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"
16 #include "MipsISelLowering.h"
17 #include "InstPrinter/MipsInstPrinter.h"
18 #include "MCTargetDesc/MipsBaseInfo.h"
19 #include "MipsMachineFunction.h"
20 #include "MipsSubtarget.h"
21 #include "MipsTargetMachine.h"
22 #include "MipsTargetObjectFile.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/SelectionDAGISel.h"
30 #include "llvm/CodeGen/ValueTypes.h"
31 #include "llvm/IR/CallingConv.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/GlobalVariable.h"
35 #include "llvm/IR/Intrinsics.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/raw_ostream.h"
43 STATISTIC(NumTailCalls, "Number of tail calls");
46 EnableMipsTailCalls("enable-mips-tail-calls", cl::Hidden,
47 cl::desc("MIPS: Enable tail calls."), cl::init(false));
50 LargeGOT("mxgot", cl::Hidden,
51 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
54 Mips16HardFloat("mips16-hard-float", cl::NotHidden,
55 cl::desc("MIPS: mips16 hard float enable."),
58 static cl::opt<bool> DontExpandCondPseudos16(
59 "mips16-dont-expand-cond-pseudo",
61 cl::desc("Dont expand conditional move related "
62 "pseudos for Mips 16"),
66 static const uint16_t O32IntRegs[4] = {
67 Mips::A0, Mips::A1, Mips::A2, Mips::A3
70 static const uint16_t Mips64IntRegs[8] = {
71 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
72 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64
75 static const uint16_t Mips64DPRegs[8] = {
76 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
77 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
80 // If I is a shifted mask, set the size (Size) and the first bit of the
81 // mask (Pos), and return true.
82 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
83 static bool IsShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
84 if (!isShiftedMask_64(I))
87 Size = CountPopulation_64(I);
88 Pos = CountTrailingZeros_64(I);
92 static SDValue GetGlobalReg(SelectionDAG &DAG, EVT Ty) {
93 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
94 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
97 static SDValue getTargetNode(SDValue Op, SelectionDAG &DAG, unsigned Flag) {
98 EVT Ty = Op.getValueType();
100 if (GlobalAddressSDNode *N = dyn_cast<GlobalAddressSDNode>(Op))
101 return DAG.getTargetGlobalAddress(N->getGlobal(), Op.getDebugLoc(), Ty, 0,
103 if (ExternalSymbolSDNode *N = dyn_cast<ExternalSymbolSDNode>(Op))
104 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
105 if (BlockAddressSDNode *N = dyn_cast<BlockAddressSDNode>(Op))
106 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
107 if (JumpTableSDNode *N = dyn_cast<JumpTableSDNode>(Op))
108 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
109 if (ConstantPoolSDNode *N = dyn_cast<ConstantPoolSDNode>(Op))
110 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
111 N->getOffset(), Flag);
113 llvm_unreachable("Unexpected node type.");
117 static SDValue getAddrNonPIC(SDValue Op, SelectionDAG &DAG) {
118 DebugLoc DL = Op.getDebugLoc();
119 EVT Ty = Op.getValueType();
120 SDValue Hi = getTargetNode(Op, DAG, MipsII::MO_ABS_HI);
121 SDValue Lo = getTargetNode(Op, DAG, MipsII::MO_ABS_LO);
122 return DAG.getNode(ISD::ADD, DL, Ty,
123 DAG.getNode(MipsISD::Hi, DL, Ty, Hi),
124 DAG.getNode(MipsISD::Lo, DL, Ty, Lo));
127 static SDValue getAddrLocal(SDValue Op, SelectionDAG &DAG, bool HasMips64) {
128 DebugLoc DL = Op.getDebugLoc();
129 EVT Ty = Op.getValueType();
130 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
131 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, GetGlobalReg(DAG, Ty),
132 getTargetNode(Op, DAG, GOTFlag));
133 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT,
134 MachinePointerInfo::getGOT(), false, false, false,
136 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
137 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty, getTargetNode(Op, DAG, LoFlag));
138 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo);
141 static SDValue getAddrGlobal(SDValue Op, SelectionDAG &DAG, unsigned Flag) {
142 DebugLoc DL = Op.getDebugLoc();
143 EVT Ty = Op.getValueType();
144 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, GetGlobalReg(DAG, Ty),
145 getTargetNode(Op, DAG, Flag));
146 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt,
147 MachinePointerInfo::getGOT(), false, false, false, 0);
150 static SDValue getAddrGlobalLargeGOT(SDValue Op, SelectionDAG &DAG,
151 unsigned HiFlag, unsigned LoFlag) {
152 DebugLoc DL = Op.getDebugLoc();
153 EVT Ty = Op.getValueType();
154 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty, getTargetNode(Op, DAG, HiFlag));
155 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, GetGlobalReg(DAG, Ty));
156 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi,
157 getTargetNode(Op, DAG, LoFlag));
158 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper,
159 MachinePointerInfo::getGOT(), false, false, false, 0);
162 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
164 case MipsISD::JmpLink: return "MipsISD::JmpLink";
165 case MipsISD::TailCall: return "MipsISD::TailCall";
166 case MipsISD::Hi: return "MipsISD::Hi";
167 case MipsISD::Lo: return "MipsISD::Lo";
168 case MipsISD::GPRel: return "MipsISD::GPRel";
169 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
170 case MipsISD::Ret: return "MipsISD::Ret";
171 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
172 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
173 case MipsISD::FPCmp: return "MipsISD::FPCmp";
174 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
175 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
176 case MipsISD::FPRound: return "MipsISD::FPRound";
177 case MipsISD::MAdd: return "MipsISD::MAdd";
178 case MipsISD::MAddu: return "MipsISD::MAddu";
179 case MipsISD::MSub: return "MipsISD::MSub";
180 case MipsISD::MSubu: return "MipsISD::MSubu";
181 case MipsISD::DivRem: return "MipsISD::DivRem";
182 case MipsISD::DivRemU: return "MipsISD::DivRemU";
183 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
184 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
185 case MipsISD::Wrapper: return "MipsISD::Wrapper";
186 case MipsISD::Sync: return "MipsISD::Sync";
187 case MipsISD::Ext: return "MipsISD::Ext";
188 case MipsISD::Ins: return "MipsISD::Ins";
189 case MipsISD::LWL: return "MipsISD::LWL";
190 case MipsISD::LWR: return "MipsISD::LWR";
191 case MipsISD::SWL: return "MipsISD::SWL";
192 case MipsISD::SWR: return "MipsISD::SWR";
193 case MipsISD::LDL: return "MipsISD::LDL";
194 case MipsISD::LDR: return "MipsISD::LDR";
195 case MipsISD::SDL: return "MipsISD::SDL";
196 case MipsISD::SDR: return "MipsISD::SDR";
197 case MipsISD::EXTP: return "MipsISD::EXTP";
198 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
199 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
200 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
201 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
202 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
203 case MipsISD::SHILO: return "MipsISD::SHILO";
204 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
205 case MipsISD::MULT: return "MipsISD::MULT";
206 case MipsISD::MULTU: return "MipsISD::MULTU";
207 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSPDSP";
208 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
209 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
210 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
211 default: return NULL;
217 bool operator()(const char *s1, const char *s2) const
219 return strcmp(s1, s2) < 0;
223 std::set<const char*, ltstr> noHelperNeeded;
226 void MipsTargetLowering::SetMips16LibcallName
227 (RTLIB::Libcall l, const char *Name) {
228 setLibcallName(l, Name);
229 noHelperNeeded.insert(Name);
232 void MipsTargetLowering::setMips16HardFloatLibCalls() {
233 SetMips16LibcallName(RTLIB::ADD_F32, "__mips16_addsf3");
234 SetMips16LibcallName(RTLIB::ADD_F64, "__mips16_adddf3");
235 SetMips16LibcallName(RTLIB::SUB_F32, "__mips16_subsf3");
236 SetMips16LibcallName(RTLIB::SUB_F64, "__mips16_subdf3");
237 SetMips16LibcallName(RTLIB::MUL_F32, "__mips16_mulsf3");
238 SetMips16LibcallName(RTLIB::MUL_F64, "__mips16_muldf3");
239 SetMips16LibcallName(RTLIB::DIV_F32, "__mips16_divsf3");
240 SetMips16LibcallName(RTLIB::DIV_F64, "__mips16_divdf3");
241 SetMips16LibcallName(RTLIB::FPEXT_F32_F64, "__mips16_extendsfdf2");
242 SetMips16LibcallName(RTLIB::FPROUND_F64_F32, "__mips16_truncdfsf2");
243 SetMips16LibcallName(RTLIB::FPTOSINT_F32_I32, "__mips16_fix_truncsfsi");
244 SetMips16LibcallName(RTLIB::FPTOSINT_F64_I32, "__mips16_fix_truncdfsi");
245 SetMips16LibcallName(RTLIB::SINTTOFP_I32_F32, "__mips16_floatsisf");
246 SetMips16LibcallName(RTLIB::SINTTOFP_I32_F64, "__mips16_floatsidf");
247 SetMips16LibcallName(RTLIB::UINTTOFP_I32_F32, "__mips16_floatunsisf");
248 SetMips16LibcallName(RTLIB::UINTTOFP_I32_F64, "__mips16_floatunsidf");
249 SetMips16LibcallName(RTLIB::OEQ_F32, "__mips16_eqsf2");
250 SetMips16LibcallName(RTLIB::OEQ_F64, "__mips16_eqdf2");
251 SetMips16LibcallName(RTLIB::UNE_F32, "__mips16_nesf2");
252 SetMips16LibcallName(RTLIB::UNE_F64, "__mips16_nedf2");
253 SetMips16LibcallName(RTLIB::OGE_F32, "__mips16_gesf2");
254 SetMips16LibcallName(RTLIB::OGE_F64, "__mips16_gedf2");
255 SetMips16LibcallName(RTLIB::OLT_F32, "__mips16_ltsf2");
256 SetMips16LibcallName(RTLIB::OLT_F64, "__mips16_ltdf2");
257 SetMips16LibcallName(RTLIB::OLE_F32, "__mips16_lesf2");
258 SetMips16LibcallName(RTLIB::OLE_F64, "__mips16_ledf2");
259 SetMips16LibcallName(RTLIB::OGT_F32, "__mips16_gtsf2");
260 SetMips16LibcallName(RTLIB::OGT_F64, "__mips16_gtdf2");
261 SetMips16LibcallName(RTLIB::UO_F32, "__mips16_unordsf2");
262 SetMips16LibcallName(RTLIB::UO_F64, "__mips16_unorddf2");
263 SetMips16LibcallName(RTLIB::O_F32, "__mips16_unordsf2");
264 SetMips16LibcallName(RTLIB::O_F64, "__mips16_unorddf2");
268 MipsTargetLowering(MipsTargetMachine &TM)
269 : TargetLowering(TM, new MipsTargetObjectFile()),
270 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
271 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
272 IsO32(Subtarget->isABI_O32()) {
274 // Mips does not have i1 type, so use i32 for
275 // setcc operations results (slt, sgt, ...).
276 setBooleanContents(ZeroOrOneBooleanContent);
277 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
279 // Set up the register classes
280 addRegisterClass(MVT::i32, &Mips::CPURegsRegClass);
283 addRegisterClass(MVT::i64, &Mips::CPU64RegsRegClass);
285 if (Subtarget->inMips16Mode()) {
286 addRegisterClass(MVT::i32, &Mips::CPU16RegsRegClass);
288 setMips16HardFloatLibCalls();
291 if (Subtarget->hasDSP()) {
292 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
294 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
295 addRegisterClass(VecTys[i], &Mips::DSPRegsRegClass);
297 // Expand all builtin opcodes.
298 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
299 setOperationAction(Opc, VecTys[i], Expand);
301 setOperationAction(ISD::LOAD, VecTys[i], Legal);
302 setOperationAction(ISD::STORE, VecTys[i], Legal);
303 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
307 if (!TM.Options.UseSoftFloat) {
308 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
310 // When dealing with single precision only, use libcalls
311 if (!Subtarget->isSingleFloat()) {
313 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
315 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
319 // Load extented operations for i1 types must be promoted
320 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
321 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
322 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
324 // MIPS doesn't have extending float->double load/store
325 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
326 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
328 // Used by legalize types to correctly generate the setcc result.
329 // Without this, every float setcc comes with a AND/OR with the result,
330 // we don't want this, since the fpcmp result goes to a flag register,
331 // which is used implicitly by brcond and select operations.
332 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
334 // Mips Custom Operations
335 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
336 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
337 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
338 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
339 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
340 setOperationAction(ISD::SELECT, MVT::f32, Custom);
341 setOperationAction(ISD::SELECT, MVT::f64, Custom);
342 setOperationAction(ISD::SELECT, MVT::i32, Custom);
343 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
344 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
345 setOperationAction(ISD::SETCC, MVT::f32, Custom);
346 setOperationAction(ISD::SETCC, MVT::f64, Custom);
347 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
348 setOperationAction(ISD::VASTART, MVT::Other, Custom);
349 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
350 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
351 if (Subtarget->inMips16Mode()) {
352 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
353 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
356 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
357 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
359 if (!Subtarget->inMips16Mode()) {
360 setOperationAction(ISD::LOAD, MVT::i32, Custom);
361 setOperationAction(ISD::STORE, MVT::i32, Custom);
364 if (!TM.Options.NoNaNsFPMath) {
365 setOperationAction(ISD::FABS, MVT::f32, Custom);
366 setOperationAction(ISD::FABS, MVT::f64, Custom);
370 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
371 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
372 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
373 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
374 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
375 setOperationAction(ISD::SELECT, MVT::i64, Custom);
376 setOperationAction(ISD::LOAD, MVT::i64, Custom);
377 setOperationAction(ISD::STORE, MVT::i64, Custom);
381 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
382 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
383 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
386 setOperationAction(ISD::ADD, MVT::i32, Custom);
388 setOperationAction(ISD::ADD, MVT::i64, Custom);
390 setOperationAction(ISD::SDIV, MVT::i32, Expand);
391 setOperationAction(ISD::SREM, MVT::i32, Expand);
392 setOperationAction(ISD::UDIV, MVT::i32, Expand);
393 setOperationAction(ISD::UREM, MVT::i32, Expand);
394 setOperationAction(ISD::SDIV, MVT::i64, Expand);
395 setOperationAction(ISD::SREM, MVT::i64, Expand);
396 setOperationAction(ISD::UDIV, MVT::i64, Expand);
397 setOperationAction(ISD::UREM, MVT::i64, Expand);
399 // Operations not directly supported by Mips.
400 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
401 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
402 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
403 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
404 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
405 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
406 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
407 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
408 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
409 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
410 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
411 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
412 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
413 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
414 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
415 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
416 setOperationAction(ISD::ROTL, MVT::i32, Expand);
417 setOperationAction(ISD::ROTL, MVT::i64, Expand);
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
421 if (!Subtarget->hasMips32r2())
422 setOperationAction(ISD::ROTR, MVT::i32, Expand);
424 if (!Subtarget->hasMips64r2())
425 setOperationAction(ISD::ROTR, MVT::i64, Expand);
427 setOperationAction(ISD::FSIN, MVT::f32, Expand);
428 setOperationAction(ISD::FSIN, MVT::f64, Expand);
429 setOperationAction(ISD::FCOS, MVT::f32, Expand);
430 setOperationAction(ISD::FCOS, MVT::f64, Expand);
431 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
432 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
433 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
434 setOperationAction(ISD::FPOW, MVT::f32, Expand);
435 setOperationAction(ISD::FPOW, MVT::f64, Expand);
436 setOperationAction(ISD::FLOG, MVT::f32, Expand);
437 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
438 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
439 setOperationAction(ISD::FEXP, MVT::f32, Expand);
440 setOperationAction(ISD::FMA, MVT::f32, Expand);
441 setOperationAction(ISD::FMA, MVT::f64, Expand);
442 setOperationAction(ISD::FREM, MVT::f32, Expand);
443 setOperationAction(ISD::FREM, MVT::f64, Expand);
445 if (!TM.Options.NoNaNsFPMath) {
446 setOperationAction(ISD::FNEG, MVT::f32, Expand);
447 setOperationAction(ISD::FNEG, MVT::f64, Expand);
450 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
451 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
452 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
453 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
455 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
457 setOperationAction(ISD::VAARG, MVT::Other, Expand);
458 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
459 setOperationAction(ISD::VAEND, MVT::Other, Expand);
461 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
462 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
464 // Use the default for now
465 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
466 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
468 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
469 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
470 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
471 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
473 if (Subtarget->inMips16Mode()) {
474 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
475 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
476 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
477 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
478 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
479 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
480 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
481 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
482 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
483 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
484 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
485 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
488 setInsertFencesForAtomic(true);
490 if (!Subtarget->hasSEInReg()) {
491 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
492 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
495 if (!Subtarget->hasBitCount()) {
496 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
497 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
500 if (!Subtarget->hasSwap()) {
501 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
502 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
506 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
507 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
508 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
509 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
512 setTargetDAGCombine(ISD::ADDE);
513 setTargetDAGCombine(ISD::SUBE);
514 setTargetDAGCombine(ISD::SDIVREM);
515 setTargetDAGCombine(ISD::UDIVREM);
516 setTargetDAGCombine(ISD::SELECT);
517 setTargetDAGCombine(ISD::AND);
518 setTargetDAGCombine(ISD::OR);
519 setTargetDAGCombine(ISD::ADD);
521 setMinFunctionAlignment(HasMips64 ? 3 : 2);
523 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP);
524 computeRegisterProperties();
526 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0);
527 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1);
529 MaxStoresPerMemcpy = 16;
533 MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
534 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
536 if (Subtarget->inMips16Mode())
550 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
553 return VT.changeVectorElementTypeToInteger();
557 // Transforms a subgraph in CurDAG if the following pattern is found:
558 // (addc multLo, Lo0), (adde multHi, Hi0),
560 // multHi/Lo: product of multiplication
561 // Lo0: initial value of Lo register
562 // Hi0: initial value of Hi register
563 // Return true if pattern matching was successful.
564 static bool SelectMadd(SDNode *ADDENode, SelectionDAG *CurDAG) {
565 // ADDENode's second operand must be a flag output of an ADDC node in order
566 // for the matching to be successful.
567 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
569 if (ADDCNode->getOpcode() != ISD::ADDC)
572 SDValue MultHi = ADDENode->getOperand(0);
573 SDValue MultLo = ADDCNode->getOperand(0);
574 SDNode *MultNode = MultHi.getNode();
575 unsigned MultOpc = MultHi.getOpcode();
577 // MultHi and MultLo must be generated by the same node,
578 if (MultLo.getNode() != MultNode)
581 // and it must be a multiplication.
582 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
585 // MultLo amd MultHi must be the first and second output of MultNode
587 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
590 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
591 // of the values of MultNode, in which case MultNode will be removed in later
593 // If there exist users other than ADDENode or ADDCNode, this function returns
594 // here, which will result in MultNode being mapped to a single MULT
595 // instruction node rather than a pair of MULT and MADD instructions being
597 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
600 SDValue Chain = CurDAG->getEntryNode();
601 DebugLoc dl = ADDENode->getDebugLoc();
603 // create MipsMAdd(u) node
604 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
606 SDValue MAdd = CurDAG->getNode(MultOpc, dl, MVT::Glue,
607 MultNode->getOperand(0),// Factor 0
608 MultNode->getOperand(1),// Factor 1
609 ADDCNode->getOperand(1),// Lo0
610 ADDENode->getOperand(1));// Hi0
612 // create CopyFromReg nodes
613 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
615 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
617 CopyFromLo.getValue(2));
619 // replace uses of adde and addc here
620 if (!SDValue(ADDCNode, 0).use_empty())
621 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
623 if (!SDValue(ADDENode, 0).use_empty())
624 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
630 // Transforms a subgraph in CurDAG if the following pattern is found:
631 // (addc Lo0, multLo), (sube Hi0, multHi),
633 // multHi/Lo: product of multiplication
634 // Lo0: initial value of Lo register
635 // Hi0: initial value of Hi register
636 // Return true if pattern matching was successful.
637 static bool SelectMsub(SDNode *SUBENode, SelectionDAG *CurDAG) {
638 // SUBENode's second operand must be a flag output of an SUBC node in order
639 // for the matching to be successful.
640 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
642 if (SUBCNode->getOpcode() != ISD::SUBC)
645 SDValue MultHi = SUBENode->getOperand(1);
646 SDValue MultLo = SUBCNode->getOperand(1);
647 SDNode *MultNode = MultHi.getNode();
648 unsigned MultOpc = MultHi.getOpcode();
650 // MultHi and MultLo must be generated by the same node,
651 if (MultLo.getNode() != MultNode)
654 // and it must be a multiplication.
655 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
658 // MultLo amd MultHi must be the first and second output of MultNode
660 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
663 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
664 // of the values of MultNode, in which case MultNode will be removed in later
666 // If there exist users other than SUBENode or SUBCNode, this function returns
667 // here, which will result in MultNode being mapped to a single MULT
668 // instruction node rather than a pair of MULT and MSUB instructions being
670 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
673 SDValue Chain = CurDAG->getEntryNode();
674 DebugLoc dl = SUBENode->getDebugLoc();
676 // create MipsSub(u) node
677 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
679 SDValue MSub = CurDAG->getNode(MultOpc, dl, MVT::Glue,
680 MultNode->getOperand(0),// Factor 0
681 MultNode->getOperand(1),// Factor 1
682 SUBCNode->getOperand(0),// Lo0
683 SUBENode->getOperand(0));// Hi0
685 // create CopyFromReg nodes
686 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
688 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
690 CopyFromLo.getValue(2));
692 // replace uses of sube and subc here
693 if (!SDValue(SUBCNode, 0).use_empty())
694 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
696 if (!SDValue(SUBENode, 0).use_empty())
697 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
702 static SDValue PerformADDECombine(SDNode *N, SelectionDAG &DAG,
703 TargetLowering::DAGCombinerInfo &DCI,
704 const MipsSubtarget *Subtarget) {
705 if (DCI.isBeforeLegalize())
708 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
710 return SDValue(N, 0);
715 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG &DAG,
716 TargetLowering::DAGCombinerInfo &DCI,
717 const MipsSubtarget *Subtarget) {
718 if (DCI.isBeforeLegalize())
721 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
723 return SDValue(N, 0);
728 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG &DAG,
729 TargetLowering::DAGCombinerInfo &DCI,
730 const MipsSubtarget *Subtarget) {
731 if (DCI.isBeforeLegalizeOps())
734 EVT Ty = N->getValueType(0);
735 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
736 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
737 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
739 DebugLoc dl = N->getDebugLoc();
741 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
742 N->getOperand(0), N->getOperand(1));
743 SDValue InChain = DAG.getEntryNode();
744 SDValue InGlue = DivRem;
747 if (N->hasAnyUseOfValue(0)) {
748 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty,
750 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
751 InChain = CopyFromLo.getValue(1);
752 InGlue = CopyFromLo.getValue(2);
756 if (N->hasAnyUseOfValue(1)) {
757 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
759 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
765 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
767 default: llvm_unreachable("Unknown fp condition code!");
769 case ISD::SETOEQ: return Mips::FCOND_OEQ;
770 case ISD::SETUNE: return Mips::FCOND_UNE;
772 case ISD::SETOLT: return Mips::FCOND_OLT;
774 case ISD::SETOGT: return Mips::FCOND_OGT;
776 case ISD::SETOLE: return Mips::FCOND_OLE;
778 case ISD::SETOGE: return Mips::FCOND_OGE;
779 case ISD::SETULT: return Mips::FCOND_ULT;
780 case ISD::SETULE: return Mips::FCOND_ULE;
781 case ISD::SETUGT: return Mips::FCOND_UGT;
782 case ISD::SETUGE: return Mips::FCOND_UGE;
783 case ISD::SETUO: return Mips::FCOND_UN;
784 case ISD::SETO: return Mips::FCOND_OR;
786 case ISD::SETONE: return Mips::FCOND_ONE;
787 case ISD::SETUEQ: return Mips::FCOND_UEQ;
792 // Returns true if condition code has to be inverted.
793 static bool InvertFPCondCode(Mips::CondCode CC) {
794 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
797 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
798 "Illegal Condition Code");
803 // Creates and returns an FPCmp node from a setcc node.
804 // Returns Op if setcc is not a floating point comparison.
805 static SDValue CreateFPCmp(SelectionDAG &DAG, const SDValue &Op) {
806 // must be a SETCC node
807 if (Op.getOpcode() != ISD::SETCC)
810 SDValue LHS = Op.getOperand(0);
812 if (!LHS.getValueType().isFloatingPoint())
815 SDValue RHS = Op.getOperand(1);
816 DebugLoc dl = Op.getDebugLoc();
818 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
819 // node if necessary.
820 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
822 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
823 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
826 // Creates and returns a CMovFPT/F node.
827 static SDValue CreateCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
828 SDValue False, DebugLoc DL) {
829 bool invert = InvertFPCondCode((Mips::CondCode)
830 cast<ConstantSDNode>(Cond.getOperand(2))
833 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
834 True.getValueType(), True, False, Cond);
837 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
838 TargetLowering::DAGCombinerInfo &DCI,
839 const MipsSubtarget *Subtarget) {
840 if (DCI.isBeforeLegalizeOps())
843 SDValue SetCC = N->getOperand(0);
845 if ((SetCC.getOpcode() != ISD::SETCC) ||
846 !SetCC.getOperand(0).getValueType().isInteger())
849 SDValue False = N->getOperand(2);
850 EVT FalseTy = False.getValueType();
852 if (!FalseTy.isInteger())
855 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
857 if (!CN || CN->getZExtValue())
860 const DebugLoc DL = N->getDebugLoc();
861 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
862 SDValue True = N->getOperand(1);
864 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
865 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
867 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
870 static SDValue PerformANDCombine(SDNode *N, SelectionDAG &DAG,
871 TargetLowering::DAGCombinerInfo &DCI,
872 const MipsSubtarget *Subtarget) {
873 // Pattern match EXT.
874 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
875 // => ext $dst, $src, size, pos
876 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
879 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
880 unsigned ShiftRightOpc = ShiftRight.getOpcode();
882 // Op's first operand must be a shift right.
883 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
886 // The second operand of the shift must be an immediate.
888 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
891 uint64_t Pos = CN->getZExtValue();
892 uint64_t SMPos, SMSize;
894 // Op's second operand must be a shifted mask.
895 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
896 !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
899 // Return if the shifted mask does not start at bit 0 or the sum of its size
900 // and Pos exceeds the word's size.
901 EVT ValTy = N->getValueType(0);
902 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
905 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
906 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
907 DAG.getConstant(SMSize, MVT::i32));
910 static SDValue PerformORCombine(SDNode *N, SelectionDAG &DAG,
911 TargetLowering::DAGCombinerInfo &DCI,
912 const MipsSubtarget *Subtarget) {
913 // Pattern match INS.
914 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
915 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
916 // => ins $dst, $src, size, pos, $src1
917 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
920 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
921 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
924 // See if Op's first operand matches (and $src1 , mask0).
925 if (And0.getOpcode() != ISD::AND)
928 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
929 !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
932 // See if Op's second operand matches (and (shl $src, pos), mask1).
933 if (And1.getOpcode() != ISD::AND)
936 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
937 !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
940 // The shift masks must have the same position and size.
941 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
944 SDValue Shl = And1.getOperand(0);
945 if (Shl.getOpcode() != ISD::SHL)
948 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
951 unsigned Shamt = CN->getZExtValue();
953 // Return if the shift amount and the first bit position of mask are not the
955 EVT ValTy = N->getValueType(0);
956 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
959 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
960 DAG.getConstant(SMPos0, MVT::i32),
961 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
964 static SDValue PerformADDCombine(SDNode *N, SelectionDAG &DAG,
965 TargetLowering::DAGCombinerInfo &DCI,
966 const MipsSubtarget *Subtarget) {
967 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
969 if (DCI.isBeforeLegalizeOps())
972 SDValue Add = N->getOperand(1);
974 if (Add.getOpcode() != ISD::ADD)
977 SDValue Lo = Add.getOperand(1);
979 if ((Lo.getOpcode() != MipsISD::Lo) ||
980 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
983 EVT ValTy = N->getValueType(0);
984 DebugLoc DL = N->getDebugLoc();
986 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
988 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
991 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
993 SelectionDAG &DAG = DCI.DAG;
994 unsigned opc = N->getOpcode();
999 return PerformADDECombine(N, DAG, DCI, Subtarget);
1001 return PerformSUBECombine(N, DAG, DCI, Subtarget);
1004 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
1006 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
1008 return PerformANDCombine(N, DAG, DCI, Subtarget);
1010 return PerformORCombine(N, DAG, DCI, Subtarget);
1012 return PerformADDCombine(N, DAG, DCI, Subtarget);
1019 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
1020 SmallVectorImpl<SDValue> &Results,
1021 SelectionDAG &DAG) const {
1022 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
1024 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
1025 Results.push_back(Res.getValue(I));
1029 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
1030 SmallVectorImpl<SDValue> &Results,
1031 SelectionDAG &DAG) const {
1032 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
1034 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
1035 Results.push_back(Res.getValue(I));
1038 SDValue MipsTargetLowering::
1039 LowerOperation(SDValue Op, SelectionDAG &DAG) const
1041 switch (Op.getOpcode())
1043 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
1044 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
1045 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
1046 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
1047 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
1048 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
1049 case ISD::SELECT: return LowerSELECT(Op, DAG);
1050 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
1051 case ISD::SETCC: return LowerSETCC(Op, DAG);
1052 case ISD::VASTART: return LowerVASTART(Op, DAG);
1053 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
1054 case ISD::FABS: return LowerFABS(Op, DAG);
1055 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
1056 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
1057 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
1058 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
1059 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
1060 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
1061 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG, true);
1062 case ISD::SRL_PARTS: return LowerShiftRightParts(Op, DAG, false);
1063 case ISD::LOAD: return LowerLOAD(Op, DAG);
1064 case ISD::STORE: return LowerSTORE(Op, DAG);
1065 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
1066 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
1067 case ISD::ADD: return LowerADD(Op, DAG);
1072 //===----------------------------------------------------------------------===//
1073 // Lower helper functions
1074 //===----------------------------------------------------------------------===//
1076 // AddLiveIn - This helper function adds the specified physical register to the
1077 // MachineFunction as a live in value. It also creates a corresponding
1078 // virtual register for it.
1080 AddLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
1082 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1083 MF.getRegInfo().addLiveIn(PReg, VReg);
1087 // Get fp branch code (not opcode) from condition code.
1088 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
1089 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
1090 return Mips::BRANCH_T;
1092 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
1093 "Invalid CondCode.");
1095 return Mips::BRANCH_F;
1099 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
1101 const MipsSubtarget *Subtarget,
1102 const TargetInstrInfo *TII,
1103 bool isFPCmp, unsigned Opc) {
1104 // There is no need to expand CMov instructions if target has
1105 // conditional moves.
1106 if (Subtarget->hasCondMov())
1109 // To "insert" a SELECT_CC instruction, we actually have to insert the
1110 // diamond control-flow pattern. The incoming instruction knows the
1111 // destination vreg to set, the condition code register to branch on, the
1112 // true/false values to select between, and a branch opcode to use.
1113 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1114 MachineFunction::iterator It = BB;
1121 // bNE r1, r0, copy1MBB
1122 // fallthrough --> copy0MBB
1123 MachineBasicBlock *thisMBB = BB;
1124 MachineFunction *F = BB->getParent();
1125 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1126 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1127 F->insert(It, copy0MBB);
1128 F->insert(It, sinkMBB);
1130 // Transfer the remainder of BB and its successor edges to sinkMBB.
1131 sinkMBB->splice(sinkMBB->begin(), BB,
1132 llvm::next(MachineBasicBlock::iterator(MI)),
1134 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1136 // Next, add the true and fallthrough blocks as its successors.
1137 BB->addSuccessor(copy0MBB);
1138 BB->addSuccessor(sinkMBB);
1140 // Emit the right instruction according to the type of the operands compared
1142 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
1144 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
1145 .addReg(Mips::ZERO).addMBB(sinkMBB);
1148 // %FalseValue = ...
1149 // # fallthrough to sinkMBB
1152 // Update machine-CFG edges
1153 BB->addSuccessor(sinkMBB);
1156 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
1161 BuildMI(*BB, BB->begin(), dl,
1162 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1163 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
1164 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
1166 BuildMI(*BB, BB->begin(), dl,
1167 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1168 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
1169 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
1171 MI->eraseFromParent(); // The pseudo instruction is gone now.
1177 MipsTargetLowering::EmitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
1179 // bposge32_pseudo $vr0
1189 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
1191 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
1192 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1193 const TargetRegisterClass *RC = &Mips::CPURegsRegClass;
1194 DebugLoc DL = MI->getDebugLoc();
1195 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1196 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
1197 MachineFunction *F = BB->getParent();
1198 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
1199 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
1200 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
1203 F->insert(It, Sink);
1205 // Transfer the remainder of BB and its successor edges to Sink.
1206 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
1208 Sink->transferSuccessorsAndUpdatePHIs(BB);
1211 BB->addSuccessor(FBB);
1212 BB->addSuccessor(TBB);
1213 FBB->addSuccessor(Sink);
1214 TBB->addSuccessor(Sink);
1216 // Insert the real bposge32 instruction to $BB.
1217 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
1220 unsigned VR2 = RegInfo.createVirtualRegister(RC);
1221 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
1222 .addReg(Mips::ZERO).addImm(0);
1223 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
1226 unsigned VR1 = RegInfo.createVirtualRegister(RC);
1227 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
1228 .addReg(Mips::ZERO).addImm(1);
1230 // Insert phi function to $Sink.
1231 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
1232 MI->getOperand(0).getReg())
1233 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
1235 MI->eraseFromParent(); // The pseudo instruction is gone now.
1239 MachineBasicBlock *MipsTargetLowering::EmitSel16(unsigned Opc, MachineInstr *MI,
1240 MachineBasicBlock *BB) const {
1241 if (DontExpandCondPseudos16)
1243 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1244 DebugLoc dl = MI->getDebugLoc();
1245 // To "insert" a SELECT_CC instruction, we actually have to insert the
1246 // diamond control-flow pattern. The incoming instruction knows the
1247 // destination vreg to set, the condition code register to branch on, the
1248 // true/false values to select between, and a branch opcode to use.
1249 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1250 MachineFunction::iterator It = BB;
1257 // bNE r1, r0, copy1MBB
1258 // fallthrough --> copy0MBB
1259 MachineBasicBlock *thisMBB = BB;
1260 MachineFunction *F = BB->getParent();
1261 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1262 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1263 F->insert(It, copy0MBB);
1264 F->insert(It, sinkMBB);
1266 // Transfer the remainder of BB and its successor edges to sinkMBB.
1267 sinkMBB->splice(sinkMBB->begin(), BB,
1268 llvm::next(MachineBasicBlock::iterator(MI)),
1270 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1272 // Next, add the true and fallthrough blocks as its successors.
1273 BB->addSuccessor(copy0MBB);
1274 BB->addSuccessor(sinkMBB);
1276 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(3).getReg())
1280 // %FalseValue = ...
1281 // # fallthrough to sinkMBB
1284 // Update machine-CFG edges
1285 BB->addSuccessor(sinkMBB);
1288 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
1292 BuildMI(*BB, BB->begin(), dl,
1293 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1294 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB)
1295 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB);
1297 MI->eraseFromParent(); // The pseudo instruction is gone now.
1301 MachineBasicBlock *MipsTargetLowering::EmitSelT16
1302 (unsigned Opc1, unsigned Opc2,
1303 MachineInstr *MI, MachineBasicBlock *BB) const {
1304 if (DontExpandCondPseudos16)
1306 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1307 DebugLoc dl = MI->getDebugLoc();
1308 // To "insert" a SELECT_CC instruction, we actually have to insert the
1309 // diamond control-flow pattern. The incoming instruction knows the
1310 // destination vreg to set, the condition code register to branch on, the
1311 // true/false values to select between, and a branch opcode to use.
1312 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1313 MachineFunction::iterator It = BB;
1320 // bNE r1, r0, copy1MBB
1321 // fallthrough --> copy0MBB
1322 MachineBasicBlock *thisMBB = BB;
1323 MachineFunction *F = BB->getParent();
1324 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1325 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1326 F->insert(It, copy0MBB);
1327 F->insert(It, sinkMBB);
1329 // Transfer the remainder of BB and its successor edges to sinkMBB.
1330 sinkMBB->splice(sinkMBB->begin(), BB,
1331 llvm::next(MachineBasicBlock::iterator(MI)),
1333 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1335 // Next, add the true and fallthrough blocks as its successors.
1336 BB->addSuccessor(copy0MBB);
1337 BB->addSuccessor(sinkMBB);
1339 BuildMI(BB, dl, TII->get(Opc2)).addReg(MI->getOperand(3).getReg())
1340 .addReg(MI->getOperand(4).getReg());
1341 BuildMI(BB, dl, TII->get(Opc1)).addMBB(sinkMBB);
1344 // %FalseValue = ...
1345 // # fallthrough to sinkMBB
1348 // Update machine-CFG edges
1349 BB->addSuccessor(sinkMBB);
1352 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
1356 BuildMI(*BB, BB->begin(), dl,
1357 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1358 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB)
1359 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB);
1361 MI->eraseFromParent(); // The pseudo instruction is gone now.
1367 MachineBasicBlock *MipsTargetLowering::EmitSeliT16
1368 (unsigned Opc1, unsigned Opc2,
1369 MachineInstr *MI, MachineBasicBlock *BB) const {
1370 if (DontExpandCondPseudos16)
1372 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1373 DebugLoc dl = MI->getDebugLoc();
1374 // To "insert" a SELECT_CC instruction, we actually have to insert the
1375 // diamond control-flow pattern. The incoming instruction knows the
1376 // destination vreg to set, the condition code register to branch on, the
1377 // true/false values to select between, and a branch opcode to use.
1378 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1379 MachineFunction::iterator It = BB;
1386 // bNE r1, r0, copy1MBB
1387 // fallthrough --> copy0MBB
1388 MachineBasicBlock *thisMBB = BB;
1389 MachineFunction *F = BB->getParent();
1390 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1391 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1392 F->insert(It, copy0MBB);
1393 F->insert(It, sinkMBB);
1395 // Transfer the remainder of BB and its successor edges to sinkMBB.
1396 sinkMBB->splice(sinkMBB->begin(), BB,
1397 llvm::next(MachineBasicBlock::iterator(MI)),
1399 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1401 // Next, add the true and fallthrough blocks as its successors.
1402 BB->addSuccessor(copy0MBB);
1403 BB->addSuccessor(sinkMBB);
1405 BuildMI(BB, dl, TII->get(Opc2)).addReg(MI->getOperand(3).getReg())
1406 .addImm(MI->getOperand(4).getImm());
1407 BuildMI(BB, dl, TII->get(Opc1)).addMBB(sinkMBB);
1410 // %FalseValue = ...
1411 // # fallthrough to sinkMBB
1414 // Update machine-CFG edges
1415 BB->addSuccessor(sinkMBB);
1418 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
1422 BuildMI(*BB, BB->begin(), dl,
1423 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1424 .addReg(MI->getOperand(1).getReg()).addMBB(thisMBB)
1425 .addReg(MI->getOperand(2).getReg()).addMBB(copy0MBB);
1427 MI->eraseFromParent(); // The pseudo instruction is gone now.
1433 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1434 MachineBasicBlock *BB) const {
1435 switch (MI->getOpcode()) {
1437 llvm_unreachable("Unexpected instr type to insert");
1438 case Mips::ATOMIC_LOAD_ADD_I8:
1439 case Mips::ATOMIC_LOAD_ADD_I8_P8:
1440 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
1441 case Mips::ATOMIC_LOAD_ADD_I16:
1442 case Mips::ATOMIC_LOAD_ADD_I16_P8:
1443 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
1444 case Mips::ATOMIC_LOAD_ADD_I32:
1445 case Mips::ATOMIC_LOAD_ADD_I32_P8:
1446 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
1447 case Mips::ATOMIC_LOAD_ADD_I64:
1448 case Mips::ATOMIC_LOAD_ADD_I64_P8:
1449 return EmitAtomicBinary(MI, BB, 8, Mips::DADDu);
1451 case Mips::ATOMIC_LOAD_AND_I8:
1452 case Mips::ATOMIC_LOAD_AND_I8_P8:
1453 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
1454 case Mips::ATOMIC_LOAD_AND_I16:
1455 case Mips::ATOMIC_LOAD_AND_I16_P8:
1456 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
1457 case Mips::ATOMIC_LOAD_AND_I32:
1458 case Mips::ATOMIC_LOAD_AND_I32_P8:
1459 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
1460 case Mips::ATOMIC_LOAD_AND_I64:
1461 case Mips::ATOMIC_LOAD_AND_I64_P8:
1462 return EmitAtomicBinary(MI, BB, 8, Mips::AND64);
1464 case Mips::ATOMIC_LOAD_OR_I8:
1465 case Mips::ATOMIC_LOAD_OR_I8_P8:
1466 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
1467 case Mips::ATOMIC_LOAD_OR_I16:
1468 case Mips::ATOMIC_LOAD_OR_I16_P8:
1469 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
1470 case Mips::ATOMIC_LOAD_OR_I32:
1471 case Mips::ATOMIC_LOAD_OR_I32_P8:
1472 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
1473 case Mips::ATOMIC_LOAD_OR_I64:
1474 case Mips::ATOMIC_LOAD_OR_I64_P8:
1475 return EmitAtomicBinary(MI, BB, 8, Mips::OR64);
1477 case Mips::ATOMIC_LOAD_XOR_I8:
1478 case Mips::ATOMIC_LOAD_XOR_I8_P8:
1479 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
1480 case Mips::ATOMIC_LOAD_XOR_I16:
1481 case Mips::ATOMIC_LOAD_XOR_I16_P8:
1482 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
1483 case Mips::ATOMIC_LOAD_XOR_I32:
1484 case Mips::ATOMIC_LOAD_XOR_I32_P8:
1485 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
1486 case Mips::ATOMIC_LOAD_XOR_I64:
1487 case Mips::ATOMIC_LOAD_XOR_I64_P8:
1488 return EmitAtomicBinary(MI, BB, 8, Mips::XOR64);
1490 case Mips::ATOMIC_LOAD_NAND_I8:
1491 case Mips::ATOMIC_LOAD_NAND_I8_P8:
1492 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
1493 case Mips::ATOMIC_LOAD_NAND_I16:
1494 case Mips::ATOMIC_LOAD_NAND_I16_P8:
1495 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
1496 case Mips::ATOMIC_LOAD_NAND_I32:
1497 case Mips::ATOMIC_LOAD_NAND_I32_P8:
1498 return EmitAtomicBinary(MI, BB, 4, 0, true);
1499 case Mips::ATOMIC_LOAD_NAND_I64:
1500 case Mips::ATOMIC_LOAD_NAND_I64_P8:
1501 return EmitAtomicBinary(MI, BB, 8, 0, true);
1503 case Mips::ATOMIC_LOAD_SUB_I8:
1504 case Mips::ATOMIC_LOAD_SUB_I8_P8:
1505 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
1506 case Mips::ATOMIC_LOAD_SUB_I16:
1507 case Mips::ATOMIC_LOAD_SUB_I16_P8:
1508 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
1509 case Mips::ATOMIC_LOAD_SUB_I32:
1510 case Mips::ATOMIC_LOAD_SUB_I32_P8:
1511 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
1512 case Mips::ATOMIC_LOAD_SUB_I64:
1513 case Mips::ATOMIC_LOAD_SUB_I64_P8:
1514 return EmitAtomicBinary(MI, BB, 8, Mips::DSUBu);
1516 case Mips::ATOMIC_SWAP_I8:
1517 case Mips::ATOMIC_SWAP_I8_P8:
1518 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
1519 case Mips::ATOMIC_SWAP_I16:
1520 case Mips::ATOMIC_SWAP_I16_P8:
1521 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
1522 case Mips::ATOMIC_SWAP_I32:
1523 case Mips::ATOMIC_SWAP_I32_P8:
1524 return EmitAtomicBinary(MI, BB, 4, 0);
1525 case Mips::ATOMIC_SWAP_I64:
1526 case Mips::ATOMIC_SWAP_I64_P8:
1527 return EmitAtomicBinary(MI, BB, 8, 0);
1529 case Mips::ATOMIC_CMP_SWAP_I8:
1530 case Mips::ATOMIC_CMP_SWAP_I8_P8:
1531 return EmitAtomicCmpSwapPartword(MI, BB, 1);
1532 case Mips::ATOMIC_CMP_SWAP_I16:
1533 case Mips::ATOMIC_CMP_SWAP_I16_P8:
1534 return EmitAtomicCmpSwapPartword(MI, BB, 2);
1535 case Mips::ATOMIC_CMP_SWAP_I32:
1536 case Mips::ATOMIC_CMP_SWAP_I32_P8:
1537 return EmitAtomicCmpSwap(MI, BB, 4);
1538 case Mips::ATOMIC_CMP_SWAP_I64:
1539 case Mips::ATOMIC_CMP_SWAP_I64_P8:
1540 return EmitAtomicCmpSwap(MI, BB, 8);
1541 case Mips::BPOSGE32_PSEUDO:
1542 return EmitBPOSGE32(MI, BB);
1544 return EmitSel16(Mips::BeqzRxImm16, MI, BB);
1546 return EmitSel16(Mips::BnezRxImm16, MI, BB);
1547 case Mips::SelTBteqZCmpi:
1548 return EmitSeliT16(Mips::BteqzX16, Mips::CmpiRxImmX16, MI, BB);
1549 case Mips::SelTBteqZSlti:
1550 return EmitSeliT16(Mips::BteqzX16, Mips::SltiRxImmX16, MI, BB);
1551 case Mips::SelTBteqZSltiu:
1552 return EmitSeliT16(Mips::BteqzX16, Mips::SltiuRxImmX16, MI, BB);
1553 case Mips::SelTBtneZCmpi:
1554 return EmitSeliT16(Mips::BtnezX16, Mips::CmpiRxImmX16, MI, BB);
1555 case Mips::SelTBtneZSlti:
1556 return EmitSeliT16(Mips::BtnezX16, Mips::SltiRxImmX16, MI, BB);
1557 case Mips::SelTBtneZSltiu:
1558 return EmitSeliT16(Mips::BtnezX16, Mips::SltiuRxImmX16, MI, BB);
1559 case Mips::SelTBteqZCmp:
1560 return EmitSelT16(Mips::BteqzX16, Mips::CmpRxRy16, MI, BB);
1561 case Mips::SelTBteqZSlt:
1562 return EmitSelT16(Mips::BteqzX16, Mips::SltRxRy16, MI, BB);
1563 case Mips::SelTBteqZSltu:
1564 return EmitSelT16(Mips::BteqzX16, Mips::SltuRxRy16, MI, BB);
1565 case Mips::SelTBtneZCmp:
1566 return EmitSelT16(Mips::BtnezX16, Mips::CmpRxRy16, MI, BB);
1567 case Mips::SelTBtneZSlt:
1568 return EmitSelT16(Mips::BtnezX16, Mips::SltRxRy16, MI, BB);
1569 case Mips::SelTBtneZSltu:
1570 return EmitSelT16(Mips::BtnezX16, Mips::SltuRxRy16, MI, BB);
1574 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1575 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1577 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
1578 unsigned Size, unsigned BinOpcode,
1580 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1582 MachineFunction *MF = BB->getParent();
1583 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1584 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1585 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1586 DebugLoc dl = MI->getDebugLoc();
1587 unsigned LL, SC, AND, NOR, ZERO, BEQ;
1590 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1591 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1598 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1599 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1602 ZERO = Mips::ZERO_64;
1606 unsigned OldVal = MI->getOperand(0).getReg();
1607 unsigned Ptr = MI->getOperand(1).getReg();
1608 unsigned Incr = MI->getOperand(2).getReg();
1610 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1611 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1612 unsigned Success = RegInfo.createVirtualRegister(RC);
1614 // insert new blocks after the current block
1615 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1616 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1617 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1618 MachineFunction::iterator It = BB;
1620 MF->insert(It, loopMBB);
1621 MF->insert(It, exitMBB);
1623 // Transfer the remainder of BB and its successor edges to exitMBB.
1624 exitMBB->splice(exitMBB->begin(), BB,
1625 llvm::next(MachineBasicBlock::iterator(MI)),
1627 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1631 // fallthrough --> loopMBB
1632 BB->addSuccessor(loopMBB);
1633 loopMBB->addSuccessor(loopMBB);
1634 loopMBB->addSuccessor(exitMBB);
1637 // ll oldval, 0(ptr)
1638 // <binop> storeval, oldval, incr
1639 // sc success, storeval, 0(ptr)
1640 // beq success, $0, loopMBB
1642 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1644 // and andres, oldval, incr
1645 // nor storeval, $0, andres
1646 BuildMI(BB, dl, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1647 BuildMI(BB, dl, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1648 } else if (BinOpcode) {
1649 // <binop> storeval, oldval, incr
1650 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1654 BuildMI(BB, dl, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1655 BuildMI(BB, dl, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1657 MI->eraseFromParent(); // The instruction is gone now.
1663 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
1664 MachineBasicBlock *BB,
1665 unsigned Size, unsigned BinOpcode,
1667 assert((Size == 1 || Size == 2) &&
1668 "Unsupported size for EmitAtomicBinaryPartial.");
1670 MachineFunction *MF = BB->getParent();
1671 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1672 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1673 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1674 DebugLoc dl = MI->getDebugLoc();
1675 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1676 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1678 unsigned Dest = MI->getOperand(0).getReg();
1679 unsigned Ptr = MI->getOperand(1).getReg();
1680 unsigned Incr = MI->getOperand(2).getReg();
1682 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1683 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1684 unsigned Mask = RegInfo.createVirtualRegister(RC);
1685 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1686 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1687 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1688 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1689 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1690 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1691 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1692 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1693 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1694 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1695 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1696 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1697 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1698 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1699 unsigned Success = RegInfo.createVirtualRegister(RC);
1701 // insert new blocks after the current block
1702 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1703 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1704 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1705 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1706 MachineFunction::iterator It = BB;
1708 MF->insert(It, loopMBB);
1709 MF->insert(It, sinkMBB);
1710 MF->insert(It, exitMBB);
1712 // Transfer the remainder of BB and its successor edges to exitMBB.
1713 exitMBB->splice(exitMBB->begin(), BB,
1714 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1715 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1717 BB->addSuccessor(loopMBB);
1718 loopMBB->addSuccessor(loopMBB);
1719 loopMBB->addSuccessor(sinkMBB);
1720 sinkMBB->addSuccessor(exitMBB);
1723 // addiu masklsb2,$0,-4 # 0xfffffffc
1724 // and alignedaddr,ptr,masklsb2
1725 // andi ptrlsb2,ptr,3
1726 // sll shiftamt,ptrlsb2,3
1727 // ori maskupper,$0,255 # 0xff
1728 // sll mask,maskupper,shiftamt
1729 // nor mask2,$0,mask
1730 // sll incr2,incr,shiftamt
1732 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1733 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1734 .addReg(Mips::ZERO).addImm(-4);
1735 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1736 .addReg(Ptr).addReg(MaskLSB2);
1737 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1738 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1739 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1740 .addReg(Mips::ZERO).addImm(MaskImm);
1741 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1742 .addReg(ShiftAmt).addReg(MaskUpper);
1743 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1744 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1746 // atomic.load.binop
1748 // ll oldval,0(alignedaddr)
1749 // binop binopres,oldval,incr2
1750 // and newval,binopres,mask
1751 // and maskedoldval0,oldval,mask2
1752 // or storeval,maskedoldval0,newval
1753 // sc success,storeval,0(alignedaddr)
1754 // beq success,$0,loopMBB
1758 // ll oldval,0(alignedaddr)
1759 // and newval,incr2,mask
1760 // and maskedoldval0,oldval,mask2
1761 // or storeval,maskedoldval0,newval
1762 // sc success,storeval,0(alignedaddr)
1763 // beq success,$0,loopMBB
1766 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1768 // and andres, oldval, incr2
1769 // nor binopres, $0, andres
1770 // and newval, binopres, mask
1771 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1772 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1773 .addReg(Mips::ZERO).addReg(AndRes);
1774 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1775 } else if (BinOpcode) {
1776 // <binop> binopres, oldval, incr2
1777 // and newval, binopres, mask
1778 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1779 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1780 } else {// atomic.swap
1781 // and newval, incr2, mask
1782 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1785 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1786 .addReg(OldVal).addReg(Mask2);
1787 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1788 .addReg(MaskedOldVal0).addReg(NewVal);
1789 BuildMI(BB, dl, TII->get(SC), Success)
1790 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1791 BuildMI(BB, dl, TII->get(Mips::BEQ))
1792 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1795 // and maskedoldval1,oldval,mask
1796 // srl srlres,maskedoldval1,shiftamt
1797 // sll sllres,srlres,24
1798 // sra dest,sllres,24
1800 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1802 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1803 .addReg(OldVal).addReg(Mask);
1804 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1805 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1806 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1807 .addReg(SrlRes).addImm(ShiftImm);
1808 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1809 .addReg(SllRes).addImm(ShiftImm);
1811 MI->eraseFromParent(); // The instruction is gone now.
1817 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1818 MachineBasicBlock *BB,
1819 unsigned Size) const {
1820 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1822 MachineFunction *MF = BB->getParent();
1823 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1824 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1825 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1826 DebugLoc dl = MI->getDebugLoc();
1827 unsigned LL, SC, ZERO, BNE, BEQ;
1830 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1831 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1837 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1838 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1839 ZERO = Mips::ZERO_64;
1844 unsigned Dest = MI->getOperand(0).getReg();
1845 unsigned Ptr = MI->getOperand(1).getReg();
1846 unsigned OldVal = MI->getOperand(2).getReg();
1847 unsigned NewVal = MI->getOperand(3).getReg();
1849 unsigned Success = RegInfo.createVirtualRegister(RC);
1851 // insert new blocks after the current block
1852 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1853 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1854 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1855 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1856 MachineFunction::iterator It = BB;
1858 MF->insert(It, loop1MBB);
1859 MF->insert(It, loop2MBB);
1860 MF->insert(It, exitMBB);
1862 // Transfer the remainder of BB and its successor edges to exitMBB.
1863 exitMBB->splice(exitMBB->begin(), BB,
1864 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1865 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1869 // fallthrough --> loop1MBB
1870 BB->addSuccessor(loop1MBB);
1871 loop1MBB->addSuccessor(exitMBB);
1872 loop1MBB->addSuccessor(loop2MBB);
1873 loop2MBB->addSuccessor(loop1MBB);
1874 loop2MBB->addSuccessor(exitMBB);
1878 // bne dest, oldval, exitMBB
1880 BuildMI(BB, dl, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1881 BuildMI(BB, dl, TII->get(BNE))
1882 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1885 // sc success, newval, 0(ptr)
1886 // beq success, $0, loop1MBB
1888 BuildMI(BB, dl, TII->get(SC), Success)
1889 .addReg(NewVal).addReg(Ptr).addImm(0);
1890 BuildMI(BB, dl, TII->get(BEQ))
1891 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1893 MI->eraseFromParent(); // The instruction is gone now.
1899 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1900 MachineBasicBlock *BB,
1901 unsigned Size) const {
1902 assert((Size == 1 || Size == 2) &&
1903 "Unsupported size for EmitAtomicCmpSwapPartial.");
1905 MachineFunction *MF = BB->getParent();
1906 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1907 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1908 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1909 DebugLoc dl = MI->getDebugLoc();
1910 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1911 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1913 unsigned Dest = MI->getOperand(0).getReg();
1914 unsigned Ptr = MI->getOperand(1).getReg();
1915 unsigned CmpVal = MI->getOperand(2).getReg();
1916 unsigned NewVal = MI->getOperand(3).getReg();
1918 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1919 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1920 unsigned Mask = RegInfo.createVirtualRegister(RC);
1921 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1922 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1923 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1924 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1925 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1926 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1927 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1928 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1929 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1930 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1931 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1932 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1933 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1934 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1935 unsigned Success = RegInfo.createVirtualRegister(RC);
1937 // insert new blocks after the current block
1938 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1939 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1940 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1941 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1942 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1943 MachineFunction::iterator It = BB;
1945 MF->insert(It, loop1MBB);
1946 MF->insert(It, loop2MBB);
1947 MF->insert(It, sinkMBB);
1948 MF->insert(It, exitMBB);
1950 // Transfer the remainder of BB and its successor edges to exitMBB.
1951 exitMBB->splice(exitMBB->begin(), BB,
1952 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1953 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1955 BB->addSuccessor(loop1MBB);
1956 loop1MBB->addSuccessor(sinkMBB);
1957 loop1MBB->addSuccessor(loop2MBB);
1958 loop2MBB->addSuccessor(loop1MBB);
1959 loop2MBB->addSuccessor(sinkMBB);
1960 sinkMBB->addSuccessor(exitMBB);
1962 // FIXME: computation of newval2 can be moved to loop2MBB.
1964 // addiu masklsb2,$0,-4 # 0xfffffffc
1965 // and alignedaddr,ptr,masklsb2
1966 // andi ptrlsb2,ptr,3
1967 // sll shiftamt,ptrlsb2,3
1968 // ori maskupper,$0,255 # 0xff
1969 // sll mask,maskupper,shiftamt
1970 // nor mask2,$0,mask
1971 // andi maskedcmpval,cmpval,255
1972 // sll shiftedcmpval,maskedcmpval,shiftamt
1973 // andi maskednewval,newval,255
1974 // sll shiftednewval,maskednewval,shiftamt
1975 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1976 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1977 .addReg(Mips::ZERO).addImm(-4);
1978 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1979 .addReg(Ptr).addReg(MaskLSB2);
1980 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1981 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1982 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1983 .addReg(Mips::ZERO).addImm(MaskImm);
1984 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1985 .addReg(ShiftAmt).addReg(MaskUpper);
1986 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1987 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1988 .addReg(CmpVal).addImm(MaskImm);
1989 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1990 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1991 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1992 .addReg(NewVal).addImm(MaskImm);
1993 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1994 .addReg(ShiftAmt).addReg(MaskedNewVal);
1997 // ll oldval,0(alginedaddr)
1998 // and maskedoldval0,oldval,mask
1999 // bne maskedoldval0,shiftedcmpval,sinkMBB
2001 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
2002 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
2003 .addReg(OldVal).addReg(Mask);
2004 BuildMI(BB, dl, TII->get(Mips::BNE))
2005 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
2008 // and maskedoldval1,oldval,mask2
2009 // or storeval,maskedoldval1,shiftednewval
2010 // sc success,storeval,0(alignedaddr)
2011 // beq success,$0,loop1MBB
2013 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
2014 .addReg(OldVal).addReg(Mask2);
2015 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
2016 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
2017 BuildMI(BB, dl, TII->get(SC), Success)
2018 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
2019 BuildMI(BB, dl, TII->get(Mips::BEQ))
2020 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
2023 // srl srlres,maskedoldval0,shiftamt
2024 // sll sllres,srlres,24
2025 // sra dest,sllres,24
2027 int64_t ShiftImm = (Size == 1) ? 24 : 16;
2029 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
2030 .addReg(ShiftAmt).addReg(MaskedOldVal0);
2031 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
2032 .addReg(SrlRes).addImm(ShiftImm);
2033 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
2034 .addReg(SllRes).addImm(ShiftImm);
2036 MI->eraseFromParent(); // The instruction is gone now.
2041 //===----------------------------------------------------------------------===//
2042 // Misc Lower Operation implementation
2043 //===----------------------------------------------------------------------===//
2044 SDValue MipsTargetLowering::
2045 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
2047 // The first operand is the chain, the second is the condition, the third is
2048 // the block to branch to if the condition is true.
2049 SDValue Chain = Op.getOperand(0);
2050 SDValue Dest = Op.getOperand(2);
2051 DebugLoc dl = Op.getDebugLoc();
2053 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
2055 // Return if flag is not set by a floating point comparison.
2056 if (CondRes.getOpcode() != MipsISD::FPCmp)
2059 SDValue CCNode = CondRes.getOperand(2);
2061 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
2062 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
2064 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
2068 SDValue MipsTargetLowering::
2069 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
2071 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
2073 // Return if flag is not set by a floating point comparison.
2074 if (Cond.getOpcode() != MipsISD::FPCmp)
2077 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
2081 SDValue MipsTargetLowering::
2082 LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
2084 DebugLoc DL = Op.getDebugLoc();
2085 EVT Ty = Op.getOperand(0).getValueType();
2086 SDValue Cond = DAG.getNode(ISD::SETCC, DL, getSetCCResultType(Ty),
2087 Op.getOperand(0), Op.getOperand(1),
2090 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
2094 SDValue MipsTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2095 SDValue Cond = CreateFPCmp(DAG, Op);
2097 assert(Cond.getOpcode() == MipsISD::FPCmp &&
2098 "Floating point operand expected.");
2100 SDValue True = DAG.getConstant(1, MVT::i32);
2101 SDValue False = DAG.getConstant(0, MVT::i32);
2103 return CreateCMovFP(DAG, Cond, True, False, Op.getDebugLoc());
2106 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
2107 SelectionDAG &DAG) const {
2108 // FIXME there isn't actually debug info here
2109 DebugLoc dl = Op.getDebugLoc();
2110 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2112 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
2113 const MipsTargetObjectFile &TLOF =
2114 (const MipsTargetObjectFile&)getObjFileLowering();
2116 // %gp_rel relocation
2117 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
2118 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
2120 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl,
2121 DAG.getVTList(MVT::i32), &GA, 1);
2122 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
2123 return DAG.getNode(ISD::ADD, dl, MVT::i32, GPReg, GPRelNode);
2126 // %hi/%lo relocation
2127 return getAddrNonPIC(Op, DAG);
2130 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
2131 return getAddrLocal(Op, DAG, HasMips64);
2134 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16,
2135 MipsII::MO_GOT_LO16);
2137 return getAddrGlobal(Op, DAG,
2138 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
2141 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
2142 SelectionDAG &DAG) const {
2143 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
2144 return getAddrNonPIC(Op, DAG);
2146 return getAddrLocal(Op, DAG, HasMips64);
2149 SDValue MipsTargetLowering::
2150 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
2152 // If the relocation model is PIC, use the General Dynamic TLS Model or
2153 // Local Dynamic TLS model, otherwise use the Initial Exec or
2154 // Local Exec TLS Model.
2156 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2157 DebugLoc dl = GA->getDebugLoc();
2158 const GlobalValue *GV = GA->getGlobal();
2159 EVT PtrVT = getPointerTy();
2161 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
2163 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
2164 // General Dynamic and Local Dynamic TLS Model.
2165 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
2168 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, Flag);
2169 SDValue Argument = DAG.getNode(MipsISD::Wrapper, dl, PtrVT,
2170 GetGlobalReg(DAG, PtrVT), TGA);
2171 unsigned PtrSize = PtrVT.getSizeInBits();
2172 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
2174 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
2178 Entry.Node = Argument;
2180 Args.push_back(Entry);
2182 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
2183 false, false, false, false, 0, CallingConv::C,
2184 /*isTailCall=*/false, /*doesNotRet=*/false,
2185 /*isReturnValueUsed=*/true,
2186 TlsGetAddr, Args, DAG, dl);
2187 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2189 SDValue Ret = CallResult.first;
2191 if (model != TLSModel::LocalDynamic)
2194 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2195 MipsII::MO_DTPREL_HI);
2196 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
2197 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2198 MipsII::MO_DTPREL_LO);
2199 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
2200 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Ret);
2201 return DAG.getNode(ISD::ADD, dl, PtrVT, Add, Lo);
2205 if (model == TLSModel::InitialExec) {
2206 // Initial Exec TLS Model
2207 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2208 MipsII::MO_GOTTPREL);
2209 TGA = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, GetGlobalReg(DAG, PtrVT),
2211 Offset = DAG.getLoad(PtrVT, dl,
2212 DAG.getEntryNode(), TGA, MachinePointerInfo(),
2213 false, false, false, 0);
2215 // Local Exec TLS Model
2216 assert(model == TLSModel::LocalExec);
2217 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2218 MipsII::MO_TPREL_HI);
2219 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2220 MipsII::MO_TPREL_LO);
2221 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
2222 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
2223 Offset = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
2226 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
2227 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2230 SDValue MipsTargetLowering::
2231 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
2233 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
2234 return getAddrNonPIC(Op, DAG);
2236 return getAddrLocal(Op, DAG, HasMips64);
2239 SDValue MipsTargetLowering::
2240 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
2242 // gp_rel relocation
2243 // FIXME: we should reference the constant pool using small data sections,
2244 // but the asm printer currently doesn't support this feature without
2245 // hacking it. This feature should come soon so we can uncomment the
2247 //if (IsInSmallSection(C->getType())) {
2248 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
2249 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
2250 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
2252 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
2253 return getAddrNonPIC(Op, DAG);
2255 return getAddrLocal(Op, DAG, HasMips64);
2258 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
2259 MachineFunction &MF = DAG.getMachineFunction();
2260 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2262 DebugLoc dl = Op.getDebugLoc();
2263 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
2266 // vastart just stores the address of the VarArgsFrameIndex slot into the
2267 // memory location argument.
2268 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2269 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
2270 MachinePointerInfo(SV), false, false, 0);
2273 static SDValue LowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2274 EVT TyX = Op.getOperand(0).getValueType();
2275 EVT TyY = Op.getOperand(1).getValueType();
2276 SDValue Const1 = DAG.getConstant(1, MVT::i32);
2277 SDValue Const31 = DAG.getConstant(31, MVT::i32);
2278 DebugLoc DL = Op.getDebugLoc();
2281 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2283 SDValue X = (TyX == MVT::f32) ?
2284 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
2285 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2287 SDValue Y = (TyY == MVT::f32) ?
2288 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
2289 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
2293 // ext E, Y, 31, 1 ; extract bit31 of Y
2294 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
2295 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
2296 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
2299 // srl SrlX, SllX, 1
2301 // sll SllY, SrlX, 31
2302 // or Or, SrlX, SllY
2303 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2304 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2305 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
2306 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
2307 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
2310 if (TyX == MVT::f32)
2311 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
2313 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2314 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
2315 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2318 static SDValue LowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2319 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
2320 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
2321 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
2322 SDValue Const1 = DAG.getConstant(1, MVT::i32);
2323 DebugLoc DL = Op.getDebugLoc();
2325 // Bitcast to integer nodes.
2326 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
2327 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
2330 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
2331 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
2332 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
2333 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
2335 if (WidthX > WidthY)
2336 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
2337 else if (WidthY > WidthX)
2338 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
2340 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
2341 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
2342 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
2345 // (d)sll SllX, X, 1
2346 // (d)srl SrlX, SllX, 1
2347 // (d)srl SrlY, Y, width(Y)-1
2348 // (d)sll SllY, SrlX, width(Y)-1
2349 // or Or, SrlX, SllY
2350 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
2351 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
2352 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
2353 DAG.getConstant(WidthY - 1, MVT::i32));
2355 if (WidthX > WidthY)
2356 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
2357 else if (WidthY > WidthX)
2358 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
2360 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
2361 DAG.getConstant(WidthX - 1, MVT::i32));
2362 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
2363 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
2367 MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2368 if (Subtarget->hasMips64())
2369 return LowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
2371 return LowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
2374 static SDValue LowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2375 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
2376 DebugLoc DL = Op.getDebugLoc();
2378 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2380 SDValue X = (Op.getValueType() == MVT::f32) ?
2381 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
2382 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2387 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
2388 DAG.getRegister(Mips::ZERO, MVT::i32),
2389 DAG.getConstant(31, MVT::i32), Const1, X);
2391 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2392 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2395 if (Op.getValueType() == MVT::f32)
2396 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
2398 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2399 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
2400 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2403 static SDValue LowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2404 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
2405 DebugLoc DL = Op.getDebugLoc();
2407 // Bitcast to integer node.
2408 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
2412 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
2413 DAG.getRegister(Mips::ZERO_64, MVT::i64),
2414 DAG.getConstant(63, MVT::i32), Const1, X);
2416 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
2417 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
2420 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
2424 MipsTargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
2425 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
2426 return LowerFABS64(Op, DAG, Subtarget->hasMips32r2());
2428 return LowerFABS32(Op, DAG, Subtarget->hasMips32r2());
2431 SDValue MipsTargetLowering::
2432 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2434 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2435 "Frame address can only be determined for current frame.");
2437 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2438 MFI->setFrameAddressIsTaken(true);
2439 EVT VT = Op.getValueType();
2440 DebugLoc dl = Op.getDebugLoc();
2441 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2442 IsN64 ? Mips::FP_64 : Mips::FP, VT);
2446 SDValue MipsTargetLowering::LowerRETURNADDR(SDValue Op,
2447 SelectionDAG &DAG) const {
2449 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2450 "Return address can be determined only for current frame.");
2452 MachineFunction &MF = DAG.getMachineFunction();
2453 MachineFrameInfo *MFI = MF.getFrameInfo();
2454 MVT VT = Op.getSimpleValueType();
2455 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
2456 MFI->setReturnAddressIsTaken(true);
2458 // Return RA, which contains the return address. Mark it an implicit live-in.
2459 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2460 return DAG.getCopyFromReg(DAG.getEntryNode(), Op.getDebugLoc(), Reg, VT);
2463 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
2464 // generated from __builtin_eh_return (offset, handler)
2465 // The effect of this is to adjust the stack pointer by "offset"
2466 // and then branch to "handler".
2467 SDValue MipsTargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
2469 MachineFunction &MF = DAG.getMachineFunction();
2470 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2472 MipsFI->setCallsEhReturn();
2473 SDValue Chain = Op.getOperand(0);
2474 SDValue Offset = Op.getOperand(1);
2475 SDValue Handler = Op.getOperand(2);
2476 DebugLoc DL = Op.getDebugLoc();
2477 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2479 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2480 // EH_RETURN nodes, so that instructions are emitted back-to-back.
2481 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1;
2482 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0;
2483 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2484 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2485 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2486 DAG.getRegister(OffsetReg, Ty),
2487 DAG.getRegister(AddrReg, getPointerTy()),
2491 // TODO: set SType according to the desired memory barrier behavior.
2493 MipsTargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const {
2495 DebugLoc dl = Op.getDebugLoc();
2496 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
2497 DAG.getConstant(SType, MVT::i32));
2500 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
2501 SelectionDAG &DAG) const {
2502 // FIXME: Need pseudo-fence for 'singlethread' fences
2503 // FIXME: Set SType for weaker fences where supported/appropriate.
2505 DebugLoc dl = Op.getDebugLoc();
2506 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
2507 DAG.getConstant(SType, MVT::i32));
2510 SDValue MipsTargetLowering::LowerShiftLeftParts(SDValue Op,
2511 SelectionDAG &DAG) const {
2512 DebugLoc DL = Op.getDebugLoc();
2513 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2514 SDValue Shamt = Op.getOperand(2);
2517 // lo = (shl lo, shamt)
2518 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2521 // hi = (shl lo, shamt[4:0])
2522 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2523 DAG.getConstant(-1, MVT::i32));
2524 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
2525 DAG.getConstant(1, MVT::i32));
2526 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
2528 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
2529 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2530 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
2531 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2532 DAG.getConstant(0x20, MVT::i32));
2533 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2534 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
2535 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
2537 SDValue Ops[2] = {Lo, Hi};
2538 return DAG.getMergeValues(Ops, 2, DL);
2541 SDValue MipsTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2543 DebugLoc DL = Op.getDebugLoc();
2544 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2545 SDValue Shamt = Op.getOperand(2);
2548 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2550 // hi = (sra hi, shamt)
2552 // hi = (srl hi, shamt)
2555 // lo = (sra hi, shamt[4:0])
2556 // hi = (sra hi, 31)
2558 // lo = (srl hi, shamt[4:0])
2560 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2561 DAG.getConstant(-1, MVT::i32));
2562 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
2563 DAG.getConstant(1, MVT::i32));
2564 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
2565 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
2566 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2567 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2569 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2570 DAG.getConstant(0x20, MVT::i32));
2571 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
2572 DAG.getConstant(31, MVT::i32));
2573 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
2574 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2575 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
2578 SDValue Ops[2] = {Lo, Hi};
2579 return DAG.getMergeValues(Ops, 2, DL);
2582 static SDValue CreateLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2583 SDValue Chain, SDValue Src, unsigned Offset) {
2584 SDValue Ptr = LD->getBasePtr();
2585 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2586 EVT BasePtrVT = Ptr.getValueType();
2587 DebugLoc DL = LD->getDebugLoc();
2588 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2591 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2592 DAG.getConstant(Offset, BasePtrVT));
2594 SDValue Ops[] = { Chain, Ptr, Src };
2595 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2596 LD->getMemOperand());
2599 // Expand an unaligned 32 or 64-bit integer load node.
2600 SDValue MipsTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2601 LoadSDNode *LD = cast<LoadSDNode>(Op);
2602 EVT MemVT = LD->getMemoryVT();
2604 // Return if load is aligned or if MemVT is neither i32 nor i64.
2605 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2606 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2609 bool IsLittle = Subtarget->isLittle();
2610 EVT VT = Op.getValueType();
2611 ISD::LoadExtType ExtType = LD->getExtensionType();
2612 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2614 assert((VT == MVT::i32) || (VT == MVT::i64));
2617 // (set dst, (i64 (load baseptr)))
2619 // (set tmp, (ldl (add baseptr, 7), undef))
2620 // (set dst, (ldr baseptr, tmp))
2621 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2622 SDValue LDL = CreateLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2624 return CreateLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2628 SDValue LWL = CreateLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2630 SDValue LWR = CreateLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2634 // (set dst, (i32 (load baseptr))) or
2635 // (set dst, (i64 (sextload baseptr))) or
2636 // (set dst, (i64 (extload baseptr)))
2638 // (set tmp, (lwl (add baseptr, 3), undef))
2639 // (set dst, (lwr baseptr, tmp))
2640 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2641 (ExtType == ISD::EXTLOAD))
2644 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2647 // (set dst, (i64 (zextload baseptr)))
2649 // (set tmp0, (lwl (add baseptr, 3), undef))
2650 // (set tmp1, (lwr baseptr, tmp0))
2651 // (set tmp2, (shl tmp1, 32))
2652 // (set dst, (srl tmp2, 32))
2653 DebugLoc DL = LD->getDebugLoc();
2654 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2655 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2656 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2657 SDValue Ops[] = { SRL, LWR.getValue(1) };
2658 return DAG.getMergeValues(Ops, 2, DL);
2661 static SDValue CreateStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2662 SDValue Chain, unsigned Offset) {
2663 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2664 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2665 DebugLoc DL = SD->getDebugLoc();
2666 SDVTList VTList = DAG.getVTList(MVT::Other);
2669 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2670 DAG.getConstant(Offset, BasePtrVT));
2672 SDValue Ops[] = { Chain, Value, Ptr };
2673 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2674 SD->getMemOperand());
2677 // Expand an unaligned 32 or 64-bit integer store node.
2678 SDValue MipsTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2679 StoreSDNode *SD = cast<StoreSDNode>(Op);
2680 EVT MemVT = SD->getMemoryVT();
2682 // Return if store is aligned or if MemVT is neither i32 nor i64.
2683 if ((SD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2684 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2687 bool IsLittle = Subtarget->isLittle();
2688 SDValue Value = SD->getValue(), Chain = SD->getChain();
2689 EVT VT = Value.getValueType();
2692 // (store val, baseptr) or
2693 // (truncstore val, baseptr)
2695 // (swl val, (add baseptr, 3))
2696 // (swr val, baseptr)
2697 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2698 SDValue SWL = CreateStoreLR(MipsISD::SWL, DAG, SD, Chain,
2700 return CreateStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2703 assert(VT == MVT::i64);
2706 // (store val, baseptr)
2708 // (sdl val, (add baseptr, 7))
2709 // (sdr val, baseptr)
2710 SDValue SDL = CreateStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2711 return CreateStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2714 // This function expands mips intrinsic nodes which have 64-bit input operands
2715 // or output values.
2717 // out64 = intrinsic-node in64
2719 // lo = copy (extract-element (in64, 0))
2720 // hi = copy (extract-element (in64, 1))
2721 // mips-specific-node
2724 // out64 = merge-values (v0, v1)
2726 static SDValue LowerDSPIntr(SDValue Op, SelectionDAG &DAG,
2727 unsigned Opc, bool HasI64In, bool HasI64Out) {
2728 DebugLoc DL = Op.getDebugLoc();
2729 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
2730 SDValue Chain = HasChainIn ? Op->getOperand(0) : DAG.getEntryNode();
2731 SmallVector<SDValue, 3> Ops;
2734 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2735 Op->getOperand(1 + HasChainIn),
2736 DAG.getConstant(0, MVT::i32));
2737 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2738 Op->getOperand(1 + HasChainIn),
2739 DAG.getConstant(1, MVT::i32));
2741 Chain = DAG.getCopyToReg(Chain, DL, Mips::LO, InLo, SDValue());
2742 Chain = DAG.getCopyToReg(Chain, DL, Mips::HI, InHi, Chain.getValue(1));
2744 Ops.push_back(Chain);
2745 Ops.append(Op->op_begin() + HasChainIn + 2, Op->op_end());
2746 Ops.push_back(Chain.getValue(1));
2748 Ops.push_back(Chain);
2749 Ops.append(Op->op_begin() + HasChainIn + 1, Op->op_end());
2753 return DAG.getNode(Opc, DL, Op->value_begin(), Op->getNumValues(),
2754 Ops.begin(), Ops.size());
2756 SDValue Intr = DAG.getNode(Opc, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2757 Ops.begin(), Ops.size());
2758 SDValue OutLo = DAG.getCopyFromReg(Intr.getValue(0), DL, Mips::LO, MVT::i32,
2760 SDValue OutHi = DAG.getCopyFromReg(OutLo.getValue(1), DL, Mips::HI, MVT::i32,
2762 SDValue Out = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, OutLo, OutHi);
2767 SDValue Vals[] = { Out, OutHi.getValue(1) };
2768 return DAG.getMergeValues(Vals, 2, DL);
2771 SDValue MipsTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2772 SelectionDAG &DAG) const {
2773 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
2776 case Intrinsic::mips_shilo:
2777 return LowerDSPIntr(Op, DAG, MipsISD::SHILO, true, true);
2778 case Intrinsic::mips_dpau_h_qbl:
2779 return LowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL, true, true);
2780 case Intrinsic::mips_dpau_h_qbr:
2781 return LowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR, true, true);
2782 case Intrinsic::mips_dpsu_h_qbl:
2783 return LowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL, true, true);
2784 case Intrinsic::mips_dpsu_h_qbr:
2785 return LowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR, true, true);
2786 case Intrinsic::mips_dpa_w_ph:
2787 return LowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH, true, true);
2788 case Intrinsic::mips_dps_w_ph:
2789 return LowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH, true, true);
2790 case Intrinsic::mips_dpax_w_ph:
2791 return LowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH, true, true);
2792 case Intrinsic::mips_dpsx_w_ph:
2793 return LowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH, true, true);
2794 case Intrinsic::mips_mulsa_w_ph:
2795 return LowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH, true, true);
2796 case Intrinsic::mips_mult:
2797 return LowerDSPIntr(Op, DAG, MipsISD::MULT, false, true);
2798 case Intrinsic::mips_multu:
2799 return LowerDSPIntr(Op, DAG, MipsISD::MULTU, false, true);
2800 case Intrinsic::mips_madd:
2801 return LowerDSPIntr(Op, DAG, MipsISD::MADD_DSP, true, true);
2802 case Intrinsic::mips_maddu:
2803 return LowerDSPIntr(Op, DAG, MipsISD::MADDU_DSP, true, true);
2804 case Intrinsic::mips_msub:
2805 return LowerDSPIntr(Op, DAG, MipsISD::MSUB_DSP, true, true);
2806 case Intrinsic::mips_msubu:
2807 return LowerDSPIntr(Op, DAG, MipsISD::MSUBU_DSP, true, true);
2811 SDValue MipsTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
2812 SelectionDAG &DAG) const {
2813 switch (cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue()) {
2816 case Intrinsic::mips_extp:
2817 return LowerDSPIntr(Op, DAG, MipsISD::EXTP, true, false);
2818 case Intrinsic::mips_extpdp:
2819 return LowerDSPIntr(Op, DAG, MipsISD::EXTPDP, true, false);
2820 case Intrinsic::mips_extr_w:
2821 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_W, true, false);
2822 case Intrinsic::mips_extr_r_w:
2823 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W, true, false);
2824 case Intrinsic::mips_extr_rs_w:
2825 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W, true, false);
2826 case Intrinsic::mips_extr_s_h:
2827 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H, true, false);
2828 case Intrinsic::mips_mthlip:
2829 return LowerDSPIntr(Op, DAG, MipsISD::MTHLIP, true, true);
2830 case Intrinsic::mips_mulsaq_s_w_ph:
2831 return LowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH, true, true);
2832 case Intrinsic::mips_maq_s_w_phl:
2833 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL, true, true);
2834 case Intrinsic::mips_maq_s_w_phr:
2835 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR, true, true);
2836 case Intrinsic::mips_maq_sa_w_phl:
2837 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL, true, true);
2838 case Intrinsic::mips_maq_sa_w_phr:
2839 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR, true, true);
2840 case Intrinsic::mips_dpaq_s_w_ph:
2841 return LowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH, true, true);
2842 case Intrinsic::mips_dpsq_s_w_ph:
2843 return LowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH, true, true);
2844 case Intrinsic::mips_dpaq_sa_l_w:
2845 return LowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W, true, true);
2846 case Intrinsic::mips_dpsq_sa_l_w:
2847 return LowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W, true, true);
2848 case Intrinsic::mips_dpaqx_s_w_ph:
2849 return LowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH, true, true);
2850 case Intrinsic::mips_dpaqx_sa_w_ph:
2851 return LowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH, true, true);
2852 case Intrinsic::mips_dpsqx_s_w_ph:
2853 return LowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH, true, true);
2854 case Intrinsic::mips_dpsqx_sa_w_ph:
2855 return LowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH, true, true);
2859 SDValue MipsTargetLowering::LowerADD(SDValue Op, SelectionDAG &DAG) const {
2860 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2861 || cast<ConstantSDNode>
2862 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2863 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2867 // (add (frameaddr 0), (frame_to_args_offset))
2868 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2869 // (add FrameObject, 0)
2870 // where FrameObject is a fixed StackObject with offset 0 which points to
2871 // the old stack pointer.
2872 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2873 EVT ValTy = Op->getValueType(0);
2874 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2875 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2876 return DAG.getNode(ISD::ADD, Op->getDebugLoc(), ValTy, InArgsAddr,
2877 DAG.getConstant(0, ValTy));
2880 //===----------------------------------------------------------------------===//
2881 // Calling Convention Implementation
2882 //===----------------------------------------------------------------------===//
2884 //===----------------------------------------------------------------------===//
2885 // TODO: Implement a generic logic using tblgen that can support this.
2886 // Mips O32 ABI rules:
2888 // i32 - Passed in A0, A1, A2, A3 and stack
2889 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2890 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2891 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2892 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2893 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2896 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2897 //===----------------------------------------------------------------------===//
2899 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2900 MVT LocVT, CCValAssign::LocInfo LocInfo,
2901 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2903 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2905 static const uint16_t IntRegs[] = {
2906 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2908 static const uint16_t F32Regs[] = {
2909 Mips::F12, Mips::F14
2911 static const uint16_t F64Regs[] = {
2915 // Do not process byval args here.
2916 if (ArgFlags.isByVal())
2919 // Promote i8 and i16
2920 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2922 if (ArgFlags.isSExt())
2923 LocInfo = CCValAssign::SExt;
2924 else if (ArgFlags.isZExt())
2925 LocInfo = CCValAssign::ZExt;
2927 LocInfo = CCValAssign::AExt;
2932 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2933 // is true: function is vararg, argument is 3rd or higher, there is previous
2934 // argument which is not f32 or f64.
2935 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2936 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2937 unsigned OrigAlign = ArgFlags.getOrigAlign();
2938 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2940 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2941 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2942 // If this is the first part of an i64 arg,
2943 // the allocated register must be either A0 or A2.
2944 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2945 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2947 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2948 // Allocate int register and shadow next int register. If first
2949 // available register is Mips::A1 or Mips::A3, shadow it too.
2950 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2951 if (Reg == Mips::A1 || Reg == Mips::A3)
2952 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2953 State.AllocateReg(IntRegs, IntRegsSize);
2955 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2956 // we are guaranteed to find an available float register
2957 if (ValVT == MVT::f32) {
2958 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2959 // Shadow int register
2960 State.AllocateReg(IntRegs, IntRegsSize);
2962 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2963 // Shadow int registers
2964 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2965 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2966 State.AllocateReg(IntRegs, IntRegsSize);
2967 State.AllocateReg(IntRegs, IntRegsSize);
2970 llvm_unreachable("Cannot handle this ValVT.");
2973 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2975 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2977 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2982 #include "MipsGenCallingConv.inc"
2984 //===----------------------------------------------------------------------===//
2985 // Call Calling Convention Implementation
2986 //===----------------------------------------------------------------------===//
2988 static const unsigned O32IntRegsSize = 4;
2990 // Return next O32 integer argument register.
2991 static unsigned getNextIntArgReg(unsigned Reg) {
2992 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2993 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2996 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2997 /// for tail call optimization.
2998 bool MipsTargetLowering::
2999 IsEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
3000 unsigned NextStackOffset,
3001 const MipsFunctionInfo& FI) const {
3002 if (!EnableMipsTailCalls)
3005 // No tail call optimization for mips16.
3006 if (Subtarget->inMips16Mode())
3009 // Return false if either the callee or caller has a byval argument.
3010 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
3013 // Return true if the callee's argument area is no larger than the
3015 return NextStackOffset <= FI.getIncomingArgSize();
3019 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
3020 SDValue Chain, SDValue Arg, DebugLoc DL,
3021 bool IsTailCall, SelectionDAG &DAG) const {
3023 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
3024 DAG.getIntPtrConstant(Offset));
3025 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
3029 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3030 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
3031 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3032 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
3033 /*isVolatile=*/ true, false, 0);
3037 // The Mips16 hard float is a crazy quilt inherited from gcc. I have a much
3038 // cleaner way to do all of this but it will have to wait until the traditional
3039 // gcc mechanism is completed.
3041 // For Pic, in order for Mips16 code to call Mips32 code which according the abi
3042 // have either arguments or returned values placed in floating point registers,
3043 // we use a set of helper functions. (This includes functions which return type
3044 // complex which on Mips are returned in a pair of floating point registers).
3046 // This is an encoding that we inherited from gcc.
3047 // In Mips traditional O32, N32 ABI, floating point numbers are passed in
3048 // floating point argument registers 1,2 only when the first and optionally
3049 // the second arguments are float (sf) or double (df).
3050 // For Mips16 we are only concerned with the situations where floating point
3051 // arguments are being passed in floating point registers by the ABI, because
3052 // Mips16 mode code cannot execute floating point instructions to load those
3053 // values and hence helper functions are needed.
3054 // The possibilities are (), (sf), (sf, sf), (sf, df), (df), (df, sf), (df, df)
3055 // the helper function suffixs for these are:
3056 // 0, 1, 5, 9, 2, 6, 10
3057 // this suffix can then be calculated as follows:
3058 // for a given argument Arg:
3059 // Arg1x, Arg2x = 1 : Arg is sf
3061 // 0: Arg is neither sf or df
3062 // So this stub is the string for number Arg1x + Arg2x*4.
3063 // However not all numbers between 0 and 10 are possible, we check anyway and
3064 // assert if the impossible exists.
3067 unsigned int MipsTargetLowering::getMips16HelperFunctionStubNumber
3068 (ArgListTy &Args) const {
3069 unsigned int resultNum = 0;
3070 if (Args.size() >= 1) {
3071 Type *t = Args[0].Ty;
3072 if (t->isFloatTy()) {
3075 else if (t->isDoubleTy()) {
3080 if (Args.size() >=2) {
3081 Type *t = Args[1].Ty;
3082 if (t->isFloatTy()) {
3085 else if (t->isDoubleTy()) {
3094 // prefixs are attached to stub numbers depending on the return type .
3095 // return type: float sf_
3097 // single complex sc_
3098 // double complext dc_
3102 // The full name of a helper function is__mips16_call_stub +
3103 // return type dependent prefix + stub number
3106 // This is something that probably should be in a different source file and
3107 // perhaps done differently but my main purpose is to not waste runtime
3108 // on something that we can enumerate in the source. Another possibility is
3109 // to have a python script to generate these mapping tables. This will do
3110 // for now. There are a whole series of helper function mapping arrays, one
3111 // for each return type class as outlined above. There there are 11 possible
3112 // entries. Ones with 0 are ones which should never be selected
3114 // All the arrays are similar except for ones which return neither
3115 // sf, df, sc, dc, in which only care about ones which have sf or df as a
3118 #define P_ "__mips16_call_stub_"
3119 #define MAX_STUB_NUMBER 10
3120 #define T1 P "1", P "2", 0, 0, P "5", P "6", 0, 0, P "9", P "10"
3121 #define T P "0" , T1
3123 static char const * vMips16Helper[MAX_STUB_NUMBER+1] =
3127 static char const * sfMips16Helper[MAX_STUB_NUMBER+1] =
3131 static char const * dfMips16Helper[MAX_STUB_NUMBER+1] =
3135 static char const * scMips16Helper[MAX_STUB_NUMBER+1] =
3139 static char const * dcMips16Helper[MAX_STUB_NUMBER+1] =
3145 const char* MipsTargetLowering::
3146 getMips16HelperFunction
3147 (Type* RetTy, ArgListTy &Args, bool &needHelper) const {
3148 const unsigned int stubNum = getMips16HelperFunctionStubNumber(Args);
3150 const unsigned int maxStubNum = 10;
3151 assert(stubNum <= maxStubNum);
3152 const bool validStubNum[maxStubNum+1] =
3153 {true, true, true, false, false, true, true, false, false, true, true};
3154 assert(validStubNum[stubNum]);
3157 if (RetTy->isFloatTy()) {
3158 result = sfMips16Helper[stubNum];
3160 else if (RetTy ->isDoubleTy()) {
3161 result = dfMips16Helper[stubNum];
3163 else if (RetTy->isStructTy()) {
3164 // check if it's complex
3165 if (RetTy->getNumContainedTypes() == 2) {
3166 if ((RetTy->getContainedType(0)->isFloatTy()) &&
3167 (RetTy->getContainedType(1)->isFloatTy())) {
3168 result = scMips16Helper[stubNum];
3170 else if ((RetTy->getContainedType(0)->isDoubleTy()) &&
3171 (RetTy->getContainedType(1)->isDoubleTy())) {
3172 result = dcMips16Helper[stubNum];
3175 llvm_unreachable("Uncovered condition");
3179 llvm_unreachable("Uncovered condition");
3187 result = vMips16Helper[stubNum];
3193 /// LowerCall - functions arguments are copied from virtual regs to
3194 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
3196 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3197 SmallVectorImpl<SDValue> &InVals) const {
3198 SelectionDAG &DAG = CLI.DAG;
3199 DebugLoc &dl = CLI.DL;
3200 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3201 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3202 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3203 SDValue Chain = CLI.Chain;
3204 SDValue Callee = CLI.Callee;
3205 bool &isTailCall = CLI.IsTailCall;
3206 CallingConv::ID CallConv = CLI.CallConv;
3207 bool isVarArg = CLI.IsVarArg;
3209 const char* mips16HelperFunction = 0;
3210 bool needMips16Helper = false;
3212 if (Subtarget->inMips16Mode() && getTargetMachine().Options.UseSoftFloat &&
3215 // currently we don't have symbols tagged with the mips16 or mips32
3216 // qualifier so we will assume that we don't know what kind it is.
3217 // and generate the helper
3219 bool lookupHelper = true;
3220 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3221 if (noHelperNeeded.find(S->getSymbol()) != noHelperNeeded.end()) {
3222 lookupHelper = false;
3225 if (lookupHelper) mips16HelperFunction =
3226 getMips16HelperFunction(CLI.RetTy, CLI.Args, needMips16Helper);
3229 MachineFunction &MF = DAG.getMachineFunction();
3230 MachineFrameInfo *MFI = MF.getFrameInfo();
3231 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
3232 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
3234 // Analyze operands of the call, assigning locations to each operand.
3235 SmallVector<CCValAssign, 16> ArgLocs;
3236 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3237 getTargetMachine(), ArgLocs, *DAG.getContext());
3238 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
3240 MipsCCInfo.analyzeCallOperands(Outs, isVarArg);
3242 // Get a count of how many bytes are to be pushed on the stack.
3243 unsigned NextStackOffset = CCInfo.getNextStackOffset();
3245 // Check if it's really possible to do a tail call.
3248 IsEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
3249 *MF.getInfo<MipsFunctionInfo>());
3254 // Chain is the output chain of the last Load/Store or CopyToReg node.
3255 // ByValChain is the output chain of the last Memcpy node created for copying
3256 // byval arguments to the stack.
3257 unsigned StackAlignment = TFL->getStackAlignment();
3258 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
3259 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
3262 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal);
3264 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl,
3265 IsN64 ? Mips::SP_64 : Mips::SP,
3268 // With EABI is it possible to have 16 args on registers.
3269 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
3270 SmallVector<SDValue, 8> MemOpChains;
3271 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
3273 // Walk the register/memloc assignments, inserting copies/loads.
3274 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3275 SDValue Arg = OutVals[i];
3276 CCValAssign &VA = ArgLocs[i];
3277 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
3278 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3281 if (Flags.isByVal()) {
3282 assert(Flags.getByValSize() &&
3283 "ByVal args of size 0 should have been ignored by front-end.");
3284 assert(ByValArg != MipsCCInfo.byval_end());
3285 assert(!isTailCall &&
3286 "Do not tail-call optimize if there is a byval argument.");
3287 passByValArg(Chain, dl, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
3288 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
3293 // Promote the value if needed.
3294 switch (VA.getLocInfo()) {
3295 default: llvm_unreachable("Unknown loc info!");
3296 case CCValAssign::Full:
3297 if (VA.isRegLoc()) {
3298 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
3299 (ValVT == MVT::f64 && LocVT == MVT::i64))
3300 Arg = DAG.getNode(ISD::BITCAST, dl, LocVT, Arg);
3301 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
3302 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
3303 Arg, DAG.getConstant(0, MVT::i32));
3304 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
3305 Arg, DAG.getConstant(1, MVT::i32));
3306 if (!Subtarget->isLittle())
3308 unsigned LocRegLo = VA.getLocReg();
3309 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
3310 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
3311 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
3316 case CCValAssign::SExt:
3317 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, LocVT, Arg);
3319 case CCValAssign::ZExt:
3320 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, LocVT, Arg);
3322 case CCValAssign::AExt:
3323 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, LocVT, Arg);
3327 // Arguments that can be passed on register must be kept at
3328 // RegsToPass vector
3329 if (VA.isRegLoc()) {
3330 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3334 // Register can't get to this point...
3335 assert(VA.isMemLoc());
3337 // emit ISD::STORE whichs stores the
3338 // parameter value to a stack Location
3339 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
3340 Chain, Arg, dl, isTailCall, DAG));
3343 // Transform all store nodes into one single node because all store
3344 // nodes are independent of each other.
3345 if (!MemOpChains.empty())
3346 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3347 &MemOpChains[0], MemOpChains.size());
3349 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3350 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3351 // node so that legalize doesn't hack it.
3352 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
3353 bool GlobalOrExternal = false, InternalLinkage = false;
3356 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3358 InternalLinkage = G->getGlobal()->hasInternalLinkage();
3360 if (InternalLinkage)
3361 Callee = getAddrLocal(Callee, DAG, HasMips64);
3363 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
3364 MipsII::MO_CALL_LO16);
3366 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
3368 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(), 0,
3369 MipsII::MO_NO_FLAG);
3370 GlobalOrExternal = true;
3372 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3373 if (!IsN64 && !IsPIC) // !N64 && static
3374 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3375 MipsII::MO_NO_FLAG);
3377 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
3378 MipsII::MO_CALL_LO16);
3380 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
3382 GlobalOrExternal = true;
3385 SDValue JumpTarget = Callee;
3387 // T9 should contain the address of the callee function if
3388 // -reloction-model=pic or it is an indirect call.
3389 if (IsPICCall || !GlobalOrExternal) {
3390 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
3391 unsigned V0Reg = Mips::V0;
3392 if (needMips16Helper) {
3393 RegsToPass.push_front(std::make_pair(V0Reg, Callee));
3394 JumpTarget = DAG.getExternalSymbol(
3395 mips16HelperFunction, getPointerTy());
3396 JumpTarget = getAddrGlobal(JumpTarget, DAG, MipsII::MO_GOT);
3399 RegsToPass.push_front(std::make_pair(T9Reg, Callee));
3401 if (!Subtarget->inMips16Mode())
3402 JumpTarget = SDValue();
3406 // Insert node "GP copy globalreg" before call to function.
3408 // R_MIPS_CALL* operators (emitted when non-internal functions are called
3409 // in PIC mode) allow symbols to be resolved via lazy binding.
3410 // The lazy binding stub requires GP to point to the GOT.
3411 if (IsPICCall && !InternalLinkage) {
3412 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
3413 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
3414 RegsToPass.push_back(std::make_pair(GPReg, GetGlobalReg(DAG, Ty)));
3417 // Build a sequence of copy-to-reg nodes chained together with token
3418 // chain and flag operands which copy the outgoing args into registers.
3419 // The InFlag in necessary since all emitted instructions must be
3423 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3424 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3425 RegsToPass[i].second, InFlag);
3426 InFlag = Chain.getValue(1);
3429 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
3430 // = Chain, Callee, Reg#1, Reg#2, ...
3432 // Returns a chain & a flag for retval copy to use.
3433 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3434 SmallVector<SDValue, 8> Ops(1, Chain);
3436 if (JumpTarget.getNode())
3437 Ops.push_back(JumpTarget);
3439 // Add argument registers to the end of the list so that they are
3440 // known live into the call.
3441 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3442 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3443 RegsToPass[i].second.getValueType()));
3445 // Add a register mask operand representing the call-preserved registers.
3446 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3447 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3448 assert(Mask && "Missing call preserved mask for calling convention");
3449 Ops.push_back(DAG.getRegisterMask(Mask));
3451 if (InFlag.getNode())
3452 Ops.push_back(InFlag);
3455 return DAG.getNode(MipsISD::TailCall, dl, MVT::Other, &Ops[0], Ops.size());
3457 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
3458 InFlag = Chain.getValue(1);
3460 // Create the CALLSEQ_END node.
3461 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
3462 DAG.getIntPtrConstant(0, true), InFlag);
3463 InFlag = Chain.getValue(1);
3465 // Handle result values, copying them out of physregs into vregs that we
3467 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3468 Ins, dl, DAG, InVals);
3471 /// LowerCallResult - Lower the result values of a call into the
3472 /// appropriate copies out of appropriate physical registers.
3474 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3475 CallingConv::ID CallConv, bool isVarArg,
3476 const SmallVectorImpl<ISD::InputArg> &Ins,
3477 DebugLoc dl, SelectionDAG &DAG,
3478 SmallVectorImpl<SDValue> &InVals) const {
3479 // Assign locations to each value returned by this call.
3480 SmallVector<CCValAssign, 16> RVLocs;
3481 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3482 getTargetMachine(), RVLocs, *DAG.getContext());
3484 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
3486 // Copy all of the result registers out of their specified physreg.
3487 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3488 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
3489 RVLocs[i].getValVT(), InFlag).getValue(1);
3490 InFlag = Chain.getValue(2);
3491 InVals.push_back(Chain.getValue(0));
3497 //===----------------------------------------------------------------------===//
3498 // Formal Arguments Calling Convention Implementation
3499 //===----------------------------------------------------------------------===//
3500 /// LowerFormalArguments - transform physical registers into virtual registers
3501 /// and generate load operations for arguments places on the stack.
3503 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
3504 CallingConv::ID CallConv,
3506 const SmallVectorImpl<ISD::InputArg> &Ins,
3507 DebugLoc dl, SelectionDAG &DAG,
3508 SmallVectorImpl<SDValue> &InVals)
3510 MachineFunction &MF = DAG.getMachineFunction();
3511 MachineFrameInfo *MFI = MF.getFrameInfo();
3512 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3514 MipsFI->setVarArgsFrameIndex(0);
3516 // Used with vargs to acumulate store chains.
3517 std::vector<SDValue> OutChains;
3519 // Assign locations to all of the incoming arguments.
3520 SmallVector<CCValAssign, 16> ArgLocs;
3521 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3522 getTargetMachine(), ArgLocs, *DAG.getContext());
3523 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo);
3525 MipsCCInfo.analyzeFormalArguments(Ins);
3526 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
3527 MipsCCInfo.hasByValArg());
3529 Function::const_arg_iterator FuncArg =
3530 DAG.getMachineFunction().getFunction()->arg_begin();
3531 unsigned CurArgIdx = 0;
3532 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
3534 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3535 CCValAssign &VA = ArgLocs[i];
3536 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
3537 CurArgIdx = Ins[i].OrigArgIndex;
3538 EVT ValVT = VA.getValVT();
3539 ISD::ArgFlagsTy Flags = Ins[i].Flags;
3540 bool IsRegLoc = VA.isRegLoc();
3542 if (Flags.isByVal()) {
3543 assert(Flags.getByValSize() &&
3544 "ByVal args of size 0 should have been ignored by front-end.");
3545 assert(ByValArg != MipsCCInfo.byval_end());
3546 copyByValRegs(Chain, dl, OutChains, DAG, Flags, InVals, &*FuncArg,
3547 MipsCCInfo, *ByValArg);
3552 // Arguments stored on registers
3554 EVT RegVT = VA.getLocVT();
3555 unsigned ArgReg = VA.getLocReg();
3556 const TargetRegisterClass *RC;
3558 if (RegVT == MVT::i32)
3559 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass :
3560 &Mips::CPURegsRegClass;
3561 else if (RegVT == MVT::i64)
3562 RC = &Mips::CPU64RegsRegClass;
3563 else if (RegVT == MVT::f32)
3564 RC = &Mips::FGR32RegClass;
3565 else if (RegVT == MVT::f64)
3566 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass;
3568 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
3570 // Transform the arguments stored on
3571 // physical registers into virtual ones
3572 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
3573 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3575 // If this is an 8 or 16-bit value, it has been passed promoted
3576 // to 32 bits. Insert an assert[sz]ext to capture this, then
3577 // truncate to the right size.
3578 if (VA.getLocInfo() != CCValAssign::Full) {
3579 unsigned Opcode = 0;
3580 if (VA.getLocInfo() == CCValAssign::SExt)
3581 Opcode = ISD::AssertSext;
3582 else if (VA.getLocInfo() == CCValAssign::ZExt)
3583 Opcode = ISD::AssertZext;
3585 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
3586 DAG.getValueType(ValVT));
3587 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue);
3590 // Handle floating point arguments passed in integer registers.
3591 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
3592 (RegVT == MVT::i64 && ValVT == MVT::f64))
3593 ArgValue = DAG.getNode(ISD::BITCAST, dl, ValVT, ArgValue);
3594 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
3595 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
3596 getNextIntArgReg(ArgReg), RC);
3597 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
3598 if (!Subtarget->isLittle())
3599 std::swap(ArgValue, ArgValue2);
3600 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
3601 ArgValue, ArgValue2);
3604 InVals.push_back(ArgValue);
3605 } else { // VA.isRegLoc()
3608 assert(VA.isMemLoc());
3610 // The stack pointer offset is relative to the caller stack frame.
3611 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
3612 VA.getLocMemOffset(), true);
3614 // Create load nodes to retrieve arguments from the stack
3615 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3616 InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN,
3617 MachinePointerInfo::getFixedStack(FI),
3618 false, false, false, 0));
3622 // The mips ABIs for returning structs by value requires that we copy
3623 // the sret argument into $v0 for the return. Save the argument into
3624 // a virtual register so that we can access it from the return points.
3625 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
3626 unsigned Reg = MipsFI->getSRetReturnReg();
3628 Reg = MF.getRegInfo().
3629 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
3630 MipsFI->setSRetReturnReg(Reg);
3632 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
3633 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
3637 writeVarArgRegs(OutChains, MipsCCInfo, Chain, dl, DAG);
3639 // All stores are grouped in one node to allow the matching between
3640 // the size of Ins and InVals. This only happens when on varg functions
3641 if (!OutChains.empty()) {
3642 OutChains.push_back(Chain);
3643 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3644 &OutChains[0], OutChains.size());
3650 //===----------------------------------------------------------------------===//
3651 // Return Value Calling Convention Implementation
3652 //===----------------------------------------------------------------------===//
3655 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3656 MachineFunction &MF, bool isVarArg,
3657 const SmallVectorImpl<ISD::OutputArg> &Outs,
3658 LLVMContext &Context) const {
3659 SmallVector<CCValAssign, 16> RVLocs;
3660 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3662 return CCInfo.CheckReturn(Outs, RetCC_Mips);
3666 MipsTargetLowering::LowerReturn(SDValue Chain,
3667 CallingConv::ID CallConv, bool isVarArg,
3668 const SmallVectorImpl<ISD::OutputArg> &Outs,
3669 const SmallVectorImpl<SDValue> &OutVals,
3670 DebugLoc dl, SelectionDAG &DAG) const {
3672 // CCValAssign - represent the assignment of
3673 // the return value to a location
3674 SmallVector<CCValAssign, 16> RVLocs;
3676 // CCState - Info about the registers and stack slot.
3677 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3678 getTargetMachine(), RVLocs, *DAG.getContext());
3680 // Analize return values.
3681 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
3684 SmallVector<SDValue, 4> RetOps(1, Chain);
3686 // Copy the result values into the output registers.
3687 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3688 CCValAssign &VA = RVLocs[i];
3689 assert(VA.isRegLoc() && "Can only return in registers!");
3691 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
3693 // Guarantee that all emitted copies are stuck together with flags.
3694 Flag = Chain.getValue(1);
3695 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3698 // The mips ABIs for returning structs by value requires that we copy
3699 // the sret argument into $v0 for the return. We saved the argument into
3700 // a virtual register in the entry block, so now we copy the value out
3702 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
3703 MachineFunction &MF = DAG.getMachineFunction();
3704 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3705 unsigned Reg = MipsFI->getSRetReturnReg();
3708 llvm_unreachable("sret virtual register not created in the entry block");
3709 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
3710 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
3712 Chain = DAG.getCopyToReg(Chain, dl, V0, Val, Flag);
3713 Flag = Chain.getValue(1);
3714 RetOps.push_back(DAG.getRegister(V0, getPointerTy()));
3717 RetOps[0] = Chain; // Update chain.
3719 // Add the flag if we have it.
3721 RetOps.push_back(Flag);
3723 // Return on Mips is always a "jr $ra"
3724 return DAG.getNode(MipsISD::Ret, dl, MVT::Other, &RetOps[0], RetOps.size());
3727 //===----------------------------------------------------------------------===//
3728 // Mips Inline Assembly Support
3729 //===----------------------------------------------------------------------===//
3731 /// getConstraintType - Given a constraint letter, return the type of
3732 /// constraint it is for this target.
3733 MipsTargetLowering::ConstraintType MipsTargetLowering::
3734 getConstraintType(const std::string &Constraint) const
3736 // Mips specific constrainy
3737 // GCC config/mips/constraints.md
3739 // 'd' : An address register. Equivalent to r
3740 // unless generating MIPS16 code.
3741 // 'y' : Equivalent to r; retained for
3742 // backwards compatibility.
3743 // 'c' : A register suitable for use in an indirect
3744 // jump. This will always be $25 for -mabicalls.
3745 // 'l' : The lo register. 1 word storage.
3746 // 'x' : The hilo register pair. Double word storage.
3747 if (Constraint.size() == 1) {
3748 switch (Constraint[0]) {
3756 return C_RegisterClass;
3759 return TargetLowering::getConstraintType(Constraint);
3762 /// Examine constraint type and operand type and determine a weight value.
3763 /// This object must already have been set up with the operand type
3764 /// and the current alternative constraint selected.
3765 TargetLowering::ConstraintWeight
3766 MipsTargetLowering::getSingleConstraintMatchWeight(
3767 AsmOperandInfo &info, const char *constraint) const {
3768 ConstraintWeight weight = CW_Invalid;
3769 Value *CallOperandVal = info.CallOperandVal;
3770 // If we don't have a value, we can't do a match,
3771 // but allow it at the lowest weight.
3772 if (CallOperandVal == NULL)
3774 Type *type = CallOperandVal->getType();
3775 // Look at the constraint type.
3776 switch (*constraint) {
3778 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3782 if (type->isIntegerTy())
3783 weight = CW_Register;
3786 if (type->isFloatTy())
3787 weight = CW_Register;
3789 case 'c': // $25 for indirect jumps
3790 case 'l': // lo register
3791 case 'x': // hilo register pair
3792 if (type->isIntegerTy())
3793 weight = CW_SpecificReg;
3795 case 'I': // signed 16 bit immediate
3796 case 'J': // integer zero
3797 case 'K': // unsigned 16 bit immediate
3798 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3799 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3800 case 'O': // signed 15 bit immediate (+- 16383)
3801 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3802 if (isa<ConstantInt>(CallOperandVal))
3803 weight = CW_Constant;
3809 /// Given a register class constraint, like 'r', if this corresponds directly
3810 /// to an LLVM register class, return a register of 0 and the register class
3812 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
3813 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
3815 if (Constraint.size() == 1) {
3816 switch (Constraint[0]) {
3817 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3818 case 'y': // Same as 'r'. Exists for compatibility.
3820 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3821 if (Subtarget->inMips16Mode())
3822 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3823 return std::make_pair(0U, &Mips::CPURegsRegClass);
3825 if (VT == MVT::i64 && !HasMips64)
3826 return std::make_pair(0U, &Mips::CPURegsRegClass);
3827 if (VT == MVT::i64 && HasMips64)
3828 return std::make_pair(0U, &Mips::CPU64RegsRegClass);
3829 // This will generate an error message
3830 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3833 return std::make_pair(0U, &Mips::FGR32RegClass);
3834 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
3835 if (Subtarget->isFP64bit())
3836 return std::make_pair(0U, &Mips::FGR64RegClass);
3837 return std::make_pair(0U, &Mips::AFGR64RegClass);
3840 case 'c': // register suitable for indirect jump
3842 return std::make_pair((unsigned)Mips::T9, &Mips::CPURegsRegClass);
3843 assert(VT == MVT::i64 && "Unexpected type.");
3844 return std::make_pair((unsigned)Mips::T9_64, &Mips::CPU64RegsRegClass);
3845 case 'l': // register suitable for indirect jump
3847 return std::make_pair((unsigned)Mips::LO, &Mips::HILORegClass);
3848 return std::make_pair((unsigned)Mips::LO64, &Mips::HILO64RegClass);
3849 case 'x': // register suitable for indirect jump
3850 // Fixme: Not triggering the use of both hi and low
3851 // This will generate an error message
3852 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3855 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3858 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3859 /// vector. If it is invalid, don't add anything to Ops.
3860 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3861 std::string &Constraint,
3862 std::vector<SDValue>&Ops,
3863 SelectionDAG &DAG) const {
3864 SDValue Result(0, 0);
3866 // Only support length 1 constraints for now.
3867 if (Constraint.length() > 1) return;
3869 char ConstraintLetter = Constraint[0];
3870 switch (ConstraintLetter) {
3871 default: break; // This will fall through to the generic implementation
3872 case 'I': // Signed 16 bit constant
3873 // If this fails, the parent routine will give an error
3874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3875 EVT Type = Op.getValueType();
3876 int64_t Val = C->getSExtValue();
3877 if (isInt<16>(Val)) {
3878 Result = DAG.getTargetConstant(Val, Type);
3883 case 'J': // integer zero
3884 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3885 EVT Type = Op.getValueType();
3886 int64_t Val = C->getZExtValue();
3888 Result = DAG.getTargetConstant(0, Type);
3893 case 'K': // unsigned 16 bit immediate
3894 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3895 EVT Type = Op.getValueType();
3896 uint64_t Val = (uint64_t)C->getZExtValue();
3897 if (isUInt<16>(Val)) {
3898 Result = DAG.getTargetConstant(Val, Type);
3903 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3904 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3905 EVT Type = Op.getValueType();
3906 int64_t Val = C->getSExtValue();
3907 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3908 Result = DAG.getTargetConstant(Val, Type);
3913 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3914 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3915 EVT Type = Op.getValueType();
3916 int64_t Val = C->getSExtValue();
3917 if ((Val >= -65535) && (Val <= -1)) {
3918 Result = DAG.getTargetConstant(Val, Type);
3923 case 'O': // signed 15 bit immediate
3924 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3925 EVT Type = Op.getValueType();
3926 int64_t Val = C->getSExtValue();
3927 if ((isInt<15>(Val))) {
3928 Result = DAG.getTargetConstant(Val, Type);
3933 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3934 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3935 EVT Type = Op.getValueType();
3936 int64_t Val = C->getSExtValue();
3937 if ((Val <= 65535) && (Val >= 1)) {
3938 Result = DAG.getTargetConstant(Val, Type);
3945 if (Result.getNode()) {
3946 Ops.push_back(Result);
3950 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3954 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
3955 // No global is ever allowed as a base.
3960 case 0: // "r+i" or just "i", depending on HasBaseReg.
3963 if (!AM.HasBaseReg) // allow "r+i".
3965 return false; // disallow "r+r" or "r+r+i".
3974 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3975 // The Mips target isn't yet aware of offsets.
3979 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3981 bool IsMemset, bool ZeroMemset,
3983 MachineFunction &MF) const {
3984 if (Subtarget->hasMips64())
3990 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3991 if (VT != MVT::f32 && VT != MVT::f64)
3993 if (Imm.isNegZero())
3995 return Imm.isZero();
3998 unsigned MipsTargetLowering::getJumpTableEncoding() const {
4000 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
4002 return TargetLowering::getJumpTableEncoding();
4005 MipsTargetLowering::MipsCC::MipsCC(CallingConv::ID CC, bool IsO32_,
4007 : CCInfo(Info), CallConv(CC), IsO32(IsO32_) {
4008 // Pre-allocate reserved argument area.
4009 CCInfo.AllocateStack(reservedArgArea(), 1);
4012 void MipsTargetLowering::MipsCC::
4013 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args,
4015 assert((CallConv != CallingConv::Fast || !IsVarArg) &&
4016 "CallingConv::Fast shouldn't be used for vararg functions.");
4018 unsigned NumOpnds = Args.size();
4019 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn();
4021 for (unsigned I = 0; I != NumOpnds; ++I) {
4022 MVT ArgVT = Args[I].VT;
4023 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
4026 if (ArgFlags.isByVal()) {
4027 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
4031 if (IsVarArg && !Args[I].IsFixed)
4032 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
4034 R = FixedFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
4038 dbgs() << "Call operand #" << I << " has unhandled type "
4039 << EVT(ArgVT).getEVTString();
4041 llvm_unreachable(0);
4046 void MipsTargetLowering::MipsCC::
4047 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args) {
4048 unsigned NumArgs = Args.size();
4049 llvm::CCAssignFn *FixedFn = fixedArgFn();
4051 for (unsigned I = 0; I != NumArgs; ++I) {
4052 MVT ArgVT = Args[I].VT;
4053 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
4055 if (ArgFlags.isByVal()) {
4056 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
4060 if (!FixedFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo))
4064 dbgs() << "Formal Arg #" << I << " has unhandled type "
4065 << EVT(ArgVT).getEVTString();
4067 llvm_unreachable(0);
4072 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
4074 CCValAssign::LocInfo LocInfo,
4075 ISD::ArgFlagsTy ArgFlags) {
4076 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
4078 struct ByValArgInfo ByVal;
4079 unsigned RegSize = regSize();
4080 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
4081 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
4084 if (useRegsForByval())
4085 allocateRegs(ByVal, ByValSize, Align);
4087 // Allocate space on caller's stack.
4088 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
4090 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
4092 ByValArgs.push_back(ByVal);
4095 unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const {
4096 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs);
4099 unsigned MipsTargetLowering::MipsCC::reservedArgArea() const {
4100 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0;
4103 const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const {
4104 return IsO32 ? O32IntRegs : Mips64IntRegs;
4107 llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const {
4108 if (CallConv == CallingConv::Fast)
4109 return CC_Mips_FastCC;
4111 return IsO32 ? CC_MipsO32 : CC_MipsN;
4114 llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const {
4115 return IsO32 ? CC_MipsO32 : CC_MipsN_VarArg;
4118 const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const {
4119 return IsO32 ? O32IntRegs : Mips64DPRegs;
4122 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
4125 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs();
4126 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs();
4127 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
4128 "Byval argument's size and alignment should be a multiple of"
4131 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
4133 // If Align > RegSize, the first arg register must be even.
4134 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
4135 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
4139 // Mark the registers allocated.
4140 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
4141 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
4142 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
4145 void MipsTargetLowering::
4146 copyByValRegs(SDValue Chain, DebugLoc DL, std::vector<SDValue> &OutChains,
4147 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
4148 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
4149 const MipsCC &CC, const ByValArgInfo &ByVal) const {
4150 MachineFunction &MF = DAG.getMachineFunction();
4151 MachineFrameInfo *MFI = MF.getFrameInfo();
4152 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
4153 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
4157 FrameObjOffset = (int)CC.reservedArgArea() -
4158 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
4160 FrameObjOffset = ByVal.Address;
4162 // Create frame object.
4163 EVT PtrTy = getPointerTy();
4164 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
4165 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4166 InVals.push_back(FIN);
4171 // Copy arg registers.
4172 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
4173 const TargetRegisterClass *RC = getRegClassFor(RegTy);
4175 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
4176 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
4177 unsigned VReg = AddLiveIn(MF, ArgReg, RC);
4178 unsigned Offset = I * CC.regSize();
4179 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
4180 DAG.getConstant(Offset, PtrTy));
4181 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
4182 StorePtr, MachinePointerInfo(FuncArg, Offset),
4184 OutChains.push_back(Store);
4188 // Copy byVal arg to registers and stack.
4189 void MipsTargetLowering::
4190 passByValArg(SDValue Chain, DebugLoc DL,
4191 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
4192 SmallVector<SDValue, 8> &MemOpChains, SDValue StackPtr,
4193 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
4194 const MipsCC &CC, const ByValArgInfo &ByVal,
4195 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
4196 unsigned ByValSize = Flags.getByValSize();
4197 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
4198 unsigned RegSize = CC.regSize();
4199 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
4200 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
4202 if (ByVal.NumRegs) {
4203 const uint16_t *ArgRegs = CC.intArgRegs();
4204 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
4207 // Copy words to registers.
4208 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
4209 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4210 DAG.getConstant(Offset, PtrTy));
4211 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
4212 MachinePointerInfo(), false, false, false,
4214 MemOpChains.push_back(LoadVal.getValue(1));
4215 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
4216 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
4219 // Return if the struct has been fully copied.
4220 if (ByValSize == Offset)
4223 // Copy the remainder of the byval argument with sub-word loads and shifts.
4224 if (LeftoverBytes) {
4225 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
4226 "Size of the remainder should be smaller than RegSize.");
4229 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
4230 Offset < ByValSize; LoadSize /= 2) {
4231 unsigned RemSize = ByValSize - Offset;
4233 if (RemSize < LoadSize)
4237 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4238 DAG.getConstant(Offset, PtrTy));
4240 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
4241 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
4242 false, false, Alignment);
4243 MemOpChains.push_back(LoadVal.getValue(1));
4245 // Shift the loaded value.
4249 Shamt = TotalSizeLoaded;
4251 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
4253 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
4254 DAG.getConstant(Shamt, MVT::i32));
4257 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
4262 TotalSizeLoaded += LoadSize;
4263 Alignment = std::min(Alignment, LoadSize);
4266 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
4267 RegsToPass.push_back(std::make_pair(ArgReg, Val));
4272 // Copy remainder of byval arg to it with memcpy.
4273 unsigned MemCpySize = ByValSize - Offset;
4274 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
4275 DAG.getConstant(Offset, PtrTy));
4276 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
4277 DAG.getIntPtrConstant(ByVal.Address));
4278 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
4279 DAG.getConstant(MemCpySize, PtrTy), Alignment,
4280 /*isVolatile=*/false, /*AlwaysInline=*/false,
4281 MachinePointerInfo(0), MachinePointerInfo(0));
4282 MemOpChains.push_back(Chain);
4286 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
4287 const MipsCC &CC, SDValue Chain,
4288 DebugLoc DL, SelectionDAG &DAG) const {
4289 unsigned NumRegs = CC.numIntArgRegs();
4290 const uint16_t *ArgRegs = CC.intArgRegs();
4291 const CCState &CCInfo = CC.getCCInfo();
4292 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
4293 unsigned RegSize = CC.regSize();
4294 MVT RegTy = MVT::getIntegerVT(RegSize * 8);
4295 const TargetRegisterClass *RC = getRegClassFor(RegTy);
4296 MachineFunction &MF = DAG.getMachineFunction();
4297 MachineFrameInfo *MFI = MF.getFrameInfo();
4298 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
4300 // Offset of the first variable argument from stack pointer.
4304 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
4307 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
4309 // Record the frame index of the first variable argument
4310 // which is a value necessary to VASTART.
4311 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
4312 MipsFI->setVarArgsFrameIndex(FI);
4314 // Copy the integer registers that have not been used for argument passing
4315 // to the argument register save area. For O32, the save area is allocated
4316 // in the caller's stack frame, while for N32/64, it is allocated in the
4317 // callee's stack frame.
4318 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
4319 unsigned Reg = AddLiveIn(MF, ArgRegs[I], RC);
4320 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
4321 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
4322 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
4323 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
4324 MachinePointerInfo(), false, false, 0);
4325 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
4326 OutChains.push_back(Store);