1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that Mips uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "mips-lower"
15 #include "MipsISelLowering.h"
16 #include "InstPrinter/MipsInstPrinter.h"
17 #include "MCTargetDesc/MipsBaseInfo.h"
18 #include "MipsMachineFunction.h"
19 #include "MipsSubtarget.h"
20 #include "MipsTargetMachine.h"
21 #include "MipsTargetObjectFile.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAGISel.h"
29 #include "llvm/CodeGen/ValueTypes.h"
30 #include "llvm/IR/CallingConv.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
41 STATISTIC(NumTailCalls, "Number of tail calls");
44 LargeGOT("mxgot", cl::Hidden,
45 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
47 static const uint16_t O32IntRegs[4] = {
48 Mips::A0, Mips::A1, Mips::A2, Mips::A3
51 static const uint16_t Mips64IntRegs[8] = {
52 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
53 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64
56 static const uint16_t Mips64DPRegs[8] = {
57 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
58 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
61 // If I is a shifted mask, set the size (Size) and the first bit of the
62 // mask (Pos), and return true.
63 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
64 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
65 if (!isShiftedMask_64(I))
68 Size = CountPopulation_64(I);
69 Pos = CountTrailingZeros_64(I);
73 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
74 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
75 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
78 static SDValue getTargetNode(SDValue Op, SelectionDAG &DAG, unsigned Flag) {
79 EVT Ty = Op.getValueType();
81 if (GlobalAddressSDNode *N = dyn_cast<GlobalAddressSDNode>(Op))
82 return DAG.getTargetGlobalAddress(N->getGlobal(), Op.getDebugLoc(), Ty, 0,
84 if (ExternalSymbolSDNode *N = dyn_cast<ExternalSymbolSDNode>(Op))
85 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
86 if (BlockAddressSDNode *N = dyn_cast<BlockAddressSDNode>(Op))
87 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
88 if (JumpTableSDNode *N = dyn_cast<JumpTableSDNode>(Op))
89 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
90 if (ConstantPoolSDNode *N = dyn_cast<ConstantPoolSDNode>(Op))
91 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
92 N->getOffset(), Flag);
94 llvm_unreachable("Unexpected node type.");
98 static SDValue getAddrNonPIC(SDValue Op, SelectionDAG &DAG) {
99 DebugLoc DL = Op.getDebugLoc();
100 EVT Ty = Op.getValueType();
101 SDValue Hi = getTargetNode(Op, DAG, MipsII::MO_ABS_HI);
102 SDValue Lo = getTargetNode(Op, DAG, MipsII::MO_ABS_LO);
103 return DAG.getNode(ISD::ADD, DL, Ty,
104 DAG.getNode(MipsISD::Hi, DL, Ty, Hi),
105 DAG.getNode(MipsISD::Lo, DL, Ty, Lo));
108 SDValue MipsTargetLowering::getAddrLocal(SDValue Op, SelectionDAG &DAG,
109 bool HasMips64) const {
110 DebugLoc DL = Op.getDebugLoc();
111 EVT Ty = Op.getValueType();
112 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
113 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
114 getTargetNode(Op, DAG, GOTFlag));
115 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT,
116 MachinePointerInfo::getGOT(), false, false, false,
118 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
119 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty, getTargetNode(Op, DAG, LoFlag));
120 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo);
123 SDValue MipsTargetLowering::getAddrGlobal(SDValue Op, SelectionDAG &DAG,
124 unsigned Flag) const {
125 DebugLoc DL = Op.getDebugLoc();
126 EVT Ty = Op.getValueType();
127 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty),
128 getTargetNode(Op, DAG, Flag));
129 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt,
130 MachinePointerInfo::getGOT(), false, false, false, 0);
133 SDValue MipsTargetLowering::getAddrGlobalLargeGOT(SDValue Op, SelectionDAG &DAG,
135 unsigned LoFlag) const {
136 DebugLoc DL = Op.getDebugLoc();
137 EVT Ty = Op.getValueType();
138 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty, getTargetNode(Op, DAG, HiFlag));
139 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, getGlobalReg(DAG, Ty));
140 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi,
141 getTargetNode(Op, DAG, LoFlag));
142 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper,
143 MachinePointerInfo::getGOT(), false, false, false, 0);
146 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
148 case MipsISD::JmpLink: return "MipsISD::JmpLink";
149 case MipsISD::TailCall: return "MipsISD::TailCall";
150 case MipsISD::Hi: return "MipsISD::Hi";
151 case MipsISD::Lo: return "MipsISD::Lo";
152 case MipsISD::GPRel: return "MipsISD::GPRel";
153 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
154 case MipsISD::Ret: return "MipsISD::Ret";
155 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
156 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
157 case MipsISD::FPCmp: return "MipsISD::FPCmp";
158 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
159 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
160 case MipsISD::FPRound: return "MipsISD::FPRound";
161 case MipsISD::ExtractLOHI: return "MipsISD::ExtractLOHI";
162 case MipsISD::InsertLOHI: return "MipsISD::InsertLOHI";
163 case MipsISD::Mult: return "MipsISD::Mult";
164 case MipsISD::Multu: return "MipsISD::Multu";
165 case MipsISD::MAdd: return "MipsISD::MAdd";
166 case MipsISD::MAddu: return "MipsISD::MAddu";
167 case MipsISD::MSub: return "MipsISD::MSub";
168 case MipsISD::MSubu: return "MipsISD::MSubu";
169 case MipsISD::DivRem: return "MipsISD::DivRem";
170 case MipsISD::DivRemU: return "MipsISD::DivRemU";
171 case MipsISD::DivRem16: return "MipsISD::DivRem16";
172 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
173 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
174 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
175 case MipsISD::Wrapper: return "MipsISD::Wrapper";
176 case MipsISD::Sync: return "MipsISD::Sync";
177 case MipsISD::Ext: return "MipsISD::Ext";
178 case MipsISD::Ins: return "MipsISD::Ins";
179 case MipsISD::LWL: return "MipsISD::LWL";
180 case MipsISD::LWR: return "MipsISD::LWR";
181 case MipsISD::SWL: return "MipsISD::SWL";
182 case MipsISD::SWR: return "MipsISD::SWR";
183 case MipsISD::LDL: return "MipsISD::LDL";
184 case MipsISD::LDR: return "MipsISD::LDR";
185 case MipsISD::SDL: return "MipsISD::SDL";
186 case MipsISD::SDR: return "MipsISD::SDR";
187 case MipsISD::EXTP: return "MipsISD::EXTP";
188 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
189 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
190 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
191 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
192 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
193 case MipsISD::SHILO: return "MipsISD::SHILO";
194 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
195 case MipsISD::MULT: return "MipsISD::MULT";
196 case MipsISD::MULTU: return "MipsISD::MULTU";
197 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
198 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
199 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
200 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
201 default: return NULL;
206 MipsTargetLowering(MipsTargetMachine &TM)
207 : TargetLowering(TM, new MipsTargetObjectFile()),
208 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
209 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
210 IsO32(Subtarget->isABI_O32()) {
211 // Mips does not have i1 type, so use i32 for
212 // setcc operations results (slt, sgt, ...).
213 setBooleanContents(ZeroOrOneBooleanContent);
214 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
216 // Load extented operations for i1 types must be promoted
217 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
218 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
219 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
221 // MIPS doesn't have extending float->double load/store
222 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
223 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
225 // Used by legalize types to correctly generate the setcc result.
226 // Without this, every float setcc comes with a AND/OR with the result,
227 // we don't want this, since the fpcmp result goes to a flag register,
228 // which is used implicitly by brcond and select operations.
229 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
231 // Mips Custom Operations
232 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
233 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
234 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
235 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
236 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
237 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
238 setOperationAction(ISD::SELECT, MVT::f32, Custom);
239 setOperationAction(ISD::SELECT, MVT::f64, Custom);
240 setOperationAction(ISD::SELECT, MVT::i32, Custom);
241 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
242 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
243 setOperationAction(ISD::SETCC, MVT::f32, Custom);
244 setOperationAction(ISD::SETCC, MVT::f64, Custom);
245 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
246 setOperationAction(ISD::VASTART, MVT::Other, Custom);
247 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
248 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
250 if (!TM.Options.NoNaNsFPMath) {
251 setOperationAction(ISD::FABS, MVT::f32, Custom);
252 setOperationAction(ISD::FABS, MVT::f64, Custom);
256 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
257 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
258 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
259 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
260 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
261 setOperationAction(ISD::SELECT, MVT::i64, Custom);
262 setOperationAction(ISD::LOAD, MVT::i64, Custom);
263 setOperationAction(ISD::STORE, MVT::i64, Custom);
267 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
268 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
269 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
272 setOperationAction(ISD::ADD, MVT::i32, Custom);
274 setOperationAction(ISD::ADD, MVT::i64, Custom);
276 setOperationAction(ISD::SDIV, MVT::i32, Expand);
277 setOperationAction(ISD::SREM, MVT::i32, Expand);
278 setOperationAction(ISD::UDIV, MVT::i32, Expand);
279 setOperationAction(ISD::UREM, MVT::i32, Expand);
280 setOperationAction(ISD::SDIV, MVT::i64, Expand);
281 setOperationAction(ISD::SREM, MVT::i64, Expand);
282 setOperationAction(ISD::UDIV, MVT::i64, Expand);
283 setOperationAction(ISD::UREM, MVT::i64, Expand);
285 // Operations not directly supported by Mips.
286 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
287 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
288 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
289 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
290 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
291 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
292 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
293 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
294 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
295 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
296 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
297 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
298 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
299 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
300 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
301 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
302 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
303 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
304 setOperationAction(ISD::ROTL, MVT::i32, Expand);
305 setOperationAction(ISD::ROTL, MVT::i64, Expand);
306 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
307 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
309 if (!Subtarget->hasMips32r2())
310 setOperationAction(ISD::ROTR, MVT::i32, Expand);
312 if (!Subtarget->hasMips64r2())
313 setOperationAction(ISD::ROTR, MVT::i64, Expand);
315 setOperationAction(ISD::FSIN, MVT::f32, Expand);
316 setOperationAction(ISD::FSIN, MVT::f64, Expand);
317 setOperationAction(ISD::FCOS, MVT::f32, Expand);
318 setOperationAction(ISD::FCOS, MVT::f64, Expand);
319 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
320 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
321 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
322 setOperationAction(ISD::FPOW, MVT::f32, Expand);
323 setOperationAction(ISD::FPOW, MVT::f64, Expand);
324 setOperationAction(ISD::FLOG, MVT::f32, Expand);
325 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
326 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
327 setOperationAction(ISD::FEXP, MVT::f32, Expand);
328 setOperationAction(ISD::FMA, MVT::f32, Expand);
329 setOperationAction(ISD::FMA, MVT::f64, Expand);
330 setOperationAction(ISD::FREM, MVT::f32, Expand);
331 setOperationAction(ISD::FREM, MVT::f64, Expand);
333 if (!TM.Options.NoNaNsFPMath) {
334 setOperationAction(ISD::FNEG, MVT::f32, Expand);
335 setOperationAction(ISD::FNEG, MVT::f64, Expand);
338 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
339 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
340 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
341 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
343 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
345 setOperationAction(ISD::VAARG, MVT::Other, Expand);
346 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
347 setOperationAction(ISD::VAEND, MVT::Other, Expand);
349 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
350 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
352 // Use the default for now
353 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
354 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
356 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
357 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
358 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
359 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
361 setInsertFencesForAtomic(true);
363 if (!Subtarget->hasSEInReg()) {
364 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
365 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
368 if (!Subtarget->hasBitCount()) {
369 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
370 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
373 if (!Subtarget->hasSwap()) {
374 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
375 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
379 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
380 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
381 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
382 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
385 setTargetDAGCombine(ISD::ADDE);
386 setTargetDAGCombine(ISD::SUBE);
387 setTargetDAGCombine(ISD::SDIVREM);
388 setTargetDAGCombine(ISD::UDIVREM);
389 setTargetDAGCombine(ISD::SELECT);
390 setTargetDAGCombine(ISD::AND);
391 setTargetDAGCombine(ISD::OR);
392 setTargetDAGCombine(ISD::ADD);
394 setMinFunctionAlignment(HasMips64 ? 3 : 2);
396 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP);
398 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0);
399 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1);
401 MaxStoresPerMemcpy = 16;
404 const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) {
405 if (TM.getSubtargetImpl()->inMips16Mode())
406 return llvm::createMips16TargetLowering(TM);
408 return llvm::createMipsSETargetLowering(TM);
411 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
414 return VT.changeVectorElementTypeToInteger();
418 // Transforms a subgraph in CurDAG if the following pattern is found:
419 // (addc multLo, Lo0), (adde multHi, Hi0),
421 // multHi/Lo: product of multiplication
422 // Lo0: initial value of Lo register
423 // Hi0: initial value of Hi register
424 // Return true if pattern matching was successful.
425 static bool selectMADD(SDNode *ADDENode, SelectionDAG *CurDAG) {
426 // ADDENode's second operand must be a flag output of an ADDC node in order
427 // for the matching to be successful.
428 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
430 if (ADDCNode->getOpcode() != ISD::ADDC)
433 SDValue MultHi = ADDENode->getOperand(0);
434 SDValue MultLo = ADDCNode->getOperand(0);
435 SDNode *MultNode = MultHi.getNode();
436 unsigned MultOpc = MultHi.getOpcode();
438 // MultHi and MultLo must be generated by the same node,
439 if (MultLo.getNode() != MultNode)
442 // and it must be a multiplication.
443 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
446 // MultLo amd MultHi must be the first and second output of MultNode
448 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
451 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
452 // of the values of MultNode, in which case MultNode will be removed in later
454 // If there exist users other than ADDENode or ADDCNode, this function returns
455 // here, which will result in MultNode being mapped to a single MULT
456 // instruction node rather than a pair of MULT and MADD instructions being
458 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
461 SDValue Chain = CurDAG->getEntryNode();
462 DebugLoc DL = ADDENode->getDebugLoc();
464 // Initialize accumulator.
465 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
466 ADDCNode->getOperand(1),
467 ADDENode->getOperand(1));
469 // create MipsMAdd(u) node
470 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
472 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Untyped,
473 MultNode->getOperand(0),// Factor 0
474 MultNode->getOperand(1),// Factor 1
477 // replace uses of adde and addc here
478 if (!SDValue(ADDCNode, 0).use_empty()) {
479 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
480 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
482 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), LoOut);
484 if (!SDValue(ADDENode, 0).use_empty()) {
485 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
486 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MAdd,
488 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), HiOut);
495 // Transforms a subgraph in CurDAG if the following pattern is found:
496 // (addc Lo0, multLo), (sube Hi0, multHi),
498 // multHi/Lo: product of multiplication
499 // Lo0: initial value of Lo register
500 // Hi0: initial value of Hi register
501 // Return true if pattern matching was successful.
502 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
503 // SUBENode's second operand must be a flag output of an SUBC node in order
504 // for the matching to be successful.
505 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
507 if (SUBCNode->getOpcode() != ISD::SUBC)
510 SDValue MultHi = SUBENode->getOperand(1);
511 SDValue MultLo = SUBCNode->getOperand(1);
512 SDNode *MultNode = MultHi.getNode();
513 unsigned MultOpc = MultHi.getOpcode();
515 // MultHi and MultLo must be generated by the same node,
516 if (MultLo.getNode() != MultNode)
519 // and it must be a multiplication.
520 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
523 // MultLo amd MultHi must be the first and second output of MultNode
525 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
528 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
529 // of the values of MultNode, in which case MultNode will be removed in later
531 // If there exist users other than SUBENode or SUBCNode, this function returns
532 // here, which will result in MultNode being mapped to a single MULT
533 // instruction node rather than a pair of MULT and MSUB instructions being
535 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
538 SDValue Chain = CurDAG->getEntryNode();
539 DebugLoc DL = SUBENode->getDebugLoc();
541 // Initialize accumulator.
542 SDValue ACCIn = CurDAG->getNode(MipsISD::InsertLOHI, DL, MVT::Untyped,
543 SUBCNode->getOperand(0),
544 SUBENode->getOperand(0));
546 // create MipsSub(u) node
547 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
549 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
550 MultNode->getOperand(0),// Factor 0
551 MultNode->getOperand(1),// Factor 1
554 // replace uses of sube and subc here
555 if (!SDValue(SUBCNode, 0).use_empty()) {
556 SDValue LoIdx = CurDAG->getConstant(Mips::sub_lo, MVT::i32);
557 SDValue LoOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
559 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), LoOut);
561 if (!SDValue(SUBENode, 0).use_empty()) {
562 SDValue HiIdx = CurDAG->getConstant(Mips::sub_hi, MVT::i32);
563 SDValue HiOut = CurDAG->getNode(MipsISD::ExtractLOHI, DL, MVT::i32, MSub,
565 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), HiOut);
571 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
572 TargetLowering::DAGCombinerInfo &DCI,
573 const MipsSubtarget *Subtarget) {
574 if (DCI.isBeforeLegalize())
577 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
579 return SDValue(N, 0);
584 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
585 TargetLowering::DAGCombinerInfo &DCI,
586 const MipsSubtarget *Subtarget) {
587 if (DCI.isBeforeLegalize())
590 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
592 return SDValue(N, 0);
597 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
598 TargetLowering::DAGCombinerInfo &DCI,
599 const MipsSubtarget *Subtarget) {
600 if (DCI.isBeforeLegalizeOps())
603 EVT Ty = N->getValueType(0);
604 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
605 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
606 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
608 DebugLoc DL = N->getDebugLoc();
610 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
611 N->getOperand(0), N->getOperand(1));
612 SDValue InChain = DAG.getEntryNode();
613 SDValue InGlue = DivRem;
616 if (N->hasAnyUseOfValue(0)) {
617 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
619 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
620 InChain = CopyFromLo.getValue(1);
621 InGlue = CopyFromLo.getValue(2);
625 if (N->hasAnyUseOfValue(1)) {
626 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
628 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
634 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
636 default: llvm_unreachable("Unknown fp condition code!");
638 case ISD::SETOEQ: return Mips::FCOND_OEQ;
639 case ISD::SETUNE: return Mips::FCOND_UNE;
641 case ISD::SETOLT: return Mips::FCOND_OLT;
643 case ISD::SETOGT: return Mips::FCOND_OGT;
645 case ISD::SETOLE: return Mips::FCOND_OLE;
647 case ISD::SETOGE: return Mips::FCOND_OGE;
648 case ISD::SETULT: return Mips::FCOND_ULT;
649 case ISD::SETULE: return Mips::FCOND_ULE;
650 case ISD::SETUGT: return Mips::FCOND_UGT;
651 case ISD::SETUGE: return Mips::FCOND_UGE;
652 case ISD::SETUO: return Mips::FCOND_UN;
653 case ISD::SETO: return Mips::FCOND_OR;
655 case ISD::SETONE: return Mips::FCOND_ONE;
656 case ISD::SETUEQ: return Mips::FCOND_UEQ;
661 /// This function returns true if the floating point conditional branches and
662 /// conditional moves which use condition code CC should be inverted.
663 static bool invertFPCondCodeUser(Mips::CondCode CC) {
664 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
667 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
668 "Illegal Condition Code");
673 // Creates and returns an FPCmp node from a setcc node.
674 // Returns Op if setcc is not a floating point comparison.
675 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
676 // must be a SETCC node
677 if (Op.getOpcode() != ISD::SETCC)
680 SDValue LHS = Op.getOperand(0);
682 if (!LHS.getValueType().isFloatingPoint())
685 SDValue RHS = Op.getOperand(1);
686 DebugLoc DL = Op.getDebugLoc();
688 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
689 // node if necessary.
690 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
692 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
693 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
696 // Creates and returns a CMovFPT/F node.
697 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
698 SDValue False, DebugLoc DL) {
699 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
700 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
702 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
703 True.getValueType(), True, False, Cond);
706 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
707 TargetLowering::DAGCombinerInfo &DCI,
708 const MipsSubtarget *Subtarget) {
709 if (DCI.isBeforeLegalizeOps())
712 SDValue SetCC = N->getOperand(0);
714 if ((SetCC.getOpcode() != ISD::SETCC) ||
715 !SetCC.getOperand(0).getValueType().isInteger())
718 SDValue False = N->getOperand(2);
719 EVT FalseTy = False.getValueType();
721 if (!FalseTy.isInteger())
724 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
726 if (!CN || CN->getZExtValue())
729 const DebugLoc DL = N->getDebugLoc();
730 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
731 SDValue True = N->getOperand(1);
733 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
734 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
736 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
739 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
740 TargetLowering::DAGCombinerInfo &DCI,
741 const MipsSubtarget *Subtarget) {
742 // Pattern match EXT.
743 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
744 // => ext $dst, $src, size, pos
745 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
748 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
749 unsigned ShiftRightOpc = ShiftRight.getOpcode();
751 // Op's first operand must be a shift right.
752 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
755 // The second operand of the shift must be an immediate.
757 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
760 uint64_t Pos = CN->getZExtValue();
761 uint64_t SMPos, SMSize;
763 // Op's second operand must be a shifted mask.
764 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
765 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
768 // Return if the shifted mask does not start at bit 0 or the sum of its size
769 // and Pos exceeds the word's size.
770 EVT ValTy = N->getValueType(0);
771 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
774 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
775 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
776 DAG.getConstant(SMSize, MVT::i32));
779 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
780 TargetLowering::DAGCombinerInfo &DCI,
781 const MipsSubtarget *Subtarget) {
782 // Pattern match INS.
783 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
784 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
785 // => ins $dst, $src, size, pos, $src1
786 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
789 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
790 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
793 // See if Op's first operand matches (and $src1 , mask0).
794 if (And0.getOpcode() != ISD::AND)
797 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
798 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
801 // See if Op's second operand matches (and (shl $src, pos), mask1).
802 if (And1.getOpcode() != ISD::AND)
805 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
806 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
809 // The shift masks must have the same position and size.
810 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
813 SDValue Shl = And1.getOperand(0);
814 if (Shl.getOpcode() != ISD::SHL)
817 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
820 unsigned Shamt = CN->getZExtValue();
822 // Return if the shift amount and the first bit position of mask are not the
824 EVT ValTy = N->getValueType(0);
825 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
828 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
829 DAG.getConstant(SMPos0, MVT::i32),
830 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
833 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
834 TargetLowering::DAGCombinerInfo &DCI,
835 const MipsSubtarget *Subtarget) {
836 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
838 if (DCI.isBeforeLegalizeOps())
841 SDValue Add = N->getOperand(1);
843 if (Add.getOpcode() != ISD::ADD)
846 SDValue Lo = Add.getOperand(1);
848 if ((Lo.getOpcode() != MipsISD::Lo) ||
849 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
852 EVT ValTy = N->getValueType(0);
853 DebugLoc DL = N->getDebugLoc();
855 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
857 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
860 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
862 SelectionDAG &DAG = DCI.DAG;
863 unsigned Opc = N->getOpcode();
868 return performADDECombine(N, DAG, DCI, Subtarget);
870 return performSUBECombine(N, DAG, DCI, Subtarget);
873 return performDivRemCombine(N, DAG, DCI, Subtarget);
875 return performSELECTCombine(N, DAG, DCI, Subtarget);
877 return performANDCombine(N, DAG, DCI, Subtarget);
879 return performORCombine(N, DAG, DCI, Subtarget);
881 return performADDCombine(N, DAG, DCI, Subtarget);
888 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
889 SmallVectorImpl<SDValue> &Results,
890 SelectionDAG &DAG) const {
891 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
893 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
894 Results.push_back(Res.getValue(I));
898 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
899 SmallVectorImpl<SDValue> &Results,
900 SelectionDAG &DAG) const {
901 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
903 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
904 Results.push_back(Res.getValue(I));
907 SDValue MipsTargetLowering::
908 LowerOperation(SDValue Op, SelectionDAG &DAG) const
910 switch (Op.getOpcode())
912 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
913 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
914 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
915 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
916 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
917 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
918 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
919 case ISD::SELECT: return lowerSELECT(Op, DAG);
920 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
921 case ISD::SETCC: return lowerSETCC(Op, DAG);
922 case ISD::VASTART: return lowerVASTART(Op, DAG);
923 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
924 case ISD::FABS: return lowerFABS(Op, DAG);
925 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
926 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
927 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
928 case ISD::MEMBARRIER: return lowerMEMBARRIER(Op, DAG);
929 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
930 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
931 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
932 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
933 case ISD::LOAD: return lowerLOAD(Op, DAG);
934 case ISD::STORE: return lowerSTORE(Op, DAG);
935 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
936 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
937 case ISD::ADD: return lowerADD(Op, DAG);
942 //===----------------------------------------------------------------------===//
943 // Lower helper functions
944 //===----------------------------------------------------------------------===//
946 // addLiveIn - This helper function adds the specified physical register to the
947 // MachineFunction as a live in value. It also creates a corresponding
948 // virtual register for it.
950 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
952 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
953 MF.getRegInfo().addLiveIn(PReg, VReg);
958 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
959 MachineBasicBlock *BB) const {
960 switch (MI->getOpcode()) {
962 llvm_unreachable("Unexpected instr type to insert");
963 case Mips::ATOMIC_LOAD_ADD_I8:
964 case Mips::ATOMIC_LOAD_ADD_I8_P8:
965 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
966 case Mips::ATOMIC_LOAD_ADD_I16:
967 case Mips::ATOMIC_LOAD_ADD_I16_P8:
968 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
969 case Mips::ATOMIC_LOAD_ADD_I32:
970 case Mips::ATOMIC_LOAD_ADD_I32_P8:
971 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
972 case Mips::ATOMIC_LOAD_ADD_I64:
973 case Mips::ATOMIC_LOAD_ADD_I64_P8:
974 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
976 case Mips::ATOMIC_LOAD_AND_I8:
977 case Mips::ATOMIC_LOAD_AND_I8_P8:
978 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
979 case Mips::ATOMIC_LOAD_AND_I16:
980 case Mips::ATOMIC_LOAD_AND_I16_P8:
981 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
982 case Mips::ATOMIC_LOAD_AND_I32:
983 case Mips::ATOMIC_LOAD_AND_I32_P8:
984 return emitAtomicBinary(MI, BB, 4, Mips::AND);
985 case Mips::ATOMIC_LOAD_AND_I64:
986 case Mips::ATOMIC_LOAD_AND_I64_P8:
987 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
989 case Mips::ATOMIC_LOAD_OR_I8:
990 case Mips::ATOMIC_LOAD_OR_I8_P8:
991 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
992 case Mips::ATOMIC_LOAD_OR_I16:
993 case Mips::ATOMIC_LOAD_OR_I16_P8:
994 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
995 case Mips::ATOMIC_LOAD_OR_I32:
996 case Mips::ATOMIC_LOAD_OR_I32_P8:
997 return emitAtomicBinary(MI, BB, 4, Mips::OR);
998 case Mips::ATOMIC_LOAD_OR_I64:
999 case Mips::ATOMIC_LOAD_OR_I64_P8:
1000 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
1002 case Mips::ATOMIC_LOAD_XOR_I8:
1003 case Mips::ATOMIC_LOAD_XOR_I8_P8:
1004 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
1005 case Mips::ATOMIC_LOAD_XOR_I16:
1006 case Mips::ATOMIC_LOAD_XOR_I16_P8:
1007 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
1008 case Mips::ATOMIC_LOAD_XOR_I32:
1009 case Mips::ATOMIC_LOAD_XOR_I32_P8:
1010 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
1011 case Mips::ATOMIC_LOAD_XOR_I64:
1012 case Mips::ATOMIC_LOAD_XOR_I64_P8:
1013 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
1015 case Mips::ATOMIC_LOAD_NAND_I8:
1016 case Mips::ATOMIC_LOAD_NAND_I8_P8:
1017 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
1018 case Mips::ATOMIC_LOAD_NAND_I16:
1019 case Mips::ATOMIC_LOAD_NAND_I16_P8:
1020 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
1021 case Mips::ATOMIC_LOAD_NAND_I32:
1022 case Mips::ATOMIC_LOAD_NAND_I32_P8:
1023 return emitAtomicBinary(MI, BB, 4, 0, true);
1024 case Mips::ATOMIC_LOAD_NAND_I64:
1025 case Mips::ATOMIC_LOAD_NAND_I64_P8:
1026 return emitAtomicBinary(MI, BB, 8, 0, true);
1028 case Mips::ATOMIC_LOAD_SUB_I8:
1029 case Mips::ATOMIC_LOAD_SUB_I8_P8:
1030 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
1031 case Mips::ATOMIC_LOAD_SUB_I16:
1032 case Mips::ATOMIC_LOAD_SUB_I16_P8:
1033 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
1034 case Mips::ATOMIC_LOAD_SUB_I32:
1035 case Mips::ATOMIC_LOAD_SUB_I32_P8:
1036 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
1037 case Mips::ATOMIC_LOAD_SUB_I64:
1038 case Mips::ATOMIC_LOAD_SUB_I64_P8:
1039 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
1041 case Mips::ATOMIC_SWAP_I8:
1042 case Mips::ATOMIC_SWAP_I8_P8:
1043 return emitAtomicBinaryPartword(MI, BB, 1, 0);
1044 case Mips::ATOMIC_SWAP_I16:
1045 case Mips::ATOMIC_SWAP_I16_P8:
1046 return emitAtomicBinaryPartword(MI, BB, 2, 0);
1047 case Mips::ATOMIC_SWAP_I32:
1048 case Mips::ATOMIC_SWAP_I32_P8:
1049 return emitAtomicBinary(MI, BB, 4, 0);
1050 case Mips::ATOMIC_SWAP_I64:
1051 case Mips::ATOMIC_SWAP_I64_P8:
1052 return emitAtomicBinary(MI, BB, 8, 0);
1054 case Mips::ATOMIC_CMP_SWAP_I8:
1055 case Mips::ATOMIC_CMP_SWAP_I8_P8:
1056 return emitAtomicCmpSwapPartword(MI, BB, 1);
1057 case Mips::ATOMIC_CMP_SWAP_I16:
1058 case Mips::ATOMIC_CMP_SWAP_I16_P8:
1059 return emitAtomicCmpSwapPartword(MI, BB, 2);
1060 case Mips::ATOMIC_CMP_SWAP_I32:
1061 case Mips::ATOMIC_CMP_SWAP_I32_P8:
1062 return emitAtomicCmpSwap(MI, BB, 4);
1063 case Mips::ATOMIC_CMP_SWAP_I64:
1064 case Mips::ATOMIC_CMP_SWAP_I64_P8:
1065 return emitAtomicCmpSwap(MI, BB, 8);
1069 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1070 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1072 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
1073 unsigned Size, unsigned BinOpcode,
1075 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1077 MachineFunction *MF = BB->getParent();
1078 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1079 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1080 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1081 DebugLoc DL = MI->getDebugLoc();
1082 unsigned LL, SC, AND, NOR, ZERO, BEQ;
1085 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1086 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1093 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1094 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1097 ZERO = Mips::ZERO_64;
1101 unsigned OldVal = MI->getOperand(0).getReg();
1102 unsigned Ptr = MI->getOperand(1).getReg();
1103 unsigned Incr = MI->getOperand(2).getReg();
1105 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1106 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1107 unsigned Success = RegInfo.createVirtualRegister(RC);
1109 // insert new blocks after the current block
1110 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1111 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1112 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1113 MachineFunction::iterator It = BB;
1115 MF->insert(It, loopMBB);
1116 MF->insert(It, exitMBB);
1118 // Transfer the remainder of BB and its successor edges to exitMBB.
1119 exitMBB->splice(exitMBB->begin(), BB,
1120 llvm::next(MachineBasicBlock::iterator(MI)),
1122 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1126 // fallthrough --> loopMBB
1127 BB->addSuccessor(loopMBB);
1128 loopMBB->addSuccessor(loopMBB);
1129 loopMBB->addSuccessor(exitMBB);
1132 // ll oldval, 0(ptr)
1133 // <binop> storeval, oldval, incr
1134 // sc success, storeval, 0(ptr)
1135 // beq success, $0, loopMBB
1137 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1139 // and andres, oldval, incr
1140 // nor storeval, $0, andres
1141 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1142 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1143 } else if (BinOpcode) {
1144 // <binop> storeval, oldval, incr
1145 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1149 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1150 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1152 MI->eraseFromParent(); // The instruction is gone now.
1158 MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI,
1159 MachineBasicBlock *BB,
1160 unsigned Size, unsigned BinOpcode,
1162 assert((Size == 1 || Size == 2) &&
1163 "Unsupported size for EmitAtomicBinaryPartial.");
1165 MachineFunction *MF = BB->getParent();
1166 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1167 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1168 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1169 DebugLoc DL = MI->getDebugLoc();
1170 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1171 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1173 unsigned Dest = MI->getOperand(0).getReg();
1174 unsigned Ptr = MI->getOperand(1).getReg();
1175 unsigned Incr = MI->getOperand(2).getReg();
1177 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1178 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1179 unsigned Mask = RegInfo.createVirtualRegister(RC);
1180 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1181 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1182 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1183 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1184 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1185 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1186 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1187 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1188 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1189 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1190 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1191 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1192 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1193 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1194 unsigned Success = RegInfo.createVirtualRegister(RC);
1196 // insert new blocks after the current block
1197 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1198 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1199 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1200 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1201 MachineFunction::iterator It = BB;
1203 MF->insert(It, loopMBB);
1204 MF->insert(It, sinkMBB);
1205 MF->insert(It, exitMBB);
1207 // Transfer the remainder of BB and its successor edges to exitMBB.
1208 exitMBB->splice(exitMBB->begin(), BB,
1209 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1210 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1212 BB->addSuccessor(loopMBB);
1213 loopMBB->addSuccessor(loopMBB);
1214 loopMBB->addSuccessor(sinkMBB);
1215 sinkMBB->addSuccessor(exitMBB);
1218 // addiu masklsb2,$0,-4 # 0xfffffffc
1219 // and alignedaddr,ptr,masklsb2
1220 // andi ptrlsb2,ptr,3
1221 // sll shiftamt,ptrlsb2,3
1222 // ori maskupper,$0,255 # 0xff
1223 // sll mask,maskupper,shiftamt
1224 // nor mask2,$0,mask
1225 // sll incr2,incr,shiftamt
1227 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1228 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1229 .addReg(Mips::ZERO).addImm(-4);
1230 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1231 .addReg(Ptr).addReg(MaskLSB2);
1232 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1233 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1234 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1235 .addReg(Mips::ZERO).addImm(MaskImm);
1236 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1237 .addReg(ShiftAmt).addReg(MaskUpper);
1238 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1239 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1241 // atomic.load.binop
1243 // ll oldval,0(alignedaddr)
1244 // binop binopres,oldval,incr2
1245 // and newval,binopres,mask
1246 // and maskedoldval0,oldval,mask2
1247 // or storeval,maskedoldval0,newval
1248 // sc success,storeval,0(alignedaddr)
1249 // beq success,$0,loopMBB
1253 // ll oldval,0(alignedaddr)
1254 // and newval,incr2,mask
1255 // and maskedoldval0,oldval,mask2
1256 // or storeval,maskedoldval0,newval
1257 // sc success,storeval,0(alignedaddr)
1258 // beq success,$0,loopMBB
1261 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1263 // and andres, oldval, incr2
1264 // nor binopres, $0, andres
1265 // and newval, binopres, mask
1266 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1267 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1268 .addReg(Mips::ZERO).addReg(AndRes);
1269 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1270 } else if (BinOpcode) {
1271 // <binop> binopres, oldval, incr2
1272 // and newval, binopres, mask
1273 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1274 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1275 } else {// atomic.swap
1276 // and newval, incr2, mask
1277 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1280 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1281 .addReg(OldVal).addReg(Mask2);
1282 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1283 .addReg(MaskedOldVal0).addReg(NewVal);
1284 BuildMI(BB, DL, TII->get(SC), Success)
1285 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1286 BuildMI(BB, DL, TII->get(Mips::BEQ))
1287 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1290 // and maskedoldval1,oldval,mask
1291 // srl srlres,maskedoldval1,shiftamt
1292 // sll sllres,srlres,24
1293 // sra dest,sllres,24
1295 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1297 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1298 .addReg(OldVal).addReg(Mask);
1299 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1300 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1301 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1302 .addReg(SrlRes).addImm(ShiftImm);
1303 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1304 .addReg(SllRes).addImm(ShiftImm);
1306 MI->eraseFromParent(); // The instruction is gone now.
1312 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1313 MachineBasicBlock *BB,
1314 unsigned Size) const {
1315 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1317 MachineFunction *MF = BB->getParent();
1318 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1319 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1320 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1321 DebugLoc DL = MI->getDebugLoc();
1322 unsigned LL, SC, ZERO, BNE, BEQ;
1325 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1326 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1332 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1333 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1334 ZERO = Mips::ZERO_64;
1339 unsigned Dest = MI->getOperand(0).getReg();
1340 unsigned Ptr = MI->getOperand(1).getReg();
1341 unsigned OldVal = MI->getOperand(2).getReg();
1342 unsigned NewVal = MI->getOperand(3).getReg();
1344 unsigned Success = RegInfo.createVirtualRegister(RC);
1346 // insert new blocks after the current block
1347 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1348 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1349 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1350 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1351 MachineFunction::iterator It = BB;
1353 MF->insert(It, loop1MBB);
1354 MF->insert(It, loop2MBB);
1355 MF->insert(It, exitMBB);
1357 // Transfer the remainder of BB and its successor edges to exitMBB.
1358 exitMBB->splice(exitMBB->begin(), BB,
1359 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1360 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1364 // fallthrough --> loop1MBB
1365 BB->addSuccessor(loop1MBB);
1366 loop1MBB->addSuccessor(exitMBB);
1367 loop1MBB->addSuccessor(loop2MBB);
1368 loop2MBB->addSuccessor(loop1MBB);
1369 loop2MBB->addSuccessor(exitMBB);
1373 // bne dest, oldval, exitMBB
1375 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1376 BuildMI(BB, DL, TII->get(BNE))
1377 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1380 // sc success, newval, 0(ptr)
1381 // beq success, $0, loop1MBB
1383 BuildMI(BB, DL, TII->get(SC), Success)
1384 .addReg(NewVal).addReg(Ptr).addImm(0);
1385 BuildMI(BB, DL, TII->get(BEQ))
1386 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1388 MI->eraseFromParent(); // The instruction is gone now.
1394 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1395 MachineBasicBlock *BB,
1396 unsigned Size) const {
1397 assert((Size == 1 || Size == 2) &&
1398 "Unsupported size for EmitAtomicCmpSwapPartial.");
1400 MachineFunction *MF = BB->getParent();
1401 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1402 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1403 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1404 DebugLoc DL = MI->getDebugLoc();
1405 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1406 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1408 unsigned Dest = MI->getOperand(0).getReg();
1409 unsigned Ptr = MI->getOperand(1).getReg();
1410 unsigned CmpVal = MI->getOperand(2).getReg();
1411 unsigned NewVal = MI->getOperand(3).getReg();
1413 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1414 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1415 unsigned Mask = RegInfo.createVirtualRegister(RC);
1416 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1417 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1418 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1419 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1420 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1421 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1422 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1423 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1424 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1425 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1426 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1427 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1428 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1429 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1430 unsigned Success = RegInfo.createVirtualRegister(RC);
1432 // insert new blocks after the current block
1433 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1434 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1435 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1436 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1437 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1438 MachineFunction::iterator It = BB;
1440 MF->insert(It, loop1MBB);
1441 MF->insert(It, loop2MBB);
1442 MF->insert(It, sinkMBB);
1443 MF->insert(It, exitMBB);
1445 // Transfer the remainder of BB and its successor edges to exitMBB.
1446 exitMBB->splice(exitMBB->begin(), BB,
1447 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1448 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1450 BB->addSuccessor(loop1MBB);
1451 loop1MBB->addSuccessor(sinkMBB);
1452 loop1MBB->addSuccessor(loop2MBB);
1453 loop2MBB->addSuccessor(loop1MBB);
1454 loop2MBB->addSuccessor(sinkMBB);
1455 sinkMBB->addSuccessor(exitMBB);
1457 // FIXME: computation of newval2 can be moved to loop2MBB.
1459 // addiu masklsb2,$0,-4 # 0xfffffffc
1460 // and alignedaddr,ptr,masklsb2
1461 // andi ptrlsb2,ptr,3
1462 // sll shiftamt,ptrlsb2,3
1463 // ori maskupper,$0,255 # 0xff
1464 // sll mask,maskupper,shiftamt
1465 // nor mask2,$0,mask
1466 // andi maskedcmpval,cmpval,255
1467 // sll shiftedcmpval,maskedcmpval,shiftamt
1468 // andi maskednewval,newval,255
1469 // sll shiftednewval,maskednewval,shiftamt
1470 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1471 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1472 .addReg(Mips::ZERO).addImm(-4);
1473 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1474 .addReg(Ptr).addReg(MaskLSB2);
1475 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1476 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1477 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1478 .addReg(Mips::ZERO).addImm(MaskImm);
1479 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1480 .addReg(ShiftAmt).addReg(MaskUpper);
1481 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1482 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1483 .addReg(CmpVal).addImm(MaskImm);
1484 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1485 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1486 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1487 .addReg(NewVal).addImm(MaskImm);
1488 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1489 .addReg(ShiftAmt).addReg(MaskedNewVal);
1492 // ll oldval,0(alginedaddr)
1493 // and maskedoldval0,oldval,mask
1494 // bne maskedoldval0,shiftedcmpval,sinkMBB
1496 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1497 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1498 .addReg(OldVal).addReg(Mask);
1499 BuildMI(BB, DL, TII->get(Mips::BNE))
1500 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1503 // and maskedoldval1,oldval,mask2
1504 // or storeval,maskedoldval1,shiftednewval
1505 // sc success,storeval,0(alignedaddr)
1506 // beq success,$0,loop1MBB
1508 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1509 .addReg(OldVal).addReg(Mask2);
1510 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1511 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1512 BuildMI(BB, DL, TII->get(SC), Success)
1513 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1514 BuildMI(BB, DL, TII->get(Mips::BEQ))
1515 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1518 // srl srlres,maskedoldval0,shiftamt
1519 // sll sllres,srlres,24
1520 // sra dest,sllres,24
1522 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1524 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1525 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1526 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1527 .addReg(SrlRes).addImm(ShiftImm);
1528 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1529 .addReg(SllRes).addImm(ShiftImm);
1531 MI->eraseFromParent(); // The instruction is gone now.
1536 //===----------------------------------------------------------------------===//
1537 // Misc Lower Operation implementation
1538 //===----------------------------------------------------------------------===//
1539 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1540 SDValue Chain = Op.getOperand(0);
1541 SDValue Table = Op.getOperand(1);
1542 SDValue Index = Op.getOperand(2);
1543 DebugLoc DL = Op.getDebugLoc();
1544 EVT PTy = getPointerTy();
1545 unsigned EntrySize =
1546 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout());
1548 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1549 DAG.getConstant(EntrySize, PTy));
1550 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1552 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1553 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1554 MachinePointerInfo::getJumpTable(), MemVT, false, false,
1556 Chain = Addr.getValue(1);
1558 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) {
1559 // For PIC, the sequence is:
1560 // BRIND(load(Jumptable + index) + RelocBase)
1561 // RelocBase can be JumpTable, GOT or some sort of global base.
1562 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1563 getPICJumpTableRelocBase(Table, DAG));
1566 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1569 SDValue MipsTargetLowering::
1570 lowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1572 // The first operand is the chain, the second is the condition, the third is
1573 // the block to branch to if the condition is true.
1574 SDValue Chain = Op.getOperand(0);
1575 SDValue Dest = Op.getOperand(2);
1576 DebugLoc DL = Op.getDebugLoc();
1578 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1580 // Return if flag is not set by a floating point comparison.
1581 if (CondRes.getOpcode() != MipsISD::FPCmp)
1584 SDValue CCNode = CondRes.getOperand(2);
1586 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1587 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1588 SDValue BrCode = DAG.getConstant(Opc, MVT::i32);
1589 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1593 SDValue MipsTargetLowering::
1594 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1596 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1598 // Return if flag is not set by a floating point comparison.
1599 if (Cond.getOpcode() != MipsISD::FPCmp)
1602 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1606 SDValue MipsTargetLowering::
1607 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1609 DebugLoc DL = Op.getDebugLoc();
1610 EVT Ty = Op.getOperand(0).getValueType();
1611 SDValue Cond = DAG.getNode(ISD::SETCC, DL, getSetCCResultType(Ty),
1612 Op.getOperand(0), Op.getOperand(1),
1615 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1619 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1620 SDValue Cond = createFPCmp(DAG, Op);
1622 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1623 "Floating point operand expected.");
1625 SDValue True = DAG.getConstant(1, MVT::i32);
1626 SDValue False = DAG.getConstant(0, MVT::i32);
1628 return createCMovFP(DAG, Cond, True, False, Op.getDebugLoc());
1631 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1632 SelectionDAG &DAG) const {
1633 // FIXME there isn't actually debug info here
1634 DebugLoc DL = Op.getDebugLoc();
1635 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1637 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1638 const MipsTargetObjectFile &TLOF =
1639 (const MipsTargetObjectFile&)getObjFileLowering();
1641 // %gp_rel relocation
1642 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1643 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0,
1645 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL,
1646 DAG.getVTList(MVT::i32), &GA, 1);
1647 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1648 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode);
1651 // %hi/%lo relocation
1652 return getAddrNonPIC(Op, DAG);
1655 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1656 return getAddrLocal(Op, DAG, HasMips64);
1659 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16,
1660 MipsII::MO_GOT_LO16);
1662 return getAddrGlobal(Op, DAG,
1663 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
1666 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1667 SelectionDAG &DAG) const {
1668 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1669 return getAddrNonPIC(Op, DAG);
1671 return getAddrLocal(Op, DAG, HasMips64);
1674 SDValue MipsTargetLowering::
1675 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1677 // If the relocation model is PIC, use the General Dynamic TLS Model or
1678 // Local Dynamic TLS model, otherwise use the Initial Exec or
1679 // Local Exec TLS Model.
1681 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1682 DebugLoc DL = GA->getDebugLoc();
1683 const GlobalValue *GV = GA->getGlobal();
1684 EVT PtrVT = getPointerTy();
1686 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1688 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1689 // General Dynamic and Local Dynamic TLS Model.
1690 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1693 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1694 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1695 getGlobalReg(DAG, PtrVT), TGA);
1696 unsigned PtrSize = PtrVT.getSizeInBits();
1697 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1699 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1703 Entry.Node = Argument;
1705 Args.push_back(Entry);
1707 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
1708 false, false, false, false, 0, CallingConv::C,
1709 /*IsTailCall=*/false, /*doesNotRet=*/false,
1710 /*isReturnValueUsed=*/true,
1711 TlsGetAddr, Args, DAG, DL);
1712 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1714 SDValue Ret = CallResult.first;
1716 if (model != TLSModel::LocalDynamic)
1719 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1720 MipsII::MO_DTPREL_HI);
1721 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1722 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1723 MipsII::MO_DTPREL_LO);
1724 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1725 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1726 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1730 if (model == TLSModel::InitialExec) {
1731 // Initial Exec TLS Model
1732 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1733 MipsII::MO_GOTTPREL);
1734 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1736 Offset = DAG.getLoad(PtrVT, DL,
1737 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1738 false, false, false, 0);
1740 // Local Exec TLS Model
1741 assert(model == TLSModel::LocalExec);
1742 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1743 MipsII::MO_TPREL_HI);
1744 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1745 MipsII::MO_TPREL_LO);
1746 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1747 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1748 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1751 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1752 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1755 SDValue MipsTargetLowering::
1756 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1758 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1759 return getAddrNonPIC(Op, DAG);
1761 return getAddrLocal(Op, DAG, HasMips64);
1764 SDValue MipsTargetLowering::
1765 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1767 // gp_rel relocation
1768 // FIXME: we should reference the constant pool using small data sections,
1769 // but the asm printer currently doesn't support this feature without
1770 // hacking it. This feature should come soon so we can uncomment the
1772 //if (IsInSmallSection(C->getType())) {
1773 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1774 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1775 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1777 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1778 return getAddrNonPIC(Op, DAG);
1780 return getAddrLocal(Op, DAG, HasMips64);
1783 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1784 MachineFunction &MF = DAG.getMachineFunction();
1785 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1787 DebugLoc DL = Op.getDebugLoc();
1788 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1791 // vastart just stores the address of the VarArgsFrameIndex slot into the
1792 // memory location argument.
1793 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1794 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1795 MachinePointerInfo(SV), false, false, 0);
1798 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1799 EVT TyX = Op.getOperand(0).getValueType();
1800 EVT TyY = Op.getOperand(1).getValueType();
1801 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1802 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1803 DebugLoc DL = Op.getDebugLoc();
1806 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1808 SDValue X = (TyX == MVT::f32) ?
1809 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1810 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1812 SDValue Y = (TyY == MVT::f32) ?
1813 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1814 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1818 // ext E, Y, 31, 1 ; extract bit31 of Y
1819 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1820 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1821 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1824 // srl SrlX, SllX, 1
1826 // sll SllY, SrlX, 31
1827 // or Or, SrlX, SllY
1828 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1829 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1830 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1831 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1832 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1835 if (TyX == MVT::f32)
1836 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1838 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1839 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1840 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1843 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1844 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1845 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1846 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1847 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1848 DebugLoc DL = Op.getDebugLoc();
1850 // Bitcast to integer nodes.
1851 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1852 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1855 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1856 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1857 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1858 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
1860 if (WidthX > WidthY)
1861 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1862 else if (WidthY > WidthX)
1863 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1865 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1866 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
1867 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1870 // (d)sll SllX, X, 1
1871 // (d)srl SrlX, SllX, 1
1872 // (d)srl SrlY, Y, width(Y)-1
1873 // (d)sll SllY, SrlX, width(Y)-1
1874 // or Or, SrlX, SllY
1875 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1876 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1877 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1878 DAG.getConstant(WidthY - 1, MVT::i32));
1880 if (WidthX > WidthY)
1881 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1882 else if (WidthY > WidthX)
1883 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1885 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1886 DAG.getConstant(WidthX - 1, MVT::i32));
1887 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
1888 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
1892 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1893 if (Subtarget->hasMips64())
1894 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
1896 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
1899 static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1900 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1901 DebugLoc DL = Op.getDebugLoc();
1903 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1905 SDValue X = (Op.getValueType() == MVT::f32) ?
1906 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1907 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1912 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
1913 DAG.getRegister(Mips::ZERO, MVT::i32),
1914 DAG.getConstant(31, MVT::i32), Const1, X);
1916 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1917 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1920 if (Op.getValueType() == MVT::f32)
1921 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
1923 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1924 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1925 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1928 static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1929 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1930 DebugLoc DL = Op.getDebugLoc();
1932 // Bitcast to integer node.
1933 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
1937 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
1938 DAG.getRegister(Mips::ZERO_64, MVT::i64),
1939 DAG.getConstant(63, MVT::i32), Const1, X);
1941 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
1942 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
1945 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
1949 MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const {
1950 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
1951 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2());
1953 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2());
1956 SDValue MipsTargetLowering::
1957 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1959 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1960 "Frame address can only be determined for current frame.");
1962 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1963 MFI->setFrameAddressIsTaken(true);
1964 EVT VT = Op.getValueType();
1965 DebugLoc DL = Op.getDebugLoc();
1966 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1967 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1971 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
1972 SelectionDAG &DAG) const {
1974 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1975 "Return address can be determined only for current frame.");
1977 MachineFunction &MF = DAG.getMachineFunction();
1978 MachineFrameInfo *MFI = MF.getFrameInfo();
1979 MVT VT = Op.getSimpleValueType();
1980 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
1981 MFI->setReturnAddressIsTaken(true);
1983 // Return RA, which contains the return address. Mark it an implicit live-in.
1984 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
1985 return DAG.getCopyFromReg(DAG.getEntryNode(), Op.getDebugLoc(), Reg, VT);
1988 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
1989 // generated from __builtin_eh_return (offset, handler)
1990 // The effect of this is to adjust the stack pointer by "offset"
1991 // and then branch to "handler".
1992 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
1994 MachineFunction &MF = DAG.getMachineFunction();
1995 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1997 MipsFI->setCallsEhReturn();
1998 SDValue Chain = Op.getOperand(0);
1999 SDValue Offset = Op.getOperand(1);
2000 SDValue Handler = Op.getOperand(2);
2001 DebugLoc DL = Op.getDebugLoc();
2002 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2004 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2005 // EH_RETURN nodes, so that instructions are emitted back-to-back.
2006 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1;
2007 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0;
2008 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2009 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2010 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2011 DAG.getRegister(OffsetReg, Ty),
2012 DAG.getRegister(AddrReg, getPointerTy()),
2016 // TODO: set SType according to the desired memory barrier behavior.
2018 MipsTargetLowering::lowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const {
2020 DebugLoc DL = Op.getDebugLoc();
2021 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2022 DAG.getConstant(SType, MVT::i32));
2025 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
2026 SelectionDAG &DAG) const {
2027 // FIXME: Need pseudo-fence for 'singlethread' fences
2028 // FIXME: Set SType for weaker fences where supported/appropriate.
2030 DebugLoc DL = Op.getDebugLoc();
2031 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2032 DAG.getConstant(SType, MVT::i32));
2035 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
2036 SelectionDAG &DAG) const {
2037 DebugLoc DL = Op.getDebugLoc();
2038 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2039 SDValue Shamt = Op.getOperand(2);
2042 // lo = (shl lo, shamt)
2043 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2046 // hi = (shl lo, shamt[4:0])
2047 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2048 DAG.getConstant(-1, MVT::i32));
2049 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
2050 DAG.getConstant(1, MVT::i32));
2051 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
2053 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
2054 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2055 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
2056 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2057 DAG.getConstant(0x20, MVT::i32));
2058 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2059 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
2060 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
2062 SDValue Ops[2] = {Lo, Hi};
2063 return DAG.getMergeValues(Ops, 2, DL);
2066 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2068 DebugLoc DL = Op.getDebugLoc();
2069 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2070 SDValue Shamt = Op.getOperand(2);
2073 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2075 // hi = (sra hi, shamt)
2077 // hi = (srl hi, shamt)
2080 // lo = (sra hi, shamt[4:0])
2081 // hi = (sra hi, 31)
2083 // lo = (srl hi, shamt[4:0])
2085 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2086 DAG.getConstant(-1, MVT::i32));
2087 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
2088 DAG.getConstant(1, MVT::i32));
2089 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
2090 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
2091 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2092 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2094 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2095 DAG.getConstant(0x20, MVT::i32));
2096 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
2097 DAG.getConstant(31, MVT::i32));
2098 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
2099 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2100 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
2103 SDValue Ops[2] = {Lo, Hi};
2104 return DAG.getMergeValues(Ops, 2, DL);
2107 static SDValue CreateLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2108 SDValue Chain, SDValue Src, unsigned Offset) {
2109 SDValue Ptr = LD->getBasePtr();
2110 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2111 EVT BasePtrVT = Ptr.getValueType();
2112 DebugLoc DL = LD->getDebugLoc();
2113 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2116 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2117 DAG.getConstant(Offset, BasePtrVT));
2119 SDValue Ops[] = { Chain, Ptr, Src };
2120 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2121 LD->getMemOperand());
2124 // Expand an unaligned 32 or 64-bit integer load node.
2125 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2126 LoadSDNode *LD = cast<LoadSDNode>(Op);
2127 EVT MemVT = LD->getMemoryVT();
2129 // Return if load is aligned or if MemVT is neither i32 nor i64.
2130 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2131 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2134 bool IsLittle = Subtarget->isLittle();
2135 EVT VT = Op.getValueType();
2136 ISD::LoadExtType ExtType = LD->getExtensionType();
2137 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2139 assert((VT == MVT::i32) || (VT == MVT::i64));
2142 // (set dst, (i64 (load baseptr)))
2144 // (set tmp, (ldl (add baseptr, 7), undef))
2145 // (set dst, (ldr baseptr, tmp))
2146 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2147 SDValue LDL = CreateLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2149 return CreateLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2153 SDValue LWL = CreateLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2155 SDValue LWR = CreateLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2159 // (set dst, (i32 (load baseptr))) or
2160 // (set dst, (i64 (sextload baseptr))) or
2161 // (set dst, (i64 (extload baseptr)))
2163 // (set tmp, (lwl (add baseptr, 3), undef))
2164 // (set dst, (lwr baseptr, tmp))
2165 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2166 (ExtType == ISD::EXTLOAD))
2169 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2172 // (set dst, (i64 (zextload baseptr)))
2174 // (set tmp0, (lwl (add baseptr, 3), undef))
2175 // (set tmp1, (lwr baseptr, tmp0))
2176 // (set tmp2, (shl tmp1, 32))
2177 // (set dst, (srl tmp2, 32))
2178 DebugLoc DL = LD->getDebugLoc();
2179 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2180 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2181 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2182 SDValue Ops[] = { SRL, LWR.getValue(1) };
2183 return DAG.getMergeValues(Ops, 2, DL);
2186 static SDValue CreateStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2187 SDValue Chain, unsigned Offset) {
2188 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2189 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2190 DebugLoc DL = SD->getDebugLoc();
2191 SDVTList VTList = DAG.getVTList(MVT::Other);
2194 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2195 DAG.getConstant(Offset, BasePtrVT));
2197 SDValue Ops[] = { Chain, Value, Ptr };
2198 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2199 SD->getMemOperand());
2202 // Expand an unaligned 32 or 64-bit integer store node.
2203 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2204 StoreSDNode *SD = cast<StoreSDNode>(Op);
2205 EVT MemVT = SD->getMemoryVT();
2207 // Return if store is aligned or if MemVT is neither i32 nor i64.
2208 if ((SD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2209 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2212 bool IsLittle = Subtarget->isLittle();
2213 SDValue Value = SD->getValue(), Chain = SD->getChain();
2214 EVT VT = Value.getValueType();
2217 // (store val, baseptr) or
2218 // (truncstore val, baseptr)
2220 // (swl val, (add baseptr, 3))
2221 // (swr val, baseptr)
2222 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2223 SDValue SWL = CreateStoreLR(MipsISD::SWL, DAG, SD, Chain,
2225 return CreateStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2228 assert(VT == MVT::i64);
2231 // (store val, baseptr)
2233 // (sdl val, (add baseptr, 7))
2234 // (sdr val, baseptr)
2235 SDValue SDL = CreateStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2236 return CreateStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2239 // This function expands mips intrinsic nodes which have 64-bit input operands
2240 // or output values.
2242 // out64 = intrinsic-node in64
2244 // lo = copy (extract-element (in64, 0))
2245 // hi = copy (extract-element (in64, 1))
2246 // mips-specific-node
2249 // out64 = merge-values (v0, v1)
2251 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG,
2252 unsigned Opc, bool HasI64In, bool HasI64Out) {
2253 DebugLoc DL = Op.getDebugLoc();
2254 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
2255 SDValue Chain = HasChainIn ? Op->getOperand(0) : DAG.getEntryNode();
2256 SmallVector<SDValue, 3> Ops;
2259 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2260 Op->getOperand(1 + HasChainIn),
2261 DAG.getConstant(0, MVT::i32));
2262 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2263 Op->getOperand(1 + HasChainIn),
2264 DAG.getConstant(1, MVT::i32));
2266 Chain = DAG.getCopyToReg(Chain, DL, Mips::LO, InLo, SDValue());
2267 Chain = DAG.getCopyToReg(Chain, DL, Mips::HI, InHi, Chain.getValue(1));
2269 Ops.push_back(Chain);
2270 Ops.append(Op->op_begin() + HasChainIn + 2, Op->op_end());
2271 Ops.push_back(Chain.getValue(1));
2273 Ops.push_back(Chain);
2274 Ops.append(Op->op_begin() + HasChainIn + 1, Op->op_end());
2278 return DAG.getNode(Opc, DL, Op->value_begin(), Op->getNumValues(),
2279 Ops.begin(), Ops.size());
2281 SDValue Intr = DAG.getNode(Opc, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2282 Ops.begin(), Ops.size());
2283 SDValue OutLo = DAG.getCopyFromReg(Intr.getValue(0), DL, Mips::LO, MVT::i32,
2285 SDValue OutHi = DAG.getCopyFromReg(OutLo.getValue(1), DL, Mips::HI, MVT::i32,
2287 SDValue Out = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, OutLo, OutHi);
2292 SDValue Vals[] = { Out, OutHi.getValue(1) };
2293 return DAG.getMergeValues(Vals, 2, DL);
2296 SDValue MipsTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
2297 SelectionDAG &DAG) const {
2298 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
2301 case Intrinsic::mips_shilo:
2302 return lowerDSPIntr(Op, DAG, MipsISD::SHILO, true, true);
2303 case Intrinsic::mips_dpau_h_qbl:
2304 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL, true, true);
2305 case Intrinsic::mips_dpau_h_qbr:
2306 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR, true, true);
2307 case Intrinsic::mips_dpsu_h_qbl:
2308 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL, true, true);
2309 case Intrinsic::mips_dpsu_h_qbr:
2310 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR, true, true);
2311 case Intrinsic::mips_dpa_w_ph:
2312 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH, true, true);
2313 case Intrinsic::mips_dps_w_ph:
2314 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH, true, true);
2315 case Intrinsic::mips_dpax_w_ph:
2316 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH, true, true);
2317 case Intrinsic::mips_dpsx_w_ph:
2318 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH, true, true);
2319 case Intrinsic::mips_mulsa_w_ph:
2320 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH, true, true);
2321 case Intrinsic::mips_mult:
2322 return lowerDSPIntr(Op, DAG, MipsISD::MULT, false, true);
2323 case Intrinsic::mips_multu:
2324 return lowerDSPIntr(Op, DAG, MipsISD::MULTU, false, true);
2325 case Intrinsic::mips_madd:
2326 return lowerDSPIntr(Op, DAG, MipsISD::MADD_DSP, true, true);
2327 case Intrinsic::mips_maddu:
2328 return lowerDSPIntr(Op, DAG, MipsISD::MADDU_DSP, true, true);
2329 case Intrinsic::mips_msub:
2330 return lowerDSPIntr(Op, DAG, MipsISD::MSUB_DSP, true, true);
2331 case Intrinsic::mips_msubu:
2332 return lowerDSPIntr(Op, DAG, MipsISD::MSUBU_DSP, true, true);
2336 SDValue MipsTargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
2337 SelectionDAG &DAG) const {
2338 switch (cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue()) {
2341 case Intrinsic::mips_extp:
2342 return lowerDSPIntr(Op, DAG, MipsISD::EXTP, true, false);
2343 case Intrinsic::mips_extpdp:
2344 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP, true, false);
2345 case Intrinsic::mips_extr_w:
2346 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W, true, false);
2347 case Intrinsic::mips_extr_r_w:
2348 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W, true, false);
2349 case Intrinsic::mips_extr_rs_w:
2350 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W, true, false);
2351 case Intrinsic::mips_extr_s_h:
2352 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H, true, false);
2353 case Intrinsic::mips_mthlip:
2354 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP, true, true);
2355 case Intrinsic::mips_mulsaq_s_w_ph:
2356 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH, true, true);
2357 case Intrinsic::mips_maq_s_w_phl:
2358 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL, true, true);
2359 case Intrinsic::mips_maq_s_w_phr:
2360 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR, true, true);
2361 case Intrinsic::mips_maq_sa_w_phl:
2362 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL, true, true);
2363 case Intrinsic::mips_maq_sa_w_phr:
2364 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR, true, true);
2365 case Intrinsic::mips_dpaq_s_w_ph:
2366 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH, true, true);
2367 case Intrinsic::mips_dpsq_s_w_ph:
2368 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH, true, true);
2369 case Intrinsic::mips_dpaq_sa_l_w:
2370 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W, true, true);
2371 case Intrinsic::mips_dpsq_sa_l_w:
2372 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W, true, true);
2373 case Intrinsic::mips_dpaqx_s_w_ph:
2374 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH, true, true);
2375 case Intrinsic::mips_dpaqx_sa_w_ph:
2376 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH, true, true);
2377 case Intrinsic::mips_dpsqx_s_w_ph:
2378 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH, true, true);
2379 case Intrinsic::mips_dpsqx_sa_w_ph:
2380 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH, true, true);
2384 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2385 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2386 || cast<ConstantSDNode>
2387 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2388 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2392 // (add (frameaddr 0), (frame_to_args_offset))
2393 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2394 // (add FrameObject, 0)
2395 // where FrameObject is a fixed StackObject with offset 0 which points to
2396 // the old stack pointer.
2397 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2398 EVT ValTy = Op->getValueType(0);
2399 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2400 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2401 return DAG.getNode(ISD::ADD, Op->getDebugLoc(), ValTy, InArgsAddr,
2402 DAG.getConstant(0, ValTy));
2405 //===----------------------------------------------------------------------===//
2406 // Calling Convention Implementation
2407 //===----------------------------------------------------------------------===//
2409 //===----------------------------------------------------------------------===//
2410 // TODO: Implement a generic logic using tblgen that can support this.
2411 // Mips O32 ABI rules:
2413 // i32 - Passed in A0, A1, A2, A3 and stack
2414 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2415 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2416 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2417 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2418 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2421 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2422 //===----------------------------------------------------------------------===//
2424 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2425 MVT LocVT, CCValAssign::LocInfo LocInfo,
2426 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2428 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2430 static const uint16_t IntRegs[] = {
2431 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2433 static const uint16_t F32Regs[] = {
2434 Mips::F12, Mips::F14
2436 static const uint16_t F64Regs[] = {
2440 // Do not process byval args here.
2441 if (ArgFlags.isByVal())
2444 // Promote i8 and i16
2445 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2447 if (ArgFlags.isSExt())
2448 LocInfo = CCValAssign::SExt;
2449 else if (ArgFlags.isZExt())
2450 LocInfo = CCValAssign::ZExt;
2452 LocInfo = CCValAssign::AExt;
2457 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2458 // is true: function is vararg, argument is 3rd or higher, there is previous
2459 // argument which is not f32 or f64.
2460 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2461 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2462 unsigned OrigAlign = ArgFlags.getOrigAlign();
2463 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2465 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2466 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2467 // If this is the first part of an i64 arg,
2468 // the allocated register must be either A0 or A2.
2469 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2470 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2472 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2473 // Allocate int register and shadow next int register. If first
2474 // available register is Mips::A1 or Mips::A3, shadow it too.
2475 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2476 if (Reg == Mips::A1 || Reg == Mips::A3)
2477 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2478 State.AllocateReg(IntRegs, IntRegsSize);
2480 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2481 // we are guaranteed to find an available float register
2482 if (ValVT == MVT::f32) {
2483 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2484 // Shadow int register
2485 State.AllocateReg(IntRegs, IntRegsSize);
2487 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2488 // Shadow int registers
2489 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2490 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2491 State.AllocateReg(IntRegs, IntRegsSize);
2492 State.AllocateReg(IntRegs, IntRegsSize);
2495 llvm_unreachable("Cannot handle this ValVT.");
2498 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2500 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2502 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2507 #include "MipsGenCallingConv.inc"
2509 //===----------------------------------------------------------------------===//
2510 // Call Calling Convention Implementation
2511 //===----------------------------------------------------------------------===//
2513 static const unsigned O32IntRegsSize = 4;
2515 // Return next O32 integer argument register.
2516 static unsigned getNextIntArgReg(unsigned Reg) {
2517 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2518 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2522 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2523 SDValue Chain, SDValue Arg, DebugLoc DL,
2524 bool IsTailCall, SelectionDAG &DAG) const {
2526 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2527 DAG.getIntPtrConstant(Offset));
2528 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2532 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2533 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2534 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2535 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2536 /*isVolatile=*/ true, false, 0);
2539 void MipsTargetLowering::
2540 getOpndList(SmallVectorImpl<SDValue> &Ops,
2541 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2542 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2543 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
2544 // Insert node "GP copy globalreg" before call to function.
2546 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2547 // in PIC mode) allow symbols to be resolved via lazy binding.
2548 // The lazy binding stub requires GP to point to the GOT.
2549 if (IsPICCall && !InternalLinkage) {
2550 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
2551 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2552 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2555 // Build a sequence of copy-to-reg nodes chained together with token
2556 // chain and flag operands which copy the outgoing args into registers.
2557 // The InFlag in necessary since all emitted instructions must be
2561 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2562 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2563 RegsToPass[i].second, InFlag);
2564 InFlag = Chain.getValue(1);
2567 // Add argument registers to the end of the list so that they are
2568 // known live into the call.
2569 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2570 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2571 RegsToPass[i].second.getValueType()));
2573 // Add a register mask operand representing the call-preserved registers.
2574 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2575 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv);
2576 assert(Mask && "Missing call preserved mask for calling convention");
2577 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2579 if (InFlag.getNode())
2580 Ops.push_back(InFlag);
2583 /// LowerCall - functions arguments are copied from virtual regs to
2584 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2586 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2587 SmallVectorImpl<SDValue> &InVals) const {
2588 SelectionDAG &DAG = CLI.DAG;
2589 DebugLoc &DL = CLI.DL;
2590 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2591 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2592 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2593 SDValue Chain = CLI.Chain;
2594 SDValue Callee = CLI.Callee;
2595 bool &IsTailCall = CLI.IsTailCall;
2596 CallingConv::ID CallConv = CLI.CallConv;
2597 bool IsVarArg = CLI.IsVarArg;
2599 MachineFunction &MF = DAG.getMachineFunction();
2600 MachineFrameInfo *MFI = MF.getFrameInfo();
2601 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2602 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2604 // Analyze operands of the call, assigning locations to each operand.
2605 SmallVector<CCValAssign, 16> ArgLocs;
2606 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2607 getTargetMachine(), ArgLocs, *DAG.getContext());
2608 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2610 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg,
2611 getTargetMachine().Options.UseSoftFloat,
2612 Callee.getNode(), CLI.Args);
2614 // Get a count of how many bytes are to be pushed on the stack.
2615 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2617 // Check if it's really possible to do a tail call.
2620 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2621 *MF.getInfo<MipsFunctionInfo>());
2626 // Chain is the output chain of the last Load/Store or CopyToReg node.
2627 // ByValChain is the output chain of the last Memcpy node created for copying
2628 // byval arguments to the stack.
2629 unsigned StackAlignment = TFL->getStackAlignment();
2630 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2631 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2634 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal);
2636 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL,
2637 IsN64 ? Mips::SP_64 : Mips::SP,
2640 // With EABI is it possible to have 16 args on registers.
2641 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2642 SmallVector<SDValue, 8> MemOpChains;
2643 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2645 // Walk the register/memloc assignments, inserting copies/loads.
2646 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2647 SDValue Arg = OutVals[i];
2648 CCValAssign &VA = ArgLocs[i];
2649 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2650 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2653 if (Flags.isByVal()) {
2654 assert(Flags.getByValSize() &&
2655 "ByVal args of size 0 should have been ignored by front-end.");
2656 assert(ByValArg != MipsCCInfo.byval_end());
2657 assert(!IsTailCall &&
2658 "Do not tail-call optimize if there is a byval argument.");
2659 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2660 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2665 // Promote the value if needed.
2666 switch (VA.getLocInfo()) {
2667 default: llvm_unreachable("Unknown loc info!");
2668 case CCValAssign::Full:
2669 if (VA.isRegLoc()) {
2670 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2671 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2672 (ValVT == MVT::i64 && LocVT == MVT::f64))
2673 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2674 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2675 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2676 Arg, DAG.getConstant(0, MVT::i32));
2677 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2678 Arg, DAG.getConstant(1, MVT::i32));
2679 if (!Subtarget->isLittle())
2681 unsigned LocRegLo = VA.getLocReg();
2682 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2683 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2684 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2689 case CCValAssign::SExt:
2690 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2692 case CCValAssign::ZExt:
2693 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2695 case CCValAssign::AExt:
2696 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2700 // Arguments that can be passed on register must be kept at
2701 // RegsToPass vector
2702 if (VA.isRegLoc()) {
2703 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2707 // Register can't get to this point...
2708 assert(VA.isMemLoc());
2710 // emit ISD::STORE whichs stores the
2711 // parameter value to a stack Location
2712 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2713 Chain, Arg, DL, IsTailCall, DAG));
2716 // Transform all store nodes into one single node because all store
2717 // nodes are independent of each other.
2718 if (!MemOpChains.empty())
2719 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2720 &MemOpChains[0], MemOpChains.size());
2722 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2723 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2724 // node so that legalize doesn't hack it.
2725 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2726 bool GlobalOrExternal = false, InternalLinkage = false;
2729 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2731 InternalLinkage = G->getGlobal()->hasInternalLinkage();
2733 if (InternalLinkage)
2734 Callee = getAddrLocal(Callee, DAG, HasMips64);
2736 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2737 MipsII::MO_CALL_LO16);
2739 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2741 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0,
2742 MipsII::MO_NO_FLAG);
2743 GlobalOrExternal = true;
2745 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2746 if (!IsN64 && !IsPIC) // !N64 && static
2747 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2748 MipsII::MO_NO_FLAG);
2750 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2751 MipsII::MO_CALL_LO16);
2753 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2755 GlobalOrExternal = true;
2758 SmallVector<SDValue, 8> Ops(1, Chain);
2759 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2761 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2762 CLI, Callee, Chain);
2765 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size());
2767 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size());
2768 SDValue InFlag = Chain.getValue(1);
2770 // Create the CALLSEQ_END node.
2771 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2772 DAG.getIntPtrConstant(0, true), InFlag);
2773 InFlag = Chain.getValue(1);
2775 // Handle result values, copying them out of physregs into vregs that we
2777 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg,
2778 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy);
2781 /// LowerCallResult - Lower the result values of a call into the
2782 /// appropriate copies out of appropriate physical registers.
2784 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2785 CallingConv::ID CallConv, bool IsVarArg,
2786 const SmallVectorImpl<ISD::InputArg> &Ins,
2787 DebugLoc DL, SelectionDAG &DAG,
2788 SmallVectorImpl<SDValue> &InVals,
2789 const SDNode *CallNode,
2790 const Type *RetTy) const {
2791 // Assign locations to each value returned by this call.
2792 SmallVector<CCValAssign, 16> RVLocs;
2793 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2794 getTargetMachine(), RVLocs, *DAG.getContext());
2795 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2797 MipsCCInfo.analyzeCallResult(Ins, getTargetMachine().Options.UseSoftFloat,
2800 // Copy all of the result registers out of their specified physreg.
2801 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2802 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2803 RVLocs[i].getLocVT(), InFlag);
2804 Chain = Val.getValue(1);
2805 InFlag = Val.getValue(2);
2807 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2808 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val);
2810 InVals.push_back(Val);
2816 //===----------------------------------------------------------------------===//
2817 // Formal Arguments Calling Convention Implementation
2818 //===----------------------------------------------------------------------===//
2819 /// LowerFormalArguments - transform physical registers into virtual registers
2820 /// and generate load operations for arguments places on the stack.
2822 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2823 CallingConv::ID CallConv,
2825 const SmallVectorImpl<ISD::InputArg> &Ins,
2826 DebugLoc DL, SelectionDAG &DAG,
2827 SmallVectorImpl<SDValue> &InVals)
2829 MachineFunction &MF = DAG.getMachineFunction();
2830 MachineFrameInfo *MFI = MF.getFrameInfo();
2831 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2833 MipsFI->setVarArgsFrameIndex(0);
2835 // Used with vargs to acumulate store chains.
2836 std::vector<SDValue> OutChains;
2838 // Assign locations to all of the incoming arguments.
2839 SmallVector<CCValAssign, 16> ArgLocs;
2840 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2841 getTargetMachine(), ArgLocs, *DAG.getContext());
2842 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2843 Function::const_arg_iterator FuncArg =
2844 DAG.getMachineFunction().getFunction()->arg_begin();
2845 bool UseSoftFloat = getTargetMachine().Options.UseSoftFloat;
2847 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg);
2848 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2849 MipsCCInfo.hasByValArg());
2851 unsigned CurArgIdx = 0;
2852 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2854 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2855 CCValAssign &VA = ArgLocs[i];
2856 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
2857 CurArgIdx = Ins[i].OrigArgIndex;
2858 EVT ValVT = VA.getValVT();
2859 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2860 bool IsRegLoc = VA.isRegLoc();
2862 if (Flags.isByVal()) {
2863 assert(Flags.getByValSize() &&
2864 "ByVal args of size 0 should have been ignored by front-end.");
2865 assert(ByValArg != MipsCCInfo.byval_end());
2866 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2867 MipsCCInfo, *ByValArg);
2872 // Arguments stored on registers
2874 EVT RegVT = VA.getLocVT();
2875 unsigned ArgReg = VA.getLocReg();
2876 const TargetRegisterClass *RC;
2878 if (RegVT == MVT::i32)
2879 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass :
2880 &Mips::CPURegsRegClass;
2881 else if (RegVT == MVT::i64)
2882 RC = &Mips::CPU64RegsRegClass;
2883 else if (RegVT == MVT::f32)
2884 RC = &Mips::FGR32RegClass;
2885 else if (RegVT == MVT::f64)
2886 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass;
2888 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2890 // Transform the arguments stored on
2891 // physical registers into virtual ones
2892 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2893 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2895 // If this is an 8 or 16-bit value, it has been passed promoted
2896 // to 32 bits. Insert an assert[sz]ext to capture this, then
2897 // truncate to the right size.
2898 if (VA.getLocInfo() != CCValAssign::Full) {
2899 unsigned Opcode = 0;
2900 if (VA.getLocInfo() == CCValAssign::SExt)
2901 Opcode = ISD::AssertSext;
2902 else if (VA.getLocInfo() == CCValAssign::ZExt)
2903 Opcode = ISD::AssertZext;
2905 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue,
2906 DAG.getValueType(ValVT));
2907 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue);
2910 // Handle floating point arguments passed in integer registers and
2911 // long double arguments passed in floating point registers.
2912 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2913 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
2914 (RegVT == MVT::f64 && ValVT == MVT::i64))
2915 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
2916 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2917 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
2918 getNextIntArgReg(ArgReg), RC);
2919 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
2920 if (!Subtarget->isLittle())
2921 std::swap(ArgValue, ArgValue2);
2922 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
2923 ArgValue, ArgValue2);
2926 InVals.push_back(ArgValue);
2927 } else { // VA.isRegLoc()
2930 assert(VA.isMemLoc());
2932 // The stack pointer offset is relative to the caller stack frame.
2933 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2934 VA.getLocMemOffset(), true);
2936 // Create load nodes to retrieve arguments from the stack
2937 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2938 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN,
2939 MachinePointerInfo::getFixedStack(FI),
2940 false, false, false, 0));
2944 // The mips ABIs for returning structs by value requires that we copy
2945 // the sret argument into $v0 for the return. Save the argument into
2946 // a virtual register so that we can access it from the return points.
2947 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2948 unsigned Reg = MipsFI->getSRetReturnReg();
2950 Reg = MF.getRegInfo().
2951 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
2952 MipsFI->setSRetReturnReg(Reg);
2954 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]);
2955 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
2959 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG);
2961 // All stores are grouped in one node to allow the matching between
2962 // the size of Ins and InVals. This only happens when on varg functions
2963 if (!OutChains.empty()) {
2964 OutChains.push_back(Chain);
2965 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2966 &OutChains[0], OutChains.size());
2972 //===----------------------------------------------------------------------===//
2973 // Return Value Calling Convention Implementation
2974 //===----------------------------------------------------------------------===//
2977 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2978 MachineFunction &MF, bool IsVarArg,
2979 const SmallVectorImpl<ISD::OutputArg> &Outs,
2980 LLVMContext &Context) const {
2981 SmallVector<CCValAssign, 16> RVLocs;
2982 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(),
2984 return CCInfo.CheckReturn(Outs, RetCC_Mips);
2988 MipsTargetLowering::LowerReturn(SDValue Chain,
2989 CallingConv::ID CallConv, bool IsVarArg,
2990 const SmallVectorImpl<ISD::OutputArg> &Outs,
2991 const SmallVectorImpl<SDValue> &OutVals,
2992 DebugLoc DL, SelectionDAG &DAG) const {
2993 // CCValAssign - represent the assignment of
2994 // the return value to a location
2995 SmallVector<CCValAssign, 16> RVLocs;
2996 MachineFunction &MF = DAG.getMachineFunction();
2998 // CCState - Info about the registers and stack slot.
2999 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs,
3001 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
3003 // Analyze return values.
3004 MipsCCInfo.analyzeReturn(Outs, getTargetMachine().Options.UseSoftFloat,
3005 MF.getFunction()->getReturnType());
3008 SmallVector<SDValue, 4> RetOps(1, Chain);
3010 // Copy the result values into the output registers.
3011 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3012 SDValue Val = OutVals[i];
3013 CCValAssign &VA = RVLocs[i];
3014 assert(VA.isRegLoc() && "Can only return in registers!");
3016 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
3017 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val);
3019 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
3021 // Guarantee that all emitted copies are stuck together with flags.
3022 Flag = Chain.getValue(1);
3023 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3026 // The mips ABIs for returning structs by value requires that we copy
3027 // the sret argument into $v0 for the return. We saved the argument into
3028 // a virtual register in the entry block, so now we copy the value out
3030 if (MF.getFunction()->hasStructRetAttr()) {
3031 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3032 unsigned Reg = MipsFI->getSRetReturnReg();
3035 llvm_unreachable("sret virtual register not created in the entry block");
3036 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
3037 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
3039 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
3040 Flag = Chain.getValue(1);
3041 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
3044 RetOps[0] = Chain; // Update chain.
3046 // Add the flag if we have it.
3048 RetOps.push_back(Flag);
3050 // Return on Mips is always a "jr $ra"
3051 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size());
3054 //===----------------------------------------------------------------------===//
3055 // Mips Inline Assembly Support
3056 //===----------------------------------------------------------------------===//
3058 /// getConstraintType - Given a constraint letter, return the type of
3059 /// constraint it is for this target.
3060 MipsTargetLowering::ConstraintType MipsTargetLowering::
3061 getConstraintType(const std::string &Constraint) const
3063 // Mips specific constrainy
3064 // GCC config/mips/constraints.md
3066 // 'd' : An address register. Equivalent to r
3067 // unless generating MIPS16 code.
3068 // 'y' : Equivalent to r; retained for
3069 // backwards compatibility.
3070 // 'c' : A register suitable for use in an indirect
3071 // jump. This will always be $25 for -mabicalls.
3072 // 'l' : The lo register. 1 word storage.
3073 // 'x' : The hilo register pair. Double word storage.
3074 if (Constraint.size() == 1) {
3075 switch (Constraint[0]) {
3083 return C_RegisterClass;
3088 return TargetLowering::getConstraintType(Constraint);
3091 /// Examine constraint type and operand type and determine a weight value.
3092 /// This object must already have been set up with the operand type
3093 /// and the current alternative constraint selected.
3094 TargetLowering::ConstraintWeight
3095 MipsTargetLowering::getSingleConstraintMatchWeight(
3096 AsmOperandInfo &info, const char *constraint) const {
3097 ConstraintWeight weight = CW_Invalid;
3098 Value *CallOperandVal = info.CallOperandVal;
3099 // If we don't have a value, we can't do a match,
3100 // but allow it at the lowest weight.
3101 if (CallOperandVal == NULL)
3103 Type *type = CallOperandVal->getType();
3104 // Look at the constraint type.
3105 switch (*constraint) {
3107 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3111 if (type->isIntegerTy())
3112 weight = CW_Register;
3115 if (type->isFloatTy())
3116 weight = CW_Register;
3118 case 'c': // $25 for indirect jumps
3119 case 'l': // lo register
3120 case 'x': // hilo register pair
3121 if (type->isIntegerTy())
3122 weight = CW_SpecificReg;
3124 case 'I': // signed 16 bit immediate
3125 case 'J': // integer zero
3126 case 'K': // unsigned 16 bit immediate
3127 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3128 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3129 case 'O': // signed 15 bit immediate (+- 16383)
3130 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3131 if (isa<ConstantInt>(CallOperandVal))
3132 weight = CW_Constant;
3141 /// Given a register class constraint, like 'r', if this corresponds directly
3142 /// to an LLVM register class, return a register of 0 and the register class
3144 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
3145 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
3147 if (Constraint.size() == 1) {
3148 switch (Constraint[0]) {
3149 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3150 case 'y': // Same as 'r'. Exists for compatibility.
3152 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3153 if (Subtarget->inMips16Mode())
3154 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3155 return std::make_pair(0U, &Mips::CPURegsRegClass);
3157 if (VT == MVT::i64 && !HasMips64)
3158 return std::make_pair(0U, &Mips::CPURegsRegClass);
3159 if (VT == MVT::i64 && HasMips64)
3160 return std::make_pair(0U, &Mips::CPU64RegsRegClass);
3161 // This will generate an error message
3162 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3165 return std::make_pair(0U, &Mips::FGR32RegClass);
3166 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
3167 if (Subtarget->isFP64bit())
3168 return std::make_pair(0U, &Mips::FGR64RegClass);
3169 return std::make_pair(0U, &Mips::AFGR64RegClass);
3172 case 'c': // register suitable for indirect jump
3174 return std::make_pair((unsigned)Mips::T9, &Mips::CPURegsRegClass);
3175 assert(VT == MVT::i64 && "Unexpected type.");
3176 return std::make_pair((unsigned)Mips::T9_64, &Mips::CPU64RegsRegClass);
3177 case 'l': // register suitable for indirect jump
3179 return std::make_pair((unsigned)Mips::LO, &Mips::HILORegClass);
3180 return std::make_pair((unsigned)Mips::LO64, &Mips::HILO64RegClass);
3181 case 'x': // register suitable for indirect jump
3182 // Fixme: Not triggering the use of both hi and low
3183 // This will generate an error message
3184 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3187 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3190 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3191 /// vector. If it is invalid, don't add anything to Ops.
3192 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3193 std::string &Constraint,
3194 std::vector<SDValue>&Ops,
3195 SelectionDAG &DAG) const {
3196 SDValue Result(0, 0);
3198 // Only support length 1 constraints for now.
3199 if (Constraint.length() > 1) return;
3201 char ConstraintLetter = Constraint[0];
3202 switch (ConstraintLetter) {
3203 default: break; // This will fall through to the generic implementation
3204 case 'I': // Signed 16 bit constant
3205 // If this fails, the parent routine will give an error
3206 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3207 EVT Type = Op.getValueType();
3208 int64_t Val = C->getSExtValue();
3209 if (isInt<16>(Val)) {
3210 Result = DAG.getTargetConstant(Val, Type);
3215 case 'J': // integer zero
3216 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3217 EVT Type = Op.getValueType();
3218 int64_t Val = C->getZExtValue();
3220 Result = DAG.getTargetConstant(0, Type);
3225 case 'K': // unsigned 16 bit immediate
3226 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3227 EVT Type = Op.getValueType();
3228 uint64_t Val = (uint64_t)C->getZExtValue();
3229 if (isUInt<16>(Val)) {
3230 Result = DAG.getTargetConstant(Val, Type);
3235 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3236 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3237 EVT Type = Op.getValueType();
3238 int64_t Val = C->getSExtValue();
3239 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3240 Result = DAG.getTargetConstant(Val, Type);
3245 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3246 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3247 EVT Type = Op.getValueType();
3248 int64_t Val = C->getSExtValue();
3249 if ((Val >= -65535) && (Val <= -1)) {
3250 Result = DAG.getTargetConstant(Val, Type);
3255 case 'O': // signed 15 bit immediate
3256 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3257 EVT Type = Op.getValueType();
3258 int64_t Val = C->getSExtValue();
3259 if ((isInt<15>(Val))) {
3260 Result = DAG.getTargetConstant(Val, Type);
3265 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3266 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3267 EVT Type = Op.getValueType();
3268 int64_t Val = C->getSExtValue();
3269 if ((Val <= 65535) && (Val >= 1)) {
3270 Result = DAG.getTargetConstant(Val, Type);
3277 if (Result.getNode()) {
3278 Ops.push_back(Result);
3282 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3286 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
3287 // No global is ever allowed as a base.
3292 case 0: // "r+i" or just "i", depending on HasBaseReg.
3295 if (!AM.HasBaseReg) // allow "r+i".
3297 return false; // disallow "r+r" or "r+r+i".
3306 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3307 // The Mips target isn't yet aware of offsets.
3311 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3313 bool IsMemset, bool ZeroMemset,
3315 MachineFunction &MF) const {
3316 if (Subtarget->hasMips64())
3322 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3323 if (VT != MVT::f32 && VT != MVT::f64)
3325 if (Imm.isNegZero())
3327 return Imm.isZero();
3330 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3332 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3334 return TargetLowering::getJumpTableEncoding();
3337 /// This function returns true if CallSym is a long double emulation routine.
3338 static bool isF128SoftLibCall(const char *CallSym) {
3339 const char *const LibCalls[] =
3340 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2",
3341 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi",
3342 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf",
3343 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2",
3344 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3",
3345 "__trunctfdf2", "__trunctfsf2", "__unordtf2",
3346 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl",
3347 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl",
3350 const char * const *End = LibCalls + array_lengthof(LibCalls);
3352 // Check that LibCalls is sorted alphabetically.
3353 MipsTargetLowering::LTStr Comp;
3356 for (const char * const *I = LibCalls; I < End - 1; ++I)
3357 assert(Comp(*I, *(I + 1)));
3360 return std::binary_search(LibCalls, End, CallSym, Comp);
3363 /// This function returns true if Ty is fp128 or i128 which was originally a
3365 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) {
3366 if (Ty->isFP128Ty())
3369 const ExternalSymbolSDNode *ES =
3370 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
3372 // If the Ty is i128 and the function being called is a long double emulation
3373 // routine, then the original type is f128.
3374 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
3377 MipsTargetLowering::MipsCC::MipsCC(CallingConv::ID CC, bool IsO32_,
3379 : CCInfo(Info), CallConv(CC), IsO32(IsO32_) {
3380 // Pre-allocate reserved argument area.
3381 CCInfo.AllocateStack(reservedArgArea(), 1);
3384 void MipsTargetLowering::MipsCC::
3385 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
3386 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode,
3387 std::vector<ArgListEntry> &FuncArgs) {
3388 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
3389 "CallingConv::Fast shouldn't be used for vararg functions.");
3391 unsigned NumOpnds = Args.size();
3392 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
3394 for (unsigned I = 0; I != NumOpnds; ++I) {
3395 MVT ArgVT = Args[I].VT;
3396 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3399 if (ArgFlags.isByVal()) {
3400 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3404 if (IsVarArg && !Args[I].IsFixed)
3405 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3407 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode,
3409 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo);
3414 dbgs() << "Call operand #" << I << " has unhandled type "
3415 << EVT(ArgVT).getEVTString();
3417 llvm_unreachable(0);
3422 void MipsTargetLowering::MipsCC::
3423 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args,
3424 bool IsSoftFloat, Function::const_arg_iterator FuncArg) {
3425 unsigned NumArgs = Args.size();
3426 llvm::CCAssignFn *FixedFn = fixedArgFn();
3427 unsigned CurArgIdx = 0;
3429 for (unsigned I = 0; I != NumArgs; ++I) {
3430 MVT ArgVT = Args[I].VT;
3431 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3432 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx);
3433 CurArgIdx = Args[I].OrigArgIndex;
3435 if (ArgFlags.isByVal()) {
3436 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3440 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat);
3442 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo))
3446 dbgs() << "Formal Arg #" << I << " has unhandled type "
3447 << EVT(ArgVT).getEVTString();
3449 llvm_unreachable(0);
3453 template<typename Ty>
3454 void MipsTargetLowering::MipsCC::
3455 analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat,
3456 const SDNode *CallNode, const Type *RetTy) const {
3459 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode))
3460 Fn = RetCC_F128Soft;
3464 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) {
3465 MVT VT = RetVals[I].VT;
3466 ISD::ArgFlagsTy Flags = RetVals[I].Flags;
3467 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat);
3469 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) {
3471 dbgs() << "Call result #" << I << " has unhandled type "
3472 << EVT(VT).getEVTString() << '\n';
3474 llvm_unreachable(0);
3479 void MipsTargetLowering::MipsCC::
3480 analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat,
3481 const SDNode *CallNode, const Type *RetTy) const {
3482 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy);
3485 void MipsTargetLowering::MipsCC::
3486 analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat,
3487 const Type *RetTy) const {
3488 analyzeReturn(Outs, IsSoftFloat, 0, RetTy);
3492 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3494 CCValAssign::LocInfo LocInfo,
3495 ISD::ArgFlagsTy ArgFlags) {
3496 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3498 struct ByValArgInfo ByVal;
3499 unsigned RegSize = regSize();
3500 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3501 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3504 if (useRegsForByval())
3505 allocateRegs(ByVal, ByValSize, Align);
3507 // Allocate space on caller's stack.
3508 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3510 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3512 ByValArgs.push_back(ByVal);
3515 unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const {
3516 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs);
3519 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
3520 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0;
3523 const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const {
3524 return IsO32 ? O32IntRegs : Mips64IntRegs;
3527 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
3528 if (CallConv == CallingConv::Fast)
3529 return CC_Mips_FastCC;
3531 return IsO32 ? CC_MipsO32 : CC_MipsN;
3534 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
3535 return IsO32 ? CC_MipsO32 : CC_MipsN_VarArg;
3538 const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const {
3539 return IsO32 ? O32IntRegs : Mips64DPRegs;
3542 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3545 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs();
3546 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs();
3547 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3548 "Byval argument's size and alignment should be a multiple of"
3551 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3553 // If Align > RegSize, the first arg register must be even.
3554 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3555 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3559 // Mark the registers allocated.
3560 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3561 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3562 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3565 MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy,
3566 const SDNode *CallNode,
3567 bool IsSoftFloat) const {
3568 if (IsSoftFloat || IsO32)
3571 // Check if the original type was fp128.
3572 if (originalTypeIsF128(OrigTy, CallNode)) {
3573 assert(VT == MVT::i64);
3580 void MipsTargetLowering::
3581 copyByValRegs(SDValue Chain, DebugLoc DL, std::vector<SDValue> &OutChains,
3582 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3583 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3584 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3585 MachineFunction &MF = DAG.getMachineFunction();
3586 MachineFrameInfo *MFI = MF.getFrameInfo();
3587 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3588 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3592 FrameObjOffset = (int)CC.reservedArgArea() -
3593 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3595 FrameObjOffset = ByVal.Address;
3597 // Create frame object.
3598 EVT PtrTy = getPointerTy();
3599 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3600 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3601 InVals.push_back(FIN);
3606 // Copy arg registers.
3607 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3608 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3610 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3611 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3612 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3613 unsigned Offset = I * CC.regSize();
3614 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3615 DAG.getConstant(Offset, PtrTy));
3616 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3617 StorePtr, MachinePointerInfo(FuncArg, Offset),
3619 OutChains.push_back(Store);
3623 // Copy byVal arg to registers and stack.
3624 void MipsTargetLowering::
3625 passByValArg(SDValue Chain, DebugLoc DL,
3626 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
3627 SmallVector<SDValue, 8> &MemOpChains, SDValue StackPtr,
3628 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3629 const MipsCC &CC, const ByValArgInfo &ByVal,
3630 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3631 unsigned ByValSize = Flags.getByValSize();
3632 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
3633 unsigned RegSize = CC.regSize();
3634 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
3635 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
3637 if (ByVal.NumRegs) {
3638 const uint16_t *ArgRegs = CC.intArgRegs();
3639 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
3642 // Copy words to registers.
3643 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
3644 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3645 DAG.getConstant(Offset, PtrTy));
3646 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3647 MachinePointerInfo(), false, false, false,
3649 MemOpChains.push_back(LoadVal.getValue(1));
3650 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3651 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3654 // Return if the struct has been fully copied.
3655 if (ByValSize == Offset)
3658 // Copy the remainder of the byval argument with sub-word loads and shifts.
3659 if (LeftoverBytes) {
3660 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
3661 "Size of the remainder should be smaller than RegSize.");
3664 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
3665 Offset < ByValSize; LoadSize /= 2) {
3666 unsigned RemSize = ByValSize - Offset;
3668 if (RemSize < LoadSize)
3672 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3673 DAG.getConstant(Offset, PtrTy));
3675 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
3676 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
3677 false, false, Alignment);
3678 MemOpChains.push_back(LoadVal.getValue(1));
3680 // Shift the loaded value.
3684 Shamt = TotalSizeLoaded;
3686 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
3688 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3689 DAG.getConstant(Shamt, MVT::i32));
3692 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3697 TotalSizeLoaded += LoadSize;
3698 Alignment = std::min(Alignment, LoadSize);
3701 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3702 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3707 // Copy remainder of byval arg to it with memcpy.
3708 unsigned MemCpySize = ByValSize - Offset;
3709 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3710 DAG.getConstant(Offset, PtrTy));
3711 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3712 DAG.getIntPtrConstant(ByVal.Address));
3713 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3714 DAG.getConstant(MemCpySize, PtrTy), Alignment,
3715 /*isVolatile=*/false, /*AlwaysInline=*/false,
3716 MachinePointerInfo(0), MachinePointerInfo(0));
3717 MemOpChains.push_back(Chain);
3721 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3722 const MipsCC &CC, SDValue Chain,
3723 DebugLoc DL, SelectionDAG &DAG) const {
3724 unsigned NumRegs = CC.numIntArgRegs();
3725 const uint16_t *ArgRegs = CC.intArgRegs();
3726 const CCState &CCInfo = CC.getCCInfo();
3727 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3728 unsigned RegSize = CC.regSize();
3729 MVT RegTy = MVT::getIntegerVT(RegSize * 8);
3730 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3731 MachineFunction &MF = DAG.getMachineFunction();
3732 MachineFrameInfo *MFI = MF.getFrameInfo();
3733 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3735 // Offset of the first variable argument from stack pointer.
3739 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3742 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3744 // Record the frame index of the first variable argument
3745 // which is a value necessary to VASTART.
3746 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3747 MipsFI->setVarArgsFrameIndex(FI);
3749 // Copy the integer registers that have not been used for argument passing
3750 // to the argument register save area. For O32, the save area is allocated
3751 // in the caller's stack frame, while for N32/64, it is allocated in the
3752 // callee's stack frame.
3753 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3754 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3755 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3756 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3757 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3758 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3759 MachinePointerInfo(), false, false, 0);
3760 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
3761 OutChains.push_back(Store);