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 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "mips-lower"
16 #include "MipsISelLowering.h"
17 #include "MipsMachineFunction.h"
18 #include "MipsTargetMachine.h"
19 #include "MipsTargetObjectFile.h"
20 #include "MipsSubtarget.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/CallingConv.h"
26 #include "InstPrinter/MipsInstPrinter.h"
27 #include "MCTargetDesc/MipsBaseInfo.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/SelectionDAGISel.h"
34 #include "llvm/CodeGen/ValueTypes.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
39 // If I is a shifted mask, set the size (Size) and the first bit of the
40 // mask (Pos), and return true.
41 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
42 static bool IsShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
43 if (!isShiftedMask_64(I))
46 Size = CountPopulation_64(I);
47 Pos = CountTrailingZeros_64(I);
51 static SDValue GetGlobalReg(SelectionDAG &DAG, EVT Ty) {
52 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
53 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
56 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
58 case MipsISD::JmpLink: return "MipsISD::JmpLink";
59 case MipsISD::Hi: return "MipsISD::Hi";
60 case MipsISD::Lo: return "MipsISD::Lo";
61 case MipsISD::GPRel: return "MipsISD::GPRel";
62 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
63 case MipsISD::Ret: return "MipsISD::Ret";
64 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
65 case MipsISD::FPCmp: return "MipsISD::FPCmp";
66 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
67 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
68 case MipsISD::FPRound: return "MipsISD::FPRound";
69 case MipsISD::MAdd: return "MipsISD::MAdd";
70 case MipsISD::MAddu: return "MipsISD::MAddu";
71 case MipsISD::MSub: return "MipsISD::MSub";
72 case MipsISD::MSubu: return "MipsISD::MSubu";
73 case MipsISD::DivRem: return "MipsISD::DivRem";
74 case MipsISD::DivRemU: return "MipsISD::DivRemU";
75 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
76 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
77 case MipsISD::Wrapper: return "MipsISD::Wrapper";
78 case MipsISD::DynAlloc: return "MipsISD::DynAlloc";
79 case MipsISD::Sync: return "MipsISD::Sync";
80 case MipsISD::Ext: return "MipsISD::Ext";
81 case MipsISD::Ins: return "MipsISD::Ins";
87 MipsTargetLowering(MipsTargetMachine &TM)
88 : TargetLowering(TM, new MipsTargetObjectFile()),
89 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
90 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
91 IsO32(Subtarget->isABI_O32()) {
93 // Mips does not have i1 type, so use i32 for
94 // setcc operations results (slt, sgt, ...).
95 setBooleanContents(ZeroOrOneBooleanContent);
96 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
98 // Set up the register classes
99 addRegisterClass(MVT::i32, Mips::CPURegsRegisterClass);
102 addRegisterClass(MVT::i64, Mips::CPU64RegsRegisterClass);
104 if (!TM.Options.UseSoftFloat) {
105 addRegisterClass(MVT::f32, Mips::FGR32RegisterClass);
107 // When dealing with single precision only, use libcalls
108 if (!Subtarget->isSingleFloat()) {
110 addRegisterClass(MVT::f64, Mips::FGR64RegisterClass);
112 addRegisterClass(MVT::f64, Mips::AFGR64RegisterClass);
116 // Load extented operations for i1 types must be promoted
117 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
118 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
119 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
121 // MIPS doesn't have extending float->double load/store
122 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
123 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
125 // Used by legalize types to correctly generate the setcc result.
126 // Without this, every float setcc comes with a AND/OR with the result,
127 // we don't want this, since the fpcmp result goes to a flag register,
128 // which is used implicitly by brcond and select operations.
129 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
131 // Mips Custom Operations
132 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
133 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
134 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
135 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
136 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
137 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
138 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
139 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
140 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
141 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
142 setOperationAction(ISD::SELECT, MVT::f32, Custom);
143 setOperationAction(ISD::SELECT, MVT::f64, Custom);
144 setOperationAction(ISD::SELECT, MVT::i32, Custom);
145 setOperationAction(ISD::SETCC, MVT::f32, Custom);
146 setOperationAction(ISD::SETCC, MVT::f64, Custom);
147 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
148 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
149 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Custom);
150 setOperationAction(ISD::VASTART, MVT::Other, Custom);
152 setOperationAction(ISD::SDIV, MVT::i32, Expand);
153 setOperationAction(ISD::SREM, MVT::i32, Expand);
154 setOperationAction(ISD::UDIV, MVT::i32, Expand);
155 setOperationAction(ISD::UREM, MVT::i32, Expand);
156 setOperationAction(ISD::SDIV, MVT::i64, Expand);
157 setOperationAction(ISD::SREM, MVT::i64, Expand);
158 setOperationAction(ISD::UDIV, MVT::i64, Expand);
159 setOperationAction(ISD::UREM, MVT::i64, Expand);
161 // Operations not directly supported by Mips.
162 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
163 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
164 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
165 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
166 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
167 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
168 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
169 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
170 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
171 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
172 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
173 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
174 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
175 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
176 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
177 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
178 setOperationAction(ISD::ROTL, MVT::i32, Expand);
179 setOperationAction(ISD::ROTL, MVT::i64, Expand);
181 if (!Subtarget->hasMips32r2())
182 setOperationAction(ISD::ROTR, MVT::i32, Expand);
184 if (!Subtarget->hasMips64r2())
185 setOperationAction(ISD::ROTR, MVT::i64, Expand);
187 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
188 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
189 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
190 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
191 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
192 setOperationAction(ISD::FSIN, MVT::f32, Expand);
193 setOperationAction(ISD::FSIN, MVT::f64, Expand);
194 setOperationAction(ISD::FCOS, MVT::f32, Expand);
195 setOperationAction(ISD::FCOS, MVT::f64, Expand);
196 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
197 setOperationAction(ISD::FPOW, MVT::f32, Expand);
198 setOperationAction(ISD::FPOW, MVT::f64, Expand);
199 setOperationAction(ISD::FLOG, MVT::f32, Expand);
200 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
201 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
202 setOperationAction(ISD::FEXP, MVT::f32, Expand);
203 setOperationAction(ISD::FMA, MVT::f32, Expand);
204 setOperationAction(ISD::FMA, MVT::f64, Expand);
206 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
207 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
208 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
209 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
211 setOperationAction(ISD::VAARG, MVT::Other, Expand);
212 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
213 setOperationAction(ISD::VAEND, MVT::Other, Expand);
215 // Use the default for now
216 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
217 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
219 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
220 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
222 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
223 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
224 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
225 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
227 setInsertFencesForAtomic(true);
229 if (Subtarget->isSingleFloat())
230 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
232 if (!Subtarget->hasSEInReg()) {
233 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
234 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
237 if (!Subtarget->hasBitCount()) {
238 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
239 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
242 if (!Subtarget->hasSwap()) {
243 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
244 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
247 setTargetDAGCombine(ISD::ADDE);
248 setTargetDAGCombine(ISD::SUBE);
249 setTargetDAGCombine(ISD::SDIVREM);
250 setTargetDAGCombine(ISD::UDIVREM);
251 setTargetDAGCombine(ISD::SELECT);
252 setTargetDAGCombine(ISD::AND);
253 setTargetDAGCombine(ISD::OR);
255 setMinFunctionAlignment(HasMips64 ? 3 : 2);
257 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP);
258 computeRegisterProperties();
260 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0);
261 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1);
264 bool MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
265 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
273 return Subtarget->hasMips32r2Or64();
279 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
284 // Transforms a subgraph in CurDAG if the following pattern is found:
285 // (addc multLo, Lo0), (adde multHi, Hi0),
287 // multHi/Lo: product of multiplication
288 // Lo0: initial value of Lo register
289 // Hi0: initial value of Hi register
290 // Return true if pattern matching was successful.
291 static bool SelectMadd(SDNode* ADDENode, SelectionDAG* CurDAG) {
292 // ADDENode's second operand must be a flag output of an ADDC node in order
293 // for the matching to be successful.
294 SDNode* ADDCNode = ADDENode->getOperand(2).getNode();
296 if (ADDCNode->getOpcode() != ISD::ADDC)
299 SDValue MultHi = ADDENode->getOperand(0);
300 SDValue MultLo = ADDCNode->getOperand(0);
301 SDNode* MultNode = MultHi.getNode();
302 unsigned MultOpc = MultHi.getOpcode();
304 // MultHi and MultLo must be generated by the same node,
305 if (MultLo.getNode() != MultNode)
308 // and it must be a multiplication.
309 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
312 // MultLo amd MultHi must be the first and second output of MultNode
314 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
317 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
318 // of the values of MultNode, in which case MultNode will be removed in later
320 // If there exist users other than ADDENode or ADDCNode, this function returns
321 // here, which will result in MultNode being mapped to a single MULT
322 // instruction node rather than a pair of MULT and MADD instructions being
324 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
327 SDValue Chain = CurDAG->getEntryNode();
328 DebugLoc dl = ADDENode->getDebugLoc();
330 // create MipsMAdd(u) node
331 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
333 SDValue MAdd = CurDAG->getNode(MultOpc, dl, MVT::Glue,
334 MultNode->getOperand(0),// Factor 0
335 MultNode->getOperand(1),// Factor 1
336 ADDCNode->getOperand(1),// Lo0
337 ADDENode->getOperand(1));// Hi0
339 // create CopyFromReg nodes
340 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
342 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
344 CopyFromLo.getValue(2));
346 // replace uses of adde and addc here
347 if (!SDValue(ADDCNode, 0).use_empty())
348 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
350 if (!SDValue(ADDENode, 0).use_empty())
351 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
357 // Transforms a subgraph in CurDAG if the following pattern is found:
358 // (addc Lo0, multLo), (sube Hi0, multHi),
360 // multHi/Lo: product of multiplication
361 // Lo0: initial value of Lo register
362 // Hi0: initial value of Hi register
363 // Return true if pattern matching was successful.
364 static bool SelectMsub(SDNode* SUBENode, SelectionDAG* CurDAG) {
365 // SUBENode's second operand must be a flag output of an SUBC node in order
366 // for the matching to be successful.
367 SDNode* SUBCNode = SUBENode->getOperand(2).getNode();
369 if (SUBCNode->getOpcode() != ISD::SUBC)
372 SDValue MultHi = SUBENode->getOperand(1);
373 SDValue MultLo = SUBCNode->getOperand(1);
374 SDNode* MultNode = MultHi.getNode();
375 unsigned MultOpc = MultHi.getOpcode();
377 // MultHi and MultLo must be generated by the same node,
378 if (MultLo.getNode() != MultNode)
381 // and it must be a multiplication.
382 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
385 // MultLo amd MultHi must be the first and second output of MultNode
387 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
390 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
391 // of the values of MultNode, in which case MultNode will be removed in later
393 // If there exist users other than SUBENode or SUBCNode, this function returns
394 // here, which will result in MultNode being mapped to a single MULT
395 // instruction node rather than a pair of MULT and MSUB instructions being
397 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
400 SDValue Chain = CurDAG->getEntryNode();
401 DebugLoc dl = SUBENode->getDebugLoc();
403 // create MipsSub(u) node
404 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
406 SDValue MSub = CurDAG->getNode(MultOpc, dl, MVT::Glue,
407 MultNode->getOperand(0),// Factor 0
408 MultNode->getOperand(1),// Factor 1
409 SUBCNode->getOperand(0),// Lo0
410 SUBENode->getOperand(0));// Hi0
412 // create CopyFromReg nodes
413 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
415 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
417 CopyFromLo.getValue(2));
419 // replace uses of sube and subc here
420 if (!SDValue(SUBCNode, 0).use_empty())
421 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
423 if (!SDValue(SUBENode, 0).use_empty())
424 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
429 static SDValue PerformADDECombine(SDNode *N, SelectionDAG& DAG,
430 TargetLowering::DAGCombinerInfo &DCI,
431 const MipsSubtarget* Subtarget) {
432 if (DCI.isBeforeLegalize())
435 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
437 return SDValue(N, 0);
442 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG& DAG,
443 TargetLowering::DAGCombinerInfo &DCI,
444 const MipsSubtarget* Subtarget) {
445 if (DCI.isBeforeLegalize())
448 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
450 return SDValue(N, 0);
455 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG& DAG,
456 TargetLowering::DAGCombinerInfo &DCI,
457 const MipsSubtarget* Subtarget) {
458 if (DCI.isBeforeLegalizeOps())
461 EVT Ty = N->getValueType(0);
462 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
463 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
464 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
466 DebugLoc dl = N->getDebugLoc();
468 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
469 N->getOperand(0), N->getOperand(1));
470 SDValue InChain = DAG.getEntryNode();
471 SDValue InGlue = DivRem;
474 if (N->hasAnyUseOfValue(0)) {
475 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty,
477 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
478 InChain = CopyFromLo.getValue(1);
479 InGlue = CopyFromLo.getValue(2);
483 if (N->hasAnyUseOfValue(1)) {
484 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
486 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
492 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
494 default: llvm_unreachable("Unknown fp condition code!");
496 case ISD::SETOEQ: return Mips::FCOND_OEQ;
497 case ISD::SETUNE: return Mips::FCOND_UNE;
499 case ISD::SETOLT: return Mips::FCOND_OLT;
501 case ISD::SETOGT: return Mips::FCOND_OGT;
503 case ISD::SETOLE: return Mips::FCOND_OLE;
505 case ISD::SETOGE: return Mips::FCOND_OGE;
506 case ISD::SETULT: return Mips::FCOND_ULT;
507 case ISD::SETULE: return Mips::FCOND_ULE;
508 case ISD::SETUGT: return Mips::FCOND_UGT;
509 case ISD::SETUGE: return Mips::FCOND_UGE;
510 case ISD::SETUO: return Mips::FCOND_UN;
511 case ISD::SETO: return Mips::FCOND_OR;
513 case ISD::SETONE: return Mips::FCOND_ONE;
514 case ISD::SETUEQ: return Mips::FCOND_UEQ;
519 // Returns true if condition code has to be inverted.
520 static bool InvertFPCondCode(Mips::CondCode CC) {
521 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
524 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
525 "Illegal Condition Code");
530 // Creates and returns an FPCmp node from a setcc node.
531 // Returns Op if setcc is not a floating point comparison.
532 static SDValue CreateFPCmp(SelectionDAG& DAG, const SDValue& Op) {
533 // must be a SETCC node
534 if (Op.getOpcode() != ISD::SETCC)
537 SDValue LHS = Op.getOperand(0);
539 if (!LHS.getValueType().isFloatingPoint())
542 SDValue RHS = Op.getOperand(1);
543 DebugLoc dl = Op.getDebugLoc();
545 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
546 // node if necessary.
547 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
549 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
550 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
553 // Creates and returns a CMovFPT/F node.
554 static SDValue CreateCMovFP(SelectionDAG& DAG, SDValue Cond, SDValue True,
555 SDValue False, DebugLoc DL) {
556 bool invert = InvertFPCondCode((Mips::CondCode)
557 cast<ConstantSDNode>(Cond.getOperand(2))
560 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
561 True.getValueType(), True, False, Cond);
564 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG& DAG,
565 TargetLowering::DAGCombinerInfo &DCI,
566 const MipsSubtarget* Subtarget) {
567 if (DCI.isBeforeLegalizeOps())
570 SDValue SetCC = N->getOperand(0);
572 if ((SetCC.getOpcode() != ISD::SETCC) ||
573 !SetCC.getOperand(0).getValueType().isInteger())
576 SDValue False = N->getOperand(2);
577 EVT FalseTy = False.getValueType();
579 if (!FalseTy.isInteger())
582 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
584 if (!CN || CN->getZExtValue())
587 const DebugLoc DL = N->getDebugLoc();
588 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
589 SDValue True = N->getOperand(1);
591 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
592 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
594 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
597 static SDValue PerformANDCombine(SDNode *N, SelectionDAG& DAG,
598 TargetLowering::DAGCombinerInfo &DCI,
599 const MipsSubtarget* Subtarget) {
600 // Pattern match EXT.
601 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
602 // => ext $dst, $src, size, pos
603 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
606 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
607 unsigned ShiftRightOpc = ShiftRight.getOpcode();
609 // Op's first operand must be a shift right.
610 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
613 // The second operand of the shift must be an immediate.
615 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
618 uint64_t Pos = CN->getZExtValue();
619 uint64_t SMPos, SMSize;
621 // Op's second operand must be a shifted mask.
622 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
623 !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
626 // Return if the shifted mask does not start at bit 0 or the sum of its size
627 // and Pos exceeds the word's size.
628 EVT ValTy = N->getValueType(0);
629 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
632 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
633 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
634 DAG.getConstant(SMSize, MVT::i32));
637 static SDValue PerformORCombine(SDNode *N, SelectionDAG& DAG,
638 TargetLowering::DAGCombinerInfo &DCI,
639 const MipsSubtarget* Subtarget) {
640 // Pattern match INS.
641 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
642 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
643 // => ins $dst, $src, size, pos, $src1
644 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
647 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
648 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
651 // See if Op's first operand matches (and $src1 , mask0).
652 if (And0.getOpcode() != ISD::AND)
655 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
656 !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
659 // See if Op's second operand matches (and (shl $src, pos), mask1).
660 if (And1.getOpcode() != ISD::AND)
663 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
664 !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
667 // The shift masks must have the same position and size.
668 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
671 SDValue Shl = And1.getOperand(0);
672 if (Shl.getOpcode() != ISD::SHL)
675 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
678 unsigned Shamt = CN->getZExtValue();
680 // Return if the shift amount and the first bit position of mask are not the
682 EVT ValTy = N->getValueType(0);
683 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
686 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
687 DAG.getConstant(SMPos0, MVT::i32),
688 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
691 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
693 SelectionDAG &DAG = DCI.DAG;
694 unsigned opc = N->getOpcode();
699 return PerformADDECombine(N, DAG, DCI, Subtarget);
701 return PerformSUBECombine(N, DAG, DCI, Subtarget);
704 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
706 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
708 return PerformANDCombine(N, DAG, DCI, Subtarget);
710 return PerformORCombine(N, DAG, DCI, Subtarget);
716 SDValue MipsTargetLowering::
717 LowerOperation(SDValue Op, SelectionDAG &DAG) const
719 switch (Op.getOpcode())
721 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
722 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
723 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
724 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
725 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
726 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
727 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
728 case ISD::SELECT: return LowerSELECT(Op, DAG);
729 case ISD::SETCC: return LowerSETCC(Op, DAG);
730 case ISD::VASTART: return LowerVASTART(Op, DAG);
731 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
732 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
733 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
734 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
739 //===----------------------------------------------------------------------===//
740 // Lower helper functions
741 //===----------------------------------------------------------------------===//
743 // AddLiveIn - This helper function adds the specified physical register to the
744 // MachineFunction as a live in value. It also creates a corresponding
745 // virtual register for it.
747 AddLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
749 assert(RC->contains(PReg) && "Not the correct regclass!");
750 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
751 MF.getRegInfo().addLiveIn(PReg, VReg);
755 // Get fp branch code (not opcode) from condition code.
756 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
757 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
758 return Mips::BRANCH_T;
760 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
761 "Invalid CondCode.");
763 return Mips::BRANCH_F;
767 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
769 const MipsSubtarget* Subtarget,
770 const TargetInstrInfo *TII,
771 bool isFPCmp, unsigned Opc) {
772 // There is no need to expand CMov instructions if target has
773 // conditional moves.
774 if (Subtarget->hasCondMov())
777 // To "insert" a SELECT_CC instruction, we actually have to insert the
778 // diamond control-flow pattern. The incoming instruction knows the
779 // destination vreg to set, the condition code register to branch on, the
780 // true/false values to select between, and a branch opcode to use.
781 const BasicBlock *LLVM_BB = BB->getBasicBlock();
782 MachineFunction::iterator It = BB;
789 // bNE r1, r0, copy1MBB
790 // fallthrough --> copy0MBB
791 MachineBasicBlock *thisMBB = BB;
792 MachineFunction *F = BB->getParent();
793 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
794 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
795 F->insert(It, copy0MBB);
796 F->insert(It, sinkMBB);
798 // Transfer the remainder of BB and its successor edges to sinkMBB.
799 sinkMBB->splice(sinkMBB->begin(), BB,
800 llvm::next(MachineBasicBlock::iterator(MI)),
802 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
804 // Next, add the true and fallthrough blocks as its successors.
805 BB->addSuccessor(copy0MBB);
806 BB->addSuccessor(sinkMBB);
808 // Emit the right instruction according to the type of the operands compared
810 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
812 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
813 .addReg(Mips::ZERO).addMBB(sinkMBB);
817 // # fallthrough to sinkMBB
820 // Update machine-CFG edges
821 BB->addSuccessor(sinkMBB);
824 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
829 BuildMI(*BB, BB->begin(), dl,
830 TII->get(Mips::PHI), MI->getOperand(0).getReg())
831 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
832 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
834 BuildMI(*BB, BB->begin(), dl,
835 TII->get(Mips::PHI), MI->getOperand(0).getReg())
836 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
837 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
839 MI->eraseFromParent(); // The pseudo instruction is gone now.
844 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
845 MachineBasicBlock *BB) const {
846 switch (MI->getOpcode()) {
847 default: llvm_unreachable("Unexpected instr type to insert");
848 case Mips::ATOMIC_LOAD_ADD_I8:
849 case Mips::ATOMIC_LOAD_ADD_I8_P8:
850 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
851 case Mips::ATOMIC_LOAD_ADD_I16:
852 case Mips::ATOMIC_LOAD_ADD_I16_P8:
853 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
854 case Mips::ATOMIC_LOAD_ADD_I32:
855 case Mips::ATOMIC_LOAD_ADD_I32_P8:
856 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
857 case Mips::ATOMIC_LOAD_ADD_I64:
858 case Mips::ATOMIC_LOAD_ADD_I64_P8:
859 return EmitAtomicBinary(MI, BB, 8, Mips::DADDu);
861 case Mips::ATOMIC_LOAD_AND_I8:
862 case Mips::ATOMIC_LOAD_AND_I8_P8:
863 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
864 case Mips::ATOMIC_LOAD_AND_I16:
865 case Mips::ATOMIC_LOAD_AND_I16_P8:
866 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
867 case Mips::ATOMIC_LOAD_AND_I32:
868 case Mips::ATOMIC_LOAD_AND_I32_P8:
869 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
870 case Mips::ATOMIC_LOAD_AND_I64:
871 case Mips::ATOMIC_LOAD_AND_I64_P8:
872 return EmitAtomicBinary(MI, BB, 8, Mips::AND64);
874 case Mips::ATOMIC_LOAD_OR_I8:
875 case Mips::ATOMIC_LOAD_OR_I8_P8:
876 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
877 case Mips::ATOMIC_LOAD_OR_I16:
878 case Mips::ATOMIC_LOAD_OR_I16_P8:
879 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
880 case Mips::ATOMIC_LOAD_OR_I32:
881 case Mips::ATOMIC_LOAD_OR_I32_P8:
882 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
883 case Mips::ATOMIC_LOAD_OR_I64:
884 case Mips::ATOMIC_LOAD_OR_I64_P8:
885 return EmitAtomicBinary(MI, BB, 8, Mips::OR64);
887 case Mips::ATOMIC_LOAD_XOR_I8:
888 case Mips::ATOMIC_LOAD_XOR_I8_P8:
889 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
890 case Mips::ATOMIC_LOAD_XOR_I16:
891 case Mips::ATOMIC_LOAD_XOR_I16_P8:
892 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
893 case Mips::ATOMIC_LOAD_XOR_I32:
894 case Mips::ATOMIC_LOAD_XOR_I32_P8:
895 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
896 case Mips::ATOMIC_LOAD_XOR_I64:
897 case Mips::ATOMIC_LOAD_XOR_I64_P8:
898 return EmitAtomicBinary(MI, BB, 8, Mips::XOR64);
900 case Mips::ATOMIC_LOAD_NAND_I8:
901 case Mips::ATOMIC_LOAD_NAND_I8_P8:
902 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
903 case Mips::ATOMIC_LOAD_NAND_I16:
904 case Mips::ATOMIC_LOAD_NAND_I16_P8:
905 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
906 case Mips::ATOMIC_LOAD_NAND_I32:
907 case Mips::ATOMIC_LOAD_NAND_I32_P8:
908 return EmitAtomicBinary(MI, BB, 4, 0, true);
909 case Mips::ATOMIC_LOAD_NAND_I64:
910 case Mips::ATOMIC_LOAD_NAND_I64_P8:
911 return EmitAtomicBinary(MI, BB, 8, 0, true);
913 case Mips::ATOMIC_LOAD_SUB_I8:
914 case Mips::ATOMIC_LOAD_SUB_I8_P8:
915 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
916 case Mips::ATOMIC_LOAD_SUB_I16:
917 case Mips::ATOMIC_LOAD_SUB_I16_P8:
918 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
919 case Mips::ATOMIC_LOAD_SUB_I32:
920 case Mips::ATOMIC_LOAD_SUB_I32_P8:
921 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
922 case Mips::ATOMIC_LOAD_SUB_I64:
923 case Mips::ATOMIC_LOAD_SUB_I64_P8:
924 return EmitAtomicBinary(MI, BB, 8, Mips::DSUBu);
926 case Mips::ATOMIC_SWAP_I8:
927 case Mips::ATOMIC_SWAP_I8_P8:
928 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
929 case Mips::ATOMIC_SWAP_I16:
930 case Mips::ATOMIC_SWAP_I16_P8:
931 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
932 case Mips::ATOMIC_SWAP_I32:
933 case Mips::ATOMIC_SWAP_I32_P8:
934 return EmitAtomicBinary(MI, BB, 4, 0);
935 case Mips::ATOMIC_SWAP_I64:
936 case Mips::ATOMIC_SWAP_I64_P8:
937 return EmitAtomicBinary(MI, BB, 8, 0);
939 case Mips::ATOMIC_CMP_SWAP_I8:
940 case Mips::ATOMIC_CMP_SWAP_I8_P8:
941 return EmitAtomicCmpSwapPartword(MI, BB, 1);
942 case Mips::ATOMIC_CMP_SWAP_I16:
943 case Mips::ATOMIC_CMP_SWAP_I16_P8:
944 return EmitAtomicCmpSwapPartword(MI, BB, 2);
945 case Mips::ATOMIC_CMP_SWAP_I32:
946 case Mips::ATOMIC_CMP_SWAP_I32_P8:
947 return EmitAtomicCmpSwap(MI, BB, 4);
948 case Mips::ATOMIC_CMP_SWAP_I64:
949 case Mips::ATOMIC_CMP_SWAP_I64_P8:
950 return EmitAtomicCmpSwap(MI, BB, 8);
954 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
955 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
957 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
958 unsigned Size, unsigned BinOpcode,
960 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
962 MachineFunction *MF = BB->getParent();
963 MachineRegisterInfo &RegInfo = MF->getRegInfo();
964 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
965 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
966 DebugLoc dl = MI->getDebugLoc();
967 unsigned LL, SC, AND, NOR, ZERO, BEQ;
970 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
971 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
978 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
979 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
982 ZERO = Mips::ZERO_64;
986 unsigned OldVal = MI->getOperand(0).getReg();
987 unsigned Ptr = MI->getOperand(1).getReg();
988 unsigned Incr = MI->getOperand(2).getReg();
990 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
991 unsigned AndRes = RegInfo.createVirtualRegister(RC);
992 unsigned Success = RegInfo.createVirtualRegister(RC);
994 // insert new blocks after the current block
995 const BasicBlock *LLVM_BB = BB->getBasicBlock();
996 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
997 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
998 MachineFunction::iterator It = BB;
1000 MF->insert(It, loopMBB);
1001 MF->insert(It, exitMBB);
1003 // Transfer the remainder of BB and its successor edges to exitMBB.
1004 exitMBB->splice(exitMBB->begin(), BB,
1005 llvm::next(MachineBasicBlock::iterator(MI)),
1007 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1011 // fallthrough --> loopMBB
1012 BB->addSuccessor(loopMBB);
1013 loopMBB->addSuccessor(loopMBB);
1014 loopMBB->addSuccessor(exitMBB);
1017 // ll oldval, 0(ptr)
1018 // <binop> storeval, oldval, incr
1019 // sc success, storeval, 0(ptr)
1020 // beq success, $0, loopMBB
1022 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1024 // and andres, oldval, incr
1025 // nor storeval, $0, andres
1026 BuildMI(BB, dl, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1027 BuildMI(BB, dl, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1028 } else if (BinOpcode) {
1029 // <binop> storeval, oldval, incr
1030 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1034 BuildMI(BB, dl, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1035 BuildMI(BB, dl, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1037 MI->eraseFromParent(); // The instruction is gone now.
1043 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
1044 MachineBasicBlock *BB,
1045 unsigned Size, unsigned BinOpcode,
1047 assert((Size == 1 || Size == 2) &&
1048 "Unsupported size for EmitAtomicBinaryPartial.");
1050 MachineFunction *MF = BB->getParent();
1051 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1052 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1053 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1054 DebugLoc dl = MI->getDebugLoc();
1055 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1056 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1058 unsigned Dest = MI->getOperand(0).getReg();
1059 unsigned Ptr = MI->getOperand(1).getReg();
1060 unsigned Incr = MI->getOperand(2).getReg();
1062 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1063 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1064 unsigned Mask = RegInfo.createVirtualRegister(RC);
1065 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1066 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1067 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1068 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1069 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1070 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1071 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1072 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1073 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1074 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1075 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1076 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1077 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1078 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1079 unsigned Success = RegInfo.createVirtualRegister(RC);
1081 // insert new blocks after the current block
1082 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1083 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1084 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1085 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1086 MachineFunction::iterator It = BB;
1088 MF->insert(It, loopMBB);
1089 MF->insert(It, sinkMBB);
1090 MF->insert(It, exitMBB);
1092 // Transfer the remainder of BB and its successor edges to exitMBB.
1093 exitMBB->splice(exitMBB->begin(), BB,
1094 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1095 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1097 BB->addSuccessor(loopMBB);
1098 loopMBB->addSuccessor(loopMBB);
1099 loopMBB->addSuccessor(sinkMBB);
1100 sinkMBB->addSuccessor(exitMBB);
1103 // addiu masklsb2,$0,-4 # 0xfffffffc
1104 // and alignedaddr,ptr,masklsb2
1105 // andi ptrlsb2,ptr,3
1106 // sll shiftamt,ptrlsb2,3
1107 // ori maskupper,$0,255 # 0xff
1108 // sll mask,maskupper,shiftamt
1109 // nor mask2,$0,mask
1110 // sll incr2,incr,shiftamt
1112 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1113 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1114 .addReg(Mips::ZERO).addImm(-4);
1115 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1116 .addReg(Ptr).addReg(MaskLSB2);
1117 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1118 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1119 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1120 .addReg(Mips::ZERO).addImm(MaskImm);
1121 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1122 .addReg(ShiftAmt).addReg(MaskUpper);
1123 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1124 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1126 // atomic.load.binop
1128 // ll oldval,0(alignedaddr)
1129 // binop binopres,oldval,incr2
1130 // and newval,binopres,mask
1131 // and maskedoldval0,oldval,mask2
1132 // or storeval,maskedoldval0,newval
1133 // sc success,storeval,0(alignedaddr)
1134 // beq success,$0,loopMBB
1138 // ll oldval,0(alignedaddr)
1139 // and newval,incr2,mask
1140 // and maskedoldval0,oldval,mask2
1141 // or storeval,maskedoldval0,newval
1142 // sc success,storeval,0(alignedaddr)
1143 // beq success,$0,loopMBB
1146 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1148 // and andres, oldval, incr2
1149 // nor binopres, $0, andres
1150 // and newval, binopres, mask
1151 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1152 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1153 .addReg(Mips::ZERO).addReg(AndRes);
1154 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1155 } else if (BinOpcode) {
1156 // <binop> binopres, oldval, incr2
1157 // and newval, binopres, mask
1158 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1159 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1160 } else {// atomic.swap
1161 // and newval, incr2, mask
1162 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1165 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1166 .addReg(OldVal).addReg(Mask2);
1167 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1168 .addReg(MaskedOldVal0).addReg(NewVal);
1169 BuildMI(BB, dl, TII->get(SC), Success)
1170 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1171 BuildMI(BB, dl, TII->get(Mips::BEQ))
1172 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1175 // and maskedoldval1,oldval,mask
1176 // srl srlres,maskedoldval1,shiftamt
1177 // sll sllres,srlres,24
1178 // sra dest,sllres,24
1180 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1182 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1183 .addReg(OldVal).addReg(Mask);
1184 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1185 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1186 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1187 .addReg(SrlRes).addImm(ShiftImm);
1188 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1189 .addReg(SllRes).addImm(ShiftImm);
1191 MI->eraseFromParent(); // The instruction is gone now.
1197 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1198 MachineBasicBlock *BB,
1199 unsigned Size) const {
1200 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1202 MachineFunction *MF = BB->getParent();
1203 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1204 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1205 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1206 DebugLoc dl = MI->getDebugLoc();
1207 unsigned LL, SC, ZERO, BNE, BEQ;
1210 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1211 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1217 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1218 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1219 ZERO = Mips::ZERO_64;
1224 unsigned Dest = MI->getOperand(0).getReg();
1225 unsigned Ptr = MI->getOperand(1).getReg();
1226 unsigned OldVal = MI->getOperand(2).getReg();
1227 unsigned NewVal = MI->getOperand(3).getReg();
1229 unsigned Success = RegInfo.createVirtualRegister(RC);
1231 // insert new blocks after the current block
1232 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1233 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1234 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1235 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1236 MachineFunction::iterator It = BB;
1238 MF->insert(It, loop1MBB);
1239 MF->insert(It, loop2MBB);
1240 MF->insert(It, exitMBB);
1242 // Transfer the remainder of BB and its successor edges to exitMBB.
1243 exitMBB->splice(exitMBB->begin(), BB,
1244 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1245 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1249 // fallthrough --> loop1MBB
1250 BB->addSuccessor(loop1MBB);
1251 loop1MBB->addSuccessor(exitMBB);
1252 loop1MBB->addSuccessor(loop2MBB);
1253 loop2MBB->addSuccessor(loop1MBB);
1254 loop2MBB->addSuccessor(exitMBB);
1258 // bne dest, oldval, exitMBB
1260 BuildMI(BB, dl, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1261 BuildMI(BB, dl, TII->get(BNE))
1262 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1265 // sc success, newval, 0(ptr)
1266 // beq success, $0, loop1MBB
1268 BuildMI(BB, dl, TII->get(SC), Success)
1269 .addReg(NewVal).addReg(Ptr).addImm(0);
1270 BuildMI(BB, dl, TII->get(BEQ))
1271 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1273 MI->eraseFromParent(); // The instruction is gone now.
1279 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1280 MachineBasicBlock *BB,
1281 unsigned Size) const {
1282 assert((Size == 1 || Size == 2) &&
1283 "Unsupported size for EmitAtomicCmpSwapPartial.");
1285 MachineFunction *MF = BB->getParent();
1286 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1287 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1288 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1289 DebugLoc dl = MI->getDebugLoc();
1290 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1291 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1293 unsigned Dest = MI->getOperand(0).getReg();
1294 unsigned Ptr = MI->getOperand(1).getReg();
1295 unsigned CmpVal = MI->getOperand(2).getReg();
1296 unsigned NewVal = MI->getOperand(3).getReg();
1298 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1299 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1300 unsigned Mask = RegInfo.createVirtualRegister(RC);
1301 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1302 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1303 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1304 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1305 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1306 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1307 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1308 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1309 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1310 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1311 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1312 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1313 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1314 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1315 unsigned Success = RegInfo.createVirtualRegister(RC);
1317 // insert new blocks after the current block
1318 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1319 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1320 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1321 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1322 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1323 MachineFunction::iterator It = BB;
1325 MF->insert(It, loop1MBB);
1326 MF->insert(It, loop2MBB);
1327 MF->insert(It, sinkMBB);
1328 MF->insert(It, exitMBB);
1330 // Transfer the remainder of BB and its successor edges to exitMBB.
1331 exitMBB->splice(exitMBB->begin(), BB,
1332 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1333 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1335 BB->addSuccessor(loop1MBB);
1336 loop1MBB->addSuccessor(sinkMBB);
1337 loop1MBB->addSuccessor(loop2MBB);
1338 loop2MBB->addSuccessor(loop1MBB);
1339 loop2MBB->addSuccessor(sinkMBB);
1340 sinkMBB->addSuccessor(exitMBB);
1342 // FIXME: computation of newval2 can be moved to loop2MBB.
1344 // addiu masklsb2,$0,-4 # 0xfffffffc
1345 // and alignedaddr,ptr,masklsb2
1346 // andi ptrlsb2,ptr,3
1347 // sll shiftamt,ptrlsb2,3
1348 // ori maskupper,$0,255 # 0xff
1349 // sll mask,maskupper,shiftamt
1350 // nor mask2,$0,mask
1351 // andi maskedcmpval,cmpval,255
1352 // sll shiftedcmpval,maskedcmpval,shiftamt
1353 // andi maskednewval,newval,255
1354 // sll shiftednewval,maskednewval,shiftamt
1355 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1356 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1357 .addReg(Mips::ZERO).addImm(-4);
1358 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1359 .addReg(Ptr).addReg(MaskLSB2);
1360 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1361 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1362 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1363 .addReg(Mips::ZERO).addImm(MaskImm);
1364 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1365 .addReg(ShiftAmt).addReg(MaskUpper);
1366 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1367 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1368 .addReg(CmpVal).addImm(MaskImm);
1369 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1370 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1371 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1372 .addReg(NewVal).addImm(MaskImm);
1373 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1374 .addReg(ShiftAmt).addReg(MaskedNewVal);
1377 // ll oldval,0(alginedaddr)
1378 // and maskedoldval0,oldval,mask
1379 // bne maskedoldval0,shiftedcmpval,sinkMBB
1381 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1382 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1383 .addReg(OldVal).addReg(Mask);
1384 BuildMI(BB, dl, TII->get(Mips::BNE))
1385 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1388 // and maskedoldval1,oldval,mask2
1389 // or storeval,maskedoldval1,shiftednewval
1390 // sc success,storeval,0(alignedaddr)
1391 // beq success,$0,loop1MBB
1393 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1394 .addReg(OldVal).addReg(Mask2);
1395 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1396 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1397 BuildMI(BB, dl, TII->get(SC), Success)
1398 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1399 BuildMI(BB, dl, TII->get(Mips::BEQ))
1400 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1403 // srl srlres,maskedoldval0,shiftamt
1404 // sll sllres,srlres,24
1405 // sra dest,sllres,24
1407 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1409 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1410 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1411 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1412 .addReg(SrlRes).addImm(ShiftImm);
1413 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1414 .addReg(SllRes).addImm(ShiftImm);
1416 MI->eraseFromParent(); // The instruction is gone now.
1421 //===----------------------------------------------------------------------===//
1422 // Misc Lower Operation implementation
1423 //===----------------------------------------------------------------------===//
1424 SDValue MipsTargetLowering::
1425 LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const
1427 MachineFunction &MF = DAG.getMachineFunction();
1428 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1429 unsigned SP = IsN64 ? Mips::SP_64 : Mips::SP;
1431 assert(getTargetMachine().getFrameLowering()->getStackAlignment() >=
1432 cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue() &&
1433 "Cannot lower if the alignment of the allocated space is larger than \
1434 that of the stack.");
1436 SDValue Chain = Op.getOperand(0);
1437 SDValue Size = Op.getOperand(1);
1438 DebugLoc dl = Op.getDebugLoc();
1440 // Get a reference from Mips stack pointer
1441 SDValue StackPointer = DAG.getCopyFromReg(Chain, dl, SP, getPointerTy());
1443 // Subtract the dynamic size from the actual stack size to
1444 // obtain the new stack size.
1445 SDValue Sub = DAG.getNode(ISD::SUB, dl, getPointerTy(), StackPointer, Size);
1447 // The Sub result contains the new stack start address, so it
1448 // must be placed in the stack pointer register.
1449 Chain = DAG.getCopyToReg(StackPointer.getValue(1), dl, SP, Sub, SDValue());
1451 // This node always has two return values: a new stack pointer
1452 // value and a chain
1453 SDVTList VTLs = DAG.getVTList(getPointerTy(), MVT::Other);
1454 SDValue Ptr = DAG.getFrameIndex(MipsFI->getDynAllocFI(), getPointerTy());
1455 SDValue Ops[] = { Chain, Ptr, Chain.getValue(1) };
1457 return DAG.getNode(MipsISD::DynAlloc, dl, VTLs, Ops, 3);
1460 SDValue MipsTargetLowering::
1461 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1463 // The first operand is the chain, the second is the condition, the third is
1464 // the block to branch to if the condition is true.
1465 SDValue Chain = Op.getOperand(0);
1466 SDValue Dest = Op.getOperand(2);
1467 DebugLoc dl = Op.getDebugLoc();
1469 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1471 // Return if flag is not set by a floating point comparison.
1472 if (CondRes.getOpcode() != MipsISD::FPCmp)
1475 SDValue CCNode = CondRes.getOperand(2);
1477 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1478 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1480 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1484 SDValue MipsTargetLowering::
1485 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1487 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1489 // Return if flag is not set by a floating point comparison.
1490 if (Cond.getOpcode() != MipsISD::FPCmp)
1493 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1497 SDValue MipsTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1498 SDValue Cond = CreateFPCmp(DAG, Op);
1500 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1501 "Floating point operand expected.");
1503 SDValue True = DAG.getConstant(1, MVT::i32);
1504 SDValue False = DAG.getConstant(0, MVT::i32);
1506 return CreateCMovFP(DAG, Cond, True, False, Op.getDebugLoc());
1509 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1510 SelectionDAG &DAG) const {
1511 // FIXME there isn't actually debug info here
1512 DebugLoc dl = Op.getDebugLoc();
1513 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1515 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1516 SDVTList VTs = DAG.getVTList(MVT::i32);
1518 MipsTargetObjectFile &TLOF = (MipsTargetObjectFile&)getObjFileLowering();
1520 // %gp_rel relocation
1521 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1522 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1524 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl, VTs, &GA, 1);
1525 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1526 return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode);
1528 // %hi/%lo relocation
1529 SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1531 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1533 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, VTs, &GAHi, 1);
1534 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1535 return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1538 EVT ValTy = Op.getValueType();
1539 bool HasGotOfst = (GV->hasInternalLinkage() ||
1540 (GV->hasLocalLinkage() && !isa<Function>(GV)));
1541 unsigned GotFlag = IsN64 ?
1542 (HasGotOfst ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT_DISP) :
1543 (HasGotOfst ? MipsII::MO_GOT : MipsII::MO_GOT16);
1544 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0, GotFlag);
1545 GA = DAG.getNode(MipsISD::Wrapper, dl, ValTy, GetGlobalReg(DAG, ValTy), GA);
1546 SDValue ResNode = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), GA,
1547 MachinePointerInfo(), false, false, false, 0);
1548 // On functions and global targets not internal linked only
1549 // a load from got/GP is necessary for PIC to work.
1552 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0,
1553 IsN64 ? MipsII::MO_GOT_OFST :
1555 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, GALo);
1556 return DAG.getNode(ISD::ADD, dl, ValTy, ResNode, Lo);
1559 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1560 SelectionDAG &DAG) const {
1561 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1562 // FIXME there isn't actually debug info here
1563 DebugLoc dl = Op.getDebugLoc();
1565 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1566 // %hi/%lo relocation
1567 SDValue BAHi = DAG.getBlockAddress(BA, MVT::i32, true, MipsII::MO_ABS_HI);
1568 SDValue BALo = DAG.getBlockAddress(BA, MVT::i32, true, MipsII::MO_ABS_LO);
1569 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, MVT::i32, BAHi);
1570 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALo);
1571 return DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1574 EVT ValTy = Op.getValueType();
1575 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1576 unsigned OFSTFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1577 SDValue BAGOTOffset = DAG.getBlockAddress(BA, ValTy, true, GOTFlag);
1578 BAGOTOffset = DAG.getNode(MipsISD::Wrapper, dl, ValTy,
1579 GetGlobalReg(DAG, ValTy), BAGOTOffset);
1580 SDValue BALOOffset = DAG.getBlockAddress(BA, ValTy, true, OFSTFlag);
1581 SDValue Load = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), BAGOTOffset,
1582 MachinePointerInfo(), false, false, false, 0);
1583 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, BALOOffset);
1584 return DAG.getNode(ISD::ADD, dl, ValTy, Load, Lo);
1587 SDValue MipsTargetLowering::
1588 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1590 // If the relocation model is PIC, use the General Dynamic TLS Model or
1591 // Local Dynamic TLS model, otherwise use the Initial Exec or
1592 // Local Exec TLS Model.
1594 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1595 DebugLoc dl = GA->getDebugLoc();
1596 const GlobalValue *GV = GA->getGlobal();
1597 EVT PtrVT = getPointerTy();
1599 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1600 // General Dynamic TLS Model
1601 bool LocalDynamic = GV->hasInternalLinkage();
1602 unsigned Flag = LocalDynamic ? MipsII::MO_TLSLDM :MipsII::MO_TLSGD;
1603 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, Flag);
1604 SDValue Argument = DAG.getNode(MipsISD::Wrapper, dl, PtrVT,
1605 GetGlobalReg(DAG, PtrVT), TGA);
1606 unsigned PtrSize = PtrVT.getSizeInBits();
1607 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1609 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1613 Entry.Node = Argument;
1615 Args.push_back(Entry);
1617 std::pair<SDValue, SDValue> CallResult =
1618 LowerCallTo(DAG.getEntryNode(), PtrTy,
1619 false, false, false, false, 0, CallingConv::C,
1620 /*isTailCall=*/false, /*doesNotRet=*/false,
1621 /*isReturnValueUsed=*/true,
1622 TlsGetAddr, Args, DAG, dl);
1624 SDValue Ret = CallResult.first;
1629 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1630 MipsII::MO_DTPREL_HI);
1631 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1632 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1633 MipsII::MO_DTPREL_LO);
1634 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1635 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Ret);
1636 return DAG.getNode(ISD::ADD, dl, PtrVT, Add, Lo);
1640 if (GV->isDeclaration()) {
1641 // Initial Exec TLS Model
1642 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1643 MipsII::MO_GOTTPREL);
1644 TGA = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, GetGlobalReg(DAG, PtrVT),
1646 Offset = DAG.getLoad(PtrVT, dl,
1647 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1648 false, false, false, 0);
1650 // Local Exec TLS Model
1651 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1652 MipsII::MO_TPREL_HI);
1653 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1654 MipsII::MO_TPREL_LO);
1655 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1656 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1657 Offset = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1660 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1661 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1664 SDValue MipsTargetLowering::
1665 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1667 SDValue HiPart, JTI, JTILo;
1668 // FIXME there isn't actually debug info here
1669 DebugLoc dl = Op.getDebugLoc();
1670 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1671 EVT PtrVT = Op.getValueType();
1672 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1674 if (!IsPIC && !IsN64) {
1675 JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MipsII::MO_ABS_HI);
1676 HiPart = DAG.getNode(MipsISD::Hi, dl, PtrVT, JTI);
1677 JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MipsII::MO_ABS_LO);
1678 } else {// Emit Load from Global Pointer
1679 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1680 unsigned OfstFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1681 JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, GOTFlag);
1682 JTI = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, GetGlobalReg(DAG, PtrVT),
1684 HiPart = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), JTI,
1685 MachinePointerInfo(), false, false, false, 0);
1686 JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OfstFlag);
1689 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, JTILo);
1690 return DAG.getNode(ISD::ADD, dl, PtrVT, HiPart, Lo);
1693 SDValue MipsTargetLowering::
1694 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1697 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1698 const Constant *C = N->getConstVal();
1699 // FIXME there isn't actually debug info here
1700 DebugLoc dl = Op.getDebugLoc();
1702 // gp_rel relocation
1703 // FIXME: we should reference the constant pool using small data sections,
1704 // but the asm printer currently doesn't support this feature without
1705 // hacking it. This feature should come soon so we can uncomment the
1707 //if (IsInSmallSection(C->getType())) {
1708 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1709 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1710 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1712 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1713 SDValue CPHi = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1714 N->getOffset(), MipsII::MO_ABS_HI);
1715 SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1716 N->getOffset(), MipsII::MO_ABS_LO);
1717 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, MVT::i32, CPHi);
1718 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1719 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1721 EVT ValTy = Op.getValueType();
1722 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1723 unsigned OFSTFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1724 SDValue CP = DAG.getTargetConstantPool(C, ValTy, N->getAlignment(),
1725 N->getOffset(), GOTFlag);
1726 CP = DAG.getNode(MipsISD::Wrapper, dl, ValTy, GetGlobalReg(DAG, ValTy), CP);
1727 SDValue Load = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), CP,
1728 MachinePointerInfo::getConstantPool(), false,
1730 SDValue CPLo = DAG.getTargetConstantPool(C, ValTy, N->getAlignment(),
1731 N->getOffset(), OFSTFlag);
1732 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, CPLo);
1733 ResNode = DAG.getNode(ISD::ADD, dl, ValTy, Load, Lo);
1739 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1740 MachineFunction &MF = DAG.getMachineFunction();
1741 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1743 DebugLoc dl = Op.getDebugLoc();
1744 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1747 // vastart just stores the address of the VarArgsFrameIndex slot into the
1748 // memory location argument.
1749 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1750 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1751 MachinePointerInfo(SV), false, false, 0);
1754 // Called if the size of integer registers is large enough to hold the whole
1755 // floating point number.
1756 static SDValue LowerFCOPYSIGNLargeIntReg(SDValue Op, SelectionDAG &DAG) {
1757 // FIXME: Use ext/ins instructions if target architecture is Mips32r2.
1758 EVT ValTy = Op.getValueType();
1759 EVT IntValTy = MVT::getIntegerVT(ValTy.getSizeInBits());
1760 uint64_t Mask = (uint64_t)1 << (ValTy.getSizeInBits() - 1);
1761 DebugLoc dl = Op.getDebugLoc();
1762 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntValTy, Op.getOperand(0));
1763 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntValTy, Op.getOperand(1));
1764 SDValue And0 = DAG.getNode(ISD::AND, dl, IntValTy, Op0,
1765 DAG.getConstant(Mask - 1, IntValTy));
1766 SDValue And1 = DAG.getNode(ISD::AND, dl, IntValTy, Op1,
1767 DAG.getConstant(Mask, IntValTy));
1768 SDValue Result = DAG.getNode(ISD::OR, dl, IntValTy, And0, And1);
1769 return DAG.getNode(ISD::BITCAST, dl, ValTy, Result);
1772 // Called if the size of integer registers is not large enough to hold the whole
1773 // floating point number (e.g. f64 & 32-bit integer register).
1775 LowerFCOPYSIGNSmallIntReg(SDValue Op, SelectionDAG &DAG, bool isLittle) {
1777 // Use ext/ins instructions if target architecture is Mips32r2.
1778 // Eliminate redundant mfc1 and mtc1 instructions.
1779 unsigned LoIdx = 0, HiIdx = 1;
1782 std::swap(LoIdx, HiIdx);
1784 DebugLoc dl = Op.getDebugLoc();
1785 SDValue Word0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1787 DAG.getConstant(LoIdx, MVT::i32));
1788 SDValue Hi0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1789 Op.getOperand(0), DAG.getConstant(HiIdx, MVT::i32));
1790 SDValue Hi1 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1791 Op.getOperand(1), DAG.getConstant(HiIdx, MVT::i32));
1792 SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi0,
1793 DAG.getConstant(0x7fffffff, MVT::i32));
1794 SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi1,
1795 DAG.getConstant(0x80000000, MVT::i32));
1796 SDValue Word1 = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1799 std::swap(Word0, Word1);
1801 return DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64, Word0, Word1);
1805 MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1806 EVT Ty = Op.getValueType();
1808 assert(Ty == MVT::f32 || Ty == MVT::f64);
1810 if (Ty == MVT::f32 || HasMips64)
1811 return LowerFCOPYSIGNLargeIntReg(Op, DAG);
1813 return LowerFCOPYSIGNSmallIntReg(Op, DAG, Subtarget->isLittle());
1816 SDValue MipsTargetLowering::
1817 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1819 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1820 "Frame address can only be determined for current frame.");
1822 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1823 MFI->setFrameAddressIsTaken(true);
1824 EVT VT = Op.getValueType();
1825 DebugLoc dl = Op.getDebugLoc();
1826 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
1827 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1831 // TODO: set SType according to the desired memory barrier behavior.
1833 MipsTargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG& DAG) const {
1835 DebugLoc dl = Op.getDebugLoc();
1836 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1837 DAG.getConstant(SType, MVT::i32));
1840 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1841 SelectionDAG& DAG) const {
1842 // FIXME: Need pseudo-fence for 'singlethread' fences
1843 // FIXME: Set SType for weaker fences where supported/appropriate.
1845 DebugLoc dl = Op.getDebugLoc();
1846 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1847 DAG.getConstant(SType, MVT::i32));
1850 //===----------------------------------------------------------------------===//
1851 // Calling Convention Implementation
1852 //===----------------------------------------------------------------------===//
1854 //===----------------------------------------------------------------------===//
1855 // TODO: Implement a generic logic using tblgen that can support this.
1856 // Mips O32 ABI rules:
1858 // i32 - Passed in A0, A1, A2, A3 and stack
1859 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
1860 // an argument. Otherwise, passed in A1, A2, A3 and stack.
1861 // f64 - Only passed in two aliased f32 registers if no int reg has been used
1862 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
1863 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
1866 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
1867 //===----------------------------------------------------------------------===//
1869 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
1870 MVT LocVT, CCValAssign::LocInfo LocInfo,
1871 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1873 static const unsigned IntRegsSize=4, FloatRegsSize=2;
1875 static const unsigned IntRegs[] = {
1876 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1878 static const unsigned F32Regs[] = {
1879 Mips::F12, Mips::F14
1881 static const unsigned F64Regs[] = {
1886 if (ArgFlags.isByVal()) {
1887 State.HandleByVal(ValNo, ValVT, LocVT, LocInfo,
1888 1 /*MinSize*/, 4 /*MinAlign*/, ArgFlags);
1889 unsigned NextReg = (State.getNextStackOffset() + 3) / 4;
1890 for (unsigned r = State.getFirstUnallocated(IntRegs, IntRegsSize);
1891 r < std::min(IntRegsSize, NextReg); ++r)
1892 State.AllocateReg(IntRegs[r]);
1896 // Promote i8 and i16
1897 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
1899 if (ArgFlags.isSExt())
1900 LocInfo = CCValAssign::SExt;
1901 else if (ArgFlags.isZExt())
1902 LocInfo = CCValAssign::ZExt;
1904 LocInfo = CCValAssign::AExt;
1909 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
1910 // is true: function is vararg, argument is 3rd or higher, there is previous
1911 // argument which is not f32 or f64.
1912 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
1913 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
1914 unsigned OrigAlign = ArgFlags.getOrigAlign();
1915 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
1917 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
1918 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1919 // If this is the first part of an i64 arg,
1920 // the allocated register must be either A0 or A2.
1921 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
1922 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1924 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
1925 // Allocate int register and shadow next int register. If first
1926 // available register is Mips::A1 or Mips::A3, shadow it too.
1927 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1928 if (Reg == Mips::A1 || Reg == Mips::A3)
1929 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1930 State.AllocateReg(IntRegs, IntRegsSize);
1932 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
1933 // we are guaranteed to find an available float register
1934 if (ValVT == MVT::f32) {
1935 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
1936 // Shadow int register
1937 State.AllocateReg(IntRegs, IntRegsSize);
1939 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
1940 // Shadow int registers
1941 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
1942 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
1943 State.AllocateReg(IntRegs, IntRegsSize);
1944 State.AllocateReg(IntRegs, IntRegsSize);
1947 llvm_unreachable("Cannot handle this ValVT.");
1949 unsigned SizeInBytes = ValVT.getSizeInBits() >> 3;
1950 unsigned Offset = State.AllocateStack(SizeInBytes, OrigAlign);
1953 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1955 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1957 return false; // CC must always match
1960 static const unsigned Mips64IntRegs[8] =
1961 {Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
1962 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64};
1963 static const unsigned Mips64DPRegs[8] =
1964 {Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
1965 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64};
1967 static bool CC_Mips64Byval(unsigned ValNo, MVT ValVT, MVT LocVT,
1968 CCValAssign::LocInfo LocInfo,
1969 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1970 unsigned Align = std::max(ArgFlags.getByValAlign(), (unsigned)8);
1971 unsigned Size = (ArgFlags.getByValSize() + 7) / 8 * 8;
1972 unsigned FirstIdx = State.getFirstUnallocated(Mips64IntRegs, 8);
1974 assert(Align <= 16 && "Cannot handle alignments larger than 16.");
1976 // If byval is 16-byte aligned, the first arg register must be even.
1977 if ((Align == 16) && (FirstIdx % 2)) {
1978 State.AllocateReg(Mips64IntRegs[FirstIdx], Mips64DPRegs[FirstIdx]);
1982 // Mark the registers allocated.
1983 for (unsigned I = FirstIdx; Size && (I < 8); Size -= 8, ++I)
1984 State.AllocateReg(Mips64IntRegs[I], Mips64DPRegs[I]);
1986 // Allocate space on caller's stack.
1987 unsigned Offset = State.AllocateStack(Size, Align);
1990 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Mips64IntRegs[FirstIdx],
1993 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1998 #include "MipsGenCallingConv.inc"
2001 AnalyzeMips64CallOperands(CCState &CCInfo,
2002 const SmallVectorImpl<ISD::OutputArg> &Outs) {
2003 unsigned NumOps = Outs.size();
2004 for (unsigned i = 0; i != NumOps; ++i) {
2005 MVT ArgVT = Outs[i].VT;
2006 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2009 if (Outs[i].IsFixed)
2010 R = CC_MipsN(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2012 R = CC_MipsN_VarArg(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2016 dbgs() << "Call operand #" << i << " has unhandled type "
2017 << EVT(ArgVT).getEVTString();
2019 llvm_unreachable(0);
2024 //===----------------------------------------------------------------------===//
2025 // Call Calling Convention Implementation
2026 //===----------------------------------------------------------------------===//
2028 static const unsigned O32IntRegsSize = 4;
2030 static const unsigned O32IntRegs[] = {
2031 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2034 // Return next O32 integer argument register.
2035 static unsigned getNextIntArgReg(unsigned Reg) {
2036 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2037 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2040 // Write ByVal Arg to arg registers and stack.
2042 WriteByValArg(SDValue& ByValChain, SDValue Chain, DebugLoc dl,
2043 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
2044 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
2045 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
2046 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
2047 MVT PtrType, bool isLittle) {
2048 unsigned LocMemOffset = VA.getLocMemOffset();
2049 unsigned Offset = 0;
2050 uint32_t RemainingSize = Flags.getByValSize();
2051 unsigned ByValAlign = Flags.getByValAlign();
2053 // Copy the first 4 words of byval arg to registers A0 - A3.
2054 // FIXME: Use a stricter alignment if it enables better optimization in passes
2056 for (; RemainingSize >= 4 && LocMemOffset < 4 * 4;
2057 Offset += 4, RemainingSize -= 4, LocMemOffset += 4) {
2058 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2059 DAG.getConstant(Offset, MVT::i32));
2060 SDValue LoadVal = DAG.getLoad(MVT::i32, dl, Chain, LoadPtr,
2061 MachinePointerInfo(), false, false, false,
2062 std::min(ByValAlign, (unsigned )4));
2063 MemOpChains.push_back(LoadVal.getValue(1));
2064 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
2065 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
2068 if (RemainingSize == 0)
2071 // If there still is a register available for argument passing, write the
2072 // remaining part of the structure to it using subword loads and shifts.
2073 if (LocMemOffset < 4 * 4) {
2074 assert(RemainingSize <= 3 && RemainingSize >= 1 &&
2075 "There must be one to three bytes remaining.");
2076 unsigned LoadSize = (RemainingSize == 3 ? 2 : RemainingSize);
2077 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2078 DAG.getConstant(Offset, MVT::i32));
2079 unsigned Alignment = std::min(ByValAlign, (unsigned )4);
2080 SDValue LoadVal = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
2081 LoadPtr, MachinePointerInfo(),
2082 MVT::getIntegerVT(LoadSize * 8), false,
2084 MemOpChains.push_back(LoadVal.getValue(1));
2086 // If target is big endian, shift it to the most significant half-word or
2089 LoadVal = DAG.getNode(ISD::SHL, dl, MVT::i32, LoadVal,
2090 DAG.getConstant(32 - LoadSize * 8, MVT::i32));
2093 RemainingSize -= LoadSize;
2095 // Read second subword if necessary.
2096 if (RemainingSize != 0) {
2097 assert(RemainingSize == 1 && "There must be one byte remaining.");
2098 LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2099 DAG.getConstant(Offset, MVT::i32));
2100 unsigned Alignment = std::min(ByValAlign, (unsigned )2);
2101 SDValue Subword = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
2102 LoadPtr, MachinePointerInfo(),
2103 MVT::i8, false, false, Alignment);
2104 MemOpChains.push_back(Subword.getValue(1));
2105 // Insert the loaded byte to LoadVal.
2106 // FIXME: Use INS if supported by target.
2107 unsigned ShiftAmt = isLittle ? 16 : 8;
2108 SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i32, Subword,
2109 DAG.getConstant(ShiftAmt, MVT::i32));
2110 LoadVal = DAG.getNode(ISD::OR, dl, MVT::i32, LoadVal, Shift);
2113 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
2114 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
2118 // Create a fixed object on stack at offset LocMemOffset and copy
2119 // remaining part of byval arg to it using memcpy.
2120 SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2121 DAG.getConstant(Offset, MVT::i32));
2122 LastFI = MFI->CreateFixedObject(RemainingSize, LocMemOffset, true);
2123 SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
2124 ByValChain = DAG.getMemcpy(ByValChain, dl, Dst, Src,
2125 DAG.getConstant(RemainingSize, MVT::i32),
2126 std::min(ByValAlign, (unsigned)4),
2127 /*isVolatile=*/false, /*AlwaysInline=*/false,
2128 MachinePointerInfo(0), MachinePointerInfo(0));
2131 // Copy Mips64 byVal arg to registers and stack.
2133 PassByValArg64(SDValue& ByValChain, SDValue Chain, DebugLoc dl,
2134 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
2135 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
2136 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
2137 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
2138 EVT PtrTy, bool isLittle) {
2139 unsigned ByValSize = Flags.getByValSize();
2140 unsigned Alignment = std::min(Flags.getByValAlign(), (unsigned)8);
2141 bool IsRegLoc = VA.isRegLoc();
2142 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
2143 unsigned LocMemOffset = 0;
2144 unsigned MemCpySize = ByValSize;
2147 LocMemOffset = VA.getLocMemOffset();
2149 const unsigned *Reg = std::find(Mips64IntRegs, Mips64IntRegs + 8,
2151 const unsigned *RegEnd = Mips64IntRegs + 8;
2153 // Copy double words to registers.
2154 for (; (Reg != RegEnd) && (ByValSize >= Offset + 8); ++Reg, Offset += 8) {
2155 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2156 DAG.getConstant(Offset, PtrTy));
2157 SDValue LoadVal = DAG.getLoad(MVT::i64, dl, Chain, LoadPtr,
2158 MachinePointerInfo(), false, false, false,
2160 MemOpChains.push_back(LoadVal.getValue(1));
2161 RegsToPass.push_back(std::make_pair(*Reg, LoadVal));
2164 // Return if the struct has been fully copied.
2165 if (!(MemCpySize = ByValSize - Offset))
2168 // If there is an argument register available, copy the remainder of the
2169 // byval argument with sub-doubleword loads and shifts.
2170 if (Reg != RegEnd) {
2171 assert((ByValSize < Offset + 8) &&
2172 "Size of the remainder should be smaller than 8-byte.");
2174 for (unsigned LoadSize = 4; Offset < ByValSize; LoadSize /= 2) {
2175 unsigned RemSize = ByValSize - Offset;
2177 if (RemSize < LoadSize)
2180 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2181 DAG.getConstant(Offset, PtrTy));
2183 DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i64, Chain, LoadPtr,
2184 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
2185 false, false, Alignment);
2186 MemOpChains.push_back(LoadVal.getValue(1));
2188 // Offset in number of bits from double word boundary.
2189 unsigned OffsetDW = (Offset % 8) * 8;
2190 unsigned Shamt = isLittle ? OffsetDW : 64 - (OffsetDW + LoadSize * 8);
2191 SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i64, LoadVal,
2192 DAG.getConstant(Shamt, MVT::i32));
2194 Val = Val.getNode() ? DAG.getNode(ISD::OR, dl, MVT::i64, Val, Shift) :
2197 Alignment = std::min(Alignment, LoadSize);
2200 RegsToPass.push_back(std::make_pair(*Reg, Val));
2205 assert(MemCpySize && "MemCpySize must not be zero.");
2207 // Create a fixed object on stack at offset LocMemOffset and copy
2208 // remainder of byval arg to it with memcpy.
2209 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2210 DAG.getConstant(Offset, PtrTy));
2211 LastFI = MFI->CreateFixedObject(MemCpySize, LocMemOffset, true);
2212 SDValue Dst = DAG.getFrameIndex(LastFI, PtrTy);
2213 ByValChain = DAG.getMemcpy(ByValChain, dl, Dst, Src,
2214 DAG.getConstant(MemCpySize, PtrTy), Alignment,
2215 /*isVolatile=*/false, /*AlwaysInline=*/false,
2216 MachinePointerInfo(0), MachinePointerInfo(0));
2219 /// LowerCall - functions arguments are copied from virtual regs to
2220 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2221 /// TODO: isTailCall.
2223 MipsTargetLowering::LowerCall(SDValue InChain, SDValue Callee,
2224 CallingConv::ID CallConv, bool isVarArg,
2225 bool doesNotRet, bool &isTailCall,
2226 const SmallVectorImpl<ISD::OutputArg> &Outs,
2227 const SmallVectorImpl<SDValue> &OutVals,
2228 const SmallVectorImpl<ISD::InputArg> &Ins,
2229 DebugLoc dl, SelectionDAG &DAG,
2230 SmallVectorImpl<SDValue> &InVals) const {
2231 // MIPs target does not yet support tail call optimization.
2234 MachineFunction &MF = DAG.getMachineFunction();
2235 MachineFrameInfo *MFI = MF.getFrameInfo();
2236 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2237 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2238 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2240 // Analyze operands of the call, assigning locations to each operand.
2241 SmallVector<CCValAssign, 16> ArgLocs;
2242 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2243 getTargetMachine(), ArgLocs, *DAG.getContext());
2246 CCInfo.AnalyzeCallOperands(Outs, CC_MipsO32);
2248 AnalyzeMips64CallOperands(CCInfo, Outs);
2250 CCInfo.AnalyzeCallOperands(Outs, CC_Mips);
2252 // Get a count of how many bytes are to be pushed on the stack.
2253 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2255 // Chain is the output chain of the last Load/Store or CopyToReg node.
2256 // ByValChain is the output chain of the last Memcpy node created for copying
2257 // byval arguments to the stack.
2258 SDValue Chain, CallSeqStart, ByValChain;
2259 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2260 Chain = CallSeqStart = DAG.getCALLSEQ_START(InChain, NextStackOffsetVal);
2261 ByValChain = InChain;
2263 // If this is the first call, create a stack frame object that points to
2264 // a location to which .cprestore saves $gp.
2265 if (IsO32 && IsPIC && MipsFI->globalBaseRegFixed() && !MipsFI->getGPFI())
2266 MipsFI->setGPFI(MFI->CreateFixedObject(4, 0, true));
2268 // Get the frame index of the stack frame object that points to the location
2269 // of dynamically allocated area on the stack.
2270 int DynAllocFI = MipsFI->getDynAllocFI();
2272 // Update size of the maximum argument space.
2273 // For O32, a minimum of four words (16 bytes) of argument space is
2276 NextStackOffset = std::max(NextStackOffset, (unsigned)16);
2278 unsigned MaxCallFrameSize = MipsFI->getMaxCallFrameSize();
2280 if (MaxCallFrameSize < NextStackOffset) {
2281 MipsFI->setMaxCallFrameSize(NextStackOffset);
2283 // Set the offsets relative to $sp of the $gp restore slot and dynamically
2284 // allocated stack space. These offsets must be aligned to a boundary
2285 // determined by the stack alignment of the ABI.
2286 unsigned StackAlignment = TFL->getStackAlignment();
2287 NextStackOffset = (NextStackOffset + StackAlignment - 1) /
2288 StackAlignment * StackAlignment;
2290 if (MipsFI->needGPSaveRestore())
2291 MFI->setObjectOffset(MipsFI->getGPFI(), NextStackOffset);
2293 MFI->setObjectOffset(DynAllocFI, NextStackOffset);
2296 // With EABI is it possible to have 16 args on registers.
2297 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
2298 SmallVector<SDValue, 8> MemOpChains;
2300 int FirstFI = -MFI->getNumFixedObjects() - 1, LastFI = 0;
2302 // Walk the register/memloc assignments, inserting copies/loads.
2303 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2304 SDValue Arg = OutVals[i];
2305 CCValAssign &VA = ArgLocs[i];
2306 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2307 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2310 if (Flags.isByVal()) {
2311 assert(Flags.getByValSize() &&
2312 "ByVal args of size 0 should have been ignored by front-end.");
2314 WriteByValArg(ByValChain, Chain, dl, RegsToPass, MemOpChains, LastFI,
2315 MFI, DAG, Arg, VA, Flags, getPointerTy(),
2316 Subtarget->isLittle());
2318 PassByValArg64(ByValChain, Chain, dl, RegsToPass, MemOpChains, LastFI,
2319 MFI, DAG, Arg, VA, Flags, getPointerTy(),
2320 Subtarget->isLittle());
2324 // Promote the value if needed.
2325 switch (VA.getLocInfo()) {
2326 default: llvm_unreachable("Unknown loc info!");
2327 case CCValAssign::Full:
2328 if (VA.isRegLoc()) {
2329 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2330 (ValVT == MVT::f64 && LocVT == MVT::i64))
2331 Arg = DAG.getNode(ISD::BITCAST, dl, LocVT, Arg);
2332 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2333 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2334 Arg, DAG.getConstant(0, MVT::i32));
2335 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2336 Arg, DAG.getConstant(1, MVT::i32));
2337 if (!Subtarget->isLittle())
2339 unsigned LocRegLo = VA.getLocReg();
2340 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2341 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2342 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2347 case CCValAssign::SExt:
2348 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, LocVT, Arg);
2350 case CCValAssign::ZExt:
2351 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, LocVT, Arg);
2353 case CCValAssign::AExt:
2354 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, LocVT, Arg);
2358 // Arguments that can be passed on register must be kept at
2359 // RegsToPass vector
2360 if (VA.isRegLoc()) {
2361 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2365 // Register can't get to this point...
2366 assert(VA.isMemLoc());
2368 // Create the frame index object for this incoming parameter
2369 LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2370 VA.getLocMemOffset(), true);
2371 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2373 // emit ISD::STORE whichs stores the
2374 // parameter value to a stack Location
2375 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2376 MachinePointerInfo(), false, false, 0));
2379 // Extend range of indices of frame objects for outgoing arguments that were
2380 // created during this function call. Skip this step if no such objects were
2383 MipsFI->extendOutArgFIRange(FirstFI, LastFI);
2385 // If a memcpy has been created to copy a byval arg to a stack, replace the
2386 // chain input of CallSeqStart with ByValChain.
2387 if (InChain != ByValChain)
2388 DAG.UpdateNodeOperands(CallSeqStart.getNode(), ByValChain,
2389 NextStackOffsetVal);
2391 // Transform all store nodes into one single node because all store
2392 // nodes are independent of each other.
2393 if (!MemOpChains.empty())
2394 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2395 &MemOpChains[0], MemOpChains.size());
2397 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2398 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2399 // node so that legalize doesn't hack it.
2400 unsigned char OpFlag;
2401 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2402 bool GlobalOrExternal = false;
2405 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2406 if (IsPICCall && G->getGlobal()->hasInternalLinkage()) {
2407 OpFlag = IsO32 ? MipsII::MO_GOT : MipsII::MO_GOT_PAGE;
2408 unsigned char LoFlag = IsO32 ? MipsII::MO_ABS_LO : MipsII::MO_GOT_OFST;
2409 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(), 0,
2411 CalleeLo = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(),
2414 OpFlag = IsPICCall ? MipsII::MO_GOT_CALL : MipsII::MO_NO_FLAG;
2415 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2416 getPointerTy(), 0, OpFlag);
2419 GlobalOrExternal = true;
2421 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2422 if (IsN64 || (!IsO32 && IsPIC))
2423 OpFlag = MipsII::MO_GOT_DISP;
2424 else if (!IsPIC) // !N64 && static
2425 OpFlag = MipsII::MO_NO_FLAG;
2427 OpFlag = MipsII::MO_GOT_CALL;
2428 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2430 GlobalOrExternal = true;
2435 // Create nodes that load address of callee and copy it to T9
2437 if (GlobalOrExternal) {
2438 // Load callee address
2439 Callee = DAG.getNode(MipsISD::Wrapper, dl, getPointerTy(),
2440 GetGlobalReg(DAG, getPointerTy()), Callee);
2441 SDValue LoadValue = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
2442 Callee, MachinePointerInfo::getGOT(),
2443 false, false, false, 0);
2445 // Use GOT+LO if callee has internal linkage.
2446 if (CalleeLo.getNode()) {
2447 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, getPointerTy(), CalleeLo);
2448 Callee = DAG.getNode(ISD::ADD, dl, getPointerTy(), LoadValue, Lo);
2454 // T9 should contain the address of the callee function if
2455 // -reloction-model=pic or it is an indirect call.
2456 if (IsPICCall || !GlobalOrExternal) {
2458 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
2459 Chain = DAG.getCopyToReg(Chain, dl, T9Reg, Callee, SDValue(0, 0));
2460 InFlag = Chain.getValue(1);
2461 Callee = DAG.getRegister(T9Reg, getPointerTy());
2464 // Build a sequence of copy-to-reg nodes chained together with token
2465 // chain and flag operands which copy the outgoing args into registers.
2466 // The InFlag in necessary since all emitted instructions must be
2468 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2469 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2470 RegsToPass[i].second, InFlag);
2471 InFlag = Chain.getValue(1);
2474 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2475 // = Chain, Callee, Reg#1, Reg#2, ...
2477 // Returns a chain & a flag for retval copy to use.
2478 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2479 SmallVector<SDValue, 8> Ops;
2480 Ops.push_back(Chain);
2481 Ops.push_back(Callee);
2483 // Add argument registers to the end of the list so that they are
2484 // known live into the call.
2485 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2486 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2487 RegsToPass[i].second.getValueType()));
2489 // Add a register mask operand representing the call-preserved registers.
2490 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2491 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2492 assert(Mask && "Missing call preserved mask for calling convention");
2493 Ops.push_back(DAG.getRegisterMask(Mask));
2495 if (InFlag.getNode())
2496 Ops.push_back(InFlag);
2498 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2499 InFlag = Chain.getValue(1);
2501 // Create the CALLSEQ_END node.
2502 Chain = DAG.getCALLSEQ_END(Chain,
2503 DAG.getIntPtrConstant(NextStackOffset, true),
2504 DAG.getIntPtrConstant(0, true), InFlag);
2505 InFlag = Chain.getValue(1);
2507 // Handle result values, copying them out of physregs into vregs that we
2509 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2510 Ins, dl, DAG, InVals);
2513 /// LowerCallResult - Lower the result values of a call into the
2514 /// appropriate copies out of appropriate physical registers.
2516 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2517 CallingConv::ID CallConv, bool isVarArg,
2518 const SmallVectorImpl<ISD::InputArg> &Ins,
2519 DebugLoc dl, SelectionDAG &DAG,
2520 SmallVectorImpl<SDValue> &InVals) const {
2521 // Assign locations to each value returned by this call.
2522 SmallVector<CCValAssign, 16> RVLocs;
2523 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2524 getTargetMachine(), RVLocs, *DAG.getContext());
2526 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2528 // Copy all of the result registers out of their specified physreg.
2529 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2530 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2531 RVLocs[i].getValVT(), InFlag).getValue(1);
2532 InFlag = Chain.getValue(2);
2533 InVals.push_back(Chain.getValue(0));
2539 //===----------------------------------------------------------------------===//
2540 // Formal Arguments Calling Convention Implementation
2541 //===----------------------------------------------------------------------===//
2542 static void ReadByValArg(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2543 std::vector<SDValue>& OutChains,
2544 SelectionDAG &DAG, unsigned NumWords, SDValue FIN,
2545 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags) {
2546 unsigned LocMem = VA.getLocMemOffset();
2547 unsigned FirstWord = LocMem / 4;
2549 // copy register A0 - A3 to frame object
2550 for (unsigned i = 0; i < NumWords; ++i) {
2551 unsigned CurWord = FirstWord + i;
2552 if (CurWord >= O32IntRegsSize)
2555 unsigned SrcReg = O32IntRegs[CurWord];
2556 unsigned Reg = AddLiveIn(MF, SrcReg, Mips::CPURegsRegisterClass);
2557 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIN,
2558 DAG.getConstant(i * 4, MVT::i32));
2559 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(Reg, MVT::i32),
2560 StorePtr, MachinePointerInfo(), false,
2562 OutChains.push_back(Store);
2566 // Create frame object on stack and copy registers used for byval passing to it.
2568 CopyMips64ByValRegs(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2569 std::vector<SDValue>& OutChains, SelectionDAG &DAG,
2570 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
2571 MachineFrameInfo *MFI, bool IsRegLoc,
2572 SmallVectorImpl<SDValue> &InVals, MipsFunctionInfo *MipsFI,
2574 const unsigned *Reg = Mips64IntRegs + 8;
2575 int FOOffset; // Frame object offset from virtual frame pointer.
2578 Reg = std::find(Mips64IntRegs, Mips64IntRegs + 8, VA.getLocReg());
2579 FOOffset = (Reg - Mips64IntRegs) * 8 - 8 * 8;
2582 FOOffset = VA.getLocMemOffset();
2584 // Create frame object.
2585 unsigned NumRegs = (Flags.getByValSize() + 7) / 8;
2586 unsigned LastFI = MFI->CreateFixedObject(NumRegs * 8, FOOffset, true);
2587 SDValue FIN = DAG.getFrameIndex(LastFI, PtrTy);
2588 InVals.push_back(FIN);
2590 // Copy arg registers.
2591 for (unsigned I = 0; (Reg != Mips64IntRegs + 8) && (I < NumRegs);
2593 unsigned VReg = AddLiveIn(MF, *Reg, Mips::CPU64RegsRegisterClass);
2594 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, PtrTy, FIN,
2595 DAG.getConstant(I * 8, PtrTy));
2596 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(VReg, MVT::i64),
2597 StorePtr, MachinePointerInfo(), false,
2599 OutChains.push_back(Store);
2605 /// LowerFormalArguments - transform physical registers into virtual registers
2606 /// and generate load operations for arguments places on the stack.
2608 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2609 CallingConv::ID CallConv,
2611 const SmallVectorImpl<ISD::InputArg> &Ins,
2612 DebugLoc dl, SelectionDAG &DAG,
2613 SmallVectorImpl<SDValue> &InVals)
2615 MachineFunction &MF = DAG.getMachineFunction();
2616 MachineFrameInfo *MFI = MF.getFrameInfo();
2617 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2619 MipsFI->setVarArgsFrameIndex(0);
2621 // Used with vargs to acumulate store chains.
2622 std::vector<SDValue> OutChains;
2624 // Assign locations to all of the incoming arguments.
2625 SmallVector<CCValAssign, 16> ArgLocs;
2626 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2627 getTargetMachine(), ArgLocs, *DAG.getContext());
2630 CCInfo.AnalyzeFormalArguments(Ins, CC_MipsO32);
2632 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips);
2634 int LastFI = 0;// MipsFI->LastInArgFI is 0 at the entry of this function.
2636 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2637 CCValAssign &VA = ArgLocs[i];
2638 EVT ValVT = VA.getValVT();
2639 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2640 bool IsRegLoc = VA.isRegLoc();
2642 if (Flags.isByVal()) {
2643 assert(Flags.getByValSize() &&
2644 "ByVal args of size 0 should have been ignored by front-end.");
2646 unsigned NumWords = (Flags.getByValSize() + 3) / 4;
2647 LastFI = MFI->CreateFixedObject(NumWords * 4, VA.getLocMemOffset(),
2649 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2650 InVals.push_back(FIN);
2651 ReadByValArg(MF, Chain, dl, OutChains, DAG, NumWords, FIN, VA, Flags);
2653 LastFI = CopyMips64ByValRegs(MF, Chain, dl, OutChains, DAG, VA, Flags,
2654 MFI, IsRegLoc, InVals, MipsFI,
2659 // Arguments stored on registers
2661 EVT RegVT = VA.getLocVT();
2662 unsigned ArgReg = VA.getLocReg();
2663 const TargetRegisterClass *RC;
2665 if (RegVT == MVT::i32)
2666 RC = Mips::CPURegsRegisterClass;
2667 else if (RegVT == MVT::i64)
2668 RC = Mips::CPU64RegsRegisterClass;
2669 else if (RegVT == MVT::f32)
2670 RC = Mips::FGR32RegisterClass;
2671 else if (RegVT == MVT::f64)
2672 RC = HasMips64 ? Mips::FGR64RegisterClass : Mips::AFGR64RegisterClass;
2674 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2676 // Transform the arguments stored on
2677 // physical registers into virtual ones
2678 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2679 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2681 // If this is an 8 or 16-bit value, it has been passed promoted
2682 // to 32 bits. Insert an assert[sz]ext to capture this, then
2683 // truncate to the right size.
2684 if (VA.getLocInfo() != CCValAssign::Full) {
2685 unsigned Opcode = 0;
2686 if (VA.getLocInfo() == CCValAssign::SExt)
2687 Opcode = ISD::AssertSext;
2688 else if (VA.getLocInfo() == CCValAssign::ZExt)
2689 Opcode = ISD::AssertZext;
2691 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
2692 DAG.getValueType(ValVT));
2693 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue);
2696 // Handle floating point arguments passed in integer registers.
2697 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2698 (RegVT == MVT::i64 && ValVT == MVT::f64))
2699 ArgValue = DAG.getNode(ISD::BITCAST, dl, ValVT, ArgValue);
2700 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2701 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
2702 getNextIntArgReg(ArgReg), RC);
2703 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
2704 if (!Subtarget->isLittle())
2705 std::swap(ArgValue, ArgValue2);
2706 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
2707 ArgValue, ArgValue2);
2710 InVals.push_back(ArgValue);
2711 } else { // VA.isRegLoc()
2714 assert(VA.isMemLoc());
2716 // The stack pointer offset is relative to the caller stack frame.
2717 LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2718 VA.getLocMemOffset(), true);
2720 // Create load nodes to retrieve arguments from the stack
2721 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2722 InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN,
2723 MachinePointerInfo::getFixedStack(LastFI),
2724 false, false, false, 0));
2728 // The mips ABIs for returning structs by value requires that we copy
2729 // the sret argument into $v0 for the return. Save the argument into
2730 // a virtual register so that we can access it from the return points.
2731 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2732 unsigned Reg = MipsFI->getSRetReturnReg();
2734 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i32));
2735 MipsFI->setSRetReturnReg(Reg);
2737 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2738 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2742 unsigned NumOfRegs = IsO32 ? 4 : 8;
2743 const unsigned *ArgRegs = IsO32 ? O32IntRegs : Mips64IntRegs;
2744 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumOfRegs);
2745 int FirstRegSlotOffset = IsO32 ? 0 : -64 ; // offset of $a0's slot.
2746 const TargetRegisterClass *RC
2747 = IsO32 ? Mips::CPURegsRegisterClass : Mips::CPU64RegsRegisterClass;
2748 unsigned RegSize = RC->getSize();
2749 int RegSlotOffset = FirstRegSlotOffset + Idx * RegSize;
2751 // Offset of the first variable argument from stack pointer.
2752 int FirstVaArgOffset;
2754 if (IsO32 || (Idx == NumOfRegs)) {
2756 (CCInfo.getNextStackOffset() + RegSize - 1) / RegSize * RegSize;
2758 FirstVaArgOffset = RegSlotOffset;
2760 // Record the frame index of the first variable argument
2761 // which is a value necessary to VASTART.
2762 LastFI = MFI->CreateFixedObject(RegSize, FirstVaArgOffset, true);
2763 MipsFI->setVarArgsFrameIndex(LastFI);
2765 // Copy the integer registers that have not been used for argument passing
2766 // to the argument register save area. For O32, the save area is allocated
2767 // in the caller's stack frame, while for N32/64, it is allocated in the
2768 // callee's stack frame.
2769 for (int StackOffset = RegSlotOffset;
2770 Idx < NumOfRegs; ++Idx, StackOffset += RegSize) {
2771 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgRegs[Idx], RC);
2772 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg,
2773 MVT::getIntegerVT(RegSize * 8));
2774 LastFI = MFI->CreateFixedObject(RegSize, StackOffset, true);
2775 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2776 OutChains.push_back(DAG.getStore(Chain, dl, ArgValue, PtrOff,
2777 MachinePointerInfo(), false, false, 0));
2781 MipsFI->setLastInArgFI(LastFI);
2783 // All stores are grouped in one node to allow the matching between
2784 // the size of Ins and InVals. This only happens when on varg functions
2785 if (!OutChains.empty()) {
2786 OutChains.push_back(Chain);
2787 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2788 &OutChains[0], OutChains.size());
2794 //===----------------------------------------------------------------------===//
2795 // Return Value Calling Convention Implementation
2796 //===----------------------------------------------------------------------===//
2799 MipsTargetLowering::LowerReturn(SDValue Chain,
2800 CallingConv::ID CallConv, bool isVarArg,
2801 const SmallVectorImpl<ISD::OutputArg> &Outs,
2802 const SmallVectorImpl<SDValue> &OutVals,
2803 DebugLoc dl, SelectionDAG &DAG) const {
2805 // CCValAssign - represent the assignment of
2806 // the return value to a location
2807 SmallVector<CCValAssign, 16> RVLocs;
2809 // CCState - Info about the registers and stack slot.
2810 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2811 getTargetMachine(), RVLocs, *DAG.getContext());
2813 // Analize return values.
2814 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
2816 // If this is the first return lowered for this function, add
2817 // the regs to the liveout set for the function.
2818 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2819 for (unsigned i = 0; i != RVLocs.size(); ++i)
2820 if (RVLocs[i].isRegLoc())
2821 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2826 // Copy the result values into the output registers.
2827 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2828 CCValAssign &VA = RVLocs[i];
2829 assert(VA.isRegLoc() && "Can only return in registers!");
2831 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
2833 // guarantee that all emitted copies are
2834 // stuck together, avoiding something bad
2835 Flag = Chain.getValue(1);
2838 // The mips ABIs for returning structs by value requires that we copy
2839 // the sret argument into $v0 for the return. We saved the argument into
2840 // a virtual register in the entry block, so now we copy the value out
2842 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2843 MachineFunction &MF = DAG.getMachineFunction();
2844 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2845 unsigned Reg = MipsFI->getSRetReturnReg();
2848 llvm_unreachable("sret virtual register not created in the entry block");
2849 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2851 Chain = DAG.getCopyToReg(Chain, dl, Mips::V0, Val, Flag);
2852 Flag = Chain.getValue(1);
2855 // Return on Mips is always a "jr $ra"
2857 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2858 Chain, DAG.getRegister(Mips::RA, MVT::i32), Flag);
2860 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2861 Chain, DAG.getRegister(Mips::RA, MVT::i32));
2864 //===----------------------------------------------------------------------===//
2865 // Mips Inline Assembly Support
2866 //===----------------------------------------------------------------------===//
2868 /// getConstraintType - Given a constraint letter, return the type of
2869 /// constraint it is for this target.
2870 MipsTargetLowering::ConstraintType MipsTargetLowering::
2871 getConstraintType(const std::string &Constraint) const
2873 // Mips specific constrainy
2874 // GCC config/mips/constraints.md
2876 // 'd' : An address register. Equivalent to r
2877 // unless generating MIPS16 code.
2878 // 'y' : Equivalent to r; retained for
2879 // backwards compatibility.
2880 // 'f' : Floating Point registers.
2881 if (Constraint.size() == 1) {
2882 switch (Constraint[0]) {
2887 return C_RegisterClass;
2890 return TargetLowering::getConstraintType(Constraint);
2893 /// Examine constraint type and operand type and determine a weight value.
2894 /// This object must already have been set up with the operand type
2895 /// and the current alternative constraint selected.
2896 TargetLowering::ConstraintWeight
2897 MipsTargetLowering::getSingleConstraintMatchWeight(
2898 AsmOperandInfo &info, const char *constraint) const {
2899 ConstraintWeight weight = CW_Invalid;
2900 Value *CallOperandVal = info.CallOperandVal;
2901 // If we don't have a value, we can't do a match,
2902 // but allow it at the lowest weight.
2903 if (CallOperandVal == NULL)
2905 Type *type = CallOperandVal->getType();
2906 // Look at the constraint type.
2907 switch (*constraint) {
2909 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2913 if (type->isIntegerTy())
2914 weight = CW_Register;
2917 if (type->isFloatTy())
2918 weight = CW_Register;
2924 /// Given a register class constraint, like 'r', if this corresponds directly
2925 /// to an LLVM register class, return a register of 0 and the register class
2927 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2928 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
2930 if (Constraint.size() == 1) {
2931 switch (Constraint[0]) {
2932 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2933 case 'y': // Same as 'r'. Exists for compatibility.
2936 return std::make_pair(0U, Mips::CPURegsRegisterClass);
2937 assert(VT == MVT::i64 && "Unexpected type.");
2938 return std::make_pair(0U, Mips::CPU64RegsRegisterClass);
2941 return std::make_pair(0U, Mips::FGR32RegisterClass);
2942 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
2943 if (Subtarget->isFP64bit())
2944 return std::make_pair(0U, Mips::FGR64RegisterClass);
2946 return std::make_pair(0U, Mips::AFGR64RegisterClass);
2950 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2954 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2955 // The Mips target isn't yet aware of offsets.
2959 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2960 if (VT != MVT::f32 && VT != MVT::f64)
2962 if (Imm.isNegZero())
2964 return Imm.isZero();
2967 unsigned MipsTargetLowering::getJumpTableEncoding() const {
2969 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
2971 return TargetLowering::getJumpTableEncoding();