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;
461 MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
462 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
464 if (Subtarget->inMips16Mode())
478 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
483 // Transforms a subgraph in CurDAG if the following pattern is found:
484 // (addc multLo, Lo0), (adde multHi, Hi0),
486 // multHi/Lo: product of multiplication
487 // Lo0: initial value of Lo register
488 // Hi0: initial value of Hi register
489 // Return true if pattern matching was successful.
490 static bool SelectMadd(SDNode *ADDENode, SelectionDAG *CurDAG) {
491 // ADDENode's second operand must be a flag output of an ADDC node in order
492 // for the matching to be successful.
493 SDNode *ADDCNode = ADDENode->getOperand(2).getNode();
495 if (ADDCNode->getOpcode() != ISD::ADDC)
498 SDValue MultHi = ADDENode->getOperand(0);
499 SDValue MultLo = ADDCNode->getOperand(0);
500 SDNode *MultNode = MultHi.getNode();
501 unsigned MultOpc = MultHi.getOpcode();
503 // MultHi and MultLo must be generated by the same node,
504 if (MultLo.getNode() != MultNode)
507 // and it must be a multiplication.
508 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
511 // MultLo amd MultHi must be the first and second output of MultNode
513 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
516 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
517 // of the values of MultNode, in which case MultNode will be removed in later
519 // If there exist users other than ADDENode or ADDCNode, this function returns
520 // here, which will result in MultNode being mapped to a single MULT
521 // instruction node rather than a pair of MULT and MADD instructions being
523 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
526 SDValue Chain = CurDAG->getEntryNode();
527 DebugLoc dl = ADDENode->getDebugLoc();
529 // create MipsMAdd(u) node
530 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
532 SDValue MAdd = CurDAG->getNode(MultOpc, dl, MVT::Glue,
533 MultNode->getOperand(0),// Factor 0
534 MultNode->getOperand(1),// Factor 1
535 ADDCNode->getOperand(1),// Lo0
536 ADDENode->getOperand(1));// Hi0
538 // create CopyFromReg nodes
539 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
541 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
543 CopyFromLo.getValue(2));
545 // replace uses of adde and addc here
546 if (!SDValue(ADDCNode, 0).use_empty())
547 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
549 if (!SDValue(ADDENode, 0).use_empty())
550 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
556 // Transforms a subgraph in CurDAG if the following pattern is found:
557 // (addc Lo0, multLo), (sube Hi0, multHi),
559 // multHi/Lo: product of multiplication
560 // Lo0: initial value of Lo register
561 // Hi0: initial value of Hi register
562 // Return true if pattern matching was successful.
563 static bool SelectMsub(SDNode *SUBENode, SelectionDAG *CurDAG) {
564 // SUBENode's second operand must be a flag output of an SUBC node in order
565 // for the matching to be successful.
566 SDNode *SUBCNode = SUBENode->getOperand(2).getNode();
568 if (SUBCNode->getOpcode() != ISD::SUBC)
571 SDValue MultHi = SUBENode->getOperand(1);
572 SDValue MultLo = SUBCNode->getOperand(1);
573 SDNode *MultNode = MultHi.getNode();
574 unsigned MultOpc = MultHi.getOpcode();
576 // MultHi and MultLo must be generated by the same node,
577 if (MultLo.getNode() != MultNode)
580 // and it must be a multiplication.
581 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
584 // MultLo amd MultHi must be the first and second output of MultNode
586 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
589 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
590 // of the values of MultNode, in which case MultNode will be removed in later
592 // If there exist users other than SUBENode or SUBCNode, this function returns
593 // here, which will result in MultNode being mapped to a single MULT
594 // instruction node rather than a pair of MULT and MSUB instructions being
596 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
599 SDValue Chain = CurDAG->getEntryNode();
600 DebugLoc dl = SUBENode->getDebugLoc();
602 // create MipsSub(u) node
603 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
605 SDValue MSub = CurDAG->getNode(MultOpc, dl, MVT::Glue,
606 MultNode->getOperand(0),// Factor 0
607 MultNode->getOperand(1),// Factor 1
608 SUBCNode->getOperand(0),// Lo0
609 SUBENode->getOperand(0));// Hi0
611 // create CopyFromReg nodes
612 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
614 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
616 CopyFromLo.getValue(2));
618 // replace uses of sube and subc here
619 if (!SDValue(SUBCNode, 0).use_empty())
620 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
622 if (!SDValue(SUBENode, 0).use_empty())
623 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
628 static SDValue PerformADDECombine(SDNode *N, SelectionDAG &DAG,
629 TargetLowering::DAGCombinerInfo &DCI,
630 const MipsSubtarget *Subtarget) {
631 if (DCI.isBeforeLegalize())
634 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
636 return SDValue(N, 0);
641 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG &DAG,
642 TargetLowering::DAGCombinerInfo &DCI,
643 const MipsSubtarget *Subtarget) {
644 if (DCI.isBeforeLegalize())
647 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
649 return SDValue(N, 0);
654 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG &DAG,
655 TargetLowering::DAGCombinerInfo &DCI,
656 const MipsSubtarget *Subtarget) {
657 if (DCI.isBeforeLegalizeOps())
660 EVT Ty = N->getValueType(0);
661 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
662 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
663 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
665 DebugLoc dl = N->getDebugLoc();
667 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
668 N->getOperand(0), N->getOperand(1));
669 SDValue InChain = DAG.getEntryNode();
670 SDValue InGlue = DivRem;
673 if (N->hasAnyUseOfValue(0)) {
674 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty,
676 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
677 InChain = CopyFromLo.getValue(1);
678 InGlue = CopyFromLo.getValue(2);
682 if (N->hasAnyUseOfValue(1)) {
683 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
685 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
691 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
693 default: llvm_unreachable("Unknown fp condition code!");
695 case ISD::SETOEQ: return Mips::FCOND_OEQ;
696 case ISD::SETUNE: return Mips::FCOND_UNE;
698 case ISD::SETOLT: return Mips::FCOND_OLT;
700 case ISD::SETOGT: return Mips::FCOND_OGT;
702 case ISD::SETOLE: return Mips::FCOND_OLE;
704 case ISD::SETOGE: return Mips::FCOND_OGE;
705 case ISD::SETULT: return Mips::FCOND_ULT;
706 case ISD::SETULE: return Mips::FCOND_ULE;
707 case ISD::SETUGT: return Mips::FCOND_UGT;
708 case ISD::SETUGE: return Mips::FCOND_UGE;
709 case ISD::SETUO: return Mips::FCOND_UN;
710 case ISD::SETO: return Mips::FCOND_OR;
712 case ISD::SETONE: return Mips::FCOND_ONE;
713 case ISD::SETUEQ: return Mips::FCOND_UEQ;
718 // Returns true if condition code has to be inverted.
719 static bool InvertFPCondCode(Mips::CondCode CC) {
720 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
723 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
724 "Illegal Condition Code");
729 // Creates and returns an FPCmp node from a setcc node.
730 // Returns Op if setcc is not a floating point comparison.
731 static SDValue CreateFPCmp(SelectionDAG &DAG, const SDValue &Op) {
732 // must be a SETCC node
733 if (Op.getOpcode() != ISD::SETCC)
736 SDValue LHS = Op.getOperand(0);
738 if (!LHS.getValueType().isFloatingPoint())
741 SDValue RHS = Op.getOperand(1);
742 DebugLoc dl = Op.getDebugLoc();
744 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
745 // node if necessary.
746 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
748 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
749 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
752 // Creates and returns a CMovFPT/F node.
753 static SDValue CreateCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
754 SDValue False, DebugLoc DL) {
755 bool invert = InvertFPCondCode((Mips::CondCode)
756 cast<ConstantSDNode>(Cond.getOperand(2))
759 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
760 True.getValueType(), True, False, Cond);
763 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
764 TargetLowering::DAGCombinerInfo &DCI,
765 const MipsSubtarget *Subtarget) {
766 if (DCI.isBeforeLegalizeOps())
769 SDValue SetCC = N->getOperand(0);
771 if ((SetCC.getOpcode() != ISD::SETCC) ||
772 !SetCC.getOperand(0).getValueType().isInteger())
775 SDValue False = N->getOperand(2);
776 EVT FalseTy = False.getValueType();
778 if (!FalseTy.isInteger())
781 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False);
783 if (!CN || CN->getZExtValue())
786 const DebugLoc DL = N->getDebugLoc();
787 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
788 SDValue True = N->getOperand(1);
790 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
791 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
793 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
796 static SDValue PerformANDCombine(SDNode *N, SelectionDAG &DAG,
797 TargetLowering::DAGCombinerInfo &DCI,
798 const MipsSubtarget *Subtarget) {
799 // Pattern match EXT.
800 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
801 // => ext $dst, $src, size, pos
802 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
805 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
806 unsigned ShiftRightOpc = ShiftRight.getOpcode();
808 // Op's first operand must be a shift right.
809 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
812 // The second operand of the shift must be an immediate.
814 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
817 uint64_t Pos = CN->getZExtValue();
818 uint64_t SMPos, SMSize;
820 // Op's second operand must be a shifted mask.
821 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
822 !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
825 // Return if the shifted mask does not start at bit 0 or the sum of its size
826 // and Pos exceeds the word's size.
827 EVT ValTy = N->getValueType(0);
828 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
831 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
832 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
833 DAG.getConstant(SMSize, MVT::i32));
836 static SDValue PerformORCombine(SDNode *N, SelectionDAG &DAG,
837 TargetLowering::DAGCombinerInfo &DCI,
838 const MipsSubtarget *Subtarget) {
839 // Pattern match INS.
840 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
841 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
842 // => ins $dst, $src, size, pos, $src1
843 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
846 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
847 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
850 // See if Op's first operand matches (and $src1 , mask0).
851 if (And0.getOpcode() != ISD::AND)
854 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
855 !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
858 // See if Op's second operand matches (and (shl $src, pos), mask1).
859 if (And1.getOpcode() != ISD::AND)
862 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
863 !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
866 // The shift masks must have the same position and size.
867 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
870 SDValue Shl = And1.getOperand(0);
871 if (Shl.getOpcode() != ISD::SHL)
874 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
877 unsigned Shamt = CN->getZExtValue();
879 // Return if the shift amount and the first bit position of mask are not the
881 EVT ValTy = N->getValueType(0);
882 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
885 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
886 DAG.getConstant(SMPos0, MVT::i32),
887 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
890 static SDValue PerformADDCombine(SDNode *N, SelectionDAG &DAG,
891 TargetLowering::DAGCombinerInfo &DCI,
892 const MipsSubtarget *Subtarget) {
893 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
895 if (DCI.isBeforeLegalizeOps())
898 SDValue Add = N->getOperand(1);
900 if (Add.getOpcode() != ISD::ADD)
903 SDValue Lo = Add.getOperand(1);
905 if ((Lo.getOpcode() != MipsISD::Lo) ||
906 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
909 EVT ValTy = N->getValueType(0);
910 DebugLoc DL = N->getDebugLoc();
912 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
914 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
917 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
919 SelectionDAG &DAG = DCI.DAG;
920 unsigned opc = N->getOpcode();
925 return PerformADDECombine(N, DAG, DCI, Subtarget);
927 return PerformSUBECombine(N, DAG, DCI, Subtarget);
930 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
932 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
934 return PerformANDCombine(N, DAG, DCI, Subtarget);
936 return PerformORCombine(N, DAG, DCI, Subtarget);
938 return PerformADDCombine(N, DAG, DCI, Subtarget);
945 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
946 SmallVectorImpl<SDValue> &Results,
947 SelectionDAG &DAG) const {
948 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
950 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
951 Results.push_back(Res.getValue(I));
955 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
956 SmallVectorImpl<SDValue> &Results,
957 SelectionDAG &DAG) const {
958 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
960 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
961 Results.push_back(Res.getValue(I));
964 SDValue MipsTargetLowering::
965 LowerOperation(SDValue Op, SelectionDAG &DAG) const
967 switch (Op.getOpcode())
969 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
970 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
971 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
972 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
973 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
974 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
975 case ISD::SELECT: return LowerSELECT(Op, DAG);
976 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
977 case ISD::SETCC: return LowerSETCC(Op, DAG);
978 case ISD::VASTART: return LowerVASTART(Op, DAG);
979 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
980 case ISD::FABS: return LowerFABS(Op, DAG);
981 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
982 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
983 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
984 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
985 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
986 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG, true);
987 case ISD::SRL_PARTS: return LowerShiftRightParts(Op, DAG, false);
988 case ISD::LOAD: return LowerLOAD(Op, DAG);
989 case ISD::STORE: return LowerSTORE(Op, DAG);
990 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
991 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
992 case ISD::ADD: return LowerADD(Op, DAG);
997 //===----------------------------------------------------------------------===//
998 // Lower helper functions
999 //===----------------------------------------------------------------------===//
1001 // AddLiveIn - This helper function adds the specified physical register to the
1002 // MachineFunction as a live in value. It also creates a corresponding
1003 // virtual register for it.
1005 AddLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
1007 assert(RC->contains(PReg) && "Not the correct regclass!");
1008 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
1009 MF.getRegInfo().addLiveIn(PReg, VReg);
1013 // Get fp branch code (not opcode) from condition code.
1014 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
1015 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
1016 return Mips::BRANCH_T;
1018 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
1019 "Invalid CondCode.");
1021 return Mips::BRANCH_F;
1025 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
1027 const MipsSubtarget *Subtarget,
1028 const TargetInstrInfo *TII,
1029 bool isFPCmp, unsigned Opc) {
1030 // There is no need to expand CMov instructions if target has
1031 // conditional moves.
1032 if (Subtarget->hasCondMov())
1035 // To "insert" a SELECT_CC instruction, we actually have to insert the
1036 // diamond control-flow pattern. The incoming instruction knows the
1037 // destination vreg to set, the condition code register to branch on, the
1038 // true/false values to select between, and a branch opcode to use.
1039 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1040 MachineFunction::iterator It = BB;
1047 // bNE r1, r0, copy1MBB
1048 // fallthrough --> copy0MBB
1049 MachineBasicBlock *thisMBB = BB;
1050 MachineFunction *F = BB->getParent();
1051 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1052 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1053 F->insert(It, copy0MBB);
1054 F->insert(It, sinkMBB);
1056 // Transfer the remainder of BB and its successor edges to sinkMBB.
1057 sinkMBB->splice(sinkMBB->begin(), BB,
1058 llvm::next(MachineBasicBlock::iterator(MI)),
1060 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1062 // Next, add the true and fallthrough blocks as its successors.
1063 BB->addSuccessor(copy0MBB);
1064 BB->addSuccessor(sinkMBB);
1066 // Emit the right instruction according to the type of the operands compared
1068 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
1070 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
1071 .addReg(Mips::ZERO).addMBB(sinkMBB);
1074 // %FalseValue = ...
1075 // # fallthrough to sinkMBB
1078 // Update machine-CFG edges
1079 BB->addSuccessor(sinkMBB);
1082 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
1087 BuildMI(*BB, BB->begin(), dl,
1088 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1089 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
1090 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
1092 BuildMI(*BB, BB->begin(), dl,
1093 TII->get(Mips::PHI), MI->getOperand(0).getReg())
1094 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
1095 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
1097 MI->eraseFromParent(); // The pseudo instruction is gone now.
1103 MipsTargetLowering::EmitBPOSGE32(MachineInstr *MI, MachineBasicBlock *BB) const{
1105 // bposge32_pseudo $vr0
1115 // $vr0 = phi($vr2, $fbb, $vr1, $tbb)
1117 MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
1118 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1119 const TargetRegisterClass *RC = &Mips::CPURegsRegClass;
1120 DebugLoc DL = MI->getDebugLoc();
1121 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1122 MachineFunction::iterator It = llvm::next(MachineFunction::iterator(BB));
1123 MachineFunction *F = BB->getParent();
1124 MachineBasicBlock *FBB = F->CreateMachineBasicBlock(LLVM_BB);
1125 MachineBasicBlock *TBB = F->CreateMachineBasicBlock(LLVM_BB);
1126 MachineBasicBlock *Sink = F->CreateMachineBasicBlock(LLVM_BB);
1129 F->insert(It, Sink);
1131 // Transfer the remainder of BB and its successor edges to Sink.
1132 Sink->splice(Sink->begin(), BB, llvm::next(MachineBasicBlock::iterator(MI)),
1134 Sink->transferSuccessorsAndUpdatePHIs(BB);
1137 BB->addSuccessor(FBB);
1138 BB->addSuccessor(TBB);
1139 FBB->addSuccessor(Sink);
1140 TBB->addSuccessor(Sink);
1142 // Insert the real bposge32 instruction to $BB.
1143 BuildMI(BB, DL, TII->get(Mips::BPOSGE32)).addMBB(TBB);
1146 unsigned VR2 = RegInfo.createVirtualRegister(RC);
1147 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::ADDiu), VR2)
1148 .addReg(Mips::ZERO).addImm(0);
1149 BuildMI(*FBB, FBB->end(), DL, TII->get(Mips::B)).addMBB(Sink);
1152 unsigned VR1 = RegInfo.createVirtualRegister(RC);
1153 BuildMI(*TBB, TBB->end(), DL, TII->get(Mips::ADDiu), VR1)
1154 .addReg(Mips::ZERO).addImm(1);
1156 // Insert phi function to $Sink.
1157 BuildMI(*Sink, Sink->begin(), DL, TII->get(Mips::PHI),
1158 MI->getOperand(0).getReg())
1159 .addReg(VR2).addMBB(FBB).addReg(VR1).addMBB(TBB);
1161 MI->eraseFromParent(); // The pseudo instruction is gone now.
1166 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1167 MachineBasicBlock *BB) const {
1168 switch (MI->getOpcode()) {
1169 default: llvm_unreachable("Unexpected instr type to insert");
1170 case Mips::ATOMIC_LOAD_ADD_I8:
1171 case Mips::ATOMIC_LOAD_ADD_I8_P8:
1172 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
1173 case Mips::ATOMIC_LOAD_ADD_I16:
1174 case Mips::ATOMIC_LOAD_ADD_I16_P8:
1175 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
1176 case Mips::ATOMIC_LOAD_ADD_I32:
1177 case Mips::ATOMIC_LOAD_ADD_I32_P8:
1178 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
1179 case Mips::ATOMIC_LOAD_ADD_I64:
1180 case Mips::ATOMIC_LOAD_ADD_I64_P8:
1181 return EmitAtomicBinary(MI, BB, 8, Mips::DADDu);
1183 case Mips::ATOMIC_LOAD_AND_I8:
1184 case Mips::ATOMIC_LOAD_AND_I8_P8:
1185 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
1186 case Mips::ATOMIC_LOAD_AND_I16:
1187 case Mips::ATOMIC_LOAD_AND_I16_P8:
1188 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
1189 case Mips::ATOMIC_LOAD_AND_I32:
1190 case Mips::ATOMIC_LOAD_AND_I32_P8:
1191 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
1192 case Mips::ATOMIC_LOAD_AND_I64:
1193 case Mips::ATOMIC_LOAD_AND_I64_P8:
1194 return EmitAtomicBinary(MI, BB, 8, Mips::AND64);
1196 case Mips::ATOMIC_LOAD_OR_I8:
1197 case Mips::ATOMIC_LOAD_OR_I8_P8:
1198 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
1199 case Mips::ATOMIC_LOAD_OR_I16:
1200 case Mips::ATOMIC_LOAD_OR_I16_P8:
1201 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
1202 case Mips::ATOMIC_LOAD_OR_I32:
1203 case Mips::ATOMIC_LOAD_OR_I32_P8:
1204 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
1205 case Mips::ATOMIC_LOAD_OR_I64:
1206 case Mips::ATOMIC_LOAD_OR_I64_P8:
1207 return EmitAtomicBinary(MI, BB, 8, Mips::OR64);
1209 case Mips::ATOMIC_LOAD_XOR_I8:
1210 case Mips::ATOMIC_LOAD_XOR_I8_P8:
1211 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
1212 case Mips::ATOMIC_LOAD_XOR_I16:
1213 case Mips::ATOMIC_LOAD_XOR_I16_P8:
1214 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
1215 case Mips::ATOMIC_LOAD_XOR_I32:
1216 case Mips::ATOMIC_LOAD_XOR_I32_P8:
1217 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
1218 case Mips::ATOMIC_LOAD_XOR_I64:
1219 case Mips::ATOMIC_LOAD_XOR_I64_P8:
1220 return EmitAtomicBinary(MI, BB, 8, Mips::XOR64);
1222 case Mips::ATOMIC_LOAD_NAND_I8:
1223 case Mips::ATOMIC_LOAD_NAND_I8_P8:
1224 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
1225 case Mips::ATOMIC_LOAD_NAND_I16:
1226 case Mips::ATOMIC_LOAD_NAND_I16_P8:
1227 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
1228 case Mips::ATOMIC_LOAD_NAND_I32:
1229 case Mips::ATOMIC_LOAD_NAND_I32_P8:
1230 return EmitAtomicBinary(MI, BB, 4, 0, true);
1231 case Mips::ATOMIC_LOAD_NAND_I64:
1232 case Mips::ATOMIC_LOAD_NAND_I64_P8:
1233 return EmitAtomicBinary(MI, BB, 8, 0, true);
1235 case Mips::ATOMIC_LOAD_SUB_I8:
1236 case Mips::ATOMIC_LOAD_SUB_I8_P8:
1237 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
1238 case Mips::ATOMIC_LOAD_SUB_I16:
1239 case Mips::ATOMIC_LOAD_SUB_I16_P8:
1240 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
1241 case Mips::ATOMIC_LOAD_SUB_I32:
1242 case Mips::ATOMIC_LOAD_SUB_I32_P8:
1243 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
1244 case Mips::ATOMIC_LOAD_SUB_I64:
1245 case Mips::ATOMIC_LOAD_SUB_I64_P8:
1246 return EmitAtomicBinary(MI, BB, 8, Mips::DSUBu);
1248 case Mips::ATOMIC_SWAP_I8:
1249 case Mips::ATOMIC_SWAP_I8_P8:
1250 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
1251 case Mips::ATOMIC_SWAP_I16:
1252 case Mips::ATOMIC_SWAP_I16_P8:
1253 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
1254 case Mips::ATOMIC_SWAP_I32:
1255 case Mips::ATOMIC_SWAP_I32_P8:
1256 return EmitAtomicBinary(MI, BB, 4, 0);
1257 case Mips::ATOMIC_SWAP_I64:
1258 case Mips::ATOMIC_SWAP_I64_P8:
1259 return EmitAtomicBinary(MI, BB, 8, 0);
1261 case Mips::ATOMIC_CMP_SWAP_I8:
1262 case Mips::ATOMIC_CMP_SWAP_I8_P8:
1263 return EmitAtomicCmpSwapPartword(MI, BB, 1);
1264 case Mips::ATOMIC_CMP_SWAP_I16:
1265 case Mips::ATOMIC_CMP_SWAP_I16_P8:
1266 return EmitAtomicCmpSwapPartword(MI, BB, 2);
1267 case Mips::ATOMIC_CMP_SWAP_I32:
1268 case Mips::ATOMIC_CMP_SWAP_I32_P8:
1269 return EmitAtomicCmpSwap(MI, BB, 4);
1270 case Mips::ATOMIC_CMP_SWAP_I64:
1271 case Mips::ATOMIC_CMP_SWAP_I64_P8:
1272 return EmitAtomicCmpSwap(MI, BB, 8);
1273 case Mips::BPOSGE32_PSEUDO:
1274 return EmitBPOSGE32(MI, BB);
1278 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1279 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1281 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
1282 unsigned Size, unsigned BinOpcode,
1284 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1286 MachineFunction *MF = BB->getParent();
1287 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1288 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1289 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1290 DebugLoc dl = MI->getDebugLoc();
1291 unsigned LL, SC, AND, NOR, ZERO, BEQ;
1294 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1295 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1302 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1303 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1306 ZERO = Mips::ZERO_64;
1310 unsigned OldVal = MI->getOperand(0).getReg();
1311 unsigned Ptr = MI->getOperand(1).getReg();
1312 unsigned Incr = MI->getOperand(2).getReg();
1314 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1315 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1316 unsigned Success = RegInfo.createVirtualRegister(RC);
1318 // insert new blocks after the current block
1319 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1320 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1321 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1322 MachineFunction::iterator It = BB;
1324 MF->insert(It, loopMBB);
1325 MF->insert(It, exitMBB);
1327 // Transfer the remainder of BB and its successor edges to exitMBB.
1328 exitMBB->splice(exitMBB->begin(), BB,
1329 llvm::next(MachineBasicBlock::iterator(MI)),
1331 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1335 // fallthrough --> loopMBB
1336 BB->addSuccessor(loopMBB);
1337 loopMBB->addSuccessor(loopMBB);
1338 loopMBB->addSuccessor(exitMBB);
1341 // ll oldval, 0(ptr)
1342 // <binop> storeval, oldval, incr
1343 // sc success, storeval, 0(ptr)
1344 // beq success, $0, loopMBB
1346 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1348 // and andres, oldval, incr
1349 // nor storeval, $0, andres
1350 BuildMI(BB, dl, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1351 BuildMI(BB, dl, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1352 } else if (BinOpcode) {
1353 // <binop> storeval, oldval, incr
1354 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1358 BuildMI(BB, dl, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1359 BuildMI(BB, dl, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1361 MI->eraseFromParent(); // The instruction is gone now.
1367 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
1368 MachineBasicBlock *BB,
1369 unsigned Size, unsigned BinOpcode,
1371 assert((Size == 1 || Size == 2) &&
1372 "Unsupported size for EmitAtomicBinaryPartial.");
1374 MachineFunction *MF = BB->getParent();
1375 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1376 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1377 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1378 DebugLoc dl = MI->getDebugLoc();
1379 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1380 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1382 unsigned Dest = MI->getOperand(0).getReg();
1383 unsigned Ptr = MI->getOperand(1).getReg();
1384 unsigned Incr = MI->getOperand(2).getReg();
1386 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1387 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1388 unsigned Mask = RegInfo.createVirtualRegister(RC);
1389 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1390 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1391 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1392 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1393 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1394 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1395 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1396 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1397 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1398 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1399 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1400 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1401 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1402 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1403 unsigned Success = RegInfo.createVirtualRegister(RC);
1405 // insert new blocks after the current block
1406 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1407 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1408 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1409 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1410 MachineFunction::iterator It = BB;
1412 MF->insert(It, loopMBB);
1413 MF->insert(It, sinkMBB);
1414 MF->insert(It, exitMBB);
1416 // Transfer the remainder of BB and its successor edges to exitMBB.
1417 exitMBB->splice(exitMBB->begin(), BB,
1418 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1419 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1421 BB->addSuccessor(loopMBB);
1422 loopMBB->addSuccessor(loopMBB);
1423 loopMBB->addSuccessor(sinkMBB);
1424 sinkMBB->addSuccessor(exitMBB);
1427 // addiu masklsb2,$0,-4 # 0xfffffffc
1428 // and alignedaddr,ptr,masklsb2
1429 // andi ptrlsb2,ptr,3
1430 // sll shiftamt,ptrlsb2,3
1431 // ori maskupper,$0,255 # 0xff
1432 // sll mask,maskupper,shiftamt
1433 // nor mask2,$0,mask
1434 // sll incr2,incr,shiftamt
1436 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1437 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1438 .addReg(Mips::ZERO).addImm(-4);
1439 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1440 .addReg(Ptr).addReg(MaskLSB2);
1441 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1442 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1443 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1444 .addReg(Mips::ZERO).addImm(MaskImm);
1445 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1446 .addReg(ShiftAmt).addReg(MaskUpper);
1447 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1448 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1450 // atomic.load.binop
1452 // ll oldval,0(alignedaddr)
1453 // binop binopres,oldval,incr2
1454 // and newval,binopres,mask
1455 // and maskedoldval0,oldval,mask2
1456 // or storeval,maskedoldval0,newval
1457 // sc success,storeval,0(alignedaddr)
1458 // beq success,$0,loopMBB
1462 // ll oldval,0(alignedaddr)
1463 // and newval,incr2,mask
1464 // and maskedoldval0,oldval,mask2
1465 // or storeval,maskedoldval0,newval
1466 // sc success,storeval,0(alignedaddr)
1467 // beq success,$0,loopMBB
1470 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1472 // and andres, oldval, incr2
1473 // nor binopres, $0, andres
1474 // and newval, binopres, mask
1475 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1476 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1477 .addReg(Mips::ZERO).addReg(AndRes);
1478 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1479 } else if (BinOpcode) {
1480 // <binop> binopres, oldval, incr2
1481 // and newval, binopres, mask
1482 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1483 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1484 } else {// atomic.swap
1485 // and newval, incr2, mask
1486 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1489 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1490 .addReg(OldVal).addReg(Mask2);
1491 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1492 .addReg(MaskedOldVal0).addReg(NewVal);
1493 BuildMI(BB, dl, TII->get(SC), Success)
1494 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1495 BuildMI(BB, dl, TII->get(Mips::BEQ))
1496 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1499 // and maskedoldval1,oldval,mask
1500 // srl srlres,maskedoldval1,shiftamt
1501 // sll sllres,srlres,24
1502 // sra dest,sllres,24
1504 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1506 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1507 .addReg(OldVal).addReg(Mask);
1508 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1509 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1510 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1511 .addReg(SrlRes).addImm(ShiftImm);
1512 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1513 .addReg(SllRes).addImm(ShiftImm);
1515 MI->eraseFromParent(); // The instruction is gone now.
1521 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1522 MachineBasicBlock *BB,
1523 unsigned Size) const {
1524 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1526 MachineFunction *MF = BB->getParent();
1527 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1528 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1529 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1530 DebugLoc dl = MI->getDebugLoc();
1531 unsigned LL, SC, ZERO, BNE, BEQ;
1534 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1535 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1541 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1542 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1543 ZERO = Mips::ZERO_64;
1548 unsigned Dest = MI->getOperand(0).getReg();
1549 unsigned Ptr = MI->getOperand(1).getReg();
1550 unsigned OldVal = MI->getOperand(2).getReg();
1551 unsigned NewVal = MI->getOperand(3).getReg();
1553 unsigned Success = RegInfo.createVirtualRegister(RC);
1555 // insert new blocks after the current block
1556 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1557 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1558 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1559 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1560 MachineFunction::iterator It = BB;
1562 MF->insert(It, loop1MBB);
1563 MF->insert(It, loop2MBB);
1564 MF->insert(It, exitMBB);
1566 // Transfer the remainder of BB and its successor edges to exitMBB.
1567 exitMBB->splice(exitMBB->begin(), BB,
1568 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1569 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1573 // fallthrough --> loop1MBB
1574 BB->addSuccessor(loop1MBB);
1575 loop1MBB->addSuccessor(exitMBB);
1576 loop1MBB->addSuccessor(loop2MBB);
1577 loop2MBB->addSuccessor(loop1MBB);
1578 loop2MBB->addSuccessor(exitMBB);
1582 // bne dest, oldval, exitMBB
1584 BuildMI(BB, dl, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1585 BuildMI(BB, dl, TII->get(BNE))
1586 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1589 // sc success, newval, 0(ptr)
1590 // beq success, $0, loop1MBB
1592 BuildMI(BB, dl, TII->get(SC), Success)
1593 .addReg(NewVal).addReg(Ptr).addImm(0);
1594 BuildMI(BB, dl, TII->get(BEQ))
1595 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1597 MI->eraseFromParent(); // The instruction is gone now.
1603 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1604 MachineBasicBlock *BB,
1605 unsigned Size) const {
1606 assert((Size == 1 || Size == 2) &&
1607 "Unsupported size for EmitAtomicCmpSwapPartial.");
1609 MachineFunction *MF = BB->getParent();
1610 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1611 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1612 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1613 DebugLoc dl = MI->getDebugLoc();
1614 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1615 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1617 unsigned Dest = MI->getOperand(0).getReg();
1618 unsigned Ptr = MI->getOperand(1).getReg();
1619 unsigned CmpVal = MI->getOperand(2).getReg();
1620 unsigned NewVal = MI->getOperand(3).getReg();
1622 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1623 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1624 unsigned Mask = RegInfo.createVirtualRegister(RC);
1625 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1626 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1627 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1628 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1629 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1630 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1631 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1632 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1633 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1634 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1635 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1636 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1637 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1638 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1639 unsigned Success = RegInfo.createVirtualRegister(RC);
1641 // insert new blocks after the current block
1642 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1643 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1644 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1645 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1646 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1647 MachineFunction::iterator It = BB;
1649 MF->insert(It, loop1MBB);
1650 MF->insert(It, loop2MBB);
1651 MF->insert(It, sinkMBB);
1652 MF->insert(It, exitMBB);
1654 // Transfer the remainder of BB and its successor edges to exitMBB.
1655 exitMBB->splice(exitMBB->begin(), BB,
1656 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1657 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1659 BB->addSuccessor(loop1MBB);
1660 loop1MBB->addSuccessor(sinkMBB);
1661 loop1MBB->addSuccessor(loop2MBB);
1662 loop2MBB->addSuccessor(loop1MBB);
1663 loop2MBB->addSuccessor(sinkMBB);
1664 sinkMBB->addSuccessor(exitMBB);
1666 // FIXME: computation of newval2 can be moved to loop2MBB.
1668 // addiu masklsb2,$0,-4 # 0xfffffffc
1669 // and alignedaddr,ptr,masklsb2
1670 // andi ptrlsb2,ptr,3
1671 // sll shiftamt,ptrlsb2,3
1672 // ori maskupper,$0,255 # 0xff
1673 // sll mask,maskupper,shiftamt
1674 // nor mask2,$0,mask
1675 // andi maskedcmpval,cmpval,255
1676 // sll shiftedcmpval,maskedcmpval,shiftamt
1677 // andi maskednewval,newval,255
1678 // sll shiftednewval,maskednewval,shiftamt
1679 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1680 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1681 .addReg(Mips::ZERO).addImm(-4);
1682 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1683 .addReg(Ptr).addReg(MaskLSB2);
1684 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1685 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1686 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1687 .addReg(Mips::ZERO).addImm(MaskImm);
1688 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1689 .addReg(ShiftAmt).addReg(MaskUpper);
1690 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1691 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1692 .addReg(CmpVal).addImm(MaskImm);
1693 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1694 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1695 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1696 .addReg(NewVal).addImm(MaskImm);
1697 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1698 .addReg(ShiftAmt).addReg(MaskedNewVal);
1701 // ll oldval,0(alginedaddr)
1702 // and maskedoldval0,oldval,mask
1703 // bne maskedoldval0,shiftedcmpval,sinkMBB
1705 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1706 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1707 .addReg(OldVal).addReg(Mask);
1708 BuildMI(BB, dl, TII->get(Mips::BNE))
1709 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1712 // and maskedoldval1,oldval,mask2
1713 // or storeval,maskedoldval1,shiftednewval
1714 // sc success,storeval,0(alignedaddr)
1715 // beq success,$0,loop1MBB
1717 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1718 .addReg(OldVal).addReg(Mask2);
1719 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1720 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1721 BuildMI(BB, dl, TII->get(SC), Success)
1722 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1723 BuildMI(BB, dl, TII->get(Mips::BEQ))
1724 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1727 // srl srlres,maskedoldval0,shiftamt
1728 // sll sllres,srlres,24
1729 // sra dest,sllres,24
1731 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1733 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1734 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1735 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1736 .addReg(SrlRes).addImm(ShiftImm);
1737 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1738 .addReg(SllRes).addImm(ShiftImm);
1740 MI->eraseFromParent(); // The instruction is gone now.
1745 //===----------------------------------------------------------------------===//
1746 // Misc Lower Operation implementation
1747 //===----------------------------------------------------------------------===//
1748 SDValue MipsTargetLowering::
1749 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1751 // The first operand is the chain, the second is the condition, the third is
1752 // the block to branch to if the condition is true.
1753 SDValue Chain = Op.getOperand(0);
1754 SDValue Dest = Op.getOperand(2);
1755 DebugLoc dl = Op.getDebugLoc();
1757 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1759 // Return if flag is not set by a floating point comparison.
1760 if (CondRes.getOpcode() != MipsISD::FPCmp)
1763 SDValue CCNode = CondRes.getOperand(2);
1765 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1766 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1768 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1772 SDValue MipsTargetLowering::
1773 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1775 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1777 // Return if flag is not set by a floating point comparison.
1778 if (Cond.getOpcode() != MipsISD::FPCmp)
1781 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1785 SDValue MipsTargetLowering::
1786 LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1788 DebugLoc DL = Op.getDebugLoc();
1789 EVT Ty = Op.getOperand(0).getValueType();
1790 SDValue Cond = DAG.getNode(ISD::SETCC, DL, getSetCCResultType(Ty),
1791 Op.getOperand(0), Op.getOperand(1),
1794 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1798 SDValue MipsTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1799 SDValue Cond = CreateFPCmp(DAG, Op);
1801 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1802 "Floating point operand expected.");
1804 SDValue True = DAG.getConstant(1, MVT::i32);
1805 SDValue False = DAG.getConstant(0, MVT::i32);
1807 return CreateCMovFP(DAG, Cond, True, False, Op.getDebugLoc());
1810 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1811 SelectionDAG &DAG) const {
1812 // FIXME there isn't actually debug info here
1813 DebugLoc dl = Op.getDebugLoc();
1814 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1816 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1817 const MipsTargetObjectFile &TLOF =
1818 (const MipsTargetObjectFile&)getObjFileLowering();
1820 // %gp_rel relocation
1821 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1822 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1824 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl,
1825 DAG.getVTList(MVT::i32), &GA, 1);
1826 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32);
1827 return DAG.getNode(ISD::ADD, dl, MVT::i32, GPReg, GPRelNode);
1830 // %hi/%lo relocation
1831 return getAddrNonPIC(Op, DAG);
1834 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1835 return getAddrLocal(Op, DAG, HasMips64);
1838 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16,
1839 MipsII::MO_GOT_LO16);
1841 return getAddrGlobal(Op, DAG,
1842 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16);
1845 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1846 SelectionDAG &DAG) const {
1847 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1848 return getAddrNonPIC(Op, DAG);
1850 return getAddrLocal(Op, DAG, HasMips64);
1853 SDValue MipsTargetLowering::
1854 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1856 // If the relocation model is PIC, use the General Dynamic TLS Model or
1857 // Local Dynamic TLS model, otherwise use the Initial Exec or
1858 // Local Exec TLS Model.
1860 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1861 DebugLoc dl = GA->getDebugLoc();
1862 const GlobalValue *GV = GA->getGlobal();
1863 EVT PtrVT = getPointerTy();
1865 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1867 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1868 // General Dynamic and Local Dynamic TLS Model.
1869 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1872 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, Flag);
1873 SDValue Argument = DAG.getNode(MipsISD::Wrapper, dl, PtrVT,
1874 GetGlobalReg(DAG, PtrVT), TGA);
1875 unsigned PtrSize = PtrVT.getSizeInBits();
1876 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1878 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1882 Entry.Node = Argument;
1884 Args.push_back(Entry);
1886 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy,
1887 false, false, false, false, 0, CallingConv::C,
1888 /*isTailCall=*/false, /*doesNotRet=*/false,
1889 /*isReturnValueUsed=*/true,
1890 TlsGetAddr, Args, DAG, dl);
1891 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1893 SDValue Ret = CallResult.first;
1895 if (model != TLSModel::LocalDynamic)
1898 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1899 MipsII::MO_DTPREL_HI);
1900 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1901 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1902 MipsII::MO_DTPREL_LO);
1903 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1904 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Ret);
1905 return DAG.getNode(ISD::ADD, dl, PtrVT, Add, Lo);
1909 if (model == TLSModel::InitialExec) {
1910 // Initial Exec TLS Model
1911 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1912 MipsII::MO_GOTTPREL);
1913 TGA = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, GetGlobalReg(DAG, PtrVT),
1915 Offset = DAG.getLoad(PtrVT, dl,
1916 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1917 false, false, false, 0);
1919 // Local Exec TLS Model
1920 assert(model == TLSModel::LocalExec);
1921 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1922 MipsII::MO_TPREL_HI);
1923 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1924 MipsII::MO_TPREL_LO);
1925 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1926 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1927 Offset = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1930 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1931 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1934 SDValue MipsTargetLowering::
1935 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1937 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1938 return getAddrNonPIC(Op, DAG);
1940 return getAddrLocal(Op, DAG, HasMips64);
1943 SDValue MipsTargetLowering::
1944 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1946 // gp_rel relocation
1947 // FIXME: we should reference the constant pool using small data sections,
1948 // but the asm printer currently doesn't support this feature without
1949 // hacking it. This feature should come soon so we can uncomment the
1951 //if (IsInSmallSection(C->getType())) {
1952 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1953 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1954 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1956 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64)
1957 return getAddrNonPIC(Op, DAG);
1959 return getAddrLocal(Op, DAG, HasMips64);
1962 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1963 MachineFunction &MF = DAG.getMachineFunction();
1964 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1966 DebugLoc dl = Op.getDebugLoc();
1967 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1970 // vastart just stores the address of the VarArgsFrameIndex slot into the
1971 // memory location argument.
1972 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1973 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1974 MachinePointerInfo(SV), false, false, 0);
1977 static SDValue LowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
1978 EVT TyX = Op.getOperand(0).getValueType();
1979 EVT TyY = Op.getOperand(1).getValueType();
1980 SDValue Const1 = DAG.getConstant(1, MVT::i32);
1981 SDValue Const31 = DAG.getConstant(31, MVT::i32);
1982 DebugLoc DL = Op.getDebugLoc();
1985 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1987 SDValue X = (TyX == MVT::f32) ?
1988 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1989 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1991 SDValue Y = (TyY == MVT::f32) ?
1992 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1993 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1997 // ext E, Y, 31, 1 ; extract bit31 of Y
1998 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1999 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
2000 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
2003 // srl SrlX, SllX, 1
2005 // sll SllY, SrlX, 31
2006 // or Or, SrlX, SllY
2007 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2008 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2009 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
2010 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
2011 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
2014 if (TyX == MVT::f32)
2015 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
2017 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2018 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
2019 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2022 static SDValue LowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2023 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
2024 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
2025 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
2026 SDValue Const1 = DAG.getConstant(1, MVT::i32);
2027 DebugLoc DL = Op.getDebugLoc();
2029 // Bitcast to integer nodes.
2030 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
2031 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
2034 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
2035 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
2036 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
2037 DAG.getConstant(WidthY - 1, MVT::i32), Const1);
2039 if (WidthX > WidthY)
2040 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
2041 else if (WidthY > WidthX)
2042 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
2044 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
2045 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X);
2046 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
2049 // (d)sll SllX, X, 1
2050 // (d)srl SrlX, SllX, 1
2051 // (d)srl SrlY, Y, width(Y)-1
2052 // (d)sll SllY, SrlX, width(Y)-1
2053 // or Or, SrlX, SllY
2054 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
2055 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
2056 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
2057 DAG.getConstant(WidthY - 1, MVT::i32));
2059 if (WidthX > WidthY)
2060 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
2061 else if (WidthY > WidthX)
2062 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
2064 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
2065 DAG.getConstant(WidthX - 1, MVT::i32));
2066 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
2067 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
2071 MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2072 if (Subtarget->hasMips64())
2073 return LowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2());
2075 return LowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2());
2078 static SDValue LowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2079 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
2080 DebugLoc DL = Op.getDebugLoc();
2082 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
2084 SDValue X = (Op.getValueType() == MVT::f32) ?
2085 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
2086 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
2091 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32,
2092 DAG.getRegister(Mips::ZERO, MVT::i32),
2093 DAG.getConstant(31, MVT::i32), Const1, X);
2095 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
2096 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
2099 if (Op.getValueType() == MVT::f32)
2100 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res);
2102 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2103 Op.getOperand(0), DAG.getConstant(0, MVT::i32));
2104 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
2107 static SDValue LowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) {
2108 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32);
2109 DebugLoc DL = Op.getDebugLoc();
2111 // Bitcast to integer node.
2112 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0));
2116 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64,
2117 DAG.getRegister(Mips::ZERO_64, MVT::i64),
2118 DAG.getConstant(63, MVT::i32), Const1, X);
2120 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1);
2121 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1);
2124 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res);
2128 MipsTargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
2129 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64))
2130 return LowerFABS64(Op, DAG, Subtarget->hasMips32r2());
2132 return LowerFABS32(Op, DAG, Subtarget->hasMips32r2());
2135 SDValue MipsTargetLowering::
2136 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2138 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2139 "Frame address can only be determined for current frame.");
2141 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2142 MFI->setFrameAddressIsTaken(true);
2143 EVT VT = Op.getValueType();
2144 DebugLoc dl = Op.getDebugLoc();
2145 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2146 IsN64 ? Mips::FP_64 : Mips::FP, VT);
2150 SDValue MipsTargetLowering::LowerRETURNADDR(SDValue Op,
2151 SelectionDAG &DAG) const {
2153 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2154 "Return address can be determined only for current frame.");
2156 MachineFunction &MF = DAG.getMachineFunction();
2157 MachineFrameInfo *MFI = MF.getFrameInfo();
2158 EVT VT = Op.getValueType();
2159 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA;
2160 MFI->setReturnAddressIsTaken(true);
2162 // Return RA, which contains the return address. Mark it an implicit live-in.
2163 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2164 return DAG.getCopyFromReg(DAG.getEntryNode(), Op.getDebugLoc(), Reg, VT);
2167 // TODO: set SType according to the desired memory barrier behavior.
2169 MipsTargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const {
2171 DebugLoc dl = Op.getDebugLoc();
2172 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
2173 DAG.getConstant(SType, MVT::i32));
2176 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
2177 SelectionDAG &DAG) const {
2178 // FIXME: Need pseudo-fence for 'singlethread' fences
2179 // FIXME: Set SType for weaker fences where supported/appropriate.
2181 DebugLoc dl = Op.getDebugLoc();
2182 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
2183 DAG.getConstant(SType, MVT::i32));
2186 SDValue MipsTargetLowering::LowerShiftLeftParts(SDValue Op,
2187 SelectionDAG &DAG) const {
2188 DebugLoc DL = Op.getDebugLoc();
2189 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2190 SDValue Shamt = Op.getOperand(2);
2193 // lo = (shl lo, shamt)
2194 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2197 // hi = (shl lo, shamt[4:0])
2198 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2199 DAG.getConstant(-1, MVT::i32));
2200 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo,
2201 DAG.getConstant(1, MVT::i32));
2202 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo,
2204 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt);
2205 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2206 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt);
2207 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2208 DAG.getConstant(0x20, MVT::i32));
2209 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2210 DAG.getConstant(0, MVT::i32), ShiftLeftLo);
2211 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or);
2213 SDValue Ops[2] = {Lo, Hi};
2214 return DAG.getMergeValues(Ops, 2, DL);
2217 SDValue MipsTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2219 DebugLoc DL = Op.getDebugLoc();
2220 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2221 SDValue Shamt = Op.getOperand(2);
2224 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2226 // hi = (sra hi, shamt)
2228 // hi = (srl hi, shamt)
2231 // lo = (sra hi, shamt[4:0])
2232 // hi = (sra hi, 31)
2234 // lo = (srl hi, shamt[4:0])
2236 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2237 DAG.getConstant(-1, MVT::i32));
2238 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi,
2239 DAG.getConstant(1, MVT::i32));
2240 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not);
2241 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt);
2242 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo);
2243 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2245 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2246 DAG.getConstant(0x20, MVT::i32));
2247 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi,
2248 DAG.getConstant(31, MVT::i32));
2249 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or);
2250 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond,
2251 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32),
2254 SDValue Ops[2] = {Lo, Hi};
2255 return DAG.getMergeValues(Ops, 2, DL);
2258 static SDValue CreateLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2259 SDValue Chain, SDValue Src, unsigned Offset) {
2260 SDValue Ptr = LD->getBasePtr();
2261 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2262 EVT BasePtrVT = Ptr.getValueType();
2263 DebugLoc DL = LD->getDebugLoc();
2264 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2267 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2268 DAG.getConstant(Offset, BasePtrVT));
2270 SDValue Ops[] = { Chain, Ptr, Src };
2271 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2272 LD->getMemOperand());
2275 // Expand an unaligned 32 or 64-bit integer load node.
2276 SDValue MipsTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2277 LoadSDNode *LD = cast<LoadSDNode>(Op);
2278 EVT MemVT = LD->getMemoryVT();
2280 // Return if load is aligned or if MemVT is neither i32 nor i64.
2281 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2282 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2285 bool IsLittle = Subtarget->isLittle();
2286 EVT VT = Op.getValueType();
2287 ISD::LoadExtType ExtType = LD->getExtensionType();
2288 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2290 assert((VT == MVT::i32) || (VT == MVT::i64));
2293 // (set dst, (i64 (load baseptr)))
2295 // (set tmp, (ldl (add baseptr, 7), undef))
2296 // (set dst, (ldr baseptr, tmp))
2297 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2298 SDValue LDL = CreateLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2300 return CreateLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2304 SDValue LWL = CreateLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2306 SDValue LWR = CreateLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2310 // (set dst, (i32 (load baseptr))) or
2311 // (set dst, (i64 (sextload baseptr))) or
2312 // (set dst, (i64 (extload baseptr)))
2314 // (set tmp, (lwl (add baseptr, 3), undef))
2315 // (set dst, (lwr baseptr, tmp))
2316 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2317 (ExtType == ISD::EXTLOAD))
2320 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2323 // (set dst, (i64 (zextload baseptr)))
2325 // (set tmp0, (lwl (add baseptr, 3), undef))
2326 // (set tmp1, (lwr baseptr, tmp0))
2327 // (set tmp2, (shl tmp1, 32))
2328 // (set dst, (srl tmp2, 32))
2329 DebugLoc DL = LD->getDebugLoc();
2330 SDValue Const32 = DAG.getConstant(32, MVT::i32);
2331 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2332 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2333 SDValue Ops[] = { SRL, LWR.getValue(1) };
2334 return DAG.getMergeValues(Ops, 2, DL);
2337 static SDValue CreateStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2338 SDValue Chain, unsigned Offset) {
2339 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2340 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2341 DebugLoc DL = SD->getDebugLoc();
2342 SDVTList VTList = DAG.getVTList(MVT::Other);
2345 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2346 DAG.getConstant(Offset, BasePtrVT));
2348 SDValue Ops[] = { Chain, Value, Ptr };
2349 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT,
2350 SD->getMemOperand());
2353 // Expand an unaligned 32 or 64-bit integer store node.
2354 SDValue MipsTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2355 StoreSDNode *SD = cast<StoreSDNode>(Op);
2356 EVT MemVT = SD->getMemoryVT();
2358 // Return if store is aligned or if MemVT is neither i32 nor i64.
2359 if ((SD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2360 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2363 bool IsLittle = Subtarget->isLittle();
2364 SDValue Value = SD->getValue(), Chain = SD->getChain();
2365 EVT VT = Value.getValueType();
2368 // (store val, baseptr) or
2369 // (truncstore val, baseptr)
2371 // (swl val, (add baseptr, 3))
2372 // (swr val, baseptr)
2373 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2374 SDValue SWL = CreateStoreLR(MipsISD::SWL, DAG, SD, Chain,
2376 return CreateStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2379 assert(VT == MVT::i64);
2382 // (store val, baseptr)
2384 // (sdl val, (add baseptr, 7))
2385 // (sdr val, baseptr)
2386 SDValue SDL = CreateStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2387 return CreateStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2390 // This function expands mips intrinsic nodes which have 64-bit input operands
2391 // or output values.
2393 // out64 = intrinsic-node in64
2395 // lo = copy (extract-element (in64, 0))
2396 // hi = copy (extract-element (in64, 1))
2397 // mips-specific-node
2400 // out64 = merge-values (v0, v1)
2402 static SDValue LowerDSPIntr(SDValue Op, SelectionDAG &DAG,
2403 unsigned Opc, bool HasI64In, bool HasI64Out) {
2404 DebugLoc DL = Op.getDebugLoc();
2405 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other;
2406 SDValue Chain = HasChainIn ? Op->getOperand(0) : DAG.getEntryNode();
2407 SmallVector<SDValue, 3> Ops;
2410 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2411 Op->getOperand(1 + HasChainIn),
2412 DAG.getConstant(0, MVT::i32));
2413 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32,
2414 Op->getOperand(1 + HasChainIn),
2415 DAG.getConstant(1, MVT::i32));
2417 Chain = DAG.getCopyToReg(Chain, DL, Mips::LO, InLo, SDValue());
2418 Chain = DAG.getCopyToReg(Chain, DL, Mips::HI, InHi, Chain.getValue(1));
2420 Ops.push_back(Chain);
2421 Ops.append(Op->op_begin() + HasChainIn + 2, Op->op_end());
2422 Ops.push_back(Chain.getValue(1));
2424 Ops.push_back(Chain);
2425 Ops.append(Op->op_begin() + HasChainIn + 1, Op->op_end());
2429 return DAG.getNode(Opc, DL, Op->value_begin(), Op->getNumValues(),
2430 Ops.begin(), Ops.size());
2432 SDValue Intr = DAG.getNode(Opc, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2433 Ops.begin(), Ops.size());
2434 SDValue OutLo = DAG.getCopyFromReg(Intr.getValue(0), DL, Mips::LO, MVT::i32,
2436 SDValue OutHi = DAG.getCopyFromReg(OutLo.getValue(1), DL, Mips::HI, MVT::i32,
2438 SDValue Out = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, OutLo, OutHi);
2443 SDValue Vals[] = { Out, OutHi.getValue(1) };
2444 return DAG.getMergeValues(Vals, 2, DL);
2447 SDValue MipsTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
2448 SelectionDAG &DAG) const {
2449 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) {
2452 case Intrinsic::mips_shilo:
2453 return LowerDSPIntr(Op, DAG, MipsISD::SHILO, true, true);
2454 case Intrinsic::mips_dpau_h_qbl:
2455 return LowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL, true, true);
2456 case Intrinsic::mips_dpau_h_qbr:
2457 return LowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR, true, true);
2458 case Intrinsic::mips_dpsu_h_qbl:
2459 return LowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL, true, true);
2460 case Intrinsic::mips_dpsu_h_qbr:
2461 return LowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR, true, true);
2462 case Intrinsic::mips_dpa_w_ph:
2463 return LowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH, true, true);
2464 case Intrinsic::mips_dps_w_ph:
2465 return LowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH, true, true);
2466 case Intrinsic::mips_dpax_w_ph:
2467 return LowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH, true, true);
2468 case Intrinsic::mips_dpsx_w_ph:
2469 return LowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH, true, true);
2470 case Intrinsic::mips_mulsa_w_ph:
2471 return LowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH, true, true);
2472 case Intrinsic::mips_mult:
2473 return LowerDSPIntr(Op, DAG, MipsISD::MULT, false, true);
2474 case Intrinsic::mips_multu:
2475 return LowerDSPIntr(Op, DAG, MipsISD::MULTU, false, true);
2476 case Intrinsic::mips_madd:
2477 return LowerDSPIntr(Op, DAG, MipsISD::MADD_DSP, true, true);
2478 case Intrinsic::mips_maddu:
2479 return LowerDSPIntr(Op, DAG, MipsISD::MADDU_DSP, true, true);
2480 case Intrinsic::mips_msub:
2481 return LowerDSPIntr(Op, DAG, MipsISD::MSUB_DSP, true, true);
2482 case Intrinsic::mips_msubu:
2483 return LowerDSPIntr(Op, DAG, MipsISD::MSUBU_DSP, true, true);
2487 SDValue MipsTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
2488 SelectionDAG &DAG) const {
2489 switch (cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue()) {
2492 case Intrinsic::mips_extp:
2493 return LowerDSPIntr(Op, DAG, MipsISD::EXTP, true, false);
2494 case Intrinsic::mips_extpdp:
2495 return LowerDSPIntr(Op, DAG, MipsISD::EXTPDP, true, false);
2496 case Intrinsic::mips_extr_w:
2497 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_W, true, false);
2498 case Intrinsic::mips_extr_r_w:
2499 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W, true, false);
2500 case Intrinsic::mips_extr_rs_w:
2501 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W, true, false);
2502 case Intrinsic::mips_extr_s_h:
2503 return LowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H, true, false);
2504 case Intrinsic::mips_mthlip:
2505 return LowerDSPIntr(Op, DAG, MipsISD::MTHLIP, true, true);
2506 case Intrinsic::mips_mulsaq_s_w_ph:
2507 return LowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH, true, true);
2508 case Intrinsic::mips_maq_s_w_phl:
2509 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL, true, true);
2510 case Intrinsic::mips_maq_s_w_phr:
2511 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR, true, true);
2512 case Intrinsic::mips_maq_sa_w_phl:
2513 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL, true, true);
2514 case Intrinsic::mips_maq_sa_w_phr:
2515 return LowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR, true, true);
2516 case Intrinsic::mips_dpaq_s_w_ph:
2517 return LowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH, true, true);
2518 case Intrinsic::mips_dpsq_s_w_ph:
2519 return LowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH, true, true);
2520 case Intrinsic::mips_dpaq_sa_l_w:
2521 return LowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W, true, true);
2522 case Intrinsic::mips_dpsq_sa_l_w:
2523 return LowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W, true, true);
2524 case Intrinsic::mips_dpaqx_s_w_ph:
2525 return LowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH, true, true);
2526 case Intrinsic::mips_dpaqx_sa_w_ph:
2527 return LowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH, true, true);
2528 case Intrinsic::mips_dpsqx_s_w_ph:
2529 return LowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH, true, true);
2530 case Intrinsic::mips_dpsqx_sa_w_ph:
2531 return LowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH, true, true);
2535 SDValue MipsTargetLowering::LowerADD(SDValue Op, SelectionDAG &DAG) const {
2536 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2537 || cast<ConstantSDNode>
2538 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2539 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2543 // (add (frameaddr 0), (frame_to_args_offset))
2544 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2545 // (add FrameObject, 0)
2546 // where FrameObject is a fixed StackObject with offset 0 which points to
2547 // the old stack pointer.
2548 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2549 EVT ValTy = Op->getValueType(0);
2550 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2551 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2552 return DAG.getNode(ISD::ADD, Op->getDebugLoc(), ValTy, InArgsAddr,
2553 DAG.getConstant(0, ValTy));
2556 //===----------------------------------------------------------------------===//
2557 // Calling Convention Implementation
2558 //===----------------------------------------------------------------------===//
2560 //===----------------------------------------------------------------------===//
2561 // TODO: Implement a generic logic using tblgen that can support this.
2562 // Mips O32 ABI rules:
2564 // i32 - Passed in A0, A1, A2, A3 and stack
2565 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2566 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2567 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2568 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2569 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
2572 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2573 //===----------------------------------------------------------------------===//
2575 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
2576 MVT LocVT, CCValAssign::LocInfo LocInfo,
2577 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2579 static const unsigned IntRegsSize=4, FloatRegsSize=2;
2581 static const uint16_t IntRegs[] = {
2582 Mips::A0, Mips::A1, Mips::A2, Mips::A3
2584 static const uint16_t F32Regs[] = {
2585 Mips::F12, Mips::F14
2587 static const uint16_t F64Regs[] = {
2591 // Do not process byval args here.
2592 if (ArgFlags.isByVal())
2595 // Promote i8 and i16
2596 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2598 if (ArgFlags.isSExt())
2599 LocInfo = CCValAssign::SExt;
2600 else if (ArgFlags.isZExt())
2601 LocInfo = CCValAssign::ZExt;
2603 LocInfo = CCValAssign::AExt;
2608 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2609 // is true: function is vararg, argument is 3rd or higher, there is previous
2610 // argument which is not f32 or f64.
2611 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
2612 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
2613 unsigned OrigAlign = ArgFlags.getOrigAlign();
2614 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2616 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2617 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2618 // If this is the first part of an i64 arg,
2619 // the allocated register must be either A0 or A2.
2620 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2621 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2623 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2624 // Allocate int register and shadow next int register. If first
2625 // available register is Mips::A1 or Mips::A3, shadow it too.
2626 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2627 if (Reg == Mips::A1 || Reg == Mips::A3)
2628 Reg = State.AllocateReg(IntRegs, IntRegsSize);
2629 State.AllocateReg(IntRegs, IntRegsSize);
2631 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2632 // we are guaranteed to find an available float register
2633 if (ValVT == MVT::f32) {
2634 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
2635 // Shadow int register
2636 State.AllocateReg(IntRegs, IntRegsSize);
2638 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
2639 // Shadow int registers
2640 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
2641 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2642 State.AllocateReg(IntRegs, IntRegsSize);
2643 State.AllocateReg(IntRegs, IntRegsSize);
2646 llvm_unreachable("Cannot handle this ValVT.");
2649 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2651 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2653 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2658 #include "MipsGenCallingConv.inc"
2660 //===----------------------------------------------------------------------===//
2661 // Call Calling Convention Implementation
2662 //===----------------------------------------------------------------------===//
2664 static const unsigned O32IntRegsSize = 4;
2666 // Return next O32 integer argument register.
2667 static unsigned getNextIntArgReg(unsigned Reg) {
2668 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2669 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2672 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2673 /// for tail call optimization.
2674 bool MipsTargetLowering::
2675 IsEligibleForTailCallOptimization(const MipsCC &MipsCCInfo,
2676 unsigned NextStackOffset,
2677 const MipsFunctionInfo& FI) const {
2678 if (!EnableMipsTailCalls)
2681 // No tail call optimization for mips16.
2682 if (Subtarget->inMips16Mode())
2685 // Return false if either the callee or caller has a byval argument.
2686 if (MipsCCInfo.hasByValArg() || FI.hasByvalArg())
2689 // Return true if the callee's argument area is no larger than the
2691 return NextStackOffset <= FI.getIncomingArgSize();
2695 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2696 SDValue Chain, SDValue Arg, DebugLoc DL,
2697 bool IsTailCall, SelectionDAG &DAG) const {
2699 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr,
2700 DAG.getIntPtrConstant(Offset));
2701 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2705 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2706 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2707 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2708 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2709 /*isVolatile=*/ true, false, 0);
2712 /// LowerCall - functions arguments are copied from virtual regs to
2713 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2715 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2716 SmallVectorImpl<SDValue> &InVals) const {
2717 SelectionDAG &DAG = CLI.DAG;
2718 DebugLoc &dl = CLI.DL;
2719 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2720 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2721 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2722 SDValue Chain = CLI.Chain;
2723 SDValue Callee = CLI.Callee;
2724 bool &isTailCall = CLI.IsTailCall;
2725 CallingConv::ID CallConv = CLI.CallConv;
2726 bool isVarArg = CLI.IsVarArg;
2728 MachineFunction &MF = DAG.getMachineFunction();
2729 MachineFrameInfo *MFI = MF.getFrameInfo();
2730 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2731 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2733 // Analyze operands of the call, assigning locations to each operand.
2734 SmallVector<CCValAssign, 16> ArgLocs;
2735 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2736 getTargetMachine(), ArgLocs, *DAG.getContext());
2737 MipsCC MipsCCInfo(CallConv, isVarArg, IsO32, CCInfo);
2739 MipsCCInfo.analyzeCallOperands(Outs);
2741 // Get a count of how many bytes are to be pushed on the stack.
2742 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2744 // Check if it's really possible to do a tail call.
2747 IsEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset,
2748 *MF.getInfo<MipsFunctionInfo>());
2753 // Chain is the output chain of the last Load/Store or CopyToReg node.
2754 // ByValChain is the output chain of the last Memcpy node created for copying
2755 // byval arguments to the stack.
2756 unsigned StackAlignment = TFL->getStackAlignment();
2757 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2758 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2761 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal);
2763 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl,
2764 IsN64 ? Mips::SP_64 : Mips::SP,
2767 // With EABI is it possible to have 16 args on registers.
2768 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
2769 SmallVector<SDValue, 8> MemOpChains;
2770 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
2772 // Walk the register/memloc assignments, inserting copies/loads.
2773 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2774 SDValue Arg = OutVals[i];
2775 CCValAssign &VA = ArgLocs[i];
2776 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2777 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2780 if (Flags.isByVal()) {
2781 assert(Flags.getByValSize() &&
2782 "ByVal args of size 0 should have been ignored by front-end.");
2783 assert(ByValArg != MipsCCInfo.byval_end());
2784 assert(!isTailCall &&
2785 "Do not tail-call optimize if there is a byval argument.");
2786 passByValArg(Chain, dl, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2787 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle());
2792 // Promote the value if needed.
2793 switch (VA.getLocInfo()) {
2794 default: llvm_unreachable("Unknown loc info!");
2795 case CCValAssign::Full:
2796 if (VA.isRegLoc()) {
2797 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2798 (ValVT == MVT::f64 && LocVT == MVT::i64))
2799 Arg = DAG.getNode(ISD::BITCAST, dl, LocVT, Arg);
2800 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2801 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2802 Arg, DAG.getConstant(0, MVT::i32));
2803 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2804 Arg, DAG.getConstant(1, MVT::i32));
2805 if (!Subtarget->isLittle())
2807 unsigned LocRegLo = VA.getLocReg();
2808 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2809 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2810 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2815 case CCValAssign::SExt:
2816 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, LocVT, Arg);
2818 case CCValAssign::ZExt:
2819 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, LocVT, Arg);
2821 case CCValAssign::AExt:
2822 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, LocVT, Arg);
2826 // Arguments that can be passed on register must be kept at
2827 // RegsToPass vector
2828 if (VA.isRegLoc()) {
2829 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2833 // Register can't get to this point...
2834 assert(VA.isMemLoc());
2836 // emit ISD::STORE whichs stores the
2837 // parameter value to a stack Location
2838 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2839 Chain, Arg, dl, isTailCall, DAG));
2842 // Transform all store nodes into one single node because all store
2843 // nodes are independent of each other.
2844 if (!MemOpChains.empty())
2845 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2846 &MemOpChains[0], MemOpChains.size());
2848 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2849 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2850 // node so that legalize doesn't hack it.
2851 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2852 bool GlobalOrExternal = false;
2855 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2857 if (G->getGlobal()->hasInternalLinkage())
2858 Callee = getAddrLocal(Callee, DAG, HasMips64);
2860 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2861 MipsII::MO_CALL_LO16);
2863 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2865 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(), 0,
2866 MipsII::MO_NO_FLAG);
2867 GlobalOrExternal = true;
2869 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2870 if (!IsN64 && !IsPIC) // !N64 && static
2871 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2872 MipsII::MO_NO_FLAG);
2874 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16,
2875 MipsII::MO_CALL_LO16);
2877 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_DISP);
2879 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL);
2881 GlobalOrExternal = true;
2886 // T9 register operand.
2889 // T9 should contain the address of the callee function if
2890 // -reloction-model=pic or it is an indirect call.
2891 if (IsPICCall || !GlobalOrExternal) {
2893 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
2894 Chain = DAG.getCopyToReg(Chain, dl, T9Reg, Callee, SDValue(0, 0));
2895 InFlag = Chain.getValue(1);
2897 if (Subtarget->inMips16Mode())
2898 T9 = DAG.getRegister(T9Reg, getPointerTy());
2900 Callee = DAG.getRegister(T9Reg, getPointerTy());
2903 // Insert node "GP copy globalreg" before call to function.
2904 // Lazy-binding stubs require GP to point to the GOT.
2906 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP;
2907 EVT Ty = IsN64 ? MVT::i64 : MVT::i32;
2908 RegsToPass.push_back(std::make_pair(GPReg, GetGlobalReg(DAG, Ty)));
2911 // Build a sequence of copy-to-reg nodes chained together with token
2912 // chain and flag operands which copy the outgoing args into registers.
2913 // The InFlag in necessary since all emitted instructions must be
2915 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2916 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2917 RegsToPass[i].second, InFlag);
2918 InFlag = Chain.getValue(1);
2921 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2922 // = Chain, Callee, Reg#1, Reg#2, ...
2924 // Returns a chain & a flag for retval copy to use.
2925 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2926 SmallVector<SDValue, 8> Ops;
2927 Ops.push_back(Chain);
2928 Ops.push_back(Callee);
2930 // Add argument registers to the end of the list so that they are
2931 // known live into the call.
2932 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2933 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2934 RegsToPass[i].second.getValueType()));
2936 // Add T9 register operand.
2940 // Add a register mask operand representing the call-preserved registers.
2941 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2942 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2943 assert(Mask && "Missing call preserved mask for calling convention");
2944 Ops.push_back(DAG.getRegisterMask(Mask));
2946 if (InFlag.getNode())
2947 Ops.push_back(InFlag);
2950 return DAG.getNode(MipsISD::TailCall, dl, MVT::Other, &Ops[0], Ops.size());
2952 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2953 InFlag = Chain.getValue(1);
2955 // Create the CALLSEQ_END node.
2956 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2957 DAG.getIntPtrConstant(0, true), InFlag);
2958 InFlag = Chain.getValue(1);
2960 // Handle result values, copying them out of physregs into vregs that we
2962 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2963 Ins, dl, DAG, InVals);
2966 /// LowerCallResult - Lower the result values of a call into the
2967 /// appropriate copies out of appropriate physical registers.
2969 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2970 CallingConv::ID CallConv, bool isVarArg,
2971 const SmallVectorImpl<ISD::InputArg> &Ins,
2972 DebugLoc dl, SelectionDAG &DAG,
2973 SmallVectorImpl<SDValue> &InVals) const {
2974 // Assign locations to each value returned by this call.
2975 SmallVector<CCValAssign, 16> RVLocs;
2976 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2977 getTargetMachine(), RVLocs, *DAG.getContext());
2979 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2981 // Copy all of the result registers out of their specified physreg.
2982 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2983 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2984 RVLocs[i].getValVT(), InFlag).getValue(1);
2985 InFlag = Chain.getValue(2);
2986 InVals.push_back(Chain.getValue(0));
2992 //===----------------------------------------------------------------------===//
2993 // Formal Arguments Calling Convention Implementation
2994 //===----------------------------------------------------------------------===//
2995 /// LowerFormalArguments - transform physical registers into virtual registers
2996 /// and generate load operations for arguments places on the stack.
2998 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2999 CallingConv::ID CallConv,
3001 const SmallVectorImpl<ISD::InputArg> &Ins,
3002 DebugLoc dl, SelectionDAG &DAG,
3003 SmallVectorImpl<SDValue> &InVals)
3005 MachineFunction &MF = DAG.getMachineFunction();
3006 MachineFrameInfo *MFI = MF.getFrameInfo();
3007 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3009 MipsFI->setVarArgsFrameIndex(0);
3011 // Used with vargs to acumulate store chains.
3012 std::vector<SDValue> OutChains;
3014 // Assign locations to all of the incoming arguments.
3015 SmallVector<CCValAssign, 16> ArgLocs;
3016 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3017 getTargetMachine(), ArgLocs, *DAG.getContext());
3018 MipsCC MipsCCInfo(CallConv, isVarArg, IsO32, CCInfo);
3020 MipsCCInfo.analyzeFormalArguments(Ins);
3021 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
3022 MipsCCInfo.hasByValArg());
3024 Function::const_arg_iterator FuncArg =
3025 DAG.getMachineFunction().getFunction()->arg_begin();
3026 unsigned CurArgIdx = 0;
3027 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin();
3029 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3030 CCValAssign &VA = ArgLocs[i];
3031 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx);
3032 CurArgIdx = Ins[i].OrigArgIndex;
3033 EVT ValVT = VA.getValVT();
3034 ISD::ArgFlagsTy Flags = Ins[i].Flags;
3035 bool IsRegLoc = VA.isRegLoc();
3037 if (Flags.isByVal()) {
3038 assert(Flags.getByValSize() &&
3039 "ByVal args of size 0 should have been ignored by front-end.");
3040 assert(ByValArg != MipsCCInfo.byval_end());
3041 copyByValRegs(Chain, dl, OutChains, DAG, Flags, InVals, &*FuncArg,
3042 MipsCCInfo, *ByValArg);
3047 // Arguments stored on registers
3049 EVT RegVT = VA.getLocVT();
3050 unsigned ArgReg = VA.getLocReg();
3051 const TargetRegisterClass *RC;
3053 if (RegVT == MVT::i32)
3054 RC = &Mips::CPURegsRegClass;
3055 else if (RegVT == MVT::i64)
3056 RC = &Mips::CPU64RegsRegClass;
3057 else if (RegVT == MVT::f32)
3058 RC = &Mips::FGR32RegClass;
3059 else if (RegVT == MVT::f64)
3060 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass;
3062 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
3064 // Transform the arguments stored on
3065 // physical registers into virtual ones
3066 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
3067 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3069 // If this is an 8 or 16-bit value, it has been passed promoted
3070 // to 32 bits. Insert an assert[sz]ext to capture this, then
3071 // truncate to the right size.
3072 if (VA.getLocInfo() != CCValAssign::Full) {
3073 unsigned Opcode = 0;
3074 if (VA.getLocInfo() == CCValAssign::SExt)
3075 Opcode = ISD::AssertSext;
3076 else if (VA.getLocInfo() == CCValAssign::ZExt)
3077 Opcode = ISD::AssertZext;
3079 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
3080 DAG.getValueType(ValVT));
3081 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue);
3084 // Handle floating point arguments passed in integer registers.
3085 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
3086 (RegVT == MVT::i64 && ValVT == MVT::f64))
3087 ArgValue = DAG.getNode(ISD::BITCAST, dl, ValVT, ArgValue);
3088 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
3089 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
3090 getNextIntArgReg(ArgReg), RC);
3091 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
3092 if (!Subtarget->isLittle())
3093 std::swap(ArgValue, ArgValue2);
3094 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
3095 ArgValue, ArgValue2);
3098 InVals.push_back(ArgValue);
3099 } else { // VA.isRegLoc()
3102 assert(VA.isMemLoc());
3104 // The stack pointer offset is relative to the caller stack frame.
3105 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
3106 VA.getLocMemOffset(), true);
3108 // Create load nodes to retrieve arguments from the stack
3109 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3110 InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN,
3111 MachinePointerInfo::getFixedStack(FI),
3112 false, false, false, 0));
3116 // The mips ABIs for returning structs by value requires that we copy
3117 // the sret argument into $v0 for the return. Save the argument into
3118 // a virtual register so that we can access it from the return points.
3119 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
3120 unsigned Reg = MipsFI->getSRetReturnReg();
3122 Reg = MF.getRegInfo().
3123 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32));
3124 MipsFI->setSRetReturnReg(Reg);
3126 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
3127 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
3131 writeVarArgRegs(OutChains, MipsCCInfo, Chain, dl, DAG);
3133 // All stores are grouped in one node to allow the matching between
3134 // the size of Ins and InVals. This only happens when on varg functions
3135 if (!OutChains.empty()) {
3136 OutChains.push_back(Chain);
3137 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3138 &OutChains[0], OutChains.size());
3144 //===----------------------------------------------------------------------===//
3145 // Return Value Calling Convention Implementation
3146 //===----------------------------------------------------------------------===//
3149 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3150 MachineFunction &MF, bool isVarArg,
3151 const SmallVectorImpl<ISD::OutputArg> &Outs,
3152 LLVMContext &Context) const {
3153 SmallVector<CCValAssign, 16> RVLocs;
3154 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3156 return CCInfo.CheckReturn(Outs, RetCC_Mips);
3160 MipsTargetLowering::LowerReturn(SDValue Chain,
3161 CallingConv::ID CallConv, bool isVarArg,
3162 const SmallVectorImpl<ISD::OutputArg> &Outs,
3163 const SmallVectorImpl<SDValue> &OutVals,
3164 DebugLoc dl, SelectionDAG &DAG) const {
3166 // CCValAssign - represent the assignment of
3167 // the return value to a location
3168 SmallVector<CCValAssign, 16> RVLocs;
3170 // CCState - Info about the registers and stack slot.
3171 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3172 getTargetMachine(), RVLocs, *DAG.getContext());
3174 // Analize return values.
3175 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
3177 // If this is the first return lowered for this function, add
3178 // the regs to the liveout set for the function.
3179 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3180 for (unsigned i = 0; i != RVLocs.size(); ++i)
3181 if (RVLocs[i].isRegLoc())
3182 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3187 // Copy the result values into the output registers.
3188 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3189 CCValAssign &VA = RVLocs[i];
3190 assert(VA.isRegLoc() && "Can only return in registers!");
3192 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
3194 // guarantee that all emitted copies are
3195 // stuck together, avoiding something bad
3196 Flag = Chain.getValue(1);
3199 // The mips ABIs for returning structs by value requires that we copy
3200 // the sret argument into $v0 for the return. We saved the argument into
3201 // a virtual register in the entry block, so now we copy the value out
3203 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
3204 MachineFunction &MF = DAG.getMachineFunction();
3205 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3206 unsigned Reg = MipsFI->getSRetReturnReg();
3209 llvm_unreachable("sret virtual register not created in the entry block");
3210 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
3211 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0;
3213 Chain = DAG.getCopyToReg(Chain, dl, V0, Val, Flag);
3214 Flag = Chain.getValue(1);
3215 MF.getRegInfo().addLiveOut(V0);
3218 // Return on Mips is always a "jr $ra"
3220 return DAG.getNode(MipsISD::Ret, dl, MVT::Other, Chain, Flag);
3223 return DAG.getNode(MipsISD::Ret, dl, MVT::Other, Chain);
3226 //===----------------------------------------------------------------------===//
3227 // Mips Inline Assembly Support
3228 //===----------------------------------------------------------------------===//
3230 /// getConstraintType - Given a constraint letter, return the type of
3231 /// constraint it is for this target.
3232 MipsTargetLowering::ConstraintType MipsTargetLowering::
3233 getConstraintType(const std::string &Constraint) const
3235 // Mips specific constrainy
3236 // GCC config/mips/constraints.md
3238 // 'd' : An address register. Equivalent to r
3239 // unless generating MIPS16 code.
3240 // 'y' : Equivalent to r; retained for
3241 // backwards compatibility.
3242 // 'c' : A register suitable for use in an indirect
3243 // jump. This will always be $25 for -mabicalls.
3244 // 'l' : The lo register. 1 word storage.
3245 // 'x' : The hilo register pair. Double word storage.
3246 if (Constraint.size() == 1) {
3247 switch (Constraint[0]) {
3255 return C_RegisterClass;
3258 return TargetLowering::getConstraintType(Constraint);
3261 /// Examine constraint type and operand type and determine a weight value.
3262 /// This object must already have been set up with the operand type
3263 /// and the current alternative constraint selected.
3264 TargetLowering::ConstraintWeight
3265 MipsTargetLowering::getSingleConstraintMatchWeight(
3266 AsmOperandInfo &info, const char *constraint) const {
3267 ConstraintWeight weight = CW_Invalid;
3268 Value *CallOperandVal = info.CallOperandVal;
3269 // If we don't have a value, we can't do a match,
3270 // but allow it at the lowest weight.
3271 if (CallOperandVal == NULL)
3273 Type *type = CallOperandVal->getType();
3274 // Look at the constraint type.
3275 switch (*constraint) {
3277 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3281 if (type->isIntegerTy())
3282 weight = CW_Register;
3285 if (type->isFloatTy())
3286 weight = CW_Register;
3288 case 'c': // $25 for indirect jumps
3289 case 'l': // lo register
3290 case 'x': // hilo register pair
3291 if (type->isIntegerTy())
3292 weight = CW_SpecificReg;
3294 case 'I': // signed 16 bit immediate
3295 case 'J': // integer zero
3296 case 'K': // unsigned 16 bit immediate
3297 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3298 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3299 case 'O': // signed 15 bit immediate (+- 16383)
3300 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3301 if (isa<ConstantInt>(CallOperandVal))
3302 weight = CW_Constant;
3308 /// Given a register class constraint, like 'r', if this corresponds directly
3309 /// to an LLVM register class, return a register of 0 and the register class
3311 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
3312 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
3314 if (Constraint.size() == 1) {
3315 switch (Constraint[0]) {
3316 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3317 case 'y': // Same as 'r'. Exists for compatibility.
3319 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3320 if (Subtarget->inMips16Mode())
3321 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3322 return std::make_pair(0U, &Mips::CPURegsRegClass);
3324 if (VT == MVT::i64 && !HasMips64)
3325 return std::make_pair(0U, &Mips::CPURegsRegClass);
3326 if (VT == MVT::i64 && HasMips64)
3327 return std::make_pair(0U, &Mips::CPU64RegsRegClass);
3328 // This will generate an error message
3329 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3332 return std::make_pair(0U, &Mips::FGR32RegClass);
3333 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) {
3334 if (Subtarget->isFP64bit())
3335 return std::make_pair(0U, &Mips::FGR64RegClass);
3336 return std::make_pair(0U, &Mips::AFGR64RegClass);
3339 case 'c': // register suitable for indirect jump
3341 return std::make_pair((unsigned)Mips::T9, &Mips::CPURegsRegClass);
3342 assert(VT == MVT::i64 && "Unexpected type.");
3343 return std::make_pair((unsigned)Mips::T9_64, &Mips::CPU64RegsRegClass);
3344 case 'l': // register suitable for indirect jump
3346 return std::make_pair((unsigned)Mips::LO, &Mips::HILORegClass);
3347 return std::make_pair((unsigned)Mips::LO64, &Mips::HILO64RegClass);
3348 case 'x': // register suitable for indirect jump
3349 // Fixme: Not triggering the use of both hi and low
3350 // This will generate an error message
3351 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
3354 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3357 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3358 /// vector. If it is invalid, don't add anything to Ops.
3359 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3360 std::string &Constraint,
3361 std::vector<SDValue>&Ops,
3362 SelectionDAG &DAG) const {
3363 SDValue Result(0, 0);
3365 // Only support length 1 constraints for now.
3366 if (Constraint.length() > 1) return;
3368 char ConstraintLetter = Constraint[0];
3369 switch (ConstraintLetter) {
3370 default: break; // This will fall through to the generic implementation
3371 case 'I': // Signed 16 bit constant
3372 // If this fails, the parent routine will give an error
3373 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3374 EVT Type = Op.getValueType();
3375 int64_t Val = C->getSExtValue();
3376 if (isInt<16>(Val)) {
3377 Result = DAG.getTargetConstant(Val, Type);
3382 case 'J': // integer zero
3383 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3384 EVT Type = Op.getValueType();
3385 int64_t Val = C->getZExtValue();
3387 Result = DAG.getTargetConstant(0, Type);
3392 case 'K': // unsigned 16 bit immediate
3393 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3394 EVT Type = Op.getValueType();
3395 uint64_t Val = (uint64_t)C->getZExtValue();
3396 if (isUInt<16>(Val)) {
3397 Result = DAG.getTargetConstant(Val, Type);
3402 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3403 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3404 EVT Type = Op.getValueType();
3405 int64_t Val = C->getSExtValue();
3406 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3407 Result = DAG.getTargetConstant(Val, Type);
3412 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3413 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3414 EVT Type = Op.getValueType();
3415 int64_t Val = C->getSExtValue();
3416 if ((Val >= -65535) && (Val <= -1)) {
3417 Result = DAG.getTargetConstant(Val, Type);
3422 case 'O': // signed 15 bit immediate
3423 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3424 EVT Type = Op.getValueType();
3425 int64_t Val = C->getSExtValue();
3426 if ((isInt<15>(Val))) {
3427 Result = DAG.getTargetConstant(Val, Type);
3432 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3433 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3434 EVT Type = Op.getValueType();
3435 int64_t Val = C->getSExtValue();
3436 if ((Val <= 65535) && (Val >= 1)) {
3437 Result = DAG.getTargetConstant(Val, Type);
3444 if (Result.getNode()) {
3445 Ops.push_back(Result);
3449 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3453 MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const {
3454 // No global is ever allowed as a base.
3459 case 0: // "r+i" or just "i", depending on HasBaseReg.
3462 if (!AM.HasBaseReg) // allow "r+i".
3464 return false; // disallow "r+r" or "r+r+i".
3473 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3474 // The Mips target isn't yet aware of offsets.
3478 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3479 unsigned SrcAlign, bool IsZeroVal,
3481 MachineFunction &MF) const {
3482 if (Subtarget->hasMips64())
3488 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3489 if (VT != MVT::f32 && VT != MVT::f64)
3491 if (Imm.isNegZero())
3493 return Imm.isZero();
3496 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3498 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3500 return TargetLowering::getJumpTableEncoding();
3503 MipsTargetLowering::MipsCC::MipsCC(CallingConv::ID CallConv, bool IsVarArg,
3504 bool IsO32, CCState &Info) : CCInfo(Info) {
3505 UseRegsForByval = true;
3509 NumIntArgRegs = array_lengthof(O32IntRegs);
3510 ReservedArgArea = 16;
3511 IntArgRegs = ShadowRegs = O32IntRegs;
3512 FixedFn = VarFn = CC_MipsO32;
3515 NumIntArgRegs = array_lengthof(Mips64IntRegs);
3516 ReservedArgArea = 0;
3517 IntArgRegs = Mips64IntRegs;
3518 ShadowRegs = Mips64DPRegs;
3520 VarFn = CC_MipsN_VarArg;
3523 if (CallConv == CallingConv::Fast) {
3525 UseRegsForByval = false;
3526 ReservedArgArea = 0;
3527 FixedFn = VarFn = CC_Mips_FastCC;
3530 // Pre-allocate reserved argument area.
3531 CCInfo.AllocateStack(ReservedArgArea, 1);
3534 void MipsTargetLowering::MipsCC::
3535 analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args) {
3536 unsigned NumOpnds = Args.size();
3538 for (unsigned I = 0; I != NumOpnds; ++I) {
3539 MVT ArgVT = Args[I].VT;
3540 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3543 if (ArgFlags.isByVal()) {
3544 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3548 if (Args[I].IsFixed)
3549 R = FixedFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3551 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
3555 dbgs() << "Call operand #" << I << " has unhandled type "
3556 << EVT(ArgVT).getEVTString();
3558 llvm_unreachable(0);
3563 void MipsTargetLowering::MipsCC::
3564 analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args) {
3565 unsigned NumArgs = Args.size();
3567 for (unsigned I = 0; I != NumArgs; ++I) {
3568 MVT ArgVT = Args[I].VT;
3569 ISD::ArgFlagsTy ArgFlags = Args[I].Flags;
3571 if (ArgFlags.isByVal()) {
3572 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags);
3576 if (!FixedFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo))
3580 dbgs() << "Formal Arg #" << I << " has unhandled type "
3581 << EVT(ArgVT).getEVTString();
3583 llvm_unreachable(0);
3588 MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT,
3590 CCValAssign::LocInfo LocInfo,
3591 ISD::ArgFlagsTy ArgFlags) {
3592 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0.");
3594 struct ByValArgInfo ByVal;
3595 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize);
3596 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize),
3599 if (UseRegsForByval)
3600 allocateRegs(ByVal, ByValSize, Align);
3602 // Allocate space on caller's stack.
3603 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs,
3605 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT,
3607 ByValArgs.push_back(ByVal);
3610 void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal,
3613 assert(!(ByValSize % RegSize) && !(Align % RegSize) &&
3614 "Byval argument's size and alignment should be a multiple of"
3617 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs);
3619 // If Align > RegSize, the first arg register must be even.
3620 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) {
3621 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]);
3625 // Mark the registers allocated.
3626 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs);
3627 ByValSize -= RegSize, ++I, ++ByVal.NumRegs)
3628 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3631 void MipsTargetLowering::
3632 copyByValRegs(SDValue Chain, DebugLoc DL, std::vector<SDValue> &OutChains,
3633 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags,
3634 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg,
3635 const MipsCC &CC, const ByValArgInfo &ByVal) const {
3636 MachineFunction &MF = DAG.getMachineFunction();
3637 MachineFrameInfo *MFI = MF.getFrameInfo();
3638 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize();
3639 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3643 FrameObjOffset = (int)CC.reservedArgArea() -
3644 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize());
3646 FrameObjOffset = ByVal.Address;
3648 // Create frame object.
3649 EVT PtrTy = getPointerTy();
3650 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3651 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3652 InVals.push_back(FIN);
3657 // Copy arg registers.
3658 EVT RegTy = MVT::getIntegerVT(CC.regSize() * 8);
3659 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3661 for (unsigned I = 0; I < ByVal.NumRegs; ++I) {
3662 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I];
3663 unsigned VReg = AddLiveIn(MF, ArgReg, RC);
3664 unsigned Offset = I * CC.regSize();
3665 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3666 DAG.getConstant(Offset, PtrTy));
3667 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3668 StorePtr, MachinePointerInfo(FuncArg, Offset),
3670 OutChains.push_back(Store);
3674 // Copy byVal arg to registers and stack.
3675 void MipsTargetLowering::
3676 passByValArg(SDValue Chain, DebugLoc DL,
3677 SmallVector<std::pair<unsigned, SDValue>, 16> &RegsToPass,
3678 SmallVector<SDValue, 8> &MemOpChains, SDValue StackPtr,
3679 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
3680 const MipsCC &CC, const ByValArgInfo &ByVal,
3681 const ISD::ArgFlagsTy &Flags, bool isLittle) const {
3682 unsigned ByValSize = Flags.getByValSize();
3683 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
3684 unsigned RegSize = CC.regSize();
3685 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize);
3686 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8);
3688 if (ByVal.NumRegs) {
3689 const uint16_t *ArgRegs = CC.intArgRegs();
3690 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize);
3693 // Copy words to registers.
3694 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) {
3695 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3696 DAG.getConstant(Offset, PtrTy));
3697 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3698 MachinePointerInfo(), false, false, false,
3700 MemOpChains.push_back(LoadVal.getValue(1));
3701 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3702 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3705 // Return if the struct has been fully copied.
3706 if (ByValSize == Offset)
3709 // Copy the remainder of the byval argument with sub-word loads and shifts.
3710 if (LeftoverBytes) {
3711 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) &&
3712 "Size of the remainder should be smaller than RegSize.");
3715 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0;
3716 Offset < ByValSize; LoadSize /= 2) {
3717 unsigned RemSize = ByValSize - Offset;
3719 if (RemSize < LoadSize)
3723 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3724 DAG.getConstant(Offset, PtrTy));
3726 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr,
3727 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
3728 false, false, Alignment);
3729 MemOpChains.push_back(LoadVal.getValue(1));
3731 // Shift the loaded value.
3735 Shamt = TotalSizeLoaded;
3737 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8;
3739 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3740 DAG.getConstant(Shamt, MVT::i32));
3743 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3748 TotalSizeLoaded += LoadSize;
3749 Alignment = std::min(Alignment, LoadSize);
3752 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I];
3753 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3758 // Copy remainder of byval arg to it with memcpy.
3759 unsigned MemCpySize = ByValSize - Offset;
3760 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3761 DAG.getConstant(Offset, PtrTy));
3762 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3763 DAG.getIntPtrConstant(ByVal.Address));
3764 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3765 DAG.getConstant(MemCpySize, PtrTy), Alignment,
3766 /*isVolatile=*/false, /*AlwaysInline=*/false,
3767 MachinePointerInfo(0), MachinePointerInfo(0));
3768 MemOpChains.push_back(Chain);
3772 MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3773 const MipsCC &CC, SDValue Chain,
3774 DebugLoc DL, SelectionDAG &DAG) const {
3775 unsigned NumRegs = CC.numIntArgRegs();
3776 const uint16_t *ArgRegs = CC.intArgRegs();
3777 const CCState &CCInfo = CC.getCCInfo();
3778 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs);
3779 unsigned RegSize = CC.regSize();
3780 EVT RegTy = MVT::getIntegerVT(RegSize * 8);
3781 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3782 MachineFunction &MF = DAG.getMachineFunction();
3783 MachineFrameInfo *MFI = MF.getFrameInfo();
3784 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3786 // Offset of the first variable argument from stack pointer.
3790 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize);
3793 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx));
3795 // Record the frame index of the first variable argument
3796 // which is a value necessary to VASTART.
3797 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3798 MipsFI->setVarArgsFrameIndex(FI);
3800 // Copy the integer registers that have not been used for argument passing
3801 // to the argument register save area. For O32, the save area is allocated
3802 // in the caller's stack frame, while for N32/64, it is allocated in the
3803 // callee's stack frame.
3804 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) {
3805 unsigned Reg = AddLiveIn(MF, ArgRegs[I], RC);
3806 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3807 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true);
3808 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy());
3809 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3810 MachinePointerInfo(), false, false, 0);
3811 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0);
3812 OutChains.push_back(Store);