1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that Mips uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "mips-lower"
16 #include "MipsISelLowering.h"
17 #include "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/CallingConv.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAGISel.h"
31 #include "llvm/CodeGen/ValueTypes.h"
32 #include "llvm/DerivedTypes.h"
33 #include "llvm/Function.h"
34 #include "llvm/GlobalVariable.h"
35 #include "llvm/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));
53 static const uint16_t O32IntRegs[4] = {
54 Mips::A0, Mips::A1, Mips::A2, Mips::A3
57 static const uint16_t Mips64IntRegs[8] = {
58 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
59 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64
62 static const uint16_t Mips64DPRegs[8] = {
63 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
64 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
67 // If I is a shifted mask, set the size (Size) and the first bit of the
68 // mask (Pos), and return true.
69 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
70 static bool IsShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
71 if (!isShiftedMask_64(I))
74 Size = CountPopulation_64(I);
75 Pos = CountTrailingZeros_64(I);
79 static SDValue GetGlobalReg(SelectionDAG &DAG, EVT Ty) {
80 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
81 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
84 static SDValue getTargetNode(SDValue Op, SelectionDAG &DAG, unsigned Flag) {
85 EVT Ty = Op.getValueType();
87 if (GlobalAddressSDNode *N = dyn_cast<GlobalAddressSDNode>(Op))
88 return DAG.getTargetGlobalAddress(N->getGlobal(), Op.getDebugLoc(), Ty, 0,
90 if (ExternalSymbolSDNode *N = dyn_cast<ExternalSymbolSDNode>(Op))
91 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
92 if (BlockAddressSDNode *N = dyn_cast<BlockAddressSDNode>(Op))
93 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
94 if (JumpTableSDNode *N = dyn_cast<JumpTableSDNode>(Op))
95 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
96 if (ConstantPoolSDNode *N = dyn_cast<ConstantPoolSDNode>(Op))
97 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
98 N->getOffset(), Flag);
100 llvm_unreachable("Unexpected node type.");
104 static SDValue getAddrNonPIC(SDValue Op, SelectionDAG &DAG) {
105 DebugLoc DL = Op.getDebugLoc();
106 EVT Ty = Op.getValueType();
107 SDValue Hi = getTargetNode(Op, DAG, MipsII::MO_ABS_HI);
108 SDValue Lo = getTargetNode(Op, DAG, MipsII::MO_ABS_LO);
109 return DAG.getNode(ISD::ADD, DL, Ty,
110 DAG.getNode(MipsISD::Hi, DL, Ty, Hi),
111 DAG.getNode(MipsISD::Lo, DL, Ty, Lo));
114 static SDValue getAddrLocal(SDValue Op, SelectionDAG &DAG, bool HasMips64) {
115 DebugLoc DL = Op.getDebugLoc();
116 EVT Ty = Op.getValueType();
117 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
118 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, GetGlobalReg(DAG, Ty),
119 getTargetNode(Op, DAG, GOTFlag));
120 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT,
121 MachinePointerInfo::getGOT(), false, false, false,
123 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
124 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty, getTargetNode(Op, DAG, LoFlag));
125 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo);
128 static SDValue getAddrGlobal(SDValue Op, SelectionDAG &DAG, unsigned Flag) {
129 DebugLoc DL = Op.getDebugLoc();
130 EVT Ty = Op.getValueType();
131 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, GetGlobalReg(DAG, Ty),
132 getTargetNode(Op, DAG, Flag));
133 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt,
134 MachinePointerInfo::getGOT(), false, false, false, 0);
137 static SDValue getAddrGlobalLargeGOT(SDValue Op, SelectionDAG &DAG,
138 unsigned HiFlag, unsigned LoFlag) {
139 DebugLoc DL = Op.getDebugLoc();
140 EVT Ty = Op.getValueType();
141 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty, getTargetNode(Op, DAG, HiFlag));
142 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, GetGlobalReg(DAG, Ty));
143 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi,
144 getTargetNode(Op, DAG, LoFlag));
145 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper,
146 MachinePointerInfo::getGOT(), false, false, false, 0);
149 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
151 case MipsISD::JmpLink: return "MipsISD::JmpLink";
152 case MipsISD::TailCall: return "MipsISD::TailCall";
153 case MipsISD::Hi: return "MipsISD::Hi";
154 case MipsISD::Lo: return "MipsISD::Lo";
155 case MipsISD::GPRel: return "MipsISD::GPRel";
156 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
157 case MipsISD::Ret: return "MipsISD::Ret";
158 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
159 case MipsISD::FPCmp: return "MipsISD::FPCmp";
160 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
161 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
162 case MipsISD::FPRound: return "MipsISD::FPRound";
163 case MipsISD::MAdd: return "MipsISD::MAdd";
164 case MipsISD::MAddu: return "MipsISD::MAddu";
165 case MipsISD::MSub: return "MipsISD::MSub";
166 case MipsISD::MSubu: return "MipsISD::MSubu";
167 case MipsISD::DivRem: return "MipsISD::DivRem";
168 case MipsISD::DivRemU: return "MipsISD::DivRemU";
169 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
170 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
171 case MipsISD::Wrapper: return "MipsISD::Wrapper";
172 case MipsISD::Sync: return "MipsISD::Sync";
173 case MipsISD::Ext: return "MipsISD::Ext";
174 case MipsISD::Ins: return "MipsISD::Ins";
175 case MipsISD::LWL: return "MipsISD::LWL";
176 case MipsISD::LWR: return "MipsISD::LWR";
177 case MipsISD::SWL: return "MipsISD::SWL";
178 case MipsISD::SWR: return "MipsISD::SWR";
179 case MipsISD::LDL: return "MipsISD::LDL";
180 case MipsISD::LDR: return "MipsISD::LDR";
181 case MipsISD::SDL: return "MipsISD::SDL";
182 case MipsISD::SDR: return "MipsISD::SDR";
183 case MipsISD::EXTP: return "MipsISD::EXTP";
184 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
185 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
186 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
187 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
188 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
189 case MipsISD::SHILO: return "MipsISD::SHILO";
190 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
191 case MipsISD::MULT: return "MipsISD::MULT";
192 case MipsISD::MULTU: return "MipsISD::MULTU";
193 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSPDSP";
194 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
195 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
196 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
197 default: return NULL;
202 MipsTargetLowering(MipsTargetMachine &TM)
203 : TargetLowering(TM, new MipsTargetObjectFile()),
204 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
205 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
206 IsO32(Subtarget->isABI_O32()) {
208 // Mips does not have i1 type, so use i32 for
209 // setcc operations results (slt, sgt, ...).
210 setBooleanContents(ZeroOrOneBooleanContent);
211 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
213 // Set up the register classes
214 addRegisterClass(MVT::i32, &Mips::CPURegsRegClass);
217 addRegisterClass(MVT::i64, &Mips::CPU64RegsRegClass);
219 if (Subtarget->inMips16Mode()) {
220 addRegisterClass(MVT::i32, &Mips::CPU16RegsRegClass);
223 if (Subtarget->hasDSP()) {
224 MVT::SimpleValueType VecTys[2] = {MVT::v2i16, MVT::v4i8};
226 for (unsigned i = 0; i < array_lengthof(VecTys); ++i) {
227 addRegisterClass(VecTys[i], &Mips::DSPRegsRegClass);
229 // Expand all builtin opcodes.
230 for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
231 setOperationAction(Opc, VecTys[i], Expand);
233 setOperationAction(ISD::LOAD, VecTys[i], Legal);
234 setOperationAction(ISD::STORE, VecTys[i], Legal);
235 setOperationAction(ISD::BITCAST, VecTys[i], Legal);
239 if (!TM.Options.UseSoftFloat) {
240 addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
242 // When dealing with single precision only, use libcalls
243 if (!Subtarget->isSingleFloat()) {
245 addRegisterClass(MVT::f64, &Mips::FGR64RegClass);
247 addRegisterClass(MVT::f64, &Mips::AFGR64RegClass);
251 // Load extented operations for i1 types must be promoted
252 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
253 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
254 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
256 // MIPS doesn't have extending float->double load/store
257 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
258 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
260 // Used by legalize types to correctly generate the setcc result.
261 // Without this, every float setcc comes with a AND/OR with the result,
262 // we don't want this, since the fpcmp result goes to a flag register,
263 // which is used implicitly by brcond and select operations.
264 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
266 // Mips Custom Operations
267 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
268 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
269 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
270 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
271 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
272 setOperationAction(ISD::SELECT, MVT::f32, Custom);
273 setOperationAction(ISD::SELECT, MVT::f64, Custom);
274 setOperationAction(ISD::SELECT, MVT::i32, Custom);
275 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
276 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
277 setOperationAction(ISD::SETCC, MVT::f32, Custom);
278 setOperationAction(ISD::SETCC, MVT::f64, Custom);
279 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
280 setOperationAction(ISD::VASTART, MVT::Other, Custom);
281 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
282 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
283 if (Subtarget->inMips16Mode()) {
284 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
285 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
288 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
289 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
291 if (!Subtarget->inMips16Mode()) {
292 setOperationAction(ISD::LOAD, MVT::i32, Custom);
293 setOperationAction(ISD::STORE, MVT::i32, Custom);
296 if (!TM.Options.NoNaNsFPMath) {
297 setOperationAction(ISD::FABS, MVT::f32, Custom);
298 setOperationAction(ISD::FABS, MVT::f64, Custom);
302 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
303 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
304 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
305 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
306 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
307 setOperationAction(ISD::SELECT, MVT::i64, Custom);
308 setOperationAction(ISD::LOAD, MVT::i64, Custom);
309 setOperationAction(ISD::STORE, MVT::i64, Custom);
313 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
314 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
315 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
318 setOperationAction(ISD::ADD, MVT::i32, Custom);
320 setOperationAction(ISD::ADD, MVT::i64, Custom);
322 setOperationAction(ISD::SDIV, MVT::i32, Expand);
323 setOperationAction(ISD::SREM, MVT::i32, Expand);
324 setOperationAction(ISD::UDIV, MVT::i32, Expand);
325 setOperationAction(ISD::UREM, MVT::i32, Expand);
326 setOperationAction(ISD::SDIV, MVT::i64, Expand);
327 setOperationAction(ISD::SREM, MVT::i64, Expand);
328 setOperationAction(ISD::UDIV, MVT::i64, Expand);
329 setOperationAction(ISD::UREM, MVT::i64, Expand);
331 // Operations not directly supported by Mips.
332 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
333 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
334 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
335 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
336 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
337 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
338 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
339 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
340 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
341 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
342 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
343 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
344 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
345 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
346 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
347 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
348 setOperationAction(ISD::ROTL, MVT::i32, Expand);
349 setOperationAction(ISD::ROTL, MVT::i64, Expand);
350 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
351 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
353 if (!Subtarget->hasMips32r2())
354 setOperationAction(ISD::ROTR, MVT::i32, Expand);
356 if (!Subtarget->hasMips64r2())
357 setOperationAction(ISD::ROTR, MVT::i64, Expand);
359 setOperationAction(ISD::FSIN, MVT::f32, Expand);
360 setOperationAction(ISD::FSIN, MVT::f64, Expand);
361 setOperationAction(ISD::FCOS, MVT::f32, Expand);
362 setOperationAction(ISD::FCOS, MVT::f64, Expand);
363 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
364 setOperationAction(ISD::FPOW, MVT::f32, Expand);
365 setOperationAction(ISD::FPOW, MVT::f64, Expand);
366 setOperationAction(ISD::FLOG, MVT::f32, Expand);
367 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
368 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
369 setOperationAction(ISD::FEXP, MVT::f32, Expand);
370 setOperationAction(ISD::FMA, MVT::f32, Expand);
371 setOperationAction(ISD::FMA, MVT::f64, Expand);
372 setOperationAction(ISD::FREM, MVT::f32, Expand);
373 setOperationAction(ISD::FREM, MVT::f64, Expand);
375 if (!TM.Options.NoNaNsFPMath) {
376 setOperationAction(ISD::FNEG, MVT::f32, Expand);
377 setOperationAction(ISD::FNEG, MVT::f64, Expand);
380 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
381 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
382 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
383 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
385 setOperationAction(ISD::VAARG, MVT::Other, Expand);
386 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
387 setOperationAction(ISD::VAEND, MVT::Other, Expand);
389 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
390 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
392 // Use the default for now
393 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
394 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
396 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
397 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
398 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
399 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
401 if (Subtarget->inMips16Mode()) {
402 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
403 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
404 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
405 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
406 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
407 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
408 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
409 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
410 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
411 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
412 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
413 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
416 setInsertFencesForAtomic(true);
418 if (!Subtarget->hasSEInReg()) {
419 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
420 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
423 if (!Subtarget->hasBitCount()) {
424 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
425 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
428 if (!Subtarget->hasSwap()) {
429 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
430 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
434 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom);
435 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom);
436 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom);
437 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
440 setTargetDAGCombine(ISD::ADDE);
441 setTargetDAGCombine(ISD::SUBE);
442 setTargetDAGCombine(ISD::SDIVREM);
443 setTargetDAGCombine(ISD::UDIVREM);
444 setTargetDAGCombine(ISD::SELECT);
445 setTargetDAGCombine(ISD::AND);
446 setTargetDAGCombine(ISD::OR);
447 setTargetDAGCombine(ISD::ADD);
449 setMinFunctionAlignment(HasMips64 ? 3 : 2);
451 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP);
452 computeRegisterProperties();
454 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0);
455 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1);
457 maxStoresPerMemcpy = 16;
460 bool MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
461 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
463 if (Subtarget->inMips16Mode())
475 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
480 // Transforms a subgraph in CurDAG if the following pattern is found:
481 // (addc multLo, Lo0), (adde multHi, Hi0),
483 // multHi/Lo: product of multiplication
484 // Lo0: initial value of Lo register
485 // Hi0: initial value of Hi register
486 // Return true if pattern matching was successful.
487 static bool SelectMadd(SDNode *ADDENode, SelectionDAG *CurDAG) {
488 // ADDENode's second operand must be a flag output of an ADDC node in order
489 // for the matching to be successful.
490 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
492 if (ADDCNode->getOpcode() != ISD::ADDC)
495 SDValue MultHi = ADDENode->getOperand(0);
496 SDValue MultLo = ADDCNode->getOperand(0);
497 SDNode *MultNode = MultHi.getNode();
498 unsigned MultOpc = MultHi.getOpcode();
500 // MultHi and MultLo must be generated by the same node,
501 if (MultLo.getNode() != MultNode)
504 // and it must be a multiplication.
505 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
508 // MultLo amd MultHi must be the first and second output of MultNode
510 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
513 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
514 // of the values of MultNode, in which case MultNode will be removed in later
516 // If there exist users other than ADDENode or ADDCNode, this function returns
517 // here, which will result in MultNode being mapped to a single MULT
518 // instruction node rather than a pair of MULT and MADD instructions being
520 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
523 SDValue Chain = CurDAG->getEntryNode();
524 DebugLoc dl = ADDENode->getDebugLoc();
526 // create MipsMAdd(u) node
527 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
529 SDValue MAdd = CurDAG->getNode(MultOpc, dl, MVT::Glue,
530 MultNode->getOperand(0),// Factor 0
531 MultNode->getOperand(1),// Factor 1
532 ADDCNode->getOperand(1),// Lo0
533 ADDENode->getOperand(1));// Hi0
535 // create CopyFromReg nodes
536 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
538 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
540 CopyFromLo.getValue(2));
542 // replace uses of adde and addc here
543 if (!SDValue(ADDCNode, 0).use_empty())
544 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
546 if (!SDValue(ADDENode, 0).use_empty())
547 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
553 // Transforms a subgraph in CurDAG if the following pattern is found:
554 // (addc Lo0, multLo), (sube Hi0, multHi),
556 // multHi/Lo: product of multiplication
557 // Lo0: initial value of Lo register
558 // Hi0: initial value of Hi register
559 // Return true if pattern matching was successful.
560 static bool SelectMsub(SDNode *SUBENode, SelectionDAG *CurDAG) {
561 // SUBENode's second operand must be a flag output of an SUBC node in order
562 // for the matching to be successful.
563 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
565 if (SUBCNode->getOpcode() != ISD::SUBC)
568 SDValue MultHi = SUBENode->getOperand(1);
569 SDValue MultLo = SUBCNode->getOperand(1);
570 SDNode *MultNode = MultHi.getNode();
571 unsigned MultOpc = MultHi.getOpcode();
573 // MultHi and MultLo must be generated by the same node,
574 if (MultLo.getNode() != MultNode)
577 // and it must be a multiplication.
578 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
581 // MultLo amd MultHi must be the first and second output of MultNode
583 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
586 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
587 // of the values of MultNode, in which case MultNode will be removed in later
589 // If there exist users other than SUBENode or SUBCNode, this function returns
590 // here, which will result in MultNode being mapped to a single MULT
591 // instruction node rather than a pair of MULT and MSUB instructions being
593 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
596 SDValue Chain = CurDAG->getEntryNode();
597 DebugLoc dl = SUBENode->getDebugLoc();
599 // create MipsSub(u) node
600 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
602 SDValue MSub = CurDAG->getNode(MultOpc, dl, MVT::Glue,
603 MultNode->getOperand(0),// Factor 0
604 MultNode->getOperand(1),// Factor 1
605 SUBCNode->getOperand(0),// Lo0
606 SUBENode->getOperand(0));// Hi0
608 // create CopyFromReg nodes
609 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
611 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
613 CopyFromLo.getValue(2));
615 // replace uses of sube and subc here
616 if (!SDValue(SUBCNode, 0).use_empty())
617 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
619 if (!SDValue(SUBENode, 0).use_empty())
620 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
625 static SDValue PerformADDECombine(SDNode *N, SelectionDAG &DAG,
626 TargetLowering::DAGCombinerInfo &DCI,
627 const MipsSubtarget *Subtarget) {
628 if (DCI.isBeforeLegalize())
631 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
633 return SDValue(N, 0);
638 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG &DAG,
639 TargetLowering::DAGCombinerInfo &DCI,
640 const MipsSubtarget *Subtarget) {
641 if (DCI.isBeforeLegalize())
644 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
646 return SDValue(N, 0);
651 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG &DAG,
652 TargetLowering::DAGCombinerInfo &DCI,
653 const MipsSubtarget *Subtarget) {
654 if (DCI.isBeforeLegalizeOps())
657 EVT Ty = N->getValueType(0);
658 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
659 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
660 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
662 DebugLoc dl = N->getDebugLoc();
664 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
665 N->getOperand(0), N->getOperand(1));
666 SDValue InChain = DAG.getEntryNode();
667 SDValue InGlue = DivRem;
670 if (N->hasAnyUseOfValue(0)) {
671 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty,
673 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
674 InChain = CopyFromLo.getValue(1);
675 InGlue = CopyFromLo.getValue(2);
679 if (N->hasAnyUseOfValue(1)) {
680 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
682 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
688 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
690 default: llvm_unreachable("Unknown fp condition code!");
692 case ISD::SETOEQ: return Mips::FCOND_OEQ;
693 case ISD::SETUNE: return Mips::FCOND_UNE;
695 case ISD::SETOLT: return Mips::FCOND_OLT;
697 case ISD::SETOGT: return Mips::FCOND_OGT;
699 case ISD::SETOLE: return Mips::FCOND_OLE;
701 case ISD::SETOGE: return Mips::FCOND_OGE;
702 case ISD::SETULT: return Mips::FCOND_ULT;
703 case ISD::SETULE: return Mips::FCOND_ULE;
704 case ISD::SETUGT: return Mips::FCOND_UGT;
705 case ISD::SETUGE: return Mips::FCOND_UGE;
706 case ISD::SETUO: return Mips::FCOND_UN;
707 case ISD::SETO: return Mips::FCOND_OR;
709 case ISD::SETONE: return Mips::FCOND_ONE;
710 case ISD::SETUEQ: return Mips::FCOND_UEQ;
715 // Returns true if condition code has to be inverted.
716 static bool InvertFPCondCode(Mips::CondCode CC) {
717 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
720 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
721 "Illegal Condition Code");
726 // Creates and returns an FPCmp node from a setcc node.
727 // Returns Op if setcc is not a floating point comparison.
728 static SDValue CreateFPCmp(SelectionDAG &DAG, const SDValue &Op) {
729 // must be a SETCC node
730 if (Op.getOpcode() != ISD::SETCC)
733 SDValue LHS = Op.getOperand(0);
735 if (!LHS.getValueType().isFloatingPoint())
738 SDValue RHS = Op.getOperand(1);
739 DebugLoc dl = Op.getDebugLoc();
741 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
742 // node if necessary.
743 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
745 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
746 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
749 // Creates and returns a CMovFPT/F node.
750 static SDValue CreateCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
751 SDValue False, DebugLoc DL) {
752 bool invert = InvertFPCondCode((Mips::CondCode)
753 cast<ConstantSDNode>(Cond.getOperand(2))
756 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
757 True.getValueType(), True, False, Cond);
760 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
761 TargetLowering::DAGCombinerInfo &DCI,
762 const MipsSubtarget *Subtarget) {
763 if (DCI.isBeforeLegalizeOps())
766 SDValue SetCC = N->getOperand(0);
768 if ((SetCC.getOpcode() != ISD::SETCC) ||
769 !SetCC.getOperand(0).getValueType().isInteger())
772 SDValue False = N->getOperand(2);
773 EVT FalseTy = False.getValueType();
775 if (!FalseTy.isInteger())
778 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
780 if (!CN || CN->getZExtValue())
783 const DebugLoc DL = N->getDebugLoc();
784 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
785 SDValue True = N->getOperand(1);
787 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
788 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
790 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
793 static SDValue PerformANDCombine(SDNode *N, SelectionDAG &DAG,
794 TargetLowering::DAGCombinerInfo &DCI,
795 const MipsSubtarget *Subtarget) {
796 // Pattern match EXT.
797 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
798 // => ext $dst, $src, size, pos
799 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
802 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
803 unsigned ShiftRightOpc = ShiftRight.getOpcode();
805 // Op's first operand must be a shift right.
806 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
809 // The second operand of the shift must be an immediate.
811 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
814 uint64_t Pos = CN->getZExtValue();
815 uint64_t SMPos, SMSize;
817 // Op's second operand must be a shifted mask.
818 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
819 !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
822 // Return if the shifted mask does not start at bit 0 or the sum of its size
823 // and Pos exceeds the word's size.
824 EVT ValTy = N->getValueType(0);
825 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
828 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
829 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
830 DAG.getConstant(SMSize, MVT::i32));
833 static SDValue PerformORCombine(SDNode *N, SelectionDAG &DAG,
834 TargetLowering::DAGCombinerInfo &DCI,
835 const MipsSubtarget *Subtarget) {
836 // Pattern match INS.
837 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
838 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
839 // => ins $dst, $src, size, pos, $src1
840 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
843 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
844 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
847 // See if Op's first operand matches (and $src1 , mask0).
848 if (And0.getOpcode() != ISD::AND)
851 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
852 !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
855 // See if Op's second operand matches (and (shl $src, pos), mask1).
856 if (And1.getOpcode() != ISD::AND)
859 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
860 !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
863 // The shift masks must have the same position and size.
864 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
867 SDValue Shl = And1.getOperand(0);
868 if (Shl.getOpcode() != ISD::SHL)
871 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
874 unsigned Shamt = CN->getZExtValue();
876 // Return if the shift amount and the first bit position of mask are not the
878 EVT ValTy = N->getValueType(0);
879 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
882 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
883 DAG.getConstant(SMPos0, MVT::i32),
884 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
887 static SDValue PerformADDCombine(SDNode *N, SelectionDAG &DAG,
888 TargetLowering::DAGCombinerInfo &DCI,
889 const MipsSubtarget *Subtarget) {
890 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
892 if (DCI.isBeforeLegalizeOps())
895 SDValue Add = N->getOperand(1);
897 if (Add.getOpcode() != ISD::ADD)
900 SDValue Lo = Add.getOperand(1);
902 if ((Lo.getOpcode() != MipsISD::Lo) ||
903 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
906 EVT ValTy = N->getValueType(0);
907 DebugLoc DL = N->getDebugLoc();
909 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
911 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
914 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
916 SelectionDAG &DAG = DCI.DAG;
917 unsigned opc = N->getOpcode();
922 return PerformADDECombine(N, DAG, DCI, Subtarget);
924 return PerformSUBECombine(N, DAG, DCI, Subtarget);
927 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
929 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
931 return PerformANDCombine(N, DAG, DCI, Subtarget);
933 return PerformORCombine(N, DAG, DCI, Subtarget);
935 return PerformADDCombine(N, DAG, DCI, Subtarget);
942 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
943 SmallVectorImpl<SDValue> &Results,
944 SelectionDAG &DAG) const {
945 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
947 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
948 Results.push_back(Res.getValue(I));
952 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
953 SmallVectorImpl<SDValue> &Results,
954 SelectionDAG &DAG) const {
955 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
957 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
958 Results.push_back(Res.getValue(I));
961 SDValue MipsTargetLowering::
962 LowerOperation(SDValue Op, SelectionDAG &DAG) const
964 switch (Op.getOpcode())
966 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
967 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
968 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
969 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
970 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
971 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
972 case ISD::SELECT: return LowerSELECT(Op, DAG);
973 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
974 case ISD::SETCC: return LowerSETCC(Op, DAG);
975 case ISD::VASTART: return LowerVASTART(Op, DAG);
976 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
977 case ISD::FABS: return LowerFABS(Op, DAG);
978 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
979 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
980 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
981 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
982 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
983 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG, true);
984 case ISD::SRL_PARTS: return LowerShiftRightParts(Op, DAG, false);
985 case ISD::LOAD: return LowerLOAD(Op, DAG);
986 case ISD::STORE: return LowerSTORE(Op, DAG);
987 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
988 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
989 case ISD::ADD: return LowerADD(Op, DAG);
994 //===----------------------------------------------------------------------===//
995 // Lower helper functions
996 //===----------------------------------------------------------------------===//
998 // AddLiveIn - This helper function adds the specified physical register to the
999 // MachineFunction as a live in value. It also creates a corresponding
1000 // virtual register for it.
1002 AddLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
1004 assert(RC->contains(PReg) && "Not the correct regclass!");
1005 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1006 MF.getRegInfo().addLiveIn(PReg, VReg);
1010 // Get fp branch code (not opcode) from condition code.
1011 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
1012 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
1013 return Mips::BRANCH_T;
1015 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
1016 "Invalid CondCode.");
1018 return Mips::BRANCH_F;
1022 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
1024 const MipsSubtarget *Subtarget,
1025 const TargetInstrInfo *TII,
1026 bool isFPCmp, unsigned Opc) {
1027 // There is no need to expand CMov instructions if target has
1028 // conditional moves.
1029 if (Subtarget->hasCondMov())
1032 // To "insert" a SELECT_CC instruction, we actually have to insert the
1033 // diamond control-flow pattern. The incoming instruction knows the
1034 // destination vreg to set, the condition code register to branch on, the
1035 // true/false values to select between, and a branch opcode to use.
1036 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1037 MachineFunction::iterator It = BB;
1044 // bNE r1, r0, copy1MBB
1045 // fallthrough --> copy0MBB
1046 MachineBasicBlock *thisMBB = BB;
1047 MachineFunction *F = BB->getParent();
1048 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1049 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1050 F->insert(It, copy0MBB);
1051 F->insert(It, sinkMBB);
1053 // Transfer the remainder of BB and its successor edges to sinkMBB.
1054 sinkMBB->splice(sinkMBB->begin(), BB,
1055 llvm::next(MachineBasicBlock::iterator(MI)),
1057 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1059 // Next, add the true and fallthrough blocks as its successors.
1060 BB->addSuccessor(copy0MBB);
1061 BB->addSuccessor(sinkMBB);
1063 // Emit the right instruction according to the type of the operands compared
1065 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
1067 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
1068 .addReg(Mips::ZERO).addMBB(sinkMBB);
1071 // %FalseValue = ...
1072 // # fallthrough to sinkMBB
1075 // Update machine-CFG edges
1076 BB->addSuccessor(sinkMBB);
1079 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
1084 BuildMI(*BB, BB->begin(), dl,
1085 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1086 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
1087 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
1089 BuildMI(*BB, BB->begin(), dl,
1090 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1091 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
1092 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
1094 MI->eraseFromParent(); // The pseudo instruction is gone now.
1100 MipsTargetLowering::EmitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
1102 // bposge32_pseudo $vr0
1112 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
1114 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
1115 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1116 const TargetRegisterClass *RC = &Mips::CPURegsRegClass;
1117 DebugLoc DL = MI->getDebugLoc();
1118 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1119 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
1120 MachineFunction *F = BB->getParent();
1121 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
1122 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
1123 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
1126 F->insert(It, Sink);
1128 // Transfer the remainder of BB and its successor edges to Sink.
1129 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
1131 Sink->transferSuccessorsAndUpdatePHIs(BB);
1134 BB->addSuccessor(FBB);
1135 BB->addSuccessor(TBB);
1136 FBB->addSuccessor(Sink);
1137 TBB->addSuccessor(Sink);
1139 // Insert the real bposge32 instruction to $BB.
1140 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
1143 unsigned VR2 = RegInfo.createVirtualRegister(RC);
1144 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
1145 .addReg(Mips::ZERO).addImm(0);
1146 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
1149 unsigned VR1 = RegInfo.createVirtualRegister(RC);
1150 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
1151 .addReg(Mips::ZERO).addImm(1);
1153 // Insert phi function to $Sink.
1154 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
1155 MI->getOperand(0).getReg())
1156 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
1158 MI->eraseFromParent(); // The pseudo instruction is gone now.
1163 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1164 MachineBasicBlock *BB) const {
1165 switch (MI->getOpcode()) {
1166 default: llvm_unreachable("Unexpected instr type to insert");
1167 case Mips::ATOMIC_LOAD_ADD_I8:
1168 case Mips::ATOMIC_LOAD_ADD_I8_P8:
1169 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
1170 case Mips::ATOMIC_LOAD_ADD_I16:
1171 case Mips::ATOMIC_LOAD_ADD_I16_P8:
1172 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
1173 case Mips::ATOMIC_LOAD_ADD_I32:
1174 case Mips::ATOMIC_LOAD_ADD_I32_P8:
1175 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
1176 case Mips::ATOMIC_LOAD_ADD_I64:
1177 case Mips::ATOMIC_LOAD_ADD_I64_P8:
1178 return EmitAtomicBinary(MI, BB, 8, Mips::DADDu);
1180 case Mips::ATOMIC_LOAD_AND_I8:
1181 case Mips::ATOMIC_LOAD_AND_I8_P8:
1182 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
1183 case Mips::ATOMIC_LOAD_AND_I16:
1184 case Mips::ATOMIC_LOAD_AND_I16_P8:
1185 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
1186 case Mips::ATOMIC_LOAD_AND_I32:
1187 case Mips::ATOMIC_LOAD_AND_I32_P8:
1188 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
1189 case Mips::ATOMIC_LOAD_AND_I64:
1190 case Mips::ATOMIC_LOAD_AND_I64_P8:
1191 return EmitAtomicBinary(MI, BB, 8, Mips::AND64);
1193 case Mips::ATOMIC_LOAD_OR_I8:
1194 case Mips::ATOMIC_LOAD_OR_I8_P8:
1195 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
1196 case Mips::ATOMIC_LOAD_OR_I16:
1197 case Mips::ATOMIC_LOAD_OR_I16_P8:
1198 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
1199 case Mips::ATOMIC_LOAD_OR_I32:
1200 case Mips::ATOMIC_LOAD_OR_I32_P8:
1201 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
1202 case Mips::ATOMIC_LOAD_OR_I64:
1203 case Mips::ATOMIC_LOAD_OR_I64_P8:
1204 return EmitAtomicBinary(MI, BB, 8, Mips::OR64);
1206 case Mips::ATOMIC_LOAD_XOR_I8:
1207 case Mips::ATOMIC_LOAD_XOR_I8_P8:
1208 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
1209 case Mips::ATOMIC_LOAD_XOR_I16:
1210 case Mips::ATOMIC_LOAD_XOR_I16_P8:
1211 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
1212 case Mips::ATOMIC_LOAD_XOR_I32:
1213 case Mips::ATOMIC_LOAD_XOR_I32_P8:
1214 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
1215 case Mips::ATOMIC_LOAD_XOR_I64:
1216 case Mips::ATOMIC_LOAD_XOR_I64_P8:
1217 return EmitAtomicBinary(MI, BB, 8, Mips::XOR64);
1219 case Mips::ATOMIC_LOAD_NAND_I8:
1220 case Mips::ATOMIC_LOAD_NAND_I8_P8:
1221 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
1222 case Mips::ATOMIC_LOAD_NAND_I16:
1223 case Mips::ATOMIC_LOAD_NAND_I16_P8:
1224 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
1225 case Mips::ATOMIC_LOAD_NAND_I32:
1226 case Mips::ATOMIC_LOAD_NAND_I32_P8:
1227 return EmitAtomicBinary(MI, BB, 4, 0, true);
1228 case Mips::ATOMIC_LOAD_NAND_I64:
1229 case Mips::ATOMIC_LOAD_NAND_I64_P8:
1230 return EmitAtomicBinary(MI, BB, 8, 0, true);
1232 case Mips::ATOMIC_LOAD_SUB_I8:
1233 case Mips::ATOMIC_LOAD_SUB_I8_P8:
1234 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
1235 case Mips::ATOMIC_LOAD_SUB_I16:
1236 case Mips::ATOMIC_LOAD_SUB_I16_P8:
1237 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
1238 case Mips::ATOMIC_LOAD_SUB_I32:
1239 case Mips::ATOMIC_LOAD_SUB_I32_P8:
1240 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
1241 case Mips::ATOMIC_LOAD_SUB_I64:
1242 case Mips::ATOMIC_LOAD_SUB_I64_P8:
1243 return EmitAtomicBinary(MI, BB, 8, Mips::DSUBu);
1245 case Mips::ATOMIC_SWAP_I8:
1246 case Mips::ATOMIC_SWAP_I8_P8:
1247 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
1248 case Mips::ATOMIC_SWAP_I16:
1249 case Mips::ATOMIC_SWAP_I16_P8:
1250 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
1251 case Mips::ATOMIC_SWAP_I32:
1252 case Mips::ATOMIC_SWAP_I32_P8:
1253 return EmitAtomicBinary(MI, BB, 4, 0);
1254 case Mips::ATOMIC_SWAP_I64:
1255 case Mips::ATOMIC_SWAP_I64_P8:
1256 return EmitAtomicBinary(MI, BB, 8, 0);
1258 case Mips::ATOMIC_CMP_SWAP_I8:
1259 case Mips::ATOMIC_CMP_SWAP_I8_P8:
1260 return EmitAtomicCmpSwapPartword(MI, BB, 1);
1261 case Mips::ATOMIC_CMP_SWAP_I16:
1262 case Mips::ATOMIC_CMP_SWAP_I16_P8:
1263 return EmitAtomicCmpSwapPartword(MI, BB, 2);
1264 case Mips::ATOMIC_CMP_SWAP_I32:
1265 case Mips::ATOMIC_CMP_SWAP_I32_P8:
1266 return EmitAtomicCmpSwap(MI, BB, 4);
1267 case Mips::ATOMIC_CMP_SWAP_I64:
1268 case Mips::ATOMIC_CMP_SWAP_I64_P8:
1269 return EmitAtomicCmpSwap(MI, BB, 8);
1270 case Mips::BPOSGE32_PSEUDO:
1271 return EmitBPOSGE32(MI, BB);
1275 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1276 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1278 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
1279 unsigned Size, unsigned BinOpcode,
1281 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1283 MachineFunction *MF = BB->getParent();
1284 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1285 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1286 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1287 DebugLoc dl = MI->getDebugLoc();
1288 unsigned LL, SC, AND, NOR, ZERO, BEQ;
1291 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1292 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1299 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1300 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1303 ZERO = Mips::ZERO_64;
1307 unsigned OldVal = MI->getOperand(0).getReg();
1308 unsigned Ptr = MI->getOperand(1).getReg();
1309 unsigned Incr = MI->getOperand(2).getReg();
1311 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1312 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1313 unsigned Success = RegInfo.createVirtualRegister(RC);
1315 // insert new blocks after the current block
1316 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1317 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1318 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1319 MachineFunction::iterator It = BB;
1321 MF->insert(It, loopMBB);
1322 MF->insert(It, exitMBB);
1324 // Transfer the remainder of BB and its successor edges to exitMBB.
1325 exitMBB->splice(exitMBB->begin(), BB,
1326 llvm::next(MachineBasicBlock::iterator(MI)),
1328 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1332 // fallthrough --> loopMBB
1333 BB->addSuccessor(loopMBB);
1334 loopMBB->addSuccessor(loopMBB);
1335 loopMBB->addSuccessor(exitMBB);
1338 // ll oldval, 0(ptr)
1339 // <binop> storeval, oldval, incr
1340 // sc success, storeval, 0(ptr)
1341 // beq success, $0, loopMBB
1343 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1345 // and andres, oldval, incr
1346 // nor storeval, $0, andres
1347 BuildMI(BB, dl, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1348 BuildMI(BB, dl, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1349 } else if (BinOpcode) {
1350 // <binop> storeval, oldval, incr
1351 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1355 BuildMI(BB, dl, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1356 BuildMI(BB, dl, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1358 MI->eraseFromParent(); // The instruction is gone now.
1364 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
1365 MachineBasicBlock *BB,
1366 unsigned Size, unsigned BinOpcode,
1368 assert((Size == 1 || Size == 2) &&
1369 "Unsupported size for EmitAtomicBinaryPartial.");
1371 MachineFunction *MF = BB->getParent();
1372 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1373 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1374 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1375 DebugLoc dl = MI->getDebugLoc();
1376 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1377 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1379 unsigned Dest = MI->getOperand(0).getReg();
1380 unsigned Ptr = MI->getOperand(1).getReg();
1381 unsigned Incr = MI->getOperand(2).getReg();
1383 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1384 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1385 unsigned Mask = RegInfo.createVirtualRegister(RC);
1386 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1387 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1388 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1389 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1390 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1391 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1392 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1393 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1394 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1395 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1396 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1397 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1398 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1399 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1400 unsigned Success = RegInfo.createVirtualRegister(RC);
1402 // insert new blocks after the current block
1403 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1404 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1405 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1406 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1407 MachineFunction::iterator It = BB;
1409 MF->insert(It, loopMBB);
1410 MF->insert(It, sinkMBB);
1411 MF->insert(It, exitMBB);
1413 // Transfer the remainder of BB and its successor edges to exitMBB.
1414 exitMBB->splice(exitMBB->begin(), BB,
1415 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1416 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1418 BB->addSuccessor(loopMBB);
1419 loopMBB->addSuccessor(loopMBB);
1420 loopMBB->addSuccessor(sinkMBB);
1421 sinkMBB->addSuccessor(exitMBB);
1424 // addiu masklsb2,$0,-4 # 0xfffffffc
1425 // and alignedaddr,ptr,masklsb2
1426 // andi ptrlsb2,ptr,3
1427 // sll shiftamt,ptrlsb2,3
1428 // ori maskupper,$0,255 # 0xff
1429 // sll mask,maskupper,shiftamt
1430 // nor mask2,$0,mask
1431 // sll incr2,incr,shiftamt
1433 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1434 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1435 .addReg(Mips::ZERO).addImm(-4);
1436 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1437 .addReg(Ptr).addReg(MaskLSB2);
1438 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1439 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1440 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1441 .addReg(Mips::ZERO).addImm(MaskImm);
1442 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1443 .addReg(ShiftAmt).addReg(MaskUpper);
1444 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1445 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1447 // atomic.load.binop
1449 // ll oldval,0(alignedaddr)
1450 // binop binopres,oldval,incr2
1451 // and newval,binopres,mask
1452 // and maskedoldval0,oldval,mask2
1453 // or storeval,maskedoldval0,newval
1454 // sc success,storeval,0(alignedaddr)
1455 // beq success,$0,loopMBB
1459 // ll oldval,0(alignedaddr)
1460 // and newval,incr2,mask
1461 // and maskedoldval0,oldval,mask2
1462 // or storeval,maskedoldval0,newval
1463 // sc success,storeval,0(alignedaddr)
1464 // beq success,$0,loopMBB
1467 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1469 // and andres, oldval, incr2
1470 // nor binopres, $0, andres
1471 // and newval, binopres, mask
1472 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1473 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1474 .addReg(Mips::ZERO).addReg(AndRes);
1475 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1476 } else if (BinOpcode) {
1477 // <binop> binopres, oldval, incr2
1478 // and newval, binopres, mask
1479 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1480 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1481 } else {// atomic.swap
1482 // and newval, incr2, mask
1483 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1486 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1487 .addReg(OldVal).addReg(Mask2);
1488 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1489 .addReg(MaskedOldVal0).addReg(NewVal);
1490 BuildMI(BB, dl, TII->get(SC), Success)
1491 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1492 BuildMI(BB, dl, TII->get(Mips::BEQ))
1493 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1496 // and maskedoldval1,oldval,mask
1497 // srl srlres,maskedoldval1,shiftamt
1498 // sll sllres,srlres,24
1499 // sra dest,sllres,24
1501 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1503 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1504 .addReg(OldVal).addReg(Mask);
1505 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1506 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1507 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1508 .addReg(SrlRes).addImm(ShiftImm);
1509 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1510 .addReg(SllRes).addImm(ShiftImm);
1512 MI->eraseFromParent(); // The instruction is gone now.
1518 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1519 MachineBasicBlock *BB,
1520 unsigned Size) const {
1521 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1523 MachineFunction *MF = BB->getParent();
1524 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1525 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1526 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1527 DebugLoc dl = MI->getDebugLoc();
1528 unsigned LL, SC, ZERO, BNE, BEQ;
1531 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1532 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1538 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1539 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1540 ZERO = Mips::ZERO_64;
1545 unsigned Dest = MI->getOperand(0).getReg();
1546 unsigned Ptr = MI->getOperand(1).getReg();
1547 unsigned OldVal = MI->getOperand(2).getReg();
1548 unsigned NewVal = MI->getOperand(3).getReg();
1550 unsigned Success = RegInfo.createVirtualRegister(RC);
1552 // insert new blocks after the current block
1553 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1554 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1555 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1556 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1557 MachineFunction::iterator It = BB;
1559 MF->insert(It, loop1MBB);
1560 MF->insert(It, loop2MBB);
1561 MF->insert(It, exitMBB);
1563 // Transfer the remainder of BB and its successor edges to exitMBB.
1564 exitMBB->splice(exitMBB->begin(), BB,
1565 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1566 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1570 // fallthrough --> loop1MBB
1571 BB->addSuccessor(loop1MBB);
1572 loop1MBB->addSuccessor(exitMBB);
1573 loop1MBB->addSuccessor(loop2MBB);
1574 loop2MBB->addSuccessor(loop1MBB);
1575 loop2MBB->addSuccessor(exitMBB);
1579 // bne dest, oldval, exitMBB
1581 BuildMI(BB, dl, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1582 BuildMI(BB, dl, TII->get(BNE))
1583 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1586 // sc success, newval, 0(ptr)
1587 // beq success, $0, loop1MBB
1589 BuildMI(BB, dl, TII->get(SC), Success)
1590 .addReg(NewVal).addReg(Ptr).addImm(0);
1591 BuildMI(BB, dl, TII->get(BEQ))
1592 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1594 MI->eraseFromParent(); // The instruction is gone now.
1600 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1601 MachineBasicBlock *BB,
1602 unsigned Size) const {
1603 assert((Size == 1 || Size == 2) &&
1604 "Unsupported size for EmitAtomicCmpSwapPartial.");
1606 MachineFunction *MF = BB->getParent();
1607 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1608 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1609 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1610 DebugLoc dl = MI->getDebugLoc();
1611 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1612 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1614 unsigned Dest = MI->getOperand(0).getReg();
1615 unsigned Ptr = MI->getOperand(1).getReg();
1616 unsigned CmpVal = MI->getOperand(2).getReg();
1617 unsigned NewVal = MI->getOperand(3).getReg();
1619 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1620 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1621 unsigned Mask = RegInfo.createVirtualRegister(RC);
1622 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1623 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1624 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1625 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1626 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1627 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1628 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1629 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1630 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1631 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1632 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1633 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1634 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1635 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1636 unsigned Success = RegInfo.createVirtualRegister(RC);
1638 // insert new blocks after the current block
1639 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1640 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1641 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1642 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1643 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1644 MachineFunction::iterator It = BB;
1646 MF->insert(It, loop1MBB);
1647 MF->insert(It, loop2MBB);
1648 MF->insert(It, sinkMBB);
1649 MF->insert(It, exitMBB);
1651 // Transfer the remainder of BB and its successor edges to exitMBB.
1652 exitMBB->splice(exitMBB->begin(), BB,
1653 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1654 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1656 BB->addSuccessor(loop1MBB);
1657 loop1MBB->addSuccessor(sinkMBB);
1658 loop1MBB->addSuccessor(loop2MBB);
1659 loop2MBB->addSuccessor(loop1MBB);
1660 loop2MBB->addSuccessor(sinkMBB);
1661 sinkMBB->addSuccessor(exitMBB);
1663 // FIXME: computation of newval2 can be moved to loop2MBB.
1665 // addiu masklsb2,$0,-4 # 0xfffffffc
1666 // and alignedaddr,ptr,masklsb2
1667 // andi ptrlsb2,ptr,3
1668 // sll shiftamt,ptrlsb2,3
1669 // ori maskupper,$0,255 # 0xff
1670 // sll mask,maskupper,shiftamt
1671 // nor mask2,$0,mask
1672 // andi maskedcmpval,cmpval,255
1673 // sll shiftedcmpval,maskedcmpval,shiftamt
1674 // andi maskednewval,newval,255
1675 // sll shiftednewval,maskednewval,shiftamt
1676 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1677 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1678 .addReg(Mips::ZERO).addImm(-4);
1679 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1680 .addReg(Ptr).addReg(MaskLSB2);
1681 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1682 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1683 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1684 .addReg(Mips::ZERO).addImm(MaskImm);
1685 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1686 .addReg(ShiftAmt).addReg(MaskUpper);
1687 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1688 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1689 .addReg(CmpVal).addImm(MaskImm);
1690 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1691 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1692 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1693 .addReg(NewVal).addImm(MaskImm);
1694 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1695 .addReg(ShiftAmt).addReg(MaskedNewVal);
1698 // ll oldval,0(alginedaddr)
1699 // and maskedoldval0,oldval,mask
1700 // bne maskedoldval0,shiftedcmpval,sinkMBB
1702 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1703 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1704 .addReg(OldVal).addReg(Mask);
1705 BuildMI(BB, dl, TII->get(Mips::BNE))
1706 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1709 // and maskedoldval1,oldval,mask2
1710 // or storeval,maskedoldval1,shiftednewval
1711 // sc success,storeval,0(alignedaddr)
1712 // beq success,$0,loop1MBB
1714 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1715 .addReg(OldVal).addReg(Mask2);
1716 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1717 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1718 BuildMI(BB, dl, TII->get(SC), Success)
1719 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1720 BuildMI(BB, dl, TII->get(Mips::BEQ))
1721 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1724 // srl srlres,maskedoldval0,shiftamt
1725 // sll sllres,srlres,24
1726 // sra dest,sllres,24
1728 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1730 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1731 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1732 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1733 .addReg(SrlRes).addImm(ShiftImm);
1734 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1735 .addReg(SllRes).addImm(ShiftImm);
1737 MI->eraseFromParent(); // The instruction is gone now.
1742 //===----------------------------------------------------------------------===//
1743 // Misc Lower Operation implementation
1744 //===----------------------------------------------------------------------===//
1745 SDValue MipsTargetLowering::
1746 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1748 // The first operand is the chain, the second is the condition, the third is
1749 // the block to branch to if the condition is true.
1750 SDValue Chain = Op.getOperand(0);
1751 SDValue Dest = Op.getOperand(2);
1752 DebugLoc dl = Op.getDebugLoc();
1754 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1756 // Return if flag is not set by a floating point comparison.
1757 if (CondRes.getOpcode() != MipsISD::FPCmp)
1760 SDValue CCNode = CondRes.getOperand(2);
1762 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1763 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1765 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1769 SDValue MipsTargetLowering::
1770 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1772 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1774 // Return if flag is not set by a floating point comparison.
1775 if (Cond.getOpcode() != MipsISD::FPCmp)
1778 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1782 SDValue MipsTargetLowering::
1783 LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1785 DebugLoc DL = Op.getDebugLoc();
1786 EVT Ty = Op.getOperand(0).getValueType();
1787 SDValue Cond = DAG.getNode(ISD::SETCC, DL, getSetCCResultType(Ty),
1788 Op.getOperand(0), Op.getOperand(1),
1791 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1795 SDValue MipsTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1796 SDValue Cond = CreateFPCmp(DAG, Op);
1798 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1799 "Floating point operand expected.");
1801 SDValue True = DAG.getConstant(1, MVT::i32);
1802 SDValue False = DAG.getConstant(0, MVT::i32);
1804 return CreateCMovFP(DAG, Cond, True, False, Op.getDebugLoc());
1807 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1808 SelectionDAG &DAG) const {
1809 // FIXME there isn't actually debug info here
1810 DebugLoc dl = Op.getDebugLoc();
1811 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1813 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1814 const MipsTargetObjectFile &TLOF =
1815 (const MipsTargetObjectFile&)getObjFileLowering();
1817 // %gp_rel relocation
1818 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1819 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1821 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl,
1822 DAG.getVTList(MVT::i32), &GA, 1);
1823 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1824 return DAG.getNode(ISD::ADD, dl, MVT::i32, GPReg, GPRelNode);
1827 // %hi/%lo relocation
1828 return getAddrNonPIC(Op, DAG);
1831 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1832 return getAddrLocal(Op, DAG, HasMips64);
1835 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16,
1836 MipsII::MO_GOT_LO16);
1838 return getAddrGlobal(Op, DAG,
1839 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
1842 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1843 SelectionDAG &DAG) const {
1844 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1845 return getAddrNonPIC(Op, DAG);
1847 return getAddrLocal(Op, DAG, HasMips64);
1850 SDValue MipsTargetLowering::
1851 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1853 // If the relocation model is PIC, use the General Dynamic TLS Model or
1854 // Local Dynamic TLS model, otherwise use the Initial Exec or
1855 // Local Exec TLS Model.
1857 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1858 DebugLoc dl = GA->getDebugLoc();
1859 const GlobalValue *GV = GA->getGlobal();
1860 EVT PtrVT = getPointerTy();
1862 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1864 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1865 // General Dynamic and Local Dynamic TLS Model.
1866 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1869 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, Flag);
1870 SDValue Argument = DAG.getNode(MipsISD::Wrapper, dl, PtrVT,
1871 GetGlobalReg(DAG, PtrVT), TGA);
1872 unsigned PtrSize = PtrVT.getSizeInBits();
1873 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1875 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1879 Entry.Node = Argument;
1881 Args.push_back(Entry);
1883 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
1884 false, false, false, false, 0, CallingConv::C,
1885 /*isTailCall=*/false, /*doesNotRet=*/false,
1886 /*isReturnValueUsed=*/true,
1887 TlsGetAddr, Args, DAG, dl);
1888 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1890 SDValue Ret = CallResult.first;
1892 if (model != TLSModel::LocalDynamic)
1895 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1896 MipsII::MO_DTPREL_HI);
1897 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1898 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1899 MipsII::MO_DTPREL_LO);
1900 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1901 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Ret);
1902 return DAG.getNode(ISD::ADD, dl, PtrVT, Add, Lo);
1906 if (model == TLSModel::InitialExec) {
1907 // Initial Exec TLS Model
1908 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1909 MipsII::MO_GOTTPREL);
1910 TGA = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, GetGlobalReg(DAG, PtrVT),
1912 Offset = DAG.getLoad(PtrVT, dl,
1913 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1914 false, false, false, 0);
1916 // Local Exec TLS Model
1917 assert(model == TLSModel::LocalExec);
1918 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1919 MipsII::MO_TPREL_HI);
1920 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1921 MipsII::MO_TPREL_LO);
1922 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1923 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1924 Offset = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1927 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1928 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1931 SDValue MipsTargetLowering::
1932 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1934 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1935 return getAddrNonPIC(Op, DAG);
1937 return getAddrLocal(Op, DAG, HasMips64);
1940 SDValue MipsTargetLowering::
1941 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1943 // gp_rel relocation
1944 // FIXME: we should reference the constant pool using small data sections,
1945 // but the asm printer currently doesn't support this feature without
1946 // hacking it. This feature should come soon so we can uncomment the
1948 //if (IsInSmallSection(C->getType())) {
1949 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1950 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1951 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1953 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1954 return getAddrNonPIC(Op, DAG);
1956 return getAddrLocal(Op, DAG, HasMips64);
1959 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1960 MachineFunction &MF = DAG.getMachineFunction();
1961 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1963 DebugLoc dl = Op.getDebugLoc();
1964 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1967 // vastart just stores the address of the VarArgsFrameIndex slot into the
1968 // memory location argument.
1969 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1970 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1971 MachinePointerInfo(SV), false, false, 0);
1974 static SDValue LowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1975 EVT TyX = Op.getOperand(0).getValueType();
1976 EVT TyY = Op.getOperand(1).getValueType();
1977 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1978 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1979 DebugLoc DL = Op.getDebugLoc();
1982 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1984 SDValue X = (TyX == MVT::f32) ?
1985 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1986 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1988 SDValue Y = (TyY == MVT::f32) ?
1989 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1990 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1994 // ext E, Y, 31, 1 ; extract bit31 of Y
1995 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1996 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1997 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
2000 // srl SrlX, SllX, 1
2002 // sll SllY, SrlX, 31
2003 // or Or, SrlX, SllY
2004 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2005 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2006 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
2007 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
2008 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
2011 if (TyX == MVT::f32)
2012 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
2014 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2015 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
2016 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2019 static SDValue LowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2020 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
2021 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
2022 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
2023 SDValue Const1 = DAG.getConstant(1, MVT::i32);
2024 DebugLoc DL = Op.getDebugLoc();
2026 // Bitcast to integer nodes.
2027 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
2028 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
2031 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
2032 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
2033 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
2034 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
2036 if (WidthX > WidthY)
2037 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
2038 else if (WidthY > WidthX)
2039 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
2041 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
2042 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
2043 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
2046 // (d)sll SllX, X, 1
2047 // (d)srl SrlX, SllX, 1
2048 // (d)srl SrlY, Y, width(Y)-1
2049 // (d)sll SllY, SrlX, width(Y)-1
2050 // or Or, SrlX, SllY
2051 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
2052 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
2053 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
2054 DAG.getConstant(WidthY - 1, MVT::i32));
2056 if (WidthX > WidthY)
2057 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
2058 else if (WidthY > WidthX)
2059 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
2061 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
2062 DAG.getConstant(WidthX - 1, MVT::i32));
2063 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
2064 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
2068 MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2069 if (Subtarget->hasMips64())
2070 return LowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
2072 return LowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
2075 static SDValue LowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2076 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
2077 DebugLoc DL = Op.getDebugLoc();
2079 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2081 SDValue X = (Op.getValueType() == MVT::f32) ?
2082 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
2083 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2088 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
2089 DAG.getRegister(Mips::ZERO, MVT::i32),
2090 DAG.getConstant(31, MVT::i32), Const1, X);
2092 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2093 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2096 if (Op.getValueType() == MVT::f32)
2097 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
2099 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2100 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
2101 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2104 static SDValue LowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2105 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
2106 DebugLoc DL = Op.getDebugLoc();
2108 // Bitcast to integer node.
2109 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
2113 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
2114 DAG.getRegister(Mips::ZERO_64, MVT::i64),
2115 DAG.getConstant(63, MVT::i32), Const1, X);
2117 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
2118 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
2121 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
2125 MipsTargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
2126 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
2127 return LowerFABS64(Op, DAG, Subtarget->hasMips32r2());
2129 return LowerFABS32(Op, DAG, Subtarget->hasMips32r2());
2132 SDValue MipsTargetLowering::
2133 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2135 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2136 "Frame address can only be determined for current frame.");
2138 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2139 MFI->setFrameAddressIsTaken(true);
2140 EVT VT = Op.getValueType();
2141 DebugLoc dl = Op.getDebugLoc();
2142 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2143 IsN64 ? Mips::FP_64 : Mips::FP, VT);
2147 SDValue MipsTargetLowering::LowerRETURNADDR(SDValue Op,
2148 SelectionDAG &DAG) const {
2150 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2151 "Return address can be determined only for current frame.");
2153 MachineFunction &MF = DAG.getMachineFunction();
2154 MachineFrameInfo *MFI = MF.getFrameInfo();
2155 EVT VT = Op.getValueType();
2156 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
2157 MFI->setReturnAddressIsTaken(true);
2159 // Return RA, which contains the return address. Mark it an implicit live-in.
2160 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2161 return DAG.getCopyFromReg(DAG.getEntryNode(), Op.getDebugLoc(), Reg, VT);
2164 // TODO: set SType according to the desired memory barrier behavior.
2166 MipsTargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const {
2168 DebugLoc dl = Op.getDebugLoc();
2169 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
2170 DAG.getConstant(SType, MVT::i32));
2173 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
2174 SelectionDAG &DAG) const {
2175 // FIXME: Need pseudo-fence for 'singlethread' fences
2176 // FIXME: Set SType for weaker fences where supported/appropriate.
2178 DebugLoc dl = Op.getDebugLoc();
2179 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
2180 DAG.getConstant(SType, MVT::i32));
2183 SDValue MipsTargetLowering::LowerShiftLeftParts(SDValue Op,
2184 SelectionDAG &DAG) const {
2185 DebugLoc DL = Op.getDebugLoc();
2186 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2187 SDValue Shamt = Op.getOperand(2);
2190 // lo = (shl lo, shamt)
2191 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2194 // hi = (shl lo, shamt[4:0])
2195 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2196 DAG.getConstant(-1, MVT::i32));
2197 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
2198 DAG.getConstant(1, MVT::i32));
2199 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
2201 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
2202 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2203 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
2204 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2205 DAG.getConstant(0x20, MVT::i32));
2206 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2207 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
2208 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
2210 SDValue Ops[2] = {Lo, Hi};
2211 return DAG.getMergeValues(Ops, 2, DL);
2214 SDValue MipsTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2216 DebugLoc DL = Op.getDebugLoc();
2217 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2218 SDValue Shamt = Op.getOperand(2);
2221 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2223 // hi = (sra hi, shamt)
2225 // hi = (srl hi, shamt)
2228 // lo = (sra hi, shamt[4:0])
2229 // hi = (sra hi, 31)
2231 // lo = (srl hi, shamt[4:0])
2233 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2234 DAG.getConstant(-1, MVT::i32));
2235 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
2236 DAG.getConstant(1, MVT::i32));
2237 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
2238 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
2239 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2240 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2242 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2243 DAG.getConstant(0x20, MVT::i32));
2244 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
2245 DAG.getConstant(31, MVT::i32));
2246 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
2247 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2248 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
2251 SDValue Ops[2] = {Lo, Hi};
2252 return DAG.getMergeValues(Ops, 2, DL);
2255 static SDValue CreateLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2256 SDValue Chain, SDValue Src, unsigned Offset) {
2257 SDValue Ptr = LD->getBasePtr();
2258 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2259 EVT BasePtrVT = Ptr.getValueType();
2260 DebugLoc DL = LD->getDebugLoc();
2261 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2264 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2265 DAG.getConstant(Offset, BasePtrVT));
2267 SDValue Ops[] = { Chain, Ptr, Src };
2268 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2269 LD->getMemOperand());
2272 // Expand an unaligned 32 or 64-bit integer load node.
2273 SDValue MipsTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2274 LoadSDNode *LD = cast<LoadSDNode>(Op);
2275 EVT MemVT = LD->getMemoryVT();
2277 // Return if load is aligned or if MemVT is neither i32 nor i64.
2278 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2279 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2282 bool IsLittle = Subtarget->isLittle();
2283 EVT VT = Op.getValueType();
2284 ISD::LoadExtType ExtType = LD->getExtensionType();
2285 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2287 assert((VT == MVT::i32) || (VT == MVT::i64));
2290 // (set dst, (i64 (load baseptr)))
2292 // (set tmp, (ldl (add baseptr, 7), undef))
2293 // (set dst, (ldr baseptr, tmp))
2294 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2295 SDValue LDL = CreateLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2297 return CreateLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2301 SDValue LWL = CreateLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2303 SDValue LWR = CreateLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2307 // (set dst, (i32 (load baseptr))) or
2308 // (set dst, (i64 (sextload baseptr))) or
2309 // (set dst, (i64 (extload baseptr)))
2311 // (set tmp, (lwl (add baseptr, 3), undef))
2312 // (set dst, (lwr baseptr, tmp))
2313 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2314 (ExtType == ISD::EXTLOAD))
2317 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2320 // (set dst, (i64 (zextload baseptr)))
2322 // (set tmp0, (lwl (add baseptr, 3), undef))
2323 // (set tmp1, (lwr baseptr, tmp0))
2324 // (set tmp2, (shl tmp1, 32))
2325 // (set dst, (srl tmp2, 32))
2326 DebugLoc DL = LD->getDebugLoc();
2327 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2328 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2329 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2330 SDValue Ops[] = { SRL, LWR.getValue(1) };
2331 return DAG.getMergeValues(Ops, 2, DL);
2334 static SDValue CreateStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2335 SDValue Chain, unsigned Offset) {
2336 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2337 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2338 DebugLoc DL = SD->getDebugLoc();
2339 SDVTList VTList = DAG.getVTList(MVT::Other);
2342 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2343 DAG.getConstant(Offset, BasePtrVT));
2345 SDValue Ops[] = { Chain, Value, Ptr };
2346 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2347 SD->getMemOperand());
2350 // Expand an unaligned 32 or 64-bit integer store node.
2351 SDValue MipsTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2352 StoreSDNode *SD = cast<StoreSDNode>(Op);
2353 EVT MemVT = SD->getMemoryVT();
2355 // Return if store is aligned or if MemVT is neither i32 nor i64.
2356 if ((SD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2357 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2360 bool IsLittle = Subtarget->isLittle();
2361 SDValue Value = SD->getValue(), Chain = SD->getChain();
2362 EVT VT = Value.getValueType();
2365 // (store val, baseptr) or
2366 // (truncstore val, baseptr)
2368 // (swl val, (add baseptr, 3))
2369 // (swr val, baseptr)
2370 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2371 SDValue SWL = CreateStoreLR(MipsISD::SWL, DAG, SD, Chain,
2373 return CreateStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2376 assert(VT == MVT::i64);
2379 // (store val, baseptr)
2381 // (sdl val, (add baseptr, 7))
2382 // (sdr val, baseptr)
2383 SDValue SDL = CreateStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2384 return CreateStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2387 // This function expands mips intrinsic nodes which have 64-bit input operands
2388 // or output values.
2390 // out64 = intrinsic-node in64
2392 // lo = copy (extract-element (in64, 0))
2393 // hi = copy (extract-element (in64, 1))
2394 // mips-specific-node
2397 // out64 = merge-values (v0, v1)
2399 static SDValue LowerDSPIntr(SDValue Op, SelectionDAG &DAG,
2400 unsigned Opc, bool HasI64In, bool HasI64Out) {
2401 DebugLoc DL = Op.getDebugLoc();
2402 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
2403 SDValue Chain = HasChainIn ? Op->getOperand(0) : DAG.getEntryNode();
2404 SmallVector<SDValue, 3> Ops;
2407 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2408 Op->getOperand(1 + HasChainIn),
2409 DAG.getConstant(0, MVT::i32));
2410 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2411 Op->getOperand(1 + HasChainIn),
2412 DAG.getConstant(1, MVT::i32));
2414 Chain = DAG.getCopyToReg(Chain, DL, Mips::LO, InLo, SDValue());
2415 Chain = DAG.getCopyToReg(Chain, DL, Mips::HI, InHi, Chain.getValue(1));
2417 Ops.push_back(Chain);
2418 Ops.append(Op->op_begin() + HasChainIn + 2, Op->op_end());
2419 Ops.push_back(Chain.getValue(1));
2421 Ops.push_back(Chain);
2422 Ops.append(Op->op_begin() + HasChainIn + 1, Op->op_end());
2426 return DAG.getNode(Opc, DL, Op->value_begin(), Op->getNumValues(),
2427 Ops.begin(), Ops.size());
2429 SDValue Intr = DAG.getNode(Opc, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2430 Ops.begin(), Ops.size());
2431 SDValue OutLo = DAG.getCopyFromReg(Intr.getValue(0), DL, Mips::LO, MVT::i32,
2433 SDValue OutHi = DAG.getCopyFromReg(OutLo.getValue(1), DL, Mips::HI, MVT::i32,
2435 SDValue Out = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, OutLo, OutHi);
2440 SDValue Vals[] = { Out, OutHi.getValue(1) };
2441 return DAG.getMergeValues(Vals, 2, DL);
2444 SDValue MipsTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2445 SelectionDAG &DAG) const {
2446 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
2449 case Intrinsic::mips_shilo:
2450 return LowerDSPIntr(Op, DAG, MipsISD::SHILO, true, true);
2451 case Intrinsic::mips_dpau_h_qbl:
2452 return LowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL, true, true);
2453 case Intrinsic::mips_dpau_h_qbr:
2454 return LowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR, true, true);
2455 case Intrinsic::mips_dpsu_h_qbl:
2456 return LowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL, true, true);
2457 case Intrinsic::mips_dpsu_h_qbr:
2458 return LowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR, true, true);
2459 case Intrinsic::mips_dpa_w_ph:
2460 return LowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH, true, true);
2461 case Intrinsic::mips_dps_w_ph:
2462 return LowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH, true, true);
2463 case Intrinsic::mips_dpax_w_ph:
2464 return LowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH, true, true);
2465 case Intrinsic::mips_dpsx_w_ph:
2466 return LowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH, true, true);
2467 case Intrinsic::mips_mulsa_w_ph:
2468 return LowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH, true, true);
2469 case Intrinsic::mips_mult:
2470 return LowerDSPIntr(Op, DAG, MipsISD::MULT, false, true);
2471 case Intrinsic::mips_multu:
2472 return LowerDSPIntr(Op, DAG, MipsISD::MULTU, false, true);
2473 case Intrinsic::mips_madd:
2474 return LowerDSPIntr(Op, DAG, MipsISD::MADD_DSP, true, true);
2475 case Intrinsic::mips_maddu:
2476 return LowerDSPIntr(Op, DAG, MipsISD::MADDU_DSP, true, true);
2477 case Intrinsic::mips_msub:
2478 return LowerDSPIntr(Op, DAG, MipsISD::MSUB_DSP, true, true);
2479 case Intrinsic::mips_msubu:
2480 return LowerDSPIntr(Op, DAG, MipsISD::MSUBU_DSP, true, true);
2484 SDValue MipsTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
2485 SelectionDAG &DAG) const {
2486 switch (cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue()) {
2489 case Intrinsic::mips_extp:
2490 return LowerDSPIntr(Op, DAG, MipsISD::EXTP, true, false);
2491 case Intrinsic::mips_extpdp:
2492 return LowerDSPIntr(Op, DAG, MipsISD::EXTPDP, true, false);
2493 case Intrinsic::mips_extr_w:
2494 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_W, true, false);
2495 case Intrinsic::mips_extr_r_w:
2496 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W, true, false);
2497 case Intrinsic::mips_extr_rs_w:
2498 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W, true, false);
2499 case Intrinsic::mips_extr_s_h:
2500 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H, true, false);
2501 case Intrinsic::mips_mthlip:
2502 return LowerDSPIntr(Op, DAG, MipsISD::MTHLIP, true, true);
2503 case Intrinsic::mips_mulsaq_s_w_ph:
2504 return LowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH, true, true);
2505 case Intrinsic::mips_maq_s_w_phl:
2506 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL, true, true);
2507 case Intrinsic::mips_maq_s_w_phr:
2508 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR, true, true);
2509 case Intrinsic::mips_maq_sa_w_phl:
2510 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL, true, true);
2511 case Intrinsic::mips_maq_sa_w_phr:
2512 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR, true, true);
2513 case Intrinsic::mips_dpaq_s_w_ph:
2514 return LowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH, true, true);
2515 case Intrinsic::mips_dpsq_s_w_ph:
2516 return LowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH, true, true);
2517 case Intrinsic::mips_dpaq_sa_l_w:
2518 return LowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W, true, true);
2519 case Intrinsic::mips_dpsq_sa_l_w:
2520 return LowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W, true, true);
2521 case Intrinsic::mips_dpaqx_s_w_ph:
2522 return LowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH, true, true);
2523 case Intrinsic::mips_dpaqx_sa_w_ph:
2524 return LowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH, true, true);
2525 case Intrinsic::mips_dpsqx_s_w_ph:
2526 return LowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH, true, true);
2527 case Intrinsic::mips_dpsqx_sa_w_ph:
2528 return LowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH, true, true);
2532 SDValue MipsTargetLowering::LowerADD(SDValue Op, SelectionDAG &DAG) const {
2533 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2534 || cast<ConstantSDNode>
2535 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2536 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2540 // (add (frameaddr 0), (frame_to_args_offset))
2541 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2542 // (add FrameObject, 0)
2543 // where FrameObject is a fixed StackObject with offset 0 which points to
2544 // the old stack pointer.
2545 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2546 EVT ValTy = Op->getValueType(0);
2547 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2548 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2549 return DAG.getNode(ISD::ADD, Op->getDebugLoc(), ValTy, InArgsAddr,
2550 DAG.getConstant(0, ValTy));
2553 //===----------------------------------------------------------------------===//
2554 // Calling Convention Implementation
2555 //===----------------------------------------------------------------------===//
2557 //===----------------------------------------------------------------------===//
2558 // TODO: Implement a generic logic using tblgen that can support this.
2559 // Mips O32 ABI rules:
2561 // i32 - Passed in A0, A1, A2, A3 and stack
2562 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2563 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2564 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2565 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2566 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2569 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2570 //===----------------------------------------------------------------------===//
2572 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2573 MVT LocVT, CCValAssign::LocInfo LocInfo,
2574 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2576 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2578 static const uint16_t IntRegs[] = {
2579 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2581 static const uint16_t F32Regs[] = {
2582 Mips::F12, Mips::F14
2584 static const uint16_t F64Regs[] = {
2588 // Do not process byval args here.
2589 if (ArgFlags.isByVal())
2592 // Promote i8 and i16
2593 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2595 if (ArgFlags.isSExt())
2596 LocInfo = CCValAssign::SExt;
2597 else if (ArgFlags.isZExt())
2598 LocInfo = CCValAssign::ZExt;
2600 LocInfo = CCValAssign::AExt;
2605 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2606 // is true: function is vararg, argument is 3rd or higher, there is previous
2607 // argument which is not f32 or f64.
2608 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2609 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2610 unsigned OrigAlign = ArgFlags.getOrigAlign();
2611 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2613 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2614 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2615 // If this is the first part of an i64 arg,
2616 // the allocated register must be either A0 or A2.
2617 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2618 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2620 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2621 // Allocate int register and shadow next int register. If first
2622 // available register is Mips::A1 or Mips::A3, shadow it too.
2623 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2624 if (Reg == Mips::A1 || Reg == Mips::A3)
2625 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2626 State.AllocateReg(IntRegs, IntRegsSize);
2628 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2629 // we are guaranteed to find an available float register
2630 if (ValVT == MVT::f32) {
2631 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2632 // Shadow int register
2633 State.AllocateReg(IntRegs, IntRegsSize);
2635 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2636 // Shadow int registers
2637 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2638 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2639 State.AllocateReg(IntRegs, IntRegsSize);
2640 State.AllocateReg(IntRegs, IntRegsSize);
2643 llvm_unreachable("Cannot handle this ValVT.");
2646 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2648 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2650 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2655 #include "MipsGenCallingConv.inc"
2657 //===----------------------------------------------------------------------===//
2658 // Call Calling Convention Implementation
2659 //===----------------------------------------------------------------------===//
2661 static const unsigned O32IntRegsSize = 4;
2663 // Return next O32 integer argument register.
2664 static unsigned getNextIntArgReg(unsigned Reg) {
2665 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2666 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2669 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2670 /// for tail call optimization.
2671 bool MipsTargetLowering::
2672 IsEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
2673 unsigned NextStackOffset,
2674 const MipsFunctionInfo& FI) const {
2675 if (!EnableMipsTailCalls)
2678 // No tail call optimization for mips16.
2679 if (Subtarget->inMips16Mode())
2682 // Return false if either the callee or caller has a byval argument.
2683 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
2686 // Return true if the callee's argument area is no larger than the
2688 return NextStackOffset <= FI.getIncomingArgSize();
2692 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2693 SDValue Chain, SDValue Arg, DebugLoc DL,
2694 bool IsTailCall, SelectionDAG &DAG) const {
2696 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2697 DAG.getIntPtrConstant(Offset));
2698 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2702 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2703 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2704 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2705 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2706 /*isVolatile=*/ true, false, 0);
2709 /// LowerCall - functions arguments are copied from virtual regs to
2710 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2712 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2713 SmallVectorImpl<SDValue> &InVals) const {
2714 SelectionDAG &DAG = CLI.DAG;
2715 DebugLoc &dl = CLI.DL;
2716 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2717 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2718 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2719 SDValue Chain = CLI.Chain;
2720 SDValue Callee = CLI.Callee;
2721 bool &isTailCall = CLI.IsTailCall;
2722 CallingConv::ID CallConv = CLI.CallConv;
2723 bool isVarArg = CLI.IsVarArg;
2725 MachineFunction &MF = DAG.getMachineFunction();
2726 MachineFrameInfo *MFI = MF.getFrameInfo();
2727 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2728 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2730 // Analyze operands of the call, assigning locations to each operand.
2731 SmallVector<CCValAssign, 16> ArgLocs;
2732 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2733 getTargetMachine(), ArgLocs, *DAG.getContext());
2734 MipsCC MipsCCInfo(CallConv, isVarArg, IsO32, CCInfo);
2736 MipsCCInfo.analyzeCallOperands(Outs);
2738 // Get a count of how many bytes are to be pushed on the stack.
2739 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2741 // Check if it's really possible to do a tail call.
2744 IsEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2745 *MF.getInfo<MipsFunctionInfo>());
2750 // Chain is the output chain of the last Load/Store or CopyToReg node.
2751 // ByValChain is the output chain of the last Memcpy node created for copying
2752 // byval arguments to the stack.
2753 unsigned StackAlignment = TFL->getStackAlignment();
2754 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2755 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2758 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal);
2760 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl,
2761 IsN64 ? Mips::SP_64 : Mips::SP,
2764 // With EABI is it possible to have 16 args on registers.
2765 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
2766 SmallVector<SDValue, 8> MemOpChains;
2767 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2769 // Walk the register/memloc assignments, inserting copies/loads.
2770 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2771 SDValue Arg = OutVals[i];
2772 CCValAssign &VA = ArgLocs[i];
2773 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2774 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2777 if (Flags.isByVal()) {
2778 assert(Flags.getByValSize() &&
2779 "ByVal args of size 0 should have been ignored by front-end.");
2780 assert(ByValArg != MipsCCInfo.byval_end());
2781 assert(!isTailCall &&
2782 "Do not tail-call optimize if there is a byval argument.");
2783 passByValArg(Chain, dl, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2784 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2789 // Promote the value if needed.
2790 switch (VA.getLocInfo()) {
2791 default: llvm_unreachable("Unknown loc info!");
2792 case CCValAssign::Full:
2793 if (VA.isRegLoc()) {
2794 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2795 (ValVT == MVT::f64 && LocVT == MVT::i64))
2796 Arg = DAG.getNode(ISD::BITCAST, dl, LocVT, Arg);
2797 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2798 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2799 Arg, DAG.getConstant(0, MVT::i32));
2800 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2801 Arg, DAG.getConstant(1, MVT::i32));
2802 if (!Subtarget->isLittle())
2804 unsigned LocRegLo = VA.getLocReg();
2805 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2806 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2807 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2812 case CCValAssign::SExt:
2813 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, LocVT, Arg);
2815 case CCValAssign::ZExt:
2816 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, LocVT, Arg);
2818 case CCValAssign::AExt:
2819 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, LocVT, Arg);
2823 // Arguments that can be passed on register must be kept at
2824 // RegsToPass vector
2825 if (VA.isRegLoc()) {
2826 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2830 // Register can't get to this point...
2831 assert(VA.isMemLoc());
2833 // emit ISD::STORE whichs stores the
2834 // parameter value to a stack Location
2835 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2836 Chain, Arg, dl, isTailCall, DAG));
2839 // Transform all store nodes into one single node because all store
2840 // nodes are independent of each other.
2841 if (!MemOpChains.empty())
2842 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2843 &MemOpChains[0], MemOpChains.size());
2845 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2846 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2847 // node so that legalize doesn't hack it.
2848 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2849 bool GlobalOrExternal = false;
2852 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2854 if (G->getGlobal()->hasInternalLinkage())
2855 Callee = getAddrLocal(Callee, DAG, HasMips64);
2857 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2858 MipsII::MO_CALL_LO16);
2860 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2862 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(), 0,
2863 MipsII::MO_NO_FLAG);
2864 GlobalOrExternal = true;
2866 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2867 if (!IsN64 && !IsPIC) // !N64 && static
2868 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2869 MipsII::MO_NO_FLAG);
2871 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2872 MipsII::MO_CALL_LO16);
2874 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_DISP);
2876 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2878 GlobalOrExternal = true;
2883 // T9 register operand.
2886 // T9 should contain the address of the callee function if
2887 // -reloction-model=pic or it is an indirect call.
2888 if (IsPICCall || !GlobalOrExternal) {
2890 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
2891 Chain = DAG.getCopyToReg(Chain, dl, T9Reg, Callee, SDValue(0, 0));
2892 InFlag = Chain.getValue(1);
2894 if (Subtarget->inMips16Mode())
2895 T9 = DAG.getRegister(T9Reg, getPointerTy());
2897 Callee = DAG.getRegister(T9Reg, getPointerTy());
2900 // Insert node "GP copy globalreg" before call to function.
2901 // Lazy-binding stubs require GP to point to the GOT.
2903 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
2904 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2905 RegsToPass.push_back(std::make_pair(GPReg, GetGlobalReg(DAG, Ty)));
2908 // Build a sequence of copy-to-reg nodes chained together with token
2909 // chain and flag operands which copy the outgoing args into registers.
2910 // The InFlag in necessary since all emitted instructions must be
2912 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2913 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2914 RegsToPass[i].second, InFlag);
2915 InFlag = Chain.getValue(1);
2918 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2919 // = Chain, Callee, Reg#1, Reg#2, ...
2921 // Returns a chain & a flag for retval copy to use.
2922 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2923 SmallVector<SDValue, 8> Ops;
2924 Ops.push_back(Chain);
2925 Ops.push_back(Callee);
2927 // Add argument registers to the end of the list so that they are
2928 // known live into the call.
2929 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2930 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2931 RegsToPass[i].second.getValueType()));
2933 // Add T9 register operand.
2937 // Add a register mask operand representing the call-preserved registers.
2938 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2939 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2940 assert(Mask && "Missing call preserved mask for calling convention");
2941 Ops.push_back(DAG.getRegisterMask(Mask));
2943 if (InFlag.getNode())
2944 Ops.push_back(InFlag);
2947 return DAG.getNode(MipsISD::TailCall, dl, MVT::Other, &Ops[0], Ops.size());
2949 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2950 InFlag = Chain.getValue(1);
2952 // Create the CALLSEQ_END node.
2953 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2954 DAG.getIntPtrConstant(0, true), InFlag);
2955 InFlag = Chain.getValue(1);
2957 // Handle result values, copying them out of physregs into vregs that we
2959 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2960 Ins, dl, DAG, InVals);
2963 /// LowerCallResult - Lower the result values of a call into the
2964 /// appropriate copies out of appropriate physical registers.
2966 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2967 CallingConv::ID CallConv, bool isVarArg,
2968 const SmallVectorImpl<ISD::InputArg> &Ins,
2969 DebugLoc dl, SelectionDAG &DAG,
2970 SmallVectorImpl<SDValue> &InVals) const {
2971 // Assign locations to each value returned by this call.
2972 SmallVector<CCValAssign, 16> RVLocs;
2973 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2974 getTargetMachine(), RVLocs, *DAG.getContext());
2976 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2978 // Copy all of the result registers out of their specified physreg.
2979 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2980 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2981 RVLocs[i].getValVT(), InFlag).getValue(1);
2982 InFlag = Chain.getValue(2);
2983 InVals.push_back(Chain.getValue(0));
2989 //===----------------------------------------------------------------------===//
2990 // Formal Arguments Calling Convention Implementation
2991 //===----------------------------------------------------------------------===//
2992 /// LowerFormalArguments - transform physical registers into virtual registers
2993 /// and generate load operations for arguments places on the stack.
2995 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2996 CallingConv::ID CallConv,
2998 const SmallVectorImpl<ISD::InputArg> &Ins,
2999 DebugLoc dl, SelectionDAG &DAG,
3000 SmallVectorImpl<SDValue> &InVals)
3002 MachineFunction &MF = DAG.getMachineFunction();
3003 MachineFrameInfo *MFI = MF.getFrameInfo();
3004 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3006 MipsFI->setVarArgsFrameIndex(0);
3008 // Used with vargs to acumulate store chains.
3009 std::vector<SDValue> OutChains;
3011 // Assign locations to all of the incoming arguments.
3012 SmallVector<CCValAssign, 16> ArgLocs;
3013 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3014 getTargetMachine(), ArgLocs, *DAG.getContext());
3015 MipsCC MipsCCInfo(CallConv, isVarArg, IsO32, CCInfo);
3017 MipsCCInfo.analyzeFormalArguments(Ins);
3018 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
3019 MipsCCInfo.hasByValArg());
3021 Function::const_arg_iterator FuncArg =
3022 DAG.getMachineFunction().getFunction()->arg_begin();
3023 unsigned CurArgIdx = 0;
3024 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
3026 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3027 CCValAssign &VA = ArgLocs[i];
3028 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
3029 CurArgIdx = Ins[i].OrigArgIndex;
3030 EVT ValVT = VA.getValVT();
3031 ISD::ArgFlagsTy Flags = Ins[i].Flags;
3032 bool IsRegLoc = VA.isRegLoc();
3034 if (Flags.isByVal()) {
3035 assert(Flags.getByValSize() &&
3036 "ByVal args of size 0 should have been ignored by front-end.");
3037 assert(ByValArg != MipsCCInfo.byval_end());
3038 copyByValRegs(Chain, dl, OutChains, DAG, Flags, InVals, &*FuncArg,
3039 MipsCCInfo, *ByValArg);
3044 // Arguments stored on registers
3046 EVT RegVT = VA.getLocVT();
3047 unsigned ArgReg = VA.getLocReg();
3048 const TargetRegisterClass *RC;
3050 if (RegVT == MVT::i32)
3051 RC = &Mips::CPURegsRegClass;
3052 else if (RegVT == MVT::i64)
3053 RC = &Mips::CPU64RegsRegClass;
3054 else if (RegVT == MVT::f32)
3055 RC = &Mips::FGR32RegClass;
3056 else if (RegVT == MVT::f64)
3057 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass;
3059 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
3061 // Transform the arguments stored on
3062 // physical registers into virtual ones
3063 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
3064 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3066 // If this is an 8 or 16-bit value, it has been passed promoted
3067 // to 32 bits. Insert an assert[sz]ext to capture this, then
3068 // truncate to the right size.
3069 if (VA.getLocInfo() != CCValAssign::Full) {
3070 unsigned Opcode = 0;
3071 if (VA.getLocInfo() == CCValAssign::SExt)
3072 Opcode = ISD::AssertSext;
3073 else if (VA.getLocInfo() == CCValAssign::ZExt)
3074 Opcode = ISD::AssertZext;
3076 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
3077 DAG.getValueType(ValVT));
3078 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue);
3081 // Handle floating point arguments passed in integer registers.
3082 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
3083 (RegVT == MVT::i64 && ValVT == MVT::f64))
3084 ArgValue = DAG.getNode(ISD::BITCAST, dl, ValVT, ArgValue);
3085 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
3086 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
3087 getNextIntArgReg(ArgReg), RC);
3088 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
3089 if (!Subtarget->isLittle())
3090 std::swap(ArgValue, ArgValue2);
3091 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
3092 ArgValue, ArgValue2);
3095 InVals.push_back(ArgValue);
3096 } else { // VA.isRegLoc()
3099 assert(VA.isMemLoc());
3101 // The stack pointer offset is relative to the caller stack frame.
3102 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
3103 VA.getLocMemOffset(), true);
3105 // Create load nodes to retrieve arguments from the stack
3106 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3107 InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN,
3108 MachinePointerInfo::getFixedStack(FI),
3109 false, false, false, 0));
3113 // The mips ABIs for returning structs by value requires that we copy
3114 // the sret argument into $v0 for the return. Save the argument into
3115 // a virtual register so that we can access it from the return points.
3116 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
3117 unsigned Reg = MipsFI->getSRetReturnReg();
3119 Reg = MF.getRegInfo().
3120 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
3121 MipsFI->setSRetReturnReg(Reg);
3123 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
3124 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
3128 writeVarArgRegs(OutChains, MipsCCInfo, Chain, dl, DAG);
3130 // All stores are grouped in one node to allow the matching between
3131 // the size of Ins and InVals. This only happens when on varg functions
3132 if (!OutChains.empty()) {
3133 OutChains.push_back(Chain);
3134 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3135 &OutChains[0], OutChains.size());
3141 //===----------------------------------------------------------------------===//
3142 // Return Value Calling Convention Implementation
3143 //===----------------------------------------------------------------------===//
3146 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3147 MachineFunction &MF, bool isVarArg,
3148 const SmallVectorImpl<ISD::OutputArg> &Outs,
3149 LLVMContext &Context) const {
3150 SmallVector<CCValAssign, 16> RVLocs;
3151 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3153 return CCInfo.CheckReturn(Outs, RetCC_Mips);
3157 MipsTargetLowering::LowerReturn(SDValue Chain,
3158 CallingConv::ID CallConv, bool isVarArg,
3159 const SmallVectorImpl<ISD::OutputArg> &Outs,
3160 const SmallVectorImpl<SDValue> &OutVals,
3161 DebugLoc dl, SelectionDAG &DAG) const {
3163 // CCValAssign - represent the assignment of
3164 // the return value to a location
3165 SmallVector<CCValAssign, 16> RVLocs;
3167 // CCState - Info about the registers and stack slot.
3168 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3169 getTargetMachine(), RVLocs, *DAG.getContext());
3171 // Analize return values.
3172 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
3174 // If this is the first return lowered for this function, add
3175 // the regs to the liveout set for the function.
3176 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3177 for (unsigned i = 0; i != RVLocs.size(); ++i)
3178 if (RVLocs[i].isRegLoc())
3179 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3184 // Copy the result values into the output registers.
3185 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3186 CCValAssign &VA = RVLocs[i];
3187 assert(VA.isRegLoc() && "Can only return in registers!");
3189 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
3191 // guarantee that all emitted copies are
3192 // stuck together, avoiding something bad
3193 Flag = Chain.getValue(1);
3196 // The mips ABIs for returning structs by value requires that we copy
3197 // the sret argument into $v0 for the return. We saved the argument into
3198 // a virtual register in the entry block, so now we copy the value out
3200 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
3201 MachineFunction &MF = DAG.getMachineFunction();
3202 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3203 unsigned Reg = MipsFI->getSRetReturnReg();
3206 llvm_unreachable("sret virtual register not created in the entry block");
3207 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
3208 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
3210 Chain = DAG.getCopyToReg(Chain, dl, V0, Val, Flag);
3211 Flag = Chain.getValue(1);
3212 MF.getRegInfo().addLiveOut(V0);
3215 // Return on Mips is always a "jr $ra"
3217 return DAG.getNode(MipsISD::Ret, dl, MVT::Other, Chain, Flag);
3220 return DAG.getNode(MipsISD::Ret, dl, MVT::Other, Chain);
3223 //===----------------------------------------------------------------------===//
3224 // Mips Inline Assembly Support
3225 //===----------------------------------------------------------------------===//
3227 /// getConstraintType - Given a constraint letter, return the type of
3228 /// constraint it is for this target.
3229 MipsTargetLowering::ConstraintType MipsTargetLowering::
3230 getConstraintType(const std::string &Constraint) const
3232 // Mips specific constrainy
3233 // GCC config/mips/constraints.md
3235 // 'd' : An address register. Equivalent to r
3236 // unless generating MIPS16 code.
3237 // 'y' : Equivalent to r; retained for
3238 // backwards compatibility.
3239 // 'c' : A register suitable for use in an indirect
3240 // jump. This will always be $25 for -mabicalls.
3241 // 'l' : The lo register. 1 word storage.
3242 // 'x' : The hilo register pair. Double word storage.
3243 if (Constraint.size() == 1) {
3244 switch (Constraint[0]) {
3252 return C_RegisterClass;
3255 return TargetLowering::getConstraintType(Constraint);
3258 /// Examine constraint type and operand type and determine a weight value.
3259 /// This object must already have been set up with the operand type
3260 /// and the current alternative constraint selected.
3261 TargetLowering::ConstraintWeight
3262 MipsTargetLowering::getSingleConstraintMatchWeight(
3263 AsmOperandInfo &info, const char *constraint) const {
3264 ConstraintWeight weight = CW_Invalid;
3265 Value *CallOperandVal = info.CallOperandVal;
3266 // If we don't have a value, we can't do a match,
3267 // but allow it at the lowest weight.
3268 if (CallOperandVal == NULL)
3270 Type *type = CallOperandVal->getType();
3271 // Look at the constraint type.
3272 switch (*constraint) {
3274 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3278 if (type->isIntegerTy())
3279 weight = CW_Register;
3282 if (type->isFloatTy())
3283 weight = CW_Register;
3285 case 'c': // $25 for indirect jumps
3286 case 'l': // lo register
3287 case 'x': // hilo register pair
3288 if (type->isIntegerTy())
3289 weight = CW_SpecificReg;
3291 case 'I': // signed 16 bit immediate
3292 case 'J': // integer zero
3293 case 'K': // unsigned 16 bit immediate
3294 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3295 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3296 case 'O': // signed 15 bit immediate (+- 16383)
3297 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3298 if (isa<ConstantInt>(CallOperandVal))
3299 weight = CW_Constant;
3305 /// Given a register class constraint, like 'r', if this corresponds directly
3306 /// to an LLVM register class, return a register of 0 and the register class
3308 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
3309 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
3311 if (Constraint.size() == 1) {
3312 switch (Constraint[0]) {
3313 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3314 case 'y': // Same as 'r'. Exists for compatibility.
3316 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3317 if (Subtarget->inMips16Mode())
3318 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3319 return std::make_pair(0U, &Mips::CPURegsRegClass);
3321 if (VT == MVT::i64 && !HasMips64)
3322 return std::make_pair(0U, &Mips::CPURegsRegClass);
3323 if (VT == MVT::i64 && HasMips64)
3324 return std::make_pair(0U, &Mips::CPU64RegsRegClass);
3325 // This will generate an error message
3326 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3329 return std::make_pair(0U, &Mips::FGR32RegClass);
3330 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
3331 if (Subtarget->isFP64bit())
3332 return std::make_pair(0U, &Mips::FGR64RegClass);
3333 return std::make_pair(0U, &Mips::AFGR64RegClass);
3336 case 'c': // register suitable for indirect jump
3338 return std::make_pair((unsigned)Mips::T9, &Mips::CPURegsRegClass);
3339 assert(VT == MVT::i64 && "Unexpected type.");
3340 return std::make_pair((unsigned)Mips::T9_64, &Mips::CPU64RegsRegClass);
3341 case 'l': // register suitable for indirect jump
3343 return std::make_pair((unsigned)Mips::LO, &Mips::HILORegClass);
3344 return std::make_pair((unsigned)Mips::LO64, &Mips::HILO64RegClass);
3345 case 'x': // register suitable for indirect jump
3346 // Fixme: Not triggering the use of both hi and low
3347 // This will generate an error message
3348 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3351 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3354 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3355 /// vector. If it is invalid, don't add anything to Ops.
3356 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3357 std::string &Constraint,
3358 std::vector<SDValue>&Ops,
3359 SelectionDAG &DAG) const {
3360 SDValue Result(0, 0);
3362 // Only support length 1 constraints for now.
3363 if (Constraint.length() > 1) return;
3365 char ConstraintLetter = Constraint[0];
3366 switch (ConstraintLetter) {
3367 default: break; // This will fall through to the generic implementation
3368 case 'I': // Signed 16 bit constant
3369 // If this fails, the parent routine will give an error
3370 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3371 EVT Type = Op.getValueType();
3372 int64_t Val = C->getSExtValue();
3373 if (isInt<16>(Val)) {
3374 Result = DAG.getTargetConstant(Val, Type);
3379 case 'J': // integer zero
3380 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3381 EVT Type = Op.getValueType();
3382 int64_t Val = C->getZExtValue();
3384 Result = DAG.getTargetConstant(0, Type);
3389 case 'K': // unsigned 16 bit immediate
3390 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3391 EVT Type = Op.getValueType();
3392 uint64_t Val = (uint64_t)C->getZExtValue();
3393 if (isUInt<16>(Val)) {
3394 Result = DAG.getTargetConstant(Val, Type);
3399 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3400 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3401 EVT Type = Op.getValueType();
3402 int64_t Val = C->getSExtValue();
3403 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3404 Result = DAG.getTargetConstant(Val, Type);
3409 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3410 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3411 EVT Type = Op.getValueType();
3412 int64_t Val = C->getSExtValue();
3413 if ((Val >= -65535) && (Val <= -1)) {
3414 Result = DAG.getTargetConstant(Val, Type);
3419 case 'O': // signed 15 bit immediate
3420 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3421 EVT Type = Op.getValueType();
3422 int64_t Val = C->getSExtValue();
3423 if ((isInt<15>(Val))) {
3424 Result = DAG.getTargetConstant(Val, Type);
3429 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3430 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3431 EVT Type = Op.getValueType();
3432 int64_t Val = C->getSExtValue();
3433 if ((Val <= 65535) && (Val >= 1)) {
3434 Result = DAG.getTargetConstant(Val, Type);
3441 if (Result.getNode()) {
3442 Ops.push_back(Result);
3446 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3450 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
3451 // No global is ever allowed as a base.
3456 case 0: // "r+i" or just "i", depending on HasBaseReg.
3459 if (!AM.HasBaseReg) // allow "r+i".
3461 return false; // disallow "r+r" or "r+r+i".
3470 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3471 // The Mips target isn't yet aware of offsets.
3475 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3476 unsigned SrcAlign, bool IsZeroVal,
3478 MachineFunction &MF) const {
3479 if (Subtarget->hasMips64())
3485 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3486 if (VT != MVT::f32 && VT != MVT::f64)
3488 if (Imm.isNegZero())
3490 return Imm.isZero();
3493 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3495 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3497 return TargetLowering::getJumpTableEncoding();
3500 MipsTargetLowering::MipsCC::MipsCC(CallingConv::ID CallConv, bool IsVarArg,
3501 bool IsO32, CCState &Info) : CCInfo(Info) {
3502 UseRegsForByval = true;
3506 NumIntArgRegs = array_lengthof(O32IntRegs);
3507 ReservedArgArea = 16;
3508 IntArgRegs = ShadowRegs = O32IntRegs;
3509 FixedFn = VarFn = CC_MipsO32;
3512 NumIntArgRegs = array_lengthof(Mips64IntRegs);
3513 ReservedArgArea = 0;
3514 IntArgRegs = Mips64IntRegs;
3515 ShadowRegs = Mips64DPRegs;
3517 VarFn = CC_MipsN_VarArg;
3520 if (CallConv == CallingConv::Fast) {
3522 UseRegsForByval = false;
3523 ReservedArgArea = 0;
3524 FixedFn = VarFn = CC_Mips_FastCC;
3527 // Pre-allocate reserved argument area.
3528 CCInfo.AllocateStack(ReservedArgArea, 1);
3531 void MipsTargetLowering::MipsCC::
3532 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args) {
3533 unsigned NumOpnds = Args.size();
3535 for (unsigned I = 0; I != NumOpnds; ++I) {
3536 MVT ArgVT = Args[I].VT;
3537 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3540 if (ArgFlags.isByVal()) {
3541 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3545 if (Args[I].IsFixed)
3546 R = FixedFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3548 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3552 dbgs() << "Call operand #" << I << " has unhandled type "
3553 << EVT(ArgVT).getEVTString();
3555 llvm_unreachable(0);
3560 void MipsTargetLowering::MipsCC::
3561 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args) {
3562 unsigned NumArgs = Args.size();
3564 for (unsigned I = 0; I != NumArgs; ++I) {
3565 MVT ArgVT = Args[I].VT;
3566 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3568 if (ArgFlags.isByVal()) {
3569 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3573 if (!FixedFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo))
3577 dbgs() << "Formal Arg #" << I << " has unhandled type "
3578 << EVT(ArgVT).getEVTString();
3580 llvm_unreachable(0);
3585 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3587 CCValAssign::LocInfo LocInfo,
3588 ISD::ArgFlagsTy ArgFlags) {
3589 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3591 struct ByValArgInfo ByVal;
3592 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3593 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3596 if (UseRegsForByval)
3597 allocateRegs(ByVal, ByValSize, Align);
3599 // Allocate space on caller's stack.
3600 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3602 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3604 ByValArgs.push_back(ByVal);
3607 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3610 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3611 "Byval argument's size and alignment should be a multiple of"
3614 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3616 // If Align > RegSize, the first arg register must be even.
3617 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3618 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3622 // Mark the registers allocated.
3623 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3624 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3625 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3628 void MipsTargetLowering::
3629 copyByValRegs(SDValue Chain, DebugLoc DL, std::vector<SDValue> &OutChains,
3630 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3631 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3632 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3633 MachineFunction &MF = DAG.getMachineFunction();
3634 MachineFrameInfo *MFI = MF.getFrameInfo();
3635 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3636 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3640 FrameObjOffset = (int)CC.reservedArgArea() -
3641 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3643 FrameObjOffset = ByVal.Address;
3645 // Create frame object.
3646 EVT PtrTy = getPointerTy();
3647 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3648 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3649 InVals.push_back(FIN);
3654 // Copy arg registers.
3655 EVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3656 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3658 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3659 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3660 unsigned VReg = AddLiveIn(MF, ArgReg, RC);
3661 unsigned Offset = I * CC.regSize();
3662 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3663 DAG.getConstant(Offset, PtrTy));
3664 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3665 StorePtr, MachinePointerInfo(FuncArg, Offset),
3667 OutChains.push_back(Store);
3671 // Copy byVal arg to registers and stack.
3672 void MipsTargetLowering::
3673 passByValArg(SDValue Chain, DebugLoc DL,
3674 SmallVector<std::pair<unsigned, SDValue>, 16> &RegsToPass,
3675 SmallVector<SDValue, 8> &MemOpChains, SDValue StackPtr,
3676 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3677 const MipsCC &CC, const ByValArgInfo &ByVal,
3678 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3679 unsigned ByValSize = Flags.getByValSize();
3680 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
3681 unsigned RegSize = CC.regSize();
3682 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
3683 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
3685 if (ByVal.NumRegs) {
3686 const uint16_t *ArgRegs = CC.intArgRegs();
3687 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
3690 // Copy words to registers.
3691 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
3692 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3693 DAG.getConstant(Offset, PtrTy));
3694 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3695 MachinePointerInfo(), false, false, false,
3697 MemOpChains.push_back(LoadVal.getValue(1));
3698 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3699 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3702 // Return if the struct has been fully copied.
3703 if (ByValSize == Offset)
3706 // Copy the remainder of the byval argument with sub-word loads and shifts.
3707 if (LeftoverBytes) {
3708 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
3709 "Size of the remainder should be smaller than RegSize.");
3712 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
3713 Offset < ByValSize; LoadSize /= 2) {
3714 unsigned RemSize = ByValSize - Offset;
3716 if (RemSize < LoadSize)
3720 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3721 DAG.getConstant(Offset, PtrTy));
3723 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
3724 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
3725 false, false, Alignment);
3726 MemOpChains.push_back(LoadVal.getValue(1));
3728 // Shift the loaded value.
3732 Shamt = TotalSizeLoaded;
3734 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
3736 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3737 DAG.getConstant(Shamt, MVT::i32));
3740 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3745 TotalSizeLoaded += LoadSize;
3746 Alignment = std::min(Alignment, LoadSize);
3749 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3750 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3755 // Copy remainder of byval arg to it with memcpy.
3756 unsigned MemCpySize = ByValSize - Offset;
3757 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3758 DAG.getConstant(Offset, PtrTy));
3759 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3760 DAG.getIntPtrConstant(ByVal.Address));
3761 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3762 DAG.getConstant(MemCpySize, PtrTy), Alignment,
3763 /*isVolatile=*/false, /*AlwaysInline=*/false,
3764 MachinePointerInfo(0), MachinePointerInfo(0));
3765 MemOpChains.push_back(Chain);
3769 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3770 const MipsCC &CC, SDValue Chain,
3771 DebugLoc DL, SelectionDAG &DAG) const {
3772 unsigned NumRegs = CC.numIntArgRegs();
3773 const uint16_t *ArgRegs = CC.intArgRegs();
3774 const CCState &CCInfo = CC.getCCInfo();
3775 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3776 unsigned RegSize = CC.regSize();
3777 EVT RegTy = MVT::getIntegerVT(RegSize * 8);
3778 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3779 MachineFunction &MF = DAG.getMachineFunction();
3780 MachineFrameInfo *MFI = MF.getFrameInfo();
3781 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3783 // Offset of the first variable argument from stack pointer.
3787 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3790 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3792 // Record the frame index of the first variable argument
3793 // which is a value necessary to VASTART.
3794 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3795 MipsFI->setVarArgsFrameIndex(FI);
3797 // Copy the integer registers that have not been used for argument passing
3798 // to the argument register save area. For O32, the save area is allocated
3799 // in the caller's stack frame, while for N32/64, it is allocated in the
3800 // callee's stack frame.
3801 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3802 unsigned Reg = AddLiveIn(MF, ArgRegs[I], RC);
3803 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3804 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3805 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3806 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3807 MachinePointerInfo(), false, false, 0);
3808 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
3809 OutChains.push_back(Store);