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"
38 // If I is a shifted mask, set the size (Size) and the first bit of the
39 // mask (Pos), and return true.
40 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
41 static bool IsShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
42 if (!isUInt<32>(I) || !isShiftedMask_32(I))
45 Size = CountPopulation_32(I);
46 Pos = CountTrailingZeros_32(I);
50 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
52 case MipsISD::JmpLink: return "MipsISD::JmpLink";
53 case MipsISD::Hi: return "MipsISD::Hi";
54 case MipsISD::Lo: return "MipsISD::Lo";
55 case MipsISD::GPRel: return "MipsISD::GPRel";
56 case MipsISD::TlsGd: return "MipsISD::TlsGd";
57 case MipsISD::TprelHi: return "MipsISD::TprelHi";
58 case MipsISD::TprelLo: return "MipsISD::TprelLo";
59 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
60 case MipsISD::Ret: return "MipsISD::Ret";
61 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
62 case MipsISD::FPCmp: return "MipsISD::FPCmp";
63 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
64 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
65 case MipsISD::FPRound: return "MipsISD::FPRound";
66 case MipsISD::MAdd: return "MipsISD::MAdd";
67 case MipsISD::MAddu: return "MipsISD::MAddu";
68 case MipsISD::MSub: return "MipsISD::MSub";
69 case MipsISD::MSubu: return "MipsISD::MSubu";
70 case MipsISD::DivRem: return "MipsISD::DivRem";
71 case MipsISD::DivRemU: return "MipsISD::DivRemU";
72 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
73 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
74 case MipsISD::WrapperPIC: return "MipsISD::WrapperPIC";
75 case MipsISD::DynAlloc: return "MipsISD::DynAlloc";
76 case MipsISD::Sync: return "MipsISD::Sync";
77 case MipsISD::Ext: return "MipsISD::Ext";
78 case MipsISD::Ins: return "MipsISD::Ins";
84 MipsTargetLowering(MipsTargetMachine &TM)
85 : TargetLowering(TM, new MipsTargetObjectFile()) {
86 Subtarget = &TM.getSubtarget<MipsSubtarget>();
88 // Mips does not have i1 type, so use i32 for
89 // setcc operations results (slt, sgt, ...).
90 setBooleanContents(ZeroOrOneBooleanContent);
92 // Set up the register classes
93 addRegisterClass(MVT::i32, Mips::CPURegsRegisterClass);
94 addRegisterClass(MVT::f32, Mips::FGR32RegisterClass);
96 // When dealing with single precision only, use libcalls
97 if (!Subtarget->isSingleFloat())
98 if (!Subtarget->isFP64bit())
99 addRegisterClass(MVT::f64, Mips::AFGR64RegisterClass);
101 // Load extented operations for i1 types must be promoted
102 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
103 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
104 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
106 // MIPS doesn't have extending float->double load/store
107 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
108 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
110 // Used by legalize types to correctly generate the setcc result.
111 // Without this, every float setcc comes with a AND/OR with the result,
112 // we don't want this, since the fpcmp result goes to a flag register,
113 // which is used implicitly by brcond and select operations.
114 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
116 // Mips Custom Operations
117 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
118 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
119 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
120 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
121 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
122 setOperationAction(ISD::SELECT, MVT::f32, Custom);
123 setOperationAction(ISD::SELECT, MVT::f64, Custom);
124 setOperationAction(ISD::SELECT, MVT::i32, Custom);
125 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
126 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
127 setOperationAction(ISD::VASTART, MVT::Other, Custom);
129 setOperationAction(ISD::SDIV, MVT::i32, Expand);
130 setOperationAction(ISD::SREM, MVT::i32, Expand);
131 setOperationAction(ISD::UDIV, MVT::i32, Expand);
132 setOperationAction(ISD::UREM, MVT::i32, Expand);
134 // Operations not directly supported by Mips.
135 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
136 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
137 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
138 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
139 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
140 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
141 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
142 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
143 setOperationAction(ISD::ROTL, MVT::i32, Expand);
145 if (!Subtarget->isMips32r2())
146 setOperationAction(ISD::ROTR, MVT::i32, Expand);
148 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
149 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
150 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
151 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
152 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
153 setOperationAction(ISD::FSIN, MVT::f32, Expand);
154 setOperationAction(ISD::FSIN, MVT::f64, Expand);
155 setOperationAction(ISD::FCOS, MVT::f32, Expand);
156 setOperationAction(ISD::FCOS, MVT::f64, Expand);
157 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
158 setOperationAction(ISD::FPOW, MVT::f32, Expand);
159 setOperationAction(ISD::FPOW, MVT::f64, Expand);
160 setOperationAction(ISD::FLOG, MVT::f32, Expand);
161 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
162 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
163 setOperationAction(ISD::FEXP, MVT::f32, Expand);
164 setOperationAction(ISD::FMA, MVT::f32, Expand);
165 setOperationAction(ISD::FMA, MVT::f64, Expand);
167 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
168 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
170 setOperationAction(ISD::VAARG, MVT::Other, Expand);
171 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
172 setOperationAction(ISD::VAEND, MVT::Other, Expand);
174 // Use the default for now
175 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
176 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
178 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
179 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
181 setInsertFencesForAtomic(true);
183 if (Subtarget->isSingleFloat())
184 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
186 if (!Subtarget->hasSEInReg()) {
187 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
188 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
191 if (!Subtarget->hasBitCount())
192 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
194 if (!Subtarget->hasSwap())
195 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
197 setTargetDAGCombine(ISD::ADDE);
198 setTargetDAGCombine(ISD::SUBE);
199 setTargetDAGCombine(ISD::SDIVREM);
200 setTargetDAGCombine(ISD::UDIVREM);
201 setTargetDAGCombine(ISD::SETCC);
202 setTargetDAGCombine(ISD::AND);
203 setTargetDAGCombine(ISD::OR);
205 setMinFunctionAlignment(2);
207 setStackPointerRegisterToSaveRestore(Mips::SP);
208 computeRegisterProperties();
210 setExceptionPointerRegister(Mips::A0);
211 setExceptionSelectorRegister(Mips::A1);
214 bool MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
215 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
216 return SVT == MVT::i32 || SVT == MVT::i16;
219 MVT::SimpleValueType MipsTargetLowering::getSetCCResultType(EVT VT) const {
224 // Transforms a subgraph in CurDAG if the following pattern is found:
225 // (addc multLo, Lo0), (adde multHi, Hi0),
227 // multHi/Lo: product of multiplication
228 // Lo0: initial value of Lo register
229 // Hi0: initial value of Hi register
230 // Return true if pattern matching was successful.
231 static bool SelectMadd(SDNode* ADDENode, SelectionDAG* CurDAG) {
232 // ADDENode's second operand must be a flag output of an ADDC node in order
233 // for the matching to be successful.
234 SDNode* ADDCNode = ADDENode->getOperand(2).getNode();
236 if (ADDCNode->getOpcode() != ISD::ADDC)
239 SDValue MultHi = ADDENode->getOperand(0);
240 SDValue MultLo = ADDCNode->getOperand(0);
241 SDNode* MultNode = MultHi.getNode();
242 unsigned MultOpc = MultHi.getOpcode();
244 // MultHi and MultLo must be generated by the same node,
245 if (MultLo.getNode() != MultNode)
248 // and it must be a multiplication.
249 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
252 // MultLo amd MultHi must be the first and second output of MultNode
254 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
257 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
258 // of the values of MultNode, in which case MultNode will be removed in later
260 // If there exist users other than ADDENode or ADDCNode, this function returns
261 // here, which will result in MultNode being mapped to a single MULT
262 // instruction node rather than a pair of MULT and MADD instructions being
264 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
267 SDValue Chain = CurDAG->getEntryNode();
268 DebugLoc dl = ADDENode->getDebugLoc();
270 // create MipsMAdd(u) node
271 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
273 SDValue MAdd = CurDAG->getNode(MultOpc, dl,
275 MultNode->getOperand(0),// Factor 0
276 MultNode->getOperand(1),// Factor 1
277 ADDCNode->getOperand(1),// Lo0
278 ADDENode->getOperand(1));// Hi0
280 // create CopyFromReg nodes
281 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
283 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
285 CopyFromLo.getValue(2));
287 // replace uses of adde and addc here
288 if (!SDValue(ADDCNode, 0).use_empty())
289 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
291 if (!SDValue(ADDENode, 0).use_empty())
292 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
298 // Transforms a subgraph in CurDAG if the following pattern is found:
299 // (addc Lo0, multLo), (sube Hi0, multHi),
301 // multHi/Lo: product of multiplication
302 // Lo0: initial value of Lo register
303 // Hi0: initial value of Hi register
304 // Return true if pattern matching was successful.
305 static bool SelectMsub(SDNode* SUBENode, SelectionDAG* CurDAG) {
306 // SUBENode's second operand must be a flag output of an SUBC node in order
307 // for the matching to be successful.
308 SDNode* SUBCNode = SUBENode->getOperand(2).getNode();
310 if (SUBCNode->getOpcode() != ISD::SUBC)
313 SDValue MultHi = SUBENode->getOperand(1);
314 SDValue MultLo = SUBCNode->getOperand(1);
315 SDNode* MultNode = MultHi.getNode();
316 unsigned MultOpc = MultHi.getOpcode();
318 // MultHi and MultLo must be generated by the same node,
319 if (MultLo.getNode() != MultNode)
322 // and it must be a multiplication.
323 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
326 // MultLo amd MultHi must be the first and second output of MultNode
328 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
331 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
332 // of the values of MultNode, in which case MultNode will be removed in later
334 // If there exist users other than SUBENode or SUBCNode, this function returns
335 // here, which will result in MultNode being mapped to a single MULT
336 // instruction node rather than a pair of MULT and MSUB instructions being
338 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
341 SDValue Chain = CurDAG->getEntryNode();
342 DebugLoc dl = SUBENode->getDebugLoc();
344 // create MipsSub(u) node
345 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
347 SDValue MSub = CurDAG->getNode(MultOpc, dl,
349 MultNode->getOperand(0),// Factor 0
350 MultNode->getOperand(1),// Factor 1
351 SUBCNode->getOperand(0),// Lo0
352 SUBENode->getOperand(0));// Hi0
354 // create CopyFromReg nodes
355 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
357 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
359 CopyFromLo.getValue(2));
361 // replace uses of sube and subc here
362 if (!SDValue(SUBCNode, 0).use_empty())
363 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
365 if (!SDValue(SUBENode, 0).use_empty())
366 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
371 static SDValue PerformADDECombine(SDNode *N, SelectionDAG& DAG,
372 TargetLowering::DAGCombinerInfo &DCI,
373 const MipsSubtarget* Subtarget) {
374 if (DCI.isBeforeLegalize())
377 if (Subtarget->isMips32() && SelectMadd(N, &DAG))
378 return SDValue(N, 0);
383 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG& DAG,
384 TargetLowering::DAGCombinerInfo &DCI,
385 const MipsSubtarget* Subtarget) {
386 if (DCI.isBeforeLegalize())
389 if (Subtarget->isMips32() && SelectMsub(N, &DAG))
390 return SDValue(N, 0);
395 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG& DAG,
396 TargetLowering::DAGCombinerInfo &DCI,
397 const MipsSubtarget* Subtarget) {
398 if (DCI.isBeforeLegalizeOps())
401 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
403 DebugLoc dl = N->getDebugLoc();
405 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
406 N->getOperand(0), N->getOperand(1));
407 SDValue InChain = DAG.getEntryNode();
408 SDValue InGlue = DivRem;
411 if (N->hasAnyUseOfValue(0)) {
412 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, Mips::LO, MVT::i32,
414 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
415 InChain = CopyFromLo.getValue(1);
416 InGlue = CopyFromLo.getValue(2);
420 if (N->hasAnyUseOfValue(1)) {
421 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
422 Mips::HI, MVT::i32, InGlue);
423 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
429 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
431 default: llvm_unreachable("Unknown fp condition code!");
433 case ISD::SETOEQ: return Mips::FCOND_OEQ;
434 case ISD::SETUNE: return Mips::FCOND_UNE;
436 case ISD::SETOLT: return Mips::FCOND_OLT;
438 case ISD::SETOGT: return Mips::FCOND_OGT;
440 case ISD::SETOLE: return Mips::FCOND_OLE;
442 case ISD::SETOGE: return Mips::FCOND_OGE;
443 case ISD::SETULT: return Mips::FCOND_ULT;
444 case ISD::SETULE: return Mips::FCOND_ULE;
445 case ISD::SETUGT: return Mips::FCOND_UGT;
446 case ISD::SETUGE: return Mips::FCOND_UGE;
447 case ISD::SETUO: return Mips::FCOND_UN;
448 case ISD::SETO: return Mips::FCOND_OR;
450 case ISD::SETONE: return Mips::FCOND_ONE;
451 case ISD::SETUEQ: return Mips::FCOND_UEQ;
456 // Returns true if condition code has to be inverted.
457 static bool InvertFPCondCode(Mips::CondCode CC) {
458 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
461 if (CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT)
464 assert(false && "Illegal Condition Code");
468 // Creates and returns an FPCmp node from a setcc node.
469 // Returns Op if setcc is not a floating point comparison.
470 static SDValue CreateFPCmp(SelectionDAG& DAG, const SDValue& Op) {
471 // must be a SETCC node
472 if (Op.getOpcode() != ISD::SETCC)
475 SDValue LHS = Op.getOperand(0);
477 if (!LHS.getValueType().isFloatingPoint())
480 SDValue RHS = Op.getOperand(1);
481 DebugLoc dl = Op.getDebugLoc();
483 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
484 // node if necessary.
485 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
487 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
488 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
491 // Creates and returns a CMovFPT/F node.
492 static SDValue CreateCMovFP(SelectionDAG& DAG, SDValue Cond, SDValue True,
493 SDValue False, DebugLoc DL) {
494 bool invert = InvertFPCondCode((Mips::CondCode)
495 cast<ConstantSDNode>(Cond.getOperand(2))
498 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
499 True.getValueType(), True, False, Cond);
502 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG& DAG,
503 TargetLowering::DAGCombinerInfo &DCI,
504 const MipsSubtarget* Subtarget) {
505 if (DCI.isBeforeLegalizeOps())
508 SDValue Cond = CreateFPCmp(DAG, SDValue(N, 0));
510 if (Cond.getOpcode() != MipsISD::FPCmp)
513 SDValue True = DAG.getConstant(1, MVT::i32);
514 SDValue False = DAG.getConstant(0, MVT::i32);
516 return CreateCMovFP(DAG, Cond, True, False, N->getDebugLoc());
519 static SDValue PerformANDCombine(SDNode *N, SelectionDAG& DAG,
520 TargetLowering::DAGCombinerInfo &DCI,
521 const MipsSubtarget* Subtarget) {
522 // Pattern match EXT.
523 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
524 // => ext $dst, $src, size, pos
525 if (DCI.isBeforeLegalizeOps() || !Subtarget->isMips32r2())
528 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
530 // Op's first operand must be a shift right.
531 if (ShiftRight.getOpcode() != ISD::SRA && ShiftRight.getOpcode() != ISD::SRL)
534 // The second operand of the shift must be an immediate.
537 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
540 Pos = CN->getZExtValue();
542 uint64_t SMPos, SMSize;
543 // Op's second operand must be a shifted mask.
544 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
545 !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
548 // Return if the shifted mask does not start at bit 0 or the sum of its size
549 // and Pos exceeds the word's size.
550 if (SMPos != 0 || Pos + SMSize > 32)
553 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), MVT::i32,
554 ShiftRight.getOperand(0),
555 DAG.getConstant(Pos, MVT::i32),
556 DAG.getConstant(SMSize, MVT::i32));
559 static SDValue PerformORCombine(SDNode *N, SelectionDAG& DAG,
560 TargetLowering::DAGCombinerInfo &DCI,
561 const MipsSubtarget* Subtarget) {
562 // Pattern match INS.
563 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
564 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
565 // => ins $dst, $src, size, pos, $src1
566 if (DCI.isBeforeLegalizeOps() || !Subtarget->isMips32r2())
569 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
570 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
573 // See if Op's first operand matches (and $src1 , mask0).
574 if (And0.getOpcode() != ISD::AND)
577 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
578 !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
581 // See if Op's second operand matches (and (shl $src, pos), mask1).
582 if (And1.getOpcode() != ISD::AND)
585 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
586 !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
589 // The shift masks must have the same position and size.
590 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
593 SDValue Shl = And1.getOperand(0);
594 if (Shl.getOpcode() != ISD::SHL)
597 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
600 unsigned Shamt = CN->getZExtValue();
602 // Return if the shift amount and the first bit position of mask are not the
607 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), MVT::i32,
609 DAG.getConstant(SMPos0, MVT::i32),
610 DAG.getConstant(SMSize0, MVT::i32),
614 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
616 SelectionDAG &DAG = DCI.DAG;
617 unsigned opc = N->getOpcode();
622 return PerformADDECombine(N, DAG, DCI, Subtarget);
624 return PerformSUBECombine(N, DAG, DCI, Subtarget);
627 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
629 return PerformSETCCCombine(N, DAG, DCI, Subtarget);
631 return PerformANDCombine(N, DAG, DCI, Subtarget);
633 return PerformORCombine(N, DAG, DCI, Subtarget);
639 SDValue MipsTargetLowering::
640 LowerOperation(SDValue Op, SelectionDAG &DAG) const
642 switch (Op.getOpcode())
644 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
645 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
646 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
647 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
648 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
649 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
650 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
651 case ISD::SELECT: return LowerSELECT(Op, DAG);
652 case ISD::VASTART: return LowerVASTART(Op, DAG);
653 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
654 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
655 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
656 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
661 //===----------------------------------------------------------------------===//
662 // Lower helper functions
663 //===----------------------------------------------------------------------===//
665 // AddLiveIn - This helper function adds the specified physical register to the
666 // MachineFunction as a live in value. It also creates a corresponding
667 // virtual register for it.
669 AddLiveIn(MachineFunction &MF, unsigned PReg, TargetRegisterClass *RC)
671 assert(RC->contains(PReg) && "Not the correct regclass!");
672 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
673 MF.getRegInfo().addLiveIn(PReg, VReg);
677 // Get fp branch code (not opcode) from condition code.
678 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
679 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
680 return Mips::BRANCH_T;
682 if (CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT)
683 return Mips::BRANCH_F;
685 return Mips::BRANCH_INVALID;
688 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
690 const MipsSubtarget* Subtarget,
691 const TargetInstrInfo *TII,
692 bool isFPCmp, unsigned Opc) {
693 // There is no need to expand CMov instructions if target has
694 // conditional moves.
695 if (Subtarget->hasCondMov())
698 // To "insert" a SELECT_CC instruction, we actually have to insert the
699 // diamond control-flow pattern. The incoming instruction knows the
700 // destination vreg to set, the condition code register to branch on, the
701 // true/false values to select between, and a branch opcode to use.
702 const BasicBlock *LLVM_BB = BB->getBasicBlock();
703 MachineFunction::iterator It = BB;
710 // bNE r1, r0, copy1MBB
711 // fallthrough --> copy0MBB
712 MachineBasicBlock *thisMBB = BB;
713 MachineFunction *F = BB->getParent();
714 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
715 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
716 F->insert(It, copy0MBB);
717 F->insert(It, sinkMBB);
719 // Transfer the remainder of BB and its successor edges to sinkMBB.
720 sinkMBB->splice(sinkMBB->begin(), BB,
721 llvm::next(MachineBasicBlock::iterator(MI)),
723 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
725 // Next, add the true and fallthrough blocks as its successors.
726 BB->addSuccessor(copy0MBB);
727 BB->addSuccessor(sinkMBB);
729 // Emit the right instruction according to the type of the operands compared
731 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
733 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
734 .addReg(Mips::ZERO).addMBB(sinkMBB);
738 // # fallthrough to sinkMBB
741 // Update machine-CFG edges
742 BB->addSuccessor(sinkMBB);
745 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
750 BuildMI(*BB, BB->begin(), dl,
751 TII->get(Mips::PHI), MI->getOperand(0).getReg())
752 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
753 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
755 BuildMI(*BB, BB->begin(), dl,
756 TII->get(Mips::PHI), MI->getOperand(0).getReg())
757 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
758 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
760 MI->eraseFromParent(); // The pseudo instruction is gone now.
765 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
766 MachineBasicBlock *BB) const {
767 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
768 DebugLoc dl = MI->getDebugLoc();
770 switch (MI->getOpcode()) {
772 assert(false && "Unexpected instr type to insert");
777 return ExpandCondMov(MI, BB, dl, Subtarget, TII, true, Mips::BC1F);
781 return ExpandCondMov(MI, BB, dl, Subtarget, TII, true, Mips::BC1T);
785 return ExpandCondMov(MI, BB, dl, Subtarget, TII, false, Mips::BNE);
789 return ExpandCondMov(MI, BB, dl, Subtarget, TII, false, Mips::BEQ);
791 case Mips::ATOMIC_LOAD_ADD_I8:
792 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
793 case Mips::ATOMIC_LOAD_ADD_I16:
794 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
795 case Mips::ATOMIC_LOAD_ADD_I32:
796 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
798 case Mips::ATOMIC_LOAD_AND_I8:
799 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
800 case Mips::ATOMIC_LOAD_AND_I16:
801 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
802 case Mips::ATOMIC_LOAD_AND_I32:
803 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
805 case Mips::ATOMIC_LOAD_OR_I8:
806 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
807 case Mips::ATOMIC_LOAD_OR_I16:
808 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
809 case Mips::ATOMIC_LOAD_OR_I32:
810 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
812 case Mips::ATOMIC_LOAD_XOR_I8:
813 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
814 case Mips::ATOMIC_LOAD_XOR_I16:
815 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
816 case Mips::ATOMIC_LOAD_XOR_I32:
817 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
819 case Mips::ATOMIC_LOAD_NAND_I8:
820 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
821 case Mips::ATOMIC_LOAD_NAND_I16:
822 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
823 case Mips::ATOMIC_LOAD_NAND_I32:
824 return EmitAtomicBinary(MI, BB, 4, 0, true);
826 case Mips::ATOMIC_LOAD_SUB_I8:
827 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
828 case Mips::ATOMIC_LOAD_SUB_I16:
829 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
830 case Mips::ATOMIC_LOAD_SUB_I32:
831 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
833 case Mips::ATOMIC_SWAP_I8:
834 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
835 case Mips::ATOMIC_SWAP_I16:
836 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
837 case Mips::ATOMIC_SWAP_I32:
838 return EmitAtomicBinary(MI, BB, 4, 0);
840 case Mips::ATOMIC_CMP_SWAP_I8:
841 return EmitAtomicCmpSwapPartword(MI, BB, 1);
842 case Mips::ATOMIC_CMP_SWAP_I16:
843 return EmitAtomicCmpSwapPartword(MI, BB, 2);
844 case Mips::ATOMIC_CMP_SWAP_I32:
845 return EmitAtomicCmpSwap(MI, BB, 4);
849 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
850 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
852 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
853 unsigned Size, unsigned BinOpcode,
855 assert(Size == 4 && "Unsupported size for EmitAtomicBinary.");
857 MachineFunction *MF = BB->getParent();
858 MachineRegisterInfo &RegInfo = MF->getRegInfo();
859 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
860 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
861 DebugLoc dl = MI->getDebugLoc();
863 unsigned OldVal = MI->getOperand(0).getReg();
864 unsigned Ptr = MI->getOperand(1).getReg();
865 unsigned Incr = MI->getOperand(2).getReg();
867 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
868 unsigned AndRes = RegInfo.createVirtualRegister(RC);
869 unsigned Success = RegInfo.createVirtualRegister(RC);
871 // insert new blocks after the current block
872 const BasicBlock *LLVM_BB = BB->getBasicBlock();
873 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
874 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
875 MachineFunction::iterator It = BB;
877 MF->insert(It, loopMBB);
878 MF->insert(It, exitMBB);
880 // Transfer the remainder of BB and its successor edges to exitMBB.
881 exitMBB->splice(exitMBB->begin(), BB,
882 llvm::next(MachineBasicBlock::iterator(MI)),
884 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
888 // fallthrough --> loopMBB
889 BB->addSuccessor(loopMBB);
890 loopMBB->addSuccessor(loopMBB);
891 loopMBB->addSuccessor(exitMBB);
895 // <binop> storeval, oldval, incr
896 // sc success, storeval, 0(ptr)
897 // beq success, $0, loopMBB
899 BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(Ptr).addImm(0);
901 // and andres, oldval, incr
902 // nor storeval, $0, andres
903 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr);
904 BuildMI(BB, dl, TII->get(Mips::NOR), StoreVal)
905 .addReg(Mips::ZERO).addReg(AndRes);
906 } else if (BinOpcode) {
907 // <binop> storeval, oldval, incr
908 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
912 BuildMI(BB, dl, TII->get(Mips::SC), Success)
913 .addReg(StoreVal).addReg(Ptr).addImm(0);
914 BuildMI(BB, dl, TII->get(Mips::BEQ))
915 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
917 MI->eraseFromParent(); // The instruction is gone now.
923 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
924 MachineBasicBlock *BB,
925 unsigned Size, unsigned BinOpcode,
927 assert((Size == 1 || Size == 2) &&
928 "Unsupported size for EmitAtomicBinaryPartial.");
930 MachineFunction *MF = BB->getParent();
931 MachineRegisterInfo &RegInfo = MF->getRegInfo();
932 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
933 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
934 DebugLoc dl = MI->getDebugLoc();
936 unsigned Dest = MI->getOperand(0).getReg();
937 unsigned Ptr = MI->getOperand(1).getReg();
938 unsigned Incr = MI->getOperand(2).getReg();
940 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
941 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
942 unsigned Mask = RegInfo.createVirtualRegister(RC);
943 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
944 unsigned NewVal = RegInfo.createVirtualRegister(RC);
945 unsigned OldVal = RegInfo.createVirtualRegister(RC);
946 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
947 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
948 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
949 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
950 unsigned AndRes = RegInfo.createVirtualRegister(RC);
951 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
952 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
953 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
954 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
955 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
956 unsigned SllRes = RegInfo.createVirtualRegister(RC);
957 unsigned Success = RegInfo.createVirtualRegister(RC);
959 // insert new blocks after the current block
960 const BasicBlock *LLVM_BB = BB->getBasicBlock();
961 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
962 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
963 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
964 MachineFunction::iterator It = BB;
966 MF->insert(It, loopMBB);
967 MF->insert(It, sinkMBB);
968 MF->insert(It, exitMBB);
970 // Transfer the remainder of BB and its successor edges to exitMBB.
971 exitMBB->splice(exitMBB->begin(), BB,
972 llvm::next(MachineBasicBlock::iterator(MI)),
974 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
976 BB->addSuccessor(loopMBB);
977 loopMBB->addSuccessor(loopMBB);
978 loopMBB->addSuccessor(sinkMBB);
979 sinkMBB->addSuccessor(exitMBB);
982 // addiu masklsb2,$0,-4 # 0xfffffffc
983 // and alignedaddr,ptr,masklsb2
984 // andi ptrlsb2,ptr,3
985 // sll shiftamt,ptrlsb2,3
986 // ori maskupper,$0,255 # 0xff
987 // sll mask,maskupper,shiftamt
989 // sll incr2,incr,shiftamt
991 int64_t MaskImm = (Size == 1) ? 255 : 65535;
992 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
993 .addReg(Mips::ZERO).addImm(-4);
994 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
995 .addReg(Ptr).addReg(MaskLSB2);
996 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
997 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
998 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
999 .addReg(Mips::ZERO).addImm(MaskImm);
1000 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1001 .addReg(ShiftAmt).addReg(MaskUpper);
1002 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1003 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1006 // atomic.load.binop
1008 // ll oldval,0(alignedaddr)
1009 // binop binopres,oldval,incr2
1010 // and newval,binopres,mask
1011 // and maskedoldval0,oldval,mask2
1012 // or storeval,maskedoldval0,newval
1013 // sc success,storeval,0(alignedaddr)
1014 // beq success,$0,loopMBB
1018 // ll oldval,0(alignedaddr)
1019 // and newval,incr2,mask
1020 // and maskedoldval0,oldval,mask2
1021 // or storeval,maskedoldval0,newval
1022 // sc success,storeval,0(alignedaddr)
1023 // beq success,$0,loopMBB
1026 BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1028 // and andres, oldval, incr2
1029 // nor binopres, $0, andres
1030 // and newval, binopres, mask
1031 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1032 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1033 .addReg(Mips::ZERO).addReg(AndRes);
1034 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1035 } else if (BinOpcode) {
1036 // <binop> binopres, oldval, incr2
1037 // and newval, binopres, mask
1038 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1039 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1040 } else {// atomic.swap
1041 // and newval, incr2, mask
1042 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1045 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1046 .addReg(OldVal).addReg(Mask2);
1047 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1048 .addReg(MaskedOldVal0).addReg(NewVal);
1049 BuildMI(BB, dl, TII->get(Mips::SC), Success)
1050 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1051 BuildMI(BB, dl, TII->get(Mips::BEQ))
1052 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1055 // and maskedoldval1,oldval,mask
1056 // srl srlres,maskedoldval1,shiftamt
1057 // sll sllres,srlres,24
1058 // sra dest,sllres,24
1060 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1062 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1063 .addReg(OldVal).addReg(Mask);
1064 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1065 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1066 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1067 .addReg(SrlRes).addImm(ShiftImm);
1068 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1069 .addReg(SllRes).addImm(ShiftImm);
1071 MI->eraseFromParent(); // The instruction is gone now.
1077 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1078 MachineBasicBlock *BB,
1079 unsigned Size) const {
1080 assert(Size == 4 && "Unsupported size for EmitAtomicCmpSwap.");
1082 MachineFunction *MF = BB->getParent();
1083 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1084 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1085 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1086 DebugLoc dl = MI->getDebugLoc();
1088 unsigned Dest = MI->getOperand(0).getReg();
1089 unsigned Ptr = MI->getOperand(1).getReg();
1090 unsigned OldVal = MI->getOperand(2).getReg();
1091 unsigned NewVal = MI->getOperand(3).getReg();
1093 unsigned Success = RegInfo.createVirtualRegister(RC);
1095 // insert new blocks after the current block
1096 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1097 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1098 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1099 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1100 MachineFunction::iterator It = BB;
1102 MF->insert(It, loop1MBB);
1103 MF->insert(It, loop2MBB);
1104 MF->insert(It, exitMBB);
1106 // Transfer the remainder of BB and its successor edges to exitMBB.
1107 exitMBB->splice(exitMBB->begin(), BB,
1108 llvm::next(MachineBasicBlock::iterator(MI)),
1110 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1114 // fallthrough --> loop1MBB
1115 BB->addSuccessor(loop1MBB);
1116 loop1MBB->addSuccessor(exitMBB);
1117 loop1MBB->addSuccessor(loop2MBB);
1118 loop2MBB->addSuccessor(loop1MBB);
1119 loop2MBB->addSuccessor(exitMBB);
1123 // bne dest, oldval, exitMBB
1125 BuildMI(BB, dl, TII->get(Mips::LL), Dest).addReg(Ptr).addImm(0);
1126 BuildMI(BB, dl, TII->get(Mips::BNE))
1127 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1130 // sc success, newval, 0(ptr)
1131 // beq success, $0, loop1MBB
1133 BuildMI(BB, dl, TII->get(Mips::SC), Success)
1134 .addReg(NewVal).addReg(Ptr).addImm(0);
1135 BuildMI(BB, dl, TII->get(Mips::BEQ))
1136 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1138 MI->eraseFromParent(); // The instruction is gone now.
1144 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1145 MachineBasicBlock *BB,
1146 unsigned Size) const {
1147 assert((Size == 1 || Size == 2) &&
1148 "Unsupported size for EmitAtomicCmpSwapPartial.");
1150 MachineFunction *MF = BB->getParent();
1151 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1152 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1153 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1154 DebugLoc dl = MI->getDebugLoc();
1156 unsigned Dest = MI->getOperand(0).getReg();
1157 unsigned Ptr = MI->getOperand(1).getReg();
1158 unsigned CmpVal = MI->getOperand(2).getReg();
1159 unsigned NewVal = MI->getOperand(3).getReg();
1161 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1162 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1163 unsigned Mask = RegInfo.createVirtualRegister(RC);
1164 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1165 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1166 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1167 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1168 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1169 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1170 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1171 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1172 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1173 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1174 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1175 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1176 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1177 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1178 unsigned Success = RegInfo.createVirtualRegister(RC);
1180 // insert new blocks after the current block
1181 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1182 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1183 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1184 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1185 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1186 MachineFunction::iterator It = BB;
1188 MF->insert(It, loop1MBB);
1189 MF->insert(It, loop2MBB);
1190 MF->insert(It, sinkMBB);
1191 MF->insert(It, exitMBB);
1193 // Transfer the remainder of BB and its successor edges to exitMBB.
1194 exitMBB->splice(exitMBB->begin(), BB,
1195 llvm::next(MachineBasicBlock::iterator(MI)),
1197 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1199 BB->addSuccessor(loop1MBB);
1200 loop1MBB->addSuccessor(sinkMBB);
1201 loop1MBB->addSuccessor(loop2MBB);
1202 loop2MBB->addSuccessor(loop1MBB);
1203 loop2MBB->addSuccessor(sinkMBB);
1204 sinkMBB->addSuccessor(exitMBB);
1206 // FIXME: computation of newval2 can be moved to loop2MBB.
1208 // addiu masklsb2,$0,-4 # 0xfffffffc
1209 // and alignedaddr,ptr,masklsb2
1210 // andi ptrlsb2,ptr,3
1211 // sll shiftamt,ptrlsb2,3
1212 // ori maskupper,$0,255 # 0xff
1213 // sll mask,maskupper,shiftamt
1214 // nor mask2,$0,mask
1215 // andi maskedcmpval,cmpval,255
1216 // sll shiftedcmpval,maskedcmpval,shiftamt
1217 // andi maskednewval,newval,255
1218 // sll shiftednewval,maskednewval,shiftamt
1219 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1220 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1221 .addReg(Mips::ZERO).addImm(-4);
1222 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1223 .addReg(Ptr).addReg(MaskLSB2);
1224 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1225 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1226 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1227 .addReg(Mips::ZERO).addImm(MaskImm);
1228 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1229 .addReg(ShiftAmt).addReg(MaskUpper);
1230 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1231 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1232 .addReg(CmpVal).addImm(MaskImm);
1233 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1234 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1235 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1236 .addReg(NewVal).addImm(MaskImm);
1237 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1238 .addReg(ShiftAmt).addReg(MaskedNewVal);
1241 // ll oldval,0(alginedaddr)
1242 // and maskedoldval0,oldval,mask
1243 // bne maskedoldval0,shiftedcmpval,sinkMBB
1245 BuildMI(BB, dl, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0);
1246 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1247 .addReg(OldVal).addReg(Mask);
1248 BuildMI(BB, dl, TII->get(Mips::BNE))
1249 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1252 // and maskedoldval1,oldval,mask2
1253 // or storeval,maskedoldval1,shiftednewval
1254 // sc success,storeval,0(alignedaddr)
1255 // beq success,$0,loop1MBB
1257 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1258 .addReg(OldVal).addReg(Mask2);
1259 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1260 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1261 BuildMI(BB, dl, TII->get(Mips::SC), Success)
1262 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1263 BuildMI(BB, dl, TII->get(Mips::BEQ))
1264 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1267 // srl srlres,maskedoldval0,shiftamt
1268 // sll sllres,srlres,24
1269 // sra dest,sllres,24
1271 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1273 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1274 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1275 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1276 .addReg(SrlRes).addImm(ShiftImm);
1277 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1278 .addReg(SllRes).addImm(ShiftImm);
1280 MI->eraseFromParent(); // The instruction is gone now.
1285 //===----------------------------------------------------------------------===//
1286 // Misc Lower Operation implementation
1287 //===----------------------------------------------------------------------===//
1288 SDValue MipsTargetLowering::
1289 LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const
1291 MachineFunction &MF = DAG.getMachineFunction();
1292 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1294 assert(getTargetMachine().getFrameLowering()->getStackAlignment() >=
1295 cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue() &&
1296 "Cannot lower if the alignment of the allocated space is larger than \
1297 that of the stack.");
1299 SDValue Chain = Op.getOperand(0);
1300 SDValue Size = Op.getOperand(1);
1301 DebugLoc dl = Op.getDebugLoc();
1303 // Get a reference from Mips stack pointer
1304 SDValue StackPointer = DAG.getCopyFromReg(Chain, dl, Mips::SP, MVT::i32);
1306 // Subtract the dynamic size from the actual stack size to
1307 // obtain the new stack size.
1308 SDValue Sub = DAG.getNode(ISD::SUB, dl, MVT::i32, StackPointer, Size);
1310 // The Sub result contains the new stack start address, so it
1311 // must be placed in the stack pointer register.
1312 Chain = DAG.getCopyToReg(StackPointer.getValue(1), dl, Mips::SP, Sub,
1315 // This node always has two return values: a new stack pointer
1316 // value and a chain
1317 SDVTList VTLs = DAG.getVTList(MVT::i32, MVT::Other);
1318 SDValue Ptr = DAG.getFrameIndex(MipsFI->getDynAllocFI(), getPointerTy());
1319 SDValue Ops[] = { Chain, Ptr, Chain.getValue(1) };
1321 return DAG.getNode(MipsISD::DynAlloc, dl, VTLs, Ops, 3);
1324 SDValue MipsTargetLowering::
1325 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1327 // The first operand is the chain, the second is the condition, the third is
1328 // the block to branch to if the condition is true.
1329 SDValue Chain = Op.getOperand(0);
1330 SDValue Dest = Op.getOperand(2);
1331 DebugLoc dl = Op.getDebugLoc();
1333 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1335 // Return if flag is not set by a floating point comparison.
1336 if (CondRes.getOpcode() != MipsISD::FPCmp)
1339 SDValue CCNode = CondRes.getOperand(2);
1341 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1342 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1344 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1348 SDValue MipsTargetLowering::
1349 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1351 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1353 // Return if flag is not set by a floating point comparison.
1354 if (Cond.getOpcode() != MipsISD::FPCmp)
1357 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1361 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1362 SelectionDAG &DAG) const {
1363 // FIXME there isn't actually debug info here
1364 DebugLoc dl = Op.getDebugLoc();
1365 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1367 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1368 SDVTList VTs = DAG.getVTList(MVT::i32);
1370 MipsTargetObjectFile &TLOF = (MipsTargetObjectFile&)getObjFileLowering();
1372 // %gp_rel relocation
1373 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1374 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1376 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl, VTs, &GA, 1);
1377 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1378 return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode);
1380 // %hi/%lo relocation
1381 SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1383 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1385 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, VTs, &GAHi, 1);
1386 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1387 return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1390 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1392 GA = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, GA);
1393 SDValue ResNode = DAG.getLoad(MVT::i32, dl,
1394 DAG.getEntryNode(), GA, MachinePointerInfo(),
1396 // On functions and global targets not internal linked only
1397 // a load from got/GP is necessary for PIC to work.
1398 if (!GV->hasInternalLinkage() &&
1399 (!GV->hasLocalLinkage() || isa<Function>(GV)))
1401 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1403 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1404 return DAG.getNode(ISD::ADD, dl, MVT::i32, ResNode, Lo);
1407 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1408 SelectionDAG &DAG) const {
1409 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1410 // FIXME there isn't actually debug info here
1411 DebugLoc dl = Op.getDebugLoc();
1413 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1414 // %hi/%lo relocation
1415 SDValue BAHi = DAG.getBlockAddress(BA, MVT::i32, true,
1417 SDValue BALo = DAG.getBlockAddress(BA, MVT::i32, true,
1419 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, MVT::i32, BAHi);
1420 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALo);
1421 return DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1424 SDValue BAGOTOffset = DAG.getBlockAddress(BA, MVT::i32, true,
1426 BAGOTOffset = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, BAGOTOffset);
1427 SDValue BALOOffset = DAG.getBlockAddress(BA, MVT::i32, true,
1429 SDValue Load = DAG.getLoad(MVT::i32, dl,
1430 DAG.getEntryNode(), BAGOTOffset,
1431 MachinePointerInfo(), false, false, 0);
1432 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALOOffset);
1433 return DAG.getNode(ISD::ADD, dl, MVT::i32, Load, Lo);
1436 SDValue MipsTargetLowering::
1437 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1439 // If the relocation model is PIC, use the General Dynamic TLS Model,
1440 // otherwise use the Initial Exec or Local Exec TLS Model.
1441 // TODO: implement Local Dynamic TLS model
1443 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1444 DebugLoc dl = GA->getDebugLoc();
1445 const GlobalValue *GV = GA->getGlobal();
1446 EVT PtrVT = getPointerTy();
1448 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1449 // General Dynamic TLS Model
1450 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32,
1451 0, MipsII::MO_TLSGD);
1452 SDValue Tlsgd = DAG.getNode(MipsISD::TlsGd, dl, MVT::i32, TGA);
1453 SDValue GP = DAG.getRegister(Mips::GP, MVT::i32);
1454 SDValue Argument = DAG.getNode(ISD::ADD, dl, MVT::i32, GP, Tlsgd);
1458 Entry.Node = Argument;
1459 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
1460 Args.push_back(Entry);
1461 std::pair<SDValue, SDValue> CallResult =
1462 LowerCallTo(DAG.getEntryNode(),
1463 (Type *) Type::getInt32Ty(*DAG.getContext()),
1464 false, false, false, false, 0, CallingConv::C, false, true,
1465 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG,
1468 return CallResult.first;
1472 if (GV->isDeclaration()) {
1473 // Initial Exec TLS Model
1474 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1475 MipsII::MO_GOTTPREL);
1476 Offset = DAG.getLoad(MVT::i32, dl,
1477 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1480 // Local Exec TLS Model
1481 SDVTList VTs = DAG.getVTList(MVT::i32);
1482 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1483 MipsII::MO_TPREL_HI);
1484 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1485 MipsII::MO_TPREL_LO);
1486 SDValue Hi = DAG.getNode(MipsISD::TprelHi, dl, VTs, &TGAHi, 1);
1487 SDValue Lo = DAG.getNode(MipsISD::TprelLo, dl, MVT::i32, TGALo);
1488 Offset = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1491 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1492 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1495 SDValue MipsTargetLowering::
1496 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1500 // FIXME there isn't actually debug info here
1501 DebugLoc dl = Op.getDebugLoc();
1502 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1503 unsigned char OpFlag = IsPIC ? MipsII::MO_GOT : MipsII::MO_ABS_HI;
1505 EVT PtrVT = Op.getValueType();
1506 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1508 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
1511 SDValue Ops[] = { JTI };
1512 HiPart = DAG.getNode(MipsISD::Hi, dl, DAG.getVTList(MVT::i32), Ops, 1);
1513 } else {// Emit Load from Global Pointer
1514 JTI = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, JTI);
1515 HiPart = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(), JTI,
1516 MachinePointerInfo(),
1520 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
1522 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, JTILo);
1523 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1528 SDValue MipsTargetLowering::
1529 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1532 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1533 const Constant *C = N->getConstVal();
1534 // FIXME there isn't actually debug info here
1535 DebugLoc dl = Op.getDebugLoc();
1537 // gp_rel relocation
1538 // FIXME: we should reference the constant pool using small data sections,
1539 // but the asm printer currently doesn't support this feature without
1540 // hacking it. This feature should come soon so we can uncomment the
1542 //if (IsInSmallSection(C->getType())) {
1543 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1544 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1545 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1547 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1548 SDValue CPHi = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1549 N->getOffset(), MipsII::MO_ABS_HI);
1550 SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1551 N->getOffset(), MipsII::MO_ABS_LO);
1552 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, MVT::i32, CPHi);
1553 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1554 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1556 SDValue CP = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1557 N->getOffset(), MipsII::MO_GOT);
1558 CP = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, CP);
1559 SDValue Load = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(),
1560 CP, MachinePointerInfo::getConstantPool(),
1562 SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1563 N->getOffset(), MipsII::MO_ABS_LO);
1564 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1565 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, Load, Lo);
1571 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1572 MachineFunction &MF = DAG.getMachineFunction();
1573 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1575 DebugLoc dl = Op.getDebugLoc();
1576 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1579 // vastart just stores the address of the VarArgsFrameIndex slot into the
1580 // memory location argument.
1581 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1582 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1583 MachinePointerInfo(SV),
1587 static SDValue LowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG) {
1588 // FIXME: Use ext/ins instructions if target architecture is Mips32r2.
1589 DebugLoc dl = Op.getDebugLoc();
1590 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op.getOperand(0));
1591 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op.getOperand(1));
1592 SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op0,
1593 DAG.getConstant(0x7fffffff, MVT::i32));
1594 SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op1,
1595 DAG.getConstant(0x80000000, MVT::i32));
1596 SDValue Result = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1597 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Result);
1600 static SDValue LowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool isLittle) {
1602 // Use ext/ins instructions if target architecture is Mips32r2.
1603 // Eliminate redundant mfc1 and mtc1 instructions.
1604 unsigned LoIdx = 0, HiIdx = 1;
1607 std::swap(LoIdx, HiIdx);
1609 DebugLoc dl = Op.getDebugLoc();
1610 SDValue Word0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1612 DAG.getConstant(LoIdx, MVT::i32));
1613 SDValue Hi0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1614 Op.getOperand(0), DAG.getConstant(HiIdx, MVT::i32));
1615 SDValue Hi1 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1616 Op.getOperand(1), DAG.getConstant(HiIdx, MVT::i32));
1617 SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi0,
1618 DAG.getConstant(0x7fffffff, MVT::i32));
1619 SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi1,
1620 DAG.getConstant(0x80000000, MVT::i32));
1621 SDValue Word1 = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1624 std::swap(Word0, Word1);
1626 return DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64, Word0, Word1);
1629 SDValue MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG)
1631 EVT Ty = Op.getValueType();
1633 assert(Ty == MVT::f32 || Ty == MVT::f64);
1636 return LowerFCOPYSIGN32(Op, DAG);
1638 return LowerFCOPYSIGN64(Op, DAG, Subtarget->isLittle());
1641 SDValue MipsTargetLowering::
1642 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1644 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1645 "Frame address can only be determined for current frame.");
1647 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1648 MFI->setFrameAddressIsTaken(true);
1649 EVT VT = Op.getValueType();
1650 DebugLoc dl = Op.getDebugLoc();
1651 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, Mips::FP, VT);
1655 // TODO: set SType according to the desired memory barrier behavior.
1656 SDValue MipsTargetLowering::LowerMEMBARRIER(SDValue Op,
1657 SelectionDAG& DAG) const {
1659 DebugLoc dl = Op.getDebugLoc();
1660 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1661 DAG.getConstant(SType, MVT::i32));
1664 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1665 SelectionDAG& DAG) const {
1666 // FIXME: Need pseudo-fence for 'singlethread' fences
1667 // FIXME: Set SType for weaker fences where supported/appropriate.
1669 DebugLoc dl = Op.getDebugLoc();
1670 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1671 DAG.getConstant(SType, MVT::i32));
1674 //===----------------------------------------------------------------------===//
1675 // Calling Convention Implementation
1676 //===----------------------------------------------------------------------===//
1678 #include "MipsGenCallingConv.inc"
1680 //===----------------------------------------------------------------------===//
1681 // TODO: Implement a generic logic using tblgen that can support this.
1682 // Mips O32 ABI rules:
1684 // i32 - Passed in A0, A1, A2, A3 and stack
1685 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
1686 // an argument. Otherwise, passed in A1, A2, A3 and stack.
1687 // f64 - Only passed in two aliased f32 registers if no int reg has been used
1688 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
1689 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
1692 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
1693 //===----------------------------------------------------------------------===//
1695 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
1696 MVT LocVT, CCValAssign::LocInfo LocInfo,
1697 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1699 static const unsigned IntRegsSize=4, FloatRegsSize=2;
1701 static const unsigned IntRegs[] = {
1702 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1704 static const unsigned F32Regs[] = {
1705 Mips::F12, Mips::F14
1707 static const unsigned F64Regs[] = {
1712 if (ArgFlags.isByVal()) {
1713 State.HandleByVal(ValNo, ValVT, LocVT, LocInfo,
1714 1 /*MinSize*/, 4 /*MinAlign*/, ArgFlags);
1715 unsigned NextReg = (State.getNextStackOffset() + 3) / 4;
1716 for (unsigned r = State.getFirstUnallocated(IntRegs, IntRegsSize);
1717 r < std::min(IntRegsSize, NextReg); ++r)
1718 State.AllocateReg(IntRegs[r]);
1722 // Promote i8 and i16
1723 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
1725 if (ArgFlags.isSExt())
1726 LocInfo = CCValAssign::SExt;
1727 else if (ArgFlags.isZExt())
1728 LocInfo = CCValAssign::ZExt;
1730 LocInfo = CCValAssign::AExt;
1735 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
1736 // is true: function is vararg, argument is 3rd or higher, there is previous
1737 // argument which is not f32 or f64.
1738 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
1739 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
1740 unsigned OrigAlign = ArgFlags.getOrigAlign();
1741 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
1743 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
1744 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1745 // If this is the first part of an i64 arg,
1746 // the allocated register must be either A0 or A2.
1747 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
1748 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1750 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
1751 // Allocate int register and shadow next int register. If first
1752 // available register is Mips::A1 or Mips::A3, shadow it too.
1753 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1754 if (Reg == Mips::A1 || Reg == Mips::A3)
1755 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1756 State.AllocateReg(IntRegs, IntRegsSize);
1758 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
1759 // we are guaranteed to find an available float register
1760 if (ValVT == MVT::f32) {
1761 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
1762 // Shadow int register
1763 State.AllocateReg(IntRegs, IntRegsSize);
1765 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
1766 // Shadow int registers
1767 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
1768 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
1769 State.AllocateReg(IntRegs, IntRegsSize);
1770 State.AllocateReg(IntRegs, IntRegsSize);
1773 llvm_unreachable("Cannot handle this ValVT.");
1775 unsigned SizeInBytes = ValVT.getSizeInBits() >> 3;
1776 unsigned Offset = State.AllocateStack(SizeInBytes, OrigAlign);
1779 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1781 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1783 return false; // CC must always match
1786 //===----------------------------------------------------------------------===//
1787 // Call Calling Convention Implementation
1788 //===----------------------------------------------------------------------===//
1790 static const unsigned O32IntRegsSize = 4;
1792 static const unsigned O32IntRegs[] = {
1793 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1796 // Write ByVal Arg to arg registers and stack.
1798 WriteByValArg(SDValue& Chain, DebugLoc dl,
1799 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
1800 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
1801 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
1802 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
1803 MVT PtrType, bool isLittle) {
1804 unsigned LocMemOffset = VA.getLocMemOffset();
1805 unsigned Offset = 0;
1806 uint32_t RemainingSize = Flags.getByValSize();
1807 unsigned ByValAlign = Flags.getByValAlign();
1809 // Copy the first 4 words of byval arg to registers A0 - A3.
1810 // FIXME: Use a stricter alignment if it enables better optimization in passes
1812 for (; RemainingSize >= 4 && LocMemOffset < 4 * 4;
1813 Offset += 4, RemainingSize -= 4, LocMemOffset += 4) {
1814 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1815 DAG.getConstant(Offset, MVT::i32));
1816 SDValue LoadVal = DAG.getLoad(MVT::i32, dl, Chain, LoadPtr,
1817 MachinePointerInfo(),
1818 false, false, std::min(ByValAlign,
1820 MemOpChains.push_back(LoadVal.getValue(1));
1821 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
1822 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
1825 if (RemainingSize == 0)
1828 // If there still is a register available for argument passing, write the
1829 // remaining part of the structure to it using subword loads and shifts.
1830 if (LocMemOffset < 4 * 4) {
1831 assert(RemainingSize <= 3 && RemainingSize >= 1 &&
1832 "There must be one to three bytes remaining.");
1833 unsigned LoadSize = (RemainingSize == 3 ? 2 : RemainingSize);
1834 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1835 DAG.getConstant(Offset, MVT::i32));
1836 unsigned Alignment = std::min(ByValAlign, (unsigned )4);
1837 SDValue LoadVal = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
1838 LoadPtr, MachinePointerInfo(),
1839 MVT::getIntegerVT(LoadSize * 8), false,
1841 MemOpChains.push_back(LoadVal.getValue(1));
1843 // If target is big endian, shift it to the most significant half-word or
1846 LoadVal = DAG.getNode(ISD::SHL, dl, MVT::i32, LoadVal,
1847 DAG.getConstant(32 - LoadSize * 8, MVT::i32));
1850 RemainingSize -= LoadSize;
1852 // Read second subword if necessary.
1853 if (RemainingSize != 0) {
1854 assert(RemainingSize == 1 && "There must be one byte remaining.");
1855 LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1856 DAG.getConstant(Offset, MVT::i32));
1857 unsigned Alignment = std::min(ByValAlign, (unsigned )2);
1858 SDValue Subword = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
1859 LoadPtr, MachinePointerInfo(),
1860 MVT::i8, false, false, Alignment);
1861 MemOpChains.push_back(Subword.getValue(1));
1862 // Insert the loaded byte to LoadVal.
1863 // FIXME: Use INS if supported by target.
1864 unsigned ShiftAmt = isLittle ? 16 : 8;
1865 SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i32, Subword,
1866 DAG.getConstant(ShiftAmt, MVT::i32));
1867 LoadVal = DAG.getNode(ISD::OR, dl, MVT::i32, LoadVal, Shift);
1870 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
1871 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
1875 // Create a fixed object on stack at offset LocMemOffset and copy
1876 // remaining part of byval arg to it using memcpy.
1877 SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1878 DAG.getConstant(Offset, MVT::i32));
1879 LastFI = MFI->CreateFixedObject(RemainingSize, LocMemOffset, true);
1880 SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
1881 Chain = DAG.getMemcpy(Chain, dl, Dst, Src,
1882 DAG.getConstant(RemainingSize, MVT::i32),
1883 std::min(ByValAlign, (unsigned)4),
1884 /*isVolatile=*/false, /*AlwaysInline=*/false,
1885 MachinePointerInfo(0), MachinePointerInfo(0));
1886 MemOpChains.push_back(Chain);
1889 /// LowerCall - functions arguments are copied from virtual regs to
1890 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
1891 /// TODO: isTailCall.
1893 MipsTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1894 CallingConv::ID CallConv, bool isVarArg,
1896 const SmallVectorImpl<ISD::OutputArg> &Outs,
1897 const SmallVectorImpl<SDValue> &OutVals,
1898 const SmallVectorImpl<ISD::InputArg> &Ins,
1899 DebugLoc dl, SelectionDAG &DAG,
1900 SmallVectorImpl<SDValue> &InVals) const {
1901 // MIPs target does not yet support tail call optimization.
1904 MachineFunction &MF = DAG.getMachineFunction();
1905 MachineFrameInfo *MFI = MF.getFrameInfo();
1906 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
1907 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1908 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1910 // Analyze operands of the call, assigning locations to each operand.
1911 SmallVector<CCValAssign, 16> ArgLocs;
1912 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1913 getTargetMachine(), ArgLocs, *DAG.getContext());
1915 if (Subtarget->isABI_O32())
1916 CCInfo.AnalyzeCallOperands(Outs, CC_MipsO32);
1918 CCInfo.AnalyzeCallOperands(Outs, CC_Mips);
1920 // Get a count of how many bytes are to be pushed on the stack.
1921 unsigned NextStackOffset = CCInfo.getNextStackOffset();
1923 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NextStackOffset,
1926 // If this is the first call, create a stack frame object that points to
1927 // a location to which .cprestore saves $gp.
1928 if (IsPIC && !MipsFI->getGPFI())
1929 MipsFI->setGPFI(MFI->CreateFixedObject(4, 0, true));
1931 // Get the frame index of the stack frame object that points to the location
1932 // of dynamically allocated area on the stack.
1933 int DynAllocFI = MipsFI->getDynAllocFI();
1935 // Update size of the maximum argument space.
1936 // For O32, a minimum of four words (16 bytes) of argument space is
1938 if (Subtarget->isABI_O32())
1939 NextStackOffset = std::max(NextStackOffset, (unsigned)16);
1941 unsigned MaxCallFrameSize = MipsFI->getMaxCallFrameSize();
1943 if (MaxCallFrameSize < NextStackOffset) {
1944 MipsFI->setMaxCallFrameSize(NextStackOffset);
1946 // Set the offsets relative to $sp of the $gp restore slot and dynamically
1947 // allocated stack space. These offsets must be aligned to a boundary
1948 // determined by the stack alignment of the ABI.
1949 unsigned StackAlignment = TFL->getStackAlignment();
1950 NextStackOffset = (NextStackOffset + StackAlignment - 1) /
1951 StackAlignment * StackAlignment;
1954 MFI->setObjectOffset(MipsFI->getGPFI(), NextStackOffset);
1956 MFI->setObjectOffset(DynAllocFI, NextStackOffset);
1959 // With EABI is it possible to have 16 args on registers.
1960 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
1961 SmallVector<SDValue, 8> MemOpChains;
1963 int FirstFI = -MFI->getNumFixedObjects() - 1, LastFI = 0;
1965 // Walk the register/memloc assignments, inserting copies/loads.
1966 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1967 SDValue Arg = OutVals[i];
1968 CCValAssign &VA = ArgLocs[i];
1970 // Promote the value if needed.
1971 switch (VA.getLocInfo()) {
1972 default: llvm_unreachable("Unknown loc info!");
1973 case CCValAssign::Full:
1974 if (Subtarget->isABI_O32() && VA.isRegLoc()) {
1975 if (VA.getValVT() == MVT::f32 && VA.getLocVT() == MVT::i32)
1976 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
1977 if (VA.getValVT() == MVT::f64 && VA.getLocVT() == MVT::i32) {
1978 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1979 Arg, DAG.getConstant(0, MVT::i32));
1980 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1981 Arg, DAG.getConstant(1, MVT::i32));
1982 if (!Subtarget->isLittle())
1984 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Lo));
1985 RegsToPass.push_back(std::make_pair(VA.getLocReg()+1, Hi));
1990 case CCValAssign::SExt:
1991 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1993 case CCValAssign::ZExt:
1994 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1996 case CCValAssign::AExt:
1997 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
2001 // Arguments that can be passed on register must be kept at
2002 // RegsToPass vector
2003 if (VA.isRegLoc()) {
2004 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2008 // Register can't get to this point...
2009 assert(VA.isMemLoc());
2012 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2013 if (Flags.isByVal()) {
2014 assert(Subtarget->isABI_O32() &&
2015 "No support for ByVal args by ABIs other than O32 yet.");
2016 assert(Flags.getByValSize() &&
2017 "ByVal args of size 0 should have been ignored by front-end.");
2018 WriteByValArg(Chain, dl, RegsToPass, MemOpChains, LastFI, MFI, DAG, Arg,
2019 VA, Flags, getPointerTy(), Subtarget->isLittle());
2023 // Create the frame index object for this incoming parameter
2024 LastFI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
2025 VA.getLocMemOffset(), true);
2026 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2028 // emit ISD::STORE whichs stores the
2029 // parameter value to a stack Location
2030 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2031 MachinePointerInfo(),
2035 // Extend range of indices of frame objects for outgoing arguments that were
2036 // created during this function call. Skip this step if no such objects were
2039 MipsFI->extendOutArgFIRange(FirstFI, LastFI);
2041 // Transform all store nodes into one single node because all store
2042 // nodes are independent of each other.
2043 if (!MemOpChains.empty())
2044 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2045 &MemOpChains[0], MemOpChains.size());
2047 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2048 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2049 // node so that legalize doesn't hack it.
2050 unsigned char OpFlag = IsPIC ? MipsII::MO_GOT_CALL : MipsII::MO_NO_FLAG;
2051 bool LoadSymAddr = false;
2054 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2055 if (IsPIC && G->getGlobal()->hasInternalLinkage()) {
2056 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2057 getPointerTy(), 0,MipsII:: MO_GOT);
2058 CalleeLo = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(),
2059 0, MipsII::MO_ABS_LO);
2061 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2062 getPointerTy(), 0, OpFlag);
2067 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2068 Callee = DAG.getTargetExternalSymbol(S->getSymbol(),
2069 getPointerTy(), OpFlag);
2075 // Create nodes that load address of callee and copy it to T9
2078 // Load callee address
2079 Callee = DAG.getNode(MipsISD::WrapperPIC, dl, MVT::i32, Callee);
2080 SDValue LoadValue = DAG.getLoad(MVT::i32, dl, DAG.getEntryNode(), Callee,
2081 MachinePointerInfo::getGOT(),
2084 // Use GOT+LO if callee has internal linkage.
2085 if (CalleeLo.getNode()) {
2086 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CalleeLo);
2087 Callee = DAG.getNode(ISD::ADD, dl, MVT::i32, LoadValue, Lo);
2093 Chain = DAG.getCopyToReg(Chain, dl, Mips::T9, Callee, SDValue(0, 0));
2094 InFlag = Chain.getValue(1);
2095 Callee = DAG.getRegister(Mips::T9, MVT::i32);
2098 // Build a sequence of copy-to-reg nodes chained together with token
2099 // chain and flag operands which copy the outgoing args into registers.
2100 // The InFlag in necessary since all emitted instructions must be
2102 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2103 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2104 RegsToPass[i].second, InFlag);
2105 InFlag = Chain.getValue(1);
2108 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2109 // = Chain, Callee, Reg#1, Reg#2, ...
2111 // Returns a chain & a flag for retval copy to use.
2112 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2113 SmallVector<SDValue, 8> Ops;
2114 Ops.push_back(Chain);
2115 Ops.push_back(Callee);
2117 // Add argument registers to the end of the list so that they are
2118 // known live into the call.
2119 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2120 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2121 RegsToPass[i].second.getValueType()));
2123 if (InFlag.getNode())
2124 Ops.push_back(InFlag);
2126 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2127 InFlag = Chain.getValue(1);
2129 // Create the CALLSEQ_END node.
2130 Chain = DAG.getCALLSEQ_END(Chain,
2131 DAG.getIntPtrConstant(NextStackOffset, true),
2132 DAG.getIntPtrConstant(0, true), InFlag);
2133 InFlag = Chain.getValue(1);
2135 // Handle result values, copying them out of physregs into vregs that we
2137 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2138 Ins, dl, DAG, InVals);
2141 /// LowerCallResult - Lower the result values of a call into the
2142 /// appropriate copies out of appropriate physical registers.
2144 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2145 CallingConv::ID CallConv, bool isVarArg,
2146 const SmallVectorImpl<ISD::InputArg> &Ins,
2147 DebugLoc dl, SelectionDAG &DAG,
2148 SmallVectorImpl<SDValue> &InVals) const {
2149 // Assign locations to each value returned by this call.
2150 SmallVector<CCValAssign, 16> RVLocs;
2151 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2152 getTargetMachine(), RVLocs, *DAG.getContext());
2154 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2156 // Copy all of the result registers out of their specified physreg.
2157 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2158 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2159 RVLocs[i].getValVT(), InFlag).getValue(1);
2160 InFlag = Chain.getValue(2);
2161 InVals.push_back(Chain.getValue(0));
2167 //===----------------------------------------------------------------------===//
2168 // Formal Arguments Calling Convention Implementation
2169 //===----------------------------------------------------------------------===//
2170 static void ReadByValArg(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2171 std::vector<SDValue>& OutChains,
2172 SelectionDAG &DAG, unsigned NumWords, SDValue FIN,
2173 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags) {
2174 unsigned LocMem = VA.getLocMemOffset();
2175 unsigned FirstWord = LocMem / 4;
2177 // copy register A0 - A3 to frame object
2178 for (unsigned i = 0; i < NumWords; ++i) {
2179 unsigned CurWord = FirstWord + i;
2180 if (CurWord >= O32IntRegsSize)
2183 unsigned SrcReg = O32IntRegs[CurWord];
2184 unsigned Reg = AddLiveIn(MF, SrcReg, Mips::CPURegsRegisterClass);
2185 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIN,
2186 DAG.getConstant(i * 4, MVT::i32));
2187 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(Reg, MVT::i32),
2188 StorePtr, MachinePointerInfo(), false,
2190 OutChains.push_back(Store);
2194 /// LowerFormalArguments - transform physical registers into virtual registers
2195 /// and generate load operations for arguments places on the stack.
2197 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2198 CallingConv::ID CallConv,
2200 const SmallVectorImpl<ISD::InputArg>
2202 DebugLoc dl, SelectionDAG &DAG,
2203 SmallVectorImpl<SDValue> &InVals)
2205 MachineFunction &MF = DAG.getMachineFunction();
2206 MachineFrameInfo *MFI = MF.getFrameInfo();
2207 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2209 MipsFI->setVarArgsFrameIndex(0);
2211 // Used with vargs to acumulate store chains.
2212 std::vector<SDValue> OutChains;
2214 // Assign locations to all of the incoming arguments.
2215 SmallVector<CCValAssign, 16> ArgLocs;
2216 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2217 getTargetMachine(), ArgLocs, *DAG.getContext());
2219 if (Subtarget->isABI_O32())
2220 CCInfo.AnalyzeFormalArguments(Ins, CC_MipsO32);
2222 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips);
2224 int LastFI = 0;// MipsFI->LastInArgFI is 0 at the entry of this function.
2226 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2227 CCValAssign &VA = ArgLocs[i];
2229 // Arguments stored on registers
2230 if (VA.isRegLoc()) {
2231 EVT RegVT = VA.getLocVT();
2232 unsigned ArgReg = VA.getLocReg();
2233 TargetRegisterClass *RC = 0;
2235 if (RegVT == MVT::i32)
2236 RC = Mips::CPURegsRegisterClass;
2237 else if (RegVT == MVT::f32)
2238 RC = Mips::FGR32RegisterClass;
2239 else if (RegVT == MVT::f64) {
2240 if (!Subtarget->isSingleFloat())
2241 RC = Mips::AFGR64RegisterClass;
2243 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2245 // Transform the arguments stored on
2246 // physical registers into virtual ones
2247 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2248 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2250 // If this is an 8 or 16-bit value, it has been passed promoted
2251 // to 32 bits. Insert an assert[sz]ext to capture this, then
2252 // truncate to the right size.
2253 if (VA.getLocInfo() != CCValAssign::Full) {
2254 unsigned Opcode = 0;
2255 if (VA.getLocInfo() == CCValAssign::SExt)
2256 Opcode = ISD::AssertSext;
2257 else if (VA.getLocInfo() == CCValAssign::ZExt)
2258 Opcode = ISD::AssertZext;
2260 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
2261 DAG.getValueType(VA.getValVT()));
2262 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2265 // Handle O32 ABI cases: i32->f32 and (i32,i32)->f64
2266 if (Subtarget->isABI_O32()) {
2267 if (RegVT == MVT::i32 && VA.getValVT() == MVT::f32)
2268 ArgValue = DAG.getNode(ISD::BITCAST, dl, MVT::f32, ArgValue);
2269 if (RegVT == MVT::i32 && VA.getValVT() == MVT::f64) {
2270 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
2271 VA.getLocReg()+1, RC);
2272 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
2273 if (!Subtarget->isLittle())
2274 std::swap(ArgValue, ArgValue2);
2275 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
2276 ArgValue, ArgValue2);
2280 InVals.push_back(ArgValue);
2281 } else { // VA.isRegLoc()
2284 assert(VA.isMemLoc());
2286 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2288 if (Flags.isByVal()) {
2289 assert(Subtarget->isABI_O32() &&
2290 "No support for ByVal args by ABIs other than O32 yet.");
2291 assert(Flags.getByValSize() &&
2292 "ByVal args of size 0 should have been ignored by front-end.");
2293 unsigned NumWords = (Flags.getByValSize() + 3) / 4;
2294 LastFI = MFI->CreateFixedObject(NumWords * 4, VA.getLocMemOffset(),
2296 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2297 InVals.push_back(FIN);
2298 ReadByValArg(MF, Chain, dl, OutChains, DAG, NumWords, FIN, VA, Flags);
2303 // The stack pointer offset is relative to the caller stack frame.
2304 LastFI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
2305 VA.getLocMemOffset(), true);
2307 // Create load nodes to retrieve arguments from the stack
2308 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2309 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2310 MachinePointerInfo::getFixedStack(LastFI),
2315 // The mips ABIs for returning structs by value requires that we copy
2316 // the sret argument into $v0 for the return. Save the argument into
2317 // a virtual register so that we can access it from the return points.
2318 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2319 unsigned Reg = MipsFI->getSRetReturnReg();
2321 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i32));
2322 MipsFI->setSRetReturnReg(Reg);
2324 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2325 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2328 if (isVarArg && Subtarget->isABI_O32()) {
2329 // Record the frame index of the first variable argument
2330 // which is a value necessary to VASTART.
2331 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2332 assert(NextStackOffset % 4 == 0 &&
2333 "NextStackOffset must be aligned to 4-byte boundaries.");
2334 LastFI = MFI->CreateFixedObject(4, NextStackOffset, true);
2335 MipsFI->setVarArgsFrameIndex(LastFI);
2337 // If NextStackOffset is smaller than o32's 16-byte reserved argument area,
2338 // copy the integer registers that have not been used for argument passing
2339 // to the caller's stack frame.
2340 for (; NextStackOffset < 16; NextStackOffset += 4) {
2341 TargetRegisterClass *RC = Mips::CPURegsRegisterClass;
2342 unsigned Idx = NextStackOffset / 4;
2343 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), O32IntRegs[Idx], RC);
2344 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, MVT::i32);
2345 LastFI = MFI->CreateFixedObject(4, NextStackOffset, true);
2346 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2347 OutChains.push_back(DAG.getStore(Chain, dl, ArgValue, PtrOff,
2348 MachinePointerInfo(),
2353 MipsFI->setLastInArgFI(LastFI);
2355 // All stores are grouped in one node to allow the matching between
2356 // the size of Ins and InVals. This only happens when on varg functions
2357 if (!OutChains.empty()) {
2358 OutChains.push_back(Chain);
2359 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2360 &OutChains[0], OutChains.size());
2366 //===----------------------------------------------------------------------===//
2367 // Return Value Calling Convention Implementation
2368 //===----------------------------------------------------------------------===//
2371 MipsTargetLowering::LowerReturn(SDValue Chain,
2372 CallingConv::ID CallConv, bool isVarArg,
2373 const SmallVectorImpl<ISD::OutputArg> &Outs,
2374 const SmallVectorImpl<SDValue> &OutVals,
2375 DebugLoc dl, SelectionDAG &DAG) const {
2377 // CCValAssign - represent the assignment of
2378 // the return value to a location
2379 SmallVector<CCValAssign, 16> RVLocs;
2381 // CCState - Info about the registers and stack slot.
2382 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2383 getTargetMachine(), RVLocs, *DAG.getContext());
2385 // Analize return values.
2386 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
2388 // If this is the first return lowered for this function, add
2389 // the regs to the liveout set for the function.
2390 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2391 for (unsigned i = 0; i != RVLocs.size(); ++i)
2392 if (RVLocs[i].isRegLoc())
2393 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2398 // Copy the result values into the output registers.
2399 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2400 CCValAssign &VA = RVLocs[i];
2401 assert(VA.isRegLoc() && "Can only return in registers!");
2403 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2406 // guarantee that all emitted copies are
2407 // stuck together, avoiding something bad
2408 Flag = Chain.getValue(1);
2411 // The mips ABIs for returning structs by value requires that we copy
2412 // the sret argument into $v0 for the return. We saved the argument into
2413 // a virtual register in the entry block, so now we copy the value out
2415 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2416 MachineFunction &MF = DAG.getMachineFunction();
2417 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2418 unsigned Reg = MipsFI->getSRetReturnReg();
2421 llvm_unreachable("sret virtual register not created in the entry block");
2422 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2424 Chain = DAG.getCopyToReg(Chain, dl, Mips::V0, Val, Flag);
2425 Flag = Chain.getValue(1);
2428 // Return on Mips is always a "jr $ra"
2430 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2431 Chain, DAG.getRegister(Mips::RA, MVT::i32), Flag);
2433 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2434 Chain, DAG.getRegister(Mips::RA, MVT::i32));
2437 //===----------------------------------------------------------------------===//
2438 // Mips Inline Assembly Support
2439 //===----------------------------------------------------------------------===//
2441 /// getConstraintType - Given a constraint letter, return the type of
2442 /// constraint it is for this target.
2443 MipsTargetLowering::ConstraintType MipsTargetLowering::
2444 getConstraintType(const std::string &Constraint) const
2446 // Mips specific constrainy
2447 // GCC config/mips/constraints.md
2449 // 'd' : An address register. Equivalent to r
2450 // unless generating MIPS16 code.
2451 // 'y' : Equivalent to r; retained for
2452 // backwards compatibility.
2453 // 'f' : Floating Point registers.
2454 if (Constraint.size() == 1) {
2455 switch (Constraint[0]) {
2460 return C_RegisterClass;
2464 return TargetLowering::getConstraintType(Constraint);
2467 /// Examine constraint type and operand type and determine a weight value.
2468 /// This object must already have been set up with the operand type
2469 /// and the current alternative constraint selected.
2470 TargetLowering::ConstraintWeight
2471 MipsTargetLowering::getSingleConstraintMatchWeight(
2472 AsmOperandInfo &info, const char *constraint) const {
2473 ConstraintWeight weight = CW_Invalid;
2474 Value *CallOperandVal = info.CallOperandVal;
2475 // If we don't have a value, we can't do a match,
2476 // but allow it at the lowest weight.
2477 if (CallOperandVal == NULL)
2479 Type *type = CallOperandVal->getType();
2480 // Look at the constraint type.
2481 switch (*constraint) {
2483 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2487 if (type->isIntegerTy())
2488 weight = CW_Register;
2491 if (type->isFloatTy())
2492 weight = CW_Register;
2498 /// Given a register class constraint, like 'r', if this corresponds directly
2499 /// to an LLVM register class, return a register of 0 and the register class
2501 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2502 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
2504 if (Constraint.size() == 1) {
2505 switch (Constraint[0]) {
2506 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2507 case 'y': // Same as 'r'. Exists for compatibility.
2509 return std::make_pair(0U, Mips::CPURegsRegisterClass);
2512 return std::make_pair(0U, Mips::FGR32RegisterClass);
2514 if ((!Subtarget->isSingleFloat()) && (!Subtarget->isFP64bit()))
2515 return std::make_pair(0U, Mips::AFGR64RegisterClass);
2519 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2523 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2524 // The Mips target isn't yet aware of offsets.
2528 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2529 if (VT != MVT::f32 && VT != MVT::f64)
2531 if (Imm.isNegZero())
2533 return Imm.isZero();