LOAD_UMIN_I64:
1546     AtomicOp = Mips::ATOMIC_LOAD_UMIN_I64_POSTRA;
1547     NeedsAdditionalReg = true;
1548     break;
1549   case Mips::ATOMIC_LOAD_UMAX_I64:
1550     AtomicOp = Mips::ATOMIC_LOAD_UMAX_I64_POSTRA;
1551     NeedsAdditionalReg = true;
1552     break;
1553   default:
1554     llvm_unreachable("Unknown pseudo atomic for replacement!");
1555   }
1556 
1557   Register OldVal = MI.getOperand(0).getReg();
1558   Register Ptr = MI.getOperand(1).getReg();
1559   Register Incr = MI.getOperand(2).getReg();
1560   Register Scratch = RegInfo.createVirtualRegister(RegInfo.getRegClass(OldVal));
1561 
1562   MachineBasicBlock::iterator II(MI);
1563 
1564   // The scratch registers here with the EarlyClobber | Define | Implicit
1565   // flags is used to persuade the register allocator and the machine
1566   // verifier to accept the usage of this register. This has to be a real
1567   // register which has an UNDEF value but is dead after the instruction which
1568   // is unique among the registers chosen for the instruction.
1569 
1570   // The EarlyClobber flag has the semantic properties that the operand it is
1571   // attached to is clobbered before the rest of the inputs are read. Hence it
1572   // must be unique among the operands to the instruction.
1573   // The Define flag is needed to coerce the machine verifier that an Undef
1574   // value isn't a problem.
1575   // The Dead flag is needed as the value in scratch isn't used by any other
1576   // instruction. Kill isn't used as Dead is more precise.
1577   // The implicit flag is here due to the interaction between the other flags
1578   // and the machine verifier.
1579 
1580   // For correctness purpose, a new pseudo is introduced here. We need this
1581   // new pseudo, so that FastRegisterAllocator does not see an ll/sc sequence
1582   // that is spread over >1 basic blocks. A register allocator which
1583   // introduces (or any codegen infact) a store, can violate the expectations
1584   // of the hardware.
1585   //
1586   // An atomic read-modify-write sequence starts with a linked load
1587   // instruction and ends with a store conditional instruction. The atomic
1588   // read-modify-write sequence fails if any of the following conditions
1589   // occur between the execution of ll and sc:
1590   //   * A coherent store is completed by another process or coherent I/O
1591   //     module into the block of synchronizable physical memory containing
1592   //     the word. The size and alignment of the block is
1593   //     implementation-dependent.
1594   //   * A coherent store is executed between an LL and SC sequence on the
1595   //     same processor to the block of synchornizable physical memory
1596   //     containing the word.
1597   //
1598 
1599   Register PtrCopy = RegInfo.createVirtualRegister(RegInfo.getRegClass(Ptr));
1600   Register IncrCopy = RegInfo.createVirtualRegister(RegInfo.getRegClass(Incr));
1601 
1602   BuildMI(*BB, II, DL, TII->get(Mips::COPY), IncrCopy).addReg(Incr);
1603   BuildMI(*BB, II, DL, TII->get(Mips::COPY), PtrCopy).addReg(Ptr);
1604 
1605   MachineInstrBuilder MIB =
1606       BuildMI(*BB, II, DL, TII->get(AtomicOp))
1607           .addReg(OldVal, RegState::Define | RegState::EarlyClobber)
1608           .addReg(PtrCopy)
1609           .addReg(IncrCopy)
1610           .addReg(Scratch, RegState::Define | RegState::EarlyClobber |
1611                                RegState::Implicit | RegState::Dead);
1612   if (NeedsAdditionalReg) {
1613     Register Scratch2 =
1614         RegInfo.createVirtualRegister(RegInfo.getRegClass(OldVal));
1615     MIB.addReg(Scratch2, RegState::Define | RegState::EarlyClobber |
1616                              RegState::Implicit | RegState::Dead);
1617   }
1618 
1619   MI.eraseFromParent();
1620 
1621   return BB;
1622 }
1623 
1624 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg(
1625     MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg,
1626     unsigned SrcReg) const {
1627   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1628   const DebugLoc &DL = MI.getDebugLoc();
1629 
1630   if (Subtarget.hasMips32r2() && Size == 1) {
1631     BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg);
1632     return BB;
1633   }
1634 
1635   if (Subtarget.hasMips32r2() && Size == 2) {
1636     BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg);
1637     return BB;
1638   }
1639 
1640   MachineFunction *MF = BB->getParent();
1641   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1642   const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1643   Register ScrReg = RegInfo.createVirtualRegister(RC);
1644 
1645   assert(Size < 32);
1646   int64_t ShiftImm = 32 - (Size * 8);
1647 
1648   BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm);
1649   BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm);
1650 
1651   return BB;
1652 }
1653 
1654 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword(
1655     MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const {
1656   assert((Size == 1 || Size == 2) &&
1657          "Unsupported size for EmitAtomicBinaryPartial.");
1658 
1659   MachineFunction *MF = BB->getParent();
1660   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1661   const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1662   const bool ArePtrs64bit = ABI.ArePtrs64bit();
1663   const TargetRegisterClass *RCp =
1664     getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32);
1665   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1666   DebugLoc DL = MI.getDebugLoc();
1667 
1668   Register Dest = MI.getOperand(0).getReg();
1669   Register Ptr = MI.getOperand(1).getReg();
1670   Register Incr = MI.getOperand(2).getReg();
1671 
1672   Register AlignedAddr = RegInfo.createVirtualRegister(RCp);
1673   Register ShiftAmt = RegInfo.createVirtualRegister(RC);
1674   Register Mask = RegInfo.createVirtualRegister(RC);
1675   Register Mask2 = RegInfo.createVirtualRegister(RC);
1676   Register Incr2 = RegInfo.createVirtualRegister(RC);
1677   Register MaskLSB2 = RegInfo.createVirtualRegister(RCp);
1678   Register PtrLSB2 = RegInfo.createVirtualRegister(RC);
1679   Register MaskUpper = RegInfo.createVirtualRegister(RC);
1680   Register Scratch = RegInfo.createVirtualRegister(RC);
1681   Register Scratch2 = RegInfo.createVirtualRegister(RC);
1682   Register Scratch3 = RegInfo.createVirtualRegister(RC);
1683 
1684   unsigned AtomicOp = 0;
1685   bool NeedsAdditionalReg = false;
1686   switch (MI.getOpcode()) {
1687   case Mips::ATOMIC_LOAD_NAND_I8:
1688     AtomicOp = Mips::ATOMIC_LOAD_NAND_I8_POSTRA;
1689     break;
1690   case Mips::ATOMIC_LOAD_NAND_I16:
1691     AtomicOp = Mips::ATOMIC_LOAD_NAND_I16_POSTRA;
1692     break;
1693   case Mips::ATOMIC_SWAP_I8:
1694     AtomicOp = Mips::ATOMIC_SWAP_I8_POSTRA;
1695     break;
1696   case Mips::ATOMIC_SWAP_I16:
1697     AtomicOp = Mips::ATOMIC_SWAP_I16_POSTRA;
1698     break;
1699   case Mips::ATOMIC_LOAD_ADD_I8:
1700     AtomicOp = Mips::ATOMIC_LOAD_ADD_I8_POSTRA;
1701     break;
1702   case Mips::ATOMIC_LOAD_ADD_I16:
1703     AtomicOp = Mips::ATOMIC_LOAD_ADD_I16_POSTRA;
1704     break;
1705   case Mips::ATOMIC_LOAD_SUB_I8:
1706     AtomicOp = Mips::ATOMIC_LOAD_SUB_I8_POSTRA;
1707     break;
1708   case Mips::ATOMIC_LOAD_SUB_I16:
1709     AtomicOp = Mips::ATOMIC_LOAD_SUB_I16_POSTRA;
1710     break;
1711   case Mips::ATOMIC_LOAD_AND_I8:
1712     AtomicOp = Mips::ATOMIC_LOAD_AND_I8_POSTRA;
1713     break;
1714   case Mips::ATOMIC_LOAD_AND_I16:
1715     AtomicOp = Mips::ATOMIC_LOAD_AND_I16_POSTRA;
1716     break;
1717   case Mips::ATOMIC_LOAD_OR_I8:
1718     AtomicOp = Mips::ATOMIC_LOAD_OR_I8_POSTRA;
1719     break;
1720   case Mips::ATOMIC_LOAD_OR_I16:
1721     AtomicOp = Mips::ATOMIC_LOAD_OR_I16_POSTRA;
1722     break;
1723   case Mips::ATOMIC_LOAD_XOR_I8:
1724     AtomicOp = Mips::ATOMIC_LOAD_XOR_I8_POSTRA;
1725     break;
1726   case Mips::ATOMIC_LOAD_XOR_I16:
1727     AtomicOp = Mips::ATOMIC_LOAD_XOR_I16_POSTRA;
1728     break;
1729   case Mips::ATOMIC_LOAD_MIN_I8:
1730     AtomicOp = Mips::ATOMIC_LOAD_MIN_I8_POSTRA;
1731     NeedsAdditionalReg = true;
1732     break;
1733   case Mips::ATOMIC_LOAD_MIN_I16:
1734     AtomicOp = Mips::ATOMIC_LOAD_MIN_I16_POSTRA;
1735     NeedsAdditionalReg = true;
1736     break;
1737   case Mips::ATOMIC_LOAD_MAX_I8:
1738     AtomicOp = Mips::ATOMIC_LOAD_MAX_I8_POSTRA;
1739     NeedsAdditionalReg = true;
1740     break;
1741   case Mips::ATOMIC_LOAD_MAX_I16:
1742     AtomicOp = Mips::ATOMIC_LOAD_MAX_I16_POSTRA;
1743     NeedsAdditionalReg = true;
1744     break;
1745   case Mips::ATOMIC_LOAD_UMIN_I8:
1746     AtomicOp = Mips::ATOMIC_LOAD_UMIN_I8_POSTRA;
1747     NeedsAdditionalReg = true;
1748     break;
1749   case Mips::ATOMIC_LOAD_UMIN_I16:
1750     AtomicOp = Mips::ATOMIC_LOAD_UMIN_I16_POSTRA;
1751     NeedsAdditionalReg = true;
1752     break;
1753   case Mips::ATOMIC_LOAD_UMAX_I8:
1754     AtomicOp = Mips::ATOMIC_LOAD_UMAX_I8_POSTRA;
1755     NeedsAdditionalReg = true;
1756     break;
1757   case Mips::ATOMIC_LOAD_UMAX_I16:
1758     AtomicOp = Mips::ATOMIC_LOAD_UMAX_I16_POSTRA;
1759     NeedsAdditionalReg = true;
1760     break;
1761   default:
1762     llvm_unreachable("Unknown subword atomic pseudo for expansion!");
1763   }
1764 
1765   // insert new blocks after the current block
1766   const BasicBlock *LLVM_BB = BB->getBasicBlock();
1767   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1768   MachineFunction::iterator It = ++BB->getIterator();
1769   MF->insert(It, exitMBB);
1770 
1771   // Transfer the remainder of BB and its successor edges to exitMBB.
1772   exitMBB->splice(exitMBB->begin(), BB,
1773                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
1774   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1775 
1776   BB->addSuccessor(exitMBB, BranchProbability::getOne());
1777 
1778   //  thisMBB:
1779   //    addiu   masklsb2,$0,-4                # 0xfffffffc
1780   //    and     alignedaddr,ptr,masklsb2
1781   //    andi    ptrlsb2,ptr,3
1782   //    sll     shiftamt,ptrlsb2,3
1783   //    ori     maskupper,$0,255               # 0xff
1784   //    sll     mask,maskupper,shiftamt
1785   //    nor     mask2,$0,mask
1786   //    sll     incr2,incr,shiftamt
1787 
1788   int64_t MaskImm = (Size == 1) ? 255 : 65535;
1789   BuildMI(BB, DL, TII->get(ABI.GetPtrAddiuOp()), MaskLSB2)
1790     .addReg(ABI.GetNullPtr()).addImm(-4);
1791   BuildMI(BB, DL, TII->get(ABI.GetPtrAndOp()), AlignedAddr)
1792     .addReg(Ptr).addReg(MaskLSB2);
1793   BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2)
1794       .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3);
1795   if (Subtarget.isLittle()) {
1796     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1797   } else {
1798     Register Off = RegInfo.createVirtualRegister(RC);
1799     BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1800       .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1801     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1802   }
1803   BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1804     .addReg(Mips::ZERO).addImm(MaskImm);
1805   BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1806     .addReg(MaskUpper).addReg(ShiftAmt);
1807   BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1808   BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1809 
1810 
1811   // The purposes of the flags on the scratch registers is explained in
1812   // emitAtomicBinary. In summary, we need a scratch register which is going to
1813   // be undef, that is unique among registers chosen for the instruction.
1814 
1815   MachineInstrBuilder MIB =
1816       BuildMI(BB, DL, TII->get(AtomicOp))
1817           .addReg(Dest, RegState::Define | RegState::EarlyClobber)
1818           .addReg(AlignedAddr)
1819           .addReg(Incr2)
1820           .addReg(Mask)
1821           .addReg(Mask2)
1822           .addReg(ShiftAmt)
1823           .addReg(Scratch, RegState::EarlyClobber | RegState::Define |
1824                                RegState::Dead | RegState::Implicit)
1825           .addReg(Scratch2, RegState::EarlyClobber | RegState::Define |
1826                                 RegState::Dead | RegState::Implicit)
1827           .addReg(Scratch3, RegState::EarlyClobber | RegState::Define |
1828                                 RegState::Dead | RegState::Implicit);
1829   if (NeedsAdditionalReg) {
1830     Register Scratch4 = RegInfo.createVirtualRegister(RC);
1831     MIB.addReg(Scratch4, RegState::EarlyClobber | RegState::Define |
1832                              RegState::Dead | RegState::Implicit);
1833   }
1834 
1835   MI.eraseFromParent(); // The instruction is gone now.
1836 
1837   return exitMBB;
1838 }
1839 
1840 // Lower atomic compare and swap to a pseudo instruction, taking care to
1841 // define a scratch register for the pseudo instruction's expansion. The
1842 // instruction is expanded after the register allocator as to prevent
1843 // the insertion of stores between the linked load and the store conditional.
1844 
1845 MachineBasicBlock *
1846 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI,
1847                                       MachineBasicBlock *BB) const {
1848 
1849   assert((MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 ||
1850           MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I64) &&
1851          "Unsupported atomic pseudo for EmitAtomicCmpSwap.");
1852 
1853   const unsigned Size = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 ? 4 : 8;
1854 
1855   MachineFunction *MF = BB->getParent();
1856   MachineRegisterInfo &MRI = MF->getRegInfo();
1857   const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1858   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1859   DebugLoc DL = MI.getDebugLoc();
1860 
1861   unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32
1862                           ? Mips::ATOMIC_CMP_SWAP_I32_POSTRA
1863                           : Mips::ATOMIC_CMP_SWAP_I64_POSTRA;
1864   Register Dest = MI.getOperand(0).getReg();
1865   Register Ptr = MI.getOperand(1).getReg();
1866   Register OldVal = MI.getOperand(2).getReg();
1867   Register NewVal = MI.getOperand(3).getReg();
1868 
1869   Register Scratch = MRI.createVirtualRegister(RC);
1870   MachineBasicBlock::iterator II(MI);
1871 
1872   // We need to create copies of the various registers and kill them at the
1873   // atomic pseudo. If the copies are not made, when the atomic is expanded
1874   // after fast register allocation, the spills will end up outside of the
1875   // blocks that their values are defined in, causing livein errors.
1876 
1877   Register PtrCopy = MRI.createVirtualRegister(MRI.getRegClass(Ptr));
1878   Register OldValCopy = MRI.createVirtualRegister(MRI.getRegClass(OldVal));
1879   Register NewValCopy = MRI.createVirtualRegister(MRI.getRegClass(NewVal));
1880 
1881   BuildMI(*BB, II, DL, TII->get(Mips::COPY), PtrCopy).addReg(Ptr);
1882   BuildMI(*BB, II, DL, TII->get(Mips::COPY), OldValCopy).addReg(OldVal);
1883   BuildMI(*BB, II, DL, TII->get(Mips::COPY), NewValCopy).addReg(NewVal);
1884 
1885   // The purposes of the flags on the scratch registers is explained in
1886   // emitAtomicBinary. In summary, we need a scratch register which is going to
1887   // be undef, that is unique among registers chosen for the instruction.
1888 
1889   BuildMI(*BB, II, DL, TII->get(AtomicOp))
1890       .addReg(Dest, RegState::Define | RegState::EarlyClobber)
1891       .addReg(PtrCopy, RegState::Kill)
1892       .addReg(OldValCopy, RegState::Kill)
1893       .addReg(NewValCopy, RegState::Kill)
1894       .addReg(Scratch, RegState::EarlyClobber | RegState::Define |
1895                            RegState::Dead | RegState::Implicit);
1896 
1897   MI.eraseFromParent(); // The instruction is gone now.
1898 
1899   return BB;
1900 }
1901 
1902 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword(
1903     MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const {
1904   assert((Size == 1 || Size == 2) &&
1905       "Unsupported size for EmitAtomicCmpSwapPartial.");
1906 
1907   MachineFunction *MF = BB->getParent();
1908   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1909   const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1910   const bool ArePtrs64bit = ABI.ArePtrs64bit();
1911   const TargetRegisterClass *RCp =
1912     getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32);
1913   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1914   DebugLoc DL = MI.getDebugLoc();
1915 
1916   Register Dest = MI.getOperand(0).getReg();
1917   Register Ptr = MI.getOperand(1).getReg();
1918   Register CmpVal = MI.getOperand(2).getReg();
1919   Register NewVal = MI.getOperand(3).getReg();
1920 
1921   Register AlignedAddr = RegInfo.createVirtualRegister(RCp);
1922   Register ShiftAmt = RegInfo.createVirtualRegister(RC);
1923   Register Mask = RegInfo.createVirtualRegister(RC);
1924   Register Mask2 = RegInfo.createVirtualRegister(RC);
1925   Register ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1926   Register ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1927   Register MaskLSB2 = RegInfo.createVirtualRegister(RCp);
1928   Register PtrLSB2 = RegInfo.createVirtualRegister(RC);
1929   Register MaskUpper = RegInfo.createVirtualRegister(RC);
1930   Register MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1931   Register MaskedNewVal = RegInfo.createVirtualRegister(RC);
1932   unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I8
1933                           ? Mips::ATOMIC_CMP_SWAP_I8_POSTRA
1934                           : Mips::ATOMIC_CMP_SWAP_I16_POSTRA;
1935 
1936   // The scratch registers here with the EarlyClobber | Define | Dead | Implicit
1937   // flags are used to coerce the register allocator and the machine verifier to
1938   // accept the usage of these registers.
1939   // The EarlyClobber flag has the semantic properties that the operand it is
1940   // attached to is clobbered before the rest of the inputs are read. Hence it
1941   // must be unique among the operands to the instruction.
1942   // The Define flag is needed to coerce the machine verifier that an Undef
1943   // value isn't a problem.
1944   // The Dead flag is needed as the value in scratch isn't used by any other
1945   // instruction. Kill isn't used as Dead is more precise.
1946   Register Scratch = RegInfo.createVirtualRegister(RC);
1947   Register Scratch2 = RegInfo.createVirtualRegister(RC);
1948 
1949   // insert new blocks after the current block
1950   const BasicBlock *LLVM_BB = BB->getBasicBlock();
1951   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1952   MachineFunction::iterator It = ++BB->getIterator();
1953   MF->insert(It, exitMBB);
1954 
1955   // Transfer the remainder of BB and its successor edges to exitMBB.
1956   exitMBB->splice(exitMBB->begin(), BB,
1957                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
1958   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1959 
1960   BB->addSuccessor(exitMBB, BranchProbability::getOne());
1961 
1962   //  thisMBB:
1963   //    addiu   masklsb2,$0,-4                # 0xfffffffc
1964   //    and     alignedaddr,ptr,masklsb2
1965   //    andi    ptrlsb2,ptr,3
1966   //    xori    ptrlsb2,ptrlsb2,3              # Only for BE
1967   //    sll     shiftamt,ptrlsb2,3
1968   //    ori     maskupper,$0,255               # 0xff
1969   //    sll     mask,maskupper,shiftamt
1970   //    nor     mask2,$0,mask
1971   //    andi    maskedcmpval,cmpval,255
1972   //    sll     shiftedcmpval,maskedcmpval,shiftamt
1973   //    andi    maskednewval,newval,255
1974   //    sll     shiftednewval,maskednewval,shiftamt
1975   int64_t MaskImm = (Size == 1) ? 255 : 65535;
1976   BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), MaskLSB2)
1977     .addReg(ABI.GetNullPtr()).addImm(-4);
1978   BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::AND64 : Mips::AND), AlignedAddr)
1979     .addReg(Ptr).addReg(MaskLSB2);
1980   BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2)
1981       .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3);
1982   if (Subtarget.isLittle()) {
1983     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1984   } else {
1985     Register Off = RegInfo.createVirtualRegister(RC);
1986     BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1987       .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1988     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1989   }
1990   BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1991     .addReg(Mips::ZERO).addImm(MaskImm);
1992   BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1993     .addReg(MaskUpper).addReg(ShiftAmt);
1994   BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1995   BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1996     .addReg(CmpVal).addImm(MaskImm);
1997   BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1998     .addReg(MaskedCmpVal).addReg(ShiftAmt);
1999   BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
2000     .addReg(NewVal).addImm(MaskImm);
2001   BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
2002     .addReg(MaskedNewVal).addReg(ShiftAmt);
2003 
2004   // The purposes of the flags on the scratch registers are explained in
2005   // emitAtomicBinary. In summary, we need a scratch register which is going to
2006   // be undef, that is unique among the register chosen for the instruction.
2007 
2008   BuildMI(BB, DL, TII->get(AtomicOp))
2009       .addReg(Dest, RegState::Define | RegState::EarlyClobber)
2010       .addReg(AlignedAddr)
2011       .addReg(Mask)
2012       .addReg(ShiftedCmpVal)
2013       .addReg(Mask2)
2014       .addReg(ShiftedNewVal)
2015       .addReg(ShiftAmt)
2016       .addReg(Scratch, RegState::EarlyClobber | RegState::Define |
2017                            RegState::Dead | RegState::Implicit)
2018       .addReg(Scratch2, RegState::EarlyClobber | RegState::Define |
2019                             RegState::Dead | RegState::Implicit);
2020 
2021   MI.eraseFromParent(); // The instruction is gone now.
2022 
2023   return exitMBB;
2024 }
2025 
2026 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
2027   // The first operand is the chain, the second is the condition, the third is
2028   // the block to branch to if the condition is true.
2029   SDValue Chain = Op.getOperand(0);
2030   SDValue Dest = Op.getOperand(2);
2031   SDLoc DL(Op);
2032 
2033   assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
2034   SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
2035 
2036   // Return if flag is not set by a floating point comparison.
2037   if (CondRes.getOpcode() != MipsISD::FPCmp)
2038     return Op;
2039 
2040   SDValue CCNode  = CondRes.getOperand(2);
2041   Mips::CondCode CC =
2042     (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
2043   unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
2044   SDValue BrCode = DAG.getConstant(Opc, DL, MVT::i32);
2045   SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
2046   return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
2047                      FCC0, Dest, CondRes);
2048 }
2049 
2050 SDValue MipsTargetLowering::
2051 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
2052 {
2053   assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
2054   SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
2055 
2056   // Return if flag is not set by a floating point comparison.
2057   if (Cond.getOpcode() != MipsISD::FPCmp)
2058     return Op;
2059 
2060   return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
2061                       SDLoc(Op));
2062 }
2063 
2064 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2065   assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
2066   SDValue Cond = createFPCmp(DAG, Op);
2067 
2068   assert(Cond.getOpcode() == MipsISD::FPCmp &&
2069          "Floating point operand expected.");
2070 
2071   SDLoc DL(Op);
2072   SDValue True  = DAG.getConstant(1, DL, MVT::i32);
2073   SDValue False = DAG.getConstant(0, DL, MVT::i32);
2074 
2075   return createCMovFP(DAG, Cond, True, False, DL);
2076 }
2077 
2078 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
2079                                                SelectionDAG &DAG) const {
2080   EVT Ty = Op.getValueType();
2081   GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
2082   const GlobalValue *GV = N->getGlobal();
2083 
2084   if (!isPositionIndependent()) {
2085     const MipsTargetObjectFile *TLOF =
2086         static_cast<const MipsTargetObjectFile *>(
2087             getTargetMachine().getObjFileLowering());
2088     const GlobalObject *GO = GV->getBaseObject();
2089     if (GO && TLOF->IsGlobalInSmallSection(GO, getTargetMachine()))
2090       // %gp_rel relocation
2091       return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64());
2092 
2093                                 // %hi/%lo relocation
2094     return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
2095                                 // %highest/%higher/%hi/%lo relocation
2096                                 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
2097   }
2098 
2099   // Every other architecture would use shouldAssumeDSOLocal in here, but
2100   // mips is special.
2101   // * In PIC code mips requires got loads even for local statics!
2102   // * To save on got entries, for local statics the got entry contains the
2103   //   page and an additional add instruction takes care of the low bits.
2104   // * It is legal to access a hidden symbol with a non hidden undefined,
2105   //   so one cannot guarantee that all access to a hidden symbol will know
2106   //   it is hidden.
2107   // * Mips linkers don't support creating a page and a full got entry for
2108   //   the same symbol.
2109   // * Given all that, we have to use a full got entry for hidden symbols :-(
2110   if (GV->hasLocalLinkage())
2111     return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
2112 
2113   if (Subtarget.useXGOT())
2114     return getAddrGlobalLargeGOT(
2115         N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, MipsII::MO_GOT_LO16,
2116         DAG.getEntryNode(),
2117         MachinePointerInfo::getGOT(DAG.getMachineFunction()));
2118 
2119   return getAddrGlobal(
2120       N, SDLoc(N), Ty, DAG,
2121       (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT,
2122       DAG.getEntryNode(), MachinePointerInfo::getGOT(DAG.getMachineFunction()));
2123 }
2124 
2125 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
2126                                               SelectionDAG &DAG) const {
2127   BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
2128   EVT Ty = Op.getValueType();
2129 
2130   if (!isPositionIndependent())
2131     return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
2132                                 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
2133 
2134   return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
2135 }
2136 
2137 SDValue MipsTargetLowering::
2138 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
2139 {
2140   // If the relocation model is PIC, use the General Dynamic TLS Model or
2141   // Local Dynamic TLS model, otherwise use the Initial Exec or
2142   // Local Exec TLS Model.
2143 
2144   GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2145   if (DAG.getTarget().useEmulatedTLS())
2146     return LowerToTLSEmulatedModel(GA, DAG);
2147 
2148   SDLoc DL(GA);
2149   const GlobalValue *GV = GA->getGlobal();
2150   EVT PtrVT = getPointerTy(DAG.getDataLayout());
2151 
2152   TLSModel::Model model = getTargetMachine().getTLSModel(GV);
2153 
2154   if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
2155     // General Dynamic and Local Dynamic TLS Model.
2156     unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
2157                                                       : MipsII::MO_TLSGD;
2158 
2159     SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
2160     SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
2161                                    getGlobalReg(DAG, PtrVT), TGA);
2162     unsigned PtrSize = PtrVT.getSizeInBits();
2163     IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
2164 
2165     SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
2166 
2167     ArgListTy Args;
2168     ArgListEntry Entry;
2169     Entry.Node = Argument;
2170     Entry.Ty = PtrTy;
2171     Args.push_back(Entry);
2172 
2173     TargetLowering::CallLoweringInfo CLI(DAG);
2174     CLI.setDebugLoc(DL)
2175         .setChain(DAG.getEntryNode())
2176         .setLibCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args));
2177     std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2178 
2179     SDValue Ret = CallResult.first;
2180 
2181     if (model != TLSModel::LocalDynamic)
2182       return Ret;
2183 
2184     SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2185                                                MipsII::MO_DTPREL_HI);
2186     SDValue Hi = DAG.getNode(MipsISD::TlsHi, DL, PtrVT, TGAHi);
2187     SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2188                                                MipsII::MO_DTPREL_LO);
2189     SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
2190     SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
2191     return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
2192   }
2193 
2194   SDValue Offset;
2195   if (model == TLSModel::InitialExec) {
2196     // Initial Exec TLS Model
2197     SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2198                                              MipsII::MO_GOTTPREL);
2199     TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
2200                       TGA);
2201     Offset =
2202         DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), TGA, MachinePointerInfo());
2203   } else {
2204     // Local Exec TLS Model
2205     assert(model == TLSModel::LocalExec);
2206     SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2207                                                MipsII::MO_TPREL_HI);
2208     SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2209                                                MipsII::MO_TPREL_LO);
2210     SDValue Hi = DAG.getNode(MipsISD::TlsHi, DL, PtrVT, TGAHi);
2211     SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
2212     Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
2213   }
2214 
2215   SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
2216   return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
2217 }
2218 
2219 SDValue MipsTargetLowering::
2220 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
2221 {
2222   JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
2223   EVT Ty = Op.getValueType();
2224 
2225   if (!isPositionIndependent())
2226     return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
2227                                 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
2228 
2229   return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
2230 }
2231 
2232 SDValue MipsTargetLowering::
2233 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
2234 {
2235   ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
2236   EVT Ty = Op.getValueType();
2237 
2238   if (!isPositionIndependent()) {
2239     const MipsTargetObjectFile *TLOF =
2240         static_cast<const MipsTargetObjectFile *>(
2241             getTargetMachine().getObjFileLowering());
2242 
2243     if (TLOF->IsConstantInSmallSection(DAG.getDataLayout(), N->getConstVal(),
2244                                        getTargetMachine()))
2245       // %gp_rel relocation
2246       return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64());
2247 
2248     return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
2249                                 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
2250   }
2251 
2252  return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
2253 }
2254 
2255 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
2256   MachineFunction &MF = DAG.getMachineFunction();
2257   MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2258 
2259   SDLoc DL(Op);
2260   SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
2261                                  getPointerTy(MF.getDataLayout()));
2262 
2263   // vastart just stores the address of the VarArgsFrameIndex slot into the
2264   // memory location argument.
2265   const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2266   return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
2267                       MachinePointerInfo(SV));
2268 }
2269 
2270 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const {
2271   SDNode *Node = Op.getNode();
2272   EVT VT = Node->getValueType(0);
2273   SDValue Chain = Node->getOperand(0);
2274   SDValue VAListPtr = Node->getOperand(1);
2275   const Align Align =
2276       llvm::MaybeAlign(Node->getConstantOperandVal(3)).valueOrOne();
2277   const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2278   SDLoc DL(Node);
2279   unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4;
2280 
2281   SDValue VAListLoad = DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL, Chain,
2282                                    VAListPtr, MachinePointerInfo(SV));
2283   SDValue VAList = VAListLoad;
2284 
2285   // Re-align the pointer if necessary.
2286   // It should only ever be necessary for 64-bit types on O32 since the minimum
2287   // argument alignment is the same as the maximum type alignment for N32/N64.
2288   //
2289   // FIXME: We currently align too often. The code generator doesn't notice
2290   //        when the pointer is still aligned from the last va_arg (or pair of
2291   //        va_args for the i64 on O32 case).
2292   if (Align > getMinStackArgumentAlignment()) {
2293     VAList = DAG.getNode(
2294         ISD::ADD, DL, VAList.getValueType(), VAList,
2295         DAG.getConstant(Align.value() - 1, DL, VAList.getValueType()));
2296 
2297     VAList = DAG.getNode(
2298         ISD::AND, DL, VAList.getValueType(), VAList,
2299         DAG.getConstant(-(int64_t)Align.value(), DL, VAList.getValueType()));
2300   }
2301 
2302   // Increment the pointer, VAList, to the next vaarg.
2303   auto &TD = DAG.getDataLayout();
2304   unsigned ArgSizeInBytes =
2305       TD.getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext()));
2306   SDValue Tmp3 =
2307       DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
2308                   DAG.getConstant(alignTo(ArgSizeInBytes, ArgSlotSizeInBytes),
2309                                   DL, VAList.getValueType()));
2310   // Store the incremented VAList to the legalized pointer
2311   Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr,
2312                        MachinePointerInfo(SV));
2313 
2314   // In big-endian mode we must adjust the pointer when the load size is smaller
2315   // than the argument slot size. We must also reduce the known alignment to
2316   // match. For example in the N64 ABI, we must add 4 bytes to the offset to get
2317   // the correct half of the slot, and reduce the alignment from 8 (slot
2318   // alignment) down to 4 (type alignment).
2319   if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) {
2320     unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes;
2321     VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList,
2322                          DAG.getIntPtrConstant(Adjustment, DL));
2323   }
2324   // Load the actual argument out of the pointer VAList
2325   return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo());
2326 }
2327 
2328 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG,
2329                                 bool HasExtractInsert) {
2330   EVT TyX = Op.getOperand(0).getValueType();
2331   EVT TyY = Op.getOperand(1).getValueType();
2332   SDLoc DL(Op);
2333   SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
2334   SDValue Const31 = DAG.getConstant(31, DL, MVT::i32);
2335   SDValue Res;
2336 
2337   // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2338   // to i32.
2339   SDValue X = (TyX == MVT::f32) ?
2340     DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
2341     DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2342                 Const1);
2343   SDValue Y = (TyY == MVT::f32) ?
2344     DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
2345     DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
2346                 Const1);
2347 
2348   if (HasExtractInsert) {
2349     // ext  E, Y, 31, 1  ; extract bit31 of Y
2350     // ins  X, E, 31, 1  ; insert extracted bit at bit31 of X
2351     SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
2352     Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
2353   } else {
2354     // sll SllX, X, 1
2355     // srl SrlX, SllX, 1
2356     // srl SrlY, Y, 31
2357     // sll SllY, SrlX, 31
2358     // or  Or, SrlX, SllY
2359     SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2360     SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2361     SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
2362     SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
2363     Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
2364   }
2365 
2366   if (TyX == MVT::f32)
2367     return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
2368 
2369   SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2370                              Op.getOperand(0),
2371                              DAG.getConstant(0, DL, MVT::i32));
2372   return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2373 }
2374 
2375 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG,
2376                                 bool HasExtractInsert) {
2377   unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
2378   unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
2379   EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
2380   SDLoc DL(Op);
2381   SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
2382 
2383   // Bitcast to integer nodes.
2384   SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
2385   SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
2386 
2387   if (HasExtractInsert) {
2388     // ext  E, Y, width(Y) - 1, 1  ; extract bit width(Y)-1 of Y
2389     // ins  X, E, width(X) - 1, 1  ; insert extracted bit at bit width(X)-1 of X
2390     SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
2391                             DAG.getConstant(WidthY - 1, DL, MVT::i32), Const1);
2392 
2393     if (WidthX > WidthY)
2394       E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
2395     else if (WidthY > WidthX)
2396       E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
2397 
2398     SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
2399                             DAG.getConstant(WidthX - 1, DL, MVT::i32), Const1,
2400                             X);
2401     return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
2402   }
2403 
2404   // (d)sll SllX, X, 1
2405   // (d)srl SrlX, SllX, 1
2406   // (d)srl SrlY, Y, width(Y)-1
2407   // (d)sll SllY, SrlX, width(Y)-1
2408   // or     Or, SrlX, SllY
2409   SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
2410   SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
2411   SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
2412                              DAG.getConstant(WidthY - 1, DL, MVT::i32));
2413 
2414   if (WidthX > WidthY)
2415     SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
2416   else if (WidthY > WidthX)
2417     SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
2418 
2419   SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
2420                              DAG.getConstant(WidthX - 1, DL, MVT::i32));
2421   SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
2422   return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
2423 }
2424 
2425 SDValue
2426 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2427   if (Subtarget.isGP64bit())
2428     return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert());
2429 
2430   return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert());
2431 }
2432 
2433 static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG,
2434                            bool HasExtractInsert) {
2435   SDLoc DL(Op);
2436   SDValue Res, Const1 = DAG.getConstant(1, DL, MVT::i32);
2437 
2438   // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2439   // to i32.
2440   SDValue X = (Op.getValueType() == MVT::f32)
2441                   ? DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0))
2442                   : DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2443                                 Op.getOperand(0), Const1);
2444 
2445   // Clear MSB.
2446   if (HasExtractInsert)
2447     Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
2448                       DAG.getRegister(Mips::ZERO, MVT::i32),
2449                       DAG.getConstant(31, DL, MVT::i32), Const1, X);
2450   else {
2451     // TODO: Provide DAG patterns which transform (and x, cst)
2452     // back to a (shl (srl x (clz cst)) (clz cst)) sequence.
2453     SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2454     Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2455   }
2456 
2457   if (Op.getValueType() == MVT::f32)
2458     return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
2459 
2460   // FIXME: For mips32r2, the sequence of (BuildPairF64 (ins (ExtractElementF64
2461   // Op 1), $zero, 31 1) (ExtractElementF64 Op 0)) and the Op has one use, we
2462   // should be able to drop the usage of mfc1/mtc1 and rewrite the register in
2463   // place.
2464   SDValue LowX =
2465       DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2466                   DAG.getConstant(0, DL, MVT::i32));
2467   return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2468 }
2469 
2470 static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG,
2471                            bool HasExtractInsert) {
2472   SDLoc DL(Op);
2473   SDValue Res, Const1 = DAG.getConstant(1, DL, MVT::i32);
2474 
2475   // Bitcast to integer node.
2476   SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
2477 
2478   // Clear MSB.
2479   if (HasExtractInsert)
2480     Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
2481                       DAG.getRegister(Mips::ZERO_64, MVT::i64),
2482                       DAG.getConstant(63, DL, MVT::i32), Const1, X);
2483   else {
2484     SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
2485     Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
2486   }
2487 
2488   return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
2489 }
2490 
2491 SDValue MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const {
2492   if ((ABI.IsN32() || ABI.IsN64()) && (Op.getValueType() == MVT::f64))
2493     return lowerFABS64(Op, DAG, Subtarget.hasExtractInsert());
2494 
2495   return lowerFABS32(Op, DAG, Subtarget.hasExtractInsert());
2496 }
2497 
2498 SDValue MipsTargetLowering::
2499 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2500   // check the depth
2501   if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() != 0) {
2502     DAG.getContext()->emitError(
2503         "return address can be determined only for current frame");
2504     return SDValue();
2505   }
2506 
2507   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2508   MFI.setFrameAddressIsTaken(true);
2509   EVT VT = Op.getValueType();
2510   SDLoc DL(Op);
2511   SDValue FrameAddr = DAG.getCopyFromReg(
2512       DAG.getEntryNode(), DL, ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT);
2513   return FrameAddr;
2514 }
2515 
2516 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
2517                                             SelectionDAG &DAG) const {
2518   if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2519     return SDValue();
2520 
2521   // check the depth
2522   if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() != 0) {
2523     DAG.getContext()->emitError(
2524         "return address can be determined only for current frame");
2525     return SDValue();
2526   }
2527 
2528   MachineFunction &MF = DAG.getMachineFunction();
2529   MachineFrameInfo &MFI = MF.getFrameInfo();
2530   MVT VT = Op.getSimpleValueType();
2531   unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA;
2532   MFI.setReturnAddressIsTaken(true);
2533 
2534   // Return RA, which contains the return address. Mark it an implicit live-in.
2535   unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2536   return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
2537 }
2538 
2539 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
2540 // generated from __builtin_eh_return (offset, handler)
2541 // The effect of this is to adjust the stack pointer by "offset"
2542 // and then branch to "handler".
2543 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
2544                                                                      const {
2545   MachineFunction &MF = DAG.getMachineFunction();
2546   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2547 
2548   MipsFI->setCallsEhReturn();
2549   SDValue Chain     = Op.getOperand(0);
2550   SDValue Offset    = Op.getOperand(1);
2551   SDValue Handler   = Op.getOperand(2);
2552   SDLoc DL(Op);
2553   EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
2554 
2555   // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2556   // EH_RETURN nodes, so that instructions are emitted back-to-back.
2557   unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1;
2558   unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
2559   Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2560   Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2561   return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2562                      DAG.getRegister(OffsetReg, Ty),
2563                      DAG.getRegister(AddrReg, getPointerTy(MF.getDataLayout())),
2564                      Chain.getValue(1));
2565 }
2566 
2567 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
2568                                               SelectionDAG &DAG) const {
2569   // FIXME: Need pseudo-fence for 'singlethread' fences
2570   // FIXME: Set SType for weaker fences where supported/appropriate.
2571   unsigned SType = 0;
2572   SDLoc DL(Op);
2573   return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2574                      DAG.getConstant(SType, DL, MVT::i32));
2575 }
2576 
2577 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
2578                                                 SelectionDAG &DAG) const {
2579   SDLoc DL(Op);
2580   MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2581 
2582   SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2583   SDValue Shamt = Op.getOperand(2);
2584   // if shamt < (VT.bits):
2585   //  lo = (shl lo, shamt)
2586   //  hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2587   // else:
2588   //  lo = 0
2589   //  hi = (shl lo, shamt[4:0])
2590   SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2591                             DAG.getConstant(-1, DL, MVT::i32));
2592   SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo,
2593                                       DAG.getConstant(1, DL, VT));
2594   SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not);
2595   SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
2596   SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2597   SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
2598   SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2599                              DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2600   Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2601                    DAG.getConstant(0, DL, VT), ShiftLeftLo);
2602   Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or);
2603 
2604   SDValue Ops[2] = {Lo, Hi};
2605   return DAG.getMergeValues(Ops, DL);
2606 }
2607 
2608 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2609                                                  bool IsSRA) const {
2610   SDLoc DL(Op);
2611   SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2612   SDValue Shamt = Op.getOperand(2);
2613   MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2614 
2615   // if shamt < (VT.bits):
2616   //  lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2617   //  if isSRA:
2618   //    hi = (sra hi, shamt)
2619   //  else:
2620   //    hi = (srl hi, shamt)
2621   // else:
2622   //  if isSRA:
2623   //   lo = (sra hi, shamt[4:0])
2624   //   hi = (sra hi, 31)
2625   //  else:
2626   //   lo = (srl hi, shamt[4:0])
2627   //   hi = 0
2628   SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2629                             DAG.getConstant(-1, DL, MVT::i32));
2630   SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi,
2631                                      DAG.getConstant(1, DL, VT));
2632   SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not);
2633   SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
2634   SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2635   SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL,
2636                                      DL, VT, Hi, Shamt);
2637   SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2638                              DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2639   SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi,
2640                             DAG.getConstant(VT.getSizeInBits() - 1, DL, VT));
2641 
2642   if (!(Subtarget.hasMips4() || Subtarget.hasMips32())) {
2643     SDVTList VTList = DAG.getVTList(VT, VT);
2644     return DAG.getNode(Subtarget.isGP64bit() ? Mips::PseudoD_SELECT_I64
2645                                              : Mips::PseudoD_SELECT_I,
2646                        DL, VTList, Cond, ShiftRightHi,
2647                        IsSRA ? Ext : DAG.getConstant(0, DL, VT), Or,
2648                        ShiftRightHi);
2649   }
2650 
2651   Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or);
2652   Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2653                    IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi);
2654 
2655   SDValue Ops[2] = {Lo, Hi};
2656   return DAG.getMergeValues(Ops, DL);
2657 }
2658 
2659 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2660                             SDValue Chain, SDValue Src, unsigned Offset) {
2661   SDValue Ptr = LD->getBasePtr();
2662   EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2663   EVT BasePtrVT = Ptr.getValueType();
2664   SDLoc DL(LD);
2665   SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2666 
2667   if (Offset)
2668     Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2669                       DAG.getConstant(Offset, DL, BasePtrVT));
2670 
2671   SDValue Ops[] = { Chain, Ptr, Src };
2672   return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2673                                  LD->getMemOperand());
2674 }
2675 
2676 // Expand an unaligned 32 or 64-bit integer load node.
2677 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2678   LoadSDNode *LD = cast<LoadSDNode>(Op);
2679   EVT MemVT = LD->getMemoryVT();
2680 
2681   if (Subtarget.systemSupportsUnalignedAccess())
2682     return Op;
2683 
2684   // Return if load is aligned or if MemVT is neither i32 nor i64.
2685   if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2686       ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2687     return SDValue();
2688 
2689   bool IsLittle = Subtarget.isLittle();
2690   EVT VT = Op.getValueType();
2691   ISD::LoadExtType ExtType = LD->getExtensionType();
2692   SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2693 
2694   assert((VT == MVT::i32) || (VT == MVT::i64));
2695 
2696   // Expand
2697   //  (set dst, (i64 (load baseptr)))
2698   // to
2699   //  (set tmp, (ldl (add baseptr, 7), undef))
2700   //  (set dst, (ldr baseptr, tmp))
2701   if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2702     SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2703                                IsLittle ? 7 : 0);
2704     return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2705                         IsLittle ? 0 : 7);
2706   }
2707 
2708   SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2709                              IsLittle ? 3 : 0);
2710   SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2711                              IsLittle ? 0 : 3);
2712 
2713   // Expand
2714   //  (set dst, (i32 (load baseptr))) or
2715   //  (set dst, (i64 (sextload baseptr))) or
2716   //  (set dst, (i64 (extload baseptr)))
2717   // to
2718   //  (set tmp, (lwl (add baseptr, 3), undef))
2719   //  (set dst, (lwr baseptr, tmp))
2720   if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2721       (ExtType == ISD::EXTLOAD))
2722     return LWR;
2723 
2724   assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2725 
2726   // Expand
2727   //  (set dst, (i64 (zextload baseptr)))
2728   // to
2729   //  (set tmp0, (lwl (add baseptr, 3), undef))
2730   //  (set tmp1, (lwr baseptr, tmp0))
2731   //  (set tmp2, (shl tmp1, 32))
2732   //  (set dst, (srl tmp2, 32))
2733   SDLoc DL(LD);
2734   SDValue Const32 = DAG.getConstant(32, DL, MVT::i32);
2735   SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2736   SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2737   SDValue Ops[] = { SRL, LWR.getValue(1) };
2738   return DAG.getMergeValues(Ops, DL);
2739 }
2740 
2741 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2742                              SDValue Chain, unsigned Offset) {
2743   SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2744   EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2745   SDLoc DL(SD);
2746   SDVTList VTList = DAG.getVTList(MVT::Other);
2747 
2748   if (Offset)
2749     Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2750                       DAG.getConstant(Offset, DL, BasePtrVT));
2751 
2752   SDValue Ops[] = { Chain, Value, Ptr };
2753   return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2754                                  SD->getMemOperand());
2755 }
2756 
2757 // Expand an unaligned 32 or 64-bit integer store node.
2758 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2759                                       bool IsLittle) {
2760   SDValue Value = SD->getValue(), Chain = SD->getChain();
2761   EVT VT = Value.getValueType();
2762 
2763   // Expand
2764   //  (store val, baseptr) or
2765   //  (truncstore val, baseptr)
2766   // to
2767   //  (swl val, (add baseptr, 3))
2768   //  (swr val, baseptr)
2769   if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2770     SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2771                                 IsLittle ? 3 : 0);
2772     return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2773   }
2774 
2775   assert(VT == MVT::i64);
2776 
2777   // Expand
2778   //  (store val, baseptr)
2779   // to
2780   //  (sdl val, (add baseptr, 7))
2781   //  (sdr val, baseptr)
2782   SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2783   return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2784 }
2785 
2786 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
2787 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG,
2788                                      bool SingleFloat) {
2789   SDValue Val = SD->getValue();
2790 
2791   if (Val.getOpcode() != ISD::FP_TO_SINT ||
2792       (Val.getValueSizeInBits() > 32 && SingleFloat))
2793     return SDValue();
2794 
2795   EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2796   SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2797                            Val.getOperand(0));
2798   return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2799                       SD->getPointerInfo(), SD->getAlignment(),
2800                       SD->getMemOperand()->getFlags());
2801 }
2802 
2803 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2804   StoreSDNode *SD = cast<StoreSDNode>(Op);
2805   EVT MemVT = SD->getMemoryVT();
2806 
2807   // Lower unaligned integer stores.
2808   if (!Subtarget.systemSupportsUnalignedAccess() &&
2809       (SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2810       ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2811     return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle());
2812 
2813   return lowerFP_TO_SINT_STORE(SD, DAG, Subtarget.isSingleFloat());
2814 }
2815 
2816 SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op,
2817                                               SelectionDAG &DAG) const {
2818 
2819   // Return a fixed StackObject with offset 0 which points to the old stack
2820   // pointer.
2821   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2822   EVT ValTy = Op->getValueType(0);
2823   int FI = MFI.CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2824   return DAG.getFrameIndex(FI, ValTy);
2825 }
2826 
2827 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2828                                             SelectionDAG &DAG) const {
2829   if (Op.getValueSizeInBits() > 32 && Subtarget.isSingleFloat())
2830     return SDValue();
2831 
2832   EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2833   SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2834                               Op.getOperand(0));
2835   return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2836 }
2837 
2838 //===----------------------------------------------------------------------===//
2839 //                      Calling Convention Implementation
2840 //===----------------------------------------------------------------------===//
2841 
2842 //===----------------------------------------------------------------------===//
2843 // TODO: Implement a generic logic using tblgen that can support this.
2844 // Mips O32 ABI rules:
2845 // ---
2846 // i32 - Passed in A0, A1, A2, A3 and stack
2847 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2848 //       an argument. Otherwise, passed in A1, A2, A3 and stack.
2849 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2850 //       yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2851 //       not used, it must be shadowed. If only A3 is available, shadow it and
2852 //       go to stack.
2853 // vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack.
2854 // vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3}
2855 //         with the remainder spilled to the stack.
2856 // vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases
2857 //         spilling the remainder to the stack.
2858 //
2859 //  For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2860 //===----------------------------------------------------------------------===//
2861 
2862 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2863                        CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2864                        CCState &State, ArrayRef<MCPhysReg> F64Regs) {
2865   const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>(
2866       State.getMachineFunction().getSubtarget());
2867 
2868   static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 };
2869 
2870   const MipsCCState * MipsState = static_cast<MipsCCState *>(&State);
2871 
2872   static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 };
2873 
2874   static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 };
2875 
2876   // Do not process byval args here.
2877   if (ArgFlags.isByVal())
2878     return true;
2879 
2880   // Promote i8 and i16
2881   if (ArgFlags.isInReg() && !Subtarget.isLittle()) {
2882     if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) {
2883       LocVT = MVT::i32;
2884       if (ArgFlags.isSExt())
2885         LocInfo = CCValAssign::SExtUpper;
2886       else if (ArgFlags.isZExt())
2887         LocInfo = CCValAssign::ZExtUpper;
2888       else
2889         LocInfo = CCValAssign::AExtUpper;
2890     }
2891   }
2892 
2893   // Promote i8 and i16
2894   if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2895     LocVT = MVT::i32;
2896     if (ArgFlags.isSExt())
2897       LocInfo = CCValAssign::SExt;
2898     else if (ArgFlags.isZExt())
2899       LocInfo = CCValAssign::ZExt;
2900     else
2901       LocInfo = CCValAssign::AExt;
2902   }
2903 
2904   unsigned Reg;
2905 
2906   // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2907   // is true: function is vararg, argument is 3rd or higher, there is previous
2908   // argument which is not f32 or f64.
2909   bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 ||
2910                                 State.getFirstUnallocated(F32Regs) != ValNo;
2911   unsigned OrigAlign = ArgFlags.getOrigAlign();
2912   bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2913   bool isVectorFloat = MipsState->WasOriginalArgVectorFloat(ValNo);
2914 
2915   // The MIPS vector ABI for floats passes them in a pair of registers
2916   if (ValVT == MVT::i32 && isVectorFloat) {
2917     // This is the start of an vector that was scalarized into an unknown number
2918     // of components. It doesn't matter how many there are. Allocate one of the
2919     // notional 8 byte aligned registers which map onto the argument stack, and
2920     // shadow the register lost to alignment requirements.
2921     if (ArgFlags.isSplit()) {
2922       Reg = State.AllocateReg(FloatVectorIntRegs);
2923       if (Reg == Mips::A2)
2924         State.AllocateReg(Mips::A1);
2925       else if (Reg == 0)
2926         State.AllocateReg(Mips::A3);
2927     } else {
2928       // If we're an intermediate component of the split, we can just attempt to
2929       // allocate a register directly.
2930       Reg = State.AllocateReg(IntRegs);
2931     }
2932   } else if (ValVT == MVT::i32 ||
2933              (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2934     Reg = State.AllocateReg(IntRegs);
2935     // If this is the first part of an i64 arg,
2936     // the allocated register must be either A0 or A2.
2937     if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2938       Reg = State.AllocateReg(IntRegs);
2939     LocVT = MVT::i32;
2940   } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2941     // Allocate int register and shadow next int register. If first
2942     // available register is Mips::A1 or Mips::A3, shadow it too.
2943     Reg = State.AllocateReg(IntRegs);
2944     if (Reg == Mips::A1 || Reg == Mips::A3)
2945       Reg = State.AllocateReg(IntRegs);
2946     State.AllocateReg(IntRegs);
2947     LocVT = MVT::i32;
2948   } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2949     // we are guaranteed to find an available float register
2950     if (ValVT == MVT::f32) {
2951       Reg = State.AllocateReg(F32Regs);
2952       // Shadow int register
2953       State.AllocateReg(IntRegs);
2954     } else {
2955       Reg = State.AllocateReg(F64Regs);
2956       // Shadow int registers
2957       unsigned Reg2 = State.AllocateReg(IntRegs);
2958       if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2959         State.AllocateReg(IntRegs);
2960       State.AllocateReg(IntRegs);
2961     }
2962   } else
2963     llvm_unreachable("Cannot handle this ValVT.");
2964 
2965   if (!Reg) {
2966     unsigned Offset = State.AllocateStack(ValVT.getStoreSize(), OrigAlign);
2967     State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2968   } else
2969     State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2970 
2971   return false;
2972 }
2973 
2974 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2975                             MVT LocVT, CCValAssign::LocInfo LocInfo,
2976                             ISD::ArgFlagsTy ArgFlags, CCState &State) {
2977   static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 };
2978 
2979   return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2980 }
2981 
2982 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2983                             MVT LocVT, CCValAssign::LocInfo LocInfo,
2984                             ISD::ArgFlagsTy ArgFlags, CCState &State) {
2985   static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 };
2986 
2987   return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2988 }
2989 
2990 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2991                        CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2992                        CCState &State) LLVM_ATTRIBUTE_UNUSED;
2993 
2994 #include "MipsGenCallingConv.inc"
2995 
2996  CCAssignFn *MipsTargetLowering::CCAssignFnForCall() const{
2997    return CC_Mips_FixedArg;
2998  }
2999 
3000  CCAssignFn *MipsTargetLowering::CCAssignFnForReturn() const{
3001    return RetCC_Mips;
3002  }
3003 //===----------------------------------------------------------------------===//
3004 //                  Call Calling Convention Implementation
3005 //===----------------------------------------------------------------------===//
3006 
3007 // Return next O32 integer argument register.
3008 static unsigned getNextIntArgReg(unsigned Reg) {
3009   assert((Reg == Mips::A0) || (Reg == Mips::A2));
3010   return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
3011 }
3012 
3013 SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
3014                                            SDValue Chain, SDValue Arg,
3015                                            const SDLoc &DL, bool IsTailCall,
3016                                            SelectionDAG &DAG) const {
3017   if (!IsTailCall) {
3018     SDValue PtrOff =
3019         DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr,
3020                     DAG.getIntPtrConstant(Offset, DL));
3021     return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo());
3022   }
3023 
3024   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
3025   int FI = MFI.CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
3026   SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3027   return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
3028                       /* Alignment = */ 0, MachineMemOperand::MOVolatile);
3029 }
3030 
3031 void MipsTargetLowering::
3032 getOpndList(SmallVectorImpl<SDValue> &Ops,
3033             std::deque<std::pair<unsigned, SDValue>> &RegsToPass,
3034             bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
3035             bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
3036             SDValue Chain) const {
3037   // Insert node "GP copy globalreg" before call to function.
3038   //
3039   // R_MIPS_CALL* operators (emitted when non-internal functions are called
3040   // in PIC mode) allow symbols to be resolved via lazy binding.
3041   // The lazy binding stub requires GP to point to the GOT.
3042   // Note that we don't need GP to point to the GOT for indirect calls
3043   // (when R_MIPS_CALL* is not used for the call) because Mips linker generates
3044   // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs
3045   // used for the function (that is, Mips linker doesn't generate lazy binding
3046   // stub for a function whose address is taken in the program).
3047   if (IsPICCall && !InternalLinkage && IsCallReloc) {
3048     unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP;
3049     EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
3050     RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
3051   }
3052 
3053   // Build a sequence of copy-to-reg nodes chained together with token
3054   // chain and flag operands which copy the outgoing args into registers.
3055   // The InFlag in necessary since all emitted instructions must be
3056   // stuck together.
3057   SDValue InFlag;
3058 
3059   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3060     Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
3061                                  RegsToPass[i].second, InFlag);
3062     InFlag = Chain.getValue(1);
3063   }
3064 
3065   // Add argument registers to the end of the list so that they are
3066   // known live into the call.
3067   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3068     Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
3069                                       RegsToPass[i].second.getValueType()));
3070 
3071   // Add a register mask operand representing the call-preserved registers.
3072   const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
3073   const uint32_t *Mask =
3074       TRI->getCallPreservedMask(CLI.DAG.getMachineFunction(), CLI.CallConv);
3075   assert(Mask && "Missing call preserved mask for calling convention");
3076   if (Subtarget.inMips16HardFloat()) {
3077     if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
3078       StringRef Sym = G->getGlobal()->getName();
3079       Function *F = G->getGlobal()->getParent()->getFunction(Sym);
3080       if (F && F->hasFnAttribute("__Mips16RetHelper")) {
3081         Mask = MipsRegisterInfo::getMips16RetHelperMask();
3082       }
3083     }
3084   }
3085   Ops.push_back(CLI.DAG.getRegisterMask(Mask));
3086 
3087   if (InFlag.getNode())
3088     Ops.push_back(InFlag);
3089 }
3090 
3091 void MipsTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
3092                                                        SDNode *Node) const {
3093   switch (MI.getOpcode()) {
3094     default:
3095       return;
3096     case Mips::JALR:
3097     case Mips::JALRPseudo:
3098     case Mips::JALR64:
3099     case Mips::JALR64Pseudo:
3100     case Mips::JALR16_MM:
3101     case Mips::JALRC16_MMR6:
3102     case Mips::TAILCALLREG:
3103     case Mips::TAILCALLREG64:
3104     case Mips::TAILCALLR6REG:
3105     case Mips::TAILCALL64R6REG:
3106     case Mips::TAILCALLREG_MM:
3107     case Mips::TAILCALLREG_MMR6: {
3108       if (!EmitJalrReloc ||
3109           Subtarget.inMips16Mode() ||
3110           !isPositionIndependent() ||
3111           Node->getNumOperands() < 1 ||
3112           Node->getOperand(0).getNumOperands() < 2) {
3113         return;
3114       }
3115       // We are after the callee address, set by LowerCall().
3116       // If added to MI, asm printer will emit .reloc R_MIPS_JALR for the
3117       // symbol.
3118       const SDValue TargetAddr = Node->getOperand(0).getOperand(1);
3119       StringRef Sym;
3120       if (const GlobalAddressSDNode *G =
3121               dyn_cast_or_null<const GlobalAddressSDNode>(TargetAddr)) {
3122         // We must not emit the R_MIPS_JALR relocation against data symbols
3123         // since this will cause run-time crashes if the linker replaces the
3124         // call instruction with a relative branch to the data symbol.
3125         if (!isa<Function>(G->getGlobal())) {
3126           LLVM_DEBUG(dbgs() << "Not adding R_MIPS_JALR against data symbol "
3127                             << G->getGlobal()->getName() << "\n");
3128           return;
3129         }
3130         Sym = G->getGlobal()->getName();
3131       }
3132       else if (const ExternalSymbolSDNode *ES =
3133                    dyn_cast_or_null<const ExternalSymbolSDNode>(TargetAddr)) {
3134         Sym = ES->getSymbol();
3135       }
3136 
3137       if (Sym.empty())
3138         return;
3139 
3140       MachineFunction *MF = MI.getParent()->getParent();
3141       MCSymbol *S = MF->getContext().getOrCreateSymbol(Sym);
3142       LLVM_DEBUG(dbgs() << "Adding R_MIPS_JALR against " << Sym << "\n");
3143       MI.addOperand(MachineOperand::CreateMCSymbol(S, MipsII::MO_JALR));
3144     }
3145   }
3146 }
3147 
3148 /// LowerCall - functions arguments are copied from virtual regs to
3149 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
3150 SDValue
3151 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3152                               SmallVectorImpl<SDValue> &InVals) const {
3153   SelectionDAG &DAG                     = CLI.DAG;
3154   SDLoc DL                              = CLI.DL;
3155   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3156   SmallVectorImpl<SDValue> &OutVals     = CLI.OutVals;
3157   SmallVectorImpl<ISD::InputArg> &Ins   = CLI.Ins;
3158   SDValue Chain                         = CLI.Chain;
3159   SDValue Callee                        = CLI.Callee;
3160   bool &IsTailCall                      = CLI.IsTailCall;
3161   CallingConv::ID CallConv              = CLI.CallConv;
3162   bool IsVarArg                         = CLI.IsVarArg;
3163 
3164   MachineFunction &MF = DAG.getMachineFunction();
3165   MachineFrameInfo &MFI = MF.getFrameInfo();
3166   const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
3167   MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
3168   bool IsPIC = isPositionIndependent();
3169 
3170   // Analyze operands of the call, assigning locations to each operand.
3171   SmallVector<CCValAssign, 16> ArgLocs;
3172   MipsCCState CCInfo(
3173       CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(),
3174       MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget));
3175 
3176   const ExternalSymbolSDNode *ES =
3177       dyn_cast_or_null<const ExternalSymbolSDNode>(Callee.getNode());
3178 
3179   // There is one case where CALLSEQ_START..CALLSEQ_END can be nested, which
3180   // is during the lowering of a call with a byval argument which produces
3181   // a call to memcpy. For the O32 case, this causes the caller to allocate
3182   // stack space for the reserved argument area for the callee, then recursively
3183   // again for the memcpy call. In the NEWABI case, this doesn't occur as those
3184   // ABIs mandate that the callee allocates the reserved argument area. We do
3185   // still produce nested CALLSEQ_START..CALLSEQ_END with zero space though.
3186   //
3187   // If the callee has a byval argument and memcpy is used, we are mandated
3188   // to already have produced a reserved argument area for the callee for O32.
3189   // Therefore, the reserved argument area can be reused for both calls.
3190   //
3191   // Other cases of calling memcpy cannot have a chain with a CALLSEQ_START
3192   // present, as we have yet to hook that node onto the chain.
3193   //
3194   // Hence, the CALLSEQ_START and CALLSEQ_END nodes can be eliminated in this
3195   // case. GCC does a similar trick, in that wherever possible, it calculates
3196   // the maximum out going argument area (including the reserved area), and
3197   // preallocates the stack space on entrance to the caller.
3198   //
3199   // FIXME: We should do the same for efficiency and space.
3200 
3201   // Note: The check on the calling convention below must match
3202   //       MipsABIInfo::GetCalleeAllocdArgSizeInBytes().
3203   bool MemcpyInByVal = ES &&
3204                        StringRef(ES->getSymbol()) == StringRef("memcpy") &&
3205                        CallConv != CallingConv::Fast &&
3206                        Chain.getOpcode() == ISD::CALLSEQ_START;
3207 
3208   // Allocate the reserved argument area. It seems strange to do this from the
3209   // caller side but removing it breaks the frame size calculation.
3210   unsigned ReservedArgArea =
3211       MemcpyInByVal ? 0 : ABI.GetCalleeAllocdArgSizeInBytes(CallConv);
3212   CCInfo.AllocateStack(ReservedArgArea, 1);
3213 
3214   CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(),
3215                              ES ? ES->getSymbol() : nullptr);
3216 
3217   // Get a count of how many bytes are to be pushed on the stack.
3218   unsigned NextStackOffset = CCInfo.getNextStackOffset();
3219 
3220   // Call site info for function parameters tracking.
3221   MachineFunction::CallSiteInfo CSInfo;
3222 
3223   // Check if it's really possible to do a tail call. Restrict it to functions
3224   // that are part of this compilation unit.
3225   bool InternalLinkage = false;
3226   if (IsTailCall) {
3227     IsTailCall = isEligibleForTailCallOptimization(
3228         CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>());
3229      if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3230       InternalLinkage = G->getGlobal()->hasInternalLinkage();
3231       IsTailCall &= (InternalLinkage || G->getGlobal()->hasLocalLinkage() ||
3232                      G->getGlobal()->hasPrivateLinkage() ||
3233                      G->getGlobal()->hasHiddenVisibility() ||
3234                      G->getGlobal()->hasProtectedVisibility());
3235      }
3236   }
3237   if (!IsTailCall && CLI.CB && CLI.CB->isMustTailCall())
3238     report_fatal_error("failed to perform tail call elimination on a call "
3239                        "site marked musttail");
3240 
3241   if (IsTailCall)
3242     ++NumTailCalls;
3243 
3244   // Chain is the output chain of the last Load/Store or CopyToReg node.
3245   // ByValChain is the output chain of the last Memcpy node created for copying
3246   // byval arguments to the stack.
3247   unsigned StackAlignment = TFL->getStackAlignment();
3248   NextStackOffset = alignTo(NextStackOffset, StackAlignment);
3249   SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, DL, true);
3250 
3251   if (!(IsTailCall || MemcpyInByVal))
3252     Chain = DAG.getCALLSEQ_START(Chain, NextStackOffset, 0, DL);
3253 
3254   SDValue StackPtr =
3255       DAG.getCopyFromReg(Chain, DL, ABI.IsN64() ? Mips::SP_64 : Mips::SP,
3256                          getPointerTy(DAG.getDataLayout()));
3257 
3258   std::deque<std::pair<unsigned, SDValue>> RegsToPass;
3259   SmallVector<SDValue, 8> MemOpChains;
3260 
3261   CCInfo.rewindByValRegsInfo();
3262 
3263   // Walk the register/memloc assignments, inserting copies/loads.
3264   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3265     SDValue Arg = OutVals[i];
3266     CCValAssign &VA = ArgLocs[i];
3267     MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
3268     ISD::ArgFlagsTy Flags = Outs[i].Flags;
3269     bool UseUpperBits = false;
3270 
3271     // ByVal Arg.
3272     if (Flags.isByVal()) {
3273       unsigned FirstByValReg, LastByValReg;
3274       unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
3275       CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
3276 
3277       assert(Flags.getByValSize() &&
3278              "ByVal args of size 0 should have been ignored by front-end.");
3279       assert(ByValIdx < CCInfo.getInRegsParamsCount());
3280       assert(!IsTailCall &&
3281              "Do not tail-call optimize if there is a byval argument.");
3282       passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
3283                    FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(),
3284                    VA);
3285       CCInfo.nextInRegsParam();
3286       continue;
3287     }
3288 
3289     // Promote the value if needed.
3290     switch (VA.getLocInfo()) {
3291     default:
3292       llvm_unreachable("Unknown loc info!");
3293     case CCValAssign::Full:
3294       if (VA.isRegLoc()) {
3295         if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
3296             (ValVT == MVT::f64 && LocVT == MVT::i64) ||
3297             (ValVT == MVT::i64 && LocVT == MVT::f64))
3298           Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
3299         else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
3300           SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
3301                                    Arg, DAG.getConstant(0, DL, MVT::i32));
3302           SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
3303                                    Arg, DAG.getConstant(1, DL, MVT::i32));
3304           if (!Subtarget.isLittle())
3305             std::swap(Lo, Hi);
3306           Register LocRegLo = VA.getLocReg();
3307           unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
3308           RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
3309           RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
3310           continue;
3311         }
3312       }
3313       break;
3314     case CCValAssign::BCvt:
3315       Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
3316       break;
3317     case CCValAssign::SExtUpper:
3318       UseUpperBits = true;
3319       LLVM_FALLTHROUGH;
3320     case CCValAssign::SExt:
3321       Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
3322       break;
3323     case CCValAssign::ZExtUpper:
3324       UseUpperBits = true;
3325       LLVM_FALLTHROUGH;
3326     case CCValAssign::ZExt:
3327       Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
3328       break;
3329     case CCValAssign::AExtUpper:
3330       UseUpperBits = true;
3331       LLVM_FALLTHROUGH;
3332     case CCValAssign::AExt:
3333       Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
3334       break;
3335     }
3336 
3337     if (UseUpperBits) {
3338       unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
3339       unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3340       Arg = DAG.getNode(
3341           ISD::SHL, DL, VA.getLocVT(), Arg,
3342           DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3343     }
3344 
3345     // Arguments that can be passed on register must be kept at
3346     // RegsToPass vector
3347     if (VA.isRegLoc()) {
3348       RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3349 
3350       // If the parameter is passed through reg $D, which splits into
3351       // two physical registers, avoid creating call site info.
3352       if (Mips::AFGR64RegClass.contains(VA.getLocReg()))
3353         continue;
3354 
3355       // Collect CSInfo about which register passes which parameter.
3356       const TargetOptions &Options = DAG.getTarget().Options;
3357       if (Options.SupportsDebugEntryValues)
3358         CSInfo.emplace_back(VA.getLocReg(), i);
3359 
3360       continue;
3361     }
3362 
3363     // Register can't get to this point...
3364     assert(VA.isMemLoc());
3365 
3366     // emit ISD::STORE whichs stores the
3367     // parameter value to a stack Location
3368     MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
3369                                          Chain, Arg, DL, IsTailCall, DAG));
3370   }
3371 
3372   // Transform all store nodes into one single node because all store
3373   // nodes are independent of each other.
3374   if (!MemOpChains.empty())
3375     Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
3376 
3377   // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3378   // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3379   // node so that legalize doesn't hack it.
3380 
3381   EVT Ty = Callee.getValueType();
3382   bool GlobalOrExternal = false, IsCallReloc = false;
3383 
3384   // The long-calls feature is ignored in case of PIC.
3385   // While we do not support -mshared / -mno-shared properly,
3386   // ignore long-calls in case of -mabicalls too.
3387   if (!Subtarget.isABICalls() && !IsPIC) {
3388     // If the function should be called using "long call",
3389     // get its address into a register to prevent using
3390     // of the `jal` instruction for the direct call.
3391     if (auto *N = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3392       if (Subtarget.useLongCalls())
3393         Callee = Subtarget.hasSym32()
3394                      ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
3395                      : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
3396     } else if (auto *N = dyn_cast<GlobalAddressSDNode>(Callee)) {
3397       bool UseLongCalls = Subtarget.useLongCalls();
3398       // If the function has long-call/far/near attribute
3399       // it overrides command line switch pased to the backend.
3400       if (auto *F = dyn_cast<Function>(N->getGlobal())) {
3401         if (F->hasFnAttribute("long-call"))
3402           UseLongCalls = true;
3403         else if (F->hasFnAttribute("short-call"))
3404           UseLongCalls = false;
3405       }
3406       if (UseLongCalls)
3407         Callee = Subtarget.hasSym32()
3408                      ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
3409                      : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
3410     }
3411   }
3412 
3413   if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3414     if (IsPIC) {
3415       const GlobalValue *Val = G->getGlobal();
3416       InternalLinkage = Val->hasInternalLinkage();
3417 
3418       if (InternalLinkage)
3419         Callee = getAddrLocal(G, DL, Ty, DAG, ABI.IsN32() || ABI.IsN64());
3420       else if (Subtarget.useXGOT()) {
3421         Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16,
3422                                        MipsII::MO_CALL_LO16, Chain,
3423                                        FuncInfo->callPtrInfo(Val));
3424         IsCallReloc = true;
3425       } else {
3426         Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
3427                                FuncInfo->callPtrInfo(Val));
3428         IsCallReloc = true;
3429       }
3430     } else
3431       Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL,
3432                                           getPointerTy(DAG.getDataLayout()), 0,
3433                                           MipsII::MO_NO_FLAG);
3434     GlobalOrExternal = true;
3435   }
3436   else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3437     const char *Sym = S->getSymbol();
3438 
3439     if (!IsPIC) // static
3440       Callee = DAG.getTargetExternalSymbol(
3441           Sym, getPointerTy(DAG.getDataLayout()), MipsII::MO_NO_FLAG);
3442     else if (Subtarget.useXGOT()) {
3443       Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16,
3444                                      MipsII::MO_CALL_LO16, Chain,
3445                                      FuncInfo->callPtrInfo(Sym));
3446       IsCallReloc = true;
3447     } else { // PIC
3448       Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
3449                              FuncInfo->callPtrInfo(Sym));
3450       IsCallReloc = true;
3451     }
3452 
3453     GlobalOrExternal = true;
3454   }
3455 
3456   SmallVector<SDValue, 8> Ops(1, Chain);
3457   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3458 
3459   getOpndList(Ops, RegsToPass, IsPIC, GlobalOrExternal, InternalLinkage,
3460               IsCallReloc, CLI, Callee, Chain);
3461 
3462   if (IsTailCall) {
3463     MF.getFrameInfo().setHasTailCall();
3464     SDValue Ret = DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops);
3465     DAG.addCallSiteInfo(Ret.getNode(), std::move(CSInfo));
3466     return Ret;
3467   }
3468 
3469   Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops);
3470   SDValue InFlag = Chain.getValue(1);
3471 
3472   DAG.addCallSiteInfo(Chain.getNode(), std::move(CSInfo));
3473 
3474   // Create the CALLSEQ_END node in the case of where it is not a call to
3475   // memcpy.
3476   if (!(MemcpyInByVal)) {
3477     Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
3478                                DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
3479     InFlag = Chain.getValue(1);
3480   }
3481 
3482   // Handle result values, copying them out of physregs into vregs that we
3483   // return.
3484   return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
3485                          InVals, CLI);
3486 }
3487 
3488 /// LowerCallResult - Lower the result values of a call into the
3489 /// appropriate copies out of appropriate physical registers.
3490 SDValue MipsTargetLowering::LowerCallResult(
3491     SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
3492     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
3493     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
3494     TargetLowering::CallLoweringInfo &CLI) const {
3495   // Assign locations to each value returned by this call.
3496   SmallVector<CCValAssign, 16> RVLocs;
3497   MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
3498                      *DAG.getContext());
3499 
3500   const ExternalSymbolSDNode *ES =
3501       dyn_cast_or_null<const ExternalSymbolSDNode>(CLI.Callee.getNode());
3502   CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI.RetTy,
3503                            ES ? ES->getSymbol() : nullptr);
3504 
3505   // Copy all of the result registers out of their specified physreg.
3506   for (unsigned i = 0; i != RVLocs.size(); ++i) {
3507     CCValAssign &VA = RVLocs[i];
3508     assert(VA.isRegLoc() && "Can only return in registers!");
3509 
3510     SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
3511                                      RVLocs[i].getLocVT(), InFlag);
3512     Chain = Val.getValue(1);
3513     InFlag = Val.getValue(2);
3514 
3515     if (VA.isUpperBitsInLoc()) {
3516       unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits();
3517       unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3518       unsigned Shift =
3519           VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
3520       Val = DAG.getNode(
3521           Shift, DL, VA.getLocVT(), Val,
3522           DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3523     }
3524 
3525     switch (VA.getLocInfo()) {
3526     default:
3527       llvm_unreachable("Unknown loc info!");
3528     case CCValAssign::Full:
3529       break;
3530     case CCValAssign::BCvt:
3531       Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
3532       break;
3533     case CCValAssign::AExt:
3534     case CCValAssign::AExtUpper:
3535       Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
3536       break;
3537     case CCValAssign::ZExt:
3538     case CCValAssign::ZExtUpper:
3539       Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val,
3540                         DAG.getValueType(VA.getValVT()));
3541       Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
3542       break;
3543     case CCValAssign::SExt:
3544     case CCValAssign::SExtUpper:
3545       Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val,
3546                         DAG.getValueType(VA.getValVT()));
3547       Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
3548       break;
3549     }
3550 
3551     InVals.push_back(Val);
3552   }
3553 
3554   return Chain;
3555 }
3556 
3557 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA,
3558                                       EVT ArgVT, const SDLoc &DL,
3559                                       SelectionDAG &DAG) {
3560   MVT LocVT = VA.getLocVT();
3561   EVT ValVT = VA.getValVT();
3562 
3563   // Shift into the upper bits if necessary.
3564   switch (VA.getLocInfo()) {
3565   default:
3566     break;
3567   case CCValAssign::AExtUpper:
3568   case CCValAssign::SExtUpper:
3569   case CCValAssign::ZExtUpper: {
3570     unsigned ValSizeInBits = ArgVT.getSizeInBits();
3571     unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3572     unsigned Opcode =
3573         VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
3574     Val = DAG.getNode(
3575         Opcode, DL, VA.getLocVT(), Val,
3576         DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3577     break;
3578   }
3579   }
3580 
3581   // If this is an value smaller than the argument slot size (32-bit for O32,
3582   // 64-bit for N32/N64), it has been promoted in some way to the argument slot
3583   // size. Extract the value and insert any appropriate assertions regarding
3584   // sign/zero extension.
3585   switch (VA.getLocInfo()) {
3586   default:
3587     llvm_unreachable("Unknown loc info!");
3588   case CCValAssign::Full:
3589     break;
3590   case CCValAssign::AExtUpper:
3591   case CCValAssign::AExt:
3592     Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
3593     break;
3594   case CCValAssign::SExtUpper:
3595   case CCValAssign::SExt:
3596     Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT));
3597     Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
3598     break;
3599   case CCValAssign::ZExtUpper:
3600   case CCValAssign::ZExt:
3601     Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT));
3602     Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
3603     break;
3604   case CCValAssign::BCvt:
3605     Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val);
3606     break;
3607   }
3608 
3609   return Val;
3610 }
3611 
3612 //===----------------------------------------------------------------------===//
3613 //             Formal Arguments Calling Convention Implementation
3614 //===----------------------------------------------------------------------===//
3615 /// LowerFormalArguments - transform physical registers into virtual registers
3616 /// and generate load operations for arguments places on the stack.
3617 SDValue MipsTargetLowering::LowerFormalArguments(
3618     SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
3619     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
3620     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
3621   MachineFunction &MF = DAG.getMachineFunction();
3622   MachineFrameInfo &MFI = MF.getFrameInfo();
3623   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3624 
3625   MipsFI->setVarArgsFrameIndex(0);
3626 
3627   // Used with vargs to acumulate store chains.
3628   std::vector<SDValue> OutChains;
3629 
3630   // Assign locations to all of the incoming arguments.
3631   SmallVector<CCValAssign, 16> ArgLocs;
3632   MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
3633                      *DAG.getContext());
3634   CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1);
3635   const Function &Func = DAG.getMachineFunction().getFunction();
3636   Function::const_arg_iterator FuncArg = Func.arg_begin();
3637 
3638   if (Func.hasFnAttribute("interrupt") && !Func.arg_empty())
3639     report_fatal_error(
3640         "Functions with the interrupt attribute cannot have arguments!");
3641 
3642   CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg);
3643   MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
3644                            CCInfo.getInRegsParamsCount() > 0);
3645 
3646   unsigned CurArgIdx = 0;
3647   CCInfo.rewindByValRegsInfo();
3648 
3649   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3650     CCValAssign &VA = ArgLocs[i];
3651     if (Ins[i].isOrigArg()) {
3652       std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx);
3653       CurArgIdx = Ins[i].getOrigArgIndex();
3654     }
3655     EVT ValVT = VA.getValVT();
3656     ISD::ArgFlagsTy Flags = Ins[i].Flags;
3657     bool IsRegLoc = VA.isRegLoc();
3658 
3659     if (Flags.isByVal()) {
3660       assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit");
3661       unsigned FirstByValReg, LastByValReg;
3662       unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
3663       CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
3664 
3665       assert(Flags.getByValSize() &&
3666              "ByVal args of size 0 should have been ignored by front-end.");
3667       assert(ByValIdx < CCInfo.getInRegsParamsCount());
3668       copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
3669                     FirstByValReg, LastByValReg, VA, CCInfo);
3670       CCInfo.nextInRegsParam();
3671       continue;
3672     }
3673 
3674     // Arguments stored on registers
3675     if (IsRegLoc) {
3676       MVT RegVT = VA.getLocVT();
3677       Register ArgReg = VA.getLocReg();
3678       const TargetRegisterClass *RC = getRegClassFor(RegVT);
3679 
3680       // Transform the arguments stored on
3681       // physical registers into virtual ones
3682       unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
3683       SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
3684 
3685       ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3686 
3687       // Handle floating point arguments passed in integer registers and
3688       // long double arguments passed in floating point registers.
3689       if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
3690           (RegVT == MVT::i64 && ValVT == MVT::f64) ||
3691           (RegVT == MVT::f64 && ValVT == MVT::i64))
3692         ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
3693       else if (ABI.IsO32() && RegVT == MVT::i32 &&
3694                ValVT == MVT::f64) {
3695         unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
3696                                   getNextIntArgReg(ArgReg), RC);
3697         SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
3698         if (!Subtarget.isLittle())
3699           std::swap(ArgValue, ArgValue2);
3700         ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
3701                                ArgValue, ArgValue2);
3702       }
3703 
3704       InVals.push_back(ArgValue);
3705     } else { // VA.isRegLoc()
3706       MVT LocVT = VA.getLocVT();
3707 
3708       if (ABI.IsO32()) {
3709         // We ought to be able to use LocVT directly but O32 sets it to i32
3710         // when allocating floating point values to integer registers.
3711         // This shouldn't influence how we load the value into registers unless
3712         // we are targeting softfloat.
3713         if (VA.getValVT().isFloatingPoint() && !Subtarget.useSoftFloat())
3714           LocVT = VA.getValVT();
3715       }
3716 
3717       // sanity check
3718       assert(VA.isMemLoc());
3719 
3720       // The stack pointer offset is relative to the caller stack frame.
3721       int FI = MFI.CreateFixedObject(LocVT.getSizeInBits() / 8,
3722                                      VA.getLocMemOffset(), true);
3723 
3724       // Create load nodes to retrieve arguments from the stack
3725       SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3726       SDValue ArgValue = DAG.getLoad(
3727           LocVT, DL, Chain, FIN,
3728           MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI));
3729       OutChains.push_back(ArgValue.getValue(1));
3730 
3731       ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3732 
3733       InVals.push_back(ArgValue);
3734     }
3735   }
3736 
3737   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3738     // The mips ABIs for returning structs by value requires that we copy
3739     // the sret argument into $v0 for the return. Save the argument into
3740     // a virtual register so that we can access it from the return points.
3741     if (Ins[i].Flags.isSRet()) {
3742       unsigned Reg = MipsFI->getSRetReturnReg();
3743       if (!Reg) {
3744         Reg = MF.getRegInfo().createVirtualRegister(
3745             getRegClassFor(ABI.IsN64() ? MVT::i64 : MVT::i32));
3746         MipsFI->setSRetReturnReg(Reg);
3747       }
3748       SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]);
3749       Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
3750       break;
3751     }
3752   }
3753 
3754   if (IsVarArg)
3755     writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo);
3756 
3757   // All stores are grouped in one node to allow the matching between
3758   // the size of Ins and InVals. This only happens when on varg functions
3759   if (!OutChains.empty()) {
3760     OutChains.push_back(Chain);
3761     Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
3762   }
3763 
3764   return Chain;
3765 }
3766 
3767 //===----------------------------------------------------------------------===//
3768 //               Return Value Calling Convention Implementation
3769 //===----------------------------------------------------------------------===//
3770 
3771 bool
3772 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3773                                    MachineFunction &MF, bool IsVarArg,
3774                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
3775                                    LLVMContext &Context) const {
3776   SmallVector<CCValAssign, 16> RVLocs;
3777   MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
3778   return CCInfo.CheckReturn(Outs, RetCC_Mips);
3779 }
3780 
3781 bool MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type,
3782                                                        bool IsSigned) const {
3783   if ((ABI.IsN32() || ABI.IsN64()) && Type == MVT::i32)
3784       return true;
3785 
3786   return IsSigned;
3787 }
3788 
3789 SDValue
3790 MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
3791                                          const SDLoc &DL,
3792                                          SelectionDAG &DAG) const {
3793   MachineFunction &MF = DAG.getMachineFunction();
3794   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3795 
3796   MipsFI->setISR();
3797 
3798   return DAG.getNode(MipsISD::ERet, DL, MVT::Other, RetOps);
3799 }
3800 
3801 SDValue
3802 MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3803                                 bool IsVarArg,
3804                                 const SmallVectorImpl<ISD::OutputArg> &Outs,
3805                                 const SmallVectorImpl<SDValue> &OutVals,
3806                                 const SDLoc &DL, SelectionDAG &DAG) const {
3807   // CCValAssign - represent the assignment of
3808   // the return value to a location
3809   SmallVector<CCValAssign, 16> RVLocs;
3810   MachineFunction &MF = DAG.getMachineFunction();
3811 
3812   // CCState - Info about the registers and stack slot.
3813   MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
3814 
3815   // Analyze return values.
3816   CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
3817 
3818   SDValue Flag;
3819   SmallVector<SDValue, 4> RetOps(1, Chain);
3820 
3821   // Copy the result values into the output registers.
3822   for (unsigned i = 0; i != RVLocs.size(); ++i) {
3823     SDValue Val = OutVals[i];
3824     CCValAssign &VA = RVLocs[i];
3825     assert(VA.isRegLoc() && "Can only return in registers!");
3826     bool UseUpperBits = false;
3827 
3828     switch (VA.getLocInfo()) {
3829     default:
3830       llvm_unreachable("Unknown loc info!");
3831     case CCValAssign::Full:
3832       break;
3833     case CCValAssign::BCvt:
3834       Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val);
3835       break;
3836     case CCValAssign::AExtUpper:
3837       UseUpperBits = true;
3838       LLVM_FALLTHROUGH;
3839     case CCValAssign::AExt:
3840       Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val);
3841       break;
3842     case CCValAssign::ZExtUpper:
3843       UseUpperBits = true;
3844       LLVM_FALLTHROUGH;
3845     case CCValAssign::ZExt:
3846       Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val);
3847       break;
3848     case CCValAssign::SExtUpper:
3849       UseUpperBits = true;
3850       LLVM_FALLTHROUGH;
3851     case CCValAssign::SExt:
3852       Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val);
3853       break;
3854     }
3855 
3856     if (UseUpperBits) {
3857       unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
3858       unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3859       Val = DAG.getNode(
3860           ISD::SHL, DL, VA.getLocVT(), Val,
3861           DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3862     }
3863 
3864     Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
3865 
3866     // Guarantee that all emitted copies are stuck together with flags.
3867     Flag = Chain.getValue(1);
3868     RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3869   }
3870 
3871   // The mips ABIs for returning structs by value requires that we copy
3872   // the sret argument into $v0 for the return. We saved the argument into
3873   // a virtual register in the entry block, so now we copy the value out
3874   // and into $v0.
3875   if (MF.getFunction().hasStructRetAttr()) {
3876     MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3877     unsigned Reg = MipsFI->getSRetReturnReg();
3878 
3879     if (!Reg)
3880       llvm_unreachable("sret virtual register not created in the entry block");
3881     SDValue Val =
3882         DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout()));
3883     unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
3884 
3885     Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
3886     Flag = Chain.getValue(1);
3887     RetOps.push_back(DAG.getRegister(V0, getPointerTy(DAG.getDataLayout())));
3888   }
3889 
3890   RetOps[0] = Chain;  // Update chain.
3891 
3892   // Add the flag if we have it.
3893   if (Flag.getNode())
3894     RetOps.push_back(Flag);
3895 
3896   // ISRs must use "eret".
3897   if (DAG.getMachineFunction().getFunction().hasFnAttribute("interrupt"))
3898     return LowerInterruptReturn(RetOps, DL, DAG);
3899 
3900   // Standard return on Mips is a "jr $ra"
3901   return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps);
3902 }
3903 
3904 //===----------------------------------------------------------------------===//
3905 //                           Mips Inline Assembly Support
3906 //===----------------------------------------------------------------------===//
3907 
3908 /// getConstraintType - Given a constraint letter, return the type of
3909 /// constraint it is for this target.
3910 MipsTargetLowering::ConstraintType
3911 MipsTargetLowering::getConstraintType(StringRef Constraint) const {
3912   // Mips specific constraints
3913   // GCC config/mips/constraints.md
3914   //
3915   // 'd' : An address register. Equivalent to r
3916   //       unless generating MIPS16 code.
3917   // 'y' : Equivalent to r; retained for
3918   //       backwards compatibility.
3919   // 'c' : A register suitable for use in an indirect
3920   //       jump. This will always be $25 for -mabicalls.
3921   // 'l' : The lo register. 1 word storage.
3922   // 'x' : The hilo register pair. Double word storage.
3923   if (Constraint.size() == 1) {
3924     switch (Constraint[0]) {
3925       default : break;
3926       case 'd':
3927       case 'y':
3928       case 'f':
3929       case 'c':
3930       case 'l':
3931       case 'x':
3932         return C_RegisterClass;
3933       case 'R':
3934         return C_Memory;
3935     }
3936   }
3937 
3938   if (Constraint == "ZC")
3939     return C_Memory;
3940 
3941   return TargetLowering::getConstraintType(Constraint);
3942 }
3943 
3944 /// Examine constraint type and operand type and determine a weight value.
3945 /// This object must already have been set up with the operand type
3946 /// and the current alternative constraint selected.
3947 TargetLowering::ConstraintWeight
3948 MipsTargetLowering::getSingleConstraintMatchWeight(
3949     AsmOperandInfo &info, const char *constraint) const {
3950   ConstraintWeight weight = CW_Invalid;
3951   Value *CallOperandVal = info.CallOperandVal;
3952     // If we don't have a value, we can't do a match,
3953     // but allow it at the lowest weight.
3954   if (!CallOperandVal)
3955     return CW_Default;
3956   Type *type = CallOperandVal->getType();
3957   // Look at the constraint type.
3958   switch (*constraint) {
3959   default:
3960     weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3961     break;
3962   case 'd':
3963   case 'y':
3964     if (type->isIntegerTy())
3965       weight = CW_Register;
3966     break;
3967   case 'f': // FPU or MSA register
3968     if (Subtarget.hasMSA() && type->isVectorTy() &&
3969         type->getPrimitiveSizeInBits().getFixedSize() == 128)
3970       weight = CW_Register;
3971     else if (type->isFloatTy())
3972       weight = CW_Register;
3973     break;
3974   case 'c': // $25 for indirect jumps
3975   case 'l': // lo register
3976   case 'x': // hilo register pair
3977     if (type->isIntegerTy())
3978       weight = CW_SpecificReg;
3979     break;
3980   case 'I': // signed 16 bit immediate
3981   case 'J': // integer zero
3982   case 'K': // unsigned 16 bit immediate
3983   case 'L': // signed 32 bit immediate where lower 16 bits are 0
3984   case 'N': // immediate in the range of -65535 to -1 (inclusive)
3985   case 'O': // signed 15 bit immediate (+- 16383)
3986   case 'P': // immediate in the range of 65535 to 1 (inclusive)
3987     if (isa<ConstantInt>(CallOperandVal))
3988       weight = CW_Constant;
3989     break;
3990   case 'R':
3991     weight = CW_Memory;
3992     break;
3993   }
3994   return weight;
3995 }
3996 
3997 /// This is a helper function to parse a physical register string and split it
3998 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
3999 /// that is returned indicates whether parsing was successful. The second flag
4000 /// is true if the numeric part exists.
4001 static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix,
4002                                               unsigned long long &Reg) {
4003   if (C.front() != '{' || C.back() != '}')
4004     return std::make_pair(false, false);
4005 
4006   // Search for the first numeric character.
4007   StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
4008   I = std::find_if(B, E, isdigit);
4009 
4010   Prefix = StringRef(B, I - B);
4011 
4012   // The second flag is set to false if no numeric characters were found.
4013   if (I == E)
4014     return std::make_pair(true, false);
4015 
4016   // Parse the numeric characters.
4017   return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
4018                         true);
4019 }
4020 
4021 EVT MipsTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT,
4022                                             ISD::NodeType) const {
4023   bool Cond = !Subtarget.isABI_O32() && VT.getSizeInBits() == 32;
4024   EVT MinVT = getRegisterType(Context, Cond ? MVT::i64 : MVT::i32);
4025   return VT.bitsLT(MinVT) ? MinVT : VT;
4026 }
4027 
4028 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
4029 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const {
4030   const TargetRegisterInfo *TRI =
4031       Subtarget.getRegisterInfo();
4032   const TargetRegisterClass *RC;
4033   StringRef Prefix;
4034   unsigned long long Reg;
4035 
4036   std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
4037 
4038   if (!R.first)
4039     return std::make_pair(0U, nullptr);
4040 
4041   if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
4042     // No numeric characters follow "hi" or "lo".
4043     if (R.second)
4044       return std::make_pair(0U, nullptr);
4045 
4046     RC = TRI->getRegClass(Prefix == "hi" ?
4047                           Mips::HI32RegClassID : Mips::LO32RegClassID);
4048     return std::make_pair(*(RC->begin()), RC);
4049   } else if (Prefix.startswith("$msa")) {
4050     // Parse $msa(ir|csr|access|save|modify|request|map|unmap)
4051 
4052     // No numeric characters follow the name.
4053     if (R.second)
4054       return std::make_pair(0U, nullptr);
4055 
4056     Reg = StringSwitch<unsigned long long>(Prefix)
4057               .Case("$msair", Mips::MSAIR)
4058               .Case("$msacsr", Mips::MSACSR)
4059               .Case("$msaaccess", Mips::MSAAccess)
4060               .Case("$msasave", Mips::MSASave)
4061               .Case("$msamodify", Mips::MSAModify)
4062               .Case("$msarequest", Mips::MSARequest)
4063               .Case("$msamap", Mips::MSAMap)
4064               .Case("$msaunmap", Mips::MSAUnmap)
4065               .Default(0);
4066 
4067     if (!Reg)
4068       return std::make_pair(0U, nullptr);
4069 
4070     RC = TRI->getRegClass(Mips::MSACtrlRegClassID);
4071     return std::make_pair(Reg, RC);
4072   }
4073 
4074   if (!R.second)
4075     return std::make_pair(0U, nullptr);
4076 
4077   if (Prefix == "$f") { // Parse $f0-$f31.
4078     // If the size of FP registers is 64-bit or Reg is an even number, select
4079     // the 64-bit register class. Otherwise, select the 32-bit register class.
4080     if (VT == MVT::Other)
4081       VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
4082 
4083     RC = getRegClassFor(VT);
4084 
4085     if (RC == &Mips::AFGR64RegClass) {
4086       assert(Reg % 2 == 0);
4087       Reg >>= 1;
4088     }
4089   } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7.
4090     RC = TRI->getRegClass(Mips::FCCRegClassID);
4091   else if (Prefix == "$w") { // Parse $w0-$w31.
4092     RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT);
4093   } else { // Parse $0-$31.
4094     assert(Prefix == "$");
4095     RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
4096   }
4097 
4098   assert(Reg < RC->getNumRegs());
4099   return std::make_pair(*(RC->begin() + Reg), RC);
4100 }
4101 
4102 /// Given a register class constraint, like 'r', if this corresponds directly
4103 /// to an LLVM register class, return a register of 0 and the register class
4104 /// pointer.
4105 std::pair<unsigned, const TargetRegisterClass *>
4106 MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
4107                                                  StringRef Constraint,
4108                                                  MVT VT) const {
4109   if (Constraint.size() == 1) {
4110     switch (Constraint[0]) {
4111     case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
4112     case 'y': // Same as 'r'. Exists for compatibility.
4113     case 'r':
4114       if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
4115         if (Subtarget.inMips16Mode())
4116           return std::make_pair(0U, &Mips::CPU16RegsRegClass);
4117         return std::make_pair(0U, &Mips::GPR32RegClass);
4118       }
4119       if (VT == MVT::i64 && !Subtarget.isGP64bit())
4120         return std::make_pair(0U, &Mips::GPR32RegClass);
4121       if (VT == MVT::i64 && Subtarget.isGP64bit())
4122         return std::make_pair(0U, &Mips::GPR64RegClass);
4123       // This will generate an error message
4124       return std::make_pair(0U, nullptr);
4125     case 'f': // FPU or MSA register
4126       if (VT == MVT::v16i8)
4127         return std::make_pair(0U, &Mips::MSA128BRegClass);
4128       else if (VT == MVT::v8i16 || VT == MVT::v8f16)
4129         return std::make_pair(0U, &Mips::MSA128HRegClass);
4130       else if (VT == MVT::v4i32 || VT == MVT::v4f32)
4131         return std::make_pair(0U, &Mips::MSA128WRegClass);
4132       else if (VT == MVT::v2i64 || VT == MVT::v2f64)
4133         return std::make_pair(0U, &Mips::MSA128DRegClass);
4134       else if (VT == MVT::f32)
4135         return std::make_pair(0U, &Mips::FGR32RegClass);
4136       else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) {
4137         if (Subtarget.isFP64bit())
4138           return std::make_pair(0U, &Mips::FGR64RegClass);
4139         return std::make_pair(0U, &Mips::AFGR64RegClass);
4140       }
4141       break;
4142     case 'c': // register suitable for indirect jump
4143       if (VT == MVT::i32)
4144         return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
4145       if (VT == MVT::i64)
4146         return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
4147       // This will generate an error message
4148       return std::make_pair(0U, nullptr);
4149     case 'l': // use the `lo` register to store values
4150               // that are no bigger than a word
4151       if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8)
4152         return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
4153       return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
4154     case 'x': // use the concatenated `hi` and `lo` registers
4155               // to store doubleword values
4156       // Fixme: Not triggering the use of both hi and low
4157       // This will generate an error message
4158       return std::make_pair(0U, nullptr);
4159     }
4160   }
4161 
4162   if (!Constraint.empty()) {
4163     std::pair<unsigned, const TargetRegisterClass *> R;
4164     R = parseRegForInlineAsmConstraint(Constraint, VT);
4165 
4166     if (R.second)
4167       return R;
4168   }
4169 
4170   return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
4171 }
4172 
4173 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4174 /// vector.  If it is invalid, don't add anything to Ops.
4175 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4176                                                      std::string &Constraint,
4177                                                      std::vector<SDValue>&Ops,
4178                                                      SelectionDAG &DAG) const {
4179   SDLoc DL(Op);
4180   SDValue Result;
4181 
4182   // Only support length 1 constraints for now.
4183   if (Constraint.length() > 1) return;
4184 
4185   char ConstraintLetter = Constraint[0];
4186   switch (ConstraintLetter) {
4187   default: break; // This will fall through to the generic implementation
4188   case 'I': // Signed 16 bit constant
4189     // If this fails, the parent routine will give an error
4190     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4191       EVT Type = Op.getValueType();
4192       int64_t Val = C->getSExtValue();
4193       if (isInt<16>(Val)) {
4194         Result = DAG.getTargetConstant(Val, DL, Type);
4195         break;
4196       }
4197     }
4198     return;
4199   case 'J': // integer zero
4200     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4201       EVT Type = Op.getValueType();
4202       int64_t Val = C->getZExtValue();
4203       if (Val == 0) {
4204         Result = DAG.getTargetConstant(0, DL, Type);
4205         break;
4206       }
4207     }
4208     return;
4209   case 'K': // unsigned 16 bit immediate
4210     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4211       EVT Type = Op.getValueType();
4212       uint64_t Val = (uint64_t)C->getZExtValue();
4213       if (isUInt<16>(Val)) {
4214         Result = DAG.getTargetConstant(Val, DL, Type);
4215         break;
4216       }
4217     }
4218     return;
4219   case 'L': // signed 32 bit immediate where lower 16 bits are 0
4220     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4221       EVT Type = Op.getValueType();
4222       int64_t Val = C->getSExtValue();
4223       if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
4224         Result = DAG.getTargetConstant(Val, DL, Type);
4225         break;
4226       }
4227     }
4228     return;
4229   case 'N': // immediate in the range of -65535 to -1 (inclusive)
4230     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4231       EVT Type = Op.getValueType();
4232       int64_t Val = C->getSExtValue();
4233       if ((Val >= -65535) && (Val <= -1)) {
4234         Result = DAG.getTargetConstant(Val, DL, Type);
4235         break;
4236       }
4237     }
4238     return;
4239   case 'O': // signed 15 bit immediate
4240     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4241       EVT Type = Op.getValueType();
4242       int64_t Val = C->getSExtValue();
4243       if ((isInt<15>(Val))) {
4244         Result = DAG.getTargetConstant(Val, DL, Type);
4245         break;
4246       }
4247     }
4248     return;
4249   case 'P': // immediate in the range of 1 to 65535 (inclusive)
4250     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
4251       EVT Type = Op.getValueType();
4252       int64_t Val = C->getSExtValue();
4253       if ((Val <= 65535) && (Val >= 1)) {
4254         Result = DAG.getTargetConstant(Val, DL, Type);
4255         break;
4256       }
4257     }
4258     return;
4259   }
4260 
4261   if (Result.getNode()) {
4262     Ops.push_back(Result);
4263     return;
4264   }
4265 
4266   TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4267 }
4268 
4269 bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL,
4270                                                const AddrMode &AM, Type *Ty,
4271                                                unsigned AS,
4272                                                Instruction *I) const {
4273   // No global is ever allowed as a base.
4274   if (AM.BaseGV)
4275     return false;
4276 
4277   switch (AM.Scale) {
4278   case 0: // "r+i" or just "i", depending on HasBaseReg.
4279     break;
4280   case 1:
4281     if (!AM.HasBaseReg) // allow "r+i".
4282       break;
4283     return false; // disallow "r+r" or "r+r+i".
4284   default:
4285     return false;
4286   }
4287 
4288   return true;
4289 }
4290 
4291 bool
4292 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4293   // The Mips target isn't yet aware of offsets.
4294   return false;
4295 }
4296 
4297 EVT MipsTargetLowering::getOptimalMemOpType(
4298     const MemOp &Op, const AttributeList &FuncAttributes) const {
4299   if (Subtarget.hasMips64())
4300     return MVT::i64;
4301 
4302   return MVT::i32;
4303 }
4304 
4305 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
4306                                       bool ForCodeSize) const {
4307   if (VT != MVT::f32 && VT != MVT::f64)
4308     return false;
4309   if (Imm.isNegZero())
4310     return false;
4311   return Imm.isZero();
4312 }
4313 
4314 unsigned MipsTargetLowering::getJumpTableEncoding() const {
4315 
4316   // FIXME: For space reasons this should be: EK_GPRel32BlockAddress.
4317   if (ABI.IsN64() && isPositionIndependent())
4318     return MachineJumpTableInfo::EK_GPRel64BlockAddress;
4319 
4320   return TargetLowering::getJumpTableEncoding();
4321 }
4322 
4323 bool MipsTargetLowering::useSoftFloat() const {
4324   return Subtarget.useSoftFloat();
4325 }
4326 
4327 void MipsTargetLowering::copyByValRegs(
4328     SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains,
4329     SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
4330     SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
4331     unsigned FirstReg, unsigned LastReg, const CCValAssign &VA,
4332     MipsCCState &State) const {
4333   MachineFunction &MF = DAG.getMachineFunction();
4334   MachineFrameInfo &MFI = MF.getFrameInfo();
4335   unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes();
4336   unsigned NumRegs = LastReg - FirstReg;
4337   unsigned RegAreaSize = NumRegs * GPRSizeInBytes;
4338   unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
4339   int FrameObjOffset;
4340   ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs();
4341 
4342   if (RegAreaSize)
4343     FrameObjOffset =
4344         (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
4345         (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes);
4346   else
4347     FrameObjOffset = VA.getLocMemOffset();
4348 
4349   // Create frame object.
4350   EVT PtrTy = getPointerTy(DAG.getDataLayout());
4351   // Make the fixed object stored to mutable so that the load instructions
4352   // referencing it have their memory dependencies added.
4353   // Set the frame object as isAliased which clears the underlying objects
4354   // vector in ScheduleDAGInstrs::buildSchedGraph() resulting in addition of all
4355   // stores as dependencies for loads referencing this fixed object.
4356   int FI = MFI.CreateFixedObject(FrameObjSize, FrameObjOffset, false, true);
4357   SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4358   InVals.push_back(FIN);
4359 
4360   if (!NumRegs)
4361     return;
4362 
4363   // Copy arg registers.
4364   MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8);
4365   const TargetRegisterClass *RC = getRegClassFor(RegTy);
4366 
4367   for (unsigned I = 0; I < NumRegs; ++I) {
4368     unsigned ArgReg = ByValArgRegs[FirstReg + I];
4369     unsigned VReg = addLiveIn(MF, ArgReg, RC);
4370     unsigned Offset = I * GPRSizeInBytes;
4371     SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
4372                                    DAG.getConstant(Offset, DL, PtrTy));
4373     SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
4374                                  StorePtr, MachinePointerInfo(FuncArg, Offset));
4375     OutChains.push_back(Store);
4376   }
4377 }
4378 
4379 // Copy byVal arg to registers and stack.
4380 void MipsTargetLowering::passByValArg(
4381     SDValue Chain, const SDLoc &DL,
4382     std::deque<std::pair<unsigned, SDValue>> &RegsToPass,
4383     SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
4384     MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg,
4385     unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle,
4386     const CCValAssign &VA) const {
4387   unsigned ByValSizeInBytes = Flags.getByValSize();
4388   unsigned OffsetInBytes = 0; // From beginning of struct
4389   unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
4390   Align Alignment =
4391       std::min(Flags.getNonZeroByValAlign(), Align(RegSizeInBytes));
4392   EVT PtrTy = getPointerTy(DAG.getDataLayout()),
4393       RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
4394   unsigned NumRegs = LastReg - FirstReg;
4395 
4396   if (NumRegs) {
4397     ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs();
4398     bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes);
4399     unsigned I = 0;
4400 
4401     // Copy words to registers.
4402     for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) {
4403       SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4404                                     DAG.getConstant(OffsetInBytes, DL, PtrTy));
4405       SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
4406                                     MachinePointerInfo(), Alignment.value());
4407       MemOpChains.push_back(LoadVal.getValue(1));
4408       unsigned ArgReg = ArgRegs[FirstReg + I];
4409       RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
4410     }
4411 
4412     // Return if the struct has been fully copied.
4413     if (ByValSizeInBytes == OffsetInBytes)
4414       return;
4415 
4416     // Copy the remainder of the byval argument with sub-word loads and shifts.
4417     if (LeftoverBytes) {
4418       SDValue Val;
4419 
4420       for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0;
4421            OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) {
4422         unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes;
4423 
4424         if (RemainingSizeInBytes < LoadSizeInBytes)
4425           continue;
4426 
4427         // Load subword.
4428         SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4429                                       DAG.getConstant(OffsetInBytes, DL,
4430                                                       PtrTy));
4431         SDValue LoadVal = DAG.getExtLoad(
4432             ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(),
4433             MVT::getIntegerVT(LoadSizeInBytes * 8), Alignment.value());
4434         MemOpChains.push_back(LoadVal.getValue(1));
4435 
4436         // Shift the loaded value.
4437         unsigned Shamt;
4438 
4439         if (isLittle)
4440           Shamt = TotalBytesLoaded * 8;
4441         else
4442           Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8;
4443 
4444         SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
4445                                     DAG.getConstant(Shamt, DL, MVT::i32));
4446 
4447         if (Val.getNode())
4448           Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
4449         else
4450           Val = Shift;
4451 
4452         OffsetInBytes += LoadSizeInBytes;
4453         TotalBytesLoaded += LoadSizeInBytes;
4454         Alignment = std::min(Alignment, Align(LoadSizeInBytes));
4455       }
4456 
4457       unsigned ArgReg = ArgRegs[FirstReg + I];
4458       RegsToPass.push_back(std::make_pair(ArgReg, Val));
4459       return;
4460     }
4461   }
4462 
4463   // Copy remainder of byval arg to it with memcpy.
4464   unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes;
4465   SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4466                             DAG.getConstant(OffsetInBytes, DL, PtrTy));
4467   SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
4468                             DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
4469   Chain = DAG.getMemcpy(
4470       Chain, DL, Dst, Src, DAG.getConstant(MemCpySize, DL, PtrTy),
4471       Align(Alignment), /*isVolatile=*/false, /*AlwaysInline=*/false,
4472       /*isTailCall=*/false, MachinePointerInfo(), MachinePointerInfo());
4473   MemOpChains.push_back(Chain);
4474 }
4475 
4476 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
4477                                          SDValue Chain, const SDLoc &DL,
4478                                          SelectionDAG &DAG,
4479                                          CCState &State) const {
4480   ArrayRef<MCPhysReg> ArgRegs = ABI.GetVarArgRegs();
4481   unsigned Idx = State.getFirstUnallocated(ArgRegs);
4482   unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
4483   MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
4484   const TargetRegisterClass *RC = getRegClassFor(RegTy);
4485   MachineFunction &MF = DAG.getMachineFunction();
4486   MachineFrameInfo &MFI = MF.getFrameInfo();
4487   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
4488 
4489   // Offset of the first variable argument from stack pointer.
4490   int VaArgOffset;
4491 
4492   if (ArgRegs.size() == Idx)
4493     VaArgOffset = alignTo(State.getNextStackOffset(), RegSizeInBytes);
4494   else {
4495     VaArgOffset =
4496         (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
4497         (int)(RegSizeInBytes * (ArgRegs.size() - Idx));
4498   }
4499 
4500   // Record the frame index of the first variable argument
4501   // which is a value necessary to VASTART.
4502   int FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
4503   MipsFI->setVarArgsFrameIndex(FI);
4504 
4505   // Copy the integer registers that have not been used for argument passing
4506   // to the argument register save area. For O32, the save area is allocated
4507   // in the caller's stack frame, while for N32/64, it is allocated in the
4508   // callee's stack frame.
4509   for (unsigned I = Idx; I < ArgRegs.size();
4510        ++I, VaArgOffset += RegSizeInBytes) {
4511     unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
4512     SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
4513     FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
4514     SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
4515     SDValue Store =
4516         DAG.getStore(Chain, DL, ArgValue, PtrOff, MachinePointerInfo());
4517     cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(
4518         (Value *)nullptr);
4519     OutChains.push_back(Store);
4520   }
4521 }
4522 
4523 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size,
4524                                      unsigned Align) const {
4525   const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
4526 
4527   assert(Size && "Byval argument's size shouldn't be 0.");
4528 
4529   Align = std::min(Align, TFL->getStackAlignment());
4530 
4531   unsigned FirstReg = 0;
4532   unsigned NumRegs = 0;
4533 
4534   if (State->getCallingConv() != CallingConv::Fast) {
4535     unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
4536     ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs();
4537     // FIXME: The O32 case actually describes no shadow registers.
4538     const MCPhysReg *ShadowRegs =
4539         ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs;
4540 
4541     // We used to check the size as well but we can't do that anymore since
4542     // CCState::HandleByVal() rounds up the size after calling this function.
4543     assert(!(Align % RegSizeInBytes) &&
4544            "Byval argument's alignment should be a multiple of"
4545            "RegSizeInBytes.");
4546 
4547     FirstReg = State->getFirstUnallocated(IntArgRegs);
4548 
4549     // If Align > RegSizeInBytes, the first arg register must be even.
4550     // FIXME: This condition happens to do the right thing but it's not the
4551     //        right way to test it. We want to check that the stack frame offset
4552     //        of the register is aligned.
4553     if ((Align > RegSizeInBytes) && (FirstReg % 2)) {
4554       State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]);
4555       ++FirstReg;
4556     }
4557 
4558     // Mark the registers allocated.
4559     Size = alignTo(Size, RegSizeInBytes);
4560     for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size());
4561          Size -= RegSizeInBytes, ++I, ++NumRegs)
4562       State->AllocateReg(IntArgRegs[I], ShadowRegs[I]);
4563   }
4564 
4565   State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs);
4566 }
4567 
4568 MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI,
4569                                                         MachineBasicBlock *BB,
4570                                                         bool isFPCmp,
4571                                                         unsigned Opc) const {
4572   assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) &&
4573          "Subtarget already supports SELECT nodes with the use of"
4574          "conditional-move instructions.");
4575 
4576   const TargetInstrInfo *TII =
4577       Subtarget.getInstrInfo();
4578   DebugLoc DL = MI.getDebugLoc();
4579 
4580   // To "insert" a SELECT instruction, we actually have to insert the
4581   // diamond control-flow pattern.  The incoming instruction knows the
4582   // destination vreg to set, the condition code register to branch on, the
4583   // true/false values to select between, and a branch opcode to use.
4584   const BasicBlock *LLVM_BB = BB->getBasicBlock();
4585   MachineFunction::iterator It = ++BB->getIterator();
4586 
4587   //  thisMBB:
4588   //  ...
4589   //   TrueVal = ...
4590   //   setcc r1, r2, r3
4591   //   bNE   r1, r0, copy1MBB
4592   //   fallthrough --> copy0MBB
4593   MachineBasicBlock *thisMBB  = BB;
4594   MachineFunction *F = BB->getParent();
4595   MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4596   MachineBasicBlock *sinkMBB  = F->CreateMachineBasicBlock(LLVM_BB);
4597   F->insert(It, copy0MBB);
4598   F->insert(It, sinkMBB);
4599 
4600   // Transfer the remainder of BB and its successor edges to sinkMBB.
4601   sinkMBB->splice(sinkMBB->begin(), BB,
4602                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
4603   sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4604 
4605   // Next, add the true and fallthrough blocks as its successors.
4606   BB->addSuccessor(copy0MBB);
4607   BB->addSuccessor(sinkMBB);
4608 
4609   if (isFPCmp) {
4610     // bc1[tf] cc, sinkMBB
4611     BuildMI(BB, DL, TII->get(Opc))
4612         .addReg(MI.getOperand(1).getReg())
4613         .addMBB(sinkMBB);
4614   } else {
4615     // bne rs, $0, sinkMBB
4616     BuildMI(BB, DL, TII->get(Opc))
4617         .addReg(MI.getOperand(1).getReg())
4618         .addReg(Mips::ZERO)
4619         .addMBB(sinkMBB);
4620   }
4621 
4622   //  copy0MBB:
4623   //   %FalseValue = ...
4624   //   # fallthrough to sinkMBB
4625   BB = copy0MBB;
4626 
4627   // Update machine-CFG edges
4628   BB->addSuccessor(sinkMBB);
4629 
4630   //  sinkMBB:
4631   //   %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
4632   //  ...
4633   BB = sinkMBB;
4634 
4635   BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg())
4636       .addReg(MI.getOperand(2).getReg())
4637       .addMBB(thisMBB)
4638       .addReg(MI.getOperand(3).getReg())
4639       .addMBB(copy0MBB);
4640 
4641   MI.eraseFromParent(); // The pseudo instruction is gone now.
4642 
4643   return BB;
4644 }
4645 
4646 MachineBasicBlock *
4647 MipsTargetLowering::emitPseudoD_SELECT(MachineInstr &MI,
4648                                        MachineBasicBlock *BB) const {
4649   assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) &&
4650          "Subtarget already supports SELECT nodes with the use of"
4651          "conditional-move instructions.");
4652 
4653   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
4654   DebugLoc DL = MI.getDebugLoc();
4655 
4656   // D_SELECT substitutes two SELECT nodes that goes one after another and
4657   // have the same condition operand. On machines which don't have
4658   // conditional-move instruction, it reduces unnecessary branch instructions
4659   // which are result of using two diamond patterns that are result of two
4660   // SELECT pseudo instructions.
4661   const BasicBlock *LLVM_BB = BB->getBasicBlock();
4662   MachineFunction::iterator It = ++BB->getIterator();
4663 
4664   //  thisMBB:
4665   //  ...
4666   //   TrueVal = ...
4667   //   setcc r1, r2, r3
4668   //   bNE   r1, r0, copy1MBB
4669   //   fallthrough --> copy0MBB
4670   MachineBasicBlock *thisMBB = BB;
4671   MachineFunction *F = BB->getParent();
4672   MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4673   MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4674   F->insert(It, copy0MBB);
4675   F->insert(It, sinkMBB);
4676 
4677   // Transfer the remainder of BB and its successor edges to sinkMBB.
4678   sinkMBB->splice(sinkMBB->begin(), BB,
4679                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
4680   sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4681 
4682   // Next, add the true and fallthrough blocks as its successors.
4683   BB->addSuccessor(copy0MBB);
4684   BB->addSuccessor(sinkMBB);
4685 
4686   // bne rs, $0, sinkMBB
4687   BuildMI(BB, DL, TII->get(Mips::BNE))
4688       .addReg(MI.getOperand(2).getReg())
4689       .addReg(Mips::ZERO)
4690       .addMBB(sinkMBB);
4691 
4692   //  copy0MBB:
4693   //   %FalseValue = ...
4694   //   # fallthrough to sinkMBB
4695   BB = copy0MBB;
4696 
4697   // Update machine-CFG edges
4698   BB->addSuccessor(sinkMBB);
4699 
4700   //  sinkMBB:
4701   //   %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
4702   //  ...
4703   BB = sinkMBB;
4704 
4705   // Use two PHI nodes to select two reults
4706   BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg())
4707       .addReg(MI.getOperand(3).getReg())
4708       .addMBB(thisMBB)
4709       .addReg(MI.getOperand(5).getReg())
4710       .addMBB(copy0MBB);
4711   BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(1).getReg())
4712       .addReg(MI.getOperand(4).getReg())
4713       .addMBB(thisMBB)
4714       .addReg(MI.getOperand(6).getReg())
4715       .addMBB(copy0MBB);
4716 
4717   MI.eraseFromParent(); // The pseudo instruction is gone now.
4718 
4719   return BB;
4720 }
4721 
4722 // FIXME? Maybe this could be a TableGen attribute on some registers and
4723 // this table could be generated automatically from RegInfo.
4724 Register
4725 MipsTargetLowering::getRegisterByName(const char *RegName, LLT VT,
4726                                       const MachineFunction &MF) const {
4727   // Named registers is expected to be fairly rare. For now, just support $28
4728   // since the linux kernel uses it.
4729   if (Subtarget.isGP64bit()) {
4730     Register Reg = StringSwitch<Register>(RegName)
4731                          .Case("$28", Mips::GP_64)
4732                          .Default(Register());
4733     if (Reg)
4734       return Reg;
4735   } else {
4736     Register Reg = StringSwitch<Register>(RegName)
4737                          .Case("$28", Mips::GP)
4738                          .Default(Register());
4739     if (Reg)
4740       return Reg;
4741   }
4742   report_fatal_error("Invalid register name global variable");
4743 }
4744 
4745 MachineBasicBlock *MipsTargetLowering::emitLDR_W(MachineInstr &MI,
4746                                                  MachineBasicBlock *BB) const {
4747   MachineFunction *MF = BB->getParent();
4748   MachineRegisterInfo &MRI = MF->getRegInfo();
4749   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
4750   const bool IsLittle = Subtarget.isLittle();
4751   DebugLoc DL = MI.getDebugLoc();
4752 
4753   Register Dest = MI.getOperand(0).getReg();
4754   Register Address = MI.getOperand(1).getReg();
4755   unsigned Imm = MI.getOperand(2).getImm();
4756 
4757   MachineBasicBlock::iterator I(MI);
4758 
4759   if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) {
4760     // Mips release 6 can load from adress that is not naturally-aligned.
4761     Register Temp = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4762     BuildMI(*BB, I, DL, TII->get(Mips::LW))
4763         .addDef(Temp)
4764         .addUse(Address)
4765         .addImm(Imm);
4766     BuildMI(*BB, I, DL, TII->get(Mips::FILL_W)).addDef(Dest).addUse(Temp);
4767   } else {
4768     // Mips release 5 needs to use instructions that can load from an unaligned
4769     // memory address.
4770     Register LoadHalf = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4771     Register LoadFull = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4772     Register Undef = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4773     BuildMI(*BB, I, DL, TII->get(Mips::IMPLICIT_DEF)).addDef(Undef);
4774     BuildMI(*BB, I, DL, TII->get(Mips::LWR))
4775         .addDef(LoadHalf)
4776         .addUse(Address)
4777         .addImm(Imm + (IsLittle ? 0 : 3))
4778         .addUse(Undef);
4779     BuildMI(*BB, I, DL, TII->get(Mips::LWL))
4780         .addDef(LoadFull)
4781         .addUse(Address)
4782         .addImm(Imm + (IsLittle ? 3 : 0))
4783         .addUse(LoadHalf);
4784     BuildMI(*BB, I, DL, TII->get(Mips::FILL_W)).addDef(Dest).addUse(LoadFull);
4785   }
4786 
4787   MI.eraseFromParent();
4788   return BB;
4789 }
4790 
4791 MachineBasicBlock *MipsTargetLowering::emitLDR_D(MachineInstr &MI,
4792                                                  MachineBasicBlock *BB) const {
4793   MachineFunction *MF = BB->getParent();
4794   MachineRegisterInfo &MRI = MF->getRegInfo();
4795   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
4796   const bool IsLittle = Subtarget.isLittle();
4797   DebugLoc DL = MI.getDebugLoc();
4798 
4799   Register Dest = MI.getOperand(0).getReg();
4800   Register Address = MI.getOperand(1).getReg();
4801   unsigned Imm = MI.getOperand(2).getImm();
4802 
4803   MachineBasicBlock::iterator I(MI);
4804 
4805   if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) {
4806     // Mips release 6 can load from adress that is not naturally-aligned.
4807     if (Subtarget.isGP64bit()) {
4808       Register Temp = MRI.createVirtualRegister(&Mips::GPR64RegClass);
4809       BuildMI(*BB, I, DL, TII->get(Mips::LD))
4810           .addDef(Temp)
4811           .addUse(Address)
4812           .addImm(Imm);
4813       BuildMI(*BB, I, DL, TII->get(Mips::FILL_D)).addDef(Dest).addUse(Temp);
4814     } else {
4815       Register Wtemp = MRI.createVirtualRegister(&Mips::MSA128WRegClass);
4816       Register Lo = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4817       Register Hi = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4818       BuildMI(*BB, I, DL, TII->get(Mips::LW))
4819           .addDef(Lo)
4820           .addUse(Address)
4821           .addImm(Imm + (IsLittle ? 0 : 4));
4822       BuildMI(*BB, I, DL, TII->get(Mips::LW))
4823           .addDef(Hi)
4824           .addUse(Address)
4825           .addImm(Imm + (IsLittle ? 4 : 0));
4826       BuildMI(*BB, I, DL, TII->get(Mips::FILL_W)).addDef(Wtemp).addUse(Lo);
4827       BuildMI(*BB, I, DL, TII->get(Mips::INSERT_W), Dest)
4828           .addUse(Wtemp)
4829           .addUse(Hi)
4830           .addImm(1);
4831     }
4832   } else {
4833     // Mips release 5 needs to use instructions that can load from an unaligned
4834     // memory address.
4835     Register LoHalf = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4836     Register LoFull = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4837     Register LoUndef = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4838     Register HiHalf = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4839     Register HiFull = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4840     Register HiUndef = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4841     Register Wtemp = MRI.createVirtualRegister(&Mips::MSA128WRegClass);
4842     BuildMI(*BB, I, DL, TII->get(Mips::IMPLICIT_DEF)).addDef(LoUndef);
4843     BuildMI(*BB, I, DL, TII->get(Mips::LWR))
4844         .addDef(LoHalf)
4845         .addUse(Address)
4846         .addImm(Imm + (IsLittle ? 0 : 7))
4847         .addUse(LoUndef);
4848     BuildMI(*BB, I, DL, TII->get(Mips::LWL))
4849         .addDef(LoFull)
4850         .addUse(Address)
4851         .addImm(Imm + (IsLittle ? 3 : 4))
4852         .addUse(LoHalf);
4853     BuildMI(*BB, I, DL, TII->get(Mips::IMPLICIT_DEF)).addDef(HiUndef);
4854     BuildMI(*BB, I, DL, TII->get(Mips::LWR))
4855         .addDef(HiHalf)
4856         .addUse(Address)
4857         .addImm(Imm + (IsLittle ? 4 : 3))
4858         .addUse(HiUndef);
4859     BuildMI(*BB, I, DL, TII->get(Mips::LWL))
4860         .addDef(HiFull)
4861         .addUse(Address)
4862         .addImm(Imm + (IsLittle ? 7 : 0))
4863         .addUse(HiHalf);
4864     BuildMI(*BB, I, DL, TII->get(Mips::FILL_W)).addDef(Wtemp).addUse(LoFull);
4865     BuildMI(*BB, I, DL, TII->get(Mips::INSERT_W), Dest)
4866         .addUse(Wtemp)
4867         .addUse(HiFull)
4868         .addImm(1);
4869   }
4870 
4871   MI.eraseFromParent();
4872   return BB;
4873 }
4874 
4875 MachineBasicBlock *MipsTargetLowering::emitSTR_W(MachineInstr &MI,
4876                                                  MachineBasicBlock *BB) const {
4877   MachineFunction *MF = BB->getParent();
4878   MachineRegisterInfo &MRI = MF->getRegInfo();
4879   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
4880   const bool IsLittle = Subtarget.isLittle();
4881   DebugLoc DL = MI.getDebugLoc();
4882 
4883   Register StoreVal = MI.getOperand(0).getReg();
4884   Register Address = MI.getOperand(1).getReg();
4885   unsigned Imm = MI.getOperand(2).getImm();
4886 
4887   MachineBasicBlock::iterator I(MI);
4888 
4889   if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) {
4890     // Mips release 6 can store to adress that is not naturally-aligned.
4891     Register BitcastW = MRI.createVirtualRegister(&Mips::MSA128WRegClass);
4892     Register Tmp = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4893     BuildMI(*BB, I, DL, TII->get(Mips::COPY)).addDef(BitcastW).addUse(StoreVal);
4894     BuildMI(*BB, I, DL, TII->get(Mips::COPY_S_W))
4895         .addDef(Tmp)
4896         .addUse(BitcastW)
4897         .addImm(0);
4898     BuildMI(*BB, I, DL, TII->get(Mips::SW))
4899         .addUse(Tmp)
4900         .addUse(Address)
4901         .addImm(Imm);
4902   } else {
4903     // Mips release 5 needs to use instructions that can store to an unaligned
4904     // memory address.
4905     Register Tmp = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4906     BuildMI(*BB, I, DL, TII->get(Mips::COPY_S_W))
4907         .addDef(Tmp)
4908         .addUse(StoreVal)
4909         .addImm(0);
4910     BuildMI(*BB, I, DL, TII->get(Mips::SWR))
4911         .addUse(Tmp)
4912         .addUse(Address)
4913         .addImm(Imm + (IsLittle ? 0 : 3));
4914     BuildMI(*BB, I, DL, TII->get(Mips::SWL))
4915         .addUse(Tmp)
4916         .addUse(Address)
4917         .addImm(Imm + (IsLittle ? 3 : 0));
4918   }
4919 
4920   MI.eraseFromParent();
4921 
4922   return BB;
4923 }
4924 
4925 MachineBasicBlock *MipsTargetLowering::emitSTR_D(MachineInstr &MI,
4926                                                  MachineBasicBlock *BB) const {
4927   MachineFunction *MF = BB->getParent();
4928   MachineRegisterInfo &MRI = MF->getRegInfo();
4929   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
4930   const bool IsLittle = Subtarget.isLittle();
4931   DebugLoc DL = MI.getDebugLoc();
4932 
4933   Register StoreVal = MI.getOperand(0).getReg();
4934   Register Address = MI.getOperand(1).getReg();
4935   unsigned Imm = MI.getOperand(2).getImm();
4936 
4937   MachineBasicBlock::iterator I(MI);
4938 
4939   if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) {
4940     // Mips release 6 can store to adress that is not naturally-aligned.
4941     if (Subtarget.isGP64bit()) {
4942       Register BitcastD = MRI.createVirtualRegister(&Mips::MSA128DRegClass);
4943       Register Lo = MRI.createVirtualRegister(&Mips::GPR64RegClass);
4944       BuildMI(*BB, I, DL, TII->get(Mips::COPY))
4945           .addDef(BitcastD)
4946           .addUse(StoreVal);
4947       BuildMI(*BB, I, DL, TII->get(Mips::COPY_S_D))
4948           .addDef(Lo)
4949           .addUse(BitcastD)
4950           .addImm(0);
4951       BuildMI(*BB, I, DL, TII->get(Mips::SD))
4952           .addUse(Lo)
4953           .addUse(Address)
4954           .addImm(Imm);
4955     } else {
4956       Register BitcastW = MRI.createVirtualRegister(&Mips::MSA128WRegClass);
4957       Register Lo = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4958       Register Hi = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4959       BuildMI(*BB, I, DL, TII->get(Mips::COPY))
4960           .addDef(BitcastW)
4961           .addUse(StoreVal);
4962       BuildMI(*BB, I, DL, TII->get(Mips::COPY_S_W))
4963           .addDef(Lo)
4964           .addUse(BitcastW)
4965           .addImm(0);
4966       BuildMI(*BB, I, DL, TII->get(Mips::COPY_S_W))
4967           .addDef(Hi)
4968           .addUse(BitcastW)
4969           .addImm(1);
4970       BuildMI(*BB, I, DL, TII->get(Mips::SW))
4971           .addUse(Lo)
4972           .addUse(Address)
4973           .addImm(Imm + (IsLittle ? 0 : 4));
4974       BuildMI(*BB, I, DL, TII->get(Mips::SW))
4975           .addUse(Hi)
4976           .addUse(Address)
4977           .addImm(Imm + (IsLittle ? 4 : 0));
4978     }
4979   } else {
4980     // Mips release 5 needs to use instructions that can store to an unaligned
4981     // memory address.
4982     Register Bitcast = MRI.createVirtualRegister(&Mips::MSA128WRegClass);
4983     Register Lo = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4984     Register Hi = MRI.createVirtualRegister(&Mips::GPR32RegClass);
4985     BuildMI(*BB, I, DL, TII->get(Mips::COPY)).addDef(Bitcast).addUse(StoreVal);
4986     BuildMI(*BB, I, DL, TII->get(Mips::COPY_S_W))
4987         .addDef(Lo)
4988         .addUse(Bitcast)
4989         .addImm(0);
4990     BuildMI(*BB, I, DL, TII->get(Mips::COPY_S_W))
4991         .addDef(Hi)
4992         .addUse(Bitcast)
4993         .addImm(1);
4994     BuildMI(*BB, I, DL, TII->get(Mips::SWR))
4995         .addUse(Lo)
4996         .addUse(Address)
4997         .addImm(Imm + (IsLittle ? 0 : 3));
4998     BuildMI(*BB, I, DL, TII->get(Mips::SWL))
4999         .addUse(Lo)
5000         .addUse(Address)
5001         .addImm(Imm + (IsLittle ? 3 : 0));
5002     BuildMI(*BB, I, DL, TII->get(Mips::SWR))
5003         .addUse(Hi)
5004         .addUse(Address)
5005         .addImm(Imm + (IsLittle ? 4 : 7));
5006     BuildMI(*BB, I, DL, TII->get(Mips::SWL))
5007         .addUse(Hi)
5008         .addUse(Address)
5009         .addImm(Imm + (IsLittle ? 7 : 4));
5010   }
5011 
5012   MI.eraseFromParent();
5013   return BB;
5014 }
5015