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 // create MipsMAdd(u) node
465 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
467 SDValue MAdd = CurDAG->getNode(MultOpc, DL, MVT::Glue,
468 MultNode->getOperand(0),// Factor 0
469 MultNode->getOperand(1),// Factor 1
470 ADDCNode->getOperand(1),// Lo0
471 ADDENode->getOperand(1));// Hi0
473 // create CopyFromReg nodes
474 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, DL, Mips::LO, MVT::i32,
476 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), DL,
478 CopyFromLo.getValue(2));
480 // replace uses of adde and addc here
481 if (!SDValue(ADDCNode, 0).use_empty())
482 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
484 if (!SDValue(ADDENode, 0).use_empty())
485 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
491 // Transforms a subgraph in CurDAG if the following pattern is found:
492 // (addc Lo0, multLo), (sube Hi0, multHi),
494 // multHi/Lo: product of multiplication
495 // Lo0: initial value of Lo register
496 // Hi0: initial value of Hi register
497 // Return true if pattern matching was successful.
498 static bool selectMSUB(SDNode *SUBENode, SelectionDAG *CurDAG) {
499 // SUBENode's second operand must be a flag output of an SUBC node in order
500 // for the matching to be successful.
501 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
503 if (SUBCNode->getOpcode() != ISD::SUBC)
506 SDValue MultHi = SUBENode->getOperand(1);
507 SDValue MultLo = SUBCNode->getOperand(1);
508 SDNode *MultNode = MultHi.getNode();
509 unsigned MultOpc = MultHi.getOpcode();
511 // MultHi and MultLo must be generated by the same node,
512 if (MultLo.getNode() != MultNode)
515 // and it must be a multiplication.
516 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
519 // MultLo amd MultHi must be the first and second output of MultNode
521 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
524 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
525 // of the values of MultNode, in which case MultNode will be removed in later
527 // If there exist users other than SUBENode or SUBCNode, this function returns
528 // here, which will result in MultNode being mapped to a single MULT
529 // instruction node rather than a pair of MULT and MSUB instructions being
531 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
534 SDValue Chain = CurDAG->getEntryNode();
535 DebugLoc DL = SUBENode->getDebugLoc();
537 // create MipsSub(u) node
538 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
540 SDValue MSub = CurDAG->getNode(MultOpc, DL, MVT::Glue,
541 MultNode->getOperand(0),// Factor 0
542 MultNode->getOperand(1),// Factor 1
543 SUBCNode->getOperand(0),// Lo0
544 SUBENode->getOperand(0));// Hi0
546 // create CopyFromReg nodes
547 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, DL, Mips::LO, MVT::i32,
549 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), DL,
551 CopyFromLo.getValue(2));
553 // replace uses of sube and subc here
554 if (!SDValue(SUBCNode, 0).use_empty())
555 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
557 if (!SDValue(SUBENode, 0).use_empty())
558 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
563 static SDValue performADDECombine(SDNode *N, SelectionDAG &DAG,
564 TargetLowering::DAGCombinerInfo &DCI,
565 const MipsSubtarget *Subtarget) {
566 if (DCI.isBeforeLegalize())
569 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
571 return SDValue(N, 0);
576 static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
577 TargetLowering::DAGCombinerInfo &DCI,
578 const MipsSubtarget *Subtarget) {
579 if (DCI.isBeforeLegalize())
582 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
584 return SDValue(N, 0);
589 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
590 TargetLowering::DAGCombinerInfo &DCI,
591 const MipsSubtarget *Subtarget) {
592 if (DCI.isBeforeLegalizeOps())
595 EVT Ty = N->getValueType(0);
596 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
597 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
598 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
600 DebugLoc DL = N->getDebugLoc();
602 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
603 N->getOperand(0), N->getOperand(1));
604 SDValue InChain = DAG.getEntryNode();
605 SDValue InGlue = DivRem;
608 if (N->hasAnyUseOfValue(0)) {
609 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
611 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
612 InChain = CopyFromLo.getValue(1);
613 InGlue = CopyFromLo.getValue(2);
617 if (N->hasAnyUseOfValue(1)) {
618 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
620 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
626 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
628 default: llvm_unreachable("Unknown fp condition code!");
630 case ISD::SETOEQ: return Mips::FCOND_OEQ;
631 case ISD::SETUNE: return Mips::FCOND_UNE;
633 case ISD::SETOLT: return Mips::FCOND_OLT;
635 case ISD::SETOGT: return Mips::FCOND_OGT;
637 case ISD::SETOLE: return Mips::FCOND_OLE;
639 case ISD::SETOGE: return Mips::FCOND_OGE;
640 case ISD::SETULT: return Mips::FCOND_ULT;
641 case ISD::SETULE: return Mips::FCOND_ULE;
642 case ISD::SETUGT: return Mips::FCOND_UGT;
643 case ISD::SETUGE: return Mips::FCOND_UGE;
644 case ISD::SETUO: return Mips::FCOND_UN;
645 case ISD::SETO: return Mips::FCOND_OR;
647 case ISD::SETONE: return Mips::FCOND_ONE;
648 case ISD::SETUEQ: return Mips::FCOND_UEQ;
653 // Returns true if condition code has to be inverted.
654 static bool invertFPCondCode(Mips::CondCode CC) {
655 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
658 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
659 "Illegal Condition Code");
664 // Creates and returns an FPCmp node from a setcc node.
665 // Returns Op if setcc is not a floating point comparison.
666 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
667 // must be a SETCC node
668 if (Op.getOpcode() != ISD::SETCC)
671 SDValue LHS = Op.getOperand(0);
673 if (!LHS.getValueType().isFloatingPoint())
676 SDValue RHS = Op.getOperand(1);
677 DebugLoc DL = Op.getDebugLoc();
679 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
680 // node if necessary.
681 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
683 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
684 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
687 // Creates and returns a CMovFPT/F node.
688 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
689 SDValue False, DebugLoc DL) {
690 bool invert = invertFPCondCode((Mips::CondCode)
691 cast<ConstantSDNode>(Cond.getOperand(2))
694 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
695 True.getValueType(), True, False, Cond);
698 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
699 TargetLowering::DAGCombinerInfo &DCI,
700 const MipsSubtarget *Subtarget) {
701 if (DCI.isBeforeLegalizeOps())
704 SDValue SetCC = N->getOperand(0);
706 if ((SetCC.getOpcode() != ISD::SETCC) ||
707 !SetCC.getOperand(0).getValueType().isInteger())
710 SDValue False = N->getOperand(2);
711 EVT FalseTy = False.getValueType();
713 if (!FalseTy.isInteger())
716 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
718 if (!CN || CN->getZExtValue())
721 const DebugLoc DL = N->getDebugLoc();
722 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
723 SDValue True = N->getOperand(1);
725 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
726 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
728 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
731 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
732 TargetLowering::DAGCombinerInfo &DCI,
733 const MipsSubtarget *Subtarget) {
734 // Pattern match EXT.
735 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
736 // => ext $dst, $src, size, pos
737 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
740 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
741 unsigned ShiftRightOpc = ShiftRight.getOpcode();
743 // Op's first operand must be a shift right.
744 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
747 // The second operand of the shift must be an immediate.
749 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
752 uint64_t Pos = CN->getZExtValue();
753 uint64_t SMPos, SMSize;
755 // Op's second operand must be a shifted mask.
756 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
757 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
760 // Return if the shifted mask does not start at bit 0 or the sum of its size
761 // and Pos exceeds the word's size.
762 EVT ValTy = N->getValueType(0);
763 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
766 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
767 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
768 DAG.getConstant(SMSize, MVT::i32));
771 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
772 TargetLowering::DAGCombinerInfo &DCI,
773 const MipsSubtarget *Subtarget) {
774 // Pattern match INS.
775 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
776 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
777 // => ins $dst, $src, size, pos, $src1
778 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
781 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
782 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
785 // See if Op's first operand matches (and $src1 , mask0).
786 if (And0.getOpcode() != ISD::AND)
789 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
790 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
793 // See if Op's second operand matches (and (shl $src, pos), mask1).
794 if (And1.getOpcode() != ISD::AND)
797 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
798 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
801 // The shift masks must have the same position and size.
802 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
805 SDValue Shl = And1.getOperand(0);
806 if (Shl.getOpcode() != ISD::SHL)
809 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
812 unsigned Shamt = CN->getZExtValue();
814 // Return if the shift amount and the first bit position of mask are not the
816 EVT ValTy = N->getValueType(0);
817 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
820 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
821 DAG.getConstant(SMPos0, MVT::i32),
822 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
825 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
826 TargetLowering::DAGCombinerInfo &DCI,
827 const MipsSubtarget *Subtarget) {
828 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
830 if (DCI.isBeforeLegalizeOps())
833 SDValue Add = N->getOperand(1);
835 if (Add.getOpcode() != ISD::ADD)
838 SDValue Lo = Add.getOperand(1);
840 if ((Lo.getOpcode() != MipsISD::Lo) ||
841 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
844 EVT ValTy = N->getValueType(0);
845 DebugLoc DL = N->getDebugLoc();
847 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
849 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
852 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
854 SelectionDAG &DAG = DCI.DAG;
855 unsigned Opc = N->getOpcode();
860 return performADDECombine(N, DAG, DCI, Subtarget);
862 return performSUBECombine(N, DAG, DCI, Subtarget);
865 return performDivRemCombine(N, DAG, DCI, Subtarget);
867 return performSELECTCombine(N, DAG, DCI, Subtarget);
869 return performANDCombine(N, DAG, DCI, Subtarget);
871 return performORCombine(N, DAG, DCI, Subtarget);
873 return performADDCombine(N, DAG, DCI, Subtarget);
880 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
881 SmallVectorImpl<SDValue> &Results,
882 SelectionDAG &DAG) const {
883 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
885 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
886 Results.push_back(Res.getValue(I));
890 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
891 SmallVectorImpl<SDValue> &Results,
892 SelectionDAG &DAG) const {
893 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
895 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
896 Results.push_back(Res.getValue(I));
899 SDValue MipsTargetLowering::
900 LowerOperation(SDValue Op, SelectionDAG &DAG) const
902 switch (Op.getOpcode())
904 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
905 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
906 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
907 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
908 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
909 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
910 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
911 case ISD::SELECT: return lowerSELECT(Op, DAG);
912 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
913 case ISD::SETCC: return lowerSETCC(Op, DAG);
914 case ISD::VASTART: return lowerVASTART(Op, DAG);
915 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
916 case ISD::FABS: return lowerFABS(Op, DAG);
917 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
918 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
919 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
920 case ISD::MEMBARRIER: return lowerMEMBARRIER(Op, DAG);
921 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
922 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
923 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
924 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
925 case ISD::LOAD: return lowerLOAD(Op, DAG);
926 case ISD::STORE: return lowerSTORE(Op, DAG);
927 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
928 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
929 case ISD::ADD: return lowerADD(Op, DAG);
934 //===----------------------------------------------------------------------===//
935 // Lower helper functions
936 //===----------------------------------------------------------------------===//
938 // addLiveIn - This helper function adds the specified physical register to the
939 // MachineFunction as a live in value. It also creates a corresponding
940 // virtual register for it.
942 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
944 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
945 MF.getRegInfo().addLiveIn(PReg, VReg);
949 // Get fp branch code (not opcode) from condition code.
950 static Mips::FPBranchCode getFPBranchCodeFromCond(Mips::CondCode CC) {
951 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
952 return Mips::BRANCH_T;
954 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
955 "Invalid CondCode.");
957 return Mips::BRANCH_F;
961 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
962 MachineBasicBlock *BB) const {
963 switch (MI->getOpcode()) {
965 llvm_unreachable("Unexpected instr type to insert");
966 case Mips::ATOMIC_LOAD_ADD_I8:
967 case Mips::ATOMIC_LOAD_ADD_I8_P8:
968 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
969 case Mips::ATOMIC_LOAD_ADD_I16:
970 case Mips::ATOMIC_LOAD_ADD_I16_P8:
971 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
972 case Mips::ATOMIC_LOAD_ADD_I32:
973 case Mips::ATOMIC_LOAD_ADD_I32_P8:
974 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
975 case Mips::ATOMIC_LOAD_ADD_I64:
976 case Mips::ATOMIC_LOAD_ADD_I64_P8:
977 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
979 case Mips::ATOMIC_LOAD_AND_I8:
980 case Mips::ATOMIC_LOAD_AND_I8_P8:
981 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
982 case Mips::ATOMIC_LOAD_AND_I16:
983 case Mips::ATOMIC_LOAD_AND_I16_P8:
984 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
985 case Mips::ATOMIC_LOAD_AND_I32:
986 case Mips::ATOMIC_LOAD_AND_I32_P8:
987 return emitAtomicBinary(MI, BB, 4, Mips::AND);
988 case Mips::ATOMIC_LOAD_AND_I64:
989 case Mips::ATOMIC_LOAD_AND_I64_P8:
990 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
992 case Mips::ATOMIC_LOAD_OR_I8:
993 case Mips::ATOMIC_LOAD_OR_I8_P8:
994 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
995 case Mips::ATOMIC_LOAD_OR_I16:
996 case Mips::ATOMIC_LOAD_OR_I16_P8:
997 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
998 case Mips::ATOMIC_LOAD_OR_I32:
999 case Mips::ATOMIC_LOAD_OR_I32_P8:
1000 return emitAtomicBinary(MI, BB, 4, Mips::OR);
1001 case Mips::ATOMIC_LOAD_OR_I64:
1002 case Mips::ATOMIC_LOAD_OR_I64_P8:
1003 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
1005 case Mips::ATOMIC_LOAD_XOR_I8:
1006 case Mips::ATOMIC_LOAD_XOR_I8_P8:
1007 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
1008 case Mips::ATOMIC_LOAD_XOR_I16:
1009 case Mips::ATOMIC_LOAD_XOR_I16_P8:
1010 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
1011 case Mips::ATOMIC_LOAD_XOR_I32:
1012 case Mips::ATOMIC_LOAD_XOR_I32_P8:
1013 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
1014 case Mips::ATOMIC_LOAD_XOR_I64:
1015 case Mips::ATOMIC_LOAD_XOR_I64_P8:
1016 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
1018 case Mips::ATOMIC_LOAD_NAND_I8:
1019 case Mips::ATOMIC_LOAD_NAND_I8_P8:
1020 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
1021 case Mips::ATOMIC_LOAD_NAND_I16:
1022 case Mips::ATOMIC_LOAD_NAND_I16_P8:
1023 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
1024 case Mips::ATOMIC_LOAD_NAND_I32:
1025 case Mips::ATOMIC_LOAD_NAND_I32_P8:
1026 return emitAtomicBinary(MI, BB, 4, 0, true);
1027 case Mips::ATOMIC_LOAD_NAND_I64:
1028 case Mips::ATOMIC_LOAD_NAND_I64_P8:
1029 return emitAtomicBinary(MI, BB, 8, 0, true);
1031 case Mips::ATOMIC_LOAD_SUB_I8:
1032 case Mips::ATOMIC_LOAD_SUB_I8_P8:
1033 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
1034 case Mips::ATOMIC_LOAD_SUB_I16:
1035 case Mips::ATOMIC_LOAD_SUB_I16_P8:
1036 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
1037 case Mips::ATOMIC_LOAD_SUB_I32:
1038 case Mips::ATOMIC_LOAD_SUB_I32_P8:
1039 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
1040 case Mips::ATOMIC_LOAD_SUB_I64:
1041 case Mips::ATOMIC_LOAD_SUB_I64_P8:
1042 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
1044 case Mips::ATOMIC_SWAP_I8:
1045 case Mips::ATOMIC_SWAP_I8_P8:
1046 return emitAtomicBinaryPartword(MI, BB, 1, 0);
1047 case Mips::ATOMIC_SWAP_I16:
1048 case Mips::ATOMIC_SWAP_I16_P8:
1049 return emitAtomicBinaryPartword(MI, BB, 2, 0);
1050 case Mips::ATOMIC_SWAP_I32:
1051 case Mips::ATOMIC_SWAP_I32_P8:
1052 return emitAtomicBinary(MI, BB, 4, 0);
1053 case Mips::ATOMIC_SWAP_I64:
1054 case Mips::ATOMIC_SWAP_I64_P8:
1055 return emitAtomicBinary(MI, BB, 8, 0);
1057 case Mips::ATOMIC_CMP_SWAP_I8:
1058 case Mips::ATOMIC_CMP_SWAP_I8_P8:
1059 return emitAtomicCmpSwapPartword(MI, BB, 1);
1060 case Mips::ATOMIC_CMP_SWAP_I16:
1061 case Mips::ATOMIC_CMP_SWAP_I16_P8:
1062 return emitAtomicCmpSwapPartword(MI, BB, 2);
1063 case Mips::ATOMIC_CMP_SWAP_I32:
1064 case Mips::ATOMIC_CMP_SWAP_I32_P8:
1065 return emitAtomicCmpSwap(MI, BB, 4);
1066 case Mips::ATOMIC_CMP_SWAP_I64:
1067 case Mips::ATOMIC_CMP_SWAP_I64_P8:
1068 return emitAtomicCmpSwap(MI, BB, 8);
1072 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1073 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1075 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
1076 unsigned Size, unsigned BinOpcode,
1078 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1080 MachineFunction *MF = BB->getParent();
1081 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1082 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1083 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1084 DebugLoc DL = MI->getDebugLoc();
1085 unsigned LL, SC, AND, NOR, ZERO, BEQ;
1088 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1089 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1096 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1097 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1100 ZERO = Mips::ZERO_64;
1104 unsigned OldVal = MI->getOperand(0).getReg();
1105 unsigned Ptr = MI->getOperand(1).getReg();
1106 unsigned Incr = MI->getOperand(2).getReg();
1108 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1109 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1110 unsigned Success = RegInfo.createVirtualRegister(RC);
1112 // insert new blocks after the current block
1113 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1114 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1115 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1116 MachineFunction::iterator It = BB;
1118 MF->insert(It, loopMBB);
1119 MF->insert(It, exitMBB);
1121 // Transfer the remainder of BB and its successor edges to exitMBB.
1122 exitMBB->splice(exitMBB->begin(), BB,
1123 llvm::next(MachineBasicBlock::iterator(MI)),
1125 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1129 // fallthrough --> loopMBB
1130 BB->addSuccessor(loopMBB);
1131 loopMBB->addSuccessor(loopMBB);
1132 loopMBB->addSuccessor(exitMBB);
1135 // ll oldval, 0(ptr)
1136 // <binop> storeval, oldval, incr
1137 // sc success, storeval, 0(ptr)
1138 // beq success, $0, loopMBB
1140 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1142 // and andres, oldval, incr
1143 // nor storeval, $0, andres
1144 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1145 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1146 } else if (BinOpcode) {
1147 // <binop> storeval, oldval, incr
1148 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1152 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1153 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1155 MI->eraseFromParent(); // The instruction is gone now.
1161 MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI,
1162 MachineBasicBlock *BB,
1163 unsigned Size, unsigned BinOpcode,
1165 assert((Size == 1 || Size == 2) &&
1166 "Unsupported size for EmitAtomicBinaryPartial.");
1168 MachineFunction *MF = BB->getParent();
1169 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1170 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1171 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1172 DebugLoc DL = MI->getDebugLoc();
1173 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1174 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1176 unsigned Dest = MI->getOperand(0).getReg();
1177 unsigned Ptr = MI->getOperand(1).getReg();
1178 unsigned Incr = MI->getOperand(2).getReg();
1180 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1181 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1182 unsigned Mask = RegInfo.createVirtualRegister(RC);
1183 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1184 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1185 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1186 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1187 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1188 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1189 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1190 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1191 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1192 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1193 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1194 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1195 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1196 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1197 unsigned Success = RegInfo.createVirtualRegister(RC);
1199 // insert new blocks after the current block
1200 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1201 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1202 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1203 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1204 MachineFunction::iterator It = BB;
1206 MF->insert(It, loopMBB);
1207 MF->insert(It, sinkMBB);
1208 MF->insert(It, exitMBB);
1210 // Transfer the remainder of BB and its successor edges to exitMBB.
1211 exitMBB->splice(exitMBB->begin(), BB,
1212 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1213 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1215 BB->addSuccessor(loopMBB);
1216 loopMBB->addSuccessor(loopMBB);
1217 loopMBB->addSuccessor(sinkMBB);
1218 sinkMBB->addSuccessor(exitMBB);
1221 // addiu masklsb2,$0,-4 # 0xfffffffc
1222 // and alignedaddr,ptr,masklsb2
1223 // andi ptrlsb2,ptr,3
1224 // sll shiftamt,ptrlsb2,3
1225 // ori maskupper,$0,255 # 0xff
1226 // sll mask,maskupper,shiftamt
1227 // nor mask2,$0,mask
1228 // sll incr2,incr,shiftamt
1230 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1231 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1232 .addReg(Mips::ZERO).addImm(-4);
1233 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1234 .addReg(Ptr).addReg(MaskLSB2);
1235 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1236 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1237 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1238 .addReg(Mips::ZERO).addImm(MaskImm);
1239 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1240 .addReg(ShiftAmt).addReg(MaskUpper);
1241 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1242 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1244 // atomic.load.binop
1246 // ll oldval,0(alignedaddr)
1247 // binop binopres,oldval,incr2
1248 // and newval,binopres,mask
1249 // and maskedoldval0,oldval,mask2
1250 // or storeval,maskedoldval0,newval
1251 // sc success,storeval,0(alignedaddr)
1252 // beq success,$0,loopMBB
1256 // ll oldval,0(alignedaddr)
1257 // and newval,incr2,mask
1258 // and maskedoldval0,oldval,mask2
1259 // or storeval,maskedoldval0,newval
1260 // sc success,storeval,0(alignedaddr)
1261 // beq success,$0,loopMBB
1264 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1266 // and andres, oldval, incr2
1267 // nor binopres, $0, andres
1268 // and newval, binopres, mask
1269 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1270 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1271 .addReg(Mips::ZERO).addReg(AndRes);
1272 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1273 } else if (BinOpcode) {
1274 // <binop> binopres, oldval, incr2
1275 // and newval, binopres, mask
1276 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1277 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1278 } else {// atomic.swap
1279 // and newval, incr2, mask
1280 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1283 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1284 .addReg(OldVal).addReg(Mask2);
1285 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1286 .addReg(MaskedOldVal0).addReg(NewVal);
1287 BuildMI(BB, DL, TII->get(SC), Success)
1288 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1289 BuildMI(BB, DL, TII->get(Mips::BEQ))
1290 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1293 // and maskedoldval1,oldval,mask
1294 // srl srlres,maskedoldval1,shiftamt
1295 // sll sllres,srlres,24
1296 // sra dest,sllres,24
1298 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1300 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1301 .addReg(OldVal).addReg(Mask);
1302 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1303 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1304 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1305 .addReg(SrlRes).addImm(ShiftImm);
1306 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1307 .addReg(SllRes).addImm(ShiftImm);
1309 MI->eraseFromParent(); // The instruction is gone now.
1315 MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1316 MachineBasicBlock *BB,
1317 unsigned Size) const {
1318 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1320 MachineFunction *MF = BB->getParent();
1321 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1322 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1323 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1324 DebugLoc DL = MI->getDebugLoc();
1325 unsigned LL, SC, ZERO, BNE, BEQ;
1328 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1329 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1335 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1336 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1337 ZERO = Mips::ZERO_64;
1342 unsigned Dest = MI->getOperand(0).getReg();
1343 unsigned Ptr = MI->getOperand(1).getReg();
1344 unsigned OldVal = MI->getOperand(2).getReg();
1345 unsigned NewVal = MI->getOperand(3).getReg();
1347 unsigned Success = RegInfo.createVirtualRegister(RC);
1349 // insert new blocks after the current block
1350 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1351 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1352 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1353 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1354 MachineFunction::iterator It = BB;
1356 MF->insert(It, loop1MBB);
1357 MF->insert(It, loop2MBB);
1358 MF->insert(It, exitMBB);
1360 // Transfer the remainder of BB and its successor edges to exitMBB.
1361 exitMBB->splice(exitMBB->begin(), BB,
1362 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1363 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1367 // fallthrough --> loop1MBB
1368 BB->addSuccessor(loop1MBB);
1369 loop1MBB->addSuccessor(exitMBB);
1370 loop1MBB->addSuccessor(loop2MBB);
1371 loop2MBB->addSuccessor(loop1MBB);
1372 loop2MBB->addSuccessor(exitMBB);
1376 // bne dest, oldval, exitMBB
1378 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1379 BuildMI(BB, DL, TII->get(BNE))
1380 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1383 // sc success, newval, 0(ptr)
1384 // beq success, $0, loop1MBB
1386 BuildMI(BB, DL, TII->get(SC), Success)
1387 .addReg(NewVal).addReg(Ptr).addImm(0);
1388 BuildMI(BB, DL, TII->get(BEQ))
1389 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1391 MI->eraseFromParent(); // The instruction is gone now.
1397 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1398 MachineBasicBlock *BB,
1399 unsigned Size) const {
1400 assert((Size == 1 || Size == 2) &&
1401 "Unsupported size for EmitAtomicCmpSwapPartial.");
1403 MachineFunction *MF = BB->getParent();
1404 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1405 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1406 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1407 DebugLoc DL = MI->getDebugLoc();
1408 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1409 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1411 unsigned Dest = MI->getOperand(0).getReg();
1412 unsigned Ptr = MI->getOperand(1).getReg();
1413 unsigned CmpVal = MI->getOperand(2).getReg();
1414 unsigned NewVal = MI->getOperand(3).getReg();
1416 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1417 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1418 unsigned Mask = RegInfo.createVirtualRegister(RC);
1419 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1420 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1421 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1422 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1423 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1424 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1425 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1426 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1427 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1428 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1429 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1430 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1431 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1432 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1433 unsigned Success = RegInfo.createVirtualRegister(RC);
1435 // insert new blocks after the current block
1436 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1437 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1438 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1439 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1440 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1441 MachineFunction::iterator It = BB;
1443 MF->insert(It, loop1MBB);
1444 MF->insert(It, loop2MBB);
1445 MF->insert(It, sinkMBB);
1446 MF->insert(It, exitMBB);
1448 // Transfer the remainder of BB and its successor edges to exitMBB.
1449 exitMBB->splice(exitMBB->begin(), BB,
1450 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1451 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1453 BB->addSuccessor(loop1MBB);
1454 loop1MBB->addSuccessor(sinkMBB);
1455 loop1MBB->addSuccessor(loop2MBB);
1456 loop2MBB->addSuccessor(loop1MBB);
1457 loop2MBB->addSuccessor(sinkMBB);
1458 sinkMBB->addSuccessor(exitMBB);
1460 // FIXME: computation of newval2 can be moved to loop2MBB.
1462 // addiu masklsb2,$0,-4 # 0xfffffffc
1463 // and alignedaddr,ptr,masklsb2
1464 // andi ptrlsb2,ptr,3
1465 // sll shiftamt,ptrlsb2,3
1466 // ori maskupper,$0,255 # 0xff
1467 // sll mask,maskupper,shiftamt
1468 // nor mask2,$0,mask
1469 // andi maskedcmpval,cmpval,255
1470 // sll shiftedcmpval,maskedcmpval,shiftamt
1471 // andi maskednewval,newval,255
1472 // sll shiftednewval,maskednewval,shiftamt
1473 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1474 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1475 .addReg(Mips::ZERO).addImm(-4);
1476 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1477 .addReg(Ptr).addReg(MaskLSB2);
1478 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1479 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1480 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1481 .addReg(Mips::ZERO).addImm(MaskImm);
1482 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1483 .addReg(ShiftAmt).addReg(MaskUpper);
1484 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1485 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1486 .addReg(CmpVal).addImm(MaskImm);
1487 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1488 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1489 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1490 .addReg(NewVal).addImm(MaskImm);
1491 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1492 .addReg(ShiftAmt).addReg(MaskedNewVal);
1495 // ll oldval,0(alginedaddr)
1496 // and maskedoldval0,oldval,mask
1497 // bne maskedoldval0,shiftedcmpval,sinkMBB
1499 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1500 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1501 .addReg(OldVal).addReg(Mask);
1502 BuildMI(BB, DL, TII->get(Mips::BNE))
1503 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1506 // and maskedoldval1,oldval,mask2
1507 // or storeval,maskedoldval1,shiftednewval
1508 // sc success,storeval,0(alignedaddr)
1509 // beq success,$0,loop1MBB
1511 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1512 .addReg(OldVal).addReg(Mask2);
1513 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1514 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1515 BuildMI(BB, DL, TII->get(SC), Success)
1516 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1517 BuildMI(BB, DL, TII->get(Mips::BEQ))
1518 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1521 // srl srlres,maskedoldval0,shiftamt
1522 // sll sllres,srlres,24
1523 // sra dest,sllres,24
1525 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1527 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1528 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1529 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes)
1530 .addReg(SrlRes).addImm(ShiftImm);
1531 BuildMI(BB, DL, TII->get(Mips::SRA), Dest)
1532 .addReg(SllRes).addImm(ShiftImm);
1534 MI->eraseFromParent(); // The instruction is gone now.
1539 //===----------------------------------------------------------------------===//
1540 // Misc Lower Operation implementation
1541 //===----------------------------------------------------------------------===//
1542 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1543 SDValue Chain = Op.getOperand(0);
1544 SDValue Table = Op.getOperand(1);
1545 SDValue Index = Op.getOperand(2);
1546 DebugLoc DL = Op.getDebugLoc();
1547 EVT PTy = getPointerTy();
1548 unsigned EntrySize =
1549 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout());
1551 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1552 DAG.getConstant(EntrySize, PTy));
1553 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1555 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1556 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1557 MachinePointerInfo::getJumpTable(), MemVT, false, false,
1559 Chain = Addr.getValue(1);
1561 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) {
1562 // For PIC, the sequence is:
1563 // BRIND(load(Jumptable + index) + RelocBase)
1564 // RelocBase can be JumpTable, GOT or some sort of global base.
1565 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1566 getPICJumpTableRelocBase(Table, DAG));
1569 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1572 SDValue MipsTargetLowering::
1573 lowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1575 // The first operand is the chain, the second is the condition, the third is
1576 // the block to branch to if the condition is true.
1577 SDValue Chain = Op.getOperand(0);
1578 SDValue Dest = Op.getOperand(2);
1579 DebugLoc DL = Op.getDebugLoc();
1581 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1583 // Return if flag is not set by a floating point comparison.
1584 if (CondRes.getOpcode() != MipsISD::FPCmp)
1587 SDValue CCNode = CondRes.getOperand(2);
1589 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1590 SDValue BrCode = DAG.getConstant(getFPBranchCodeFromCond(CC), MVT::i32);
1592 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1596 SDValue MipsTargetLowering::
1597 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1599 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1601 // Return if flag is not set by a floating point comparison.
1602 if (Cond.getOpcode() != MipsISD::FPCmp)
1605 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1609 SDValue MipsTargetLowering::
1610 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1612 DebugLoc DL = Op.getDebugLoc();
1613 EVT Ty = Op.getOperand(0).getValueType();
1614 SDValue Cond = DAG.getNode(ISD::SETCC, DL, getSetCCResultType(Ty),
1615 Op.getOperand(0), Op.getOperand(1),
1618 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1622 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1623 SDValue Cond = createFPCmp(DAG, Op);
1625 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1626 "Floating point operand expected.");
1628 SDValue True = DAG.getConstant(1, MVT::i32);
1629 SDValue False = DAG.getConstant(0, MVT::i32);
1631 return createCMovFP(DAG, Cond, True, False, Op.getDebugLoc());
1634 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1635 SelectionDAG &DAG) const {
1636 // FIXME there isn't actually debug info here
1637 DebugLoc DL = Op.getDebugLoc();
1638 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1640 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1641 const MipsTargetObjectFile &TLOF =
1642 (const MipsTargetObjectFile&)getObjFileLowering();
1644 // %gp_rel relocation
1645 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1646 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0,
1648 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL,
1649 DAG.getVTList(MVT::i32), &GA, 1);
1650 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1651 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode);
1654 // %hi/%lo relocation
1655 return getAddrNonPIC(Op, DAG);
1658 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1659 return getAddrLocal(Op, DAG, HasMips64);
1662 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16,
1663 MipsII::MO_GOT_LO16);
1665 return getAddrGlobal(Op, DAG,
1666 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
1669 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1670 SelectionDAG &DAG) const {
1671 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1672 return getAddrNonPIC(Op, DAG);
1674 return getAddrLocal(Op, DAG, HasMips64);
1677 SDValue MipsTargetLowering::
1678 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1680 // If the relocation model is PIC, use the General Dynamic TLS Model or
1681 // Local Dynamic TLS model, otherwise use the Initial Exec or
1682 // Local Exec TLS Model.
1684 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1685 DebugLoc DL = GA->getDebugLoc();
1686 const GlobalValue *GV = GA->getGlobal();
1687 EVT PtrVT = getPointerTy();
1689 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1691 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1692 // General Dynamic and Local Dynamic TLS Model.
1693 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1696 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1697 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1698 getGlobalReg(DAG, PtrVT), TGA);
1699 unsigned PtrSize = PtrVT.getSizeInBits();
1700 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1702 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1706 Entry.Node = Argument;
1708 Args.push_back(Entry);
1710 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
1711 false, false, false, false, 0, CallingConv::C,
1712 /*IsTailCall=*/false, /*doesNotRet=*/false,
1713 /*isReturnValueUsed=*/true,
1714 TlsGetAddr, Args, DAG, DL);
1715 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1717 SDValue Ret = CallResult.first;
1719 if (model != TLSModel::LocalDynamic)
1722 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1723 MipsII::MO_DTPREL_HI);
1724 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1725 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1726 MipsII::MO_DTPREL_LO);
1727 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1728 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1729 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1733 if (model == TLSModel::InitialExec) {
1734 // Initial Exec TLS Model
1735 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1736 MipsII::MO_GOTTPREL);
1737 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1739 Offset = DAG.getLoad(PtrVT, DL,
1740 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1741 false, false, false, 0);
1743 // Local Exec TLS Model
1744 assert(model == TLSModel::LocalExec);
1745 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1746 MipsII::MO_TPREL_HI);
1747 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1748 MipsII::MO_TPREL_LO);
1749 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1750 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1751 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1754 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1755 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1758 SDValue MipsTargetLowering::
1759 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1761 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1762 return getAddrNonPIC(Op, DAG);
1764 return getAddrLocal(Op, DAG, HasMips64);
1767 SDValue MipsTargetLowering::
1768 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1770 // gp_rel relocation
1771 // FIXME: we should reference the constant pool using small data sections,
1772 // but the asm printer currently doesn't support this feature without
1773 // hacking it. This feature should come soon so we can uncomment the
1775 //if (IsInSmallSection(C->getType())) {
1776 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1777 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1778 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1780 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1781 return getAddrNonPIC(Op, DAG);
1783 return getAddrLocal(Op, DAG, HasMips64);
1786 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1787 MachineFunction &MF = DAG.getMachineFunction();
1788 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1790 DebugLoc DL = Op.getDebugLoc();
1791 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1794 // vastart just stores the address of the VarArgsFrameIndex slot into the
1795 // memory location argument.
1796 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1797 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1798 MachinePointerInfo(SV), false, false, 0);
1801 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1802 EVT TyX = Op.getOperand(0).getValueType();
1803 EVT TyY = Op.getOperand(1).getValueType();
1804 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1805 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1806 DebugLoc DL = Op.getDebugLoc();
1809 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1811 SDValue X = (TyX == MVT::f32) ?
1812 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1813 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1815 SDValue Y = (TyY == MVT::f32) ?
1816 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1817 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1821 // ext E, Y, 31, 1 ; extract bit31 of Y
1822 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1823 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1824 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1827 // srl SrlX, SllX, 1
1829 // sll SllY, SrlX, 31
1830 // or Or, SrlX, SllY
1831 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1832 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1833 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1834 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1835 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1838 if (TyX == MVT::f32)
1839 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1841 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1842 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1843 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1846 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1847 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1848 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1849 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1850 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1851 DebugLoc DL = Op.getDebugLoc();
1853 // Bitcast to integer nodes.
1854 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1855 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1858 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1859 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1860 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1861 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
1863 if (WidthX > WidthY)
1864 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1865 else if (WidthY > WidthX)
1866 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1868 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1869 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
1870 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1873 // (d)sll SllX, X, 1
1874 // (d)srl SrlX, SllX, 1
1875 // (d)srl SrlY, Y, width(Y)-1
1876 // (d)sll SllY, SrlX, width(Y)-1
1877 // or Or, SrlX, SllY
1878 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1879 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1880 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1881 DAG.getConstant(WidthY - 1, MVT::i32));
1883 if (WidthX > WidthY)
1884 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1885 else if (WidthY > WidthX)
1886 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1888 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1889 DAG.getConstant(WidthX - 1, MVT::i32));
1890 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
1891 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
1895 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1896 if (Subtarget->hasMips64())
1897 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
1899 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
1902 static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1903 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1904 DebugLoc DL = Op.getDebugLoc();
1906 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1908 SDValue X = (Op.getValueType() == MVT::f32) ?
1909 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1910 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1915 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
1916 DAG.getRegister(Mips::ZERO, MVT::i32),
1917 DAG.getConstant(31, MVT::i32), Const1, X);
1919 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1920 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1923 if (Op.getValueType() == MVT::f32)
1924 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
1926 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1927 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
1928 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1931 static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1932 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
1933 DebugLoc DL = Op.getDebugLoc();
1935 // Bitcast to integer node.
1936 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
1940 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
1941 DAG.getRegister(Mips::ZERO_64, MVT::i64),
1942 DAG.getConstant(63, MVT::i32), Const1, X);
1944 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
1945 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
1948 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
1952 MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const {
1953 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
1954 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2());
1956 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2());
1959 SDValue MipsTargetLowering::
1960 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1962 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1963 "Frame address can only be determined for current frame.");
1965 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1966 MFI->setFrameAddressIsTaken(true);
1967 EVT VT = Op.getValueType();
1968 DebugLoc DL = Op.getDebugLoc();
1969 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1970 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1974 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
1975 SelectionDAG &DAG) const {
1977 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1978 "Return address can be determined only for current frame.");
1980 MachineFunction &MF = DAG.getMachineFunction();
1981 MachineFrameInfo *MFI = MF.getFrameInfo();
1982 MVT VT = Op.getSimpleValueType();
1983 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
1984 MFI->setReturnAddressIsTaken(true);
1986 // Return RA, which contains the return address. Mark it an implicit live-in.
1987 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
1988 return DAG.getCopyFromReg(DAG.getEntryNode(), Op.getDebugLoc(), Reg, VT);
1991 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
1992 // generated from __builtin_eh_return (offset, handler)
1993 // The effect of this is to adjust the stack pointer by "offset"
1994 // and then branch to "handler".
1995 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
1997 MachineFunction &MF = DAG.getMachineFunction();
1998 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2000 MipsFI->setCallsEhReturn();
2001 SDValue Chain = Op.getOperand(0);
2002 SDValue Offset = Op.getOperand(1);
2003 SDValue Handler = Op.getOperand(2);
2004 DebugLoc DL = Op.getDebugLoc();
2005 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2007 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2008 // EH_RETURN nodes, so that instructions are emitted back-to-back.
2009 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1;
2010 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0;
2011 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2012 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2013 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2014 DAG.getRegister(OffsetReg, Ty),
2015 DAG.getRegister(AddrReg, getPointerTy()),
2019 // TODO: set SType according to the desired memory barrier behavior.
2021 MipsTargetLowering::lowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const {
2023 DebugLoc DL = Op.getDebugLoc();
2024 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2025 DAG.getConstant(SType, MVT::i32));
2028 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
2029 SelectionDAG &DAG) const {
2030 // FIXME: Need pseudo-fence for 'singlethread' fences
2031 // FIXME: Set SType for weaker fences where supported/appropriate.
2033 DebugLoc DL = Op.getDebugLoc();
2034 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2035 DAG.getConstant(SType, MVT::i32));
2038 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
2039 SelectionDAG &DAG) const {
2040 DebugLoc DL = Op.getDebugLoc();
2041 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2042 SDValue Shamt = Op.getOperand(2);
2045 // lo = (shl lo, shamt)
2046 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2049 // hi = (shl lo, shamt[4:0])
2050 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2051 DAG.getConstant(-1, MVT::i32));
2052 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
2053 DAG.getConstant(1, MVT::i32));
2054 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
2056 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
2057 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2058 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
2059 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2060 DAG.getConstant(0x20, MVT::i32));
2061 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2062 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
2063 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
2065 SDValue Ops[2] = {Lo, Hi};
2066 return DAG.getMergeValues(Ops, 2, DL);
2069 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2071 DebugLoc DL = Op.getDebugLoc();
2072 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2073 SDValue Shamt = Op.getOperand(2);
2076 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2078 // hi = (sra hi, shamt)
2080 // hi = (srl hi, shamt)
2083 // lo = (sra hi, shamt[4:0])
2084 // hi = (sra hi, 31)
2086 // lo = (srl hi, shamt[4:0])
2088 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2089 DAG.getConstant(-1, MVT::i32));
2090 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
2091 DAG.getConstant(1, MVT::i32));
2092 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
2093 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
2094 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2095 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2097 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2098 DAG.getConstant(0x20, MVT::i32));
2099 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
2100 DAG.getConstant(31, MVT::i32));
2101 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
2102 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2103 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
2106 SDValue Ops[2] = {Lo, Hi};
2107 return DAG.getMergeValues(Ops, 2, DL);
2110 static SDValue CreateLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2111 SDValue Chain, SDValue Src, unsigned Offset) {
2112 SDValue Ptr = LD->getBasePtr();
2113 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2114 EVT BasePtrVT = Ptr.getValueType();
2115 DebugLoc DL = LD->getDebugLoc();
2116 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2119 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2120 DAG.getConstant(Offset, BasePtrVT));
2122 SDValue Ops[] = { Chain, Ptr, Src };
2123 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2124 LD->getMemOperand());
2127 // Expand an unaligned 32 or 64-bit integer load node.
2128 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2129 LoadSDNode *LD = cast<LoadSDNode>(Op);
2130 EVT MemVT = LD->getMemoryVT();
2132 // Return if load is aligned or if MemVT is neither i32 nor i64.
2133 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2134 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2137 bool IsLittle = Subtarget->isLittle();
2138 EVT VT = Op.getValueType();
2139 ISD::LoadExtType ExtType = LD->getExtensionType();
2140 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2142 assert((VT == MVT::i32) || (VT == MVT::i64));
2145 // (set dst, (i64 (load baseptr)))
2147 // (set tmp, (ldl (add baseptr, 7), undef))
2148 // (set dst, (ldr baseptr, tmp))
2149 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2150 SDValue LDL = CreateLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2152 return CreateLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2156 SDValue LWL = CreateLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2158 SDValue LWR = CreateLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2162 // (set dst, (i32 (load baseptr))) or
2163 // (set dst, (i64 (sextload baseptr))) or
2164 // (set dst, (i64 (extload baseptr)))
2166 // (set tmp, (lwl (add baseptr, 3), undef))
2167 // (set dst, (lwr baseptr, tmp))
2168 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2169 (ExtType == ISD::EXTLOAD))
2172 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2175 // (set dst, (i64 (zextload baseptr)))
2177 // (set tmp0, (lwl (add baseptr, 3), undef))
2178 // (set tmp1, (lwr baseptr, tmp0))
2179 // (set tmp2, (shl tmp1, 32))
2180 // (set dst, (srl tmp2, 32))
2181 DebugLoc DL = LD->getDebugLoc();
2182 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2183 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2184 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2185 SDValue Ops[] = { SRL, LWR.getValue(1) };
2186 return DAG.getMergeValues(Ops, 2, DL);
2189 static SDValue CreateStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2190 SDValue Chain, unsigned Offset) {
2191 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2192 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2193 DebugLoc DL = SD->getDebugLoc();
2194 SDVTList VTList = DAG.getVTList(MVT::Other);
2197 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2198 DAG.getConstant(Offset, BasePtrVT));
2200 SDValue Ops[] = { Chain, Value, Ptr };
2201 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2202 SD->getMemOperand());
2205 // Expand an unaligned 32 or 64-bit integer store node.
2206 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2207 StoreSDNode *SD = cast<StoreSDNode>(Op);
2208 EVT MemVT = SD->getMemoryVT();
2210 // Return if store is aligned or if MemVT is neither i32 nor i64.
2211 if ((SD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2212 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2215 bool IsLittle = Subtarget->isLittle();
2216 SDValue Value = SD->getValue(), Chain = SD->getChain();
2217 EVT VT = Value.getValueType();
2220 // (store val, baseptr) or
2221 // (truncstore val, baseptr)
2223 // (swl val, (add baseptr, 3))
2224 // (swr val, baseptr)
2225 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2226 SDValue SWL = CreateStoreLR(MipsISD::SWL, DAG, SD, Chain,
2228 return CreateStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2231 assert(VT == MVT::i64);
2234 // (store val, baseptr)
2236 // (sdl val, (add baseptr, 7))
2237 // (sdr val, baseptr)
2238 SDValue SDL = CreateStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2239 return CreateStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2242 // This function expands mips intrinsic nodes which have 64-bit input operands
2243 // or output values.
2245 // out64 = intrinsic-node in64
2247 // lo = copy (extract-element (in64, 0))
2248 // hi = copy (extract-element (in64, 1))
2249 // mips-specific-node
2252 // out64 = merge-values (v0, v1)
2254 static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG,
2255 unsigned Opc, bool HasI64In, bool HasI64Out) {
2256 DebugLoc DL = Op.getDebugLoc();
2257 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
2258 SDValue Chain = HasChainIn ? Op->getOperand(0) : DAG.getEntryNode();
2259 SmallVector<SDValue, 3> Ops;
2262 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2263 Op->getOperand(1 + HasChainIn),
2264 DAG.getConstant(0, MVT::i32));
2265 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2266 Op->getOperand(1 + HasChainIn),
2267 DAG.getConstant(1, MVT::i32));
2269 Chain = DAG.getCopyToReg(Chain, DL, Mips::LO, InLo, SDValue());
2270 Chain = DAG.getCopyToReg(Chain, DL, Mips::HI, InHi, Chain.getValue(1));
2272 Ops.push_back(Chain);
2273 Ops.append(Op->op_begin() + HasChainIn + 2, Op->op_end());
2274 Ops.push_back(Chain.getValue(1));
2276 Ops.push_back(Chain);
2277 Ops.append(Op->op_begin() + HasChainIn + 1, Op->op_end());
2281 return DAG.getNode(Opc, DL, Op->value_begin(), Op->getNumValues(),
2282 Ops.begin(), Ops.size());
2284 SDValue Intr = DAG.getNode(Opc, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2285 Ops.begin(), Ops.size());
2286 SDValue OutLo = DAG.getCopyFromReg(Intr.getValue(0), DL, Mips::LO, MVT::i32,
2288 SDValue OutHi = DAG.getCopyFromReg(OutLo.getValue(1), DL, Mips::HI, MVT::i32,
2290 SDValue Out = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, OutLo, OutHi);
2295 SDValue Vals[] = { Out, OutHi.getValue(1) };
2296 return DAG.getMergeValues(Vals, 2, DL);
2299 SDValue MipsTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
2300 SelectionDAG &DAG) const {
2301 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
2304 case Intrinsic::mips_shilo:
2305 return lowerDSPIntr(Op, DAG, MipsISD::SHILO, true, true);
2306 case Intrinsic::mips_dpau_h_qbl:
2307 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL, true, true);
2308 case Intrinsic::mips_dpau_h_qbr:
2309 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR, true, true);
2310 case Intrinsic::mips_dpsu_h_qbl:
2311 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL, true, true);
2312 case Intrinsic::mips_dpsu_h_qbr:
2313 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR, true, true);
2314 case Intrinsic::mips_dpa_w_ph:
2315 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH, true, true);
2316 case Intrinsic::mips_dps_w_ph:
2317 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH, true, true);
2318 case Intrinsic::mips_dpax_w_ph:
2319 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH, true, true);
2320 case Intrinsic::mips_dpsx_w_ph:
2321 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH, true, true);
2322 case Intrinsic::mips_mulsa_w_ph:
2323 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH, true, true);
2324 case Intrinsic::mips_mult:
2325 return lowerDSPIntr(Op, DAG, MipsISD::MULT, false, true);
2326 case Intrinsic::mips_multu:
2327 return lowerDSPIntr(Op, DAG, MipsISD::MULTU, false, true);
2328 case Intrinsic::mips_madd:
2329 return lowerDSPIntr(Op, DAG, MipsISD::MADD_DSP, true, true);
2330 case Intrinsic::mips_maddu:
2331 return lowerDSPIntr(Op, DAG, MipsISD::MADDU_DSP, true, true);
2332 case Intrinsic::mips_msub:
2333 return lowerDSPIntr(Op, DAG, MipsISD::MSUB_DSP, true, true);
2334 case Intrinsic::mips_msubu:
2335 return lowerDSPIntr(Op, DAG, MipsISD::MSUBU_DSP, true, true);
2339 SDValue MipsTargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op,
2340 SelectionDAG &DAG) const {
2341 switch (cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue()) {
2344 case Intrinsic::mips_extp:
2345 return lowerDSPIntr(Op, DAG, MipsISD::EXTP, true, false);
2346 case Intrinsic::mips_extpdp:
2347 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP, true, false);
2348 case Intrinsic::mips_extr_w:
2349 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W, true, false);
2350 case Intrinsic::mips_extr_r_w:
2351 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W, true, false);
2352 case Intrinsic::mips_extr_rs_w:
2353 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W, true, false);
2354 case Intrinsic::mips_extr_s_h:
2355 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H, true, false);
2356 case Intrinsic::mips_mthlip:
2357 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP, true, true);
2358 case Intrinsic::mips_mulsaq_s_w_ph:
2359 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH, true, true);
2360 case Intrinsic::mips_maq_s_w_phl:
2361 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL, true, true);
2362 case Intrinsic::mips_maq_s_w_phr:
2363 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR, true, true);
2364 case Intrinsic::mips_maq_sa_w_phl:
2365 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL, true, true);
2366 case Intrinsic::mips_maq_sa_w_phr:
2367 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR, true, true);
2368 case Intrinsic::mips_dpaq_s_w_ph:
2369 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH, true, true);
2370 case Intrinsic::mips_dpsq_s_w_ph:
2371 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH, true, true);
2372 case Intrinsic::mips_dpaq_sa_l_w:
2373 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W, true, true);
2374 case Intrinsic::mips_dpsq_sa_l_w:
2375 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W, true, true);
2376 case Intrinsic::mips_dpaqx_s_w_ph:
2377 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH, true, true);
2378 case Intrinsic::mips_dpaqx_sa_w_ph:
2379 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH, true, true);
2380 case Intrinsic::mips_dpsqx_s_w_ph:
2381 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH, true, true);
2382 case Intrinsic::mips_dpsqx_sa_w_ph:
2383 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH, true, true);
2387 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2388 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2389 || cast<ConstantSDNode>
2390 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2391 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2395 // (add (frameaddr 0), (frame_to_args_offset))
2396 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2397 // (add FrameObject, 0)
2398 // where FrameObject is a fixed StackObject with offset 0 which points to
2399 // the old stack pointer.
2400 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2401 EVT ValTy = Op->getValueType(0);
2402 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2403 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2404 return DAG.getNode(ISD::ADD, Op->getDebugLoc(), ValTy, InArgsAddr,
2405 DAG.getConstant(0, ValTy));
2408 //===----------------------------------------------------------------------===//
2409 // Calling Convention Implementation
2410 //===----------------------------------------------------------------------===//
2412 //===----------------------------------------------------------------------===//
2413 // TODO: Implement a generic logic using tblgen that can support this.
2414 // Mips O32 ABI rules:
2416 // i32 - Passed in A0, A1, A2, A3 and stack
2417 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2418 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2419 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2420 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2421 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2424 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2425 //===----------------------------------------------------------------------===//
2427 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2428 MVT LocVT, CCValAssign::LocInfo LocInfo,
2429 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2431 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2433 static const uint16_t IntRegs[] = {
2434 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2436 static const uint16_t F32Regs[] = {
2437 Mips::F12, Mips::F14
2439 static const uint16_t F64Regs[] = {
2443 // Do not process byval args here.
2444 if (ArgFlags.isByVal())
2447 // Promote i8 and i16
2448 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2450 if (ArgFlags.isSExt())
2451 LocInfo = CCValAssign::SExt;
2452 else if (ArgFlags.isZExt())
2453 LocInfo = CCValAssign::ZExt;
2455 LocInfo = CCValAssign::AExt;
2460 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2461 // is true: function is vararg, argument is 3rd or higher, there is previous
2462 // argument which is not f32 or f64.
2463 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2464 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2465 unsigned OrigAlign = ArgFlags.getOrigAlign();
2466 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2468 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2469 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2470 // If this is the first part of an i64 arg,
2471 // the allocated register must be either A0 or A2.
2472 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2473 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2475 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2476 // Allocate int register and shadow next int register. If first
2477 // available register is Mips::A1 or Mips::A3, shadow it too.
2478 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2479 if (Reg == Mips::A1 || Reg == Mips::A3)
2480 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2481 State.AllocateReg(IntRegs, IntRegsSize);
2483 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2484 // we are guaranteed to find an available float register
2485 if (ValVT == MVT::f32) {
2486 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2487 // Shadow int register
2488 State.AllocateReg(IntRegs, IntRegsSize);
2490 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2491 // Shadow int registers
2492 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2493 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2494 State.AllocateReg(IntRegs, IntRegsSize);
2495 State.AllocateReg(IntRegs, IntRegsSize);
2498 llvm_unreachable("Cannot handle this ValVT.");
2501 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2503 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2505 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2510 #include "MipsGenCallingConv.inc"
2512 //===----------------------------------------------------------------------===//
2513 // Call Calling Convention Implementation
2514 //===----------------------------------------------------------------------===//
2516 static const unsigned O32IntRegsSize = 4;
2518 // Return next O32 integer argument register.
2519 static unsigned getNextIntArgReg(unsigned Reg) {
2520 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2521 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2525 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2526 SDValue Chain, SDValue Arg, DebugLoc DL,
2527 bool IsTailCall, SelectionDAG &DAG) const {
2529 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2530 DAG.getIntPtrConstant(Offset));
2531 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2535 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2536 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2537 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2538 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2539 /*isVolatile=*/ true, false, 0);
2542 void MipsTargetLowering::
2543 getOpndList(SmallVectorImpl<SDValue> &Ops,
2544 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2545 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2546 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const {
2547 // Insert node "GP copy globalreg" before call to function.
2549 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2550 // in PIC mode) allow symbols to be resolved via lazy binding.
2551 // The lazy binding stub requires GP to point to the GOT.
2552 if (IsPICCall && !InternalLinkage) {
2553 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
2554 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2555 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2558 // Build a sequence of copy-to-reg nodes chained together with token
2559 // chain and flag operands which copy the outgoing args into registers.
2560 // The InFlag in necessary since all emitted instructions must be
2564 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2565 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2566 RegsToPass[i].second, InFlag);
2567 InFlag = Chain.getValue(1);
2570 // Add argument registers to the end of the list so that they are
2571 // known live into the call.
2572 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2573 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2574 RegsToPass[i].second.getValueType()));
2576 // Add a register mask operand representing the call-preserved registers.
2577 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2578 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv);
2579 assert(Mask && "Missing call preserved mask for calling convention");
2580 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2582 if (InFlag.getNode())
2583 Ops.push_back(InFlag);
2586 /// LowerCall - functions arguments are copied from virtual regs to
2587 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2589 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2590 SmallVectorImpl<SDValue> &InVals) const {
2591 SelectionDAG &DAG = CLI.DAG;
2592 DebugLoc &DL = CLI.DL;
2593 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2594 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2595 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2596 SDValue Chain = CLI.Chain;
2597 SDValue Callee = CLI.Callee;
2598 bool &IsTailCall = CLI.IsTailCall;
2599 CallingConv::ID CallConv = CLI.CallConv;
2600 bool IsVarArg = CLI.IsVarArg;
2602 MachineFunction &MF = DAG.getMachineFunction();
2603 MachineFrameInfo *MFI = MF.getFrameInfo();
2604 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2605 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2607 // Analyze operands of the call, assigning locations to each operand.
2608 SmallVector<CCValAssign, 16> ArgLocs;
2609 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2610 getTargetMachine(), ArgLocs, *DAG.getContext());
2611 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2613 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg,
2614 getTargetMachine().Options.UseSoftFloat,
2615 Callee.getNode(), CLI.Args);
2617 // Get a count of how many bytes are to be pushed on the stack.
2618 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2620 // Check if it's really possible to do a tail call.
2623 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2624 *MF.getInfo<MipsFunctionInfo>());
2629 // Chain is the output chain of the last Load/Store or CopyToReg node.
2630 // ByValChain is the output chain of the last Memcpy node created for copying
2631 // byval arguments to the stack.
2632 unsigned StackAlignment = TFL->getStackAlignment();
2633 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2634 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2637 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal);
2639 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL,
2640 IsN64 ? Mips::SP_64 : Mips::SP,
2643 // With EABI is it possible to have 16 args on registers.
2644 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2645 SmallVector<SDValue, 8> MemOpChains;
2646 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2648 // Walk the register/memloc assignments, inserting copies/loads.
2649 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2650 SDValue Arg = OutVals[i];
2651 CCValAssign &VA = ArgLocs[i];
2652 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2653 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2656 if (Flags.isByVal()) {
2657 assert(Flags.getByValSize() &&
2658 "ByVal args of size 0 should have been ignored by front-end.");
2659 assert(ByValArg != MipsCCInfo.byval_end());
2660 assert(!IsTailCall &&
2661 "Do not tail-call optimize if there is a byval argument.");
2662 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2663 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2668 // Promote the value if needed.
2669 switch (VA.getLocInfo()) {
2670 default: llvm_unreachable("Unknown loc info!");
2671 case CCValAssign::Full:
2672 if (VA.isRegLoc()) {
2673 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2674 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2675 (ValVT == MVT::i64 && LocVT == MVT::f64))
2676 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2677 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2678 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2679 Arg, DAG.getConstant(0, MVT::i32));
2680 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2681 Arg, DAG.getConstant(1, MVT::i32));
2682 if (!Subtarget->isLittle())
2684 unsigned LocRegLo = VA.getLocReg();
2685 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2686 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2687 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2692 case CCValAssign::SExt:
2693 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2695 case CCValAssign::ZExt:
2696 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2698 case CCValAssign::AExt:
2699 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2703 // Arguments that can be passed on register must be kept at
2704 // RegsToPass vector
2705 if (VA.isRegLoc()) {
2706 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2710 // Register can't get to this point...
2711 assert(VA.isMemLoc());
2713 // emit ISD::STORE whichs stores the
2714 // parameter value to a stack Location
2715 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2716 Chain, Arg, DL, IsTailCall, DAG));
2719 // Transform all store nodes into one single node because all store
2720 // nodes are independent of each other.
2721 if (!MemOpChains.empty())
2722 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2723 &MemOpChains[0], MemOpChains.size());
2725 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2726 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2727 // node so that legalize doesn't hack it.
2728 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2729 bool GlobalOrExternal = false, InternalLinkage = false;
2732 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2734 InternalLinkage = G->getGlobal()->hasInternalLinkage();
2736 if (InternalLinkage)
2737 Callee = getAddrLocal(Callee, DAG, HasMips64);
2739 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2740 MipsII::MO_CALL_LO16);
2742 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2744 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0,
2745 MipsII::MO_NO_FLAG);
2746 GlobalOrExternal = true;
2748 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2749 if (!IsN64 && !IsPIC) // !N64 && static
2750 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2751 MipsII::MO_NO_FLAG);
2753 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2754 MipsII::MO_CALL_LO16);
2756 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2758 GlobalOrExternal = true;
2761 SmallVector<SDValue, 8> Ops(1, Chain);
2762 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2764 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2765 CLI, Callee, Chain);
2768 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size());
2770 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size());
2771 SDValue InFlag = Chain.getValue(1);
2773 // Create the CALLSEQ_END node.
2774 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2775 DAG.getIntPtrConstant(0, true), InFlag);
2776 InFlag = Chain.getValue(1);
2778 // Handle result values, copying them out of physregs into vregs that we
2780 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg,
2781 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy);
2784 /// LowerCallResult - Lower the result values of a call into the
2785 /// appropriate copies out of appropriate physical registers.
2787 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2788 CallingConv::ID CallConv, bool IsVarArg,
2789 const SmallVectorImpl<ISD::InputArg> &Ins,
2790 DebugLoc DL, SelectionDAG &DAG,
2791 SmallVectorImpl<SDValue> &InVals,
2792 const SDNode *CallNode,
2793 const Type *RetTy) const {
2794 // Assign locations to each value returned by this call.
2795 SmallVector<CCValAssign, 16> RVLocs;
2796 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2797 getTargetMachine(), RVLocs, *DAG.getContext());
2798 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2800 MipsCCInfo.analyzeCallResult(Ins, getTargetMachine().Options.UseSoftFloat,
2803 // Copy all of the result registers out of their specified physreg.
2804 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2805 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2806 RVLocs[i].getLocVT(), InFlag);
2807 Chain = Val.getValue(1);
2808 InFlag = Val.getValue(2);
2810 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
2811 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val);
2813 InVals.push_back(Val);
2819 //===----------------------------------------------------------------------===//
2820 // Formal Arguments Calling Convention Implementation
2821 //===----------------------------------------------------------------------===//
2822 /// LowerFormalArguments - transform physical registers into virtual registers
2823 /// and generate load operations for arguments places on the stack.
2825 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2826 CallingConv::ID CallConv,
2828 const SmallVectorImpl<ISD::InputArg> &Ins,
2829 DebugLoc DL, SelectionDAG &DAG,
2830 SmallVectorImpl<SDValue> &InVals)
2832 MachineFunction &MF = DAG.getMachineFunction();
2833 MachineFrameInfo *MFI = MF.getFrameInfo();
2834 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2836 MipsFI->setVarArgsFrameIndex(0);
2838 // Used with vargs to acumulate store chains.
2839 std::vector<SDValue> OutChains;
2841 // Assign locations to all of the incoming arguments.
2842 SmallVector<CCValAssign, 16> ArgLocs;
2843 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2844 getTargetMachine(), ArgLocs, *DAG.getContext());
2845 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
2846 Function::const_arg_iterator FuncArg =
2847 DAG.getMachineFunction().getFunction()->arg_begin();
2848 bool UseSoftFloat = getTargetMachine().Options.UseSoftFloat;
2850 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg);
2851 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2852 MipsCCInfo.hasByValArg());
2854 unsigned CurArgIdx = 0;
2855 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2857 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2858 CCValAssign &VA = ArgLocs[i];
2859 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
2860 CurArgIdx = Ins[i].OrigArgIndex;
2861 EVT ValVT = VA.getValVT();
2862 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2863 bool IsRegLoc = VA.isRegLoc();
2865 if (Flags.isByVal()) {
2866 assert(Flags.getByValSize() &&
2867 "ByVal args of size 0 should have been ignored by front-end.");
2868 assert(ByValArg != MipsCCInfo.byval_end());
2869 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2870 MipsCCInfo, *ByValArg);
2875 // Arguments stored on registers
2877 EVT RegVT = VA.getLocVT();
2878 unsigned ArgReg = VA.getLocReg();
2879 const TargetRegisterClass *RC;
2881 if (RegVT == MVT::i32)
2882 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass :
2883 &Mips::CPURegsRegClass;
2884 else if (RegVT == MVT::i64)
2885 RC = &Mips::CPU64RegsRegClass;
2886 else if (RegVT == MVT::f32)
2887 RC = &Mips::FGR32RegClass;
2888 else if (RegVT == MVT::f64)
2889 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass;
2891 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2893 // Transform the arguments stored on
2894 // physical registers into virtual ones
2895 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2896 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2898 // If this is an 8 or 16-bit value, it has been passed promoted
2899 // to 32 bits. Insert an assert[sz]ext to capture this, then
2900 // truncate to the right size.
2901 if (VA.getLocInfo() != CCValAssign::Full) {
2902 unsigned Opcode = 0;
2903 if (VA.getLocInfo() == CCValAssign::SExt)
2904 Opcode = ISD::AssertSext;
2905 else if (VA.getLocInfo() == CCValAssign::ZExt)
2906 Opcode = ISD::AssertZext;
2908 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue,
2909 DAG.getValueType(ValVT));
2910 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue);
2913 // Handle floating point arguments passed in integer registers and
2914 // long double arguments passed in floating point registers.
2915 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2916 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
2917 (RegVT == MVT::f64 && ValVT == MVT::i64))
2918 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
2919 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2920 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
2921 getNextIntArgReg(ArgReg), RC);
2922 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
2923 if (!Subtarget->isLittle())
2924 std::swap(ArgValue, ArgValue2);
2925 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
2926 ArgValue, ArgValue2);
2929 InVals.push_back(ArgValue);
2930 } else { // VA.isRegLoc()
2933 assert(VA.isMemLoc());
2935 // The stack pointer offset is relative to the caller stack frame.
2936 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2937 VA.getLocMemOffset(), true);
2939 // Create load nodes to retrieve arguments from the stack
2940 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2941 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN,
2942 MachinePointerInfo::getFixedStack(FI),
2943 false, false, false, 0));
2947 // The mips ABIs for returning structs by value requires that we copy
2948 // the sret argument into $v0 for the return. Save the argument into
2949 // a virtual register so that we can access it from the return points.
2950 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2951 unsigned Reg = MipsFI->getSRetReturnReg();
2953 Reg = MF.getRegInfo().
2954 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
2955 MipsFI->setSRetReturnReg(Reg);
2957 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]);
2958 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
2962 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG);
2964 // All stores are grouped in one node to allow the matching between
2965 // the size of Ins and InVals. This only happens when on varg functions
2966 if (!OutChains.empty()) {
2967 OutChains.push_back(Chain);
2968 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
2969 &OutChains[0], OutChains.size());
2975 //===----------------------------------------------------------------------===//
2976 // Return Value Calling Convention Implementation
2977 //===----------------------------------------------------------------------===//
2980 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2981 MachineFunction &MF, bool IsVarArg,
2982 const SmallVectorImpl<ISD::OutputArg> &Outs,
2983 LLVMContext &Context) const {
2984 SmallVector<CCValAssign, 16> RVLocs;
2985 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(),
2987 return CCInfo.CheckReturn(Outs, RetCC_Mips);
2991 MipsTargetLowering::LowerReturn(SDValue Chain,
2992 CallingConv::ID CallConv, bool IsVarArg,
2993 const SmallVectorImpl<ISD::OutputArg> &Outs,
2994 const SmallVectorImpl<SDValue> &OutVals,
2995 DebugLoc DL, SelectionDAG &DAG) const {
2996 // CCValAssign - represent the assignment of
2997 // the return value to a location
2998 SmallVector<CCValAssign, 16> RVLocs;
2999 MachineFunction &MF = DAG.getMachineFunction();
3001 // CCState - Info about the registers and stack slot.
3002 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs,
3004 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
3006 // Analyze return values.
3007 MipsCCInfo.analyzeReturn(Outs, getTargetMachine().Options.UseSoftFloat,
3008 MF.getFunction()->getReturnType());
3011 SmallVector<SDValue, 4> RetOps(1, Chain);
3013 // Copy the result values into the output registers.
3014 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3015 SDValue Val = OutVals[i];
3016 CCValAssign &VA = RVLocs[i];
3017 assert(VA.isRegLoc() && "Can only return in registers!");
3019 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT())
3020 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val);
3022 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
3024 // Guarantee that all emitted copies are stuck together with flags.
3025 Flag = Chain.getValue(1);
3026 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3029 // The mips ABIs for returning structs by value requires that we copy
3030 // the sret argument into $v0 for the return. We saved the argument into
3031 // a virtual register in the entry block, so now we copy the value out
3033 if (MF.getFunction()->hasStructRetAttr()) {
3034 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3035 unsigned Reg = MipsFI->getSRetReturnReg();
3038 llvm_unreachable("sret virtual register not created in the entry block");
3039 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
3040 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
3042 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
3043 Flag = Chain.getValue(1);
3044 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
3047 RetOps[0] = Chain; // Update chain.
3049 // Add the flag if we have it.
3051 RetOps.push_back(Flag);
3053 // Return on Mips is always a "jr $ra"
3054 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size());
3057 //===----------------------------------------------------------------------===//
3058 // Mips Inline Assembly Support
3059 //===----------------------------------------------------------------------===//
3061 /// getConstraintType - Given a constraint letter, return the type of
3062 /// constraint it is for this target.
3063 MipsTargetLowering::ConstraintType MipsTargetLowering::
3064 getConstraintType(const std::string &Constraint) const
3066 // Mips specific constrainy
3067 // GCC config/mips/constraints.md
3069 // 'd' : An address register. Equivalent to r
3070 // unless generating MIPS16 code.
3071 // 'y' : Equivalent to r; retained for
3072 // backwards compatibility.
3073 // 'c' : A register suitable for use in an indirect
3074 // jump. This will always be $25 for -mabicalls.
3075 // 'l' : The lo register. 1 word storage.
3076 // 'x' : The hilo register pair. Double word storage.
3077 if (Constraint.size() == 1) {
3078 switch (Constraint[0]) {
3086 return C_RegisterClass;
3091 return TargetLowering::getConstraintType(Constraint);
3094 /// Examine constraint type and operand type and determine a weight value.
3095 /// This object must already have been set up with the operand type
3096 /// and the current alternative constraint selected.
3097 TargetLowering::ConstraintWeight
3098 MipsTargetLowering::getSingleConstraintMatchWeight(
3099 AsmOperandInfo &info, const char *constraint) const {
3100 ConstraintWeight weight = CW_Invalid;
3101 Value *CallOperandVal = info.CallOperandVal;
3102 // If we don't have a value, we can't do a match,
3103 // but allow it at the lowest weight.
3104 if (CallOperandVal == NULL)
3106 Type *type = CallOperandVal->getType();
3107 // Look at the constraint type.
3108 switch (*constraint) {
3110 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3114 if (type->isIntegerTy())
3115 weight = CW_Register;
3118 if (type->isFloatTy())
3119 weight = CW_Register;
3121 case 'c': // $25 for indirect jumps
3122 case 'l': // lo register
3123 case 'x': // hilo register pair
3124 if (type->isIntegerTy())
3125 weight = CW_SpecificReg;
3127 case 'I': // signed 16 bit immediate
3128 case 'J': // integer zero
3129 case 'K': // unsigned 16 bit immediate
3130 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3131 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3132 case 'O': // signed 15 bit immediate (+- 16383)
3133 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3134 if (isa<ConstantInt>(CallOperandVal))
3135 weight = CW_Constant;
3144 /// Given a register class constraint, like 'r', if this corresponds directly
3145 /// to an LLVM register class, return a register of 0 and the register class
3147 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
3148 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
3150 if (Constraint.size() == 1) {
3151 switch (Constraint[0]) {
3152 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3153 case 'y': // Same as 'r'. Exists for compatibility.
3155 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3156 if (Subtarget->inMips16Mode())
3157 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3158 return std::make_pair(0U, &Mips::CPURegsRegClass);
3160 if (VT == MVT::i64 && !HasMips64)
3161 return std::make_pair(0U, &Mips::CPURegsRegClass);
3162 if (VT == MVT::i64 && HasMips64)
3163 return std::make_pair(0U, &Mips::CPU64RegsRegClass);
3164 // This will generate an error message
3165 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3168 return std::make_pair(0U, &Mips::FGR32RegClass);
3169 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
3170 if (Subtarget->isFP64bit())
3171 return std::make_pair(0U, &Mips::FGR64RegClass);
3172 return std::make_pair(0U, &Mips::AFGR64RegClass);
3175 case 'c': // register suitable for indirect jump
3177 return std::make_pair((unsigned)Mips::T9, &Mips::CPURegsRegClass);
3178 assert(VT == MVT::i64 && "Unexpected type.");
3179 return std::make_pair((unsigned)Mips::T9_64, &Mips::CPU64RegsRegClass);
3180 case 'l': // register suitable for indirect jump
3182 return std::make_pair((unsigned)Mips::LO, &Mips::HILORegClass);
3183 return std::make_pair((unsigned)Mips::LO64, &Mips::HILO64RegClass);
3184 case 'x': // register suitable for indirect jump
3185 // Fixme: Not triggering the use of both hi and low
3186 // This will generate an error message
3187 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3190 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3193 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3194 /// vector. If it is invalid, don't add anything to Ops.
3195 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3196 std::string &Constraint,
3197 std::vector<SDValue>&Ops,
3198 SelectionDAG &DAG) const {
3199 SDValue Result(0, 0);
3201 // Only support length 1 constraints for now.
3202 if (Constraint.length() > 1) return;
3204 char ConstraintLetter = Constraint[0];
3205 switch (ConstraintLetter) {
3206 default: break; // This will fall through to the generic implementation
3207 case 'I': // Signed 16 bit constant
3208 // If this fails, the parent routine will give an error
3209 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3210 EVT Type = Op.getValueType();
3211 int64_t Val = C->getSExtValue();
3212 if (isInt<16>(Val)) {
3213 Result = DAG.getTargetConstant(Val, Type);
3218 case 'J': // integer zero
3219 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3220 EVT Type = Op.getValueType();
3221 int64_t Val = C->getZExtValue();
3223 Result = DAG.getTargetConstant(0, Type);
3228 case 'K': // unsigned 16 bit immediate
3229 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3230 EVT Type = Op.getValueType();
3231 uint64_t Val = (uint64_t)C->getZExtValue();
3232 if (isUInt<16>(Val)) {
3233 Result = DAG.getTargetConstant(Val, Type);
3238 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3239 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3240 EVT Type = Op.getValueType();
3241 int64_t Val = C->getSExtValue();
3242 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3243 Result = DAG.getTargetConstant(Val, Type);
3248 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3249 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3250 EVT Type = Op.getValueType();
3251 int64_t Val = C->getSExtValue();
3252 if ((Val >= -65535) && (Val <= -1)) {
3253 Result = DAG.getTargetConstant(Val, Type);
3258 case 'O': // signed 15 bit immediate
3259 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3260 EVT Type = Op.getValueType();
3261 int64_t Val = C->getSExtValue();
3262 if ((isInt<15>(Val))) {
3263 Result = DAG.getTargetConstant(Val, Type);
3268 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3269 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3270 EVT Type = Op.getValueType();
3271 int64_t Val = C->getSExtValue();
3272 if ((Val <= 65535) && (Val >= 1)) {
3273 Result = DAG.getTargetConstant(Val, Type);
3280 if (Result.getNode()) {
3281 Ops.push_back(Result);
3285 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3289 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
3290 // No global is ever allowed as a base.
3295 case 0: // "r+i" or just "i", depending on HasBaseReg.
3298 if (!AM.HasBaseReg) // allow "r+i".
3300 return false; // disallow "r+r" or "r+r+i".
3309 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3310 // The Mips target isn't yet aware of offsets.
3314 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3316 bool IsMemset, bool ZeroMemset,
3318 MachineFunction &MF) const {
3319 if (Subtarget->hasMips64())
3325 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3326 if (VT != MVT::f32 && VT != MVT::f64)
3328 if (Imm.isNegZero())
3330 return Imm.isZero();
3333 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3335 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3337 return TargetLowering::getJumpTableEncoding();
3340 /// This function returns true if CallSym is a long double emulation routine.
3341 static bool isF128SoftLibCall(const char *CallSym) {
3342 const char *const LibCalls[] =
3343 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2",
3344 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi",
3345 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf",
3346 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2",
3347 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3",
3348 "__trunctfdf2", "__trunctfsf2", "__unordtf2",
3349 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl",
3350 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl",
3353 const char * const *End = LibCalls + array_lengthof(LibCalls);
3355 // Check that LibCalls is sorted alphabetically.
3356 MipsTargetLowering::LTStr Comp;
3359 for (const char * const *I = LibCalls; I < End - 1; ++I)
3360 assert(Comp(*I, *(I + 1)));
3363 return std::binary_search(LibCalls, End, CallSym, Comp);
3366 /// This function returns true if Ty is fp128 or i128 which was originally a
3368 static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) {
3369 if (Ty->isFP128Ty())
3372 const ExternalSymbolSDNode *ES =
3373 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode);
3375 // If the Ty is i128 and the function being called is a long double emulation
3376 // routine, then the original type is f128.
3377 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol()));
3380 MipsTargetLowering::MipsCC::MipsCC(CallingConv::ID CC, bool IsO32_,
3382 : CCInfo(Info), CallConv(CC), IsO32(IsO32_) {
3383 // Pre-allocate reserved argument area.
3384 CCInfo.AllocateStack(reservedArgArea(), 1);
3387 void MipsTargetLowering::MipsCC::
3388 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
3389 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode,
3390 std::vector<ArgListEntry> &FuncArgs) {
3391 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
3392 "CallingConv::Fast shouldn't be used for vararg functions.");
3394 unsigned NumOpnds = Args.size();
3395 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
3397 for (unsigned I = 0; I != NumOpnds; ++I) {
3398 MVT ArgVT = Args[I].VT;
3399 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3402 if (ArgFlags.isByVal()) {
3403 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3407 if (IsVarArg && !Args[I].IsFixed)
3408 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3410 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode,
3412 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo);
3417 dbgs() << "Call operand #" << I << " has unhandled type "
3418 << EVT(ArgVT).getEVTString();
3420 llvm_unreachable(0);
3425 void MipsTargetLowering::MipsCC::
3426 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args,
3427 bool IsSoftFloat, Function::const_arg_iterator FuncArg) {
3428 unsigned NumArgs = Args.size();
3429 llvm::CCAssignFn *FixedFn = fixedArgFn();
3430 unsigned CurArgIdx = 0;
3432 for (unsigned I = 0; I != NumArgs; ++I) {
3433 MVT ArgVT = Args[I].VT;
3434 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3435 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx);
3436 CurArgIdx = Args[I].OrigArgIndex;
3438 if (ArgFlags.isByVal()) {
3439 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3443 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat);
3445 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo))
3449 dbgs() << "Formal Arg #" << I << " has unhandled type "
3450 << EVT(ArgVT).getEVTString();
3452 llvm_unreachable(0);
3456 template<typename Ty>
3457 void MipsTargetLowering::MipsCC::
3458 analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat,
3459 const SDNode *CallNode, const Type *RetTy) const {
3462 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode))
3463 Fn = RetCC_F128Soft;
3467 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) {
3468 MVT VT = RetVals[I].VT;
3469 ISD::ArgFlagsTy Flags = RetVals[I].Flags;
3470 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat);
3472 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) {
3474 dbgs() << "Call result #" << I << " has unhandled type "
3475 << EVT(VT).getEVTString() << '\n';
3477 llvm_unreachable(0);
3482 void MipsTargetLowering::MipsCC::
3483 analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat,
3484 const SDNode *CallNode, const Type *RetTy) const {
3485 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy);
3488 void MipsTargetLowering::MipsCC::
3489 analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat,
3490 const Type *RetTy) const {
3491 analyzeReturn(Outs, IsSoftFloat, 0, RetTy);
3495 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3497 CCValAssign::LocInfo LocInfo,
3498 ISD::ArgFlagsTy ArgFlags) {
3499 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3501 struct ByValArgInfo ByVal;
3502 unsigned RegSize = regSize();
3503 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3504 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3507 if (useRegsForByval())
3508 allocateRegs(ByVal, ByValSize, Align);
3510 // Allocate space on caller's stack.
3511 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3513 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3515 ByValArgs.push_back(ByVal);
3518 unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const {
3519 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs);
3522 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
3523 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0;
3526 const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const {
3527 return IsO32 ? O32IntRegs : Mips64IntRegs;
3530 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
3531 if (CallConv == CallingConv::Fast)
3532 return CC_Mips_FastCC;
3534 return IsO32 ? CC_MipsO32 : CC_MipsN;
3537 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
3538 return IsO32 ? CC_MipsO32 : CC_MipsN_VarArg;
3541 const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const {
3542 return IsO32 ? O32IntRegs : Mips64DPRegs;
3545 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3548 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs();
3549 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs();
3550 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3551 "Byval argument's size and alignment should be a multiple of"
3554 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3556 // If Align > RegSize, the first arg register must be even.
3557 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3558 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3562 // Mark the registers allocated.
3563 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3564 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3565 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3568 MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy,
3569 const SDNode *CallNode,
3570 bool IsSoftFloat) const {
3571 if (IsSoftFloat || IsO32)
3574 // Check if the original type was fp128.
3575 if (originalTypeIsF128(OrigTy, CallNode)) {
3576 assert(VT == MVT::i64);
3583 void MipsTargetLowering::
3584 copyByValRegs(SDValue Chain, DebugLoc DL, std::vector<SDValue> &OutChains,
3585 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3586 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3587 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3588 MachineFunction &MF = DAG.getMachineFunction();
3589 MachineFrameInfo *MFI = MF.getFrameInfo();
3590 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3591 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3595 FrameObjOffset = (int)CC.reservedArgArea() -
3596 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3598 FrameObjOffset = ByVal.Address;
3600 // Create frame object.
3601 EVT PtrTy = getPointerTy();
3602 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3603 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3604 InVals.push_back(FIN);
3609 // Copy arg registers.
3610 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3611 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3613 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3614 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3615 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3616 unsigned Offset = I * CC.regSize();
3617 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3618 DAG.getConstant(Offset, PtrTy));
3619 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3620 StorePtr, MachinePointerInfo(FuncArg, Offset),
3622 OutChains.push_back(Store);
3626 // Copy byVal arg to registers and stack.
3627 void MipsTargetLowering::
3628 passByValArg(SDValue Chain, DebugLoc DL,
3629 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
3630 SmallVector<SDValue, 8> &MemOpChains, SDValue StackPtr,
3631 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3632 const MipsCC &CC, const ByValArgInfo &ByVal,
3633 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3634 unsigned ByValSize = Flags.getByValSize();
3635 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
3636 unsigned RegSize = CC.regSize();
3637 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
3638 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
3640 if (ByVal.NumRegs) {
3641 const uint16_t *ArgRegs = CC.intArgRegs();
3642 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
3645 // Copy words to registers.
3646 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
3647 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3648 DAG.getConstant(Offset, PtrTy));
3649 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3650 MachinePointerInfo(), false, false, false,
3652 MemOpChains.push_back(LoadVal.getValue(1));
3653 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3654 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3657 // Return if the struct has been fully copied.
3658 if (ByValSize == Offset)
3661 // Copy the remainder of the byval argument with sub-word loads and shifts.
3662 if (LeftoverBytes) {
3663 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
3664 "Size of the remainder should be smaller than RegSize.");
3667 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
3668 Offset < ByValSize; LoadSize /= 2) {
3669 unsigned RemSize = ByValSize - Offset;
3671 if (RemSize < LoadSize)
3675 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3676 DAG.getConstant(Offset, PtrTy));
3678 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
3679 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
3680 false, false, Alignment);
3681 MemOpChains.push_back(LoadVal.getValue(1));
3683 // Shift the loaded value.
3687 Shamt = TotalSizeLoaded;
3689 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
3691 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3692 DAG.getConstant(Shamt, MVT::i32));
3695 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3700 TotalSizeLoaded += LoadSize;
3701 Alignment = std::min(Alignment, LoadSize);
3704 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3705 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3710 // Copy remainder of byval arg to it with memcpy.
3711 unsigned MemCpySize = ByValSize - Offset;
3712 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3713 DAG.getConstant(Offset, PtrTy));
3714 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3715 DAG.getIntPtrConstant(ByVal.Address));
3716 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3717 DAG.getConstant(MemCpySize, PtrTy), Alignment,
3718 /*isVolatile=*/false, /*AlwaysInline=*/false,
3719 MachinePointerInfo(0), MachinePointerInfo(0));
3720 MemOpChains.push_back(Chain);
3724 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3725 const MipsCC &CC, SDValue Chain,
3726 DebugLoc DL, SelectionDAG &DAG) const {
3727 unsigned NumRegs = CC.numIntArgRegs();
3728 const uint16_t *ArgRegs = CC.intArgRegs();
3729 const CCState &CCInfo = CC.getCCInfo();
3730 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3731 unsigned RegSize = CC.regSize();
3732 MVT RegTy = MVT::getIntegerVT(RegSize * 8);
3733 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3734 MachineFunction &MF = DAG.getMachineFunction();
3735 MachineFrameInfo *MFI = MF.getFrameInfo();
3736 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3738 // Offset of the first variable argument from stack pointer.
3742 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3745 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3747 // Record the frame index of the first variable argument
3748 // which is a value necessary to VASTART.
3749 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3750 MipsFI->setVarArgsFrameIndex(FI);
3752 // Copy the integer registers that have not been used for argument passing
3753 // to the argument register save area. For O32, the save area is allocated
3754 // in the caller's stack frame, while for N32/64, it is allocated in the
3755 // callee's stack frame.
3756 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3757 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3758 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3759 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3760 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3761 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3762 MachinePointerInfo(), false, false, 0);
3763 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
3764 OutChains.push_back(Store);