1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that Mips uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
14 #include "MipsISelLowering.h"
15 #include "InstPrinter/MipsInstPrinter.h"
16 #include "MCTargetDesc/MipsBaseInfo.h"
17 #include "MipsMachineFunction.h"
18 #include "MipsSubtarget.h"
19 #include "MipsTargetMachine.h"
20 #include "MipsTargetObjectFile.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/ADT/StringSwitch.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAGISel.h"
29 #include "llvm/CodeGen/ValueTypes.h"
30 #include "llvm/IR/CallingConv.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
41 #define DEBUG_TYPE "mips-lower"
43 STATISTIC(NumTailCalls, "Number of tail calls");
46 LargeGOT("mxgot", cl::Hidden,
47 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
50 NoZeroDivCheck("mno-check-zero-division", cl::Hidden,
51 cl::desc("MIPS: Don't trap on integer division by zero."),
55 EnableMipsFastISel("mips-fast-isel", cl::Hidden,
56 cl::desc("Allow mips-fast-isel to be used"),
59 static const MCPhysReg O32IntRegs[4] = {
60 Mips::A0, Mips::A1, Mips::A2, Mips::A3
63 static const MCPhysReg Mips64IntRegs[8] = {
64 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
65 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64
68 static const MCPhysReg Mips64DPRegs[8] = {
69 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
70 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
73 // If I is a shifted mask, set the size (Size) and the first bit of the
74 // mask (Pos), and return true.
75 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
76 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
77 if (!isShiftedMask_64(I))
80 Size = CountPopulation_64(I);
81 Pos = countTrailingZeros(I);
85 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
86 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
87 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
90 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
92 unsigned Flag) const {
93 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag);
96 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty,
98 unsigned Flag) const {
99 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
102 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty,
104 unsigned Flag) const {
105 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
108 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty,
110 unsigned Flag) const {
111 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
114 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty,
116 unsigned Flag) const {
117 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
118 N->getOffset(), Flag);
121 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
123 case MipsISD::JmpLink: return "MipsISD::JmpLink";
124 case MipsISD::TailCall: return "MipsISD::TailCall";
125 case MipsISD::Hi: return "MipsISD::Hi";
126 case MipsISD::Lo: return "MipsISD::Lo";
127 case MipsISD::GPRel: return "MipsISD::GPRel";
128 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
129 case MipsISD::Ret: return "MipsISD::Ret";
130 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
131 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
132 case MipsISD::FPCmp: return "MipsISD::FPCmp";
133 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
134 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
135 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP";
136 case MipsISD::MFHI: return "MipsISD::MFHI";
137 case MipsISD::MFLO: return "MipsISD::MFLO";
138 case MipsISD::MTLOHI: return "MipsISD::MTLOHI";
139 case MipsISD::Mult: return "MipsISD::Mult";
140 case MipsISD::Multu: return "MipsISD::Multu";
141 case MipsISD::MAdd: return "MipsISD::MAdd";
142 case MipsISD::MAddu: return "MipsISD::MAddu";
143 case MipsISD::MSub: return "MipsISD::MSub";
144 case MipsISD::MSubu: return "MipsISD::MSubu";
145 case MipsISD::DivRem: return "MipsISD::DivRem";
146 case MipsISD::DivRemU: return "MipsISD::DivRemU";
147 case MipsISD::DivRem16: return "MipsISD::DivRem16";
148 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
149 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
150 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
151 case MipsISD::Wrapper: return "MipsISD::Wrapper";
152 case MipsISD::Sync: return "MipsISD::Sync";
153 case MipsISD::Ext: return "MipsISD::Ext";
154 case MipsISD::Ins: return "MipsISD::Ins";
155 case MipsISD::LWL: return "MipsISD::LWL";
156 case MipsISD::LWR: return "MipsISD::LWR";
157 case MipsISD::SWL: return "MipsISD::SWL";
158 case MipsISD::SWR: return "MipsISD::SWR";
159 case MipsISD::LDL: return "MipsISD::LDL";
160 case MipsISD::LDR: return "MipsISD::LDR";
161 case MipsISD::SDL: return "MipsISD::SDL";
162 case MipsISD::SDR: return "MipsISD::SDR";
163 case MipsISD::EXTP: return "MipsISD::EXTP";
164 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
165 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
166 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
167 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
168 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
169 case MipsISD::SHILO: return "MipsISD::SHILO";
170 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
171 case MipsISD::MULT: return "MipsISD::MULT";
172 case MipsISD::MULTU: return "MipsISD::MULTU";
173 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
174 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
175 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
176 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
177 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP";
178 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP";
179 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP";
180 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP";
181 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP";
182 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO";
183 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO";
184 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO";
185 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO";
186 case MipsISD::VCEQ: return "MipsISD::VCEQ";
187 case MipsISD::VCLE_S: return "MipsISD::VCLE_S";
188 case MipsISD::VCLE_U: return "MipsISD::VCLE_U";
189 case MipsISD::VCLT_S: return "MipsISD::VCLT_S";
190 case MipsISD::VCLT_U: return "MipsISD::VCLT_U";
191 case MipsISD::VSMAX: return "MipsISD::VSMAX";
192 case MipsISD::VSMIN: return "MipsISD::VSMIN";
193 case MipsISD::VUMAX: return "MipsISD::VUMAX";
194 case MipsISD::VUMIN: return "MipsISD::VUMIN";
195 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
196 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
197 case MipsISD::VNOR: return "MipsISD::VNOR";
198 case MipsISD::VSHF: return "MipsISD::VSHF";
199 case MipsISD::SHF: return "MipsISD::SHF";
200 case MipsISD::ILVEV: return "MipsISD::ILVEV";
201 case MipsISD::ILVOD: return "MipsISD::ILVOD";
202 case MipsISD::ILVL: return "MipsISD::ILVL";
203 case MipsISD::ILVR: return "MipsISD::ILVR";
204 case MipsISD::PCKEV: return "MipsISD::PCKEV";
205 case MipsISD::PCKOD: return "MipsISD::PCKOD";
206 case MipsISD::INSVE: return "MipsISD::INSVE";
207 default: return nullptr;
211 MipsTargetLowering::MipsTargetLowering(MipsTargetMachine &TM)
212 : TargetLowering(TM, new MipsTargetObjectFile()),
213 Subtarget(&TM.getSubtarget<MipsSubtarget>()) {
214 // Mips does not have i1 type, so use i32 for
215 // setcc operations results (slt, sgt, ...).
216 setBooleanContents(ZeroOrOneBooleanContent);
217 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
219 // Load extented operations for i1 types must be promoted
220 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
221 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
222 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
224 // MIPS doesn't have extending float->double load/store
225 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
226 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
228 // Used by legalize types to correctly generate the setcc result.
229 // Without this, every float setcc comes with a AND/OR with the result,
230 // we don't want this, since the fpcmp result goes to a flag register,
231 // which is used implicitly by brcond and select operations.
232 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
234 // Mips Custom Operations
235 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
236 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
237 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
238 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
239 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
240 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
241 setOperationAction(ISD::SELECT, MVT::f32, Custom);
242 setOperationAction(ISD::SELECT, MVT::f64, Custom);
243 setOperationAction(ISD::SELECT, MVT::i32, Custom);
244 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
245 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
246 setOperationAction(ISD::SETCC, MVT::f32, Custom);
247 setOperationAction(ISD::SETCC, MVT::f64, Custom);
248 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
249 setOperationAction(ISD::VASTART, MVT::Other, Custom);
250 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
251 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
252 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
255 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
256 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
257 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
258 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
259 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
260 setOperationAction(ISD::SELECT, MVT::i64, Custom);
261 setOperationAction(ISD::LOAD, MVT::i64, Custom);
262 setOperationAction(ISD::STORE, MVT::i64, Custom);
263 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
267 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
268 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
269 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
272 setOperationAction(ISD::ADD, MVT::i32, Custom);
274 setOperationAction(ISD::ADD, MVT::i64, Custom);
276 setOperationAction(ISD::SDIV, MVT::i32, Expand);
277 setOperationAction(ISD::SREM, MVT::i32, Expand);
278 setOperationAction(ISD::UDIV, MVT::i32, Expand);
279 setOperationAction(ISD::UREM, MVT::i32, Expand);
280 setOperationAction(ISD::SDIV, MVT::i64, Expand);
281 setOperationAction(ISD::SREM, MVT::i64, Expand);
282 setOperationAction(ISD::UDIV, MVT::i64, Expand);
283 setOperationAction(ISD::UREM, MVT::i64, Expand);
285 // Operations not directly supported by Mips.
286 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
287 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
288 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
289 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
290 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
291 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
292 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
293 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
294 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
295 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
296 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
297 if (Subtarget->hasCnMips()) {
298 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
299 setOperationAction(ISD::CTPOP, MVT::i64, Legal);
301 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
302 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
304 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
305 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
306 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
307 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
308 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
309 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
310 setOperationAction(ISD::ROTL, MVT::i32, Expand);
311 setOperationAction(ISD::ROTL, MVT::i64, Expand);
312 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
313 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
315 if (!Subtarget->hasMips32r2())
316 setOperationAction(ISD::ROTR, MVT::i32, Expand);
318 if (!Subtarget->hasMips64r2())
319 setOperationAction(ISD::ROTR, MVT::i64, Expand);
321 setOperationAction(ISD::FSIN, MVT::f32, Expand);
322 setOperationAction(ISD::FSIN, MVT::f64, Expand);
323 setOperationAction(ISD::FCOS, MVT::f32, Expand);
324 setOperationAction(ISD::FCOS, MVT::f64, Expand);
325 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
326 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
327 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
328 setOperationAction(ISD::FPOW, MVT::f32, Expand);
329 setOperationAction(ISD::FPOW, MVT::f64, Expand);
330 setOperationAction(ISD::FLOG, MVT::f32, Expand);
331 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
332 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
333 setOperationAction(ISD::FEXP, MVT::f32, Expand);
334 setOperationAction(ISD::FMA, MVT::f32, Expand);
335 setOperationAction(ISD::FMA, MVT::f64, Expand);
336 setOperationAction(ISD::FREM, MVT::f32, Expand);
337 setOperationAction(ISD::FREM, MVT::f64, Expand);
339 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
341 setOperationAction(ISD::VAARG, MVT::Other, Expand);
342 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
343 setOperationAction(ISD::VAEND, MVT::Other, Expand);
345 // Use the default for now
346 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
347 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
349 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
350 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
351 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
352 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
354 setInsertFencesForAtomic(true);
356 if (!Subtarget->hasMips32r2()) {
357 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
358 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
361 // MIPS16 lacks MIPS32's clz and clo instructions.
362 if (!Subtarget->hasMips32() || Subtarget->inMips16Mode())
363 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
364 if (!Subtarget->hasMips64())
365 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
367 if (!Subtarget->hasMips32r2())
368 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
369 if (!Subtarget->hasMips64r2())
370 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
373 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
374 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
375 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
376 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
379 setOperationAction(ISD::TRAP, MVT::Other, Legal);
381 setTargetDAGCombine(ISD::SDIVREM);
382 setTargetDAGCombine(ISD::UDIVREM);
383 setTargetDAGCombine(ISD::SELECT);
384 setTargetDAGCombine(ISD::AND);
385 setTargetDAGCombine(ISD::OR);
386 setTargetDAGCombine(ISD::ADD);
388 setMinFunctionAlignment(isGP64bit() ? 3 : 2);
390 setStackPointerRegisterToSaveRestore(isN64() ? Mips::SP_64 : Mips::SP);
392 setExceptionPointerRegister(isN64() ? Mips::A0_64 : Mips::A0);
393 setExceptionSelectorRegister(isN64() ? Mips::A1_64 : Mips::A1);
395 MaxStoresPerMemcpy = 16;
397 isMicroMips = Subtarget->inMicroMipsMode();
400 const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) {
401 if (TM.getSubtargetImpl()->inMips16Mode())
402 return llvm::createMips16TargetLowering(TM);
404 return llvm::createMipsSETargetLowering(TM);
407 // Create a fast isel object.
409 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
410 const TargetLibraryInfo *libInfo) const {
411 if (!EnableMipsFastISel)
412 return TargetLowering::createFastISel(funcInfo, libInfo);
413 return Mips::createFastISel(funcInfo, libInfo);
416 EVT MipsTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
419 return VT.changeVectorElementTypeToInteger();
422 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
423 TargetLowering::DAGCombinerInfo &DCI,
424 const MipsSubtarget *Subtarget) {
425 if (DCI.isBeforeLegalizeOps())
428 EVT Ty = N->getValueType(0);
429 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
430 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
431 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
435 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
436 N->getOperand(0), N->getOperand(1));
437 SDValue InChain = DAG.getEntryNode();
438 SDValue InGlue = DivRem;
441 if (N->hasAnyUseOfValue(0)) {
442 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
444 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
445 InChain = CopyFromLo.getValue(1);
446 InGlue = CopyFromLo.getValue(2);
450 if (N->hasAnyUseOfValue(1)) {
451 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
453 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
459 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
461 default: llvm_unreachable("Unknown fp condition code!");
463 case ISD::SETOEQ: return Mips::FCOND_OEQ;
464 case ISD::SETUNE: return Mips::FCOND_UNE;
466 case ISD::SETOLT: return Mips::FCOND_OLT;
468 case ISD::SETOGT: return Mips::FCOND_OGT;
470 case ISD::SETOLE: return Mips::FCOND_OLE;
472 case ISD::SETOGE: return Mips::FCOND_OGE;
473 case ISD::SETULT: return Mips::FCOND_ULT;
474 case ISD::SETULE: return Mips::FCOND_ULE;
475 case ISD::SETUGT: return Mips::FCOND_UGT;
476 case ISD::SETUGE: return Mips::FCOND_UGE;
477 case ISD::SETUO: return Mips::FCOND_UN;
478 case ISD::SETO: return Mips::FCOND_OR;
480 case ISD::SETONE: return Mips::FCOND_ONE;
481 case ISD::SETUEQ: return Mips::FCOND_UEQ;
486 /// This function returns true if the floating point conditional branches and
487 /// conditional moves which use condition code CC should be inverted.
488 static bool invertFPCondCodeUser(Mips::CondCode CC) {
489 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
492 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
493 "Illegal Condition Code");
498 // Creates and returns an FPCmp node from a setcc node.
499 // Returns Op if setcc is not a floating point comparison.
500 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
501 // must be a SETCC node
502 if (Op.getOpcode() != ISD::SETCC)
505 SDValue LHS = Op.getOperand(0);
507 if (!LHS.getValueType().isFloatingPoint())
510 SDValue RHS = Op.getOperand(1);
513 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
514 // node if necessary.
515 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
517 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
518 DAG.getConstant(condCodeToFCC(CC), MVT::i32));
521 // Creates and returns a CMovFPT/F node.
522 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
523 SDValue False, SDLoc DL) {
524 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
525 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
526 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
528 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
529 True.getValueType(), True, FCC0, False, Cond);
532 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
533 TargetLowering::DAGCombinerInfo &DCI,
534 const MipsSubtarget *Subtarget) {
535 if (DCI.isBeforeLegalizeOps())
538 SDValue SetCC = N->getOperand(0);
540 if ((SetCC.getOpcode() != ISD::SETCC) ||
541 !SetCC.getOperand(0).getValueType().isInteger())
544 SDValue False = N->getOperand(2);
545 EVT FalseTy = False.getValueType();
547 if (!FalseTy.isInteger())
550 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False);
552 // If the RHS (False) is 0, we swap the order of the operands
553 // of ISD::SELECT (obviously also inverting the condition) so that we can
554 // take advantage of conditional moves using the $0 register.
556 // return (a != 0) ? x : 0;
564 if (!FalseC->getZExtValue()) {
565 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
566 SDValue True = N->getOperand(1);
568 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
569 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
571 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
574 // If both operands are integer constants there's a possibility that we
575 // can do some interesting optimizations.
576 SDValue True = N->getOperand(1);
577 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True);
579 if (!TrueC || !True.getValueType().isInteger())
582 // We'll also ignore MVT::i64 operands as this optimizations proves
583 // to be ineffective because of the required sign extensions as the result
584 // of a SETCC operator is always MVT::i32 for non-vector types.
585 if (True.getValueType() == MVT::i64)
588 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue();
590 // 1) (a < x) ? y : y-1
592 // addiu $reg2, $reg1, y-1
594 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False);
596 // 2) (a < x) ? y-1 : y
598 // xor $reg1, $reg1, 1
599 // addiu $reg2, $reg1, y-1
601 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
602 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
603 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
604 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True);
607 // Couldn't optimize.
611 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
612 TargetLowering::DAGCombinerInfo &DCI,
613 const MipsSubtarget *Subtarget) {
614 // Pattern match EXT.
615 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
616 // => ext $dst, $src, size, pos
617 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasExtractInsert())
620 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
621 unsigned ShiftRightOpc = ShiftRight.getOpcode();
623 // Op's first operand must be a shift right.
624 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
627 // The second operand of the shift must be an immediate.
629 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
632 uint64_t Pos = CN->getZExtValue();
633 uint64_t SMPos, SMSize;
635 // Op's second operand must be a shifted mask.
636 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
637 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
640 // Return if the shifted mask does not start at bit 0 or the sum of its size
641 // and Pos exceeds the word's size.
642 EVT ValTy = N->getValueType(0);
643 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
646 return DAG.getNode(MipsISD::Ext, SDLoc(N), ValTy,
647 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
648 DAG.getConstant(SMSize, MVT::i32));
651 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
652 TargetLowering::DAGCombinerInfo &DCI,
653 const MipsSubtarget *Subtarget) {
654 // Pattern match INS.
655 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
656 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
657 // => ins $dst, $src, size, pos, $src1
658 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasExtractInsert())
661 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
662 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
665 // See if Op's first operand matches (and $src1 , mask0).
666 if (And0.getOpcode() != ISD::AND)
669 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
670 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
673 // See if Op's second operand matches (and (shl $src, pos), mask1).
674 if (And1.getOpcode() != ISD::AND)
677 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
678 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
681 // The shift masks must have the same position and size.
682 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
685 SDValue Shl = And1.getOperand(0);
686 if (Shl.getOpcode() != ISD::SHL)
689 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
692 unsigned Shamt = CN->getZExtValue();
694 // Return if the shift amount and the first bit position of mask are not the
696 EVT ValTy = N->getValueType(0);
697 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
700 return DAG.getNode(MipsISD::Ins, SDLoc(N), ValTy, Shl.getOperand(0),
701 DAG.getConstant(SMPos0, MVT::i32),
702 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
705 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
706 TargetLowering::DAGCombinerInfo &DCI,
707 const MipsSubtarget *Subtarget) {
708 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
710 if (DCI.isBeforeLegalizeOps())
713 SDValue Add = N->getOperand(1);
715 if (Add.getOpcode() != ISD::ADD)
718 SDValue Lo = Add.getOperand(1);
720 if ((Lo.getOpcode() != MipsISD::Lo) ||
721 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
724 EVT ValTy = N->getValueType(0);
727 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
729 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
732 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
734 SelectionDAG &DAG = DCI.DAG;
735 unsigned Opc = N->getOpcode();
741 return performDivRemCombine(N, DAG, DCI, Subtarget);
743 return performSELECTCombine(N, DAG, DCI, Subtarget);
745 return performANDCombine(N, DAG, DCI, Subtarget);
747 return performORCombine(N, DAG, DCI, Subtarget);
749 return performADDCombine(N, DAG, DCI, Subtarget);
756 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
757 SmallVectorImpl<SDValue> &Results,
758 SelectionDAG &DAG) const {
759 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
761 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
762 Results.push_back(Res.getValue(I));
766 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
767 SmallVectorImpl<SDValue> &Results,
768 SelectionDAG &DAG) const {
769 return LowerOperationWrapper(N, Results, DAG);
772 SDValue MipsTargetLowering::
773 LowerOperation(SDValue Op, SelectionDAG &DAG) const
775 switch (Op.getOpcode())
777 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
778 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
779 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
780 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
781 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
782 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
783 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
784 case ISD::SELECT: return lowerSELECT(Op, DAG);
785 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
786 case ISD::SETCC: return lowerSETCC(Op, DAG);
787 case ISD::VASTART: return lowerVASTART(Op, DAG);
788 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
789 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
790 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
791 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
792 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
793 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
794 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
795 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
796 case ISD::LOAD: return lowerLOAD(Op, DAG);
797 case ISD::STORE: return lowerSTORE(Op, DAG);
798 case ISD::ADD: return lowerADD(Op, DAG);
799 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG);
804 //===----------------------------------------------------------------------===//
805 // Lower helper functions
806 //===----------------------------------------------------------------------===//
808 // addLiveIn - This helper function adds the specified physical register to the
809 // MachineFunction as a live in value. It also creates a corresponding
810 // virtual register for it.
812 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
814 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
815 MF.getRegInfo().addLiveIn(PReg, VReg);
819 static MachineBasicBlock *expandPseudoDIV(MachineInstr *MI,
820 MachineBasicBlock &MBB,
821 const TargetInstrInfo &TII,
826 // Insert instruction "teq $divisor_reg, $zero, 7".
827 MachineBasicBlock::iterator I(MI);
828 MachineInstrBuilder MIB;
829 MachineOperand &Divisor = MI->getOperand(2);
830 MIB = BuildMI(MBB, std::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ))
831 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill()))
832 .addReg(Mips::ZERO).addImm(7);
834 // Use the 32-bit sub-register if this is a 64-bit division.
836 MIB->getOperand(0).setSubReg(Mips::sub_32);
838 // Clear Divisor's kill flag.
839 Divisor.setIsKill(false);
844 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
845 MachineBasicBlock *BB) const {
846 switch (MI->getOpcode()) {
848 llvm_unreachable("Unexpected instr type to insert");
849 case Mips::ATOMIC_LOAD_ADD_I8:
850 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
851 case Mips::ATOMIC_LOAD_ADD_I16:
852 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
853 case Mips::ATOMIC_LOAD_ADD_I32:
854 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
855 case Mips::ATOMIC_LOAD_ADD_I64:
856 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
858 case Mips::ATOMIC_LOAD_AND_I8:
859 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
860 case Mips::ATOMIC_LOAD_AND_I16:
861 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
862 case Mips::ATOMIC_LOAD_AND_I32:
863 return emitAtomicBinary(MI, BB, 4, Mips::AND);
864 case Mips::ATOMIC_LOAD_AND_I64:
865 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
867 case Mips::ATOMIC_LOAD_OR_I8:
868 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
869 case Mips::ATOMIC_LOAD_OR_I16:
870 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
871 case Mips::ATOMIC_LOAD_OR_I32:
872 return emitAtomicBinary(MI, BB, 4, Mips::OR);
873 case Mips::ATOMIC_LOAD_OR_I64:
874 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
876 case Mips::ATOMIC_LOAD_XOR_I8:
877 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
878 case Mips::ATOMIC_LOAD_XOR_I16:
879 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
880 case Mips::ATOMIC_LOAD_XOR_I32:
881 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
882 case Mips::ATOMIC_LOAD_XOR_I64:
883 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
885 case Mips::ATOMIC_LOAD_NAND_I8:
886 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
887 case Mips::ATOMIC_LOAD_NAND_I16:
888 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
889 case Mips::ATOMIC_LOAD_NAND_I32:
890 return emitAtomicBinary(MI, BB, 4, 0, true);
891 case Mips::ATOMIC_LOAD_NAND_I64:
892 return emitAtomicBinary(MI, BB, 8, 0, true);
894 case Mips::ATOMIC_LOAD_SUB_I8:
895 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
896 case Mips::ATOMIC_LOAD_SUB_I16:
897 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
898 case Mips::ATOMIC_LOAD_SUB_I32:
899 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
900 case Mips::ATOMIC_LOAD_SUB_I64:
901 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
903 case Mips::ATOMIC_SWAP_I8:
904 return emitAtomicBinaryPartword(MI, BB, 1, 0);
905 case Mips::ATOMIC_SWAP_I16:
906 return emitAtomicBinaryPartword(MI, BB, 2, 0);
907 case Mips::ATOMIC_SWAP_I32:
908 return emitAtomicBinary(MI, BB, 4, 0);
909 case Mips::ATOMIC_SWAP_I64:
910 return emitAtomicBinary(MI, BB, 8, 0);
912 case Mips::ATOMIC_CMP_SWAP_I8:
913 return emitAtomicCmpSwapPartword(MI, BB, 1);
914 case Mips::ATOMIC_CMP_SWAP_I16:
915 return emitAtomicCmpSwapPartword(MI, BB, 2);
916 case Mips::ATOMIC_CMP_SWAP_I32:
917 return emitAtomicCmpSwap(MI, BB, 4);
918 case Mips::ATOMIC_CMP_SWAP_I64:
919 return emitAtomicCmpSwap(MI, BB, 8);
920 case Mips::PseudoSDIV:
921 case Mips::PseudoUDIV:
922 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), false);
923 case Mips::PseudoDSDIV:
924 case Mips::PseudoDUDIV:
925 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), true);
929 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
930 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
932 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
933 unsigned Size, unsigned BinOpcode,
935 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
937 MachineFunction *MF = BB->getParent();
938 MachineRegisterInfo &RegInfo = MF->getRegInfo();
939 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
940 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
941 DebugLoc DL = MI->getDebugLoc();
942 unsigned LL, SC, AND, NOR, ZERO, BEQ;
945 LL = isMicroMips ? Mips::LL_MM : Mips::LL;
946 SC = isMicroMips ? Mips::SC_MM : Mips::SC;
957 ZERO = Mips::ZERO_64;
961 unsigned OldVal = MI->getOperand(0).getReg();
962 unsigned Ptr = MI->getOperand(1).getReg();
963 unsigned Incr = MI->getOperand(2).getReg();
965 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
966 unsigned AndRes = RegInfo.createVirtualRegister(RC);
967 unsigned Success = RegInfo.createVirtualRegister(RC);
969 // insert new blocks after the current block
970 const BasicBlock *LLVM_BB = BB->getBasicBlock();
971 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
972 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
973 MachineFunction::iterator It = BB;
975 MF->insert(It, loopMBB);
976 MF->insert(It, exitMBB);
978 // Transfer the remainder of BB and its successor edges to exitMBB.
979 exitMBB->splice(exitMBB->begin(), BB,
980 std::next(MachineBasicBlock::iterator(MI)), BB->end());
981 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
985 // fallthrough --> loopMBB
986 BB->addSuccessor(loopMBB);
987 loopMBB->addSuccessor(loopMBB);
988 loopMBB->addSuccessor(exitMBB);
992 // <binop> storeval, oldval, incr
993 // sc success, storeval, 0(ptr)
994 // beq success, $0, loopMBB
996 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
998 // and andres, oldval, incr
999 // nor storeval, $0, andres
1000 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1001 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1002 } else if (BinOpcode) {
1003 // <binop> storeval, oldval, incr
1004 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1008 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1009 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1011 MI->eraseFromParent(); // The instruction is gone now.
1017 MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI,
1018 MachineBasicBlock *BB,
1019 unsigned Size, unsigned BinOpcode,
1021 assert((Size == 1 || Size == 2) &&
1022 "Unsupported size for EmitAtomicBinaryPartial.");
1024 MachineFunction *MF = BB->getParent();
1025 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1026 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1027 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1028 DebugLoc DL = MI->getDebugLoc();
1030 unsigned Dest = MI->getOperand(0).getReg();
1031 unsigned Ptr = MI->getOperand(1).getReg();
1032 unsigned Incr = MI->getOperand(2).getReg();
1034 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1035 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1036 unsigned Mask = RegInfo.createVirtualRegister(RC);
1037 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1038 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1039 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1040 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1041 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1042 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1043 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1044 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1045 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1046 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1047 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1048 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1049 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1050 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1051 unsigned Success = RegInfo.createVirtualRegister(RC);
1053 // insert new blocks after the current block
1054 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1055 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1056 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1057 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1058 MachineFunction::iterator It = BB;
1060 MF->insert(It, loopMBB);
1061 MF->insert(It, sinkMBB);
1062 MF->insert(It, exitMBB);
1064 // Transfer the remainder of BB and its successor edges to exitMBB.
1065 exitMBB->splice(exitMBB->begin(), BB,
1066 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1067 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1069 BB->addSuccessor(loopMBB);
1070 loopMBB->addSuccessor(loopMBB);
1071 loopMBB->addSuccessor(sinkMBB);
1072 sinkMBB->addSuccessor(exitMBB);
1075 // addiu masklsb2,$0,-4 # 0xfffffffc
1076 // and alignedaddr,ptr,masklsb2
1077 // andi ptrlsb2,ptr,3
1078 // sll shiftamt,ptrlsb2,3
1079 // ori maskupper,$0,255 # 0xff
1080 // sll mask,maskupper,shiftamt
1081 // nor mask2,$0,mask
1082 // sll incr2,incr,shiftamt
1084 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1085 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1086 .addReg(Mips::ZERO).addImm(-4);
1087 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1088 .addReg(Ptr).addReg(MaskLSB2);
1089 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1090 if (Subtarget->isLittle()) {
1091 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1093 unsigned Off = RegInfo.createVirtualRegister(RC);
1094 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1095 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1096 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1098 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1099 .addReg(Mips::ZERO).addImm(MaskImm);
1100 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1101 .addReg(MaskUpper).addReg(ShiftAmt);
1102 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1103 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1105 // atomic.load.binop
1107 // ll oldval,0(alignedaddr)
1108 // binop binopres,oldval,incr2
1109 // and newval,binopres,mask
1110 // and maskedoldval0,oldval,mask2
1111 // or storeval,maskedoldval0,newval
1112 // sc success,storeval,0(alignedaddr)
1113 // beq success,$0,loopMBB
1117 // ll oldval,0(alignedaddr)
1118 // and newval,incr2,mask
1119 // and maskedoldval0,oldval,mask2
1120 // or storeval,maskedoldval0,newval
1121 // sc success,storeval,0(alignedaddr)
1122 // beq success,$0,loopMBB
1125 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1127 // and andres, oldval, incr2
1128 // nor binopres, $0, andres
1129 // and newval, binopres, mask
1130 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1131 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1132 .addReg(Mips::ZERO).addReg(AndRes);
1133 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1134 } else if (BinOpcode) {
1135 // <binop> binopres, oldval, incr2
1136 // and newval, binopres, mask
1137 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1138 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1139 } else { // atomic.swap
1140 // and newval, incr2, mask
1141 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1144 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1145 .addReg(OldVal).addReg(Mask2);
1146 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1147 .addReg(MaskedOldVal0).addReg(NewVal);
1148 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1149 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1150 BuildMI(BB, DL, TII->get(Mips::BEQ))
1151 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1154 // and maskedoldval1,oldval,mask
1155 // srl srlres,maskedoldval1,shiftamt
1156 // sll sllres,srlres,24
1157 // sra dest,sllres,24
1159 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1161 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1162 .addReg(OldVal).addReg(Mask);
1163 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1164 .addReg(MaskedOldVal1).addReg(ShiftAmt);
1165 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1166 .addReg(SrlRes).addImm(ShiftImm);
1167 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1168 .addReg(SllRes).addImm(ShiftImm);
1170 MI->eraseFromParent(); // The instruction is gone now.
1175 MachineBasicBlock * MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1176 MachineBasicBlock *BB,
1177 unsigned Size) const {
1178 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1180 MachineFunction *MF = BB->getParent();
1181 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1182 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1183 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1184 DebugLoc DL = MI->getDebugLoc();
1185 unsigned LL, SC, ZERO, BNE, BEQ;
1188 LL = isMicroMips ? Mips::LL_MM : Mips::LL;
1189 SC = isMicroMips ? Mips::SC_MM : Mips::SC;
1196 ZERO = Mips::ZERO_64;
1201 unsigned Dest = MI->getOperand(0).getReg();
1202 unsigned Ptr = MI->getOperand(1).getReg();
1203 unsigned OldVal = MI->getOperand(2).getReg();
1204 unsigned NewVal = MI->getOperand(3).getReg();
1206 unsigned Success = RegInfo.createVirtualRegister(RC);
1208 // insert new blocks after the current block
1209 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1210 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1211 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1212 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1213 MachineFunction::iterator It = BB;
1215 MF->insert(It, loop1MBB);
1216 MF->insert(It, loop2MBB);
1217 MF->insert(It, exitMBB);
1219 // Transfer the remainder of BB and its successor edges to exitMBB.
1220 exitMBB->splice(exitMBB->begin(), BB,
1221 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1222 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1226 // fallthrough --> loop1MBB
1227 BB->addSuccessor(loop1MBB);
1228 loop1MBB->addSuccessor(exitMBB);
1229 loop1MBB->addSuccessor(loop2MBB);
1230 loop2MBB->addSuccessor(loop1MBB);
1231 loop2MBB->addSuccessor(exitMBB);
1235 // bne dest, oldval, exitMBB
1237 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1238 BuildMI(BB, DL, TII->get(BNE))
1239 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1242 // sc success, newval, 0(ptr)
1243 // beq success, $0, loop1MBB
1245 BuildMI(BB, DL, TII->get(SC), Success)
1246 .addReg(NewVal).addReg(Ptr).addImm(0);
1247 BuildMI(BB, DL, TII->get(BEQ))
1248 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1250 MI->eraseFromParent(); // The instruction is gone now.
1256 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1257 MachineBasicBlock *BB,
1258 unsigned Size) const {
1259 assert((Size == 1 || Size == 2) &&
1260 "Unsupported size for EmitAtomicCmpSwapPartial.");
1262 MachineFunction *MF = BB->getParent();
1263 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1264 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1265 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1266 DebugLoc DL = MI->getDebugLoc();
1268 unsigned Dest = MI->getOperand(0).getReg();
1269 unsigned Ptr = MI->getOperand(1).getReg();
1270 unsigned CmpVal = MI->getOperand(2).getReg();
1271 unsigned NewVal = MI->getOperand(3).getReg();
1273 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1274 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1275 unsigned Mask = RegInfo.createVirtualRegister(RC);
1276 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1277 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1278 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1279 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1280 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1281 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1282 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1283 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1284 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1285 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1286 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1287 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1288 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1289 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1290 unsigned Success = RegInfo.createVirtualRegister(RC);
1292 // insert new blocks after the current block
1293 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1294 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1295 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1296 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1297 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1298 MachineFunction::iterator It = BB;
1300 MF->insert(It, loop1MBB);
1301 MF->insert(It, loop2MBB);
1302 MF->insert(It, sinkMBB);
1303 MF->insert(It, exitMBB);
1305 // Transfer the remainder of BB and its successor edges to exitMBB.
1306 exitMBB->splice(exitMBB->begin(), BB,
1307 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1308 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1310 BB->addSuccessor(loop1MBB);
1311 loop1MBB->addSuccessor(sinkMBB);
1312 loop1MBB->addSuccessor(loop2MBB);
1313 loop2MBB->addSuccessor(loop1MBB);
1314 loop2MBB->addSuccessor(sinkMBB);
1315 sinkMBB->addSuccessor(exitMBB);
1317 // FIXME: computation of newval2 can be moved to loop2MBB.
1319 // addiu masklsb2,$0,-4 # 0xfffffffc
1320 // and alignedaddr,ptr,masklsb2
1321 // andi ptrlsb2,ptr,3
1322 // sll shiftamt,ptrlsb2,3
1323 // ori maskupper,$0,255 # 0xff
1324 // sll mask,maskupper,shiftamt
1325 // nor mask2,$0,mask
1326 // andi maskedcmpval,cmpval,255
1327 // sll shiftedcmpval,maskedcmpval,shiftamt
1328 // andi maskednewval,newval,255
1329 // sll shiftednewval,maskednewval,shiftamt
1330 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1331 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1332 .addReg(Mips::ZERO).addImm(-4);
1333 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1334 .addReg(Ptr).addReg(MaskLSB2);
1335 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1336 if (Subtarget->isLittle()) {
1337 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1339 unsigned Off = RegInfo.createVirtualRegister(RC);
1340 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1341 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1342 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1344 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1345 .addReg(Mips::ZERO).addImm(MaskImm);
1346 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1347 .addReg(MaskUpper).addReg(ShiftAmt);
1348 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1349 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1350 .addReg(CmpVal).addImm(MaskImm);
1351 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1352 .addReg(MaskedCmpVal).addReg(ShiftAmt);
1353 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1354 .addReg(NewVal).addImm(MaskImm);
1355 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1356 .addReg(MaskedNewVal).addReg(ShiftAmt);
1359 // ll oldval,0(alginedaddr)
1360 // and maskedoldval0,oldval,mask
1361 // bne maskedoldval0,shiftedcmpval,sinkMBB
1363 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1364 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1365 .addReg(OldVal).addReg(Mask);
1366 BuildMI(BB, DL, TII->get(Mips::BNE))
1367 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1370 // and maskedoldval1,oldval,mask2
1371 // or storeval,maskedoldval1,shiftednewval
1372 // sc success,storeval,0(alignedaddr)
1373 // beq success,$0,loop1MBB
1375 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1376 .addReg(OldVal).addReg(Mask2);
1377 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1378 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1379 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1380 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1381 BuildMI(BB, DL, TII->get(Mips::BEQ))
1382 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1385 // srl srlres,maskedoldval0,shiftamt
1386 // sll sllres,srlres,24
1387 // sra dest,sllres,24
1389 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1391 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1392 .addReg(MaskedOldVal0).addReg(ShiftAmt);
1393 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1394 .addReg(SrlRes).addImm(ShiftImm);
1395 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1396 .addReg(SllRes).addImm(ShiftImm);
1398 MI->eraseFromParent(); // The instruction is gone now.
1403 //===----------------------------------------------------------------------===//
1404 // Misc Lower Operation implementation
1405 //===----------------------------------------------------------------------===//
1406 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1407 SDValue Chain = Op.getOperand(0);
1408 SDValue Table = Op.getOperand(1);
1409 SDValue Index = Op.getOperand(2);
1411 EVT PTy = getPointerTy();
1412 unsigned EntrySize =
1413 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout());
1415 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1416 DAG.getConstant(EntrySize, PTy));
1417 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1419 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1420 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1421 MachinePointerInfo::getJumpTable(), MemVT, false, false,
1423 Chain = Addr.getValue(1);
1425 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || isN64()) {
1426 // For PIC, the sequence is:
1427 // BRIND(load(Jumptable + index) + RelocBase)
1428 // RelocBase can be JumpTable, GOT or some sort of global base.
1429 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1430 getPICJumpTableRelocBase(Table, DAG));
1433 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1436 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1437 // The first operand is the chain, the second is the condition, the third is
1438 // the block to branch to if the condition is true.
1439 SDValue Chain = Op.getOperand(0);
1440 SDValue Dest = Op.getOperand(2);
1443 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1445 // Return if flag is not set by a floating point comparison.
1446 if (CondRes.getOpcode() != MipsISD::FPCmp)
1449 SDValue CCNode = CondRes.getOperand(2);
1451 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1452 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1453 SDValue BrCode = DAG.getConstant(Opc, MVT::i32);
1454 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
1455 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1456 FCC0, Dest, CondRes);
1459 SDValue MipsTargetLowering::
1460 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1462 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1464 // Return if flag is not set by a floating point comparison.
1465 if (Cond.getOpcode() != MipsISD::FPCmp)
1468 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1472 SDValue MipsTargetLowering::
1473 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1476 EVT Ty = Op.getOperand(0).getValueType();
1477 SDValue Cond = DAG.getNode(ISD::SETCC, DL,
1478 getSetCCResultType(*DAG.getContext(), Ty),
1479 Op.getOperand(0), Op.getOperand(1),
1482 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1486 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1487 SDValue Cond = createFPCmp(DAG, Op);
1489 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1490 "Floating point operand expected.");
1492 SDValue True = DAG.getConstant(1, MVT::i32);
1493 SDValue False = DAG.getConstant(0, MVT::i32);
1495 return createCMovFP(DAG, Cond, True, False, SDLoc(Op));
1498 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1499 SelectionDAG &DAG) const {
1500 // FIXME there isn't actually debug info here
1502 EVT Ty = Op.getValueType();
1503 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
1504 const GlobalValue *GV = N->getGlobal();
1506 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !isN64()) {
1507 const MipsTargetObjectFile &TLOF =
1508 (const MipsTargetObjectFile&)getObjFileLowering();
1510 // %gp_rel relocation
1511 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1512 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0,
1514 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL,
1515 DAG.getVTList(MVT::i32), GA);
1516 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1517 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode);
1520 // %hi/%lo relocation
1521 return getAddrNonPIC(N, Ty, DAG);
1524 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1525 return getAddrLocal(N, Ty, DAG, isN32() || isN64());
1528 return getAddrGlobalLargeGOT(N, Ty, DAG, MipsII::MO_GOT_HI16,
1529 MipsII::MO_GOT_LO16, DAG.getEntryNode(),
1530 MachinePointerInfo::getGOT());
1532 return getAddrGlobal(N, Ty, DAG, (isN32() || isN64()) ? MipsII::MO_GOT_DISP
1534 DAG.getEntryNode(), MachinePointerInfo::getGOT());
1537 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1538 SelectionDAG &DAG) const {
1539 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
1540 EVT Ty = Op.getValueType();
1542 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !isN64())
1543 return getAddrNonPIC(N, Ty, DAG);
1545 return getAddrLocal(N, Ty, DAG, isN32() || isN64());
1548 SDValue MipsTargetLowering::
1549 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1551 // If the relocation model is PIC, use the General Dynamic TLS Model or
1552 // Local Dynamic TLS model, otherwise use the Initial Exec or
1553 // Local Exec TLS Model.
1555 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1557 const GlobalValue *GV = GA->getGlobal();
1558 EVT PtrVT = getPointerTy();
1560 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1562 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1563 // General Dynamic and Local Dynamic TLS Model.
1564 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1567 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1568 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1569 getGlobalReg(DAG, PtrVT), TGA);
1570 unsigned PtrSize = PtrVT.getSizeInBits();
1571 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1573 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1577 Entry.Node = Argument;
1579 Args.push_back(Entry);
1581 TargetLowering::CallLoweringInfo CLI(DAG);
1582 CLI.setDebugLoc(DL).setChain(DAG.getEntryNode())
1583 .setCallee(CallingConv::C, PtrTy, TlsGetAddr, &Args, 0);
1584 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1586 SDValue Ret = CallResult.first;
1588 if (model != TLSModel::LocalDynamic)
1591 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1592 MipsII::MO_DTPREL_HI);
1593 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1594 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1595 MipsII::MO_DTPREL_LO);
1596 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1597 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1598 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1602 if (model == TLSModel::InitialExec) {
1603 // Initial Exec TLS Model
1604 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1605 MipsII::MO_GOTTPREL);
1606 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1608 Offset = DAG.getLoad(PtrVT, DL,
1609 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1610 false, false, false, 0);
1612 // Local Exec TLS Model
1613 assert(model == TLSModel::LocalExec);
1614 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1615 MipsII::MO_TPREL_HI);
1616 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1617 MipsII::MO_TPREL_LO);
1618 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1619 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1620 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1623 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1624 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1627 SDValue MipsTargetLowering::
1628 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1630 JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
1631 EVT Ty = Op.getValueType();
1633 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !isN64())
1634 return getAddrNonPIC(N, Ty, DAG);
1636 return getAddrLocal(N, Ty, DAG, isN32() || isN64());
1639 SDValue MipsTargetLowering::
1640 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1642 // gp_rel relocation
1643 // FIXME: we should reference the constant pool using small data sections,
1644 // but the asm printer currently doesn't support this feature without
1645 // hacking it. This feature should come soon so we can uncomment the
1647 //if (IsInSmallSection(C->getType())) {
1648 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1649 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1650 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1651 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1652 EVT Ty = Op.getValueType();
1654 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !isN64())
1655 return getAddrNonPIC(N, Ty, DAG);
1657 return getAddrLocal(N, Ty, DAG, isN32() || isN64());
1660 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1661 MachineFunction &MF = DAG.getMachineFunction();
1662 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1665 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1668 // vastart just stores the address of the VarArgsFrameIndex slot into the
1669 // memory location argument.
1670 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1671 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1672 MachinePointerInfo(SV), false, false, 0);
1675 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG,
1676 bool HasExtractInsert) {
1677 EVT TyX = Op.getOperand(0).getValueType();
1678 EVT TyY = Op.getOperand(1).getValueType();
1679 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1680 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1684 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1686 SDValue X = (TyX == MVT::f32) ?
1687 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1688 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1690 SDValue Y = (TyY == MVT::f32) ?
1691 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1692 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1695 if (HasExtractInsert) {
1696 // ext E, Y, 31, 1 ; extract bit31 of Y
1697 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1698 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1699 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1702 // srl SrlX, SllX, 1
1704 // sll SllY, SrlX, 31
1705 // or Or, SrlX, SllY
1706 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1707 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1708 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1709 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1710 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1713 if (TyX == MVT::f32)
1714 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1716 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1717 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1718 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1721 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG,
1722 bool HasExtractInsert) {
1723 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1724 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1725 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1726 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1729 // Bitcast to integer nodes.
1730 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1731 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1733 if (HasExtractInsert) {
1734 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1735 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1736 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1737 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
1739 if (WidthX > WidthY)
1740 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1741 else if (WidthY > WidthX)
1742 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1744 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1745 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
1746 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1749 // (d)sll SllX, X, 1
1750 // (d)srl SrlX, SllX, 1
1751 // (d)srl SrlY, Y, width(Y)-1
1752 // (d)sll SllY, SrlX, width(Y)-1
1753 // or Or, SrlX, SllY
1754 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1755 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1756 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1757 DAG.getConstant(WidthY - 1, MVT::i32));
1759 if (WidthX > WidthY)
1760 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1761 else if (WidthY > WidthX)
1762 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1764 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1765 DAG.getConstant(WidthX - 1, MVT::i32));
1766 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
1767 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
1771 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1772 if (Subtarget->isGP64bit())
1773 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasExtractInsert());
1775 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasExtractInsert());
1778 SDValue MipsTargetLowering::
1779 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1781 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1782 "Frame address can only be determined for current frame.");
1784 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1785 MFI->setFrameAddressIsTaken(true);
1786 EVT VT = Op.getValueType();
1788 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1789 isN64() ? Mips::FP_64 : Mips::FP, VT);
1793 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
1794 SelectionDAG &DAG) const {
1795 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
1799 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1800 "Return address can be determined only for current frame.");
1802 MachineFunction &MF = DAG.getMachineFunction();
1803 MachineFrameInfo *MFI = MF.getFrameInfo();
1804 MVT VT = Op.getSimpleValueType();
1805 unsigned RA = isN64() ? Mips::RA_64 : Mips::RA;
1806 MFI->setReturnAddressIsTaken(true);
1808 // Return RA, which contains the return address. Mark it an implicit live-in.
1809 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
1810 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
1813 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
1814 // generated from __builtin_eh_return (offset, handler)
1815 // The effect of this is to adjust the stack pointer by "offset"
1816 // and then branch to "handler".
1817 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
1819 MachineFunction &MF = DAG.getMachineFunction();
1820 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1822 MipsFI->setCallsEhReturn();
1823 SDValue Chain = Op.getOperand(0);
1824 SDValue Offset = Op.getOperand(1);
1825 SDValue Handler = Op.getOperand(2);
1827 EVT Ty = isN64() ? MVT::i64 : MVT::i32;
1829 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
1830 // EH_RETURN nodes, so that instructions are emitted back-to-back.
1831 unsigned OffsetReg = isN64() ? Mips::V1_64 : Mips::V1;
1832 unsigned AddrReg = isN64() ? Mips::V0_64 : Mips::V0;
1833 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
1834 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
1835 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
1836 DAG.getRegister(OffsetReg, Ty),
1837 DAG.getRegister(AddrReg, getPointerTy()),
1841 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
1842 SelectionDAG &DAG) const {
1843 // FIXME: Need pseudo-fence for 'singlethread' fences
1844 // FIXME: Set SType for weaker fences where supported/appropriate.
1847 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
1848 DAG.getConstant(SType, MVT::i32));
1851 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
1852 SelectionDAG &DAG) const {
1854 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1855 SDValue Shamt = Op.getOperand(2);
1858 // lo = (shl lo, shamt)
1859 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
1862 // hi = (shl lo, shamt[4:0])
1863 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1864 DAG.getConstant(-1, MVT::i32));
1865 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
1866 DAG.getConstant(1, MVT::i32));
1867 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
1869 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
1870 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1871 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
1872 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1873 DAG.getConstant(0x20, MVT::i32));
1874 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1875 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
1876 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
1878 SDValue Ops[2] = {Lo, Hi};
1879 return DAG.getMergeValues(Ops, DL);
1882 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
1885 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1886 SDValue Shamt = Op.getOperand(2);
1889 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
1891 // hi = (sra hi, shamt)
1893 // hi = (srl hi, shamt)
1896 // lo = (sra hi, shamt[4:0])
1897 // hi = (sra hi, 31)
1899 // lo = (srl hi, shamt[4:0])
1901 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1902 DAG.getConstant(-1, MVT::i32));
1903 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
1904 DAG.getConstant(1, MVT::i32));
1905 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
1906 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
1907 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1908 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1910 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1911 DAG.getConstant(0x20, MVT::i32));
1912 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
1913 DAG.getConstant(31, MVT::i32));
1914 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
1915 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1916 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
1919 SDValue Ops[2] = {Lo, Hi};
1920 return DAG.getMergeValues(Ops, DL);
1923 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
1924 SDValue Chain, SDValue Src, unsigned Offset) {
1925 SDValue Ptr = LD->getBasePtr();
1926 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
1927 EVT BasePtrVT = Ptr.getValueType();
1929 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1932 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
1933 DAG.getConstant(Offset, BasePtrVT));
1935 SDValue Ops[] = { Chain, Ptr, Src };
1936 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
1937 LD->getMemOperand());
1940 // Expand an unaligned 32 or 64-bit integer load node.
1941 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1942 LoadSDNode *LD = cast<LoadSDNode>(Op);
1943 EVT MemVT = LD->getMemoryVT();
1945 if (Subtarget->systemSupportsUnalignedAccess())
1948 // Return if load is aligned or if MemVT is neither i32 nor i64.
1949 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
1950 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
1953 bool IsLittle = Subtarget->isLittle();
1954 EVT VT = Op.getValueType();
1955 ISD::LoadExtType ExtType = LD->getExtensionType();
1956 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
1958 assert((VT == MVT::i32) || (VT == MVT::i64));
1961 // (set dst, (i64 (load baseptr)))
1963 // (set tmp, (ldl (add baseptr, 7), undef))
1964 // (set dst, (ldr baseptr, tmp))
1965 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
1966 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
1968 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
1972 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
1974 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
1978 // (set dst, (i32 (load baseptr))) or
1979 // (set dst, (i64 (sextload baseptr))) or
1980 // (set dst, (i64 (extload baseptr)))
1982 // (set tmp, (lwl (add baseptr, 3), undef))
1983 // (set dst, (lwr baseptr, tmp))
1984 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
1985 (ExtType == ISD::EXTLOAD))
1988 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
1991 // (set dst, (i64 (zextload baseptr)))
1993 // (set tmp0, (lwl (add baseptr, 3), undef))
1994 // (set tmp1, (lwr baseptr, tmp0))
1995 // (set tmp2, (shl tmp1, 32))
1996 // (set dst, (srl tmp2, 32))
1998 SDValue Const32 = DAG.getConstant(32, MVT::i32);
1999 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2000 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2001 SDValue Ops[] = { SRL, LWR.getValue(1) };
2002 return DAG.getMergeValues(Ops, DL);
2005 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2006 SDValue Chain, unsigned Offset) {
2007 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2008 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2010 SDVTList VTList = DAG.getVTList(MVT::Other);
2013 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2014 DAG.getConstant(Offset, BasePtrVT));
2016 SDValue Ops[] = { Chain, Value, Ptr };
2017 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2018 SD->getMemOperand());
2021 // Expand an unaligned 32 or 64-bit integer store node.
2022 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2024 SDValue Value = SD->getValue(), Chain = SD->getChain();
2025 EVT VT = Value.getValueType();
2028 // (store val, baseptr) or
2029 // (truncstore val, baseptr)
2031 // (swl val, (add baseptr, 3))
2032 // (swr val, baseptr)
2033 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2034 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2036 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2039 assert(VT == MVT::i64);
2042 // (store val, baseptr)
2044 // (sdl val, (add baseptr, 7))
2045 // (sdr val, baseptr)
2046 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2047 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2050 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
2051 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) {
2052 SDValue Val = SD->getValue();
2054 if (Val.getOpcode() != ISD::FP_TO_SINT)
2057 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2058 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2061 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2062 SD->getPointerInfo(), SD->isVolatile(),
2063 SD->isNonTemporal(), SD->getAlignment());
2066 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2067 StoreSDNode *SD = cast<StoreSDNode>(Op);
2068 EVT MemVT = SD->getMemoryVT();
2070 // Lower unaligned integer stores.
2071 if (!Subtarget->systemSupportsUnalignedAccess() &&
2072 (SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2073 ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2074 return lowerUnalignedIntStore(SD, DAG, Subtarget->isLittle());
2076 return lowerFP_TO_SINT_STORE(SD, DAG);
2079 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2080 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2081 || cast<ConstantSDNode>
2082 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2083 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2087 // (add (frameaddr 0), (frame_to_args_offset))
2088 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2089 // (add FrameObject, 0)
2090 // where FrameObject is a fixed StackObject with offset 0 which points to
2091 // the old stack pointer.
2092 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2093 EVT ValTy = Op->getValueType(0);
2094 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2095 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2096 return DAG.getNode(ISD::ADD, SDLoc(Op), ValTy, InArgsAddr,
2097 DAG.getConstant(0, ValTy));
2100 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2101 SelectionDAG &DAG) const {
2102 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2103 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2105 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2108 //===----------------------------------------------------------------------===//
2109 // Calling Convention Implementation
2110 //===----------------------------------------------------------------------===//
2112 //===----------------------------------------------------------------------===//
2113 // TODO: Implement a generic logic using tblgen that can support this.
2114 // Mips O32 ABI rules:
2116 // i32 - Passed in A0, A1, A2, A3 and stack
2117 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2118 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2119 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2120 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2121 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2124 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2125 //===----------------------------------------------------------------------===//
2127 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2128 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2129 CCState &State, const MCPhysReg *F64Regs) {
2131 static const unsigned IntRegsSize = 4, FloatRegsSize = 2;
2133 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 };
2134 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 };
2136 // Do not process byval args here.
2137 if (ArgFlags.isByVal())
2140 // Promote i8 and i16
2141 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2143 if (ArgFlags.isSExt())
2144 LocInfo = CCValAssign::SExt;
2145 else if (ArgFlags.isZExt())
2146 LocInfo = CCValAssign::ZExt;
2148 LocInfo = CCValAssign::AExt;
2153 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2154 // is true: function is vararg, argument is 3rd or higher, there is previous
2155 // argument which is not f32 or f64.
2156 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2157 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2158 unsigned OrigAlign = ArgFlags.getOrigAlign();
2159 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2161 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2162 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2163 // If this is the first part of an i64 arg,
2164 // the allocated register must be either A0 or A2.
2165 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2166 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2168 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2169 // Allocate int register and shadow next int register. If first
2170 // available register is Mips::A1 or Mips::A3, shadow it too.
2171 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2172 if (Reg == Mips::A1 || Reg == Mips::A3)
2173 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2174 State.AllocateReg(IntRegs, IntRegsSize);
2176 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2177 // we are guaranteed to find an available float register
2178 if (ValVT == MVT::f32) {
2179 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2180 // Shadow int register
2181 State.AllocateReg(IntRegs, IntRegsSize);
2183 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2184 // Shadow int registers
2185 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2186 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2187 State.AllocateReg(IntRegs, IntRegsSize);
2188 State.AllocateReg(IntRegs, IntRegsSize);
2191 llvm_unreachable("Cannot handle this ValVT.");
2194 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2196 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2198 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2203 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2204 MVT LocVT, CCValAssign::LocInfo LocInfo,
2205 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2206 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 };
2208 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2211 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2212 MVT LocVT, CCValAssign::LocInfo LocInfo,
2213 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2214 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 };
2216 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2219 #include "MipsGenCallingConv.inc"
2221 //===----------------------------------------------------------------------===//
2222 // Call Calling Convention Implementation
2223 //===----------------------------------------------------------------------===//
2225 // Return next O32 integer argument register.
2226 static unsigned getNextIntArgReg(unsigned Reg) {
2227 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2228 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2232 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2233 SDValue Chain, SDValue Arg, SDLoc DL,
2234 bool IsTailCall, SelectionDAG &DAG) const {
2236 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2237 DAG.getIntPtrConstant(Offset));
2238 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2242 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2243 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2244 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2245 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2246 /*isVolatile=*/ true, false, 0);
2249 void MipsTargetLowering::
2250 getOpndList(SmallVectorImpl<SDValue> &Ops,
2251 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2252 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2253 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
2254 // Insert node "GP copy globalreg" before call to function.
2256 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2257 // in PIC mode) allow symbols to be resolved via lazy binding.
2258 // The lazy binding stub requires GP to point to the GOT.
2259 if (IsPICCall && !InternalLinkage) {
2260 unsigned GPReg = isN64() ? Mips::GP_64 : Mips::GP;
2261 EVT Ty = isN64() ? MVT::i64 : MVT::i32;
2262 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2265 // Build a sequence of copy-to-reg nodes chained together with token
2266 // chain and flag operands which copy the outgoing args into registers.
2267 // The InFlag in necessary since all emitted instructions must be
2271 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2272 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2273 RegsToPass[i].second, InFlag);
2274 InFlag = Chain.getValue(1);
2277 // Add argument registers to the end of the list so that they are
2278 // known live into the call.
2279 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2280 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2281 RegsToPass[i].second.getValueType()));
2283 // Add a register mask operand representing the call-preserved registers.
2284 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2285 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv);
2286 assert(Mask && "Missing call preserved mask for calling convention");
2287 if (Subtarget->inMips16HardFloat()) {
2288 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2289 llvm::StringRef Sym = G->getGlobal()->getName();
2290 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2291 if (F && F->hasFnAttribute("__Mips16RetHelper")) {
2292 Mask = MipsRegisterInfo::getMips16RetHelperMask();
2296 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2298 if (InFlag.getNode())
2299 Ops.push_back(InFlag);
2302 /// LowerCall - functions arguments are copied from virtual regs to
2303 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2305 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2306 SmallVectorImpl<SDValue> &InVals) const {
2307 SelectionDAG &DAG = CLI.DAG;
2309 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2310 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2311 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2312 SDValue Chain = CLI.Chain;
2313 SDValue Callee = CLI.Callee;
2314 bool &IsTailCall = CLI.IsTailCall;
2315 CallingConv::ID CallConv = CLI.CallConv;
2316 bool IsVarArg = CLI.IsVarArg;
2318 MachineFunction &MF = DAG.getMachineFunction();
2319 MachineFrameInfo *MFI = MF.getFrameInfo();
2320 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2321 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2322 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2324 // Analyze operands of the call, assigning locations to each operand.
2325 SmallVector<CCValAssign, 16> ArgLocs;
2326 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2327 getTargetMachine(), ArgLocs, *DAG.getContext());
2328 MipsCC::SpecialCallingConvType SpecialCallingConv =
2329 getSpecialCallingConv(Callee);
2330 MipsCC MipsCCInfo(CallConv, isO32(), Subtarget->isFP64bit(), CCInfo,
2331 SpecialCallingConv);
2333 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg,
2334 Subtarget->mipsSEUsesSoftFloat(),
2335 Callee.getNode(), CLI.getArgs());
2337 // Get a count of how many bytes are to be pushed on the stack.
2338 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2340 // Check if it's really possible to do a tail call.
2343 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2344 *MF.getInfo<MipsFunctionInfo>());
2346 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2347 report_fatal_error("failed to perform tail call elimination on a call "
2348 "site marked musttail");
2353 // Chain is the output chain of the last Load/Store or CopyToReg node.
2354 // ByValChain is the output chain of the last Memcpy node created for copying
2355 // byval arguments to the stack.
2356 unsigned StackAlignment = TFL->getStackAlignment();
2357 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2358 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2361 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL);
2363 SDValue StackPtr = DAG.getCopyFromReg(
2364 Chain, DL, isN64() ? Mips::SP_64 : Mips::SP, getPointerTy());
2366 // With EABI is it possible to have 16 args on registers.
2367 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2368 SmallVector<SDValue, 8> MemOpChains;
2369 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2371 // Walk the register/memloc assignments, inserting copies/loads.
2372 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2373 SDValue Arg = OutVals[i];
2374 CCValAssign &VA = ArgLocs[i];
2375 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2376 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2379 if (Flags.isByVal()) {
2380 assert(Flags.getByValSize() &&
2381 "ByVal args of size 0 should have been ignored by front-end.");
2382 assert(ByValArg != MipsCCInfo.byval_end());
2383 assert(!IsTailCall &&
2384 "Do not tail-call optimize if there is a byval argument.");
2385 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2386 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2391 // Promote the value if needed.
2392 switch (VA.getLocInfo()) {
2393 default: llvm_unreachable("Unknown loc info!");
2394 case CCValAssign::Full:
2395 if (VA.isRegLoc()) {
2396 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2397 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2398 (ValVT == MVT::i64 && LocVT == MVT::f64))
2399 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2400 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2401 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2402 Arg, DAG.getConstant(0, MVT::i32));
2403 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2404 Arg, DAG.getConstant(1, MVT::i32));
2405 if (!Subtarget->isLittle())
2407 unsigned LocRegLo = VA.getLocReg();
2408 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2409 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2410 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2415 case CCValAssign::SExt:
2416 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2418 case CCValAssign::ZExt:
2419 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2421 case CCValAssign::AExt:
2422 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2426 // Arguments that can be passed on register must be kept at
2427 // RegsToPass vector
2428 if (VA.isRegLoc()) {
2429 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2433 // Register can't get to this point...
2434 assert(VA.isMemLoc());
2436 // emit ISD::STORE whichs stores the
2437 // parameter value to a stack Location
2438 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2439 Chain, Arg, DL, IsTailCall, DAG));
2442 // Transform all store nodes into one single node because all store
2443 // nodes are independent of each other.
2444 if (!MemOpChains.empty())
2445 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2447 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2448 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2449 // node so that legalize doesn't hack it.
2450 bool IsPICCall = (isN64() || IsPIC); // true if calls are translated to
2452 bool GlobalOrExternal = false, InternalLinkage = false;
2454 EVT Ty = Callee.getValueType();
2456 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2458 const GlobalValue *Val = G->getGlobal();
2459 InternalLinkage = Val->hasInternalLinkage();
2461 if (InternalLinkage)
2462 Callee = getAddrLocal(G, Ty, DAG, isN32() || isN64());
2464 Callee = getAddrGlobalLargeGOT(G, Ty, DAG, MipsII::MO_CALL_HI16,
2465 MipsII::MO_CALL_LO16, Chain,
2466 FuncInfo->callPtrInfo(Val));
2468 Callee = getAddrGlobal(G, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2469 FuncInfo->callPtrInfo(Val));
2471 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0,
2472 MipsII::MO_NO_FLAG);
2473 GlobalOrExternal = true;
2475 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2476 const char *Sym = S->getSymbol();
2478 if (!isN64() && !IsPIC) // !N64 && static
2479 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(),
2480 MipsII::MO_NO_FLAG);
2482 Callee = getAddrGlobalLargeGOT(S, Ty, DAG, MipsII::MO_CALL_HI16,
2483 MipsII::MO_CALL_LO16, Chain,
2484 FuncInfo->callPtrInfo(Sym));
2486 Callee = getAddrGlobal(S, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2487 FuncInfo->callPtrInfo(Sym));
2489 GlobalOrExternal = true;
2492 SmallVector<SDValue, 8> Ops(1, Chain);
2493 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2495 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2496 CLI, Callee, Chain);
2499 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops);
2501 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops);
2502 SDValue InFlag = Chain.getValue(1);
2504 // Create the CALLSEQ_END node.
2505 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2506 DAG.getIntPtrConstant(0, true), InFlag, DL);
2507 InFlag = Chain.getValue(1);
2509 // Handle result values, copying them out of physregs into vregs that we
2511 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg,
2512 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy);
2515 /// LowerCallResult - Lower the result values of a call into the
2516 /// appropriate copies out of appropriate physical registers.
2518 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2519 CallingConv::ID CallConv, bool IsVarArg,
2520 const SmallVectorImpl<ISD::InputArg> &Ins,
2521 SDLoc DL, SelectionDAG &DAG,
2522 SmallVectorImpl<SDValue> &InVals,
2523 const SDNode *CallNode,
2524 const Type *RetTy) const {
2525 // Assign locations to each value returned by this call.
2526 SmallVector<CCValAssign, 16> RVLocs;
2527 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2528 getTargetMachine(), RVLocs, *DAG.getContext());
2529 MipsCC MipsCCInfo(CallConv, isO32(), Subtarget->isFP64bit(), CCInfo);
2531 MipsCCInfo.analyzeCallResult(Ins, Subtarget->mipsSEUsesSoftFloat(),
2534 // Copy all of the result registers out of their specified physreg.
2535 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2536 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2537 RVLocs[i].getLocVT(), InFlag);
2538 Chain = Val.getValue(1);
2539 InFlag = Val.getValue(2);
2541 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2542 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val);
2544 InVals.push_back(Val);
2550 //===----------------------------------------------------------------------===//
2551 // Formal Arguments Calling Convention Implementation
2552 //===----------------------------------------------------------------------===//
2553 /// LowerFormalArguments - transform physical registers into virtual registers
2554 /// and generate load operations for arguments places on the stack.
2556 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2557 CallingConv::ID CallConv,
2559 const SmallVectorImpl<ISD::InputArg> &Ins,
2560 SDLoc DL, SelectionDAG &DAG,
2561 SmallVectorImpl<SDValue> &InVals)
2563 MachineFunction &MF = DAG.getMachineFunction();
2564 MachineFrameInfo *MFI = MF.getFrameInfo();
2565 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2567 MipsFI->setVarArgsFrameIndex(0);
2569 // Used with vargs to acumulate store chains.
2570 std::vector<SDValue> OutChains;
2572 // Assign locations to all of the incoming arguments.
2573 SmallVector<CCValAssign, 16> ArgLocs;
2574 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2575 getTargetMachine(), ArgLocs, *DAG.getContext());
2576 MipsCC MipsCCInfo(CallConv, isO32(), Subtarget->isFP64bit(), CCInfo);
2577 Function::const_arg_iterator FuncArg =
2578 DAG.getMachineFunction().getFunction()->arg_begin();
2579 bool UseSoftFloat = Subtarget->mipsSEUsesSoftFloat();
2581 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg);
2582 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2583 MipsCCInfo.hasByValArg());
2585 unsigned CurArgIdx = 0;
2586 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2588 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2589 CCValAssign &VA = ArgLocs[i];
2590 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
2591 CurArgIdx = Ins[i].OrigArgIndex;
2592 EVT ValVT = VA.getValVT();
2593 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2594 bool IsRegLoc = VA.isRegLoc();
2596 if (Flags.isByVal()) {
2597 assert(Flags.getByValSize() &&
2598 "ByVal args of size 0 should have been ignored by front-end.");
2599 assert(ByValArg != MipsCCInfo.byval_end());
2600 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2601 MipsCCInfo, *ByValArg);
2606 // Arguments stored on registers
2608 MVT RegVT = VA.getLocVT();
2609 unsigned ArgReg = VA.getLocReg();
2610 const TargetRegisterClass *RC = getRegClassFor(RegVT);
2612 // Transform the arguments stored on
2613 // physical registers into virtual ones
2614 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2615 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2617 // If this is an 8 or 16-bit value, it has been passed promoted
2618 // to 32 bits. Insert an assert[sz]ext to capture this, then
2619 // truncate to the right size.
2620 if (VA.getLocInfo() != CCValAssign::Full) {
2621 unsigned Opcode = 0;
2622 if (VA.getLocInfo() == CCValAssign::SExt)
2623 Opcode = ISD::AssertSext;
2624 else if (VA.getLocInfo() == CCValAssign::ZExt)
2625 Opcode = ISD::AssertZext;
2627 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue,
2628 DAG.getValueType(ValVT));
2629 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue);
2632 // Handle floating point arguments passed in integer registers and
2633 // long double arguments passed in floating point registers.
2634 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2635 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
2636 (RegVT == MVT::f64 && ValVT == MVT::i64))
2637 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
2638 else if (isO32() && RegVT == MVT::i32 && ValVT == MVT::f64) {
2639 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
2640 getNextIntArgReg(ArgReg), RC);
2641 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
2642 if (!Subtarget->isLittle())
2643 std::swap(ArgValue, ArgValue2);
2644 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
2645 ArgValue, ArgValue2);
2648 InVals.push_back(ArgValue);
2649 } else { // VA.isRegLoc()
2652 assert(VA.isMemLoc());
2654 // The stack pointer offset is relative to the caller stack frame.
2655 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2656 VA.getLocMemOffset(), true);
2658 // Create load nodes to retrieve arguments from the stack
2659 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2660 SDValue Load = DAG.getLoad(ValVT, DL, Chain, FIN,
2661 MachinePointerInfo::getFixedStack(FI),
2662 false, false, false, 0);
2663 InVals.push_back(Load);
2664 OutChains.push_back(Load.getValue(1));
2668 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2669 // The mips ABIs for returning structs by value requires that we copy
2670 // the sret argument into $v0 for the return. Save the argument into
2671 // a virtual register so that we can access it from the return points.
2672 if (Ins[i].Flags.isSRet()) {
2673 unsigned Reg = MipsFI->getSRetReturnReg();
2675 Reg = MF.getRegInfo().createVirtualRegister(
2676 getRegClassFor(isN64() ? MVT::i64 : MVT::i32));
2677 MipsFI->setSRetReturnReg(Reg);
2679 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]);
2680 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
2686 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG);
2688 // All stores are grouped in one node to allow the matching between
2689 // the size of Ins and InVals. This only happens when on varg functions
2690 if (!OutChains.empty()) {
2691 OutChains.push_back(Chain);
2692 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
2698 //===----------------------------------------------------------------------===//
2699 // Return Value Calling Convention Implementation
2700 //===----------------------------------------------------------------------===//
2703 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2704 MachineFunction &MF, bool IsVarArg,
2705 const SmallVectorImpl<ISD::OutputArg> &Outs,
2706 LLVMContext &Context) const {
2707 SmallVector<CCValAssign, 16> RVLocs;
2708 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(),
2710 return CCInfo.CheckReturn(Outs, RetCC_Mips);
2714 MipsTargetLowering::LowerReturn(SDValue Chain,
2715 CallingConv::ID CallConv, bool IsVarArg,
2716 const SmallVectorImpl<ISD::OutputArg> &Outs,
2717 const SmallVectorImpl<SDValue> &OutVals,
2718 SDLoc DL, SelectionDAG &DAG) const {
2719 // CCValAssign - represent the assignment of
2720 // the return value to a location
2721 SmallVector<CCValAssign, 16> RVLocs;
2722 MachineFunction &MF = DAG.getMachineFunction();
2724 // CCState - Info about the registers and stack slot.
2725 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs,
2727 MipsCC MipsCCInfo(CallConv, isO32(), Subtarget->isFP64bit(), CCInfo);
2729 // Analyze return values.
2730 MipsCCInfo.analyzeReturn(Outs, Subtarget->mipsSEUsesSoftFloat(),
2731 MF.getFunction()->getReturnType());
2734 SmallVector<SDValue, 4> RetOps(1, Chain);
2736 // Copy the result values into the output registers.
2737 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2738 SDValue Val = OutVals[i];
2739 CCValAssign &VA = RVLocs[i];
2740 assert(VA.isRegLoc() && "Can only return in registers!");
2742 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2743 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val);
2745 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
2747 // Guarantee that all emitted copies are stuck together with flags.
2748 Flag = Chain.getValue(1);
2749 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2752 // The mips ABIs for returning structs by value requires that we copy
2753 // the sret argument into $v0 for the return. We saved the argument into
2754 // a virtual register in the entry block, so now we copy the value out
2756 if (MF.getFunction()->hasStructRetAttr()) {
2757 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2758 unsigned Reg = MipsFI->getSRetReturnReg();
2761 llvm_unreachable("sret virtual register not created in the entry block");
2762 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
2763 unsigned V0 = isN64() ? Mips::V0_64 : Mips::V0;
2765 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
2766 Flag = Chain.getValue(1);
2767 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
2770 RetOps[0] = Chain; // Update chain.
2772 // Add the flag if we have it.
2774 RetOps.push_back(Flag);
2776 // Return on Mips is always a "jr $ra"
2777 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps);
2780 //===----------------------------------------------------------------------===//
2781 // Mips Inline Assembly Support
2782 //===----------------------------------------------------------------------===//
2784 /// getConstraintType - Given a constraint letter, return the type of
2785 /// constraint it is for this target.
2786 MipsTargetLowering::ConstraintType MipsTargetLowering::
2787 getConstraintType(const std::string &Constraint) const
2789 // Mips specific constraints
2790 // GCC config/mips/constraints.md
2792 // 'd' : An address register. Equivalent to r
2793 // unless generating MIPS16 code.
2794 // 'y' : Equivalent to r; retained for
2795 // backwards compatibility.
2796 // 'c' : A register suitable for use in an indirect
2797 // jump. This will always be $25 for -mabicalls.
2798 // 'l' : The lo register. 1 word storage.
2799 // 'x' : The hilo register pair. Double word storage.
2800 if (Constraint.size() == 1) {
2801 switch (Constraint[0]) {
2809 return C_RegisterClass;
2814 return TargetLowering::getConstraintType(Constraint);
2817 /// Examine constraint type and operand type and determine a weight value.
2818 /// This object must already have been set up with the operand type
2819 /// and the current alternative constraint selected.
2820 TargetLowering::ConstraintWeight
2821 MipsTargetLowering::getSingleConstraintMatchWeight(
2822 AsmOperandInfo &info, const char *constraint) const {
2823 ConstraintWeight weight = CW_Invalid;
2824 Value *CallOperandVal = info.CallOperandVal;
2825 // If we don't have a value, we can't do a match,
2826 // but allow it at the lowest weight.
2827 if (!CallOperandVal)
2829 Type *type = CallOperandVal->getType();
2830 // Look at the constraint type.
2831 switch (*constraint) {
2833 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2837 if (type->isIntegerTy())
2838 weight = CW_Register;
2840 case 'f': // FPU or MSA register
2841 if (Subtarget->hasMSA() && type->isVectorTy() &&
2842 cast<VectorType>(type)->getBitWidth() == 128)
2843 weight = CW_Register;
2844 else if (type->isFloatTy())
2845 weight = CW_Register;
2847 case 'c': // $25 for indirect jumps
2848 case 'l': // lo register
2849 case 'x': // hilo register pair
2850 if (type->isIntegerTy())
2851 weight = CW_SpecificReg;
2853 case 'I': // signed 16 bit immediate
2854 case 'J': // integer zero
2855 case 'K': // unsigned 16 bit immediate
2856 case 'L': // signed 32 bit immediate where lower 16 bits are 0
2857 case 'N': // immediate in the range of -65535 to -1 (inclusive)
2858 case 'O': // signed 15 bit immediate (+- 16383)
2859 case 'P': // immediate in the range of 65535 to 1 (inclusive)
2860 if (isa<ConstantInt>(CallOperandVal))
2861 weight = CW_Constant;
2870 /// This is a helper function to parse a physical register string and split it
2871 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
2872 /// that is returned indicates whether parsing was successful. The second flag
2873 /// is true if the numeric part exists.
2874 static std::pair<bool, bool>
2875 parsePhysicalReg(const StringRef &C, std::string &Prefix,
2876 unsigned long long &Reg) {
2877 if (C.front() != '{' || C.back() != '}')
2878 return std::make_pair(false, false);
2880 // Search for the first numeric character.
2881 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
2882 I = std::find_if(B, E, std::ptr_fun(isdigit));
2884 Prefix.assign(B, I - B);
2886 // The second flag is set to false if no numeric characters were found.
2888 return std::make_pair(true, false);
2890 // Parse the numeric characters.
2891 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
2895 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
2896 parseRegForInlineAsmConstraint(const StringRef &C, MVT VT) const {
2897 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2898 const TargetRegisterClass *RC;
2900 unsigned long long Reg;
2902 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
2905 return std::make_pair(0U, nullptr);
2907 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
2908 // No numeric characters follow "hi" or "lo".
2910 return std::make_pair(0U, nullptr);
2912 RC = TRI->getRegClass(Prefix == "hi" ?
2913 Mips::HI32RegClassID : Mips::LO32RegClassID);
2914 return std::make_pair(*(RC->begin()), RC);
2915 } else if (Prefix.compare(0, 4, "$msa") == 0) {
2916 // Parse $msa(ir|csr|access|save|modify|request|map|unmap)
2918 // No numeric characters follow the name.
2920 return std::make_pair(0U, nullptr);
2922 Reg = StringSwitch<unsigned long long>(Prefix)
2923 .Case("$msair", Mips::MSAIR)
2924 .Case("$msacsr", Mips::MSACSR)
2925 .Case("$msaaccess", Mips::MSAAccess)
2926 .Case("$msasave", Mips::MSASave)
2927 .Case("$msamodify", Mips::MSAModify)
2928 .Case("$msarequest", Mips::MSARequest)
2929 .Case("$msamap", Mips::MSAMap)
2930 .Case("$msaunmap", Mips::MSAUnmap)
2934 return std::make_pair(0U, nullptr);
2936 RC = TRI->getRegClass(Mips::MSACtrlRegClassID);
2937 return std::make_pair(Reg, RC);
2941 return std::make_pair(0U, nullptr);
2943 if (Prefix == "$f") { // Parse $f0-$f31.
2944 // If the size of FP registers is 64-bit or Reg is an even number, select
2945 // the 64-bit register class. Otherwise, select the 32-bit register class.
2946 if (VT == MVT::Other)
2947 VT = (Subtarget->isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
2949 RC = getRegClassFor(VT);
2951 if (RC == &Mips::AFGR64RegClass) {
2952 assert(Reg % 2 == 0);
2955 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7.
2956 RC = TRI->getRegClass(Mips::FCCRegClassID);
2957 else if (Prefix == "$w") { // Parse $w0-$w31.
2958 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT);
2959 } else { // Parse $0-$31.
2960 assert(Prefix == "$");
2961 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
2964 assert(Reg < RC->getNumRegs());
2965 return std::make_pair(*(RC->begin() + Reg), RC);
2968 /// Given a register class constraint, like 'r', if this corresponds directly
2969 /// to an LLVM register class, return a register of 0 and the register class
2971 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2972 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const
2974 if (Constraint.size() == 1) {
2975 switch (Constraint[0]) {
2976 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2977 case 'y': // Same as 'r'. Exists for compatibility.
2979 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
2980 if (Subtarget->inMips16Mode())
2981 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
2982 return std::make_pair(0U, &Mips::GPR32RegClass);
2984 if (VT == MVT::i64 && !isGP64bit())
2985 return std::make_pair(0U, &Mips::GPR32RegClass);
2986 if (VT == MVT::i64 && isGP64bit())
2987 return std::make_pair(0U, &Mips::GPR64RegClass);
2988 // This will generate an error message
2989 return std::make_pair(0U, nullptr);
2990 case 'f': // FPU or MSA register
2991 if (VT == MVT::v16i8)
2992 return std::make_pair(0U, &Mips::MSA128BRegClass);
2993 else if (VT == MVT::v8i16 || VT == MVT::v8f16)
2994 return std::make_pair(0U, &Mips::MSA128HRegClass);
2995 else if (VT == MVT::v4i32 || VT == MVT::v4f32)
2996 return std::make_pair(0U, &Mips::MSA128WRegClass);
2997 else if (VT == MVT::v2i64 || VT == MVT::v2f64)
2998 return std::make_pair(0U, &Mips::MSA128DRegClass);
2999 else if (VT == MVT::f32)
3000 return std::make_pair(0U, &Mips::FGR32RegClass);
3001 else if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
3002 if (Subtarget->isFP64bit())
3003 return std::make_pair(0U, &Mips::FGR64RegClass);
3004 return std::make_pair(0U, &Mips::AFGR64RegClass);
3007 case 'c': // register suitable for indirect jump
3009 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
3010 assert(VT == MVT::i64 && "Unexpected type.");
3011 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
3012 case 'l': // register suitable for indirect jump
3014 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
3015 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
3016 case 'x': // register suitable for indirect jump
3017 // Fixme: Not triggering the use of both hi and low
3018 // This will generate an error message
3019 return std::make_pair(0U, nullptr);
3023 std::pair<unsigned, const TargetRegisterClass *> R;
3024 R = parseRegForInlineAsmConstraint(Constraint, VT);
3029 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3032 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3033 /// vector. If it is invalid, don't add anything to Ops.
3034 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3035 std::string &Constraint,
3036 std::vector<SDValue>&Ops,
3037 SelectionDAG &DAG) const {
3040 // Only support length 1 constraints for now.
3041 if (Constraint.length() > 1) return;
3043 char ConstraintLetter = Constraint[0];
3044 switch (ConstraintLetter) {
3045 default: break; // This will fall through to the generic implementation
3046 case 'I': // Signed 16 bit constant
3047 // If this fails, the parent routine will give an error
3048 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3049 EVT Type = Op.getValueType();
3050 int64_t Val = C->getSExtValue();
3051 if (isInt<16>(Val)) {
3052 Result = DAG.getTargetConstant(Val, Type);
3057 case 'J': // integer zero
3058 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3059 EVT Type = Op.getValueType();
3060 int64_t Val = C->getZExtValue();
3062 Result = DAG.getTargetConstant(0, Type);
3067 case 'K': // unsigned 16 bit immediate
3068 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3069 EVT Type = Op.getValueType();
3070 uint64_t Val = (uint64_t)C->getZExtValue();
3071 if (isUInt<16>(Val)) {
3072 Result = DAG.getTargetConstant(Val, Type);
3077 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3078 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3079 EVT Type = Op.getValueType();
3080 int64_t Val = C->getSExtValue();
3081 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3082 Result = DAG.getTargetConstant(Val, Type);
3087 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3088 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3089 EVT Type = Op.getValueType();
3090 int64_t Val = C->getSExtValue();
3091 if ((Val >= -65535) && (Val <= -1)) {
3092 Result = DAG.getTargetConstant(Val, Type);
3097 case 'O': // signed 15 bit immediate
3098 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3099 EVT Type = Op.getValueType();
3100 int64_t Val = C->getSExtValue();
3101 if ((isInt<15>(Val))) {
3102 Result = DAG.getTargetConstant(Val, Type);
3107 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3108 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3109 EVT Type = Op.getValueType();
3110 int64_t Val = C->getSExtValue();
3111 if ((Val <= 65535) && (Val >= 1)) {
3112 Result = DAG.getTargetConstant(Val, Type);
3119 if (Result.getNode()) {
3120 Ops.push_back(Result);
3124 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3127 bool MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3129 // No global is ever allowed as a base.
3134 case 0: // "r+i" or just "i", depending on HasBaseReg.
3137 if (!AM.HasBaseReg) // allow "r+i".
3139 return false; // disallow "r+r" or "r+r+i".
3148 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3149 // The Mips target isn't yet aware of offsets.
3153 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3155 bool IsMemset, bool ZeroMemset,
3157 MachineFunction &MF) const {
3158 if (Subtarget->hasMips64())
3164 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3165 if (VT != MVT::f32 && VT != MVT::f64)
3167 if (Imm.isNegZero())
3169 return Imm.isZero();
3172 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3174 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3176 return TargetLowering::getJumpTableEncoding();
3179 /// This function returns true if CallSym is a long double emulation routine.
3180 static bool isF128SoftLibCall(const char *CallSym) {
3181 const char *const LibCalls[] =
3182 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2",
3183 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi",
3184 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf",
3185 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2",
3186 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3",
3187 "__trunctfdf2", "__trunctfsf2", "__unordtf2",
3188 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl",
3189 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl",
3192 const char *const *End = LibCalls + array_lengthof(LibCalls);
3194 // Check that LibCalls is sorted alphabetically.
3195 MipsTargetLowering::LTStr Comp;
3198 for (const char *const *I = LibCalls; I < End - 1; ++I)
3199 assert(Comp(*I, *(I + 1)));
3202 return std::binary_search(LibCalls, End, CallSym, Comp);
3205 /// This function returns true if Ty is fp128 or i128 which was originally a
3207 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) {
3208 if (Ty->isFP128Ty())
3211 const ExternalSymbolSDNode *ES =
3212 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
3214 // If the Ty is i128 and the function being called is a long double emulation
3215 // routine, then the original type is f128.
3216 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
3219 MipsTargetLowering::MipsCC::SpecialCallingConvType
3220 MipsTargetLowering::getSpecialCallingConv(SDValue Callee) const {
3221 MipsCC::SpecialCallingConvType SpecialCallingConv =
3222 MipsCC::NoSpecialCallingConv;
3223 if (Subtarget->inMips16HardFloat()) {
3224 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3225 llvm::StringRef Sym = G->getGlobal()->getName();
3226 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
3227 if (F && F->hasFnAttribute("__Mips16RetHelper")) {
3228 SpecialCallingConv = MipsCC::Mips16RetHelperConv;
3232 return SpecialCallingConv;
3235 MipsTargetLowering::MipsCC::MipsCC(
3236 CallingConv::ID CC, bool IsO32_, bool IsFP64_, CCState &Info,
3237 MipsCC::SpecialCallingConvType SpecialCallingConv_)
3238 : CCInfo(Info), CallConv(CC), IsO32(IsO32_), IsFP64(IsFP64_),
3239 SpecialCallingConv(SpecialCallingConv_){
3240 // Pre-allocate reserved argument area.
3241 CCInfo.AllocateStack(reservedArgArea(), 1);
3245 void MipsTargetLowering::MipsCC::
3246 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
3247 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode,
3248 std::vector<ArgListEntry> &FuncArgs) {
3249 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
3250 "CallingConv::Fast shouldn't be used for vararg functions.");
3252 unsigned NumOpnds = Args.size();
3253 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
3255 for (unsigned I = 0; I != NumOpnds; ++I) {
3256 MVT ArgVT = Args[I].VT;
3257 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3260 if (ArgFlags.isByVal()) {
3261 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3265 if (IsVarArg && !Args[I].IsFixed)
3266 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3268 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode,
3270 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo);
3275 dbgs() << "Call operand #" << I << " has unhandled type "
3276 << EVT(ArgVT).getEVTString();
3278 llvm_unreachable(nullptr);
3283 void MipsTargetLowering::MipsCC::
3284 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args,
3285 bool IsSoftFloat, Function::const_arg_iterator FuncArg) {
3286 unsigned NumArgs = Args.size();
3287 llvm::CCAssignFn *FixedFn = fixedArgFn();
3288 unsigned CurArgIdx = 0;
3290 for (unsigned I = 0; I != NumArgs; ++I) {
3291 MVT ArgVT = Args[I].VT;
3292 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3293 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx);
3294 CurArgIdx = Args[I].OrigArgIndex;
3296 if (ArgFlags.isByVal()) {
3297 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3301 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), nullptr, IsSoftFloat);
3303 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo))
3307 dbgs() << "Formal Arg #" << I << " has unhandled type "
3308 << EVT(ArgVT).getEVTString();
3310 llvm_unreachable(nullptr);
3314 template<typename Ty>
3315 void MipsTargetLowering::MipsCC::
3316 analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat,
3317 const SDNode *CallNode, const Type *RetTy) const {
3320 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode))
3321 Fn = RetCC_F128Soft;
3325 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) {
3326 MVT VT = RetVals[I].VT;
3327 ISD::ArgFlagsTy Flags = RetVals[I].Flags;
3328 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat);
3330 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) {
3332 dbgs() << "Call result #" << I << " has unhandled type "
3333 << EVT(VT).getEVTString() << '\n';
3335 llvm_unreachable(nullptr);
3340 void MipsTargetLowering::MipsCC::
3341 analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat,
3342 const SDNode *CallNode, const Type *RetTy) const {
3343 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy);
3346 void MipsTargetLowering::MipsCC::
3347 analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat,
3348 const Type *RetTy) const {
3349 analyzeReturn(Outs, IsSoftFloat, nullptr, RetTy);
3352 void MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3354 CCValAssign::LocInfo LocInfo,
3355 ISD::ArgFlagsTy ArgFlags) {
3356 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3358 struct ByValArgInfo ByVal;
3359 unsigned RegSize = regSize();
3360 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3361 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3364 if (useRegsForByval())
3365 allocateRegs(ByVal, ByValSize, Align);
3367 // Allocate space on caller's stack.
3368 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3370 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3372 ByValArgs.push_back(ByVal);
3375 unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const {
3376 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs);
3379 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
3380 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0;
3383 const MCPhysReg *MipsTargetLowering::MipsCC::intArgRegs() const {
3384 return IsO32 ? O32IntRegs : Mips64IntRegs;
3387 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
3388 if (CallConv == CallingConv::Fast)
3389 return CC_Mips_FastCC;
3391 if (SpecialCallingConv == Mips16RetHelperConv)
3392 return CC_Mips16RetHelper;
3393 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN;
3396 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
3397 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN_VarArg;
3400 const MCPhysReg *MipsTargetLowering::MipsCC::shadowRegs() const {
3401 return IsO32 ? O32IntRegs : Mips64DPRegs;
3404 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3407 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs();
3408 const MCPhysReg *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs();
3409 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3410 "Byval argument's size and alignment should be a multiple of"
3413 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3415 // If Align > RegSize, the first arg register must be even.
3416 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3417 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3421 // Mark the registers allocated.
3422 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3423 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3424 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3427 MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy,
3428 const SDNode *CallNode,
3429 bool IsSoftFloat) const {
3430 if (IsSoftFloat || IsO32)
3433 // Check if the original type was fp128.
3434 if (originalTypeIsF128(OrigTy, CallNode)) {
3435 assert(VT == MVT::i64);
3442 void MipsTargetLowering::
3443 copyByValRegs(SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains,
3444 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3445 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3446 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3447 MachineFunction &MF = DAG.getMachineFunction();
3448 MachineFrameInfo *MFI = MF.getFrameInfo();
3449 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3450 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3454 FrameObjOffset = (int)CC.reservedArgArea() -
3455 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3457 FrameObjOffset = ByVal.Address;
3459 // Create frame object.
3460 EVT PtrTy = getPointerTy();
3461 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3462 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3463 InVals.push_back(FIN);
3468 // Copy arg registers.
3469 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3470 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3472 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3473 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3474 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3475 unsigned Offset = I * CC.regSize();
3476 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3477 DAG.getConstant(Offset, PtrTy));
3478 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3479 StorePtr, MachinePointerInfo(FuncArg, Offset),
3481 OutChains.push_back(Store);
3485 // Copy byVal arg to registers and stack.
3486 void MipsTargetLowering::
3487 passByValArg(SDValue Chain, SDLoc DL,
3488 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
3489 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
3490 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3491 const MipsCC &CC, const ByValArgInfo &ByVal,
3492 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3493 unsigned ByValSizeInBytes = Flags.getByValSize();
3494 unsigned OffsetInBytes = 0; // From beginning of struct
3495 unsigned RegSizeInBytes = CC.regSize();
3496 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes);
3497 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
3499 if (ByVal.NumRegs) {
3500 const MCPhysReg *ArgRegs = CC.intArgRegs();
3501 bool LeftoverBytes = (ByVal.NumRegs * RegSizeInBytes > ByValSizeInBytes);
3504 // Copy words to registers.
3505 for (; I < ByVal.NumRegs - LeftoverBytes;
3506 ++I, OffsetInBytes += RegSizeInBytes) {
3507 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3508 DAG.getConstant(OffsetInBytes, PtrTy));
3509 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3510 MachinePointerInfo(), false, false, false,
3512 MemOpChains.push_back(LoadVal.getValue(1));
3513 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3514 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3517 // Return if the struct has been fully copied.
3518 if (ByValSizeInBytes == OffsetInBytes)
3521 // Copy the remainder of the byval argument with sub-word loads and shifts.
3522 if (LeftoverBytes) {
3523 assert((ByValSizeInBytes > OffsetInBytes) &&
3524 (ByValSizeInBytes < OffsetInBytes + RegSizeInBytes) &&
3525 "Size of the remainder should be smaller than RegSizeInBytes.");
3528 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0;
3529 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) {
3530 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes;
3532 if (RemainingSizeInBytes < LoadSizeInBytes)
3536 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3537 DAG.getConstant(OffsetInBytes, PtrTy));
3538 SDValue LoadVal = DAG.getExtLoad(
3539 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(),
3540 MVT::getIntegerVT(LoadSizeInBytes * 8), false, false, Alignment);
3541 MemOpChains.push_back(LoadVal.getValue(1));
3543 // Shift the loaded value.
3547 Shamt = TotalBytesLoaded * 8;
3549 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8;
3551 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3552 DAG.getConstant(Shamt, MVT::i32));
3555 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3559 OffsetInBytes += LoadSizeInBytes;
3560 TotalBytesLoaded += LoadSizeInBytes;
3561 Alignment = std::min(Alignment, LoadSizeInBytes);
3564 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3565 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3570 // Copy remainder of byval arg to it with memcpy.
3571 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes;
3572 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3573 DAG.getConstant(OffsetInBytes, PtrTy));
3574 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3575 DAG.getIntPtrConstant(ByVal.Address));
3576 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, DAG.getConstant(MemCpySize, PtrTy),
3577 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false,
3578 MachinePointerInfo(), MachinePointerInfo());
3579 MemOpChains.push_back(Chain);
3582 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3583 const MipsCC &CC, SDValue Chain,
3584 SDLoc DL, SelectionDAG &DAG) const {
3585 unsigned NumRegs = CC.numIntArgRegs();
3586 const MCPhysReg *ArgRegs = CC.intArgRegs();
3587 const CCState &CCInfo = CC.getCCInfo();
3588 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3589 unsigned RegSize = CC.regSize();
3590 MVT RegTy = MVT::getIntegerVT(RegSize * 8);
3591 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3592 MachineFunction &MF = DAG.getMachineFunction();
3593 MachineFrameInfo *MFI = MF.getFrameInfo();
3594 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3596 // Offset of the first variable argument from stack pointer.
3600 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3602 VaArgOffset = (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3604 // Record the frame index of the first variable argument
3605 // which is a value necessary to VASTART.
3606 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3607 MipsFI->setVarArgsFrameIndex(FI);
3609 // Copy the integer registers that have not been used for argument passing
3610 // to the argument register save area. For O32, the save area is allocated
3611 // in the caller's stack frame, while for N32/64, it is allocated in the
3612 // callee's stack frame.
3613 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3614 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3615 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3616 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3617 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3618 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3619 MachinePointerInfo(), false, false, 0);
3620 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue((Value*)nullptr);
3621 OutChains.push_back(Store);