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 static SDValue getTargetNode(SDValue Op, SelectionDAG &DAG, unsigned Flag) {
84 EVT Ty = Op.getValueType();
86 if (GlobalAddressSDNode *N = dyn_cast<GlobalAddressSDNode>(Op))
87 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(Op), Ty, 0,
89 if (ExternalSymbolSDNode *N = dyn_cast<ExternalSymbolSDNode>(Op))
90 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
91 if (BlockAddressSDNode *N = dyn_cast<BlockAddressSDNode>(Op))
92 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
93 if (JumpTableSDNode *N = dyn_cast<JumpTableSDNode>(Op))
94 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
95 if (ConstantPoolSDNode *N = dyn_cast<ConstantPoolSDNode>(Op))
96 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
97 N->getOffset(), Flag);
99 llvm_unreachable("Unexpected node type.");
103 static SDValue getAddrNonPIC(SDValue Op, SelectionDAG &DAG) {
105 EVT Ty = Op.getValueType();
106 SDValue Hi = getTargetNode(Op, DAG, MipsII::MO_ABS_HI);
107 SDValue Lo = getTargetNode(Op, DAG, MipsII::MO_ABS_LO);
108 return DAG.getNode(ISD::ADD, DL, Ty,
109 DAG.getNode(MipsISD::Hi, DL, Ty, Hi),
110 DAG.getNode(MipsISD::Lo, DL, Ty, Lo));
113 SDValue MipsTargetLowering::getAddrLocal(SDValue Op, SelectionDAG &DAG,
114 bool HasMips64) const {
116 EVT Ty = Op.getValueType();
117 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
118 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
119 getTargetNode(Op, DAG, GOTFlag));
120 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT,
121 MachinePointerInfo::getGOT(), false, false, false,
123 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
124 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty, getTargetNode(Op, DAG, LoFlag));
125 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo);
128 SDValue MipsTargetLowering::getAddrGlobal(SDValue Op, SelectionDAG &DAG,
129 unsigned Flag) const {
131 EVT Ty = Op.getValueType();
132 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
133 getTargetNode(Op, DAG, Flag));
134 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt,
135 MachinePointerInfo::getGOT(), false, false, false, 0);
138 SDValue MipsTargetLowering::getAddrGlobalLargeGOT(SDValue Op, SelectionDAG &DAG,
140 unsigned LoFlag) const {
142 EVT Ty = Op.getValueType();
143 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty, getTargetNode(Op, DAG, HiFlag));
144 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, getGlobalReg(DAG, Ty));
145 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi,
146 getTargetNode(Op, DAG, LoFlag));
147 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper,
148 MachinePointerInfo::getGOT(), false, false, false, 0);
151 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
153 case MipsISD::JmpLink: return "MipsISD::JmpLink";
154 case MipsISD::TailCall: return "MipsISD::TailCall";
155 case MipsISD::Hi: return "MipsISD::Hi";
156 case MipsISD::Lo: return "MipsISD::Lo";
157 case MipsISD::GPRel: return "MipsISD::GPRel";
158 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
159 case MipsISD::Ret: return "MipsISD::Ret";
160 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
161 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
162 case MipsISD::FPCmp: return "MipsISD::FPCmp";
163 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
164 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
165 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP";
166 case MipsISD::ExtractLOHI: return "MipsISD::ExtractLOHI";
167 case MipsISD::InsertLOHI: return "MipsISD::InsertLOHI";
168 case MipsISD::Mult: return "MipsISD::Mult";
169 case MipsISD::Multu: return "MipsISD::Multu";
170 case MipsISD::MAdd: return "MipsISD::MAdd";
171 case MipsISD::MAddu: return "MipsISD::MAddu";
172 case MipsISD::MSub: return "MipsISD::MSub";
173 case MipsISD::MSubu: return "MipsISD::MSubu";
174 case MipsISD::DivRem: return "MipsISD::DivRem";
175 case MipsISD::DivRemU: return "MipsISD::DivRemU";
176 case MipsISD::DivRem16: return "MipsISD::DivRem16";
177 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
178 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
179 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
180 case MipsISD::Wrapper: return "MipsISD::Wrapper";
181 case MipsISD::Sync: return "MipsISD::Sync";
182 case MipsISD::Ext: return "MipsISD::Ext";
183 case MipsISD::Ins: return "MipsISD::Ins";
184 case MipsISD::LWL: return "MipsISD::LWL";
185 case MipsISD::LWR: return "MipsISD::LWR";
186 case MipsISD::SWL: return "MipsISD::SWL";
187 case MipsISD::SWR: return "MipsISD::SWR";
188 case MipsISD::LDL: return "MipsISD::LDL";
189 case MipsISD::LDR: return "MipsISD::LDR";
190 case MipsISD::SDL: return "MipsISD::SDL";
191 case MipsISD::SDR: return "MipsISD::SDR";
192 case MipsISD::EXTP: return "MipsISD::EXTP";
193 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
194 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
195 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
196 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
197 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
198 case MipsISD::SHILO: return "MipsISD::SHILO";
199 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
200 case MipsISD::MULT: return "MipsISD::MULT";
201 case MipsISD::MULTU: return "MipsISD::MULTU";
202 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
203 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
204 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
205 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
206 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP";
207 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP";
208 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP";
209 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP";
210 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP";
211 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO";
212 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO";
213 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO";
214 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO";
215 case MipsISD::VCEQ: return "MipsISD::VCEQ";
216 case MipsISD::VCLE_S: return "MipsISD::VCLE_S";
217 case MipsISD::VCLE_U: return "MipsISD::VCLE_U";
218 case MipsISD::VCLT_S: return "MipsISD::VCLT_S";
219 case MipsISD::VCLT_U: return "MipsISD::VCLT_U";
220 case MipsISD::VSMAX: return "MipsISD::VSMAX";
221 case MipsISD::VSMIN: return "MipsISD::VSMIN";
222 case MipsISD::VUMAX: return "MipsISD::VUMAX";
223 case MipsISD::VUMIN: return "MipsISD::VUMIN";
224 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
225 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
226 case MipsISD::VNOR: return "MipsISD::VNOR";
227 case MipsISD::VSHF: return "MipsISD::VSHF";
228 case MipsISD::SHF: return "MipsISD::SHF";
229 case MipsISD::ILVEV: return "MipsISD::ILVEV";
230 case MipsISD::ILVOD: return "MipsISD::ILVOD";
231 case MipsISD::ILVL: return "MipsISD::ILVL";
232 case MipsISD::ILVR: return "MipsISD::ILVR";
233 default: return NULL;
238 MipsTargetLowering(MipsTargetMachine &TM)
239 : TargetLowering(TM, new MipsTargetObjectFile()),
240 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
241 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
242 IsO32(Subtarget->isABI_O32()) {
243 // Mips does not have i1 type, so use i32 for
244 // setcc operations results (slt, sgt, ...).
245 setBooleanContents(ZeroOrOneBooleanContent);
246 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
248 // Load extented operations for i1 types must be promoted
249 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
250 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
251 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
253 // MIPS doesn't have extending float->double load/store
254 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
255 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
257 // Used by legalize types to correctly generate the setcc result.
258 // Without this, every float setcc comes with a AND/OR with the result,
259 // we don't want this, since the fpcmp result goes to a flag register,
260 // which is used implicitly by brcond and select operations.
261 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
263 // Mips Custom Operations
264 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
265 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
266 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
267 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
268 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
269 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
270 setOperationAction(ISD::SELECT, MVT::f32, Custom);
271 setOperationAction(ISD::SELECT, MVT::f64, Custom);
272 setOperationAction(ISD::SELECT, MVT::i32, Custom);
273 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
274 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
275 setOperationAction(ISD::SETCC, MVT::f32, Custom);
276 setOperationAction(ISD::SETCC, MVT::f64, Custom);
277 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
278 setOperationAction(ISD::VASTART, MVT::Other, Custom);
279 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
280 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
281 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
283 if (!TM.Options.NoNaNsFPMath) {
284 setOperationAction(ISD::FABS, MVT::f32, Custom);
285 setOperationAction(ISD::FABS, MVT::f64, Custom);
289 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
290 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
291 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
292 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
293 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
294 setOperationAction(ISD::SELECT, MVT::i64, Custom);
295 setOperationAction(ISD::LOAD, MVT::i64, Custom);
296 setOperationAction(ISD::STORE, MVT::i64, Custom);
297 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
301 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
302 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
303 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
306 setOperationAction(ISD::ADD, MVT::i32, Custom);
308 setOperationAction(ISD::ADD, MVT::i64, Custom);
310 setOperationAction(ISD::SDIV, MVT::i32, Expand);
311 setOperationAction(ISD::SREM, MVT::i32, Expand);
312 setOperationAction(ISD::UDIV, MVT::i32, Expand);
313 setOperationAction(ISD::UREM, MVT::i32, Expand);
314 setOperationAction(ISD::SDIV, MVT::i64, Expand);
315 setOperationAction(ISD::SREM, MVT::i64, Expand);
316 setOperationAction(ISD::UDIV, MVT::i64, Expand);
317 setOperationAction(ISD::UREM, MVT::i64, Expand);
319 // Operations not directly supported by Mips.
320 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
321 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
322 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
323 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
324 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
325 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
326 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
327 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
328 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
329 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
330 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
331 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
332 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
333 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
334 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
335 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
336 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
337 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
338 setOperationAction(ISD::ROTL, MVT::i32, Expand);
339 setOperationAction(ISD::ROTL, MVT::i64, Expand);
340 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
341 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
343 if (!Subtarget->hasMips32r2())
344 setOperationAction(ISD::ROTR, MVT::i32, Expand);
346 if (!Subtarget->hasMips64r2())
347 setOperationAction(ISD::ROTR, MVT::i64, Expand);
349 setOperationAction(ISD::FSIN, MVT::f32, Expand);
350 setOperationAction(ISD::FSIN, MVT::f64, Expand);
351 setOperationAction(ISD::FCOS, MVT::f32, Expand);
352 setOperationAction(ISD::FCOS, MVT::f64, Expand);
353 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
354 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
355 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
356 setOperationAction(ISD::FPOW, MVT::f32, Expand);
357 setOperationAction(ISD::FPOW, MVT::f64, Expand);
358 setOperationAction(ISD::FLOG, MVT::f32, Expand);
359 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
360 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
361 setOperationAction(ISD::FEXP, MVT::f32, Expand);
362 setOperationAction(ISD::FMA, MVT::f32, Expand);
363 setOperationAction(ISD::FMA, MVT::f64, Expand);
364 setOperationAction(ISD::FREM, MVT::f32, Expand);
365 setOperationAction(ISD::FREM, MVT::f64, Expand);
367 if (!TM.Options.NoNaNsFPMath) {
368 setOperationAction(ISD::FNEG, MVT::f32, Expand);
369 setOperationAction(ISD::FNEG, MVT::f64, Expand);
372 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
374 setOperationAction(ISD::VAARG, MVT::Other, Expand);
375 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
376 setOperationAction(ISD::VAEND, MVT::Other, Expand);
378 // Use the default for now
379 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
380 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
382 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
383 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
384 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
385 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
387 setInsertFencesForAtomic(true);
389 if (!Subtarget->hasSEInReg()) {
390 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
391 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
394 if (!Subtarget->hasBitCount()) {
395 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
396 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
399 if (!Subtarget->hasSwap()) {
400 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
401 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
405 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
406 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
407 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
408 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
411 setOperationAction(ISD::TRAP, MVT::Other, Legal);
413 setTargetDAGCombine(ISD::SDIVREM);
414 setTargetDAGCombine(ISD::UDIVREM);
415 setTargetDAGCombine(ISD::SELECT);
416 setTargetDAGCombine(ISD::AND);
417 setTargetDAGCombine(ISD::OR);
418 setTargetDAGCombine(ISD::ADD);
420 setMinFunctionAlignment(HasMips64 ? 3 : 2);
422 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP);
424 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0);
425 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1);
427 MaxStoresPerMemcpy = 16;
430 const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) {
431 if (TM.getSubtargetImpl()->inMips16Mode())
432 return llvm::createMips16TargetLowering(TM);
434 return llvm::createMipsSETargetLowering(TM);
437 EVT MipsTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
440 return VT.changeVectorElementTypeToInteger();
443 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
444 TargetLowering::DAGCombinerInfo &DCI,
445 const MipsSubtarget *Subtarget) {
446 if (DCI.isBeforeLegalizeOps())
449 EVT Ty = N->getValueType(0);
450 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
451 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
452 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
456 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
457 N->getOperand(0), N->getOperand(1));
458 SDValue InChain = DAG.getEntryNode();
459 SDValue InGlue = DivRem;
462 if (N->hasAnyUseOfValue(0)) {
463 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
465 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
466 InChain = CopyFromLo.getValue(1);
467 InGlue = CopyFromLo.getValue(2);
471 if (N->hasAnyUseOfValue(1)) {
472 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
474 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
480 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
482 default: llvm_unreachable("Unknown fp condition code!");
484 case ISD::SETOEQ: return Mips::FCOND_OEQ;
485 case ISD::SETUNE: return Mips::FCOND_UNE;
487 case ISD::SETOLT: return Mips::FCOND_OLT;
489 case ISD::SETOGT: return Mips::FCOND_OGT;
491 case ISD::SETOLE: return Mips::FCOND_OLE;
493 case ISD::SETOGE: return Mips::FCOND_OGE;
494 case ISD::SETULT: return Mips::FCOND_ULT;
495 case ISD::SETULE: return Mips::FCOND_ULE;
496 case ISD::SETUGT: return Mips::FCOND_UGT;
497 case ISD::SETUGE: return Mips::FCOND_UGE;
498 case ISD::SETUO: return Mips::FCOND_UN;
499 case ISD::SETO: return Mips::FCOND_OR;
501 case ISD::SETONE: return Mips::FCOND_ONE;
502 case ISD::SETUEQ: return Mips::FCOND_UEQ;
507 /// This function returns true if the floating point conditional branches and
508 /// conditional moves which use condition code CC should be inverted.
509 static bool invertFPCondCodeUser(Mips::CondCode CC) {
510 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
513 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
514 "Illegal Condition Code");
519 // Creates and returns an FPCmp node from a setcc node.
520 // Returns Op if setcc is not a floating point comparison.
521 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
522 // must be a SETCC node
523 if (Op.getOpcode() != ISD::SETCC)
526 SDValue LHS = Op.getOperand(0);
528 if (!LHS.getValueType().isFloatingPoint())
531 SDValue RHS = Op.getOperand(1);
534 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
535 // node if necessary.
536 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
538 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
539 DAG.getConstant(condCodeToFCC(CC), MVT::i32));
542 // Creates and returns a CMovFPT/F node.
543 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
544 SDValue False, SDLoc DL) {
545 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
546 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
547 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
549 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
550 True.getValueType(), True, FCC0, False, Cond);
553 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
554 TargetLowering::DAGCombinerInfo &DCI,
555 const MipsSubtarget *Subtarget) {
556 if (DCI.isBeforeLegalizeOps())
559 SDValue SetCC = N->getOperand(0);
561 if ((SetCC.getOpcode() != ISD::SETCC) ||
562 !SetCC.getOperand(0).getValueType().isInteger())
565 SDValue False = N->getOperand(2);
566 EVT FalseTy = False.getValueType();
568 if (!FalseTy.isInteger())
571 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
573 if (!CN || CN->getZExtValue())
577 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
578 SDValue True = N->getOperand(1);
580 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
581 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
583 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
586 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
587 TargetLowering::DAGCombinerInfo &DCI,
588 const MipsSubtarget *Subtarget) {
589 // Pattern match EXT.
590 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
591 // => ext $dst, $src, size, pos
592 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
595 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
596 unsigned ShiftRightOpc = ShiftRight.getOpcode();
598 // Op's first operand must be a shift right.
599 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
602 // The second operand of the shift must be an immediate.
604 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
607 uint64_t Pos = CN->getZExtValue();
608 uint64_t SMPos, SMSize;
610 // Op's second operand must be a shifted mask.
611 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
612 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
615 // Return if the shifted mask does not start at bit 0 or the sum of its size
616 // and Pos exceeds the word's size.
617 EVT ValTy = N->getValueType(0);
618 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
621 return DAG.getNode(MipsISD::Ext, SDLoc(N), ValTy,
622 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
623 DAG.getConstant(SMSize, MVT::i32));
626 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
627 TargetLowering::DAGCombinerInfo &DCI,
628 const MipsSubtarget *Subtarget) {
629 // Pattern match INS.
630 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
631 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
632 // => ins $dst, $src, size, pos, $src1
633 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
636 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
637 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
640 // See if Op's first operand matches (and $src1 , mask0).
641 if (And0.getOpcode() != ISD::AND)
644 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
645 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
648 // See if Op's second operand matches (and (shl $src, pos), mask1).
649 if (And1.getOpcode() != ISD::AND)
652 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
653 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
656 // The shift masks must have the same position and size.
657 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
660 SDValue Shl = And1.getOperand(0);
661 if (Shl.getOpcode() != ISD::SHL)
664 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
667 unsigned Shamt = CN->getZExtValue();
669 // Return if the shift amount and the first bit position of mask are not the
671 EVT ValTy = N->getValueType(0);
672 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
675 return DAG.getNode(MipsISD::Ins, SDLoc(N), ValTy, Shl.getOperand(0),
676 DAG.getConstant(SMPos0, MVT::i32),
677 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
680 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
681 TargetLowering::DAGCombinerInfo &DCI,
682 const MipsSubtarget *Subtarget) {
683 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
685 if (DCI.isBeforeLegalizeOps())
688 SDValue Add = N->getOperand(1);
690 if (Add.getOpcode() != ISD::ADD)
693 SDValue Lo = Add.getOperand(1);
695 if ((Lo.getOpcode() != MipsISD::Lo) ||
696 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
699 EVT ValTy = N->getValueType(0);
702 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
704 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
707 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
709 SelectionDAG &DAG = DCI.DAG;
710 unsigned Opc = N->getOpcode();
716 return performDivRemCombine(N, DAG, DCI, Subtarget);
718 return performSELECTCombine(N, DAG, DCI, Subtarget);
720 return performANDCombine(N, DAG, DCI, Subtarget);
722 return performORCombine(N, DAG, DCI, Subtarget);
724 return performADDCombine(N, DAG, DCI, Subtarget);
731 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
732 SmallVectorImpl<SDValue> &Results,
733 SelectionDAG &DAG) const {
734 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
736 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
737 Results.push_back(Res.getValue(I));
741 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
742 SmallVectorImpl<SDValue> &Results,
743 SelectionDAG &DAG) const {
744 return LowerOperationWrapper(N, Results, DAG);
747 SDValue MipsTargetLowering::
748 LowerOperation(SDValue Op, SelectionDAG &DAG) const
750 switch (Op.getOpcode())
752 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
753 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
754 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
755 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
756 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
757 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
758 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
759 case ISD::SELECT: return lowerSELECT(Op, DAG);
760 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
761 case ISD::SETCC: return lowerSETCC(Op, DAG);
762 case ISD::VASTART: return lowerVASTART(Op, DAG);
763 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
764 case ISD::FABS: return lowerFABS(Op, DAG);
765 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
766 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
767 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
768 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
769 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
770 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
771 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
772 case ISD::LOAD: return lowerLOAD(Op, DAG);
773 case ISD::STORE: return lowerSTORE(Op, DAG);
774 case ISD::ADD: return lowerADD(Op, DAG);
775 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG);
780 //===----------------------------------------------------------------------===//
781 // Lower helper functions
782 //===----------------------------------------------------------------------===//
784 // addLiveIn - This helper function adds the specified physical register to the
785 // MachineFunction as a live in value. It also creates a corresponding
786 // virtual register for it.
788 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
790 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
791 MF.getRegInfo().addLiveIn(PReg, VReg);
795 static MachineBasicBlock *expandPseudoDIV(MachineInstr *MI,
796 MachineBasicBlock &MBB,
797 const TargetInstrInfo &TII,
802 // Insert instruction "teq $divisor_reg, $zero, 7".
803 MachineBasicBlock::iterator I(MI);
804 MachineInstrBuilder MIB;
805 MIB = BuildMI(MBB, llvm::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ))
806 .addOperand(MI->getOperand(2)).addReg(Mips::ZERO).addImm(7);
808 // Use the 32-bit sub-register if this is a 64-bit division.
810 MIB->getOperand(0).setSubReg(Mips::sub_32);
816 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
817 MachineBasicBlock *BB) const {
818 switch (MI->getOpcode()) {
820 llvm_unreachable("Unexpected instr type to insert");
821 case Mips::ATOMIC_LOAD_ADD_I8:
822 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
823 case Mips::ATOMIC_LOAD_ADD_I16:
824 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
825 case Mips::ATOMIC_LOAD_ADD_I32:
826 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
827 case Mips::ATOMIC_LOAD_ADD_I64:
828 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
830 case Mips::ATOMIC_LOAD_AND_I8:
831 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
832 case Mips::ATOMIC_LOAD_AND_I16:
833 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
834 case Mips::ATOMIC_LOAD_AND_I32:
835 return emitAtomicBinary(MI, BB, 4, Mips::AND);
836 case Mips::ATOMIC_LOAD_AND_I64:
837 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
839 case Mips::ATOMIC_LOAD_OR_I8:
840 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
841 case Mips::ATOMIC_LOAD_OR_I16:
842 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
843 case Mips::ATOMIC_LOAD_OR_I32:
844 return emitAtomicBinary(MI, BB, 4, Mips::OR);
845 case Mips::ATOMIC_LOAD_OR_I64:
846 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
848 case Mips::ATOMIC_LOAD_XOR_I8:
849 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
850 case Mips::ATOMIC_LOAD_XOR_I16:
851 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
852 case Mips::ATOMIC_LOAD_XOR_I32:
853 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
854 case Mips::ATOMIC_LOAD_XOR_I64:
855 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
857 case Mips::ATOMIC_LOAD_NAND_I8:
858 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
859 case Mips::ATOMIC_LOAD_NAND_I16:
860 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
861 case Mips::ATOMIC_LOAD_NAND_I32:
862 return emitAtomicBinary(MI, BB, 4, 0, true);
863 case Mips::ATOMIC_LOAD_NAND_I64:
864 return emitAtomicBinary(MI, BB, 8, 0, true);
866 case Mips::ATOMIC_LOAD_SUB_I8:
867 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
868 case Mips::ATOMIC_LOAD_SUB_I16:
869 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
870 case Mips::ATOMIC_LOAD_SUB_I32:
871 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
872 case Mips::ATOMIC_LOAD_SUB_I64:
873 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
875 case Mips::ATOMIC_SWAP_I8:
876 return emitAtomicBinaryPartword(MI, BB, 1, 0);
877 case Mips::ATOMIC_SWAP_I16:
878 return emitAtomicBinaryPartword(MI, BB, 2, 0);
879 case Mips::ATOMIC_SWAP_I32:
880 return emitAtomicBinary(MI, BB, 4, 0);
881 case Mips::ATOMIC_SWAP_I64:
882 return emitAtomicBinary(MI, BB, 8, 0);
884 case Mips::ATOMIC_CMP_SWAP_I8:
885 return emitAtomicCmpSwapPartword(MI, BB, 1);
886 case Mips::ATOMIC_CMP_SWAP_I16:
887 return emitAtomicCmpSwapPartword(MI, BB, 2);
888 case Mips::ATOMIC_CMP_SWAP_I32:
889 return emitAtomicCmpSwap(MI, BB, 4);
890 case Mips::ATOMIC_CMP_SWAP_I64:
891 return emitAtomicCmpSwap(MI, BB, 8);
892 case Mips::PseudoSDIV:
893 case Mips::PseudoUDIV:
894 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), false);
895 case Mips::PseudoDSDIV:
896 case Mips::PseudoDUDIV:
897 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), true);
901 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
902 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
904 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
905 unsigned Size, unsigned BinOpcode,
907 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
909 MachineFunction *MF = BB->getParent();
910 MachineRegisterInfo &RegInfo = MF->getRegInfo();
911 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
912 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
913 DebugLoc DL = MI->getDebugLoc();
914 unsigned LL, SC, AND, NOR, ZERO, BEQ;
929 ZERO = Mips::ZERO_64;
933 unsigned OldVal = MI->getOperand(0).getReg();
934 unsigned Ptr = MI->getOperand(1).getReg();
935 unsigned Incr = MI->getOperand(2).getReg();
937 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
938 unsigned AndRes = RegInfo.createVirtualRegister(RC);
939 unsigned Success = RegInfo.createVirtualRegister(RC);
941 // insert new blocks after the current block
942 const BasicBlock *LLVM_BB = BB->getBasicBlock();
943 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
944 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
945 MachineFunction::iterator It = BB;
947 MF->insert(It, loopMBB);
948 MF->insert(It, exitMBB);
950 // Transfer the remainder of BB and its successor edges to exitMBB.
951 exitMBB->splice(exitMBB->begin(), BB,
952 llvm::next(MachineBasicBlock::iterator(MI)),
954 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
958 // fallthrough --> loopMBB
959 BB->addSuccessor(loopMBB);
960 loopMBB->addSuccessor(loopMBB);
961 loopMBB->addSuccessor(exitMBB);
965 // <binop> storeval, oldval, incr
966 // sc success, storeval, 0(ptr)
967 // beq success, $0, loopMBB
969 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
971 // and andres, oldval, incr
972 // nor storeval, $0, andres
973 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
974 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
975 } else if (BinOpcode) {
976 // <binop> storeval, oldval, incr
977 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
981 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
982 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
984 MI->eraseFromParent(); // The instruction is gone now.
990 MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI,
991 MachineBasicBlock *BB,
992 unsigned Size, unsigned BinOpcode,
994 assert((Size == 1 || Size == 2) &&
995 "Unsupported size for EmitAtomicBinaryPartial.");
997 MachineFunction *MF = BB->getParent();
998 MachineRegisterInfo &RegInfo = MF->getRegInfo();
999 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1000 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1001 DebugLoc DL = MI->getDebugLoc();
1003 unsigned Dest = MI->getOperand(0).getReg();
1004 unsigned Ptr = MI->getOperand(1).getReg();
1005 unsigned Incr = MI->getOperand(2).getReg();
1007 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1008 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1009 unsigned Mask = RegInfo.createVirtualRegister(RC);
1010 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1011 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1012 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1013 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1014 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1015 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1016 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1017 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1018 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1019 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1020 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1021 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1022 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1023 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1024 unsigned Success = RegInfo.createVirtualRegister(RC);
1026 // insert new blocks after the current block
1027 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1028 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1029 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1030 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1031 MachineFunction::iterator It = BB;
1033 MF->insert(It, loopMBB);
1034 MF->insert(It, sinkMBB);
1035 MF->insert(It, exitMBB);
1037 // Transfer the remainder of BB and its successor edges to exitMBB.
1038 exitMBB->splice(exitMBB->begin(), BB,
1039 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1040 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1042 BB->addSuccessor(loopMBB);
1043 loopMBB->addSuccessor(loopMBB);
1044 loopMBB->addSuccessor(sinkMBB);
1045 sinkMBB->addSuccessor(exitMBB);
1048 // addiu masklsb2,$0,-4 # 0xfffffffc
1049 // and alignedaddr,ptr,masklsb2
1050 // andi ptrlsb2,ptr,3
1051 // sll shiftamt,ptrlsb2,3
1052 // ori maskupper,$0,255 # 0xff
1053 // sll mask,maskupper,shiftamt
1054 // nor mask2,$0,mask
1055 // sll incr2,incr,shiftamt
1057 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1058 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1059 .addReg(Mips::ZERO).addImm(-4);
1060 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1061 .addReg(Ptr).addReg(MaskLSB2);
1062 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1063 if (Subtarget->isLittle()) {
1064 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1066 unsigned Off = RegInfo.createVirtualRegister(RC);
1067 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1068 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1069 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1071 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1072 .addReg(Mips::ZERO).addImm(MaskImm);
1073 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1074 .addReg(MaskUpper).addReg(ShiftAmt);
1075 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1076 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1078 // atomic.load.binop
1080 // ll oldval,0(alignedaddr)
1081 // binop binopres,oldval,incr2
1082 // and newval,binopres,mask
1083 // and maskedoldval0,oldval,mask2
1084 // or storeval,maskedoldval0,newval
1085 // sc success,storeval,0(alignedaddr)
1086 // beq success,$0,loopMBB
1090 // ll oldval,0(alignedaddr)
1091 // and newval,incr2,mask
1092 // and maskedoldval0,oldval,mask2
1093 // or storeval,maskedoldval0,newval
1094 // sc success,storeval,0(alignedaddr)
1095 // beq success,$0,loopMBB
1098 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1100 // and andres, oldval, incr2
1101 // nor binopres, $0, andres
1102 // and newval, binopres, mask
1103 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1104 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1105 .addReg(Mips::ZERO).addReg(AndRes);
1106 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1107 } else if (BinOpcode) {
1108 // <binop> binopres, oldval, incr2
1109 // and newval, binopres, mask
1110 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1111 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1112 } else {// atomic.swap
1113 // and newval, incr2, mask
1114 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1117 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1118 .addReg(OldVal).addReg(Mask2);
1119 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1120 .addReg(MaskedOldVal0).addReg(NewVal);
1121 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1122 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1123 BuildMI(BB, DL, TII->get(Mips::BEQ))
1124 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1127 // and maskedoldval1,oldval,mask
1128 // srl srlres,maskedoldval1,shiftamt
1129 // sll sllres,srlres,24
1130 // sra dest,sllres,24
1132 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1134 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1135 .addReg(OldVal).addReg(Mask);
1136 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1137 .addReg(MaskedOldVal1).addReg(ShiftAmt);
1138 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1139 .addReg(SrlRes).addImm(ShiftImm);
1140 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1141 .addReg(SllRes).addImm(ShiftImm);
1143 MI->eraseFromParent(); // The instruction is gone now.
1149 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1150 MachineBasicBlock *BB,
1151 unsigned Size) const {
1152 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1154 MachineFunction *MF = BB->getParent();
1155 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1156 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1157 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1158 DebugLoc DL = MI->getDebugLoc();
1159 unsigned LL, SC, ZERO, BNE, BEQ;
1171 ZERO = Mips::ZERO_64;
1176 unsigned Dest = MI->getOperand(0).getReg();
1177 unsigned Ptr = MI->getOperand(1).getReg();
1178 unsigned OldVal = MI->getOperand(2).getReg();
1179 unsigned NewVal = MI->getOperand(3).getReg();
1181 unsigned Success = RegInfo.createVirtualRegister(RC);
1183 // insert new blocks after the current block
1184 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1185 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1186 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1187 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1188 MachineFunction::iterator It = BB;
1190 MF->insert(It, loop1MBB);
1191 MF->insert(It, loop2MBB);
1192 MF->insert(It, exitMBB);
1194 // Transfer the remainder of BB and its successor edges to exitMBB.
1195 exitMBB->splice(exitMBB->begin(), BB,
1196 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1197 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1201 // fallthrough --> loop1MBB
1202 BB->addSuccessor(loop1MBB);
1203 loop1MBB->addSuccessor(exitMBB);
1204 loop1MBB->addSuccessor(loop2MBB);
1205 loop2MBB->addSuccessor(loop1MBB);
1206 loop2MBB->addSuccessor(exitMBB);
1210 // bne dest, oldval, exitMBB
1212 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1213 BuildMI(BB, DL, TII->get(BNE))
1214 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1217 // sc success, newval, 0(ptr)
1218 // beq success, $0, loop1MBB
1220 BuildMI(BB, DL, TII->get(SC), Success)
1221 .addReg(NewVal).addReg(Ptr).addImm(0);
1222 BuildMI(BB, DL, TII->get(BEQ))
1223 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1225 MI->eraseFromParent(); // The instruction is gone now.
1231 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1232 MachineBasicBlock *BB,
1233 unsigned Size) const {
1234 assert((Size == 1 || Size == 2) &&
1235 "Unsupported size for EmitAtomicCmpSwapPartial.");
1237 MachineFunction *MF = BB->getParent();
1238 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1239 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1240 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1241 DebugLoc DL = MI->getDebugLoc();
1243 unsigned Dest = MI->getOperand(0).getReg();
1244 unsigned Ptr = MI->getOperand(1).getReg();
1245 unsigned CmpVal = MI->getOperand(2).getReg();
1246 unsigned NewVal = MI->getOperand(3).getReg();
1248 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1249 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1250 unsigned Mask = RegInfo.createVirtualRegister(RC);
1251 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1252 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1253 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1254 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1255 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1256 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1257 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1258 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1259 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1260 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1261 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1262 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1263 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1264 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1265 unsigned Success = RegInfo.createVirtualRegister(RC);
1267 // insert new blocks after the current block
1268 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1269 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1270 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1271 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1272 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1273 MachineFunction::iterator It = BB;
1275 MF->insert(It, loop1MBB);
1276 MF->insert(It, loop2MBB);
1277 MF->insert(It, sinkMBB);
1278 MF->insert(It, exitMBB);
1280 // Transfer the remainder of BB and its successor edges to exitMBB.
1281 exitMBB->splice(exitMBB->begin(), BB,
1282 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1283 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1285 BB->addSuccessor(loop1MBB);
1286 loop1MBB->addSuccessor(sinkMBB);
1287 loop1MBB->addSuccessor(loop2MBB);
1288 loop2MBB->addSuccessor(loop1MBB);
1289 loop2MBB->addSuccessor(sinkMBB);
1290 sinkMBB->addSuccessor(exitMBB);
1292 // FIXME: computation of newval2 can be moved to loop2MBB.
1294 // addiu masklsb2,$0,-4 # 0xfffffffc
1295 // and alignedaddr,ptr,masklsb2
1296 // andi ptrlsb2,ptr,3
1297 // sll shiftamt,ptrlsb2,3
1298 // ori maskupper,$0,255 # 0xff
1299 // sll mask,maskupper,shiftamt
1300 // nor mask2,$0,mask
1301 // andi maskedcmpval,cmpval,255
1302 // sll shiftedcmpval,maskedcmpval,shiftamt
1303 // andi maskednewval,newval,255
1304 // sll shiftednewval,maskednewval,shiftamt
1305 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1306 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1307 .addReg(Mips::ZERO).addImm(-4);
1308 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1309 .addReg(Ptr).addReg(MaskLSB2);
1310 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1311 if (Subtarget->isLittle()) {
1312 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1314 unsigned Off = RegInfo.createVirtualRegister(RC);
1315 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1316 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1317 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1319 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1320 .addReg(Mips::ZERO).addImm(MaskImm);
1321 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1322 .addReg(MaskUpper).addReg(ShiftAmt);
1323 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1324 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1325 .addReg(CmpVal).addImm(MaskImm);
1326 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1327 .addReg(MaskedCmpVal).addReg(ShiftAmt);
1328 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1329 .addReg(NewVal).addImm(MaskImm);
1330 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1331 .addReg(MaskedNewVal).addReg(ShiftAmt);
1334 // ll oldval,0(alginedaddr)
1335 // and maskedoldval0,oldval,mask
1336 // bne maskedoldval0,shiftedcmpval,sinkMBB
1338 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1339 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1340 .addReg(OldVal).addReg(Mask);
1341 BuildMI(BB, DL, TII->get(Mips::BNE))
1342 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1345 // and maskedoldval1,oldval,mask2
1346 // or storeval,maskedoldval1,shiftednewval
1347 // sc success,storeval,0(alignedaddr)
1348 // beq success,$0,loop1MBB
1350 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1351 .addReg(OldVal).addReg(Mask2);
1352 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1353 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1354 BuildMI(BB, DL, TII->get(Mips::SC), Success)
1355 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1356 BuildMI(BB, DL, TII->get(Mips::BEQ))
1357 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1360 // srl srlres,maskedoldval0,shiftamt
1361 // sll sllres,srlres,24
1362 // sra dest,sllres,24
1364 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1366 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1367 .addReg(MaskedOldVal0).addReg(ShiftAmt);
1368 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1369 .addReg(SrlRes).addImm(ShiftImm);
1370 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1371 .addReg(SllRes).addImm(ShiftImm);
1373 MI->eraseFromParent(); // The instruction is gone now.
1378 //===----------------------------------------------------------------------===//
1379 // Misc Lower Operation implementation
1380 //===----------------------------------------------------------------------===//
1381 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1382 SDValue Chain = Op.getOperand(0);
1383 SDValue Table = Op.getOperand(1);
1384 SDValue Index = Op.getOperand(2);
1386 EVT PTy = getPointerTy();
1387 unsigned EntrySize =
1388 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout());
1390 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1391 DAG.getConstant(EntrySize, PTy));
1392 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1394 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1395 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1396 MachinePointerInfo::getJumpTable(), MemVT, false, false,
1398 Chain = Addr.getValue(1);
1400 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) {
1401 // For PIC, the sequence is:
1402 // BRIND(load(Jumptable + index) + RelocBase)
1403 // RelocBase can be JumpTable, GOT or some sort of global base.
1404 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1405 getPICJumpTableRelocBase(Table, DAG));
1408 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1411 SDValue MipsTargetLowering::
1412 lowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1414 // The first operand is the chain, the second is the condition, the third is
1415 // the block to branch to if the condition is true.
1416 SDValue Chain = Op.getOperand(0);
1417 SDValue Dest = Op.getOperand(2);
1420 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1422 // Return if flag is not set by a floating point comparison.
1423 if (CondRes.getOpcode() != MipsISD::FPCmp)
1426 SDValue CCNode = CondRes.getOperand(2);
1428 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1429 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1430 SDValue BrCode = DAG.getConstant(Opc, MVT::i32);
1431 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
1432 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1433 FCC0, Dest, CondRes);
1436 SDValue MipsTargetLowering::
1437 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1439 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1441 // Return if flag is not set by a floating point comparison.
1442 if (Cond.getOpcode() != MipsISD::FPCmp)
1445 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1449 SDValue MipsTargetLowering::
1450 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1453 EVT Ty = Op.getOperand(0).getValueType();
1454 SDValue Cond = DAG.getNode(ISD::SETCC, DL,
1455 getSetCCResultType(*DAG.getContext(), Ty),
1456 Op.getOperand(0), Op.getOperand(1),
1459 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1463 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1464 SDValue Cond = createFPCmp(DAG, Op);
1466 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1467 "Floating point operand expected.");
1469 SDValue True = DAG.getConstant(1, MVT::i32);
1470 SDValue False = DAG.getConstant(0, MVT::i32);
1472 return createCMovFP(DAG, Cond, True, False, SDLoc(Op));
1475 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1476 SelectionDAG &DAG) const {
1477 // FIXME there isn't actually debug info here
1479 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1481 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1482 const MipsTargetObjectFile &TLOF =
1483 (const MipsTargetObjectFile&)getObjFileLowering();
1485 // %gp_rel relocation
1486 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1487 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0,
1489 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL,
1490 DAG.getVTList(MVT::i32), &GA, 1);
1491 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1492 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode);
1495 // %hi/%lo relocation
1496 return getAddrNonPIC(Op, DAG);
1499 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1500 return getAddrLocal(Op, DAG, HasMips64);
1503 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16,
1504 MipsII::MO_GOT_LO16);
1506 return getAddrGlobal(Op, DAG,
1507 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
1510 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1511 SelectionDAG &DAG) const {
1512 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1513 return getAddrNonPIC(Op, DAG);
1515 return getAddrLocal(Op, DAG, HasMips64);
1518 SDValue MipsTargetLowering::
1519 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1521 // If the relocation model is PIC, use the General Dynamic TLS Model or
1522 // Local Dynamic TLS model, otherwise use the Initial Exec or
1523 // Local Exec TLS Model.
1525 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1527 const GlobalValue *GV = GA->getGlobal();
1528 EVT PtrVT = getPointerTy();
1530 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1532 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1533 // General Dynamic and Local Dynamic TLS Model.
1534 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1537 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1538 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1539 getGlobalReg(DAG, PtrVT), TGA);
1540 unsigned PtrSize = PtrVT.getSizeInBits();
1541 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1543 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1547 Entry.Node = Argument;
1549 Args.push_back(Entry);
1551 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
1552 false, false, false, false, 0, CallingConv::C,
1553 /*IsTailCall=*/false, /*doesNotRet=*/false,
1554 /*isReturnValueUsed=*/true,
1555 TlsGetAddr, Args, DAG, DL);
1556 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1558 SDValue Ret = CallResult.first;
1560 if (model != TLSModel::LocalDynamic)
1563 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1564 MipsII::MO_DTPREL_HI);
1565 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1566 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1567 MipsII::MO_DTPREL_LO);
1568 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1569 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1570 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1574 if (model == TLSModel::InitialExec) {
1575 // Initial Exec TLS Model
1576 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1577 MipsII::MO_GOTTPREL);
1578 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1580 Offset = DAG.getLoad(PtrVT, DL,
1581 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1582 false, false, false, 0);
1584 // Local Exec TLS Model
1585 assert(model == TLSModel::LocalExec);
1586 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1587 MipsII::MO_TPREL_HI);
1588 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1589 MipsII::MO_TPREL_LO);
1590 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1591 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1592 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1595 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1596 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1599 SDValue MipsTargetLowering::
1600 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1602 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1603 return getAddrNonPIC(Op, DAG);
1605 return getAddrLocal(Op, DAG, HasMips64);
1608 SDValue MipsTargetLowering::
1609 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1611 // gp_rel relocation
1612 // FIXME: we should reference the constant pool using small data sections,
1613 // but the asm printer currently doesn't support this feature without
1614 // hacking it. This feature should come soon so we can uncomment the
1616 //if (IsInSmallSection(C->getType())) {
1617 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1618 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1619 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1621 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1622 return getAddrNonPIC(Op, DAG);
1624 return getAddrLocal(Op, DAG, HasMips64);
1627 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1628 MachineFunction &MF = DAG.getMachineFunction();
1629 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1632 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1635 // vastart just stores the address of the VarArgsFrameIndex slot into the
1636 // memory location argument.
1637 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1638 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1639 MachinePointerInfo(SV), false, false, 0);
1642 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1643 EVT TyX = Op.getOperand(0).getValueType();
1644 EVT TyY = Op.getOperand(1).getValueType();
1645 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1646 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1650 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1652 SDValue X = (TyX == MVT::f32) ?
1653 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1654 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1656 SDValue Y = (TyY == MVT::f32) ?
1657 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1658 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1662 // ext E, Y, 31, 1 ; extract bit31 of Y
1663 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1664 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1665 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1668 // srl SrlX, SllX, 1
1670 // sll SllY, SrlX, 31
1671 // or Or, SrlX, SllY
1672 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1673 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1674 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1675 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1676 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1679 if (TyX == MVT::f32)
1680 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1682 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1683 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1684 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1687 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1688 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1689 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1690 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1691 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1694 // Bitcast to integer nodes.
1695 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1696 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1699 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1700 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1701 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1702 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
1704 if (WidthX > WidthY)
1705 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1706 else if (WidthY > WidthX)
1707 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1709 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1710 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
1711 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1714 // (d)sll SllX, X, 1
1715 // (d)srl SrlX, SllX, 1
1716 // (d)srl SrlY, Y, width(Y)-1
1717 // (d)sll SllY, SrlX, width(Y)-1
1718 // or Or, SrlX, SllY
1719 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1720 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1721 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1722 DAG.getConstant(WidthY - 1, MVT::i32));
1724 if (WidthX > WidthY)
1725 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1726 else if (WidthY > WidthX)
1727 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1729 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1730 DAG.getConstant(WidthX - 1, MVT::i32));
1731 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
1732 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
1736 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1737 if (Subtarget->hasMips64())
1738 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
1740 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
1743 static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1744 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1747 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1749 SDValue X = (Op.getValueType() == MVT::f32) ?
1750 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1751 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1756 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
1757 DAG.getRegister(Mips::ZERO, MVT::i32),
1758 DAG.getConstant(31, MVT::i32), Const1, X);
1760 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1761 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1764 if (Op.getValueType() == MVT::f32)
1765 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
1767 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1768 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1769 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1772 static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1773 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1776 // Bitcast to integer node.
1777 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
1781 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
1782 DAG.getRegister(Mips::ZERO_64, MVT::i64),
1783 DAG.getConstant(63, MVT::i32), Const1, X);
1785 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
1786 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
1789 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
1793 MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const {
1794 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
1795 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2());
1797 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2());
1800 SDValue MipsTargetLowering::
1801 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1803 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1804 "Frame address can only be determined for current frame.");
1806 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1807 MFI->setFrameAddressIsTaken(true);
1808 EVT VT = Op.getValueType();
1810 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1811 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1815 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
1816 SelectionDAG &DAG) const {
1818 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1819 "Return address can be determined only for current frame.");
1821 MachineFunction &MF = DAG.getMachineFunction();
1822 MachineFrameInfo *MFI = MF.getFrameInfo();
1823 MVT VT = Op.getSimpleValueType();
1824 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
1825 MFI->setReturnAddressIsTaken(true);
1827 // Return RA, which contains the return address. Mark it an implicit live-in.
1828 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
1829 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
1832 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
1833 // generated from __builtin_eh_return (offset, handler)
1834 // The effect of this is to adjust the stack pointer by "offset"
1835 // and then branch to "handler".
1836 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
1838 MachineFunction &MF = DAG.getMachineFunction();
1839 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1841 MipsFI->setCallsEhReturn();
1842 SDValue Chain = Op.getOperand(0);
1843 SDValue Offset = Op.getOperand(1);
1844 SDValue Handler = Op.getOperand(2);
1846 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
1848 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
1849 // EH_RETURN nodes, so that instructions are emitted back-to-back.
1850 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1;
1851 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0;
1852 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
1853 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
1854 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
1855 DAG.getRegister(OffsetReg, Ty),
1856 DAG.getRegister(AddrReg, getPointerTy()),
1860 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
1861 SelectionDAG &DAG) const {
1862 // FIXME: Need pseudo-fence for 'singlethread' fences
1863 // FIXME: Set SType for weaker fences where supported/appropriate.
1866 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
1867 DAG.getConstant(SType, MVT::i32));
1870 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
1871 SelectionDAG &DAG) const {
1873 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1874 SDValue Shamt = Op.getOperand(2);
1877 // lo = (shl lo, shamt)
1878 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
1881 // hi = (shl lo, shamt[4:0])
1882 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1883 DAG.getConstant(-1, MVT::i32));
1884 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
1885 DAG.getConstant(1, MVT::i32));
1886 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
1888 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
1889 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1890 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
1891 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1892 DAG.getConstant(0x20, MVT::i32));
1893 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1894 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
1895 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
1897 SDValue Ops[2] = {Lo, Hi};
1898 return DAG.getMergeValues(Ops, 2, DL);
1901 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
1904 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1905 SDValue Shamt = Op.getOperand(2);
1908 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
1910 // hi = (sra hi, shamt)
1912 // hi = (srl hi, shamt)
1915 // lo = (sra hi, shamt[4:0])
1916 // hi = (sra hi, 31)
1918 // lo = (srl hi, shamt[4:0])
1920 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1921 DAG.getConstant(-1, MVT::i32));
1922 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
1923 DAG.getConstant(1, MVT::i32));
1924 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
1925 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
1926 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1927 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1929 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1930 DAG.getConstant(0x20, MVT::i32));
1931 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
1932 DAG.getConstant(31, MVT::i32));
1933 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
1934 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1935 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
1938 SDValue Ops[2] = {Lo, Hi};
1939 return DAG.getMergeValues(Ops, 2, DL);
1942 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
1943 SDValue Chain, SDValue Src, unsigned Offset) {
1944 SDValue Ptr = LD->getBasePtr();
1945 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
1946 EVT BasePtrVT = Ptr.getValueType();
1948 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1951 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
1952 DAG.getConstant(Offset, BasePtrVT));
1954 SDValue Ops[] = { Chain, Ptr, Src };
1955 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
1956 LD->getMemOperand());
1959 // Expand an unaligned 32 or 64-bit integer load node.
1960 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1961 LoadSDNode *LD = cast<LoadSDNode>(Op);
1962 EVT MemVT = LD->getMemoryVT();
1964 // Return if load is aligned or if MemVT is neither i32 nor i64.
1965 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
1966 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
1969 bool IsLittle = Subtarget->isLittle();
1970 EVT VT = Op.getValueType();
1971 ISD::LoadExtType ExtType = LD->getExtensionType();
1972 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
1974 assert((VT == MVT::i32) || (VT == MVT::i64));
1977 // (set dst, (i64 (load baseptr)))
1979 // (set tmp, (ldl (add baseptr, 7), undef))
1980 // (set dst, (ldr baseptr, tmp))
1981 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
1982 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
1984 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
1988 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
1990 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
1994 // (set dst, (i32 (load baseptr))) or
1995 // (set dst, (i64 (sextload baseptr))) or
1996 // (set dst, (i64 (extload baseptr)))
1998 // (set tmp, (lwl (add baseptr, 3), undef))
1999 // (set dst, (lwr baseptr, tmp))
2000 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2001 (ExtType == ISD::EXTLOAD))
2004 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2007 // (set dst, (i64 (zextload baseptr)))
2009 // (set tmp0, (lwl (add baseptr, 3), undef))
2010 // (set tmp1, (lwr baseptr, tmp0))
2011 // (set tmp2, (shl tmp1, 32))
2012 // (set dst, (srl tmp2, 32))
2014 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2015 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2016 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2017 SDValue Ops[] = { SRL, LWR.getValue(1) };
2018 return DAG.getMergeValues(Ops, 2, DL);
2021 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2022 SDValue Chain, unsigned Offset) {
2023 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2024 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2026 SDVTList VTList = DAG.getVTList(MVT::Other);
2029 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2030 DAG.getConstant(Offset, BasePtrVT));
2032 SDValue Ops[] = { Chain, Value, Ptr };
2033 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2034 SD->getMemOperand());
2037 // Expand an unaligned 32 or 64-bit integer store node.
2038 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2040 SDValue Value = SD->getValue(), Chain = SD->getChain();
2041 EVT VT = Value.getValueType();
2044 // (store val, baseptr) or
2045 // (truncstore val, baseptr)
2047 // (swl val, (add baseptr, 3))
2048 // (swr val, baseptr)
2049 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2050 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2052 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2055 assert(VT == MVT::i64);
2058 // (store val, baseptr)
2060 // (sdl val, (add baseptr, 7))
2061 // (sdr val, baseptr)
2062 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2063 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2066 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
2067 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) {
2068 SDValue Val = SD->getValue();
2070 if (Val.getOpcode() != ISD::FP_TO_SINT)
2073 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2074 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2077 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2078 SD->getPointerInfo(), SD->isVolatile(),
2079 SD->isNonTemporal(), SD->getAlignment());
2082 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2083 StoreSDNode *SD = cast<StoreSDNode>(Op);
2084 EVT MemVT = SD->getMemoryVT();
2086 // Lower unaligned integer stores.
2087 if ((SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2088 ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2089 return lowerUnalignedIntStore(SD, DAG, Subtarget->isLittle());
2091 return lowerFP_TO_SINT_STORE(SD, DAG);
2094 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2095 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2096 || cast<ConstantSDNode>
2097 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2098 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2102 // (add (frameaddr 0), (frame_to_args_offset))
2103 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2104 // (add FrameObject, 0)
2105 // where FrameObject is a fixed StackObject with offset 0 which points to
2106 // the old stack pointer.
2107 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2108 EVT ValTy = Op->getValueType(0);
2109 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2110 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2111 return DAG.getNode(ISD::ADD, SDLoc(Op), ValTy, InArgsAddr,
2112 DAG.getConstant(0, ValTy));
2115 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2116 SelectionDAG &DAG) const {
2117 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2118 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2120 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2123 //===----------------------------------------------------------------------===//
2124 // Calling Convention Implementation
2125 //===----------------------------------------------------------------------===//
2127 //===----------------------------------------------------------------------===//
2128 // TODO: Implement a generic logic using tblgen that can support this.
2129 // Mips O32 ABI rules:
2131 // i32 - Passed in A0, A1, A2, A3 and stack
2132 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2133 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2134 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2135 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2136 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2139 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2140 //===----------------------------------------------------------------------===//
2142 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2143 MVT LocVT, CCValAssign::LocInfo LocInfo,
2144 ISD::ArgFlagsTy ArgFlags, CCState &State,
2145 const uint16_t *F64Regs) {
2147 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2149 static const uint16_t IntRegs[] = {
2150 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2152 static const uint16_t F32Regs[] = {
2153 Mips::F12, Mips::F14
2156 // Do not process byval args here.
2157 if (ArgFlags.isByVal())
2160 // Promote i8 and i16
2161 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2163 if (ArgFlags.isSExt())
2164 LocInfo = CCValAssign::SExt;
2165 else if (ArgFlags.isZExt())
2166 LocInfo = CCValAssign::ZExt;
2168 LocInfo = CCValAssign::AExt;
2173 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2174 // is true: function is vararg, argument is 3rd or higher, there is previous
2175 // argument which is not f32 or f64.
2176 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2177 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2178 unsigned OrigAlign = ArgFlags.getOrigAlign();
2179 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2181 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2182 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2183 // If this is the first part of an i64 arg,
2184 // the allocated register must be either A0 or A2.
2185 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2186 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2188 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2189 // Allocate int register and shadow next int register. If first
2190 // available register is Mips::A1 or Mips::A3, shadow it too.
2191 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2192 if (Reg == Mips::A1 || Reg == Mips::A3)
2193 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2194 State.AllocateReg(IntRegs, IntRegsSize);
2196 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2197 // we are guaranteed to find an available float register
2198 if (ValVT == MVT::f32) {
2199 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2200 // Shadow int register
2201 State.AllocateReg(IntRegs, IntRegsSize);
2203 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2204 // Shadow int registers
2205 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2206 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2207 State.AllocateReg(IntRegs, IntRegsSize);
2208 State.AllocateReg(IntRegs, IntRegsSize);
2211 llvm_unreachable("Cannot handle this ValVT.");
2214 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2216 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2218 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2223 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2224 MVT LocVT, CCValAssign::LocInfo LocInfo,
2225 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2226 static const uint16_t F64Regs[] = { Mips::D6, Mips::D7 };
2228 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2231 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2232 MVT LocVT, CCValAssign::LocInfo LocInfo,
2233 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2234 static const uint16_t F64Regs[] = { Mips::D12_64, Mips::D12_64 };
2236 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2239 #include "MipsGenCallingConv.inc"
2241 //===----------------------------------------------------------------------===//
2242 // Call Calling Convention Implementation
2243 //===----------------------------------------------------------------------===//
2245 // Return next O32 integer argument register.
2246 static unsigned getNextIntArgReg(unsigned Reg) {
2247 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2248 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2252 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2253 SDValue Chain, SDValue Arg, SDLoc DL,
2254 bool IsTailCall, SelectionDAG &DAG) const {
2256 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2257 DAG.getIntPtrConstant(Offset));
2258 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2262 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2263 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2264 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2265 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2266 /*isVolatile=*/ true, false, 0);
2269 void MipsTargetLowering::
2270 getOpndList(SmallVectorImpl<SDValue> &Ops,
2271 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2272 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2273 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
2274 // Insert node "GP copy globalreg" before call to function.
2276 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2277 // in PIC mode) allow symbols to be resolved via lazy binding.
2278 // The lazy binding stub requires GP to point to the GOT.
2279 if (IsPICCall && !InternalLinkage) {
2280 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
2281 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2282 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2285 // Build a sequence of copy-to-reg nodes chained together with token
2286 // chain and flag operands which copy the outgoing args into registers.
2287 // The InFlag in necessary since all emitted instructions must be
2291 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2292 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2293 RegsToPass[i].second, InFlag);
2294 InFlag = Chain.getValue(1);
2297 // Add argument registers to the end of the list so that they are
2298 // known live into the call.
2299 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2300 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2301 RegsToPass[i].second.getValueType()));
2303 // Add a register mask operand representing the call-preserved registers.
2304 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2305 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv);
2306 assert(Mask && "Missing call preserved mask for calling convention");
2307 if (Subtarget->inMips16HardFloat()) {
2308 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2309 llvm::StringRef Sym = G->getGlobal()->getName();
2310 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2311 if (F->hasFnAttribute("__Mips16RetHelper")) {
2312 Mask = MipsRegisterInfo::getMips16RetHelperMask();
2316 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2318 if (InFlag.getNode())
2319 Ops.push_back(InFlag);
2322 /// LowerCall - functions arguments are copied from virtual regs to
2323 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2325 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2326 SmallVectorImpl<SDValue> &InVals) const {
2327 SelectionDAG &DAG = CLI.DAG;
2329 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2330 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2331 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2332 SDValue Chain = CLI.Chain;
2333 SDValue Callee = CLI.Callee;
2334 bool &IsTailCall = CLI.IsTailCall;
2335 CallingConv::ID CallConv = CLI.CallConv;
2336 bool IsVarArg = CLI.IsVarArg;
2338 MachineFunction &MF = DAG.getMachineFunction();
2339 MachineFrameInfo *MFI = MF.getFrameInfo();
2340 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2341 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2343 // Analyze operands of the call, assigning locations to each operand.
2344 SmallVector<CCValAssign, 16> ArgLocs;
2345 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2346 getTargetMachine(), ArgLocs, *DAG.getContext());
2347 MipsCC::SpecialCallingConvType SpecialCallingConv =
2348 getSpecialCallingConv(Callee);
2349 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo,
2350 SpecialCallingConv);
2352 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg,
2353 Subtarget->mipsSEUsesSoftFloat(),
2354 Callee.getNode(), CLI.Args);
2356 // Get a count of how many bytes are to be pushed on the stack.
2357 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2359 // Check if it's really possible to do a tail call.
2362 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2363 *MF.getInfo<MipsFunctionInfo>());
2368 // Chain is the output chain of the last Load/Store or CopyToReg node.
2369 // ByValChain is the output chain of the last Memcpy node created for copying
2370 // byval arguments to the stack.
2371 unsigned StackAlignment = TFL->getStackAlignment();
2372 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2373 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2376 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL);
2378 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL,
2379 IsN64 ? Mips::SP_64 : Mips::SP,
2382 // With EABI is it possible to have 16 args on registers.
2383 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2384 SmallVector<SDValue, 8> MemOpChains;
2385 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2387 // Walk the register/memloc assignments, inserting copies/loads.
2388 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2389 SDValue Arg = OutVals[i];
2390 CCValAssign &VA = ArgLocs[i];
2391 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2392 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2395 if (Flags.isByVal()) {
2396 assert(Flags.getByValSize() &&
2397 "ByVal args of size 0 should have been ignored by front-end.");
2398 assert(ByValArg != MipsCCInfo.byval_end());
2399 assert(!IsTailCall &&
2400 "Do not tail-call optimize if there is a byval argument.");
2401 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2402 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2407 // Promote the value if needed.
2408 switch (VA.getLocInfo()) {
2409 default: llvm_unreachable("Unknown loc info!");
2410 case CCValAssign::Full:
2411 if (VA.isRegLoc()) {
2412 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2413 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2414 (ValVT == MVT::i64 && LocVT == MVT::f64))
2415 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2416 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2417 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2418 Arg, DAG.getConstant(0, MVT::i32));
2419 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2420 Arg, DAG.getConstant(1, MVT::i32));
2421 if (!Subtarget->isLittle())
2423 unsigned LocRegLo = VA.getLocReg();
2424 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2425 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2426 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2431 case CCValAssign::SExt:
2432 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2434 case CCValAssign::ZExt:
2435 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2437 case CCValAssign::AExt:
2438 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2442 // Arguments that can be passed on register must be kept at
2443 // RegsToPass vector
2444 if (VA.isRegLoc()) {
2445 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2449 // Register can't get to this point...
2450 assert(VA.isMemLoc());
2452 // emit ISD::STORE whichs stores the
2453 // parameter value to a stack Location
2454 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2455 Chain, Arg, DL, IsTailCall, DAG));
2458 // Transform all store nodes into one single node because all store
2459 // nodes are independent of each other.
2460 if (!MemOpChains.empty())
2461 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2462 &MemOpChains[0], MemOpChains.size());
2464 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2465 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2466 // node so that legalize doesn't hack it.
2467 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2468 bool GlobalOrExternal = false, InternalLinkage = false;
2471 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2473 InternalLinkage = G->getGlobal()->hasInternalLinkage();
2475 if (InternalLinkage)
2476 Callee = getAddrLocal(Callee, DAG, HasMips64);
2478 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2479 MipsII::MO_CALL_LO16);
2481 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2483 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0,
2484 MipsII::MO_NO_FLAG);
2485 GlobalOrExternal = true;
2487 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2488 if (!IsN64 && !IsPIC) // !N64 && static
2489 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2490 MipsII::MO_NO_FLAG);
2492 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2493 MipsII::MO_CALL_LO16);
2495 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2497 GlobalOrExternal = true;
2500 SmallVector<SDValue, 8> Ops(1, Chain);
2501 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2503 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2504 CLI, Callee, Chain);
2507 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size());
2509 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size());
2510 SDValue InFlag = Chain.getValue(1);
2512 // Create the CALLSEQ_END node.
2513 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2514 DAG.getIntPtrConstant(0, true), InFlag, DL);
2515 InFlag = Chain.getValue(1);
2517 // Handle result values, copying them out of physregs into vregs that we
2519 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg,
2520 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy);
2523 /// LowerCallResult - Lower the result values of a call into the
2524 /// appropriate copies out of appropriate physical registers.
2526 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2527 CallingConv::ID CallConv, bool IsVarArg,
2528 const SmallVectorImpl<ISD::InputArg> &Ins,
2529 SDLoc DL, SelectionDAG &DAG,
2530 SmallVectorImpl<SDValue> &InVals,
2531 const SDNode *CallNode,
2532 const Type *RetTy) const {
2533 // Assign locations to each value returned by this call.
2534 SmallVector<CCValAssign, 16> RVLocs;
2535 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2536 getTargetMachine(), RVLocs, *DAG.getContext());
2537 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo);
2539 MipsCCInfo.analyzeCallResult(Ins, Subtarget->mipsSEUsesSoftFloat(),
2542 // Copy all of the result registers out of their specified physreg.
2543 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2544 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2545 RVLocs[i].getLocVT(), InFlag);
2546 Chain = Val.getValue(1);
2547 InFlag = Val.getValue(2);
2549 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2550 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val);
2552 InVals.push_back(Val);
2558 //===----------------------------------------------------------------------===//
2559 // Formal Arguments Calling Convention Implementation
2560 //===----------------------------------------------------------------------===//
2561 /// LowerFormalArguments - transform physical registers into virtual registers
2562 /// and generate load operations for arguments places on the stack.
2564 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2565 CallingConv::ID CallConv,
2567 const SmallVectorImpl<ISD::InputArg> &Ins,
2568 SDLoc DL, SelectionDAG &DAG,
2569 SmallVectorImpl<SDValue> &InVals)
2571 MachineFunction &MF = DAG.getMachineFunction();
2572 MachineFrameInfo *MFI = MF.getFrameInfo();
2573 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2575 MipsFI->setVarArgsFrameIndex(0);
2577 // Used with vargs to acumulate store chains.
2578 std::vector<SDValue> OutChains;
2580 // Assign locations to all of the incoming arguments.
2581 SmallVector<CCValAssign, 16> ArgLocs;
2582 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2583 getTargetMachine(), ArgLocs, *DAG.getContext());
2584 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo);
2585 Function::const_arg_iterator FuncArg =
2586 DAG.getMachineFunction().getFunction()->arg_begin();
2587 bool UseSoftFloat = Subtarget->mipsSEUsesSoftFloat();
2589 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg);
2590 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2591 MipsCCInfo.hasByValArg());
2593 unsigned CurArgIdx = 0;
2594 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2596 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2597 CCValAssign &VA = ArgLocs[i];
2598 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
2599 CurArgIdx = Ins[i].OrigArgIndex;
2600 EVT ValVT = VA.getValVT();
2601 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2602 bool IsRegLoc = VA.isRegLoc();
2604 if (Flags.isByVal()) {
2605 assert(Flags.getByValSize() &&
2606 "ByVal args of size 0 should have been ignored by front-end.");
2607 assert(ByValArg != MipsCCInfo.byval_end());
2608 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2609 MipsCCInfo, *ByValArg);
2614 // Arguments stored on registers
2616 EVT RegVT = VA.getLocVT();
2617 unsigned ArgReg = VA.getLocReg();
2618 const TargetRegisterClass *RC;
2620 if (RegVT == MVT::i32)
2621 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass :
2622 &Mips::GPR32RegClass;
2623 else if (RegVT == MVT::i64)
2624 RC = &Mips::GPR64RegClass;
2625 else if (RegVT == MVT::f32)
2626 RC = &Mips::FGR32RegClass;
2627 else if (RegVT == MVT::f64)
2628 RC = Subtarget->isFP64bit() ? &Mips::FGR64RegClass :
2629 &Mips::AFGR64RegClass;
2631 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2633 // Transform the arguments stored on
2634 // physical registers into virtual ones
2635 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2636 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2638 // If this is an 8 or 16-bit value, it has been passed promoted
2639 // to 32 bits. Insert an assert[sz]ext to capture this, then
2640 // truncate to the right size.
2641 if (VA.getLocInfo() != CCValAssign::Full) {
2642 unsigned Opcode = 0;
2643 if (VA.getLocInfo() == CCValAssign::SExt)
2644 Opcode = ISD::AssertSext;
2645 else if (VA.getLocInfo() == CCValAssign::ZExt)
2646 Opcode = ISD::AssertZext;
2648 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue,
2649 DAG.getValueType(ValVT));
2650 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue);
2653 // Handle floating point arguments passed in integer registers and
2654 // long double arguments passed in floating point registers.
2655 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2656 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
2657 (RegVT == MVT::f64 && ValVT == MVT::i64))
2658 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
2659 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2660 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
2661 getNextIntArgReg(ArgReg), RC);
2662 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
2663 if (!Subtarget->isLittle())
2664 std::swap(ArgValue, ArgValue2);
2665 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
2666 ArgValue, ArgValue2);
2669 InVals.push_back(ArgValue);
2670 } else { // VA.isRegLoc()
2673 assert(VA.isMemLoc());
2675 // The stack pointer offset is relative to the caller stack frame.
2676 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2677 VA.getLocMemOffset(), true);
2679 // Create load nodes to retrieve arguments from the stack
2680 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2681 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN,
2682 MachinePointerInfo::getFixedStack(FI),
2683 false, false, false, 0));
2687 // The mips ABIs for returning structs by value requires that we copy
2688 // the sret argument into $v0 for the return. Save the argument into
2689 // a virtual register so that we can access it from the return points.
2690 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2691 unsigned Reg = MipsFI->getSRetReturnReg();
2693 Reg = MF.getRegInfo().
2694 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
2695 MipsFI->setSRetReturnReg(Reg);
2697 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]);
2698 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
2702 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG);
2704 // All stores are grouped in one node to allow the matching between
2705 // the size of Ins and InVals. This only happens when on varg functions
2706 if (!OutChains.empty()) {
2707 OutChains.push_back(Chain);
2708 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2709 &OutChains[0], OutChains.size());
2715 //===----------------------------------------------------------------------===//
2716 // Return Value Calling Convention Implementation
2717 //===----------------------------------------------------------------------===//
2720 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2721 MachineFunction &MF, bool IsVarArg,
2722 const SmallVectorImpl<ISD::OutputArg> &Outs,
2723 LLVMContext &Context) const {
2724 SmallVector<CCValAssign, 16> RVLocs;
2725 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(),
2727 return CCInfo.CheckReturn(Outs, RetCC_Mips);
2731 MipsTargetLowering::LowerReturn(SDValue Chain,
2732 CallingConv::ID CallConv, bool IsVarArg,
2733 const SmallVectorImpl<ISD::OutputArg> &Outs,
2734 const SmallVectorImpl<SDValue> &OutVals,
2735 SDLoc DL, SelectionDAG &DAG) const {
2736 // CCValAssign - represent the assignment of
2737 // the return value to a location
2738 SmallVector<CCValAssign, 16> RVLocs;
2739 MachineFunction &MF = DAG.getMachineFunction();
2741 // CCState - Info about the registers and stack slot.
2742 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs,
2744 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo);
2746 // Analyze return values.
2747 MipsCCInfo.analyzeReturn(Outs, Subtarget->mipsSEUsesSoftFloat(),
2748 MF.getFunction()->getReturnType());
2751 SmallVector<SDValue, 4> RetOps(1, Chain);
2753 // Copy the result values into the output registers.
2754 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2755 SDValue Val = OutVals[i];
2756 CCValAssign &VA = RVLocs[i];
2757 assert(VA.isRegLoc() && "Can only return in registers!");
2759 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2760 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val);
2762 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
2764 // Guarantee that all emitted copies are stuck together with flags.
2765 Flag = Chain.getValue(1);
2766 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2769 // The mips ABIs for returning structs by value requires that we copy
2770 // the sret argument into $v0 for the return. We saved the argument into
2771 // a virtual register in the entry block, so now we copy the value out
2773 if (MF.getFunction()->hasStructRetAttr()) {
2774 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2775 unsigned Reg = MipsFI->getSRetReturnReg();
2778 llvm_unreachable("sret virtual register not created in the entry block");
2779 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
2780 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
2782 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
2783 Flag = Chain.getValue(1);
2784 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
2787 RetOps[0] = Chain; // Update chain.
2789 // Add the flag if we have it.
2791 RetOps.push_back(Flag);
2793 // Return on Mips is always a "jr $ra"
2794 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size());
2797 //===----------------------------------------------------------------------===//
2798 // Mips Inline Assembly Support
2799 //===----------------------------------------------------------------------===//
2801 /// getConstraintType - Given a constraint letter, return the type of
2802 /// constraint it is for this target.
2803 MipsTargetLowering::ConstraintType MipsTargetLowering::
2804 getConstraintType(const std::string &Constraint) const
2806 // Mips specific constrainy
2807 // GCC config/mips/constraints.md
2809 // 'd' : An address register. Equivalent to r
2810 // unless generating MIPS16 code.
2811 // 'y' : Equivalent to r; retained for
2812 // backwards compatibility.
2813 // 'c' : A register suitable for use in an indirect
2814 // jump. This will always be $25 for -mabicalls.
2815 // 'l' : The lo register. 1 word storage.
2816 // 'x' : The hilo register pair. Double word storage.
2817 if (Constraint.size() == 1) {
2818 switch (Constraint[0]) {
2826 return C_RegisterClass;
2831 return TargetLowering::getConstraintType(Constraint);
2834 /// Examine constraint type and operand type and determine a weight value.
2835 /// This object must already have been set up with the operand type
2836 /// and the current alternative constraint selected.
2837 TargetLowering::ConstraintWeight
2838 MipsTargetLowering::getSingleConstraintMatchWeight(
2839 AsmOperandInfo &info, const char *constraint) const {
2840 ConstraintWeight weight = CW_Invalid;
2841 Value *CallOperandVal = info.CallOperandVal;
2842 // If we don't have a value, we can't do a match,
2843 // but allow it at the lowest weight.
2844 if (CallOperandVal == NULL)
2846 Type *type = CallOperandVal->getType();
2847 // Look at the constraint type.
2848 switch (*constraint) {
2850 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2854 if (type->isIntegerTy())
2855 weight = CW_Register;
2858 if (type->isFloatTy())
2859 weight = CW_Register;
2861 case 'c': // $25 for indirect jumps
2862 case 'l': // lo register
2863 case 'x': // hilo register pair
2864 if (type->isIntegerTy())
2865 weight = CW_SpecificReg;
2867 case 'I': // signed 16 bit immediate
2868 case 'J': // integer zero
2869 case 'K': // unsigned 16 bit immediate
2870 case 'L': // signed 32 bit immediate where lower 16 bits are 0
2871 case 'N': // immediate in the range of -65535 to -1 (inclusive)
2872 case 'O': // signed 15 bit immediate (+- 16383)
2873 case 'P': // immediate in the range of 65535 to 1 (inclusive)
2874 if (isa<ConstantInt>(CallOperandVal))
2875 weight = CW_Constant;
2884 /// This is a helper function to parse a physical register string and split it
2885 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
2886 /// that is returned indicates whether parsing was successful. The second flag
2887 /// is true if the numeric part exists.
2888 static std::pair<bool, bool>
2889 parsePhysicalReg(const StringRef &C, std::string &Prefix,
2890 unsigned long long &Reg) {
2891 if (C.front() != '{' || C.back() != '}')
2892 return std::make_pair(false, false);
2894 // Search for the first numeric character.
2895 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
2896 I = std::find_if(B, E, std::ptr_fun(isdigit));
2898 Prefix.assign(B, I - B);
2900 // The second flag is set to false if no numeric characters were found.
2902 return std::make_pair(true, false);
2904 // Parse the numeric characters.
2905 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
2909 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
2910 parseRegForInlineAsmConstraint(const StringRef &C, MVT VT) const {
2911 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2912 const TargetRegisterClass *RC;
2914 unsigned long long Reg;
2916 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
2919 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0);
2921 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
2922 // No numeric characters follow "hi" or "lo".
2924 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0);
2926 RC = TRI->getRegClass(Prefix == "hi" ?
2927 Mips::HI32RegClassID : Mips::LO32RegClassID);
2928 return std::make_pair(*(RC->begin()), RC);
2932 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0);
2934 if (Prefix == "$f") { // Parse $f0-$f31.
2935 // If the size of FP registers is 64-bit or Reg is an even number, select
2936 // the 64-bit register class. Otherwise, select the 32-bit register class.
2937 if (VT == MVT::Other)
2938 VT = (Subtarget->isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
2940 RC= getRegClassFor(VT);
2942 if (RC == &Mips::AFGR64RegClass) {
2943 assert(Reg % 2 == 0);
2946 } else if (Prefix == "$fcc") { // Parse $fcc0-$fcc7.
2947 RC = TRI->getRegClass(Mips::FCCRegClassID);
2948 } else { // Parse $0-$31.
2949 assert(Prefix == "$");
2950 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
2953 assert(Reg < RC->getNumRegs());
2954 return std::make_pair(*(RC->begin() + Reg), RC);
2957 /// Given a register class constraint, like 'r', if this corresponds directly
2958 /// to an LLVM register class, return a register of 0 and the register class
2960 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2961 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const
2963 if (Constraint.size() == 1) {
2964 switch (Constraint[0]) {
2965 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2966 case 'y': // Same as 'r'. Exists for compatibility.
2968 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
2969 if (Subtarget->inMips16Mode())
2970 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
2971 return std::make_pair(0U, &Mips::GPR32RegClass);
2973 if (VT == MVT::i64 && !HasMips64)
2974 return std::make_pair(0U, &Mips::GPR32RegClass);
2975 if (VT == MVT::i64 && HasMips64)
2976 return std::make_pair(0U, &Mips::GPR64RegClass);
2977 // This will generate an error message
2978 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2981 return std::make_pair(0U, &Mips::FGR32RegClass);
2982 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
2983 if (Subtarget->isFP64bit())
2984 return std::make_pair(0U, &Mips::FGR64RegClass);
2985 return std::make_pair(0U, &Mips::AFGR64RegClass);
2988 case 'c': // register suitable for indirect jump
2990 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
2991 assert(VT == MVT::i64 && "Unexpected type.");
2992 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
2993 case 'l': // register suitable for indirect jump
2995 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
2996 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
2997 case 'x': // register suitable for indirect jump
2998 // Fixme: Not triggering the use of both hi and low
2999 // This will generate an error message
3000 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3004 std::pair<unsigned, const TargetRegisterClass *> R;
3005 R = parseRegForInlineAsmConstraint(Constraint, VT);
3010 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3013 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3014 /// vector. If it is invalid, don't add anything to Ops.
3015 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3016 std::string &Constraint,
3017 std::vector<SDValue>&Ops,
3018 SelectionDAG &DAG) const {
3019 SDValue Result(0, 0);
3021 // Only support length 1 constraints for now.
3022 if (Constraint.length() > 1) return;
3024 char ConstraintLetter = Constraint[0];
3025 switch (ConstraintLetter) {
3026 default: break; // This will fall through to the generic implementation
3027 case 'I': // Signed 16 bit constant
3028 // If this fails, the parent routine will give an error
3029 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3030 EVT Type = Op.getValueType();
3031 int64_t Val = C->getSExtValue();
3032 if (isInt<16>(Val)) {
3033 Result = DAG.getTargetConstant(Val, Type);
3038 case 'J': // integer zero
3039 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3040 EVT Type = Op.getValueType();
3041 int64_t Val = C->getZExtValue();
3043 Result = DAG.getTargetConstant(0, Type);
3048 case 'K': // unsigned 16 bit immediate
3049 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3050 EVT Type = Op.getValueType();
3051 uint64_t Val = (uint64_t)C->getZExtValue();
3052 if (isUInt<16>(Val)) {
3053 Result = DAG.getTargetConstant(Val, Type);
3058 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3059 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3060 EVT Type = Op.getValueType();
3061 int64_t Val = C->getSExtValue();
3062 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3063 Result = DAG.getTargetConstant(Val, Type);
3068 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3069 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3070 EVT Type = Op.getValueType();
3071 int64_t Val = C->getSExtValue();
3072 if ((Val >= -65535) && (Val <= -1)) {
3073 Result = DAG.getTargetConstant(Val, Type);
3078 case 'O': // signed 15 bit immediate
3079 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3080 EVT Type = Op.getValueType();
3081 int64_t Val = C->getSExtValue();
3082 if ((isInt<15>(Val))) {
3083 Result = DAG.getTargetConstant(Val, Type);
3088 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3089 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3090 EVT Type = Op.getValueType();
3091 int64_t Val = C->getSExtValue();
3092 if ((Val <= 65535) && (Val >= 1)) {
3093 Result = DAG.getTargetConstant(Val, Type);
3100 if (Result.getNode()) {
3101 Ops.push_back(Result);
3105 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3109 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
3110 // No global is ever allowed as a base.
3115 case 0: // "r+i" or just "i", depending on HasBaseReg.
3118 if (!AM.HasBaseReg) // allow "r+i".
3120 return false; // disallow "r+r" or "r+r+i".
3129 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3130 // The Mips target isn't yet aware of offsets.
3134 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3136 bool IsMemset, bool ZeroMemset,
3138 MachineFunction &MF) const {
3139 if (Subtarget->hasMips64())
3145 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3146 if (VT != MVT::f32 && VT != MVT::f64)
3148 if (Imm.isNegZero())
3150 return Imm.isZero();
3153 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3155 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3157 return TargetLowering::getJumpTableEncoding();
3160 /// This function returns true if CallSym is a long double emulation routine.
3161 static bool isF128SoftLibCall(const char *CallSym) {
3162 const char *const LibCalls[] =
3163 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2",
3164 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi",
3165 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf",
3166 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2",
3167 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3",
3168 "__trunctfdf2", "__trunctfsf2", "__unordtf2",
3169 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl",
3170 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl",
3173 const char * const *End = LibCalls + array_lengthof(LibCalls);
3175 // Check that LibCalls is sorted alphabetically.
3176 MipsTargetLowering::LTStr Comp;
3179 for (const char * const *I = LibCalls; I < End - 1; ++I)
3180 assert(Comp(*I, *(I + 1)));
3183 return std::binary_search(LibCalls, End, CallSym, Comp);
3186 /// This function returns true if Ty is fp128 or i128 which was originally a
3188 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) {
3189 if (Ty->isFP128Ty())
3192 const ExternalSymbolSDNode *ES =
3193 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
3195 // If the Ty is i128 and the function being called is a long double emulation
3196 // routine, then the original type is f128.
3197 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
3200 MipsTargetLowering::MipsCC::SpecialCallingConvType
3201 MipsTargetLowering::getSpecialCallingConv(SDValue Callee) const {
3202 MipsCC::SpecialCallingConvType SpecialCallingConv =
3203 MipsCC::NoSpecialCallingConv;;
3204 if (Subtarget->inMips16HardFloat()) {
3205 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3206 llvm::StringRef Sym = G->getGlobal()->getName();
3207 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
3208 if (F->hasFnAttribute("__Mips16RetHelper")) {
3209 SpecialCallingConv = MipsCC::Mips16RetHelperConv;
3213 return SpecialCallingConv;
3216 MipsTargetLowering::MipsCC::MipsCC(
3217 CallingConv::ID CC, bool IsO32_, bool IsFP64_, CCState &Info,
3218 MipsCC::SpecialCallingConvType SpecialCallingConv_)
3219 : CCInfo(Info), CallConv(CC), IsO32(IsO32_), IsFP64(IsFP64_),
3220 SpecialCallingConv(SpecialCallingConv_){
3221 // Pre-allocate reserved argument area.
3222 CCInfo.AllocateStack(reservedArgArea(), 1);
3226 void MipsTargetLowering::MipsCC::
3227 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
3228 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode,
3229 std::vector<ArgListEntry> &FuncArgs) {
3230 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
3231 "CallingConv::Fast shouldn't be used for vararg functions.");
3233 unsigned NumOpnds = Args.size();
3234 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
3236 for (unsigned I = 0; I != NumOpnds; ++I) {
3237 MVT ArgVT = Args[I].VT;
3238 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3241 if (ArgFlags.isByVal()) {
3242 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3246 if (IsVarArg && !Args[I].IsFixed)
3247 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3249 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode,
3251 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo);
3256 dbgs() << "Call operand #" << I << " has unhandled type "
3257 << EVT(ArgVT).getEVTString();
3259 llvm_unreachable(0);
3264 void MipsTargetLowering::MipsCC::
3265 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args,
3266 bool IsSoftFloat, Function::const_arg_iterator FuncArg) {
3267 unsigned NumArgs = Args.size();
3268 llvm::CCAssignFn *FixedFn = fixedArgFn();
3269 unsigned CurArgIdx = 0;
3271 for (unsigned I = 0; I != NumArgs; ++I) {
3272 MVT ArgVT = Args[I].VT;
3273 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3274 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx);
3275 CurArgIdx = Args[I].OrigArgIndex;
3277 if (ArgFlags.isByVal()) {
3278 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3282 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat);
3284 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo))
3288 dbgs() << "Formal Arg #" << I << " has unhandled type "
3289 << EVT(ArgVT).getEVTString();
3291 llvm_unreachable(0);
3295 template<typename Ty>
3296 void MipsTargetLowering::MipsCC::
3297 analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat,
3298 const SDNode *CallNode, const Type *RetTy) const {
3301 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode))
3302 Fn = RetCC_F128Soft;
3306 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) {
3307 MVT VT = RetVals[I].VT;
3308 ISD::ArgFlagsTy Flags = RetVals[I].Flags;
3309 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat);
3311 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) {
3313 dbgs() << "Call result #" << I << " has unhandled type "
3314 << EVT(VT).getEVTString() << '\n';
3316 llvm_unreachable(0);
3321 void MipsTargetLowering::MipsCC::
3322 analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat,
3323 const SDNode *CallNode, const Type *RetTy) const {
3324 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy);
3327 void MipsTargetLowering::MipsCC::
3328 analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat,
3329 const Type *RetTy) const {
3330 analyzeReturn(Outs, IsSoftFloat, 0, RetTy);
3334 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3336 CCValAssign::LocInfo LocInfo,
3337 ISD::ArgFlagsTy ArgFlags) {
3338 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3340 struct ByValArgInfo ByVal;
3341 unsigned RegSize = regSize();
3342 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3343 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3346 if (useRegsForByval())
3347 allocateRegs(ByVal, ByValSize, Align);
3349 // Allocate space on caller's stack.
3350 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3352 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3354 ByValArgs.push_back(ByVal);
3357 unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const {
3358 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs);
3361 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
3362 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0;
3365 const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const {
3366 return IsO32 ? O32IntRegs : Mips64IntRegs;
3369 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
3370 if (CallConv == CallingConv::Fast)
3371 return CC_Mips_FastCC;
3373 if (SpecialCallingConv == Mips16RetHelperConv)
3374 return CC_Mips16RetHelper;
3375 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN;
3378 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
3379 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN_VarArg;
3382 const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const {
3383 return IsO32 ? O32IntRegs : Mips64DPRegs;
3386 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3389 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs();
3390 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs();
3391 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3392 "Byval argument's size and alignment should be a multiple of"
3395 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3397 // If Align > RegSize, the first arg register must be even.
3398 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3399 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3403 // Mark the registers allocated.
3404 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3405 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3406 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3409 MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy,
3410 const SDNode *CallNode,
3411 bool IsSoftFloat) const {
3412 if (IsSoftFloat || IsO32)
3415 // Check if the original type was fp128.
3416 if (originalTypeIsF128(OrigTy, CallNode)) {
3417 assert(VT == MVT::i64);
3424 void MipsTargetLowering::
3425 copyByValRegs(SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains,
3426 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3427 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3428 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3429 MachineFunction &MF = DAG.getMachineFunction();
3430 MachineFrameInfo *MFI = MF.getFrameInfo();
3431 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3432 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3436 FrameObjOffset = (int)CC.reservedArgArea() -
3437 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3439 FrameObjOffset = ByVal.Address;
3441 // Create frame object.
3442 EVT PtrTy = getPointerTy();
3443 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3444 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3445 InVals.push_back(FIN);
3450 // Copy arg registers.
3451 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3452 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3454 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3455 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3456 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3457 unsigned Offset = I * CC.regSize();
3458 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3459 DAG.getConstant(Offset, PtrTy));
3460 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3461 StorePtr, MachinePointerInfo(FuncArg, Offset),
3463 OutChains.push_back(Store);
3467 // Copy byVal arg to registers and stack.
3468 void MipsTargetLowering::
3469 passByValArg(SDValue Chain, SDLoc DL,
3470 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
3471 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
3472 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3473 const MipsCC &CC, const ByValArgInfo &ByVal,
3474 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3475 unsigned ByValSize = Flags.getByValSize();
3476 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
3477 unsigned RegSize = CC.regSize();
3478 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
3479 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
3481 if (ByVal.NumRegs) {
3482 const uint16_t *ArgRegs = CC.intArgRegs();
3483 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
3486 // Copy words to registers.
3487 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
3488 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3489 DAG.getConstant(Offset, PtrTy));
3490 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3491 MachinePointerInfo(), false, false, false,
3493 MemOpChains.push_back(LoadVal.getValue(1));
3494 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3495 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3498 // Return if the struct has been fully copied.
3499 if (ByValSize == Offset)
3502 // Copy the remainder of the byval argument with sub-word loads and shifts.
3503 if (LeftoverBytes) {
3504 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
3505 "Size of the remainder should be smaller than RegSize.");
3508 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
3509 Offset < ByValSize; LoadSize /= 2) {
3510 unsigned RemSize = ByValSize - Offset;
3512 if (RemSize < LoadSize)
3516 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3517 DAG.getConstant(Offset, PtrTy));
3519 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
3520 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
3521 false, false, Alignment);
3522 MemOpChains.push_back(LoadVal.getValue(1));
3524 // Shift the loaded value.
3528 Shamt = TotalSizeLoaded;
3530 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
3532 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3533 DAG.getConstant(Shamt, MVT::i32));
3536 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3541 TotalSizeLoaded += LoadSize;
3542 Alignment = std::min(Alignment, LoadSize);
3545 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3546 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3551 // Copy remainder of byval arg to it with memcpy.
3552 unsigned MemCpySize = ByValSize - Offset;
3553 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3554 DAG.getConstant(Offset, PtrTy));
3555 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3556 DAG.getIntPtrConstant(ByVal.Address));
3557 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3558 DAG.getConstant(MemCpySize, PtrTy), Alignment,
3559 /*isVolatile=*/false, /*AlwaysInline=*/false,
3560 MachinePointerInfo(0), MachinePointerInfo(0));
3561 MemOpChains.push_back(Chain);
3565 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3566 const MipsCC &CC, SDValue Chain,
3567 SDLoc DL, SelectionDAG &DAG) const {
3568 unsigned NumRegs = CC.numIntArgRegs();
3569 const uint16_t *ArgRegs = CC.intArgRegs();
3570 const CCState &CCInfo = CC.getCCInfo();
3571 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3572 unsigned RegSize = CC.regSize();
3573 MVT RegTy = MVT::getIntegerVT(RegSize * 8);
3574 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3575 MachineFunction &MF = DAG.getMachineFunction();
3576 MachineFrameInfo *MFI = MF.getFrameInfo();
3577 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3579 // Offset of the first variable argument from stack pointer.
3583 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3586 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3588 // Record the frame index of the first variable argument
3589 // which is a value necessary to VASTART.
3590 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3591 MipsFI->setVarArgsFrameIndex(FI);
3593 // Copy the integer registers that have not been used for argument passing
3594 // to the argument register save area. For O32, the save area is allocated
3595 // in the caller's stack frame, while for N32/64, it is allocated in the
3596 // callee's stack frame.
3597 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3598 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3599 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3600 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3601 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3602 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3603 MachinePointerInfo(), false, false, 0);
3604 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
3605 OutChains.push_back(Store);