1 //===- MipsISelLowering.cpp - Mips DAG Lowering Implementation ------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the interfaces that Mips uses to lower LLVM code into a
11 // selection DAG.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "MipsISelLowering.h"
16 #include "InstPrinter/MipsInstPrinter.h"
17 #include "MCTargetDesc/MipsBaseInfo.h"
18 #include "MCTargetDesc/MipsMCTargetDesc.h"
19 #include "MipsCCState.h"
20 #include "MipsInstrInfo.h"
21 #include "MipsMachineFunction.h"
22 #include "MipsRegisterInfo.h"
23 #include "MipsSubtarget.h"
24 #include "MipsTargetMachine.h"
25 #include "MipsTargetObjectFile.h"
26 #include "llvm/ADT/APFloat.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/ADT/StringRef.h"
31 #include "llvm/ADT/StringSwitch.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/FunctionLoweringInfo.h"
34 #include "llvm/CodeGen/ISDOpcodes.h"
35 #include "llvm/CodeGen/MachineBasicBlock.h"
36 #include "llvm/CodeGen/MachineFrameInfo.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineInstr.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineJumpTableInfo.h"
41 #include "llvm/CodeGen/MachineMemOperand.h"
42 #include "llvm/CodeGen/MachineOperand.h"
43 #include "llvm/CodeGen/MachineRegisterInfo.h"
44 #include "llvm/CodeGen/RuntimeLibcalls.h"
45 #include "llvm/CodeGen/SelectionDAG.h"
46 #include "llvm/CodeGen/SelectionDAGNodes.h"
47 #include "llvm/CodeGen/TargetFrameLowering.h"
48 #include "llvm/CodeGen/TargetInstrInfo.h"
49 #include "llvm/CodeGen/TargetRegisterInfo.h"
50 #include "llvm/CodeGen/ValueTypes.h"
51 #include "llvm/IR/CallingConv.h"
52 #include "llvm/IR/Constants.h"
53 #include "llvm/IR/DataLayout.h"
54 #include "llvm/IR/DebugLoc.h"
55 #include "llvm/IR/DerivedTypes.h"
56 #include "llvm/IR/Function.h"
57 #include "llvm/IR/GlobalValue.h"
58 #include "llvm/IR/Type.h"
59 #include "llvm/IR/Value.h"
60 #include "llvm/MC/MCRegisterInfo.h"
61 #include "llvm/Support/Casting.h"
62 #include "llvm/Support/CodeGen.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/Compiler.h"
65 #include "llvm/Support/ErrorHandling.h"
66 #include "llvm/Support/MachineValueType.h"
67 #include "llvm/Support/MathExtras.h"
68 #include "llvm/Target/TargetMachine.h"
69 #include "llvm/Target/TargetOptions.h"
70 #include <algorithm>
71 #include <cassert>
72 #include <cctype>
73 #include <cstdint>
74 #include <deque>
75 #include <iterator>
76 #include <utility>
77 #include <vector>
78
79 using namespace llvm;
80
81 #define DEBUG_TYPE "mips-lower"
82
83 STATISTIC(NumTailCalls, "Number of tail calls");
84
85 static cl::opt<bool>
86 LargeGOT("mxgot", cl::Hidden,
87 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
88
89 static cl::opt<bool>
90 NoZeroDivCheck("mno-check-zero-division", cl::Hidden,
91 cl::desc("MIPS: Don't trap on integer division by zero."),
92 cl::init(false));
93
94 static const MCPhysReg Mips64DPRegs[8] = {
95 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
96 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
97 };
98
99 // If I is a shifted mask, set the size (Size) and the first bit of the
100 // mask (Pos), and return true.
101 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
isShiftedMask(uint64_t I,uint64_t & Pos,uint64_t & Size)102 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
103 if (!isShiftedMask_64(I))
104 return false;
105
106 Size = countPopulation(I);
107 Pos = countTrailingZeros(I);
108 return true;
109 }
110
111 // The MIPS MSA ABI passes vector arguments in the integer register set.
112 // The number of integer registers used is dependant on the ABI used.
getRegisterTypeForCallingConv(LLVMContext & Context,CallingConv::ID CC,EVT VT) const113 MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
114 CallingConv::ID CC,
115 EVT VT) const {
116 if (VT.isVector()) {
117 if (Subtarget.isABI_O32()) {
118 return MVT::i32;
119 } else {
120 return (VT.getSizeInBits() == 32) ? MVT::i32 : MVT::i64;
121 }
122 }
123 return MipsTargetLowering::getRegisterType(Context, VT);
124 }
125
getNumRegistersForCallingConv(LLVMContext & Context,CallingConv::ID CC,EVT VT) const126 unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context,
127 CallingConv::ID CC,
128 EVT VT) const {
129 if (VT.isVector())
130 return std::max((VT.getSizeInBits() / (Subtarget.isABI_O32() ? 32 : 64)),
131 1U);
132 return MipsTargetLowering::getNumRegisters(Context, VT);
133 }
134
getVectorTypeBreakdownForCallingConv(LLVMContext & Context,CallingConv::ID CC,EVT VT,EVT & IntermediateVT,unsigned & NumIntermediates,MVT & RegisterVT) const135 unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv(
136 LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
137 unsigned &NumIntermediates, MVT &RegisterVT) const {
138 // Break down vector types to either 2 i64s or 4 i32s.
139 RegisterVT = getRegisterTypeForCallingConv(Context, CC, VT);
140 IntermediateVT = RegisterVT;
141 NumIntermediates = VT.getSizeInBits() < RegisterVT.getSizeInBits()
142 ? VT.getVectorNumElements()
143 : VT.getSizeInBits() / RegisterVT.getSizeInBits();
144
145 return NumIntermediates;
146 }
147
getGlobalReg(SelectionDAG & DAG,EVT Ty) const148 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
149 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
150 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
151 }
152
getTargetNode(GlobalAddressSDNode * N,EVT Ty,SelectionDAG & DAG,unsigned Flag) const153 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
154 SelectionDAG &DAG,
155 unsigned Flag) const {
156 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag);
157 }
158
getTargetNode(ExternalSymbolSDNode * N,EVT Ty,SelectionDAG & DAG,unsigned Flag) const159 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty,
160 SelectionDAG &DAG,
161 unsigned Flag) const {
162 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
163 }
164
getTargetNode(BlockAddressSDNode * N,EVT Ty,SelectionDAG & DAG,unsigned Flag) const165 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty,
166 SelectionDAG &DAG,
167 unsigned Flag) const {
168 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
169 }
170
getTargetNode(JumpTableSDNode * N,EVT Ty,SelectionDAG & DAG,unsigned Flag) const171 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty,
172 SelectionDAG &DAG,
173 unsigned Flag) const {
174 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
175 }
176
getTargetNode(ConstantPoolSDNode * N,EVT Ty,SelectionDAG & DAG,unsigned Flag) const177 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty,
178 SelectionDAG &DAG,
179 unsigned Flag) const {
180 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
181 N->getOffset(), Flag);
182 }
183
getTargetNodeName(unsigned Opcode) const184 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
185 switch ((MipsISD::NodeType)Opcode) {
186 case MipsISD::FIRST_NUMBER: break;
187 case MipsISD::JmpLink: return "MipsISD::JmpLink";
188 case MipsISD::TailCall: return "MipsISD::TailCall";
189 case MipsISD::Highest: return "MipsISD::Highest";
190 case MipsISD::Higher: return "MipsISD::Higher";
191 case MipsISD::Hi: return "MipsISD::Hi";
192 case MipsISD::Lo: return "MipsISD::Lo";
193 case MipsISD::GotHi: return "MipsISD::GotHi";
194 case MipsISD::TlsHi: return "MipsISD::TlsHi";
195 case MipsISD::GPRel: return "MipsISD::GPRel";
196 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
197 case MipsISD::Ret: return "MipsISD::Ret";
198 case MipsISD::ERet: return "MipsISD::ERet";
199 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
200 case MipsISD::FMS: return "MipsISD::FMS";
201 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
202 case MipsISD::FPCmp: return "MipsISD::FPCmp";
203 case MipsISD::FSELECT: return "MipsISD::FSELECT";
204 case MipsISD::MTC1_D64: return "MipsISD::MTC1_D64";
205 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
206 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
207 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP";
208 case MipsISD::MFHI: return "MipsISD::MFHI";
209 case MipsISD::MFLO: return "MipsISD::MFLO";
210 case MipsISD::MTLOHI: return "MipsISD::MTLOHI";
211 case MipsISD::Mult: return "MipsISD::Mult";
212 case MipsISD::Multu: return "MipsISD::Multu";
213 case MipsISD::MAdd: return "MipsISD::MAdd";
214 case MipsISD::MAddu: return "MipsISD::MAddu";
215 case MipsISD::MSub: return "MipsISD::MSub";
216 case MipsISD::MSubu: return "MipsISD::MSubu";
217 case MipsISD::DivRem: return "MipsISD::DivRem";
218 case MipsISD::DivRemU: return "MipsISD::DivRemU";
219 case MipsISD::DivRem16: return "MipsISD::DivRem16";
220 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
221 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
222 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
223 case MipsISD::Wrapper: return "MipsISD::Wrapper";
224 case MipsISD::DynAlloc: return "MipsISD::DynAlloc";
225 case MipsISD::Sync: return "MipsISD::Sync";
226 case MipsISD::Ext: return "MipsISD::Ext";
227 case MipsISD::Ins: return "MipsISD::Ins";
228 case MipsISD::CIns: return "MipsISD::CIns";
229 case MipsISD::LWL: return "MipsISD::LWL";
230 case MipsISD::LWR: return "MipsISD::LWR";
231 case MipsISD::SWL: return "MipsISD::SWL";
232 case MipsISD::SWR: return "MipsISD::SWR";
233 case MipsISD::LDL: return "MipsISD::LDL";
234 case MipsISD::LDR: return "MipsISD::LDR";
235 case MipsISD::SDL: return "MipsISD::SDL";
236 case MipsISD::SDR: return "MipsISD::SDR";
237 case MipsISD::EXTP: return "MipsISD::EXTP";
238 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
239 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
240 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
241 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
242 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
243 case MipsISD::SHILO: return "MipsISD::SHILO";
244 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
245 case MipsISD::MULSAQ_S_W_PH: return "MipsISD::MULSAQ_S_W_PH";
246 case MipsISD::MAQ_S_W_PHL: return "MipsISD::MAQ_S_W_PHL";
247 case MipsISD::MAQ_S_W_PHR: return "MipsISD::MAQ_S_W_PHR";
248 case MipsISD::MAQ_SA_W_PHL: return "MipsISD::MAQ_SA_W_PHL";
249 case MipsISD::MAQ_SA_W_PHR: return "MipsISD::MAQ_SA_W_PHR";
250 case MipsISD::DPAU_H_QBL: return "MipsISD::DPAU_H_QBL";
251 case MipsISD::DPAU_H_QBR: return "MipsISD::DPAU_H_QBR";
252 case MipsISD::DPSU_H_QBL: return "MipsISD::DPSU_H_QBL";
253 case MipsISD::DPSU_H_QBR: return "MipsISD::DPSU_H_QBR";
254 case MipsISD::DPAQ_S_W_PH: return "MipsISD::DPAQ_S_W_PH";
255 case MipsISD::DPSQ_S_W_PH: return "MipsISD::DPSQ_S_W_PH";
256 case MipsISD::DPAQ_SA_L_W: return "MipsISD::DPAQ_SA_L_W";
257 case MipsISD::DPSQ_SA_L_W: return "MipsISD::DPSQ_SA_L_W";
258 case MipsISD::DPA_W_PH: return "MipsISD::DPA_W_PH";
259 case MipsISD::DPS_W_PH: return "MipsISD::DPS_W_PH";
260 case MipsISD::DPAQX_S_W_PH: return "MipsISD::DPAQX_S_W_PH";
261 case MipsISD::DPAQX_SA_W_PH: return "MipsISD::DPAQX_SA_W_PH";
262 case MipsISD::DPAX_W_PH: return "MipsISD::DPAX_W_PH";
263 case MipsISD::DPSX_W_PH: return "MipsISD::DPSX_W_PH";
264 case MipsISD::DPSQX_S_W_PH: return "MipsISD::DPSQX_S_W_PH";
265 case MipsISD::DPSQX_SA_W_PH: return "MipsISD::DPSQX_SA_W_PH";
266 case MipsISD::MULSA_W_PH: return "MipsISD::MULSA_W_PH";
267 case MipsISD::MULT: return "MipsISD::MULT";
268 case MipsISD::MULTU: return "MipsISD::MULTU";
269 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
270 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
271 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
272 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
273 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP";
274 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP";
275 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP";
276 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP";
277 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP";
278 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO";
279 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO";
280 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO";
281 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO";
282 case MipsISD::VCEQ: return "MipsISD::VCEQ";
283 case MipsISD::VCLE_S: return "MipsISD::VCLE_S";
284 case MipsISD::VCLE_U: return "MipsISD::VCLE_U";
285 case MipsISD::VCLT_S: return "MipsISD::VCLT_S";
286 case MipsISD::VCLT_U: return "MipsISD::VCLT_U";
287 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
288 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
289 case MipsISD::VNOR: return "MipsISD::VNOR";
290 case MipsISD::VSHF: return "MipsISD::VSHF";
291 case MipsISD::SHF: return "MipsISD::SHF";
292 case MipsISD::ILVEV: return "MipsISD::ILVEV";
293 case MipsISD::ILVOD: return "MipsISD::ILVOD";
294 case MipsISD::ILVL: return "MipsISD::ILVL";
295 case MipsISD::ILVR: return "MipsISD::ILVR";
296 case MipsISD::PCKEV: return "MipsISD::PCKEV";
297 case MipsISD::PCKOD: return "MipsISD::PCKOD";
298 case MipsISD::INSVE: return "MipsISD::INSVE";
299 }
300 return nullptr;
301 }
302
MipsTargetLowering(const MipsTargetMachine & TM,const MipsSubtarget & STI)303 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM,
304 const MipsSubtarget &STI)
305 : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) {
306 // Mips does not have i1 type, so use i32 for
307 // setcc operations results (slt, sgt, ...).
308 setBooleanContents(ZeroOrOneBooleanContent);
309 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
310 // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA
311 // does. Integer booleans still use 0 and 1.
312 if (Subtarget.hasMips32r6())
313 setBooleanContents(ZeroOrOneBooleanContent,
314 ZeroOrNegativeOneBooleanContent);
315
316 // Load extented operations for i1 types must be promoted
317 for (MVT VT : MVT::integer_valuetypes()) {
318 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
319 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
320 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
321 }
322
323 // MIPS doesn't have extending float->double load/store. Set LoadExtAction
324 // for f32, f16
325 for (MVT VT : MVT::fp_valuetypes()) {
326 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
327 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
328 }
329
330 // Set LoadExtAction for f16 vectors to Expand
331 for (MVT VT : MVT::fp_vector_valuetypes()) {
332 MVT F16VT = MVT::getVectorVT(MVT::f16, VT.getVectorNumElements());
333 if (F16VT.isValid())
334 setLoadExtAction(ISD::EXTLOAD, VT, F16VT, Expand);
335 }
336
337 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
338 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
339
340 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
341
342 // Used by legalize types to correctly generate the setcc result.
343 // Without this, every float setcc comes with a AND/OR with the result,
344 // we don't want this, since the fpcmp result goes to a flag register,
345 // which is used implicitly by brcond and select operations.
346 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
347
348 // Mips Custom Operations
349 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
350 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
351 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
352 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
353 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
354 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
355 setOperationAction(ISD::SELECT, MVT::f32, Custom);
356 setOperationAction(ISD::SELECT, MVT::f64, Custom);
357 setOperationAction(ISD::SELECT, MVT::i32, Custom);
358 setOperationAction(ISD::SETCC, MVT::f32, Custom);
359 setOperationAction(ISD::SETCC, MVT::f64, Custom);
360 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
361 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
362 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
363 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
364
365 if (Subtarget.isGP64bit()) {
366 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
367 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
368 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
369 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
370 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
371 setOperationAction(ISD::SELECT, MVT::i64, Custom);
372 setOperationAction(ISD::LOAD, MVT::i64, Custom);
373 setOperationAction(ISD::STORE, MVT::i64, Custom);
374 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
375 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
376 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
377 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
378 }
379
380 if (!Subtarget.isGP64bit()) {
381 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
382 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
383 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
384 }
385
386 setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom);
387 if (Subtarget.isGP64bit())
388 setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom);
389
390 setOperationAction(ISD::SDIV, MVT::i32, Expand);
391 setOperationAction(ISD::SREM, MVT::i32, Expand);
392 setOperationAction(ISD::UDIV, MVT::i32, Expand);
393 setOperationAction(ISD::UREM, MVT::i32, Expand);
394 setOperationAction(ISD::SDIV, MVT::i64, Expand);
395 setOperationAction(ISD::SREM, MVT::i64, Expand);
396 setOperationAction(ISD::UDIV, MVT::i64, Expand);
397 setOperationAction(ISD::UREM, MVT::i64, Expand);
398
399 // Operations not directly supported by Mips.
400 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
401 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
402 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
403 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
404 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
405 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
406 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
407 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
408 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
409 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
410 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
411 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
412 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
413 if (Subtarget.hasCnMips()) {
414 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
415 setOperationAction(ISD::CTPOP, MVT::i64, Legal);
416 } else {
417 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
418 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
419 }
420 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
421 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
422 setOperationAction(ISD::ROTL, MVT::i32, Expand);
423 setOperationAction(ISD::ROTL, MVT::i64, Expand);
424 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
425 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
426
427 if (!Subtarget.hasMips32r2())
428 setOperationAction(ISD::ROTR, MVT::i32, Expand);
429
430 if (!Subtarget.hasMips64r2())
431 setOperationAction(ISD::ROTR, MVT::i64, Expand);
432
433 setOperationAction(ISD::FSIN, MVT::f32, Expand);
434 setOperationAction(ISD::FSIN, MVT::f64, Expand);
435 setOperationAction(ISD::FCOS, MVT::f32, Expand);
436 setOperationAction(ISD::FCOS, MVT::f64, Expand);
437 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
438 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
439 setOperationAction(ISD::FPOW, MVT::f32, Expand);
440 setOperationAction(ISD::FPOW, MVT::f64, Expand);
441 setOperationAction(ISD::FLOG, MVT::f32, Expand);
442 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
443 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
444 setOperationAction(ISD::FEXP, MVT::f32, Expand);
445 setOperationAction(ISD::FMA, MVT::f32, Expand);
446 setOperationAction(ISD::FMA, MVT::f64, Expand);
447 setOperationAction(ISD::FREM, MVT::f32, Expand);
448 setOperationAction(ISD::FREM, MVT::f64, Expand);
449
450 // Lower f16 conversion operations into library calls
451 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
452 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
453 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
454 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
455
456 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
457
458 setOperationAction(ISD::VASTART, MVT::Other, Custom);
459 setOperationAction(ISD::VAARG, MVT::Other, Custom);
460 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
461 setOperationAction(ISD::VAEND, MVT::Other, Expand);
462
463 // Use the default for now
464 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
465 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
466
467 if (!Subtarget.isGP64bit()) {
468 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
469 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
470 }
471
472 if (!Subtarget.hasMips32r2()) {
473 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
474 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
475 }
476
477 // MIPS16 lacks MIPS32's clz and clo instructions.
478 if (!Subtarget.hasMips32() || Subtarget.inMips16Mode())
479 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
480 if (!Subtarget.hasMips64())
481 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
482
483 if (!Subtarget.hasMips32r2())
484 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
485 if (!Subtarget.hasMips64r2())
486 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
487
488 if (Subtarget.isGP64bit()) {
489 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom);
490 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom);
491 setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom);
492 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
493 }
494
495 setOperationAction(ISD::TRAP, MVT::Other, Legal);
496
497 setTargetDAGCombine(ISD::SDIVREM);
498 setTargetDAGCombine(ISD::UDIVREM);
499 setTargetDAGCombine(ISD::SELECT);
500 setTargetDAGCombine(ISD::AND);
501 setTargetDAGCombine(ISD::OR);
502 setTargetDAGCombine(ISD::ADD);
503 setTargetDAGCombine(ISD::SUB);
504 setTargetDAGCombine(ISD::AssertZext);
505 setTargetDAGCombine(ISD::SHL);
506
507 if (ABI.IsO32()) {
508 // These libcalls are not available in 32-bit.
509 setLibcallName(RTLIB::SHL_I128, nullptr);
510 setLibcallName(RTLIB::SRL_I128, nullptr);
511 setLibcallName(RTLIB::SRA_I128, nullptr);
512 }
513
514 setMinFunctionAlignment(Subtarget.isGP64bit() ? 3 : 2);
515
516 // The arguments on the stack are defined in terms of 4-byte slots on O32
517 // and 8-byte slots on N32/N64.
518 setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? 8 : 4);
519
520 setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP);
521
522 MaxStoresPerMemcpy = 16;
523
524 isMicroMips = Subtarget.inMicroMipsMode();
525 }
526
create(const MipsTargetMachine & TM,const MipsSubtarget & STI)527 const MipsTargetLowering *MipsTargetLowering::create(const MipsTargetMachine &TM,
528 const MipsSubtarget &STI) {
529 if (STI.inMips16Mode())
530 return createMips16TargetLowering(TM, STI);
531
532 return createMipsSETargetLowering(TM, STI);
533 }
534
535 // Create a fast isel object.
536 FastISel *
createFastISel(FunctionLoweringInfo & funcInfo,const TargetLibraryInfo * libInfo) const537 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
538 const TargetLibraryInfo *libInfo) const {
539 const MipsTargetMachine &TM =
540 static_cast<const MipsTargetMachine &>(funcInfo.MF->getTarget());
541
542 // We support only the standard encoding [MIPS32,MIPS32R5] ISAs.
543 bool UseFastISel = TM.Options.EnableFastISel && Subtarget.hasMips32() &&
544 !Subtarget.hasMips32r6() && !Subtarget.inMips16Mode() &&
545 !Subtarget.inMicroMipsMode();
546
547 // Disable if either of the following is true:
548 // We do not generate PIC, the ABI is not O32, LargeGOT is being used.
549 if (!TM.isPositionIndependent() || !TM.getABI().IsO32() || LargeGOT)
550 UseFastISel = false;
551
552 return UseFastISel ? Mips::createFastISel(funcInfo, libInfo) : nullptr;
553 }
554
getSetCCResultType(const DataLayout &,LLVMContext &,EVT VT) const555 EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
556 EVT VT) const {
557 if (!VT.isVector())
558 return MVT::i32;
559 return VT.changeVectorElementTypeToInteger();
560 }
561
performDivRemCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)562 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
563 TargetLowering::DAGCombinerInfo &DCI,
564 const MipsSubtarget &Subtarget) {
565 if (DCI.isBeforeLegalizeOps())
566 return SDValue();
567
568 EVT Ty = N->getValueType(0);
569 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
570 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
571 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
572 MipsISD::DivRemU16;
573 SDLoc DL(N);
574
575 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
576 N->getOperand(0), N->getOperand(1));
577 SDValue InChain = DAG.getEntryNode();
578 SDValue InGlue = DivRem;
579
580 // insert MFLO
581 if (N->hasAnyUseOfValue(0)) {
582 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
583 InGlue);
584 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
585 InChain = CopyFromLo.getValue(1);
586 InGlue = CopyFromLo.getValue(2);
587 }
588
589 // insert MFHI
590 if (N->hasAnyUseOfValue(1)) {
591 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
592 HI, Ty, InGlue);
593 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
594 }
595
596 return SDValue();
597 }
598
condCodeToFCC(ISD::CondCode CC)599 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
600 switch (CC) {
601 default: llvm_unreachable("Unknown fp condition code!");
602 case ISD::SETEQ:
603 case ISD::SETOEQ: return Mips::FCOND_OEQ;
604 case ISD::SETUNE: return Mips::FCOND_UNE;
605 case ISD::SETLT:
606 case ISD::SETOLT: return Mips::FCOND_OLT;
607 case ISD::SETGT:
608 case ISD::SETOGT: return Mips::FCOND_OGT;
609 case ISD::SETLE:
610 case ISD::SETOLE: return Mips::FCOND_OLE;
611 case ISD::SETGE:
612 case ISD::SETOGE: return Mips::FCOND_OGE;
613 case ISD::SETULT: return Mips::FCOND_ULT;
614 case ISD::SETULE: return Mips::FCOND_ULE;
615 case ISD::SETUGT: return Mips::FCOND_UGT;
616 case ISD::SETUGE: return Mips::FCOND_UGE;
617 case ISD::SETUO: return Mips::FCOND_UN;
618 case ISD::SETO: return Mips::FCOND_OR;
619 case ISD::SETNE:
620 case ISD::SETONE: return Mips::FCOND_ONE;
621 case ISD::SETUEQ: return Mips::FCOND_UEQ;
622 }
623 }
624
625 /// This function returns true if the floating point conditional branches and
626 /// conditional moves which use condition code CC should be inverted.
invertFPCondCodeUser(Mips::CondCode CC)627 static bool invertFPCondCodeUser(Mips::CondCode CC) {
628 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
629 return false;
630
631 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
632 "Illegal Condition Code");
633
634 return true;
635 }
636
637 // Creates and returns an FPCmp node from a setcc node.
638 // Returns Op if setcc is not a floating point comparison.
createFPCmp(SelectionDAG & DAG,const SDValue & Op)639 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
640 // must be a SETCC node
641 if (Op.getOpcode() != ISD::SETCC)
642 return Op;
643
644 SDValue LHS = Op.getOperand(0);
645
646 if (!LHS.getValueType().isFloatingPoint())
647 return Op;
648
649 SDValue RHS = Op.getOperand(1);
650 SDLoc DL(Op);
651
652 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
653 // node if necessary.
654 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
655
656 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
657 DAG.getConstant(condCodeToFCC(CC), DL, MVT::i32));
658 }
659
660 // Creates and returns a CMovFPT/F node.
createCMovFP(SelectionDAG & DAG,SDValue Cond,SDValue True,SDValue False,const SDLoc & DL)661 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
662 SDValue False, const SDLoc &DL) {
663 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
664 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
665 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
666
667 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
668 True.getValueType(), True, FCC0, False, Cond);
669 }
670
performSELECTCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)671 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
672 TargetLowering::DAGCombinerInfo &DCI,
673 const MipsSubtarget &Subtarget) {
674 if (DCI.isBeforeLegalizeOps())
675 return SDValue();
676
677 SDValue SetCC = N->getOperand(0);
678
679 if ((SetCC.getOpcode() != ISD::SETCC) ||
680 !SetCC.getOperand(0).getValueType().isInteger())
681 return SDValue();
682
683 SDValue False = N->getOperand(2);
684 EVT FalseTy = False.getValueType();
685
686 if (!FalseTy.isInteger())
687 return SDValue();
688
689 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False);
690
691 // If the RHS (False) is 0, we swap the order of the operands
692 // of ISD::SELECT (obviously also inverting the condition) so that we can
693 // take advantage of conditional moves using the $0 register.
694 // Example:
695 // return (a != 0) ? x : 0;
696 // load $reg, x
697 // movz $reg, $0, a
698 if (!FalseC)
699 return SDValue();
700
701 const SDLoc DL(N);
702
703 if (!FalseC->getZExtValue()) {
704 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
705 SDValue True = N->getOperand(1);
706
707 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
708 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
709
710 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
711 }
712
713 // If both operands are integer constants there's a possibility that we
714 // can do some interesting optimizations.
715 SDValue True = N->getOperand(1);
716 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True);
717
718 if (!TrueC || !True.getValueType().isInteger())
719 return SDValue();
720
721 // We'll also ignore MVT::i64 operands as this optimizations proves
722 // to be ineffective because of the required sign extensions as the result
723 // of a SETCC operator is always MVT::i32 for non-vector types.
724 if (True.getValueType() == MVT::i64)
725 return SDValue();
726
727 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue();
728
729 // 1) (a < x) ? y : y-1
730 // slti $reg1, a, x
731 // addiu $reg2, $reg1, y-1
732 if (Diff == 1)
733 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False);
734
735 // 2) (a < x) ? y-1 : y
736 // slti $reg1, a, x
737 // xor $reg1, $reg1, 1
738 // addiu $reg2, $reg1, y-1
739 if (Diff == -1) {
740 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
741 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
742 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
743 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True);
744 }
745
746 // Could not optimize.
747 return SDValue();
748 }
749
performCMovFPCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)750 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG,
751 TargetLowering::DAGCombinerInfo &DCI,
752 const MipsSubtarget &Subtarget) {
753 if (DCI.isBeforeLegalizeOps())
754 return SDValue();
755
756 SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2);
757
758 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse);
759 if (!FalseC || FalseC->getZExtValue())
760 return SDValue();
761
762 // Since RHS (False) is 0, we swap the order of the True/False operands
763 // (obviously also inverting the condition) so that we can
764 // take advantage of conditional moves using the $0 register.
765 // Example:
766 // return (a != 0) ? x : 0;
767 // load $reg, x
768 // movz $reg, $0, a
769 unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F :
770 MipsISD::CMovFP_T;
771
772 SDValue FCC = N->getOperand(1), Glue = N->getOperand(3);
773 return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(),
774 ValueIfFalse, FCC, ValueIfTrue, Glue);
775 }
776
performANDCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)777 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
778 TargetLowering::DAGCombinerInfo &DCI,
779 const MipsSubtarget &Subtarget) {
780 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
781 return SDValue();
782
783 SDValue FirstOperand = N->getOperand(0);
784 unsigned FirstOperandOpc = FirstOperand.getOpcode();
785 SDValue Mask = N->getOperand(1);
786 EVT ValTy = N->getValueType(0);
787 SDLoc DL(N);
788
789 uint64_t Pos = 0, SMPos, SMSize;
790 ConstantSDNode *CN;
791 SDValue NewOperand;
792 unsigned Opc;
793
794 // Op's second operand must be a shifted mask.
795 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
796 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
797 return SDValue();
798
799 if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) {
800 // Pattern match EXT.
801 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
802 // => ext $dst, $src, pos, size
803
804 // The second operand of the shift must be an immediate.
805 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))))
806 return SDValue();
807
808 Pos = CN->getZExtValue();
809
810 // Return if the shifted mask does not start at bit 0 or the sum of its size
811 // and Pos exceeds the word's size.
812 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
813 return SDValue();
814
815 Opc = MipsISD::Ext;
816 NewOperand = FirstOperand.getOperand(0);
817 } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) {
818 // Pattern match CINS.
819 // $dst = and (shl $src , pos), mask
820 // => cins $dst, $src, pos, size
821 // mask is a shifted mask with consecutive 1's, pos = shift amount,
822 // size = population count.
823
824 // The second operand of the shift must be an immediate.
825 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))))
826 return SDValue();
827
828 Pos = CN->getZExtValue();
829
830 if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 ||
831 Pos + SMSize > ValTy.getSizeInBits())
832 return SDValue();
833
834 NewOperand = FirstOperand.getOperand(0);
835 // SMSize is 'location' (position) in this case, not size.
836 SMSize--;
837 Opc = MipsISD::CIns;
838 } else {
839 // Pattern match EXT.
840 // $dst = and $src, (2**size - 1) , if size > 16
841 // => ext $dst, $src, pos, size , pos = 0
842
843 // If the mask is <= 0xffff, andi can be used instead.
844 if (CN->getZExtValue() <= 0xffff)
845 return SDValue();
846
847 // Return if the mask doesn't start at position 0.
848 if (SMPos)
849 return SDValue();
850
851 Opc = MipsISD::Ext;
852 NewOperand = FirstOperand;
853 }
854 return DAG.getNode(Opc, DL, ValTy, NewOperand,
855 DAG.getConstant(Pos, DL, MVT::i32),
856 DAG.getConstant(SMSize, DL, MVT::i32));
857 }
858
performORCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)859 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
860 TargetLowering::DAGCombinerInfo &DCI,
861 const MipsSubtarget &Subtarget) {
862 // Pattern match INS.
863 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
864 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
865 // => ins $dst, $src, size, pos, $src1
866 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
867 return SDValue();
868
869 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
870 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
871 ConstantSDNode *CN, *CN1;
872
873 // See if Op's first operand matches (and $src1 , mask0).
874 if (And0.getOpcode() != ISD::AND)
875 return SDValue();
876
877 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
878 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
879 return SDValue();
880
881 // See if Op's second operand matches (and (shl $src, pos), mask1).
882 if (And1.getOpcode() == ISD::AND &&
883 And1.getOperand(0).getOpcode() == ISD::SHL) {
884
885 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
886 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
887 return SDValue();
888
889 // The shift masks must have the same position and size.
890 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
891 return SDValue();
892
893 SDValue Shl = And1.getOperand(0);
894
895 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
896 return SDValue();
897
898 unsigned Shamt = CN->getZExtValue();
899
900 // Return if the shift amount and the first bit position of mask are not the
901 // same.
902 EVT ValTy = N->getValueType(0);
903 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
904 return SDValue();
905
906 SDLoc DL(N);
907 return DAG.getNode(MipsISD::Ins, DL, ValTy, Shl.getOperand(0),
908 DAG.getConstant(SMPos0, DL, MVT::i32),
909 DAG.getConstant(SMSize0, DL, MVT::i32),
910 And0.getOperand(0));
911 } else {
912 // Pattern match DINS.
913 // $dst = or (and $src, mask0), mask1
914 // where mask0 = ((1 << SMSize0) -1) << SMPos0
915 // => dins $dst, $src, pos, size
916 if (~CN->getSExtValue() == ((((int64_t)1 << SMSize0) - 1) << SMPos0) &&
917 ((SMSize0 + SMPos0 <= 64 && Subtarget.hasMips64r2()) ||
918 (SMSize0 + SMPos0 <= 32))) {
919 // Check if AND instruction has constant as argument
920 bool isConstCase = And1.getOpcode() != ISD::AND;
921 if (And1.getOpcode() == ISD::AND) {
922 if (!(CN1 = dyn_cast<ConstantSDNode>(And1->getOperand(1))))
923 return SDValue();
924 } else {
925 if (!(CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1))))
926 return SDValue();
927 }
928 // Don't generate INS if constant OR operand doesn't fit into bits
929 // cleared by constant AND operand.
930 if (CN->getSExtValue() & CN1->getSExtValue())
931 return SDValue();
932
933 SDLoc DL(N);
934 EVT ValTy = N->getOperand(0)->getValueType(0);
935 SDValue Const1;
936 SDValue SrlX;
937 if (!isConstCase) {
938 Const1 = DAG.getConstant(SMPos0, DL, MVT::i32);
939 SrlX = DAG.getNode(ISD::SRL, DL, And1->getValueType(0), And1, Const1);
940 }
941 return DAG.getNode(
942 MipsISD::Ins, DL, N->getValueType(0),
943 isConstCase
944 ? DAG.getConstant(CN1->getSExtValue() >> SMPos0, DL, ValTy)
945 : SrlX,
946 DAG.getConstant(SMPos0, DL, MVT::i32),
947 DAG.getConstant(ValTy.getSizeInBits() / 8 < 8 ? SMSize0 & 31
948 : SMSize0,
949 DL, MVT::i32),
950 And0->getOperand(0));
951
952 }
953 return SDValue();
954 }
955 }
956
performMADD_MSUBCombine(SDNode * ROOTNode,SelectionDAG & CurDAG,const MipsSubtarget & Subtarget)957 static SDValue performMADD_MSUBCombine(SDNode *ROOTNode, SelectionDAG &CurDAG,
958 const MipsSubtarget &Subtarget) {
959 // ROOTNode must have a multiplication as an operand for the match to be
960 // successful.
961 if (ROOTNode->getOperand(0).getOpcode() != ISD::MUL &&
962 ROOTNode->getOperand(1).getOpcode() != ISD::MUL)
963 return SDValue();
964
965 // We don't handle vector types here.
966 if (ROOTNode->getValueType(0).isVector())
967 return SDValue();
968
969 // For MIPS64, madd / msub instructions are inefficent to use with 64 bit
970 // arithmetic. E.g.
971 // (add (mul a b) c) =>
972 // let res = (madd (mthi (drotr c 32))x(mtlo c) a b) in
973 // MIPS64: (or (dsll (mfhi res) 32) (dsrl (dsll (mflo res) 32) 32)
974 // or
975 // MIPS64R2: (dins (mflo res) (mfhi res) 32 32)
976 //
977 // The overhead of setting up the Hi/Lo registers and reassembling the
978 // result makes this a dubious optimzation for MIPS64. The core of the
979 // problem is that Hi/Lo contain the upper and lower 32 bits of the
980 // operand and result.
981 //
982 // It requires a chain of 4 add/mul for MIPS64R2 to get better code
983 // density than doing it naively, 5 for MIPS64. Additionally, using
984 // madd/msub on MIPS64 requires the operands actually be 32 bit sign
985 // extended operands, not true 64 bit values.
986 //
987 // FIXME: For the moment, disable this completely for MIPS64.
988 if (Subtarget.hasMips64())
989 return SDValue();
990
991 SDValue Mult = ROOTNode->getOperand(0).getOpcode() == ISD::MUL
992 ? ROOTNode->getOperand(0)
993 : ROOTNode->getOperand(1);
994
995 SDValue AddOperand = ROOTNode->getOperand(0).getOpcode() == ISD::MUL
996 ? ROOTNode->getOperand(1)
997 : ROOTNode->getOperand(0);
998
999 // Transform this to a MADD only if the user of this node is the add.
1000 // If there are other users of the mul, this function returns here.
1001 if (!Mult.hasOneUse())
1002 return SDValue();
1003
1004 // maddu and madd are unusual instructions in that on MIPS64 bits 63..31
1005 // must be in canonical form, i.e. sign extended. For MIPS32, the operands
1006 // of the multiply must have 32 or more sign bits, otherwise we cannot
1007 // perform this optimization. We have to check this here as we're performing
1008 // this optimization pre-legalization.
1009 SDValue MultLHS = Mult->getOperand(0);
1010 SDValue MultRHS = Mult->getOperand(1);
1011
1012 bool IsSigned = MultLHS->getOpcode() == ISD::SIGN_EXTEND &&
1013 MultRHS->getOpcode() == ISD::SIGN_EXTEND;
1014 bool IsUnsigned = MultLHS->getOpcode() == ISD::ZERO_EXTEND &&
1015 MultRHS->getOpcode() == ISD::ZERO_EXTEND;
1016
1017 if (!IsSigned && !IsUnsigned)
1018 return SDValue();
1019
1020 // Initialize accumulator.
1021 SDLoc DL(ROOTNode);
1022 SDValue TopHalf;
1023 SDValue BottomHalf;
1024 BottomHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand,
1025 CurDAG.getIntPtrConstant(0, DL));
1026
1027 TopHalf = CurDAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, AddOperand,
1028 CurDAG.getIntPtrConstant(1, DL));
1029 SDValue ACCIn = CurDAG.getNode(MipsISD::MTLOHI, DL, MVT::Untyped,
1030 BottomHalf,
1031 TopHalf);
1032
1033 // Create MipsMAdd(u) / MipsMSub(u) node.
1034 bool IsAdd = ROOTNode->getOpcode() == ISD::ADD;
1035 unsigned Opcode = IsAdd ? (IsUnsigned ? MipsISD::MAddu : MipsISD::MAdd)
1036 : (IsUnsigned ? MipsISD::MSubu : MipsISD::MSub);
1037 SDValue MAddOps[3] = {
1038 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(0)),
1039 CurDAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mult->getOperand(1)), ACCIn};
1040 EVT VTs[2] = {MVT::i32, MVT::i32};
1041 SDValue MAdd = CurDAG.getNode(Opcode, DL, VTs, MAddOps);
1042
1043 SDValue ResLo = CurDAG.getNode(MipsISD::MFLO, DL, MVT::i32, MAdd);
1044 SDValue ResHi = CurDAG.getNode(MipsISD::MFHI, DL, MVT::i32, MAdd);
1045 SDValue Combined =
1046 CurDAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResLo, ResHi);
1047 return Combined;
1048 }
1049
performSUBCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)1050 static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG,
1051 TargetLowering::DAGCombinerInfo &DCI,
1052 const MipsSubtarget &Subtarget) {
1053 // (sub v0 (mul v1, v2)) => (msub v1, v2, v0)
1054 if (DCI.isBeforeLegalizeOps()) {
1055 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
1056 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64)
1057 return performMADD_MSUBCombine(N, DAG, Subtarget);
1058
1059 return SDValue();
1060 }
1061
1062 return SDValue();
1063 }
1064
performADDCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)1065 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
1066 TargetLowering::DAGCombinerInfo &DCI,
1067 const MipsSubtarget &Subtarget) {
1068 // (add v0 (mul v1, v2)) => (madd v1, v2, v0)
1069 if (DCI.isBeforeLegalizeOps()) {
1070 if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() &&
1071 !Subtarget.inMips16Mode() && N->getValueType(0) == MVT::i64)
1072 return performMADD_MSUBCombine(N, DAG, Subtarget);
1073
1074 return SDValue();
1075 }
1076
1077 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
1078 SDValue Add = N->getOperand(1);
1079
1080 if (Add.getOpcode() != ISD::ADD)
1081 return SDValue();
1082
1083 SDValue Lo = Add.getOperand(1);
1084
1085 if ((Lo.getOpcode() != MipsISD::Lo) ||
1086 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
1087 return SDValue();
1088
1089 EVT ValTy = N->getValueType(0);
1090 SDLoc DL(N);
1091
1092 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
1093 Add.getOperand(0));
1094 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
1095 }
1096
performSHLCombine(SDNode * N,SelectionDAG & DAG,TargetLowering::DAGCombinerInfo & DCI,const MipsSubtarget & Subtarget)1097 static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
1098 TargetLowering::DAGCombinerInfo &DCI,
1099 const MipsSubtarget &Subtarget) {
1100 // Pattern match CINS.
1101 // $dst = shl (and $src , imm), pos
1102 // => cins $dst, $src, pos, size
1103
1104 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips())
1105 return SDValue();
1106
1107 SDValue FirstOperand = N->getOperand(0);
1108 unsigned FirstOperandOpc = FirstOperand.getOpcode();
1109 SDValue SecondOperand = N->getOperand(1);
1110 EVT ValTy = N->getValueType(0);
1111 SDLoc DL(N);
1112
1113 uint64_t Pos = 0, SMPos, SMSize;
1114 ConstantSDNode *CN;
1115 SDValue NewOperand;
1116
1117 // The second operand of the shift must be an immediate.
1118 if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand)))
1119 return SDValue();
1120
1121 Pos = CN->getZExtValue();
1122
1123 if (Pos >= ValTy.getSizeInBits())
1124 return SDValue();
1125
1126 if (FirstOperandOpc != ISD::AND)
1127 return SDValue();
1128
1129 // AND's second operand must be a shifted mask.
1130 if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) ||
1131 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
1132 return SDValue();
1133
1134 // Return if the shifted mask does not start at bit 0 or the sum of its size
1135 // and Pos exceeds the word's size.
1136 if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits())
1137 return SDValue();
1138
1139 NewOperand = FirstOperand.getOperand(0);
1140 // SMSize is 'location' (position) in this case, not size.
1141 SMSize--;
1142
1143 return DAG.getNode(MipsISD::CIns, DL, ValTy, NewOperand,
1144 DAG.getConstant(Pos, DL, MVT::i32),
1145 DAG.getConstant(SMSize, DL, MVT::i32));
1146 }
1147
PerformDAGCombine(SDNode * N,DAGCombinerInfo & DCI) const1148 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
1149 const {
1150 SelectionDAG &DAG = DCI.DAG;
1151 unsigned Opc = N->getOpcode();
1152
1153 switch (Opc) {
1154 default: break;
1155 case ISD::SDIVREM:
1156 case ISD::UDIVREM:
1157 return performDivRemCombine(N, DAG, DCI, Subtarget);
1158 case ISD::SELECT:
1159 return performSELECTCombine(N, DAG, DCI, Subtarget);
1160 case MipsISD::CMovFP_F:
1161 case MipsISD::CMovFP_T:
1162 return performCMovFPCombine(N, DAG, DCI, Subtarget);
1163 case ISD::AND:
1164 return performANDCombine(N, DAG, DCI, Subtarget);
1165 case ISD::OR:
1166 return performORCombine(N, DAG, DCI, Subtarget);
1167 case ISD::ADD:
1168 return performADDCombine(N, DAG, DCI, Subtarget);
1169 case ISD::SHL:
1170 return performSHLCombine(N, DAG, DCI, Subtarget);
1171 case ISD::SUB:
1172 return performSUBCombine(N, DAG, DCI, Subtarget);
1173 }
1174
1175 return SDValue();
1176 }
1177
isCheapToSpeculateCttz() const1178 bool MipsTargetLowering::isCheapToSpeculateCttz() const {
1179 return Subtarget.hasMips32();
1180 }
1181
isCheapToSpeculateCtlz() const1182 bool MipsTargetLowering::isCheapToSpeculateCtlz() const {
1183 return Subtarget.hasMips32();
1184 }
1185
1186 void
LowerOperationWrapper(SDNode * N,SmallVectorImpl<SDValue> & Results,SelectionDAG & DAG) const1187 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
1188 SmallVectorImpl<SDValue> &Results,
1189 SelectionDAG &DAG) const {
1190 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
1191
1192 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
1193 Results.push_back(Res.getValue(I));
1194 }
1195
1196 void
ReplaceNodeResults(SDNode * N,SmallVectorImpl<SDValue> & Results,SelectionDAG & DAG) const1197 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
1198 SmallVectorImpl<SDValue> &Results,
1199 SelectionDAG &DAG) const {
1200 return LowerOperationWrapper(N, Results, DAG);
1201 }
1202
1203 SDValue MipsTargetLowering::
LowerOperation(SDValue Op,SelectionDAG & DAG) const1204 LowerOperation(SDValue Op, SelectionDAG &DAG) const
1205 {
1206 switch (Op.getOpcode())
1207 {
1208 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
1209 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
1210 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
1211 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
1212 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
1213 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
1214 case ISD::SELECT: return lowerSELECT(Op, DAG);
1215 case ISD::SETCC: return lowerSETCC(Op, DAG);
1216 case ISD::VASTART: return lowerVASTART(Op, DAG);
1217 case ISD::VAARG: return lowerVAARG(Op, DAG);
1218 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
1219 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
1220 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
1221 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
1222 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
1223 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
1224 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
1225 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
1226 case ISD::LOAD: return lowerLOAD(Op, DAG);
1227 case ISD::STORE: return lowerSTORE(Op, DAG);
1228 case ISD::EH_DWARF_CFA: return lowerEH_DWARF_CFA(Op, DAG);
1229 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG);
1230 }
1231 return SDValue();
1232 }
1233
1234 //===----------------------------------------------------------------------===//
1235 // Lower helper functions
1236 //===----------------------------------------------------------------------===//
1237
1238 // addLiveIn - This helper function adds the specified physical register to the
1239 // MachineFunction as a live in value. It also creates a corresponding
1240 // virtual register for it.
1241 static unsigned
addLiveIn(MachineFunction & MF,unsigned PReg,const TargetRegisterClass * RC)1242 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
1243 {
1244 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1245 MF.getRegInfo().addLiveIn(PReg, VReg);
1246 return VReg;
1247 }
1248
insertDivByZeroTrap(MachineInstr & MI,MachineBasicBlock & MBB,const TargetInstrInfo & TII,bool Is64Bit,bool IsMicroMips)1249 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI,
1250 MachineBasicBlock &MBB,
1251 const TargetInstrInfo &TII,
1252 bool Is64Bit, bool IsMicroMips) {
1253 if (NoZeroDivCheck)
1254 return &MBB;
1255
1256 // Insert instruction "teq $divisor_reg, $zero, 7".
1257 MachineBasicBlock::iterator I(MI);
1258 MachineInstrBuilder MIB;
1259 MachineOperand &Divisor = MI.getOperand(2);
1260 MIB = BuildMI(MBB, std::next(I), MI.getDebugLoc(),
1261 TII.get(IsMicroMips ? Mips::TEQ_MM : Mips::TEQ))
1262 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill()))
1263 .addReg(Mips::ZERO)
1264 .addImm(7);
1265
1266 // Use the 32-bit sub-register if this is a 64-bit division.
1267 if (Is64Bit)
1268 MIB->getOperand(0).setSubReg(Mips::sub_32);
1269
1270 // Clear Divisor's kill flag.
1271 Divisor.setIsKill(false);
1272
1273 // We would normally delete the original instruction here but in this case
1274 // we only needed to inject an additional instruction rather than replace it.
1275
1276 return &MBB;
1277 }
1278
1279 MachineBasicBlock *
EmitInstrWithCustomInserter(MachineInstr & MI,MachineBasicBlock * BB) const1280 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
1281 MachineBasicBlock *BB) const {
1282 switch (MI.getOpcode()) {
1283 default:
1284 llvm_unreachable("Unexpected instr type to insert");
1285 case Mips::ATOMIC_LOAD_ADD_I8:
1286 return emitAtomicBinaryPartword(MI, BB, 1);
1287 case Mips::ATOMIC_LOAD_ADD_I16:
1288 return emitAtomicBinaryPartword(MI, BB, 2);
1289 case Mips::ATOMIC_LOAD_ADD_I32:
1290 return emitAtomicBinary(MI, BB);
1291 case Mips::ATOMIC_LOAD_ADD_I64:
1292 return emitAtomicBinary(MI, BB);
1293
1294 case Mips::ATOMIC_LOAD_AND_I8:
1295 return emitAtomicBinaryPartword(MI, BB, 1);
1296 case Mips::ATOMIC_LOAD_AND_I16:
1297 return emitAtomicBinaryPartword(MI, BB, 2);
1298 case Mips::ATOMIC_LOAD_AND_I32:
1299 return emitAtomicBinary(MI, BB);
1300 case Mips::ATOMIC_LOAD_AND_I64:
1301 return emitAtomicBinary(MI, BB);
1302
1303 case Mips::ATOMIC_LOAD_OR_I8:
1304 return emitAtomicBinaryPartword(MI, BB, 1);
1305 case Mips::ATOMIC_LOAD_OR_I16:
1306 return emitAtomicBinaryPartword(MI, BB, 2);
1307 case Mips::ATOMIC_LOAD_OR_I32:
1308 return emitAtomicBinary(MI, BB);
1309 case Mips::ATOMIC_LOAD_OR_I64:
1310 return emitAtomicBinary(MI, BB);
1311
1312 case Mips::ATOMIC_LOAD_XOR_I8:
1313 return emitAtomicBinaryPartword(MI, BB, 1);
1314 case Mips::ATOMIC_LOAD_XOR_I16:
1315 return emitAtomicBinaryPartword(MI, BB, 2);
1316 case Mips::ATOMIC_LOAD_XOR_I32:
1317 return emitAtomicBinary(MI, BB);
1318 case Mips::ATOMIC_LOAD_XOR_I64:
1319 return emitAtomicBinary(MI, BB);
1320
1321 case Mips::ATOMIC_LOAD_NAND_I8:
1322 return emitAtomicBinaryPartword(MI, BB, 1);
1323 case Mips::ATOMIC_LOAD_NAND_I16:
1324 return emitAtomicBinaryPartword(MI, BB, 2);
1325 case Mips::ATOMIC_LOAD_NAND_I32:
1326 return emitAtomicBinary(MI, BB);
1327 case Mips::ATOMIC_LOAD_NAND_I64:
1328 return emitAtomicBinary(MI, BB);
1329
1330 case Mips::ATOMIC_LOAD_SUB_I8:
1331 return emitAtomicBinaryPartword(MI, BB, 1);
1332 case Mips::ATOMIC_LOAD_SUB_I16:
1333 return emitAtomicBinaryPartword(MI, BB, 2);
1334 case Mips::ATOMIC_LOAD_SUB_I32:
1335 return emitAtomicBinary(MI, BB);
1336 case Mips::ATOMIC_LOAD_SUB_I64:
1337 return emitAtomicBinary(MI, BB);
1338
1339 case Mips::ATOMIC_SWAP_I8:
1340 return emitAtomicBinaryPartword(MI, BB, 1);
1341 case Mips::ATOMIC_SWAP_I16:
1342 return emitAtomicBinaryPartword(MI, BB, 2);
1343 case Mips::ATOMIC_SWAP_I32:
1344 return emitAtomicBinary(MI, BB);
1345 case Mips::ATOMIC_SWAP_I64:
1346 return emitAtomicBinary(MI, BB);
1347
1348 case Mips::ATOMIC_CMP_SWAP_I8:
1349 return emitAtomicCmpSwapPartword(MI, BB, 1);
1350 case Mips::ATOMIC_CMP_SWAP_I16:
1351 return emitAtomicCmpSwapPartword(MI, BB, 2);
1352 case Mips::ATOMIC_CMP_SWAP_I32:
1353 return emitAtomicCmpSwap(MI, BB);
1354 case Mips::ATOMIC_CMP_SWAP_I64:
1355 return emitAtomicCmpSwap(MI, BB);
1356 case Mips::PseudoSDIV:
1357 case Mips::PseudoUDIV:
1358 case Mips::DIV:
1359 case Mips::DIVU:
1360 case Mips::MOD:
1361 case Mips::MODU:
1362 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false,
1363 false);
1364 case Mips::SDIV_MM_Pseudo:
1365 case Mips::UDIV_MM_Pseudo:
1366 case Mips::SDIV_MM:
1367 case Mips::UDIV_MM:
1368 case Mips::DIV_MMR6:
1369 case Mips::DIVU_MMR6:
1370 case Mips::MOD_MMR6:
1371 case Mips::MODU_MMR6:
1372 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false, true);
1373 case Mips::PseudoDSDIV:
1374 case Mips::PseudoDUDIV:
1375 case Mips::DDIV:
1376 case Mips::DDIVU:
1377 case Mips::DMOD:
1378 case Mips::DMODU:
1379 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true, false);
1380
1381 case Mips::PseudoSELECT_I:
1382 case Mips::PseudoSELECT_I64:
1383 case Mips::PseudoSELECT_S:
1384 case Mips::PseudoSELECT_D32:
1385 case Mips::PseudoSELECT_D64:
1386 return emitPseudoSELECT(MI, BB, false, Mips::BNE);
1387 case Mips::PseudoSELECTFP_F_I:
1388 case Mips::PseudoSELECTFP_F_I64:
1389 case Mips::PseudoSELECTFP_F_S:
1390 case Mips::PseudoSELECTFP_F_D32:
1391 case Mips::PseudoSELECTFP_F_D64:
1392 return emitPseudoSELECT(MI, BB, true, Mips::BC1F);
1393 case Mips::PseudoSELECTFP_T_I:
1394 case Mips::PseudoSELECTFP_T_I64:
1395 case Mips::PseudoSELECTFP_T_S:
1396 case Mips::PseudoSELECTFP_T_D32:
1397 case Mips::PseudoSELECTFP_T_D64:
1398 return emitPseudoSELECT(MI, BB, true, Mips::BC1T);
1399 }
1400 }
1401
1402 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1403 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1404 MachineBasicBlock *
emitAtomicBinary(MachineInstr & MI,MachineBasicBlock * BB) const1405 MipsTargetLowering::emitAtomicBinary(MachineInstr &MI,
1406 MachineBasicBlock *BB) const {
1407
1408 MachineFunction *MF = BB->getParent();
1409 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1410 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1411 DebugLoc DL = MI.getDebugLoc();
1412
1413 unsigned AtomicOp;
1414 switch (MI.getOpcode()) {
1415 case Mips::ATOMIC_LOAD_ADD_I32:
1416 AtomicOp = Mips::ATOMIC_LOAD_ADD_I32_POSTRA;
1417 break;
1418 case Mips::ATOMIC_LOAD_SUB_I32:
1419 AtomicOp = Mips::ATOMIC_LOAD_SUB_I32_POSTRA;
1420 break;
1421 case Mips::ATOMIC_LOAD_AND_I32:
1422 AtomicOp = Mips::ATOMIC_LOAD_AND_I32_POSTRA;
1423 break;
1424 case Mips::ATOMIC_LOAD_OR_I32:
1425 AtomicOp = Mips::ATOMIC_LOAD_OR_I32_POSTRA;
1426 break;
1427 case Mips::ATOMIC_LOAD_XOR_I32:
1428 AtomicOp = Mips::ATOMIC_LOAD_XOR_I32_POSTRA;
1429 break;
1430 case Mips::ATOMIC_LOAD_NAND_I32:
1431 AtomicOp = Mips::ATOMIC_LOAD_NAND_I32_POSTRA;
1432 break;
1433 case Mips::ATOMIC_SWAP_I32:
1434 AtomicOp = Mips::ATOMIC_SWAP_I32_POSTRA;
1435 break;
1436 case Mips::ATOMIC_LOAD_ADD_I64:
1437 AtomicOp = Mips::ATOMIC_LOAD_ADD_I64_POSTRA;
1438 break;
1439 case Mips::ATOMIC_LOAD_SUB_I64:
1440 AtomicOp = Mips::ATOMIC_LOAD_SUB_I64_POSTRA;
1441 break;
1442 case Mips::ATOMIC_LOAD_AND_I64:
1443 AtomicOp = Mips::ATOMIC_LOAD_AND_I64_POSTRA;
1444 break;
1445 case Mips::ATOMIC_LOAD_OR_I64:
1446 AtomicOp = Mips::ATOMIC_LOAD_OR_I64_POSTRA;
1447 break;
1448 case Mips::ATOMIC_LOAD_XOR_I64:
1449 AtomicOp = Mips::ATOMIC_LOAD_XOR_I64_POSTRA;
1450 break;
1451 case Mips::ATOMIC_LOAD_NAND_I64:
1452 AtomicOp = Mips::ATOMIC_LOAD_NAND_I64_POSTRA;
1453 break;
1454 case Mips::ATOMIC_SWAP_I64:
1455 AtomicOp = Mips::ATOMIC_SWAP_I64_POSTRA;
1456 break;
1457 default:
1458 llvm_unreachable("Unknown pseudo atomic for replacement!");
1459 }
1460
1461 unsigned OldVal = MI.getOperand(0).getReg();
1462 unsigned Ptr = MI.getOperand(1).getReg();
1463 unsigned Incr = MI.getOperand(2).getReg();
1464 unsigned Scratch = RegInfo.createVirtualRegister(RegInfo.getRegClass(OldVal));
1465
1466 MachineBasicBlock::iterator II(MI);
1467
1468 // The scratch registers here with the EarlyClobber | Define | Implicit
1469 // flags is used to persuade the register allocator and the machine
1470 // verifier to accept the usage of this register. This has to be a real
1471 // register which has an UNDEF value but is dead after the instruction which
1472 // is unique among the registers chosen for the instruction.
1473
1474 // The EarlyClobber flag has the semantic properties that the operand it is
1475 // attached to is clobbered before the rest of the inputs are read. Hence it
1476 // must be unique among the operands to the instruction.
1477 // The Define flag is needed to coerce the machine verifier that an Undef
1478 // value isn't a problem.
1479 // The Dead flag is needed as the value in scratch isn't used by any other
1480 // instruction. Kill isn't used as Dead is more precise.
1481 // The implicit flag is here due to the interaction between the other flags
1482 // and the machine verifier.
1483
1484 // For correctness purpose, a new pseudo is introduced here. We need this
1485 // new pseudo, so that FastRegisterAllocator does not see an ll/sc sequence
1486 // that is spread over >1 basic blocks. A register allocator which
1487 // introduces (or any codegen infact) a store, can violate the expectations
1488 // of the hardware.
1489 //
1490 // An atomic read-modify-write sequence starts with a linked load
1491 // instruction and ends with a store conditional instruction. The atomic
1492 // read-modify-write sequence fails if any of the following conditions
1493 // occur between the execution of ll and sc:
1494 // * A coherent store is completed by another process or coherent I/O
1495 // module into the block of synchronizable physical memory containing
1496 // the word. The size and alignment of the block is
1497 // implementation-dependent.
1498 // * A coherent store is executed between an LL and SC sequence on the
1499 // same processor to the block of synchornizable physical memory
1500 // containing the word.
1501 //
1502
1503 unsigned PtrCopy = RegInfo.createVirtualRegister(RegInfo.getRegClass(Ptr));
1504 unsigned IncrCopy = RegInfo.createVirtualRegister(RegInfo.getRegClass(Incr));
1505
1506 BuildMI(*BB, II, DL, TII->get(Mips::COPY), IncrCopy).addReg(Incr);
1507 BuildMI(*BB, II, DL, TII->get(Mips::COPY), PtrCopy).addReg(Ptr);
1508
1509 BuildMI(*BB, II, DL, TII->get(AtomicOp))
1510 .addReg(OldVal, RegState::Define | RegState::EarlyClobber)
1511 .addReg(PtrCopy)
1512 .addReg(IncrCopy)
1513 .addReg(Scratch, RegState::Define | RegState::EarlyClobber |
1514 RegState::Implicit | RegState::Dead);
1515
1516 MI.eraseFromParent();
1517
1518 return BB;
1519 }
1520
emitSignExtendToI32InReg(MachineInstr & MI,MachineBasicBlock * BB,unsigned Size,unsigned DstReg,unsigned SrcReg) const1521 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg(
1522 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg,
1523 unsigned SrcReg) const {
1524 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1525 const DebugLoc &DL = MI.getDebugLoc();
1526
1527 if (Subtarget.hasMips32r2() && Size == 1) {
1528 BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg);
1529 return BB;
1530 }
1531
1532 if (Subtarget.hasMips32r2() && Size == 2) {
1533 BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg);
1534 return BB;
1535 }
1536
1537 MachineFunction *MF = BB->getParent();
1538 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1539 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1540 unsigned ScrReg = RegInfo.createVirtualRegister(RC);
1541
1542 assert(Size < 32);
1543 int64_t ShiftImm = 32 - (Size * 8);
1544
1545 BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm);
1546 BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm);
1547
1548 return BB;
1549 }
1550
emitAtomicBinaryPartword(MachineInstr & MI,MachineBasicBlock * BB,unsigned Size) const1551 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword(
1552 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const {
1553 assert((Size == 1 || Size == 2) &&
1554 "Unsupported size for EmitAtomicBinaryPartial.");
1555
1556 MachineFunction *MF = BB->getParent();
1557 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1558 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1559 const bool ArePtrs64bit = ABI.ArePtrs64bit();
1560 const TargetRegisterClass *RCp =
1561 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32);
1562 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1563 DebugLoc DL = MI.getDebugLoc();
1564
1565 unsigned Dest = MI.getOperand(0).getReg();
1566 unsigned Ptr = MI.getOperand(1).getReg();
1567 unsigned Incr = MI.getOperand(2).getReg();
1568
1569 unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp);
1570 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1571 unsigned Mask = RegInfo.createVirtualRegister(RC);
1572 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1573 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1574 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp);
1575 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1576 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1577 unsigned Scratch = RegInfo.createVirtualRegister(RC);
1578 unsigned Scratch2 = RegInfo.createVirtualRegister(RC);
1579 unsigned Scratch3 = RegInfo.createVirtualRegister(RC);
1580
1581 unsigned AtomicOp = 0;
1582 switch (MI.getOpcode()) {
1583 case Mips::ATOMIC_LOAD_NAND_I8:
1584 AtomicOp = Mips::ATOMIC_LOAD_NAND_I8_POSTRA;
1585 break;
1586 case Mips::ATOMIC_LOAD_NAND_I16:
1587 AtomicOp = Mips::ATOMIC_LOAD_NAND_I16_POSTRA;
1588 break;
1589 case Mips::ATOMIC_SWAP_I8:
1590 AtomicOp = Mips::ATOMIC_SWAP_I8_POSTRA;
1591 break;
1592 case Mips::ATOMIC_SWAP_I16:
1593 AtomicOp = Mips::ATOMIC_SWAP_I16_POSTRA;
1594 break;
1595 case Mips::ATOMIC_LOAD_ADD_I8:
1596 AtomicOp = Mips::ATOMIC_LOAD_ADD_I8_POSTRA;
1597 break;
1598 case Mips::ATOMIC_LOAD_ADD_I16:
1599 AtomicOp = Mips::ATOMIC_LOAD_ADD_I16_POSTRA;
1600 break;
1601 case Mips::ATOMIC_LOAD_SUB_I8:
1602 AtomicOp = Mips::ATOMIC_LOAD_SUB_I8_POSTRA;
1603 break;
1604 case Mips::ATOMIC_LOAD_SUB_I16:
1605 AtomicOp = Mips::ATOMIC_LOAD_SUB_I16_POSTRA;
1606 break;
1607 case Mips::ATOMIC_LOAD_AND_I8:
1608 AtomicOp = Mips::ATOMIC_LOAD_AND_I8_POSTRA;
1609 break;
1610 case Mips::ATOMIC_LOAD_AND_I16:
1611 AtomicOp = Mips::ATOMIC_LOAD_AND_I16_POSTRA;
1612 break;
1613 case Mips::ATOMIC_LOAD_OR_I8:
1614 AtomicOp = Mips::ATOMIC_LOAD_OR_I8_POSTRA;
1615 break;
1616 case Mips::ATOMIC_LOAD_OR_I16:
1617 AtomicOp = Mips::ATOMIC_LOAD_OR_I16_POSTRA;
1618 break;
1619 case Mips::ATOMIC_LOAD_XOR_I8:
1620 AtomicOp = Mips::ATOMIC_LOAD_XOR_I8_POSTRA;
1621 break;
1622 case Mips::ATOMIC_LOAD_XOR_I16:
1623 AtomicOp = Mips::ATOMIC_LOAD_XOR_I16_POSTRA;
1624 break;
1625 default:
1626 llvm_unreachable("Unknown subword atomic pseudo for expansion!");
1627 }
1628
1629 // insert new blocks after the current block
1630 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1631 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1632 MachineFunction::iterator It = ++BB->getIterator();
1633 MF->insert(It, exitMBB);
1634
1635 // Transfer the remainder of BB and its successor edges to exitMBB.
1636 exitMBB->splice(exitMBB->begin(), BB,
1637 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1638 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1639
1640 BB->addSuccessor(exitMBB, BranchProbability::getOne());
1641
1642 // thisMBB:
1643 // addiu masklsb2,$0,-4 # 0xfffffffc
1644 // and alignedaddr,ptr,masklsb2
1645 // andi ptrlsb2,ptr,3
1646 // sll shiftamt,ptrlsb2,3
1647 // ori maskupper,$0,255 # 0xff
1648 // sll mask,maskupper,shiftamt
1649 // nor mask2,$0,mask
1650 // sll incr2,incr,shiftamt
1651
1652 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1653 BuildMI(BB, DL, TII->get(ABI.GetPtrAddiuOp()), MaskLSB2)
1654 .addReg(ABI.GetNullPtr()).addImm(-4);
1655 BuildMI(BB, DL, TII->get(ABI.GetPtrAndOp()), AlignedAddr)
1656 .addReg(Ptr).addReg(MaskLSB2);
1657 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2)
1658 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3);
1659 if (Subtarget.isLittle()) {
1660 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1661 } else {
1662 unsigned Off = RegInfo.createVirtualRegister(RC);
1663 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1664 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1665 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1666 }
1667 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1668 .addReg(Mips::ZERO).addImm(MaskImm);
1669 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1670 .addReg(MaskUpper).addReg(ShiftAmt);
1671 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1672 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1673
1674
1675 // The purposes of the flags on the scratch registers is explained in
1676 // emitAtomicBinary. In summary, we need a scratch register which is going to
1677 // be undef, that is unique among registers chosen for the instruction.
1678
1679 BuildMI(BB, DL, TII->get(AtomicOp))
1680 .addReg(Dest, RegState::Define | RegState::EarlyClobber)
1681 .addReg(AlignedAddr)
1682 .addReg(Incr2)
1683 .addReg(Mask)
1684 .addReg(Mask2)
1685 .addReg(ShiftAmt)
1686 .addReg(Scratch, RegState::EarlyClobber | RegState::Define |
1687 RegState::Dead | RegState::Implicit)
1688 .addReg(Scratch2, RegState::EarlyClobber | RegState::Define |
1689 RegState::Dead | RegState::Implicit)
1690 .addReg(Scratch3, RegState::EarlyClobber | RegState::Define |
1691 RegState::Dead | RegState::Implicit);
1692
1693 MI.eraseFromParent(); // The instruction is gone now.
1694
1695 return exitMBB;
1696 }
1697
1698 // Lower atomic compare and swap to a pseudo instruction, taking care to
1699 // define a scratch register for the pseudo instruction's expansion. The
1700 // instruction is expanded after the register allocator as to prevent
1701 // the insertion of stores between the linked load and the store conditional.
1702
1703 MachineBasicBlock *
emitAtomicCmpSwap(MachineInstr & MI,MachineBasicBlock * BB) const1704 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI,
1705 MachineBasicBlock *BB) const {
1706
1707 assert((MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 ||
1708 MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I64) &&
1709 "Unsupported atomic psseudo for EmitAtomicCmpSwap.");
1710
1711 const unsigned Size = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 ? 4 : 8;
1712
1713 MachineFunction *MF = BB->getParent();
1714 MachineRegisterInfo &MRI = MF->getRegInfo();
1715 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1716 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1717 DebugLoc DL = MI.getDebugLoc();
1718
1719 unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32
1720 ? Mips::ATOMIC_CMP_SWAP_I32_POSTRA
1721 : Mips::ATOMIC_CMP_SWAP_I64_POSTRA;
1722 unsigned Dest = MI.getOperand(0).getReg();
1723 unsigned Ptr = MI.getOperand(1).getReg();
1724 unsigned OldVal = MI.getOperand(2).getReg();
1725 unsigned NewVal = MI.getOperand(3).getReg();
1726
1727 unsigned Scratch = MRI.createVirtualRegister(RC);
1728 MachineBasicBlock::iterator II(MI);
1729
1730 // We need to create copies of the various registers and kill them at the
1731 // atomic pseudo. If the copies are not made, when the atomic is expanded
1732 // after fast register allocation, the spills will end up outside of the
1733 // blocks that their values are defined in, causing livein errors.
1734
1735 unsigned DestCopy = MRI.createVirtualRegister(MRI.getRegClass(Dest));
1736 unsigned PtrCopy = MRI.createVirtualRegister(MRI.getRegClass(Ptr));
1737 unsigned OldValCopy = MRI.createVirtualRegister(MRI.getRegClass(OldVal));
1738 unsigned NewValCopy = MRI.createVirtualRegister(MRI.getRegClass(NewVal));
1739
1740 BuildMI(*BB, II, DL, TII->get(Mips::COPY), DestCopy).addReg(Dest);
1741 BuildMI(*BB, II, DL, TII->get(Mips::COPY), PtrCopy).addReg(Ptr);
1742 BuildMI(*BB, II, DL, TII->get(Mips::COPY), OldValCopy).addReg(OldVal);
1743 BuildMI(*BB, II, DL, TII->get(Mips::COPY), NewValCopy).addReg(NewVal);
1744
1745 // The purposes of the flags on the scratch registers is explained in
1746 // emitAtomicBinary. In summary, we need a scratch register which is going to
1747 // be undef, that is unique among registers chosen for the instruction.
1748
1749 BuildMI(*BB, II, DL, TII->get(AtomicOp))
1750 .addReg(Dest, RegState::Define | RegState::EarlyClobber)
1751 .addReg(PtrCopy, RegState::Kill)
1752 .addReg(OldValCopy, RegState::Kill)
1753 .addReg(NewValCopy, RegState::Kill)
1754 .addReg(Scratch, RegState::EarlyClobber | RegState::Define |
1755 RegState::Dead | RegState::Implicit);
1756
1757 MI.eraseFromParent(); // The instruction is gone now.
1758
1759 return BB;
1760 }
1761
emitAtomicCmpSwapPartword(MachineInstr & MI,MachineBasicBlock * BB,unsigned Size) const1762 MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword(
1763 MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const {
1764 assert((Size == 1 || Size == 2) &&
1765 "Unsupported size for EmitAtomicCmpSwapPartial.");
1766
1767 MachineFunction *MF = BB->getParent();
1768 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1769 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1770 const bool ArePtrs64bit = ABI.ArePtrs64bit();
1771 const TargetRegisterClass *RCp =
1772 getRegClassFor(ArePtrs64bit ? MVT::i64 : MVT::i32);
1773 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1774 DebugLoc DL = MI.getDebugLoc();
1775
1776 unsigned Dest = MI.getOperand(0).getReg();
1777 unsigned Ptr = MI.getOperand(1).getReg();
1778 unsigned CmpVal = MI.getOperand(2).getReg();
1779 unsigned NewVal = MI.getOperand(3).getReg();
1780
1781 unsigned AlignedAddr = RegInfo.createVirtualRegister(RCp);
1782 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1783 unsigned Mask = RegInfo.createVirtualRegister(RC);
1784 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1785 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1786 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1787 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RCp);
1788 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1789 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1790 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1791 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1792 unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I8
1793 ? Mips::ATOMIC_CMP_SWAP_I8_POSTRA
1794 : Mips::ATOMIC_CMP_SWAP_I16_POSTRA;
1795
1796 // The scratch registers here with the EarlyClobber | Define | Dead | Implicit
1797 // flags are used to coerce the register allocator and the machine verifier to
1798 // accept the usage of these registers.
1799 // The EarlyClobber flag has the semantic properties that the operand it is
1800 // attached to is clobbered before the rest of the inputs are read. Hence it
1801 // must be unique among the operands to the instruction.
1802 // The Define flag is needed to coerce the machine verifier that an Undef
1803 // value isn't a problem.
1804 // The Dead flag is needed as the value in scratch isn't used by any other
1805 // instruction. Kill isn't used as Dead is more precise.
1806 unsigned Scratch = RegInfo.createVirtualRegister(RC);
1807 unsigned Scratch2 = RegInfo.createVirtualRegister(RC);
1808
1809 // insert new blocks after the current block
1810 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1811 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1812 MachineFunction::iterator It = ++BB->getIterator();
1813 MF->insert(It, exitMBB);
1814
1815 // Transfer the remainder of BB and its successor edges to exitMBB.
1816 exitMBB->splice(exitMBB->begin(), BB,
1817 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1818 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1819
1820 BB->addSuccessor(exitMBB, BranchProbability::getOne());
1821
1822 // thisMBB:
1823 // addiu masklsb2,$0,-4 # 0xfffffffc
1824 // and alignedaddr,ptr,masklsb2
1825 // andi ptrlsb2,ptr,3
1826 // xori ptrlsb2,ptrlsb2,3 # Only for BE
1827 // sll shiftamt,ptrlsb2,3
1828 // ori maskupper,$0,255 # 0xff
1829 // sll mask,maskupper,shiftamt
1830 // nor mask2,$0,mask
1831 // andi maskedcmpval,cmpval,255
1832 // sll shiftedcmpval,maskedcmpval,shiftamt
1833 // andi maskednewval,newval,255
1834 // sll shiftednewval,maskednewval,shiftamt
1835 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1836 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), MaskLSB2)
1837 .addReg(ABI.GetNullPtr()).addImm(-4);
1838 BuildMI(BB, DL, TII->get(ArePtrs64bit ? Mips::AND64 : Mips::AND), AlignedAddr)
1839 .addReg(Ptr).addReg(MaskLSB2);
1840 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2)
1841 .addReg(Ptr, 0, ArePtrs64bit ? Mips::sub_32 : 0).addImm(3);
1842 if (Subtarget.isLittle()) {
1843 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1844 } else {
1845 unsigned Off = RegInfo.createVirtualRegister(RC);
1846 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1847 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1848 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1849 }
1850 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1851 .addReg(Mips::ZERO).addImm(MaskImm);
1852 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1853 .addReg(MaskUpper).addReg(ShiftAmt);
1854 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1855 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1856 .addReg(CmpVal).addImm(MaskImm);
1857 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1858 .addReg(MaskedCmpVal).addReg(ShiftAmt);
1859 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1860 .addReg(NewVal).addImm(MaskImm);
1861 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1862 .addReg(MaskedNewVal).addReg(ShiftAmt);
1863
1864 // The purposes of the flags on the scratch registers are explained in
1865 // emitAtomicBinary. In summary, we need a scratch register which is going to
1866 // be undef, that is unique among the register chosen for the instruction.
1867
1868 BuildMI(BB, DL, TII->get(AtomicOp))
1869 .addReg(Dest, RegState::Define | RegState::EarlyClobber)
1870 .addReg(AlignedAddr)
1871 .addReg(Mask)
1872 .addReg(ShiftedCmpVal)
1873 .addReg(Mask2)
1874 .addReg(ShiftedNewVal)
1875 .addReg(ShiftAmt)
1876 .addReg(Scratch, RegState::EarlyClobber | RegState::Define |
1877 RegState::Dead | RegState::Implicit)
1878 .addReg(Scratch2, RegState::EarlyClobber | RegState::Define |
1879 RegState::Dead | RegState::Implicit);
1880
1881 MI.eraseFromParent(); // The instruction is gone now.
1882
1883 return exitMBB;
1884 }
1885
lowerBRCOND(SDValue Op,SelectionDAG & DAG) const1886 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1887 // The first operand is the chain, the second is the condition, the third is
1888 // the block to branch to if the condition is true.
1889 SDValue Chain = Op.getOperand(0);
1890 SDValue Dest = Op.getOperand(2);
1891 SDLoc DL(Op);
1892
1893 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1894 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1895
1896 // Return if flag is not set by a floating point comparison.
1897 if (CondRes.getOpcode() != MipsISD::FPCmp)
1898 return Op;
1899
1900 SDValue CCNode = CondRes.getOperand(2);
1901 Mips::CondCode CC =
1902 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1903 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1904 SDValue BrCode = DAG.getConstant(Opc, DL, MVT::i32);
1905 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
1906 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1907 FCC0, Dest, CondRes);
1908 }
1909
1910 SDValue MipsTargetLowering::
lowerSELECT(SDValue Op,SelectionDAG & DAG) const1911 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1912 {
1913 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1914 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1915
1916 // Return if flag is not set by a floating point comparison.
1917 if (Cond.getOpcode() != MipsISD::FPCmp)
1918 return Op;
1919
1920 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1921 SDLoc(Op));
1922 }
1923
lowerSETCC(SDValue Op,SelectionDAG & DAG) const1924 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1925 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1926 SDValue Cond = createFPCmp(DAG, Op);
1927
1928 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1929 "Floating point operand expected.");
1930
1931 SDLoc DL(Op);
1932 SDValue True = DAG.getConstant(1, DL, MVT::i32);
1933 SDValue False = DAG.getConstant(0, DL, MVT::i32);
1934
1935 return createCMovFP(DAG, Cond, True, False, DL);
1936 }
1937
lowerGlobalAddress(SDValue Op,SelectionDAG & DAG) const1938 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1939 SelectionDAG &DAG) const {
1940 EVT Ty = Op.getValueType();
1941 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
1942 const GlobalValue *GV = N->getGlobal();
1943
1944 if (!isPositionIndependent()) {
1945 const MipsTargetObjectFile *TLOF =
1946 static_cast<const MipsTargetObjectFile *>(
1947 getTargetMachine().getObjFileLowering());
1948 const GlobalObject *GO = GV->getBaseObject();
1949 if (GO && TLOF->IsGlobalInSmallSection(GO, getTargetMachine()))
1950 // %gp_rel relocation
1951 return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64());
1952
1953 // %hi/%lo relocation
1954 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
1955 // %highest/%higher/%hi/%lo relocation
1956 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
1957 }
1958
1959 // Every other architecture would use shouldAssumeDSOLocal in here, but
1960 // mips is special.
1961 // * In PIC code mips requires got loads even for local statics!
1962 // * To save on got entries, for local statics the got entry contains the
1963 // page and an additional add instruction takes care of the low bits.
1964 // * It is legal to access a hidden symbol with a non hidden undefined,
1965 // so one cannot guarantee that all access to a hidden symbol will know
1966 // it is hidden.
1967 // * Mips linkers don't support creating a page and a full got entry for
1968 // the same symbol.
1969 // * Given all that, we have to use a full got entry for hidden symbols :-(
1970 if (GV->hasLocalLinkage())
1971 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1972
1973 if (LargeGOT)
1974 return getAddrGlobalLargeGOT(
1975 N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, MipsII::MO_GOT_LO16,
1976 DAG.getEntryNode(),
1977 MachinePointerInfo::getGOT(DAG.getMachineFunction()));
1978
1979 return getAddrGlobal(
1980 N, SDLoc(N), Ty, DAG,
1981 (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT,
1982 DAG.getEntryNode(), MachinePointerInfo::getGOT(DAG.getMachineFunction()));
1983 }
1984
lowerBlockAddress(SDValue Op,SelectionDAG & DAG) const1985 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1986 SelectionDAG &DAG) const {
1987 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
1988 EVT Ty = Op.getValueType();
1989
1990 if (!isPositionIndependent())
1991 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
1992 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
1993
1994 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1995 }
1996
1997 SDValue MipsTargetLowering::
lowerGlobalTLSAddress(SDValue Op,SelectionDAG & DAG) const1998 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1999 {
2000 // If the relocation model is PIC, use the General Dynamic TLS Model or
2001 // Local Dynamic TLS model, otherwise use the Initial Exec or
2002 // Local Exec TLS Model.
2003
2004 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2005 if (DAG.getTarget().useEmulatedTLS())
2006 return LowerToTLSEmulatedModel(GA, DAG);
2007
2008 SDLoc DL(GA);
2009 const GlobalValue *GV = GA->getGlobal();
2010 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2011
2012 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
2013
2014 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
2015 // General Dynamic and Local Dynamic TLS Model.
2016 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
2017 : MipsII::MO_TLSGD;
2018
2019 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
2020 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
2021 getGlobalReg(DAG, PtrVT), TGA);
2022 unsigned PtrSize = PtrVT.getSizeInBits();
2023 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
2024
2025 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
2026
2027 ArgListTy Args;
2028 ArgListEntry Entry;
2029 Entry.Node = Argument;
2030 Entry.Ty = PtrTy;
2031 Args.push_back(Entry);
2032
2033 TargetLowering::CallLoweringInfo CLI(DAG);
2034 CLI.setDebugLoc(DL)
2035 .setChain(DAG.getEntryNode())
2036 .setLibCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args));
2037 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2038
2039 SDValue Ret = CallResult.first;
2040
2041 if (model != TLSModel::LocalDynamic)
2042 return Ret;
2043
2044 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2045 MipsII::MO_DTPREL_HI);
2046 SDValue Hi = DAG.getNode(MipsISD::TlsHi, DL, PtrVT, TGAHi);
2047 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2048 MipsII::MO_DTPREL_LO);
2049 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
2050 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
2051 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
2052 }
2053
2054 SDValue Offset;
2055 if (model == TLSModel::InitialExec) {
2056 // Initial Exec TLS Model
2057 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2058 MipsII::MO_GOTTPREL);
2059 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
2060 TGA);
2061 Offset =
2062 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), TGA, MachinePointerInfo());
2063 } else {
2064 // Local Exec TLS Model
2065 assert(model == TLSModel::LocalExec);
2066 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2067 MipsII::MO_TPREL_HI);
2068 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2069 MipsII::MO_TPREL_LO);
2070 SDValue Hi = DAG.getNode(MipsISD::TlsHi, DL, PtrVT, TGAHi);
2071 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
2072 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
2073 }
2074
2075 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
2076 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
2077 }
2078
2079 SDValue MipsTargetLowering::
lowerJumpTable(SDValue Op,SelectionDAG & DAG) const2080 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
2081 {
2082 JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
2083 EVT Ty = Op.getValueType();
2084
2085 if (!isPositionIndependent())
2086 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
2087 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
2088
2089 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
2090 }
2091
2092 SDValue MipsTargetLowering::
lowerConstantPool(SDValue Op,SelectionDAG & DAG) const2093 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
2094 {
2095 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
2096 EVT Ty = Op.getValueType();
2097
2098 if (!isPositionIndependent()) {
2099 const MipsTargetObjectFile *TLOF =
2100 static_cast<const MipsTargetObjectFile *>(
2101 getTargetMachine().getObjFileLowering());
2102
2103 if (TLOF->IsConstantInSmallSection(DAG.getDataLayout(), N->getConstVal(),
2104 getTargetMachine()))
2105 // %gp_rel relocation
2106 return getAddrGPRel(N, SDLoc(N), Ty, DAG, ABI.IsN64());
2107
2108 return Subtarget.hasSym32() ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
2109 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
2110 }
2111
2112 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
2113 }
2114
lowerVASTART(SDValue Op,SelectionDAG & DAG) const2115 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
2116 MachineFunction &MF = DAG.getMachineFunction();
2117 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2118
2119 SDLoc DL(Op);
2120 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
2121 getPointerTy(MF.getDataLayout()));
2122
2123 // vastart just stores the address of the VarArgsFrameIndex slot into the
2124 // memory location argument.
2125 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2126 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
2127 MachinePointerInfo(SV));
2128 }
2129
lowerVAARG(SDValue Op,SelectionDAG & DAG) const2130 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const {
2131 SDNode *Node = Op.getNode();
2132 EVT VT = Node->getValueType(0);
2133 SDValue Chain = Node->getOperand(0);
2134 SDValue VAListPtr = Node->getOperand(1);
2135 unsigned Align = Node->getConstantOperandVal(3);
2136 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2137 SDLoc DL(Node);
2138 unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4;
2139
2140 SDValue VAListLoad = DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL, Chain,
2141 VAListPtr, MachinePointerInfo(SV));
2142 SDValue VAList = VAListLoad;
2143
2144 // Re-align the pointer if necessary.
2145 // It should only ever be necessary for 64-bit types on O32 since the minimum
2146 // argument alignment is the same as the maximum type alignment for N32/N64.
2147 //
2148 // FIXME: We currently align too often. The code generator doesn't notice
2149 // when the pointer is still aligned from the last va_arg (or pair of
2150 // va_args for the i64 on O32 case).
2151 if (Align > getMinStackArgumentAlignment()) {
2152 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
2153
2154 VAList = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
2155 DAG.getConstant(Align - 1, DL, VAList.getValueType()));
2156
2157 VAList = DAG.getNode(ISD::AND, DL, VAList.getValueType(), VAList,
2158 DAG.getConstant(-(int64_t)Align, DL,
2159 VAList.getValueType()));
2160 }
2161
2162 // Increment the pointer, VAList, to the next vaarg.
2163 auto &TD = DAG.getDataLayout();
2164 unsigned ArgSizeInBytes =
2165 TD.getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext()));
2166 SDValue Tmp3 =
2167 DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
2168 DAG.getConstant(alignTo(ArgSizeInBytes, ArgSlotSizeInBytes),
2169 DL, VAList.getValueType()));
2170 // Store the incremented VAList to the legalized pointer
2171 Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr,
2172 MachinePointerInfo(SV));
2173
2174 // In big-endian mode we must adjust the pointer when the load size is smaller
2175 // than the argument slot size. We must also reduce the known alignment to
2176 // match. For example in the N64 ABI, we must add 4 bytes to the offset to get
2177 // the correct half of the slot, and reduce the alignment from 8 (slot
2178 // alignment) down to 4 (type alignment).
2179 if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) {
2180 unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes;
2181 VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList,
2182 DAG.getIntPtrConstant(Adjustment, DL));
2183 }
2184 // Load the actual argument out of the pointer VAList
2185 return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo());
2186 }
2187
lowerFCOPYSIGN32(SDValue Op,SelectionDAG & DAG,bool HasExtractInsert)2188 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG,
2189 bool HasExtractInsert) {
2190 EVT TyX = Op.getOperand(0).getValueType();
2191 EVT TyY = Op.getOperand(1).getValueType();
2192 SDLoc DL(Op);
2193 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
2194 SDValue Const31 = DAG.getConstant(31, DL, MVT::i32);
2195 SDValue Res;
2196
2197 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2198 // to i32.
2199 SDValue X = (TyX == MVT::f32) ?
2200 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
2201 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2202 Const1);
2203 SDValue Y = (TyY == MVT::f32) ?
2204 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
2205 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
2206 Const1);
2207
2208 if (HasExtractInsert) {
2209 // ext E, Y, 31, 1 ; extract bit31 of Y
2210 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
2211 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
2212 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
2213 } else {
2214 // sll SllX, X, 1
2215 // srl SrlX, SllX, 1
2216 // srl SrlY, Y, 31
2217 // sll SllY, SrlX, 31
2218 // or Or, SrlX, SllY
2219 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2220 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2221 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
2222 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
2223 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
2224 }
2225
2226 if (TyX == MVT::f32)
2227 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
2228
2229 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2230 Op.getOperand(0),
2231 DAG.getConstant(0, DL, MVT::i32));
2232 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2233 }
2234
lowerFCOPYSIGN64(SDValue Op,SelectionDAG & DAG,bool HasExtractInsert)2235 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG,
2236 bool HasExtractInsert) {
2237 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
2238 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
2239 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
2240 SDLoc DL(Op);
2241 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
2242
2243 // Bitcast to integer nodes.
2244 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
2245 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
2246
2247 if (HasExtractInsert) {
2248 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
2249 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
2250 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
2251 DAG.getConstant(WidthY - 1, DL, MVT::i32), Const1);
2252
2253 if (WidthX > WidthY)
2254 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
2255 else if (WidthY > WidthX)
2256 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
2257
2258 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
2259 DAG.getConstant(WidthX - 1, DL, MVT::i32), Const1,
2260 X);
2261 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
2262 }
2263
2264 // (d)sll SllX, X, 1
2265 // (d)srl SrlX, SllX, 1
2266 // (d)srl SrlY, Y, width(Y)-1
2267 // (d)sll SllY, SrlX, width(Y)-1
2268 // or Or, SrlX, SllY
2269 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
2270 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
2271 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
2272 DAG.getConstant(WidthY - 1, DL, MVT::i32));
2273
2274 if (WidthX > WidthY)
2275 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
2276 else if (WidthY > WidthX)
2277 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
2278
2279 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
2280 DAG.getConstant(WidthX - 1, DL, MVT::i32));
2281 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
2282 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
2283 }
2284
2285 SDValue
lowerFCOPYSIGN(SDValue Op,SelectionDAG & DAG) const2286 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2287 if (Subtarget.isGP64bit())
2288 return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert());
2289
2290 return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert());
2291 }
2292
2293 SDValue MipsTargetLowering::
lowerFRAMEADDR(SDValue Op,SelectionDAG & DAG) const2294 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2295 // check the depth
2296 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2297 "Frame address can only be determined for current frame.");
2298
2299 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2300 MFI.setFrameAddressIsTaken(true);
2301 EVT VT = Op.getValueType();
2302 SDLoc DL(Op);
2303 SDValue FrameAddr = DAG.getCopyFromReg(
2304 DAG.getEntryNode(), DL, ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT);
2305 return FrameAddr;
2306 }
2307
lowerRETURNADDR(SDValue Op,SelectionDAG & DAG) const2308 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
2309 SelectionDAG &DAG) const {
2310 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2311 return SDValue();
2312
2313 // check the depth
2314 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2315 "Return address can be determined only for current frame.");
2316
2317 MachineFunction &MF = DAG.getMachineFunction();
2318 MachineFrameInfo &MFI = MF.getFrameInfo();
2319 MVT VT = Op.getSimpleValueType();
2320 unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA;
2321 MFI.setReturnAddressIsTaken(true);
2322
2323 // Return RA, which contains the return address. Mark it an implicit live-in.
2324 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2325 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
2326 }
2327
2328 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
2329 // generated from __builtin_eh_return (offset, handler)
2330 // The effect of this is to adjust the stack pointer by "offset"
2331 // and then branch to "handler".
lowerEH_RETURN(SDValue Op,SelectionDAG & DAG) const2332 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
2333 const {
2334 MachineFunction &MF = DAG.getMachineFunction();
2335 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2336
2337 MipsFI->setCallsEhReturn();
2338 SDValue Chain = Op.getOperand(0);
2339 SDValue Offset = Op.getOperand(1);
2340 SDValue Handler = Op.getOperand(2);
2341 SDLoc DL(Op);
2342 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
2343
2344 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2345 // EH_RETURN nodes, so that instructions are emitted back-to-back.
2346 unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1;
2347 unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
2348 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2349 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2350 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2351 DAG.getRegister(OffsetReg, Ty),
2352 DAG.getRegister(AddrReg, getPointerTy(MF.getDataLayout())),
2353 Chain.getValue(1));
2354 }
2355
lowerATOMIC_FENCE(SDValue Op,SelectionDAG & DAG) const2356 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
2357 SelectionDAG &DAG) const {
2358 // FIXME: Need pseudo-fence for 'singlethread' fences
2359 // FIXME: Set SType for weaker fences where supported/appropriate.
2360 unsigned SType = 0;
2361 SDLoc DL(Op);
2362 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2363 DAG.getConstant(SType, DL, MVT::i32));
2364 }
2365
lowerShiftLeftParts(SDValue Op,SelectionDAG & DAG) const2366 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
2367 SelectionDAG &DAG) const {
2368 SDLoc DL(Op);
2369 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2370
2371 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2372 SDValue Shamt = Op.getOperand(2);
2373 // if shamt < (VT.bits):
2374 // lo = (shl lo, shamt)
2375 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2376 // else:
2377 // lo = 0
2378 // hi = (shl lo, shamt[4:0])
2379 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2380 DAG.getConstant(-1, DL, MVT::i32));
2381 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo,
2382 DAG.getConstant(1, DL, VT));
2383 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not);
2384 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
2385 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2386 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
2387 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2388 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2389 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2390 DAG.getConstant(0, DL, VT), ShiftLeftLo);
2391 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or);
2392
2393 SDValue Ops[2] = {Lo, Hi};
2394 return DAG.getMergeValues(Ops, DL);
2395 }
2396
lowerShiftRightParts(SDValue Op,SelectionDAG & DAG,bool IsSRA) const2397 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2398 bool IsSRA) const {
2399 SDLoc DL(Op);
2400 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2401 SDValue Shamt = Op.getOperand(2);
2402 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2403
2404 // if shamt < (VT.bits):
2405 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2406 // if isSRA:
2407 // hi = (sra hi, shamt)
2408 // else:
2409 // hi = (srl hi, shamt)
2410 // else:
2411 // if isSRA:
2412 // lo = (sra hi, shamt[4:0])
2413 // hi = (sra hi, 31)
2414 // else:
2415 // lo = (srl hi, shamt[4:0])
2416 // hi = 0
2417 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2418 DAG.getConstant(-1, DL, MVT::i32));
2419 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi,
2420 DAG.getConstant(1, DL, VT));
2421 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not);
2422 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
2423 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2424 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL,
2425 DL, VT, Hi, Shamt);
2426 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2427 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2428 SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi,
2429 DAG.getConstant(VT.getSizeInBits() - 1, DL, VT));
2430 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or);
2431 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2432 IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi);
2433
2434 SDValue Ops[2] = {Lo, Hi};
2435 return DAG.getMergeValues(Ops, DL);
2436 }
2437
createLoadLR(unsigned Opc,SelectionDAG & DAG,LoadSDNode * LD,SDValue Chain,SDValue Src,unsigned Offset)2438 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2439 SDValue Chain, SDValue Src, unsigned Offset) {
2440 SDValue Ptr = LD->getBasePtr();
2441 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2442 EVT BasePtrVT = Ptr.getValueType();
2443 SDLoc DL(LD);
2444 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2445
2446 if (Offset)
2447 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2448 DAG.getConstant(Offset, DL, BasePtrVT));
2449
2450 SDValue Ops[] = { Chain, Ptr, Src };
2451 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2452 LD->getMemOperand());
2453 }
2454
2455 // Expand an unaligned 32 or 64-bit integer load node.
lowerLOAD(SDValue Op,SelectionDAG & DAG) const2456 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2457 LoadSDNode *LD = cast<LoadSDNode>(Op);
2458 EVT MemVT = LD->getMemoryVT();
2459
2460 if (Subtarget.systemSupportsUnalignedAccess())
2461 return Op;
2462
2463 // Return if load is aligned or if MemVT is neither i32 nor i64.
2464 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2465 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2466 return SDValue();
2467
2468 bool IsLittle = Subtarget.isLittle();
2469 EVT VT = Op.getValueType();
2470 ISD::LoadExtType ExtType = LD->getExtensionType();
2471 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2472
2473 assert((VT == MVT::i32) || (VT == MVT::i64));
2474
2475 // Expand
2476 // (set dst, (i64 (load baseptr)))
2477 // to
2478 // (set tmp, (ldl (add baseptr, 7), undef))
2479 // (set dst, (ldr baseptr, tmp))
2480 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2481 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2482 IsLittle ? 7 : 0);
2483 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2484 IsLittle ? 0 : 7);
2485 }
2486
2487 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2488 IsLittle ? 3 : 0);
2489 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2490 IsLittle ? 0 : 3);
2491
2492 // Expand
2493 // (set dst, (i32 (load baseptr))) or
2494 // (set dst, (i64 (sextload baseptr))) or
2495 // (set dst, (i64 (extload baseptr)))
2496 // to
2497 // (set tmp, (lwl (add baseptr, 3), undef))
2498 // (set dst, (lwr baseptr, tmp))
2499 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2500 (ExtType == ISD::EXTLOAD))
2501 return LWR;
2502
2503 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2504
2505 // Expand
2506 // (set dst, (i64 (zextload baseptr)))
2507 // to
2508 // (set tmp0, (lwl (add baseptr, 3), undef))
2509 // (set tmp1, (lwr baseptr, tmp0))
2510 // (set tmp2, (shl tmp1, 32))
2511 // (set dst, (srl tmp2, 32))
2512 SDLoc DL(LD);
2513 SDValue Const32 = DAG.getConstant(32, DL, MVT::i32);
2514 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2515 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2516 SDValue Ops[] = { SRL, LWR.getValue(1) };
2517 return DAG.getMergeValues(Ops, DL);
2518 }
2519
createStoreLR(unsigned Opc,SelectionDAG & DAG,StoreSDNode * SD,SDValue Chain,unsigned Offset)2520 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2521 SDValue Chain, unsigned Offset) {
2522 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2523 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2524 SDLoc DL(SD);
2525 SDVTList VTList = DAG.getVTList(MVT::Other);
2526
2527 if (Offset)
2528 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2529 DAG.getConstant(Offset, DL, BasePtrVT));
2530
2531 SDValue Ops[] = { Chain, Value, Ptr };
2532 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2533 SD->getMemOperand());
2534 }
2535
2536 // Expand an unaligned 32 or 64-bit integer store node.
lowerUnalignedIntStore(StoreSDNode * SD,SelectionDAG & DAG,bool IsLittle)2537 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2538 bool IsLittle) {
2539 SDValue Value = SD->getValue(), Chain = SD->getChain();
2540 EVT VT = Value.getValueType();
2541
2542 // Expand
2543 // (store val, baseptr) or
2544 // (truncstore val, baseptr)
2545 // to
2546 // (swl val, (add baseptr, 3))
2547 // (swr val, baseptr)
2548 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2549 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2550 IsLittle ? 3 : 0);
2551 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2552 }
2553
2554 assert(VT == MVT::i64);
2555
2556 // Expand
2557 // (store val, baseptr)
2558 // to
2559 // (sdl val, (add baseptr, 7))
2560 // (sdr val, baseptr)
2561 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2562 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2563 }
2564
2565 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
lowerFP_TO_SINT_STORE(StoreSDNode * SD,SelectionDAG & DAG)2566 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) {
2567 SDValue Val = SD->getValue();
2568
2569 if (Val.getOpcode() != ISD::FP_TO_SINT)
2570 return SDValue();
2571
2572 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2573 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2574 Val.getOperand(0));
2575 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2576 SD->getPointerInfo(), SD->getAlignment(),
2577 SD->getMemOperand()->getFlags());
2578 }
2579
lowerSTORE(SDValue Op,SelectionDAG & DAG) const2580 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2581 StoreSDNode *SD = cast<StoreSDNode>(Op);
2582 EVT MemVT = SD->getMemoryVT();
2583
2584 // Lower unaligned integer stores.
2585 if (!Subtarget.systemSupportsUnalignedAccess() &&
2586 (SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2587 ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2588 return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle());
2589
2590 return lowerFP_TO_SINT_STORE(SD, DAG);
2591 }
2592
lowerEH_DWARF_CFA(SDValue Op,SelectionDAG & DAG) const2593 SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op,
2594 SelectionDAG &DAG) const {
2595
2596 // Return a fixed StackObject with offset 0 which points to the old stack
2597 // pointer.
2598 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2599 EVT ValTy = Op->getValueType(0);
2600 int FI = MFI.CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2601 return DAG.getFrameIndex(FI, ValTy);
2602 }
2603
lowerFP_TO_SINT(SDValue Op,SelectionDAG & DAG) const2604 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2605 SelectionDAG &DAG) const {
2606 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2607 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2608 Op.getOperand(0));
2609 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2610 }
2611
2612 //===----------------------------------------------------------------------===//
2613 // Calling Convention Implementation
2614 //===----------------------------------------------------------------------===//
2615
2616 //===----------------------------------------------------------------------===//
2617 // TODO: Implement a generic logic using tblgen that can support this.
2618 // Mips O32 ABI rules:
2619 // ---
2620 // i32 - Passed in A0, A1, A2, A3 and stack
2621 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2622 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2623 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2624 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2625 // not used, it must be shadowed. If only A3 is available, shadow it and
2626 // go to stack.
2627 // vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack.
2628 // vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3}
2629 // with the remainder spilled to the stack.
2630 // vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases
2631 // spilling the remainder to the stack.
2632 //
2633 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2634 //===----------------------------------------------------------------------===//
2635
CC_MipsO32(unsigned ValNo,MVT ValVT,MVT LocVT,CCValAssign::LocInfo LocInfo,ISD::ArgFlagsTy ArgFlags,CCState & State,ArrayRef<MCPhysReg> F64Regs)2636 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2637 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2638 CCState &State, ArrayRef<MCPhysReg> F64Regs) {
2639 const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>(
2640 State.getMachineFunction().getSubtarget());
2641
2642 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 };
2643
2644 const MipsCCState * MipsState = static_cast<MipsCCState *>(&State);
2645
2646 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 };
2647
2648 static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 };
2649
2650 // Do not process byval args here.
2651 if (ArgFlags.isByVal())
2652 return true;
2653
2654 // Promote i8 and i16
2655 if (ArgFlags.isInReg() && !Subtarget.isLittle()) {
2656 if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) {
2657 LocVT = MVT::i32;
2658 if (ArgFlags.isSExt())
2659 LocInfo = CCValAssign::SExtUpper;
2660 else if (ArgFlags.isZExt())
2661 LocInfo = CCValAssign::ZExtUpper;
2662 else
2663 LocInfo = CCValAssign::AExtUpper;
2664 }
2665 }
2666
2667 // Promote i8 and i16
2668 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2669 LocVT = MVT::i32;
2670 if (ArgFlags.isSExt())
2671 LocInfo = CCValAssign::SExt;
2672 else if (ArgFlags.isZExt())
2673 LocInfo = CCValAssign::ZExt;
2674 else
2675 LocInfo = CCValAssign::AExt;
2676 }
2677
2678 unsigned Reg;
2679
2680 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2681 // is true: function is vararg, argument is 3rd or higher, there is previous
2682 // argument which is not f32 or f64.
2683 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 ||
2684 State.getFirstUnallocated(F32Regs) != ValNo;
2685 unsigned OrigAlign = ArgFlags.getOrigAlign();
2686 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2687 bool isVectorFloat = MipsState->WasOriginalArgVectorFloat(ValNo);
2688
2689 // The MIPS vector ABI for floats passes them in a pair of registers
2690 if (ValVT == MVT::i32 && isVectorFloat) {
2691 // This is the start of an vector that was scalarized into an unknown number
2692 // of components. It doesn't matter how many there are. Allocate one of the
2693 // notional 8 byte aligned registers which map onto the argument stack, and
2694 // shadow the register lost to alignment requirements.
2695 if (ArgFlags.isSplit()) {
2696 Reg = State.AllocateReg(FloatVectorIntRegs);
2697 if (Reg == Mips::A2)
2698 State.AllocateReg(Mips::A1);
2699 else if (Reg == 0)
2700 State.AllocateReg(Mips::A3);
2701 } else {
2702 // If we're an intermediate component of the split, we can just attempt to
2703 // allocate a register directly.
2704 Reg = State.AllocateReg(IntRegs);
2705 }
2706 } else if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2707 Reg = State.AllocateReg(IntRegs);
2708 // If this is the first part of an i64 arg,
2709 // the allocated register must be either A0 or A2.
2710 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2711 Reg = State.AllocateReg(IntRegs);
2712 LocVT = MVT::i32;
2713 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2714 // Allocate int register and shadow next int register. If first
2715 // available register is Mips::A1 or Mips::A3, shadow it too.
2716 Reg = State.AllocateReg(IntRegs);
2717 if (Reg == Mips::A1 || Reg == Mips::A3)
2718 Reg = State.AllocateReg(IntRegs);
2719 State.AllocateReg(IntRegs);
2720 LocVT = MVT::i32;
2721 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2722 // we are guaranteed to find an available float register
2723 if (ValVT == MVT::f32) {
2724 Reg = State.AllocateReg(F32Regs);
2725 // Shadow int register
2726 State.AllocateReg(IntRegs);
2727 } else {
2728 Reg = State.AllocateReg(F64Regs);
2729 // Shadow int registers
2730 unsigned Reg2 = State.AllocateReg(IntRegs);
2731 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2732 State.AllocateReg(IntRegs);
2733 State.AllocateReg(IntRegs);
2734 }
2735 } else
2736 llvm_unreachable("Cannot handle this ValVT.");
2737
2738 if (!Reg) {
2739 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(), OrigAlign);
2740 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2741 } else
2742 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2743
2744 return false;
2745 }
2746
CC_MipsO32_FP32(unsigned ValNo,MVT ValVT,MVT LocVT,CCValAssign::LocInfo LocInfo,ISD::ArgFlagsTy ArgFlags,CCState & State)2747 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2748 MVT LocVT, CCValAssign::LocInfo LocInfo,
2749 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2750 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 };
2751
2752 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2753 }
2754
CC_MipsO32_FP64(unsigned ValNo,MVT ValVT,MVT LocVT,CCValAssign::LocInfo LocInfo,ISD::ArgFlagsTy ArgFlags,CCState & State)2755 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2756 MVT LocVT, CCValAssign::LocInfo LocInfo,
2757 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2758 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 };
2759
2760 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2761 }
2762
2763 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2764 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2765 CCState &State) LLVM_ATTRIBUTE_UNUSED;
2766
2767 #include "MipsGenCallingConv.inc"
2768
CCAssignFnForCall() const2769 CCAssignFn *MipsTargetLowering::CCAssignFnForCall() const{
2770 return CC_Mips;
2771 }
2772
CCAssignFnForReturn() const2773 CCAssignFn *MipsTargetLowering::CCAssignFnForReturn() const{
2774 return RetCC_Mips;
2775 }
2776 //===----------------------------------------------------------------------===//
2777 // Call Calling Convention Implementation
2778 //===----------------------------------------------------------------------===//
2779
2780 // Return next O32 integer argument register.
getNextIntArgReg(unsigned Reg)2781 static unsigned getNextIntArgReg(unsigned Reg) {
2782 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2783 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2784 }
2785
passArgOnStack(SDValue StackPtr,unsigned Offset,SDValue Chain,SDValue Arg,const SDLoc & DL,bool IsTailCall,SelectionDAG & DAG) const2786 SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2787 SDValue Chain, SDValue Arg,
2788 const SDLoc &DL, bool IsTailCall,
2789 SelectionDAG &DAG) const {
2790 if (!IsTailCall) {
2791 SDValue PtrOff =
2792 DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr,
2793 DAG.getIntPtrConstant(Offset, DL));
2794 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo());
2795 }
2796
2797 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
2798 int FI = MFI.CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2799 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2800 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2801 /* Alignment = */ 0, MachineMemOperand::MOVolatile);
2802 }
2803
2804 void MipsTargetLowering::
getOpndList(SmallVectorImpl<SDValue> & Ops,std::deque<std::pair<unsigned,SDValue>> & RegsToPass,bool IsPICCall,bool GlobalOrExternal,bool InternalLinkage,bool IsCallReloc,CallLoweringInfo & CLI,SDValue Callee,SDValue Chain) const2805 getOpndList(SmallVectorImpl<SDValue> &Ops,
2806 std::deque<std::pair<unsigned, SDValue>> &RegsToPass,
2807 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2808 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
2809 SDValue Chain) const {
2810 // Insert node "GP copy globalreg" before call to function.
2811 //
2812 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2813 // in PIC mode) allow symbols to be resolved via lazy binding.
2814 // The lazy binding stub requires GP to point to the GOT.
2815 // Note that we don't need GP to point to the GOT for indirect calls
2816 // (when R_MIPS_CALL* is not used for the call) because Mips linker generates
2817 // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs
2818 // used for the function (that is, Mips linker doesn't generate lazy binding
2819 // stub for a function whose address is taken in the program).
2820 if (IsPICCall && !InternalLinkage && IsCallReloc) {
2821 unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP;
2822 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
2823 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2824 }
2825
2826 // Build a sequence of copy-to-reg nodes chained together with token
2827 // chain and flag operands which copy the outgoing args into registers.
2828 // The InFlag in necessary since all emitted instructions must be
2829 // stuck together.
2830 SDValue InFlag;
2831
2832 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2833 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2834 RegsToPass[i].second, InFlag);
2835 InFlag = Chain.getValue(1);
2836 }
2837
2838 // Add argument registers to the end of the list so that they are
2839 // known live into the call.
2840 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2841 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2842 RegsToPass[i].second.getValueType()));
2843
2844 // Add a register mask operand representing the call-preserved registers.
2845 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
2846 const uint32_t *Mask =
2847 TRI->getCallPreservedMask(CLI.DAG.getMachineFunction(), CLI.CallConv);
2848 assert(Mask && "Missing call preserved mask for calling convention");
2849 if (Subtarget.inMips16HardFloat()) {
2850 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2851 StringRef Sym = G->getGlobal()->getName();
2852 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2853 if (F && F->hasFnAttribute("__Mips16RetHelper")) {
2854 Mask = MipsRegisterInfo::getMips16RetHelperMask();
2855 }
2856 }
2857 }
2858 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2859
2860 if (InFlag.getNode())
2861 Ops.push_back(InFlag);
2862 }
2863
2864 /// LowerCall - functions arguments are copied from virtual regs to
2865 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2866 SDValue
LowerCall(TargetLowering::CallLoweringInfo & CLI,SmallVectorImpl<SDValue> & InVals) const2867 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2868 SmallVectorImpl<SDValue> &InVals) const {
2869 SelectionDAG &DAG = CLI.DAG;
2870 SDLoc DL = CLI.DL;
2871 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2872 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2873 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2874 SDValue Chain = CLI.Chain;
2875 SDValue Callee = CLI.Callee;
2876 bool &IsTailCall = CLI.IsTailCall;
2877 CallingConv::ID CallConv = CLI.CallConv;
2878 bool IsVarArg = CLI.IsVarArg;
2879
2880 MachineFunction &MF = DAG.getMachineFunction();
2881 MachineFrameInfo &MFI = MF.getFrameInfo();
2882 const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
2883 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2884 bool IsPIC = isPositionIndependent();
2885
2886 // Analyze operands of the call, assigning locations to each operand.
2887 SmallVector<CCValAssign, 16> ArgLocs;
2888 MipsCCState CCInfo(
2889 CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(),
2890 MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget));
2891
2892 const ExternalSymbolSDNode *ES =
2893 dyn_cast_or_null<const ExternalSymbolSDNode>(Callee.getNode());
2894
2895 // There is one case where CALLSEQ_START..CALLSEQ_END can be nested, which
2896 // is during the lowering of a call with a byval argument which produces
2897 // a call to memcpy. For the O32 case, this causes the caller to allocate
2898 // stack space for the reserved argument area for the callee, then recursively
2899 // again for the memcpy call. In the NEWABI case, this doesn't occur as those
2900 // ABIs mandate that the callee allocates the reserved argument area. We do
2901 // still produce nested CALLSEQ_START..CALLSEQ_END with zero space though.
2902 //
2903 // If the callee has a byval argument and memcpy is used, we are mandated
2904 // to already have produced a reserved argument area for the callee for O32.
2905 // Therefore, the reserved argument area can be reused for both calls.
2906 //
2907 // Other cases of calling memcpy cannot have a chain with a CALLSEQ_START
2908 // present, as we have yet to hook that node onto the chain.
2909 //
2910 // Hence, the CALLSEQ_START and CALLSEQ_END nodes can be eliminated in this
2911 // case. GCC does a similar trick, in that wherever possible, it calculates
2912 // the maximum out going argument area (including the reserved area), and
2913 // preallocates the stack space on entrance to the caller.
2914 //
2915 // FIXME: We should do the same for efficency and space.
2916
2917 // Note: The check on the calling convention below must match
2918 // MipsABIInfo::GetCalleeAllocdArgSizeInBytes().
2919 bool MemcpyInByVal = ES &&
2920 StringRef(ES->getSymbol()) == StringRef("memcpy") &&
2921 CallConv != CallingConv::Fast &&
2922 Chain.getOpcode() == ISD::CALLSEQ_START;
2923
2924 // Allocate the reserved argument area. It seems strange to do this from the
2925 // caller side but removing it breaks the frame size calculation.
2926 unsigned ReservedArgArea =
2927 MemcpyInByVal ? 0 : ABI.GetCalleeAllocdArgSizeInBytes(CallConv);
2928 CCInfo.AllocateStack(ReservedArgArea, 1);
2929
2930 CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(),
2931 ES ? ES->getSymbol() : nullptr);
2932
2933 // Get a count of how many bytes are to be pushed on the stack.
2934 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2935
2936 // Check if it's really possible to do a tail call. Restrict it to functions
2937 // that are part of this compilation unit.
2938 bool InternalLinkage = false;
2939 if (IsTailCall) {
2940 IsTailCall = isEligibleForTailCallOptimization(
2941 CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>());
2942 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2943 InternalLinkage = G->getGlobal()->hasInternalLinkage();
2944 IsTailCall &= (InternalLinkage || G->getGlobal()->hasLocalLinkage() ||
2945 G->getGlobal()->hasPrivateLinkage() ||
2946 G->getGlobal()->hasHiddenVisibility() ||
2947 G->getGlobal()->hasProtectedVisibility());
2948 }
2949 }
2950 if (!IsTailCall && CLI.CS && CLI.CS.isMustTailCall())
2951 report_fatal_error("failed to perform tail call elimination on a call "
2952 "site marked musttail");
2953
2954 if (IsTailCall)
2955 ++NumTailCalls;
2956
2957 // Chain is the output chain of the last Load/Store or CopyToReg node.
2958 // ByValChain is the output chain of the last Memcpy node created for copying
2959 // byval arguments to the stack.
2960 unsigned StackAlignment = TFL->getStackAlignment();
2961 NextStackOffset = alignTo(NextStackOffset, StackAlignment);
2962 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, DL, true);
2963
2964 if (!(IsTailCall || MemcpyInByVal))
2965 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffset, 0, DL);
2966
2967 SDValue StackPtr =
2968 DAG.getCopyFromReg(Chain, DL, ABI.IsN64() ? Mips::SP_64 : Mips::SP,
2969 getPointerTy(DAG.getDataLayout()));
2970
2971 std::deque<std::pair<unsigned, SDValue>> RegsToPass;
2972 SmallVector<SDValue, 8> MemOpChains;
2973
2974 CCInfo.rewindByValRegsInfo();
2975
2976 // Walk the register/memloc assignments, inserting copies/loads.
2977 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2978 SDValue Arg = OutVals[i];
2979 CCValAssign &VA = ArgLocs[i];
2980 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2981 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2982 bool UseUpperBits = false;
2983
2984 // ByVal Arg.
2985 if (Flags.isByVal()) {
2986 unsigned FirstByValReg, LastByValReg;
2987 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
2988 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
2989
2990 assert(Flags.getByValSize() &&
2991 "ByVal args of size 0 should have been ignored by front-end.");
2992 assert(ByValIdx < CCInfo.getInRegsParamsCount());
2993 assert(!IsTailCall &&
2994 "Do not tail-call optimize if there is a byval argument.");
2995 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2996 FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(),
2997 VA);
2998 CCInfo.nextInRegsParam();
2999 continue;
3000 }
3001
3002 // Promote the value if needed.
3003 switch (VA.getLocInfo()) {
3004 default:
3005 llvm_unreachable("Unknown loc info!");
3006 case CCValAssign::Full:
3007 if (VA.isRegLoc()) {
3008 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
3009 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
3010 (ValVT == MVT::i64 && LocVT == MVT::f64))
3011 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
3012 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
3013 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
3014 Arg, DAG.getConstant(0, DL, MVT::i32));
3015 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
3016 Arg, DAG.getConstant(1, DL, MVT::i32));
3017 if (!Subtarget.isLittle())
3018 std::swap(Lo, Hi);
3019 unsigned LocRegLo = VA.getLocReg();
3020 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
3021 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
3022 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
3023 continue;
3024 }
3025 }
3026 break;
3027 case CCValAssign::BCvt:
3028 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
3029 break;
3030 case CCValAssign::SExtUpper:
3031 UseUpperBits = true;
3032 LLVM_FALLTHROUGH;
3033 case CCValAssign::SExt:
3034 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
3035 break;
3036 case CCValAssign::ZExtUpper:
3037 UseUpperBits = true;
3038 LLVM_FALLTHROUGH;
3039 case CCValAssign::ZExt:
3040 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
3041 break;
3042 case CCValAssign::AExtUpper:
3043 UseUpperBits = true;
3044 LLVM_FALLTHROUGH;
3045 case CCValAssign::AExt:
3046 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
3047 break;
3048 }
3049
3050 if (UseUpperBits) {
3051 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
3052 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3053 Arg = DAG.getNode(
3054 ISD::SHL, DL, VA.getLocVT(), Arg,
3055 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3056 }
3057
3058 // Arguments that can be passed on register must be kept at
3059 // RegsToPass vector
3060 if (VA.isRegLoc()) {
3061 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3062 continue;
3063 }
3064
3065 // Register can't get to this point...
3066 assert(VA.isMemLoc());
3067
3068 // emit ISD::STORE whichs stores the
3069 // parameter value to a stack Location
3070 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
3071 Chain, Arg, DL, IsTailCall, DAG));
3072 }
3073
3074 // Transform all store nodes into one single node because all store
3075 // nodes are independent of each other.
3076 if (!MemOpChains.empty())
3077 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
3078
3079 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3080 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3081 // node so that legalize doesn't hack it.
3082
3083 EVT Ty = Callee.getValueType();
3084 bool GlobalOrExternal = false, IsCallReloc = false;
3085
3086 // The long-calls feature is ignored in case of PIC.
3087 // While we do not support -mshared / -mno-shared properly,
3088 // ignore long-calls in case of -mabicalls too.
3089 if (!Subtarget.isABICalls() && !IsPIC) {
3090 // If the function should be called using "long call",
3091 // get its address into a register to prevent using
3092 // of the `jal` instruction for the direct call.
3093 if (auto *N = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3094 if (Subtarget.useLongCalls())
3095 Callee = Subtarget.hasSym32()
3096 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
3097 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
3098 } else if (auto *N = dyn_cast<GlobalAddressSDNode>(Callee)) {
3099 bool UseLongCalls = Subtarget.useLongCalls();
3100 // If the function has long-call/far/near attribute
3101 // it overrides command line switch pased to the backend.
3102 if (auto *F = dyn_cast<Function>(N->getGlobal())) {
3103 if (F->hasFnAttribute("long-call"))
3104 UseLongCalls = true;
3105 else if (F->hasFnAttribute("short-call"))
3106 UseLongCalls = false;
3107 }
3108 if (UseLongCalls)
3109 Callee = Subtarget.hasSym32()
3110 ? getAddrNonPIC(N, SDLoc(N), Ty, DAG)
3111 : getAddrNonPICSym64(N, SDLoc(N), Ty, DAG);
3112 }
3113 }
3114
3115 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3116 if (IsPIC) {
3117 const GlobalValue *Val = G->getGlobal();
3118 InternalLinkage = Val->hasInternalLinkage();
3119
3120 if (InternalLinkage)
3121 Callee = getAddrLocal(G, DL, Ty, DAG, ABI.IsN32() || ABI.IsN64());
3122 else if (LargeGOT) {
3123 Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16,
3124 MipsII::MO_CALL_LO16, Chain,
3125 FuncInfo->callPtrInfo(Val));
3126 IsCallReloc = true;
3127 } else {
3128 Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
3129 FuncInfo->callPtrInfo(Val));
3130 IsCallReloc = true;
3131 }
3132 } else
3133 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL,
3134 getPointerTy(DAG.getDataLayout()), 0,
3135 MipsII::MO_NO_FLAG);
3136 GlobalOrExternal = true;
3137 }
3138 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3139 const char *Sym = S->getSymbol();
3140
3141 if (!IsPIC) // static
3142 Callee = DAG.getTargetExternalSymbol(
3143 Sym, getPointerTy(DAG.getDataLayout()), MipsII::MO_NO_FLAG);
3144 else if (LargeGOT) {
3145 Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16,
3146 MipsII::MO_CALL_LO16, Chain,
3147 FuncInfo->callPtrInfo(Sym));
3148 IsCallReloc = true;
3149 } else { // PIC
3150 Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
3151 FuncInfo->callPtrInfo(Sym));
3152 IsCallReloc = true;
3153 }
3154
3155 GlobalOrExternal = true;
3156 }
3157
3158 SmallVector<SDValue, 8> Ops(1, Chain);
3159 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3160
3161 getOpndList(Ops, RegsToPass, IsPIC, GlobalOrExternal, InternalLinkage,
3162 IsCallReloc, CLI, Callee, Chain);
3163
3164 if (IsTailCall) {
3165 MF.getFrameInfo().setHasTailCall();
3166 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops);
3167 }
3168
3169 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops);
3170 SDValue InFlag = Chain.getValue(1);
3171
3172 // Create the CALLSEQ_END node in the case of where it is not a call to
3173 // memcpy.
3174 if (!(MemcpyInByVal)) {
3175 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
3176 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
3177 InFlag = Chain.getValue(1);
3178 }
3179
3180 // Handle result values, copying them out of physregs into vregs that we
3181 // return.
3182 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
3183 InVals, CLI);
3184 }
3185
3186 /// LowerCallResult - Lower the result values of a call into the
3187 /// appropriate copies out of appropriate physical registers.
LowerCallResult(SDValue Chain,SDValue InFlag,CallingConv::ID CallConv,bool IsVarArg,const SmallVectorImpl<ISD::InputArg> & Ins,const SDLoc & DL,SelectionDAG & DAG,SmallVectorImpl<SDValue> & InVals,TargetLowering::CallLoweringInfo & CLI) const3188 SDValue MipsTargetLowering::LowerCallResult(
3189 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
3190 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
3191 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
3192 TargetLowering::CallLoweringInfo &CLI) const {
3193 // Assign locations to each value returned by this call.
3194 SmallVector<CCValAssign, 16> RVLocs;
3195 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
3196 *DAG.getContext());
3197
3198 const ExternalSymbolSDNode *ES =
3199 dyn_cast_or_null<const ExternalSymbolSDNode>(CLI.Callee.getNode());
3200 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI.RetTy,
3201 ES ? ES->getSymbol() : nullptr);
3202
3203 // Copy all of the result registers out of their specified physreg.
3204 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3205 CCValAssign &VA = RVLocs[i];
3206 assert(VA.isRegLoc() && "Can only return in registers!");
3207
3208 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
3209 RVLocs[i].getLocVT(), InFlag);
3210 Chain = Val.getValue(1);
3211 InFlag = Val.getValue(2);
3212
3213 if (VA.isUpperBitsInLoc()) {
3214 unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits();
3215 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3216 unsigned Shift =
3217 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
3218 Val = DAG.getNode(
3219 Shift, DL, VA.getLocVT(), Val,
3220 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3221 }
3222
3223 switch (VA.getLocInfo()) {
3224 default:
3225 llvm_unreachable("Unknown loc info!");
3226 case CCValAssign::Full:
3227 break;
3228 case CCValAssign::BCvt:
3229 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
3230 break;
3231 case CCValAssign::AExt:
3232 case CCValAssign::AExtUpper:
3233 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
3234 break;
3235 case CCValAssign::ZExt:
3236 case CCValAssign::ZExtUpper:
3237 Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val,
3238 DAG.getValueType(VA.getValVT()));
3239 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
3240 break;
3241 case CCValAssign::SExt:
3242 case CCValAssign::SExtUpper:
3243 Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val,
3244 DAG.getValueType(VA.getValVT()));
3245 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
3246 break;
3247 }
3248
3249 InVals.push_back(Val);
3250 }
3251
3252 return Chain;
3253 }
3254
UnpackFromArgumentSlot(SDValue Val,const CCValAssign & VA,EVT ArgVT,const SDLoc & DL,SelectionDAG & DAG)3255 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA,
3256 EVT ArgVT, const SDLoc &DL,
3257 SelectionDAG &DAG) {
3258 MVT LocVT = VA.getLocVT();
3259 EVT ValVT = VA.getValVT();
3260
3261 // Shift into the upper bits if necessary.
3262 switch (VA.getLocInfo()) {
3263 default:
3264 break;
3265 case CCValAssign::AExtUpper:
3266 case CCValAssign::SExtUpper:
3267 case CCValAssign::ZExtUpper: {
3268 unsigned ValSizeInBits = ArgVT.getSizeInBits();
3269 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3270 unsigned Opcode =
3271 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
3272 Val = DAG.getNode(
3273 Opcode, DL, VA.getLocVT(), Val,
3274 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3275 break;
3276 }
3277 }
3278
3279 // If this is an value smaller than the argument slot size (32-bit for O32,
3280 // 64-bit for N32/N64), it has been promoted in some way to the argument slot
3281 // size. Extract the value and insert any appropriate assertions regarding
3282 // sign/zero extension.
3283 switch (VA.getLocInfo()) {
3284 default:
3285 llvm_unreachable("Unknown loc info!");
3286 case CCValAssign::Full:
3287 break;
3288 case CCValAssign::AExtUpper:
3289 case CCValAssign::AExt:
3290 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
3291 break;
3292 case CCValAssign::SExtUpper:
3293 case CCValAssign::SExt:
3294 Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT));
3295 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
3296 break;
3297 case CCValAssign::ZExtUpper:
3298 case CCValAssign::ZExt:
3299 Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT));
3300 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
3301 break;
3302 case CCValAssign::BCvt:
3303 Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val);
3304 break;
3305 }
3306
3307 return Val;
3308 }
3309
3310 //===----------------------------------------------------------------------===//
3311 // Formal Arguments Calling Convention Implementation
3312 //===----------------------------------------------------------------------===//
3313 /// LowerFormalArguments - transform physical registers into virtual registers
3314 /// and generate load operations for arguments places on the stack.
LowerFormalArguments(SDValue Chain,CallingConv::ID CallConv,bool IsVarArg,const SmallVectorImpl<ISD::InputArg> & Ins,const SDLoc & DL,SelectionDAG & DAG,SmallVectorImpl<SDValue> & InVals) const3315 SDValue MipsTargetLowering::LowerFormalArguments(
3316 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
3317 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
3318 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
3319 MachineFunction &MF = DAG.getMachineFunction();
3320 MachineFrameInfo &MFI = MF.getFrameInfo();
3321 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3322
3323 MipsFI->setVarArgsFrameIndex(0);
3324
3325 // Used with vargs to acumulate store chains.
3326 std::vector<SDValue> OutChains;
3327
3328 // Assign locations to all of the incoming arguments.
3329 SmallVector<CCValAssign, 16> ArgLocs;
3330 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
3331 *DAG.getContext());
3332 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1);
3333 const Function &Func = DAG.getMachineFunction().getFunction();
3334 Function::const_arg_iterator FuncArg = Func.arg_begin();
3335
3336 if (Func.hasFnAttribute("interrupt") && !Func.arg_empty())
3337 report_fatal_error(
3338 "Functions with the interrupt attribute cannot have arguments!");
3339
3340 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg);
3341 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
3342 CCInfo.getInRegsParamsCount() > 0);
3343
3344 unsigned CurArgIdx = 0;
3345 CCInfo.rewindByValRegsInfo();
3346
3347 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3348 CCValAssign &VA = ArgLocs[i];
3349 if (Ins[i].isOrigArg()) {
3350 std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx);
3351 CurArgIdx = Ins[i].getOrigArgIndex();
3352 }
3353 EVT ValVT = VA.getValVT();
3354 ISD::ArgFlagsTy Flags = Ins[i].Flags;
3355 bool IsRegLoc = VA.isRegLoc();
3356
3357 if (Flags.isByVal()) {
3358 assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit");
3359 unsigned FirstByValReg, LastByValReg;
3360 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
3361 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
3362
3363 assert(Flags.getByValSize() &&
3364 "ByVal args of size 0 should have been ignored by front-end.");
3365 assert(ByValIdx < CCInfo.getInRegsParamsCount());
3366 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
3367 FirstByValReg, LastByValReg, VA, CCInfo);
3368 CCInfo.nextInRegsParam();
3369 continue;
3370 }
3371
3372 // Arguments stored on registers
3373 if (IsRegLoc) {
3374 MVT RegVT = VA.getLocVT();
3375 unsigned ArgReg = VA.getLocReg();
3376 const TargetRegisterClass *RC = getRegClassFor(RegVT);
3377
3378 // Transform the arguments stored on
3379 // physical registers into virtual ones
3380 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
3381 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
3382
3383 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3384
3385 // Handle floating point arguments passed in integer registers and
3386 // long double arguments passed in floating point registers.
3387 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
3388 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
3389 (RegVT == MVT::f64 && ValVT == MVT::i64))
3390 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
3391 else if (ABI.IsO32() && RegVT == MVT::i32 &&
3392 ValVT == MVT::f64) {
3393 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
3394 getNextIntArgReg(ArgReg), RC);
3395 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
3396 if (!Subtarget.isLittle())
3397 std::swap(ArgValue, ArgValue2);
3398 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
3399 ArgValue, ArgValue2);
3400 }
3401
3402 InVals.push_back(ArgValue);
3403 } else { // VA.isRegLoc()
3404 MVT LocVT = VA.getLocVT();
3405
3406 if (ABI.IsO32()) {
3407 // We ought to be able to use LocVT directly but O32 sets it to i32
3408 // when allocating floating point values to integer registers.
3409 // This shouldn't influence how we load the value into registers unless
3410 // we are targeting softfloat.
3411 if (VA.getValVT().isFloatingPoint() && !Subtarget.useSoftFloat())
3412 LocVT = VA.getValVT();
3413 }
3414
3415 // sanity check
3416 assert(VA.isMemLoc());
3417
3418 // The stack pointer offset is relative to the caller stack frame.
3419 int FI = MFI.CreateFixedObject(LocVT.getSizeInBits() / 8,
3420 VA.getLocMemOffset(), true);
3421
3422 // Create load nodes to retrieve arguments from the stack
3423 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3424 SDValue ArgValue = DAG.getLoad(
3425 LocVT, DL, Chain, FIN,
3426 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI));
3427 OutChains.push_back(ArgValue.getValue(1));
3428
3429 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3430
3431 InVals.push_back(ArgValue);
3432 }
3433 }
3434
3435 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3436 // The mips ABIs for returning structs by value requires that we copy
3437 // the sret argument into $v0 for the return. Save the argument into
3438 // a virtual register so that we can access it from the return points.
3439 if (Ins[i].Flags.isSRet()) {
3440 unsigned Reg = MipsFI->getSRetReturnReg();
3441 if (!Reg) {
3442 Reg = MF.getRegInfo().createVirtualRegister(
3443 getRegClassFor(ABI.IsN64() ? MVT::i64 : MVT::i32));
3444 MipsFI->setSRetReturnReg(Reg);
3445 }
3446 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]);
3447 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
3448 break;
3449 }
3450 }
3451
3452 if (IsVarArg)
3453 writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo);
3454
3455 // All stores are grouped in one node to allow the matching between
3456 // the size of Ins and InVals. This only happens when on varg functions
3457 if (!OutChains.empty()) {
3458 OutChains.push_back(Chain);
3459 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
3460 }
3461
3462 return Chain;
3463 }
3464
3465 //===----------------------------------------------------------------------===//
3466 // Return Value Calling Convention Implementation
3467 //===----------------------------------------------------------------------===//
3468
3469 bool
CanLowerReturn(CallingConv::ID CallConv,MachineFunction & MF,bool IsVarArg,const SmallVectorImpl<ISD::OutputArg> & Outs,LLVMContext & Context) const3470 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3471 MachineFunction &MF, bool IsVarArg,
3472 const SmallVectorImpl<ISD::OutputArg> &Outs,
3473 LLVMContext &Context) const {
3474 SmallVector<CCValAssign, 16> RVLocs;
3475 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
3476 return CCInfo.CheckReturn(Outs, RetCC_Mips);
3477 }
3478
3479 bool
shouldSignExtendTypeInLibCall(EVT Type,bool IsSigned) const3480 MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const {
3481 if ((ABI.IsN32() || ABI.IsN64()) && Type == MVT::i32)
3482 return true;
3483
3484 return IsSigned;
3485 }
3486
3487 SDValue
LowerInterruptReturn(SmallVectorImpl<SDValue> & RetOps,const SDLoc & DL,SelectionDAG & DAG) const3488 MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
3489 const SDLoc &DL,
3490 SelectionDAG &DAG) const {
3491 MachineFunction &MF = DAG.getMachineFunction();
3492 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3493
3494 MipsFI->setISR();
3495
3496 return DAG.getNode(MipsISD::ERet, DL, MVT::Other, RetOps);
3497 }
3498
3499 SDValue
LowerReturn(SDValue Chain,CallingConv::ID CallConv,bool IsVarArg,const SmallVectorImpl<ISD::OutputArg> & Outs,const SmallVectorImpl<SDValue> & OutVals,const SDLoc & DL,SelectionDAG & DAG) const3500 MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3501 bool IsVarArg,
3502 const SmallVectorImpl<ISD::OutputArg> &Outs,
3503 const SmallVectorImpl<SDValue> &OutVals,
3504 const SDLoc &DL, SelectionDAG &DAG) const {
3505 // CCValAssign - represent the assignment of
3506 // the return value to a location
3507 SmallVector<CCValAssign, 16> RVLocs;
3508 MachineFunction &MF = DAG.getMachineFunction();
3509
3510 // CCState - Info about the registers and stack slot.
3511 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
3512
3513 // Analyze return values.
3514 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
3515
3516 SDValue Flag;
3517 SmallVector<SDValue, 4> RetOps(1, Chain);
3518
3519 // Copy the result values into the output registers.
3520 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3521 SDValue Val = OutVals[i];
3522 CCValAssign &VA = RVLocs[i];
3523 assert(VA.isRegLoc() && "Can only return in registers!");
3524 bool UseUpperBits = false;
3525
3526 switch (VA.getLocInfo()) {
3527 default:
3528 llvm_unreachable("Unknown loc info!");
3529 case CCValAssign::Full:
3530 break;
3531 case CCValAssign::BCvt:
3532 Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val);
3533 break;
3534 case CCValAssign::AExtUpper:
3535 UseUpperBits = true;
3536 LLVM_FALLTHROUGH;
3537 case CCValAssign::AExt:
3538 Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val);
3539 break;
3540 case CCValAssign::ZExtUpper:
3541 UseUpperBits = true;
3542 LLVM_FALLTHROUGH;
3543 case CCValAssign::ZExt:
3544 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val);
3545 break;
3546 case CCValAssign::SExtUpper:
3547 UseUpperBits = true;
3548 LLVM_FALLTHROUGH;
3549 case CCValAssign::SExt:
3550 Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val);
3551 break;
3552 }
3553
3554 if (UseUpperBits) {
3555 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
3556 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3557 Val = DAG.getNode(
3558 ISD::SHL, DL, VA.getLocVT(), Val,
3559 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3560 }
3561
3562 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
3563
3564 // Guarantee that all emitted copies are stuck together with flags.
3565 Flag = Chain.getValue(1);
3566 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3567 }
3568
3569 // The mips ABIs for returning structs by value requires that we copy
3570 // the sret argument into $v0 for the return. We saved the argument into
3571 // a virtual register in the entry block, so now we copy the value out
3572 // and into $v0.
3573 if (MF.getFunction().hasStructRetAttr()) {
3574 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3575 unsigned Reg = MipsFI->getSRetReturnReg();
3576
3577 if (!Reg)
3578 llvm_unreachable("sret virtual register not created in the entry block");
3579 SDValue Val =
3580 DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout()));
3581 unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
3582
3583 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
3584 Flag = Chain.getValue(1);
3585 RetOps.push_back(DAG.getRegister(V0, getPointerTy(DAG.getDataLayout())));
3586 }
3587
3588 RetOps[0] = Chain; // Update chain.
3589
3590 // Add the flag if we have it.
3591 if (Flag.getNode())
3592 RetOps.push_back(Flag);
3593
3594 // ISRs must use "eret".
3595 if (DAG.getMachineFunction().getFunction().hasFnAttribute("interrupt"))
3596 return LowerInterruptReturn(RetOps, DL, DAG);
3597
3598 // Standard return on Mips is a "jr $ra"
3599 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps);
3600 }
3601
3602 //===----------------------------------------------------------------------===//
3603 // Mips Inline Assembly Support
3604 //===----------------------------------------------------------------------===//
3605
3606 /// getConstraintType - Given a constraint letter, return the type of
3607 /// constraint it is for this target.
3608 MipsTargetLowering::ConstraintType
getConstraintType(StringRef Constraint) const3609 MipsTargetLowering::getConstraintType(StringRef Constraint) const {
3610 // Mips specific constraints
3611 // GCC config/mips/constraints.md
3612 //
3613 // 'd' : An address register. Equivalent to r
3614 // unless generating MIPS16 code.
3615 // 'y' : Equivalent to r; retained for
3616 // backwards compatibility.
3617 // 'c' : A register suitable for use in an indirect
3618 // jump. This will always be $25 for -mabicalls.
3619 // 'l' : The lo register. 1 word storage.
3620 // 'x' : The hilo register pair. Double word storage.
3621 if (Constraint.size() == 1) {
3622 switch (Constraint[0]) {
3623 default : break;
3624 case 'd':
3625 case 'y':
3626 case 'f':
3627 case 'c':
3628 case 'l':
3629 case 'x':
3630 return C_RegisterClass;
3631 case 'R':
3632 return C_Memory;
3633 }
3634 }
3635
3636 if (Constraint == "ZC")
3637 return C_Memory;
3638
3639 return TargetLowering::getConstraintType(Constraint);
3640 }
3641
3642 /// Examine constraint type and operand type and determine a weight value.
3643 /// This object must already have been set up with the operand type
3644 /// and the current alternative constraint selected.
3645 TargetLowering::ConstraintWeight
getSingleConstraintMatchWeight(AsmOperandInfo & info,const char * constraint) const3646 MipsTargetLowering::getSingleConstraintMatchWeight(
3647 AsmOperandInfo &info, const char *constraint) const {
3648 ConstraintWeight weight = CW_Invalid;
3649 Value *CallOperandVal = info.CallOperandVal;
3650 // If we don't have a value, we can't do a match,
3651 // but allow it at the lowest weight.
3652 if (!CallOperandVal)
3653 return CW_Default;
3654 Type *type = CallOperandVal->getType();
3655 // Look at the constraint type.
3656 switch (*constraint) {
3657 default:
3658 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3659 break;
3660 case 'd':
3661 case 'y':
3662 if (type->isIntegerTy())
3663 weight = CW_Register;
3664 break;
3665 case 'f': // FPU or MSA register
3666 if (Subtarget.hasMSA() && type->isVectorTy() &&
3667 cast<VectorType>(type)->getBitWidth() == 128)
3668 weight = CW_Register;
3669 else if (type->isFloatTy())
3670 weight = CW_Register;
3671 break;
3672 case 'c': // $25 for indirect jumps
3673 case 'l': // lo register
3674 case 'x': // hilo register pair
3675 if (type->isIntegerTy())
3676 weight = CW_SpecificReg;
3677 break;
3678 case 'I': // signed 16 bit immediate
3679 case 'J': // integer zero
3680 case 'K': // unsigned 16 bit immediate
3681 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3682 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3683 case 'O': // signed 15 bit immediate (+- 16383)
3684 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3685 if (isa<ConstantInt>(CallOperandVal))
3686 weight = CW_Constant;
3687 break;
3688 case 'R':
3689 weight = CW_Memory;
3690 break;
3691 }
3692 return weight;
3693 }
3694
3695 /// This is a helper function to parse a physical register string and split it
3696 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
3697 /// that is returned indicates whether parsing was successful. The second flag
3698 /// is true if the numeric part exists.
parsePhysicalReg(StringRef C,StringRef & Prefix,unsigned long long & Reg)3699 static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix,
3700 unsigned long long &Reg) {
3701 if (C.front() != '{' || C.back() != '}')
3702 return std::make_pair(false, false);
3703
3704 // Search for the first numeric character.
3705 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
3706 I = std::find_if(B, E, isdigit);
3707
3708 Prefix = StringRef(B, I - B);
3709
3710 // The second flag is set to false if no numeric characters were found.
3711 if (I == E)
3712 return std::make_pair(true, false);
3713
3714 // Parse the numeric characters.
3715 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
3716 true);
3717 }
3718
getTypeForExtReturn(LLVMContext & Context,EVT VT,ISD::NodeType) const3719 EVT MipsTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT,
3720 ISD::NodeType) const {
3721 bool Cond = !Subtarget.isABI_O32() && VT.getSizeInBits() == 32;
3722 EVT MinVT = getRegisterType(Context, Cond ? MVT::i64 : MVT::i32);
3723 return VT.bitsLT(MinVT) ? MinVT : VT;
3724 }
3725
3726 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
parseRegForInlineAsmConstraint(StringRef C,MVT VT) const3727 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const {
3728 const TargetRegisterInfo *TRI =
3729 Subtarget.getRegisterInfo();
3730 const TargetRegisterClass *RC;
3731 StringRef Prefix;
3732 unsigned long long Reg;
3733
3734 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
3735
3736 if (!R.first)
3737 return std::make_pair(0U, nullptr);
3738
3739 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
3740 // No numeric characters follow "hi" or "lo".
3741 if (R.second)
3742 return std::make_pair(0U, nullptr);
3743
3744 RC = TRI->getRegClass(Prefix == "hi" ?
3745 Mips::HI32RegClassID : Mips::LO32RegClassID);
3746 return std::make_pair(*(RC->begin()), RC);
3747 } else if (Prefix.startswith("$msa")) {
3748 // Parse $msa(ir|csr|access|save|modify|request|map|unmap)
3749
3750 // No numeric characters follow the name.
3751 if (R.second)
3752 return std::make_pair(0U, nullptr);
3753
3754 Reg = StringSwitch<unsigned long long>(Prefix)
3755 .Case("$msair", Mips::MSAIR)
3756 .Case("$msacsr", Mips::MSACSR)
3757 .Case("$msaaccess", Mips::MSAAccess)
3758 .Case("$msasave", Mips::MSASave)
3759 .Case("$msamodify", Mips::MSAModify)
3760 .Case("$msarequest", Mips::MSARequest)
3761 .Case("$msamap", Mips::MSAMap)
3762 .Case("$msaunmap", Mips::MSAUnmap)
3763 .Default(0);
3764
3765 if (!Reg)
3766 return std::make_pair(0U, nullptr);
3767
3768 RC = TRI->getRegClass(Mips::MSACtrlRegClassID);
3769 return std::make_pair(Reg, RC);
3770 }
3771
3772 if (!R.second)
3773 return std::make_pair(0U, nullptr);
3774
3775 if (Prefix == "$f") { // Parse $f0-$f31.
3776 // If the size of FP registers is 64-bit or Reg is an even number, select
3777 // the 64-bit register class. Otherwise, select the 32-bit register class.
3778 if (VT == MVT::Other)
3779 VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
3780
3781 RC = getRegClassFor(VT);
3782
3783 if (RC == &Mips::AFGR64RegClass) {
3784 assert(Reg % 2 == 0);
3785 Reg >>= 1;
3786 }
3787 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7.
3788 RC = TRI->getRegClass(Mips::FCCRegClassID);
3789 else if (Prefix == "$w") { // Parse $w0-$w31.
3790 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT);
3791 } else { // Parse $0-$31.
3792 assert(Prefix == "$");
3793 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
3794 }
3795
3796 assert(Reg < RC->getNumRegs());
3797 return std::make_pair(*(RC->begin() + Reg), RC);
3798 }
3799
3800 /// Given a register class constraint, like 'r', if this corresponds directly
3801 /// to an LLVM register class, return a register of 0 and the register class
3802 /// pointer.
3803 std::pair<unsigned, const TargetRegisterClass *>
getRegForInlineAsmConstraint(const TargetRegisterInfo * TRI,StringRef Constraint,MVT VT) const3804 MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
3805 StringRef Constraint,
3806 MVT VT) const {
3807 if (Constraint.size() == 1) {
3808 switch (Constraint[0]) {
3809 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3810 case 'y': // Same as 'r'. Exists for compatibility.
3811 case 'r':
3812 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3813 if (Subtarget.inMips16Mode())
3814 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3815 return std::make_pair(0U, &Mips::GPR32RegClass);
3816 }
3817 if (VT == MVT::i64 && !Subtarget.isGP64bit())
3818 return std::make_pair(0U, &Mips::GPR32RegClass);
3819 if (VT == MVT::i64 && Subtarget.isGP64bit())
3820 return std::make_pair(0U, &Mips::GPR64RegClass);
3821 // This will generate an error message
3822 return std::make_pair(0U, nullptr);
3823 case 'f': // FPU or MSA register
3824 if (VT == MVT::v16i8)
3825 return std::make_pair(0U, &Mips::MSA128BRegClass);
3826 else if (VT == MVT::v8i16 || VT == MVT::v8f16)
3827 return std::make_pair(0U, &Mips::MSA128HRegClass);
3828 else if (VT == MVT::v4i32 || VT == MVT::v4f32)
3829 return std::make_pair(0U, &Mips::MSA128WRegClass);
3830 else if (VT == MVT::v2i64 || VT == MVT::v2f64)
3831 return std::make_pair(0U, &Mips::MSA128DRegClass);
3832 else if (VT == MVT::f32)
3833 return std::make_pair(0U, &Mips::FGR32RegClass);
3834 else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) {
3835 if (Subtarget.isFP64bit())
3836 return std::make_pair(0U, &Mips::FGR64RegClass);
3837 return std::make_pair(0U, &Mips::AFGR64RegClass);
3838 }
3839 break;
3840 case 'c': // register suitable for indirect jump
3841 if (VT == MVT::i32)
3842 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
3843 if (VT == MVT::i64)
3844 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
3845 // This will generate an error message
3846 return std::make_pair(0U, nullptr);
3847 case 'l': // use the `lo` register to store values
3848 // that are no bigger than a word
3849 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8)
3850 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
3851 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
3852 case 'x': // use the concatenated `hi` and `lo` registers
3853 // to store doubleword values
3854 // Fixme: Not triggering the use of both hi and low
3855 // This will generate an error message
3856 return std::make_pair(0U, nullptr);
3857 }
3858 }
3859
3860 std::pair<unsigned, const TargetRegisterClass *> R;
3861 R = parseRegForInlineAsmConstraint(Constraint, VT);
3862
3863 if (R.second)
3864 return R;
3865
3866 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
3867 }
3868
3869 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3870 /// vector. If it is invalid, don't add anything to Ops.
LowerAsmOperandForConstraint(SDValue Op,std::string & Constraint,std::vector<SDValue> & Ops,SelectionDAG & DAG) const3871 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3872 std::string &Constraint,
3873 std::vector<SDValue>&Ops,
3874 SelectionDAG &DAG) const {
3875 SDLoc DL(Op);
3876 SDValue Result;
3877
3878 // Only support length 1 constraints for now.
3879 if (Constraint.length() > 1) return;
3880
3881 char ConstraintLetter = Constraint[0];
3882 switch (ConstraintLetter) {
3883 default: break; // This will fall through to the generic implementation
3884 case 'I': // Signed 16 bit constant
3885 // If this fails, the parent routine will give an error
3886 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3887 EVT Type = Op.getValueType();
3888 int64_t Val = C->getSExtValue();
3889 if (isInt<16>(Val)) {
3890 Result = DAG.getTargetConstant(Val, DL, Type);
3891 break;
3892 }
3893 }
3894 return;
3895 case 'J': // integer zero
3896 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3897 EVT Type = Op.getValueType();
3898 int64_t Val = C->getZExtValue();
3899 if (Val == 0) {
3900 Result = DAG.getTargetConstant(0, DL, Type);
3901 break;
3902 }
3903 }
3904 return;
3905 case 'K': // unsigned 16 bit immediate
3906 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3907 EVT Type = Op.getValueType();
3908 uint64_t Val = (uint64_t)C->getZExtValue();
3909 if (isUInt<16>(Val)) {
3910 Result = DAG.getTargetConstant(Val, DL, Type);
3911 break;
3912 }
3913 }
3914 return;
3915 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3916 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3917 EVT Type = Op.getValueType();
3918 int64_t Val = C->getSExtValue();
3919 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3920 Result = DAG.getTargetConstant(Val, DL, Type);
3921 break;
3922 }
3923 }
3924 return;
3925 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3926 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3927 EVT Type = Op.getValueType();
3928 int64_t Val = C->getSExtValue();
3929 if ((Val >= -65535) && (Val <= -1)) {
3930 Result = DAG.getTargetConstant(Val, DL, Type);
3931 break;
3932 }
3933 }
3934 return;
3935 case 'O': // signed 15 bit immediate
3936 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3937 EVT Type = Op.getValueType();
3938 int64_t Val = C->getSExtValue();
3939 if ((isInt<15>(Val))) {
3940 Result = DAG.getTargetConstant(Val, DL, Type);
3941 break;
3942 }
3943 }
3944 return;
3945 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3946 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3947 EVT Type = Op.getValueType();
3948 int64_t Val = C->getSExtValue();
3949 if ((Val <= 65535) && (Val >= 1)) {
3950 Result = DAG.getTargetConstant(Val, DL, Type);
3951 break;
3952 }
3953 }
3954 return;
3955 }
3956
3957 if (Result.getNode()) {
3958 Ops.push_back(Result);
3959 return;
3960 }
3961
3962 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3963 }
3964
isLegalAddressingMode(const DataLayout & DL,const AddrMode & AM,Type * Ty,unsigned AS,Instruction * I) const3965 bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL,
3966 const AddrMode &AM, Type *Ty,
3967 unsigned AS, Instruction *I) const {
3968 // No global is ever allowed as a base.
3969 if (AM.BaseGV)
3970 return false;
3971
3972 switch (AM.Scale) {
3973 case 0: // "r+i" or just "i", depending on HasBaseReg.
3974 break;
3975 case 1:
3976 if (!AM.HasBaseReg) // allow "r+i".
3977 break;
3978 return false; // disallow "r+r" or "r+r+i".
3979 default:
3980 return false;
3981 }
3982
3983 return true;
3984 }
3985
3986 bool
isOffsetFoldingLegal(const GlobalAddressSDNode * GA) const3987 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3988 // The Mips target isn't yet aware of offsets.
3989 return false;
3990 }
3991
getOptimalMemOpType(uint64_t Size,unsigned DstAlign,unsigned SrcAlign,bool IsMemset,bool ZeroMemset,bool MemcpyStrSrc,MachineFunction & MF) const3992 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3993 unsigned SrcAlign,
3994 bool IsMemset, bool ZeroMemset,
3995 bool MemcpyStrSrc,
3996 MachineFunction &MF) const {
3997 if (Subtarget.hasMips64())
3998 return MVT::i64;
3999
4000 return MVT::i32;
4001 }
4002
isFPImmLegal(const APFloat & Imm,EVT VT) const4003 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4004 if (VT != MVT::f32 && VT != MVT::f64)
4005 return false;
4006 if (Imm.isNegZero())
4007 return false;
4008 return Imm.isZero();
4009 }
4010
getJumpTableEncoding() const4011 unsigned MipsTargetLowering::getJumpTableEncoding() const {
4012
4013 // FIXME: For space reasons this should be: EK_GPRel32BlockAddress.
4014 if (ABI.IsN64() && isPositionIndependent())
4015 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
4016
4017 return TargetLowering::getJumpTableEncoding();
4018 }
4019
useSoftFloat() const4020 bool MipsTargetLowering::useSoftFloat() const {
4021 return Subtarget.useSoftFloat();
4022 }
4023
copyByValRegs(SDValue Chain,const SDLoc & DL,std::vector<SDValue> & OutChains,SelectionDAG & DAG,const ISD::ArgFlagsTy & Flags,SmallVectorImpl<SDValue> & InVals,const Argument * FuncArg,unsigned FirstReg,unsigned LastReg,const CCValAssign & VA,MipsCCState & State) const4024 void MipsTargetLowering::copyByValRegs(
4025 SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains,
4026 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
4027 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
4028 unsigned FirstReg, unsigned LastReg, const CCValAssign &VA,
4029 MipsCCState &State) const {
4030 MachineFunction &MF = DAG.getMachineFunction();
4031 MachineFrameInfo &MFI = MF.getFrameInfo();
4032 unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes();
4033 unsigned NumRegs = LastReg - FirstReg;
4034 unsigned RegAreaSize = NumRegs * GPRSizeInBytes;
4035 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
4036 int FrameObjOffset;
4037 ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs();
4038
4039 if (RegAreaSize)
4040 FrameObjOffset =
4041 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
4042 (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes);
4043 else
4044 FrameObjOffset = VA.getLocMemOffset();
4045
4046 // Create frame object.
4047 EVT PtrTy = getPointerTy(DAG.getDataLayout());
4048 // Make the fixed object stored to mutable so that the load instructions
4049 // referencing it have their memory dependencies added.
4050 // Set the frame object as isAliased which clears the underlying objects
4051 // vector in ScheduleDAGInstrs::buildSchedGraph() resulting in addition of all
4052 // stores as dependencies for loads referencing this fixed object.
4053 int FI = MFI.CreateFixedObject(FrameObjSize, FrameObjOffset, false, true);
4054 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4055 InVals.push_back(FIN);
4056
4057 if (!NumRegs)
4058 return;
4059
4060 // Copy arg registers.
4061 MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8);
4062 const TargetRegisterClass *RC = getRegClassFor(RegTy);
4063
4064 for (unsigned I = 0; I < NumRegs; ++I) {
4065 unsigned ArgReg = ByValArgRegs[FirstReg + I];
4066 unsigned VReg = addLiveIn(MF, ArgReg, RC);
4067 unsigned Offset = I * GPRSizeInBytes;
4068 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
4069 DAG.getConstant(Offset, DL, PtrTy));
4070 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
4071 StorePtr, MachinePointerInfo(FuncArg, Offset));
4072 OutChains.push_back(Store);
4073 }
4074 }
4075
4076 // Copy byVal arg to registers and stack.
passByValArg(SDValue Chain,const SDLoc & DL,std::deque<std::pair<unsigned,SDValue>> & RegsToPass,SmallVectorImpl<SDValue> & MemOpChains,SDValue StackPtr,MachineFrameInfo & MFI,SelectionDAG & DAG,SDValue Arg,unsigned FirstReg,unsigned LastReg,const ISD::ArgFlagsTy & Flags,bool isLittle,const CCValAssign & VA) const4077 void MipsTargetLowering::passByValArg(
4078 SDValue Chain, const SDLoc &DL,
4079 std::deque<std::pair<unsigned, SDValue>> &RegsToPass,
4080 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
4081 MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg,
4082 unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle,
4083 const CCValAssign &VA) const {
4084 unsigned ByValSizeInBytes = Flags.getByValSize();
4085 unsigned OffsetInBytes = 0; // From beginning of struct
4086 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
4087 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes);
4088 EVT PtrTy = getPointerTy(DAG.getDataLayout()),
4089 RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
4090 unsigned NumRegs = LastReg - FirstReg;
4091
4092 if (NumRegs) {
4093 ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs();
4094 bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes);
4095 unsigned I = 0;
4096
4097 // Copy words to registers.
4098 for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) {
4099 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4100 DAG.getConstant(OffsetInBytes, DL, PtrTy));
4101 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
4102 MachinePointerInfo(), Alignment);
4103 MemOpChains.push_back(LoadVal.getValue(1));
4104 unsigned ArgReg = ArgRegs[FirstReg + I];
4105 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
4106 }
4107
4108 // Return if the struct has been fully copied.
4109 if (ByValSizeInBytes == OffsetInBytes)
4110 return;
4111
4112 // Copy the remainder of the byval argument with sub-word loads and shifts.
4113 if (LeftoverBytes) {
4114 SDValue Val;
4115
4116 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0;
4117 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) {
4118 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes;
4119
4120 if (RemainingSizeInBytes < LoadSizeInBytes)
4121 continue;
4122
4123 // Load subword.
4124 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4125 DAG.getConstant(OffsetInBytes, DL,
4126 PtrTy));
4127 SDValue LoadVal = DAG.getExtLoad(
4128 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(),
4129 MVT::getIntegerVT(LoadSizeInBytes * 8), Alignment);
4130 MemOpChains.push_back(LoadVal.getValue(1));
4131
4132 // Shift the loaded value.
4133 unsigned Shamt;
4134
4135 if (isLittle)
4136 Shamt = TotalBytesLoaded * 8;
4137 else
4138 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8;
4139
4140 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
4141 DAG.getConstant(Shamt, DL, MVT::i32));
4142
4143 if (Val.getNode())
4144 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
4145 else
4146 Val = Shift;
4147
4148 OffsetInBytes += LoadSizeInBytes;
4149 TotalBytesLoaded += LoadSizeInBytes;
4150 Alignment = std::min(Alignment, LoadSizeInBytes);
4151 }
4152
4153 unsigned ArgReg = ArgRegs[FirstReg + I];
4154 RegsToPass.push_back(std::make_pair(ArgReg, Val));
4155 return;
4156 }
4157 }
4158
4159 // Copy remainder of byval arg to it with memcpy.
4160 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes;
4161 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4162 DAG.getConstant(OffsetInBytes, DL, PtrTy));
4163 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
4164 DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
4165 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
4166 DAG.getConstant(MemCpySize, DL, PtrTy),
4167 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false,
4168 /*isTailCall=*/false,
4169 MachinePointerInfo(), MachinePointerInfo());
4170 MemOpChains.push_back(Chain);
4171 }
4172
writeVarArgRegs(std::vector<SDValue> & OutChains,SDValue Chain,const SDLoc & DL,SelectionDAG & DAG,CCState & State) const4173 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
4174 SDValue Chain, const SDLoc &DL,
4175 SelectionDAG &DAG,
4176 CCState &State) const {
4177 ArrayRef<MCPhysReg> ArgRegs = ABI.GetVarArgRegs();
4178 unsigned Idx = State.getFirstUnallocated(ArgRegs);
4179 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
4180 MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
4181 const TargetRegisterClass *RC = getRegClassFor(RegTy);
4182 MachineFunction &MF = DAG.getMachineFunction();
4183 MachineFrameInfo &MFI = MF.getFrameInfo();
4184 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
4185
4186 // Offset of the first variable argument from stack pointer.
4187 int VaArgOffset;
4188
4189 if (ArgRegs.size() == Idx)
4190 VaArgOffset = alignTo(State.getNextStackOffset(), RegSizeInBytes);
4191 else {
4192 VaArgOffset =
4193 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
4194 (int)(RegSizeInBytes * (ArgRegs.size() - Idx));
4195 }
4196
4197 // Record the frame index of the first variable argument
4198 // which is a value necessary to VASTART.
4199 int FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
4200 MipsFI->setVarArgsFrameIndex(FI);
4201
4202 // Copy the integer registers that have not been used for argument passing
4203 // to the argument register save area. For O32, the save area is allocated
4204 // in the caller's stack frame, while for N32/64, it is allocated in the
4205 // callee's stack frame.
4206 for (unsigned I = Idx; I < ArgRegs.size();
4207 ++I, VaArgOffset += RegSizeInBytes) {
4208 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
4209 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
4210 FI = MFI.CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
4211 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
4212 SDValue Store =
4213 DAG.getStore(Chain, DL, ArgValue, PtrOff, MachinePointerInfo());
4214 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(
4215 (Value *)nullptr);
4216 OutChains.push_back(Store);
4217 }
4218 }
4219
HandleByVal(CCState * State,unsigned & Size,unsigned Align) const4220 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size,
4221 unsigned Align) const {
4222 const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
4223
4224 assert(Size && "Byval argument's size shouldn't be 0.");
4225
4226 Align = std::min(Align, TFL->getStackAlignment());
4227
4228 unsigned FirstReg = 0;
4229 unsigned NumRegs = 0;
4230
4231 if (State->getCallingConv() != CallingConv::Fast) {
4232 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
4233 ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs();
4234 // FIXME: The O32 case actually describes no shadow registers.
4235 const MCPhysReg *ShadowRegs =
4236 ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs;
4237
4238 // We used to check the size as well but we can't do that anymore since
4239 // CCState::HandleByVal() rounds up the size after calling this function.
4240 assert(!(Align % RegSizeInBytes) &&
4241 "Byval argument's alignment should be a multiple of"
4242 "RegSizeInBytes.");
4243
4244 FirstReg = State->getFirstUnallocated(IntArgRegs);
4245
4246 // If Align > RegSizeInBytes, the first arg register must be even.
4247 // FIXME: This condition happens to do the right thing but it's not the
4248 // right way to test it. We want to check that the stack frame offset
4249 // of the register is aligned.
4250 if ((Align > RegSizeInBytes) && (FirstReg % 2)) {
4251 State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]);
4252 ++FirstReg;
4253 }
4254
4255 // Mark the registers allocated.
4256 Size = alignTo(Size, RegSizeInBytes);
4257 for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size());
4258 Size -= RegSizeInBytes, ++I, ++NumRegs)
4259 State->AllocateReg(IntArgRegs[I], ShadowRegs[I]);
4260 }
4261
4262 State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs);
4263 }
4264
emitPseudoSELECT(MachineInstr & MI,MachineBasicBlock * BB,bool isFPCmp,unsigned Opc) const4265 MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI,
4266 MachineBasicBlock *BB,
4267 bool isFPCmp,
4268 unsigned Opc) const {
4269 assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) &&
4270 "Subtarget already supports SELECT nodes with the use of"
4271 "conditional-move instructions.");
4272
4273 const TargetInstrInfo *TII =
4274 Subtarget.getInstrInfo();
4275 DebugLoc DL = MI.getDebugLoc();
4276
4277 // To "insert" a SELECT instruction, we actually have to insert the
4278 // diamond control-flow pattern. The incoming instruction knows the
4279 // destination vreg to set, the condition code register to branch on, the
4280 // true/false values to select between, and a branch opcode to use.
4281 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4282 MachineFunction::iterator It = ++BB->getIterator();
4283
4284 // thisMBB:
4285 // ...
4286 // TrueVal = ...
4287 // setcc r1, r2, r3
4288 // bNE r1, r0, copy1MBB
4289 // fallthrough --> copy0MBB
4290 MachineBasicBlock *thisMBB = BB;
4291 MachineFunction *F = BB->getParent();
4292 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4293 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4294 F->insert(It, copy0MBB);
4295 F->insert(It, sinkMBB);
4296
4297 // Transfer the remainder of BB and its successor edges to sinkMBB.
4298 sinkMBB->splice(sinkMBB->begin(), BB,
4299 std::next(MachineBasicBlock::iterator(MI)), BB->end());
4300 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4301
4302 // Next, add the true and fallthrough blocks as its successors.
4303 BB->addSuccessor(copy0MBB);
4304 BB->addSuccessor(sinkMBB);
4305
4306 if (isFPCmp) {
4307 // bc1[tf] cc, sinkMBB
4308 BuildMI(BB, DL, TII->get(Opc))
4309 .addReg(MI.getOperand(1).getReg())
4310 .addMBB(sinkMBB);
4311 } else {
4312 // bne rs, $0, sinkMBB
4313 BuildMI(BB, DL, TII->get(Opc))
4314 .addReg(MI.getOperand(1).getReg())
4315 .addReg(Mips::ZERO)
4316 .addMBB(sinkMBB);
4317 }
4318
4319 // copy0MBB:
4320 // %FalseValue = ...
4321 // # fallthrough to sinkMBB
4322 BB = copy0MBB;
4323
4324 // Update machine-CFG edges
4325 BB->addSuccessor(sinkMBB);
4326
4327 // sinkMBB:
4328 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
4329 // ...
4330 BB = sinkMBB;
4331
4332 BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg())
4333 .addReg(MI.getOperand(2).getReg())
4334 .addMBB(thisMBB)
4335 .addReg(MI.getOperand(3).getReg())
4336 .addMBB(copy0MBB);
4337
4338 MI.eraseFromParent(); // The pseudo instruction is gone now.
4339
4340 return BB;
4341 }
4342
4343 // FIXME? Maybe this could be a TableGen attribute on some registers and
4344 // this table could be generated automatically from RegInfo.
getRegisterByName(const char * RegName,EVT VT,SelectionDAG & DAG) const4345 unsigned MipsTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4346 SelectionDAG &DAG) const {
4347 // Named registers is expected to be fairly rare. For now, just support $28
4348 // since the linux kernel uses it.
4349 if (Subtarget.isGP64bit()) {
4350 unsigned Reg = StringSwitch<unsigned>(RegName)
4351 .Case("$28", Mips::GP_64)
4352 .Default(0);
4353 if (Reg)
4354 return Reg;
4355 } else {
4356 unsigned Reg = StringSwitch<unsigned>(RegName)
4357 .Case("$28", Mips::GP)
4358 .Default(0);
4359 if (Reg)
4360 return Reg;
4361 }
4362 report_fatal_error("Invalid register name global variable");
4363 }
4364