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/MachineJumpTableInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/SelectionDAGISel.h"
30 #include "llvm/CodeGen/ValueTypes.h"
31 #include "llvm/IR/CallingConv.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/GlobalVariable.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
42 #define DEBUG_TYPE "mips-lower"
44 STATISTIC(NumTailCalls, "Number of tail calls");
47 LargeGOT("mxgot", cl::Hidden,
48 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
51 NoZeroDivCheck("mno-check-zero-division", cl::Hidden,
52 cl::desc("MIPS: Don't trap on integer division by zero."),
56 EnableMipsFastISel("mips-fast-isel", cl::Hidden,
57 cl::desc("Allow mips-fast-isel to be used"),
60 static const MCPhysReg O32IntRegs[4] = {
61 Mips::A0, Mips::A1, Mips::A2, Mips::A3
64 static const MCPhysReg Mips64IntRegs[8] = {
65 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
66 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64
69 static const MCPhysReg Mips64DPRegs[8] = {
70 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
71 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
74 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode);
77 class MipsCCState : public CCState {
79 /// Identify lowered values that originated from f128 arguments and record
80 /// this for use by RetCC_MipsN.
82 PreAnalyzeCallResultForF128(const SmallVectorImpl<ISD::InputArg> &Ins,
83 const TargetLowering::CallLoweringInfo &CLI) {
84 const MachineFunction &MF = getMachineFunction();
85 for (unsigned i = 0; i < Ins.size(); ++i)
86 OriginalArgWasF128.push_back(
87 originalTypeIsF128(CLI.RetTy, CLI.Callee.getNode()));
90 /// Identify lowered values that originated from f128 arguments and record
91 /// this for use by RetCC_MipsN.
92 void PreAnalyzeReturnForF128(const SmallVectorImpl<ISD::OutputArg> &Outs) {
93 const MachineFunction &MF = getMachineFunction();
94 for (unsigned i = 0; i < Outs.size(); ++i)
95 OriginalArgWasF128.push_back(
96 originalTypeIsF128(MF.getFunction()->getReturnType(), nullptr));
99 /// Records whether the value has been lowered from an f128.
100 SmallVector<bool, 4> OriginalArgWasF128;
103 MipsCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
104 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C)
105 : CCState(CC, isVarArg, MF, locs, C) {}
107 void AnalyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins,
109 const TargetLowering::CallLoweringInfo &CLI) {
110 PreAnalyzeCallResultForF128(Ins, CLI);
111 CCState::AnalyzeCallResult(Ins, Fn);
112 OriginalArgWasF128.clear();
115 void AnalyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs,
117 PreAnalyzeReturnForF128(Outs);
118 CCState::AnalyzeReturn(Outs, Fn);
119 OriginalArgWasF128.clear();
122 bool CheckReturn(const SmallVectorImpl<ISD::OutputArg> &ArgsFlags,
124 PreAnalyzeReturnForF128(ArgsFlags);
125 bool Return = CCState::CheckReturn(ArgsFlags, Fn);
126 OriginalArgWasF128.clear();
130 bool WasOriginalArgF128(unsigned ValNo) { return OriginalArgWasF128[ValNo]; }
134 // If I is a shifted mask, set the size (Size) and the first bit of the
135 // mask (Pos), and return true.
136 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
137 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
138 if (!isShiftedMask_64(I))
141 Size = CountPopulation_64(I);
142 Pos = countTrailingZeros(I);
146 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
147 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
148 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
151 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
153 unsigned Flag) const {
154 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag);
157 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty,
159 unsigned Flag) const {
160 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
163 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty,
165 unsigned Flag) const {
166 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
169 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty,
171 unsigned Flag) const {
172 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
175 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty,
177 unsigned Flag) const {
178 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
179 N->getOffset(), Flag);
182 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
184 case MipsISD::JmpLink: return "MipsISD::JmpLink";
185 case MipsISD::TailCall: return "MipsISD::TailCall";
186 case MipsISD::Hi: return "MipsISD::Hi";
187 case MipsISD::Lo: return "MipsISD::Lo";
188 case MipsISD::GPRel: return "MipsISD::GPRel";
189 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
190 case MipsISD::Ret: return "MipsISD::Ret";
191 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
192 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
193 case MipsISD::FPCmp: return "MipsISD::FPCmp";
194 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
195 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
196 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP";
197 case MipsISD::MFHI: return "MipsISD::MFHI";
198 case MipsISD::MFLO: return "MipsISD::MFLO";
199 case MipsISD::MTLOHI: return "MipsISD::MTLOHI";
200 case MipsISD::Mult: return "MipsISD::Mult";
201 case MipsISD::Multu: return "MipsISD::Multu";
202 case MipsISD::MAdd: return "MipsISD::MAdd";
203 case MipsISD::MAddu: return "MipsISD::MAddu";
204 case MipsISD::MSub: return "MipsISD::MSub";
205 case MipsISD::MSubu: return "MipsISD::MSubu";
206 case MipsISD::DivRem: return "MipsISD::DivRem";
207 case MipsISD::DivRemU: return "MipsISD::DivRemU";
208 case MipsISD::DivRem16: return "MipsISD::DivRem16";
209 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
210 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
211 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
212 case MipsISD::Wrapper: return "MipsISD::Wrapper";
213 case MipsISD::Sync: return "MipsISD::Sync";
214 case MipsISD::Ext: return "MipsISD::Ext";
215 case MipsISD::Ins: return "MipsISD::Ins";
216 case MipsISD::LWL: return "MipsISD::LWL";
217 case MipsISD::LWR: return "MipsISD::LWR";
218 case MipsISD::SWL: return "MipsISD::SWL";
219 case MipsISD::SWR: return "MipsISD::SWR";
220 case MipsISD::LDL: return "MipsISD::LDL";
221 case MipsISD::LDR: return "MipsISD::LDR";
222 case MipsISD::SDL: return "MipsISD::SDL";
223 case MipsISD::SDR: return "MipsISD::SDR";
224 case MipsISD::EXTP: return "MipsISD::EXTP";
225 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
226 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
227 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
228 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
229 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
230 case MipsISD::SHILO: return "MipsISD::SHILO";
231 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
232 case MipsISD::MULT: return "MipsISD::MULT";
233 case MipsISD::MULTU: return "MipsISD::MULTU";
234 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
235 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
236 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
237 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
238 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP";
239 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP";
240 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP";
241 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP";
242 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP";
243 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO";
244 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO";
245 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO";
246 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO";
247 case MipsISD::VCEQ: return "MipsISD::VCEQ";
248 case MipsISD::VCLE_S: return "MipsISD::VCLE_S";
249 case MipsISD::VCLE_U: return "MipsISD::VCLE_U";
250 case MipsISD::VCLT_S: return "MipsISD::VCLT_S";
251 case MipsISD::VCLT_U: return "MipsISD::VCLT_U";
252 case MipsISD::VSMAX: return "MipsISD::VSMAX";
253 case MipsISD::VSMIN: return "MipsISD::VSMIN";
254 case MipsISD::VUMAX: return "MipsISD::VUMAX";
255 case MipsISD::VUMIN: return "MipsISD::VUMIN";
256 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
257 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
258 case MipsISD::VNOR: return "MipsISD::VNOR";
259 case MipsISD::VSHF: return "MipsISD::VSHF";
260 case MipsISD::SHF: return "MipsISD::SHF";
261 case MipsISD::ILVEV: return "MipsISD::ILVEV";
262 case MipsISD::ILVOD: return "MipsISD::ILVOD";
263 case MipsISD::ILVL: return "MipsISD::ILVL";
264 case MipsISD::ILVR: return "MipsISD::ILVR";
265 case MipsISD::PCKEV: return "MipsISD::PCKEV";
266 case MipsISD::PCKOD: return "MipsISD::PCKOD";
267 case MipsISD::INSVE: return "MipsISD::INSVE";
268 default: return nullptr;
272 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM,
273 const MipsSubtarget &STI)
274 : TargetLowering(TM, new MipsTargetObjectFile()), Subtarget(STI) {
275 // Mips does not have i1 type, so use i32 for
276 // setcc operations results (slt, sgt, ...).
277 setBooleanContents(ZeroOrOneBooleanContent);
278 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
279 // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA
280 // does. Integer booleans still use 0 and 1.
281 if (Subtarget.hasMips32r6())
282 setBooleanContents(ZeroOrOneBooleanContent,
283 ZeroOrNegativeOneBooleanContent);
285 // Load extented operations for i1 types must be promoted
286 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
287 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
288 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
290 // MIPS doesn't have extending float->double load/store
291 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
292 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
294 // Used by legalize types to correctly generate the setcc result.
295 // Without this, every float setcc comes with a AND/OR with the result,
296 // we don't want this, since the fpcmp result goes to a flag register,
297 // which is used implicitly by brcond and select operations.
298 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
300 // Mips Custom Operations
301 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
302 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
303 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
304 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
305 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
306 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
307 setOperationAction(ISD::SELECT, MVT::f32, Custom);
308 setOperationAction(ISD::SELECT, MVT::f64, Custom);
309 setOperationAction(ISD::SELECT, MVT::i32, Custom);
310 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
311 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
312 setOperationAction(ISD::SETCC, MVT::f32, Custom);
313 setOperationAction(ISD::SETCC, MVT::f64, Custom);
314 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
315 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
316 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
317 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
319 if (Subtarget.isGP64bit()) {
320 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
321 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
322 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
323 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
324 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
325 setOperationAction(ISD::SELECT, MVT::i64, Custom);
326 setOperationAction(ISD::LOAD, MVT::i64, Custom);
327 setOperationAction(ISD::STORE, MVT::i64, Custom);
328 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
331 if (!Subtarget.isGP64bit()) {
332 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
333 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
334 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
337 setOperationAction(ISD::ADD, MVT::i32, Custom);
338 if (Subtarget.isGP64bit())
339 setOperationAction(ISD::ADD, MVT::i64, Custom);
341 setOperationAction(ISD::SDIV, MVT::i32, Expand);
342 setOperationAction(ISD::SREM, MVT::i32, Expand);
343 setOperationAction(ISD::UDIV, MVT::i32, Expand);
344 setOperationAction(ISD::UREM, MVT::i32, Expand);
345 setOperationAction(ISD::SDIV, MVT::i64, Expand);
346 setOperationAction(ISD::SREM, MVT::i64, Expand);
347 setOperationAction(ISD::UDIV, MVT::i64, Expand);
348 setOperationAction(ISD::UREM, MVT::i64, Expand);
350 // Operations not directly supported by Mips.
351 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
352 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
353 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
354 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
355 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
356 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
357 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
358 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
359 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
360 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
361 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
362 if (Subtarget.hasCnMips()) {
363 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
364 setOperationAction(ISD::CTPOP, MVT::i64, Legal);
366 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
367 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
369 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
370 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
371 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
372 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
373 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
374 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
375 setOperationAction(ISD::ROTL, MVT::i32, Expand);
376 setOperationAction(ISD::ROTL, MVT::i64, Expand);
377 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
378 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
380 if (!Subtarget.hasMips32r2())
381 setOperationAction(ISD::ROTR, MVT::i32, Expand);
383 if (!Subtarget.hasMips64r2())
384 setOperationAction(ISD::ROTR, MVT::i64, Expand);
386 setOperationAction(ISD::FSIN, MVT::f32, Expand);
387 setOperationAction(ISD::FSIN, MVT::f64, Expand);
388 setOperationAction(ISD::FCOS, MVT::f32, Expand);
389 setOperationAction(ISD::FCOS, MVT::f64, Expand);
390 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
391 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
392 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
393 setOperationAction(ISD::FPOW, MVT::f32, Expand);
394 setOperationAction(ISD::FPOW, MVT::f64, Expand);
395 setOperationAction(ISD::FLOG, MVT::f32, Expand);
396 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
397 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
398 setOperationAction(ISD::FEXP, MVT::f32, Expand);
399 setOperationAction(ISD::FMA, MVT::f32, Expand);
400 setOperationAction(ISD::FMA, MVT::f64, Expand);
401 setOperationAction(ISD::FREM, MVT::f32, Expand);
402 setOperationAction(ISD::FREM, MVT::f64, Expand);
404 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
406 setOperationAction(ISD::VASTART, MVT::Other, Custom);
407 setOperationAction(ISD::VAARG, MVT::Other, Custom);
408 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
409 setOperationAction(ISD::VAEND, MVT::Other, Expand);
411 // Use the default for now
412 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
413 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
415 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
416 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
417 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
418 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
420 setInsertFencesForAtomic(true);
422 if (!Subtarget.hasMips32r2()) {
423 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
424 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
427 // MIPS16 lacks MIPS32's clz and clo instructions.
428 if (!Subtarget.hasMips32() || Subtarget.inMips16Mode())
429 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
430 if (!Subtarget.hasMips64())
431 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
433 if (!Subtarget.hasMips32r2())
434 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
435 if (!Subtarget.hasMips64r2())
436 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
438 if (Subtarget.isGP64bit()) {
439 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
440 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
441 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
442 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
445 setOperationAction(ISD::TRAP, MVT::Other, Legal);
447 setTargetDAGCombine(ISD::SDIVREM);
448 setTargetDAGCombine(ISD::UDIVREM);
449 setTargetDAGCombine(ISD::SELECT);
450 setTargetDAGCombine(ISD::AND);
451 setTargetDAGCombine(ISD::OR);
452 setTargetDAGCombine(ISD::ADD);
454 setMinFunctionAlignment(Subtarget.isGP64bit() ? 3 : 2);
456 // The arguments on the stack are defined in terms of 4-byte slots on O32
457 // and 8-byte slots on N32/N64.
458 setMinStackArgumentAlignment(
459 (Subtarget.isABI_N32() || Subtarget.isABI_N64()) ? 8 : 4);
461 setStackPointerRegisterToSaveRestore(Subtarget.isABI_N64() ? Mips::SP_64
464 setExceptionPointerRegister(Subtarget.isABI_N64() ? Mips::A0_64 : Mips::A0);
465 setExceptionSelectorRegister(Subtarget.isABI_N64() ? Mips::A1_64 : Mips::A1);
467 MaxStoresPerMemcpy = 16;
469 isMicroMips = Subtarget.inMicroMipsMode();
472 const MipsTargetLowering *MipsTargetLowering::create(const MipsTargetMachine &TM,
473 const MipsSubtarget &STI) {
474 if (STI.inMips16Mode())
475 return llvm::createMips16TargetLowering(TM, STI);
477 return llvm::createMipsSETargetLowering(TM, STI);
480 // Create a fast isel object.
482 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
483 const TargetLibraryInfo *libInfo) const {
484 if (!EnableMipsFastISel)
485 return TargetLowering::createFastISel(funcInfo, libInfo);
486 return Mips::createFastISel(funcInfo, libInfo);
489 EVT MipsTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
492 return VT.changeVectorElementTypeToInteger();
495 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
496 TargetLowering::DAGCombinerInfo &DCI,
497 const MipsSubtarget &Subtarget) {
498 if (DCI.isBeforeLegalizeOps())
501 EVT Ty = N->getValueType(0);
502 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
503 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
504 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
508 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
509 N->getOperand(0), N->getOperand(1));
510 SDValue InChain = DAG.getEntryNode();
511 SDValue InGlue = DivRem;
514 if (N->hasAnyUseOfValue(0)) {
515 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
517 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
518 InChain = CopyFromLo.getValue(1);
519 InGlue = CopyFromLo.getValue(2);
523 if (N->hasAnyUseOfValue(1)) {
524 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
526 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
532 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
534 default: llvm_unreachable("Unknown fp condition code!");
536 case ISD::SETOEQ: return Mips::FCOND_OEQ;
537 case ISD::SETUNE: return Mips::FCOND_UNE;
539 case ISD::SETOLT: return Mips::FCOND_OLT;
541 case ISD::SETOGT: return Mips::FCOND_OGT;
543 case ISD::SETOLE: return Mips::FCOND_OLE;
545 case ISD::SETOGE: return Mips::FCOND_OGE;
546 case ISD::SETULT: return Mips::FCOND_ULT;
547 case ISD::SETULE: return Mips::FCOND_ULE;
548 case ISD::SETUGT: return Mips::FCOND_UGT;
549 case ISD::SETUGE: return Mips::FCOND_UGE;
550 case ISD::SETUO: return Mips::FCOND_UN;
551 case ISD::SETO: return Mips::FCOND_OR;
553 case ISD::SETONE: return Mips::FCOND_ONE;
554 case ISD::SETUEQ: return Mips::FCOND_UEQ;
559 /// This function returns true if the floating point conditional branches and
560 /// conditional moves which use condition code CC should be inverted.
561 static bool invertFPCondCodeUser(Mips::CondCode CC) {
562 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
565 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
566 "Illegal Condition Code");
571 // Creates and returns an FPCmp node from a setcc node.
572 // Returns Op if setcc is not a floating point comparison.
573 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
574 // must be a SETCC node
575 if (Op.getOpcode() != ISD::SETCC)
578 SDValue LHS = Op.getOperand(0);
580 if (!LHS.getValueType().isFloatingPoint())
583 SDValue RHS = Op.getOperand(1);
586 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
587 // node if necessary.
588 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
590 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
591 DAG.getConstant(condCodeToFCC(CC), MVT::i32));
594 // Creates and returns a CMovFPT/F node.
595 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
596 SDValue False, SDLoc DL) {
597 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
598 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
599 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
601 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
602 True.getValueType(), True, FCC0, False, Cond);
605 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
606 TargetLowering::DAGCombinerInfo &DCI,
607 const MipsSubtarget &Subtarget) {
608 if (DCI.isBeforeLegalizeOps())
611 SDValue SetCC = N->getOperand(0);
613 if ((SetCC.getOpcode() != ISD::SETCC) ||
614 !SetCC.getOperand(0).getValueType().isInteger())
617 SDValue False = N->getOperand(2);
618 EVT FalseTy = False.getValueType();
620 if (!FalseTy.isInteger())
623 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False);
625 // If the RHS (False) is 0, we swap the order of the operands
626 // of ISD::SELECT (obviously also inverting the condition) so that we can
627 // take advantage of conditional moves using the $0 register.
629 // return (a != 0) ? x : 0;
637 if (!FalseC->getZExtValue()) {
638 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
639 SDValue True = N->getOperand(1);
641 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
642 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
644 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
647 // If both operands are integer constants there's a possibility that we
648 // can do some interesting optimizations.
649 SDValue True = N->getOperand(1);
650 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True);
652 if (!TrueC || !True.getValueType().isInteger())
655 // We'll also ignore MVT::i64 operands as this optimizations proves
656 // to be ineffective because of the required sign extensions as the result
657 // of a SETCC operator is always MVT::i32 for non-vector types.
658 if (True.getValueType() == MVT::i64)
661 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue();
663 // 1) (a < x) ? y : y-1
665 // addiu $reg2, $reg1, y-1
667 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False);
669 // 2) (a < x) ? y-1 : y
671 // xor $reg1, $reg1, 1
672 // addiu $reg2, $reg1, y-1
674 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
675 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
676 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
677 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True);
680 // Couldn't optimize.
684 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
685 TargetLowering::DAGCombinerInfo &DCI,
686 const MipsSubtarget &Subtarget) {
687 // Pattern match EXT.
688 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
689 // => ext $dst, $src, size, pos
690 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
693 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
694 unsigned ShiftRightOpc = ShiftRight.getOpcode();
696 // Op's first operand must be a shift right.
697 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
700 // The second operand of the shift must be an immediate.
702 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
705 uint64_t Pos = CN->getZExtValue();
706 uint64_t SMPos, SMSize;
708 // Op's second operand must be a shifted mask.
709 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
710 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
713 // Return if the shifted mask does not start at bit 0 or the sum of its size
714 // and Pos exceeds the word's size.
715 EVT ValTy = N->getValueType(0);
716 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
719 return DAG.getNode(MipsISD::Ext, SDLoc(N), ValTy,
720 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
721 DAG.getConstant(SMSize, MVT::i32));
724 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
725 TargetLowering::DAGCombinerInfo &DCI,
726 const MipsSubtarget &Subtarget) {
727 // Pattern match INS.
728 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
729 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
730 // => ins $dst, $src, size, pos, $src1
731 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
734 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
735 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
738 // See if Op's first operand matches (and $src1 , mask0).
739 if (And0.getOpcode() != ISD::AND)
742 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
743 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
746 // See if Op's second operand matches (and (shl $src, pos), mask1).
747 if (And1.getOpcode() != ISD::AND)
750 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
751 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
754 // The shift masks must have the same position and size.
755 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
758 SDValue Shl = And1.getOperand(0);
759 if (Shl.getOpcode() != ISD::SHL)
762 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
765 unsigned Shamt = CN->getZExtValue();
767 // Return if the shift amount and the first bit position of mask are not the
769 EVT ValTy = N->getValueType(0);
770 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
773 return DAG.getNode(MipsISD::Ins, SDLoc(N), ValTy, Shl.getOperand(0),
774 DAG.getConstant(SMPos0, MVT::i32),
775 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
778 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
779 TargetLowering::DAGCombinerInfo &DCI,
780 const MipsSubtarget &Subtarget) {
781 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
783 if (DCI.isBeforeLegalizeOps())
786 SDValue Add = N->getOperand(1);
788 if (Add.getOpcode() != ISD::ADD)
791 SDValue Lo = Add.getOperand(1);
793 if ((Lo.getOpcode() != MipsISD::Lo) ||
794 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
797 EVT ValTy = N->getValueType(0);
800 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
802 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
805 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
807 SelectionDAG &DAG = DCI.DAG;
808 unsigned Opc = N->getOpcode();
814 return performDivRemCombine(N, DAG, DCI, Subtarget);
816 return performSELECTCombine(N, DAG, DCI, Subtarget);
818 return performANDCombine(N, DAG, DCI, Subtarget);
820 return performORCombine(N, DAG, DCI, Subtarget);
822 return performADDCombine(N, DAG, DCI, Subtarget);
829 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
830 SmallVectorImpl<SDValue> &Results,
831 SelectionDAG &DAG) const {
832 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
834 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
835 Results.push_back(Res.getValue(I));
839 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
840 SmallVectorImpl<SDValue> &Results,
841 SelectionDAG &DAG) const {
842 return LowerOperationWrapper(N, Results, DAG);
845 SDValue MipsTargetLowering::
846 LowerOperation(SDValue Op, SelectionDAG &DAG) const
848 switch (Op.getOpcode())
850 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
851 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
852 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
853 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
854 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
855 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
856 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
857 case ISD::SELECT: return lowerSELECT(Op, DAG);
858 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
859 case ISD::SETCC: return lowerSETCC(Op, DAG);
860 case ISD::VASTART: return lowerVASTART(Op, DAG);
861 case ISD::VAARG: return lowerVAARG(Op, DAG);
862 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
863 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
864 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
865 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
866 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
867 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
868 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
869 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
870 case ISD::LOAD: return lowerLOAD(Op, DAG);
871 case ISD::STORE: return lowerSTORE(Op, DAG);
872 case ISD::ADD: return lowerADD(Op, DAG);
873 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG);
878 //===----------------------------------------------------------------------===//
879 // Lower helper functions
880 //===----------------------------------------------------------------------===//
882 // addLiveIn - This helper function adds the specified physical register to the
883 // MachineFunction as a live in value. It also creates a corresponding
884 // virtual register for it.
886 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
888 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
889 MF.getRegInfo().addLiveIn(PReg, VReg);
893 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr *MI,
894 MachineBasicBlock &MBB,
895 const TargetInstrInfo &TII,
900 // Insert instruction "teq $divisor_reg, $zero, 7".
901 MachineBasicBlock::iterator I(MI);
902 MachineInstrBuilder MIB;
903 MachineOperand &Divisor = MI->getOperand(2);
904 MIB = BuildMI(MBB, std::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ))
905 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill()))
906 .addReg(Mips::ZERO).addImm(7);
908 // Use the 32-bit sub-register if this is a 64-bit division.
910 MIB->getOperand(0).setSubReg(Mips::sub_32);
912 // Clear Divisor's kill flag.
913 Divisor.setIsKill(false);
915 // We would normally delete the original instruction here but in this case
916 // we only needed to inject an additional instruction rather than replace it.
922 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
923 MachineBasicBlock *BB) const {
924 switch (MI->getOpcode()) {
926 llvm_unreachable("Unexpected instr type to insert");
927 case Mips::ATOMIC_LOAD_ADD_I8:
928 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
929 case Mips::ATOMIC_LOAD_ADD_I16:
930 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
931 case Mips::ATOMIC_LOAD_ADD_I32:
932 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
933 case Mips::ATOMIC_LOAD_ADD_I64:
934 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
936 case Mips::ATOMIC_LOAD_AND_I8:
937 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
938 case Mips::ATOMIC_LOAD_AND_I16:
939 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
940 case Mips::ATOMIC_LOAD_AND_I32:
941 return emitAtomicBinary(MI, BB, 4, Mips::AND);
942 case Mips::ATOMIC_LOAD_AND_I64:
943 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
945 case Mips::ATOMIC_LOAD_OR_I8:
946 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
947 case Mips::ATOMIC_LOAD_OR_I16:
948 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
949 case Mips::ATOMIC_LOAD_OR_I32:
950 return emitAtomicBinary(MI, BB, 4, Mips::OR);
951 case Mips::ATOMIC_LOAD_OR_I64:
952 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
954 case Mips::ATOMIC_LOAD_XOR_I8:
955 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
956 case Mips::ATOMIC_LOAD_XOR_I16:
957 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
958 case Mips::ATOMIC_LOAD_XOR_I32:
959 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
960 case Mips::ATOMIC_LOAD_XOR_I64:
961 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
963 case Mips::ATOMIC_LOAD_NAND_I8:
964 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
965 case Mips::ATOMIC_LOAD_NAND_I16:
966 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
967 case Mips::ATOMIC_LOAD_NAND_I32:
968 return emitAtomicBinary(MI, BB, 4, 0, true);
969 case Mips::ATOMIC_LOAD_NAND_I64:
970 return emitAtomicBinary(MI, BB, 8, 0, true);
972 case Mips::ATOMIC_LOAD_SUB_I8:
973 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
974 case Mips::ATOMIC_LOAD_SUB_I16:
975 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
976 case Mips::ATOMIC_LOAD_SUB_I32:
977 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
978 case Mips::ATOMIC_LOAD_SUB_I64:
979 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
981 case Mips::ATOMIC_SWAP_I8:
982 return emitAtomicBinaryPartword(MI, BB, 1, 0);
983 case Mips::ATOMIC_SWAP_I16:
984 return emitAtomicBinaryPartword(MI, BB, 2, 0);
985 case Mips::ATOMIC_SWAP_I32:
986 return emitAtomicBinary(MI, BB, 4, 0);
987 case Mips::ATOMIC_SWAP_I64:
988 return emitAtomicBinary(MI, BB, 8, 0);
990 case Mips::ATOMIC_CMP_SWAP_I8:
991 return emitAtomicCmpSwapPartword(MI, BB, 1);
992 case Mips::ATOMIC_CMP_SWAP_I16:
993 return emitAtomicCmpSwapPartword(MI, BB, 2);
994 case Mips::ATOMIC_CMP_SWAP_I32:
995 return emitAtomicCmpSwap(MI, BB, 4);
996 case Mips::ATOMIC_CMP_SWAP_I64:
997 return emitAtomicCmpSwap(MI, BB, 8);
998 case Mips::PseudoSDIV:
999 case Mips::PseudoUDIV:
1004 return insertDivByZeroTrap(
1005 MI, *BB, *getTargetMachine().getSubtargetImpl()->getInstrInfo(), false);
1006 case Mips::PseudoDSDIV:
1007 case Mips::PseudoDUDIV:
1012 return insertDivByZeroTrap(
1013 MI, *BB, *getTargetMachine().getSubtargetImpl()->getInstrInfo(), true);
1015 return emitSEL_D(MI, BB);
1019 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1020 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1022 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
1023 unsigned Size, unsigned BinOpcode,
1025 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1027 MachineFunction *MF = BB->getParent();
1028 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1029 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1030 const TargetInstrInfo *TII =
1031 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1032 DebugLoc DL = MI->getDebugLoc();
1033 unsigned LL, SC, AND, NOR, ZERO, BEQ;
1040 LL = Subtarget.hasMips32r6() ? Mips::LL_R6 : Mips::LL;
1041 SC = Subtarget.hasMips32r6() ? Mips::SC_R6 : Mips::SC;
1048 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD;
1049 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD;
1052 ZERO = Mips::ZERO_64;
1056 unsigned OldVal = MI->getOperand(0).getReg();
1057 unsigned Ptr = MI->getOperand(1).getReg();
1058 unsigned Incr = MI->getOperand(2).getReg();
1060 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1061 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1062 unsigned Success = RegInfo.createVirtualRegister(RC);
1064 // insert new blocks after the current block
1065 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1066 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1067 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1068 MachineFunction::iterator It = BB;
1070 MF->insert(It, loopMBB);
1071 MF->insert(It, exitMBB);
1073 // Transfer the remainder of BB and its successor edges to exitMBB.
1074 exitMBB->splice(exitMBB->begin(), BB,
1075 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1076 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1080 // fallthrough --> loopMBB
1081 BB->addSuccessor(loopMBB);
1082 loopMBB->addSuccessor(loopMBB);
1083 loopMBB->addSuccessor(exitMBB);
1086 // ll oldval, 0(ptr)
1087 // <binop> storeval, oldval, incr
1088 // sc success, storeval, 0(ptr)
1089 // beq success, $0, loopMBB
1091 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1093 // and andres, oldval, incr
1094 // nor storeval, $0, andres
1095 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1096 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1097 } else if (BinOpcode) {
1098 // <binop> storeval, oldval, incr
1099 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1103 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1104 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1106 MI->eraseFromParent(); // The instruction is gone now.
1111 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg(
1112 MachineInstr *MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg,
1113 unsigned SrcReg) const {
1114 const TargetInstrInfo *TII =
1115 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1116 DebugLoc DL = MI->getDebugLoc();
1118 if (Subtarget.hasMips32r2() && Size == 1) {
1119 BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg);
1123 if (Subtarget.hasMips32r2() && Size == 2) {
1124 BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg);
1128 MachineFunction *MF = BB->getParent();
1129 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1130 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1131 unsigned ScrReg = RegInfo.createVirtualRegister(RC);
1134 int64_t ShiftImm = 32 - (Size * 8);
1136 BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm);
1137 BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm);
1142 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword(
1143 MachineInstr *MI, MachineBasicBlock *BB, unsigned Size, unsigned BinOpcode,
1145 assert((Size == 1 || Size == 2) &&
1146 "Unsupported size for EmitAtomicBinaryPartial.");
1148 MachineFunction *MF = BB->getParent();
1149 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1150 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1151 const TargetInstrInfo *TII =
1152 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1153 DebugLoc DL = MI->getDebugLoc();
1155 unsigned Dest = MI->getOperand(0).getReg();
1156 unsigned Ptr = MI->getOperand(1).getReg();
1157 unsigned Incr = MI->getOperand(2).getReg();
1159 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1160 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1161 unsigned Mask = RegInfo.createVirtualRegister(RC);
1162 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1163 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1164 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1165 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1166 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1167 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1168 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1169 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1170 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1171 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1172 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1173 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1174 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1175 unsigned Success = RegInfo.createVirtualRegister(RC);
1177 // insert new blocks after the current block
1178 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1179 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1180 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1181 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1182 MachineFunction::iterator It = BB;
1184 MF->insert(It, loopMBB);
1185 MF->insert(It, sinkMBB);
1186 MF->insert(It, exitMBB);
1188 // Transfer the remainder of BB and its successor edges to exitMBB.
1189 exitMBB->splice(exitMBB->begin(), BB,
1190 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1191 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1193 BB->addSuccessor(loopMBB);
1194 loopMBB->addSuccessor(loopMBB);
1195 loopMBB->addSuccessor(sinkMBB);
1196 sinkMBB->addSuccessor(exitMBB);
1199 // addiu masklsb2,$0,-4 # 0xfffffffc
1200 // and alignedaddr,ptr,masklsb2
1201 // andi ptrlsb2,ptr,3
1202 // sll shiftamt,ptrlsb2,3
1203 // ori maskupper,$0,255 # 0xff
1204 // sll mask,maskupper,shiftamt
1205 // nor mask2,$0,mask
1206 // sll incr2,incr,shiftamt
1208 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1209 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1210 .addReg(Mips::ZERO).addImm(-4);
1211 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1212 .addReg(Ptr).addReg(MaskLSB2);
1213 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1214 if (Subtarget.isLittle()) {
1215 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1217 unsigned Off = RegInfo.createVirtualRegister(RC);
1218 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1219 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1220 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1222 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1223 .addReg(Mips::ZERO).addImm(MaskImm);
1224 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1225 .addReg(MaskUpper).addReg(ShiftAmt);
1226 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1227 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1229 // atomic.load.binop
1231 // ll oldval,0(alignedaddr)
1232 // binop binopres,oldval,incr2
1233 // and newval,binopres,mask
1234 // and maskedoldval0,oldval,mask2
1235 // or storeval,maskedoldval0,newval
1236 // sc success,storeval,0(alignedaddr)
1237 // beq success,$0,loopMBB
1241 // ll oldval,0(alignedaddr)
1242 // and newval,incr2,mask
1243 // and maskedoldval0,oldval,mask2
1244 // or storeval,maskedoldval0,newval
1245 // sc success,storeval,0(alignedaddr)
1246 // beq success,$0,loopMBB
1249 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1251 // and andres, oldval, incr2
1252 // nor binopres, $0, andres
1253 // and newval, binopres, mask
1254 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1255 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1256 .addReg(Mips::ZERO).addReg(AndRes);
1257 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1258 } else if (BinOpcode) {
1259 // <binop> binopres, oldval, incr2
1260 // and newval, binopres, mask
1261 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1262 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1263 } else { // atomic.swap
1264 // and newval, incr2, mask
1265 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1268 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1269 .addReg(OldVal).addReg(Mask2);
1270 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1271 .addReg(MaskedOldVal0).addReg(NewVal);
1272 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1273 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1274 BuildMI(BB, DL, TII->get(Mips::BEQ))
1275 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1278 // and maskedoldval1,oldval,mask
1279 // srl srlres,maskedoldval1,shiftamt
1280 // sign_extend dest,srlres
1283 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1284 .addReg(OldVal).addReg(Mask);
1285 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1286 .addReg(MaskedOldVal1).addReg(ShiftAmt);
1287 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes);
1289 MI->eraseFromParent(); // The instruction is gone now.
1294 MachineBasicBlock * MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1295 MachineBasicBlock *BB,
1296 unsigned Size) const {
1297 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1299 MachineFunction *MF = BB->getParent();
1300 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1301 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1302 const TargetInstrInfo *TII =
1303 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1304 DebugLoc DL = MI->getDebugLoc();
1305 unsigned LL, SC, ZERO, BNE, BEQ;
1308 LL = isMicroMips ? Mips::LL_MM : Mips::LL;
1309 SC = isMicroMips ? Mips::SC_MM : Mips::SC;
1316 ZERO = Mips::ZERO_64;
1321 unsigned Dest = MI->getOperand(0).getReg();
1322 unsigned Ptr = MI->getOperand(1).getReg();
1323 unsigned OldVal = MI->getOperand(2).getReg();
1324 unsigned NewVal = MI->getOperand(3).getReg();
1326 unsigned Success = RegInfo.createVirtualRegister(RC);
1328 // insert new blocks after the current block
1329 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1330 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1331 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1332 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1333 MachineFunction::iterator It = BB;
1335 MF->insert(It, loop1MBB);
1336 MF->insert(It, loop2MBB);
1337 MF->insert(It, exitMBB);
1339 // Transfer the remainder of BB and its successor edges to exitMBB.
1340 exitMBB->splice(exitMBB->begin(), BB,
1341 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1342 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1346 // fallthrough --> loop1MBB
1347 BB->addSuccessor(loop1MBB);
1348 loop1MBB->addSuccessor(exitMBB);
1349 loop1MBB->addSuccessor(loop2MBB);
1350 loop2MBB->addSuccessor(loop1MBB);
1351 loop2MBB->addSuccessor(exitMBB);
1355 // bne dest, oldval, exitMBB
1357 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1358 BuildMI(BB, DL, TII->get(BNE))
1359 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1362 // sc success, newval, 0(ptr)
1363 // beq success, $0, loop1MBB
1365 BuildMI(BB, DL, TII->get(SC), Success)
1366 .addReg(NewVal).addReg(Ptr).addImm(0);
1367 BuildMI(BB, DL, TII->get(BEQ))
1368 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1370 MI->eraseFromParent(); // The instruction is gone now.
1376 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1377 MachineBasicBlock *BB,
1378 unsigned Size) const {
1379 assert((Size == 1 || Size == 2) &&
1380 "Unsupported size for EmitAtomicCmpSwapPartial.");
1382 MachineFunction *MF = BB->getParent();
1383 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1384 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1385 const TargetInstrInfo *TII =
1386 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1387 DebugLoc DL = MI->getDebugLoc();
1389 unsigned Dest = MI->getOperand(0).getReg();
1390 unsigned Ptr = MI->getOperand(1).getReg();
1391 unsigned CmpVal = MI->getOperand(2).getReg();
1392 unsigned NewVal = MI->getOperand(3).getReg();
1394 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1395 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1396 unsigned Mask = RegInfo.createVirtualRegister(RC);
1397 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1398 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1399 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1400 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1401 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1402 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1403 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1404 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1405 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1406 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1407 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1408 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1409 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1410 unsigned Success = RegInfo.createVirtualRegister(RC);
1412 // insert new blocks after the current block
1413 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1414 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1415 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1416 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1417 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1418 MachineFunction::iterator It = BB;
1420 MF->insert(It, loop1MBB);
1421 MF->insert(It, loop2MBB);
1422 MF->insert(It, sinkMBB);
1423 MF->insert(It, exitMBB);
1425 // Transfer the remainder of BB and its successor edges to exitMBB.
1426 exitMBB->splice(exitMBB->begin(), BB,
1427 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1428 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1430 BB->addSuccessor(loop1MBB);
1431 loop1MBB->addSuccessor(sinkMBB);
1432 loop1MBB->addSuccessor(loop2MBB);
1433 loop2MBB->addSuccessor(loop1MBB);
1434 loop2MBB->addSuccessor(sinkMBB);
1435 sinkMBB->addSuccessor(exitMBB);
1437 // FIXME: computation of newval2 can be moved to loop2MBB.
1439 // addiu masklsb2,$0,-4 # 0xfffffffc
1440 // and alignedaddr,ptr,masklsb2
1441 // andi ptrlsb2,ptr,3
1442 // sll shiftamt,ptrlsb2,3
1443 // ori maskupper,$0,255 # 0xff
1444 // sll mask,maskupper,shiftamt
1445 // nor mask2,$0,mask
1446 // andi maskedcmpval,cmpval,255
1447 // sll shiftedcmpval,maskedcmpval,shiftamt
1448 // andi maskednewval,newval,255
1449 // sll shiftednewval,maskednewval,shiftamt
1450 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1451 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1452 .addReg(Mips::ZERO).addImm(-4);
1453 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1454 .addReg(Ptr).addReg(MaskLSB2);
1455 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1456 if (Subtarget.isLittle()) {
1457 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1459 unsigned Off = RegInfo.createVirtualRegister(RC);
1460 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1461 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1462 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1464 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1465 .addReg(Mips::ZERO).addImm(MaskImm);
1466 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1467 .addReg(MaskUpper).addReg(ShiftAmt);
1468 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1469 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1470 .addReg(CmpVal).addImm(MaskImm);
1471 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1472 .addReg(MaskedCmpVal).addReg(ShiftAmt);
1473 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1474 .addReg(NewVal).addImm(MaskImm);
1475 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1476 .addReg(MaskedNewVal).addReg(ShiftAmt);
1479 // ll oldval,0(alginedaddr)
1480 // and maskedoldval0,oldval,mask
1481 // bne maskedoldval0,shiftedcmpval,sinkMBB
1483 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1484 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1485 .addReg(OldVal).addReg(Mask);
1486 BuildMI(BB, DL, TII->get(Mips::BNE))
1487 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1490 // and maskedoldval1,oldval,mask2
1491 // or storeval,maskedoldval1,shiftednewval
1492 // sc success,storeval,0(alignedaddr)
1493 // beq success,$0,loop1MBB
1495 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1496 .addReg(OldVal).addReg(Mask2);
1497 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1498 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1499 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1500 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1501 BuildMI(BB, DL, TII->get(Mips::BEQ))
1502 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1505 // srl srlres,maskedoldval0,shiftamt
1506 // sign_extend dest,srlres
1509 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1510 .addReg(MaskedOldVal0).addReg(ShiftAmt);
1511 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes);
1513 MI->eraseFromParent(); // The instruction is gone now.
1518 MachineBasicBlock *MipsTargetLowering::emitSEL_D(MachineInstr *MI,
1519 MachineBasicBlock *BB) const {
1520 MachineFunction *MF = BB->getParent();
1521 const TargetRegisterInfo *TRI =
1522 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
1523 const TargetInstrInfo *TII =
1524 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1525 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1526 DebugLoc DL = MI->getDebugLoc();
1527 MachineBasicBlock::iterator II(MI);
1529 unsigned Fc = MI->getOperand(1).getReg();
1530 const auto &FGR64RegClass = TRI->getRegClass(Mips::FGR64RegClassID);
1532 unsigned Fc2 = RegInfo.createVirtualRegister(FGR64RegClass);
1534 BuildMI(*BB, II, DL, TII->get(Mips::SUBREG_TO_REG), Fc2)
1537 .addImm(Mips::sub_lo);
1539 // We don't erase the original instruction, we just replace the condition
1540 // register with the 64-bit super-register.
1541 MI->getOperand(1).setReg(Fc2);
1546 //===----------------------------------------------------------------------===//
1547 // Misc Lower Operation implementation
1548 //===----------------------------------------------------------------------===//
1549 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1550 SDValue Chain = Op.getOperand(0);
1551 SDValue Table = Op.getOperand(1);
1552 SDValue Index = Op.getOperand(2);
1554 EVT PTy = getPointerTy();
1555 unsigned EntrySize =
1556 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout());
1558 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1559 DAG.getConstant(EntrySize, PTy));
1560 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1562 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1563 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1564 MachinePointerInfo::getJumpTable(), MemVT, false, false,
1566 Chain = Addr.getValue(1);
1568 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) ||
1569 Subtarget.isABI_N64()) {
1570 // For PIC, the sequence is:
1571 // BRIND(load(Jumptable + index) + RelocBase)
1572 // RelocBase can be JumpTable, GOT or some sort of global base.
1573 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1574 getPICJumpTableRelocBase(Table, DAG));
1577 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1580 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1581 // The first operand is the chain, the second is the condition, the third is
1582 // the block to branch to if the condition is true.
1583 SDValue Chain = Op.getOperand(0);
1584 SDValue Dest = Op.getOperand(2);
1587 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1588 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1590 // Return if flag is not set by a floating point comparison.
1591 if (CondRes.getOpcode() != MipsISD::FPCmp)
1594 SDValue CCNode = CondRes.getOperand(2);
1596 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1597 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1598 SDValue BrCode = DAG.getConstant(Opc, MVT::i32);
1599 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
1600 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1601 FCC0, Dest, CondRes);
1604 SDValue MipsTargetLowering::
1605 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1607 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1608 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1610 // Return if flag is not set by a floating point comparison.
1611 if (Cond.getOpcode() != MipsISD::FPCmp)
1614 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1618 SDValue MipsTargetLowering::
1619 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1622 EVT Ty = Op.getOperand(0).getValueType();
1623 SDValue Cond = DAG.getNode(ISD::SETCC, DL,
1624 getSetCCResultType(*DAG.getContext(), Ty),
1625 Op.getOperand(0), Op.getOperand(1),
1628 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1632 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1633 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1634 SDValue Cond = createFPCmp(DAG, Op);
1636 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1637 "Floating point operand expected.");
1639 SDValue True = DAG.getConstant(1, MVT::i32);
1640 SDValue False = DAG.getConstant(0, MVT::i32);
1642 return createCMovFP(DAG, Cond, True, False, SDLoc(Op));
1645 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1646 SelectionDAG &DAG) const {
1647 // FIXME there isn't actually debug info here
1649 EVT Ty = Op.getValueType();
1650 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
1651 const GlobalValue *GV = N->getGlobal();
1653 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ &&
1654 !Subtarget.isABI_N64()) {
1655 const MipsTargetObjectFile &TLOF =
1656 (const MipsTargetObjectFile&)getObjFileLowering();
1658 // %gp_rel relocation
1659 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1660 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0,
1662 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL,
1663 DAG.getVTList(MVT::i32), GA);
1664 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1665 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode);
1668 // %hi/%lo relocation
1669 return getAddrNonPIC(N, Ty, DAG);
1672 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1673 return getAddrLocal(N, Ty, DAG,
1674 Subtarget.isABI_N32() || Subtarget.isABI_N64());
1677 return getAddrGlobalLargeGOT(N, Ty, DAG, MipsII::MO_GOT_HI16,
1678 MipsII::MO_GOT_LO16, DAG.getEntryNode(),
1679 MachinePointerInfo::getGOT());
1681 return getAddrGlobal(N, Ty, DAG,
1682 (Subtarget.isABI_N32() || Subtarget.isABI_N64())
1683 ? MipsII::MO_GOT_DISP
1685 DAG.getEntryNode(), MachinePointerInfo::getGOT());
1688 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1689 SelectionDAG &DAG) const {
1690 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
1691 EVT Ty = Op.getValueType();
1693 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ &&
1694 !Subtarget.isABI_N64())
1695 return getAddrNonPIC(N, Ty, DAG);
1697 return getAddrLocal(N, Ty, DAG,
1698 Subtarget.isABI_N32() || Subtarget.isABI_N64());
1701 SDValue MipsTargetLowering::
1702 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1704 // If the relocation model is PIC, use the General Dynamic TLS Model or
1705 // Local Dynamic TLS model, otherwise use the Initial Exec or
1706 // Local Exec TLS Model.
1708 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1710 const GlobalValue *GV = GA->getGlobal();
1711 EVT PtrVT = getPointerTy();
1713 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1715 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1716 // General Dynamic and Local Dynamic TLS Model.
1717 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1720 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1721 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1722 getGlobalReg(DAG, PtrVT), TGA);
1723 unsigned PtrSize = PtrVT.getSizeInBits();
1724 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1726 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1730 Entry.Node = Argument;
1732 Args.push_back(Entry);
1734 TargetLowering::CallLoweringInfo CLI(DAG);
1735 CLI.setDebugLoc(DL).setChain(DAG.getEntryNode())
1736 .setCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args), 0);
1737 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1739 SDValue Ret = CallResult.first;
1741 if (model != TLSModel::LocalDynamic)
1744 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1745 MipsII::MO_DTPREL_HI);
1746 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1747 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1748 MipsII::MO_DTPREL_LO);
1749 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1750 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1751 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1755 if (model == TLSModel::InitialExec) {
1756 // Initial Exec TLS Model
1757 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1758 MipsII::MO_GOTTPREL);
1759 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1761 Offset = DAG.getLoad(PtrVT, DL,
1762 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1763 false, false, false, 0);
1765 // Local Exec TLS Model
1766 assert(model == TLSModel::LocalExec);
1767 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1768 MipsII::MO_TPREL_HI);
1769 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1770 MipsII::MO_TPREL_LO);
1771 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1772 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1773 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1776 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1777 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1780 SDValue MipsTargetLowering::
1781 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1783 JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
1784 EVT Ty = Op.getValueType();
1786 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ &&
1787 !Subtarget.isABI_N64())
1788 return getAddrNonPIC(N, Ty, DAG);
1790 return getAddrLocal(N, Ty, DAG,
1791 Subtarget.isABI_N32() || Subtarget.isABI_N64());
1794 SDValue MipsTargetLowering::
1795 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1797 // gp_rel relocation
1798 // FIXME: we should reference the constant pool using small data sections,
1799 // but the asm printer currently doesn't support this feature without
1800 // hacking it. This feature should come soon so we can uncomment the
1802 //if (IsInSmallSection(C->getType())) {
1803 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1804 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1805 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1806 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1807 EVT Ty = Op.getValueType();
1809 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ &&
1810 !Subtarget.isABI_N64())
1811 return getAddrNonPIC(N, Ty, DAG);
1813 return getAddrLocal(N, Ty, DAG,
1814 Subtarget.isABI_N32() || Subtarget.isABI_N64());
1817 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1818 MachineFunction &MF = DAG.getMachineFunction();
1819 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1822 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1825 // vastart just stores the address of the VarArgsFrameIndex slot into the
1826 // memory location argument.
1827 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1828 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1829 MachinePointerInfo(SV), false, false, 0);
1832 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const {
1833 SDNode *Node = Op.getNode();
1834 EVT VT = Node->getValueType(0);
1835 SDValue Chain = Node->getOperand(0);
1836 SDValue VAListPtr = Node->getOperand(1);
1837 unsigned Align = Node->getConstantOperandVal(3);
1838 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1840 unsigned ArgSlotSizeInBytes =
1841 (Subtarget.isABI_N32() || Subtarget.isABI_N64()) ? 8 : 4;
1843 SDValue VAListLoad = DAG.getLoad(getPointerTy(), DL, Chain, VAListPtr,
1844 MachinePointerInfo(SV), false, false, false,
1846 SDValue VAList = VAListLoad;
1848 // Re-align the pointer if necessary.
1849 // It should only ever be necessary for 64-bit types on O32 since the minimum
1850 // argument alignment is the same as the maximum type alignment for N32/N64.
1852 // FIXME: We currently align too often. The code generator doesn't notice
1853 // when the pointer is still aligned from the last va_arg (or pair of
1854 // va_args for the i64 on O32 case).
1855 if (Align > getMinStackArgumentAlignment()) {
1856 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
1858 VAList = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
1859 DAG.getConstant(Align - 1,
1860 VAList.getValueType()));
1862 VAList = DAG.getNode(ISD::AND, DL, VAList.getValueType(), VAList,
1863 DAG.getConstant(-(int64_t)Align,
1864 VAList.getValueType()));
1867 // Increment the pointer, VAList, to the next vaarg.
1868 unsigned ArgSizeInBytes = getDataLayout()->getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext()));
1869 SDValue Tmp3 = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
1870 DAG.getConstant(RoundUpToAlignment(ArgSizeInBytes, ArgSlotSizeInBytes),
1871 VAList.getValueType()));
1872 // Store the incremented VAList to the legalized pointer
1873 Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr,
1874 MachinePointerInfo(SV), false, false, 0);
1876 // In big-endian mode we must adjust the pointer when the load size is smaller
1877 // than the argument slot size. We must also reduce the known alignment to
1878 // match. For example in the N64 ABI, we must add 4 bytes to the offset to get
1879 // the correct half of the slot, and reduce the alignment from 8 (slot
1880 // alignment) down to 4 (type alignment).
1881 if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) {
1882 unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes;
1883 VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList,
1884 DAG.getIntPtrConstant(Adjustment));
1886 // Load the actual argument out of the pointer VAList
1887 return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo(), false, false,
1891 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG,
1892 bool HasExtractInsert) {
1893 EVT TyX = Op.getOperand(0).getValueType();
1894 EVT TyY = Op.getOperand(1).getValueType();
1895 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1896 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1900 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1902 SDValue X = (TyX == MVT::f32) ?
1903 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1904 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1906 SDValue Y = (TyY == MVT::f32) ?
1907 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1908 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1911 if (HasExtractInsert) {
1912 // ext E, Y, 31, 1 ; extract bit31 of Y
1913 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1914 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1915 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1918 // srl SrlX, SllX, 1
1920 // sll SllY, SrlX, 31
1921 // or Or, SrlX, SllY
1922 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1923 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1924 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1925 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1926 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1929 if (TyX == MVT::f32)
1930 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1932 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1933 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1934 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1937 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG,
1938 bool HasExtractInsert) {
1939 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1940 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1941 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1942 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1945 // Bitcast to integer nodes.
1946 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1947 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1949 if (HasExtractInsert) {
1950 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1951 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1952 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1953 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
1955 if (WidthX > WidthY)
1956 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1957 else if (WidthY > WidthX)
1958 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1960 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1961 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
1962 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1965 // (d)sll SllX, X, 1
1966 // (d)srl SrlX, SllX, 1
1967 // (d)srl SrlY, Y, width(Y)-1
1968 // (d)sll SllY, SrlX, width(Y)-1
1969 // or Or, SrlX, SllY
1970 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1971 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1972 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1973 DAG.getConstant(WidthY - 1, MVT::i32));
1975 if (WidthX > WidthY)
1976 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1977 else if (WidthY > WidthX)
1978 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1980 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1981 DAG.getConstant(WidthX - 1, MVT::i32));
1982 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
1983 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
1987 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1988 if (Subtarget.isGP64bit())
1989 return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert());
1991 return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert());
1994 SDValue MipsTargetLowering::
1995 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1997 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1998 "Frame address can only be determined for current frame.");
2000 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2001 MFI->setFrameAddressIsTaken(true);
2002 EVT VT = Op.getValueType();
2005 DAG.getCopyFromReg(DAG.getEntryNode(), DL,
2006 Subtarget.isABI_N64() ? Mips::FP_64 : Mips::FP, VT);
2010 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
2011 SelectionDAG &DAG) const {
2012 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2016 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2017 "Return address can be determined only for current frame.");
2019 MachineFunction &MF = DAG.getMachineFunction();
2020 MachineFrameInfo *MFI = MF.getFrameInfo();
2021 MVT VT = Op.getSimpleValueType();
2022 unsigned RA = Subtarget.isABI_N64() ? Mips::RA_64 : Mips::RA;
2023 MFI->setReturnAddressIsTaken(true);
2025 // Return RA, which contains the return address. Mark it an implicit live-in.
2026 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2027 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
2030 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
2031 // generated from __builtin_eh_return (offset, handler)
2032 // The effect of this is to adjust the stack pointer by "offset"
2033 // and then branch to "handler".
2034 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
2036 MachineFunction &MF = DAG.getMachineFunction();
2037 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2039 MipsFI->setCallsEhReturn();
2040 SDValue Chain = Op.getOperand(0);
2041 SDValue Offset = Op.getOperand(1);
2042 SDValue Handler = Op.getOperand(2);
2044 EVT Ty = Subtarget.isABI_N64() ? MVT::i64 : MVT::i32;
2046 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2047 // EH_RETURN nodes, so that instructions are emitted back-to-back.
2048 unsigned OffsetReg = Subtarget.isABI_N64() ? Mips::V1_64 : Mips::V1;
2049 unsigned AddrReg = Subtarget.isABI_N64() ? Mips::V0_64 : Mips::V0;
2050 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2051 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2052 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2053 DAG.getRegister(OffsetReg, Ty),
2054 DAG.getRegister(AddrReg, getPointerTy()),
2058 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
2059 SelectionDAG &DAG) const {
2060 // FIXME: Need pseudo-fence for 'singlethread' fences
2061 // FIXME: Set SType for weaker fences where supported/appropriate.
2064 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2065 DAG.getConstant(SType, MVT::i32));
2068 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
2069 SelectionDAG &DAG) const {
2071 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2072 SDValue Shamt = Op.getOperand(2);
2075 // lo = (shl lo, shamt)
2076 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2079 // hi = (shl lo, shamt[4:0])
2080 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2081 DAG.getConstant(-1, MVT::i32));
2082 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
2083 DAG.getConstant(1, MVT::i32));
2084 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
2086 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
2087 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2088 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
2089 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2090 DAG.getConstant(0x20, MVT::i32));
2091 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2092 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
2093 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
2095 SDValue Ops[2] = {Lo, Hi};
2096 return DAG.getMergeValues(Ops, DL);
2099 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2102 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2103 SDValue Shamt = Op.getOperand(2);
2106 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2108 // hi = (sra hi, shamt)
2110 // hi = (srl hi, shamt)
2113 // lo = (sra hi, shamt[4:0])
2114 // hi = (sra hi, 31)
2116 // lo = (srl hi, shamt[4:0])
2118 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2119 DAG.getConstant(-1, MVT::i32));
2120 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
2121 DAG.getConstant(1, MVT::i32));
2122 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
2123 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
2124 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2125 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2127 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2128 DAG.getConstant(0x20, MVT::i32));
2129 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
2130 DAG.getConstant(31, MVT::i32));
2131 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
2132 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2133 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
2136 SDValue Ops[2] = {Lo, Hi};
2137 return DAG.getMergeValues(Ops, DL);
2140 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2141 SDValue Chain, SDValue Src, unsigned Offset) {
2142 SDValue Ptr = LD->getBasePtr();
2143 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2144 EVT BasePtrVT = Ptr.getValueType();
2146 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2149 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2150 DAG.getConstant(Offset, BasePtrVT));
2152 SDValue Ops[] = { Chain, Ptr, Src };
2153 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2154 LD->getMemOperand());
2157 // Expand an unaligned 32 or 64-bit integer load node.
2158 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2159 LoadSDNode *LD = cast<LoadSDNode>(Op);
2160 EVT MemVT = LD->getMemoryVT();
2162 if (Subtarget.systemSupportsUnalignedAccess())
2165 // Return if load is aligned or if MemVT is neither i32 nor i64.
2166 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2167 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2170 bool IsLittle = Subtarget.isLittle();
2171 EVT VT = Op.getValueType();
2172 ISD::LoadExtType ExtType = LD->getExtensionType();
2173 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2175 assert((VT == MVT::i32) || (VT == MVT::i64));
2178 // (set dst, (i64 (load baseptr)))
2180 // (set tmp, (ldl (add baseptr, 7), undef))
2181 // (set dst, (ldr baseptr, tmp))
2182 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2183 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2185 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2189 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2191 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2195 // (set dst, (i32 (load baseptr))) or
2196 // (set dst, (i64 (sextload baseptr))) or
2197 // (set dst, (i64 (extload baseptr)))
2199 // (set tmp, (lwl (add baseptr, 3), undef))
2200 // (set dst, (lwr baseptr, tmp))
2201 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2202 (ExtType == ISD::EXTLOAD))
2205 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2208 // (set dst, (i64 (zextload baseptr)))
2210 // (set tmp0, (lwl (add baseptr, 3), undef))
2211 // (set tmp1, (lwr baseptr, tmp0))
2212 // (set tmp2, (shl tmp1, 32))
2213 // (set dst, (srl tmp2, 32))
2215 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2216 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2217 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2218 SDValue Ops[] = { SRL, LWR.getValue(1) };
2219 return DAG.getMergeValues(Ops, DL);
2222 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2223 SDValue Chain, unsigned Offset) {
2224 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2225 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2227 SDVTList VTList = DAG.getVTList(MVT::Other);
2230 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2231 DAG.getConstant(Offset, BasePtrVT));
2233 SDValue Ops[] = { Chain, Value, Ptr };
2234 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2235 SD->getMemOperand());
2238 // Expand an unaligned 32 or 64-bit integer store node.
2239 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2241 SDValue Value = SD->getValue(), Chain = SD->getChain();
2242 EVT VT = Value.getValueType();
2245 // (store val, baseptr) or
2246 // (truncstore val, baseptr)
2248 // (swl val, (add baseptr, 3))
2249 // (swr val, baseptr)
2250 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2251 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2253 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2256 assert(VT == MVT::i64);
2259 // (store val, baseptr)
2261 // (sdl val, (add baseptr, 7))
2262 // (sdr val, baseptr)
2263 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2264 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2267 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
2268 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) {
2269 SDValue Val = SD->getValue();
2271 if (Val.getOpcode() != ISD::FP_TO_SINT)
2274 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2275 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2278 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2279 SD->getPointerInfo(), SD->isVolatile(),
2280 SD->isNonTemporal(), SD->getAlignment());
2283 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2284 StoreSDNode *SD = cast<StoreSDNode>(Op);
2285 EVT MemVT = SD->getMemoryVT();
2287 // Lower unaligned integer stores.
2288 if (!Subtarget.systemSupportsUnalignedAccess() &&
2289 (SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2290 ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2291 return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle());
2293 return lowerFP_TO_SINT_STORE(SD, DAG);
2296 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2297 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2298 || cast<ConstantSDNode>
2299 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2300 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2304 // (add (frameaddr 0), (frame_to_args_offset))
2305 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2306 // (add FrameObject, 0)
2307 // where FrameObject is a fixed StackObject with offset 0 which points to
2308 // the old stack pointer.
2309 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2310 EVT ValTy = Op->getValueType(0);
2311 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2312 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2313 return DAG.getNode(ISD::ADD, SDLoc(Op), ValTy, InArgsAddr,
2314 DAG.getConstant(0, ValTy));
2317 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2318 SelectionDAG &DAG) const {
2319 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2320 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2322 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2325 //===----------------------------------------------------------------------===//
2326 // Calling Convention Implementation
2327 //===----------------------------------------------------------------------===//
2329 //===----------------------------------------------------------------------===//
2330 // TODO: Implement a generic logic using tblgen that can support this.
2331 // Mips O32 ABI rules:
2333 // i32 - Passed in A0, A1, A2, A3 and stack
2334 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2335 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2336 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2337 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2338 // not used, it must be shadowed. If only A3 is available, shadow it and
2341 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2342 //===----------------------------------------------------------------------===//
2344 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2345 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2346 CCState &State, const MCPhysReg *F64Regs) {
2348 static const unsigned IntRegsSize = 4, FloatRegsSize = 2;
2350 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 };
2351 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 };
2353 // Do not process byval args here.
2354 if (ArgFlags.isByVal())
2357 // Promote i8 and i16
2358 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2360 if (ArgFlags.isSExt())
2361 LocInfo = CCValAssign::SExt;
2362 else if (ArgFlags.isZExt())
2363 LocInfo = CCValAssign::ZExt;
2365 LocInfo = CCValAssign::AExt;
2370 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2371 // is true: function is vararg, argument is 3rd or higher, there is previous
2372 // argument which is not f32 or f64.
2373 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2374 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2375 unsigned OrigAlign = ArgFlags.getOrigAlign();
2376 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2378 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2379 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2380 // If this is the first part of an i64 arg,
2381 // the allocated register must be either A0 or A2.
2382 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2383 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2385 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2386 // Allocate int register and shadow next int register. If first
2387 // available register is Mips::A1 or Mips::A3, shadow it too.
2388 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2389 if (Reg == Mips::A1 || Reg == Mips::A3)
2390 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2391 State.AllocateReg(IntRegs, IntRegsSize);
2393 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2394 // we are guaranteed to find an available float register
2395 if (ValVT == MVT::f32) {
2396 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2397 // Shadow int register
2398 State.AllocateReg(IntRegs, IntRegsSize);
2400 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2401 // Shadow int registers
2402 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2403 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2404 State.AllocateReg(IntRegs, IntRegsSize);
2405 State.AllocateReg(IntRegs, IntRegsSize);
2408 llvm_unreachable("Cannot handle this ValVT.");
2411 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2413 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2415 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2420 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2421 MVT LocVT, CCValAssign::LocInfo LocInfo,
2422 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2423 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 };
2425 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2428 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2429 MVT LocVT, CCValAssign::LocInfo LocInfo,
2430 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2431 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 };
2433 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2436 #include "MipsGenCallingConv.inc"
2438 //===----------------------------------------------------------------------===//
2439 // Call Calling Convention Implementation
2440 //===----------------------------------------------------------------------===//
2442 // Return next O32 integer argument register.
2443 static unsigned getNextIntArgReg(unsigned Reg) {
2444 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2445 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2449 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2450 SDValue Chain, SDValue Arg, SDLoc DL,
2451 bool IsTailCall, SelectionDAG &DAG) const {
2453 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2454 DAG.getIntPtrConstant(Offset));
2455 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2459 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2460 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2461 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2462 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2463 /*isVolatile=*/ true, false, 0);
2466 void MipsTargetLowering::
2467 getOpndList(SmallVectorImpl<SDValue> &Ops,
2468 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2469 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2470 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
2471 // Insert node "GP copy globalreg" before call to function.
2473 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2474 // in PIC mode) allow symbols to be resolved via lazy binding.
2475 // The lazy binding stub requires GP to point to the GOT.
2476 if (IsPICCall && !InternalLinkage) {
2477 unsigned GPReg = Subtarget.isABI_N64() ? Mips::GP_64 : Mips::GP;
2478 EVT Ty = Subtarget.isABI_N64() ? MVT::i64 : MVT::i32;
2479 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2482 // Build a sequence of copy-to-reg nodes chained together with token
2483 // chain and flag operands which copy the outgoing args into registers.
2484 // The InFlag in necessary since all emitted instructions must be
2488 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2489 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2490 RegsToPass[i].second, InFlag);
2491 InFlag = Chain.getValue(1);
2494 // Add argument registers to the end of the list so that they are
2495 // known live into the call.
2496 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2497 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2498 RegsToPass[i].second.getValueType()));
2500 // Add a register mask operand representing the call-preserved registers.
2501 const TargetRegisterInfo *TRI =
2502 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
2503 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv);
2504 assert(Mask && "Missing call preserved mask for calling convention");
2505 if (Subtarget.inMips16HardFloat()) {
2506 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2507 llvm::StringRef Sym = G->getGlobal()->getName();
2508 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2509 if (F && F->hasFnAttribute("__Mips16RetHelper")) {
2510 Mask = MipsRegisterInfo::getMips16RetHelperMask();
2514 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2516 if (InFlag.getNode())
2517 Ops.push_back(InFlag);
2520 /// LowerCall - functions arguments are copied from virtual regs to
2521 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2523 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2524 SmallVectorImpl<SDValue> &InVals) const {
2525 SelectionDAG &DAG = CLI.DAG;
2527 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2528 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2529 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2530 SDValue Chain = CLI.Chain;
2531 SDValue Callee = CLI.Callee;
2532 bool &IsTailCall = CLI.IsTailCall;
2533 CallingConv::ID CallConv = CLI.CallConv;
2534 bool IsVarArg = CLI.IsVarArg;
2536 MachineFunction &MF = DAG.getMachineFunction();
2537 MachineFrameInfo *MFI = MF.getFrameInfo();
2538 const TargetFrameLowering *TFL = MF.getSubtarget().getFrameLowering();
2539 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2540 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2542 // Analyze operands of the call, assigning locations to each operand.
2543 SmallVector<CCValAssign, 16> ArgLocs;
2544 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2546 MipsCC::SpecialCallingConvType SpecialCallingConv =
2547 getSpecialCallingConv(Callee);
2548 MipsCC MipsCCInfo(CallConv, Subtarget, CCInfo, SpecialCallingConv);
2550 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg,
2551 Subtarget.abiUsesSoftFloat(),
2552 Callee.getNode(), CLI.getArgs());
2554 // Get a count of how many bytes are to be pushed on the stack.
2555 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2557 // Check if it's really possible to do a tail call.
2560 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2561 *MF.getInfo<MipsFunctionInfo>());
2563 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2564 report_fatal_error("failed to perform tail call elimination on a call "
2565 "site marked musttail");
2570 // Chain is the output chain of the last Load/Store or CopyToReg node.
2571 // ByValChain is the output chain of the last Memcpy node created for copying
2572 // byval arguments to the stack.
2573 unsigned StackAlignment = TFL->getStackAlignment();
2574 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2575 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2578 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL);
2580 SDValue StackPtr = DAG.getCopyFromReg(
2581 Chain, DL, Subtarget.isABI_N64() ? Mips::SP_64 : Mips::SP,
2584 // With EABI is it possible to have 16 args on registers.
2585 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2586 SmallVector<SDValue, 8> MemOpChains;
2587 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2589 // Walk the register/memloc assignments, inserting copies/loads.
2590 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2591 SDValue Arg = OutVals[i];
2592 CCValAssign &VA = ArgLocs[i];
2593 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2594 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2597 if (Flags.isByVal()) {
2598 assert(Flags.getByValSize() &&
2599 "ByVal args of size 0 should have been ignored by front-end.");
2600 assert(ByValArg != MipsCCInfo.byval_end());
2601 assert(!IsTailCall &&
2602 "Do not tail-call optimize if there is a byval argument.");
2603 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2604 MipsCCInfo, *ByValArg, Flags, Subtarget.isLittle());
2609 // Promote the value if needed.
2610 switch (VA.getLocInfo()) {
2611 default: llvm_unreachable("Unknown loc info!");
2612 case CCValAssign::Full:
2613 if (VA.isRegLoc()) {
2614 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2615 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2616 (ValVT == MVT::i64 && LocVT == MVT::f64))
2617 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2618 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2619 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2620 Arg, DAG.getConstant(0, MVT::i32));
2621 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2622 Arg, DAG.getConstant(1, MVT::i32));
2623 if (!Subtarget.isLittle())
2625 unsigned LocRegLo = VA.getLocReg();
2626 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2627 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2628 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2633 case CCValAssign::SExt:
2634 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2636 case CCValAssign::ZExt:
2637 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2639 case CCValAssign::AExt:
2640 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2644 // Arguments that can be passed on register must be kept at
2645 // RegsToPass vector
2646 if (VA.isRegLoc()) {
2647 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2651 // Register can't get to this point...
2652 assert(VA.isMemLoc());
2654 // emit ISD::STORE whichs stores the
2655 // parameter value to a stack Location
2656 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2657 Chain, Arg, DL, IsTailCall, DAG));
2660 // Transform all store nodes into one single node because all store
2661 // nodes are independent of each other.
2662 if (!MemOpChains.empty())
2663 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2665 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2666 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2667 // node so that legalize doesn't hack it.
2669 (Subtarget.isABI_N64() || IsPIC); // true if calls are translated to
2671 bool GlobalOrExternal = false, InternalLinkage = false;
2673 EVT Ty = Callee.getValueType();
2675 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2677 const GlobalValue *Val = G->getGlobal();
2678 InternalLinkage = Val->hasInternalLinkage();
2680 if (InternalLinkage)
2681 Callee = getAddrLocal(G, Ty, DAG,
2682 Subtarget.isABI_N32() || Subtarget.isABI_N64());
2684 Callee = getAddrGlobalLargeGOT(G, Ty, DAG, MipsII::MO_CALL_HI16,
2685 MipsII::MO_CALL_LO16, Chain,
2686 FuncInfo->callPtrInfo(Val));
2688 Callee = getAddrGlobal(G, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2689 FuncInfo->callPtrInfo(Val));
2691 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0,
2692 MipsII::MO_NO_FLAG);
2693 GlobalOrExternal = true;
2695 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2696 const char *Sym = S->getSymbol();
2698 if (!Subtarget.isABI_N64() && !IsPIC) // !N64 && static
2699 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(),
2700 MipsII::MO_NO_FLAG);
2702 Callee = getAddrGlobalLargeGOT(S, Ty, DAG, MipsII::MO_CALL_HI16,
2703 MipsII::MO_CALL_LO16, Chain,
2704 FuncInfo->callPtrInfo(Sym));
2706 Callee = getAddrGlobal(S, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2707 FuncInfo->callPtrInfo(Sym));
2709 GlobalOrExternal = true;
2712 SmallVector<SDValue, 8> Ops(1, Chain);
2713 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2715 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2716 CLI, Callee, Chain);
2719 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops);
2721 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops);
2722 SDValue InFlag = Chain.getValue(1);
2724 // Create the CALLSEQ_END node.
2725 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2726 DAG.getIntPtrConstant(0, true), InFlag, DL);
2727 InFlag = Chain.getValue(1);
2729 // Handle result values, copying them out of physregs into vregs that we
2731 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2735 /// LowerCallResult - Lower the result values of a call into the
2736 /// appropriate copies out of appropriate physical registers.
2737 SDValue MipsTargetLowering::LowerCallResult(
2738 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
2739 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2740 SmallVectorImpl<SDValue> &InVals,
2741 TargetLowering::CallLoweringInfo &CLI) const {
2742 // Assign locations to each value returned by this call.
2743 SmallVector<CCValAssign, 16> RVLocs;
2744 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
2746 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI);
2748 // Copy all of the result registers out of their specified physreg.
2749 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2750 CCValAssign &VA = RVLocs[i];
2751 assert(VA.isRegLoc() && "Can only return in registers!");
2753 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2754 RVLocs[i].getLocVT(), InFlag);
2755 Chain = Val.getValue(1);
2756 InFlag = Val.getValue(2);
2758 if (VA.isUpperBitsInLoc()) {
2759 unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits();
2760 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
2762 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
2764 Shift, DL, VA.getLocVT(), Val,
2765 DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT()));
2768 switch (VA.getLocInfo()) {
2770 llvm_unreachable("Unknown loc info!");
2771 case CCValAssign::Full:
2773 case CCValAssign::BCvt:
2774 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2776 case CCValAssign::AExt:
2777 case CCValAssign::AExtUpper:
2778 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2780 case CCValAssign::ZExt:
2781 case CCValAssign::ZExtUpper:
2782 Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val,
2783 DAG.getValueType(VA.getValVT()));
2784 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2786 case CCValAssign::SExt:
2787 case CCValAssign::SExtUpper:
2788 Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val,
2789 DAG.getValueType(VA.getValVT()));
2790 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2794 InVals.push_back(Val);
2800 //===----------------------------------------------------------------------===//
2801 // Formal Arguments Calling Convention Implementation
2802 //===----------------------------------------------------------------------===//
2803 /// LowerFormalArguments - transform physical registers into virtual registers
2804 /// and generate load operations for arguments places on the stack.
2806 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2807 CallingConv::ID CallConv,
2809 const SmallVectorImpl<ISD::InputArg> &Ins,
2810 SDLoc DL, SelectionDAG &DAG,
2811 SmallVectorImpl<SDValue> &InVals)
2813 MachineFunction &MF = DAG.getMachineFunction();
2814 MachineFrameInfo *MFI = MF.getFrameInfo();
2815 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2817 MipsFI->setVarArgsFrameIndex(0);
2819 // Used with vargs to acumulate store chains.
2820 std::vector<SDValue> OutChains;
2822 // Assign locations to all of the incoming arguments.
2823 SmallVector<CCValAssign, 16> ArgLocs;
2824 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2826 MipsCC MipsCCInfo(CallConv, Subtarget, CCInfo);
2827 Function::const_arg_iterator FuncArg =
2828 DAG.getMachineFunction().getFunction()->arg_begin();
2829 bool UseSoftFloat = Subtarget.abiUsesSoftFloat();
2831 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg);
2832 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2833 MipsCCInfo.hasByValArg());
2835 unsigned CurArgIdx = 0;
2836 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2838 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2839 CCValAssign &VA = ArgLocs[i];
2840 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
2841 CurArgIdx = Ins[i].OrigArgIndex;
2842 EVT ValVT = VA.getValVT();
2843 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2844 bool IsRegLoc = VA.isRegLoc();
2846 if (Flags.isByVal()) {
2847 assert(Flags.getByValSize() &&
2848 "ByVal args of size 0 should have been ignored by front-end.");
2849 assert(ByValArg != MipsCCInfo.byval_end());
2850 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2851 MipsCCInfo, *ByValArg);
2856 // Arguments stored on registers
2858 MVT RegVT = VA.getLocVT();
2859 unsigned ArgReg = VA.getLocReg();
2860 const TargetRegisterClass *RC = getRegClassFor(RegVT);
2862 // Transform the arguments stored on
2863 // physical registers into virtual ones
2864 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2865 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2867 // If this is an 8 or 16-bit value, it has been passed promoted
2868 // to 32 bits. Insert an assert[sz]ext to capture this, then
2869 // truncate to the right size.
2870 if (VA.getLocInfo() != CCValAssign::Full) {
2871 unsigned Opcode = 0;
2872 if (VA.getLocInfo() == CCValAssign::SExt)
2873 Opcode = ISD::AssertSext;
2874 else if (VA.getLocInfo() == CCValAssign::ZExt)
2875 Opcode = ISD::AssertZext;
2877 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue,
2878 DAG.getValueType(ValVT));
2879 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue);
2882 // Handle floating point arguments passed in integer registers and
2883 // long double arguments passed in floating point registers.
2884 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2885 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
2886 (RegVT == MVT::f64 && ValVT == MVT::i64))
2887 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
2888 else if (Subtarget.isABI_O32() && RegVT == MVT::i32 &&
2889 ValVT == MVT::f64) {
2890 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
2891 getNextIntArgReg(ArgReg), RC);
2892 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
2893 if (!Subtarget.isLittle())
2894 std::swap(ArgValue, ArgValue2);
2895 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
2896 ArgValue, ArgValue2);
2899 InVals.push_back(ArgValue);
2900 } else { // VA.isRegLoc()
2903 assert(VA.isMemLoc());
2905 // The stack pointer offset is relative to the caller stack frame.
2906 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2907 VA.getLocMemOffset(), true);
2909 // Create load nodes to retrieve arguments from the stack
2910 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2911 SDValue Load = DAG.getLoad(ValVT, DL, Chain, FIN,
2912 MachinePointerInfo::getFixedStack(FI),
2913 false, false, false, 0);
2914 InVals.push_back(Load);
2915 OutChains.push_back(Load.getValue(1));
2919 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2920 // The mips ABIs for returning structs by value requires that we copy
2921 // the sret argument into $v0 for the return. Save the argument into
2922 // a virtual register so that we can access it from the return points.
2923 if (Ins[i].Flags.isSRet()) {
2924 unsigned Reg = MipsFI->getSRetReturnReg();
2926 Reg = MF.getRegInfo().createVirtualRegister(
2927 getRegClassFor(Subtarget.isABI_N64() ? MVT::i64 : MVT::i32));
2928 MipsFI->setSRetReturnReg(Reg);
2930 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]);
2931 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
2937 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG);
2939 // All stores are grouped in one node to allow the matching between
2940 // the size of Ins and InVals. This only happens when on varg functions
2941 if (!OutChains.empty()) {
2942 OutChains.push_back(Chain);
2943 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
2949 //===----------------------------------------------------------------------===//
2950 // Return Value Calling Convention Implementation
2951 //===----------------------------------------------------------------------===//
2954 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2955 MachineFunction &MF, bool IsVarArg,
2956 const SmallVectorImpl<ISD::OutputArg> &Outs,
2957 LLVMContext &Context) const {
2958 SmallVector<CCValAssign, 16> RVLocs;
2959 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
2960 return CCInfo.CheckReturn(Outs, RetCC_Mips);
2964 MipsTargetLowering::LowerReturn(SDValue Chain,
2965 CallingConv::ID CallConv, bool IsVarArg,
2966 const SmallVectorImpl<ISD::OutputArg> &Outs,
2967 const SmallVectorImpl<SDValue> &OutVals,
2968 SDLoc DL, SelectionDAG &DAG) const {
2969 // CCValAssign - represent the assignment of
2970 // the return value to a location
2971 SmallVector<CCValAssign, 16> RVLocs;
2972 MachineFunction &MF = DAG.getMachineFunction();
2974 // CCState - Info about the registers and stack slot.
2975 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
2976 MipsCC MipsCCInfo(CallConv, Subtarget, CCInfo);
2978 // Analyze return values.
2979 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
2982 SmallVector<SDValue, 4> RetOps(1, Chain);
2984 // Copy the result values into the output registers.
2985 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2986 SDValue Val = OutVals[i];
2987 CCValAssign &VA = RVLocs[i];
2988 assert(VA.isRegLoc() && "Can only return in registers!");
2989 bool UseUpperBits = false;
2991 switch (VA.getLocInfo()) {
2993 llvm_unreachable("Unknown loc info!");
2994 case CCValAssign::Full:
2996 case CCValAssign::BCvt:
2997 Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val);
2999 case CCValAssign::AExtUpper:
3000 UseUpperBits = true;
3002 case CCValAssign::AExt:
3003 Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val);
3005 case CCValAssign::ZExtUpper:
3006 UseUpperBits = true;
3008 case CCValAssign::ZExt:
3009 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val);
3011 case CCValAssign::SExtUpper:
3012 UseUpperBits = true;
3014 case CCValAssign::SExt:
3015 Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val);
3020 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
3021 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3023 ISD::SHL, DL, VA.getLocVT(), Val,
3024 DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT()));
3027 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
3029 // Guarantee that all emitted copies are stuck together with flags.
3030 Flag = Chain.getValue(1);
3031 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3034 // The mips ABIs for returning structs by value requires that we copy
3035 // the sret argument into $v0 for the return. We saved the argument into
3036 // a virtual register in the entry block, so now we copy the value out
3038 if (MF.getFunction()->hasStructRetAttr()) {
3039 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3040 unsigned Reg = MipsFI->getSRetReturnReg();
3043 llvm_unreachable("sret virtual register not created in the entry block");
3044 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
3045 unsigned V0 = Subtarget.isABI_N64() ? Mips::V0_64 : Mips::V0;
3047 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
3048 Flag = Chain.getValue(1);
3049 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
3052 RetOps[0] = Chain; // Update chain.
3054 // Add the flag if we have it.
3056 RetOps.push_back(Flag);
3058 // Return on Mips is always a "jr $ra"
3059 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps);
3062 //===----------------------------------------------------------------------===//
3063 // Mips Inline Assembly Support
3064 //===----------------------------------------------------------------------===//
3066 /// getConstraintType - Given a constraint letter, return the type of
3067 /// constraint it is for this target.
3068 MipsTargetLowering::ConstraintType MipsTargetLowering::
3069 getConstraintType(const std::string &Constraint) const
3071 // Mips specific constraints
3072 // GCC config/mips/constraints.md
3074 // 'd' : An address register. Equivalent to r
3075 // unless generating MIPS16 code.
3076 // 'y' : Equivalent to r; retained for
3077 // backwards compatibility.
3078 // 'c' : A register suitable for use in an indirect
3079 // jump. This will always be $25 for -mabicalls.
3080 // 'l' : The lo register. 1 word storage.
3081 // 'x' : The hilo register pair. Double word storage.
3082 if (Constraint.size() == 1) {
3083 switch (Constraint[0]) {
3091 return C_RegisterClass;
3096 return TargetLowering::getConstraintType(Constraint);
3099 /// Examine constraint type and operand type and determine a weight value.
3100 /// This object must already have been set up with the operand type
3101 /// and the current alternative constraint selected.
3102 TargetLowering::ConstraintWeight
3103 MipsTargetLowering::getSingleConstraintMatchWeight(
3104 AsmOperandInfo &info, const char *constraint) const {
3105 ConstraintWeight weight = CW_Invalid;
3106 Value *CallOperandVal = info.CallOperandVal;
3107 // If we don't have a value, we can't do a match,
3108 // but allow it at the lowest weight.
3109 if (!CallOperandVal)
3111 Type *type = CallOperandVal->getType();
3112 // Look at the constraint type.
3113 switch (*constraint) {
3115 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3119 if (type->isIntegerTy())
3120 weight = CW_Register;
3122 case 'f': // FPU or MSA register
3123 if (Subtarget.hasMSA() && type->isVectorTy() &&
3124 cast<VectorType>(type)->getBitWidth() == 128)
3125 weight = CW_Register;
3126 else if (type->isFloatTy())
3127 weight = CW_Register;
3129 case 'c': // $25 for indirect jumps
3130 case 'l': // lo register
3131 case 'x': // hilo register pair
3132 if (type->isIntegerTy())
3133 weight = CW_SpecificReg;
3135 case 'I': // signed 16 bit immediate
3136 case 'J': // integer zero
3137 case 'K': // unsigned 16 bit immediate
3138 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3139 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3140 case 'O': // signed 15 bit immediate (+- 16383)
3141 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3142 if (isa<ConstantInt>(CallOperandVal))
3143 weight = CW_Constant;
3152 /// This is a helper function to parse a physical register string and split it
3153 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
3154 /// that is returned indicates whether parsing was successful. The second flag
3155 /// is true if the numeric part exists.
3156 static std::pair<bool, bool>
3157 parsePhysicalReg(StringRef C, std::string &Prefix,
3158 unsigned long long &Reg) {
3159 if (C.front() != '{' || C.back() != '}')
3160 return std::make_pair(false, false);
3162 // Search for the first numeric character.
3163 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
3164 I = std::find_if(B, E, std::ptr_fun(isdigit));
3166 Prefix.assign(B, I - B);
3168 // The second flag is set to false if no numeric characters were found.
3170 return std::make_pair(true, false);
3172 // Parse the numeric characters.
3173 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
3177 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
3178 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const {
3179 const TargetRegisterInfo *TRI =
3180 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
3181 const TargetRegisterClass *RC;
3183 unsigned long long Reg;
3185 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
3188 return std::make_pair(0U, nullptr);
3190 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
3191 // No numeric characters follow "hi" or "lo".
3193 return std::make_pair(0U, nullptr);
3195 RC = TRI->getRegClass(Prefix == "hi" ?
3196 Mips::HI32RegClassID : Mips::LO32RegClassID);
3197 return std::make_pair(*(RC->begin()), RC);
3198 } else if (Prefix.compare(0, 4, "$msa") == 0) {
3199 // Parse $msa(ir|csr|access|save|modify|request|map|unmap)
3201 // No numeric characters follow the name.
3203 return std::make_pair(0U, nullptr);
3205 Reg = StringSwitch<unsigned long long>(Prefix)
3206 .Case("$msair", Mips::MSAIR)
3207 .Case("$msacsr", Mips::MSACSR)
3208 .Case("$msaaccess", Mips::MSAAccess)
3209 .Case("$msasave", Mips::MSASave)
3210 .Case("$msamodify", Mips::MSAModify)
3211 .Case("$msarequest", Mips::MSARequest)
3212 .Case("$msamap", Mips::MSAMap)
3213 .Case("$msaunmap", Mips::MSAUnmap)
3217 return std::make_pair(0U, nullptr);
3219 RC = TRI->getRegClass(Mips::MSACtrlRegClassID);
3220 return std::make_pair(Reg, RC);
3224 return std::make_pair(0U, nullptr);
3226 if (Prefix == "$f") { // Parse $f0-$f31.
3227 // If the size of FP registers is 64-bit or Reg is an even number, select
3228 // the 64-bit register class. Otherwise, select the 32-bit register class.
3229 if (VT == MVT::Other)
3230 VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
3232 RC = getRegClassFor(VT);
3234 if (RC == &Mips::AFGR64RegClass) {
3235 assert(Reg % 2 == 0);
3238 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7.
3239 RC = TRI->getRegClass(Mips::FCCRegClassID);
3240 else if (Prefix == "$w") { // Parse $w0-$w31.
3241 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT);
3242 } else { // Parse $0-$31.
3243 assert(Prefix == "$");
3244 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
3247 assert(Reg < RC->getNumRegs());
3248 return std::make_pair(*(RC->begin() + Reg), RC);
3251 /// Given a register class constraint, like 'r', if this corresponds directly
3252 /// to an LLVM register class, return a register of 0 and the register class
3254 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
3255 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const
3257 if (Constraint.size() == 1) {
3258 switch (Constraint[0]) {
3259 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3260 case 'y': // Same as 'r'. Exists for compatibility.
3262 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3263 if (Subtarget.inMips16Mode())
3264 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3265 return std::make_pair(0U, &Mips::GPR32RegClass);
3267 if (VT == MVT::i64 && !Subtarget.isGP64bit())
3268 return std::make_pair(0U, &Mips::GPR32RegClass);
3269 if (VT == MVT::i64 && Subtarget.isGP64bit())
3270 return std::make_pair(0U, &Mips::GPR64RegClass);
3271 // This will generate an error message
3272 return std::make_pair(0U, nullptr);
3273 case 'f': // FPU or MSA register
3274 if (VT == MVT::v16i8)
3275 return std::make_pair(0U, &Mips::MSA128BRegClass);
3276 else if (VT == MVT::v8i16 || VT == MVT::v8f16)
3277 return std::make_pair(0U, &Mips::MSA128HRegClass);
3278 else if (VT == MVT::v4i32 || VT == MVT::v4f32)
3279 return std::make_pair(0U, &Mips::MSA128WRegClass);
3280 else if (VT == MVT::v2i64 || VT == MVT::v2f64)
3281 return std::make_pair(0U, &Mips::MSA128DRegClass);
3282 else if (VT == MVT::f32)
3283 return std::make_pair(0U, &Mips::FGR32RegClass);
3284 else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) {
3285 if (Subtarget.isFP64bit())
3286 return std::make_pair(0U, &Mips::FGR64RegClass);
3287 return std::make_pair(0U, &Mips::AFGR64RegClass);
3290 case 'c': // register suitable for indirect jump
3292 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
3293 assert(VT == MVT::i64 && "Unexpected type.");
3294 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
3295 case 'l': // register suitable for indirect jump
3297 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
3298 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
3299 case 'x': // register suitable for indirect jump
3300 // Fixme: Not triggering the use of both hi and low
3301 // This will generate an error message
3302 return std::make_pair(0U, nullptr);
3306 std::pair<unsigned, const TargetRegisterClass *> R;
3307 R = parseRegForInlineAsmConstraint(Constraint, VT);
3312 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3315 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3316 /// vector. If it is invalid, don't add anything to Ops.
3317 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3318 std::string &Constraint,
3319 std::vector<SDValue>&Ops,
3320 SelectionDAG &DAG) const {
3323 // Only support length 1 constraints for now.
3324 if (Constraint.length() > 1) return;
3326 char ConstraintLetter = Constraint[0];
3327 switch (ConstraintLetter) {
3328 default: break; // This will fall through to the generic implementation
3329 case 'I': // Signed 16 bit constant
3330 // If this fails, the parent routine will give an error
3331 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3332 EVT Type = Op.getValueType();
3333 int64_t Val = C->getSExtValue();
3334 if (isInt<16>(Val)) {
3335 Result = DAG.getTargetConstant(Val, Type);
3340 case 'J': // integer zero
3341 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3342 EVT Type = Op.getValueType();
3343 int64_t Val = C->getZExtValue();
3345 Result = DAG.getTargetConstant(0, Type);
3350 case 'K': // unsigned 16 bit immediate
3351 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3352 EVT Type = Op.getValueType();
3353 uint64_t Val = (uint64_t)C->getZExtValue();
3354 if (isUInt<16>(Val)) {
3355 Result = DAG.getTargetConstant(Val, Type);
3360 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3361 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3362 EVT Type = Op.getValueType();
3363 int64_t Val = C->getSExtValue();
3364 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3365 Result = DAG.getTargetConstant(Val, Type);
3370 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3371 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3372 EVT Type = Op.getValueType();
3373 int64_t Val = C->getSExtValue();
3374 if ((Val >= -65535) && (Val <= -1)) {
3375 Result = DAG.getTargetConstant(Val, Type);
3380 case 'O': // signed 15 bit immediate
3381 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3382 EVT Type = Op.getValueType();
3383 int64_t Val = C->getSExtValue();
3384 if ((isInt<15>(Val))) {
3385 Result = DAG.getTargetConstant(Val, Type);
3390 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3391 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3392 EVT Type = Op.getValueType();
3393 int64_t Val = C->getSExtValue();
3394 if ((Val <= 65535) && (Val >= 1)) {
3395 Result = DAG.getTargetConstant(Val, Type);
3402 if (Result.getNode()) {
3403 Ops.push_back(Result);
3407 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3410 bool MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3412 // No global is ever allowed as a base.
3417 case 0: // "r+i" or just "i", depending on HasBaseReg.
3420 if (!AM.HasBaseReg) // allow "r+i".
3422 return false; // disallow "r+r" or "r+r+i".
3431 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3432 // The Mips target isn't yet aware of offsets.
3436 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3438 bool IsMemset, bool ZeroMemset,
3440 MachineFunction &MF) const {
3441 if (Subtarget.hasMips64())
3447 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3448 if (VT != MVT::f32 && VT != MVT::f64)
3450 if (Imm.isNegZero())
3452 return Imm.isZero();
3455 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3456 if (Subtarget.isABI_N64())
3457 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3459 return TargetLowering::getJumpTableEncoding();
3462 /// This function returns true if CallSym is a long double emulation routine.
3463 static bool isF128SoftLibCall(const char *CallSym) {
3464 const char *const LibCalls[] =
3465 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2",
3466 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi",
3467 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf",
3468 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2",
3469 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3",
3470 "__trunctfdf2", "__trunctfsf2", "__unordtf2",
3471 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl",
3472 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl",
3475 const char *const *End = LibCalls + array_lengthof(LibCalls);
3477 // Check that LibCalls is sorted alphabetically.
3478 MipsTargetLowering::LTStr Comp;
3481 for (const char *const *I = LibCalls; I < End - 1; ++I)
3482 assert(Comp(*I, *(I + 1)));
3485 return std::binary_search(LibCalls, End, CallSym, Comp);
3488 /// This function returns true if Ty is fp128 or i128 which was originally a
3490 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) {
3491 if (Ty->isFP128Ty())
3494 const ExternalSymbolSDNode *ES =
3495 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
3497 // If the Ty is i128 and the function being called is a long double emulation
3498 // routine, then the original type is f128.
3499 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
3502 MipsTargetLowering::MipsCC::SpecialCallingConvType
3503 MipsTargetLowering::getSpecialCallingConv(SDValue Callee) const {
3504 MipsCC::SpecialCallingConvType SpecialCallingConv =
3505 MipsCC::NoSpecialCallingConv;
3506 if (Subtarget.inMips16HardFloat()) {
3507 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3508 llvm::StringRef Sym = G->getGlobal()->getName();
3509 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
3510 if (F && F->hasFnAttribute("__Mips16RetHelper")) {
3511 SpecialCallingConv = MipsCC::Mips16RetHelperConv;
3515 return SpecialCallingConv;
3518 MipsTargetLowering::MipsCC::MipsCC(
3519 CallingConv::ID CC, const MipsSubtarget &Subtarget_, CCState &Info,
3520 MipsCC::SpecialCallingConvType SpecialCallingConv_)
3521 : CCInfo(Info), CallConv(CC), Subtarget(Subtarget_),
3522 SpecialCallingConv(SpecialCallingConv_) {
3523 // Pre-allocate reserved argument area.
3524 CCInfo.AllocateStack(reservedArgArea(), 1);
3528 void MipsTargetLowering::MipsCC::
3529 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
3530 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode,
3531 std::vector<ArgListEntry> &FuncArgs) {
3532 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
3533 "CallingConv::Fast shouldn't be used for vararg functions.");
3535 unsigned NumOpnds = Args.size();
3536 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
3538 for (unsigned I = 0; I != NumOpnds; ++I) {
3539 MVT ArgVT = Args[I].VT;
3540 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3543 if (ArgFlags.isByVal()) {
3544 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3548 if (IsVarArg && !Args[I].IsFixed)
3549 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3551 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode,
3553 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo);
3558 dbgs() << "Call operand #" << I << " has unhandled type "
3559 << EVT(ArgVT).getEVTString();
3561 llvm_unreachable(nullptr);
3566 void MipsTargetLowering::MipsCC::
3567 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args,
3568 bool IsSoftFloat, Function::const_arg_iterator FuncArg) {
3569 unsigned NumArgs = Args.size();
3570 llvm::CCAssignFn *FixedFn = fixedArgFn();
3571 unsigned CurArgIdx = 0;
3573 for (unsigned I = 0; I != NumArgs; ++I) {
3574 MVT ArgVT = Args[I].VT;
3575 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3576 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx);
3577 CurArgIdx = Args[I].OrigArgIndex;
3579 if (ArgFlags.isByVal()) {
3580 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3584 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), nullptr, IsSoftFloat);
3586 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo))
3590 dbgs() << "Formal Arg #" << I << " has unhandled type "
3591 << EVT(ArgVT).getEVTString();
3593 llvm_unreachable(nullptr);
3597 void MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3599 CCValAssign::LocInfo LocInfo,
3600 ISD::ArgFlagsTy ArgFlags) {
3601 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3603 struct ByValArgInfo ByVal;
3604 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3605 unsigned ByValSize =
3606 RoundUpToAlignment(ArgFlags.getByValSize(), RegSizeInBytes);
3607 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSizeInBytes),
3608 RegSizeInBytes * 2);
3610 if (useRegsForByval())
3611 allocateRegs(ByVal, ByValSize, Align);
3613 // Allocate space on caller's stack.
3615 CCInfo.AllocateStack(ByValSize - RegSizeInBytes * ByVal.NumRegs, Align);
3616 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3618 ByValArgs.push_back(ByVal);
3621 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
3622 return (Subtarget.isABI_O32() && (CallConv != CallingConv::Fast)) ? 16 : 0;
3625 const ArrayRef<MCPhysReg> MipsTargetLowering::MipsCC::intArgRegs() const {
3626 if (Subtarget.isABI_O32())
3627 return makeArrayRef(O32IntRegs);
3628 return makeArrayRef(Mips64IntRegs);
3631 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
3632 if (CallConv == CallingConv::Fast)
3633 return CC_Mips_FastCC;
3635 if (SpecialCallingConv == Mips16RetHelperConv)
3636 return CC_Mips16RetHelper;
3637 return Subtarget.isABI_O32()
3638 ? (Subtarget.isFP64bit() ? CC_MipsO32_FP64 : CC_MipsO32_FP32)
3642 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
3643 return Subtarget.isABI_O32()
3644 ? (Subtarget.isFP64bit() ? CC_MipsO32_FP64 : CC_MipsO32_FP32)
3648 const MCPhysReg *MipsTargetLowering::MipsCC::shadowRegs() const {
3649 return Subtarget.isABI_O32() ? O32IntRegs : Mips64DPRegs;
3652 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3655 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3656 const ArrayRef<MCPhysReg> IntArgRegs = intArgRegs();
3657 const MCPhysReg *ShadowRegs = shadowRegs();
3658 assert(!(ByValSize % RegSizeInBytes) && !(Align % RegSizeInBytes) &&
3659 "Byval argument's size and alignment should be a multiple of"
3663 CCInfo.getFirstUnallocated(IntArgRegs.data(), IntArgRegs.size());
3665 // If Align > RegSizeInBytes, the first arg register must be even.
3666 if ((Align > RegSizeInBytes) && (ByVal.FirstIdx % 2)) {
3667 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3671 // Mark the registers allocated.
3672 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < IntArgRegs.size());
3673 ByValSize -= RegSizeInBytes, ++I, ++ByVal.NumRegs)
3674 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3677 MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy,
3678 const SDNode *CallNode,
3679 bool IsSoftFloat) const {
3680 if (IsSoftFloat || Subtarget.isABI_O32())
3683 // Check if the original type was fp128.
3684 if (originalTypeIsF128(OrigTy, CallNode)) {
3685 assert(VT == MVT::i64);
3692 void MipsTargetLowering::
3693 copyByValRegs(SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains,
3694 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3695 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3696 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3697 MachineFunction &MF = DAG.getMachineFunction();
3698 MachineFrameInfo *MFI = MF.getFrameInfo();
3699 unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes();
3700 unsigned RegAreaSize = ByVal.NumRegs * GPRSizeInBytes;
3701 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3706 (int)CC.reservedArgArea() -
3707 (int)((CC.intArgRegs().size() - ByVal.FirstIdx) * GPRSizeInBytes);
3709 FrameObjOffset = ByVal.Address;
3711 // Create frame object.
3712 EVT PtrTy = getPointerTy();
3713 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3714 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3715 InVals.push_back(FIN);
3720 // Copy arg registers.
3721 MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8);
3722 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3724 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3725 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3726 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3727 unsigned Offset = I * GPRSizeInBytes;
3728 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3729 DAG.getConstant(Offset, PtrTy));
3730 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3731 StorePtr, MachinePointerInfo(FuncArg, Offset),
3733 OutChains.push_back(Store);
3737 // Copy byVal arg to registers and stack.
3738 void MipsTargetLowering::
3739 passByValArg(SDValue Chain, SDLoc DL,
3740 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
3741 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
3742 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3743 const MipsCC &CC, const ByValArgInfo &ByVal,
3744 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3745 unsigned ByValSizeInBytes = Flags.getByValSize();
3746 unsigned OffsetInBytes = 0; // From beginning of struct
3747 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3748 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes);
3749 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
3751 if (ByVal.NumRegs) {
3752 const ArrayRef<MCPhysReg> ArgRegs = CC.intArgRegs();
3753 bool LeftoverBytes = (ByVal.NumRegs * RegSizeInBytes > ByValSizeInBytes);
3756 // Copy words to registers.
3757 for (; I < ByVal.NumRegs - LeftoverBytes;
3758 ++I, OffsetInBytes += RegSizeInBytes) {
3759 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3760 DAG.getConstant(OffsetInBytes, PtrTy));
3761 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3762 MachinePointerInfo(), false, false, false,
3764 MemOpChains.push_back(LoadVal.getValue(1));
3765 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3766 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3769 // Return if the struct has been fully copied.
3770 if (ByValSizeInBytes == OffsetInBytes)
3773 // Copy the remainder of the byval argument with sub-word loads and shifts.
3774 if (LeftoverBytes) {
3775 assert((ByValSizeInBytes > OffsetInBytes) &&
3776 (ByValSizeInBytes < OffsetInBytes + RegSizeInBytes) &&
3777 "Size of the remainder should be smaller than RegSizeInBytes.");
3780 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0;
3781 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) {
3782 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes;
3784 if (RemainingSizeInBytes < LoadSizeInBytes)
3788 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3789 DAG.getConstant(OffsetInBytes, PtrTy));
3790 SDValue LoadVal = DAG.getExtLoad(
3791 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(),
3792 MVT::getIntegerVT(LoadSizeInBytes * 8), false, false, false,
3794 MemOpChains.push_back(LoadVal.getValue(1));
3796 // Shift the loaded value.
3800 Shamt = TotalBytesLoaded * 8;
3802 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8;
3804 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3805 DAG.getConstant(Shamt, MVT::i32));
3808 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3812 OffsetInBytes += LoadSizeInBytes;
3813 TotalBytesLoaded += LoadSizeInBytes;
3814 Alignment = std::min(Alignment, LoadSizeInBytes);
3817 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3818 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3823 // Copy remainder of byval arg to it with memcpy.
3824 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes;
3825 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3826 DAG.getConstant(OffsetInBytes, PtrTy));
3827 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3828 DAG.getIntPtrConstant(ByVal.Address));
3829 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, DAG.getConstant(MemCpySize, PtrTy),
3830 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false,
3831 MachinePointerInfo(), MachinePointerInfo());
3832 MemOpChains.push_back(Chain);
3835 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3836 const MipsCC &CC, SDValue Chain,
3837 SDLoc DL, SelectionDAG &DAG) const {
3838 const ArrayRef<MCPhysReg> ArgRegs = CC.intArgRegs();
3839 const CCState &CCInfo = CC.getCCInfo();
3840 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs.data(), ArgRegs.size());
3841 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3842 MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
3843 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3844 MachineFunction &MF = DAG.getMachineFunction();
3845 MachineFrameInfo *MFI = MF.getFrameInfo();
3846 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3848 // Offset of the first variable argument from stack pointer.
3851 if (ArgRegs.size() == Idx)
3853 RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSizeInBytes);
3855 VaArgOffset = (int)CC.reservedArgArea() -
3856 (int)(RegSizeInBytes * (ArgRegs.size() - Idx));
3858 // Record the frame index of the first variable argument
3859 // which is a value necessary to VASTART.
3860 int FI = MFI->CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
3861 MipsFI->setVarArgsFrameIndex(FI);
3863 // Copy the integer registers that have not been used for argument passing
3864 // to the argument register save area. For O32, the save area is allocated
3865 // in the caller's stack frame, while for N32/64, it is allocated in the
3866 // callee's stack frame.
3867 for (unsigned I = Idx; I < ArgRegs.size();
3868 ++I, VaArgOffset += RegSizeInBytes) {
3869 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3870 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3871 FI = MFI->CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
3872 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3873 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3874 MachinePointerInfo(), false, false, 0);
3875 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(
3877 OutChains.push_back(Store);