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"
40 STATISTIC(NumTailCalls, "Number of tail calls");
43 LargeGOT("mxgot", cl::Hidden,
44 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
46 static const uint16_t O32IntRegs[4] = {
47 Mips::A0, Mips::A1, Mips::A2, Mips::A3
50 static const uint16_t Mips64IntRegs[8] = {
51 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
52 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64
55 static const uint16_t Mips64DPRegs[8] = {
56 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
57 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
60 // If I is a shifted mask, set the size (Size) and the first bit of the
61 // mask (Pos), and return true.
62 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
63 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
64 if (!isShiftedMask_64(I))
67 Size = CountPopulation_64(I);
68 Pos = CountTrailingZeros_64(I);
72 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
73 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
74 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
77 static SDValue getTargetNode(SDValue Op, SelectionDAG &DAG, unsigned Flag) {
78 EVT Ty = Op.getValueType();
80 if (GlobalAddressSDNode *N = dyn_cast<GlobalAddressSDNode>(Op))
81 return DAG.getTargetGlobalAddress(N->getGlobal(), Op.getDebugLoc(), Ty, 0,
83 if (ExternalSymbolSDNode *N = dyn_cast<ExternalSymbolSDNode>(Op))
84 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
85 if (BlockAddressSDNode *N = dyn_cast<BlockAddressSDNode>(Op))
86 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
87 if (JumpTableSDNode *N = dyn_cast<JumpTableSDNode>(Op))
88 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
89 if (ConstantPoolSDNode *N = dyn_cast<ConstantPoolSDNode>(Op))
90 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
91 N->getOffset(), Flag);
93 llvm_unreachable("Unexpected node type.");
97 static SDValue getAddrNonPIC(SDValue Op, SelectionDAG &DAG) {
98 DebugLoc DL = Op.getDebugLoc();
99 EVT Ty = Op.getValueType();
100 SDValue Hi = getTargetNode(Op, DAG, MipsII::MO_ABS_HI);
101 SDValue Lo = getTargetNode(Op, DAG, MipsII::MO_ABS_LO);
102 return DAG.getNode(ISD::ADD, DL, Ty,
103 DAG.getNode(MipsISD::Hi, DL, Ty, Hi),
104 DAG.getNode(MipsISD::Lo, DL, Ty, Lo));
107 SDValue MipsTargetLowering::getAddrLocal(SDValue Op, SelectionDAG &DAG,
108 bool HasMips64) const {
109 DebugLoc DL = Op.getDebugLoc();
110 EVT Ty = Op.getValueType();
111 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
112 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
113 getTargetNode(Op, DAG, GOTFlag));
114 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT,
115 MachinePointerInfo::getGOT(), false, false, false,
117 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
118 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty, getTargetNode(Op, DAG, LoFlag));
119 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo);
122 SDValue MipsTargetLowering::getAddrGlobal(SDValue Op, SelectionDAG &DAG,
123 unsigned Flag) const {
124 DebugLoc DL = Op.getDebugLoc();
125 EVT Ty = Op.getValueType();
126 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
127 getTargetNode(Op, DAG, Flag));
128 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt,
129 MachinePointerInfo::getGOT(), false, false, false, 0);
132 SDValue MipsTargetLowering::getAddrGlobalLargeGOT(SDValue Op, SelectionDAG &DAG,
134 unsigned LoFlag) const {
135 DebugLoc DL = Op.getDebugLoc();
136 EVT Ty = Op.getValueType();
137 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty, getTargetNode(Op, DAG, HiFlag));
138 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, getGlobalReg(DAG, Ty));
139 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi,
140 getTargetNode(Op, DAG, LoFlag));
141 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper,
142 MachinePointerInfo::getGOT(), false, false, false, 0);
145 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
147 case MipsISD::JmpLink: return "MipsISD::JmpLink";
148 case MipsISD::TailCall: return "MipsISD::TailCall";
149 case MipsISD::Hi: return "MipsISD::Hi";
150 case MipsISD::Lo: return "MipsISD::Lo";
151 case MipsISD::GPRel: return "MipsISD::GPRel";
152 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
153 case MipsISD::Ret: return "MipsISD::Ret";
154 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
155 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
156 case MipsISD::FPCmp: return "MipsISD::FPCmp";
157 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
158 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
159 case MipsISD::ExtractLOHI: return "MipsISD::ExtractLOHI";
160 case MipsISD::InsertLOHI: return "MipsISD::InsertLOHI";
161 case MipsISD::Mult: return "MipsISD::Mult";
162 case MipsISD::Multu: return "MipsISD::Multu";
163 case MipsISD::MAdd: return "MipsISD::MAdd";
164 case MipsISD::MAddu: return "MipsISD::MAddu";
165 case MipsISD::MSub: return "MipsISD::MSub";
166 case MipsISD::MSubu: return "MipsISD::MSubu";
167 case MipsISD::DivRem: return "MipsISD::DivRem";
168 case MipsISD::DivRemU: return "MipsISD::DivRemU";
169 case MipsISD::DivRem16: return "MipsISD::DivRem16";
170 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
171 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
172 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
173 case MipsISD::Wrapper: return "MipsISD::Wrapper";
174 case MipsISD::Sync: return "MipsISD::Sync";
175 case MipsISD::Ext: return "MipsISD::Ext";
176 case MipsISD::Ins: return "MipsISD::Ins";
177 case MipsISD::LWL: return "MipsISD::LWL";
178 case MipsISD::LWR: return "MipsISD::LWR";
179 case MipsISD::SWL: return "MipsISD::SWL";
180 case MipsISD::SWR: return "MipsISD::SWR";
181 case MipsISD::LDL: return "MipsISD::LDL";
182 case MipsISD::LDR: return "MipsISD::LDR";
183 case MipsISD::SDL: return "MipsISD::SDL";
184 case MipsISD::SDR: return "MipsISD::SDR";
185 case MipsISD::EXTP: return "MipsISD::EXTP";
186 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
187 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
188 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
189 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
190 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
191 case MipsISD::SHILO: return "MipsISD::SHILO";
192 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
193 case MipsISD::MULT: return "MipsISD::MULT";
194 case MipsISD::MULTU: return "MipsISD::MULTU";
195 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
196 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
197 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
198 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
199 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP";
200 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP";
201 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP";
202 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP";
203 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP";
204 default: return NULL;
209 MipsTargetLowering(MipsTargetMachine &TM)
210 : TargetLowering(TM, new MipsTargetObjectFile()),
211 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
212 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
213 IsO32(Subtarget->isABI_O32()) {
214 // Mips does not have i1 type, so use i32 for
215 // setcc operations results (slt, sgt, ...).
216 setBooleanContents(ZeroOrOneBooleanContent);
217 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
219 // Load extented operations for i1 types must be promoted
220 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
221 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
222 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
224 // MIPS doesn't have extending float->double load/store
225 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
226 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
228 // Used by legalize types to correctly generate the setcc result.
229 // Without this, every float setcc comes with a AND/OR with the result,
230 // we don't want this, since the fpcmp result goes to a flag register,
231 // which is used implicitly by brcond and select operations.
232 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
234 // Mips Custom Operations
235 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
236 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
237 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
238 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
239 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
240 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
241 setOperationAction(ISD::SELECT, MVT::f32, Custom);
242 setOperationAction(ISD::SELECT, MVT::f64, Custom);
243 setOperationAction(ISD::SELECT, MVT::i32, Custom);
244 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
245 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
246 setOperationAction(ISD::SETCC, MVT::f32, Custom);
247 setOperationAction(ISD::SETCC, MVT::f64, Custom);
248 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
249 setOperationAction(ISD::VASTART, MVT::Other, Custom);
250 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
251 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
253 if (!TM.Options.NoNaNsFPMath) {
254 setOperationAction(ISD::FABS, MVT::f32, Custom);
255 setOperationAction(ISD::FABS, MVT::f64, Custom);
259 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
260 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
261 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
262 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
263 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
264 setOperationAction(ISD::SELECT, MVT::i64, Custom);
265 setOperationAction(ISD::LOAD, MVT::i64, Custom);
266 setOperationAction(ISD::STORE, MVT::i64, Custom);
270 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
271 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
272 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
275 setOperationAction(ISD::ADD, MVT::i32, Custom);
277 setOperationAction(ISD::ADD, MVT::i64, Custom);
279 setOperationAction(ISD::SDIV, MVT::i32, Expand);
280 setOperationAction(ISD::SREM, MVT::i32, Expand);
281 setOperationAction(ISD::UDIV, MVT::i32, Expand);
282 setOperationAction(ISD::UREM, MVT::i32, Expand);
283 setOperationAction(ISD::SDIV, MVT::i64, Expand);
284 setOperationAction(ISD::SREM, MVT::i64, Expand);
285 setOperationAction(ISD::UDIV, MVT::i64, Expand);
286 setOperationAction(ISD::UREM, MVT::i64, Expand);
288 // Operations not directly supported by Mips.
289 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
290 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
291 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
292 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
293 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
294 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
295 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
296 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
297 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
298 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
299 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
300 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
301 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
302 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
303 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
304 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
305 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
306 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
307 setOperationAction(ISD::ROTL, MVT::i32, Expand);
308 setOperationAction(ISD::ROTL, MVT::i64, Expand);
309 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
310 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
312 if (!Subtarget->hasMips32r2())
313 setOperationAction(ISD::ROTR, MVT::i32, Expand);
315 if (!Subtarget->hasMips64r2())
316 setOperationAction(ISD::ROTR, MVT::i64, Expand);
318 setOperationAction(ISD::FSIN, MVT::f32, Expand);
319 setOperationAction(ISD::FSIN, MVT::f64, Expand);
320 setOperationAction(ISD::FCOS, MVT::f32, Expand);
321 setOperationAction(ISD::FCOS, MVT::f64, Expand);
322 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
323 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
324 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
325 setOperationAction(ISD::FPOW, MVT::f32, Expand);
326 setOperationAction(ISD::FPOW, MVT::f64, Expand);
327 setOperationAction(ISD::FLOG, MVT::f32, Expand);
328 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
329 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
330 setOperationAction(ISD::FEXP, MVT::f32, Expand);
331 setOperationAction(ISD::FMA, MVT::f32, Expand);
332 setOperationAction(ISD::FMA, MVT::f64, Expand);
333 setOperationAction(ISD::FREM, MVT::f32, Expand);
334 setOperationAction(ISD::FREM, MVT::f64, Expand);
336 if (!TM.Options.NoNaNsFPMath) {
337 setOperationAction(ISD::FNEG, MVT::f32, Expand);
338 setOperationAction(ISD::FNEG, MVT::f64, Expand);
341 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
342 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
343 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
344 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
346 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
348 setOperationAction(ISD::VAARG, MVT::Other, Expand);
349 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
350 setOperationAction(ISD::VAEND, MVT::Other, Expand);
352 // Use the default for now
353 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
354 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
356 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
357 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
358 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
359 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
361 setInsertFencesForAtomic(true);
363 if (!Subtarget->hasSEInReg()) {
364 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
365 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
368 if (!Subtarget->hasBitCount()) {
369 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
370 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
373 if (!Subtarget->hasSwap()) {
374 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
375 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
379 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
380 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
381 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
382 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
385 setTargetDAGCombine(ISD::SDIVREM);
386 setTargetDAGCombine(ISD::UDIVREM);
387 setTargetDAGCombine(ISD::SELECT);
388 setTargetDAGCombine(ISD::AND);
389 setTargetDAGCombine(ISD::OR);
390 setTargetDAGCombine(ISD::ADD);
392 setMinFunctionAlignment(HasMips64 ? 3 : 2);
394 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP);
396 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0);
397 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1);
399 MaxStoresPerMemcpy = 16;
402 const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) {
403 if (TM.getSubtargetImpl()->inMips16Mode())
404 return llvm::createMips16TargetLowering(TM);
406 return llvm::createMipsSETargetLowering(TM);
409 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
412 return VT.changeVectorElementTypeToInteger();
415 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
416 TargetLowering::DAGCombinerInfo &DCI,
417 const MipsSubtarget *Subtarget) {
418 if (DCI.isBeforeLegalizeOps())
421 EVT Ty = N->getValueType(0);
422 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
423 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
424 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
426 DebugLoc DL = N->getDebugLoc();
428 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
429 N->getOperand(0), N->getOperand(1));
430 SDValue InChain = DAG.getEntryNode();
431 SDValue InGlue = DivRem;
434 if (N->hasAnyUseOfValue(0)) {
435 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
437 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
438 InChain = CopyFromLo.getValue(1);
439 InGlue = CopyFromLo.getValue(2);
443 if (N->hasAnyUseOfValue(1)) {
444 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
446 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
452 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
454 default: llvm_unreachable("Unknown fp condition code!");
456 case ISD::SETOEQ: return Mips::FCOND_OEQ;
457 case ISD::SETUNE: return Mips::FCOND_UNE;
459 case ISD::SETOLT: return Mips::FCOND_OLT;
461 case ISD::SETOGT: return Mips::FCOND_OGT;
463 case ISD::SETOLE: return Mips::FCOND_OLE;
465 case ISD::SETOGE: return Mips::FCOND_OGE;
466 case ISD::SETULT: return Mips::FCOND_ULT;
467 case ISD::SETULE: return Mips::FCOND_ULE;
468 case ISD::SETUGT: return Mips::FCOND_UGT;
469 case ISD::SETUGE: return Mips::FCOND_UGE;
470 case ISD::SETUO: return Mips::FCOND_UN;
471 case ISD::SETO: return Mips::FCOND_OR;
473 case ISD::SETONE: return Mips::FCOND_ONE;
474 case ISD::SETUEQ: return Mips::FCOND_UEQ;
479 /// This function returns true if the floating point conditional branches and
480 /// conditional moves which use condition code CC should be inverted.
481 static bool invertFPCondCodeUser(Mips::CondCode CC) {
482 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
485 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
486 "Illegal Condition Code");
491 // Creates and returns an FPCmp node from a setcc node.
492 // Returns Op if setcc is not a floating point comparison.
493 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
494 // must be a SETCC node
495 if (Op.getOpcode() != ISD::SETCC)
498 SDValue LHS = Op.getOperand(0);
500 if (!LHS.getValueType().isFloatingPoint())
503 SDValue RHS = Op.getOperand(1);
504 DebugLoc DL = Op.getDebugLoc();
506 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
507 // node if necessary.
508 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
510 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
511 DAG.getConstant(condCodeToFCC(CC), MVT::i32));
514 // Creates and returns a CMovFPT/F node.
515 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
516 SDValue False, DebugLoc DL) {
517 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
518 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
520 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
521 True.getValueType(), True, False, Cond);
524 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
525 TargetLowering::DAGCombinerInfo &DCI,
526 const MipsSubtarget *Subtarget) {
527 if (DCI.isBeforeLegalizeOps())
530 SDValue SetCC = N->getOperand(0);
532 if ((SetCC.getOpcode() != ISD::SETCC) ||
533 !SetCC.getOperand(0).getValueType().isInteger())
536 SDValue False = N->getOperand(2);
537 EVT FalseTy = False.getValueType();
539 if (!FalseTy.isInteger())
542 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
544 if (!CN || CN->getZExtValue())
547 const DebugLoc DL = N->getDebugLoc();
548 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
549 SDValue True = N->getOperand(1);
551 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
552 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
554 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
557 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
558 TargetLowering::DAGCombinerInfo &DCI,
559 const MipsSubtarget *Subtarget) {
560 // Pattern match EXT.
561 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
562 // => ext $dst, $src, size, pos
563 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
566 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
567 unsigned ShiftRightOpc = ShiftRight.getOpcode();
569 // Op's first operand must be a shift right.
570 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
573 // The second operand of the shift must be an immediate.
575 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
578 uint64_t Pos = CN->getZExtValue();
579 uint64_t SMPos, SMSize;
581 // Op's second operand must be a shifted mask.
582 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
583 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
586 // Return if the shifted mask does not start at bit 0 or the sum of its size
587 // and Pos exceeds the word's size.
588 EVT ValTy = N->getValueType(0);
589 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
592 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
593 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
594 DAG.getConstant(SMSize, MVT::i32));
597 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
598 TargetLowering::DAGCombinerInfo &DCI,
599 const MipsSubtarget *Subtarget) {
600 // Pattern match INS.
601 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
602 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
603 // => ins $dst, $src, size, pos, $src1
604 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
607 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
608 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
611 // See if Op's first operand matches (and $src1 , mask0).
612 if (And0.getOpcode() != ISD::AND)
615 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
616 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
619 // See if Op's second operand matches (and (shl $src, pos), mask1).
620 if (And1.getOpcode() != ISD::AND)
623 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
624 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
627 // The shift masks must have the same position and size.
628 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
631 SDValue Shl = And1.getOperand(0);
632 if (Shl.getOpcode() != ISD::SHL)
635 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
638 unsigned Shamt = CN->getZExtValue();
640 // Return if the shift amount and the first bit position of mask are not the
642 EVT ValTy = N->getValueType(0);
643 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
646 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
647 DAG.getConstant(SMPos0, MVT::i32),
648 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
651 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
652 TargetLowering::DAGCombinerInfo &DCI,
653 const MipsSubtarget *Subtarget) {
654 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
656 if (DCI.isBeforeLegalizeOps())
659 SDValue Add = N->getOperand(1);
661 if (Add.getOpcode() != ISD::ADD)
664 SDValue Lo = Add.getOperand(1);
666 if ((Lo.getOpcode() != MipsISD::Lo) ||
667 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
670 EVT ValTy = N->getValueType(0);
671 DebugLoc DL = N->getDebugLoc();
673 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
675 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
678 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
680 SelectionDAG &DAG = DCI.DAG;
681 unsigned Opc = N->getOpcode();
687 return performDivRemCombine(N, DAG, DCI, Subtarget);
689 return performSELECTCombine(N, DAG, DCI, Subtarget);
691 return performANDCombine(N, DAG, DCI, Subtarget);
693 return performORCombine(N, DAG, DCI, Subtarget);
695 return performADDCombine(N, DAG, DCI, Subtarget);
702 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
703 SmallVectorImpl<SDValue> &Results,
704 SelectionDAG &DAG) const {
705 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
707 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
708 Results.push_back(Res.getValue(I));
712 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
713 SmallVectorImpl<SDValue> &Results,
714 SelectionDAG &DAG) const {
715 return LowerOperationWrapper(N, Results, DAG);
718 SDValue MipsTargetLowering::
719 LowerOperation(SDValue Op, SelectionDAG &DAG) const
721 switch (Op.getOpcode())
723 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
724 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
725 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
726 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
727 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
728 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
729 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
730 case ISD::SELECT: return lowerSELECT(Op, DAG);
731 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
732 case ISD::SETCC: return lowerSETCC(Op, DAG);
733 case ISD::VASTART: return lowerVASTART(Op, DAG);
734 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
735 case ISD::FABS: return lowerFABS(Op, DAG);
736 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
737 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
738 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
739 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
740 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
741 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
742 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
743 case ISD::LOAD: return lowerLOAD(Op, DAG);
744 case ISD::STORE: return lowerSTORE(Op, DAG);
745 case ISD::ADD: return lowerADD(Op, DAG);
750 //===----------------------------------------------------------------------===//
751 // Lower helper functions
752 //===----------------------------------------------------------------------===//
754 // addLiveIn - This helper function adds the specified physical register to the
755 // MachineFunction as a live in value. It also creates a corresponding
756 // virtual register for it.
758 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
760 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
761 MF.getRegInfo().addLiveIn(PReg, VReg);
766 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
767 MachineBasicBlock *BB) const {
768 switch (MI->getOpcode()) {
770 llvm_unreachable("Unexpected instr type to insert");
771 case Mips::ATOMIC_LOAD_ADD_I8:
772 case Mips::ATOMIC_LOAD_ADD_I8_P8:
773 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
774 case Mips::ATOMIC_LOAD_ADD_I16:
775 case Mips::ATOMIC_LOAD_ADD_I16_P8:
776 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
777 case Mips::ATOMIC_LOAD_ADD_I32:
778 case Mips::ATOMIC_LOAD_ADD_I32_P8:
779 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
780 case Mips::ATOMIC_LOAD_ADD_I64:
781 case Mips::ATOMIC_LOAD_ADD_I64_P8:
782 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
784 case Mips::ATOMIC_LOAD_AND_I8:
785 case Mips::ATOMIC_LOAD_AND_I8_P8:
786 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
787 case Mips::ATOMIC_LOAD_AND_I16:
788 case Mips::ATOMIC_LOAD_AND_I16_P8:
789 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
790 case Mips::ATOMIC_LOAD_AND_I32:
791 case Mips::ATOMIC_LOAD_AND_I32_P8:
792 return emitAtomicBinary(MI, BB, 4, Mips::AND);
793 case Mips::ATOMIC_LOAD_AND_I64:
794 case Mips::ATOMIC_LOAD_AND_I64_P8:
795 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
797 case Mips::ATOMIC_LOAD_OR_I8:
798 case Mips::ATOMIC_LOAD_OR_I8_P8:
799 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
800 case Mips::ATOMIC_LOAD_OR_I16:
801 case Mips::ATOMIC_LOAD_OR_I16_P8:
802 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
803 case Mips::ATOMIC_LOAD_OR_I32:
804 case Mips::ATOMIC_LOAD_OR_I32_P8:
805 return emitAtomicBinary(MI, BB, 4, Mips::OR);
806 case Mips::ATOMIC_LOAD_OR_I64:
807 case Mips::ATOMIC_LOAD_OR_I64_P8:
808 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
810 case Mips::ATOMIC_LOAD_XOR_I8:
811 case Mips::ATOMIC_LOAD_XOR_I8_P8:
812 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
813 case Mips::ATOMIC_LOAD_XOR_I16:
814 case Mips::ATOMIC_LOAD_XOR_I16_P8:
815 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
816 case Mips::ATOMIC_LOAD_XOR_I32:
817 case Mips::ATOMIC_LOAD_XOR_I32_P8:
818 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
819 case Mips::ATOMIC_LOAD_XOR_I64:
820 case Mips::ATOMIC_LOAD_XOR_I64_P8:
821 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
823 case Mips::ATOMIC_LOAD_NAND_I8:
824 case Mips::ATOMIC_LOAD_NAND_I8_P8:
825 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
826 case Mips::ATOMIC_LOAD_NAND_I16:
827 case Mips::ATOMIC_LOAD_NAND_I16_P8:
828 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
829 case Mips::ATOMIC_LOAD_NAND_I32:
830 case Mips::ATOMIC_LOAD_NAND_I32_P8:
831 return emitAtomicBinary(MI, BB, 4, 0, true);
832 case Mips::ATOMIC_LOAD_NAND_I64:
833 case Mips::ATOMIC_LOAD_NAND_I64_P8:
834 return emitAtomicBinary(MI, BB, 8, 0, true);
836 case Mips::ATOMIC_LOAD_SUB_I8:
837 case Mips::ATOMIC_LOAD_SUB_I8_P8:
838 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
839 case Mips::ATOMIC_LOAD_SUB_I16:
840 case Mips::ATOMIC_LOAD_SUB_I16_P8:
841 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
842 case Mips::ATOMIC_LOAD_SUB_I32:
843 case Mips::ATOMIC_LOAD_SUB_I32_P8:
844 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
845 case Mips::ATOMIC_LOAD_SUB_I64:
846 case Mips::ATOMIC_LOAD_SUB_I64_P8:
847 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
849 case Mips::ATOMIC_SWAP_I8:
850 case Mips::ATOMIC_SWAP_I8_P8:
851 return emitAtomicBinaryPartword(MI, BB, 1, 0);
852 case Mips::ATOMIC_SWAP_I16:
853 case Mips::ATOMIC_SWAP_I16_P8:
854 return emitAtomicBinaryPartword(MI, BB, 2, 0);
855 case Mips::ATOMIC_SWAP_I32:
856 case Mips::ATOMIC_SWAP_I32_P8:
857 return emitAtomicBinary(MI, BB, 4, 0);
858 case Mips::ATOMIC_SWAP_I64:
859 case Mips::ATOMIC_SWAP_I64_P8:
860 return emitAtomicBinary(MI, BB, 8, 0);
862 case Mips::ATOMIC_CMP_SWAP_I8:
863 case Mips::ATOMIC_CMP_SWAP_I8_P8:
864 return emitAtomicCmpSwapPartword(MI, BB, 1);
865 case Mips::ATOMIC_CMP_SWAP_I16:
866 case Mips::ATOMIC_CMP_SWAP_I16_P8:
867 return emitAtomicCmpSwapPartword(MI, BB, 2);
868 case Mips::ATOMIC_CMP_SWAP_I32:
869 case Mips::ATOMIC_CMP_SWAP_I32_P8:
870 return emitAtomicCmpSwap(MI, BB, 4);
871 case Mips::ATOMIC_CMP_SWAP_I64:
872 case Mips::ATOMIC_CMP_SWAP_I64_P8:
873 return emitAtomicCmpSwap(MI, BB, 8);
877 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
878 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
880 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
881 unsigned Size, unsigned BinOpcode,
883 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
885 MachineFunction *MF = BB->getParent();
886 MachineRegisterInfo &RegInfo = MF->getRegInfo();
887 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
888 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
889 DebugLoc DL = MI->getDebugLoc();
890 unsigned LL, SC, AND, NOR, ZERO, BEQ;
893 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
894 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
901 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
902 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
905 ZERO = Mips::ZERO_64;
909 unsigned OldVal = MI->getOperand(0).getReg();
910 unsigned Ptr = MI->getOperand(1).getReg();
911 unsigned Incr = MI->getOperand(2).getReg();
913 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
914 unsigned AndRes = RegInfo.createVirtualRegister(RC);
915 unsigned Success = RegInfo.createVirtualRegister(RC);
917 // insert new blocks after the current block
918 const BasicBlock *LLVM_BB = BB->getBasicBlock();
919 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
920 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
921 MachineFunction::iterator It = BB;
923 MF->insert(It, loopMBB);
924 MF->insert(It, exitMBB);
926 // Transfer the remainder of BB and its successor edges to exitMBB.
927 exitMBB->splice(exitMBB->begin(), BB,
928 llvm::next(MachineBasicBlock::iterator(MI)),
930 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
934 // fallthrough --> loopMBB
935 BB->addSuccessor(loopMBB);
936 loopMBB->addSuccessor(loopMBB);
937 loopMBB->addSuccessor(exitMBB);
941 // <binop> storeval, oldval, incr
942 // sc success, storeval, 0(ptr)
943 // beq success, $0, loopMBB
945 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
947 // and andres, oldval, incr
948 // nor storeval, $0, andres
949 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
950 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
951 } else if (BinOpcode) {
952 // <binop> storeval, oldval, incr
953 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
957 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
958 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
960 MI->eraseFromParent(); // The instruction is gone now.
966 MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI,
967 MachineBasicBlock *BB,
968 unsigned Size, unsigned BinOpcode,
970 assert((Size == 1 || Size == 2) &&
971 "Unsupported size for EmitAtomicBinaryPartial.");
973 MachineFunction *MF = BB->getParent();
974 MachineRegisterInfo &RegInfo = MF->getRegInfo();
975 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
976 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
977 DebugLoc DL = MI->getDebugLoc();
978 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
979 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
981 unsigned Dest = MI->getOperand(0).getReg();
982 unsigned Ptr = MI->getOperand(1).getReg();
983 unsigned Incr = MI->getOperand(2).getReg();
985 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
986 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
987 unsigned Mask = RegInfo.createVirtualRegister(RC);
988 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
989 unsigned NewVal = RegInfo.createVirtualRegister(RC);
990 unsigned OldVal = RegInfo.createVirtualRegister(RC);
991 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
992 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
993 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
994 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
995 unsigned AndRes = RegInfo.createVirtualRegister(RC);
996 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
997 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
998 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
999 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1000 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1001 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1002 unsigned Success = RegInfo.createVirtualRegister(RC);
1004 // insert new blocks after the current block
1005 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1006 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1007 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1008 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1009 MachineFunction::iterator It = BB;
1011 MF->insert(It, loopMBB);
1012 MF->insert(It, sinkMBB);
1013 MF->insert(It, exitMBB);
1015 // Transfer the remainder of BB and its successor edges to exitMBB.
1016 exitMBB->splice(exitMBB->begin(), BB,
1017 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1018 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1020 BB->addSuccessor(loopMBB);
1021 loopMBB->addSuccessor(loopMBB);
1022 loopMBB->addSuccessor(sinkMBB);
1023 sinkMBB->addSuccessor(exitMBB);
1026 // addiu masklsb2,$0,-4 # 0xfffffffc
1027 // and alignedaddr,ptr,masklsb2
1028 // andi ptrlsb2,ptr,3
1029 // sll shiftamt,ptrlsb2,3
1030 // ori maskupper,$0,255 # 0xff
1031 // sll mask,maskupper,shiftamt
1032 // nor mask2,$0,mask
1033 // sll incr2,incr,shiftamt
1035 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1036 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1037 .addReg(Mips::ZERO).addImm(-4);
1038 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1039 .addReg(Ptr).addReg(MaskLSB2);
1040 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1041 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1042 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1043 .addReg(Mips::ZERO).addImm(MaskImm);
1044 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1045 .addReg(ShiftAmt).addReg(MaskUpper);
1046 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1047 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1049 // atomic.load.binop
1051 // ll oldval,0(alignedaddr)
1052 // binop binopres,oldval,incr2
1053 // and newval,binopres,mask
1054 // and maskedoldval0,oldval,mask2
1055 // or storeval,maskedoldval0,newval
1056 // sc success,storeval,0(alignedaddr)
1057 // beq success,$0,loopMBB
1061 // ll oldval,0(alignedaddr)
1062 // and newval,incr2,mask
1063 // and maskedoldval0,oldval,mask2
1064 // or storeval,maskedoldval0,newval
1065 // sc success,storeval,0(alignedaddr)
1066 // beq success,$0,loopMBB
1069 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1071 // and andres, oldval, incr2
1072 // nor binopres, $0, andres
1073 // and newval, binopres, mask
1074 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1075 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1076 .addReg(Mips::ZERO).addReg(AndRes);
1077 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1078 } else if (BinOpcode) {
1079 // <binop> binopres, oldval, incr2
1080 // and newval, binopres, mask
1081 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1082 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1083 } else {// atomic.swap
1084 // and newval, incr2, mask
1085 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1088 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1089 .addReg(OldVal).addReg(Mask2);
1090 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1091 .addReg(MaskedOldVal0).addReg(NewVal);
1092 BuildMI(BB, DL, TII->get(SC), Success)
1093 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1094 BuildMI(BB, DL, TII->get(Mips::BEQ))
1095 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1098 // and maskedoldval1,oldval,mask
1099 // srl srlres,maskedoldval1,shiftamt
1100 // sll sllres,srlres,24
1101 // sra dest,sllres,24
1103 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1105 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1106 .addReg(OldVal).addReg(Mask);
1107 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1108 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1109 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1110 .addReg(SrlRes).addImm(ShiftImm);
1111 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1112 .addReg(SllRes).addImm(ShiftImm);
1114 MI->eraseFromParent(); // The instruction is gone now.
1120 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1121 MachineBasicBlock *BB,
1122 unsigned Size) const {
1123 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1125 MachineFunction *MF = BB->getParent();
1126 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1127 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1128 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1129 DebugLoc DL = MI->getDebugLoc();
1130 unsigned LL, SC, ZERO, BNE, BEQ;
1133 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1134 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1140 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1141 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1142 ZERO = Mips::ZERO_64;
1147 unsigned Dest = MI->getOperand(0).getReg();
1148 unsigned Ptr = MI->getOperand(1).getReg();
1149 unsigned OldVal = MI->getOperand(2).getReg();
1150 unsigned NewVal = MI->getOperand(3).getReg();
1152 unsigned Success = RegInfo.createVirtualRegister(RC);
1154 // insert new blocks after the current block
1155 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1156 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1157 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1158 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1159 MachineFunction::iterator It = BB;
1161 MF->insert(It, loop1MBB);
1162 MF->insert(It, loop2MBB);
1163 MF->insert(It, exitMBB);
1165 // Transfer the remainder of BB and its successor edges to exitMBB.
1166 exitMBB->splice(exitMBB->begin(), BB,
1167 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1168 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1172 // fallthrough --> loop1MBB
1173 BB->addSuccessor(loop1MBB);
1174 loop1MBB->addSuccessor(exitMBB);
1175 loop1MBB->addSuccessor(loop2MBB);
1176 loop2MBB->addSuccessor(loop1MBB);
1177 loop2MBB->addSuccessor(exitMBB);
1181 // bne dest, oldval, exitMBB
1183 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1184 BuildMI(BB, DL, TII->get(BNE))
1185 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1188 // sc success, newval, 0(ptr)
1189 // beq success, $0, loop1MBB
1191 BuildMI(BB, DL, TII->get(SC), Success)
1192 .addReg(NewVal).addReg(Ptr).addImm(0);
1193 BuildMI(BB, DL, TII->get(BEQ))
1194 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1196 MI->eraseFromParent(); // The instruction is gone now.
1202 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1203 MachineBasicBlock *BB,
1204 unsigned Size) const {
1205 assert((Size == 1 || Size == 2) &&
1206 "Unsupported size for EmitAtomicCmpSwapPartial.");
1208 MachineFunction *MF = BB->getParent();
1209 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1210 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1211 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1212 DebugLoc DL = MI->getDebugLoc();
1213 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1214 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1216 unsigned Dest = MI->getOperand(0).getReg();
1217 unsigned Ptr = MI->getOperand(1).getReg();
1218 unsigned CmpVal = MI->getOperand(2).getReg();
1219 unsigned NewVal = MI->getOperand(3).getReg();
1221 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1222 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1223 unsigned Mask = RegInfo.createVirtualRegister(RC);
1224 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1225 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1226 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1227 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1228 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1229 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1230 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1231 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1232 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1233 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1234 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1235 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1236 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1237 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1238 unsigned Success = RegInfo.createVirtualRegister(RC);
1240 // insert new blocks after the current block
1241 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1242 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1243 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1244 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1245 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1246 MachineFunction::iterator It = BB;
1248 MF->insert(It, loop1MBB);
1249 MF->insert(It, loop2MBB);
1250 MF->insert(It, sinkMBB);
1251 MF->insert(It, exitMBB);
1253 // Transfer the remainder of BB and its successor edges to exitMBB.
1254 exitMBB->splice(exitMBB->begin(), BB,
1255 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1256 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1258 BB->addSuccessor(loop1MBB);
1259 loop1MBB->addSuccessor(sinkMBB);
1260 loop1MBB->addSuccessor(loop2MBB);
1261 loop2MBB->addSuccessor(loop1MBB);
1262 loop2MBB->addSuccessor(sinkMBB);
1263 sinkMBB->addSuccessor(exitMBB);
1265 // FIXME: computation of newval2 can be moved to loop2MBB.
1267 // addiu masklsb2,$0,-4 # 0xfffffffc
1268 // and alignedaddr,ptr,masklsb2
1269 // andi ptrlsb2,ptr,3
1270 // sll shiftamt,ptrlsb2,3
1271 // ori maskupper,$0,255 # 0xff
1272 // sll mask,maskupper,shiftamt
1273 // nor mask2,$0,mask
1274 // andi maskedcmpval,cmpval,255
1275 // sll shiftedcmpval,maskedcmpval,shiftamt
1276 // andi maskednewval,newval,255
1277 // sll shiftednewval,maskednewval,shiftamt
1278 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1279 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1280 .addReg(Mips::ZERO).addImm(-4);
1281 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1282 .addReg(Ptr).addReg(MaskLSB2);
1283 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1284 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1285 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1286 .addReg(Mips::ZERO).addImm(MaskImm);
1287 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1288 .addReg(ShiftAmt).addReg(MaskUpper);
1289 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1290 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1291 .addReg(CmpVal).addImm(MaskImm);
1292 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1293 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1294 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1295 .addReg(NewVal).addImm(MaskImm);
1296 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1297 .addReg(ShiftAmt).addReg(MaskedNewVal);
1300 // ll oldval,0(alginedaddr)
1301 // and maskedoldval0,oldval,mask
1302 // bne maskedoldval0,shiftedcmpval,sinkMBB
1304 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1305 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1306 .addReg(OldVal).addReg(Mask);
1307 BuildMI(BB, DL, TII->get(Mips::BNE))
1308 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1311 // and maskedoldval1,oldval,mask2
1312 // or storeval,maskedoldval1,shiftednewval
1313 // sc success,storeval,0(alignedaddr)
1314 // beq success,$0,loop1MBB
1316 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1317 .addReg(OldVal).addReg(Mask2);
1318 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1319 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1320 BuildMI(BB, DL, TII->get(SC), Success)
1321 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1322 BuildMI(BB, DL, TII->get(Mips::BEQ))
1323 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1326 // srl srlres,maskedoldval0,shiftamt
1327 // sll sllres,srlres,24
1328 // sra dest,sllres,24
1330 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1332 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1333 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1334 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1335 .addReg(SrlRes).addImm(ShiftImm);
1336 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1337 .addReg(SllRes).addImm(ShiftImm);
1339 MI->eraseFromParent(); // The instruction is gone now.
1344 //===----------------------------------------------------------------------===//
1345 // Misc Lower Operation implementation
1346 //===----------------------------------------------------------------------===//
1347 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1348 SDValue Chain = Op.getOperand(0);
1349 SDValue Table = Op.getOperand(1);
1350 SDValue Index = Op.getOperand(2);
1351 DebugLoc DL = Op.getDebugLoc();
1352 EVT PTy = getPointerTy();
1353 unsigned EntrySize =
1354 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout());
1356 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1357 DAG.getConstant(EntrySize, PTy));
1358 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1360 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1361 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1362 MachinePointerInfo::getJumpTable(), MemVT, false, false,
1364 Chain = Addr.getValue(1);
1366 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) {
1367 // For PIC, the sequence is:
1368 // BRIND(load(Jumptable + index) + RelocBase)
1369 // RelocBase can be JumpTable, GOT or some sort of global base.
1370 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1371 getPICJumpTableRelocBase(Table, DAG));
1374 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1377 SDValue MipsTargetLowering::
1378 lowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1380 // The first operand is the chain, the second is the condition, the third is
1381 // the block to branch to if the condition is true.
1382 SDValue Chain = Op.getOperand(0);
1383 SDValue Dest = Op.getOperand(2);
1384 DebugLoc DL = Op.getDebugLoc();
1386 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1388 // Return if flag is not set by a floating point comparison.
1389 if (CondRes.getOpcode() != MipsISD::FPCmp)
1392 SDValue CCNode = CondRes.getOperand(2);
1394 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1395 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1396 SDValue BrCode = DAG.getConstant(Opc, MVT::i32);
1397 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1401 SDValue MipsTargetLowering::
1402 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1404 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1406 // Return if flag is not set by a floating point comparison.
1407 if (Cond.getOpcode() != MipsISD::FPCmp)
1410 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1414 SDValue MipsTargetLowering::
1415 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1417 DebugLoc DL = Op.getDebugLoc();
1418 EVT Ty = Op.getOperand(0).getValueType();
1419 SDValue Cond = DAG.getNode(ISD::SETCC, DL, getSetCCResultType(Ty),
1420 Op.getOperand(0), Op.getOperand(1),
1423 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1427 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1428 SDValue Cond = createFPCmp(DAG, Op);
1430 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1431 "Floating point operand expected.");
1433 SDValue True = DAG.getConstant(1, MVT::i32);
1434 SDValue False = DAG.getConstant(0, MVT::i32);
1436 return createCMovFP(DAG, Cond, True, False, Op.getDebugLoc());
1439 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1440 SelectionDAG &DAG) const {
1441 // FIXME there isn't actually debug info here
1442 DebugLoc DL = Op.getDebugLoc();
1443 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1445 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1446 const MipsTargetObjectFile &TLOF =
1447 (const MipsTargetObjectFile&)getObjFileLowering();
1449 // %gp_rel relocation
1450 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1451 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0,
1453 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL,
1454 DAG.getVTList(MVT::i32), &GA, 1);
1455 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1456 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode);
1459 // %hi/%lo relocation
1460 return getAddrNonPIC(Op, DAG);
1463 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1464 return getAddrLocal(Op, DAG, HasMips64);
1467 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16,
1468 MipsII::MO_GOT_LO16);
1470 return getAddrGlobal(Op, DAG,
1471 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
1474 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1475 SelectionDAG &DAG) const {
1476 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1477 return getAddrNonPIC(Op, DAG);
1479 return getAddrLocal(Op, DAG, HasMips64);
1482 SDValue MipsTargetLowering::
1483 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1485 // If the relocation model is PIC, use the General Dynamic TLS Model or
1486 // Local Dynamic TLS model, otherwise use the Initial Exec or
1487 // Local Exec TLS Model.
1489 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1490 DebugLoc DL = GA->getDebugLoc();
1491 const GlobalValue *GV = GA->getGlobal();
1492 EVT PtrVT = getPointerTy();
1494 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1496 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1497 // General Dynamic and Local Dynamic TLS Model.
1498 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1501 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1502 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1503 getGlobalReg(DAG, PtrVT), TGA);
1504 unsigned PtrSize = PtrVT.getSizeInBits();
1505 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1507 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1511 Entry.Node = Argument;
1513 Args.push_back(Entry);
1515 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
1516 false, false, false, false, 0, CallingConv::C,
1517 /*IsTailCall=*/false, /*doesNotRet=*/false,
1518 /*isReturnValueUsed=*/true,
1519 TlsGetAddr, Args, DAG, DL);
1520 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1522 SDValue Ret = CallResult.first;
1524 if (model != TLSModel::LocalDynamic)
1527 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1528 MipsII::MO_DTPREL_HI);
1529 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1530 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1531 MipsII::MO_DTPREL_LO);
1532 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1533 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1534 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1538 if (model == TLSModel::InitialExec) {
1539 // Initial Exec TLS Model
1540 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1541 MipsII::MO_GOTTPREL);
1542 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1544 Offset = DAG.getLoad(PtrVT, DL,
1545 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1546 false, false, false, 0);
1548 // Local Exec TLS Model
1549 assert(model == TLSModel::LocalExec);
1550 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1551 MipsII::MO_TPREL_HI);
1552 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1553 MipsII::MO_TPREL_LO);
1554 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1555 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1556 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1559 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1560 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1563 SDValue MipsTargetLowering::
1564 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1566 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1567 return getAddrNonPIC(Op, DAG);
1569 return getAddrLocal(Op, DAG, HasMips64);
1572 SDValue MipsTargetLowering::
1573 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1575 // gp_rel relocation
1576 // FIXME: we should reference the constant pool using small data sections,
1577 // but the asm printer currently doesn't support this feature without
1578 // hacking it. This feature should come soon so we can uncomment the
1580 //if (IsInSmallSection(C->getType())) {
1581 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1582 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1583 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1585 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1586 return getAddrNonPIC(Op, DAG);
1588 return getAddrLocal(Op, DAG, HasMips64);
1591 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1592 MachineFunction &MF = DAG.getMachineFunction();
1593 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1595 DebugLoc DL = Op.getDebugLoc();
1596 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1599 // vastart just stores the address of the VarArgsFrameIndex slot into the
1600 // memory location argument.
1601 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1602 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1603 MachinePointerInfo(SV), false, false, 0);
1606 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1607 EVT TyX = Op.getOperand(0).getValueType();
1608 EVT TyY = Op.getOperand(1).getValueType();
1609 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1610 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1611 DebugLoc DL = Op.getDebugLoc();
1614 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1616 SDValue X = (TyX == MVT::f32) ?
1617 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1618 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1620 SDValue Y = (TyY == MVT::f32) ?
1621 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1622 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1626 // ext E, Y, 31, 1 ; extract bit31 of Y
1627 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1628 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1629 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1632 // srl SrlX, SllX, 1
1634 // sll SllY, SrlX, 31
1635 // or Or, SrlX, SllY
1636 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1637 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1638 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1639 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1640 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1643 if (TyX == MVT::f32)
1644 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1646 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1647 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1648 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1651 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1652 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1653 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1654 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1655 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1656 DebugLoc DL = Op.getDebugLoc();
1658 // Bitcast to integer nodes.
1659 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1660 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1663 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1664 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1665 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1666 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
1668 if (WidthX > WidthY)
1669 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1670 else if (WidthY > WidthX)
1671 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1673 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1674 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
1675 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1678 // (d)sll SllX, X, 1
1679 // (d)srl SrlX, SllX, 1
1680 // (d)srl SrlY, Y, width(Y)-1
1681 // (d)sll SllY, SrlX, width(Y)-1
1682 // or Or, SrlX, SllY
1683 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1684 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1685 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1686 DAG.getConstant(WidthY - 1, MVT::i32));
1688 if (WidthX > WidthY)
1689 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1690 else if (WidthY > WidthX)
1691 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1693 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1694 DAG.getConstant(WidthX - 1, MVT::i32));
1695 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
1696 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
1700 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1701 if (Subtarget->hasMips64())
1702 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
1704 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
1707 static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1708 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1709 DebugLoc DL = Op.getDebugLoc();
1711 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1713 SDValue X = (Op.getValueType() == MVT::f32) ?
1714 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1715 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1720 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
1721 DAG.getRegister(Mips::ZERO, MVT::i32),
1722 DAG.getConstant(31, MVT::i32), Const1, X);
1724 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1725 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1728 if (Op.getValueType() == MVT::f32)
1729 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
1731 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1732 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1733 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1736 static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1737 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1738 DebugLoc DL = Op.getDebugLoc();
1740 // Bitcast to integer node.
1741 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
1745 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
1746 DAG.getRegister(Mips::ZERO_64, MVT::i64),
1747 DAG.getConstant(63, MVT::i32), Const1, X);
1749 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
1750 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
1753 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
1757 MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const {
1758 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
1759 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2());
1761 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2());
1764 SDValue MipsTargetLowering::
1765 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1767 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1768 "Frame address can only be determined for current frame.");
1770 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1771 MFI->setFrameAddressIsTaken(true);
1772 EVT VT = Op.getValueType();
1773 DebugLoc DL = Op.getDebugLoc();
1774 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1775 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1779 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
1780 SelectionDAG &DAG) const {
1782 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1783 "Return address can be determined only for current frame.");
1785 MachineFunction &MF = DAG.getMachineFunction();
1786 MachineFrameInfo *MFI = MF.getFrameInfo();
1787 MVT VT = Op.getSimpleValueType();
1788 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
1789 MFI->setReturnAddressIsTaken(true);
1791 // Return RA, which contains the return address. Mark it an implicit live-in.
1792 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
1793 return DAG.getCopyFromReg(DAG.getEntryNode(), Op.getDebugLoc(), Reg, VT);
1796 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
1797 // generated from __builtin_eh_return (offset, handler)
1798 // The effect of this is to adjust the stack pointer by "offset"
1799 // and then branch to "handler".
1800 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
1802 MachineFunction &MF = DAG.getMachineFunction();
1803 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1805 MipsFI->setCallsEhReturn();
1806 SDValue Chain = Op.getOperand(0);
1807 SDValue Offset = Op.getOperand(1);
1808 SDValue Handler = Op.getOperand(2);
1809 DebugLoc DL = Op.getDebugLoc();
1810 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
1812 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
1813 // EH_RETURN nodes, so that instructions are emitted back-to-back.
1814 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1;
1815 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0;
1816 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
1817 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
1818 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
1819 DAG.getRegister(OffsetReg, Ty),
1820 DAG.getRegister(AddrReg, getPointerTy()),
1824 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
1825 SelectionDAG &DAG) const {
1826 // FIXME: Need pseudo-fence for 'singlethread' fences
1827 // FIXME: Set SType for weaker fences where supported/appropriate.
1829 DebugLoc DL = Op.getDebugLoc();
1830 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
1831 DAG.getConstant(SType, MVT::i32));
1834 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
1835 SelectionDAG &DAG) const {
1836 DebugLoc DL = Op.getDebugLoc();
1837 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1838 SDValue Shamt = Op.getOperand(2);
1841 // lo = (shl lo, shamt)
1842 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
1845 // hi = (shl lo, shamt[4:0])
1846 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1847 DAG.getConstant(-1, MVT::i32));
1848 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
1849 DAG.getConstant(1, MVT::i32));
1850 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
1852 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
1853 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1854 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
1855 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1856 DAG.getConstant(0x20, MVT::i32));
1857 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1858 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
1859 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
1861 SDValue Ops[2] = {Lo, Hi};
1862 return DAG.getMergeValues(Ops, 2, DL);
1865 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
1867 DebugLoc DL = Op.getDebugLoc();
1868 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
1869 SDValue Shamt = Op.getOperand(2);
1872 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
1874 // hi = (sra hi, shamt)
1876 // hi = (srl hi, shamt)
1879 // lo = (sra hi, shamt[4:0])
1880 // hi = (sra hi, 31)
1882 // lo = (srl hi, shamt[4:0])
1884 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
1885 DAG.getConstant(-1, MVT::i32));
1886 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
1887 DAG.getConstant(1, MVT::i32));
1888 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
1889 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
1890 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
1891 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1893 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
1894 DAG.getConstant(0x20, MVT::i32));
1895 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
1896 DAG.getConstant(31, MVT::i32));
1897 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
1898 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
1899 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
1902 SDValue Ops[2] = {Lo, Hi};
1903 return DAG.getMergeValues(Ops, 2, DL);
1906 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
1907 SDValue Chain, SDValue Src, unsigned Offset) {
1908 SDValue Ptr = LD->getBasePtr();
1909 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
1910 EVT BasePtrVT = Ptr.getValueType();
1911 DebugLoc DL = LD->getDebugLoc();
1912 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1915 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
1916 DAG.getConstant(Offset, BasePtrVT));
1918 SDValue Ops[] = { Chain, Ptr, Src };
1919 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
1920 LD->getMemOperand());
1923 // Expand an unaligned 32 or 64-bit integer load node.
1924 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1925 LoadSDNode *LD = cast<LoadSDNode>(Op);
1926 EVT MemVT = LD->getMemoryVT();
1928 // Return if load is aligned or if MemVT is neither i32 nor i64.
1929 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
1930 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
1933 bool IsLittle = Subtarget->isLittle();
1934 EVT VT = Op.getValueType();
1935 ISD::LoadExtType ExtType = LD->getExtensionType();
1936 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
1938 assert((VT == MVT::i32) || (VT == MVT::i64));
1941 // (set dst, (i64 (load baseptr)))
1943 // (set tmp, (ldl (add baseptr, 7), undef))
1944 // (set dst, (ldr baseptr, tmp))
1945 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
1946 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
1948 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
1952 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
1954 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
1958 // (set dst, (i32 (load baseptr))) or
1959 // (set dst, (i64 (sextload baseptr))) or
1960 // (set dst, (i64 (extload baseptr)))
1962 // (set tmp, (lwl (add baseptr, 3), undef))
1963 // (set dst, (lwr baseptr, tmp))
1964 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
1965 (ExtType == ISD::EXTLOAD))
1968 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
1971 // (set dst, (i64 (zextload baseptr)))
1973 // (set tmp0, (lwl (add baseptr, 3), undef))
1974 // (set tmp1, (lwr baseptr, tmp0))
1975 // (set tmp2, (shl tmp1, 32))
1976 // (set dst, (srl tmp2, 32))
1977 DebugLoc DL = LD->getDebugLoc();
1978 SDValue Const32 = DAG.getConstant(32, MVT::i32);
1979 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
1980 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
1981 SDValue Ops[] = { SRL, LWR.getValue(1) };
1982 return DAG.getMergeValues(Ops, 2, DL);
1985 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
1986 SDValue Chain, unsigned Offset) {
1987 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
1988 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
1989 DebugLoc DL = SD->getDebugLoc();
1990 SDVTList VTList = DAG.getVTList(MVT::Other);
1993 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
1994 DAG.getConstant(Offset, BasePtrVT));
1996 SDValue Ops[] = { Chain, Value, Ptr };
1997 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
1998 SD->getMemOperand());
2001 // Expand an unaligned 32 or 64-bit integer store node.
2002 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2003 StoreSDNode *SD = cast<StoreSDNode>(Op);
2004 EVT MemVT = SD->getMemoryVT();
2006 // Return if store is aligned or if MemVT is neither i32 nor i64.
2007 if ((SD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2008 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2011 bool IsLittle = Subtarget->isLittle();
2012 SDValue Value = SD->getValue(), Chain = SD->getChain();
2013 EVT VT = Value.getValueType();
2016 // (store val, baseptr) or
2017 // (truncstore val, baseptr)
2019 // (swl val, (add baseptr, 3))
2020 // (swr val, baseptr)
2021 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2022 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2024 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2027 assert(VT == MVT::i64);
2030 // (store val, baseptr)
2032 // (sdl val, (add baseptr, 7))
2033 // (sdr val, baseptr)
2034 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2035 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2038 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2039 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2040 || cast<ConstantSDNode>
2041 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2042 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2046 // (add (frameaddr 0), (frame_to_args_offset))
2047 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2048 // (add FrameObject, 0)
2049 // where FrameObject is a fixed StackObject with offset 0 which points to
2050 // the old stack pointer.
2051 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2052 EVT ValTy = Op->getValueType(0);
2053 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2054 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2055 return DAG.getNode(ISD::ADD, Op->getDebugLoc(), ValTy, InArgsAddr,
2056 DAG.getConstant(0, ValTy));
2059 //===----------------------------------------------------------------------===//
2060 // Calling Convention Implementation
2061 //===----------------------------------------------------------------------===//
2063 //===----------------------------------------------------------------------===//
2064 // TODO: Implement a generic logic using tblgen that can support this.
2065 // Mips O32 ABI rules:
2067 // i32 - Passed in A0, A1, A2, A3 and stack
2068 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2069 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2070 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2071 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2072 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2075 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2076 //===----------------------------------------------------------------------===//
2078 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2079 MVT LocVT, CCValAssign::LocInfo LocInfo,
2080 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2082 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2084 static const uint16_t IntRegs[] = {
2085 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2087 static const uint16_t F32Regs[] = {
2088 Mips::F12, Mips::F14
2090 static const uint16_t F64Regs[] = {
2094 // Do not process byval args here.
2095 if (ArgFlags.isByVal())
2098 // Promote i8 and i16
2099 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2101 if (ArgFlags.isSExt())
2102 LocInfo = CCValAssign::SExt;
2103 else if (ArgFlags.isZExt())
2104 LocInfo = CCValAssign::ZExt;
2106 LocInfo = CCValAssign::AExt;
2111 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2112 // is true: function is vararg, argument is 3rd or higher, there is previous
2113 // argument which is not f32 or f64.
2114 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2115 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2116 unsigned OrigAlign = ArgFlags.getOrigAlign();
2117 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2119 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2120 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2121 // If this is the first part of an i64 arg,
2122 // the allocated register must be either A0 or A2.
2123 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2124 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2126 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2127 // Allocate int register and shadow next int register. If first
2128 // available register is Mips::A1 or Mips::A3, shadow it too.
2129 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2130 if (Reg == Mips::A1 || Reg == Mips::A3)
2131 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2132 State.AllocateReg(IntRegs, IntRegsSize);
2134 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2135 // we are guaranteed to find an available float register
2136 if (ValVT == MVT::f32) {
2137 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2138 // Shadow int register
2139 State.AllocateReg(IntRegs, IntRegsSize);
2141 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2142 // Shadow int registers
2143 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2144 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2145 State.AllocateReg(IntRegs, IntRegsSize);
2146 State.AllocateReg(IntRegs, IntRegsSize);
2149 llvm_unreachable("Cannot handle this ValVT.");
2152 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2154 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2156 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2161 #include "MipsGenCallingConv.inc"
2163 //===----------------------------------------------------------------------===//
2164 // Call Calling Convention Implementation
2165 //===----------------------------------------------------------------------===//
2167 static const unsigned O32IntRegsSize = 4;
2169 // Return next O32 integer argument register.
2170 static unsigned getNextIntArgReg(unsigned Reg) {
2171 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2172 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2176 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2177 SDValue Chain, SDValue Arg, DebugLoc DL,
2178 bool IsTailCall, SelectionDAG &DAG) const {
2180 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2181 DAG.getIntPtrConstant(Offset));
2182 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2186 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2187 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2188 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2189 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2190 /*isVolatile=*/ true, false, 0);
2193 void MipsTargetLowering::
2194 getOpndList(SmallVectorImpl<SDValue> &Ops,
2195 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2196 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2197 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
2198 // Insert node "GP copy globalreg" before call to function.
2200 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2201 // in PIC mode) allow symbols to be resolved via lazy binding.
2202 // The lazy binding stub requires GP to point to the GOT.
2203 if (IsPICCall && !InternalLinkage) {
2204 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
2205 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2206 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2209 // Build a sequence of copy-to-reg nodes chained together with token
2210 // chain and flag operands which copy the outgoing args into registers.
2211 // The InFlag in necessary since all emitted instructions must be
2215 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2216 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2217 RegsToPass[i].second, InFlag);
2218 InFlag = Chain.getValue(1);
2221 // Add argument registers to the end of the list so that they are
2222 // known live into the call.
2223 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2224 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2225 RegsToPass[i].second.getValueType()));
2227 // Add a register mask operand representing the call-preserved registers.
2228 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2229 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv);
2230 assert(Mask && "Missing call preserved mask for calling convention");
2231 if (Subtarget->inMips16HardFloat()) {
2232 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2233 llvm::StringRef Sym = G->getGlobal()->getName();
2234 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2235 if (F->hasFnAttribute("__Mips16RetHelper")) {
2236 Mask = MipsRegisterInfo::getMips16RetHelperMask();
2240 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2242 if (InFlag.getNode())
2243 Ops.push_back(InFlag);
2246 /// LowerCall - functions arguments are copied from virtual regs to
2247 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2249 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2250 SmallVectorImpl<SDValue> &InVals) const {
2251 SelectionDAG &DAG = CLI.DAG;
2252 DebugLoc &DL = CLI.DL;
2253 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2254 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2255 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2256 SDValue Chain = CLI.Chain;
2257 SDValue Callee = CLI.Callee;
2258 bool &IsTailCall = CLI.IsTailCall;
2259 CallingConv::ID CallConv = CLI.CallConv;
2260 bool IsVarArg = CLI.IsVarArg;
2262 MachineFunction &MF = DAG.getMachineFunction();
2263 MachineFrameInfo *MFI = MF.getFrameInfo();
2264 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2265 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2267 // Analyze operands of the call, assigning locations to each operand.
2268 SmallVector<CCValAssign, 16> ArgLocs;
2269 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2270 getTargetMachine(), ArgLocs, *DAG.getContext());
2271 MipsCC::SpecialCallingConvType SpecialCallingConv =
2272 getSpecialCallingConv(Callee);
2273 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo, SpecialCallingConv);
2275 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg,
2276 getTargetMachine().Options.UseSoftFloat,
2277 Callee.getNode(), CLI.Args);
2279 // Get a count of how many bytes are to be pushed on the stack.
2280 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2282 // Check if it's really possible to do a tail call.
2285 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2286 *MF.getInfo<MipsFunctionInfo>());
2291 // Chain is the output chain of the last Load/Store or CopyToReg node.
2292 // ByValChain is the output chain of the last Memcpy node created for copying
2293 // byval arguments to the stack.
2294 unsigned StackAlignment = TFL->getStackAlignment();
2295 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2296 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2299 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal);
2301 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL,
2302 IsN64 ? Mips::SP_64 : Mips::SP,
2305 // With EABI is it possible to have 16 args on registers.
2306 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2307 SmallVector<SDValue, 8> MemOpChains;
2308 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2310 // Walk the register/memloc assignments, inserting copies/loads.
2311 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2312 SDValue Arg = OutVals[i];
2313 CCValAssign &VA = ArgLocs[i];
2314 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2315 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2318 if (Flags.isByVal()) {
2319 assert(Flags.getByValSize() &&
2320 "ByVal args of size 0 should have been ignored by front-end.");
2321 assert(ByValArg != MipsCCInfo.byval_end());
2322 assert(!IsTailCall &&
2323 "Do not tail-call optimize if there is a byval argument.");
2324 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2325 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2330 // Promote the value if needed.
2331 switch (VA.getLocInfo()) {
2332 default: llvm_unreachable("Unknown loc info!");
2333 case CCValAssign::Full:
2334 if (VA.isRegLoc()) {
2335 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2336 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2337 (ValVT == MVT::i64 && LocVT == MVT::f64))
2338 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2339 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2340 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2341 Arg, DAG.getConstant(0, MVT::i32));
2342 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2343 Arg, DAG.getConstant(1, MVT::i32));
2344 if (!Subtarget->isLittle())
2346 unsigned LocRegLo = VA.getLocReg();
2347 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2348 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2349 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2354 case CCValAssign::SExt:
2355 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2357 case CCValAssign::ZExt:
2358 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2360 case CCValAssign::AExt:
2361 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2365 // Arguments that can be passed on register must be kept at
2366 // RegsToPass vector
2367 if (VA.isRegLoc()) {
2368 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2372 // Register can't get to this point...
2373 assert(VA.isMemLoc());
2375 // emit ISD::STORE whichs stores the
2376 // parameter value to a stack Location
2377 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2378 Chain, Arg, DL, IsTailCall, DAG));
2381 // Transform all store nodes into one single node because all store
2382 // nodes are independent of each other.
2383 if (!MemOpChains.empty())
2384 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2385 &MemOpChains[0], MemOpChains.size());
2387 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2388 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2389 // node so that legalize doesn't hack it.
2390 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2391 bool GlobalOrExternal = false, InternalLinkage = false;
2394 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2396 InternalLinkage = G->getGlobal()->hasInternalLinkage();
2398 if (InternalLinkage)
2399 Callee = getAddrLocal(Callee, DAG, HasMips64);
2401 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2402 MipsII::MO_CALL_LO16);
2404 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2406 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0,
2407 MipsII::MO_NO_FLAG);
2408 GlobalOrExternal = true;
2410 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2411 if (!IsN64 && !IsPIC) // !N64 && static
2412 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2413 MipsII::MO_NO_FLAG);
2415 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2416 MipsII::MO_CALL_LO16);
2418 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2420 GlobalOrExternal = true;
2423 SmallVector<SDValue, 8> Ops(1, Chain);
2424 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2426 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2427 CLI, Callee, Chain);
2430 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size());
2432 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size());
2433 SDValue InFlag = Chain.getValue(1);
2435 // Create the CALLSEQ_END node.
2436 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2437 DAG.getIntPtrConstant(0, true), InFlag);
2438 InFlag = Chain.getValue(1);
2440 // Handle result values, copying them out of physregs into vregs that we
2442 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg,
2443 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy);
2446 /// LowerCallResult - Lower the result values of a call into the
2447 /// appropriate copies out of appropriate physical registers.
2449 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2450 CallingConv::ID CallConv, bool IsVarArg,
2451 const SmallVectorImpl<ISD::InputArg> &Ins,
2452 DebugLoc DL, SelectionDAG &DAG,
2453 SmallVectorImpl<SDValue> &InVals,
2454 const SDNode *CallNode,
2455 const Type *RetTy) const {
2456 // Assign locations to each value returned by this call.
2457 SmallVector<CCValAssign, 16> RVLocs;
2458 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2459 getTargetMachine(), RVLocs, *DAG.getContext());
2460 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2462 MipsCCInfo.analyzeCallResult(Ins, getTargetMachine().Options.UseSoftFloat,
2465 // Copy all of the result registers out of their specified physreg.
2466 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2467 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2468 RVLocs[i].getLocVT(), InFlag);
2469 Chain = Val.getValue(1);
2470 InFlag = Val.getValue(2);
2472 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2473 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val);
2475 InVals.push_back(Val);
2481 //===----------------------------------------------------------------------===//
2482 // Formal Arguments Calling Convention Implementation
2483 //===----------------------------------------------------------------------===//
2484 /// LowerFormalArguments - transform physical registers into virtual registers
2485 /// and generate load operations for arguments places on the stack.
2487 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2488 CallingConv::ID CallConv,
2490 const SmallVectorImpl<ISD::InputArg> &Ins,
2491 DebugLoc DL, SelectionDAG &DAG,
2492 SmallVectorImpl<SDValue> &InVals)
2494 MachineFunction &MF = DAG.getMachineFunction();
2495 MachineFrameInfo *MFI = MF.getFrameInfo();
2496 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2498 MipsFI->setVarArgsFrameIndex(0);
2500 // Used with vargs to acumulate store chains.
2501 std::vector<SDValue> OutChains;
2503 // Assign locations to all of the incoming arguments.
2504 SmallVector<CCValAssign, 16> ArgLocs;
2505 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2506 getTargetMachine(), ArgLocs, *DAG.getContext());
2507 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2508 Function::const_arg_iterator FuncArg =
2509 DAG.getMachineFunction().getFunction()->arg_begin();
2510 bool UseSoftFloat = getTargetMachine().Options.UseSoftFloat;
2512 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg);
2513 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2514 MipsCCInfo.hasByValArg());
2516 unsigned CurArgIdx = 0;
2517 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2519 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2520 CCValAssign &VA = ArgLocs[i];
2521 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
2522 CurArgIdx = Ins[i].OrigArgIndex;
2523 EVT ValVT = VA.getValVT();
2524 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2525 bool IsRegLoc = VA.isRegLoc();
2527 if (Flags.isByVal()) {
2528 assert(Flags.getByValSize() &&
2529 "ByVal args of size 0 should have been ignored by front-end.");
2530 assert(ByValArg != MipsCCInfo.byval_end());
2531 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2532 MipsCCInfo, *ByValArg);
2537 // Arguments stored on registers
2539 EVT RegVT = VA.getLocVT();
2540 unsigned ArgReg = VA.getLocReg();
2541 const TargetRegisterClass *RC;
2543 if (RegVT == MVT::i32)
2544 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass :
2545 &Mips::CPURegsRegClass;
2546 else if (RegVT == MVT::i64)
2547 RC = &Mips::CPU64RegsRegClass;
2548 else if (RegVT == MVT::f32)
2549 RC = &Mips::FGR32RegClass;
2550 else if (RegVT == MVT::f64)
2551 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass;
2553 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2555 // Transform the arguments stored on
2556 // physical registers into virtual ones
2557 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2558 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2560 // If this is an 8 or 16-bit value, it has been passed promoted
2561 // to 32 bits. Insert an assert[sz]ext to capture this, then
2562 // truncate to the right size.
2563 if (VA.getLocInfo() != CCValAssign::Full) {
2564 unsigned Opcode = 0;
2565 if (VA.getLocInfo() == CCValAssign::SExt)
2566 Opcode = ISD::AssertSext;
2567 else if (VA.getLocInfo() == CCValAssign::ZExt)
2568 Opcode = ISD::AssertZext;
2570 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue,
2571 DAG.getValueType(ValVT));
2572 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue);
2575 // Handle floating point arguments passed in integer registers and
2576 // long double arguments passed in floating point registers.
2577 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2578 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
2579 (RegVT == MVT::f64 && ValVT == MVT::i64))
2580 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
2581 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2582 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
2583 getNextIntArgReg(ArgReg), RC);
2584 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
2585 if (!Subtarget->isLittle())
2586 std::swap(ArgValue, ArgValue2);
2587 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
2588 ArgValue, ArgValue2);
2591 InVals.push_back(ArgValue);
2592 } else { // VA.isRegLoc()
2595 assert(VA.isMemLoc());
2597 // The stack pointer offset is relative to the caller stack frame.
2598 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2599 VA.getLocMemOffset(), true);
2601 // Create load nodes to retrieve arguments from the stack
2602 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2603 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN,
2604 MachinePointerInfo::getFixedStack(FI),
2605 false, false, false, 0));
2609 // The mips ABIs for returning structs by value requires that we copy
2610 // the sret argument into $v0 for the return. Save the argument into
2611 // a virtual register so that we can access it from the return points.
2612 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2613 unsigned Reg = MipsFI->getSRetReturnReg();
2615 Reg = MF.getRegInfo().
2616 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
2617 MipsFI->setSRetReturnReg(Reg);
2619 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]);
2620 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
2624 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG);
2626 // All stores are grouped in one node to allow the matching between
2627 // the size of Ins and InVals. This only happens when on varg functions
2628 if (!OutChains.empty()) {
2629 OutChains.push_back(Chain);
2630 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2631 &OutChains[0], OutChains.size());
2637 //===----------------------------------------------------------------------===//
2638 // Return Value Calling Convention Implementation
2639 //===----------------------------------------------------------------------===//
2642 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2643 MachineFunction &MF, bool IsVarArg,
2644 const SmallVectorImpl<ISD::OutputArg> &Outs,
2645 LLVMContext &Context) const {
2646 SmallVector<CCValAssign, 16> RVLocs;
2647 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(),
2649 return CCInfo.CheckReturn(Outs, RetCC_Mips);
2653 MipsTargetLowering::LowerReturn(SDValue Chain,
2654 CallingConv::ID CallConv, bool IsVarArg,
2655 const SmallVectorImpl<ISD::OutputArg> &Outs,
2656 const SmallVectorImpl<SDValue> &OutVals,
2657 DebugLoc DL, SelectionDAG &DAG) const {
2658 // CCValAssign - represent the assignment of
2659 // the return value to a location
2660 SmallVector<CCValAssign, 16> RVLocs;
2661 MachineFunction &MF = DAG.getMachineFunction();
2663 // CCState - Info about the registers and stack slot.
2664 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs,
2666 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2668 // Analyze return values.
2669 MipsCCInfo.analyzeReturn(Outs, getTargetMachine().Options.UseSoftFloat,
2670 MF.getFunction()->getReturnType());
2673 SmallVector<SDValue, 4> RetOps(1, Chain);
2675 // Copy the result values into the output registers.
2676 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2677 SDValue Val = OutVals[i];
2678 CCValAssign &VA = RVLocs[i];
2679 assert(VA.isRegLoc() && "Can only return in registers!");
2681 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2682 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val);
2684 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
2686 // Guarantee that all emitted copies are stuck together with flags.
2687 Flag = Chain.getValue(1);
2688 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2691 // The mips ABIs for returning structs by value requires that we copy
2692 // the sret argument into $v0 for the return. We saved the argument into
2693 // a virtual register in the entry block, so now we copy the value out
2695 if (MF.getFunction()->hasStructRetAttr()) {
2696 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2697 unsigned Reg = MipsFI->getSRetReturnReg();
2700 llvm_unreachable("sret virtual register not created in the entry block");
2701 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
2702 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
2704 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
2705 Flag = Chain.getValue(1);
2706 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
2709 RetOps[0] = Chain; // Update chain.
2711 // Add the flag if we have it.
2713 RetOps.push_back(Flag);
2715 // Return on Mips is always a "jr $ra"
2716 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size());
2719 //===----------------------------------------------------------------------===//
2720 // Mips Inline Assembly Support
2721 //===----------------------------------------------------------------------===//
2723 /// getConstraintType - Given a constraint letter, return the type of
2724 /// constraint it is for this target.
2725 MipsTargetLowering::ConstraintType MipsTargetLowering::
2726 getConstraintType(const std::string &Constraint) const
2728 // Mips specific constrainy
2729 // GCC config/mips/constraints.md
2731 // 'd' : An address register. Equivalent to r
2732 // unless generating MIPS16 code.
2733 // 'y' : Equivalent to r; retained for
2734 // backwards compatibility.
2735 // 'c' : A register suitable for use in an indirect
2736 // jump. This will always be $25 for -mabicalls.
2737 // 'l' : The lo register. 1 word storage.
2738 // 'x' : The hilo register pair. Double word storage.
2739 if (Constraint.size() == 1) {
2740 switch (Constraint[0]) {
2748 return C_RegisterClass;
2753 return TargetLowering::getConstraintType(Constraint);
2756 /// Examine constraint type and operand type and determine a weight value.
2757 /// This object must already have been set up with the operand type
2758 /// and the current alternative constraint selected.
2759 TargetLowering::ConstraintWeight
2760 MipsTargetLowering::getSingleConstraintMatchWeight(
2761 AsmOperandInfo &info, const char *constraint) const {
2762 ConstraintWeight weight = CW_Invalid;
2763 Value *CallOperandVal = info.CallOperandVal;
2764 // If we don't have a value, we can't do a match,
2765 // but allow it at the lowest weight.
2766 if (CallOperandVal == NULL)
2768 Type *type = CallOperandVal->getType();
2769 // Look at the constraint type.
2770 switch (*constraint) {
2772 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2776 if (type->isIntegerTy())
2777 weight = CW_Register;
2780 if (type->isFloatTy())
2781 weight = CW_Register;
2783 case 'c': // $25 for indirect jumps
2784 case 'l': // lo register
2785 case 'x': // hilo register pair
2786 if (type->isIntegerTy())
2787 weight = CW_SpecificReg;
2789 case 'I': // signed 16 bit immediate
2790 case 'J': // integer zero
2791 case 'K': // unsigned 16 bit immediate
2792 case 'L': // signed 32 bit immediate where lower 16 bits are 0
2793 case 'N': // immediate in the range of -65535 to -1 (inclusive)
2794 case 'O': // signed 15 bit immediate (+- 16383)
2795 case 'P': // immediate in the range of 65535 to 1 (inclusive)
2796 if (isa<ConstantInt>(CallOperandVal))
2797 weight = CW_Constant;
2806 /// Given a register class constraint, like 'r', if this corresponds directly
2807 /// to an LLVM register class, return a register of 0 and the register class
2809 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2810 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
2812 if (Constraint.size() == 1) {
2813 switch (Constraint[0]) {
2814 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2815 case 'y': // Same as 'r'. Exists for compatibility.
2817 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
2818 if (Subtarget->inMips16Mode())
2819 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
2820 return std::make_pair(0U, &Mips::CPURegsRegClass);
2822 if (VT == MVT::i64 && !HasMips64)
2823 return std::make_pair(0U, &Mips::CPURegsRegClass);
2824 if (VT == MVT::i64 && HasMips64)
2825 return std::make_pair(0U, &Mips::CPU64RegsRegClass);
2826 // This will generate an error message
2827 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2830 return std::make_pair(0U, &Mips::FGR32RegClass);
2831 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
2832 if (Subtarget->isFP64bit())
2833 return std::make_pair(0U, &Mips::FGR64RegClass);
2834 return std::make_pair(0U, &Mips::AFGR64RegClass);
2837 case 'c': // register suitable for indirect jump
2839 return std::make_pair((unsigned)Mips::T9, &Mips::CPURegsRegClass);
2840 assert(VT == MVT::i64 && "Unexpected type.");
2841 return std::make_pair((unsigned)Mips::T9_64, &Mips::CPU64RegsRegClass);
2842 case 'l': // register suitable for indirect jump
2844 return std::make_pair((unsigned)Mips::LO, &Mips::LORegsRegClass);
2845 return std::make_pair((unsigned)Mips::LO64, &Mips::LORegs64RegClass);
2846 case 'x': // register suitable for indirect jump
2847 // Fixme: Not triggering the use of both hi and low
2848 // This will generate an error message
2849 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2852 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2855 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2856 /// vector. If it is invalid, don't add anything to Ops.
2857 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2858 std::string &Constraint,
2859 std::vector<SDValue>&Ops,
2860 SelectionDAG &DAG) const {
2861 SDValue Result(0, 0);
2863 // Only support length 1 constraints for now.
2864 if (Constraint.length() > 1) return;
2866 char ConstraintLetter = Constraint[0];
2867 switch (ConstraintLetter) {
2868 default: break; // This will fall through to the generic implementation
2869 case 'I': // Signed 16 bit constant
2870 // If this fails, the parent routine will give an error
2871 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2872 EVT Type = Op.getValueType();
2873 int64_t Val = C->getSExtValue();
2874 if (isInt<16>(Val)) {
2875 Result = DAG.getTargetConstant(Val, Type);
2880 case 'J': // integer zero
2881 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2882 EVT Type = Op.getValueType();
2883 int64_t Val = C->getZExtValue();
2885 Result = DAG.getTargetConstant(0, Type);
2890 case 'K': // unsigned 16 bit immediate
2891 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2892 EVT Type = Op.getValueType();
2893 uint64_t Val = (uint64_t)C->getZExtValue();
2894 if (isUInt<16>(Val)) {
2895 Result = DAG.getTargetConstant(Val, Type);
2900 case 'L': // signed 32 bit immediate where lower 16 bits are 0
2901 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2902 EVT Type = Op.getValueType();
2903 int64_t Val = C->getSExtValue();
2904 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
2905 Result = DAG.getTargetConstant(Val, Type);
2910 case 'N': // immediate in the range of -65535 to -1 (inclusive)
2911 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2912 EVT Type = Op.getValueType();
2913 int64_t Val = C->getSExtValue();
2914 if ((Val >= -65535) && (Val <= -1)) {
2915 Result = DAG.getTargetConstant(Val, Type);
2920 case 'O': // signed 15 bit immediate
2921 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2922 EVT Type = Op.getValueType();
2923 int64_t Val = C->getSExtValue();
2924 if ((isInt<15>(Val))) {
2925 Result = DAG.getTargetConstant(Val, Type);
2930 case 'P': // immediate in the range of 1 to 65535 (inclusive)
2931 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2932 EVT Type = Op.getValueType();
2933 int64_t Val = C->getSExtValue();
2934 if ((Val <= 65535) && (Val >= 1)) {
2935 Result = DAG.getTargetConstant(Val, Type);
2942 if (Result.getNode()) {
2943 Ops.push_back(Result);
2947 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2951 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
2952 // No global is ever allowed as a base.
2957 case 0: // "r+i" or just "i", depending on HasBaseReg.
2960 if (!AM.HasBaseReg) // allow "r+i".
2962 return false; // disallow "r+r" or "r+r+i".
2971 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2972 // The Mips target isn't yet aware of offsets.
2976 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
2978 bool IsMemset, bool ZeroMemset,
2980 MachineFunction &MF) const {
2981 if (Subtarget->hasMips64())
2987 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2988 if (VT != MVT::f32 && VT != MVT::f64)
2990 if (Imm.isNegZero())
2992 return Imm.isZero();
2995 unsigned MipsTargetLowering::getJumpTableEncoding() const {
2997 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
2999 return TargetLowering::getJumpTableEncoding();
3002 /// This function returns true if CallSym is a long double emulation routine.
3003 static bool isF128SoftLibCall(const char *CallSym) {
3004 const char *const LibCalls[] =
3005 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2",
3006 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi",
3007 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf",
3008 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2",
3009 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3",
3010 "__trunctfdf2", "__trunctfsf2", "__unordtf2",
3011 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl",
3012 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl",
3015 const char * const *End = LibCalls + array_lengthof(LibCalls);
3017 // Check that LibCalls is sorted alphabetically.
3018 MipsTargetLowering::LTStr Comp;
3021 for (const char * const *I = LibCalls; I < End - 1; ++I)
3022 assert(Comp(*I, *(I + 1)));
3025 return std::binary_search(LibCalls, End, CallSym, Comp);
3028 /// This function returns true if Ty is fp128 or i128 which was originally a
3030 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) {
3031 if (Ty->isFP128Ty())
3034 const ExternalSymbolSDNode *ES =
3035 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
3037 // If the Ty is i128 and the function being called is a long double emulation
3038 // routine, then the original type is f128.
3039 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
3042 MipsTargetLowering::MipsCC::SpecialCallingConvType
3043 MipsTargetLowering::getSpecialCallingConv(SDValue Callee) const {
3044 MipsCC::SpecialCallingConvType SpecialCallingConv =
3045 MipsCC::NoSpecialCallingConv;;
3046 if (Subtarget->inMips16HardFloat()) {
3047 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3048 llvm::StringRef Sym = G->getGlobal()->getName();
3049 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
3050 if (F->hasFnAttribute("__Mips16RetHelper")) {
3051 SpecialCallingConv = MipsCC::Mips16RetHelperConv;
3055 return SpecialCallingConv;
3058 MipsTargetLowering::MipsCC::MipsCC(
3059 CallingConv::ID CC, bool IsO32_, CCState &Info,
3060 MipsCC::SpecialCallingConvType SpecialCallingConv_)
3061 : CCInfo(Info), CallConv(CC), IsO32(IsO32_),
3062 SpecialCallingConv(SpecialCallingConv_){
3063 // Pre-allocate reserved argument area.
3064 CCInfo.AllocateStack(reservedArgArea(), 1);
3068 void MipsTargetLowering::MipsCC::
3069 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
3070 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode,
3071 std::vector<ArgListEntry> &FuncArgs) {
3072 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
3073 "CallingConv::Fast shouldn't be used for vararg functions.");
3075 unsigned NumOpnds = Args.size();
3076 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
3078 for (unsigned I = 0; I != NumOpnds; ++I) {
3079 MVT ArgVT = Args[I].VT;
3080 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3083 if (ArgFlags.isByVal()) {
3084 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3088 if (IsVarArg && !Args[I].IsFixed)
3089 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3091 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode,
3093 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo);
3098 dbgs() << "Call operand #" << I << " has unhandled type "
3099 << EVT(ArgVT).getEVTString();
3101 llvm_unreachable(0);
3106 void MipsTargetLowering::MipsCC::
3107 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args,
3108 bool IsSoftFloat, Function::const_arg_iterator FuncArg) {
3109 unsigned NumArgs = Args.size();
3110 llvm::CCAssignFn *FixedFn = fixedArgFn();
3111 unsigned CurArgIdx = 0;
3113 for (unsigned I = 0; I != NumArgs; ++I) {
3114 MVT ArgVT = Args[I].VT;
3115 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3116 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx);
3117 CurArgIdx = Args[I].OrigArgIndex;
3119 if (ArgFlags.isByVal()) {
3120 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3124 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat);
3126 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo))
3130 dbgs() << "Formal Arg #" << I << " has unhandled type "
3131 << EVT(ArgVT).getEVTString();
3133 llvm_unreachable(0);
3137 template<typename Ty>
3138 void MipsTargetLowering::MipsCC::
3139 analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat,
3140 const SDNode *CallNode, const Type *RetTy) const {
3143 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode))
3144 Fn = RetCC_F128Soft;
3148 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) {
3149 MVT VT = RetVals[I].VT;
3150 ISD::ArgFlagsTy Flags = RetVals[I].Flags;
3151 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat);
3153 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) {
3155 dbgs() << "Call result #" << I << " has unhandled type "
3156 << EVT(VT).getEVTString() << '\n';
3158 llvm_unreachable(0);
3163 void MipsTargetLowering::MipsCC::
3164 analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat,
3165 const SDNode *CallNode, const Type *RetTy) const {
3166 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy);
3169 void MipsTargetLowering::MipsCC::
3170 analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat,
3171 const Type *RetTy) const {
3172 analyzeReturn(Outs, IsSoftFloat, 0, RetTy);
3176 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3178 CCValAssign::LocInfo LocInfo,
3179 ISD::ArgFlagsTy ArgFlags) {
3180 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3182 struct ByValArgInfo ByVal;
3183 unsigned RegSize = regSize();
3184 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3185 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3188 if (useRegsForByval())
3189 allocateRegs(ByVal, ByValSize, Align);
3191 // Allocate space on caller's stack.
3192 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3194 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3196 ByValArgs.push_back(ByVal);
3199 unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const {
3200 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs);
3203 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
3204 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0;
3207 const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const {
3208 return IsO32 ? O32IntRegs : Mips64IntRegs;
3211 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
3212 if (CallConv == CallingConv::Fast)
3213 return CC_Mips_FastCC;
3215 if (SpecialCallingConv == Mips16RetHelperConv)
3216 return CC_Mips16RetHelper;
3217 return IsO32 ? CC_MipsO32 : CC_MipsN;
3220 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
3221 return IsO32 ? CC_MipsO32 : CC_MipsN_VarArg;
3224 const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const {
3225 return IsO32 ? O32IntRegs : Mips64DPRegs;
3228 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3231 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs();
3232 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs();
3233 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3234 "Byval argument's size and alignment should be a multiple of"
3237 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3239 // If Align > RegSize, the first arg register must be even.
3240 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3241 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3245 // Mark the registers allocated.
3246 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3247 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3248 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3251 MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy,
3252 const SDNode *CallNode,
3253 bool IsSoftFloat) const {
3254 if (IsSoftFloat || IsO32)
3257 // Check if the original type was fp128.
3258 if (originalTypeIsF128(OrigTy, CallNode)) {
3259 assert(VT == MVT::i64);
3266 void MipsTargetLowering::
3267 copyByValRegs(SDValue Chain, DebugLoc DL, std::vector<SDValue> &OutChains,
3268 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3269 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3270 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3271 MachineFunction &MF = DAG.getMachineFunction();
3272 MachineFrameInfo *MFI = MF.getFrameInfo();
3273 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3274 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3278 FrameObjOffset = (int)CC.reservedArgArea() -
3279 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3281 FrameObjOffset = ByVal.Address;
3283 // Create frame object.
3284 EVT PtrTy = getPointerTy();
3285 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3286 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3287 InVals.push_back(FIN);
3292 // Copy arg registers.
3293 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3294 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3296 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3297 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3298 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3299 unsigned Offset = I * CC.regSize();
3300 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3301 DAG.getConstant(Offset, PtrTy));
3302 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3303 StorePtr, MachinePointerInfo(FuncArg, Offset),
3305 OutChains.push_back(Store);
3309 // Copy byVal arg to registers and stack.
3310 void MipsTargetLowering::
3311 passByValArg(SDValue Chain, DebugLoc DL,
3312 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
3313 SmallVector<SDValue, 8> &MemOpChains, SDValue StackPtr,
3314 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3315 const MipsCC &CC, const ByValArgInfo &ByVal,
3316 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3317 unsigned ByValSize = Flags.getByValSize();
3318 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
3319 unsigned RegSize = CC.regSize();
3320 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
3321 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
3323 if (ByVal.NumRegs) {
3324 const uint16_t *ArgRegs = CC.intArgRegs();
3325 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
3328 // Copy words to registers.
3329 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
3330 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3331 DAG.getConstant(Offset, PtrTy));
3332 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3333 MachinePointerInfo(), false, false, false,
3335 MemOpChains.push_back(LoadVal.getValue(1));
3336 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3337 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3340 // Return if the struct has been fully copied.
3341 if (ByValSize == Offset)
3344 // Copy the remainder of the byval argument with sub-word loads and shifts.
3345 if (LeftoverBytes) {
3346 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
3347 "Size of the remainder should be smaller than RegSize.");
3350 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
3351 Offset < ByValSize; LoadSize /= 2) {
3352 unsigned RemSize = ByValSize - Offset;
3354 if (RemSize < LoadSize)
3358 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3359 DAG.getConstant(Offset, PtrTy));
3361 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
3362 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
3363 false, false, Alignment);
3364 MemOpChains.push_back(LoadVal.getValue(1));
3366 // Shift the loaded value.
3370 Shamt = TotalSizeLoaded;
3372 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
3374 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3375 DAG.getConstant(Shamt, MVT::i32));
3378 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3383 TotalSizeLoaded += LoadSize;
3384 Alignment = std::min(Alignment, LoadSize);
3387 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3388 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3393 // Copy remainder of byval arg to it with memcpy.
3394 unsigned MemCpySize = ByValSize - Offset;
3395 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3396 DAG.getConstant(Offset, PtrTy));
3397 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3398 DAG.getIntPtrConstant(ByVal.Address));
3399 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3400 DAG.getConstant(MemCpySize, PtrTy), Alignment,
3401 /*isVolatile=*/false, /*AlwaysInline=*/false,
3402 MachinePointerInfo(0), MachinePointerInfo(0));
3403 MemOpChains.push_back(Chain);
3407 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3408 const MipsCC &CC, SDValue Chain,
3409 DebugLoc DL, SelectionDAG &DAG) const {
3410 unsigned NumRegs = CC.numIntArgRegs();
3411 const uint16_t *ArgRegs = CC.intArgRegs();
3412 const CCState &CCInfo = CC.getCCInfo();
3413 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3414 unsigned RegSize = CC.regSize();
3415 MVT RegTy = MVT::getIntegerVT(RegSize * 8);
3416 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3417 MachineFunction &MF = DAG.getMachineFunction();
3418 MachineFrameInfo *MFI = MF.getFrameInfo();
3419 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3421 // Offset of the first variable argument from stack pointer.
3425 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3428 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3430 // Record the frame index of the first variable argument
3431 // which is a value necessary to VASTART.
3432 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3433 MipsFI->setVarArgsFrameIndex(FI);
3435 // Copy the integer registers that have not been used for argument passing
3436 // to the argument register save area. For O32, the save area is allocated
3437 // in the caller's stack frame, while for N32/64, it is allocated in the
3438 // callee's stack frame.
3439 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3440 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3441 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3442 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3443 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3444 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3445 MachinePointerInfo(), false, false, 0);
3446 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
3447 OutChains.push_back(Store);