1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the interfaces that Mips uses to lower LLVM code into a
11 // selection DAG.
12 //
13 //===----------------------------------------------------------------------===//
14 #include "MipsISelLowering.h"
15 #include "InstPrinter/MipsInstPrinter.h"
16 #include "MCTargetDesc/MipsBaseInfo.h"
17 #include "MipsCCState.h"
18 #include "MipsMachineFunction.h"
19 #include "MipsSubtarget.h"
20 #include "MipsTargetMachine.h"
21 #include "MipsTargetObjectFile.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/StringSwitch.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineJumpTableInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/FunctionLoweringInfo.h"
31 #include "llvm/CodeGen/SelectionDAGISel.h"
32 #include "llvm/CodeGen/ValueTypes.h"
33 #include "llvm/IR/CallingConv.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/GlobalVariable.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include <cctype>
41 
42 using namespace llvm;
43 
44 #define DEBUG_TYPE "mips-lower"
45 
46 STATISTIC(NumTailCalls, "Number of tail calls");
47 
48 static cl::opt<bool>
49 LargeGOT("mxgot", cl::Hidden,
50          cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
51 
52 static cl::opt<bool>
53 NoZeroDivCheck("mno-check-zero-division", cl::Hidden,
54                cl::desc("MIPS: Don't trap on integer division by zero."),
55                cl::init(false));
56 
57 static const MCPhysReg Mips64DPRegs[8] = {
58   Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
59   Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
60 };
61 
62 // If I is a shifted mask, set the size (Size) and the first bit of the
63 // mask (Pos), and return true.
64 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
65 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
66   if (!isShiftedMask_64(I))
67     return false;
68 
69   Size = countPopulation(I);
70   Pos = countTrailingZeros(I);
71   return true;
72 }
73 
74 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
75   MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
76   return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
77 }
78 
79 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
80                                           SelectionDAG &DAG,
81                                           unsigned Flag) const {
82   return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag);
83 }
84 
85 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty,
86                                           SelectionDAG &DAG,
87                                           unsigned Flag) const {
88   return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
89 }
90 
91 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty,
92                                           SelectionDAG &DAG,
93                                           unsigned Flag) const {
94   return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
95 }
96 
97 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty,
98                                           SelectionDAG &DAG,
99                                           unsigned Flag) const {
100   return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
101 }
102 
103 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty,
104                                           SelectionDAG &DAG,
105                                           unsigned Flag) const {
106   return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
107                                    N->getOffset(), Flag);
108 }
109 
110 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
111   switch ((MipsISD::NodeType)Opcode) {
112   case MipsISD::FIRST_NUMBER:      break;
113   case MipsISD::JmpLink:           return "MipsISD::JmpLink";
114   case MipsISD::TailCall:          return "MipsISD::TailCall";
115   case MipsISD::Hi:                return "MipsISD::Hi";
116   case MipsISD::Lo:                return "MipsISD::Lo";
117   case MipsISD::GPRel:             return "MipsISD::GPRel";
118   case MipsISD::ThreadPointer:     return "MipsISD::ThreadPointer";
119   case MipsISD::Ret:               return "MipsISD::Ret";
120   case MipsISD::ERet:              return "MipsISD::ERet";
121   case MipsISD::EH_RETURN:         return "MipsISD::EH_RETURN";
122   case MipsISD::FPBrcond:          return "MipsISD::FPBrcond";
123   case MipsISD::FPCmp:             return "MipsISD::FPCmp";
124   case MipsISD::CMovFP_T:          return "MipsISD::CMovFP_T";
125   case MipsISD::CMovFP_F:          return "MipsISD::CMovFP_F";
126   case MipsISD::TruncIntFP:        return "MipsISD::TruncIntFP";
127   case MipsISD::MFHI:              return "MipsISD::MFHI";
128   case MipsISD::MFLO:              return "MipsISD::MFLO";
129   case MipsISD::MTLOHI:            return "MipsISD::MTLOHI";
130   case MipsISD::Mult:              return "MipsISD::Mult";
131   case MipsISD::Multu:             return "MipsISD::Multu";
132   case MipsISD::MAdd:              return "MipsISD::MAdd";
133   case MipsISD::MAddu:             return "MipsISD::MAddu";
134   case MipsISD::MSub:              return "MipsISD::MSub";
135   case MipsISD::MSubu:             return "MipsISD::MSubu";
136   case MipsISD::DivRem:            return "MipsISD::DivRem";
137   case MipsISD::DivRemU:           return "MipsISD::DivRemU";
138   case MipsISD::DivRem16:          return "MipsISD::DivRem16";
139   case MipsISD::DivRemU16:         return "MipsISD::DivRemU16";
140   case MipsISD::BuildPairF64:      return "MipsISD::BuildPairF64";
141   case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
142   case MipsISD::Wrapper:           return "MipsISD::Wrapper";
143   case MipsISD::DynAlloc:          return "MipsISD::DynAlloc";
144   case MipsISD::Sync:              return "MipsISD::Sync";
145   case MipsISD::Ext:               return "MipsISD::Ext";
146   case MipsISD::Ins:               return "MipsISD::Ins";
147   case MipsISD::LWL:               return "MipsISD::LWL";
148   case MipsISD::LWR:               return "MipsISD::LWR";
149   case MipsISD::SWL:               return "MipsISD::SWL";
150   case MipsISD::SWR:               return "MipsISD::SWR";
151   case MipsISD::LDL:               return "MipsISD::LDL";
152   case MipsISD::LDR:               return "MipsISD::LDR";
153   case MipsISD::SDL:               return "MipsISD::SDL";
154   case MipsISD::SDR:               return "MipsISD::SDR";
155   case MipsISD::EXTP:              return "MipsISD::EXTP";
156   case MipsISD::EXTPDP:            return "MipsISD::EXTPDP";
157   case MipsISD::EXTR_S_H:          return "MipsISD::EXTR_S_H";
158   case MipsISD::EXTR_W:            return "MipsISD::EXTR_W";
159   case MipsISD::EXTR_R_W:          return "MipsISD::EXTR_R_W";
160   case MipsISD::EXTR_RS_W:         return "MipsISD::EXTR_RS_W";
161   case MipsISD::SHILO:             return "MipsISD::SHILO";
162   case MipsISD::MTHLIP:            return "MipsISD::MTHLIP";
163   case MipsISD::MULSAQ_S_W_PH:     return "MipsISD::MULSAQ_S_W_PH";
164   case MipsISD::MAQ_S_W_PHL:       return "MipsISD::MAQ_S_W_PHL";
165   case MipsISD::MAQ_S_W_PHR:       return "MipsISD::MAQ_S_W_PHR";
166   case MipsISD::MAQ_SA_W_PHL:      return "MipsISD::MAQ_SA_W_PHL";
167   case MipsISD::MAQ_SA_W_PHR:      return "MipsISD::MAQ_SA_W_PHR";
168   case MipsISD::DPAU_H_QBL:        return "MipsISD::DPAU_H_QBL";
169   case MipsISD::DPAU_H_QBR:        return "MipsISD::DPAU_H_QBR";
170   case MipsISD::DPSU_H_QBL:        return "MipsISD::DPSU_H_QBL";
171   case MipsISD::DPSU_H_QBR:        return "MipsISD::DPSU_H_QBR";
172   case MipsISD::DPAQ_S_W_PH:       return "MipsISD::DPAQ_S_W_PH";
173   case MipsISD::DPSQ_S_W_PH:       return "MipsISD::DPSQ_S_W_PH";
174   case MipsISD::DPAQ_SA_L_W:       return "MipsISD::DPAQ_SA_L_W";
175   case MipsISD::DPSQ_SA_L_W:       return "MipsISD::DPSQ_SA_L_W";
176   case MipsISD::DPA_W_PH:          return "MipsISD::DPA_W_PH";
177   case MipsISD::DPS_W_PH:          return "MipsISD::DPS_W_PH";
178   case MipsISD::DPAQX_S_W_PH:      return "MipsISD::DPAQX_S_W_PH";
179   case MipsISD::DPAQX_SA_W_PH:     return "MipsISD::DPAQX_SA_W_PH";
180   case MipsISD::DPAX_W_PH:         return "MipsISD::DPAX_W_PH";
181   case MipsISD::DPSX_W_PH:         return "MipsISD::DPSX_W_PH";
182   case MipsISD::DPSQX_S_W_PH:      return "MipsISD::DPSQX_S_W_PH";
183   case MipsISD::DPSQX_SA_W_PH:     return "MipsISD::DPSQX_SA_W_PH";
184   case MipsISD::MULSA_W_PH:        return "MipsISD::MULSA_W_PH";
185   case MipsISD::MULT:              return "MipsISD::MULT";
186   case MipsISD::MULTU:             return "MipsISD::MULTU";
187   case MipsISD::MADD_DSP:          return "MipsISD::MADD_DSP";
188   case MipsISD::MADDU_DSP:         return "MipsISD::MADDU_DSP";
189   case MipsISD::MSUB_DSP:          return "MipsISD::MSUB_DSP";
190   case MipsISD::MSUBU_DSP:         return "MipsISD::MSUBU_DSP";
191   case MipsISD::SHLL_DSP:          return "MipsISD::SHLL_DSP";
192   case MipsISD::SHRA_DSP:          return "MipsISD::SHRA_DSP";
193   case MipsISD::SHRL_DSP:          return "MipsISD::SHRL_DSP";
194   case MipsISD::SETCC_DSP:         return "MipsISD::SETCC_DSP";
195   case MipsISD::SELECT_CC_DSP:     return "MipsISD::SELECT_CC_DSP";
196   case MipsISD::VALL_ZERO:         return "MipsISD::VALL_ZERO";
197   case MipsISD::VANY_ZERO:         return "MipsISD::VANY_ZERO";
198   case MipsISD::VALL_NONZERO:      return "MipsISD::VALL_NONZERO";
199   case MipsISD::VANY_NONZERO:      return "MipsISD::VANY_NONZERO";
200   case MipsISD::VCEQ:              return "MipsISD::VCEQ";
201   case MipsISD::VCLE_S:            return "MipsISD::VCLE_S";
202   case MipsISD::VCLE_U:            return "MipsISD::VCLE_U";
203   case MipsISD::VCLT_S:            return "MipsISD::VCLT_S";
204   case MipsISD::VCLT_U:            return "MipsISD::VCLT_U";
205   case MipsISD::VSMAX:             return "MipsISD::VSMAX";
206   case MipsISD::VSMIN:             return "MipsISD::VSMIN";
207   case MipsISD::VUMAX:             return "MipsISD::VUMAX";
208   case MipsISD::VUMIN:             return "MipsISD::VUMIN";
209   case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
210   case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
211   case MipsISD::VNOR:              return "MipsISD::VNOR";
212   case MipsISD::VSHF:              return "MipsISD::VSHF";
213   case MipsISD::SHF:               return "MipsISD::SHF";
214   case MipsISD::ILVEV:             return "MipsISD::ILVEV";
215   case MipsISD::ILVOD:             return "MipsISD::ILVOD";
216   case MipsISD::ILVL:              return "MipsISD::ILVL";
217   case MipsISD::ILVR:              return "MipsISD::ILVR";
218   case MipsISD::PCKEV:             return "MipsISD::PCKEV";
219   case MipsISD::PCKOD:             return "MipsISD::PCKOD";
220   case MipsISD::INSVE:             return "MipsISD::INSVE";
221   }
222   return nullptr;
223 }
224 
225 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM,
226                                        const MipsSubtarget &STI)
227     : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) {
228   // Mips does not have i1 type, so use i32 for
229   // setcc operations results (slt, sgt, ...).
230   setBooleanContents(ZeroOrOneBooleanContent);
231   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
232   // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA
233   // does. Integer booleans still use 0 and 1.
234   if (Subtarget.hasMips32r6())
235     setBooleanContents(ZeroOrOneBooleanContent,
236                        ZeroOrNegativeOneBooleanContent);
237 
238   // Load extented operations for i1 types must be promoted
239   for (MVT VT : MVT::integer_valuetypes()) {
240     setLoadExtAction(ISD::EXTLOAD,  VT, MVT::i1,  Promote);
241     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1,  Promote);
242     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1,  Promote);
243   }
244 
245   // MIPS doesn't have extending float->double load/store.  Set LoadExtAction
246   // for f32, f16
247   for (MVT VT : MVT::fp_valuetypes()) {
248     setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
249     setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
250   }
251 
252   // Set LoadExtAction for f16 vectors to Expand
253   for (MVT VT : MVT::fp_vector_valuetypes()) {
254     MVT F16VT = MVT::getVectorVT(MVT::f16, VT.getVectorNumElements());
255     if (F16VT.isValid())
256       setLoadExtAction(ISD::EXTLOAD, VT, F16VT, Expand);
257   }
258 
259   setTruncStoreAction(MVT::f32, MVT::f16, Expand);
260   setTruncStoreAction(MVT::f64, MVT::f16, Expand);
261 
262   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
263 
264   // Used by legalize types to correctly generate the setcc result.
265   // Without this, every float setcc comes with a AND/OR with the result,
266   // we don't want this, since the fpcmp result goes to a flag register,
267   // which is used implicitly by brcond and select operations.
268   AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
269 
270   // Mips Custom Operations
271   setOperationAction(ISD::BR_JT,              MVT::Other, Custom);
272   setOperationAction(ISD::GlobalAddress,      MVT::i32,   Custom);
273   setOperationAction(ISD::BlockAddress,       MVT::i32,   Custom);
274   setOperationAction(ISD::GlobalTLSAddress,   MVT::i32,   Custom);
275   setOperationAction(ISD::JumpTable,          MVT::i32,   Custom);
276   setOperationAction(ISD::ConstantPool,       MVT::i32,   Custom);
277   setOperationAction(ISD::SELECT,             MVT::f32,   Custom);
278   setOperationAction(ISD::SELECT,             MVT::f64,   Custom);
279   setOperationAction(ISD::SELECT,             MVT::i32,   Custom);
280   setOperationAction(ISD::SETCC,              MVT::f32,   Custom);
281   setOperationAction(ISD::SETCC,              MVT::f64,   Custom);
282   setOperationAction(ISD::BRCOND,             MVT::Other, Custom);
283   setOperationAction(ISD::FCOPYSIGN,          MVT::f32,   Custom);
284   setOperationAction(ISD::FCOPYSIGN,          MVT::f64,   Custom);
285   setOperationAction(ISD::FP_TO_SINT,         MVT::i32,   Custom);
286 
287   if (Subtarget.isGP64bit()) {
288     setOperationAction(ISD::GlobalAddress,      MVT::i64,   Custom);
289     setOperationAction(ISD::BlockAddress,       MVT::i64,   Custom);
290     setOperationAction(ISD::GlobalTLSAddress,   MVT::i64,   Custom);
291     setOperationAction(ISD::JumpTable,          MVT::i64,   Custom);
292     setOperationAction(ISD::ConstantPool,       MVT::i64,   Custom);
293     setOperationAction(ISD::SELECT,             MVT::i64,   Custom);
294     setOperationAction(ISD::LOAD,               MVT::i64,   Custom);
295     setOperationAction(ISD::STORE,              MVT::i64,   Custom);
296     setOperationAction(ISD::FP_TO_SINT,         MVT::i64,   Custom);
297     setOperationAction(ISD::SHL_PARTS,          MVT::i64,   Custom);
298     setOperationAction(ISD::SRA_PARTS,          MVT::i64,   Custom);
299     setOperationAction(ISD::SRL_PARTS,          MVT::i64,   Custom);
300   }
301 
302   if (!Subtarget.isGP64bit()) {
303     setOperationAction(ISD::SHL_PARTS,          MVT::i32,   Custom);
304     setOperationAction(ISD::SRA_PARTS,          MVT::i32,   Custom);
305     setOperationAction(ISD::SRL_PARTS,          MVT::i32,   Custom);
306   }
307 
308   setOperationAction(ISD::EH_DWARF_CFA,         MVT::i32,   Custom);
309   if (Subtarget.isGP64bit())
310     setOperationAction(ISD::EH_DWARF_CFA,       MVT::i64,   Custom);
311 
312   setOperationAction(ISD::SDIV, MVT::i32, Expand);
313   setOperationAction(ISD::SREM, MVT::i32, Expand);
314   setOperationAction(ISD::UDIV, MVT::i32, Expand);
315   setOperationAction(ISD::UREM, MVT::i32, Expand);
316   setOperationAction(ISD::SDIV, MVT::i64, Expand);
317   setOperationAction(ISD::SREM, MVT::i64, Expand);
318   setOperationAction(ISD::UDIV, MVT::i64, Expand);
319   setOperationAction(ISD::UREM, MVT::i64, Expand);
320 
321   // Operations not directly supported by Mips.
322   setOperationAction(ISD::BR_CC,             MVT::f32,   Expand);
323   setOperationAction(ISD::BR_CC,             MVT::f64,   Expand);
324   setOperationAction(ISD::BR_CC,             MVT::i32,   Expand);
325   setOperationAction(ISD::BR_CC,             MVT::i64,   Expand);
326   setOperationAction(ISD::SELECT_CC,         MVT::i32,   Expand);
327   setOperationAction(ISD::SELECT_CC,         MVT::i64,   Expand);
328   setOperationAction(ISD::SELECT_CC,         MVT::f32,   Expand);
329   setOperationAction(ISD::SELECT_CC,         MVT::f64,   Expand);
330   setOperationAction(ISD::UINT_TO_FP,        MVT::i32,   Expand);
331   setOperationAction(ISD::UINT_TO_FP,        MVT::i64,   Expand);
332   setOperationAction(ISD::FP_TO_UINT,        MVT::i32,   Expand);
333   setOperationAction(ISD::FP_TO_UINT,        MVT::i64,   Expand);
334   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1,    Expand);
335   if (Subtarget.hasCnMips()) {
336     setOperationAction(ISD::CTPOP,           MVT::i32,   Legal);
337     setOperationAction(ISD::CTPOP,           MVT::i64,   Legal);
338   } else {
339     setOperationAction(ISD::CTPOP,           MVT::i32,   Expand);
340     setOperationAction(ISD::CTPOP,           MVT::i64,   Expand);
341   }
342   setOperationAction(ISD::CTTZ,              MVT::i32,   Expand);
343   setOperationAction(ISD::CTTZ,              MVT::i64,   Expand);
344   setOperationAction(ISD::ROTL,              MVT::i32,   Expand);
345   setOperationAction(ISD::ROTL,              MVT::i64,   Expand);
346   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32,  Expand);
347   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64,  Expand);
348 
349   if (!Subtarget.hasMips32r2())
350     setOperationAction(ISD::ROTR, MVT::i32,   Expand);
351 
352   if (!Subtarget.hasMips64r2())
353     setOperationAction(ISD::ROTR, MVT::i64,   Expand);
354 
355   setOperationAction(ISD::FSIN,              MVT::f32,   Expand);
356   setOperationAction(ISD::FSIN,              MVT::f64,   Expand);
357   setOperationAction(ISD::FCOS,              MVT::f32,   Expand);
358   setOperationAction(ISD::FCOS,              MVT::f64,   Expand);
359   setOperationAction(ISD::FSINCOS,           MVT::f32,   Expand);
360   setOperationAction(ISD::FSINCOS,           MVT::f64,   Expand);
361   setOperationAction(ISD::FPOWI,             MVT::f32,   Expand);
362   setOperationAction(ISD::FPOW,              MVT::f32,   Expand);
363   setOperationAction(ISD::FPOW,              MVT::f64,   Expand);
364   setOperationAction(ISD::FLOG,              MVT::f32,   Expand);
365   setOperationAction(ISD::FLOG2,             MVT::f32,   Expand);
366   setOperationAction(ISD::FLOG10,            MVT::f32,   Expand);
367   setOperationAction(ISD::FEXP,              MVT::f32,   Expand);
368   setOperationAction(ISD::FMA,               MVT::f32,   Expand);
369   setOperationAction(ISD::FMA,               MVT::f64,   Expand);
370   setOperationAction(ISD::FREM,              MVT::f32,   Expand);
371   setOperationAction(ISD::FREM,              MVT::f64,   Expand);
372 
373   // Lower f16 conversion operations into library calls
374   setOperationAction(ISD::FP16_TO_FP,        MVT::f32,   Expand);
375   setOperationAction(ISD::FP_TO_FP16,        MVT::f32,   Expand);
376   setOperationAction(ISD::FP16_TO_FP,        MVT::f64,   Expand);
377   setOperationAction(ISD::FP_TO_FP16,        MVT::f64,   Expand);
378 
379   setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
380 
381   setOperationAction(ISD::VASTART,           MVT::Other, Custom);
382   setOperationAction(ISD::VAARG,             MVT::Other, Custom);
383   setOperationAction(ISD::VACOPY,            MVT::Other, Expand);
384   setOperationAction(ISD::VAEND,             MVT::Other, Expand);
385 
386   // Use the default for now
387   setOperationAction(ISD::STACKSAVE,         MVT::Other, Expand);
388   setOperationAction(ISD::STACKRESTORE,      MVT::Other, Expand);
389 
390   if (!Subtarget.isGP64bit()) {
391     setOperationAction(ISD::ATOMIC_LOAD,     MVT::i64,   Expand);
392     setOperationAction(ISD::ATOMIC_STORE,    MVT::i64,   Expand);
393   }
394 
395 
396   if (!Subtarget.hasMips32r2()) {
397     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
398     setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
399   }
400 
401   // MIPS16 lacks MIPS32's clz and clo instructions.
402   if (!Subtarget.hasMips32() || Subtarget.inMips16Mode())
403     setOperationAction(ISD::CTLZ, MVT::i32, Expand);
404   if (!Subtarget.hasMips64())
405     setOperationAction(ISD::CTLZ, MVT::i64, Expand);
406 
407   if (!Subtarget.hasMips32r2())
408     setOperationAction(ISD::BSWAP, MVT::i32, Expand);
409   if (!Subtarget.hasMips64r2())
410     setOperationAction(ISD::BSWAP, MVT::i64, Expand);
411 
412   if (Subtarget.isGP64bit()) {
413     setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom);
414     setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom);
415     setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom);
416     setTruncStoreAction(MVT::i64, MVT::i32, Custom);
417   }
418 
419   setOperationAction(ISD::TRAP, MVT::Other, Legal);
420 
421   setTargetDAGCombine(ISD::SDIVREM);
422   setTargetDAGCombine(ISD::UDIVREM);
423   setTargetDAGCombine(ISD::SELECT);
424   setTargetDAGCombine(ISD::AND);
425   setTargetDAGCombine(ISD::OR);
426   setTargetDAGCombine(ISD::ADD);
427   setTargetDAGCombine(ISD::AssertZext);
428 
429   if (ABI.IsO32()) {
430     // These libcalls are not available in 32-bit.
431     setLibcallName(RTLIB::SHL_I128, nullptr);
432     setLibcallName(RTLIB::SRL_I128, nullptr);
433     setLibcallName(RTLIB::SRA_I128, nullptr);
434   }
435 
436   setMinFunctionAlignment(Subtarget.isGP64bit() ? 3 : 2);
437 
438   // The arguments on the stack are defined in terms of 4-byte slots on O32
439   // and 8-byte slots on N32/N64.
440   setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? 8 : 4);
441 
442   setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP);
443 
444   MaxStoresPerMemcpy = 16;
445 
446   isMicroMips = Subtarget.inMicroMipsMode();
447 }
448 
449 const MipsTargetLowering *MipsTargetLowering::create(const MipsTargetMachine &TM,
450                                                      const MipsSubtarget &STI) {
451   if (STI.inMips16Mode())
452     return llvm::createMips16TargetLowering(TM, STI);
453 
454   return llvm::createMipsSETargetLowering(TM, STI);
455 }
456 
457 // Create a fast isel object.
458 FastISel *
459 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
460                                   const TargetLibraryInfo *libInfo) const {
461   if (!funcInfo.MF->getTarget().Options.EnableFastISel)
462     return TargetLowering::createFastISel(funcInfo, libInfo);
463   return Mips::createFastISel(funcInfo, libInfo);
464 }
465 
466 EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
467                                            EVT VT) const {
468   if (!VT.isVector())
469     return MVT::i32;
470   return VT.changeVectorElementTypeToInteger();
471 }
472 
473 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
474                                     TargetLowering::DAGCombinerInfo &DCI,
475                                     const MipsSubtarget &Subtarget) {
476   if (DCI.isBeforeLegalizeOps())
477     return SDValue();
478 
479   EVT Ty = N->getValueType(0);
480   unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
481   unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
482   unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
483                                                   MipsISD::DivRemU16;
484   SDLoc DL(N);
485 
486   SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
487                                N->getOperand(0), N->getOperand(1));
488   SDValue InChain = DAG.getEntryNode();
489   SDValue InGlue = DivRem;
490 
491   // insert MFLO
492   if (N->hasAnyUseOfValue(0)) {
493     SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
494                                             InGlue);
495     DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
496     InChain = CopyFromLo.getValue(1);
497     InGlue = CopyFromLo.getValue(2);
498   }
499 
500   // insert MFHI
501   if (N->hasAnyUseOfValue(1)) {
502     SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
503                                             HI, Ty, InGlue);
504     DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
505   }
506 
507   return SDValue();
508 }
509 
510 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
511   switch (CC) {
512   default: llvm_unreachable("Unknown fp condition code!");
513   case ISD::SETEQ:
514   case ISD::SETOEQ: return Mips::FCOND_OEQ;
515   case ISD::SETUNE: return Mips::FCOND_UNE;
516   case ISD::SETLT:
517   case ISD::SETOLT: return Mips::FCOND_OLT;
518   case ISD::SETGT:
519   case ISD::SETOGT: return Mips::FCOND_OGT;
520   case ISD::SETLE:
521   case ISD::SETOLE: return Mips::FCOND_OLE;
522   case ISD::SETGE:
523   case ISD::SETOGE: return Mips::FCOND_OGE;
524   case ISD::SETULT: return Mips::FCOND_ULT;
525   case ISD::SETULE: return Mips::FCOND_ULE;
526   case ISD::SETUGT: return Mips::FCOND_UGT;
527   case ISD::SETUGE: return Mips::FCOND_UGE;
528   case ISD::SETUO:  return Mips::FCOND_UN;
529   case ISD::SETO:   return Mips::FCOND_OR;
530   case ISD::SETNE:
531   case ISD::SETONE: return Mips::FCOND_ONE;
532   case ISD::SETUEQ: return Mips::FCOND_UEQ;
533   }
534 }
535 
536 
537 /// This function returns true if the floating point conditional branches and
538 /// conditional moves which use condition code CC should be inverted.
539 static bool invertFPCondCodeUser(Mips::CondCode CC) {
540   if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
541     return false;
542 
543   assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
544          "Illegal Condition Code");
545 
546   return true;
547 }
548 
549 // Creates and returns an FPCmp node from a setcc node.
550 // Returns Op if setcc is not a floating point comparison.
551 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
552   // must be a SETCC node
553   if (Op.getOpcode() != ISD::SETCC)
554     return Op;
555 
556   SDValue LHS = Op.getOperand(0);
557 
558   if (!LHS.getValueType().isFloatingPoint())
559     return Op;
560 
561   SDValue RHS = Op.getOperand(1);
562   SDLoc DL(Op);
563 
564   // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
565   // node if necessary.
566   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
567 
568   return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
569                      DAG.getConstant(condCodeToFCC(CC), DL, MVT::i32));
570 }
571 
572 // Creates and returns a CMovFPT/F node.
573 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
574                             SDValue False, const SDLoc &DL) {
575   ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
576   bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
577   SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
578 
579   return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
580                      True.getValueType(), True, FCC0, False, Cond);
581 }
582 
583 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
584                                     TargetLowering::DAGCombinerInfo &DCI,
585                                     const MipsSubtarget &Subtarget) {
586   if (DCI.isBeforeLegalizeOps())
587     return SDValue();
588 
589   SDValue SetCC = N->getOperand(0);
590 
591   if ((SetCC.getOpcode() != ISD::SETCC) ||
592       !SetCC.getOperand(0).getValueType().isInteger())
593     return SDValue();
594 
595   SDValue False = N->getOperand(2);
596   EVT FalseTy = False.getValueType();
597 
598   if (!FalseTy.isInteger())
599     return SDValue();
600 
601   ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False);
602 
603   // If the RHS (False) is 0, we swap the order of the operands
604   // of ISD::SELECT (obviously also inverting the condition) so that we can
605   // take advantage of conditional moves using the $0 register.
606   // Example:
607   //   return (a != 0) ? x : 0;
608   //     load $reg, x
609   //     movz $reg, $0, a
610   if (!FalseC)
611     return SDValue();
612 
613   const SDLoc DL(N);
614 
615   if (!FalseC->getZExtValue()) {
616     ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
617     SDValue True = N->getOperand(1);
618 
619     SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
620                          SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
621 
622     return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
623   }
624 
625   // If both operands are integer constants there's a possibility that we
626   // can do some interesting optimizations.
627   SDValue True = N->getOperand(1);
628   ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True);
629 
630   if (!TrueC || !True.getValueType().isInteger())
631     return SDValue();
632 
633   // We'll also ignore MVT::i64 operands as this optimizations proves
634   // to be ineffective because of the required sign extensions as the result
635   // of a SETCC operator is always MVT::i32 for non-vector types.
636   if (True.getValueType() == MVT::i64)
637     return SDValue();
638 
639   int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue();
640 
641   // 1)  (a < x) ? y : y-1
642   //  slti $reg1, a, x
643   //  addiu $reg2, $reg1, y-1
644   if (Diff == 1)
645     return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False);
646 
647   // 2)  (a < x) ? y-1 : y
648   //  slti $reg1, a, x
649   //  xor $reg1, $reg1, 1
650   //  addiu $reg2, $reg1, y-1
651   if (Diff == -1) {
652     ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
653     SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
654                          SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
655     return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True);
656   }
657 
658   // Couldn't optimize.
659   return SDValue();
660 }
661 
662 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG,
663                                     TargetLowering::DAGCombinerInfo &DCI,
664                                     const MipsSubtarget &Subtarget) {
665   if (DCI.isBeforeLegalizeOps())
666     return SDValue();
667 
668   SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2);
669 
670   ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse);
671   if (!FalseC || FalseC->getZExtValue())
672     return SDValue();
673 
674   // Since RHS (False) is 0, we swap the order of the True/False operands
675   // (obviously also inverting the condition) so that we can
676   // take advantage of conditional moves using the $0 register.
677   // Example:
678   //   return (a != 0) ? x : 0;
679   //     load $reg, x
680   //     movz $reg, $0, a
681   unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F :
682                                                          MipsISD::CMovFP_T;
683 
684   SDValue FCC = N->getOperand(1), Glue = N->getOperand(3);
685   return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(),
686                      ValueIfFalse, FCC, ValueIfTrue, Glue);
687 }
688 
689 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
690                                  TargetLowering::DAGCombinerInfo &DCI,
691                                  const MipsSubtarget &Subtarget) {
692   // Pattern match EXT.
693   //  $dst = and ((sra or srl) $src , pos), (2**size - 1)
694   //  => ext $dst, $src, size, pos
695   if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
696     return SDValue();
697 
698   SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
699   unsigned ShiftRightOpc = ShiftRight.getOpcode();
700 
701   // Op's first operand must be a shift right.
702   if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
703     return SDValue();
704 
705   // The second operand of the shift must be an immediate.
706   ConstantSDNode *CN;
707   if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
708     return SDValue();
709 
710   uint64_t Pos = CN->getZExtValue();
711   uint64_t SMPos, SMSize;
712 
713   // Op's second operand must be a shifted mask.
714   if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
715       !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
716     return SDValue();
717 
718   // Return if the shifted mask does not start at bit 0 or the sum of its size
719   // and Pos exceeds the word's size.
720   EVT ValTy = N->getValueType(0);
721   if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
722     return SDValue();
723 
724   SDLoc DL(N);
725   return DAG.getNode(MipsISD::Ext, DL, ValTy,
726                      ShiftRight.getOperand(0),
727                      DAG.getConstant(Pos, DL, MVT::i32),
728                      DAG.getConstant(SMSize, DL, MVT::i32));
729 }
730 
731 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
732                                 TargetLowering::DAGCombinerInfo &DCI,
733                                 const MipsSubtarget &Subtarget) {
734   // Pattern match INS.
735   //  $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
736   //  where mask1 = (2**size - 1) << pos, mask0 = ~mask1
737   //  => ins $dst, $src, size, pos, $src1
738   if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
739     return SDValue();
740 
741   SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
742   uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
743   ConstantSDNode *CN;
744 
745   // See if Op's first operand matches (and $src1 , mask0).
746   if (And0.getOpcode() != ISD::AND)
747     return SDValue();
748 
749   if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
750       !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
751     return SDValue();
752 
753   // See if Op's second operand matches (and (shl $src, pos), mask1).
754   if (And1.getOpcode() != ISD::AND)
755     return SDValue();
756 
757   if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
758       !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
759     return SDValue();
760 
761   // The shift masks must have the same position and size.
762   if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
763     return SDValue();
764 
765   SDValue Shl = And1.getOperand(0);
766   if (Shl.getOpcode() != ISD::SHL)
767     return SDValue();
768 
769   if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
770     return SDValue();
771 
772   unsigned Shamt = CN->getZExtValue();
773 
774   // Return if the shift amount and the first bit position of mask are not the
775   // same.
776   EVT ValTy = N->getValueType(0);
777   if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
778     return SDValue();
779 
780   SDLoc DL(N);
781   return DAG.getNode(MipsISD::Ins, DL, ValTy, Shl.getOperand(0),
782                      DAG.getConstant(SMPos0, DL, MVT::i32),
783                      DAG.getConstant(SMSize0, DL, MVT::i32),
784                      And0.getOperand(0));
785 }
786 
787 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
788                                  TargetLowering::DAGCombinerInfo &DCI,
789                                  const MipsSubtarget &Subtarget) {
790   // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
791 
792   if (DCI.isBeforeLegalizeOps())
793     return SDValue();
794 
795   SDValue Add = N->getOperand(1);
796 
797   if (Add.getOpcode() != ISD::ADD)
798     return SDValue();
799 
800   SDValue Lo = Add.getOperand(1);
801 
802   if ((Lo.getOpcode() != MipsISD::Lo) ||
803       (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
804     return SDValue();
805 
806   EVT ValTy = N->getValueType(0);
807   SDLoc DL(N);
808 
809   SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
810                              Add.getOperand(0));
811   return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
812 }
813 
814 static SDValue performAssertZextCombine(SDNode *N, SelectionDAG &DAG,
815                                         TargetLowering::DAGCombinerInfo &DCI,
816                                         const MipsSubtarget &Subtarget) {
817   SDValue N0 = N->getOperand(0);
818   EVT NarrowerVT = cast<VTSDNode>(N->getOperand(1))->getVT();
819 
820   if (N0.getOpcode() != ISD::TRUNCATE)
821     return SDValue();
822 
823   if (N0.getOperand(0).getOpcode() != ISD::AssertZext)
824     return SDValue();
825 
826   // fold (AssertZext (trunc (AssertZext x))) -> (trunc (AssertZext x))
827   // if the type of the extension of the innermost AssertZext node is
828   // smaller from that of the outermost node, eg:
829   // (AssertZext:i32 (trunc:i32 (AssertZext:i64 X, i32)), i8)
830   //   -> (trunc:i32 (AssertZext X, i8))
831   SDValue WiderAssertZext = N0.getOperand(0);
832   EVT WiderVT = cast<VTSDNode>(WiderAssertZext->getOperand(1))->getVT();
833 
834   if (NarrowerVT.bitsLT(WiderVT)) {
835     SDValue NewAssertZext = DAG.getNode(
836         ISD::AssertZext, SDLoc(N), WiderAssertZext.getValueType(),
837         WiderAssertZext.getOperand(0), DAG.getValueType(NarrowerVT));
838     return DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0),
839                        NewAssertZext);
840   }
841 
842   return SDValue();
843 }
844 
845 SDValue  MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
846   const {
847   SelectionDAG &DAG = DCI.DAG;
848   unsigned Opc = N->getOpcode();
849 
850   switch (Opc) {
851   default: break;
852   case ISD::SDIVREM:
853   case ISD::UDIVREM:
854     return performDivRemCombine(N, DAG, DCI, Subtarget);
855   case ISD::SELECT:
856     return performSELECTCombine(N, DAG, DCI, Subtarget);
857   case MipsISD::CMovFP_F:
858   case MipsISD::CMovFP_T:
859     return performCMovFPCombine(N, DAG, DCI, Subtarget);
860   case ISD::AND:
861     return performANDCombine(N, DAG, DCI, Subtarget);
862   case ISD::OR:
863     return performORCombine(N, DAG, DCI, Subtarget);
864   case ISD::ADD:
865     return performADDCombine(N, DAG, DCI, Subtarget);
866   case ISD::AssertZext:
867     return performAssertZextCombine(N, DAG, DCI, Subtarget);
868   }
869 
870   return SDValue();
871 }
872 
873 bool MipsTargetLowering::isCheapToSpeculateCttz() const {
874   return Subtarget.hasMips32();
875 }
876 
877 bool MipsTargetLowering::isCheapToSpeculateCtlz() const {
878   return Subtarget.hasMips32();
879 }
880 
881 void
882 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
883                                           SmallVectorImpl<SDValue> &Results,
884                                           SelectionDAG &DAG) const {
885   SDValue Res = LowerOperation(SDValue(N, 0), DAG);
886 
887   for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
888     Results.push_back(Res.getValue(I));
889 }
890 
891 void
892 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
893                                        SmallVectorImpl<SDValue> &Results,
894                                        SelectionDAG &DAG) const {
895   return LowerOperationWrapper(N, Results, DAG);
896 }
897 
898 SDValue MipsTargetLowering::
899 LowerOperation(SDValue Op, SelectionDAG &DAG) const
900 {
901   switch (Op.getOpcode())
902   {
903   case ISD::BR_JT:              return lowerBR_JT(Op, DAG);
904   case ISD::BRCOND:             return lowerBRCOND(Op, DAG);
905   case ISD::ConstantPool:       return lowerConstantPool(Op, DAG);
906   case ISD::GlobalAddress:      return lowerGlobalAddress(Op, DAG);
907   case ISD::BlockAddress:       return lowerBlockAddress(Op, DAG);
908   case ISD::GlobalTLSAddress:   return lowerGlobalTLSAddress(Op, DAG);
909   case ISD::JumpTable:          return lowerJumpTable(Op, DAG);
910   case ISD::SELECT:             return lowerSELECT(Op, DAG);
911   case ISD::SETCC:              return lowerSETCC(Op, DAG);
912   case ISD::VASTART:            return lowerVASTART(Op, DAG);
913   case ISD::VAARG:              return lowerVAARG(Op, DAG);
914   case ISD::FCOPYSIGN:          return lowerFCOPYSIGN(Op, DAG);
915   case ISD::FRAMEADDR:          return lowerFRAMEADDR(Op, DAG);
916   case ISD::RETURNADDR:         return lowerRETURNADDR(Op, DAG);
917   case ISD::EH_RETURN:          return lowerEH_RETURN(Op, DAG);
918   case ISD::ATOMIC_FENCE:       return lowerATOMIC_FENCE(Op, DAG);
919   case ISD::SHL_PARTS:          return lowerShiftLeftParts(Op, DAG);
920   case ISD::SRA_PARTS:          return lowerShiftRightParts(Op, DAG, true);
921   case ISD::SRL_PARTS:          return lowerShiftRightParts(Op, DAG, false);
922   case ISD::LOAD:               return lowerLOAD(Op, DAG);
923   case ISD::STORE:              return lowerSTORE(Op, DAG);
924   case ISD::EH_DWARF_CFA:       return lowerEH_DWARF_CFA(Op, DAG);
925   case ISD::FP_TO_SINT:         return lowerFP_TO_SINT(Op, DAG);
926   }
927   return SDValue();
928 }
929 
930 //===----------------------------------------------------------------------===//
931 //  Lower helper functions
932 //===----------------------------------------------------------------------===//
933 
934 // addLiveIn - This helper function adds the specified physical register to the
935 // MachineFunction as a live in value.  It also creates a corresponding
936 // virtual register for it.
937 static unsigned
938 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
939 {
940   unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
941   MF.getRegInfo().addLiveIn(PReg, VReg);
942   return VReg;
943 }
944 
945 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI,
946                                               MachineBasicBlock &MBB,
947                                               const TargetInstrInfo &TII,
948                                               bool Is64Bit, bool IsMicroMips) {
949   if (NoZeroDivCheck)
950     return &MBB;
951 
952   // Insert instruction "teq $divisor_reg, $zero, 7".
953   MachineBasicBlock::iterator I(MI);
954   MachineInstrBuilder MIB;
955   MachineOperand &Divisor = MI.getOperand(2);
956   MIB = BuildMI(MBB, std::next(I), MI.getDebugLoc(),
957                 TII.get(IsMicroMips ? Mips::TEQ_MM : Mips::TEQ))
958             .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill()))
959             .addReg(Mips::ZERO)
960             .addImm(7);
961 
962   // Use the 32-bit sub-register if this is a 64-bit division.
963   if (Is64Bit)
964     MIB->getOperand(0).setSubReg(Mips::sub_32);
965 
966   // Clear Divisor's kill flag.
967   Divisor.setIsKill(false);
968 
969   // We would normally delete the original instruction here but in this case
970   // we only needed to inject an additional instruction rather than replace it.
971 
972   return &MBB;
973 }
974 
975 MachineBasicBlock *
976 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
977                                                 MachineBasicBlock *BB) const {
978   switch (MI.getOpcode()) {
979   default:
980     llvm_unreachable("Unexpected instr type to insert");
981   case Mips::ATOMIC_LOAD_ADD_I8:
982     return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
983   case Mips::ATOMIC_LOAD_ADD_I16:
984     return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
985   case Mips::ATOMIC_LOAD_ADD_I32:
986     return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
987   case Mips::ATOMIC_LOAD_ADD_I64:
988     return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
989 
990   case Mips::ATOMIC_LOAD_AND_I8:
991     return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
992   case Mips::ATOMIC_LOAD_AND_I16:
993     return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
994   case Mips::ATOMIC_LOAD_AND_I32:
995     return emitAtomicBinary(MI, BB, 4, Mips::AND);
996   case Mips::ATOMIC_LOAD_AND_I64:
997     return emitAtomicBinary(MI, BB, 8, Mips::AND64);
998 
999   case Mips::ATOMIC_LOAD_OR_I8:
1000     return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
1001   case Mips::ATOMIC_LOAD_OR_I16:
1002     return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
1003   case Mips::ATOMIC_LOAD_OR_I32:
1004     return emitAtomicBinary(MI, BB, 4, Mips::OR);
1005   case Mips::ATOMIC_LOAD_OR_I64:
1006     return emitAtomicBinary(MI, BB, 8, Mips::OR64);
1007 
1008   case Mips::ATOMIC_LOAD_XOR_I8:
1009     return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
1010   case Mips::ATOMIC_LOAD_XOR_I16:
1011     return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
1012   case Mips::ATOMIC_LOAD_XOR_I32:
1013     return emitAtomicBinary(MI, BB, 4, Mips::XOR);
1014   case Mips::ATOMIC_LOAD_XOR_I64:
1015     return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
1016 
1017   case Mips::ATOMIC_LOAD_NAND_I8:
1018     return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
1019   case Mips::ATOMIC_LOAD_NAND_I16:
1020     return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
1021   case Mips::ATOMIC_LOAD_NAND_I32:
1022     return emitAtomicBinary(MI, BB, 4, 0, true);
1023   case Mips::ATOMIC_LOAD_NAND_I64:
1024     return emitAtomicBinary(MI, BB, 8, 0, true);
1025 
1026   case Mips::ATOMIC_LOAD_SUB_I8:
1027     return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
1028   case Mips::ATOMIC_LOAD_SUB_I16:
1029     return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
1030   case Mips::ATOMIC_LOAD_SUB_I32:
1031     return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
1032   case Mips::ATOMIC_LOAD_SUB_I64:
1033     return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
1034 
1035   case Mips::ATOMIC_SWAP_I8:
1036     return emitAtomicBinaryPartword(MI, BB, 1, 0);
1037   case Mips::ATOMIC_SWAP_I16:
1038     return emitAtomicBinaryPartword(MI, BB, 2, 0);
1039   case Mips::ATOMIC_SWAP_I32:
1040     return emitAtomicBinary(MI, BB, 4, 0);
1041   case Mips::ATOMIC_SWAP_I64:
1042     return emitAtomicBinary(MI, BB, 8, 0);
1043 
1044   case Mips::ATOMIC_CMP_SWAP_I8:
1045     return emitAtomicCmpSwapPartword(MI, BB, 1);
1046   case Mips::ATOMIC_CMP_SWAP_I16:
1047     return emitAtomicCmpSwapPartword(MI, BB, 2);
1048   case Mips::ATOMIC_CMP_SWAP_I32:
1049     return emitAtomicCmpSwap(MI, BB, 4);
1050   case Mips::ATOMIC_CMP_SWAP_I64:
1051     return emitAtomicCmpSwap(MI, BB, 8);
1052   case Mips::PseudoSDIV:
1053   case Mips::PseudoUDIV:
1054   case Mips::DIV:
1055   case Mips::DIVU:
1056   case Mips::MOD:
1057   case Mips::MODU:
1058     return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false,
1059                                false);
1060   case Mips::SDIV_MM_Pseudo:
1061   case Mips::UDIV_MM_Pseudo:
1062   case Mips::SDIV_MM:
1063   case Mips::UDIV_MM:
1064   case Mips::DIV_MMR6:
1065   case Mips::DIVU_MMR6:
1066   case Mips::MOD_MMR6:
1067   case Mips::MODU_MMR6:
1068     return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, true);
1069   case Mips::PseudoDSDIV:
1070   case Mips::PseudoDUDIV:
1071   case Mips::DDIV:
1072   case Mips::DDIVU:
1073   case Mips::DMOD:
1074   case Mips::DMODU:
1075     return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, false);
1076   case Mips::DDIV_MM64R6:
1077   case Mips::DDIVU_MM64R6:
1078   case Mips::DMOD_MM64R6:
1079   case Mips::DMODU_MM64R6:
1080     return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, true);
1081   case Mips::SEL_D:
1082   case Mips::SEL_D_MMR6:
1083     return emitSEL_D(MI, BB);
1084 
1085   case Mips::PseudoSELECT_I:
1086   case Mips::PseudoSELECT_I64:
1087   case Mips::PseudoSELECT_S:
1088   case Mips::PseudoSELECT_D32:
1089   case Mips::PseudoSELECT_D64:
1090     return emitPseudoSELECT(MI, BB, false, Mips::BNE);
1091   case Mips::PseudoSELECTFP_F_I:
1092   case Mips::PseudoSELECTFP_F_I64:
1093   case Mips::PseudoSELECTFP_F_S:
1094   case Mips::PseudoSELECTFP_F_D32:
1095   case Mips::PseudoSELECTFP_F_D64:
1096     return emitPseudoSELECT(MI, BB, true, Mips::BC1F);
1097   case Mips::PseudoSELECTFP_T_I:
1098   case Mips::PseudoSELECTFP_T_I64:
1099   case Mips::PseudoSELECTFP_T_S:
1100   case Mips::PseudoSELECTFP_T_D32:
1101   case Mips::PseudoSELECTFP_T_D64:
1102     return emitPseudoSELECT(MI, BB, true, Mips::BC1T);
1103   }
1104 }
1105 
1106 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1107 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1108 MachineBasicBlock *MipsTargetLowering::emitAtomicBinary(MachineInstr &MI,
1109                                                         MachineBasicBlock *BB,
1110                                                         unsigned Size,
1111                                                         unsigned BinOpcode,
1112                                                         bool Nand) const {
1113   assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1114 
1115   MachineFunction *MF = BB->getParent();
1116   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1117   const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1118   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1119   const bool ArePtrs64bit = ABI.ArePtrs64bit();
1120   DebugLoc DL = MI.getDebugLoc();
1121   unsigned LL, SC, AND, NOR, ZERO, BEQ;
1122 
1123   if (Size == 4) {
1124     if (isMicroMips) {
1125       LL = Mips::LL_MM;
1126       SC = Mips::SC_MM;
1127     } else {
1128       LL = Subtarget.hasMips32r6()
1129                ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6)
1130                : (ArePtrs64bit ? Mips::LL64 : Mips::LL);
1131       SC = Subtarget.hasMips32r6()
1132                ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6)
1133                : (ArePtrs64bit ? Mips::SC64 : Mips::SC);
1134     }
1135 
1136     AND = Mips::AND;
1137     NOR = Mips::NOR;
1138     ZERO = Mips::ZERO;
1139     BEQ = Mips::BEQ;
1140   } else {
1141     LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD;
1142     SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD;
1143     AND = Mips::AND64;
1144     NOR = Mips::NOR64;
1145     ZERO = Mips::ZERO_64;
1146     BEQ = Mips::BEQ64;
1147   }
1148 
1149   unsigned OldVal = MI.getOperand(0).getReg();
1150   unsigned Ptr = MI.getOperand(1).getReg();
1151   unsigned Incr = MI.getOperand(2).getReg();
1152 
1153   unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1154   unsigned AndRes = RegInfo.createVirtualRegister(RC);
1155   unsigned Success = RegInfo.createVirtualRegister(RC);
1156 
1157   // insert new blocks after the current block
1158   const BasicBlock *LLVM_BB = BB->getBasicBlock();
1159   MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1160   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1161   MachineFunction::iterator It = ++BB->getIterator();
1162   MF->insert(It, loopMBB);
1163   MF->insert(It, exitMBB);
1164 
1165   // Transfer the remainder of BB and its successor edges to exitMBB.
1166   exitMBB->splice(exitMBB->begin(), BB,
1167                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
1168   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1169 
1170   //  thisMBB:
1171   //    ...
1172   //    fallthrough --> loopMBB
1173   BB->addSuccessor(loopMBB);
1174   loopMBB->addSuccessor(loopMBB);
1175   loopMBB->addSuccessor(exitMBB);
1176 
1177   //  loopMBB:
1178   //    ll oldval, 0(ptr)
1179   //    <binop> storeval, oldval, incr
1180   //    sc success, storeval, 0(ptr)
1181   //    beq success, $0, loopMBB
1182   BB = loopMBB;
1183   BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1184   if (Nand) {
1185     //  and andres, oldval, incr
1186     //  nor storeval, $0, andres
1187     BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1188     BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1189   } else if (BinOpcode) {
1190     //  <binop> storeval, oldval, incr
1191     BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1192   } else {
1193     StoreVal = Incr;
1194   }
1195   BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1196   BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1197 
1198   MI.eraseFromParent(); // The instruction is gone now.
1199 
1200   return exitMBB;
1201 }
1202 
1203 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg(
1204     MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg,
1205     unsigned SrcReg) const {
1206   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1207   const DebugLoc &DL = MI.getDebugLoc();
1208 
1209   if (Subtarget.hasMips32r2() && Size == 1) {
1210     BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg);
1211     return BB;
1212   }
1213 
1214   if (Subtarget.hasMips32r2() && Size == 2) {
1215     BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg);
1216     return BB;
1217   }
1218 
1219   MachineFunction *MF = BB->getParent();
1220   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1221   const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1222   unsigned ScrReg = RegInfo.createVirtualRegister(RC);
1223 
1224   assert(Size < 32);
1225   int64_t ShiftImm = 32 - (Size * 8);
1226 
1227   BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm);
1228   BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm);
1229 
1230   return BB;
1231 }
1232 
1233 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword(
1234     MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned BinOpcode,
1235     bool Nand) const {
1236   assert((Size == 1 || Size == 2) &&
1237          "Unsupported size for EmitAtomicBinaryPartial.");
1238 
1239   MachineFunction *MF = BB->getParent();
1240   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1241   const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1242   const bool ArePtrs64bit = ABI.ArePtrs64bit();
1243   const TargetRegisterClass *RCp =
1244     getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32);
1245   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1246   DebugLoc DL = MI.getDebugLoc();
1247 
1248   unsigned Dest = MI.getOperand(0).getReg();
1249   unsigned Ptr = MI.getOperand(1).getReg();
1250   unsigned Incr = MI.getOperand(2).getReg();
1251 
1252   unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp);
1253   unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1254   unsigned Mask = RegInfo.createVirtualRegister(RC);
1255   unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1256   unsigned NewVal = RegInfo.createVirtualRegister(RC);
1257   unsigned OldVal = RegInfo.createVirtualRegister(RC);
1258   unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1259   unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp);
1260   unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1261   unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1262   unsigned AndRes = RegInfo.createVirtualRegister(RC);
1263   unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1264   unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1265   unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1266   unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1267   unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1268   unsigned Success = RegInfo.createVirtualRegister(RC);
1269 
1270   unsigned LL, SC;
1271   if (isMicroMips) {
1272     LL = Mips::LL_MM;
1273     SC = Mips::SC_MM;
1274   } else {
1275     LL = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6)
1276                                  : (ArePtrs64bit ? Mips::LL64 : Mips::LL);
1277     SC = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6)
1278                                  : (ArePtrs64bit ? Mips::SC64 : Mips::SC);
1279   }
1280 
1281   // insert new blocks after the current block
1282   const BasicBlock *LLVM_BB = BB->getBasicBlock();
1283   MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1284   MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1285   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1286   MachineFunction::iterator It = ++BB->getIterator();
1287   MF->insert(It, loopMBB);
1288   MF->insert(It, sinkMBB);
1289   MF->insert(It, exitMBB);
1290 
1291   // Transfer the remainder of BB and its successor edges to exitMBB.
1292   exitMBB->splice(exitMBB->begin(), BB,
1293                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
1294   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1295 
1296   BB->addSuccessor(loopMBB);
1297   loopMBB->addSuccessor(loopMBB);
1298   loopMBB->addSuccessor(sinkMBB);
1299   sinkMBB->addSuccessor(exitMBB);
1300 
1301   //  thisMBB:
1302   //    addiu   masklsb2,$0,-4                # 0xfffffffc
1303   //    and     alignedaddr,ptr,masklsb2
1304   //    andi    ptrlsb2,ptr,3
1305   //    sll     shiftamt,ptrlsb2,3
1306   //    ori     maskupper,$0,255               # 0xff
1307   //    sll     mask,maskupper,shiftamt
1308   //    nor     mask2,$0,mask
1309   //    sll     incr2,incr,shiftamt
1310 
1311   int64_t MaskImm = (Size == 1) ? 255 : 65535;
1312   BuildMI(BB, DL, TII->get(ABI.GetPtrAddiuOp()), MaskLSB2)
1313     .addReg(ABI.GetNullPtr()).addImm(-4);
1314   BuildMI(BB, DL, TII->get(ABI.GetPtrAndOp()), AlignedAddr)
1315     .addReg(Ptr).addReg(MaskLSB2);
1316   BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2)
1317       .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3);
1318   if (Subtarget.isLittle()) {
1319     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1320   } else {
1321     unsigned Off = RegInfo.createVirtualRegister(RC);
1322     BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1323       .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1324     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1325   }
1326   BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1327     .addReg(Mips::ZERO).addImm(MaskImm);
1328   BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1329     .addReg(MaskUpper).addReg(ShiftAmt);
1330   BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1331   BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1332 
1333   // atomic.load.binop
1334   // loopMBB:
1335   //   ll      oldval,0(alignedaddr)
1336   //   binop   binopres,oldval,incr2
1337   //   and     newval,binopres,mask
1338   //   and     maskedoldval0,oldval,mask2
1339   //   or      storeval,maskedoldval0,newval
1340   //   sc      success,storeval,0(alignedaddr)
1341   //   beq     success,$0,loopMBB
1342 
1343   // atomic.swap
1344   // loopMBB:
1345   //   ll      oldval,0(alignedaddr)
1346   //   and     newval,incr2,mask
1347   //   and     maskedoldval0,oldval,mask2
1348   //   or      storeval,maskedoldval0,newval
1349   //   sc      success,storeval,0(alignedaddr)
1350   //   beq     success,$0,loopMBB
1351 
1352   BB = loopMBB;
1353   BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1354   if (Nand) {
1355     //  and andres, oldval, incr2
1356     //  nor binopres, $0, andres
1357     //  and newval, binopres, mask
1358     BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1359     BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1360       .addReg(Mips::ZERO).addReg(AndRes);
1361     BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1362   } else if (BinOpcode) {
1363     //  <binop> binopres, oldval, incr2
1364     //  and newval, binopres, mask
1365     BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1366     BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1367   } else { // atomic.swap
1368     //  and newval, incr2, mask
1369     BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1370   }
1371 
1372   BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1373     .addReg(OldVal).addReg(Mask2);
1374   BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1375     .addReg(MaskedOldVal0).addReg(NewVal);
1376   BuildMI(BB, DL, TII->get(SC), Success)
1377     .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1378   BuildMI(BB, DL, TII->get(Mips::BEQ))
1379     .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1380 
1381   //  sinkMBB:
1382   //    and     maskedoldval1,oldval,mask
1383   //    srl     srlres,maskedoldval1,shiftamt
1384   //    sign_extend dest,srlres
1385   BB = sinkMBB;
1386 
1387   BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1388     .addReg(OldVal).addReg(Mask);
1389   BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1390       .addReg(MaskedOldVal1).addReg(ShiftAmt);
1391   BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes);
1392 
1393   MI.eraseFromParent(); // The instruction is gone now.
1394 
1395   return exitMBB;
1396 }
1397 
1398 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI,
1399                                                          MachineBasicBlock *BB,
1400                                                          unsigned Size) const {
1401   assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1402 
1403   MachineFunction *MF = BB->getParent();
1404   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1405   const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1406   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1407   const bool ArePtrs64bit = ABI.ArePtrs64bit();
1408   DebugLoc DL = MI.getDebugLoc();
1409   unsigned LL, SC, ZERO, BNE, BEQ;
1410 
1411   if (Size == 4) {
1412     if (isMicroMips) {
1413       LL = Mips::LL_MM;
1414       SC = Mips::SC_MM;
1415     } else {
1416       LL = Subtarget.hasMips32r6()
1417                ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6)
1418                : (ArePtrs64bit ? Mips::LL64 : Mips::LL);
1419       SC = Subtarget.hasMips32r6()
1420                ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6)
1421                : (ArePtrs64bit ? Mips::SC64 : Mips::SC);
1422     }
1423 
1424     ZERO = Mips::ZERO;
1425     BNE = Mips::BNE;
1426     BEQ = Mips::BEQ;
1427   } else {
1428     LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD;
1429     SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD;
1430     ZERO = Mips::ZERO_64;
1431     BNE = Mips::BNE64;
1432     BEQ = Mips::BEQ64;
1433   }
1434 
1435   unsigned Dest = MI.getOperand(0).getReg();
1436   unsigned Ptr = MI.getOperand(1).getReg();
1437   unsigned OldVal = MI.getOperand(2).getReg();
1438   unsigned NewVal = MI.getOperand(3).getReg();
1439 
1440   unsigned Success = RegInfo.createVirtualRegister(RC);
1441 
1442   // insert new blocks after the current block
1443   const BasicBlock *LLVM_BB = BB->getBasicBlock();
1444   MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1445   MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1446   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1447   MachineFunction::iterator It = ++BB->getIterator();
1448   MF->insert(It, loop1MBB);
1449   MF->insert(It, loop2MBB);
1450   MF->insert(It, exitMBB);
1451 
1452   // Transfer the remainder of BB and its successor edges to exitMBB.
1453   exitMBB->splice(exitMBB->begin(), BB,
1454                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
1455   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1456 
1457   //  thisMBB:
1458   //    ...
1459   //    fallthrough --> loop1MBB
1460   BB->addSuccessor(loop1MBB);
1461   loop1MBB->addSuccessor(exitMBB);
1462   loop1MBB->addSuccessor(loop2MBB);
1463   loop2MBB->addSuccessor(loop1MBB);
1464   loop2MBB->addSuccessor(exitMBB);
1465 
1466   // loop1MBB:
1467   //   ll dest, 0(ptr)
1468   //   bne dest, oldval, exitMBB
1469   BB = loop1MBB;
1470   BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1471   BuildMI(BB, DL, TII->get(BNE))
1472     .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1473 
1474   // loop2MBB:
1475   //   sc success, newval, 0(ptr)
1476   //   beq success, $0, loop1MBB
1477   BB = loop2MBB;
1478   BuildMI(BB, DL, TII->get(SC), Success)
1479     .addReg(NewVal).addReg(Ptr).addImm(0);
1480   BuildMI(BB, DL, TII->get(BEQ))
1481     .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1482 
1483   MI.eraseFromParent(); // The instruction is gone now.
1484 
1485   return exitMBB;
1486 }
1487 
1488 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword(
1489     MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const {
1490   assert((Size == 1 || Size == 2) &&
1491       "Unsupported size for EmitAtomicCmpSwapPartial.");
1492 
1493   MachineFunction *MF = BB->getParent();
1494   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1495   const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1496   const bool ArePtrs64bit = ABI.ArePtrs64bit();
1497   const TargetRegisterClass *RCp =
1498     getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32);
1499   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1500   DebugLoc DL = MI.getDebugLoc();
1501 
1502   unsigned Dest = MI.getOperand(0).getReg();
1503   unsigned Ptr = MI.getOperand(1).getReg();
1504   unsigned CmpVal = MI.getOperand(2).getReg();
1505   unsigned NewVal = MI.getOperand(3).getReg();
1506 
1507   unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp);
1508   unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1509   unsigned Mask = RegInfo.createVirtualRegister(RC);
1510   unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1511   unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1512   unsigned OldVal = RegInfo.createVirtualRegister(RC);
1513   unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1514   unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1515   unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp);
1516   unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1517   unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1518   unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1519   unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1520   unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1521   unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1522   unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1523   unsigned Success = RegInfo.createVirtualRegister(RC);
1524   unsigned LL, SC;
1525 
1526   if (isMicroMips) {
1527     LL = Mips::LL_MM;
1528     SC = Mips::SC_MM;
1529   } else {
1530     LL = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::LL64_R6 : Mips::LL_R6)
1531                                  : (ArePtrs64bit ? Mips::LL64 : Mips::LL);
1532     SC = Subtarget.hasMips32r6() ? (ArePtrs64bit ? Mips::SC64_R6 : Mips::SC_R6)
1533                                  : (ArePtrs64bit ? Mips::SC64 : Mips::SC);
1534   }
1535 
1536   // insert new blocks after the current block
1537   const BasicBlock *LLVM_BB = BB->getBasicBlock();
1538   MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1539   MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1540   MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1541   MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1542   MachineFunction::iterator It = ++BB->getIterator();
1543   MF->insert(It, loop1MBB);
1544   MF->insert(It, loop2MBB);
1545   MF->insert(It, sinkMBB);
1546   MF->insert(It, exitMBB);
1547 
1548   // Transfer the remainder of BB and its successor edges to exitMBB.
1549   exitMBB->splice(exitMBB->begin(), BB,
1550                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
1551   exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1552 
1553   BB->addSuccessor(loop1MBB);
1554   loop1MBB->addSuccessor(sinkMBB);
1555   loop1MBB->addSuccessor(loop2MBB);
1556   loop2MBB->addSuccessor(loop1MBB);
1557   loop2MBB->addSuccessor(sinkMBB);
1558   sinkMBB->addSuccessor(exitMBB);
1559 
1560   // FIXME: computation of newval2 can be moved to loop2MBB.
1561   //  thisMBB:
1562   //    addiu   masklsb2,$0,-4                # 0xfffffffc
1563   //    and     alignedaddr,ptr,masklsb2
1564   //    andi    ptrlsb2,ptr,3
1565   //    xori    ptrlsb2,ptrlsb2,3              # Only for BE
1566   //    sll     shiftamt,ptrlsb2,3
1567   //    ori     maskupper,$0,255               # 0xff
1568   //    sll     mask,maskupper,shiftamt
1569   //    nor     mask2,$0,mask
1570   //    andi    maskedcmpval,cmpval,255
1571   //    sll     shiftedcmpval,maskedcmpval,shiftamt
1572   //    andi    maskednewval,newval,255
1573   //    sll     shiftednewval,maskednewval,shiftamt
1574   int64_t MaskImm = (Size == 1) ? 255 : 65535;
1575   BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), MaskLSB2)
1576     .addReg(ABI.GetNullPtr()).addImm(-4);
1577   BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::AND64 : Mips::AND), AlignedAddr)
1578     .addReg(Ptr).addReg(MaskLSB2);
1579   BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2)
1580       .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3);
1581   if (Subtarget.isLittle()) {
1582     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1583   } else {
1584     unsigned Off = RegInfo.createVirtualRegister(RC);
1585     BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1586       .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1587     BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1588   }
1589   BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1590     .addReg(Mips::ZERO).addImm(MaskImm);
1591   BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1592     .addReg(MaskUpper).addReg(ShiftAmt);
1593   BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1594   BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1595     .addReg(CmpVal).addImm(MaskImm);
1596   BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1597     .addReg(MaskedCmpVal).addReg(ShiftAmt);
1598   BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1599     .addReg(NewVal).addImm(MaskImm);
1600   BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1601     .addReg(MaskedNewVal).addReg(ShiftAmt);
1602 
1603   //  loop1MBB:
1604   //    ll      oldval,0(alginedaddr)
1605   //    and     maskedoldval0,oldval,mask
1606   //    bne     maskedoldval0,shiftedcmpval,sinkMBB
1607   BB = loop1MBB;
1608   BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1609   BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1610     .addReg(OldVal).addReg(Mask);
1611   BuildMI(BB, DL, TII->get(Mips::BNE))
1612     .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1613 
1614   //  loop2MBB:
1615   //    and     maskedoldval1,oldval,mask2
1616   //    or      storeval,maskedoldval1,shiftednewval
1617   //    sc      success,storeval,0(alignedaddr)
1618   //    beq     success,$0,loop1MBB
1619   BB = loop2MBB;
1620   BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1621     .addReg(OldVal).addReg(Mask2);
1622   BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1623     .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1624   BuildMI(BB, DL, TII->get(SC), Success)
1625       .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1626   BuildMI(BB, DL, TII->get(Mips::BEQ))
1627       .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1628 
1629   //  sinkMBB:
1630   //    srl     srlres,maskedoldval0,shiftamt
1631   //    sign_extend dest,srlres
1632   BB = sinkMBB;
1633 
1634   BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1635       .addReg(MaskedOldVal0).addReg(ShiftAmt);
1636   BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes);
1637 
1638   MI.eraseFromParent(); // The instruction is gone now.
1639 
1640   return exitMBB;
1641 }
1642 
1643 MachineBasicBlock *MipsTargetLowering::emitSEL_D(MachineInstr &MI,
1644                                                  MachineBasicBlock *BB) const {
1645   MachineFunction *MF = BB->getParent();
1646   const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
1647   const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1648   MachineRegisterInfo &RegInfo = MF->getRegInfo();
1649   DebugLoc DL = MI.getDebugLoc();
1650   MachineBasicBlock::iterator II(MI);
1651 
1652   unsigned Fc = MI.getOperand(1).getReg();
1653   const auto &FGR64RegClass = TRI->getRegClass(Mips::FGR64RegClassID);
1654 
1655   unsigned Fc2 = RegInfo.createVirtualRegister(FGR64RegClass);
1656 
1657   BuildMI(*BB, II, DL, TII->get(Mips::SUBREG_TO_REG), Fc2)
1658       .addImm(0)
1659       .addReg(Fc)
1660       .addImm(Mips::sub_lo);
1661 
1662   // We don't erase the original instruction, we just replace the condition
1663   // register with the 64-bit super-register.
1664   MI.getOperand(1).setReg(Fc2);
1665 
1666   return BB;
1667 }
1668 
1669 //===----------------------------------------------------------------------===//
1670 //  Misc Lower Operation implementation
1671 //===----------------------------------------------------------------------===//
1672 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1673   SDValue Chain = Op.getOperand(0);
1674   SDValue Table = Op.getOperand(1);
1675   SDValue Index = Op.getOperand(2);
1676   SDLoc DL(Op);
1677   auto &TD = DAG.getDataLayout();
1678   EVT PTy = getPointerTy(TD);
1679   unsigned EntrySize =
1680       DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
1681 
1682   Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1683                       DAG.getConstant(EntrySize, DL, PTy));
1684   SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1685 
1686   EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1687   Addr = DAG.getExtLoad(
1688       ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1689       MachinePointerInfo::getJumpTable(DAG.getMachineFunction()), MemVT);
1690   Chain = Addr.getValue(1);
1691 
1692   if (isPositionIndependent() || ABI.IsN64()) {
1693     // For PIC, the sequence is:
1694     // BRIND(load(Jumptable + index) + RelocBase)
1695     // RelocBase can be JumpTable, GOT or some sort of global base.
1696     Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1697                        getPICJumpTableRelocBase(Table, DAG));
1698   }
1699 
1700   return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1701 }
1702 
1703 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1704   // The first operand is the chain, the second is the condition, the third is
1705   // the block to branch to if the condition is true.
1706   SDValue Chain = Op.getOperand(0);
1707   SDValue Dest = Op.getOperand(2);
1708   SDLoc DL(Op);
1709 
1710   assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1711   SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1712 
1713   // Return if flag is not set by a floating point comparison.
1714   if (CondRes.getOpcode() != MipsISD::FPCmp)
1715     return Op;
1716 
1717   SDValue CCNode  = CondRes.getOperand(2);
1718   Mips::CondCode CC =
1719     (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1720   unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1721   SDValue BrCode = DAG.getConstant(Opc, DL, MVT::i32);
1722   SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
1723   return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1724                      FCC0, Dest, CondRes);
1725 }
1726 
1727 SDValue MipsTargetLowering::
1728 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1729 {
1730   assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1731   SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1732 
1733   // Return if flag is not set by a floating point comparison.
1734   if (Cond.getOpcode() != MipsISD::FPCmp)
1735     return Op;
1736 
1737   return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1738                       SDLoc(Op));
1739 }
1740 
1741 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1742   assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1743   SDValue Cond = createFPCmp(DAG, Op);
1744 
1745   assert(Cond.getOpcode() == MipsISD::FPCmp &&
1746          "Floating point operand expected.");
1747 
1748   SDLoc DL(Op);
1749   SDValue True  = DAG.getConstant(1, DL, MVT::i32);
1750   SDValue False = DAG.getConstant(0, DL, MVT::i32);
1751 
1752   return createCMovFP(DAG, Cond, True, False, DL);
1753 }
1754 
1755 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1756                                                SelectionDAG &DAG) const {
1757   EVT Ty = Op.getValueType();
1758   GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
1759   const GlobalValue *GV = N->getGlobal();
1760 
1761   if (!isPositionIndependent() && !ABI.IsN64()) {
1762     const MipsTargetObjectFile *TLOF =
1763         static_cast<const MipsTargetObjectFile *>(
1764             getTargetMachine().getObjFileLowering());
1765     if (TLOF->IsGlobalInSmallSection(GV, getTargetMachine()))
1766       // %gp_rel relocation
1767       return getAddrGPRel(N, SDLoc(N), Ty, DAG);
1768 
1769     // %hi/%lo relocation
1770     return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1771   }
1772 
1773   // Every other architecture would use shouldAssumeDSOLocal in here, but
1774   // mips is special.
1775   // * In PIC code mips requires got loads even for local statics!
1776   // * To save on got entries, for local statics the got entry contains the
1777   //   page and an additional add instruction takes care of the low bits.
1778   // * It is legal to access a hidden symbol with a non hidden undefined,
1779   //   so one cannot guarantee that all access to a hidden symbol will know
1780   //   it is hidden.
1781   // * Mips linkers don't support creating a page and a full got entry for
1782   //   the same symbol.
1783   // * Given all that, we have to use a full got entry for hidden symbols :-(
1784   if (GV->hasLocalLinkage())
1785     return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1786 
1787   if (LargeGOT)
1788     return getAddrGlobalLargeGOT(
1789         N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, MipsII::MO_GOT_LO16,
1790         DAG.getEntryNode(),
1791         MachinePointerInfo::getGOT(DAG.getMachineFunction()));
1792 
1793   return getAddrGlobal(
1794       N, SDLoc(N), Ty, DAG,
1795       (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT,
1796       DAG.getEntryNode(), MachinePointerInfo::getGOT(DAG.getMachineFunction()));
1797 }
1798 
1799 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1800                                               SelectionDAG &DAG) const {
1801   BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
1802   EVT Ty = Op.getValueType();
1803 
1804   if (!isPositionIndependent() && !ABI.IsN64())
1805     return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1806 
1807   return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1808 }
1809 
1810 SDValue MipsTargetLowering::
1811 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1812 {
1813   // If the relocation model is PIC, use the General Dynamic TLS Model or
1814   // Local Dynamic TLS model, otherwise use the Initial Exec or
1815   // Local Exec TLS Model.
1816 
1817   GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1818   if (DAG.getTarget().Options.EmulatedTLS)
1819     return LowerToTLSEmulatedModel(GA, DAG);
1820 
1821   SDLoc DL(GA);
1822   const GlobalValue *GV = GA->getGlobal();
1823   EVT PtrVT = getPointerTy(DAG.getDataLayout());
1824 
1825   TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1826 
1827   if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1828     // General Dynamic and Local Dynamic TLS Model.
1829     unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1830                                                       : MipsII::MO_TLSGD;
1831 
1832     SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1833     SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1834                                    getGlobalReg(DAG, PtrVT), TGA);
1835     unsigned PtrSize = PtrVT.getSizeInBits();
1836     IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1837 
1838     SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1839 
1840     ArgListTy Args;
1841     ArgListEntry Entry;
1842     Entry.Node = Argument;
1843     Entry.Ty = PtrTy;
1844     Args.push_back(Entry);
1845 
1846     TargetLowering::CallLoweringInfo CLI(DAG);
1847     CLI.setDebugLoc(DL).setChain(DAG.getEntryNode())
1848       .setCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args));
1849     std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1850 
1851     SDValue Ret = CallResult.first;
1852 
1853     if (model != TLSModel::LocalDynamic)
1854       return Ret;
1855 
1856     SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1857                                                MipsII::MO_DTPREL_HI);
1858     SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1859     SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1860                                                MipsII::MO_DTPREL_LO);
1861     SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1862     SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1863     return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1864   }
1865 
1866   SDValue Offset;
1867   if (model == TLSModel::InitialExec) {
1868     // Initial Exec TLS Model
1869     SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1870                                              MipsII::MO_GOTTPREL);
1871     TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1872                       TGA);
1873     Offset =
1874         DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), TGA, MachinePointerInfo());
1875   } else {
1876     // Local Exec TLS Model
1877     assert(model == TLSModel::LocalExec);
1878     SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1879                                                MipsII::MO_TPREL_HI);
1880     SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1881                                                MipsII::MO_TPREL_LO);
1882     SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1883     SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1884     Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1885   }
1886 
1887   SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1888   return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1889 }
1890 
1891 SDValue MipsTargetLowering::
1892 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1893 {
1894   JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
1895   EVT Ty = Op.getValueType();
1896 
1897   if (!isPositionIndependent() && !ABI.IsN64())
1898     return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1899 
1900   return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1901 }
1902 
1903 SDValue MipsTargetLowering::
1904 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1905 {
1906   ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1907   EVT Ty = Op.getValueType();
1908 
1909   if (!isPositionIndependent() && !ABI.IsN64()) {
1910     const MipsTargetObjectFile *TLOF =
1911         static_cast<const MipsTargetObjectFile *>(
1912             getTargetMachine().getObjFileLowering());
1913 
1914     if (TLOF->IsConstantInSmallSection(DAG.getDataLayout(), N->getConstVal(),
1915                                        getTargetMachine()))
1916       // %gp_rel relocation
1917       return getAddrGPRel(N, SDLoc(N), Ty, DAG);
1918 
1919     return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1920   }
1921 
1922   return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1923 }
1924 
1925 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1926   MachineFunction &MF = DAG.getMachineFunction();
1927   MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1928 
1929   SDLoc DL(Op);
1930   SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1931                                  getPointerTy(MF.getDataLayout()));
1932 
1933   // vastart just stores the address of the VarArgsFrameIndex slot into the
1934   // memory location argument.
1935   const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1936   return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1937                       MachinePointerInfo(SV));
1938 }
1939 
1940 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const {
1941   SDNode *Node = Op.getNode();
1942   EVT VT = Node->getValueType(0);
1943   SDValue Chain = Node->getOperand(0);
1944   SDValue VAListPtr = Node->getOperand(1);
1945   unsigned Align = Node->getConstantOperandVal(3);
1946   const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1947   SDLoc DL(Node);
1948   unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4;
1949 
1950   SDValue VAListLoad = DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL, Chain,
1951                                    VAListPtr, MachinePointerInfo(SV));
1952   SDValue VAList = VAListLoad;
1953 
1954   // Re-align the pointer if necessary.
1955   // It should only ever be necessary for 64-bit types on O32 since the minimum
1956   // argument alignment is the same as the maximum type alignment for N32/N64.
1957   //
1958   // FIXME: We currently align too often. The code generator doesn't notice
1959   //        when the pointer is still aligned from the last va_arg (or pair of
1960   //        va_args for the i64 on O32 case).
1961   if (Align > getMinStackArgumentAlignment()) {
1962     assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
1963 
1964     VAList = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
1965                          DAG.getConstant(Align - 1, DL, VAList.getValueType()));
1966 
1967     VAList = DAG.getNode(ISD::AND, DL, VAList.getValueType(), VAList,
1968                          DAG.getConstant(-(int64_t)Align, DL,
1969                                          VAList.getValueType()));
1970   }
1971 
1972   // Increment the pointer, VAList, to the next vaarg.
1973   auto &TD = DAG.getDataLayout();
1974   unsigned ArgSizeInBytes =
1975       TD.getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext()));
1976   SDValue Tmp3 =
1977       DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
1978                   DAG.getConstant(alignTo(ArgSizeInBytes, ArgSlotSizeInBytes),
1979                                   DL, VAList.getValueType()));
1980   // Store the incremented VAList to the legalized pointer
1981   Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr,
1982                        MachinePointerInfo(SV));
1983 
1984   // In big-endian mode we must adjust the pointer when the load size is smaller
1985   // than the argument slot size. We must also reduce the known alignment to
1986   // match. For example in the N64 ABI, we must add 4 bytes to the offset to get
1987   // the correct half of the slot, and reduce the alignment from 8 (slot
1988   // alignment) down to 4 (type alignment).
1989   if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) {
1990     unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes;
1991     VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList,
1992                          DAG.getIntPtrConstant(Adjustment, DL));
1993   }
1994   // Load the actual argument out of the pointer VAList
1995   return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo());
1996 }
1997 
1998 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG,
1999                                 bool HasExtractInsert) {
2000   EVT TyX = Op.getOperand(0).getValueType();
2001   EVT TyY = Op.getOperand(1).getValueType();
2002   SDLoc DL(Op);
2003   SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
2004   SDValue Const31 = DAG.getConstant(31, DL, MVT::i32);
2005   SDValue Res;
2006 
2007   // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2008   // to i32.
2009   SDValue X = (TyX == MVT::f32) ?
2010     DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
2011     DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2012                 Const1);
2013   SDValue Y = (TyY == MVT::f32) ?
2014     DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
2015     DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
2016                 Const1);
2017 
2018   if (HasExtractInsert) {
2019     // ext  E, Y, 31, 1  ; extract bit31 of Y
2020     // ins  X, E, 31, 1  ; insert extracted bit at bit31 of X
2021     SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
2022     Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
2023   } else {
2024     // sll SllX, X, 1
2025     // srl SrlX, SllX, 1
2026     // srl SrlY, Y, 31
2027     // sll SllY, SrlX, 31
2028     // or  Or, SrlX, SllY
2029     SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2030     SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2031     SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
2032     SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
2033     Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
2034   }
2035 
2036   if (TyX == MVT::f32)
2037     return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
2038 
2039   SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2040                              Op.getOperand(0),
2041                              DAG.getConstant(0, DL, MVT::i32));
2042   return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2043 }
2044 
2045 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG,
2046                                 bool HasExtractInsert) {
2047   unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
2048   unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
2049   EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
2050   SDLoc DL(Op);
2051   SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
2052 
2053   // Bitcast to integer nodes.
2054   SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
2055   SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
2056 
2057   if (HasExtractInsert) {
2058     // ext  E, Y, width(Y) - 1, 1  ; extract bit width(Y)-1 of Y
2059     // ins  X, E, width(X) - 1, 1  ; insert extracted bit at bit width(X)-1 of X
2060     SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
2061                             DAG.getConstant(WidthY - 1, DL, MVT::i32), Const1);
2062 
2063     if (WidthX > WidthY)
2064       E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
2065     else if (WidthY > WidthX)
2066       E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
2067 
2068     SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
2069                             DAG.getConstant(WidthX - 1, DL, MVT::i32), Const1,
2070                             X);
2071     return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
2072   }
2073 
2074   // (d)sll SllX, X, 1
2075   // (d)srl SrlX, SllX, 1
2076   // (d)srl SrlY, Y, width(Y)-1
2077   // (d)sll SllY, SrlX, width(Y)-1
2078   // or     Or, SrlX, SllY
2079   SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
2080   SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
2081   SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
2082                              DAG.getConstant(WidthY - 1, DL, MVT::i32));
2083 
2084   if (WidthX > WidthY)
2085     SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
2086   else if (WidthY > WidthX)
2087     SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
2088 
2089   SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
2090                              DAG.getConstant(WidthX - 1, DL, MVT::i32));
2091   SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
2092   return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
2093 }
2094 
2095 SDValue
2096 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2097   if (Subtarget.isGP64bit())
2098     return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert());
2099 
2100   return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert());
2101 }
2102 
2103 SDValue MipsTargetLowering::
2104 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2105   // check the depth
2106   assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2107          "Frame address can only be determined for current frame.");
2108 
2109   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2110   MFI.setFrameAddressIsTaken(true);
2111   EVT VT = Op.getValueType();
2112   SDLoc DL(Op);
2113   SDValue FrameAddr = DAG.getCopyFromReg(
2114       DAG.getEntryNode(), DL, ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT);
2115   return FrameAddr;
2116 }
2117 
2118 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
2119                                             SelectionDAG &DAG) const {
2120   if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2121     return SDValue();
2122 
2123   // check the depth
2124   assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2125          "Return address can be determined only for current frame.");
2126 
2127   MachineFunction &MF = DAG.getMachineFunction();
2128   MachineFrameInfo &MFI = MF.getFrameInfo();
2129   MVT VT = Op.getSimpleValueType();
2130   unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA;
2131   MFI.setReturnAddressIsTaken(true);
2132 
2133   // Return RA, which contains the return address. Mark it an implicit live-in.
2134   unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2135   return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
2136 }
2137 
2138 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
2139 // generated from __builtin_eh_return (offset, handler)
2140 // The effect of this is to adjust the stack pointer by "offset"
2141 // and then branch to "handler".
2142 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
2143                                                                      const {
2144   MachineFunction &MF = DAG.getMachineFunction();
2145   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2146 
2147   MipsFI->setCallsEhReturn();
2148   SDValue Chain     = Op.getOperand(0);
2149   SDValue Offset    = Op.getOperand(1);
2150   SDValue Handler   = Op.getOperand(2);
2151   SDLoc DL(Op);
2152   EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
2153 
2154   // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2155   // EH_RETURN nodes, so that instructions are emitted back-to-back.
2156   unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1;
2157   unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
2158   Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2159   Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2160   return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2161                      DAG.getRegister(OffsetReg, Ty),
2162                      DAG.getRegister(AddrReg, getPointerTy(MF.getDataLayout())),
2163                      Chain.getValue(1));
2164 }
2165 
2166 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
2167                                               SelectionDAG &DAG) const {
2168   // FIXME: Need pseudo-fence for 'singlethread' fences
2169   // FIXME: Set SType for weaker fences where supported/appropriate.
2170   unsigned SType = 0;
2171   SDLoc DL(Op);
2172   return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2173                      DAG.getConstant(SType, DL, MVT::i32));
2174 }
2175 
2176 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
2177                                                 SelectionDAG &DAG) const {
2178   SDLoc DL(Op);
2179   MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2180 
2181   SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2182   SDValue Shamt = Op.getOperand(2);
2183   // if shamt < (VT.bits):
2184   //  lo = (shl lo, shamt)
2185   //  hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2186   // else:
2187   //  lo = 0
2188   //  hi = (shl lo, shamt[4:0])
2189   SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2190                             DAG.getConstant(-1, DL, MVT::i32));
2191   SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo,
2192                                       DAG.getConstant(1, DL, VT));
2193   SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not);
2194   SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
2195   SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2196   SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
2197   SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2198                              DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2199   Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2200                    DAG.getConstant(0, DL, VT), ShiftLeftLo);
2201   Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or);
2202 
2203   SDValue Ops[2] = {Lo, Hi};
2204   return DAG.getMergeValues(Ops, DL);
2205 }
2206 
2207 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2208                                                  bool IsSRA) const {
2209   SDLoc DL(Op);
2210   SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2211   SDValue Shamt = Op.getOperand(2);
2212   MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2213 
2214   // if shamt < (VT.bits):
2215   //  lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2216   //  if isSRA:
2217   //    hi = (sra hi, shamt)
2218   //  else:
2219   //    hi = (srl hi, shamt)
2220   // else:
2221   //  if isSRA:
2222   //   lo = (sra hi, shamt[4:0])
2223   //   hi = (sra hi, 31)
2224   //  else:
2225   //   lo = (srl hi, shamt[4:0])
2226   //   hi = 0
2227   SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2228                             DAG.getConstant(-1, DL, MVT::i32));
2229   SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi,
2230                                      DAG.getConstant(1, DL, VT));
2231   SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not);
2232   SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
2233   SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2234   SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL,
2235                                      DL, VT, Hi, Shamt);
2236   SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2237                              DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2238   SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi,
2239                             DAG.getConstant(VT.getSizeInBits() - 1, DL, VT));
2240   Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or);
2241   Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2242                    IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi);
2243 
2244   SDValue Ops[2] = {Lo, Hi};
2245   return DAG.getMergeValues(Ops, DL);
2246 }
2247 
2248 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2249                             SDValue Chain, SDValue Src, unsigned Offset) {
2250   SDValue Ptr = LD->getBasePtr();
2251   EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2252   EVT BasePtrVT = Ptr.getValueType();
2253   SDLoc DL(LD);
2254   SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2255 
2256   if (Offset)
2257     Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2258                       DAG.getConstant(Offset, DL, BasePtrVT));
2259 
2260   SDValue Ops[] = { Chain, Ptr, Src };
2261   return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2262                                  LD->getMemOperand());
2263 }
2264 
2265 // Expand an unaligned 32 or 64-bit integer load node.
2266 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2267   LoadSDNode *LD = cast<LoadSDNode>(Op);
2268   EVT MemVT = LD->getMemoryVT();
2269 
2270   if (Subtarget.systemSupportsUnalignedAccess())
2271     return Op;
2272 
2273   // Return if load is aligned or if MemVT is neither i32 nor i64.
2274   if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2275       ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2276     return SDValue();
2277 
2278   bool IsLittle = Subtarget.isLittle();
2279   EVT VT = Op.getValueType();
2280   ISD::LoadExtType ExtType = LD->getExtensionType();
2281   SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2282 
2283   assert((VT == MVT::i32) || (VT == MVT::i64));
2284 
2285   // Expand
2286   //  (set dst, (i64 (load baseptr)))
2287   // to
2288   //  (set tmp, (ldl (add baseptr, 7), undef))
2289   //  (set dst, (ldr baseptr, tmp))
2290   if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2291     SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2292                                IsLittle ? 7 : 0);
2293     return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2294                         IsLittle ? 0 : 7);
2295   }
2296 
2297   SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2298                              IsLittle ? 3 : 0);
2299   SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2300                              IsLittle ? 0 : 3);
2301 
2302   // Expand
2303   //  (set dst, (i32 (load baseptr))) or
2304   //  (set dst, (i64 (sextload baseptr))) or
2305   //  (set dst, (i64 (extload baseptr)))
2306   // to
2307   //  (set tmp, (lwl (add baseptr, 3), undef))
2308   //  (set dst, (lwr baseptr, tmp))
2309   if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2310       (ExtType == ISD::EXTLOAD))
2311     return LWR;
2312 
2313   assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2314 
2315   // Expand
2316   //  (set dst, (i64 (zextload baseptr)))
2317   // to
2318   //  (set tmp0, (lwl (add baseptr, 3), undef))
2319   //  (set tmp1, (lwr baseptr, tmp0))
2320   //  (set tmp2, (shl tmp1, 32))
2321   //  (set dst, (srl tmp2, 32))
2322   SDLoc DL(LD);
2323   SDValue Const32 = DAG.getConstant(32, DL, MVT::i32);
2324   SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2325   SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2326   SDValue Ops[] = { SRL, LWR.getValue(1) };
2327   return DAG.getMergeValues(Ops, DL);
2328 }
2329 
2330 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2331                              SDValue Chain, unsigned Offset) {
2332   SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2333   EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2334   SDLoc DL(SD);
2335   SDVTList VTList = DAG.getVTList(MVT::Other);
2336 
2337   if (Offset)
2338     Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2339                       DAG.getConstant(Offset, DL, BasePtrVT));
2340 
2341   SDValue Ops[] = { Chain, Value, Ptr };
2342   return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2343                                  SD->getMemOperand());
2344 }
2345 
2346 // Expand an unaligned 32 or 64-bit integer store node.
2347 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2348                                       bool IsLittle) {
2349   SDValue Value = SD->getValue(), Chain = SD->getChain();
2350   EVT VT = Value.getValueType();
2351 
2352   // Expand
2353   //  (store val, baseptr) or
2354   //  (truncstore val, baseptr)
2355   // to
2356   //  (swl val, (add baseptr, 3))
2357   //  (swr val, baseptr)
2358   if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2359     SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2360                                 IsLittle ? 3 : 0);
2361     return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2362   }
2363 
2364   assert(VT == MVT::i64);
2365 
2366   // Expand
2367   //  (store val, baseptr)
2368   // to
2369   //  (sdl val, (add baseptr, 7))
2370   //  (sdr val, baseptr)
2371   SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2372   return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2373 }
2374 
2375 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
2376 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) {
2377   SDValue Val = SD->getValue();
2378 
2379   if (Val.getOpcode() != ISD::FP_TO_SINT)
2380     return SDValue();
2381 
2382   EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2383   SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2384                            Val.getOperand(0));
2385   return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2386                       SD->getPointerInfo(), SD->getAlignment(),
2387                       SD->getMemOperand()->getFlags());
2388 }
2389 
2390 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2391   StoreSDNode *SD = cast<StoreSDNode>(Op);
2392   EVT MemVT = SD->getMemoryVT();
2393 
2394   // Lower unaligned integer stores.
2395   if (!Subtarget.systemSupportsUnalignedAccess() &&
2396       (SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2397       ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2398     return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle());
2399 
2400   return lowerFP_TO_SINT_STORE(SD, DAG);
2401 }
2402 
2403 SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op,
2404                                               SelectionDAG &DAG) const {
2405 
2406   // Return a fixed StackObject with offset 0 which points to the old stack
2407   // pointer.
2408   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2409   EVT ValTy = Op->getValueType(0);
2410   int FI = MFI.CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2411   return DAG.getFrameIndex(FI, ValTy);
2412 }
2413 
2414 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2415                                             SelectionDAG &DAG) const {
2416   EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2417   SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2418                               Op.getOperand(0));
2419   return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2420 }
2421 
2422 //===----------------------------------------------------------------------===//
2423 //                      Calling Convention Implementation
2424 //===----------------------------------------------------------------------===//
2425 
2426 //===----------------------------------------------------------------------===//
2427 // TODO: Implement a generic logic using tblgen that can support this.
2428 // Mips O32 ABI rules:
2429 // ---
2430 // i32 - Passed in A0, A1, A2, A3 and stack
2431 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2432 //       an argument. Otherwise, passed in A1, A2, A3 and stack.
2433 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2434 //       yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2435 //       not used, it must be shadowed. If only A3 is available, shadow it and
2436 //       go to stack.
2437 //
2438 //  For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2439 //===----------------------------------------------------------------------===//
2440 
2441 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2442                        CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2443                        CCState &State, ArrayRef<MCPhysReg> F64Regs) {
2444   const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>(
2445       State.getMachineFunction().getSubtarget());
2446 
2447   static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 };
2448   static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 };
2449 
2450   // Do not process byval args here.
2451   if (ArgFlags.isByVal())
2452     return true;
2453 
2454   // Promote i8 and i16
2455   if (ArgFlags.isInReg() && !Subtarget.isLittle()) {
2456     if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) {
2457       LocVT = MVT::i32;
2458       if (ArgFlags.isSExt())
2459         LocInfo = CCValAssign::SExtUpper;
2460       else if (ArgFlags.isZExt())
2461         LocInfo = CCValAssign::ZExtUpper;
2462       else
2463         LocInfo = CCValAssign::AExtUpper;
2464     }
2465   }
2466 
2467   // Promote i8 and i16
2468   if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2469     LocVT = MVT::i32;
2470     if (ArgFlags.isSExt())
2471       LocInfo = CCValAssign::SExt;
2472     else if (ArgFlags.isZExt())
2473       LocInfo = CCValAssign::ZExt;
2474     else
2475       LocInfo = CCValAssign::AExt;
2476   }
2477 
2478   unsigned Reg;
2479 
2480   // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2481   // is true: function is vararg, argument is 3rd or higher, there is previous
2482   // argument which is not f32 or f64.
2483   bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 ||
2484                                 State.getFirstUnallocated(F32Regs) != ValNo;
2485   unsigned OrigAlign = ArgFlags.getOrigAlign();
2486   bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2487 
2488   if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2489     Reg = State.AllocateReg(IntRegs);
2490     // If this is the first part of an i64 arg,
2491     // the allocated register must be either A0 or A2.
2492     if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2493       Reg = State.AllocateReg(IntRegs);
2494     LocVT = MVT::i32;
2495   } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2496     // Allocate int register and shadow next int register. If first
2497     // available register is Mips::A1 or Mips::A3, shadow it too.
2498     Reg = State.AllocateReg(IntRegs);
2499     if (Reg == Mips::A1 || Reg == Mips::A3)
2500       Reg = State.AllocateReg(IntRegs);
2501     State.AllocateReg(IntRegs);
2502     LocVT = MVT::i32;
2503   } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2504     // we are guaranteed to find an available float register
2505     if (ValVT == MVT::f32) {
2506       Reg = State.AllocateReg(F32Regs);
2507       // Shadow int register
2508       State.AllocateReg(IntRegs);
2509     } else {
2510       Reg = State.AllocateReg(F64Regs);
2511       // Shadow int registers
2512       unsigned Reg2 = State.AllocateReg(IntRegs);
2513       if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2514         State.AllocateReg(IntRegs);
2515       State.AllocateReg(IntRegs);
2516     }
2517   } else
2518     llvm_unreachable("Cannot handle this ValVT.");
2519 
2520   if (!Reg) {
2521     unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2522                                           OrigAlign);
2523     State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2524   } else
2525     State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2526 
2527   return false;
2528 }
2529 
2530 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2531                             MVT LocVT, CCValAssign::LocInfo LocInfo,
2532                             ISD::ArgFlagsTy ArgFlags, CCState &State) {
2533   static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 };
2534 
2535   return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2536 }
2537 
2538 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2539                             MVT LocVT, CCValAssign::LocInfo LocInfo,
2540                             ISD::ArgFlagsTy ArgFlags, CCState &State) {
2541   static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 };
2542 
2543   return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2544 }
2545 
2546 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2547                        CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2548                        CCState &State) LLVM_ATTRIBUTE_UNUSED;
2549 
2550 #include "MipsGenCallingConv.inc"
2551 
2552 //===----------------------------------------------------------------------===//
2553 //                  Call Calling Convention Implementation
2554 //===----------------------------------------------------------------------===//
2555 
2556 // Return next O32 integer argument register.
2557 static unsigned getNextIntArgReg(unsigned Reg) {
2558   assert((Reg == Mips::A0) || (Reg == Mips::A2));
2559   return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2560 }
2561 
2562 SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2563                                            SDValue Chain, SDValue Arg,
2564                                            const SDLoc &DL, bool IsTailCall,
2565                                            SelectionDAG &DAG) const {
2566   if (!IsTailCall) {
2567     SDValue PtrOff =
2568         DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr,
2569                     DAG.getIntPtrConstant(Offset, DL));
2570     return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo());
2571   }
2572 
2573   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2574   int FI = MFI.CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2575   SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2576   return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2577                       /* Alignment = */ 0, MachineMemOperand::MOVolatile);
2578 }
2579 
2580 void MipsTargetLowering::
2581 getOpndList(SmallVectorImpl<SDValue> &Ops,
2582             std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2583             bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2584             bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
2585             SDValue Chain) const {
2586   // Insert node "GP copy globalreg" before call to function.
2587   //
2588   // R_MIPS_CALL* operators (emitted when non-internal functions are called
2589   // in PIC mode) allow symbols to be resolved via lazy binding.
2590   // The lazy binding stub requires GP to point to the GOT.
2591   // Note that we don't need GP to point to the GOT for indirect calls
2592   // (when R_MIPS_CALL* is not used for the call) because Mips linker generates
2593   // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs
2594   // used for the function (that is, Mips linker doesn't generate lazy binding
2595   // stub for a function whose address is taken in the program).
2596   if (IsPICCall && !InternalLinkage && IsCallReloc) {
2597     unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP;
2598     EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
2599     RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2600   }
2601 
2602   // Build a sequence of copy-to-reg nodes chained together with token
2603   // chain and flag operands which copy the outgoing args into registers.
2604   // The InFlag in necessary since all emitted instructions must be
2605   // stuck together.
2606   SDValue InFlag;
2607 
2608   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2609     Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2610                                  RegsToPass[i].second, InFlag);
2611     InFlag = Chain.getValue(1);
2612   }
2613 
2614   // Add argument registers to the end of the list so that they are
2615   // known live into the call.
2616   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2617     Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2618                                       RegsToPass[i].second.getValueType()));
2619 
2620   // Add a register mask operand representing the call-preserved registers.
2621   const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
2622   const uint32_t *Mask =
2623       TRI->getCallPreservedMask(CLI.DAG.getMachineFunction(), CLI.CallConv);
2624   assert(Mask && "Missing call preserved mask for calling convention");
2625   if (Subtarget.inMips16HardFloat()) {
2626     if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2627       llvm::StringRef Sym = G->getGlobal()->getName();
2628       Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2629       if (F && F->hasFnAttribute("__Mips16RetHelper")) {
2630         Mask = MipsRegisterInfo::getMips16RetHelperMask();
2631       }
2632     }
2633   }
2634   Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2635 
2636   if (InFlag.getNode())
2637     Ops.push_back(InFlag);
2638 }
2639 
2640 /// LowerCall - functions arguments are copied from virtual regs to
2641 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2642 SDValue
2643 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2644                               SmallVectorImpl<SDValue> &InVals) const {
2645   SelectionDAG &DAG                     = CLI.DAG;
2646   SDLoc DL                              = CLI.DL;
2647   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2648   SmallVectorImpl<SDValue> &OutVals     = CLI.OutVals;
2649   SmallVectorImpl<ISD::InputArg> &Ins   = CLI.Ins;
2650   SDValue Chain                         = CLI.Chain;
2651   SDValue Callee                        = CLI.Callee;
2652   bool &IsTailCall                      = CLI.IsTailCall;
2653   CallingConv::ID CallConv              = CLI.CallConv;
2654   bool IsVarArg                         = CLI.IsVarArg;
2655 
2656   MachineFunction &MF = DAG.getMachineFunction();
2657   MachineFrameInfo &MFI = MF.getFrameInfo();
2658   const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
2659   MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2660   bool IsPIC = isPositionIndependent();
2661 
2662   // Analyze operands of the call, assigning locations to each operand.
2663   SmallVector<CCValAssign, 16> ArgLocs;
2664   MipsCCState CCInfo(
2665       CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(),
2666       MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget));
2667 
2668   // Allocate the reserved argument area. It seems strange to do this from the
2669   // caller side but removing it breaks the frame size calculation.
2670   CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1);
2671 
2672   CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(), Callee.getNode());
2673 
2674   // Get a count of how many bytes are to be pushed on the stack.
2675   unsigned NextStackOffset = CCInfo.getNextStackOffset();
2676 
2677   // Check if it's really possible to do a tail call.
2678   if (IsTailCall)
2679     IsTailCall = isEligibleForTailCallOptimization(
2680         CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>());
2681 
2682   if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2683     report_fatal_error("failed to perform tail call elimination on a call "
2684                        "site marked musttail");
2685 
2686   if (IsTailCall)
2687     ++NumTailCalls;
2688 
2689   // Chain is the output chain of the last Load/Store or CopyToReg node.
2690   // ByValChain is the output chain of the last Memcpy node created for copying
2691   // byval arguments to the stack.
2692   unsigned StackAlignment = TFL->getStackAlignment();
2693   NextStackOffset = alignTo(NextStackOffset, StackAlignment);
2694   SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, DL, true);
2695 
2696   if (!IsTailCall)
2697     Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL);
2698 
2699   SDValue StackPtr =
2700       DAG.getCopyFromReg(Chain, DL, ABI.IsN64() ? Mips::SP_64 : Mips::SP,
2701                          getPointerTy(DAG.getDataLayout()));
2702 
2703   std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2704   SmallVector<SDValue, 8> MemOpChains;
2705 
2706   CCInfo.rewindByValRegsInfo();
2707 
2708   // Walk the register/memloc assignments, inserting copies/loads.
2709   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2710     SDValue Arg = OutVals[i];
2711     CCValAssign &VA = ArgLocs[i];
2712     MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2713     ISD::ArgFlagsTy Flags = Outs[i].Flags;
2714     bool UseUpperBits = false;
2715 
2716     // ByVal Arg.
2717     if (Flags.isByVal()) {
2718       unsigned FirstByValReg, LastByValReg;
2719       unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
2720       CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
2721 
2722       assert(Flags.getByValSize() &&
2723              "ByVal args of size 0 should have been ignored by front-end.");
2724       assert(ByValIdx < CCInfo.getInRegsParamsCount());
2725       assert(!IsTailCall &&
2726              "Do not tail-call optimize if there is a byval argument.");
2727       passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2728                    FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(),
2729                    VA);
2730       CCInfo.nextInRegsParam();
2731       continue;
2732     }
2733 
2734     // Promote the value if needed.
2735     switch (VA.getLocInfo()) {
2736     default:
2737       llvm_unreachable("Unknown loc info!");
2738     case CCValAssign::Full:
2739       if (VA.isRegLoc()) {
2740         if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2741             (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2742             (ValVT == MVT::i64 && LocVT == MVT::f64))
2743           Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2744         else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2745           SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2746                                    Arg, DAG.getConstant(0, DL, MVT::i32));
2747           SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2748                                    Arg, DAG.getConstant(1, DL, MVT::i32));
2749           if (!Subtarget.isLittle())
2750             std::swap(Lo, Hi);
2751           unsigned LocRegLo = VA.getLocReg();
2752           unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2753           RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2754           RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2755           continue;
2756         }
2757       }
2758       break;
2759     case CCValAssign::BCvt:
2760       Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2761       break;
2762     case CCValAssign::SExtUpper:
2763       UseUpperBits = true;
2764       LLVM_FALLTHROUGH;
2765     case CCValAssign::SExt:
2766       Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2767       break;
2768     case CCValAssign::ZExtUpper:
2769       UseUpperBits = true;
2770       LLVM_FALLTHROUGH;
2771     case CCValAssign::ZExt:
2772       Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2773       break;
2774     case CCValAssign::AExtUpper:
2775       UseUpperBits = true;
2776       LLVM_FALLTHROUGH;
2777     case CCValAssign::AExt:
2778       Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2779       break;
2780     }
2781 
2782     if (UseUpperBits) {
2783       unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
2784       unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
2785       Arg = DAG.getNode(
2786           ISD::SHL, DL, VA.getLocVT(), Arg,
2787           DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
2788     }
2789 
2790     // Arguments that can be passed on register must be kept at
2791     // RegsToPass vector
2792     if (VA.isRegLoc()) {
2793       RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2794       continue;
2795     }
2796 
2797     // Register can't get to this point...
2798     assert(VA.isMemLoc());
2799 
2800     // emit ISD::STORE whichs stores the
2801     // parameter value to a stack Location
2802     MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2803                                          Chain, Arg, DL, IsTailCall, DAG));
2804   }
2805 
2806   // Transform all store nodes into one single node because all store
2807   // nodes are independent of each other.
2808   if (!MemOpChains.empty())
2809     Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2810 
2811   // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2812   // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2813   // node so that legalize doesn't hack it.
2814   bool IsPICCall = (ABI.IsN64() || IsPIC); // true if calls are translated to
2815                                            // jalr $25
2816   bool GlobalOrExternal = false, InternalLinkage = false, IsCallReloc = false;
2817   SDValue CalleeLo;
2818   EVT Ty = Callee.getValueType();
2819 
2820   if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2821     if (IsPICCall) {
2822       const GlobalValue *Val = G->getGlobal();
2823       InternalLinkage = Val->hasInternalLinkage();
2824 
2825       if (InternalLinkage)
2826         Callee = getAddrLocal(G, DL, Ty, DAG, ABI.IsN32() || ABI.IsN64());
2827       else if (LargeGOT) {
2828         Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16,
2829                                        MipsII::MO_CALL_LO16, Chain,
2830                                        FuncInfo->callPtrInfo(Val));
2831         IsCallReloc = true;
2832       } else {
2833         Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2834                                FuncInfo->callPtrInfo(Val));
2835         IsCallReloc = true;
2836       }
2837     } else
2838       Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL,
2839                                           getPointerTy(DAG.getDataLayout()), 0,
2840                                           MipsII::MO_NO_FLAG);
2841     GlobalOrExternal = true;
2842   }
2843   else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2844     const char *Sym = S->getSymbol();
2845 
2846     if (!ABI.IsN64() && !IsPIC) // !N64 && static
2847       Callee = DAG.getTargetExternalSymbol(
2848           Sym, getPointerTy(DAG.getDataLayout()), MipsII::MO_NO_FLAG);
2849     else if (LargeGOT) {
2850       Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16,
2851                                      MipsII::MO_CALL_LO16, Chain,
2852                                      FuncInfo->callPtrInfo(Sym));
2853       IsCallReloc = true;
2854     } else { // N64 || PIC
2855       Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2856                              FuncInfo->callPtrInfo(Sym));
2857       IsCallReloc = true;
2858     }
2859 
2860     GlobalOrExternal = true;
2861   }
2862 
2863   SmallVector<SDValue, 8> Ops(1, Chain);
2864   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2865 
2866   getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2867               IsCallReloc, CLI, Callee, Chain);
2868 
2869   if (IsTailCall) {
2870     MF.getFrameInfo().setHasTailCall();
2871     return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops);
2872   }
2873 
2874   Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops);
2875   SDValue InFlag = Chain.getValue(1);
2876 
2877   // Create the CALLSEQ_END node.
2878   Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2879                              DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
2880   InFlag = Chain.getValue(1);
2881 
2882   // Handle result values, copying them out of physregs into vregs that we
2883   // return.
2884   return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2885                          InVals, CLI);
2886 }
2887 
2888 /// LowerCallResult - Lower the result values of a call into the
2889 /// appropriate copies out of appropriate physical registers.
2890 SDValue MipsTargetLowering::LowerCallResult(
2891     SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
2892     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
2893     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
2894     TargetLowering::CallLoweringInfo &CLI) const {
2895   // Assign locations to each value returned by this call.
2896   SmallVector<CCValAssign, 16> RVLocs;
2897   MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
2898                      *DAG.getContext());
2899   CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI);
2900 
2901   // Copy all of the result registers out of their specified physreg.
2902   for (unsigned i = 0; i != RVLocs.size(); ++i) {
2903     CCValAssign &VA = RVLocs[i];
2904     assert(VA.isRegLoc() && "Can only return in registers!");
2905 
2906     SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2907                                      RVLocs[i].getLocVT(), InFlag);
2908     Chain = Val.getValue(1);
2909     InFlag = Val.getValue(2);
2910 
2911     if (VA.isUpperBitsInLoc()) {
2912       unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits();
2913       unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
2914       unsigned Shift =
2915           VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
2916       Val = DAG.getNode(
2917           Shift, DL, VA.getLocVT(), Val,
2918           DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
2919     }
2920 
2921     switch (VA.getLocInfo()) {
2922     default:
2923       llvm_unreachable("Unknown loc info!");
2924     case CCValAssign::Full:
2925       break;
2926     case CCValAssign::BCvt:
2927       Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2928       break;
2929     case CCValAssign::AExt:
2930     case CCValAssign::AExtUpper:
2931       Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2932       break;
2933     case CCValAssign::ZExt:
2934     case CCValAssign::ZExtUpper:
2935       Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val,
2936                         DAG.getValueType(VA.getValVT()));
2937       Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2938       break;
2939     case CCValAssign::SExt:
2940     case CCValAssign::SExtUpper:
2941       Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val,
2942                         DAG.getValueType(VA.getValVT()));
2943       Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2944       break;
2945     }
2946 
2947     InVals.push_back(Val);
2948   }
2949 
2950   return Chain;
2951 }
2952 
2953 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA,
2954                                       EVT ArgVT, const SDLoc &DL,
2955                                       SelectionDAG &DAG) {
2956   MVT LocVT = VA.getLocVT();
2957   EVT ValVT = VA.getValVT();
2958 
2959   // Shift into the upper bits if necessary.
2960   switch (VA.getLocInfo()) {
2961   default:
2962     break;
2963   case CCValAssign::AExtUpper:
2964   case CCValAssign::SExtUpper:
2965   case CCValAssign::ZExtUpper: {
2966     unsigned ValSizeInBits = ArgVT.getSizeInBits();
2967     unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
2968     unsigned Opcode =
2969         VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
2970     Val = DAG.getNode(
2971         Opcode, DL, VA.getLocVT(), Val,
2972         DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
2973     break;
2974   }
2975   }
2976 
2977   // If this is an value smaller than the argument slot size (32-bit for O32,
2978   // 64-bit for N32/N64), it has been promoted in some way to the argument slot
2979   // size. Extract the value and insert any appropriate assertions regarding
2980   // sign/zero extension.
2981   switch (VA.getLocInfo()) {
2982   default:
2983     llvm_unreachable("Unknown loc info!");
2984   case CCValAssign::Full:
2985     break;
2986   case CCValAssign::AExtUpper:
2987   case CCValAssign::AExt:
2988     Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
2989     break;
2990   case CCValAssign::SExtUpper:
2991   case CCValAssign::SExt:
2992     Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT));
2993     Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
2994     break;
2995   case CCValAssign::ZExtUpper:
2996   case CCValAssign::ZExt:
2997     Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT));
2998     Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
2999     break;
3000   case CCValAssign::BCvt:
3001     Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val);
3002     break;
3003   }
3004 
3005   return Val;
3006 }
3007 
3008 //===----------------------------------------------------------------------===//
3009 //             Formal Arguments Calling Convention Implementation
3010 //===----------------------------------------------------------------------===//
3011 /// LowerFormalArguments - transform physical registers into virtual registers
3012 /// and generate load operations for arguments places on the stack.
3013 SDValue MipsTargetLowering::LowerFormalArguments(
3014     SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
3015     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
3016     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
3017   MachineFunction &MF = DAG.getMachineFunction();
3018   MachineFrameInfo &MFI = MF.getFrameInfo();
3019   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3020 
3021   MipsFI->setVarArgsFrameIndex(0);
3022 
3023   // Used with vargs to acumulate store chains.
3024   std::vector<SDValue> OutChains;
3025 
3026   // Assign locations to all of the incoming arguments.
3027   SmallVector<CCValAssign, 16> ArgLocs;
3028   MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
3029                      *DAG.getContext());
3030   CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1);
3031   const Function *Func = DAG.getMachineFunction().getFunction();
3032   Function::const_arg_iterator FuncArg = Func->arg_begin();
3033 
3034   if (Func->hasFnAttribute("interrupt") && !Func->arg_empty())
3035     report_fatal_error(
3036         "Functions with the interrupt attribute cannot have arguments!");
3037 
3038   CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg);
3039   MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
3040                            CCInfo.getInRegsParamsCount() > 0);
3041 
3042   unsigned CurArgIdx = 0;
3043   CCInfo.rewindByValRegsInfo();
3044 
3045   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3046     CCValAssign &VA = ArgLocs[i];
3047     if (Ins[i].isOrigArg()) {
3048       std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx);
3049       CurArgIdx = Ins[i].getOrigArgIndex();
3050     }
3051     EVT ValVT = VA.getValVT();
3052     ISD::ArgFlagsTy Flags = Ins[i].Flags;
3053     bool IsRegLoc = VA.isRegLoc();
3054 
3055     if (Flags.isByVal()) {
3056       assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit");
3057       unsigned FirstByValReg, LastByValReg;
3058       unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
3059       CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
3060 
3061       assert(Flags.getByValSize() &&
3062              "ByVal args of size 0 should have been ignored by front-end.");
3063       assert(ByValIdx < CCInfo.getInRegsParamsCount());
3064       copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
3065                     FirstByValReg, LastByValReg, VA, CCInfo);
3066       CCInfo.nextInRegsParam();
3067       continue;
3068     }
3069 
3070     // Arguments stored on registers
3071     if (IsRegLoc) {
3072       MVT RegVT = VA.getLocVT();
3073       unsigned ArgReg = VA.getLocReg();
3074       const TargetRegisterClass *RC = getRegClassFor(RegVT);
3075 
3076       // Transform the arguments stored on
3077       // physical registers into virtual ones
3078       unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
3079       SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
3080 
3081       ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3082 
3083       // Handle floating point arguments passed in integer registers and
3084       // long double arguments passed in floating point registers.
3085       if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
3086           (RegVT == MVT::i64 && ValVT == MVT::f64) ||
3087           (RegVT == MVT::f64 && ValVT == MVT::i64))
3088         ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
3089       else if (ABI.IsO32() && RegVT == MVT::i32 &&
3090                ValVT == MVT::f64) {
3091         unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
3092                                   getNextIntArgReg(ArgReg), RC);
3093         SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
3094         if (!Subtarget.isLittle())
3095           std::swap(ArgValue, ArgValue2);
3096         ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
3097                                ArgValue, ArgValue2);
3098       }
3099 
3100       InVals.push_back(ArgValue);
3101     } else { // VA.isRegLoc()
3102       MVT LocVT = VA.getLocVT();
3103 
3104       if (ABI.IsO32()) {
3105         // We ought to be able to use LocVT directly but O32 sets it to i32
3106         // when allocating floating point values to integer registers.
3107         // This shouldn't influence how we load the value into registers unless
3108         // we are targeting softfloat.
3109         if (VA.getValVT().isFloatingPoint() && !Subtarget.useSoftFloat())
3110           LocVT = VA.getValVT();
3111       }
3112 
3113       // sanity check
3114       assert(VA.isMemLoc());
3115 
3116       // The stack pointer offset is relative to the caller stack frame.
3117       int FI = MFI.CreateFixedObject(LocVT.getSizeInBits() / 8,
3118                                      VA.getLocMemOffset(), true);
3119 
3120       // Create load nodes to retrieve arguments from the stack
3121       SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3122       SDValue ArgValue = DAG.getLoad(
3123           LocVT, DL, Chain, FIN,
3124           MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI));
3125       OutChains.push_back(ArgValue.getValue(1));
3126 
3127       ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3128 
3129       InVals.push_back(ArgValue);
3130     }
3131   }
3132 
3133   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3134     // The mips ABIs for returning structs by value requires that we copy
3135     // the sret argument into $v0 for the return. Save the argument into
3136     // a virtual register so that we can access it from the return points.
3137     if (Ins[i].Flags.isSRet()) {
3138       unsigned Reg = MipsFI->getSRetReturnReg();
3139       if (!Reg) {
3140         Reg = MF.getRegInfo().createVirtualRegister(
3141             getRegClassFor(ABI.IsN64() ? MVT::i64 : MVT::i32));
3142         MipsFI->setSRetReturnReg(Reg);
3143       }
3144       SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]);
3145       Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
3146       break;
3147     }
3148   }
3149 
3150   if (IsVarArg)
3151     writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo);
3152 
3153   // All stores are grouped in one node to allow the matching between
3154   // the size of Ins and InVals. This only happens when on varg functions
3155   if (!OutChains.empty()) {
3156     OutChains.push_back(Chain);
3157     Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
3158   }
3159 
3160   return Chain;
3161 }
3162 
3163 //===----------------------------------------------------------------------===//
3164 //               Return Value Calling Convention Implementation
3165 //===----------------------------------------------------------------------===//
3166 
3167 bool
3168 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3169                                    MachineFunction &MF, bool IsVarArg,
3170                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
3171                                    LLVMContext &Context) const {
3172   SmallVector<CCValAssign, 16> RVLocs;
3173   MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
3174   return CCInfo.CheckReturn(Outs, RetCC_Mips);
3175 }
3176 
3177 bool
3178 MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const {
3179   if (Subtarget.hasMips3() && Subtarget.useSoftFloat()) {
3180     if (Type == MVT::i32)
3181       return true;
3182   }
3183   return IsSigned;
3184 }
3185 
3186 SDValue
3187 MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
3188                                          const SDLoc &DL,
3189                                          SelectionDAG &DAG) const {
3190 
3191   MachineFunction &MF = DAG.getMachineFunction();
3192   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3193 
3194   MipsFI->setISR();
3195 
3196   return DAG.getNode(MipsISD::ERet, DL, MVT::Other, RetOps);
3197 }
3198 
3199 SDValue
3200 MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3201                                 bool IsVarArg,
3202                                 const SmallVectorImpl<ISD::OutputArg> &Outs,
3203                                 const SmallVectorImpl<SDValue> &OutVals,
3204                                 const SDLoc &DL, SelectionDAG &DAG) const {
3205   // CCValAssign - represent the assignment of
3206   // the return value to a location
3207   SmallVector<CCValAssign, 16> RVLocs;
3208   MachineFunction &MF = DAG.getMachineFunction();
3209 
3210   // CCState - Info about the registers and stack slot.
3211   MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
3212 
3213   // Analyze return values.
3214   CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
3215 
3216   SDValue Flag;
3217   SmallVector<SDValue, 4> RetOps(1, Chain);
3218 
3219   // Copy the result values into the output registers.
3220   for (unsigned i = 0; i != RVLocs.size(); ++i) {
3221     SDValue Val = OutVals[i];
3222     CCValAssign &VA = RVLocs[i];
3223     assert(VA.isRegLoc() && "Can only return in registers!");
3224     bool UseUpperBits = false;
3225 
3226     switch (VA.getLocInfo()) {
3227     default:
3228       llvm_unreachable("Unknown loc info!");
3229     case CCValAssign::Full:
3230       break;
3231     case CCValAssign::BCvt:
3232       Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val);
3233       break;
3234     case CCValAssign::AExtUpper:
3235       UseUpperBits = true;
3236       LLVM_FALLTHROUGH;
3237     case CCValAssign::AExt:
3238       Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val);
3239       break;
3240     case CCValAssign::ZExtUpper:
3241       UseUpperBits = true;
3242       LLVM_FALLTHROUGH;
3243     case CCValAssign::ZExt:
3244       Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val);
3245       break;
3246     case CCValAssign::SExtUpper:
3247       UseUpperBits = true;
3248       LLVM_FALLTHROUGH;
3249     case CCValAssign::SExt:
3250       Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val);
3251       break;
3252     }
3253 
3254     if (UseUpperBits) {
3255       unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
3256       unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3257       Val = DAG.getNode(
3258           ISD::SHL, DL, VA.getLocVT(), Val,
3259           DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3260     }
3261 
3262     Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
3263 
3264     // Guarantee that all emitted copies are stuck together with flags.
3265     Flag = Chain.getValue(1);
3266     RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3267   }
3268 
3269   // The mips ABIs for returning structs by value requires that we copy
3270   // the sret argument into $v0 for the return. We saved the argument into
3271   // a virtual register in the entry block, so now we copy the value out
3272   // and into $v0.
3273   if (MF.getFunction()->hasStructRetAttr()) {
3274     MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3275     unsigned Reg = MipsFI->getSRetReturnReg();
3276 
3277     if (!Reg)
3278       llvm_unreachable("sret virtual register not created in the entry block");
3279     SDValue Val =
3280         DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout()));
3281     unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
3282 
3283     Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
3284     Flag = Chain.getValue(1);
3285     RetOps.push_back(DAG.getRegister(V0, getPointerTy(DAG.getDataLayout())));
3286   }
3287 
3288   RetOps[0] = Chain;  // Update chain.
3289 
3290   // Add the flag if we have it.
3291   if (Flag.getNode())
3292     RetOps.push_back(Flag);
3293 
3294   // ISRs must use "eret".
3295   if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt"))
3296     return LowerInterruptReturn(RetOps, DL, DAG);
3297 
3298   // Standard return on Mips is a "jr $ra"
3299   return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps);
3300 }
3301 
3302 //===----------------------------------------------------------------------===//
3303 //                           Mips Inline Assembly Support
3304 //===----------------------------------------------------------------------===//
3305 
3306 /// getConstraintType - Given a constraint letter, return the type of
3307 /// constraint it is for this target.
3308 MipsTargetLowering::ConstraintType
3309 MipsTargetLowering::getConstraintType(StringRef Constraint) const {
3310   // Mips specific constraints
3311   // GCC config/mips/constraints.md
3312   //
3313   // 'd' : An address register. Equivalent to r
3314   //       unless generating MIPS16 code.
3315   // 'y' : Equivalent to r; retained for
3316   //       backwards compatibility.
3317   // 'c' : A register suitable for use in an indirect
3318   //       jump. This will always be $25 for -mabicalls.
3319   // 'l' : The lo register. 1 word storage.
3320   // 'x' : The hilo register pair. Double word storage.
3321   if (Constraint.size() == 1) {
3322     switch (Constraint[0]) {
3323       default : break;
3324       case 'd':
3325       case 'y':
3326       case 'f':
3327       case 'c':
3328       case 'l':
3329       case 'x':
3330         return C_RegisterClass;
3331       case 'R':
3332         return C_Memory;
3333     }
3334   }
3335 
3336   if (Constraint == "ZC")
3337     return C_Memory;
3338 
3339   return TargetLowering::getConstraintType(Constraint);
3340 }
3341 
3342 /// Examine constraint type and operand type and determine a weight value.
3343 /// This object must already have been set up with the operand type
3344 /// and the current alternative constraint selected.
3345 TargetLowering::ConstraintWeight
3346 MipsTargetLowering::getSingleConstraintMatchWeight(
3347     AsmOperandInfo &info, const char *constraint) const {
3348   ConstraintWeight weight = CW_Invalid;
3349   Value *CallOperandVal = info.CallOperandVal;
3350     // If we don't have a value, we can't do a match,
3351     // but allow it at the lowest weight.
3352   if (!CallOperandVal)
3353     return CW_Default;
3354   Type *type = CallOperandVal->getType();
3355   // Look at the constraint type.
3356   switch (*constraint) {
3357   default:
3358     weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3359     break;
3360   case 'd':
3361   case 'y':
3362     if (type->isIntegerTy())
3363       weight = CW_Register;
3364     break;
3365   case 'f': // FPU or MSA register
3366     if (Subtarget.hasMSA() && type->isVectorTy() &&
3367         cast<VectorType>(type)->getBitWidth() == 128)
3368       weight = CW_Register;
3369     else if (type->isFloatTy())
3370       weight = CW_Register;
3371     break;
3372   case 'c': // $25 for indirect jumps
3373   case 'l': // lo register
3374   case 'x': // hilo register pair
3375     if (type->isIntegerTy())
3376       weight = CW_SpecificReg;
3377     break;
3378   case 'I': // signed 16 bit immediate
3379   case 'J': // integer zero
3380   case 'K': // unsigned 16 bit immediate
3381   case 'L': // signed 32 bit immediate where lower 16 bits are 0
3382   case 'N': // immediate in the range of -65535 to -1 (inclusive)
3383   case 'O': // signed 15 bit immediate (+- 16383)
3384   case 'P': // immediate in the range of 65535 to 1 (inclusive)
3385     if (isa<ConstantInt>(CallOperandVal))
3386       weight = CW_Constant;
3387     break;
3388   case 'R':
3389     weight = CW_Memory;
3390     break;
3391   }
3392   return weight;
3393 }
3394 
3395 /// This is a helper function to parse a physical register string and split it
3396 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
3397 /// that is returned indicates whether parsing was successful. The second flag
3398 /// is true if the numeric part exists.
3399 static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix,
3400                                               unsigned long long &Reg) {
3401   if (C.front() != '{' || C.back() != '}')
3402     return std::make_pair(false, false);
3403 
3404   // Search for the first numeric character.
3405   StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
3406   I = std::find_if(B, E, isdigit);
3407 
3408   Prefix = StringRef(B, I - B);
3409 
3410   // The second flag is set to false if no numeric characters were found.
3411   if (I == E)
3412     return std::make_pair(true, false);
3413 
3414   // Parse the numeric characters.
3415   return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
3416                         true);
3417 }
3418 
3419 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
3420 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const {
3421   const TargetRegisterInfo *TRI =
3422       Subtarget.getRegisterInfo();
3423   const TargetRegisterClass *RC;
3424   StringRef Prefix;
3425   unsigned long long Reg;
3426 
3427   std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
3428 
3429   if (!R.first)
3430     return std::make_pair(0U, nullptr);
3431 
3432   if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
3433     // No numeric characters follow "hi" or "lo".
3434     if (R.second)
3435       return std::make_pair(0U, nullptr);
3436 
3437     RC = TRI->getRegClass(Prefix == "hi" ?
3438                           Mips::HI32RegClassID : Mips::LO32RegClassID);
3439     return std::make_pair(*(RC->begin()), RC);
3440   } else if (Prefix.startswith("$msa")) {
3441     // Parse $msa(ir|csr|access|save|modify|request|map|unmap)
3442 
3443     // No numeric characters follow the name.
3444     if (R.second)
3445       return std::make_pair(0U, nullptr);
3446 
3447     Reg = StringSwitch<unsigned long long>(Prefix)
3448               .Case("$msair", Mips::MSAIR)
3449               .Case("$msacsr", Mips::MSACSR)
3450               .Case("$msaaccess", Mips::MSAAccess)
3451               .Case("$msasave", Mips::MSASave)
3452               .Case("$msamodify", Mips::MSAModify)
3453               .Case("$msarequest", Mips::MSARequest)
3454               .Case("$msamap", Mips::MSAMap)
3455               .Case("$msaunmap", Mips::MSAUnmap)
3456               .Default(0);
3457 
3458     if (!Reg)
3459       return std::make_pair(0U, nullptr);
3460 
3461     RC = TRI->getRegClass(Mips::MSACtrlRegClassID);
3462     return std::make_pair(Reg, RC);
3463   }
3464 
3465   if (!R.second)
3466     return std::make_pair(0U, nullptr);
3467 
3468   if (Prefix == "$f") { // Parse $f0-$f31.
3469     // If the size of FP registers is 64-bit or Reg is an even number, select
3470     // the 64-bit register class. Otherwise, select the 32-bit register class.
3471     if (VT == MVT::Other)
3472       VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
3473 
3474     RC = getRegClassFor(VT);
3475 
3476     if (RC == &Mips::AFGR64RegClass) {
3477       assert(Reg % 2 == 0);
3478       Reg >>= 1;
3479     }
3480   } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7.
3481     RC = TRI->getRegClass(Mips::FCCRegClassID);
3482   else if (Prefix == "$w") { // Parse $w0-$w31.
3483     RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT);
3484   } else { // Parse $0-$31.
3485     assert(Prefix == "$");
3486     RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
3487   }
3488 
3489   assert(Reg < RC->getNumRegs());
3490   return std::make_pair(*(RC->begin() + Reg), RC);
3491 }
3492 
3493 /// Given a register class constraint, like 'r', if this corresponds directly
3494 /// to an LLVM register class, return a register of 0 and the register class
3495 /// pointer.
3496 std::pair<unsigned, const TargetRegisterClass *>
3497 MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
3498                                                  StringRef Constraint,
3499                                                  MVT VT) const {
3500   if (Constraint.size() == 1) {
3501     switch (Constraint[0]) {
3502     case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3503     case 'y': // Same as 'r'. Exists for compatibility.
3504     case 'r':
3505       if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3506         if (Subtarget.inMips16Mode())
3507           return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3508         return std::make_pair(0U, &Mips::GPR32RegClass);
3509       }
3510       if (VT == MVT::i64 && !Subtarget.isGP64bit())
3511         return std::make_pair(0U, &Mips::GPR32RegClass);
3512       if (VT == MVT::i64 && Subtarget.isGP64bit())
3513         return std::make_pair(0U, &Mips::GPR64RegClass);
3514       // This will generate an error message
3515       return std::make_pair(0U, nullptr);
3516     case 'f': // FPU or MSA register
3517       if (VT == MVT::v16i8)
3518         return std::make_pair(0U, &Mips::MSA128BRegClass);
3519       else if (VT == MVT::v8i16 || VT == MVT::v8f16)
3520         return std::make_pair(0U, &Mips::MSA128HRegClass);
3521       else if (VT == MVT::v4i32 || VT == MVT::v4f32)
3522         return std::make_pair(0U, &Mips::MSA128WRegClass);
3523       else if (VT == MVT::v2i64 || VT == MVT::v2f64)
3524         return std::make_pair(0U, &Mips::MSA128DRegClass);
3525       else if (VT == MVT::f32)
3526         return std::make_pair(0U, &Mips::FGR32RegClass);
3527       else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) {
3528         if (Subtarget.isFP64bit())
3529           return std::make_pair(0U, &Mips::FGR64RegClass);
3530         return std::make_pair(0U, &Mips::AFGR64RegClass);
3531       }
3532       break;
3533     case 'c': // register suitable for indirect jump
3534       if (VT == MVT::i32)
3535         return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
3536       assert(VT == MVT::i64 && "Unexpected type.");
3537       return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
3538     case 'l': // register suitable for indirect jump
3539       if (VT == MVT::i32)
3540         return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
3541       return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
3542     case 'x': // register suitable for indirect jump
3543       // Fixme: Not triggering the use of both hi and low
3544       // This will generate an error message
3545       return std::make_pair(0U, nullptr);
3546     }
3547   }
3548 
3549   std::pair<unsigned, const TargetRegisterClass *> R;
3550   R = parseRegForInlineAsmConstraint(Constraint, VT);
3551 
3552   if (R.second)
3553     return R;
3554 
3555   return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
3556 }
3557 
3558 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3559 /// vector.  If it is invalid, don't add anything to Ops.
3560 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3561                                                      std::string &Constraint,
3562                                                      std::vector<SDValue>&Ops,
3563                                                      SelectionDAG &DAG) const {
3564   SDLoc DL(Op);
3565   SDValue Result;
3566 
3567   // Only support length 1 constraints for now.
3568   if (Constraint.length() > 1) return;
3569 
3570   char ConstraintLetter = Constraint[0];
3571   switch (ConstraintLetter) {
3572   default: break; // This will fall through to the generic implementation
3573   case 'I': // Signed 16 bit constant
3574     // If this fails, the parent routine will give an error
3575     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3576       EVT Type = Op.getValueType();
3577       int64_t Val = C->getSExtValue();
3578       if (isInt<16>(Val)) {
3579         Result = DAG.getTargetConstant(Val, DL, Type);
3580         break;
3581       }
3582     }
3583     return;
3584   case 'J': // integer zero
3585     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3586       EVT Type = Op.getValueType();
3587       int64_t Val = C->getZExtValue();
3588       if (Val == 0) {
3589         Result = DAG.getTargetConstant(0, DL, Type);
3590         break;
3591       }
3592     }
3593     return;
3594   case 'K': // unsigned 16 bit immediate
3595     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3596       EVT Type = Op.getValueType();
3597       uint64_t Val = (uint64_t)C->getZExtValue();
3598       if (isUInt<16>(Val)) {
3599         Result = DAG.getTargetConstant(Val, DL, Type);
3600         break;
3601       }
3602     }
3603     return;
3604   case 'L': // signed 32 bit immediate where lower 16 bits are 0
3605     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3606       EVT Type = Op.getValueType();
3607       int64_t Val = C->getSExtValue();
3608       if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3609         Result = DAG.getTargetConstant(Val, DL, Type);
3610         break;
3611       }
3612     }
3613     return;
3614   case 'N': // immediate in the range of -65535 to -1 (inclusive)
3615     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3616       EVT Type = Op.getValueType();
3617       int64_t Val = C->getSExtValue();
3618       if ((Val >= -65535) && (Val <= -1)) {
3619         Result = DAG.getTargetConstant(Val, DL, Type);
3620         break;
3621       }
3622     }
3623     return;
3624   case 'O': // signed 15 bit immediate
3625     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3626       EVT Type = Op.getValueType();
3627       int64_t Val = C->getSExtValue();
3628       if ((isInt<15>(Val))) {
3629         Result = DAG.getTargetConstant(Val, DL, Type);
3630         break;
3631       }
3632     }
3633     return;
3634   case 'P': // immediate in the range of 1 to 65535 (inclusive)
3635     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3636       EVT Type = Op.getValueType();
3637       int64_t Val = C->getSExtValue();
3638       if ((Val <= 65535) && (Val >= 1)) {
3639         Result = DAG.getTargetConstant(Val, DL, Type);
3640         break;
3641       }
3642     }
3643     return;
3644   }
3645 
3646   if (Result.getNode()) {
3647     Ops.push_back(Result);
3648     return;
3649   }
3650 
3651   TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3652 }
3653 
3654 bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL,
3655                                                const AddrMode &AM, Type *Ty,
3656                                                unsigned AS) const {
3657   // No global is ever allowed as a base.
3658   if (AM.BaseGV)
3659     return false;
3660 
3661   switch (AM.Scale) {
3662   case 0: // "r+i" or just "i", depending on HasBaseReg.
3663     break;
3664   case 1:
3665     if (!AM.HasBaseReg) // allow "r+i".
3666       break;
3667     return false; // disallow "r+r" or "r+r+i".
3668   default:
3669     return false;
3670   }
3671 
3672   return true;
3673 }
3674 
3675 bool
3676 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3677   // The Mips target isn't yet aware of offsets.
3678   return false;
3679 }
3680 
3681 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3682                                             unsigned SrcAlign,
3683                                             bool IsMemset, bool ZeroMemset,
3684                                             bool MemcpyStrSrc,
3685                                             MachineFunction &MF) const {
3686   if (Subtarget.hasMips64())
3687     return MVT::i64;
3688 
3689   return MVT::i32;
3690 }
3691 
3692 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3693   if (VT != MVT::f32 && VT != MVT::f64)
3694     return false;
3695   if (Imm.isNegZero())
3696     return false;
3697   return Imm.isZero();
3698 }
3699 
3700 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3701   if (ABI.IsN64())
3702     return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3703 
3704   return TargetLowering::getJumpTableEncoding();
3705 }
3706 
3707 bool MipsTargetLowering::useSoftFloat() const {
3708   return Subtarget.useSoftFloat();
3709 }
3710 
3711 void MipsTargetLowering::copyByValRegs(
3712     SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains,
3713     SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3714     SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3715     unsigned FirstReg, unsigned LastReg, const CCValAssign &VA,
3716     MipsCCState &State) const {
3717   MachineFunction &MF = DAG.getMachineFunction();
3718   MachineFrameInfo &MFI = MF.getFrameInfo();
3719   unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes();
3720   unsigned NumRegs = LastReg - FirstReg;
3721   unsigned RegAreaSize = NumRegs * GPRSizeInBytes;
3722   unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3723   int FrameObjOffset;
3724   ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs();
3725 
3726   if (RegAreaSize)
3727     FrameObjOffset =
3728         (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
3729         (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes);
3730   else
3731     FrameObjOffset = VA.getLocMemOffset();
3732 
3733   // Create frame object.
3734   EVT PtrTy = getPointerTy(DAG.getDataLayout());
3735   int FI = MFI.CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3736   SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3737   InVals.push_back(FIN);
3738 
3739   if (!NumRegs)
3740     return;
3741 
3742   // Copy arg registers.
3743   MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8);
3744   const TargetRegisterClass *RC = getRegClassFor(RegTy);
3745 
3746   for (unsigned I = 0; I < NumRegs; ++I) {
3747     unsigned ArgReg = ByValArgRegs[FirstReg + I];
3748     unsigned VReg = addLiveIn(MF, ArgReg, RC);
3749     unsigned Offset = I * GPRSizeInBytes;
3750     SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3751                                    DAG.getConstant(Offset, DL, PtrTy));
3752     SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3753                                  StorePtr, MachinePointerInfo(FuncArg, Offset));
3754     OutChains.push_back(Store);
3755   }
3756 }
3757 
3758 // Copy byVal arg to registers and stack.
3759 void MipsTargetLowering::passByValArg(
3760     SDValue Chain, const SDLoc &DL,
3761     std::deque<std::pair<unsigned, SDValue>> &RegsToPass,
3762     SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
3763     MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg,
3764     unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle,
3765     const CCValAssign &VA) const {
3766   unsigned ByValSizeInBytes = Flags.getByValSize();
3767   unsigned OffsetInBytes = 0; // From beginning of struct
3768   unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3769   unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes);
3770   EVT PtrTy = getPointerTy(DAG.getDataLayout()),
3771       RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
3772   unsigned NumRegs = LastReg - FirstReg;
3773 
3774   if (NumRegs) {
3775     ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs();
3776     bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes);
3777     unsigned I = 0;
3778 
3779     // Copy words to registers.
3780     for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) {
3781       SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3782                                     DAG.getConstant(OffsetInBytes, DL, PtrTy));
3783       SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3784                                     MachinePointerInfo(), Alignment);
3785       MemOpChains.push_back(LoadVal.getValue(1));
3786       unsigned ArgReg = ArgRegs[FirstReg + I];
3787       RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3788     }
3789 
3790     // Return if the struct has been fully copied.
3791     if (ByValSizeInBytes == OffsetInBytes)
3792       return;
3793 
3794     // Copy the remainder of the byval argument with sub-word loads and shifts.
3795     if (LeftoverBytes) {
3796       SDValue Val;
3797 
3798       for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0;
3799            OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) {
3800         unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes;
3801 
3802         if (RemainingSizeInBytes < LoadSizeInBytes)
3803           continue;
3804 
3805         // Load subword.
3806         SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3807                                       DAG.getConstant(OffsetInBytes, DL,
3808                                                       PtrTy));
3809         SDValue LoadVal = DAG.getExtLoad(
3810             ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(),
3811             MVT::getIntegerVT(LoadSizeInBytes * 8), Alignment);
3812         MemOpChains.push_back(LoadVal.getValue(1));
3813 
3814         // Shift the loaded value.
3815         unsigned Shamt;
3816 
3817         if (isLittle)
3818           Shamt = TotalBytesLoaded * 8;
3819         else
3820           Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8;
3821 
3822         SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3823                                     DAG.getConstant(Shamt, DL, MVT::i32));
3824 
3825         if (Val.getNode())
3826           Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3827         else
3828           Val = Shift;
3829 
3830         OffsetInBytes += LoadSizeInBytes;
3831         TotalBytesLoaded += LoadSizeInBytes;
3832         Alignment = std::min(Alignment, LoadSizeInBytes);
3833       }
3834 
3835       unsigned ArgReg = ArgRegs[FirstReg + I];
3836       RegsToPass.push_back(std::make_pair(ArgReg, Val));
3837       return;
3838     }
3839   }
3840 
3841   // Copy remainder of byval arg to it with memcpy.
3842   unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes;
3843   SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3844                             DAG.getConstant(OffsetInBytes, DL, PtrTy));
3845   SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3846                             DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
3847   Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3848                         DAG.getConstant(MemCpySize, DL, PtrTy),
3849                         Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false,
3850                         /*isTailCall=*/false,
3851                         MachinePointerInfo(), MachinePointerInfo());
3852   MemOpChains.push_back(Chain);
3853 }
3854 
3855 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3856                                          SDValue Chain, const SDLoc &DL,
3857                                          SelectionDAG &DAG,
3858                                          CCState &State) const {
3859   ArrayRef<MCPhysReg> ArgRegs = ABI.GetVarArgRegs();
3860   unsigned Idx = State.getFirstUnallocated(ArgRegs);
3861   unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3862   MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
3863   const TargetRegisterClass *RC = getRegClassFor(RegTy);
3864   MachineFunction &MF = DAG.getMachineFunction();
3865   MachineFrameInfo &MFI = MF.getFrameInfo();
3866   MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3867 
3868   // Offset of the first variable argument from stack pointer.
3869   int VaArgOffset;
3870 
3871   if (ArgRegs.size() == Idx)
3872     VaArgOffset = alignTo(State.getNextStackOffset(), RegSizeInBytes);
3873   else {
3874     VaArgOffset =
3875         (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
3876         (int)(RegSizeInBytes * (ArgRegs.size() - Idx));
3877   }
3878 
3879   // Record the frame index of the first variable argument
3880   // which is a value necessary to VASTART.
3881   int FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
3882   MipsFI->setVarArgsFrameIndex(FI);
3883 
3884   // Copy the integer registers that have not been used for argument passing
3885   // to the argument register save area. For O32, the save area is allocated
3886   // in the caller's stack frame, while for N32/64, it is allocated in the
3887   // callee's stack frame.
3888   for (unsigned I = Idx; I < ArgRegs.size();
3889        ++I, VaArgOffset += RegSizeInBytes) {
3890     unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3891     SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3892     FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
3893     SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3894     SDValue Store =
3895         DAG.getStore(Chain, DL, ArgValue, PtrOff, MachinePointerInfo());
3896     cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(
3897         (Value *)nullptr);
3898     OutChains.push_back(Store);
3899   }
3900 }
3901 
3902 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size,
3903                                      unsigned Align) const {
3904   const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
3905 
3906   assert(Size && "Byval argument's size shouldn't be 0.");
3907 
3908   Align = std::min(Align, TFL->getStackAlignment());
3909 
3910   unsigned FirstReg = 0;
3911   unsigned NumRegs = 0;
3912 
3913   if (State->getCallingConv() != CallingConv::Fast) {
3914     unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3915     ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs();
3916     // FIXME: The O32 case actually describes no shadow registers.
3917     const MCPhysReg *ShadowRegs =
3918         ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs;
3919 
3920     // We used to check the size as well but we can't do that anymore since
3921     // CCState::HandleByVal() rounds up the size after calling this function.
3922     assert(!(Align % RegSizeInBytes) &&
3923            "Byval argument's alignment should be a multiple of"
3924            "RegSizeInBytes.");
3925 
3926     FirstReg = State->getFirstUnallocated(IntArgRegs);
3927 
3928     // If Align > RegSizeInBytes, the first arg register must be even.
3929     // FIXME: This condition happens to do the right thing but it's not the
3930     //        right way to test it. We want to check that the stack frame offset
3931     //        of the register is aligned.
3932     if ((Align > RegSizeInBytes) && (FirstReg % 2)) {
3933       State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]);
3934       ++FirstReg;
3935     }
3936 
3937     // Mark the registers allocated.
3938     Size = alignTo(Size, RegSizeInBytes);
3939     for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size());
3940          Size -= RegSizeInBytes, ++I, ++NumRegs)
3941       State->AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3942   }
3943 
3944   State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs);
3945 }
3946 
3947 MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI,
3948                                                         MachineBasicBlock *BB,
3949                                                         bool isFPCmp,
3950                                                         unsigned Opc) const {
3951   assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) &&
3952          "Subtarget already supports SELECT nodes with the use of"
3953          "conditional-move instructions.");
3954 
3955   const TargetInstrInfo *TII =
3956       Subtarget.getInstrInfo();
3957   DebugLoc DL = MI.getDebugLoc();
3958 
3959   // To "insert" a SELECT instruction, we actually have to insert the
3960   // diamond control-flow pattern.  The incoming instruction knows the
3961   // destination vreg to set, the condition code register to branch on, the
3962   // true/false values to select between, and a branch opcode to use.
3963   const BasicBlock *LLVM_BB = BB->getBasicBlock();
3964   MachineFunction::iterator It = ++BB->getIterator();
3965 
3966   //  thisMBB:
3967   //  ...
3968   //   TrueVal = ...
3969   //   setcc r1, r2, r3
3970   //   bNE   r1, r0, copy1MBB
3971   //   fallthrough --> copy0MBB
3972   MachineBasicBlock *thisMBB  = BB;
3973   MachineFunction *F = BB->getParent();
3974   MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3975   MachineBasicBlock *sinkMBB  = F->CreateMachineBasicBlock(LLVM_BB);
3976   F->insert(It, copy0MBB);
3977   F->insert(It, sinkMBB);
3978 
3979   // Transfer the remainder of BB and its successor edges to sinkMBB.
3980   sinkMBB->splice(sinkMBB->begin(), BB,
3981                   std::next(MachineBasicBlock::iterator(MI)), BB->end());
3982   sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
3983 
3984   // Next, add the true and fallthrough blocks as its successors.
3985   BB->addSuccessor(copy0MBB);
3986   BB->addSuccessor(sinkMBB);
3987 
3988   if (isFPCmp) {
3989     // bc1[tf] cc, sinkMBB
3990     BuildMI(BB, DL, TII->get(Opc))
3991         .addReg(MI.getOperand(1).getReg())
3992         .addMBB(sinkMBB);
3993   } else {
3994     // bne rs, $0, sinkMBB
3995     BuildMI(BB, DL, TII->get(Opc))
3996         .addReg(MI.getOperand(1).getReg())
3997         .addReg(Mips::ZERO)
3998         .addMBB(sinkMBB);
3999   }
4000 
4001   //  copy0MBB:
4002   //   %FalseValue = ...
4003   //   # fallthrough to sinkMBB
4004   BB = copy0MBB;
4005 
4006   // Update machine-CFG edges
4007   BB->addSuccessor(sinkMBB);
4008 
4009   //  sinkMBB:
4010   //   %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
4011   //  ...
4012   BB = sinkMBB;
4013 
4014   BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg())
4015       .addReg(MI.getOperand(2).getReg())
4016       .addMBB(thisMBB)
4017       .addReg(MI.getOperand(3).getReg())
4018       .addMBB(copy0MBB);
4019 
4020   MI.eraseFromParent(); // The pseudo instruction is gone now.
4021 
4022   return BB;
4023 }
4024 
4025 // FIXME? Maybe this could be a TableGen attribute on some registers and
4026 // this table could be generated automatically from RegInfo.
4027 unsigned MipsTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4028                                                SelectionDAG &DAG) const {
4029   // Named registers is expected to be fairly rare. For now, just support $28
4030   // since the linux kernel uses it.
4031   if (Subtarget.isGP64bit()) {
4032     unsigned Reg = StringSwitch<unsigned>(RegName)
4033                          .Case("$28", Mips::GP_64)
4034                          .Default(0);
4035     if (Reg)
4036       return Reg;
4037   } else {
4038     unsigned Reg = StringSwitch<unsigned>(RegName)
4039                          .Case("$28", Mips::GP)
4040                          .Default(0);
4041     if (Reg)
4042       return Reg;
4043   }
4044   report_fatal_error("Invalid register name global variable");
4045 }
4046