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 "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/SelectionDAGISel.h"
33 #include "llvm/CodeGen/ValueTypes.h"
34 #include "llvm/Support/Debug.h"
35 #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 bool IsShiftedMask(uint64_t I, unsigned SizeInBits, uint64_t &Pos,
43 assert(SizeInBits == 32 || SizeInBits == 64);
44 bool Is32Bits = (SizeInBits == 32);
46 if ((Is32Bits == 32 && !isShiftedMask_32(I)) ||
47 (!Is32Bits && !isShiftedMask_64(I)))
50 Size = Is32Bits ? CountPopulation_32(I) : CountPopulation_64(I);
51 Pos = Is32Bits ? CountTrailingZeros_32(I) : CountTrailingZeros_64(I);
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::TlsGd: return "MipsISD::TlsGd";
63 case MipsISD::TprelHi: return "MipsISD::TprelHi";
64 case MipsISD::TprelLo: return "MipsISD::TprelLo";
65 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
66 case MipsISD::Ret: return "MipsISD::Ret";
67 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
68 case MipsISD::FPCmp: return "MipsISD::FPCmp";
69 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
70 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
71 case MipsISD::FPRound: return "MipsISD::FPRound";
72 case MipsISD::MAdd: return "MipsISD::MAdd";
73 case MipsISD::MAddu: return "MipsISD::MAddu";
74 case MipsISD::MSub: return "MipsISD::MSub";
75 case MipsISD::MSubu: return "MipsISD::MSubu";
76 case MipsISD::DivRem: return "MipsISD::DivRem";
77 case MipsISD::DivRemU: return "MipsISD::DivRemU";
78 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
79 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
80 case MipsISD::WrapperPIC: return "MipsISD::WrapperPIC";
81 case MipsISD::DynAlloc: return "MipsISD::DynAlloc";
82 case MipsISD::Sync: return "MipsISD::Sync";
83 case MipsISD::Ext: return "MipsISD::Ext";
84 case MipsISD::Ins: return "MipsISD::Ins";
90 MipsTargetLowering(MipsTargetMachine &TM)
91 : TargetLowering(TM, new MipsTargetObjectFile()) {
92 Subtarget = &TM.getSubtarget<MipsSubtarget>();
94 // Mips does not have i1 type, so use i32 for
95 // setcc operations results (slt, sgt, ...).
96 setBooleanContents(ZeroOrOneBooleanContent);
98 // Set up the register classes
99 addRegisterClass(MVT::i32, Mips::CPURegsRegisterClass);
100 addRegisterClass(MVT::f32, Mips::FGR32RegisterClass);
102 // When dealing with single precision only, use libcalls
103 if (!Subtarget->isSingleFloat())
104 if (!Subtarget->isFP64bit())
105 addRegisterClass(MVT::f64, Mips::AFGR64RegisterClass);
107 // Load extented operations for i1 types must be promoted
108 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
109 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
110 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
112 // MIPS doesn't have extending float->double load/store
113 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
114 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
116 // Used by legalize types to correctly generate the setcc result.
117 // Without this, every float setcc comes with a AND/OR with the result,
118 // we don't want this, since the fpcmp result goes to a flag register,
119 // which is used implicitly by brcond and select operations.
120 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
122 // Mips Custom Operations
123 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
124 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
125 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
126 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
127 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
128 setOperationAction(ISD::SELECT, MVT::f32, Custom);
129 setOperationAction(ISD::SELECT, MVT::f64, Custom);
130 setOperationAction(ISD::SELECT, MVT::i32, Custom);
131 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
132 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
133 setOperationAction(ISD::VASTART, MVT::Other, Custom);
135 setOperationAction(ISD::SDIV, MVT::i32, Expand);
136 setOperationAction(ISD::SREM, MVT::i32, Expand);
137 setOperationAction(ISD::UDIV, MVT::i32, Expand);
138 setOperationAction(ISD::UREM, MVT::i32, Expand);
140 // Operations not directly supported by Mips.
141 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
142 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
143 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
144 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
145 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
146 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
147 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
148 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
149 setOperationAction(ISD::ROTL, MVT::i32, Expand);
151 if (!Subtarget->isMips32r2())
152 setOperationAction(ISD::ROTR, MVT::i32, Expand);
154 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
155 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
156 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
157 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
158 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
159 setOperationAction(ISD::FSIN, MVT::f32, Expand);
160 setOperationAction(ISD::FSIN, MVT::f64, Expand);
161 setOperationAction(ISD::FCOS, MVT::f32, Expand);
162 setOperationAction(ISD::FCOS, MVT::f64, Expand);
163 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
164 setOperationAction(ISD::FPOW, MVT::f32, Expand);
165 setOperationAction(ISD::FPOW, MVT::f64, Expand);
166 setOperationAction(ISD::FLOG, MVT::f32, Expand);
167 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
168 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
169 setOperationAction(ISD::FEXP, MVT::f32, Expand);
170 setOperationAction(ISD::FMA, MVT::f32, Expand);
171 setOperationAction(ISD::FMA, MVT::f64, Expand);
173 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
174 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
176 setOperationAction(ISD::VAARG, MVT::Other, Expand);
177 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
178 setOperationAction(ISD::VAEND, MVT::Other, Expand);
180 // Use the default for now
181 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
182 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
184 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
185 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
187 setInsertFencesForAtomic(true);
189 if (Subtarget->isSingleFloat())
190 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
192 if (!Subtarget->hasSEInReg()) {
193 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
194 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
197 if (!Subtarget->hasBitCount())
198 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
200 if (!Subtarget->hasSwap())
201 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
203 setTargetDAGCombine(ISD::ADDE);
204 setTargetDAGCombine(ISD::SUBE);
205 setTargetDAGCombine(ISD::SDIVREM);
206 setTargetDAGCombine(ISD::UDIVREM);
207 setTargetDAGCombine(ISD::SETCC);
208 setTargetDAGCombine(ISD::AND);
209 setTargetDAGCombine(ISD::OR);
211 setMinFunctionAlignment(2);
213 setStackPointerRegisterToSaveRestore(Mips::SP);
214 computeRegisterProperties();
216 setExceptionPointerRegister(Mips::A0);
217 setExceptionSelectorRegister(Mips::A1);
220 bool MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
221 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
222 return SVT == MVT::i32 || SVT == MVT::i16;
225 MVT::SimpleValueType MipsTargetLowering::getSetCCResultType(EVT VT) const {
230 // Transforms a subgraph in CurDAG if the following pattern is found:
231 // (addc multLo, Lo0), (adde multHi, Hi0),
233 // multHi/Lo: product of multiplication
234 // Lo0: initial value of Lo register
235 // Hi0: initial value of Hi register
236 // Return true if pattern matching was successful.
237 static bool SelectMadd(SDNode* ADDENode, SelectionDAG* CurDAG) {
238 // ADDENode's second operand must be a flag output of an ADDC node in order
239 // for the matching to be successful.
240 SDNode* ADDCNode = ADDENode->getOperand(2).getNode();
242 if (ADDCNode->getOpcode() != ISD::ADDC)
245 SDValue MultHi = ADDENode->getOperand(0);
246 SDValue MultLo = ADDCNode->getOperand(0);
247 SDNode* MultNode = MultHi.getNode();
248 unsigned MultOpc = MultHi.getOpcode();
250 // MultHi and MultLo must be generated by the same node,
251 if (MultLo.getNode() != MultNode)
254 // and it must be a multiplication.
255 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
258 // MultLo amd MultHi must be the first and second output of MultNode
260 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
263 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
264 // of the values of MultNode, in which case MultNode will be removed in later
266 // If there exist users other than ADDENode or ADDCNode, this function returns
267 // here, which will result in MultNode being mapped to a single MULT
268 // instruction node rather than a pair of MULT and MADD instructions being
270 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
273 SDValue Chain = CurDAG->getEntryNode();
274 DebugLoc dl = ADDENode->getDebugLoc();
276 // create MipsMAdd(u) node
277 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
279 SDValue MAdd = CurDAG->getNode(MultOpc, dl,
281 MultNode->getOperand(0),// Factor 0
282 MultNode->getOperand(1),// Factor 1
283 ADDCNode->getOperand(1),// Lo0
284 ADDENode->getOperand(1));// Hi0
286 // create CopyFromReg nodes
287 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
289 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
291 CopyFromLo.getValue(2));
293 // replace uses of adde and addc here
294 if (!SDValue(ADDCNode, 0).use_empty())
295 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
297 if (!SDValue(ADDENode, 0).use_empty())
298 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
304 // Transforms a subgraph in CurDAG if the following pattern is found:
305 // (addc Lo0, multLo), (sube Hi0, multHi),
307 // multHi/Lo: product of multiplication
308 // Lo0: initial value of Lo register
309 // Hi0: initial value of Hi register
310 // Return true if pattern matching was successful.
311 static bool SelectMsub(SDNode* SUBENode, SelectionDAG* CurDAG) {
312 // SUBENode's second operand must be a flag output of an SUBC node in order
313 // for the matching to be successful.
314 SDNode* SUBCNode = SUBENode->getOperand(2).getNode();
316 if (SUBCNode->getOpcode() != ISD::SUBC)
319 SDValue MultHi = SUBENode->getOperand(1);
320 SDValue MultLo = SUBCNode->getOperand(1);
321 SDNode* MultNode = MultHi.getNode();
322 unsigned MultOpc = MultHi.getOpcode();
324 // MultHi and MultLo must be generated by the same node,
325 if (MultLo.getNode() != MultNode)
328 // and it must be a multiplication.
329 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
332 // MultLo amd MultHi must be the first and second output of MultNode
334 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
337 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
338 // of the values of MultNode, in which case MultNode will be removed in later
340 // If there exist users other than SUBENode or SUBCNode, this function returns
341 // here, which will result in MultNode being mapped to a single MULT
342 // instruction node rather than a pair of MULT and MSUB instructions being
344 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
347 SDValue Chain = CurDAG->getEntryNode();
348 DebugLoc dl = SUBENode->getDebugLoc();
350 // create MipsSub(u) node
351 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
353 SDValue MSub = CurDAG->getNode(MultOpc, dl,
355 MultNode->getOperand(0),// Factor 0
356 MultNode->getOperand(1),// Factor 1
357 SUBCNode->getOperand(0),// Lo0
358 SUBENode->getOperand(0));// Hi0
360 // create CopyFromReg nodes
361 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
363 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
365 CopyFromLo.getValue(2));
367 // replace uses of sube and subc here
368 if (!SDValue(SUBCNode, 0).use_empty())
369 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
371 if (!SDValue(SUBENode, 0).use_empty())
372 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
377 static SDValue PerformADDECombine(SDNode *N, SelectionDAG& DAG,
378 TargetLowering::DAGCombinerInfo &DCI,
379 const MipsSubtarget* Subtarget) {
380 if (DCI.isBeforeLegalize())
383 if (Subtarget->isMips32() && SelectMadd(N, &DAG))
384 return SDValue(N, 0);
389 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG& DAG,
390 TargetLowering::DAGCombinerInfo &DCI,
391 const MipsSubtarget* Subtarget) {
392 if (DCI.isBeforeLegalize())
395 if (Subtarget->isMips32() && SelectMsub(N, &DAG))
396 return SDValue(N, 0);
401 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG& DAG,
402 TargetLowering::DAGCombinerInfo &DCI,
403 const MipsSubtarget* Subtarget) {
404 if (DCI.isBeforeLegalizeOps())
407 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
409 DebugLoc dl = N->getDebugLoc();
411 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
412 N->getOperand(0), N->getOperand(1));
413 SDValue InChain = DAG.getEntryNode();
414 SDValue InGlue = DivRem;
417 if (N->hasAnyUseOfValue(0)) {
418 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, Mips::LO, MVT::i32,
420 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
421 InChain = CopyFromLo.getValue(1);
422 InGlue = CopyFromLo.getValue(2);
426 if (N->hasAnyUseOfValue(1)) {
427 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
428 Mips::HI, MVT::i32, InGlue);
429 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
435 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
437 default: llvm_unreachable("Unknown fp condition code!");
439 case ISD::SETOEQ: return Mips::FCOND_OEQ;
440 case ISD::SETUNE: return Mips::FCOND_UNE;
442 case ISD::SETOLT: return Mips::FCOND_OLT;
444 case ISD::SETOGT: return Mips::FCOND_OGT;
446 case ISD::SETOLE: return Mips::FCOND_OLE;
448 case ISD::SETOGE: return Mips::FCOND_OGE;
449 case ISD::SETULT: return Mips::FCOND_ULT;
450 case ISD::SETULE: return Mips::FCOND_ULE;
451 case ISD::SETUGT: return Mips::FCOND_UGT;
452 case ISD::SETUGE: return Mips::FCOND_UGE;
453 case ISD::SETUO: return Mips::FCOND_UN;
454 case ISD::SETO: return Mips::FCOND_OR;
456 case ISD::SETONE: return Mips::FCOND_ONE;
457 case ISD::SETUEQ: return Mips::FCOND_UEQ;
462 // Returns true if condition code has to be inverted.
463 static bool InvertFPCondCode(Mips::CondCode CC) {
464 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
467 if (CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT)
470 assert(false && "Illegal Condition Code");
474 // Creates and returns an FPCmp node from a setcc node.
475 // Returns Op if setcc is not a floating point comparison.
476 static SDValue CreateFPCmp(SelectionDAG& DAG, const SDValue& Op) {
477 // must be a SETCC node
478 if (Op.getOpcode() != ISD::SETCC)
481 SDValue LHS = Op.getOperand(0);
483 if (!LHS.getValueType().isFloatingPoint())
486 SDValue RHS = Op.getOperand(1);
487 DebugLoc dl = Op.getDebugLoc();
489 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
490 // node if necessary.
491 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
493 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
494 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
497 // Creates and returns a CMovFPT/F node.
498 static SDValue CreateCMovFP(SelectionDAG& DAG, SDValue Cond, SDValue True,
499 SDValue False, DebugLoc DL) {
500 bool invert = InvertFPCondCode((Mips::CondCode)
501 cast<ConstantSDNode>(Cond.getOperand(2))
504 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
505 True.getValueType(), True, False, Cond);
508 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG& DAG,
509 TargetLowering::DAGCombinerInfo &DCI,
510 const MipsSubtarget* Subtarget) {
511 if (DCI.isBeforeLegalizeOps())
514 SDValue Cond = CreateFPCmp(DAG, SDValue(N, 0));
516 if (Cond.getOpcode() != MipsISD::FPCmp)
519 SDValue True = DAG.getConstant(1, MVT::i32);
520 SDValue False = DAG.getConstant(0, MVT::i32);
522 return CreateCMovFP(DAG, Cond, True, False, N->getDebugLoc());
525 static SDValue PerformANDCombine(SDNode *N, SelectionDAG& DAG,
526 TargetLowering::DAGCombinerInfo &DCI,
527 const MipsSubtarget* Subtarget) {
528 // Pattern match EXT.
529 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
530 // => ext $dst, $src, size, pos
531 if (DCI.isBeforeLegalizeOps() || !Subtarget->isMips32r2())
534 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
536 // Op's first operand must be a shift right.
537 if (ShiftRight.getOpcode() != ISD::SRA && ShiftRight.getOpcode() != ISD::SRL)
540 // The second operand of the shift must be an immediate.
543 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
546 Pos = CN->getZExtValue();
548 uint64_t SMPos, SMSize;
549 // Op's second operand must be a shifted mask.
550 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
551 !IsShiftedMask(CN->getZExtValue(), 32, SMPos, SMSize))
554 // Return if the shifted mask does not start at bit 0 or the sum of its size
555 // and Pos exceeds the word's size.
556 if (SMPos != 0 || Pos + SMSize > 32)
559 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), MVT::i32,
560 ShiftRight.getOperand(0),
561 DAG.getConstant(Pos, MVT::i32),
562 DAG.getConstant(SMSize, MVT::i32));
565 static SDValue PerformORCombine(SDNode *N, SelectionDAG& DAG,
566 TargetLowering::DAGCombinerInfo &DCI,
567 const MipsSubtarget* Subtarget) {
568 // Pattern match INS.
569 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
570 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
571 // => ins $dst, $src, size, pos, $src1
572 if (DCI.isBeforeLegalizeOps() || !Subtarget->isMips32r2())
575 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
576 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
579 // See if Op's first operand matches (and $src1 , mask0).
580 if (And0.getOpcode() != ISD::AND)
583 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
584 !IsShiftedMask(~CN->getZExtValue(), 32, SMPos0, SMSize0))
587 // See if Op's second operand matches (and (shl $src, pos), mask1).
588 if (And1.getOpcode() != ISD::AND)
591 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
592 !IsShiftedMask(CN->getZExtValue(), CN->getValueSizeInBits(0), SMPos1,
596 // The shift masks must have the same position and size.
597 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
600 SDValue Shl = And1.getOperand(0);
601 if (Shl.getOpcode() != ISD::SHL)
604 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
607 unsigned Shamt = CN->getZExtValue();
609 // Return if the shift amount and the first bit position of mask are not the
614 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), MVT::i32,
616 DAG.getConstant(SMPos0, MVT::i32),
617 DAG.getConstant(SMSize0, MVT::i32),
621 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
623 SelectionDAG &DAG = DCI.DAG;
624 unsigned opc = N->getOpcode();
629 return PerformADDECombine(N, DAG, DCI, Subtarget);
631 return PerformSUBECombine(N, DAG, DCI, Subtarget);
634 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
636 return PerformSETCCCombine(N, DAG, DCI, Subtarget);
638 return PerformANDCombine(N, DAG, DCI, Subtarget);
640 return PerformORCombine(N, DAG, DCI, Subtarget);
646 SDValue MipsTargetLowering::
647 LowerOperation(SDValue Op, SelectionDAG &DAG) const
649 switch (Op.getOpcode())
651 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
652 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
653 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
654 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
655 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
656 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
657 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
658 case ISD::SELECT: return LowerSELECT(Op, DAG);
659 case ISD::VASTART: return LowerVASTART(Op, DAG);
660 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
661 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
662 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
663 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
668 //===----------------------------------------------------------------------===//
669 // Lower helper functions
670 //===----------------------------------------------------------------------===//
672 // AddLiveIn - This helper function adds the specified physical register to the
673 // MachineFunction as a live in value. It also creates a corresponding
674 // virtual register for it.
676 AddLiveIn(MachineFunction &MF, unsigned PReg, TargetRegisterClass *RC)
678 assert(RC->contains(PReg) && "Not the correct regclass!");
679 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
680 MF.getRegInfo().addLiveIn(PReg, VReg);
684 // Get fp branch code (not opcode) from condition code.
685 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
686 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
687 return Mips::BRANCH_T;
689 if (CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT)
690 return Mips::BRANCH_F;
692 return Mips::BRANCH_INVALID;
695 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
697 const MipsSubtarget* Subtarget,
698 const TargetInstrInfo *TII,
699 bool isFPCmp, unsigned Opc) {
700 // There is no need to expand CMov instructions if target has
701 // conditional moves.
702 if (Subtarget->hasCondMov())
705 // To "insert" a SELECT_CC instruction, we actually have to insert the
706 // diamond control-flow pattern. The incoming instruction knows the
707 // destination vreg to set, the condition code register to branch on, the
708 // true/false values to select between, and a branch opcode to use.
709 const BasicBlock *LLVM_BB = BB->getBasicBlock();
710 MachineFunction::iterator It = BB;
717 // bNE r1, r0, copy1MBB
718 // fallthrough --> copy0MBB
719 MachineBasicBlock *thisMBB = BB;
720 MachineFunction *F = BB->getParent();
721 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
722 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
723 F->insert(It, copy0MBB);
724 F->insert(It, sinkMBB);
726 // Transfer the remainder of BB and its successor edges to sinkMBB.
727 sinkMBB->splice(sinkMBB->begin(), BB,
728 llvm::next(MachineBasicBlock::iterator(MI)),
730 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
732 // Next, add the true and fallthrough blocks as its successors.
733 BB->addSuccessor(copy0MBB);
734 BB->addSuccessor(sinkMBB);
736 // Emit the right instruction according to the type of the operands compared
738 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
740 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
741 .addReg(Mips::ZERO).addMBB(sinkMBB);
745 // # fallthrough to sinkMBB
748 // Update machine-CFG edges
749 BB->addSuccessor(sinkMBB);
752 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
757 BuildMI(*BB, BB->begin(), dl,
758 TII->get(Mips::PHI), MI->getOperand(0).getReg())
759 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
760 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
762 BuildMI(*BB, BB->begin(), dl,
763 TII->get(Mips::PHI), MI->getOperand(0).getReg())
764 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
765 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
767 MI->eraseFromParent(); // The pseudo instruction is gone now.
772 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
773 MachineBasicBlock *BB) const {
774 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
775 DebugLoc dl = MI->getDebugLoc();
777 switch (MI->getOpcode()) {
779 assert(false && "Unexpected instr type to insert");
784 return ExpandCondMov(MI, BB, dl, Subtarget, TII, true, Mips::BC1F);
788 return ExpandCondMov(MI, BB, dl, Subtarget, TII, true, Mips::BC1T);
792 return ExpandCondMov(MI, BB, dl, Subtarget, TII, false, Mips::BNE);
796 return ExpandCondMov(MI, BB, dl, Subtarget, TII, false, Mips::BEQ);
798 case Mips::ATOMIC_LOAD_ADD_I8:
799 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
800 case Mips::ATOMIC_LOAD_ADD_I16:
801 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
802 case Mips::ATOMIC_LOAD_ADD_I32:
803 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
805 case Mips::ATOMIC_LOAD_AND_I8:
806 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
807 case Mips::ATOMIC_LOAD_AND_I16:
808 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
809 case Mips::ATOMIC_LOAD_AND_I32:
810 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
812 case Mips::ATOMIC_LOAD_OR_I8:
813 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
814 case Mips::ATOMIC_LOAD_OR_I16:
815 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
816 case Mips::ATOMIC_LOAD_OR_I32:
817 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
819 case Mips::ATOMIC_LOAD_XOR_I8:
820 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
821 case Mips::ATOMIC_LOAD_XOR_I16:
822 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
823 case Mips::ATOMIC_LOAD_XOR_I32:
824 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
826 case Mips::ATOMIC_LOAD_NAND_I8:
827 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
828 case Mips::ATOMIC_LOAD_NAND_I16:
829 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
830 case Mips::ATOMIC_LOAD_NAND_I32:
831 return EmitAtomicBinary(MI, BB, 4, 0, true);
833 case Mips::ATOMIC_LOAD_SUB_I8:
834 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
835 case Mips::ATOMIC_LOAD_SUB_I16:
836 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
837 case Mips::ATOMIC_LOAD_SUB_I32:
838 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
840 case Mips::ATOMIC_SWAP_I8:
841 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
842 case Mips::ATOMIC_SWAP_I16:
843 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
844 case Mips::ATOMIC_SWAP_I32:
845 return EmitAtomicBinary(MI, BB, 4, 0);
847 case Mips::ATOMIC_CMP_SWAP_I8:
848 return EmitAtomicCmpSwapPartword(MI, BB, 1);
849 case Mips::ATOMIC_CMP_SWAP_I16:
850 return EmitAtomicCmpSwapPartword(MI, BB, 2);
851 case Mips::ATOMIC_CMP_SWAP_I32:
852 return EmitAtomicCmpSwap(MI, BB, 4);
856 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
857 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
859 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
860 unsigned Size, unsigned BinOpcode,
862 assert(Size == 4 && "Unsupported size for EmitAtomicBinary.");
864 MachineFunction *MF = BB->getParent();
865 MachineRegisterInfo &RegInfo = MF->getRegInfo();
866 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
867 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
868 DebugLoc dl = MI->getDebugLoc();
870 unsigned OldVal = MI->getOperand(0).getReg();
871 unsigned Ptr = MI->getOperand(1).getReg();
872 unsigned Incr = MI->getOperand(2).getReg();
874 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
875 unsigned AndRes = RegInfo.createVirtualRegister(RC);
876 unsigned Success = RegInfo.createVirtualRegister(RC);
878 // insert new blocks after the current block
879 const BasicBlock *LLVM_BB = BB->getBasicBlock();
880 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
881 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
882 MachineFunction::iterator It = BB;
884 MF->insert(It, loopMBB);
885 MF->insert(It, exitMBB);
887 // Transfer the remainder of BB and its successor edges to exitMBB.
888 exitMBB->splice(exitMBB->begin(), BB,
889 llvm::next(MachineBasicBlock::iterator(MI)),
891 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
895 // fallthrough --> loopMBB
896 BB->addSuccessor(loopMBB);
897 loopMBB->addSuccessor(loopMBB);
898 loopMBB->addSuccessor(exitMBB);
902 // <binop> storeval, oldval, incr
903 // sc success, storeval, 0(ptr)
904 // beq success, $0, loopMBB
906 BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(Ptr).addImm(0);
908 // and andres, oldval, incr
909 // nor storeval, $0, andres
910 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr);
911 BuildMI(BB, dl, TII->get(Mips::NOR), StoreVal)
912 .addReg(Mips::ZERO).addReg(AndRes);
913 } else if (BinOpcode) {
914 // <binop> storeval, oldval, incr
915 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
919 BuildMI(BB, dl, TII->get(Mips::SC), Success)
920 .addReg(StoreVal).addReg(Ptr).addImm(0);
921 BuildMI(BB, dl, TII->get(Mips::BEQ))
922 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
924 MI->eraseFromParent(); // The instruction is gone now.
930 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
931 MachineBasicBlock *BB,
932 unsigned Size, unsigned BinOpcode,
934 assert((Size == 1 || Size == 2) &&
935 "Unsupported size for EmitAtomicBinaryPartial.");
937 MachineFunction *MF = BB->getParent();
938 MachineRegisterInfo &RegInfo = MF->getRegInfo();
939 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
940 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
941 DebugLoc dl = MI->getDebugLoc();
943 unsigned Dest = MI->getOperand(0).getReg();
944 unsigned Ptr = MI->getOperand(1).getReg();
945 unsigned Incr = MI->getOperand(2).getReg();
947 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
948 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
949 unsigned Mask = RegInfo.createVirtualRegister(RC);
950 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
951 unsigned NewVal = RegInfo.createVirtualRegister(RC);
952 unsigned OldVal = RegInfo.createVirtualRegister(RC);
953 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
954 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
955 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
956 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
957 unsigned AndRes = RegInfo.createVirtualRegister(RC);
958 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
959 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
960 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
961 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
962 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
963 unsigned SllRes = RegInfo.createVirtualRegister(RC);
964 unsigned Success = RegInfo.createVirtualRegister(RC);
966 // insert new blocks after the current block
967 const BasicBlock *LLVM_BB = BB->getBasicBlock();
968 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
969 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
970 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
971 MachineFunction::iterator It = BB;
973 MF->insert(It, loopMBB);
974 MF->insert(It, sinkMBB);
975 MF->insert(It, exitMBB);
977 // Transfer the remainder of BB and its successor edges to exitMBB.
978 exitMBB->splice(exitMBB->begin(), BB,
979 llvm::next(MachineBasicBlock::iterator(MI)),
981 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
983 BB->addSuccessor(loopMBB);
984 loopMBB->addSuccessor(loopMBB);
985 loopMBB->addSuccessor(sinkMBB);
986 sinkMBB->addSuccessor(exitMBB);
989 // addiu masklsb2,$0,-4 # 0xfffffffc
990 // and alignedaddr,ptr,masklsb2
991 // andi ptrlsb2,ptr,3
992 // sll shiftamt,ptrlsb2,3
993 // ori maskupper,$0,255 # 0xff
994 // sll mask,maskupper,shiftamt
996 // sll incr2,incr,shiftamt
998 int64_t MaskImm = (Size == 1) ? 255 : 65535;
999 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1000 .addReg(Mips::ZERO).addImm(-4);
1001 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1002 .addReg(Ptr).addReg(MaskLSB2);
1003 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1004 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1005 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1006 .addReg(Mips::ZERO).addImm(MaskImm);
1007 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1008 .addReg(ShiftAmt).addReg(MaskUpper);
1009 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1010 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1013 // atomic.load.binop
1015 // ll oldval,0(alignedaddr)
1016 // binop binopres,oldval,incr2
1017 // and newval,binopres,mask
1018 // and maskedoldval0,oldval,mask2
1019 // or storeval,maskedoldval0,newval
1020 // sc success,storeval,0(alignedaddr)
1021 // beq success,$0,loopMBB
1025 // ll oldval,0(alignedaddr)
1026 // and newval,incr2,mask
1027 // and maskedoldval0,oldval,mask2
1028 // or storeval,maskedoldval0,newval
1029 // sc success,storeval,0(alignedaddr)
1030 // beq success,$0,loopMBB
1033 BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1035 // and andres, oldval, incr2
1036 // nor binopres, $0, andres
1037 // and newval, binopres, mask
1038 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1039 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1040 .addReg(Mips::ZERO).addReg(AndRes);
1041 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1042 } else if (BinOpcode) {
1043 // <binop> binopres, oldval, incr2
1044 // and newval, binopres, mask
1045 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1046 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1047 } else {// atomic.swap
1048 // and newval, incr2, mask
1049 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1052 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1053 .addReg(OldVal).addReg(Mask2);
1054 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1055 .addReg(MaskedOldVal0).addReg(NewVal);
1056 BuildMI(BB, dl, TII->get(Mips::SC), Success)
1057 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1058 BuildMI(BB, dl, TII->get(Mips::BEQ))
1059 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1062 // and maskedoldval1,oldval,mask
1063 // srl srlres,maskedoldval1,shiftamt
1064 // sll sllres,srlres,24
1065 // sra dest,sllres,24
1067 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1069 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1070 .addReg(OldVal).addReg(Mask);
1071 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1072 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1073 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1074 .addReg(SrlRes).addImm(ShiftImm);
1075 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1076 .addReg(SllRes).addImm(ShiftImm);
1078 MI->eraseFromParent(); // The instruction is gone now.
1084 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1085 MachineBasicBlock *BB,
1086 unsigned Size) const {
1087 assert(Size == 4 && "Unsupported size for EmitAtomicCmpSwap.");
1089 MachineFunction *MF = BB->getParent();
1090 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1091 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1092 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1093 DebugLoc dl = MI->getDebugLoc();
1095 unsigned Dest = MI->getOperand(0).getReg();
1096 unsigned Ptr = MI->getOperand(1).getReg();
1097 unsigned OldVal = MI->getOperand(2).getReg();
1098 unsigned NewVal = MI->getOperand(3).getReg();
1100 unsigned Success = RegInfo.createVirtualRegister(RC);
1102 // insert new blocks after the current block
1103 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1104 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1105 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1106 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1107 MachineFunction::iterator It = BB;
1109 MF->insert(It, loop1MBB);
1110 MF->insert(It, loop2MBB);
1111 MF->insert(It, exitMBB);
1113 // Transfer the remainder of BB and its successor edges to exitMBB.
1114 exitMBB->splice(exitMBB->begin(), BB,
1115 llvm::next(MachineBasicBlock::iterator(MI)),
1117 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1121 // fallthrough --> loop1MBB
1122 BB->addSuccessor(loop1MBB);
1123 loop1MBB->addSuccessor(exitMBB);
1124 loop1MBB->addSuccessor(loop2MBB);
1125 loop2MBB->addSuccessor(loop1MBB);
1126 loop2MBB->addSuccessor(exitMBB);
1130 // bne dest, oldval, exitMBB
1132 BuildMI(BB, dl, TII->get(Mips::LL), Dest).addReg(Ptr).addImm(0);
1133 BuildMI(BB, dl, TII->get(Mips::BNE))
1134 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1137 // sc success, newval, 0(ptr)
1138 // beq success, $0, loop1MBB
1140 BuildMI(BB, dl, TII->get(Mips::SC), Success)
1141 .addReg(NewVal).addReg(Ptr).addImm(0);
1142 BuildMI(BB, dl, TII->get(Mips::BEQ))
1143 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1145 MI->eraseFromParent(); // The instruction is gone now.
1151 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1152 MachineBasicBlock *BB,
1153 unsigned Size) const {
1154 assert((Size == 1 || Size == 2) &&
1155 "Unsupported size for EmitAtomicCmpSwapPartial.");
1157 MachineFunction *MF = BB->getParent();
1158 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1159 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1160 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1161 DebugLoc dl = MI->getDebugLoc();
1163 unsigned Dest = MI->getOperand(0).getReg();
1164 unsigned Ptr = MI->getOperand(1).getReg();
1165 unsigned CmpVal = MI->getOperand(2).getReg();
1166 unsigned NewVal = MI->getOperand(3).getReg();
1168 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1169 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1170 unsigned Mask = RegInfo.createVirtualRegister(RC);
1171 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1172 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1173 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1174 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1175 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1176 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1177 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1178 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1179 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1180 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1181 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1182 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1183 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1184 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1185 unsigned Success = RegInfo.createVirtualRegister(RC);
1187 // insert new blocks after the current block
1188 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1189 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1190 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1191 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1192 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1193 MachineFunction::iterator It = BB;
1195 MF->insert(It, loop1MBB);
1196 MF->insert(It, loop2MBB);
1197 MF->insert(It, sinkMBB);
1198 MF->insert(It, exitMBB);
1200 // Transfer the remainder of BB and its successor edges to exitMBB.
1201 exitMBB->splice(exitMBB->begin(), BB,
1202 llvm::next(MachineBasicBlock::iterator(MI)),
1204 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1206 BB->addSuccessor(loop1MBB);
1207 loop1MBB->addSuccessor(sinkMBB);
1208 loop1MBB->addSuccessor(loop2MBB);
1209 loop2MBB->addSuccessor(loop1MBB);
1210 loop2MBB->addSuccessor(sinkMBB);
1211 sinkMBB->addSuccessor(exitMBB);
1213 // FIXME: computation of newval2 can be moved to loop2MBB.
1215 // addiu masklsb2,$0,-4 # 0xfffffffc
1216 // and alignedaddr,ptr,masklsb2
1217 // andi ptrlsb2,ptr,3
1218 // sll shiftamt,ptrlsb2,3
1219 // ori maskupper,$0,255 # 0xff
1220 // sll mask,maskupper,shiftamt
1221 // nor mask2,$0,mask
1222 // andi maskedcmpval,cmpval,255
1223 // sll shiftedcmpval,maskedcmpval,shiftamt
1224 // andi maskednewval,newval,255
1225 // sll shiftednewval,maskednewval,shiftamt
1226 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1227 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1228 .addReg(Mips::ZERO).addImm(-4);
1229 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1230 .addReg(Ptr).addReg(MaskLSB2);
1231 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1232 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1233 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1234 .addReg(Mips::ZERO).addImm(MaskImm);
1235 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1236 .addReg(ShiftAmt).addReg(MaskUpper);
1237 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1238 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1239 .addReg(CmpVal).addImm(MaskImm);
1240 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1241 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1242 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1243 .addReg(NewVal).addImm(MaskImm);
1244 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1245 .addReg(ShiftAmt).addReg(MaskedNewVal);
1248 // ll oldval,0(alginedaddr)
1249 // and maskedoldval0,oldval,mask
1250 // bne maskedoldval0,shiftedcmpval,sinkMBB
1252 BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1253 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1254 .addReg(OldVal).addReg(Mask);
1255 BuildMI(BB, dl, TII->get(Mips::BNE))
1256 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1259 // and maskedoldval1,oldval,mask2
1260 // or storeval,maskedoldval1,shiftednewval
1261 // sc success,storeval,0(alignedaddr)
1262 // beq success,$0,loop1MBB
1264 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1265 .addReg(OldVal).addReg(Mask2);
1266 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1267 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1268 BuildMI(BB, dl, TII->get(Mips::SC), Success)
1269 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1270 BuildMI(BB, dl, TII->get(Mips::BEQ))
1271 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1274 // srl srlres,maskedoldval0,shiftamt
1275 // sll sllres,srlres,24
1276 // sra dest,sllres,24
1278 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1280 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1281 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1282 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1283 .addReg(SrlRes).addImm(ShiftImm);
1284 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1285 .addReg(SllRes).addImm(ShiftImm);
1287 MI->eraseFromParent(); // The instruction is gone now.
1292 //===----------------------------------------------------------------------===//
1293 // Misc Lower Operation implementation
1294 //===----------------------------------------------------------------------===//
1295 SDValue MipsTargetLowering::
1296 LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const
1298 MachineFunction &MF = DAG.getMachineFunction();
1299 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1301 assert(getTargetMachine().getFrameLowering()->getStackAlignment() >=
1302 cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue() &&
1303 "Cannot lower if the alignment of the allocated space is larger than \
1304 that of the stack.");
1306 SDValue Chain = Op.getOperand(0);
1307 SDValue Size = Op.getOperand(1);
1308 DebugLoc dl = Op.getDebugLoc();
1310 // Get a reference from Mips stack pointer
1311 SDValue StackPointer = DAG.getCopyFromReg(Chain, dl, Mips::SP, MVT::i32);
1313 // Subtract the dynamic size from the actual stack size to
1314 // obtain the new stack size.
1315 SDValue Sub = DAG.getNode(ISD::SUB, dl, MVT::i32, StackPointer, Size);
1317 // The Sub result contains the new stack start address, so it
1318 // must be placed in the stack pointer register.
1319 Chain = DAG.getCopyToReg(StackPointer.getValue(1), dl, Mips::SP, Sub,
1322 // This node always has two return values: a new stack pointer
1323 // value and a chain
1324 SDVTList VTLs = DAG.getVTList(MVT::i32, MVT::Other);
1325 SDValue Ptr = DAG.getFrameIndex(MipsFI->getDynAllocFI(), getPointerTy());
1326 SDValue Ops[] = { Chain, Ptr, Chain.getValue(1) };
1328 return DAG.getNode(MipsISD::DynAlloc, dl, VTLs, Ops, 3);
1331 SDValue MipsTargetLowering::
1332 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1334 // The first operand is the chain, the second is the condition, the third is
1335 // the block to branch to if the condition is true.
1336 SDValue Chain = Op.getOperand(0);
1337 SDValue Dest = Op.getOperand(2);
1338 DebugLoc dl = Op.getDebugLoc();
1340 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1342 // Return if flag is not set by a floating point comparison.
1343 if (CondRes.getOpcode() != MipsISD::FPCmp)
1346 SDValue CCNode = CondRes.getOperand(2);
1348 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1349 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1351 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1355 SDValue MipsTargetLowering::
1356 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1358 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1360 // Return if flag is not set by a floating point comparison.
1361 if (Cond.getOpcode() != MipsISD::FPCmp)
1364 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1368 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1369 SelectionDAG &DAG) const {
1370 // FIXME there isn't actually debug info here
1371 DebugLoc dl = Op.getDebugLoc();
1372 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1374 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1375 SDVTList VTs = DAG.getVTList(MVT::i32);
1377 MipsTargetObjectFile &TLOF = (MipsTargetObjectFile&)getObjFileLowering();
1379 // %gp_rel relocation
1380 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1381 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1383 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl, VTs, &GA, 1);
1384 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1385 return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode);
1387 // %hi/%lo relocation
1388 SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1390 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1392 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, VTs, &GAHi, 1);
1393 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1394 return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1397 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1399 GA = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, GA);
1400 SDValue ResNode = DAG.getLoad(MVT::i32, dl,
1401 DAG.getEntryNode(), GA, MachinePointerInfo(),
1403 // On functions and global targets not internal linked only
1404 // a load from got/GP is necessary for PIC to work.
1405 if (!GV->hasInternalLinkage() &&
1406 (!GV->hasLocalLinkage() || isa<Function>(GV)))
1408 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1410 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1411 return DAG.getNode(ISD::ADD, dl, MVT::i32, ResNode, Lo);
1414 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1415 SelectionDAG &DAG) const {
1416 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1417 // FIXME there isn't actually debug info here
1418 DebugLoc dl = Op.getDebugLoc();
1420 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1421 // %hi/%lo relocation
1422 SDValue BAHi = DAG.getBlockAddress(BA, MVT::i32, true,
1424 SDValue BALo = DAG.getBlockAddress(BA, MVT::i32, true,
1426 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, MVT::i32, BAHi);
1427 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALo);
1428 return DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1431 SDValue BAGOTOffset = DAG.getBlockAddress(BA, MVT::i32, true,
1433 BAGOTOffset = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, BAGOTOffset);
1434 SDValue BALOOffset = DAG.getBlockAddress(BA, MVT::i32, true,
1436 SDValue Load = DAG.getLoad(MVT::i32, dl,
1437 DAG.getEntryNode(), BAGOTOffset,
1438 MachinePointerInfo(), false, false, 0);
1439 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALOOffset);
1440 return DAG.getNode(ISD::ADD, dl, MVT::i32, Load, Lo);
1443 SDValue MipsTargetLowering::
1444 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1446 // If the relocation model is PIC, use the General Dynamic TLS Model,
1447 // otherwise use the Initial Exec or Local Exec TLS Model.
1448 // TODO: implement Local Dynamic TLS model
1450 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1451 DebugLoc dl = GA->getDebugLoc();
1452 const GlobalValue *GV = GA->getGlobal();
1453 EVT PtrVT = getPointerTy();
1455 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1456 // General Dynamic TLS Model
1457 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32,
1458 0, MipsII::MO_TLSGD);
1459 SDValue Tlsgd = DAG.getNode(MipsISD::TlsGd, dl, MVT::i32, TGA);
1460 SDValue GP = DAG.getRegister(Mips::GP, MVT::i32);
1461 SDValue Argument = DAG.getNode(ISD::ADD, dl, MVT::i32, GP, Tlsgd);
1465 Entry.Node = Argument;
1466 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
1467 Args.push_back(Entry);
1468 std::pair<SDValue, SDValue> CallResult =
1469 LowerCallTo(DAG.getEntryNode(),
1470 (Type *) Type::getInt32Ty(*DAG.getContext()),
1471 false, false, false, false, 0, CallingConv::C, false, true,
1472 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG,
1475 return CallResult.first;
1479 if (GV->isDeclaration()) {
1480 // Initial Exec TLS Model
1481 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1482 MipsII::MO_GOTTPREL);
1483 Offset = DAG.getLoad(MVT::i32, dl,
1484 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1487 // Local Exec TLS Model
1488 SDVTList VTs = DAG.getVTList(MVT::i32);
1489 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1490 MipsII::MO_TPREL_HI);
1491 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1492 MipsII::MO_TPREL_LO);
1493 SDValue Hi = DAG.getNode(MipsISD::TprelHi, dl, VTs, &TGAHi, 1);
1494 SDValue Lo = DAG.getNode(MipsISD::TprelLo, dl, MVT::i32, TGALo);
1495 Offset = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1498 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1499 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1502 SDValue MipsTargetLowering::
1503 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1507 // FIXME there isn't actually debug info here
1508 DebugLoc dl = Op.getDebugLoc();
1509 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1510 unsigned char OpFlag = IsPIC ? MipsII::MO_GOT : MipsII::MO_ABS_HI;
1512 EVT PtrVT = Op.getValueType();
1513 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1515 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
1518 SDValue Ops[] = { JTI };
1519 HiPart = DAG.getNode(MipsISD::Hi, dl, DAG.getVTList(MVT::i32), Ops, 1);
1520 } else {// Emit Load from Global Pointer
1521 JTI = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, JTI);
1522 HiPart = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(), JTI,
1523 MachinePointerInfo(),
1527 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
1529 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, JTILo);
1530 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1535 SDValue MipsTargetLowering::
1536 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1539 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1540 const Constant *C = N->getConstVal();
1541 // FIXME there isn't actually debug info here
1542 DebugLoc dl = Op.getDebugLoc();
1544 // gp_rel relocation
1545 // FIXME: we should reference the constant pool using small data sections,
1546 // but the asm printer currently doesn't support this feature without
1547 // hacking it. This feature should come soon so we can uncomment the
1549 //if (IsInSmallSection(C->getType())) {
1550 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1551 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1552 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1554 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1555 SDValue CPHi = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1556 N->getOffset(), MipsII::MO_ABS_HI);
1557 SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1558 N->getOffset(), MipsII::MO_ABS_LO);
1559 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, MVT::i32, CPHi);
1560 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1561 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1563 SDValue CP = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1564 N->getOffset(), MipsII::MO_GOT);
1565 CP = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, CP);
1566 SDValue Load = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(),
1567 CP, MachinePointerInfo::getConstantPool(),
1569 SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1570 N->getOffset(), MipsII::MO_ABS_LO);
1571 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1572 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, Load, Lo);
1578 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1579 MachineFunction &MF = DAG.getMachineFunction();
1580 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1582 DebugLoc dl = Op.getDebugLoc();
1583 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1586 // vastart just stores the address of the VarArgsFrameIndex slot into the
1587 // memory location argument.
1588 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1589 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1590 MachinePointerInfo(SV),
1594 static SDValue LowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG) {
1595 // FIXME: Use ext/ins instructions if target architecture is Mips32r2.
1596 DebugLoc dl = Op.getDebugLoc();
1597 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op.getOperand(0));
1598 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op.getOperand(1));
1599 SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op0,
1600 DAG.getConstant(0x7fffffff, MVT::i32));
1601 SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op1,
1602 DAG.getConstant(0x80000000, MVT::i32));
1603 SDValue Result = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1604 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Result);
1607 static SDValue LowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool isLittle) {
1609 // Use ext/ins instructions if target architecture is Mips32r2.
1610 // Eliminate redundant mfc1 and mtc1 instructions.
1611 unsigned LoIdx = 0, HiIdx = 1;
1614 std::swap(LoIdx, HiIdx);
1616 DebugLoc dl = Op.getDebugLoc();
1617 SDValue Word0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1619 DAG.getConstant(LoIdx, MVT::i32));
1620 SDValue Hi0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1621 Op.getOperand(0), DAG.getConstant(HiIdx, MVT::i32));
1622 SDValue Hi1 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1623 Op.getOperand(1), DAG.getConstant(HiIdx, MVT::i32));
1624 SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi0,
1625 DAG.getConstant(0x7fffffff, MVT::i32));
1626 SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi1,
1627 DAG.getConstant(0x80000000, MVT::i32));
1628 SDValue Word1 = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1631 std::swap(Word0, Word1);
1633 return DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64, Word0, Word1);
1636 SDValue MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG)
1638 EVT Ty = Op.getValueType();
1640 assert(Ty == MVT::f32 || Ty == MVT::f64);
1643 return LowerFCOPYSIGN32(Op, DAG);
1645 return LowerFCOPYSIGN64(Op, DAG, Subtarget->isLittle());
1648 SDValue MipsTargetLowering::
1649 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1651 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1652 "Frame address can only be determined for current frame.");
1654 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1655 MFI->setFrameAddressIsTaken(true);
1656 EVT VT = Op.getValueType();
1657 DebugLoc dl = Op.getDebugLoc();
1658 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, Mips::FP, VT);
1662 // TODO: set SType according to the desired memory barrier behavior.
1663 SDValue MipsTargetLowering::LowerMEMBARRIER(SDValue Op,
1664 SelectionDAG& DAG) const {
1666 DebugLoc dl = Op.getDebugLoc();
1667 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1668 DAG.getConstant(SType, MVT::i32));
1671 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1672 SelectionDAG& DAG) const {
1673 // FIXME: Need pseudo-fence for 'singlethread' fences
1674 // FIXME: Set SType for weaker fences where supported/appropriate.
1676 DebugLoc dl = Op.getDebugLoc();
1677 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1678 DAG.getConstant(SType, MVT::i32));
1681 //===----------------------------------------------------------------------===//
1682 // Calling Convention Implementation
1683 //===----------------------------------------------------------------------===//
1685 #include "MipsGenCallingConv.inc"
1687 //===----------------------------------------------------------------------===//
1688 // TODO: Implement a generic logic using tblgen that can support this.
1689 // Mips O32 ABI rules:
1691 // i32 - Passed in A0, A1, A2, A3 and stack
1692 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
1693 // an argument. Otherwise, passed in A1, A2, A3 and stack.
1694 // f64 - Only passed in two aliased f32 registers if no int reg has been used
1695 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
1696 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
1699 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
1700 //===----------------------------------------------------------------------===//
1702 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
1703 MVT LocVT, CCValAssign::LocInfo LocInfo,
1704 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1706 static const unsigned IntRegsSize=4, FloatRegsSize=2;
1708 static const unsigned IntRegs[] = {
1709 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1711 static const unsigned F32Regs[] = {
1712 Mips::F12, Mips::F14
1714 static const unsigned F64Regs[] = {
1719 if (ArgFlags.isByVal()) {
1720 State.HandleByVal(ValNo, ValVT, LocVT, LocInfo,
1721 1 /*MinSize*/, 4 /*MinAlign*/, ArgFlags);
1722 unsigned NextReg = (State.getNextStackOffset() + 3) / 4;
1723 for (unsigned r = State.getFirstUnallocated(IntRegs, IntRegsSize);
1724 r < std::min(IntRegsSize, NextReg); ++r)
1725 State.AllocateReg(IntRegs[r]);
1729 // Promote i8 and i16
1730 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
1732 if (ArgFlags.isSExt())
1733 LocInfo = CCValAssign::SExt;
1734 else if (ArgFlags.isZExt())
1735 LocInfo = CCValAssign::ZExt;
1737 LocInfo = CCValAssign::AExt;
1742 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
1743 // is true: function is vararg, argument is 3rd or higher, there is previous
1744 // argument which is not f32 or f64.
1745 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
1746 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
1747 unsigned OrigAlign = ArgFlags.getOrigAlign();
1748 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
1750 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
1751 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1752 // If this is the first part of an i64 arg,
1753 // the allocated register must be either A0 or A2.
1754 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
1755 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1757 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
1758 // Allocate int register and shadow next int register. If first
1759 // available register is Mips::A1 or Mips::A3, shadow it too.
1760 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1761 if (Reg == Mips::A1 || Reg == Mips::A3)
1762 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1763 State.AllocateReg(IntRegs, IntRegsSize);
1765 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
1766 // we are guaranteed to find an available float register
1767 if (ValVT == MVT::f32) {
1768 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
1769 // Shadow int register
1770 State.AllocateReg(IntRegs, IntRegsSize);
1772 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
1773 // Shadow int registers
1774 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
1775 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
1776 State.AllocateReg(IntRegs, IntRegsSize);
1777 State.AllocateReg(IntRegs, IntRegsSize);
1780 llvm_unreachable("Cannot handle this ValVT.");
1782 unsigned SizeInBytes = ValVT.getSizeInBits() >> 3;
1783 unsigned Offset = State.AllocateStack(SizeInBytes, OrigAlign);
1786 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1788 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1790 return false; // CC must always match
1793 //===----------------------------------------------------------------------===//
1794 // Call Calling Convention Implementation
1795 //===----------------------------------------------------------------------===//
1797 static const unsigned O32IntRegsSize = 4;
1799 static const unsigned O32IntRegs[] = {
1800 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1803 // Write ByVal Arg to arg registers and stack.
1805 WriteByValArg(SDValue& Chain, DebugLoc dl,
1806 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
1807 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
1808 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
1809 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
1811 unsigned FirstWord = VA.getLocMemOffset() / 4;
1812 unsigned NumWords = (Flags.getByValSize() + 3) / 4;
1813 unsigned LastWord = FirstWord + NumWords;
1815 unsigned ByValAlign = Flags.getByValAlign();
1817 // copy the first 4 words of byval arg to registers A0 - A3
1818 for (CurWord = FirstWord; CurWord < std::min(LastWord, O32IntRegsSize);
1820 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1821 DAG.getConstant((CurWord - FirstWord) * 4,
1823 SDValue LoadVal = DAG.getLoad(MVT::i32, dl, Chain, LoadPtr,
1824 MachinePointerInfo(),
1825 false, false, std::min(ByValAlign,
1827 MemOpChains.push_back(LoadVal.getValue(1));
1828 unsigned DstReg = O32IntRegs[CurWord];
1829 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
1832 // copy remaining part of byval arg to stack.
1833 if (CurWord < LastWord) {
1834 unsigned SizeInBytes = (LastWord - CurWord) * 4;
1835 SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1836 DAG.getConstant((CurWord - FirstWord) * 4,
1838 LastFI = MFI->CreateFixedObject(SizeInBytes, CurWord * 4, true);
1839 SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
1840 Chain = DAG.getMemcpy(Chain, dl, Dst, Src,
1841 DAG.getConstant(SizeInBytes, MVT::i32),
1842 /*Align*/ByValAlign,
1843 /*isVolatile=*/false, /*AlwaysInline=*/false,
1844 MachinePointerInfo(0), MachinePointerInfo(0));
1845 MemOpChains.push_back(Chain);
1849 /// LowerCall - functions arguments are copied from virtual regs to
1850 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
1851 /// TODO: isTailCall.
1853 MipsTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1854 CallingConv::ID CallConv, bool isVarArg,
1856 const SmallVectorImpl<ISD::OutputArg> &Outs,
1857 const SmallVectorImpl<SDValue> &OutVals,
1858 const SmallVectorImpl<ISD::InputArg> &Ins,
1859 DebugLoc dl, SelectionDAG &DAG,
1860 SmallVectorImpl<SDValue> &InVals) const {
1861 // MIPs target does not yet support tail call optimization.
1864 MachineFunction &MF = DAG.getMachineFunction();
1865 MachineFrameInfo *MFI = MF.getFrameInfo();
1866 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
1867 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1868 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1870 // Analyze operands of the call, assigning locations to each operand.
1871 SmallVector<CCValAssign, 16> ArgLocs;
1872 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1873 getTargetMachine(), ArgLocs, *DAG.getContext());
1875 if (Subtarget->isABI_O32())
1876 CCInfo.AnalyzeCallOperands(Outs, CC_MipsO32);
1878 CCInfo.AnalyzeCallOperands(Outs, CC_Mips);
1880 // Get a count of how many bytes are to be pushed on the stack.
1881 unsigned NextStackOffset = CCInfo.getNextStackOffset();
1883 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NextStackOffset,
1886 // If this is the first call, create a stack frame object that points to
1887 // a location to which .cprestore saves $gp.
1888 if (IsPIC && !MipsFI->getGPFI())
1889 MipsFI->setGPFI(MFI->CreateFixedObject(4, 0, true));
1891 // Get the frame index of the stack frame object that points to the location
1892 // of dynamically allocated area on the stack.
1893 int DynAllocFI = MipsFI->getDynAllocFI();
1895 // Update size of the maximum argument space.
1896 // For O32, a minimum of four words (16 bytes) of argument space is
1898 if (Subtarget->isABI_O32())
1899 NextStackOffset = std::max(NextStackOffset, (unsigned)16);
1901 unsigned MaxCallFrameSize = MipsFI->getMaxCallFrameSize();
1903 if (MaxCallFrameSize < NextStackOffset) {
1904 MipsFI->setMaxCallFrameSize(NextStackOffset);
1906 // Set the offsets relative to $sp of the $gp restore slot and dynamically
1907 // allocated stack space. These offsets must be aligned to a boundary
1908 // determined by the stack alignment of the ABI.
1909 unsigned StackAlignment = TFL->getStackAlignment();
1910 NextStackOffset = (NextStackOffset + StackAlignment - 1) /
1911 StackAlignment * StackAlignment;
1914 MFI->setObjectOffset(MipsFI->getGPFI(), NextStackOffset);
1916 MFI->setObjectOffset(DynAllocFI, NextStackOffset);
1919 // With EABI is it possible to have 16 args on registers.
1920 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
1921 SmallVector<SDValue, 8> MemOpChains;
1923 int FirstFI = -MFI->getNumFixedObjects() - 1, LastFI = 0;
1925 // Walk the register/memloc assignments, inserting copies/loads.
1926 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1927 SDValue Arg = OutVals[i];
1928 CCValAssign &VA = ArgLocs[i];
1930 // Promote the value if needed.
1931 switch (VA.getLocInfo()) {
1932 default: llvm_unreachable("Unknown loc info!");
1933 case CCValAssign::Full:
1934 if (Subtarget->isABI_O32() && VA.isRegLoc()) {
1935 if (VA.getValVT() == MVT::f32 && VA.getLocVT() == MVT::i32)
1936 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
1937 if (VA.getValVT() == MVT::f64 && VA.getLocVT() == MVT::i32) {
1938 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1939 Arg, DAG.getConstant(0, MVT::i32));
1940 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1941 Arg, DAG.getConstant(1, MVT::i32));
1942 if (!Subtarget->isLittle())
1944 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Lo));
1945 RegsToPass.push_back(std::make_pair(VA.getLocReg()+1, Hi));
1950 case CCValAssign::SExt:
1951 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1953 case CCValAssign::ZExt:
1954 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1956 case CCValAssign::AExt:
1957 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1961 // Arguments that can be passed on register must be kept at
1962 // RegsToPass vector
1963 if (VA.isRegLoc()) {
1964 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1968 // Register can't get to this point...
1969 assert(VA.isMemLoc());
1972 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1973 if (Flags.isByVal()) {
1974 assert(Subtarget->isABI_O32() &&
1975 "No support for ByVal args by ABIs other than O32 yet.");
1976 assert(Flags.getByValSize() &&
1977 "ByVal args of size 0 should have been ignored by front-end.");
1978 WriteByValArg(Chain, dl, RegsToPass, MemOpChains, LastFI, MFI, DAG, Arg,
1979 VA, Flags, getPointerTy());
1983 // Create the frame index object for this incoming parameter
1984 LastFI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1985 VA.getLocMemOffset(), true);
1986 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
1988 // emit ISD::STORE whichs stores the
1989 // parameter value to a stack Location
1990 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
1991 MachinePointerInfo(),
1995 // Extend range of indices of frame objects for outgoing arguments that were
1996 // created during this function call. Skip this step if no such objects were
1999 MipsFI->extendOutArgFIRange(FirstFI, LastFI);
2001 // Transform all store nodes into one single node because all store
2002 // nodes are independent of each other.
2003 if (!MemOpChains.empty())
2004 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2005 &MemOpChains[0], MemOpChains.size());
2007 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2008 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2009 // node so that legalize doesn't hack it.
2010 unsigned char OpFlag = IsPIC ? MipsII::MO_GOT_CALL : MipsII::MO_NO_FLAG;
2011 bool LoadSymAddr = false;
2014 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2015 if (IsPIC && G->getGlobal()->hasInternalLinkage()) {
2016 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2017 getPointerTy(), 0,MipsII:: MO_GOT);
2018 CalleeLo = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(),
2019 0, MipsII::MO_ABS_LO);
2021 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2022 getPointerTy(), 0, OpFlag);
2027 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2028 Callee = DAG.getTargetExternalSymbol(S->getSymbol(),
2029 getPointerTy(), OpFlag);
2035 // Create nodes that load address of callee and copy it to T9
2038 // Load callee address
2039 Callee = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, Callee);
2040 SDValue LoadValue = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(), Callee,
2041 MachinePointerInfo::getGOT(),
2044 // Use GOT+LO if callee has internal linkage.
2045 if (CalleeLo.getNode()) {
2046 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CalleeLo);
2047 Callee = DAG.getNode(ISD::ADD, dl, MVT::i32, LoadValue, Lo);
2053 Chain = DAG.getCopyToReg(Chain, dl, Mips::T9, Callee, SDValue(0, 0));
2054 InFlag = Chain.getValue(1);
2055 Callee = DAG.getRegister(Mips::T9, MVT::i32);
2058 // Build a sequence of copy-to-reg nodes chained together with token
2059 // chain and flag operands which copy the outgoing args into registers.
2060 // The InFlag in necessary since all emitted instructions must be
2062 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2063 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2064 RegsToPass[i].second, InFlag);
2065 InFlag = Chain.getValue(1);
2068 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2069 // = Chain, Callee, Reg#1, Reg#2, ...
2071 // Returns a chain & a flag for retval copy to use.
2072 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2073 SmallVector<SDValue, 8> Ops;
2074 Ops.push_back(Chain);
2075 Ops.push_back(Callee);
2077 // Add argument registers to the end of the list so that they are
2078 // known live into the call.
2079 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2080 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2081 RegsToPass[i].second.getValueType()));
2083 if (InFlag.getNode())
2084 Ops.push_back(InFlag);
2086 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2087 InFlag = Chain.getValue(1);
2089 // Create the CALLSEQ_END node.
2090 Chain = DAG.getCALLSEQ_END(Chain,
2091 DAG.getIntPtrConstant(NextStackOffset, true),
2092 DAG.getIntPtrConstant(0, true), InFlag);
2093 InFlag = Chain.getValue(1);
2095 // Handle result values, copying them out of physregs into vregs that we
2097 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2098 Ins, dl, DAG, InVals);
2101 /// LowerCallResult - Lower the result values of a call into the
2102 /// appropriate copies out of appropriate physical registers.
2104 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2105 CallingConv::ID CallConv, bool isVarArg,
2106 const SmallVectorImpl<ISD::InputArg> &Ins,
2107 DebugLoc dl, SelectionDAG &DAG,
2108 SmallVectorImpl<SDValue> &InVals) const {
2109 // Assign locations to each value returned by this call.
2110 SmallVector<CCValAssign, 16> RVLocs;
2111 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2112 getTargetMachine(), RVLocs, *DAG.getContext());
2114 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2116 // Copy all of the result registers out of their specified physreg.
2117 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2118 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2119 RVLocs[i].getValVT(), InFlag).getValue(1);
2120 InFlag = Chain.getValue(2);
2121 InVals.push_back(Chain.getValue(0));
2127 //===----------------------------------------------------------------------===//
2128 // Formal Arguments Calling Convention Implementation
2129 //===----------------------------------------------------------------------===//
2130 static void ReadByValArg(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2131 std::vector<SDValue>& OutChains,
2132 SelectionDAG &DAG, unsigned NumWords, SDValue FIN,
2133 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags) {
2134 unsigned LocMem = VA.getLocMemOffset();
2135 unsigned FirstWord = LocMem / 4;
2137 // copy register A0 - A3 to frame object
2138 for (unsigned i = 0; i < NumWords; ++i) {
2139 unsigned CurWord = FirstWord + i;
2140 if (CurWord >= O32IntRegsSize)
2143 unsigned SrcReg = O32IntRegs[CurWord];
2144 unsigned Reg = AddLiveIn(MF, SrcReg, Mips::CPURegsRegisterClass);
2145 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIN,
2146 DAG.getConstant(i * 4, MVT::i32));
2147 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(Reg, MVT::i32),
2148 StorePtr, MachinePointerInfo(), false,
2150 OutChains.push_back(Store);
2154 /// LowerFormalArguments - transform physical registers into virtual registers
2155 /// and generate load operations for arguments places on the stack.
2157 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2158 CallingConv::ID CallConv,
2160 const SmallVectorImpl<ISD::InputArg>
2162 DebugLoc dl, SelectionDAG &DAG,
2163 SmallVectorImpl<SDValue> &InVals)
2165 MachineFunction &MF = DAG.getMachineFunction();
2166 MachineFrameInfo *MFI = MF.getFrameInfo();
2167 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2169 MipsFI->setVarArgsFrameIndex(0);
2171 // Used with vargs to acumulate store chains.
2172 std::vector<SDValue> OutChains;
2174 // Assign locations to all of the incoming arguments.
2175 SmallVector<CCValAssign, 16> ArgLocs;
2176 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2177 getTargetMachine(), ArgLocs, *DAG.getContext());
2179 if (Subtarget->isABI_O32())
2180 CCInfo.AnalyzeFormalArguments(Ins, CC_MipsO32);
2182 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips);
2184 int LastFI = 0;// MipsFI->LastInArgFI is 0 at the entry of this function.
2186 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2187 CCValAssign &VA = ArgLocs[i];
2189 // Arguments stored on registers
2190 if (VA.isRegLoc()) {
2191 EVT RegVT = VA.getLocVT();
2192 unsigned ArgReg = VA.getLocReg();
2193 TargetRegisterClass *RC = 0;
2195 if (RegVT == MVT::i32)
2196 RC = Mips::CPURegsRegisterClass;
2197 else if (RegVT == MVT::f32)
2198 RC = Mips::FGR32RegisterClass;
2199 else if (RegVT == MVT::f64) {
2200 if (!Subtarget->isSingleFloat())
2201 RC = Mips::AFGR64RegisterClass;
2203 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2205 // Transform the arguments stored on
2206 // physical registers into virtual ones
2207 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2208 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2210 // If this is an 8 or 16-bit value, it has been passed promoted
2211 // to 32 bits. Insert an assert[sz]ext to capture this, then
2212 // truncate to the right size.
2213 if (VA.getLocInfo() != CCValAssign::Full) {
2214 unsigned Opcode = 0;
2215 if (VA.getLocInfo() == CCValAssign::SExt)
2216 Opcode = ISD::AssertSext;
2217 else if (VA.getLocInfo() == CCValAssign::ZExt)
2218 Opcode = ISD::AssertZext;
2220 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
2221 DAG.getValueType(VA.getValVT()));
2222 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2225 // Handle O32 ABI cases: i32->f32 and (i32,i32)->f64
2226 if (Subtarget->isABI_O32()) {
2227 if (RegVT == MVT::i32 && VA.getValVT() == MVT::f32)
2228 ArgValue = DAG.getNode(ISD::BITCAST, dl, MVT::f32, ArgValue);
2229 if (RegVT == MVT::i32 && VA.getValVT() == MVT::f64) {
2230 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
2231 VA.getLocReg()+1, RC);
2232 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
2233 if (!Subtarget->isLittle())
2234 std::swap(ArgValue, ArgValue2);
2235 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
2236 ArgValue, ArgValue2);
2240 InVals.push_back(ArgValue);
2241 } else { // VA.isRegLoc()
2244 assert(VA.isMemLoc());
2246 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2248 if (Flags.isByVal()) {
2249 assert(Subtarget->isABI_O32() &&
2250 "No support for ByVal args by ABIs other than O32 yet.");
2251 assert(Flags.getByValSize() &&
2252 "ByVal args of size 0 should have been ignored by front-end.");
2253 unsigned NumWords = (Flags.getByValSize() + 3) / 4;
2254 LastFI = MFI->CreateFixedObject(NumWords * 4, VA.getLocMemOffset(),
2256 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2257 InVals.push_back(FIN);
2258 ReadByValArg(MF, Chain, dl, OutChains, DAG, NumWords, FIN, VA, Flags);
2263 // The stack pointer offset is relative to the caller stack frame.
2264 LastFI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
2265 VA.getLocMemOffset(), true);
2267 // Create load nodes to retrieve arguments from the stack
2268 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2269 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2270 MachinePointerInfo::getFixedStack(LastFI),
2275 // The mips ABIs for returning structs by value requires that we copy
2276 // the sret argument into $v0 for the return. Save the argument into
2277 // a virtual register so that we can access it from the return points.
2278 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2279 unsigned Reg = MipsFI->getSRetReturnReg();
2281 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i32));
2282 MipsFI->setSRetReturnReg(Reg);
2284 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2285 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2288 if (isVarArg && Subtarget->isABI_O32()) {
2289 // Record the frame index of the first variable argument
2290 // which is a value necessary to VASTART.
2291 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2292 assert(NextStackOffset % 4 == 0 &&
2293 "NextStackOffset must be aligned to 4-byte boundaries.");
2294 LastFI = MFI->CreateFixedObject(4, NextStackOffset, true);
2295 MipsFI->setVarArgsFrameIndex(LastFI);
2297 // If NextStackOffset is smaller than o32's 16-byte reserved argument area,
2298 // copy the integer registers that have not been used for argument passing
2299 // to the caller's stack frame.
2300 for (; NextStackOffset < 16; NextStackOffset += 4) {
2301 TargetRegisterClass *RC = Mips::CPURegsRegisterClass;
2302 unsigned Idx = NextStackOffset / 4;
2303 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), O32IntRegs[Idx], RC);
2304 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, MVT::i32);
2305 LastFI = MFI->CreateFixedObject(4, NextStackOffset, true);
2306 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2307 OutChains.push_back(DAG.getStore(Chain, dl, ArgValue, PtrOff,
2308 MachinePointerInfo(),
2313 MipsFI->setLastInArgFI(LastFI);
2315 // All stores are grouped in one node to allow the matching between
2316 // the size of Ins and InVals. This only happens when on varg functions
2317 if (!OutChains.empty()) {
2318 OutChains.push_back(Chain);
2319 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2320 &OutChains[0], OutChains.size());
2326 //===----------------------------------------------------------------------===//
2327 // Return Value Calling Convention Implementation
2328 //===----------------------------------------------------------------------===//
2331 MipsTargetLowering::LowerReturn(SDValue Chain,
2332 CallingConv::ID CallConv, bool isVarArg,
2333 const SmallVectorImpl<ISD::OutputArg> &Outs,
2334 const SmallVectorImpl<SDValue> &OutVals,
2335 DebugLoc dl, SelectionDAG &DAG) const {
2337 // CCValAssign - represent the assignment of
2338 // the return value to a location
2339 SmallVector<CCValAssign, 16> RVLocs;
2341 // CCState - Info about the registers and stack slot.
2342 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2343 getTargetMachine(), RVLocs, *DAG.getContext());
2345 // Analize return values.
2346 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
2348 // If this is the first return lowered for this function, add
2349 // the regs to the liveout set for the function.
2350 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2351 for (unsigned i = 0; i != RVLocs.size(); ++i)
2352 if (RVLocs[i].isRegLoc())
2353 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2358 // Copy the result values into the output registers.
2359 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2360 CCValAssign &VA = RVLocs[i];
2361 assert(VA.isRegLoc() && "Can only return in registers!");
2363 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2366 // guarantee that all emitted copies are
2367 // stuck together, avoiding something bad
2368 Flag = Chain.getValue(1);
2371 // The mips ABIs for returning structs by value requires that we copy
2372 // the sret argument into $v0 for the return. We saved the argument into
2373 // a virtual register in the entry block, so now we copy the value out
2375 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2376 MachineFunction &MF = DAG.getMachineFunction();
2377 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2378 unsigned Reg = MipsFI->getSRetReturnReg();
2381 llvm_unreachable("sret virtual register not created in the entry block");
2382 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2384 Chain = DAG.getCopyToReg(Chain, dl, Mips::V0, Val, Flag);
2385 Flag = Chain.getValue(1);
2388 // Return on Mips is always a "jr $ra"
2390 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2391 Chain, DAG.getRegister(Mips::RA, MVT::i32), Flag);
2393 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2394 Chain, DAG.getRegister(Mips::RA, MVT::i32));
2397 //===----------------------------------------------------------------------===//
2398 // Mips Inline Assembly Support
2399 //===----------------------------------------------------------------------===//
2401 /// getConstraintType - Given a constraint letter, return the type of
2402 /// constraint it is for this target.
2403 MipsTargetLowering::ConstraintType MipsTargetLowering::
2404 getConstraintType(const std::string &Constraint) const
2406 // Mips specific constrainy
2407 // GCC config/mips/constraints.md
2409 // 'd' : An address register. Equivalent to r
2410 // unless generating MIPS16 code.
2411 // 'y' : Equivalent to r; retained for
2412 // backwards compatibility.
2413 // 'f' : Floating Point registers.
2414 if (Constraint.size() == 1) {
2415 switch (Constraint[0]) {
2420 return C_RegisterClass;
2424 return TargetLowering::getConstraintType(Constraint);
2427 /// Examine constraint type and operand type and determine a weight value.
2428 /// This object must already have been set up with the operand type
2429 /// and the current alternative constraint selected.
2430 TargetLowering::ConstraintWeight
2431 MipsTargetLowering::getSingleConstraintMatchWeight(
2432 AsmOperandInfo &info, const char *constraint) const {
2433 ConstraintWeight weight = CW_Invalid;
2434 Value *CallOperandVal = info.CallOperandVal;
2435 // If we don't have a value, we can't do a match,
2436 // but allow it at the lowest weight.
2437 if (CallOperandVal == NULL)
2439 Type *type = CallOperandVal->getType();
2440 // Look at the constraint type.
2441 switch (*constraint) {
2443 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2447 if (type->isIntegerTy())
2448 weight = CW_Register;
2451 if (type->isFloatTy())
2452 weight = CW_Register;
2458 /// Given a register class constraint, like 'r', if this corresponds directly
2459 /// to an LLVM register class, return a register of 0 and the register class
2461 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2462 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
2464 if (Constraint.size() == 1) {
2465 switch (Constraint[0]) {
2466 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2467 case 'y': // Same as 'r'. Exists for compatibility.
2469 return std::make_pair(0U, Mips::CPURegsRegisterClass);
2472 return std::make_pair(0U, Mips::FGR32RegisterClass);
2474 if ((!Subtarget->isSingleFloat()) && (!Subtarget->isFP64bit()))
2475 return std::make_pair(0U, Mips::AFGR64RegisterClass);
2479 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2483 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2484 // The Mips target isn't yet aware of offsets.
2488 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2489 if (VT != MVT::f32 && VT != MVT::f64)
2491 if (Imm.isNegZero())
2493 return Imm.isZero();