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 #define DEBUG_TYPE "mips-lower"
15 #include "MipsISelLowering.h"
16 #include "InstPrinter/MipsInstPrinter.h"
17 #include "MCTargetDesc/MipsBaseInfo.h"
18 #include "MipsMachineFunction.h"
19 #include "MipsSubtarget.h"
20 #include "MipsTargetMachine.h"
21 #include "MipsTargetObjectFile.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAGISel.h"
29 #include "llvm/CodeGen/ValueTypes.h"
30 #include "llvm/IR/CallingConv.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
41 STATISTIC(NumTailCalls, "Number of tail calls");
44 LargeGOT("mxgot", cl::Hidden,
45 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
48 NoZeroDivCheck("mno-check-zero-division", cl::Hidden,
49 cl::desc("MIPS: Don't trap on integer division by zero."),
52 static const uint16_t O32IntRegs[4] = {
53 Mips::A0, Mips::A1, Mips::A2, Mips::A3
56 static const uint16_t Mips64IntRegs[8] = {
57 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
58 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64
61 static const uint16_t Mips64DPRegs[8] = {
62 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
63 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
66 // If I is a shifted mask, set the size (Size) and the first bit of the
67 // mask (Pos), and return true.
68 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
69 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
70 if (!isShiftedMask_64(I))
73 Size = CountPopulation_64(I);
74 Pos = countTrailingZeros(I);
78 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
79 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
80 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
83 template<class NodeTy>
84 static SDValue getTargetNode(NodeTy *Node, EVT Ty, SelectionDAG &DAG,
86 llvm_unreachable("Unexpected node type.");
91 SDValue getTargetNode<GlobalAddressSDNode>(GlobalAddressSDNode *N, EVT Ty,
92 SelectionDAG &DAG, unsigned Flag) {
93 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag);
97 SDValue getTargetNode<ExternalSymbolSDNode>(ExternalSymbolSDNode *N, EVT Ty,
98 SelectionDAG &DAG, unsigned Flag) {
99 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
103 SDValue getTargetNode<BlockAddressSDNode>(BlockAddressSDNode *N, EVT Ty,
104 SelectionDAG &DAG, unsigned Flag) {
105 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
109 SDValue getTargetNode<JumpTableSDNode>(JumpTableSDNode *N, EVT Ty,
110 SelectionDAG &DAG, unsigned Flag) {
111 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
115 SDValue getTargetNode<ConstantPoolSDNode>(ConstantPoolSDNode *N, EVT Ty,
116 SelectionDAG &DAG, unsigned Flag) {
117 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
118 N->getOffset(), Flag);
121 template<class NodeTy>
122 static SDValue getAddrNonPIC(NodeTy *N, EVT Ty, SelectionDAG &DAG) {
124 SDValue Hi = getTargetNode(N, Ty, DAG, MipsII::MO_ABS_HI);
125 SDValue Lo = getTargetNode(N, Ty, DAG, MipsII::MO_ABS_LO);
126 return DAG.getNode(ISD::ADD, DL, Ty,
127 DAG.getNode(MipsISD::Hi, DL, Ty, Hi),
128 DAG.getNode(MipsISD::Lo, DL, Ty, Lo));
131 template<class NodeTy>
132 SDValue MipsTargetLowering::getAddrLocal(NodeTy *N, EVT Ty, SelectionDAG &DAG,
133 bool HasMips64) const {
135 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
136 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
137 getTargetNode(N, Ty, DAG, GOTFlag));
138 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT,
139 MachinePointerInfo::getGOT(), false, false, false,
141 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
142 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty,
143 getTargetNode(N, Ty, DAG, LoFlag));
144 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo);
147 template<class NodeTy>
148 SDValue MipsTargetLowering::getAddrGlobal(NodeTy *N, EVT Ty, SelectionDAG &DAG,
149 unsigned Flag) const {
151 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
152 getTargetNode(N, Ty, DAG, Flag));
153 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt,
154 MachinePointerInfo::getGOT(), false, false, false, 0);
157 template<class NodeTy>
158 SDValue MipsTargetLowering::getAddrGlobalLargeGOT(NodeTy *N, EVT Ty,
161 unsigned LoFlag) const {
163 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty,
164 getTargetNode(N, Ty, DAG, HiFlag));
165 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, getGlobalReg(DAG, Ty));
166 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi,
167 getTargetNode(N, Ty, DAG, LoFlag));
168 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper,
169 MachinePointerInfo::getGOT(), false, false, false, 0);
172 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
174 case MipsISD::JmpLink: return "MipsISD::JmpLink";
175 case MipsISD::TailCall: return "MipsISD::TailCall";
176 case MipsISD::Hi: return "MipsISD::Hi";
177 case MipsISD::Lo: return "MipsISD::Lo";
178 case MipsISD::GPRel: return "MipsISD::GPRel";
179 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
180 case MipsISD::Ret: return "MipsISD::Ret";
181 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
182 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
183 case MipsISD::FPCmp: return "MipsISD::FPCmp";
184 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
185 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
186 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP";
187 case MipsISD::ExtractLOHI: return "MipsISD::ExtractLOHI";
188 case MipsISD::InsertLOHI: return "MipsISD::InsertLOHI";
189 case MipsISD::Mult: return "MipsISD::Mult";
190 case MipsISD::Multu: return "MipsISD::Multu";
191 case MipsISD::MAdd: return "MipsISD::MAdd";
192 case MipsISD::MAddu: return "MipsISD::MAddu";
193 case MipsISD::MSub: return "MipsISD::MSub";
194 case MipsISD::MSubu: return "MipsISD::MSubu";
195 case MipsISD::DivRem: return "MipsISD::DivRem";
196 case MipsISD::DivRemU: return "MipsISD::DivRemU";
197 case MipsISD::DivRem16: return "MipsISD::DivRem16";
198 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
199 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
200 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
201 case MipsISD::Wrapper: return "MipsISD::Wrapper";
202 case MipsISD::Sync: return "MipsISD::Sync";
203 case MipsISD::Ext: return "MipsISD::Ext";
204 case MipsISD::Ins: return "MipsISD::Ins";
205 case MipsISD::LWL: return "MipsISD::LWL";
206 case MipsISD::LWR: return "MipsISD::LWR";
207 case MipsISD::SWL: return "MipsISD::SWL";
208 case MipsISD::SWR: return "MipsISD::SWR";
209 case MipsISD::LDL: return "MipsISD::LDL";
210 case MipsISD::LDR: return "MipsISD::LDR";
211 case MipsISD::SDL: return "MipsISD::SDL";
212 case MipsISD::SDR: return "MipsISD::SDR";
213 case MipsISD::EXTP: return "MipsISD::EXTP";
214 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
215 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
216 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
217 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
218 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
219 case MipsISD::SHILO: return "MipsISD::SHILO";
220 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
221 case MipsISD::MULT: return "MipsISD::MULT";
222 case MipsISD::MULTU: return "MipsISD::MULTU";
223 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
224 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
225 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
226 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
227 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP";
228 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP";
229 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP";
230 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP";
231 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP";
232 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO";
233 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO";
234 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO";
235 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO";
236 case MipsISD::VCEQ: return "MipsISD::VCEQ";
237 case MipsISD::VCLE_S: return "MipsISD::VCLE_S";
238 case MipsISD::VCLE_U: return "MipsISD::VCLE_U";
239 case MipsISD::VCLT_S: return "MipsISD::VCLT_S";
240 case MipsISD::VCLT_U: return "MipsISD::VCLT_U";
241 case MipsISD::VSMAX: return "MipsISD::VSMAX";
242 case MipsISD::VSMIN: return "MipsISD::VSMIN";
243 case MipsISD::VUMAX: return "MipsISD::VUMAX";
244 case MipsISD::VUMIN: return "MipsISD::VUMIN";
245 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
246 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
247 case MipsISD::VNOR: return "MipsISD::VNOR";
248 case MipsISD::VSHF: return "MipsISD::VSHF";
249 case MipsISD::SHF: return "MipsISD::SHF";
250 case MipsISD::ILVEV: return "MipsISD::ILVEV";
251 case MipsISD::ILVOD: return "MipsISD::ILVOD";
252 case MipsISD::ILVL: return "MipsISD::ILVL";
253 case MipsISD::ILVR: return "MipsISD::ILVR";
254 case MipsISD::PCKEV: return "MipsISD::PCKEV";
255 case MipsISD::PCKOD: return "MipsISD::PCKOD";
256 default: return NULL;
261 MipsTargetLowering(MipsTargetMachine &TM)
262 : TargetLowering(TM, new MipsTargetObjectFile()),
263 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
264 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
265 IsO32(Subtarget->isABI_O32()) {
266 // Mips does not have i1 type, so use i32 for
267 // setcc operations results (slt, sgt, ...).
268 setBooleanContents(ZeroOrOneBooleanContent);
269 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
271 // Load extented operations for i1 types must be promoted
272 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
273 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
274 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
276 // MIPS doesn't have extending float->double load/store
277 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
278 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
280 // Used by legalize types to correctly generate the setcc result.
281 // Without this, every float setcc comes with a AND/OR with the result,
282 // we don't want this, since the fpcmp result goes to a flag register,
283 // which is used implicitly by brcond and select operations.
284 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
286 // Mips Custom Operations
287 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
288 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
289 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
290 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
291 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
292 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
293 setOperationAction(ISD::SELECT, MVT::f32, Custom);
294 setOperationAction(ISD::SELECT, MVT::f64, Custom);
295 setOperationAction(ISD::SELECT, MVT::i32, Custom);
296 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
297 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
298 setOperationAction(ISD::SETCC, MVT::f32, Custom);
299 setOperationAction(ISD::SETCC, MVT::f64, Custom);
300 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
301 setOperationAction(ISD::VASTART, MVT::Other, Custom);
302 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
303 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
304 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
306 if (!TM.Options.NoNaNsFPMath) {
307 setOperationAction(ISD::FABS, MVT::f32, Custom);
308 setOperationAction(ISD::FABS, MVT::f64, Custom);
312 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
313 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
314 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
315 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
316 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
317 setOperationAction(ISD::SELECT, MVT::i64, Custom);
318 setOperationAction(ISD::LOAD, MVT::i64, Custom);
319 setOperationAction(ISD::STORE, MVT::i64, Custom);
320 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
324 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
325 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
326 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
329 setOperationAction(ISD::ADD, MVT::i32, Custom);
331 setOperationAction(ISD::ADD, MVT::i64, Custom);
333 setOperationAction(ISD::SDIV, MVT::i32, Expand);
334 setOperationAction(ISD::SREM, MVT::i32, Expand);
335 setOperationAction(ISD::UDIV, MVT::i32, Expand);
336 setOperationAction(ISD::UREM, MVT::i32, Expand);
337 setOperationAction(ISD::SDIV, MVT::i64, Expand);
338 setOperationAction(ISD::SREM, MVT::i64, Expand);
339 setOperationAction(ISD::UDIV, MVT::i64, Expand);
340 setOperationAction(ISD::UREM, MVT::i64, Expand);
342 // Operations not directly supported by Mips.
343 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
344 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
345 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
346 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
347 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
348 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
349 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
350 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
351 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
352 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
353 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
354 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
355 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
356 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
357 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
358 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
361 setOperationAction(ISD::ROTL, MVT::i32, Expand);
362 setOperationAction(ISD::ROTL, MVT::i64, Expand);
363 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
364 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
366 if (!Subtarget->hasMips32r2())
367 setOperationAction(ISD::ROTR, MVT::i32, Expand);
369 if (!Subtarget->hasMips64r2())
370 setOperationAction(ISD::ROTR, MVT::i64, Expand);
372 setOperationAction(ISD::FSIN, MVT::f32, Expand);
373 setOperationAction(ISD::FSIN, MVT::f64, Expand);
374 setOperationAction(ISD::FCOS, MVT::f32, Expand);
375 setOperationAction(ISD::FCOS, MVT::f64, Expand);
376 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
377 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
378 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
379 setOperationAction(ISD::FPOW, MVT::f32, Expand);
380 setOperationAction(ISD::FPOW, MVT::f64, Expand);
381 setOperationAction(ISD::FLOG, MVT::f32, Expand);
382 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
383 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
384 setOperationAction(ISD::FEXP, MVT::f32, Expand);
385 setOperationAction(ISD::FMA, MVT::f32, Expand);
386 setOperationAction(ISD::FMA, MVT::f64, Expand);
387 setOperationAction(ISD::FREM, MVT::f32, Expand);
388 setOperationAction(ISD::FREM, MVT::f64, Expand);
390 if (!TM.Options.NoNaNsFPMath) {
391 setOperationAction(ISD::FNEG, MVT::f32, Expand);
392 setOperationAction(ISD::FNEG, MVT::f64, Expand);
395 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
397 setOperationAction(ISD::VAARG, MVT::Other, Expand);
398 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
399 setOperationAction(ISD::VAEND, MVT::Other, Expand);
401 // Use the default for now
402 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
403 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
405 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
406 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
407 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
408 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
410 setInsertFencesForAtomic(true);
412 if (!Subtarget->hasSEInReg()) {
413 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
414 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
417 if (!Subtarget->hasBitCount()) {
418 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
419 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
422 if (!Subtarget->hasSwap()) {
423 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
424 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
428 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
429 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
430 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
431 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
434 setOperationAction(ISD::TRAP, MVT::Other, Legal);
436 setTargetDAGCombine(ISD::SDIVREM);
437 setTargetDAGCombine(ISD::UDIVREM);
438 setTargetDAGCombine(ISD::SELECT);
439 setTargetDAGCombine(ISD::AND);
440 setTargetDAGCombine(ISD::OR);
441 setTargetDAGCombine(ISD::ADD);
443 setMinFunctionAlignment(HasMips64 ? 3 : 2);
445 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP);
447 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0);
448 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1);
450 MaxStoresPerMemcpy = 16;
453 const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) {
454 if (TM.getSubtargetImpl()->inMips16Mode())
455 return llvm::createMips16TargetLowering(TM);
457 return llvm::createMipsSETargetLowering(TM);
460 EVT MipsTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
463 return VT.changeVectorElementTypeToInteger();
466 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
467 TargetLowering::DAGCombinerInfo &DCI,
468 const MipsSubtarget *Subtarget) {
469 if (DCI.isBeforeLegalizeOps())
472 EVT Ty = N->getValueType(0);
473 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
474 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
475 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
479 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
480 N->getOperand(0), N->getOperand(1));
481 SDValue InChain = DAG.getEntryNode();
482 SDValue InGlue = DivRem;
485 if (N->hasAnyUseOfValue(0)) {
486 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
488 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
489 InChain = CopyFromLo.getValue(1);
490 InGlue = CopyFromLo.getValue(2);
494 if (N->hasAnyUseOfValue(1)) {
495 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
497 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
503 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
505 default: llvm_unreachable("Unknown fp condition code!");
507 case ISD::SETOEQ: return Mips::FCOND_OEQ;
508 case ISD::SETUNE: return Mips::FCOND_UNE;
510 case ISD::SETOLT: return Mips::FCOND_OLT;
512 case ISD::SETOGT: return Mips::FCOND_OGT;
514 case ISD::SETOLE: return Mips::FCOND_OLE;
516 case ISD::SETOGE: return Mips::FCOND_OGE;
517 case ISD::SETULT: return Mips::FCOND_ULT;
518 case ISD::SETULE: return Mips::FCOND_ULE;
519 case ISD::SETUGT: return Mips::FCOND_UGT;
520 case ISD::SETUGE: return Mips::FCOND_UGE;
521 case ISD::SETUO: return Mips::FCOND_UN;
522 case ISD::SETO: return Mips::FCOND_OR;
524 case ISD::SETONE: return Mips::FCOND_ONE;
525 case ISD::SETUEQ: return Mips::FCOND_UEQ;
530 /// This function returns true if the floating point conditional branches and
531 /// conditional moves which use condition code CC should be inverted.
532 static bool invertFPCondCodeUser(Mips::CondCode CC) {
533 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
536 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
537 "Illegal Condition Code");
542 // Creates and returns an FPCmp node from a setcc node.
543 // Returns Op if setcc is not a floating point comparison.
544 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
545 // must be a SETCC node
546 if (Op.getOpcode() != ISD::SETCC)
549 SDValue LHS = Op.getOperand(0);
551 if (!LHS.getValueType().isFloatingPoint())
554 SDValue RHS = Op.getOperand(1);
557 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
558 // node if necessary.
559 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
561 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
562 DAG.getConstant(condCodeToFCC(CC), MVT::i32));
565 // Creates and returns a CMovFPT/F node.
566 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
567 SDValue False, SDLoc DL) {
568 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
569 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
570 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
572 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
573 True.getValueType(), True, FCC0, False, Cond);
576 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
577 TargetLowering::DAGCombinerInfo &DCI,
578 const MipsSubtarget *Subtarget) {
579 if (DCI.isBeforeLegalizeOps())
582 SDValue SetCC = N->getOperand(0);
584 if ((SetCC.getOpcode() != ISD::SETCC) ||
585 !SetCC.getOperand(0).getValueType().isInteger())
588 SDValue False = N->getOperand(2);
589 EVT FalseTy = False.getValueType();
591 if (!FalseTy.isInteger())
594 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
596 if (!CN || CN->getZExtValue())
600 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
601 SDValue True = N->getOperand(1);
603 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
604 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
606 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
609 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
610 TargetLowering::DAGCombinerInfo &DCI,
611 const MipsSubtarget *Subtarget) {
612 // Pattern match EXT.
613 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
614 // => ext $dst, $src, size, pos
615 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
618 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
619 unsigned ShiftRightOpc = ShiftRight.getOpcode();
621 // Op's first operand must be a shift right.
622 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
625 // The second operand of the shift must be an immediate.
627 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
630 uint64_t Pos = CN->getZExtValue();
631 uint64_t SMPos, SMSize;
633 // Op's second operand must be a shifted mask.
634 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
635 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
638 // Return if the shifted mask does not start at bit 0 or the sum of its size
639 // and Pos exceeds the word's size.
640 EVT ValTy = N->getValueType(0);
641 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
644 return DAG.getNode(MipsISD::Ext, SDLoc(N), ValTy,
645 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
646 DAG.getConstant(SMSize, MVT::i32));
649 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
650 TargetLowering::DAGCombinerInfo &DCI,
651 const MipsSubtarget *Subtarget) {
652 // Pattern match INS.
653 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
654 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
655 // => ins $dst, $src, size, pos, $src1
656 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
659 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
660 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
663 // See if Op's first operand matches (and $src1 , mask0).
664 if (And0.getOpcode() != ISD::AND)
667 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
668 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
671 // See if Op's second operand matches (and (shl $src, pos), mask1).
672 if (And1.getOpcode() != ISD::AND)
675 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
676 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
679 // The shift masks must have the same position and size.
680 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
683 SDValue Shl = And1.getOperand(0);
684 if (Shl.getOpcode() != ISD::SHL)
687 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
690 unsigned Shamt = CN->getZExtValue();
692 // Return if the shift amount and the first bit position of mask are not the
694 EVT ValTy = N->getValueType(0);
695 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
698 return DAG.getNode(MipsISD::Ins, SDLoc(N), ValTy, Shl.getOperand(0),
699 DAG.getConstant(SMPos0, MVT::i32),
700 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
703 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
704 TargetLowering::DAGCombinerInfo &DCI,
705 const MipsSubtarget *Subtarget) {
706 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
708 if (DCI.isBeforeLegalizeOps())
711 SDValue Add = N->getOperand(1);
713 if (Add.getOpcode() != ISD::ADD)
716 SDValue Lo = Add.getOperand(1);
718 if ((Lo.getOpcode() != MipsISD::Lo) ||
719 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
722 EVT ValTy = N->getValueType(0);
725 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
727 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
730 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
732 SelectionDAG &DAG = DCI.DAG;
733 unsigned Opc = N->getOpcode();
739 return performDivRemCombine(N, DAG, DCI, Subtarget);
741 return performSELECTCombine(N, DAG, DCI, Subtarget);
743 return performANDCombine(N, DAG, DCI, Subtarget);
745 return performORCombine(N, DAG, DCI, Subtarget);
747 return performADDCombine(N, DAG, DCI, Subtarget);
754 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
755 SmallVectorImpl<SDValue> &Results,
756 SelectionDAG &DAG) const {
757 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
759 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
760 Results.push_back(Res.getValue(I));
764 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
765 SmallVectorImpl<SDValue> &Results,
766 SelectionDAG &DAG) const {
767 return LowerOperationWrapper(N, Results, DAG);
770 SDValue MipsTargetLowering::
771 LowerOperation(SDValue Op, SelectionDAG &DAG) const
773 switch (Op.getOpcode())
775 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
776 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
777 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
778 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
779 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
780 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
781 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
782 case ISD::SELECT: return lowerSELECT(Op, DAG);
783 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
784 case ISD::SETCC: return lowerSETCC(Op, DAG);
785 case ISD::VASTART: return lowerVASTART(Op, DAG);
786 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
787 case ISD::FABS: return lowerFABS(Op, DAG);
788 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
789 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
790 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
791 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
792 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
793 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
794 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
795 case ISD::LOAD: return lowerLOAD(Op, DAG);
796 case ISD::STORE: return lowerSTORE(Op, DAG);
797 case ISD::ADD: return lowerADD(Op, DAG);
798 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG);
803 //===----------------------------------------------------------------------===//
804 // Lower helper functions
805 //===----------------------------------------------------------------------===//
807 // addLiveIn - This helper function adds the specified physical register to the
808 // MachineFunction as a live in value. It also creates a corresponding
809 // virtual register for it.
811 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
813 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
814 MF.getRegInfo().addLiveIn(PReg, VReg);
818 static MachineBasicBlock *expandPseudoDIV(MachineInstr *MI,
819 MachineBasicBlock &MBB,
820 const TargetInstrInfo &TII,
825 // Insert instruction "teq $divisor_reg, $zero, 7".
826 MachineBasicBlock::iterator I(MI);
827 MachineInstrBuilder MIB;
828 MIB = BuildMI(MBB, llvm::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ))
829 .addOperand(MI->getOperand(2)).addReg(Mips::ZERO).addImm(7);
831 // Use the 32-bit sub-register if this is a 64-bit division.
833 MIB->getOperand(0).setSubReg(Mips::sub_32);
839 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
840 MachineBasicBlock *BB) const {
841 switch (MI->getOpcode()) {
843 llvm_unreachable("Unexpected instr type to insert");
844 case Mips::ATOMIC_LOAD_ADD_I8:
845 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
846 case Mips::ATOMIC_LOAD_ADD_I16:
847 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
848 case Mips::ATOMIC_LOAD_ADD_I32:
849 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
850 case Mips::ATOMIC_LOAD_ADD_I64:
851 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
853 case Mips::ATOMIC_LOAD_AND_I8:
854 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
855 case Mips::ATOMIC_LOAD_AND_I16:
856 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
857 case Mips::ATOMIC_LOAD_AND_I32:
858 return emitAtomicBinary(MI, BB, 4, Mips::AND);
859 case Mips::ATOMIC_LOAD_AND_I64:
860 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
862 case Mips::ATOMIC_LOAD_OR_I8:
863 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
864 case Mips::ATOMIC_LOAD_OR_I16:
865 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
866 case Mips::ATOMIC_LOAD_OR_I32:
867 return emitAtomicBinary(MI, BB, 4, Mips::OR);
868 case Mips::ATOMIC_LOAD_OR_I64:
869 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
871 case Mips::ATOMIC_LOAD_XOR_I8:
872 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
873 case Mips::ATOMIC_LOAD_XOR_I16:
874 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
875 case Mips::ATOMIC_LOAD_XOR_I32:
876 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
877 case Mips::ATOMIC_LOAD_XOR_I64:
878 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
880 case Mips::ATOMIC_LOAD_NAND_I8:
881 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
882 case Mips::ATOMIC_LOAD_NAND_I16:
883 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
884 case Mips::ATOMIC_LOAD_NAND_I32:
885 return emitAtomicBinary(MI, BB, 4, 0, true);
886 case Mips::ATOMIC_LOAD_NAND_I64:
887 return emitAtomicBinary(MI, BB, 8, 0, true);
889 case Mips::ATOMIC_LOAD_SUB_I8:
890 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
891 case Mips::ATOMIC_LOAD_SUB_I16:
892 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
893 case Mips::ATOMIC_LOAD_SUB_I32:
894 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
895 case Mips::ATOMIC_LOAD_SUB_I64:
896 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
898 case Mips::ATOMIC_SWAP_I8:
899 return emitAtomicBinaryPartword(MI, BB, 1, 0);
900 case Mips::ATOMIC_SWAP_I16:
901 return emitAtomicBinaryPartword(MI, BB, 2, 0);
902 case Mips::ATOMIC_SWAP_I32:
903 return emitAtomicBinary(MI, BB, 4, 0);
904 case Mips::ATOMIC_SWAP_I64:
905 return emitAtomicBinary(MI, BB, 8, 0);
907 case Mips::ATOMIC_CMP_SWAP_I8:
908 return emitAtomicCmpSwapPartword(MI, BB, 1);
909 case Mips::ATOMIC_CMP_SWAP_I16:
910 return emitAtomicCmpSwapPartword(MI, BB, 2);
911 case Mips::ATOMIC_CMP_SWAP_I32:
912 return emitAtomicCmpSwap(MI, BB, 4);
913 case Mips::ATOMIC_CMP_SWAP_I64:
914 return emitAtomicCmpSwap(MI, BB, 8);
915 case Mips::PseudoSDIV:
916 case Mips::PseudoUDIV:
917 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), false);
918 case Mips::PseudoDSDIV:
919 case Mips::PseudoDUDIV:
920 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), true);
924 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
925 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
927 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
928 unsigned Size, unsigned BinOpcode,
930 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
932 MachineFunction *MF = BB->getParent();
933 MachineRegisterInfo &RegInfo = MF->getRegInfo();
934 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
935 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
936 DebugLoc DL = MI->getDebugLoc();
937 unsigned LL, SC, AND, NOR, ZERO, BEQ;
952 ZERO = Mips::ZERO_64;
956 unsigned OldVal = MI->getOperand(0).getReg();
957 unsigned Ptr = MI->getOperand(1).getReg();
958 unsigned Incr = MI->getOperand(2).getReg();
960 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
961 unsigned AndRes = RegInfo.createVirtualRegister(RC);
962 unsigned Success = RegInfo.createVirtualRegister(RC);
964 // insert new blocks after the current block
965 const BasicBlock *LLVM_BB = BB->getBasicBlock();
966 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
967 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
968 MachineFunction::iterator It = BB;
970 MF->insert(It, loopMBB);
971 MF->insert(It, exitMBB);
973 // Transfer the remainder of BB and its successor edges to exitMBB.
974 exitMBB->splice(exitMBB->begin(), BB,
975 llvm::next(MachineBasicBlock::iterator(MI)),
977 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
981 // fallthrough --> loopMBB
982 BB->addSuccessor(loopMBB);
983 loopMBB->addSuccessor(loopMBB);
984 loopMBB->addSuccessor(exitMBB);
988 // <binop> storeval, oldval, incr
989 // sc success, storeval, 0(ptr)
990 // beq success, $0, loopMBB
992 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
994 // and andres, oldval, incr
995 // nor storeval, $0, andres
996 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
997 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
998 } else if (BinOpcode) {
999 // <binop> storeval, oldval, incr
1000 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1004 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1005 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1007 MI->eraseFromParent(); // The instruction is gone now.
1013 MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI,
1014 MachineBasicBlock *BB,
1015 unsigned Size, unsigned BinOpcode,
1017 assert((Size == 1 || Size == 2) &&
1018 "Unsupported size for EmitAtomicBinaryPartial.");
1020 MachineFunction *MF = BB->getParent();
1021 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1022 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1023 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1024 DebugLoc DL = MI->getDebugLoc();
1026 unsigned Dest = MI->getOperand(0).getReg();
1027 unsigned Ptr = MI->getOperand(1).getReg();
1028 unsigned Incr = MI->getOperand(2).getReg();
1030 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1031 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1032 unsigned Mask = RegInfo.createVirtualRegister(RC);
1033 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1034 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1035 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1036 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1037 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1038 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1039 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1040 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1041 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1042 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1043 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1044 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1045 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1046 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1047 unsigned Success = RegInfo.createVirtualRegister(RC);
1049 // insert new blocks after the current block
1050 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1051 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1052 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1053 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1054 MachineFunction::iterator It = BB;
1056 MF->insert(It, loopMBB);
1057 MF->insert(It, sinkMBB);
1058 MF->insert(It, exitMBB);
1060 // Transfer the remainder of BB and its successor edges to exitMBB.
1061 exitMBB->splice(exitMBB->begin(), BB,
1062 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1063 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1065 BB->addSuccessor(loopMBB);
1066 loopMBB->addSuccessor(loopMBB);
1067 loopMBB->addSuccessor(sinkMBB);
1068 sinkMBB->addSuccessor(exitMBB);
1071 // addiu masklsb2,$0,-4 # 0xfffffffc
1072 // and alignedaddr,ptr,masklsb2
1073 // andi ptrlsb2,ptr,3
1074 // sll shiftamt,ptrlsb2,3
1075 // ori maskupper,$0,255 # 0xff
1076 // sll mask,maskupper,shiftamt
1077 // nor mask2,$0,mask
1078 // sll incr2,incr,shiftamt
1080 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1081 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1082 .addReg(Mips::ZERO).addImm(-4);
1083 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1084 .addReg(Ptr).addReg(MaskLSB2);
1085 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1086 if (Subtarget->isLittle()) {
1087 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1089 unsigned Off = RegInfo.createVirtualRegister(RC);
1090 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1091 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1092 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1094 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1095 .addReg(Mips::ZERO).addImm(MaskImm);
1096 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1097 .addReg(MaskUpper).addReg(ShiftAmt);
1098 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1099 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1101 // atomic.load.binop
1103 // ll oldval,0(alignedaddr)
1104 // binop binopres,oldval,incr2
1105 // and newval,binopres,mask
1106 // and maskedoldval0,oldval,mask2
1107 // or storeval,maskedoldval0,newval
1108 // sc success,storeval,0(alignedaddr)
1109 // beq success,$0,loopMBB
1113 // ll oldval,0(alignedaddr)
1114 // and newval,incr2,mask
1115 // and maskedoldval0,oldval,mask2
1116 // or storeval,maskedoldval0,newval
1117 // sc success,storeval,0(alignedaddr)
1118 // beq success,$0,loopMBB
1121 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1123 // and andres, oldval, incr2
1124 // nor binopres, $0, andres
1125 // and newval, binopres, mask
1126 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1127 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1128 .addReg(Mips::ZERO).addReg(AndRes);
1129 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1130 } else if (BinOpcode) {
1131 // <binop> binopres, oldval, incr2
1132 // and newval, binopres, mask
1133 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1134 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1135 } else {// atomic.swap
1136 // and newval, incr2, mask
1137 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1140 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1141 .addReg(OldVal).addReg(Mask2);
1142 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1143 .addReg(MaskedOldVal0).addReg(NewVal);
1144 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1145 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1146 BuildMI(BB, DL, TII->get(Mips::BEQ))
1147 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1150 // and maskedoldval1,oldval,mask
1151 // srl srlres,maskedoldval1,shiftamt
1152 // sll sllres,srlres,24
1153 // sra dest,sllres,24
1155 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1157 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1158 .addReg(OldVal).addReg(Mask);
1159 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1160 .addReg(MaskedOldVal1).addReg(ShiftAmt);
1161 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1162 .addReg(SrlRes).addImm(ShiftImm);
1163 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1164 .addReg(SllRes).addImm(ShiftImm);
1166 MI->eraseFromParent(); // The instruction is gone now.
1172 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1173 MachineBasicBlock *BB,
1174 unsigned Size) const {
1175 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1177 MachineFunction *MF = BB->getParent();
1178 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1179 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1180 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1181 DebugLoc DL = MI->getDebugLoc();
1182 unsigned LL, SC, ZERO, BNE, BEQ;
1194 ZERO = Mips::ZERO_64;
1199 unsigned Dest = MI->getOperand(0).getReg();
1200 unsigned Ptr = MI->getOperand(1).getReg();
1201 unsigned OldVal = MI->getOperand(2).getReg();
1202 unsigned NewVal = MI->getOperand(3).getReg();
1204 unsigned Success = RegInfo.createVirtualRegister(RC);
1206 // insert new blocks after the current block
1207 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1208 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1209 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1210 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1211 MachineFunction::iterator It = BB;
1213 MF->insert(It, loop1MBB);
1214 MF->insert(It, loop2MBB);
1215 MF->insert(It, exitMBB);
1217 // Transfer the remainder of BB and its successor edges to exitMBB.
1218 exitMBB->splice(exitMBB->begin(), BB,
1219 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1220 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1224 // fallthrough --> loop1MBB
1225 BB->addSuccessor(loop1MBB);
1226 loop1MBB->addSuccessor(exitMBB);
1227 loop1MBB->addSuccessor(loop2MBB);
1228 loop2MBB->addSuccessor(loop1MBB);
1229 loop2MBB->addSuccessor(exitMBB);
1233 // bne dest, oldval, exitMBB
1235 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1236 BuildMI(BB, DL, TII->get(BNE))
1237 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1240 // sc success, newval, 0(ptr)
1241 // beq success, $0, loop1MBB
1243 BuildMI(BB, DL, TII->get(SC), Success)
1244 .addReg(NewVal).addReg(Ptr).addImm(0);
1245 BuildMI(BB, DL, TII->get(BEQ))
1246 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1248 MI->eraseFromParent(); // The instruction is gone now.
1254 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1255 MachineBasicBlock *BB,
1256 unsigned Size) const {
1257 assert((Size == 1 || Size == 2) &&
1258 "Unsupported size for EmitAtomicCmpSwapPartial.");
1260 MachineFunction *MF = BB->getParent();
1261 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1262 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1263 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1264 DebugLoc DL = MI->getDebugLoc();
1266 unsigned Dest = MI->getOperand(0).getReg();
1267 unsigned Ptr = MI->getOperand(1).getReg();
1268 unsigned CmpVal = MI->getOperand(2).getReg();
1269 unsigned NewVal = MI->getOperand(3).getReg();
1271 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1272 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1273 unsigned Mask = RegInfo.createVirtualRegister(RC);
1274 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1275 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1276 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1277 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1278 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1279 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1280 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1281 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1282 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1283 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1284 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1285 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1286 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1287 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1288 unsigned Success = RegInfo.createVirtualRegister(RC);
1290 // insert new blocks after the current block
1291 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1292 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1293 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1294 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1295 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1296 MachineFunction::iterator It = BB;
1298 MF->insert(It, loop1MBB);
1299 MF->insert(It, loop2MBB);
1300 MF->insert(It, sinkMBB);
1301 MF->insert(It, exitMBB);
1303 // Transfer the remainder of BB and its successor edges to exitMBB.
1304 exitMBB->splice(exitMBB->begin(), BB,
1305 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1306 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1308 BB->addSuccessor(loop1MBB);
1309 loop1MBB->addSuccessor(sinkMBB);
1310 loop1MBB->addSuccessor(loop2MBB);
1311 loop2MBB->addSuccessor(loop1MBB);
1312 loop2MBB->addSuccessor(sinkMBB);
1313 sinkMBB->addSuccessor(exitMBB);
1315 // FIXME: computation of newval2 can be moved to loop2MBB.
1317 // addiu masklsb2,$0,-4 # 0xfffffffc
1318 // and alignedaddr,ptr,masklsb2
1319 // andi ptrlsb2,ptr,3
1320 // sll shiftamt,ptrlsb2,3
1321 // ori maskupper,$0,255 # 0xff
1322 // sll mask,maskupper,shiftamt
1323 // nor mask2,$0,mask
1324 // andi maskedcmpval,cmpval,255
1325 // sll shiftedcmpval,maskedcmpval,shiftamt
1326 // andi maskednewval,newval,255
1327 // sll shiftednewval,maskednewval,shiftamt
1328 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1329 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1330 .addReg(Mips::ZERO).addImm(-4);
1331 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1332 .addReg(Ptr).addReg(MaskLSB2);
1333 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1334 if (Subtarget->isLittle()) {
1335 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1337 unsigned Off = RegInfo.createVirtualRegister(RC);
1338 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1339 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1340 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1342 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1343 .addReg(Mips::ZERO).addImm(MaskImm);
1344 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1345 .addReg(MaskUpper).addReg(ShiftAmt);
1346 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1347 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1348 .addReg(CmpVal).addImm(MaskImm);
1349 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1350 .addReg(MaskedCmpVal).addReg(ShiftAmt);
1351 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1352 .addReg(NewVal).addImm(MaskImm);
1353 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1354 .addReg(MaskedNewVal).addReg(ShiftAmt);
1357 // ll oldval,0(alginedaddr)
1358 // and maskedoldval0,oldval,mask
1359 // bne maskedoldval0,shiftedcmpval,sinkMBB
1361 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1362 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1363 .addReg(OldVal).addReg(Mask);
1364 BuildMI(BB, DL, TII->get(Mips::BNE))
1365 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1368 // and maskedoldval1,oldval,mask2
1369 // or storeval,maskedoldval1,shiftednewval
1370 // sc success,storeval,0(alignedaddr)
1371 // beq success,$0,loop1MBB
1373 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1374 .addReg(OldVal).addReg(Mask2);
1375 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1376 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1377 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1378 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1379 BuildMI(BB, DL, TII->get(Mips::BEQ))
1380 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1383 // srl srlres,maskedoldval0,shiftamt
1384 // sll sllres,srlres,24
1385 // sra dest,sllres,24
1387 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1389 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1390 .addReg(MaskedOldVal0).addReg(ShiftAmt);
1391 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1392 .addReg(SrlRes).addImm(ShiftImm);
1393 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1394 .addReg(SllRes).addImm(ShiftImm);
1396 MI->eraseFromParent(); // The instruction is gone now.
1401 //===----------------------------------------------------------------------===//
1402 // Misc Lower Operation implementation
1403 //===----------------------------------------------------------------------===//
1404 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1405 SDValue Chain = Op.getOperand(0);
1406 SDValue Table = Op.getOperand(1);
1407 SDValue Index = Op.getOperand(2);
1409 EVT PTy = getPointerTy();
1410 unsigned EntrySize =
1411 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout());
1413 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1414 DAG.getConstant(EntrySize, PTy));
1415 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1417 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1418 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1419 MachinePointerInfo::getJumpTable(), MemVT, false, false,
1421 Chain = Addr.getValue(1);
1423 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) {
1424 // For PIC, the sequence is:
1425 // BRIND(load(Jumptable + index) + RelocBase)
1426 // RelocBase can be JumpTable, GOT or some sort of global base.
1427 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1428 getPICJumpTableRelocBase(Table, DAG));
1431 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1434 SDValue MipsTargetLowering::
1435 lowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1437 // The first operand is the chain, the second is the condition, the third is
1438 // the block to branch to if the condition is true.
1439 SDValue Chain = Op.getOperand(0);
1440 SDValue Dest = Op.getOperand(2);
1443 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1445 // Return if flag is not set by a floating point comparison.
1446 if (CondRes.getOpcode() != MipsISD::FPCmp)
1449 SDValue CCNode = CondRes.getOperand(2);
1451 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1452 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1453 SDValue BrCode = DAG.getConstant(Opc, MVT::i32);
1454 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
1455 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1456 FCC0, Dest, CondRes);
1459 SDValue MipsTargetLowering::
1460 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1462 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1464 // Return if flag is not set by a floating point comparison.
1465 if (Cond.getOpcode() != MipsISD::FPCmp)
1468 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1472 SDValue MipsTargetLowering::
1473 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1476 EVT Ty = Op.getOperand(0).getValueType();
1477 SDValue Cond = DAG.getNode(ISD::SETCC, DL,
1478 getSetCCResultType(*DAG.getContext(), Ty),
1479 Op.getOperand(0), Op.getOperand(1),
1482 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1486 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1487 SDValue Cond = createFPCmp(DAG, Op);
1489 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1490 "Floating point operand expected.");
1492 SDValue True = DAG.getConstant(1, MVT::i32);
1493 SDValue False = DAG.getConstant(0, MVT::i32);
1495 return createCMovFP(DAG, Cond, True, False, SDLoc(Op));
1498 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1499 SelectionDAG &DAG) const {
1500 // FIXME there isn't actually debug info here
1502 EVT Ty = Op.getValueType();
1503 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
1504 const GlobalValue *GV = N->getGlobal();
1506 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1507 const MipsTargetObjectFile &TLOF =
1508 (const MipsTargetObjectFile&)getObjFileLowering();
1510 // %gp_rel relocation
1511 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1512 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0,
1514 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL,
1515 DAG.getVTList(MVT::i32), &GA, 1);
1516 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1517 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode);
1520 // %hi/%lo relocation
1521 return getAddrNonPIC(N, Ty, DAG);
1524 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1525 return getAddrLocal(N, Ty, DAG, HasMips64);
1528 return getAddrGlobalLargeGOT(N, Ty, DAG, MipsII::MO_GOT_HI16,
1529 MipsII::MO_GOT_LO16);
1531 return getAddrGlobal(N, Ty, DAG,
1532 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
1535 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1536 SelectionDAG &DAG) const {
1537 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
1538 EVT Ty = Op.getValueType();
1540 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1541 return getAddrNonPIC(N, Ty, DAG);
1543 return getAddrLocal(N, Ty, DAG, HasMips64);
1546 SDValue MipsTargetLowering::
1547 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1549 // If the relocation model is PIC, use the General Dynamic TLS Model or
1550 // Local Dynamic TLS model, otherwise use the Initial Exec or
1551 // Local Exec TLS Model.
1553 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1555 const GlobalValue *GV = GA->getGlobal();
1556 EVT PtrVT = getPointerTy();
1558 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1560 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1561 // General Dynamic and Local Dynamic TLS Model.
1562 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1565 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1566 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1567 getGlobalReg(DAG, PtrVT), TGA);
1568 unsigned PtrSize = PtrVT.getSizeInBits();
1569 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1571 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1575 Entry.Node = Argument;
1577 Args.push_back(Entry);
1579 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
1580 false, false, false, false, 0, CallingConv::C,
1581 /*IsTailCall=*/false, /*doesNotRet=*/false,
1582 /*isReturnValueUsed=*/true,
1583 TlsGetAddr, Args, DAG, DL);
1584 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1586 SDValue Ret = CallResult.first;
1588 if (model != TLSModel::LocalDynamic)
1591 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1592 MipsII::MO_DTPREL_HI);
1593 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1594 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1595 MipsII::MO_DTPREL_LO);
1596 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1597 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1598 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1602 if (model == TLSModel::InitialExec) {
1603 // Initial Exec TLS Model
1604 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1605 MipsII::MO_GOTTPREL);
1606 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1608 Offset = DAG.getLoad(PtrVT, DL,
1609 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1610 false, false, false, 0);
1612 // Local Exec TLS Model
1613 assert(model == TLSModel::LocalExec);
1614 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1615 MipsII::MO_TPREL_HI);
1616 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1617 MipsII::MO_TPREL_LO);
1618 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1619 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1620 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1623 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1624 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1627 SDValue MipsTargetLowering::
1628 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1630 JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
1631 EVT Ty = Op.getValueType();
1633 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1634 return getAddrNonPIC(N, Ty, DAG);
1636 return getAddrLocal(N, Ty, DAG, HasMips64);
1639 SDValue MipsTargetLowering::
1640 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1642 // gp_rel relocation
1643 // FIXME: we should reference the constant pool using small data sections,
1644 // but the asm printer currently doesn't support this feature without
1645 // hacking it. This feature should come soon so we can uncomment the
1647 //if (IsInSmallSection(C->getType())) {
1648 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1649 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1650 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1651 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1652 EVT Ty = Op.getValueType();
1654 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1655 return getAddrNonPIC(N, Ty, DAG);
1657 return getAddrLocal(N, Ty, DAG, HasMips64);
1660 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1661 MachineFunction &MF = DAG.getMachineFunction();
1662 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1665 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1668 // vastart just stores the address of the VarArgsFrameIndex slot into the
1669 // memory location argument.
1670 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1671 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1672 MachinePointerInfo(SV), false, false, 0);
1675 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1676 EVT TyX = Op.getOperand(0).getValueType();
1677 EVT TyY = Op.getOperand(1).getValueType();
1678 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1679 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1683 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1685 SDValue X = (TyX == MVT::f32) ?
1686 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1687 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1689 SDValue Y = (TyY == MVT::f32) ?
1690 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1691 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1695 // ext E, Y, 31, 1 ; extract bit31 of Y
1696 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1697 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1698 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1701 // srl SrlX, SllX, 1
1703 // sll SllY, SrlX, 31
1704 // or Or, SrlX, SllY
1705 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1706 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1707 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1708 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1709 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1712 if (TyX == MVT::f32)
1713 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1715 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1716 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1717 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1720 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1721 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1722 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1723 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1724 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1727 // Bitcast to integer nodes.
1728 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1729 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1732 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1733 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1734 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1735 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
1737 if (WidthX > WidthY)
1738 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1739 else if (WidthY > WidthX)
1740 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1742 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1743 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
1744 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1747 // (d)sll SllX, X, 1
1748 // (d)srl SrlX, SllX, 1
1749 // (d)srl SrlY, Y, width(Y)-1
1750 // (d)sll SllY, SrlX, width(Y)-1
1751 // or Or, SrlX, SllY
1752 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1753 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1754 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1755 DAG.getConstant(WidthY - 1, MVT::i32));
1757 if (WidthX > WidthY)
1758 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1759 else if (WidthY > WidthX)
1760 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1762 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1763 DAG.getConstant(WidthX - 1, MVT::i32));
1764 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
1765 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
1769 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1770 if (Subtarget->hasMips64())
1771 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
1773 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
1776 static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1777 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1780 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1782 SDValue X = (Op.getValueType() == MVT::f32) ?
1783 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1784 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1789 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
1790 DAG.getRegister(Mips::ZERO, MVT::i32),
1791 DAG.getConstant(31, MVT::i32), Const1, X);
1793 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1794 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1797 if (Op.getValueType() == MVT::f32)
1798 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
1800 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1801 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1802 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1805 static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1806 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1809 // Bitcast to integer node.
1810 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
1814 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
1815 DAG.getRegister(Mips::ZERO_64, MVT::i64),
1816 DAG.getConstant(63, MVT::i32), Const1, X);
1818 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
1819 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
1822 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
1826 MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const {
1827 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
1828 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2());
1830 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2());
1833 SDValue MipsTargetLowering::
1834 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1836 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1837 "Frame address can only be determined for current frame.");
1839 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1840 MFI->setFrameAddressIsTaken(true);
1841 EVT VT = Op.getValueType();
1843 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1844 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1848 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
1849 SelectionDAG &DAG) const {
1851 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1852 "Return address can be determined only for current frame.");
1854 MachineFunction &MF = DAG.getMachineFunction();
1855 MachineFrameInfo *MFI = MF.getFrameInfo();
1856 MVT VT = Op.getSimpleValueType();
1857 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
1858 MFI->setReturnAddressIsTaken(true);
1860 // Return RA, which contains the return address. Mark it an implicit live-in.
1861 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
1862 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
1865 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
1866 // generated from __builtin_eh_return (offset, handler)
1867 // The effect of this is to adjust the stack pointer by "offset"
1868 // and then branch to "handler".
1869 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
1871 MachineFunction &MF = DAG.getMachineFunction();
1872 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1874 MipsFI->setCallsEhReturn();
1875 SDValue Chain = Op.getOperand(0);
1876 SDValue Offset = Op.getOperand(1);
1877 SDValue Handler = Op.getOperand(2);
1879 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
1881 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
1882 // EH_RETURN nodes, so that instructions are emitted back-to-back.
1883 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1;
1884 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0;
1885 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
1886 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
1887 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
1888 DAG.getRegister(OffsetReg, Ty),
1889 DAG.getRegister(AddrReg, getPointerTy()),
1893 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
1894 SelectionDAG &DAG) const {
1895 // FIXME: Need pseudo-fence for 'singlethread' fences
1896 // FIXME: Set SType for weaker fences where supported/appropriate.
1899 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
1900 DAG.getConstant(SType, MVT::i32));
1903 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
1904 SelectionDAG &DAG) const {
1906 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1907 SDValue Shamt = Op.getOperand(2);
1910 // lo = (shl lo, shamt)
1911 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
1914 // hi = (shl lo, shamt[4:0])
1915 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1916 DAG.getConstant(-1, MVT::i32));
1917 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
1918 DAG.getConstant(1, MVT::i32));
1919 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
1921 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
1922 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1923 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
1924 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1925 DAG.getConstant(0x20, MVT::i32));
1926 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1927 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
1928 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
1930 SDValue Ops[2] = {Lo, Hi};
1931 return DAG.getMergeValues(Ops, 2, DL);
1934 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
1937 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1938 SDValue Shamt = Op.getOperand(2);
1941 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
1943 // hi = (sra hi, shamt)
1945 // hi = (srl hi, shamt)
1948 // lo = (sra hi, shamt[4:0])
1949 // hi = (sra hi, 31)
1951 // lo = (srl hi, shamt[4:0])
1953 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1954 DAG.getConstant(-1, MVT::i32));
1955 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
1956 DAG.getConstant(1, MVT::i32));
1957 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
1958 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
1959 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1960 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1962 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1963 DAG.getConstant(0x20, MVT::i32));
1964 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
1965 DAG.getConstant(31, MVT::i32));
1966 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
1967 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1968 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
1971 SDValue Ops[2] = {Lo, Hi};
1972 return DAG.getMergeValues(Ops, 2, DL);
1975 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
1976 SDValue Chain, SDValue Src, unsigned Offset) {
1977 SDValue Ptr = LD->getBasePtr();
1978 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
1979 EVT BasePtrVT = Ptr.getValueType();
1981 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1984 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
1985 DAG.getConstant(Offset, BasePtrVT));
1987 SDValue Ops[] = { Chain, Ptr, Src };
1988 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
1989 LD->getMemOperand());
1992 // Expand an unaligned 32 or 64-bit integer load node.
1993 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1994 LoadSDNode *LD = cast<LoadSDNode>(Op);
1995 EVT MemVT = LD->getMemoryVT();
1997 // Return if load is aligned or if MemVT is neither i32 nor i64.
1998 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
1999 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2002 bool IsLittle = Subtarget->isLittle();
2003 EVT VT = Op.getValueType();
2004 ISD::LoadExtType ExtType = LD->getExtensionType();
2005 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2007 assert((VT == MVT::i32) || (VT == MVT::i64));
2010 // (set dst, (i64 (load baseptr)))
2012 // (set tmp, (ldl (add baseptr, 7), undef))
2013 // (set dst, (ldr baseptr, tmp))
2014 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2015 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2017 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2021 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2023 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2027 // (set dst, (i32 (load baseptr))) or
2028 // (set dst, (i64 (sextload baseptr))) or
2029 // (set dst, (i64 (extload baseptr)))
2031 // (set tmp, (lwl (add baseptr, 3), undef))
2032 // (set dst, (lwr baseptr, tmp))
2033 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2034 (ExtType == ISD::EXTLOAD))
2037 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2040 // (set dst, (i64 (zextload baseptr)))
2042 // (set tmp0, (lwl (add baseptr, 3), undef))
2043 // (set tmp1, (lwr baseptr, tmp0))
2044 // (set tmp2, (shl tmp1, 32))
2045 // (set dst, (srl tmp2, 32))
2047 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2048 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2049 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2050 SDValue Ops[] = { SRL, LWR.getValue(1) };
2051 return DAG.getMergeValues(Ops, 2, DL);
2054 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2055 SDValue Chain, unsigned Offset) {
2056 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2057 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2059 SDVTList VTList = DAG.getVTList(MVT::Other);
2062 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2063 DAG.getConstant(Offset, BasePtrVT));
2065 SDValue Ops[] = { Chain, Value, Ptr };
2066 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2067 SD->getMemOperand());
2070 // Expand an unaligned 32 or 64-bit integer store node.
2071 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2073 SDValue Value = SD->getValue(), Chain = SD->getChain();
2074 EVT VT = Value.getValueType();
2077 // (store val, baseptr) or
2078 // (truncstore val, baseptr)
2080 // (swl val, (add baseptr, 3))
2081 // (swr val, baseptr)
2082 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2083 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2085 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2088 assert(VT == MVT::i64);
2091 // (store val, baseptr)
2093 // (sdl val, (add baseptr, 7))
2094 // (sdr val, baseptr)
2095 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2096 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2099 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
2100 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) {
2101 SDValue Val = SD->getValue();
2103 if (Val.getOpcode() != ISD::FP_TO_SINT)
2106 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2107 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2110 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2111 SD->getPointerInfo(), SD->isVolatile(),
2112 SD->isNonTemporal(), SD->getAlignment());
2115 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2116 StoreSDNode *SD = cast<StoreSDNode>(Op);
2117 EVT MemVT = SD->getMemoryVT();
2119 // Lower unaligned integer stores.
2120 if ((SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2121 ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2122 return lowerUnalignedIntStore(SD, DAG, Subtarget->isLittle());
2124 return lowerFP_TO_SINT_STORE(SD, DAG);
2127 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2128 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2129 || cast<ConstantSDNode>
2130 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2131 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2135 // (add (frameaddr 0), (frame_to_args_offset))
2136 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2137 // (add FrameObject, 0)
2138 // where FrameObject is a fixed StackObject with offset 0 which points to
2139 // the old stack pointer.
2140 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2141 EVT ValTy = Op->getValueType(0);
2142 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2143 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2144 return DAG.getNode(ISD::ADD, SDLoc(Op), ValTy, InArgsAddr,
2145 DAG.getConstant(0, ValTy));
2148 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2149 SelectionDAG &DAG) const {
2150 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2151 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2153 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2156 //===----------------------------------------------------------------------===//
2157 // Calling Convention Implementation
2158 //===----------------------------------------------------------------------===//
2160 //===----------------------------------------------------------------------===//
2161 // TODO: Implement a generic logic using tblgen that can support this.
2162 // Mips O32 ABI rules:
2164 // i32 - Passed in A0, A1, A2, A3 and stack
2165 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2166 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2167 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2168 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2169 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2172 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2173 //===----------------------------------------------------------------------===//
2175 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2176 MVT LocVT, CCValAssign::LocInfo LocInfo,
2177 ISD::ArgFlagsTy ArgFlags, CCState &State,
2178 const uint16_t *F64Regs) {
2180 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2182 static const uint16_t IntRegs[] = {
2183 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2185 static const uint16_t F32Regs[] = {
2186 Mips::F12, Mips::F14
2189 // Do not process byval args here.
2190 if (ArgFlags.isByVal())
2193 // Promote i8 and i16
2194 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2196 if (ArgFlags.isSExt())
2197 LocInfo = CCValAssign::SExt;
2198 else if (ArgFlags.isZExt())
2199 LocInfo = CCValAssign::ZExt;
2201 LocInfo = CCValAssign::AExt;
2206 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2207 // is true: function is vararg, argument is 3rd or higher, there is previous
2208 // argument which is not f32 or f64.
2209 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2210 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2211 unsigned OrigAlign = ArgFlags.getOrigAlign();
2212 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2214 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2215 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2216 // If this is the first part of an i64 arg,
2217 // the allocated register must be either A0 or A2.
2218 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2219 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2221 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2222 // Allocate int register and shadow next int register. If first
2223 // available register is Mips::A1 or Mips::A3, shadow it too.
2224 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2225 if (Reg == Mips::A1 || Reg == Mips::A3)
2226 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2227 State.AllocateReg(IntRegs, IntRegsSize);
2229 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2230 // we are guaranteed to find an available float register
2231 if (ValVT == MVT::f32) {
2232 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2233 // Shadow int register
2234 State.AllocateReg(IntRegs, IntRegsSize);
2236 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2237 // Shadow int registers
2238 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2239 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2240 State.AllocateReg(IntRegs, IntRegsSize);
2241 State.AllocateReg(IntRegs, IntRegsSize);
2244 llvm_unreachable("Cannot handle this ValVT.");
2247 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2249 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2251 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2256 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2257 MVT LocVT, CCValAssign::LocInfo LocInfo,
2258 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2259 static const uint16_t F64Regs[] = { Mips::D6, Mips::D7 };
2261 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2264 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2265 MVT LocVT, CCValAssign::LocInfo LocInfo,
2266 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2267 static const uint16_t F64Regs[] = { Mips::D12_64, Mips::D12_64 };
2269 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2272 #include "MipsGenCallingConv.inc"
2274 //===----------------------------------------------------------------------===//
2275 // Call Calling Convention Implementation
2276 //===----------------------------------------------------------------------===//
2278 // Return next O32 integer argument register.
2279 static unsigned getNextIntArgReg(unsigned Reg) {
2280 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2281 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2285 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2286 SDValue Chain, SDValue Arg, SDLoc DL,
2287 bool IsTailCall, SelectionDAG &DAG) const {
2289 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2290 DAG.getIntPtrConstant(Offset));
2291 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2295 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2296 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2297 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2298 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2299 /*isVolatile=*/ true, false, 0);
2302 void MipsTargetLowering::
2303 getOpndList(SmallVectorImpl<SDValue> &Ops,
2304 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2305 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2306 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
2307 // Insert node "GP copy globalreg" before call to function.
2309 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2310 // in PIC mode) allow symbols to be resolved via lazy binding.
2311 // The lazy binding stub requires GP to point to the GOT.
2312 if (IsPICCall && !InternalLinkage) {
2313 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
2314 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2315 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2318 // Build a sequence of copy-to-reg nodes chained together with token
2319 // chain and flag operands which copy the outgoing args into registers.
2320 // The InFlag in necessary since all emitted instructions must be
2324 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2325 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2326 RegsToPass[i].second, InFlag);
2327 InFlag = Chain.getValue(1);
2330 // Add argument registers to the end of the list so that they are
2331 // known live into the call.
2332 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2333 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2334 RegsToPass[i].second.getValueType()));
2336 // Add a register mask operand representing the call-preserved registers.
2337 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2338 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv);
2339 assert(Mask && "Missing call preserved mask for calling convention");
2340 if (Subtarget->inMips16HardFloat()) {
2341 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2342 llvm::StringRef Sym = G->getGlobal()->getName();
2343 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2344 if (F->hasFnAttribute("__Mips16RetHelper")) {
2345 Mask = MipsRegisterInfo::getMips16RetHelperMask();
2349 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2351 if (InFlag.getNode())
2352 Ops.push_back(InFlag);
2355 /// LowerCall - functions arguments are copied from virtual regs to
2356 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2358 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2359 SmallVectorImpl<SDValue> &InVals) const {
2360 SelectionDAG &DAG = CLI.DAG;
2362 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2363 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2364 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2365 SDValue Chain = CLI.Chain;
2366 SDValue Callee = CLI.Callee;
2367 bool &IsTailCall = CLI.IsTailCall;
2368 CallingConv::ID CallConv = CLI.CallConv;
2369 bool IsVarArg = CLI.IsVarArg;
2371 MachineFunction &MF = DAG.getMachineFunction();
2372 MachineFrameInfo *MFI = MF.getFrameInfo();
2373 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2374 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2376 // Analyze operands of the call, assigning locations to each operand.
2377 SmallVector<CCValAssign, 16> ArgLocs;
2378 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2379 getTargetMachine(), ArgLocs, *DAG.getContext());
2380 MipsCC::SpecialCallingConvType SpecialCallingConv =
2381 getSpecialCallingConv(Callee);
2382 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo,
2383 SpecialCallingConv);
2385 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg,
2386 Subtarget->mipsSEUsesSoftFloat(),
2387 Callee.getNode(), CLI.Args);
2389 // Get a count of how many bytes are to be pushed on the stack.
2390 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2392 // Check if it's really possible to do a tail call.
2395 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2396 *MF.getInfo<MipsFunctionInfo>());
2401 // Chain is the output chain of the last Load/Store or CopyToReg node.
2402 // ByValChain is the output chain of the last Memcpy node created for copying
2403 // byval arguments to the stack.
2404 unsigned StackAlignment = TFL->getStackAlignment();
2405 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2406 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2409 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL);
2411 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL,
2412 IsN64 ? Mips::SP_64 : Mips::SP,
2415 // With EABI is it possible to have 16 args on registers.
2416 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2417 SmallVector<SDValue, 8> MemOpChains;
2418 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2420 // Walk the register/memloc assignments, inserting copies/loads.
2421 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2422 SDValue Arg = OutVals[i];
2423 CCValAssign &VA = ArgLocs[i];
2424 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2425 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2428 if (Flags.isByVal()) {
2429 assert(Flags.getByValSize() &&
2430 "ByVal args of size 0 should have been ignored by front-end.");
2431 assert(ByValArg != MipsCCInfo.byval_end());
2432 assert(!IsTailCall &&
2433 "Do not tail-call optimize if there is a byval argument.");
2434 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2435 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2440 // Promote the value if needed.
2441 switch (VA.getLocInfo()) {
2442 default: llvm_unreachable("Unknown loc info!");
2443 case CCValAssign::Full:
2444 if (VA.isRegLoc()) {
2445 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2446 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2447 (ValVT == MVT::i64 && LocVT == MVT::f64))
2448 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2449 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2450 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2451 Arg, DAG.getConstant(0, MVT::i32));
2452 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2453 Arg, DAG.getConstant(1, MVT::i32));
2454 if (!Subtarget->isLittle())
2456 unsigned LocRegLo = VA.getLocReg();
2457 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2458 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2459 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2464 case CCValAssign::SExt:
2465 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2467 case CCValAssign::ZExt:
2468 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2470 case CCValAssign::AExt:
2471 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2475 // Arguments that can be passed on register must be kept at
2476 // RegsToPass vector
2477 if (VA.isRegLoc()) {
2478 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2482 // Register can't get to this point...
2483 assert(VA.isMemLoc());
2485 // emit ISD::STORE whichs stores the
2486 // parameter value to a stack Location
2487 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2488 Chain, Arg, DL, IsTailCall, DAG));
2491 // Transform all store nodes into one single node because all store
2492 // nodes are independent of each other.
2493 if (!MemOpChains.empty())
2494 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2495 &MemOpChains[0], MemOpChains.size());
2497 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2498 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2499 // node so that legalize doesn't hack it.
2500 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2501 bool GlobalOrExternal = false, InternalLinkage = false;
2503 EVT Ty = Callee.getValueType();
2505 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2507 InternalLinkage = G->getGlobal()->hasInternalLinkage();
2509 if (InternalLinkage)
2510 Callee = getAddrLocal(G, Ty, DAG, HasMips64);
2512 Callee = getAddrGlobalLargeGOT(G, Ty, DAG, MipsII::MO_CALL_HI16,
2513 MipsII::MO_CALL_LO16);
2515 Callee = getAddrGlobal(G, Ty, DAG, MipsII::MO_GOT_CALL);
2517 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0,
2518 MipsII::MO_NO_FLAG);
2519 GlobalOrExternal = true;
2521 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2522 if (!IsN64 && !IsPIC) // !N64 && static
2523 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2524 MipsII::MO_NO_FLAG);
2526 Callee = getAddrGlobalLargeGOT(S, Ty, DAG, MipsII::MO_CALL_HI16,
2527 MipsII::MO_CALL_LO16);
2529 Callee = getAddrGlobal(S, Ty, DAG, MipsII::MO_GOT_CALL);
2531 GlobalOrExternal = true;
2534 SmallVector<SDValue, 8> Ops(1, Chain);
2535 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2537 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2538 CLI, Callee, Chain);
2541 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size());
2543 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size());
2544 SDValue InFlag = Chain.getValue(1);
2546 // Create the CALLSEQ_END node.
2547 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2548 DAG.getIntPtrConstant(0, true), InFlag, DL);
2549 InFlag = Chain.getValue(1);
2551 // Handle result values, copying them out of physregs into vregs that we
2553 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg,
2554 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy);
2557 /// LowerCallResult - Lower the result values of a call into the
2558 /// appropriate copies out of appropriate physical registers.
2560 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2561 CallingConv::ID CallConv, bool IsVarArg,
2562 const SmallVectorImpl<ISD::InputArg> &Ins,
2563 SDLoc DL, SelectionDAG &DAG,
2564 SmallVectorImpl<SDValue> &InVals,
2565 const SDNode *CallNode,
2566 const Type *RetTy) const {
2567 // Assign locations to each value returned by this call.
2568 SmallVector<CCValAssign, 16> RVLocs;
2569 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2570 getTargetMachine(), RVLocs, *DAG.getContext());
2571 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo);
2573 MipsCCInfo.analyzeCallResult(Ins, Subtarget->mipsSEUsesSoftFloat(),
2576 // Copy all of the result registers out of their specified physreg.
2577 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2578 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2579 RVLocs[i].getLocVT(), InFlag);
2580 Chain = Val.getValue(1);
2581 InFlag = Val.getValue(2);
2583 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2584 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val);
2586 InVals.push_back(Val);
2592 //===----------------------------------------------------------------------===//
2593 // Formal Arguments Calling Convention Implementation
2594 //===----------------------------------------------------------------------===//
2595 /// LowerFormalArguments - transform physical registers into virtual registers
2596 /// and generate load operations for arguments places on the stack.
2598 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2599 CallingConv::ID CallConv,
2601 const SmallVectorImpl<ISD::InputArg> &Ins,
2602 SDLoc DL, SelectionDAG &DAG,
2603 SmallVectorImpl<SDValue> &InVals)
2605 MachineFunction &MF = DAG.getMachineFunction();
2606 MachineFrameInfo *MFI = MF.getFrameInfo();
2607 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2609 MipsFI->setVarArgsFrameIndex(0);
2611 // Used with vargs to acumulate store chains.
2612 std::vector<SDValue> OutChains;
2614 // Assign locations to all of the incoming arguments.
2615 SmallVector<CCValAssign, 16> ArgLocs;
2616 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2617 getTargetMachine(), ArgLocs, *DAG.getContext());
2618 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo);
2619 Function::const_arg_iterator FuncArg =
2620 DAG.getMachineFunction().getFunction()->arg_begin();
2621 bool UseSoftFloat = Subtarget->mipsSEUsesSoftFloat();
2623 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg);
2624 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2625 MipsCCInfo.hasByValArg());
2627 unsigned CurArgIdx = 0;
2628 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2630 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2631 CCValAssign &VA = ArgLocs[i];
2632 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
2633 CurArgIdx = Ins[i].OrigArgIndex;
2634 EVT ValVT = VA.getValVT();
2635 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2636 bool IsRegLoc = VA.isRegLoc();
2638 if (Flags.isByVal()) {
2639 assert(Flags.getByValSize() &&
2640 "ByVal args of size 0 should have been ignored by front-end.");
2641 assert(ByValArg != MipsCCInfo.byval_end());
2642 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2643 MipsCCInfo, *ByValArg);
2648 // Arguments stored on registers
2650 EVT RegVT = VA.getLocVT();
2651 unsigned ArgReg = VA.getLocReg();
2652 const TargetRegisterClass *RC;
2654 if (RegVT == MVT::i32)
2655 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass :
2656 &Mips::GPR32RegClass;
2657 else if (RegVT == MVT::i64)
2658 RC = &Mips::GPR64RegClass;
2659 else if (RegVT == MVT::f32)
2660 RC = &Mips::FGR32RegClass;
2661 else if (RegVT == MVT::f64)
2662 RC = Subtarget->isFP64bit() ? &Mips::FGR64RegClass :
2663 &Mips::AFGR64RegClass;
2665 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2667 // Transform the arguments stored on
2668 // physical registers into virtual ones
2669 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2670 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2672 // If this is an 8 or 16-bit value, it has been passed promoted
2673 // to 32 bits. Insert an assert[sz]ext to capture this, then
2674 // truncate to the right size.
2675 if (VA.getLocInfo() != CCValAssign::Full) {
2676 unsigned Opcode = 0;
2677 if (VA.getLocInfo() == CCValAssign::SExt)
2678 Opcode = ISD::AssertSext;
2679 else if (VA.getLocInfo() == CCValAssign::ZExt)
2680 Opcode = ISD::AssertZext;
2682 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue,
2683 DAG.getValueType(ValVT));
2684 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue);
2687 // Handle floating point arguments passed in integer registers and
2688 // long double arguments passed in floating point registers.
2689 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2690 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
2691 (RegVT == MVT::f64 && ValVT == MVT::i64))
2692 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
2693 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2694 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
2695 getNextIntArgReg(ArgReg), RC);
2696 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
2697 if (!Subtarget->isLittle())
2698 std::swap(ArgValue, ArgValue2);
2699 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
2700 ArgValue, ArgValue2);
2703 InVals.push_back(ArgValue);
2704 } else { // VA.isRegLoc()
2707 assert(VA.isMemLoc());
2709 // The stack pointer offset is relative to the caller stack frame.
2710 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2711 VA.getLocMemOffset(), true);
2713 // Create load nodes to retrieve arguments from the stack
2714 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2715 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN,
2716 MachinePointerInfo::getFixedStack(FI),
2717 false, false, false, 0));
2721 // The mips ABIs for returning structs by value requires that we copy
2722 // the sret argument into $v0 for the return. Save the argument into
2723 // a virtual register so that we can access it from the return points.
2724 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2725 unsigned Reg = MipsFI->getSRetReturnReg();
2727 Reg = MF.getRegInfo().
2728 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
2729 MipsFI->setSRetReturnReg(Reg);
2731 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]);
2732 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
2736 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG);
2738 // All stores are grouped in one node to allow the matching between
2739 // the size of Ins and InVals. This only happens when on varg functions
2740 if (!OutChains.empty()) {
2741 OutChains.push_back(Chain);
2742 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2743 &OutChains[0], OutChains.size());
2749 //===----------------------------------------------------------------------===//
2750 // Return Value Calling Convention Implementation
2751 //===----------------------------------------------------------------------===//
2754 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2755 MachineFunction &MF, bool IsVarArg,
2756 const SmallVectorImpl<ISD::OutputArg> &Outs,
2757 LLVMContext &Context) const {
2758 SmallVector<CCValAssign, 16> RVLocs;
2759 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(),
2761 return CCInfo.CheckReturn(Outs, RetCC_Mips);
2765 MipsTargetLowering::LowerReturn(SDValue Chain,
2766 CallingConv::ID CallConv, bool IsVarArg,
2767 const SmallVectorImpl<ISD::OutputArg> &Outs,
2768 const SmallVectorImpl<SDValue> &OutVals,
2769 SDLoc DL, SelectionDAG &DAG) const {
2770 // CCValAssign - represent the assignment of
2771 // the return value to a location
2772 SmallVector<CCValAssign, 16> RVLocs;
2773 MachineFunction &MF = DAG.getMachineFunction();
2775 // CCState - Info about the registers and stack slot.
2776 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs,
2778 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo);
2780 // Analyze return values.
2781 MipsCCInfo.analyzeReturn(Outs, Subtarget->mipsSEUsesSoftFloat(),
2782 MF.getFunction()->getReturnType());
2785 SmallVector<SDValue, 4> RetOps(1, Chain);
2787 // Copy the result values into the output registers.
2788 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2789 SDValue Val = OutVals[i];
2790 CCValAssign &VA = RVLocs[i];
2791 assert(VA.isRegLoc() && "Can only return in registers!");
2793 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2794 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val);
2796 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
2798 // Guarantee that all emitted copies are stuck together with flags.
2799 Flag = Chain.getValue(1);
2800 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2803 // The mips ABIs for returning structs by value requires that we copy
2804 // the sret argument into $v0 for the return. We saved the argument into
2805 // a virtual register in the entry block, so now we copy the value out
2807 if (MF.getFunction()->hasStructRetAttr()) {
2808 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2809 unsigned Reg = MipsFI->getSRetReturnReg();
2812 llvm_unreachable("sret virtual register not created in the entry block");
2813 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
2814 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
2816 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
2817 Flag = Chain.getValue(1);
2818 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
2821 RetOps[0] = Chain; // Update chain.
2823 // Add the flag if we have it.
2825 RetOps.push_back(Flag);
2827 // Return on Mips is always a "jr $ra"
2828 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size());
2831 //===----------------------------------------------------------------------===//
2832 // Mips Inline Assembly Support
2833 //===----------------------------------------------------------------------===//
2835 /// getConstraintType - Given a constraint letter, return the type of
2836 /// constraint it is for this target.
2837 MipsTargetLowering::ConstraintType MipsTargetLowering::
2838 getConstraintType(const std::string &Constraint) const
2840 // Mips specific constrainy
2841 // GCC config/mips/constraints.md
2843 // 'd' : An address register. Equivalent to r
2844 // unless generating MIPS16 code.
2845 // 'y' : Equivalent to r; retained for
2846 // backwards compatibility.
2847 // 'c' : A register suitable for use in an indirect
2848 // jump. This will always be $25 for -mabicalls.
2849 // 'l' : The lo register. 1 word storage.
2850 // 'x' : The hilo register pair. Double word storage.
2851 if (Constraint.size() == 1) {
2852 switch (Constraint[0]) {
2860 return C_RegisterClass;
2865 return TargetLowering::getConstraintType(Constraint);
2868 /// Examine constraint type and operand type and determine a weight value.
2869 /// This object must already have been set up with the operand type
2870 /// and the current alternative constraint selected.
2871 TargetLowering::ConstraintWeight
2872 MipsTargetLowering::getSingleConstraintMatchWeight(
2873 AsmOperandInfo &info, const char *constraint) const {
2874 ConstraintWeight weight = CW_Invalid;
2875 Value *CallOperandVal = info.CallOperandVal;
2876 // If we don't have a value, we can't do a match,
2877 // but allow it at the lowest weight.
2878 if (CallOperandVal == NULL)
2880 Type *type = CallOperandVal->getType();
2881 // Look at the constraint type.
2882 switch (*constraint) {
2884 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2888 if (type->isIntegerTy())
2889 weight = CW_Register;
2892 if (type->isFloatTy())
2893 weight = CW_Register;
2895 case 'c': // $25 for indirect jumps
2896 case 'l': // lo register
2897 case 'x': // hilo register pair
2898 if (type->isIntegerTy())
2899 weight = CW_SpecificReg;
2901 case 'I': // signed 16 bit immediate
2902 case 'J': // integer zero
2903 case 'K': // unsigned 16 bit immediate
2904 case 'L': // signed 32 bit immediate where lower 16 bits are 0
2905 case 'N': // immediate in the range of -65535 to -1 (inclusive)
2906 case 'O': // signed 15 bit immediate (+- 16383)
2907 case 'P': // immediate in the range of 65535 to 1 (inclusive)
2908 if (isa<ConstantInt>(CallOperandVal))
2909 weight = CW_Constant;
2918 /// This is a helper function to parse a physical register string and split it
2919 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
2920 /// that is returned indicates whether parsing was successful. The second flag
2921 /// is true if the numeric part exists.
2922 static std::pair<bool, bool>
2923 parsePhysicalReg(const StringRef &C, std::string &Prefix,
2924 unsigned long long &Reg) {
2925 if (C.front() != '{' || C.back() != '}')
2926 return std::make_pair(false, false);
2928 // Search for the first numeric character.
2929 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
2930 I = std::find_if(B, E, std::ptr_fun(isdigit));
2932 Prefix.assign(B, I - B);
2934 // The second flag is set to false if no numeric characters were found.
2936 return std::make_pair(true, false);
2938 // Parse the numeric characters.
2939 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
2943 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
2944 parseRegForInlineAsmConstraint(const StringRef &C, MVT VT) const {
2945 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2946 const TargetRegisterClass *RC;
2948 unsigned long long Reg;
2950 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
2953 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0);
2955 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
2956 // No numeric characters follow "hi" or "lo".
2958 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0);
2960 RC = TRI->getRegClass(Prefix == "hi" ?
2961 Mips::HI32RegClassID : Mips::LO32RegClassID);
2962 return std::make_pair(*(RC->begin()), RC);
2966 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0);
2968 if (Prefix == "$f") { // Parse $f0-$f31.
2969 // If the size of FP registers is 64-bit or Reg is an even number, select
2970 // the 64-bit register class. Otherwise, select the 32-bit register class.
2971 if (VT == MVT::Other)
2972 VT = (Subtarget->isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
2974 RC= getRegClassFor(VT);
2976 if (RC == &Mips::AFGR64RegClass) {
2977 assert(Reg % 2 == 0);
2980 } else if (Prefix == "$fcc") { // Parse $fcc0-$fcc7.
2981 RC = TRI->getRegClass(Mips::FCCRegClassID);
2982 } else { // Parse $0-$31.
2983 assert(Prefix == "$");
2984 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
2987 assert(Reg < RC->getNumRegs());
2988 return std::make_pair(*(RC->begin() + Reg), RC);
2991 /// Given a register class constraint, like 'r', if this corresponds directly
2992 /// to an LLVM register class, return a register of 0 and the register class
2994 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2995 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const
2997 if (Constraint.size() == 1) {
2998 switch (Constraint[0]) {
2999 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3000 case 'y': // Same as 'r'. Exists for compatibility.
3002 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3003 if (Subtarget->inMips16Mode())
3004 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3005 return std::make_pair(0U, &Mips::GPR32RegClass);
3007 if (VT == MVT::i64 && !HasMips64)
3008 return std::make_pair(0U, &Mips::GPR32RegClass);
3009 if (VT == MVT::i64 && HasMips64)
3010 return std::make_pair(0U, &Mips::GPR64RegClass);
3011 // This will generate an error message
3012 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3015 return std::make_pair(0U, &Mips::FGR32RegClass);
3016 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
3017 if (Subtarget->isFP64bit())
3018 return std::make_pair(0U, &Mips::FGR64RegClass);
3019 return std::make_pair(0U, &Mips::AFGR64RegClass);
3022 case 'c': // register suitable for indirect jump
3024 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
3025 assert(VT == MVT::i64 && "Unexpected type.");
3026 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
3027 case 'l': // register suitable for indirect jump
3029 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
3030 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
3031 case 'x': // register suitable for indirect jump
3032 // Fixme: Not triggering the use of both hi and low
3033 // This will generate an error message
3034 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3038 std::pair<unsigned, const TargetRegisterClass *> R;
3039 R = parseRegForInlineAsmConstraint(Constraint, VT);
3044 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3047 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3048 /// vector. If it is invalid, don't add anything to Ops.
3049 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3050 std::string &Constraint,
3051 std::vector<SDValue>&Ops,
3052 SelectionDAG &DAG) const {
3053 SDValue Result(0, 0);
3055 // Only support length 1 constraints for now.
3056 if (Constraint.length() > 1) return;
3058 char ConstraintLetter = Constraint[0];
3059 switch (ConstraintLetter) {
3060 default: break; // This will fall through to the generic implementation
3061 case 'I': // Signed 16 bit constant
3062 // If this fails, the parent routine will give an error
3063 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3064 EVT Type = Op.getValueType();
3065 int64_t Val = C->getSExtValue();
3066 if (isInt<16>(Val)) {
3067 Result = DAG.getTargetConstant(Val, Type);
3072 case 'J': // integer zero
3073 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3074 EVT Type = Op.getValueType();
3075 int64_t Val = C->getZExtValue();
3077 Result = DAG.getTargetConstant(0, Type);
3082 case 'K': // unsigned 16 bit immediate
3083 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3084 EVT Type = Op.getValueType();
3085 uint64_t Val = (uint64_t)C->getZExtValue();
3086 if (isUInt<16>(Val)) {
3087 Result = DAG.getTargetConstant(Val, Type);
3092 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3093 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3094 EVT Type = Op.getValueType();
3095 int64_t Val = C->getSExtValue();
3096 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3097 Result = DAG.getTargetConstant(Val, Type);
3102 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3103 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3104 EVT Type = Op.getValueType();
3105 int64_t Val = C->getSExtValue();
3106 if ((Val >= -65535) && (Val <= -1)) {
3107 Result = DAG.getTargetConstant(Val, Type);
3112 case 'O': // signed 15 bit immediate
3113 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3114 EVT Type = Op.getValueType();
3115 int64_t Val = C->getSExtValue();
3116 if ((isInt<15>(Val))) {
3117 Result = DAG.getTargetConstant(Val, Type);
3122 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3123 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3124 EVT Type = Op.getValueType();
3125 int64_t Val = C->getSExtValue();
3126 if ((Val <= 65535) && (Val >= 1)) {
3127 Result = DAG.getTargetConstant(Val, Type);
3134 if (Result.getNode()) {
3135 Ops.push_back(Result);
3139 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3143 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
3144 // No global is ever allowed as a base.
3149 case 0: // "r+i" or just "i", depending on HasBaseReg.
3152 if (!AM.HasBaseReg) // allow "r+i".
3154 return false; // disallow "r+r" or "r+r+i".
3163 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3164 // The Mips target isn't yet aware of offsets.
3168 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3170 bool IsMemset, bool ZeroMemset,
3172 MachineFunction &MF) const {
3173 if (Subtarget->hasMips64())
3179 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3180 if (VT != MVT::f32 && VT != MVT::f64)
3182 if (Imm.isNegZero())
3184 return Imm.isZero();
3187 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3189 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3191 return TargetLowering::getJumpTableEncoding();
3194 /// This function returns true if CallSym is a long double emulation routine.
3195 static bool isF128SoftLibCall(const char *CallSym) {
3196 const char *const LibCalls[] =
3197 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2",
3198 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi",
3199 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf",
3200 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2",
3201 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3",
3202 "__trunctfdf2", "__trunctfsf2", "__unordtf2",
3203 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl",
3204 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl",
3207 const char * const *End = LibCalls + array_lengthof(LibCalls);
3209 // Check that LibCalls is sorted alphabetically.
3210 MipsTargetLowering::LTStr Comp;
3213 for (const char * const *I = LibCalls; I < End - 1; ++I)
3214 assert(Comp(*I, *(I + 1)));
3217 return std::binary_search(LibCalls, End, CallSym, Comp);
3220 /// This function returns true if Ty is fp128 or i128 which was originally a
3222 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) {
3223 if (Ty->isFP128Ty())
3226 const ExternalSymbolSDNode *ES =
3227 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
3229 // If the Ty is i128 and the function being called is a long double emulation
3230 // routine, then the original type is f128.
3231 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
3234 MipsTargetLowering::MipsCC::SpecialCallingConvType
3235 MipsTargetLowering::getSpecialCallingConv(SDValue Callee) const {
3236 MipsCC::SpecialCallingConvType SpecialCallingConv =
3237 MipsCC::NoSpecialCallingConv;;
3238 if (Subtarget->inMips16HardFloat()) {
3239 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3240 llvm::StringRef Sym = G->getGlobal()->getName();
3241 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
3242 if (F->hasFnAttribute("__Mips16RetHelper")) {
3243 SpecialCallingConv = MipsCC::Mips16RetHelperConv;
3247 return SpecialCallingConv;
3250 MipsTargetLowering::MipsCC::MipsCC(
3251 CallingConv::ID CC, bool IsO32_, bool IsFP64_, CCState &Info,
3252 MipsCC::SpecialCallingConvType SpecialCallingConv_)
3253 : CCInfo(Info), CallConv(CC), IsO32(IsO32_), IsFP64(IsFP64_),
3254 SpecialCallingConv(SpecialCallingConv_){
3255 // Pre-allocate reserved argument area.
3256 CCInfo.AllocateStack(reservedArgArea(), 1);
3260 void MipsTargetLowering::MipsCC::
3261 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
3262 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode,
3263 std::vector<ArgListEntry> &FuncArgs) {
3264 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
3265 "CallingConv::Fast shouldn't be used for vararg functions.");
3267 unsigned NumOpnds = Args.size();
3268 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
3270 for (unsigned I = 0; I != NumOpnds; ++I) {
3271 MVT ArgVT = Args[I].VT;
3272 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3275 if (ArgFlags.isByVal()) {
3276 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3280 if (IsVarArg && !Args[I].IsFixed)
3281 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3283 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode,
3285 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo);
3290 dbgs() << "Call operand #" << I << " has unhandled type "
3291 << EVT(ArgVT).getEVTString();
3293 llvm_unreachable(0);
3298 void MipsTargetLowering::MipsCC::
3299 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args,
3300 bool IsSoftFloat, Function::const_arg_iterator FuncArg) {
3301 unsigned NumArgs = Args.size();
3302 llvm::CCAssignFn *FixedFn = fixedArgFn();
3303 unsigned CurArgIdx = 0;
3305 for (unsigned I = 0; I != NumArgs; ++I) {
3306 MVT ArgVT = Args[I].VT;
3307 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3308 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx);
3309 CurArgIdx = Args[I].OrigArgIndex;
3311 if (ArgFlags.isByVal()) {
3312 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3316 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat);
3318 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo))
3322 dbgs() << "Formal Arg #" << I << " has unhandled type "
3323 << EVT(ArgVT).getEVTString();
3325 llvm_unreachable(0);
3329 template<typename Ty>
3330 void MipsTargetLowering::MipsCC::
3331 analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat,
3332 const SDNode *CallNode, const Type *RetTy) const {
3335 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode))
3336 Fn = RetCC_F128Soft;
3340 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) {
3341 MVT VT = RetVals[I].VT;
3342 ISD::ArgFlagsTy Flags = RetVals[I].Flags;
3343 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat);
3345 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) {
3347 dbgs() << "Call result #" << I << " has unhandled type "
3348 << EVT(VT).getEVTString() << '\n';
3350 llvm_unreachable(0);
3355 void MipsTargetLowering::MipsCC::
3356 analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat,
3357 const SDNode *CallNode, const Type *RetTy) const {
3358 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy);
3361 void MipsTargetLowering::MipsCC::
3362 analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat,
3363 const Type *RetTy) const {
3364 analyzeReturn(Outs, IsSoftFloat, 0, RetTy);
3368 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3370 CCValAssign::LocInfo LocInfo,
3371 ISD::ArgFlagsTy ArgFlags) {
3372 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3374 struct ByValArgInfo ByVal;
3375 unsigned RegSize = regSize();
3376 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3377 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3380 if (useRegsForByval())
3381 allocateRegs(ByVal, ByValSize, Align);
3383 // Allocate space on caller's stack.
3384 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3386 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3388 ByValArgs.push_back(ByVal);
3391 unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const {
3392 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs);
3395 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
3396 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0;
3399 const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const {
3400 return IsO32 ? O32IntRegs : Mips64IntRegs;
3403 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
3404 if (CallConv == CallingConv::Fast)
3405 return CC_Mips_FastCC;
3407 if (SpecialCallingConv == Mips16RetHelperConv)
3408 return CC_Mips16RetHelper;
3409 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN;
3412 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
3413 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN_VarArg;
3416 const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const {
3417 return IsO32 ? O32IntRegs : Mips64DPRegs;
3420 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3423 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs();
3424 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs();
3425 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3426 "Byval argument's size and alignment should be a multiple of"
3429 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3431 // If Align > RegSize, the first arg register must be even.
3432 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3433 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3437 // Mark the registers allocated.
3438 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3439 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3440 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3443 MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy,
3444 const SDNode *CallNode,
3445 bool IsSoftFloat) const {
3446 if (IsSoftFloat || IsO32)
3449 // Check if the original type was fp128.
3450 if (originalTypeIsF128(OrigTy, CallNode)) {
3451 assert(VT == MVT::i64);
3458 void MipsTargetLowering::
3459 copyByValRegs(SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains,
3460 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3461 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3462 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3463 MachineFunction &MF = DAG.getMachineFunction();
3464 MachineFrameInfo *MFI = MF.getFrameInfo();
3465 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3466 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3470 FrameObjOffset = (int)CC.reservedArgArea() -
3471 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3473 FrameObjOffset = ByVal.Address;
3475 // Create frame object.
3476 EVT PtrTy = getPointerTy();
3477 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3478 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3479 InVals.push_back(FIN);
3484 // Copy arg registers.
3485 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3486 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3488 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3489 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3490 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3491 unsigned Offset = I * CC.regSize();
3492 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3493 DAG.getConstant(Offset, PtrTy));
3494 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3495 StorePtr, MachinePointerInfo(FuncArg, Offset),
3497 OutChains.push_back(Store);
3501 // Copy byVal arg to registers and stack.
3502 void MipsTargetLowering::
3503 passByValArg(SDValue Chain, SDLoc DL,
3504 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
3505 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
3506 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3507 const MipsCC &CC, const ByValArgInfo &ByVal,
3508 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3509 unsigned ByValSize = Flags.getByValSize();
3510 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
3511 unsigned RegSize = CC.regSize();
3512 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
3513 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
3515 if (ByVal.NumRegs) {
3516 const uint16_t *ArgRegs = CC.intArgRegs();
3517 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
3520 // Copy words to registers.
3521 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
3522 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3523 DAG.getConstant(Offset, PtrTy));
3524 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3525 MachinePointerInfo(), false, false, false,
3527 MemOpChains.push_back(LoadVal.getValue(1));
3528 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3529 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3532 // Return if the struct has been fully copied.
3533 if (ByValSize == Offset)
3536 // Copy the remainder of the byval argument with sub-word loads and shifts.
3537 if (LeftoverBytes) {
3538 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
3539 "Size of the remainder should be smaller than RegSize.");
3542 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
3543 Offset < ByValSize; LoadSize /= 2) {
3544 unsigned RemSize = ByValSize - Offset;
3546 if (RemSize < LoadSize)
3550 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3551 DAG.getConstant(Offset, PtrTy));
3553 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
3554 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
3555 false, false, Alignment);
3556 MemOpChains.push_back(LoadVal.getValue(1));
3558 // Shift the loaded value.
3562 Shamt = TotalSizeLoaded;
3564 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
3566 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3567 DAG.getConstant(Shamt, MVT::i32));
3570 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3575 TotalSizeLoaded += LoadSize;
3576 Alignment = std::min(Alignment, LoadSize);
3579 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3580 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3585 // Copy remainder of byval arg to it with memcpy.
3586 unsigned MemCpySize = ByValSize - Offset;
3587 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3588 DAG.getConstant(Offset, PtrTy));
3589 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3590 DAG.getIntPtrConstant(ByVal.Address));
3591 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3592 DAG.getConstant(MemCpySize, PtrTy), Alignment,
3593 /*isVolatile=*/false, /*AlwaysInline=*/false,
3594 MachinePointerInfo(0), MachinePointerInfo(0));
3595 MemOpChains.push_back(Chain);
3599 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3600 const MipsCC &CC, SDValue Chain,
3601 SDLoc DL, SelectionDAG &DAG) const {
3602 unsigned NumRegs = CC.numIntArgRegs();
3603 const uint16_t *ArgRegs = CC.intArgRegs();
3604 const CCState &CCInfo = CC.getCCInfo();
3605 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3606 unsigned RegSize = CC.regSize();
3607 MVT RegTy = MVT::getIntegerVT(RegSize * 8);
3608 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3609 MachineFunction &MF = DAG.getMachineFunction();
3610 MachineFrameInfo *MFI = MF.getFrameInfo();
3611 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3613 // Offset of the first variable argument from stack pointer.
3617 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3620 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3622 // Record the frame index of the first variable argument
3623 // which is a value necessary to VASTART.
3624 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3625 MipsFI->setVarArgsFrameIndex(FI);
3627 // Copy the integer registers that have not been used for argument passing
3628 // to the argument register save area. For O32, the save area is allocated
3629 // in the caller's stack frame, while for N32/64, it is allocated in the
3630 // callee's stack frame.
3631 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3632 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3633 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3634 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3635 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3636 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3637 MachinePointerInfo(), false, false, 0);
3638 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
3639 OutChains.push_back(Store);