1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that Mips uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "mips-lower"
16 #include "MipsISelLowering.h"
17 #include "MipsMachineFunction.h"
18 #include "MipsTargetMachine.h"
19 #include "MipsTargetObjectFile.h"
20 #include "MipsSubtarget.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/CallingConv.h"
26 #include "InstPrinter/MipsInstPrinter.h"
27 #include "MCTargetDesc/MipsBaseInfo.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/SelectionDAGISel.h"
34 #include "llvm/CodeGen/ValueTypes.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
39 // If I is a shifted mask, set the size (Size) and the first bit of the
40 // mask (Pos), and return true.
41 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
42 static bool IsShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
43 if (!isShiftedMask_64(I))
46 Size = CountPopulation_64(I);
47 Pos = CountTrailingZeros_64(I);
51 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
53 case MipsISD::JmpLink: return "MipsISD::JmpLink";
54 case MipsISD::Hi: return "MipsISD::Hi";
55 case MipsISD::Lo: return "MipsISD::Lo";
56 case MipsISD::GPRel: return "MipsISD::GPRel";
57 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
58 case MipsISD::Ret: return "MipsISD::Ret";
59 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
60 case MipsISD::FPCmp: return "MipsISD::FPCmp";
61 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
62 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
63 case MipsISD::FPRound: return "MipsISD::FPRound";
64 case MipsISD::MAdd: return "MipsISD::MAdd";
65 case MipsISD::MAddu: return "MipsISD::MAddu";
66 case MipsISD::MSub: return "MipsISD::MSub";
67 case MipsISD::MSubu: return "MipsISD::MSubu";
68 case MipsISD::DivRem: return "MipsISD::DivRem";
69 case MipsISD::DivRemU: return "MipsISD::DivRemU";
70 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
71 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
72 case MipsISD::Wrapper: return "MipsISD::Wrapper";
73 case MipsISD::DynAlloc: return "MipsISD::DynAlloc";
74 case MipsISD::Sync: return "MipsISD::Sync";
75 case MipsISD::Ext: return "MipsISD::Ext";
76 case MipsISD::Ins: return "MipsISD::Ins";
82 MipsTargetLowering(MipsTargetMachine &TM)
83 : TargetLowering(TM, new MipsTargetObjectFile()),
84 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
85 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
86 IsO32(Subtarget->isABI_O32()) {
88 // Mips does not have i1 type, so use i32 for
89 // setcc operations results (slt, sgt, ...).
90 setBooleanContents(ZeroOrOneBooleanContent);
91 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
93 // Set up the register classes
94 addRegisterClass(MVT::i32, Mips::CPURegsRegisterClass);
95 addRegisterClass(MVT::f32, Mips::FGR32RegisterClass);
98 addRegisterClass(MVT::i64, Mips::CPU64RegsRegisterClass);
100 // When dealing with single precision only, use libcalls
101 if (!Subtarget->isSingleFloat()) {
103 addRegisterClass(MVT::f64, Mips::FGR64RegisterClass);
105 addRegisterClass(MVT::f64, Mips::AFGR64RegisterClass);
108 // Load extented operations for i1 types must be promoted
109 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
110 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
111 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
113 // MIPS doesn't have extending float->double load/store
114 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
115 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
117 // Used by legalize types to correctly generate the setcc result.
118 // Without this, every float setcc comes with a AND/OR with the result,
119 // we don't want this, since the fpcmp result goes to a flag register,
120 // which is used implicitly by brcond and select operations.
121 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
123 // Mips Custom Operations
124 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
125 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
126 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
127 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
128 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
129 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
130 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
131 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
132 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
133 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
134 setOperationAction(ISD::SELECT, MVT::f32, Custom);
135 setOperationAction(ISD::SELECT, MVT::f64, Custom);
136 setOperationAction(ISD::SELECT, MVT::i32, Custom);
137 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
138 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
139 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Custom);
140 setOperationAction(ISD::VASTART, MVT::Other, Custom);
142 setOperationAction(ISD::SDIV, MVT::i32, Expand);
143 setOperationAction(ISD::SREM, MVT::i32, Expand);
144 setOperationAction(ISD::UDIV, MVT::i32, Expand);
145 setOperationAction(ISD::UREM, MVT::i32, Expand);
146 setOperationAction(ISD::SDIV, MVT::i64, Expand);
147 setOperationAction(ISD::SREM, MVT::i64, Expand);
148 setOperationAction(ISD::UDIV, MVT::i64, Expand);
149 setOperationAction(ISD::UREM, MVT::i64, Expand);
151 // Operations not directly supported by Mips.
152 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
153 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
154 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
155 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
156 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
157 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
158 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
159 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
160 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
161 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
162 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
163 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
164 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
165 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
166 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
167 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
168 setOperationAction(ISD::ROTL, MVT::i32, Expand);
169 setOperationAction(ISD::ROTL, MVT::i64, Expand);
171 if (!Subtarget->hasMips32r2())
172 setOperationAction(ISD::ROTR, MVT::i32, Expand);
174 if (!Subtarget->hasMips64r2())
175 setOperationAction(ISD::ROTR, MVT::i64, Expand);
177 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
178 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
179 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
180 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
181 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
182 setOperationAction(ISD::FSIN, MVT::f32, Expand);
183 setOperationAction(ISD::FSIN, MVT::f64, Expand);
184 setOperationAction(ISD::FCOS, MVT::f32, Expand);
185 setOperationAction(ISD::FCOS, MVT::f64, Expand);
186 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
187 setOperationAction(ISD::FPOW, MVT::f32, Expand);
188 setOperationAction(ISD::FPOW, MVT::f64, Expand);
189 setOperationAction(ISD::FLOG, MVT::f32, Expand);
190 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
191 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
192 setOperationAction(ISD::FEXP, MVT::f32, Expand);
193 setOperationAction(ISD::FMA, MVT::f32, Expand);
194 setOperationAction(ISD::FMA, MVT::f64, Expand);
196 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
197 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
199 setOperationAction(ISD::VAARG, MVT::Other, Expand);
200 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
201 setOperationAction(ISD::VAEND, MVT::Other, Expand);
203 // Use the default for now
204 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
205 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
207 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
208 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
210 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
211 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
212 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
213 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
215 setInsertFencesForAtomic(true);
217 if (Subtarget->isSingleFloat())
218 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
220 if (!Subtarget->hasSEInReg()) {
221 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
222 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
225 if (!Subtarget->hasBitCount()) {
226 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
227 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
230 if (!Subtarget->hasSwap()) {
231 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
232 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
235 setTargetDAGCombine(ISD::ADDE);
236 setTargetDAGCombine(ISD::SUBE);
237 setTargetDAGCombine(ISD::SDIVREM);
238 setTargetDAGCombine(ISD::UDIVREM);
239 setTargetDAGCombine(ISD::SETCC);
240 setTargetDAGCombine(ISD::AND);
241 setTargetDAGCombine(ISD::OR);
243 setMinFunctionAlignment(2);
245 setStackPointerRegisterToSaveRestore(HasMips64 ? Mips::SP_64 : Mips::SP);
246 computeRegisterProperties();
248 setExceptionPointerRegister(Mips::A0);
249 setExceptionSelectorRegister(Mips::A1);
252 bool MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
253 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
254 return SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16;
257 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
262 // Transforms a subgraph in CurDAG if the following pattern is found:
263 // (addc multLo, Lo0), (adde multHi, Hi0),
265 // multHi/Lo: product of multiplication
266 // Lo0: initial value of Lo register
267 // Hi0: initial value of Hi register
268 // Return true if pattern matching was successful.
269 static bool SelectMadd(SDNode* ADDENode, SelectionDAG* CurDAG) {
270 // ADDENode's second operand must be a flag output of an ADDC node in order
271 // for the matching to be successful.
272 SDNode* ADDCNode = ADDENode->getOperand(2).getNode();
274 if (ADDCNode->getOpcode() != ISD::ADDC)
277 SDValue MultHi = ADDENode->getOperand(0);
278 SDValue MultLo = ADDCNode->getOperand(0);
279 SDNode* MultNode = MultHi.getNode();
280 unsigned MultOpc = MultHi.getOpcode();
282 // MultHi and MultLo must be generated by the same node,
283 if (MultLo.getNode() != MultNode)
286 // and it must be a multiplication.
287 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
290 // MultLo amd MultHi must be the first and second output of MultNode
292 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
295 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
296 // of the values of MultNode, in which case MultNode will be removed in later
298 // If there exist users other than ADDENode or ADDCNode, this function returns
299 // here, which will result in MultNode being mapped to a single MULT
300 // instruction node rather than a pair of MULT and MADD instructions being
302 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
305 SDValue Chain = CurDAG->getEntryNode();
306 DebugLoc dl = ADDENode->getDebugLoc();
308 // create MipsMAdd(u) node
309 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
311 SDValue MAdd = CurDAG->getNode(MultOpc, dl, MVT::Glue,
312 MultNode->getOperand(0),// Factor 0
313 MultNode->getOperand(1),// Factor 1
314 ADDCNode->getOperand(1),// Lo0
315 ADDENode->getOperand(1));// Hi0
317 // create CopyFromReg nodes
318 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
320 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
322 CopyFromLo.getValue(2));
324 // replace uses of adde and addc here
325 if (!SDValue(ADDCNode, 0).use_empty())
326 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
328 if (!SDValue(ADDENode, 0).use_empty())
329 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
335 // Transforms a subgraph in CurDAG if the following pattern is found:
336 // (addc Lo0, multLo), (sube Hi0, multHi),
338 // multHi/Lo: product of multiplication
339 // Lo0: initial value of Lo register
340 // Hi0: initial value of Hi register
341 // Return true if pattern matching was successful.
342 static bool SelectMsub(SDNode* SUBENode, SelectionDAG* CurDAG) {
343 // SUBENode's second operand must be a flag output of an SUBC node in order
344 // for the matching to be successful.
345 SDNode* SUBCNode = SUBENode->getOperand(2).getNode();
347 if (SUBCNode->getOpcode() != ISD::SUBC)
350 SDValue MultHi = SUBENode->getOperand(1);
351 SDValue MultLo = SUBCNode->getOperand(1);
352 SDNode* MultNode = MultHi.getNode();
353 unsigned MultOpc = MultHi.getOpcode();
355 // MultHi and MultLo must be generated by the same node,
356 if (MultLo.getNode() != MultNode)
359 // and it must be a multiplication.
360 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
363 // MultLo amd MultHi must be the first and second output of MultNode
365 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
368 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
369 // of the values of MultNode, in which case MultNode will be removed in later
371 // If there exist users other than SUBENode or SUBCNode, this function returns
372 // here, which will result in MultNode being mapped to a single MULT
373 // instruction node rather than a pair of MULT and MSUB instructions being
375 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
378 SDValue Chain = CurDAG->getEntryNode();
379 DebugLoc dl = SUBENode->getDebugLoc();
381 // create MipsSub(u) node
382 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
384 SDValue MSub = CurDAG->getNode(MultOpc, dl, MVT::Glue,
385 MultNode->getOperand(0),// Factor 0
386 MultNode->getOperand(1),// Factor 1
387 SUBCNode->getOperand(0),// Lo0
388 SUBENode->getOperand(0));// Hi0
390 // create CopyFromReg nodes
391 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
393 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
395 CopyFromLo.getValue(2));
397 // replace uses of sube and subc here
398 if (!SDValue(SUBCNode, 0).use_empty())
399 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
401 if (!SDValue(SUBENode, 0).use_empty())
402 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
407 static SDValue PerformADDECombine(SDNode *N, SelectionDAG& DAG,
408 TargetLowering::DAGCombinerInfo &DCI,
409 const MipsSubtarget* Subtarget) {
410 if (DCI.isBeforeLegalize())
413 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
415 return SDValue(N, 0);
420 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG& DAG,
421 TargetLowering::DAGCombinerInfo &DCI,
422 const MipsSubtarget* Subtarget) {
423 if (DCI.isBeforeLegalize())
426 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
428 return SDValue(N, 0);
433 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG& DAG,
434 TargetLowering::DAGCombinerInfo &DCI,
435 const MipsSubtarget* Subtarget) {
436 if (DCI.isBeforeLegalizeOps())
439 EVT Ty = N->getValueType(0);
440 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
441 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
442 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
444 DebugLoc dl = N->getDebugLoc();
446 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
447 N->getOperand(0), N->getOperand(1));
448 SDValue InChain = DAG.getEntryNode();
449 SDValue InGlue = DivRem;
452 if (N->hasAnyUseOfValue(0)) {
453 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty,
455 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
456 InChain = CopyFromLo.getValue(1);
457 InGlue = CopyFromLo.getValue(2);
461 if (N->hasAnyUseOfValue(1)) {
462 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
464 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
470 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
472 default: llvm_unreachable("Unknown fp condition code!");
474 case ISD::SETOEQ: return Mips::FCOND_OEQ;
475 case ISD::SETUNE: return Mips::FCOND_UNE;
477 case ISD::SETOLT: return Mips::FCOND_OLT;
479 case ISD::SETOGT: return Mips::FCOND_OGT;
481 case ISD::SETOLE: return Mips::FCOND_OLE;
483 case ISD::SETOGE: return Mips::FCOND_OGE;
484 case ISD::SETULT: return Mips::FCOND_ULT;
485 case ISD::SETULE: return Mips::FCOND_ULE;
486 case ISD::SETUGT: return Mips::FCOND_UGT;
487 case ISD::SETUGE: return Mips::FCOND_UGE;
488 case ISD::SETUO: return Mips::FCOND_UN;
489 case ISD::SETO: return Mips::FCOND_OR;
491 case ISD::SETONE: return Mips::FCOND_ONE;
492 case ISD::SETUEQ: return Mips::FCOND_UEQ;
497 // Returns true if condition code has to be inverted.
498 static bool InvertFPCondCode(Mips::CondCode CC) {
499 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
502 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
503 "Illegal Condition Code");
508 // Creates and returns an FPCmp node from a setcc node.
509 // Returns Op if setcc is not a floating point comparison.
510 static SDValue CreateFPCmp(SelectionDAG& DAG, const SDValue& Op) {
511 // must be a SETCC node
512 if (Op.getOpcode() != ISD::SETCC)
515 SDValue LHS = Op.getOperand(0);
517 if (!LHS.getValueType().isFloatingPoint())
520 SDValue RHS = Op.getOperand(1);
521 DebugLoc dl = Op.getDebugLoc();
523 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
524 // node if necessary.
525 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
527 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
528 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
531 // Creates and returns a CMovFPT/F node.
532 static SDValue CreateCMovFP(SelectionDAG& DAG, SDValue Cond, SDValue True,
533 SDValue False, DebugLoc DL) {
534 bool invert = InvertFPCondCode((Mips::CondCode)
535 cast<ConstantSDNode>(Cond.getOperand(2))
538 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
539 True.getValueType(), True, False, Cond);
542 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG& DAG,
543 TargetLowering::DAGCombinerInfo &DCI,
544 const MipsSubtarget* Subtarget) {
545 if (DCI.isBeforeLegalizeOps())
548 SDValue Cond = CreateFPCmp(DAG, SDValue(N, 0));
550 if (Cond.getOpcode() != MipsISD::FPCmp)
553 SDValue True = DAG.getConstant(1, MVT::i32);
554 SDValue False = DAG.getConstant(0, MVT::i32);
556 return CreateCMovFP(DAG, Cond, True, False, N->getDebugLoc());
559 static SDValue PerformANDCombine(SDNode *N, SelectionDAG& DAG,
560 TargetLowering::DAGCombinerInfo &DCI,
561 const MipsSubtarget* Subtarget) {
562 // Pattern match EXT.
563 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
564 // => ext $dst, $src, size, pos
565 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
568 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
569 unsigned ShiftRightOpc = ShiftRight.getOpcode();
571 // Op's first operand must be a shift right.
572 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
575 // The second operand of the shift must be an immediate.
577 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
580 uint64_t Pos = CN->getZExtValue();
581 uint64_t SMPos, SMSize;
583 // Op's second operand must be a shifted mask.
584 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
585 !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
588 // Return if the shifted mask does not start at bit 0 or the sum of its size
589 // and Pos exceeds the word's size.
590 EVT ValTy = N->getValueType(0);
591 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
594 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
595 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
596 DAG.getConstant(SMSize, MVT::i32));
599 static SDValue PerformORCombine(SDNode *N, SelectionDAG& DAG,
600 TargetLowering::DAGCombinerInfo &DCI,
601 const MipsSubtarget* Subtarget) {
602 // Pattern match INS.
603 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
604 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
605 // => ins $dst, $src, size, pos, $src1
606 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
609 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
610 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
613 // See if Op's first operand matches (and $src1 , mask0).
614 if (And0.getOpcode() != ISD::AND)
617 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
618 !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
621 // See if Op's second operand matches (and (shl $src, pos), mask1).
622 if (And1.getOpcode() != ISD::AND)
625 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
626 !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
629 // The shift masks must have the same position and size.
630 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
633 SDValue Shl = And1.getOperand(0);
634 if (Shl.getOpcode() != ISD::SHL)
637 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
640 unsigned Shamt = CN->getZExtValue();
642 // Return if the shift amount and the first bit position of mask are not the
644 EVT ValTy = N->getValueType(0);
645 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
648 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
649 DAG.getConstant(SMPos0, MVT::i32),
650 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
653 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
655 SelectionDAG &DAG = DCI.DAG;
656 unsigned opc = N->getOpcode();
661 return PerformADDECombine(N, DAG, DCI, Subtarget);
663 return PerformSUBECombine(N, DAG, DCI, Subtarget);
666 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
668 return PerformSETCCCombine(N, DAG, DCI, Subtarget);
670 return PerformANDCombine(N, DAG, DCI, Subtarget);
672 return PerformORCombine(N, DAG, DCI, Subtarget);
678 SDValue MipsTargetLowering::
679 LowerOperation(SDValue Op, SelectionDAG &DAG) const
681 switch (Op.getOpcode())
683 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
684 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
685 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
686 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
687 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
688 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
689 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
690 case ISD::SELECT: return LowerSELECT(Op, DAG);
691 case ISD::VASTART: return LowerVASTART(Op, DAG);
692 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
693 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
694 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
695 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
700 //===----------------------------------------------------------------------===//
701 // Lower helper functions
702 //===----------------------------------------------------------------------===//
704 // AddLiveIn - This helper function adds the specified physical register to the
705 // MachineFunction as a live in value. It also creates a corresponding
706 // virtual register for it.
708 AddLiveIn(MachineFunction &MF, unsigned PReg, TargetRegisterClass *RC)
710 assert(RC->contains(PReg) && "Not the correct regclass!");
711 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
712 MF.getRegInfo().addLiveIn(PReg, VReg);
716 // Get fp branch code (not opcode) from condition code.
717 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
718 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
719 return Mips::BRANCH_T;
721 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
722 "Invalid CondCode.");
724 return Mips::BRANCH_F;
728 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
730 const MipsSubtarget* Subtarget,
731 const TargetInstrInfo *TII,
732 bool isFPCmp, unsigned Opc) {
733 // There is no need to expand CMov instructions if target has
734 // conditional moves.
735 if (Subtarget->hasCondMov())
738 // To "insert" a SELECT_CC instruction, we actually have to insert the
739 // diamond control-flow pattern. The incoming instruction knows the
740 // destination vreg to set, the condition code register to branch on, the
741 // true/false values to select between, and a branch opcode to use.
742 const BasicBlock *LLVM_BB = BB->getBasicBlock();
743 MachineFunction::iterator It = BB;
750 // bNE r1, r0, copy1MBB
751 // fallthrough --> copy0MBB
752 MachineBasicBlock *thisMBB = BB;
753 MachineFunction *F = BB->getParent();
754 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
755 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
756 F->insert(It, copy0MBB);
757 F->insert(It, sinkMBB);
759 // Transfer the remainder of BB and its successor edges to sinkMBB.
760 sinkMBB->splice(sinkMBB->begin(), BB,
761 llvm::next(MachineBasicBlock::iterator(MI)),
763 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
765 // Next, add the true and fallthrough blocks as its successors.
766 BB->addSuccessor(copy0MBB);
767 BB->addSuccessor(sinkMBB);
769 // Emit the right instruction according to the type of the operands compared
771 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
773 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
774 .addReg(Mips::ZERO).addMBB(sinkMBB);
778 // # fallthrough to sinkMBB
781 // Update machine-CFG edges
782 BB->addSuccessor(sinkMBB);
785 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
790 BuildMI(*BB, BB->begin(), dl,
791 TII->get(Mips::PHI), MI->getOperand(0).getReg())
792 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
793 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
795 BuildMI(*BB, BB->begin(), dl,
796 TII->get(Mips::PHI), MI->getOperand(0).getReg())
797 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
798 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
800 MI->eraseFromParent(); // The pseudo instruction is gone now.
805 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
806 MachineBasicBlock *BB) const {
807 switch (MI->getOpcode()) {
809 assert(false && "Unexpected instr type to insert");
811 case Mips::ATOMIC_LOAD_ADD_I8:
812 case Mips::ATOMIC_LOAD_ADD_I8_P8:
813 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
814 case Mips::ATOMIC_LOAD_ADD_I16:
815 case Mips::ATOMIC_LOAD_ADD_I16_P8:
816 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
817 case Mips::ATOMIC_LOAD_ADD_I32:
818 case Mips::ATOMIC_LOAD_ADD_I32_P8:
819 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
820 case Mips::ATOMIC_LOAD_ADD_I64:
821 case Mips::ATOMIC_LOAD_ADD_I64_P8:
822 return EmitAtomicBinary(MI, BB, 8, Mips::DADDu);
824 case Mips::ATOMIC_LOAD_AND_I8:
825 case Mips::ATOMIC_LOAD_AND_I8_P8:
826 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
827 case Mips::ATOMIC_LOAD_AND_I16:
828 case Mips::ATOMIC_LOAD_AND_I16_P8:
829 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
830 case Mips::ATOMIC_LOAD_AND_I32:
831 case Mips::ATOMIC_LOAD_AND_I32_P8:
832 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
833 case Mips::ATOMIC_LOAD_AND_I64:
834 case Mips::ATOMIC_LOAD_AND_I64_P8:
835 return EmitAtomicBinary(MI, BB, 8, Mips::AND64);
837 case Mips::ATOMIC_LOAD_OR_I8:
838 case Mips::ATOMIC_LOAD_OR_I8_P8:
839 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
840 case Mips::ATOMIC_LOAD_OR_I16:
841 case Mips::ATOMIC_LOAD_OR_I16_P8:
842 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
843 case Mips::ATOMIC_LOAD_OR_I32:
844 case Mips::ATOMIC_LOAD_OR_I32_P8:
845 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
846 case Mips::ATOMIC_LOAD_OR_I64:
847 case Mips::ATOMIC_LOAD_OR_I64_P8:
848 return EmitAtomicBinary(MI, BB, 8, Mips::OR64);
850 case Mips::ATOMIC_LOAD_XOR_I8:
851 case Mips::ATOMIC_LOAD_XOR_I8_P8:
852 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
853 case Mips::ATOMIC_LOAD_XOR_I16:
854 case Mips::ATOMIC_LOAD_XOR_I16_P8:
855 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
856 case Mips::ATOMIC_LOAD_XOR_I32:
857 case Mips::ATOMIC_LOAD_XOR_I32_P8:
858 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
859 case Mips::ATOMIC_LOAD_XOR_I64:
860 case Mips::ATOMIC_LOAD_XOR_I64_P8:
861 return EmitAtomicBinary(MI, BB, 8, Mips::XOR64);
863 case Mips::ATOMIC_LOAD_NAND_I8:
864 case Mips::ATOMIC_LOAD_NAND_I8_P8:
865 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
866 case Mips::ATOMIC_LOAD_NAND_I16:
867 case Mips::ATOMIC_LOAD_NAND_I16_P8:
868 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
869 case Mips::ATOMIC_LOAD_NAND_I32:
870 case Mips::ATOMIC_LOAD_NAND_I32_P8:
871 return EmitAtomicBinary(MI, BB, 4, 0, true);
872 case Mips::ATOMIC_LOAD_NAND_I64:
873 case Mips::ATOMIC_LOAD_NAND_I64_P8:
874 return EmitAtomicBinary(MI, BB, 8, 0, true);
876 case Mips::ATOMIC_LOAD_SUB_I8:
877 case Mips::ATOMIC_LOAD_SUB_I8_P8:
878 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
879 case Mips::ATOMIC_LOAD_SUB_I16:
880 case Mips::ATOMIC_LOAD_SUB_I16_P8:
881 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
882 case Mips::ATOMIC_LOAD_SUB_I32:
883 case Mips::ATOMIC_LOAD_SUB_I32_P8:
884 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
885 case Mips::ATOMIC_LOAD_SUB_I64:
886 case Mips::ATOMIC_LOAD_SUB_I64_P8:
887 return EmitAtomicBinary(MI, BB, 8, Mips::DSUBu);
889 case Mips::ATOMIC_SWAP_I8:
890 case Mips::ATOMIC_SWAP_I8_P8:
891 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
892 case Mips::ATOMIC_SWAP_I16:
893 case Mips::ATOMIC_SWAP_I16_P8:
894 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
895 case Mips::ATOMIC_SWAP_I32:
896 case Mips::ATOMIC_SWAP_I32_P8:
897 return EmitAtomicBinary(MI, BB, 4, 0);
898 case Mips::ATOMIC_SWAP_I64:
899 case Mips::ATOMIC_SWAP_I64_P8:
900 return EmitAtomicBinary(MI, BB, 8, 0);
902 case Mips::ATOMIC_CMP_SWAP_I8:
903 case Mips::ATOMIC_CMP_SWAP_I8_P8:
904 return EmitAtomicCmpSwapPartword(MI, BB, 1);
905 case Mips::ATOMIC_CMP_SWAP_I16:
906 case Mips::ATOMIC_CMP_SWAP_I16_P8:
907 return EmitAtomicCmpSwapPartword(MI, BB, 2);
908 case Mips::ATOMIC_CMP_SWAP_I32:
909 case Mips::ATOMIC_CMP_SWAP_I32_P8:
910 return EmitAtomicCmpSwap(MI, BB, 4);
911 case Mips::ATOMIC_CMP_SWAP_I64:
912 case Mips::ATOMIC_CMP_SWAP_I64_P8:
913 return EmitAtomicCmpSwap(MI, BB, 8);
917 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
918 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
920 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
921 unsigned Size, unsigned BinOpcode,
923 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
925 MachineFunction *MF = BB->getParent();
926 MachineRegisterInfo &RegInfo = MF->getRegInfo();
927 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
928 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
929 DebugLoc dl = MI->getDebugLoc();
930 unsigned LL, SC, AND, NOR, ZERO, BEQ;
933 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
934 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
941 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
942 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
945 ZERO = Mips::ZERO_64;
949 unsigned OldVal = MI->getOperand(0).getReg();
950 unsigned Ptr = MI->getOperand(1).getReg();
951 unsigned Incr = MI->getOperand(2).getReg();
953 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
954 unsigned AndRes = RegInfo.createVirtualRegister(RC);
955 unsigned Success = RegInfo.createVirtualRegister(RC);
957 // insert new blocks after the current block
958 const BasicBlock *LLVM_BB = BB->getBasicBlock();
959 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
960 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
961 MachineFunction::iterator It = BB;
963 MF->insert(It, loopMBB);
964 MF->insert(It, exitMBB);
966 // Transfer the remainder of BB and its successor edges to exitMBB.
967 exitMBB->splice(exitMBB->begin(), BB,
968 llvm::next(MachineBasicBlock::iterator(MI)),
970 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
974 // fallthrough --> loopMBB
975 BB->addSuccessor(loopMBB);
976 loopMBB->addSuccessor(loopMBB);
977 loopMBB->addSuccessor(exitMBB);
981 // <binop> storeval, oldval, incr
982 // sc success, storeval, 0(ptr)
983 // beq success, $0, loopMBB
985 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
987 // and andres, oldval, incr
988 // nor storeval, $0, andres
989 BuildMI(BB, dl, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
990 BuildMI(BB, dl, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
991 } else if (BinOpcode) {
992 // <binop> storeval, oldval, incr
993 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
997 BuildMI(BB, dl, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
998 BuildMI(BB, dl, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1000 MI->eraseFromParent(); // The instruction is gone now.
1006 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
1007 MachineBasicBlock *BB,
1008 unsigned Size, unsigned BinOpcode,
1010 assert((Size == 1 || Size == 2) &&
1011 "Unsupported size for EmitAtomicBinaryPartial.");
1013 MachineFunction *MF = BB->getParent();
1014 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1015 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1016 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1017 DebugLoc dl = MI->getDebugLoc();
1018 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1019 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1021 unsigned Dest = MI->getOperand(0).getReg();
1022 unsigned Ptr = MI->getOperand(1).getReg();
1023 unsigned Incr = MI->getOperand(2).getReg();
1025 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1026 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1027 unsigned Mask = RegInfo.createVirtualRegister(RC);
1028 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1029 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1030 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1031 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1032 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1033 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1034 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1035 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1036 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1037 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1038 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1039 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1040 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1041 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1042 unsigned Success = RegInfo.createVirtualRegister(RC);
1044 // insert new blocks after the current block
1045 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1046 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1047 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1048 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1049 MachineFunction::iterator It = BB;
1051 MF->insert(It, loopMBB);
1052 MF->insert(It, sinkMBB);
1053 MF->insert(It, exitMBB);
1055 // Transfer the remainder of BB and its successor edges to exitMBB.
1056 exitMBB->splice(exitMBB->begin(), BB,
1057 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1058 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1060 BB->addSuccessor(loopMBB);
1061 loopMBB->addSuccessor(loopMBB);
1062 loopMBB->addSuccessor(sinkMBB);
1063 sinkMBB->addSuccessor(exitMBB);
1066 // addiu masklsb2,$0,-4 # 0xfffffffc
1067 // and alignedaddr,ptr,masklsb2
1068 // andi ptrlsb2,ptr,3
1069 // sll shiftamt,ptrlsb2,3
1070 // ori maskupper,$0,255 # 0xff
1071 // sll mask,maskupper,shiftamt
1072 // nor mask2,$0,mask
1073 // sll incr2,incr,shiftamt
1075 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1076 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1077 .addReg(Mips::ZERO).addImm(-4);
1078 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1079 .addReg(Ptr).addReg(MaskLSB2);
1080 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1081 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1082 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1083 .addReg(Mips::ZERO).addImm(MaskImm);
1084 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1085 .addReg(ShiftAmt).addReg(MaskUpper);
1086 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1087 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1089 // atomic.load.binop
1091 // ll oldval,0(alignedaddr)
1092 // binop binopres,oldval,incr2
1093 // and newval,binopres,mask
1094 // and maskedoldval0,oldval,mask2
1095 // or storeval,maskedoldval0,newval
1096 // sc success,storeval,0(alignedaddr)
1097 // beq success,$0,loopMBB
1101 // ll oldval,0(alignedaddr)
1102 // and newval,incr2,mask
1103 // and maskedoldval0,oldval,mask2
1104 // or storeval,maskedoldval0,newval
1105 // sc success,storeval,0(alignedaddr)
1106 // beq success,$0,loopMBB
1109 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1111 // and andres, oldval, incr2
1112 // nor binopres, $0, andres
1113 // and newval, binopres, mask
1114 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1115 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1116 .addReg(Mips::ZERO).addReg(AndRes);
1117 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1118 } else if (BinOpcode) {
1119 // <binop> binopres, oldval, incr2
1120 // and newval, binopres, mask
1121 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1122 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1123 } else {// atomic.swap
1124 // and newval, incr2, mask
1125 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1128 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1129 .addReg(OldVal).addReg(Mask2);
1130 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1131 .addReg(MaskedOldVal0).addReg(NewVal);
1132 BuildMI(BB, dl, TII->get(SC), Success)
1133 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1134 BuildMI(BB, dl, TII->get(Mips::BEQ))
1135 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1138 // and maskedoldval1,oldval,mask
1139 // srl srlres,maskedoldval1,shiftamt
1140 // sll sllres,srlres,24
1141 // sra dest,sllres,24
1143 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1145 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1146 .addReg(OldVal).addReg(Mask);
1147 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1148 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1149 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1150 .addReg(SrlRes).addImm(ShiftImm);
1151 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1152 .addReg(SllRes).addImm(ShiftImm);
1154 MI->eraseFromParent(); // The instruction is gone now.
1160 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1161 MachineBasicBlock *BB,
1162 unsigned Size) const {
1163 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1165 MachineFunction *MF = BB->getParent();
1166 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1167 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1168 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1169 DebugLoc dl = MI->getDebugLoc();
1170 unsigned LL, SC, ZERO, BNE, BEQ;
1173 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1174 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1180 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1181 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1182 ZERO = Mips::ZERO_64;
1187 unsigned Dest = MI->getOperand(0).getReg();
1188 unsigned Ptr = MI->getOperand(1).getReg();
1189 unsigned OldVal = MI->getOperand(2).getReg();
1190 unsigned NewVal = MI->getOperand(3).getReg();
1192 unsigned Success = RegInfo.createVirtualRegister(RC);
1194 // insert new blocks after the current block
1195 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1196 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1197 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1198 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1199 MachineFunction::iterator It = BB;
1201 MF->insert(It, loop1MBB);
1202 MF->insert(It, loop2MBB);
1203 MF->insert(It, exitMBB);
1205 // Transfer the remainder of BB and its successor edges to exitMBB.
1206 exitMBB->splice(exitMBB->begin(), BB,
1207 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1208 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1212 // fallthrough --> loop1MBB
1213 BB->addSuccessor(loop1MBB);
1214 loop1MBB->addSuccessor(exitMBB);
1215 loop1MBB->addSuccessor(loop2MBB);
1216 loop2MBB->addSuccessor(loop1MBB);
1217 loop2MBB->addSuccessor(exitMBB);
1221 // bne dest, oldval, exitMBB
1223 BuildMI(BB, dl, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1224 BuildMI(BB, dl, TII->get(BNE))
1225 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1228 // sc success, newval, 0(ptr)
1229 // beq success, $0, loop1MBB
1231 BuildMI(BB, dl, TII->get(SC), Success)
1232 .addReg(NewVal).addReg(Ptr).addImm(0);
1233 BuildMI(BB, dl, TII->get(BEQ))
1234 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1236 MI->eraseFromParent(); // The instruction is gone now.
1242 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1243 MachineBasicBlock *BB,
1244 unsigned Size) const {
1245 assert((Size == 1 || Size == 2) &&
1246 "Unsupported size for EmitAtomicCmpSwapPartial.");
1248 MachineFunction *MF = BB->getParent();
1249 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1250 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1251 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1252 DebugLoc dl = MI->getDebugLoc();
1253 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1254 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1256 unsigned Dest = MI->getOperand(0).getReg();
1257 unsigned Ptr = MI->getOperand(1).getReg();
1258 unsigned CmpVal = MI->getOperand(2).getReg();
1259 unsigned NewVal = MI->getOperand(3).getReg();
1261 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1262 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1263 unsigned Mask = RegInfo.createVirtualRegister(RC);
1264 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1265 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1266 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1267 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1268 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1269 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1270 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1271 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1272 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1273 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1274 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1275 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1276 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1277 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1278 unsigned Success = RegInfo.createVirtualRegister(RC);
1280 // insert new blocks after the current block
1281 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1282 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1283 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1284 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1285 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1286 MachineFunction::iterator It = BB;
1288 MF->insert(It, loop1MBB);
1289 MF->insert(It, loop2MBB);
1290 MF->insert(It, sinkMBB);
1291 MF->insert(It, exitMBB);
1293 // Transfer the remainder of BB and its successor edges to exitMBB.
1294 exitMBB->splice(exitMBB->begin(), BB,
1295 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1296 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1298 BB->addSuccessor(loop1MBB);
1299 loop1MBB->addSuccessor(sinkMBB);
1300 loop1MBB->addSuccessor(loop2MBB);
1301 loop2MBB->addSuccessor(loop1MBB);
1302 loop2MBB->addSuccessor(sinkMBB);
1303 sinkMBB->addSuccessor(exitMBB);
1305 // FIXME: computation of newval2 can be moved to loop2MBB.
1307 // addiu masklsb2,$0,-4 # 0xfffffffc
1308 // and alignedaddr,ptr,masklsb2
1309 // andi ptrlsb2,ptr,3
1310 // sll shiftamt,ptrlsb2,3
1311 // ori maskupper,$0,255 # 0xff
1312 // sll mask,maskupper,shiftamt
1313 // nor mask2,$0,mask
1314 // andi maskedcmpval,cmpval,255
1315 // sll shiftedcmpval,maskedcmpval,shiftamt
1316 // andi maskednewval,newval,255
1317 // sll shiftednewval,maskednewval,shiftamt
1318 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1319 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1320 .addReg(Mips::ZERO).addImm(-4);
1321 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1322 .addReg(Ptr).addReg(MaskLSB2);
1323 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1324 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1325 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1326 .addReg(Mips::ZERO).addImm(MaskImm);
1327 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1328 .addReg(ShiftAmt).addReg(MaskUpper);
1329 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1330 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1331 .addReg(CmpVal).addImm(MaskImm);
1332 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1333 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1334 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1335 .addReg(NewVal).addImm(MaskImm);
1336 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1337 .addReg(ShiftAmt).addReg(MaskedNewVal);
1340 // ll oldval,0(alginedaddr)
1341 // and maskedoldval0,oldval,mask
1342 // bne maskedoldval0,shiftedcmpval,sinkMBB
1344 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1345 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1346 .addReg(OldVal).addReg(Mask);
1347 BuildMI(BB, dl, TII->get(Mips::BNE))
1348 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1351 // and maskedoldval1,oldval,mask2
1352 // or storeval,maskedoldval1,shiftednewval
1353 // sc success,storeval,0(alignedaddr)
1354 // beq success,$0,loop1MBB
1356 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1357 .addReg(OldVal).addReg(Mask2);
1358 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1359 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1360 BuildMI(BB, dl, TII->get(SC), Success)
1361 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1362 BuildMI(BB, dl, TII->get(Mips::BEQ))
1363 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1366 // srl srlres,maskedoldval0,shiftamt
1367 // sll sllres,srlres,24
1368 // sra dest,sllres,24
1370 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1372 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1373 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1374 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1375 .addReg(SrlRes).addImm(ShiftImm);
1376 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1377 .addReg(SllRes).addImm(ShiftImm);
1379 MI->eraseFromParent(); // The instruction is gone now.
1384 //===----------------------------------------------------------------------===//
1385 // Misc Lower Operation implementation
1386 //===----------------------------------------------------------------------===//
1387 SDValue MipsTargetLowering::
1388 LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const
1390 MachineFunction &MF = DAG.getMachineFunction();
1391 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1392 unsigned SP = IsN64 ? Mips::SP_64 : Mips::SP;
1394 assert(getTargetMachine().getFrameLowering()->getStackAlignment() >=
1395 cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue() &&
1396 "Cannot lower if the alignment of the allocated space is larger than \
1397 that of the stack.");
1399 SDValue Chain = Op.getOperand(0);
1400 SDValue Size = Op.getOperand(1);
1401 DebugLoc dl = Op.getDebugLoc();
1403 // Get a reference from Mips stack pointer
1404 SDValue StackPointer = DAG.getCopyFromReg(Chain, dl, SP, getPointerTy());
1406 // Subtract the dynamic size from the actual stack size to
1407 // obtain the new stack size.
1408 SDValue Sub = DAG.getNode(ISD::SUB, dl, getPointerTy(), StackPointer, Size);
1410 // The Sub result contains the new stack start address, so it
1411 // must be placed in the stack pointer register.
1412 Chain = DAG.getCopyToReg(StackPointer.getValue(1), dl, SP, Sub, SDValue());
1414 // This node always has two return values: a new stack pointer
1415 // value and a chain
1416 SDVTList VTLs = DAG.getVTList(getPointerTy(), MVT::Other);
1417 SDValue Ptr = DAG.getFrameIndex(MipsFI->getDynAllocFI(), getPointerTy());
1418 SDValue Ops[] = { Chain, Ptr, Chain.getValue(1) };
1420 return DAG.getNode(MipsISD::DynAlloc, dl, VTLs, Ops, 3);
1423 SDValue MipsTargetLowering::
1424 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1426 // The first operand is the chain, the second is the condition, the third is
1427 // the block to branch to if the condition is true.
1428 SDValue Chain = Op.getOperand(0);
1429 SDValue Dest = Op.getOperand(2);
1430 DebugLoc dl = Op.getDebugLoc();
1432 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1434 // Return if flag is not set by a floating point comparison.
1435 if (CondRes.getOpcode() != MipsISD::FPCmp)
1438 SDValue CCNode = CondRes.getOperand(2);
1440 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1441 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1443 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1447 SDValue MipsTargetLowering::
1448 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1450 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1452 // Return if flag is not set by a floating point comparison.
1453 if (Cond.getOpcode() != MipsISD::FPCmp)
1456 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1460 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1461 SelectionDAG &DAG) const {
1462 // FIXME there isn't actually debug info here
1463 DebugLoc dl = Op.getDebugLoc();
1464 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1466 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1467 SDVTList VTs = DAG.getVTList(MVT::i32);
1469 MipsTargetObjectFile &TLOF = (MipsTargetObjectFile&)getObjFileLowering();
1471 // %gp_rel relocation
1472 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1473 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1475 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl, VTs, &GA, 1);
1476 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1477 return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode);
1479 // %hi/%lo relocation
1480 SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1482 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1484 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, VTs, &GAHi, 1);
1485 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1486 return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1489 EVT ValTy = Op.getValueType();
1490 bool HasGotOfst = (GV->hasInternalLinkage() ||
1491 (GV->hasLocalLinkage() && !isa<Function>(GV)));
1492 unsigned GotFlag = IsN64 ?
1493 (HasGotOfst ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT_DISP) :
1494 (HasGotOfst ? MipsII::MO_GOT : MipsII::MO_GOT16);
1495 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0, GotFlag);
1496 GA = DAG.getNode(MipsISD::Wrapper, dl, ValTy, GA);
1497 SDValue ResNode = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), GA,
1498 MachinePointerInfo(), false, false, false, 0);
1499 // On functions and global targets not internal linked only
1500 // a load from got/GP is necessary for PIC to work.
1503 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0,
1504 IsN64 ? MipsII::MO_GOT_OFST :
1506 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, GALo);
1507 return DAG.getNode(ISD::ADD, dl, ValTy, ResNode, Lo);
1510 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1511 SelectionDAG &DAG) const {
1512 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1513 // FIXME there isn't actually debug info here
1514 DebugLoc dl = Op.getDebugLoc();
1516 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1517 // %hi/%lo relocation
1518 SDValue BAHi = DAG.getBlockAddress(BA, MVT::i32, true, MipsII::MO_ABS_HI);
1519 SDValue BALo = DAG.getBlockAddress(BA, MVT::i32, true, MipsII::MO_ABS_LO);
1520 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, MVT::i32, BAHi);
1521 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALo);
1522 return DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1525 EVT ValTy = Op.getValueType();
1526 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1527 unsigned OFSTFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1528 SDValue BAGOTOffset = DAG.getBlockAddress(BA, ValTy, true, GOTFlag);
1529 BAGOTOffset = DAG.getNode(MipsISD::Wrapper, dl, ValTy, BAGOTOffset);
1530 SDValue BALOOffset = DAG.getBlockAddress(BA, ValTy, true, OFSTFlag);
1531 SDValue Load = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), BAGOTOffset,
1532 MachinePointerInfo(), false, false, false, 0);
1533 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, BALOOffset);
1534 return DAG.getNode(ISD::ADD, dl, ValTy, Load, Lo);
1537 SDValue MipsTargetLowering::
1538 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1540 // If the relocation model is PIC, use the General Dynamic TLS Model or
1541 // Local Dynamic TLS model, otherwise use the Initial Exec or
1542 // Local Exec TLS Model.
1544 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1545 DebugLoc dl = GA->getDebugLoc();
1546 const GlobalValue *GV = GA->getGlobal();
1547 EVT PtrVT = getPointerTy();
1549 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1550 // General Dynamic TLS Model
1551 bool LocalDynamic = GV->hasInternalLinkage();
1552 unsigned Flag = LocalDynamic ? MipsII::MO_TLSLDM :MipsII::MO_TLSGD;
1553 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, Flag);
1554 SDValue Argument = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, TGA);
1555 unsigned PtrSize = PtrVT.getSizeInBits();
1556 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1558 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1562 Entry.Node = Argument;
1564 Args.push_back(Entry);
1566 std::pair<SDValue, SDValue> CallResult =
1567 LowerCallTo(DAG.getEntryNode(), PtrTy,
1568 false, false, false, false, 0, CallingConv::C, false, true,
1569 TlsGetAddr, Args, DAG, dl);
1571 SDValue Ret = CallResult.first;
1576 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1577 MipsII::MO_DTPREL_HI);
1578 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1579 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1580 MipsII::MO_DTPREL_LO);
1581 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1582 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Ret);
1583 return DAG.getNode(ISD::ADD, dl, PtrVT, Add, Lo);
1587 if (GV->isDeclaration()) {
1588 // Initial Exec TLS Model
1589 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1590 MipsII::MO_GOTTPREL);
1591 TGA = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, TGA);
1592 Offset = DAG.getLoad(PtrVT, dl,
1593 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1594 false, false, false, 0);
1596 // Local Exec TLS Model
1597 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1598 MipsII::MO_TPREL_HI);
1599 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1600 MipsII::MO_TPREL_LO);
1601 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1602 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1603 Offset = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1606 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1607 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1610 SDValue MipsTargetLowering::
1611 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1613 SDValue HiPart, JTI, JTILo;
1614 // FIXME there isn't actually debug info here
1615 DebugLoc dl = Op.getDebugLoc();
1616 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1617 EVT PtrVT = Op.getValueType();
1618 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1620 if (!IsPIC && !IsN64) {
1621 JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MipsII::MO_ABS_HI);
1622 HiPart = DAG.getNode(MipsISD::Hi, dl, PtrVT, JTI);
1623 JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MipsII::MO_ABS_LO);
1624 } else {// Emit Load from Global Pointer
1625 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1626 unsigned OfstFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1627 JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, GOTFlag);
1628 JTI = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, JTI);
1629 HiPart = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), JTI,
1630 MachinePointerInfo(), false, false, false, 0);
1631 JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OfstFlag);
1634 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, JTILo);
1635 return DAG.getNode(ISD::ADD, dl, PtrVT, HiPart, Lo);
1638 SDValue MipsTargetLowering::
1639 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1642 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1643 const Constant *C = N->getConstVal();
1644 // FIXME there isn't actually debug info here
1645 DebugLoc dl = Op.getDebugLoc();
1647 // gp_rel relocation
1648 // FIXME: we should reference the constant pool using small data sections,
1649 // but the asm printer currently doesn't support this feature without
1650 // hacking it. This feature should come soon so we can uncomment the
1652 //if (IsInSmallSection(C->getType())) {
1653 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1654 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1655 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1657 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1658 SDValue CPHi = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1659 N->getOffset(), MipsII::MO_ABS_HI);
1660 SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1661 N->getOffset(), MipsII::MO_ABS_LO);
1662 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, MVT::i32, CPHi);
1663 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1664 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1666 EVT ValTy = Op.getValueType();
1667 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1668 unsigned OFSTFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1669 SDValue CP = DAG.getTargetConstantPool(C, ValTy, N->getAlignment(),
1670 N->getOffset(), GOTFlag);
1671 CP = DAG.getNode(MipsISD::Wrapper, dl, ValTy, CP);
1672 SDValue Load = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), CP,
1673 MachinePointerInfo::getConstantPool(), false,
1675 SDValue CPLo = DAG.getTargetConstantPool(C, ValTy, N->getAlignment(),
1676 N->getOffset(), OFSTFlag);
1677 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, CPLo);
1678 ResNode = DAG.getNode(ISD::ADD, dl, ValTy, Load, Lo);
1684 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1685 MachineFunction &MF = DAG.getMachineFunction();
1686 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1688 DebugLoc dl = Op.getDebugLoc();
1689 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1692 // vastart just stores the address of the VarArgsFrameIndex slot into the
1693 // memory location argument.
1694 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1695 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1696 MachinePointerInfo(SV), false, false, 0);
1699 // Called if the size of integer registers is large enough to hold the whole
1700 // floating point number.
1701 static SDValue LowerFCOPYSIGNLargeIntReg(SDValue Op, SelectionDAG &DAG) {
1702 // FIXME: Use ext/ins instructions if target architecture is Mips32r2.
1703 EVT ValTy = Op.getValueType();
1704 EVT IntValTy = MVT::getIntegerVT(ValTy.getSizeInBits());
1705 uint64_t Mask = (uint64_t)1 << (ValTy.getSizeInBits() - 1);
1706 DebugLoc dl = Op.getDebugLoc();
1707 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntValTy, Op.getOperand(0));
1708 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntValTy, Op.getOperand(1));
1709 SDValue And0 = DAG.getNode(ISD::AND, dl, IntValTy, Op0,
1710 DAG.getConstant(Mask - 1, IntValTy));
1711 SDValue And1 = DAG.getNode(ISD::AND, dl, IntValTy, Op1,
1712 DAG.getConstant(Mask, IntValTy));
1713 SDValue Result = DAG.getNode(ISD::OR, dl, IntValTy, And0, And1);
1714 return DAG.getNode(ISD::BITCAST, dl, ValTy, Result);
1717 // Called if the size of integer registers is not large enough to hold the whole
1718 // floating point number (e.g. f64 & 32-bit integer register).
1720 LowerFCOPYSIGNSmallIntReg(SDValue Op, SelectionDAG &DAG, bool isLittle) {
1722 // Use ext/ins instructions if target architecture is Mips32r2.
1723 // Eliminate redundant mfc1 and mtc1 instructions.
1724 unsigned LoIdx = 0, HiIdx = 1;
1727 std::swap(LoIdx, HiIdx);
1729 DebugLoc dl = Op.getDebugLoc();
1730 SDValue Word0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1732 DAG.getConstant(LoIdx, MVT::i32));
1733 SDValue Hi0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1734 Op.getOperand(0), DAG.getConstant(HiIdx, MVT::i32));
1735 SDValue Hi1 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1736 Op.getOperand(1), DAG.getConstant(HiIdx, MVT::i32));
1737 SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi0,
1738 DAG.getConstant(0x7fffffff, MVT::i32));
1739 SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi1,
1740 DAG.getConstant(0x80000000, MVT::i32));
1741 SDValue Word1 = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1744 std::swap(Word0, Word1);
1746 return DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64, Word0, Word1);
1750 MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1751 EVT Ty = Op.getValueType();
1753 assert(Ty == MVT::f32 || Ty == MVT::f64);
1755 if (Ty == MVT::f32 || HasMips64)
1756 return LowerFCOPYSIGNLargeIntReg(Op, DAG);
1758 return LowerFCOPYSIGNSmallIntReg(Op, DAG, Subtarget->isLittle());
1761 SDValue MipsTargetLowering::
1762 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1764 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1765 "Frame address can only be determined for current frame.");
1767 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1768 MFI->setFrameAddressIsTaken(true);
1769 EVT VT = Op.getValueType();
1770 DebugLoc dl = Op.getDebugLoc();
1771 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
1772 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1776 // TODO: set SType according to the desired memory barrier behavior.
1778 MipsTargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG& DAG) const {
1780 DebugLoc dl = Op.getDebugLoc();
1781 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1782 DAG.getConstant(SType, MVT::i32));
1785 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1786 SelectionDAG& DAG) const {
1787 // FIXME: Need pseudo-fence for 'singlethread' fences
1788 // FIXME: Set SType for weaker fences where supported/appropriate.
1790 DebugLoc dl = Op.getDebugLoc();
1791 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1792 DAG.getConstant(SType, MVT::i32));
1795 //===----------------------------------------------------------------------===//
1796 // Calling Convention Implementation
1797 //===----------------------------------------------------------------------===//
1799 //===----------------------------------------------------------------------===//
1800 // TODO: Implement a generic logic using tblgen that can support this.
1801 // Mips O32 ABI rules:
1803 // i32 - Passed in A0, A1, A2, A3 and stack
1804 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
1805 // an argument. Otherwise, passed in A1, A2, A3 and stack.
1806 // f64 - Only passed in two aliased f32 registers if no int reg has been used
1807 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
1808 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
1811 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
1812 //===----------------------------------------------------------------------===//
1814 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
1815 MVT LocVT, CCValAssign::LocInfo LocInfo,
1816 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1818 static const unsigned IntRegsSize=4, FloatRegsSize=2;
1820 static const unsigned IntRegs[] = {
1821 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1823 static const unsigned F32Regs[] = {
1824 Mips::F12, Mips::F14
1826 static const unsigned F64Regs[] = {
1831 if (ArgFlags.isByVal()) {
1832 State.HandleByVal(ValNo, ValVT, LocVT, LocInfo,
1833 1 /*MinSize*/, 4 /*MinAlign*/, ArgFlags);
1834 unsigned NextReg = (State.getNextStackOffset() + 3) / 4;
1835 for (unsigned r = State.getFirstUnallocated(IntRegs, IntRegsSize);
1836 r < std::min(IntRegsSize, NextReg); ++r)
1837 State.AllocateReg(IntRegs[r]);
1841 // Promote i8 and i16
1842 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
1844 if (ArgFlags.isSExt())
1845 LocInfo = CCValAssign::SExt;
1846 else if (ArgFlags.isZExt())
1847 LocInfo = CCValAssign::ZExt;
1849 LocInfo = CCValAssign::AExt;
1854 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
1855 // is true: function is vararg, argument is 3rd or higher, there is previous
1856 // argument which is not f32 or f64.
1857 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
1858 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
1859 unsigned OrigAlign = ArgFlags.getOrigAlign();
1860 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
1862 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
1863 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1864 // If this is the first part of an i64 arg,
1865 // the allocated register must be either A0 or A2.
1866 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
1867 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1869 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
1870 // Allocate int register and shadow next int register. If first
1871 // available register is Mips::A1 or Mips::A3, shadow it too.
1872 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1873 if (Reg == Mips::A1 || Reg == Mips::A3)
1874 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1875 State.AllocateReg(IntRegs, IntRegsSize);
1877 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
1878 // we are guaranteed to find an available float register
1879 if (ValVT == MVT::f32) {
1880 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
1881 // Shadow int register
1882 State.AllocateReg(IntRegs, IntRegsSize);
1884 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
1885 // Shadow int registers
1886 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
1887 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
1888 State.AllocateReg(IntRegs, IntRegsSize);
1889 State.AllocateReg(IntRegs, IntRegsSize);
1892 llvm_unreachable("Cannot handle this ValVT.");
1894 unsigned SizeInBytes = ValVT.getSizeInBits() >> 3;
1895 unsigned Offset = State.AllocateStack(SizeInBytes, OrigAlign);
1898 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1900 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1902 return false; // CC must always match
1905 static const unsigned Mips64IntRegs[8] =
1906 {Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
1907 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64};
1908 static const unsigned Mips64DPRegs[8] =
1909 {Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
1910 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64};
1912 static bool CC_Mips64Byval(unsigned ValNo, MVT ValVT, MVT LocVT,
1913 CCValAssign::LocInfo LocInfo,
1914 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1915 unsigned Align = std::max(ArgFlags.getByValAlign(), (unsigned)8);
1916 unsigned Size = (ArgFlags.getByValSize() + 7) / 8 * 8;
1917 unsigned FirstIdx = State.getFirstUnallocated(Mips64IntRegs, 8);
1919 assert(Align <= 16 && "Cannot handle alignments larger than 16.");
1921 // If byval is 16-byte aligned, the first arg register must be even.
1922 if ((Align == 16) && (FirstIdx % 2)) {
1923 State.AllocateReg(Mips64IntRegs[FirstIdx], Mips64DPRegs[FirstIdx]);
1927 // Mark the registers allocated.
1928 for (unsigned I = FirstIdx; Size && (I < 8); Size -= 8, ++I)
1929 State.AllocateReg(Mips64IntRegs[I], Mips64DPRegs[I]);
1931 // Allocate space on caller's stack.
1932 unsigned Offset = State.AllocateStack(Size, Align);
1935 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Mips64IntRegs[FirstIdx],
1938 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1943 #include "MipsGenCallingConv.inc"
1946 AnalyzeMips64CallOperands(CCState CCInfo,
1947 const SmallVectorImpl<ISD::OutputArg> &Outs) {
1948 unsigned NumOps = Outs.size();
1949 for (unsigned i = 0; i != NumOps; ++i) {
1950 MVT ArgVT = Outs[i].VT;
1951 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
1954 if (Outs[i].IsFixed)
1955 R = CC_MipsN(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
1957 R = CC_MipsN_VarArg(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
1961 dbgs() << "Call operand #" << i << " has unhandled type "
1962 << EVT(ArgVT).getEVTString();
1964 llvm_unreachable(0);
1969 //===----------------------------------------------------------------------===//
1970 // Call Calling Convention Implementation
1971 //===----------------------------------------------------------------------===//
1973 static const unsigned O32IntRegsSize = 4;
1975 static const unsigned O32IntRegs[] = {
1976 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1979 // Return next O32 integer argument register.
1980 static unsigned getNextIntArgReg(unsigned Reg) {
1981 assert((Reg == Mips::A0) || (Reg == Mips::A2));
1982 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
1985 // Write ByVal Arg to arg registers and stack.
1987 WriteByValArg(SDValue& ByValChain, SDValue Chain, DebugLoc dl,
1988 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
1989 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
1990 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
1991 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
1992 MVT PtrType, bool isLittle) {
1993 unsigned LocMemOffset = VA.getLocMemOffset();
1994 unsigned Offset = 0;
1995 uint32_t RemainingSize = Flags.getByValSize();
1996 unsigned ByValAlign = Flags.getByValAlign();
1998 // Copy the first 4 words of byval arg to registers A0 - A3.
1999 // FIXME: Use a stricter alignment if it enables better optimization in passes
2001 for (; RemainingSize >= 4 && LocMemOffset < 4 * 4;
2002 Offset += 4, RemainingSize -= 4, LocMemOffset += 4) {
2003 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2004 DAG.getConstant(Offset, MVT::i32));
2005 SDValue LoadVal = DAG.getLoad(MVT::i32, dl, Chain, LoadPtr,
2006 MachinePointerInfo(), false, false, false,
2007 std::min(ByValAlign, (unsigned )4));
2008 MemOpChains.push_back(LoadVal.getValue(1));
2009 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
2010 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
2013 if (RemainingSize == 0)
2016 // If there still is a register available for argument passing, write the
2017 // remaining part of the structure to it using subword loads and shifts.
2018 if (LocMemOffset < 4 * 4) {
2019 assert(RemainingSize <= 3 && RemainingSize >= 1 &&
2020 "There must be one to three bytes remaining.");
2021 unsigned LoadSize = (RemainingSize == 3 ? 2 : RemainingSize);
2022 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2023 DAG.getConstant(Offset, MVT::i32));
2024 unsigned Alignment = std::min(ByValAlign, (unsigned )4);
2025 SDValue LoadVal = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
2026 LoadPtr, MachinePointerInfo(),
2027 MVT::getIntegerVT(LoadSize * 8), false,
2029 MemOpChains.push_back(LoadVal.getValue(1));
2031 // If target is big endian, shift it to the most significant half-word or
2034 LoadVal = DAG.getNode(ISD::SHL, dl, MVT::i32, LoadVal,
2035 DAG.getConstant(32 - LoadSize * 8, MVT::i32));
2038 RemainingSize -= LoadSize;
2040 // Read second subword if necessary.
2041 if (RemainingSize != 0) {
2042 assert(RemainingSize == 1 && "There must be one byte remaining.");
2043 LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2044 DAG.getConstant(Offset, MVT::i32));
2045 unsigned Alignment = std::min(ByValAlign, (unsigned )2);
2046 SDValue Subword = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
2047 LoadPtr, MachinePointerInfo(),
2048 MVT::i8, false, false, Alignment);
2049 MemOpChains.push_back(Subword.getValue(1));
2050 // Insert the loaded byte to LoadVal.
2051 // FIXME: Use INS if supported by target.
2052 unsigned ShiftAmt = isLittle ? 16 : 8;
2053 SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i32, Subword,
2054 DAG.getConstant(ShiftAmt, MVT::i32));
2055 LoadVal = DAG.getNode(ISD::OR, dl, MVT::i32, LoadVal, Shift);
2058 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
2059 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
2063 // Create a fixed object on stack at offset LocMemOffset and copy
2064 // remaining part of byval arg to it using memcpy.
2065 SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2066 DAG.getConstant(Offset, MVT::i32));
2067 LastFI = MFI->CreateFixedObject(RemainingSize, LocMemOffset, true);
2068 SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
2069 ByValChain = DAG.getMemcpy(ByValChain, dl, Dst, Src,
2070 DAG.getConstant(RemainingSize, MVT::i32),
2071 std::min(ByValAlign, (unsigned)4),
2072 /*isVolatile=*/false, /*AlwaysInline=*/false,
2073 MachinePointerInfo(0), MachinePointerInfo(0));
2076 // Copy Mips64 byVal arg to registers and stack.
2078 PassByValArg64(SDValue& ByValChain, SDValue Chain, DebugLoc dl,
2079 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
2080 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
2081 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
2082 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
2083 EVT PtrTy, bool isLittle) {
2084 unsigned ByValSize = Flags.getByValSize();
2085 unsigned Alignment = std::min(Flags.getByValAlign(), (unsigned)8);
2086 bool IsRegLoc = VA.isRegLoc();
2087 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
2088 unsigned LocMemOffset = 0;
2089 unsigned MemCpySize = ByValSize;
2092 LocMemOffset = VA.getLocMemOffset();
2094 const unsigned *Reg = std::find(Mips64IntRegs, Mips64IntRegs + 8,
2096 const unsigned *RegEnd = Mips64IntRegs + 8;
2098 // Copy double words to registers.
2099 for (; (Reg != RegEnd) && (ByValSize >= Offset + 8); ++Reg, Offset += 8) {
2100 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2101 DAG.getConstant(Offset, PtrTy));
2102 SDValue LoadVal = DAG.getLoad(MVT::i64, dl, Chain, LoadPtr,
2103 MachinePointerInfo(), false, false, false,
2105 MemOpChains.push_back(LoadVal.getValue(1));
2106 RegsToPass.push_back(std::make_pair(*Reg, LoadVal));
2109 // Return if the struct has been fully copied.
2110 if (!(MemCpySize = ByValSize - Offset))
2113 // If there is an argument register available, copy the remainder of the
2114 // byval argument with sub-doubleword loads and shifts.
2115 if (Reg != RegEnd) {
2116 assert((ByValSize < Offset + 8) &&
2117 "Size of the remainder should be smaller than 8-byte.");
2119 for (unsigned LoadSize = 4; Offset < ByValSize; LoadSize /= 2) {
2120 unsigned RemSize = ByValSize - Offset;
2122 if (RemSize < LoadSize)
2125 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2126 DAG.getConstant(Offset, PtrTy));
2128 DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i64, Chain, LoadPtr,
2129 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
2130 false, false, Alignment);
2131 MemOpChains.push_back(LoadVal.getValue(1));
2133 // Offset in number of bits from double word boundary.
2134 unsigned OffsetDW = (Offset % 8) * 8;
2135 unsigned Shamt = isLittle ? OffsetDW : 64 - (OffsetDW + LoadSize * 8);
2136 SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i64, LoadVal,
2137 DAG.getConstant(Shamt, MVT::i32));
2139 Val = Val.getNode() ? DAG.getNode(ISD::OR, dl, MVT::i64, Val, Shift) :
2142 Alignment = std::min(Alignment, LoadSize);
2145 RegsToPass.push_back(std::make_pair(*Reg, Val));
2150 assert(MemCpySize && "MemCpySize must not be zero.");
2152 // Create a fixed object on stack at offset LocMemOffset and copy
2153 // remainder of byval arg to it with memcpy.
2154 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2155 DAG.getConstant(Offset, PtrTy));
2156 LastFI = MFI->CreateFixedObject(MemCpySize, LocMemOffset, true);
2157 SDValue Dst = DAG.getFrameIndex(LastFI, PtrTy);
2158 ByValChain = DAG.getMemcpy(ByValChain, dl, Dst, Src,
2159 DAG.getConstant(MemCpySize, PtrTy), Alignment,
2160 /*isVolatile=*/false, /*AlwaysInline=*/false,
2161 MachinePointerInfo(0), MachinePointerInfo(0));
2164 /// LowerCall - functions arguments are copied from virtual regs to
2165 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2166 /// TODO: isTailCall.
2168 MipsTargetLowering::LowerCall(SDValue InChain, SDValue Callee,
2169 CallingConv::ID CallConv, bool isVarArg,
2171 const SmallVectorImpl<ISD::OutputArg> &Outs,
2172 const SmallVectorImpl<SDValue> &OutVals,
2173 const SmallVectorImpl<ISD::InputArg> &Ins,
2174 DebugLoc dl, SelectionDAG &DAG,
2175 SmallVectorImpl<SDValue> &InVals) const {
2176 // MIPs target does not yet support tail call optimization.
2179 MachineFunction &MF = DAG.getMachineFunction();
2180 MachineFrameInfo *MFI = MF.getFrameInfo();
2181 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2182 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2183 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2185 // Analyze operands of the call, assigning locations to each operand.
2186 SmallVector<CCValAssign, 16> ArgLocs;
2187 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2188 getTargetMachine(), ArgLocs, *DAG.getContext());
2191 CCInfo.AnalyzeCallOperands(Outs, CC_MipsO32);
2193 AnalyzeMips64CallOperands(CCInfo, Outs);
2195 CCInfo.AnalyzeCallOperands(Outs, CC_Mips);
2197 // Get a count of how many bytes are to be pushed on the stack.
2198 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2200 // Chain is the output chain of the last Load/Store or CopyToReg node.
2201 // ByValChain is the output chain of the last Memcpy node created for copying
2202 // byval arguments to the stack.
2203 SDValue Chain, CallSeqStart, ByValChain;
2204 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2205 Chain = CallSeqStart = DAG.getCALLSEQ_START(InChain, NextStackOffsetVal);
2206 ByValChain = InChain;
2208 // If this is the first call, create a stack frame object that points to
2209 // a location to which .cprestore saves $gp.
2210 if (IsO32 && IsPIC && !MipsFI->getGPFI())
2211 MipsFI->setGPFI(MFI->CreateFixedObject(4, 0, true));
2213 // Get the frame index of the stack frame object that points to the location
2214 // of dynamically allocated area on the stack.
2215 int DynAllocFI = MipsFI->getDynAllocFI();
2217 // Update size of the maximum argument space.
2218 // For O32, a minimum of four words (16 bytes) of argument space is
2221 NextStackOffset = std::max(NextStackOffset, (unsigned)16);
2223 unsigned MaxCallFrameSize = MipsFI->getMaxCallFrameSize();
2225 if (MaxCallFrameSize < NextStackOffset) {
2226 MipsFI->setMaxCallFrameSize(NextStackOffset);
2228 // Set the offsets relative to $sp of the $gp restore slot and dynamically
2229 // allocated stack space. These offsets must be aligned to a boundary
2230 // determined by the stack alignment of the ABI.
2231 unsigned StackAlignment = TFL->getStackAlignment();
2232 NextStackOffset = (NextStackOffset + StackAlignment - 1) /
2233 StackAlignment * StackAlignment;
2235 if (MipsFI->needGPSaveRestore())
2236 MFI->setObjectOffset(MipsFI->getGPFI(), NextStackOffset);
2238 MFI->setObjectOffset(DynAllocFI, NextStackOffset);
2241 // With EABI is it possible to have 16 args on registers.
2242 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
2243 SmallVector<SDValue, 8> MemOpChains;
2245 int FirstFI = -MFI->getNumFixedObjects() - 1, LastFI = 0;
2247 // Walk the register/memloc assignments, inserting copies/loads.
2248 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2249 SDValue Arg = OutVals[i];
2250 CCValAssign &VA = ArgLocs[i];
2251 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2252 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2255 if (Flags.isByVal()) {
2256 assert(Flags.getByValSize() &&
2257 "ByVal args of size 0 should have been ignored by front-end.");
2259 WriteByValArg(ByValChain, Chain, dl, RegsToPass, MemOpChains, LastFI,
2260 MFI, DAG, Arg, VA, Flags, getPointerTy(),
2261 Subtarget->isLittle());
2263 PassByValArg64(ByValChain, Chain, dl, RegsToPass, MemOpChains, LastFI,
2264 MFI, DAG, Arg, VA, Flags, getPointerTy(),
2265 Subtarget->isLittle());
2269 // Promote the value if needed.
2270 switch (VA.getLocInfo()) {
2271 default: llvm_unreachable("Unknown loc info!");
2272 case CCValAssign::Full:
2273 if (VA.isRegLoc()) {
2274 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2275 (ValVT == MVT::f64 && LocVT == MVT::i64))
2276 Arg = DAG.getNode(ISD::BITCAST, dl, LocVT, Arg);
2277 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2278 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2279 Arg, DAG.getConstant(0, MVT::i32));
2280 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2281 Arg, DAG.getConstant(1, MVT::i32));
2282 if (!Subtarget->isLittle())
2284 unsigned LocRegLo = VA.getLocReg();
2285 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2286 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2287 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2292 case CCValAssign::SExt:
2293 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, LocVT, Arg);
2295 case CCValAssign::ZExt:
2296 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, LocVT, Arg);
2298 case CCValAssign::AExt:
2299 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, LocVT, Arg);
2303 // Arguments that can be passed on register must be kept at
2304 // RegsToPass vector
2305 if (VA.isRegLoc()) {
2306 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2310 // Register can't get to this point...
2311 assert(VA.isMemLoc());
2313 // Create the frame index object for this incoming parameter
2314 LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2315 VA.getLocMemOffset(), true);
2316 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2318 // emit ISD::STORE whichs stores the
2319 // parameter value to a stack Location
2320 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2321 MachinePointerInfo(), false, false, 0));
2324 // Extend range of indices of frame objects for outgoing arguments that were
2325 // created during this function call. Skip this step if no such objects were
2328 MipsFI->extendOutArgFIRange(FirstFI, LastFI);
2330 // If a memcpy has been created to copy a byval arg to a stack, replace the
2331 // chain input of CallSeqStart with ByValChain.
2332 if (InChain != ByValChain)
2333 DAG.UpdateNodeOperands(CallSeqStart.getNode(), ByValChain,
2334 NextStackOffsetVal);
2336 // Transform all store nodes into one single node because all store
2337 // nodes are independent of each other.
2338 if (!MemOpChains.empty())
2339 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2340 &MemOpChains[0], MemOpChains.size());
2342 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2343 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2344 // node so that legalize doesn't hack it.
2345 unsigned char OpFlag;
2346 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2347 bool GlobalOrExternal = false;
2350 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2351 if (IsPICCall && G->getGlobal()->hasInternalLinkage()) {
2352 OpFlag = IsO32 ? MipsII::MO_GOT : MipsII::MO_GOT_PAGE;
2353 unsigned char LoFlag = IsO32 ? MipsII::MO_ABS_LO : MipsII::MO_GOT_OFST;
2354 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(), 0,
2356 CalleeLo = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(),
2359 OpFlag = IsPICCall ? MipsII::MO_GOT_CALL : MipsII::MO_NO_FLAG;
2360 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2361 getPointerTy(), 0, OpFlag);
2364 GlobalOrExternal = true;
2366 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2367 if (IsN64 || (!IsO32 && IsPIC))
2368 OpFlag = MipsII::MO_GOT_DISP;
2369 else if (!IsPIC) // !N64 && static
2370 OpFlag = MipsII::MO_NO_FLAG;
2372 OpFlag = MipsII::MO_GOT_CALL;
2373 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2375 GlobalOrExternal = true;
2380 // Create nodes that load address of callee and copy it to T9
2382 if (GlobalOrExternal) {
2383 // Load callee address
2384 Callee = DAG.getNode(MipsISD::Wrapper, dl, getPointerTy(), Callee);
2385 SDValue LoadValue = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
2386 Callee, MachinePointerInfo::getGOT(),
2387 false, false, false, 0);
2389 // Use GOT+LO if callee has internal linkage.
2390 if (CalleeLo.getNode()) {
2391 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, getPointerTy(), CalleeLo);
2392 Callee = DAG.getNode(ISD::ADD, dl, getPointerTy(), LoadValue, Lo);
2398 // T9 should contain the address of the callee function if
2399 // -reloction-model=pic or it is an indirect call.
2400 if (IsPICCall || !GlobalOrExternal) {
2402 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
2403 Chain = DAG.getCopyToReg(Chain, dl, T9Reg, Callee, SDValue(0, 0));
2404 InFlag = Chain.getValue(1);
2405 Callee = DAG.getRegister(T9Reg, getPointerTy());
2408 // Build a sequence of copy-to-reg nodes chained together with token
2409 // chain and flag operands which copy the outgoing args into registers.
2410 // The InFlag in necessary since all emitted instructions must be
2412 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2413 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2414 RegsToPass[i].second, InFlag);
2415 InFlag = Chain.getValue(1);
2418 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2419 // = Chain, Callee, Reg#1, Reg#2, ...
2421 // Returns a chain & a flag for retval copy to use.
2422 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2423 SmallVector<SDValue, 8> Ops;
2424 Ops.push_back(Chain);
2425 Ops.push_back(Callee);
2427 // Add argument registers to the end of the list so that they are
2428 // known live into the call.
2429 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2430 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2431 RegsToPass[i].second.getValueType()));
2433 if (InFlag.getNode())
2434 Ops.push_back(InFlag);
2436 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2437 InFlag = Chain.getValue(1);
2439 // Create the CALLSEQ_END node.
2440 Chain = DAG.getCALLSEQ_END(Chain,
2441 DAG.getIntPtrConstant(NextStackOffset, true),
2442 DAG.getIntPtrConstant(0, true), InFlag);
2443 InFlag = Chain.getValue(1);
2445 // Handle result values, copying them out of physregs into vregs that we
2447 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2448 Ins, dl, DAG, InVals);
2451 /// LowerCallResult - Lower the result values of a call into the
2452 /// appropriate copies out of appropriate physical registers.
2454 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2455 CallingConv::ID CallConv, bool isVarArg,
2456 const SmallVectorImpl<ISD::InputArg> &Ins,
2457 DebugLoc dl, SelectionDAG &DAG,
2458 SmallVectorImpl<SDValue> &InVals) const {
2459 // Assign locations to each value returned by this call.
2460 SmallVector<CCValAssign, 16> RVLocs;
2461 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2462 getTargetMachine(), RVLocs, *DAG.getContext());
2464 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2466 // Copy all of the result registers out of their specified physreg.
2467 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2468 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2469 RVLocs[i].getValVT(), InFlag).getValue(1);
2470 InFlag = Chain.getValue(2);
2471 InVals.push_back(Chain.getValue(0));
2477 //===----------------------------------------------------------------------===//
2478 // Formal Arguments Calling Convention Implementation
2479 //===----------------------------------------------------------------------===//
2480 static void ReadByValArg(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2481 std::vector<SDValue>& OutChains,
2482 SelectionDAG &DAG, unsigned NumWords, SDValue FIN,
2483 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags) {
2484 unsigned LocMem = VA.getLocMemOffset();
2485 unsigned FirstWord = LocMem / 4;
2487 // copy register A0 - A3 to frame object
2488 for (unsigned i = 0; i < NumWords; ++i) {
2489 unsigned CurWord = FirstWord + i;
2490 if (CurWord >= O32IntRegsSize)
2493 unsigned SrcReg = O32IntRegs[CurWord];
2494 unsigned Reg = AddLiveIn(MF, SrcReg, Mips::CPURegsRegisterClass);
2495 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIN,
2496 DAG.getConstant(i * 4, MVT::i32));
2497 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(Reg, MVT::i32),
2498 StorePtr, MachinePointerInfo(), false,
2500 OutChains.push_back(Store);
2504 // Create frame object on stack and copy registers used for byval passing to it.
2506 CopyMips64ByValRegs(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2507 std::vector<SDValue>& OutChains, SelectionDAG &DAG,
2508 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
2509 MachineFrameInfo *MFI, bool IsRegLoc,
2510 SmallVectorImpl<SDValue> &InVals, MipsFunctionInfo *MipsFI,
2512 const unsigned *Reg = Mips64IntRegs + 8;
2513 int FOOffset; // Frame object offset from virtual frame pointer.
2516 Reg = std::find(Mips64IntRegs, Mips64IntRegs + 8, VA.getLocReg());
2517 FOOffset = (Reg - Mips64IntRegs) * 8 - 8 * 8;
2520 FOOffset = VA.getLocMemOffset();
2522 // Create frame object.
2523 unsigned NumRegs = (Flags.getByValSize() + 7) / 8;
2524 unsigned LastFI = MFI->CreateFixedObject(NumRegs * 8, FOOffset, true);
2525 SDValue FIN = DAG.getFrameIndex(LastFI, PtrTy);
2526 InVals.push_back(FIN);
2528 // Copy arg registers.
2529 for (unsigned I = 0; (Reg != Mips64IntRegs + 8) && (I < NumRegs);
2531 unsigned VReg = AddLiveIn(MF, *Reg, Mips::CPU64RegsRegisterClass);
2532 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, PtrTy, FIN,
2533 DAG.getConstant(I * 8, PtrTy));
2534 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(VReg, MVT::i64),
2535 StorePtr, MachinePointerInfo(), false,
2537 OutChains.push_back(Store);
2543 /// LowerFormalArguments - transform physical registers into virtual registers
2544 /// and generate load operations for arguments places on the stack.
2546 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2547 CallingConv::ID CallConv,
2549 const SmallVectorImpl<ISD::InputArg> &Ins,
2550 DebugLoc dl, SelectionDAG &DAG,
2551 SmallVectorImpl<SDValue> &InVals)
2553 MachineFunction &MF = DAG.getMachineFunction();
2554 MachineFrameInfo *MFI = MF.getFrameInfo();
2555 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2557 MipsFI->setVarArgsFrameIndex(0);
2559 // Used with vargs to acumulate store chains.
2560 std::vector<SDValue> OutChains;
2562 // Assign locations to all of the incoming arguments.
2563 SmallVector<CCValAssign, 16> ArgLocs;
2564 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2565 getTargetMachine(), ArgLocs, *DAG.getContext());
2568 CCInfo.AnalyzeFormalArguments(Ins, CC_MipsO32);
2570 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips);
2572 int LastFI = 0;// MipsFI->LastInArgFI is 0 at the entry of this function.
2574 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2575 CCValAssign &VA = ArgLocs[i];
2576 EVT ValVT = VA.getValVT();
2577 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2578 bool IsRegLoc = VA.isRegLoc();
2580 if (Flags.isByVal()) {
2581 assert(Flags.getByValSize() &&
2582 "ByVal args of size 0 should have been ignored by front-end.");
2584 unsigned NumWords = (Flags.getByValSize() + 3) / 4;
2585 LastFI = MFI->CreateFixedObject(NumWords * 4, VA.getLocMemOffset(),
2587 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2588 InVals.push_back(FIN);
2589 ReadByValArg(MF, Chain, dl, OutChains, DAG, NumWords, FIN, VA, Flags);
2591 LastFI = CopyMips64ByValRegs(MF, Chain, dl, OutChains, DAG, VA, Flags,
2592 MFI, IsRegLoc, InVals, MipsFI,
2597 // Arguments stored on registers
2599 EVT RegVT = VA.getLocVT();
2600 unsigned ArgReg = VA.getLocReg();
2601 TargetRegisterClass *RC = 0;
2603 if (RegVT == MVT::i32)
2604 RC = Mips::CPURegsRegisterClass;
2605 else if (RegVT == MVT::i64)
2606 RC = Mips::CPU64RegsRegisterClass;
2607 else if (RegVT == MVT::f32)
2608 RC = Mips::FGR32RegisterClass;
2609 else if (RegVT == MVT::f64)
2610 RC = HasMips64 ? Mips::FGR64RegisterClass : Mips::AFGR64RegisterClass;
2612 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2614 // Transform the arguments stored on
2615 // physical registers into virtual ones
2616 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2617 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2619 // If this is an 8 or 16-bit value, it has been passed promoted
2620 // to 32 bits. Insert an assert[sz]ext to capture this, then
2621 // truncate to the right size.
2622 if (VA.getLocInfo() != CCValAssign::Full) {
2623 unsigned Opcode = 0;
2624 if (VA.getLocInfo() == CCValAssign::SExt)
2625 Opcode = ISD::AssertSext;
2626 else if (VA.getLocInfo() == CCValAssign::ZExt)
2627 Opcode = ISD::AssertZext;
2629 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
2630 DAG.getValueType(ValVT));
2631 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue);
2634 // Handle floating point arguments passed in integer registers.
2635 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2636 (RegVT == MVT::i64 && ValVT == MVT::f64))
2637 ArgValue = DAG.getNode(ISD::BITCAST, dl, ValVT, ArgValue);
2638 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2639 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
2640 getNextIntArgReg(ArgReg), RC);
2641 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
2642 if (!Subtarget->isLittle())
2643 std::swap(ArgValue, ArgValue2);
2644 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
2645 ArgValue, ArgValue2);
2648 InVals.push_back(ArgValue);
2649 } else { // VA.isRegLoc()
2652 assert(VA.isMemLoc());
2654 // The stack pointer offset is relative to the caller stack frame.
2655 LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2656 VA.getLocMemOffset(), true);
2658 // Create load nodes to retrieve arguments from the stack
2659 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2660 InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN,
2661 MachinePointerInfo::getFixedStack(LastFI),
2662 false, false, false, 0));
2666 // The mips ABIs for returning structs by value requires that we copy
2667 // the sret argument into $v0 for the return. Save the argument into
2668 // a virtual register so that we can access it from the return points.
2669 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2670 unsigned Reg = MipsFI->getSRetReturnReg();
2672 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i32));
2673 MipsFI->setSRetReturnReg(Reg);
2675 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2676 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2680 unsigned NumOfRegs = IsO32 ? 4 : 8;
2681 const unsigned *ArgRegs = IsO32 ? O32IntRegs : Mips64IntRegs;
2682 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumOfRegs);
2683 int FirstRegSlotOffset = IsO32 ? 0 : -64 ; // offset of $a0's slot.
2684 TargetRegisterClass *RC
2685 = IsO32 ? Mips::CPURegsRegisterClass : Mips::CPU64RegsRegisterClass;
2686 unsigned RegSize = RC->getSize();
2687 int RegSlotOffset = FirstRegSlotOffset + Idx * RegSize;
2689 // Offset of the first variable argument from stack pointer.
2690 int FirstVaArgOffset;
2692 if (IsO32 || (Idx == NumOfRegs)) {
2694 (CCInfo.getNextStackOffset() + RegSize - 1) / RegSize * RegSize;
2696 FirstVaArgOffset = RegSlotOffset;
2698 // Record the frame index of the first variable argument
2699 // which is a value necessary to VASTART.
2700 LastFI = MFI->CreateFixedObject(RegSize, FirstVaArgOffset, true);
2701 MipsFI->setVarArgsFrameIndex(LastFI);
2703 // Copy the integer registers that have not been used for argument passing
2704 // to the argument register save area. For O32, the save area is allocated
2705 // in the caller's stack frame, while for N32/64, it is allocated in the
2706 // callee's stack frame.
2707 for (int StackOffset = RegSlotOffset;
2708 Idx < NumOfRegs; ++Idx, StackOffset += RegSize) {
2709 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgRegs[Idx], RC);
2710 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg,
2711 MVT::getIntegerVT(RegSize * 8));
2712 LastFI = MFI->CreateFixedObject(RegSize, StackOffset, true);
2713 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2714 OutChains.push_back(DAG.getStore(Chain, dl, ArgValue, PtrOff,
2715 MachinePointerInfo(), false, false, 0));
2719 MipsFI->setLastInArgFI(LastFI);
2721 // All stores are grouped in one node to allow the matching between
2722 // the size of Ins and InVals. This only happens when on varg functions
2723 if (!OutChains.empty()) {
2724 OutChains.push_back(Chain);
2725 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2726 &OutChains[0], OutChains.size());
2732 //===----------------------------------------------------------------------===//
2733 // Return Value Calling Convention Implementation
2734 //===----------------------------------------------------------------------===//
2737 MipsTargetLowering::LowerReturn(SDValue Chain,
2738 CallingConv::ID CallConv, bool isVarArg,
2739 const SmallVectorImpl<ISD::OutputArg> &Outs,
2740 const SmallVectorImpl<SDValue> &OutVals,
2741 DebugLoc dl, SelectionDAG &DAG) const {
2743 // CCValAssign - represent the assignment of
2744 // the return value to a location
2745 SmallVector<CCValAssign, 16> RVLocs;
2747 // CCState - Info about the registers and stack slot.
2748 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2749 getTargetMachine(), RVLocs, *DAG.getContext());
2751 // Analize return values.
2752 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
2754 // If this is the first return lowered for this function, add
2755 // the regs to the liveout set for the function.
2756 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2757 for (unsigned i = 0; i != RVLocs.size(); ++i)
2758 if (RVLocs[i].isRegLoc())
2759 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2764 // Copy the result values into the output registers.
2765 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2766 CCValAssign &VA = RVLocs[i];
2767 assert(VA.isRegLoc() && "Can only return in registers!");
2769 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
2771 // guarantee that all emitted copies are
2772 // stuck together, avoiding something bad
2773 Flag = Chain.getValue(1);
2776 // The mips ABIs for returning structs by value requires that we copy
2777 // the sret argument into $v0 for the return. We saved the argument into
2778 // a virtual register in the entry block, so now we copy the value out
2780 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2781 MachineFunction &MF = DAG.getMachineFunction();
2782 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2783 unsigned Reg = MipsFI->getSRetReturnReg();
2786 llvm_unreachable("sret virtual register not created in the entry block");
2787 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2789 Chain = DAG.getCopyToReg(Chain, dl, Mips::V0, Val, Flag);
2790 Flag = Chain.getValue(1);
2793 // Return on Mips is always a "jr $ra"
2795 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2796 Chain, DAG.getRegister(Mips::RA, MVT::i32), Flag);
2798 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2799 Chain, DAG.getRegister(Mips::RA, MVT::i32));
2802 //===----------------------------------------------------------------------===//
2803 // Mips Inline Assembly Support
2804 //===----------------------------------------------------------------------===//
2806 /// getConstraintType - Given a constraint letter, return the type of
2807 /// constraint it is for this target.
2808 MipsTargetLowering::ConstraintType MipsTargetLowering::
2809 getConstraintType(const std::string &Constraint) const
2811 // Mips specific constrainy
2812 // GCC config/mips/constraints.md
2814 // 'd' : An address register. Equivalent to r
2815 // unless generating MIPS16 code.
2816 // 'y' : Equivalent to r; retained for
2817 // backwards compatibility.
2818 // 'f' : Floating Point registers.
2819 if (Constraint.size() == 1) {
2820 switch (Constraint[0]) {
2825 return C_RegisterClass;
2829 return TargetLowering::getConstraintType(Constraint);
2832 /// Examine constraint type and operand type and determine a weight value.
2833 /// This object must already have been set up with the operand type
2834 /// and the current alternative constraint selected.
2835 TargetLowering::ConstraintWeight
2836 MipsTargetLowering::getSingleConstraintMatchWeight(
2837 AsmOperandInfo &info, const char *constraint) const {
2838 ConstraintWeight weight = CW_Invalid;
2839 Value *CallOperandVal = info.CallOperandVal;
2840 // If we don't have a value, we can't do a match,
2841 // but allow it at the lowest weight.
2842 if (CallOperandVal == NULL)
2844 Type *type = CallOperandVal->getType();
2845 // Look at the constraint type.
2846 switch (*constraint) {
2848 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2852 if (type->isIntegerTy())
2853 weight = CW_Register;
2856 if (type->isFloatTy())
2857 weight = CW_Register;
2863 /// Given a register class constraint, like 'r', if this corresponds directly
2864 /// to an LLVM register class, return a register of 0 and the register class
2866 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2867 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
2869 if (Constraint.size() == 1) {
2870 switch (Constraint[0]) {
2871 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2872 case 'y': // Same as 'r'. Exists for compatibility.
2874 return std::make_pair(0U, Mips::CPURegsRegisterClass);
2877 return std::make_pair(0U, Mips::FGR32RegisterClass);
2879 if ((!Subtarget->isSingleFloat()) && (!Subtarget->isFP64bit()))
2880 return std::make_pair(0U, Mips::AFGR64RegisterClass);
2884 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2888 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2889 // The Mips target isn't yet aware of offsets.
2893 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2894 if (VT != MVT::f32 && VT != MVT::f64)
2896 if (Imm.isNegZero())
2898 return Imm.isZero();