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 "MipsMachineFunction.h"
18 #include "MipsTargetMachine.h"
19 #include "MipsTargetObjectFile.h"
20 #include "MipsSubtarget.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/CallingConv.h"
26 #include "InstPrinter/MipsInstPrinter.h"
27 #include "MCTargetDesc/MipsBaseInfo.h"
28 #include "llvm/CodeGen/CallingConvLower.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/SelectionDAGISel.h"
34 #include "llvm/CodeGen/ValueTypes.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
39 // If I is a shifted mask, set the size (Size) and the first bit of the
40 // mask (Pos), and return true.
41 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
42 static bool IsShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
43 if (!isShiftedMask_64(I))
46 Size = CountPopulation_64(I);
47 Pos = CountTrailingZeros_64(I);
51 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
53 case MipsISD::JmpLink: return "MipsISD::JmpLink";
54 case MipsISD::Hi: return "MipsISD::Hi";
55 case MipsISD::Lo: return "MipsISD::Lo";
56 case MipsISD::GPRel: return "MipsISD::GPRel";
57 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
58 case MipsISD::Ret: return "MipsISD::Ret";
59 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
60 case MipsISD::FPCmp: return "MipsISD::FPCmp";
61 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
62 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
63 case MipsISD::FPRound: return "MipsISD::FPRound";
64 case MipsISD::MAdd: return "MipsISD::MAdd";
65 case MipsISD::MAddu: return "MipsISD::MAddu";
66 case MipsISD::MSub: return "MipsISD::MSub";
67 case MipsISD::MSubu: return "MipsISD::MSubu";
68 case MipsISD::DivRem: return "MipsISD::DivRem";
69 case MipsISD::DivRemU: return "MipsISD::DivRemU";
70 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
71 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
72 case MipsISD::Wrapper: return "MipsISD::Wrapper";
73 case MipsISD::DynAlloc: return "MipsISD::DynAlloc";
74 case MipsISD::Sync: return "MipsISD::Sync";
75 case MipsISD::Ext: return "MipsISD::Ext";
76 case MipsISD::Ins: return "MipsISD::Ins";
82 MipsTargetLowering(MipsTargetMachine &TM)
83 : TargetLowering(TM, new MipsTargetObjectFile()),
84 Subtarget(&TM.getSubtarget<MipsSubtarget>()),
85 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()),
86 IsO32(Subtarget->isABI_O32()) {
88 // Mips does not have i1 type, so use i32 for
89 // setcc operations results (slt, sgt, ...).
90 setBooleanContents(ZeroOrOneBooleanContent);
91 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
93 // Set up the register classes
94 addRegisterClass(MVT::i32, Mips::CPURegsRegisterClass);
95 addRegisterClass(MVT::f32, Mips::FGR32RegisterClass);
98 addRegisterClass(MVT::i64, Mips::CPU64RegsRegisterClass);
100 // When dealing with single precision only, use libcalls
101 if (!Subtarget->isSingleFloat()) {
103 addRegisterClass(MVT::f64, Mips::FGR64RegisterClass);
105 addRegisterClass(MVT::f64, Mips::AFGR64RegisterClass);
108 // Load extented operations for i1 types must be promoted
109 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
110 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
111 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
113 // MIPS doesn't have extending float->double load/store
114 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
115 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
117 // Used by legalize types to correctly generate the setcc result.
118 // Without this, every float setcc comes with a AND/OR with the result,
119 // we don't want this, since the fpcmp result goes to a flag register,
120 // which is used implicitly by brcond and select operations.
121 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
123 // Mips Custom Operations
124 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
125 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
126 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
127 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
128 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
129 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
130 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
131 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
132 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
133 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
134 setOperationAction(ISD::SELECT, MVT::f32, Custom);
135 setOperationAction(ISD::SELECT, MVT::f64, Custom);
136 setOperationAction(ISD::SELECT, MVT::i32, Custom);
137 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
138 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
139 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Custom);
140 setOperationAction(ISD::VASTART, MVT::Other, Custom);
142 setOperationAction(ISD::SDIV, MVT::i32, Expand);
143 setOperationAction(ISD::SREM, MVT::i32, Expand);
144 setOperationAction(ISD::UDIV, MVT::i32, Expand);
145 setOperationAction(ISD::UREM, MVT::i32, Expand);
146 setOperationAction(ISD::SDIV, MVT::i64, Expand);
147 setOperationAction(ISD::SREM, MVT::i64, Expand);
148 setOperationAction(ISD::UDIV, MVT::i64, Expand);
149 setOperationAction(ISD::UREM, MVT::i64, Expand);
151 // Operations not directly supported by Mips.
152 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
153 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
154 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
155 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
156 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
157 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
158 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
159 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
160 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
161 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
162 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
163 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
164 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
165 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
166 setOperationAction(ISD::ROTL, MVT::i32, Expand);
167 setOperationAction(ISD::ROTL, MVT::i64, Expand);
169 if (!Subtarget->hasMips32r2())
170 setOperationAction(ISD::ROTR, MVT::i32, Expand);
172 if (!Subtarget->hasMips64r2())
173 setOperationAction(ISD::ROTR, MVT::i64, Expand);
175 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
176 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
177 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
178 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
179 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
180 setOperationAction(ISD::FSIN, MVT::f32, Expand);
181 setOperationAction(ISD::FSIN, MVT::f64, Expand);
182 setOperationAction(ISD::FCOS, MVT::f32, Expand);
183 setOperationAction(ISD::FCOS, MVT::f64, Expand);
184 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
185 setOperationAction(ISD::FPOW, MVT::f32, Expand);
186 setOperationAction(ISD::FPOW, MVT::f64, Expand);
187 setOperationAction(ISD::FLOG, MVT::f32, Expand);
188 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
189 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
190 setOperationAction(ISD::FEXP, MVT::f32, Expand);
191 setOperationAction(ISD::FMA, MVT::f32, Expand);
192 setOperationAction(ISD::FMA, MVT::f64, Expand);
194 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
195 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
197 setOperationAction(ISD::VAARG, MVT::Other, Expand);
198 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
199 setOperationAction(ISD::VAEND, MVT::Other, Expand);
201 // Use the default for now
202 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
203 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
205 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
206 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
208 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
209 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
211 setInsertFencesForAtomic(true);
213 if (Subtarget->isSingleFloat())
214 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
216 if (!Subtarget->hasSEInReg()) {
217 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
218 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
221 if (!Subtarget->hasBitCount())
222 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
224 if (!Subtarget->hasSwap())
225 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
227 setTargetDAGCombine(ISD::ADDE);
228 setTargetDAGCombine(ISD::SUBE);
229 setTargetDAGCombine(ISD::SDIVREM);
230 setTargetDAGCombine(ISD::UDIVREM);
231 setTargetDAGCombine(ISD::SETCC);
232 setTargetDAGCombine(ISD::AND);
233 setTargetDAGCombine(ISD::OR);
235 setMinFunctionAlignment(2);
237 setStackPointerRegisterToSaveRestore(HasMips64 ? Mips::SP_64 : Mips::SP);
238 computeRegisterProperties();
240 setExceptionPointerRegister(Mips::A0);
241 setExceptionSelectorRegister(Mips::A1);
244 bool MipsTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
245 MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
246 return SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16;
249 EVT MipsTargetLowering::getSetCCResultType(EVT VT) const {
254 // Transforms a subgraph in CurDAG if the following pattern is found:
255 // (addc multLo, Lo0), (adde multHi, Hi0),
257 // multHi/Lo: product of multiplication
258 // Lo0: initial value of Lo register
259 // Hi0: initial value of Hi register
260 // Return true if pattern matching was successful.
261 static bool SelectMadd(SDNode* ADDENode, SelectionDAG* CurDAG) {
262 // ADDENode's second operand must be a flag output of an ADDC node in order
263 // for the matching to be successful.
264 SDNode* ADDCNode = ADDENode->getOperand(2).getNode();
266 if (ADDCNode->getOpcode() != ISD::ADDC)
269 SDValue MultHi = ADDENode->getOperand(0);
270 SDValue MultLo = ADDCNode->getOperand(0);
271 SDNode* MultNode = MultHi.getNode();
272 unsigned MultOpc = MultHi.getOpcode();
274 // MultHi and MultLo must be generated by the same node,
275 if (MultLo.getNode() != MultNode)
278 // and it must be a multiplication.
279 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
282 // MultLo amd MultHi must be the first and second output of MultNode
284 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
287 // Transform this to a MADD only if ADDENode and ADDCNode are the only users
288 // of the values of MultNode, in which case MultNode will be removed in later
290 // If there exist users other than ADDENode or ADDCNode, this function returns
291 // here, which will result in MultNode being mapped to a single MULT
292 // instruction node rather than a pair of MULT and MADD instructions being
294 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
297 SDValue Chain = CurDAG->getEntryNode();
298 DebugLoc dl = ADDENode->getDebugLoc();
300 // create MipsMAdd(u) node
301 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MAddu : MipsISD::MAdd;
303 SDValue MAdd = CurDAG->getNode(MultOpc, dl, MVT::Glue,
304 MultNode->getOperand(0),// Factor 0
305 MultNode->getOperand(1),// Factor 1
306 ADDCNode->getOperand(1),// Lo0
307 ADDENode->getOperand(1));// Hi0
309 // create CopyFromReg nodes
310 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
312 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
314 CopyFromLo.getValue(2));
316 // replace uses of adde and addc here
317 if (!SDValue(ADDCNode, 0).use_empty())
318 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDCNode, 0), CopyFromLo);
320 if (!SDValue(ADDENode, 0).use_empty())
321 CurDAG->ReplaceAllUsesOfValueWith(SDValue(ADDENode, 0), CopyFromHi);
327 // Transforms a subgraph in CurDAG if the following pattern is found:
328 // (addc Lo0, multLo), (sube Hi0, multHi),
330 // multHi/Lo: product of multiplication
331 // Lo0: initial value of Lo register
332 // Hi0: initial value of Hi register
333 // Return true if pattern matching was successful.
334 static bool SelectMsub(SDNode* SUBENode, SelectionDAG* CurDAG) {
335 // SUBENode's second operand must be a flag output of an SUBC node in order
336 // for the matching to be successful.
337 SDNode* SUBCNode = SUBENode->getOperand(2).getNode();
339 if (SUBCNode->getOpcode() != ISD::SUBC)
342 SDValue MultHi = SUBENode->getOperand(1);
343 SDValue MultLo = SUBCNode->getOperand(1);
344 SDNode* MultNode = MultHi.getNode();
345 unsigned MultOpc = MultHi.getOpcode();
347 // MultHi and MultLo must be generated by the same node,
348 if (MultLo.getNode() != MultNode)
351 // and it must be a multiplication.
352 if (MultOpc != ISD::SMUL_LOHI && MultOpc != ISD::UMUL_LOHI)
355 // MultLo amd MultHi must be the first and second output of MultNode
357 if (MultHi.getResNo() != 1 || MultLo.getResNo() != 0)
360 // Transform this to a MSUB only if SUBENode and SUBCNode are the only users
361 // of the values of MultNode, in which case MultNode will be removed in later
363 // If there exist users other than SUBENode or SUBCNode, this function returns
364 // here, which will result in MultNode being mapped to a single MULT
365 // instruction node rather than a pair of MULT and MSUB instructions being
367 if (!MultHi.hasOneUse() || !MultLo.hasOneUse())
370 SDValue Chain = CurDAG->getEntryNode();
371 DebugLoc dl = SUBENode->getDebugLoc();
373 // create MipsSub(u) node
374 MultOpc = MultOpc == ISD::UMUL_LOHI ? MipsISD::MSubu : MipsISD::MSub;
376 SDValue MSub = CurDAG->getNode(MultOpc, dl, MVT::Glue,
377 MultNode->getOperand(0),// Factor 0
378 MultNode->getOperand(1),// Factor 1
379 SUBCNode->getOperand(0),// Lo0
380 SUBENode->getOperand(0));// Hi0
382 // create CopyFromReg nodes
383 SDValue CopyFromLo = CurDAG->getCopyFromReg(Chain, dl, Mips::LO, MVT::i32,
385 SDValue CopyFromHi = CurDAG->getCopyFromReg(CopyFromLo.getValue(1), dl,
387 CopyFromLo.getValue(2));
389 // replace uses of sube and subc here
390 if (!SDValue(SUBCNode, 0).use_empty())
391 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBCNode, 0), CopyFromLo);
393 if (!SDValue(SUBENode, 0).use_empty())
394 CurDAG->ReplaceAllUsesOfValueWith(SDValue(SUBENode, 0), CopyFromHi);
399 static SDValue PerformADDECombine(SDNode *N, SelectionDAG& DAG,
400 TargetLowering::DAGCombinerInfo &DCI,
401 const MipsSubtarget* Subtarget) {
402 if (DCI.isBeforeLegalize())
405 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
407 return SDValue(N, 0);
412 static SDValue PerformSUBECombine(SDNode *N, SelectionDAG& DAG,
413 TargetLowering::DAGCombinerInfo &DCI,
414 const MipsSubtarget* Subtarget) {
415 if (DCI.isBeforeLegalize())
418 if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
420 return SDValue(N, 0);
425 static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG& DAG,
426 TargetLowering::DAGCombinerInfo &DCI,
427 const MipsSubtarget* Subtarget) {
428 if (DCI.isBeforeLegalizeOps())
431 EVT Ty = N->getValueType(0);
432 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64;
433 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64;
434 unsigned opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem :
436 DebugLoc dl = N->getDebugLoc();
438 SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue,
439 N->getOperand(0), N->getOperand(1));
440 SDValue InChain = DAG.getEntryNode();
441 SDValue InGlue = DivRem;
444 if (N->hasAnyUseOfValue(0)) {
445 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty,
447 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
448 InChain = CopyFromLo.getValue(1);
449 InGlue = CopyFromLo.getValue(2);
453 if (N->hasAnyUseOfValue(1)) {
454 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl,
456 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
462 static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) {
464 default: llvm_unreachable("Unknown fp condition code!");
466 case ISD::SETOEQ: return Mips::FCOND_OEQ;
467 case ISD::SETUNE: return Mips::FCOND_UNE;
469 case ISD::SETOLT: return Mips::FCOND_OLT;
471 case ISD::SETOGT: return Mips::FCOND_OGT;
473 case ISD::SETOLE: return Mips::FCOND_OLE;
475 case ISD::SETOGE: return Mips::FCOND_OGE;
476 case ISD::SETULT: return Mips::FCOND_ULT;
477 case ISD::SETULE: return Mips::FCOND_ULE;
478 case ISD::SETUGT: return Mips::FCOND_UGT;
479 case ISD::SETUGE: return Mips::FCOND_UGE;
480 case ISD::SETUO: return Mips::FCOND_UN;
481 case ISD::SETO: return Mips::FCOND_OR;
483 case ISD::SETONE: return Mips::FCOND_ONE;
484 case ISD::SETUEQ: return Mips::FCOND_UEQ;
489 // Returns true if condition code has to be inverted.
490 static bool InvertFPCondCode(Mips::CondCode CC) {
491 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
494 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
495 "Illegal Condition Code");
500 // Creates and returns an FPCmp node from a setcc node.
501 // Returns Op if setcc is not a floating point comparison.
502 static SDValue CreateFPCmp(SelectionDAG& DAG, const SDValue& Op) {
503 // must be a SETCC node
504 if (Op.getOpcode() != ISD::SETCC)
507 SDValue LHS = Op.getOperand(0);
509 if (!LHS.getValueType().isFloatingPoint())
512 SDValue RHS = Op.getOperand(1);
513 DebugLoc dl = Op.getDebugLoc();
515 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
516 // node if necessary.
517 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
519 return DAG.getNode(MipsISD::FPCmp, dl, MVT::Glue, LHS, RHS,
520 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32));
523 // Creates and returns a CMovFPT/F node.
524 static SDValue CreateCMovFP(SelectionDAG& DAG, SDValue Cond, SDValue True,
525 SDValue False, DebugLoc DL) {
526 bool invert = InvertFPCondCode((Mips::CondCode)
527 cast<ConstantSDNode>(Cond.getOperand(2))
530 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
531 True.getValueType(), True, False, Cond);
534 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG& DAG,
535 TargetLowering::DAGCombinerInfo &DCI,
536 const MipsSubtarget* Subtarget) {
537 if (DCI.isBeforeLegalizeOps())
540 SDValue Cond = CreateFPCmp(DAG, SDValue(N, 0));
542 if (Cond.getOpcode() != MipsISD::FPCmp)
545 SDValue True = DAG.getConstant(1, MVT::i32);
546 SDValue False = DAG.getConstant(0, MVT::i32);
548 return CreateCMovFP(DAG, Cond, True, False, N->getDebugLoc());
551 static SDValue PerformANDCombine(SDNode *N, SelectionDAG& DAG,
552 TargetLowering::DAGCombinerInfo &DCI,
553 const MipsSubtarget* Subtarget) {
554 // Pattern match EXT.
555 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
556 // => ext $dst, $src, size, pos
557 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
560 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
561 unsigned ShiftRightOpc = ShiftRight.getOpcode();
563 // Op's first operand must be a shift right.
564 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
567 // The second operand of the shift must be an immediate.
569 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
572 uint64_t Pos = CN->getZExtValue();
573 uint64_t SMPos, SMSize;
575 // Op's second operand must be a shifted mask.
576 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
577 !IsShiftedMask(CN->getZExtValue(), SMPos, SMSize))
580 // Return if the shifted mask does not start at bit 0 or the sum of its size
581 // and Pos exceeds the word's size.
582 EVT ValTy = N->getValueType(0);
583 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
586 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy,
587 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32),
588 DAG.getConstant(SMSize, MVT::i32));
591 static SDValue PerformORCombine(SDNode *N, SelectionDAG& DAG,
592 TargetLowering::DAGCombinerInfo &DCI,
593 const MipsSubtarget* Subtarget) {
594 // Pattern match INS.
595 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
596 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
597 // => ins $dst, $src, size, pos, $src1
598 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2())
601 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
602 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
605 // See if Op's first operand matches (and $src1 , mask0).
606 if (And0.getOpcode() != ISD::AND)
609 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
610 !IsShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
613 // See if Op's second operand matches (and (shl $src, pos), mask1).
614 if (And1.getOpcode() != ISD::AND)
617 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
618 !IsShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
621 // The shift masks must have the same position and size.
622 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
625 SDValue Shl = And1.getOperand(0);
626 if (Shl.getOpcode() != ISD::SHL)
629 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
632 unsigned Shamt = CN->getZExtValue();
634 // Return if the shift amount and the first bit position of mask are not the
636 EVT ValTy = N->getValueType(0);
637 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
640 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0),
641 DAG.getConstant(SMPos0, MVT::i32),
642 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0));
645 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
647 SelectionDAG &DAG = DCI.DAG;
648 unsigned opc = N->getOpcode();
653 return PerformADDECombine(N, DAG, DCI, Subtarget);
655 return PerformSUBECombine(N, DAG, DCI, Subtarget);
658 return PerformDivRemCombine(N, DAG, DCI, Subtarget);
660 return PerformSETCCCombine(N, DAG, DCI, Subtarget);
662 return PerformANDCombine(N, DAG, DCI, Subtarget);
664 return PerformORCombine(N, DAG, DCI, Subtarget);
670 SDValue MipsTargetLowering::
671 LowerOperation(SDValue Op, SelectionDAG &DAG) const
673 switch (Op.getOpcode())
675 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
676 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
677 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
678 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
679 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
680 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
681 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
682 case ISD::SELECT: return LowerSELECT(Op, DAG);
683 case ISD::VASTART: return LowerVASTART(Op, DAG);
684 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
685 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
686 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG);
687 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
692 //===----------------------------------------------------------------------===//
693 // Lower helper functions
694 //===----------------------------------------------------------------------===//
696 // AddLiveIn - This helper function adds the specified physical register to the
697 // MachineFunction as a live in value. It also creates a corresponding
698 // virtual register for it.
700 AddLiveIn(MachineFunction &MF, unsigned PReg, TargetRegisterClass *RC)
702 assert(RC->contains(PReg) && "Not the correct regclass!");
703 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
704 MF.getRegInfo().addLiveIn(PReg, VReg);
708 // Get fp branch code (not opcode) from condition code.
709 static Mips::FPBranchCode GetFPBranchCodeFromCond(Mips::CondCode CC) {
710 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
711 return Mips::BRANCH_T;
713 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
714 "Invalid CondCode.");
716 return Mips::BRANCH_F;
720 static MachineBasicBlock* ExpandCondMov(MachineInstr *MI, MachineBasicBlock *BB,
722 const MipsSubtarget* Subtarget,
723 const TargetInstrInfo *TII,
724 bool isFPCmp, unsigned Opc) {
725 // There is no need to expand CMov instructions if target has
726 // conditional moves.
727 if (Subtarget->hasCondMov())
730 // To "insert" a SELECT_CC instruction, we actually have to insert the
731 // diamond control-flow pattern. The incoming instruction knows the
732 // destination vreg to set, the condition code register to branch on, the
733 // true/false values to select between, and a branch opcode to use.
734 const BasicBlock *LLVM_BB = BB->getBasicBlock();
735 MachineFunction::iterator It = BB;
742 // bNE r1, r0, copy1MBB
743 // fallthrough --> copy0MBB
744 MachineBasicBlock *thisMBB = BB;
745 MachineFunction *F = BB->getParent();
746 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
747 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
748 F->insert(It, copy0MBB);
749 F->insert(It, sinkMBB);
751 // Transfer the remainder of BB and its successor edges to sinkMBB.
752 sinkMBB->splice(sinkMBB->begin(), BB,
753 llvm::next(MachineBasicBlock::iterator(MI)),
755 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
757 // Next, add the true and fallthrough blocks as its successors.
758 BB->addSuccessor(copy0MBB);
759 BB->addSuccessor(sinkMBB);
761 // Emit the right instruction according to the type of the operands compared
763 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
765 BuildMI(BB, dl, TII->get(Opc)).addReg(MI->getOperand(2).getReg())
766 .addReg(Mips::ZERO).addMBB(sinkMBB);
770 // # fallthrough to sinkMBB
773 // Update machine-CFG edges
774 BB->addSuccessor(sinkMBB);
777 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
782 BuildMI(*BB, BB->begin(), dl,
783 TII->get(Mips::PHI), MI->getOperand(0).getReg())
784 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
785 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
787 BuildMI(*BB, BB->begin(), dl,
788 TII->get(Mips::PHI), MI->getOperand(0).getReg())
789 .addReg(MI->getOperand(3).getReg()).addMBB(thisMBB)
790 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB);
792 MI->eraseFromParent(); // The pseudo instruction is gone now.
797 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
798 MachineBasicBlock *BB) const {
799 switch (MI->getOpcode()) {
801 assert(false && "Unexpected instr type to insert");
803 case Mips::ATOMIC_LOAD_ADD_I8:
804 case Mips::ATOMIC_LOAD_ADD_I8_P8:
805 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
806 case Mips::ATOMIC_LOAD_ADD_I16:
807 case Mips::ATOMIC_LOAD_ADD_I16_P8:
808 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
809 case Mips::ATOMIC_LOAD_ADD_I32:
810 case Mips::ATOMIC_LOAD_ADD_I32_P8:
811 return EmitAtomicBinary(MI, BB, 4, Mips::ADDu);
812 case Mips::ATOMIC_LOAD_ADD_I64:
813 case Mips::ATOMIC_LOAD_ADD_I64_P8:
814 return EmitAtomicBinary(MI, BB, 8, Mips::DADDu);
816 case Mips::ATOMIC_LOAD_AND_I8:
817 case Mips::ATOMIC_LOAD_AND_I8_P8:
818 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
819 case Mips::ATOMIC_LOAD_AND_I16:
820 case Mips::ATOMIC_LOAD_AND_I16_P8:
821 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
822 case Mips::ATOMIC_LOAD_AND_I32:
823 case Mips::ATOMIC_LOAD_AND_I32_P8:
824 return EmitAtomicBinary(MI, BB, 4, Mips::AND);
825 case Mips::ATOMIC_LOAD_AND_I64:
826 case Mips::ATOMIC_LOAD_AND_I64_P8:
827 return EmitAtomicBinary(MI, BB, 8, Mips::AND64);
829 case Mips::ATOMIC_LOAD_OR_I8:
830 case Mips::ATOMIC_LOAD_OR_I8_P8:
831 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
832 case Mips::ATOMIC_LOAD_OR_I16:
833 case Mips::ATOMIC_LOAD_OR_I16_P8:
834 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
835 case Mips::ATOMIC_LOAD_OR_I32:
836 case Mips::ATOMIC_LOAD_OR_I32_P8:
837 return EmitAtomicBinary(MI, BB, 4, Mips::OR);
838 case Mips::ATOMIC_LOAD_OR_I64:
839 case Mips::ATOMIC_LOAD_OR_I64_P8:
840 return EmitAtomicBinary(MI, BB, 8, Mips::OR64);
842 case Mips::ATOMIC_LOAD_XOR_I8:
843 case Mips::ATOMIC_LOAD_XOR_I8_P8:
844 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
845 case Mips::ATOMIC_LOAD_XOR_I16:
846 case Mips::ATOMIC_LOAD_XOR_I16_P8:
847 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
848 case Mips::ATOMIC_LOAD_XOR_I32:
849 case Mips::ATOMIC_LOAD_XOR_I32_P8:
850 return EmitAtomicBinary(MI, BB, 4, Mips::XOR);
851 case Mips::ATOMIC_LOAD_XOR_I64:
852 case Mips::ATOMIC_LOAD_XOR_I64_P8:
853 return EmitAtomicBinary(MI, BB, 8, Mips::XOR64);
855 case Mips::ATOMIC_LOAD_NAND_I8:
856 case Mips::ATOMIC_LOAD_NAND_I8_P8:
857 return EmitAtomicBinaryPartword(MI, BB, 1, 0, true);
858 case Mips::ATOMIC_LOAD_NAND_I16:
859 case Mips::ATOMIC_LOAD_NAND_I16_P8:
860 return EmitAtomicBinaryPartword(MI, BB, 2, 0, true);
861 case Mips::ATOMIC_LOAD_NAND_I32:
862 case Mips::ATOMIC_LOAD_NAND_I32_P8:
863 return EmitAtomicBinary(MI, BB, 4, 0, true);
864 case Mips::ATOMIC_LOAD_NAND_I64:
865 case Mips::ATOMIC_LOAD_NAND_I64_P8:
866 return EmitAtomicBinary(MI, BB, 8, 0, true);
868 case Mips::ATOMIC_LOAD_SUB_I8:
869 case Mips::ATOMIC_LOAD_SUB_I8_P8:
870 return EmitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
871 case Mips::ATOMIC_LOAD_SUB_I16:
872 case Mips::ATOMIC_LOAD_SUB_I16_P8:
873 return EmitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
874 case Mips::ATOMIC_LOAD_SUB_I32:
875 case Mips::ATOMIC_LOAD_SUB_I32_P8:
876 return EmitAtomicBinary(MI, BB, 4, Mips::SUBu);
877 case Mips::ATOMIC_LOAD_SUB_I64:
878 case Mips::ATOMIC_LOAD_SUB_I64_P8:
879 return EmitAtomicBinary(MI, BB, 8, Mips::DSUBu);
881 case Mips::ATOMIC_SWAP_I8:
882 case Mips::ATOMIC_SWAP_I8_P8:
883 return EmitAtomicBinaryPartword(MI, BB, 1, 0);
884 case Mips::ATOMIC_SWAP_I16:
885 case Mips::ATOMIC_SWAP_I16_P8:
886 return EmitAtomicBinaryPartword(MI, BB, 2, 0);
887 case Mips::ATOMIC_SWAP_I32:
888 case Mips::ATOMIC_SWAP_I32_P8:
889 return EmitAtomicBinary(MI, BB, 4, 0);
890 case Mips::ATOMIC_SWAP_I64:
891 case Mips::ATOMIC_SWAP_I64_P8:
892 return EmitAtomicBinary(MI, BB, 8, 0);
894 case Mips::ATOMIC_CMP_SWAP_I8:
895 case Mips::ATOMIC_CMP_SWAP_I8_P8:
896 return EmitAtomicCmpSwapPartword(MI, BB, 1);
897 case Mips::ATOMIC_CMP_SWAP_I16:
898 case Mips::ATOMIC_CMP_SWAP_I16_P8:
899 return EmitAtomicCmpSwapPartword(MI, BB, 2);
900 case Mips::ATOMIC_CMP_SWAP_I32:
901 case Mips::ATOMIC_CMP_SWAP_I32_P8:
902 return EmitAtomicCmpSwap(MI, BB, 4);
903 case Mips::ATOMIC_CMP_SWAP_I64:
904 case Mips::ATOMIC_CMP_SWAP_I64_P8:
905 return EmitAtomicCmpSwap(MI, BB, 8);
909 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
910 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
912 MipsTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
913 unsigned Size, unsigned BinOpcode,
915 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
917 MachineFunction *MF = BB->getParent();
918 MachineRegisterInfo &RegInfo = MF->getRegInfo();
919 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
920 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
921 DebugLoc dl = MI->getDebugLoc();
922 unsigned LL, SC, AND, NOR, ZERO, BEQ;
925 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
926 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
933 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
934 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
937 ZERO = Mips::ZERO_64;
941 unsigned OldVal = MI->getOperand(0).getReg();
942 unsigned Ptr = MI->getOperand(1).getReg();
943 unsigned Incr = MI->getOperand(2).getReg();
945 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
946 unsigned AndRes = RegInfo.createVirtualRegister(RC);
947 unsigned Success = RegInfo.createVirtualRegister(RC);
949 // insert new blocks after the current block
950 const BasicBlock *LLVM_BB = BB->getBasicBlock();
951 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
952 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
953 MachineFunction::iterator It = BB;
955 MF->insert(It, loopMBB);
956 MF->insert(It, exitMBB);
958 // Transfer the remainder of BB and its successor edges to exitMBB.
959 exitMBB->splice(exitMBB->begin(), BB,
960 llvm::next(MachineBasicBlock::iterator(MI)),
962 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
966 // fallthrough --> loopMBB
967 BB->addSuccessor(loopMBB);
968 loopMBB->addSuccessor(loopMBB);
969 loopMBB->addSuccessor(exitMBB);
973 // <binop> storeval, oldval, incr
974 // sc success, storeval, 0(ptr)
975 // beq success, $0, loopMBB
977 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
979 // and andres, oldval, incr
980 // nor storeval, $0, andres
981 BuildMI(BB, dl, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
982 BuildMI(BB, dl, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
983 } else if (BinOpcode) {
984 // <binop> storeval, oldval, incr
985 BuildMI(BB, dl, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
989 BuildMI(BB, dl, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
990 BuildMI(BB, dl, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
992 MI->eraseFromParent(); // The instruction is gone now.
998 MipsTargetLowering::EmitAtomicBinaryPartword(MachineInstr *MI,
999 MachineBasicBlock *BB,
1000 unsigned Size, unsigned BinOpcode,
1002 assert((Size == 1 || Size == 2) &&
1003 "Unsupported size for EmitAtomicBinaryPartial.");
1005 MachineFunction *MF = BB->getParent();
1006 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1007 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1008 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1009 DebugLoc dl = MI->getDebugLoc();
1010 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1011 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1013 unsigned Dest = MI->getOperand(0).getReg();
1014 unsigned Ptr = MI->getOperand(1).getReg();
1015 unsigned Incr = MI->getOperand(2).getReg();
1017 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1018 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1019 unsigned Mask = RegInfo.createVirtualRegister(RC);
1020 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1021 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1022 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1023 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1024 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1025 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1026 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1027 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1028 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1029 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1030 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1031 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1032 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1033 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1034 unsigned Success = RegInfo.createVirtualRegister(RC);
1036 // insert new blocks after the current block
1037 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1038 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1039 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1040 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1041 MachineFunction::iterator It = BB;
1043 MF->insert(It, loopMBB);
1044 MF->insert(It, sinkMBB);
1045 MF->insert(It, exitMBB);
1047 // Transfer the remainder of BB and its successor edges to exitMBB.
1048 exitMBB->splice(exitMBB->begin(), BB,
1049 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1050 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1052 BB->addSuccessor(loopMBB);
1053 loopMBB->addSuccessor(loopMBB);
1054 loopMBB->addSuccessor(sinkMBB);
1055 sinkMBB->addSuccessor(exitMBB);
1058 // addiu masklsb2,$0,-4 # 0xfffffffc
1059 // and alignedaddr,ptr,masklsb2
1060 // andi ptrlsb2,ptr,3
1061 // sll shiftamt,ptrlsb2,3
1062 // ori maskupper,$0,255 # 0xff
1063 // sll mask,maskupper,shiftamt
1064 // nor mask2,$0,mask
1065 // sll incr2,incr,shiftamt
1067 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1068 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1069 .addReg(Mips::ZERO).addImm(-4);
1070 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1071 .addReg(Ptr).addReg(MaskLSB2);
1072 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1073 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1074 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1075 .addReg(Mips::ZERO).addImm(MaskImm);
1076 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1077 .addReg(ShiftAmt).addReg(MaskUpper);
1078 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1079 BuildMI(BB, dl, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr);
1081 // atomic.load.binop
1083 // ll oldval,0(alignedaddr)
1084 // binop binopres,oldval,incr2
1085 // and newval,binopres,mask
1086 // and maskedoldval0,oldval,mask2
1087 // or storeval,maskedoldval0,newval
1088 // sc success,storeval,0(alignedaddr)
1089 // beq success,$0,loopMBB
1093 // ll oldval,0(alignedaddr)
1094 // and newval,incr2,mask
1095 // and maskedoldval0,oldval,mask2
1096 // or storeval,maskedoldval0,newval
1097 // sc success,storeval,0(alignedaddr)
1098 // beq success,$0,loopMBB
1101 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1103 // and andres, oldval, incr2
1104 // nor binopres, $0, andres
1105 // and newval, binopres, mask
1106 BuildMI(BB, dl, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1107 BuildMI(BB, dl, TII->get(Mips::NOR), BinOpRes)
1108 .addReg(Mips::ZERO).addReg(AndRes);
1109 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1110 } else if (BinOpcode) {
1111 // <binop> binopres, oldval, incr2
1112 // and newval, binopres, mask
1113 BuildMI(BB, dl, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1114 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1115 } else {// atomic.swap
1116 // and newval, incr2, mask
1117 BuildMI(BB, dl, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1120 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1121 .addReg(OldVal).addReg(Mask2);
1122 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1123 .addReg(MaskedOldVal0).addReg(NewVal);
1124 BuildMI(BB, dl, TII->get(SC), Success)
1125 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1126 BuildMI(BB, dl, TII->get(Mips::BEQ))
1127 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1130 // and maskedoldval1,oldval,mask
1131 // srl srlres,maskedoldval1,shiftamt
1132 // sll sllres,srlres,24
1133 // sra dest,sllres,24
1135 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1137 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1138 .addReg(OldVal).addReg(Mask);
1139 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1140 .addReg(ShiftAmt).addReg(MaskedOldVal1);
1141 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1142 .addReg(SrlRes).addImm(ShiftImm);
1143 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1144 .addReg(SllRes).addImm(ShiftImm);
1146 MI->eraseFromParent(); // The instruction is gone now.
1152 MipsTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
1153 MachineBasicBlock *BB,
1154 unsigned Size) const {
1155 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1157 MachineFunction *MF = BB->getParent();
1158 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1159 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1160 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1161 DebugLoc dl = MI->getDebugLoc();
1162 unsigned LL, SC, ZERO, BNE, BEQ;
1165 LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1166 SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1172 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD;
1173 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD;
1174 ZERO = Mips::ZERO_64;
1179 unsigned Dest = MI->getOperand(0).getReg();
1180 unsigned Ptr = MI->getOperand(1).getReg();
1181 unsigned OldVal = MI->getOperand(2).getReg();
1182 unsigned NewVal = MI->getOperand(3).getReg();
1184 unsigned Success = RegInfo.createVirtualRegister(RC);
1186 // insert new blocks after the current block
1187 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1188 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1189 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1190 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1191 MachineFunction::iterator It = BB;
1193 MF->insert(It, loop1MBB);
1194 MF->insert(It, loop2MBB);
1195 MF->insert(It, exitMBB);
1197 // Transfer the remainder of BB and its successor edges to exitMBB.
1198 exitMBB->splice(exitMBB->begin(), BB,
1199 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1200 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1204 // fallthrough --> loop1MBB
1205 BB->addSuccessor(loop1MBB);
1206 loop1MBB->addSuccessor(exitMBB);
1207 loop1MBB->addSuccessor(loop2MBB);
1208 loop2MBB->addSuccessor(loop1MBB);
1209 loop2MBB->addSuccessor(exitMBB);
1213 // bne dest, oldval, exitMBB
1215 BuildMI(BB, dl, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1216 BuildMI(BB, dl, TII->get(BNE))
1217 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1220 // sc success, newval, 0(ptr)
1221 // beq success, $0, loop1MBB
1223 BuildMI(BB, dl, TII->get(SC), Success)
1224 .addReg(NewVal).addReg(Ptr).addImm(0);
1225 BuildMI(BB, dl, TII->get(BEQ))
1226 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1228 MI->eraseFromParent(); // The instruction is gone now.
1234 MipsTargetLowering::EmitAtomicCmpSwapPartword(MachineInstr *MI,
1235 MachineBasicBlock *BB,
1236 unsigned Size) const {
1237 assert((Size == 1 || Size == 2) &&
1238 "Unsupported size for EmitAtomicCmpSwapPartial.");
1240 MachineFunction *MF = BB->getParent();
1241 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1242 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1243 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1244 DebugLoc dl = MI->getDebugLoc();
1245 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL;
1246 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC;
1248 unsigned Dest = MI->getOperand(0).getReg();
1249 unsigned Ptr = MI->getOperand(1).getReg();
1250 unsigned CmpVal = MI->getOperand(2).getReg();
1251 unsigned NewVal = MI->getOperand(3).getReg();
1253 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1254 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1255 unsigned Mask = RegInfo.createVirtualRegister(RC);
1256 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1257 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1258 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1259 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1260 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1261 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1262 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1263 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1264 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1265 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1266 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1267 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1268 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1269 unsigned SllRes = RegInfo.createVirtualRegister(RC);
1270 unsigned Success = RegInfo.createVirtualRegister(RC);
1272 // insert new blocks after the current block
1273 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1274 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1275 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1276 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1277 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1278 MachineFunction::iterator It = BB;
1280 MF->insert(It, loop1MBB);
1281 MF->insert(It, loop2MBB);
1282 MF->insert(It, sinkMBB);
1283 MF->insert(It, exitMBB);
1285 // Transfer the remainder of BB and its successor edges to exitMBB.
1286 exitMBB->splice(exitMBB->begin(), BB,
1287 llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
1288 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1290 BB->addSuccessor(loop1MBB);
1291 loop1MBB->addSuccessor(sinkMBB);
1292 loop1MBB->addSuccessor(loop2MBB);
1293 loop2MBB->addSuccessor(loop1MBB);
1294 loop2MBB->addSuccessor(sinkMBB);
1295 sinkMBB->addSuccessor(exitMBB);
1297 // FIXME: computation of newval2 can be moved to loop2MBB.
1299 // addiu masklsb2,$0,-4 # 0xfffffffc
1300 // and alignedaddr,ptr,masklsb2
1301 // andi ptrlsb2,ptr,3
1302 // sll shiftamt,ptrlsb2,3
1303 // ori maskupper,$0,255 # 0xff
1304 // sll mask,maskupper,shiftamt
1305 // nor mask2,$0,mask
1306 // andi maskedcmpval,cmpval,255
1307 // sll shiftedcmpval,maskedcmpval,shiftamt
1308 // andi maskednewval,newval,255
1309 // sll shiftednewval,maskednewval,shiftamt
1310 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1311 BuildMI(BB, dl, TII->get(Mips::ADDiu), MaskLSB2)
1312 .addReg(Mips::ZERO).addImm(-4);
1313 BuildMI(BB, dl, TII->get(Mips::AND), AlignedAddr)
1314 .addReg(Ptr).addReg(MaskLSB2);
1315 BuildMI(BB, dl, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1316 BuildMI(BB, dl, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1317 BuildMI(BB, dl, TII->get(Mips::ORi), MaskUpper)
1318 .addReg(Mips::ZERO).addImm(MaskImm);
1319 BuildMI(BB, dl, TII->get(Mips::SLLV), Mask)
1320 .addReg(ShiftAmt).addReg(MaskUpper);
1321 BuildMI(BB, dl, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1322 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedCmpVal)
1323 .addReg(CmpVal).addImm(MaskImm);
1324 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedCmpVal)
1325 .addReg(ShiftAmt).addReg(MaskedCmpVal);
1326 BuildMI(BB, dl, TII->get(Mips::ANDi), MaskedNewVal)
1327 .addReg(NewVal).addImm(MaskImm);
1328 BuildMI(BB, dl, TII->get(Mips::SLLV), ShiftedNewVal)
1329 .addReg(ShiftAmt).addReg(MaskedNewVal);
1332 // ll oldval,0(alginedaddr)
1333 // and maskedoldval0,oldval,mask
1334 // bne maskedoldval0,shiftedcmpval,sinkMBB
1336 BuildMI(BB, dl, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1337 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal0)
1338 .addReg(OldVal).addReg(Mask);
1339 BuildMI(BB, dl, TII->get(Mips::BNE))
1340 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1343 // and maskedoldval1,oldval,mask2
1344 // or storeval,maskedoldval1,shiftednewval
1345 // sc success,storeval,0(alignedaddr)
1346 // beq success,$0,loop1MBB
1348 BuildMI(BB, dl, TII->get(Mips::AND), MaskedOldVal1)
1349 .addReg(OldVal).addReg(Mask2);
1350 BuildMI(BB, dl, TII->get(Mips::OR), StoreVal)
1351 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1352 BuildMI(BB, dl, TII->get(SC), Success)
1353 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1354 BuildMI(BB, dl, TII->get(Mips::BEQ))
1355 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1358 // srl srlres,maskedoldval0,shiftamt
1359 // sll sllres,srlres,24
1360 // sra dest,sllres,24
1362 int64_t ShiftImm = (Size == 1) ? 24 : 16;
1364 BuildMI(BB, dl, TII->get(Mips::SRLV), SrlRes)
1365 .addReg(ShiftAmt).addReg(MaskedOldVal0);
1366 BuildMI(BB, dl, TII->get(Mips::SLL), SllRes)
1367 .addReg(SrlRes).addImm(ShiftImm);
1368 BuildMI(BB, dl, TII->get(Mips::SRA), Dest)
1369 .addReg(SllRes).addImm(ShiftImm);
1371 MI->eraseFromParent(); // The instruction is gone now.
1376 //===----------------------------------------------------------------------===//
1377 // Misc Lower Operation implementation
1378 //===----------------------------------------------------------------------===//
1379 SDValue MipsTargetLowering::
1380 LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const
1382 MachineFunction &MF = DAG.getMachineFunction();
1383 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
1384 unsigned SP = IsN64 ? Mips::SP_64 : Mips::SP;
1386 assert(getTargetMachine().getFrameLowering()->getStackAlignment() >=
1387 cast<ConstantSDNode>(Op.getOperand(2).getNode())->getZExtValue() &&
1388 "Cannot lower if the alignment of the allocated space is larger than \
1389 that of the stack.");
1391 SDValue Chain = Op.getOperand(0);
1392 SDValue Size = Op.getOperand(1);
1393 DebugLoc dl = Op.getDebugLoc();
1395 // Get a reference from Mips stack pointer
1396 SDValue StackPointer = DAG.getCopyFromReg(Chain, dl, SP, getPointerTy());
1398 // Subtract the dynamic size from the actual stack size to
1399 // obtain the new stack size.
1400 SDValue Sub = DAG.getNode(ISD::SUB, dl, getPointerTy(), StackPointer, Size);
1402 // The Sub result contains the new stack start address, so it
1403 // must be placed in the stack pointer register.
1404 Chain = DAG.getCopyToReg(StackPointer.getValue(1), dl, SP, Sub, SDValue());
1406 // This node always has two return values: a new stack pointer
1407 // value and a chain
1408 SDVTList VTLs = DAG.getVTList(getPointerTy(), MVT::Other);
1409 SDValue Ptr = DAG.getFrameIndex(MipsFI->getDynAllocFI(), getPointerTy());
1410 SDValue Ops[] = { Chain, Ptr, Chain.getValue(1) };
1412 return DAG.getNode(MipsISD::DynAlloc, dl, VTLs, Ops, 3);
1415 SDValue MipsTargetLowering::
1416 LowerBRCOND(SDValue Op, SelectionDAG &DAG) const
1418 // The first operand is the chain, the second is the condition, the third is
1419 // the block to branch to if the condition is true.
1420 SDValue Chain = Op.getOperand(0);
1421 SDValue Dest = Op.getOperand(2);
1422 DebugLoc dl = Op.getDebugLoc();
1424 SDValue CondRes = CreateFPCmp(DAG, Op.getOperand(1));
1426 // Return if flag is not set by a floating point comparison.
1427 if (CondRes.getOpcode() != MipsISD::FPCmp)
1430 SDValue CCNode = CondRes.getOperand(2);
1432 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1433 SDValue BrCode = DAG.getConstant(GetFPBranchCodeFromCond(CC), MVT::i32);
1435 return DAG.getNode(MipsISD::FPBrcond, dl, Op.getValueType(), Chain, BrCode,
1439 SDValue MipsTargetLowering::
1440 LowerSELECT(SDValue Op, SelectionDAG &DAG) const
1442 SDValue Cond = CreateFPCmp(DAG, Op.getOperand(0));
1444 // Return if flag is not set by a floating point comparison.
1445 if (Cond.getOpcode() != MipsISD::FPCmp)
1448 return CreateCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1452 SDValue MipsTargetLowering::LowerGlobalAddress(SDValue Op,
1453 SelectionDAG &DAG) const {
1454 // FIXME there isn't actually debug info here
1455 DebugLoc dl = Op.getDebugLoc();
1456 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1458 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1459 SDVTList VTs = DAG.getVTList(MVT::i32);
1461 MipsTargetObjectFile &TLOF = (MipsTargetObjectFile&)getObjFileLowering();
1463 // %gp_rel relocation
1464 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) {
1465 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1467 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, dl, VTs, &GA, 1);
1468 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1469 return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode);
1471 // %hi/%lo relocation
1472 SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1474 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0,
1476 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, VTs, &GAHi, 1);
1477 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, GALo);
1478 return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1481 EVT ValTy = Op.getValueType();
1482 bool HasGotOfst = (GV->hasInternalLinkage() ||
1483 (GV->hasLocalLinkage() && !isa<Function>(GV)));
1484 unsigned GotFlag = IsN64 ?
1485 (HasGotOfst ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT_DISP) :
1486 (HasGotOfst ? MipsII::MO_GOT : MipsII::MO_GOT16);
1487 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0, GotFlag);
1488 GA = DAG.getNode(MipsISD::Wrapper, dl, ValTy, GA);
1489 SDValue ResNode = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), GA,
1490 MachinePointerInfo(), false, false, false, 0);
1491 // On functions and global targets not internal linked only
1492 // a load from got/GP is necessary for PIC to work.
1495 SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0,
1496 IsN64 ? MipsII::MO_GOT_OFST :
1498 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, GALo);
1499 return DAG.getNode(ISD::ADD, dl, ValTy, ResNode, Lo);
1502 SDValue MipsTargetLowering::LowerBlockAddress(SDValue Op,
1503 SelectionDAG &DAG) const {
1504 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1505 // FIXME there isn't actually debug info here
1506 DebugLoc dl = Op.getDebugLoc();
1508 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) {
1509 // %hi/%lo relocation
1510 SDValue BAHi = DAG.getBlockAddress(BA, MVT::i32, true, MipsII::MO_ABS_HI);
1511 SDValue BALo = DAG.getBlockAddress(BA, MVT::i32, true, MipsII::MO_ABS_LO);
1512 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, MVT::i32, BAHi);
1513 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, BALo);
1514 return DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, Lo);
1517 EVT ValTy = Op.getValueType();
1518 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1519 unsigned OFSTFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1520 SDValue BAGOTOffset = DAG.getBlockAddress(BA, ValTy, true, GOTFlag);
1521 BAGOTOffset = DAG.getNode(MipsISD::Wrapper, dl, ValTy, BAGOTOffset);
1522 SDValue BALOOffset = DAG.getBlockAddress(BA, ValTy, true, OFSTFlag);
1523 SDValue Load = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), BAGOTOffset,
1524 MachinePointerInfo(), false, false, false, 0);
1525 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, BALOOffset);
1526 return DAG.getNode(ISD::ADD, dl, ValTy, Load, Lo);
1529 SDValue MipsTargetLowering::
1530 LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1532 // If the relocation model is PIC, use the General Dynamic TLS Model or
1533 // Local Dynamic TLS model, otherwise use the Initial Exec or
1534 // Local Exec TLS Model.
1536 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1537 DebugLoc dl = GA->getDebugLoc();
1538 const GlobalValue *GV = GA->getGlobal();
1539 EVT PtrVT = getPointerTy();
1541 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1542 // General Dynamic TLS Model
1543 bool LocalDynamic = GV->hasInternalLinkage();
1544 unsigned Flag = LocalDynamic ? MipsII::MO_TLSLDM :MipsII::MO_TLSGD;
1545 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, Flag);
1546 SDValue Argument = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, TGA);
1547 unsigned PtrSize = PtrVT.getSizeInBits();
1548 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1550 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1554 Entry.Node = Argument;
1556 Args.push_back(Entry);
1558 std::pair<SDValue, SDValue> CallResult =
1559 LowerCallTo(DAG.getEntryNode(), PtrTy,
1560 false, false, false, false, 0, CallingConv::C, false, true,
1561 TlsGetAddr, Args, DAG, dl);
1563 SDValue Ret = CallResult.first;
1568 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1569 MipsII::MO_DTPREL_HI);
1570 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1571 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1572 MipsII::MO_DTPREL_LO);
1573 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1574 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Ret);
1575 return DAG.getNode(ISD::ADD, dl, PtrVT, Add, Lo);
1579 if (GV->isDeclaration()) {
1580 // Initial Exec TLS Model
1581 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1582 MipsII::MO_GOTTPREL);
1583 TGA = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, TGA);
1584 Offset = DAG.getLoad(PtrVT, dl,
1585 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1586 false, false, false, 0);
1588 // Local Exec TLS Model
1589 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1590 MipsII::MO_TPREL_HI);
1591 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1592 MipsII::MO_TPREL_LO);
1593 SDValue Hi = DAG.getNode(MipsISD::Hi, dl, PtrVT, TGAHi);
1594 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, TGALo);
1595 Offset = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1598 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, dl, PtrVT);
1599 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1602 SDValue MipsTargetLowering::
1603 LowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1605 SDValue HiPart, JTI, JTILo;
1606 // FIXME there isn't actually debug info here
1607 DebugLoc dl = Op.getDebugLoc();
1608 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1609 EVT PtrVT = Op.getValueType();
1610 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1612 if (!IsPIC && !IsN64) {
1613 JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MipsII::MO_ABS_HI);
1614 HiPart = DAG.getNode(MipsISD::Hi, dl, PtrVT, JTI);
1615 JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MipsII::MO_ABS_LO);
1616 } else {// Emit Load from Global Pointer
1617 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1618 unsigned OfstFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1619 JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, GOTFlag);
1620 JTI = DAG.getNode(MipsISD::Wrapper, dl, PtrVT, JTI);
1621 HiPart = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), JTI,
1622 MachinePointerInfo(), false, false, false, 0);
1623 JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OfstFlag);
1626 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, PtrVT, JTILo);
1627 return DAG.getNode(ISD::ADD, dl, PtrVT, HiPart, Lo);
1630 SDValue MipsTargetLowering::
1631 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1634 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1635 const Constant *C = N->getConstVal();
1636 // FIXME there isn't actually debug info here
1637 DebugLoc dl = Op.getDebugLoc();
1639 // gp_rel relocation
1640 // FIXME: we should reference the constant pool using small data sections,
1641 // but the asm printer currently doesn't support this feature without
1642 // hacking it. This feature should come soon so we can uncomment the
1644 //if (IsInSmallSection(C->getType())) {
1645 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP);
1646 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32);
1647 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode);
1649 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
1650 SDValue CPHi = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1651 N->getOffset(), MipsII::MO_ABS_HI);
1652 SDValue CPLo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1653 N->getOffset(), MipsII::MO_ABS_LO);
1654 SDValue HiPart = DAG.getNode(MipsISD::Hi, dl, MVT::i32, CPHi);
1655 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, MVT::i32, CPLo);
1656 ResNode = DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);
1658 EVT ValTy = Op.getValueType();
1659 unsigned GOTFlag = IsN64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT;
1660 unsigned OFSTFlag = IsN64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO;
1661 SDValue CP = DAG.getTargetConstantPool(C, ValTy, N->getAlignment(),
1662 N->getOffset(), GOTFlag);
1663 CP = DAG.getNode(MipsISD::Wrapper, dl, ValTy, CP);
1664 SDValue Load = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), CP,
1665 MachinePointerInfo::getConstantPool(), false,
1667 SDValue CPLo = DAG.getTargetConstantPool(C, ValTy, N->getAlignment(),
1668 N->getOffset(), OFSTFlag);
1669 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, ValTy, CPLo);
1670 ResNode = DAG.getNode(ISD::ADD, dl, ValTy, Load, Lo);
1676 SDValue MipsTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1677 MachineFunction &MF = DAG.getMachineFunction();
1678 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1680 DebugLoc dl = Op.getDebugLoc();
1681 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1684 // vastart just stores the address of the VarArgsFrameIndex slot into the
1685 // memory location argument.
1686 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1687 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
1688 MachinePointerInfo(SV), false, false, 0);
1691 // Called if the size of integer registers is large enough to hold the whole
1692 // floating point number.
1693 static SDValue LowerFCOPYSIGNLargeIntReg(SDValue Op, SelectionDAG &DAG) {
1694 // FIXME: Use ext/ins instructions if target architecture is Mips32r2.
1695 EVT ValTy = Op.getValueType();
1696 EVT IntValTy = MVT::getIntegerVT(ValTy.getSizeInBits());
1697 uint64_t Mask = (uint64_t)1 << (ValTy.getSizeInBits() - 1);
1698 DebugLoc dl = Op.getDebugLoc();
1699 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntValTy, Op.getOperand(0));
1700 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntValTy, Op.getOperand(1));
1701 SDValue And0 = DAG.getNode(ISD::AND, dl, IntValTy, Op0,
1702 DAG.getConstant(Mask - 1, IntValTy));
1703 SDValue And1 = DAG.getNode(ISD::AND, dl, IntValTy, Op1,
1704 DAG.getConstant(Mask, IntValTy));
1705 SDValue Result = DAG.getNode(ISD::OR, dl, IntValTy, And0, And1);
1706 return DAG.getNode(ISD::BITCAST, dl, ValTy, Result);
1709 // Called if the size of integer registers is not large enough to hold the whole
1710 // floating point number (e.g. f64 & 32-bit integer register).
1712 LowerFCOPYSIGNSmallIntReg(SDValue Op, SelectionDAG &DAG, bool isLittle) {
1714 // Use ext/ins instructions if target architecture is Mips32r2.
1715 // Eliminate redundant mfc1 and mtc1 instructions.
1716 unsigned LoIdx = 0, HiIdx = 1;
1719 std::swap(LoIdx, HiIdx);
1721 DebugLoc dl = Op.getDebugLoc();
1722 SDValue Word0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1724 DAG.getConstant(LoIdx, MVT::i32));
1725 SDValue Hi0 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1726 Op.getOperand(0), DAG.getConstant(HiIdx, MVT::i32));
1727 SDValue Hi1 = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
1728 Op.getOperand(1), DAG.getConstant(HiIdx, MVT::i32));
1729 SDValue And0 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi0,
1730 DAG.getConstant(0x7fffffff, MVT::i32));
1731 SDValue And1 = DAG.getNode(ISD::AND, dl, MVT::i32, Hi1,
1732 DAG.getConstant(0x80000000, MVT::i32));
1733 SDValue Word1 = DAG.getNode(ISD::OR, dl, MVT::i32, And0, And1);
1736 std::swap(Word0, Word1);
1738 return DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64, Word0, Word1);
1742 MipsTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
1743 EVT Ty = Op.getValueType();
1745 assert(Ty == MVT::f32 || Ty == MVT::f64);
1747 if (Ty == MVT::f32 || HasMips64)
1748 return LowerFCOPYSIGNLargeIntReg(Op, DAG);
1750 return LowerFCOPYSIGNSmallIntReg(Op, DAG, Subtarget->isLittle());
1753 SDValue MipsTargetLowering::
1754 LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1756 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
1757 "Frame address can only be determined for current frame.");
1759 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1760 MFI->setFrameAddressIsTaken(true);
1761 EVT VT = Op.getValueType();
1762 DebugLoc dl = Op.getDebugLoc();
1763 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
1764 IsN64 ? Mips::FP_64 : Mips::FP, VT);
1768 // TODO: set SType according to the desired memory barrier behavior.
1770 MipsTargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG& DAG) const {
1772 DebugLoc dl = Op.getDebugLoc();
1773 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1774 DAG.getConstant(SType, MVT::i32));
1777 SDValue MipsTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1778 SelectionDAG& DAG) const {
1779 // FIXME: Need pseudo-fence for 'singlethread' fences
1780 // FIXME: Set SType for weaker fences where supported/appropriate.
1782 DebugLoc dl = Op.getDebugLoc();
1783 return DAG.getNode(MipsISD::Sync, dl, MVT::Other, Op.getOperand(0),
1784 DAG.getConstant(SType, MVT::i32));
1787 //===----------------------------------------------------------------------===//
1788 // Calling Convention Implementation
1789 //===----------------------------------------------------------------------===//
1791 //===----------------------------------------------------------------------===//
1792 // TODO: Implement a generic logic using tblgen that can support this.
1793 // Mips O32 ABI rules:
1795 // i32 - Passed in A0, A1, A2, A3 and stack
1796 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
1797 // an argument. Otherwise, passed in A1, A2, A3 and stack.
1798 // f64 - Only passed in two aliased f32 registers if no int reg has been used
1799 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
1800 // not used, it must be shadowed. If only A3 is avaiable, shadow it and
1803 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
1804 //===----------------------------------------------------------------------===//
1806 static bool CC_MipsO32(unsigned ValNo, MVT ValVT,
1807 MVT LocVT, CCValAssign::LocInfo LocInfo,
1808 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1810 static const unsigned IntRegsSize=4, FloatRegsSize=2;
1812 static const unsigned IntRegs[] = {
1813 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1815 static const unsigned F32Regs[] = {
1816 Mips::F12, Mips::F14
1818 static const unsigned F64Regs[] = {
1823 if (ArgFlags.isByVal()) {
1824 State.HandleByVal(ValNo, ValVT, LocVT, LocInfo,
1825 1 /*MinSize*/, 4 /*MinAlign*/, ArgFlags);
1826 unsigned NextReg = (State.getNextStackOffset() + 3) / 4;
1827 for (unsigned r = State.getFirstUnallocated(IntRegs, IntRegsSize);
1828 r < std::min(IntRegsSize, NextReg); ++r)
1829 State.AllocateReg(IntRegs[r]);
1833 // Promote i8 and i16
1834 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
1836 if (ArgFlags.isSExt())
1837 LocInfo = CCValAssign::SExt;
1838 else if (ArgFlags.isZExt())
1839 LocInfo = CCValAssign::ZExt;
1841 LocInfo = CCValAssign::AExt;
1846 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
1847 // is true: function is vararg, argument is 3rd or higher, there is previous
1848 // argument which is not f32 or f64.
1849 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1
1850 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo;
1851 unsigned OrigAlign = ArgFlags.getOrigAlign();
1852 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
1854 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
1855 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1856 // If this is the first part of an i64 arg,
1857 // the allocated register must be either A0 or A2.
1858 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
1859 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1861 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
1862 // Allocate int register and shadow next int register. If first
1863 // available register is Mips::A1 or Mips::A3, shadow it too.
1864 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1865 if (Reg == Mips::A1 || Reg == Mips::A3)
1866 Reg = State.AllocateReg(IntRegs, IntRegsSize);
1867 State.AllocateReg(IntRegs, IntRegsSize);
1869 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
1870 // we are guaranteed to find an available float register
1871 if (ValVT == MVT::f32) {
1872 Reg = State.AllocateReg(F32Regs, FloatRegsSize);
1873 // Shadow int register
1874 State.AllocateReg(IntRegs, IntRegsSize);
1876 Reg = State.AllocateReg(F64Regs, FloatRegsSize);
1877 // Shadow int registers
1878 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize);
1879 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
1880 State.AllocateReg(IntRegs, IntRegsSize);
1881 State.AllocateReg(IntRegs, IntRegsSize);
1884 llvm_unreachable("Cannot handle this ValVT.");
1886 unsigned SizeInBytes = ValVT.getSizeInBits() >> 3;
1887 unsigned Offset = State.AllocateStack(SizeInBytes, OrigAlign);
1890 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1892 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1894 return false; // CC must always match
1897 static const unsigned Mips64IntRegs[8] =
1898 {Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64,
1899 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64};
1900 static const unsigned Mips64DPRegs[8] =
1901 {Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
1902 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64};
1904 static bool CC_Mips64Byval(unsigned ValNo, MVT ValVT, MVT LocVT,
1905 CCValAssign::LocInfo LocInfo,
1906 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1907 unsigned Align = std::max(ArgFlags.getByValAlign(), (unsigned)8);
1908 unsigned Size = (ArgFlags.getByValSize() + 7) / 8 * 8;
1909 unsigned FirstIdx = State.getFirstUnallocated(Mips64IntRegs, 8);
1911 assert(Align <= 16 && "Cannot handle alignments larger than 16.");
1913 // If byval is 16-byte aligned, the first arg register must be even.
1914 if ((Align == 16) && (FirstIdx % 2)) {
1915 State.AllocateReg(Mips64IntRegs[FirstIdx], Mips64DPRegs[FirstIdx]);
1919 // Mark the registers allocated.
1920 for (unsigned I = FirstIdx; Size && (I < 8); Size -= 8, ++I)
1921 State.AllocateReg(Mips64IntRegs[I], Mips64DPRegs[I]);
1923 // Allocate space on caller's stack.
1924 unsigned Offset = State.AllocateStack(Size, Align);
1927 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Mips64IntRegs[FirstIdx],
1930 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
1935 #include "MipsGenCallingConv.inc"
1938 AnalyzeMips64CallOperands(CCState CCInfo,
1939 const SmallVectorImpl<ISD::OutputArg> &Outs) {
1940 unsigned NumOps = Outs.size();
1941 for (unsigned i = 0; i != NumOps; ++i) {
1942 MVT ArgVT = Outs[i].VT;
1943 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
1946 if (Outs[i].IsFixed)
1947 R = CC_MipsN(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
1949 R = CC_MipsN_VarArg(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
1953 dbgs() << "Call operand #" << i << " has unhandled type "
1954 << EVT(ArgVT).getEVTString();
1956 llvm_unreachable(0);
1961 //===----------------------------------------------------------------------===//
1962 // Call Calling Convention Implementation
1963 //===----------------------------------------------------------------------===//
1965 static const unsigned O32IntRegsSize = 4;
1967 static const unsigned O32IntRegs[] = {
1968 Mips::A0, Mips::A1, Mips::A2, Mips::A3
1971 // Return next O32 integer argument register.
1972 static unsigned getNextIntArgReg(unsigned Reg) {
1973 assert((Reg == Mips::A0) || (Reg == Mips::A2));
1974 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
1977 // Write ByVal Arg to arg registers and stack.
1979 WriteByValArg(SDValue& ByValChain, SDValue Chain, DebugLoc dl,
1980 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
1981 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
1982 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
1983 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
1984 MVT PtrType, bool isLittle) {
1985 unsigned LocMemOffset = VA.getLocMemOffset();
1986 unsigned Offset = 0;
1987 uint32_t RemainingSize = Flags.getByValSize();
1988 unsigned ByValAlign = Flags.getByValAlign();
1990 // Copy the first 4 words of byval arg to registers A0 - A3.
1991 // FIXME: Use a stricter alignment if it enables better optimization in passes
1993 for (; RemainingSize >= 4 && LocMemOffset < 4 * 4;
1994 Offset += 4, RemainingSize -= 4, LocMemOffset += 4) {
1995 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
1996 DAG.getConstant(Offset, MVT::i32));
1997 SDValue LoadVal = DAG.getLoad(MVT::i32, dl, Chain, LoadPtr,
1998 MachinePointerInfo(), false, false, false,
1999 std::min(ByValAlign, (unsigned )4));
2000 MemOpChains.push_back(LoadVal.getValue(1));
2001 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
2002 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
2005 if (RemainingSize == 0)
2008 // If there still is a register available for argument passing, write the
2009 // remaining part of the structure to it using subword loads and shifts.
2010 if (LocMemOffset < 4 * 4) {
2011 assert(RemainingSize <= 3 && RemainingSize >= 1 &&
2012 "There must be one to three bytes remaining.");
2013 unsigned LoadSize = (RemainingSize == 3 ? 2 : RemainingSize);
2014 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2015 DAG.getConstant(Offset, MVT::i32));
2016 unsigned Alignment = std::min(ByValAlign, (unsigned )4);
2017 SDValue LoadVal = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
2018 LoadPtr, MachinePointerInfo(),
2019 MVT::getIntegerVT(LoadSize * 8), false,
2021 MemOpChains.push_back(LoadVal.getValue(1));
2023 // If target is big endian, shift it to the most significant half-word or
2026 LoadVal = DAG.getNode(ISD::SHL, dl, MVT::i32, LoadVal,
2027 DAG.getConstant(32 - LoadSize * 8, MVT::i32));
2030 RemainingSize -= LoadSize;
2032 // Read second subword if necessary.
2033 if (RemainingSize != 0) {
2034 assert(RemainingSize == 1 && "There must be one byte remaining.");
2035 LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2036 DAG.getConstant(Offset, MVT::i32));
2037 unsigned Alignment = std::min(ByValAlign, (unsigned )2);
2038 SDValue Subword = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
2039 LoadPtr, MachinePointerInfo(),
2040 MVT::i8, false, false, Alignment);
2041 MemOpChains.push_back(Subword.getValue(1));
2042 // Insert the loaded byte to LoadVal.
2043 // FIXME: Use INS if supported by target.
2044 unsigned ShiftAmt = isLittle ? 16 : 8;
2045 SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i32, Subword,
2046 DAG.getConstant(ShiftAmt, MVT::i32));
2047 LoadVal = DAG.getNode(ISD::OR, dl, MVT::i32, LoadVal, Shift);
2050 unsigned DstReg = O32IntRegs[LocMemOffset / 4];
2051 RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
2055 // Create a fixed object on stack at offset LocMemOffset and copy
2056 // remaining part of byval arg to it using memcpy.
2057 SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
2058 DAG.getConstant(Offset, MVT::i32));
2059 LastFI = MFI->CreateFixedObject(RemainingSize, LocMemOffset, true);
2060 SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
2061 ByValChain = DAG.getMemcpy(ByValChain, dl, Dst, Src,
2062 DAG.getConstant(RemainingSize, MVT::i32),
2063 std::min(ByValAlign, (unsigned)4),
2064 /*isVolatile=*/false, /*AlwaysInline=*/false,
2065 MachinePointerInfo(0), MachinePointerInfo(0));
2068 // Copy Mips64 byVal arg to registers and stack.
2070 PassByValArg64(SDValue& ByValChain, SDValue Chain, DebugLoc dl,
2071 SmallVector<std::pair<unsigned, SDValue>, 16>& RegsToPass,
2072 SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
2073 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
2074 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
2075 EVT PtrTy, bool isLittle) {
2076 unsigned ByValSize = Flags.getByValSize();
2077 unsigned Alignment = std::min(Flags.getByValAlign(), (unsigned)8);
2078 bool IsRegLoc = VA.isRegLoc();
2079 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct.
2080 unsigned LocMemOffset = 0;
2081 unsigned MemCpySize = ByValSize;
2084 LocMemOffset = VA.getLocMemOffset();
2086 const unsigned *Reg = std::find(Mips64IntRegs, Mips64IntRegs + 8,
2088 const unsigned *RegEnd = Mips64IntRegs + 8;
2090 // Copy double words to registers.
2091 for (; (Reg != RegEnd) && (ByValSize >= Offset + 8); ++Reg, Offset += 8) {
2092 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2093 DAG.getConstant(Offset, PtrTy));
2094 SDValue LoadVal = DAG.getLoad(MVT::i64, dl, Chain, LoadPtr,
2095 MachinePointerInfo(), false, false, false,
2097 MemOpChains.push_back(LoadVal.getValue(1));
2098 RegsToPass.push_back(std::make_pair(*Reg, LoadVal));
2101 // Return if the struct has been fully copied.
2102 if (!(MemCpySize = ByValSize - Offset))
2105 // If there is an argument register available, copy the remainder of the
2106 // byval argument with sub-doubleword loads and shifts.
2107 if (Reg != RegEnd) {
2108 assert((ByValSize < Offset + 8) &&
2109 "Size of the remainder should be smaller than 8-byte.");
2111 for (unsigned LoadSize = 4; Offset < ByValSize; LoadSize /= 2) {
2112 unsigned RemSize = ByValSize - Offset;
2114 if (RemSize < LoadSize)
2117 SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2118 DAG.getConstant(Offset, PtrTy));
2120 DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i64, Chain, LoadPtr,
2121 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8),
2122 false, false, Alignment);
2123 MemOpChains.push_back(LoadVal.getValue(1));
2125 // Offset in number of bits from double word boundary.
2126 unsigned OffsetDW = (Offset % 8) * 8;
2127 unsigned Shamt = isLittle ? OffsetDW : 64 - (OffsetDW + LoadSize * 8);
2128 SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i64, LoadVal,
2129 DAG.getConstant(Shamt, MVT::i32));
2131 Val = Val.getNode() ? DAG.getNode(ISD::OR, dl, MVT::i64, Val, Shift) :
2134 Alignment = std::min(Alignment, LoadSize);
2137 RegsToPass.push_back(std::make_pair(*Reg, Val));
2142 assert(MemCpySize && "MemCpySize must not be zero.");
2144 // Create a fixed object on stack at offset LocMemOffset and copy
2145 // remainder of byval arg to it with memcpy.
2146 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrTy, Arg,
2147 DAG.getConstant(Offset, PtrTy));
2148 LastFI = MFI->CreateFixedObject(MemCpySize, LocMemOffset, true);
2149 SDValue Dst = DAG.getFrameIndex(LastFI, PtrTy);
2150 ByValChain = DAG.getMemcpy(ByValChain, dl, Dst, Src,
2151 DAG.getConstant(MemCpySize, PtrTy), Alignment,
2152 /*isVolatile=*/false, /*AlwaysInline=*/false,
2153 MachinePointerInfo(0), MachinePointerInfo(0));
2156 /// LowerCall - functions arguments are copied from virtual regs to
2157 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2158 /// TODO: isTailCall.
2160 MipsTargetLowering::LowerCall(SDValue InChain, SDValue Callee,
2161 CallingConv::ID CallConv, bool isVarArg,
2163 const SmallVectorImpl<ISD::OutputArg> &Outs,
2164 const SmallVectorImpl<SDValue> &OutVals,
2165 const SmallVectorImpl<ISD::InputArg> &Ins,
2166 DebugLoc dl, SelectionDAG &DAG,
2167 SmallVectorImpl<SDValue> &InVals) const {
2168 // MIPs target does not yet support tail call optimization.
2171 MachineFunction &MF = DAG.getMachineFunction();
2172 MachineFrameInfo *MFI = MF.getFrameInfo();
2173 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering();
2174 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2175 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2177 // Analyze operands of the call, assigning locations to each operand.
2178 SmallVector<CCValAssign, 16> ArgLocs;
2179 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2180 getTargetMachine(), ArgLocs, *DAG.getContext());
2183 CCInfo.AnalyzeCallOperands(Outs, CC_MipsO32);
2185 AnalyzeMips64CallOperands(CCInfo, Outs);
2187 CCInfo.AnalyzeCallOperands(Outs, CC_Mips);
2189 // Get a count of how many bytes are to be pushed on the stack.
2190 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2192 // Chain is the output chain of the last Load/Store or CopyToReg node.
2193 // ByValChain is the output chain of the last Memcpy node created for copying
2194 // byval arguments to the stack.
2195 SDValue Chain, CallSeqStart, ByValChain;
2196 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true);
2197 Chain = CallSeqStart = DAG.getCALLSEQ_START(InChain, NextStackOffsetVal);
2198 ByValChain = InChain;
2200 // If this is the first call, create a stack frame object that points to
2201 // a location to which .cprestore saves $gp.
2202 if (IsO32 && IsPIC && !MipsFI->getGPFI())
2203 MipsFI->setGPFI(MFI->CreateFixedObject(4, 0, true));
2205 // Get the frame index of the stack frame object that points to the location
2206 // of dynamically allocated area on the stack.
2207 int DynAllocFI = MipsFI->getDynAllocFI();
2209 // Update size of the maximum argument space.
2210 // For O32, a minimum of four words (16 bytes) of argument space is
2213 NextStackOffset = std::max(NextStackOffset, (unsigned)16);
2215 unsigned MaxCallFrameSize = MipsFI->getMaxCallFrameSize();
2217 if (MaxCallFrameSize < NextStackOffset) {
2218 MipsFI->setMaxCallFrameSize(NextStackOffset);
2220 // Set the offsets relative to $sp of the $gp restore slot and dynamically
2221 // allocated stack space. These offsets must be aligned to a boundary
2222 // determined by the stack alignment of the ABI.
2223 unsigned StackAlignment = TFL->getStackAlignment();
2224 NextStackOffset = (NextStackOffset + StackAlignment - 1) /
2225 StackAlignment * StackAlignment;
2227 if (MipsFI->needGPSaveRestore())
2228 MFI->setObjectOffset(MipsFI->getGPFI(), NextStackOffset);
2230 MFI->setObjectOffset(DynAllocFI, NextStackOffset);
2233 // With EABI is it possible to have 16 args on registers.
2234 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
2235 SmallVector<SDValue, 8> MemOpChains;
2237 int FirstFI = -MFI->getNumFixedObjects() - 1, LastFI = 0;
2239 // Walk the register/memloc assignments, inserting copies/loads.
2240 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2241 SDValue Arg = OutVals[i];
2242 CCValAssign &VA = ArgLocs[i];
2243 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2244 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2247 if (Flags.isByVal()) {
2248 assert(Flags.getByValSize() &&
2249 "ByVal args of size 0 should have been ignored by front-end.");
2251 WriteByValArg(ByValChain, Chain, dl, RegsToPass, MemOpChains, LastFI,
2252 MFI, DAG, Arg, VA, Flags, getPointerTy(),
2253 Subtarget->isLittle());
2255 PassByValArg64(ByValChain, Chain, dl, RegsToPass, MemOpChains, LastFI,
2256 MFI, DAG, Arg, VA, Flags, getPointerTy(),
2257 Subtarget->isLittle());
2261 // Promote the value if needed.
2262 switch (VA.getLocInfo()) {
2263 default: llvm_unreachable("Unknown loc info!");
2264 case CCValAssign::Full:
2265 if (VA.isRegLoc()) {
2266 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2267 (ValVT == MVT::f64 && LocVT == MVT::i64))
2268 Arg = DAG.getNode(ISD::BITCAST, dl, LocVT, Arg);
2269 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2270 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2271 Arg, DAG.getConstant(0, MVT::i32));
2272 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, dl, MVT::i32,
2273 Arg, DAG.getConstant(1, MVT::i32));
2274 if (!Subtarget->isLittle())
2276 unsigned LocRegLo = VA.getLocReg();
2277 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2278 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2279 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2284 case CCValAssign::SExt:
2285 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, LocVT, Arg);
2287 case CCValAssign::ZExt:
2288 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, LocVT, Arg);
2290 case CCValAssign::AExt:
2291 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, LocVT, Arg);
2295 // Arguments that can be passed on register must be kept at
2296 // RegsToPass vector
2297 if (VA.isRegLoc()) {
2298 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2302 // Register can't get to this point...
2303 assert(VA.isMemLoc());
2305 // Create the frame index object for this incoming parameter
2306 LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2307 VA.getLocMemOffset(), true);
2308 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2310 // emit ISD::STORE whichs stores the
2311 // parameter value to a stack Location
2312 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2313 MachinePointerInfo(), false, false, 0));
2316 // Extend range of indices of frame objects for outgoing arguments that were
2317 // created during this function call. Skip this step if no such objects were
2320 MipsFI->extendOutArgFIRange(FirstFI, LastFI);
2322 // If a memcpy has been created to copy a byval arg to a stack, replace the
2323 // chain input of CallSeqStart with ByValChain.
2324 if (InChain != ByValChain)
2325 DAG.UpdateNodeOperands(CallSeqStart.getNode(), ByValChain,
2326 NextStackOffsetVal);
2328 // Transform all store nodes into one single node because all store
2329 // nodes are independent of each other.
2330 if (!MemOpChains.empty())
2331 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2332 &MemOpChains[0], MemOpChains.size());
2334 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2335 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2336 // node so that legalize doesn't hack it.
2337 unsigned char OpFlag;
2338 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25
2339 bool GlobalOrExternal = false;
2342 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2343 if (IsPICCall && G->getGlobal()->hasInternalLinkage()) {
2344 OpFlag = IsO32 ? MipsII::MO_GOT : MipsII::MO_GOT_PAGE;
2345 unsigned char LoFlag = IsO32 ? MipsII::MO_ABS_LO : MipsII::MO_GOT_OFST;
2346 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(), 0,
2348 CalleeLo = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(),
2351 OpFlag = IsPICCall ? MipsII::MO_GOT_CALL : MipsII::MO_NO_FLAG;
2352 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2353 getPointerTy(), 0, OpFlag);
2356 GlobalOrExternal = true;
2358 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2359 if (IsN64 || (!IsO32 && IsPIC))
2360 OpFlag = MipsII::MO_GOT_DISP;
2361 else if (!IsPIC) // !N64 && static
2362 OpFlag = MipsII::MO_NO_FLAG;
2364 OpFlag = MipsII::MO_GOT_CALL;
2365 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2367 GlobalOrExternal = true;
2372 // Create nodes that load address of callee and copy it to T9
2374 if (GlobalOrExternal) {
2375 // Load callee address
2376 Callee = DAG.getNode(MipsISD::Wrapper, dl, getPointerTy(), Callee);
2377 SDValue LoadValue = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
2378 Callee, MachinePointerInfo::getGOT(),
2379 false, false, false, 0);
2381 // Use GOT+LO if callee has internal linkage.
2382 if (CalleeLo.getNode()) {
2383 SDValue Lo = DAG.getNode(MipsISD::Lo, dl, getPointerTy(), CalleeLo);
2384 Callee = DAG.getNode(ISD::ADD, dl, getPointerTy(), LoadValue, Lo);
2390 // T9 should contain the address of the callee function if
2391 // -reloction-model=pic or it is an indirect call.
2392 if (IsPICCall || !GlobalOrExternal) {
2394 unsigned T9Reg = IsN64 ? Mips::T9_64 : Mips::T9;
2395 Chain = DAG.getCopyToReg(Chain, dl, T9Reg, Callee, SDValue(0, 0));
2396 InFlag = Chain.getValue(1);
2397 Callee = DAG.getRegister(T9Reg, getPointerTy());
2400 // Build a sequence of copy-to-reg nodes chained together with token
2401 // chain and flag operands which copy the outgoing args into registers.
2402 // The InFlag in necessary since all emitted instructions must be
2404 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2405 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2406 RegsToPass[i].second, InFlag);
2407 InFlag = Chain.getValue(1);
2410 // MipsJmpLink = #chain, #target_address, #opt_in_flags...
2411 // = Chain, Callee, Reg#1, Reg#2, ...
2413 // Returns a chain & a flag for retval copy to use.
2414 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2415 SmallVector<SDValue, 8> Ops;
2416 Ops.push_back(Chain);
2417 Ops.push_back(Callee);
2419 // Add argument registers to the end of the list so that they are
2420 // known live into the call.
2421 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2422 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2423 RegsToPass[i].second.getValueType()));
2425 if (InFlag.getNode())
2426 Ops.push_back(InFlag);
2428 Chain = DAG.getNode(MipsISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size());
2429 InFlag = Chain.getValue(1);
2431 // Create the CALLSEQ_END node.
2432 Chain = DAG.getCALLSEQ_END(Chain,
2433 DAG.getIntPtrConstant(NextStackOffset, true),
2434 DAG.getIntPtrConstant(0, true), InFlag);
2435 InFlag = Chain.getValue(1);
2437 // Handle result values, copying them out of physregs into vregs that we
2439 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2440 Ins, dl, DAG, InVals);
2443 /// LowerCallResult - Lower the result values of a call into the
2444 /// appropriate copies out of appropriate physical registers.
2446 MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2447 CallingConv::ID CallConv, bool isVarArg,
2448 const SmallVectorImpl<ISD::InputArg> &Ins,
2449 DebugLoc dl, SelectionDAG &DAG,
2450 SmallVectorImpl<SDValue> &InVals) const {
2451 // Assign locations to each value returned by this call.
2452 SmallVector<CCValAssign, 16> RVLocs;
2453 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2454 getTargetMachine(), RVLocs, *DAG.getContext());
2456 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips);
2458 // Copy all of the result registers out of their specified physreg.
2459 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2460 Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(),
2461 RVLocs[i].getValVT(), InFlag).getValue(1);
2462 InFlag = Chain.getValue(2);
2463 InVals.push_back(Chain.getValue(0));
2469 //===----------------------------------------------------------------------===//
2470 // Formal Arguments Calling Convention Implementation
2471 //===----------------------------------------------------------------------===//
2472 static void ReadByValArg(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2473 std::vector<SDValue>& OutChains,
2474 SelectionDAG &DAG, unsigned NumWords, SDValue FIN,
2475 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags) {
2476 unsigned LocMem = VA.getLocMemOffset();
2477 unsigned FirstWord = LocMem / 4;
2479 // copy register A0 - A3 to frame object
2480 for (unsigned i = 0; i < NumWords; ++i) {
2481 unsigned CurWord = FirstWord + i;
2482 if (CurWord >= O32IntRegsSize)
2485 unsigned SrcReg = O32IntRegs[CurWord];
2486 unsigned Reg = AddLiveIn(MF, SrcReg, Mips::CPURegsRegisterClass);
2487 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, MVT::i32, FIN,
2488 DAG.getConstant(i * 4, MVT::i32));
2489 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(Reg, MVT::i32),
2490 StorePtr, MachinePointerInfo(), false,
2492 OutChains.push_back(Store);
2496 // Create frame object on stack and copy registers used for byval passing to it.
2498 CopyMips64ByValRegs(MachineFunction &MF, SDValue Chain, DebugLoc dl,
2499 std::vector<SDValue>& OutChains, SelectionDAG &DAG,
2500 const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
2501 MachineFrameInfo *MFI, bool IsRegLoc,
2502 SmallVectorImpl<SDValue> &InVals, MipsFunctionInfo *MipsFI,
2504 const unsigned *Reg = Mips64IntRegs + 8;
2505 int FOOffset; // Frame object offset from virtual frame pointer.
2508 Reg = std::find(Mips64IntRegs, Mips64IntRegs + 8, VA.getLocReg());
2509 FOOffset = (Reg - Mips64IntRegs) * 8 - 8 * 8;
2512 FOOffset = VA.getLocMemOffset();
2514 // Create frame object.
2515 unsigned NumRegs = (Flags.getByValSize() + 7) / 8;
2516 unsigned LastFI = MFI->CreateFixedObject(NumRegs * 8, FOOffset, true);
2517 SDValue FIN = DAG.getFrameIndex(LastFI, PtrTy);
2518 InVals.push_back(FIN);
2520 // Copy arg registers.
2521 for (unsigned I = 0; (Reg != Mips64IntRegs + 8) && (I < NumRegs);
2523 unsigned VReg = AddLiveIn(MF, *Reg, Mips::CPU64RegsRegisterClass);
2524 SDValue StorePtr = DAG.getNode(ISD::ADD, dl, PtrTy, FIN,
2525 DAG.getConstant(I * 8, PtrTy));
2526 SDValue Store = DAG.getStore(Chain, dl, DAG.getRegister(VReg, MVT::i64),
2527 StorePtr, MachinePointerInfo(), false,
2529 OutChains.push_back(Store);
2535 /// LowerFormalArguments - transform physical registers into virtual registers
2536 /// and generate load operations for arguments places on the stack.
2538 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2539 CallingConv::ID CallConv,
2541 const SmallVectorImpl<ISD::InputArg> &Ins,
2542 DebugLoc dl, SelectionDAG &DAG,
2543 SmallVectorImpl<SDValue> &InVals)
2545 MachineFunction &MF = DAG.getMachineFunction();
2546 MachineFrameInfo *MFI = MF.getFrameInfo();
2547 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2549 MipsFI->setVarArgsFrameIndex(0);
2551 // Used with vargs to acumulate store chains.
2552 std::vector<SDValue> OutChains;
2554 // Assign locations to all of the incoming arguments.
2555 SmallVector<CCValAssign, 16> ArgLocs;
2556 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2557 getTargetMachine(), ArgLocs, *DAG.getContext());
2560 CCInfo.AnalyzeFormalArguments(Ins, CC_MipsO32);
2562 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips);
2564 int LastFI = 0;// MipsFI->LastInArgFI is 0 at the entry of this function.
2566 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2567 CCValAssign &VA = ArgLocs[i];
2568 EVT ValVT = VA.getValVT();
2569 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2570 bool IsRegLoc = VA.isRegLoc();
2572 if (Flags.isByVal()) {
2573 assert(Flags.getByValSize() &&
2574 "ByVal args of size 0 should have been ignored by front-end.");
2576 unsigned NumWords = (Flags.getByValSize() + 3) / 4;
2577 LastFI = MFI->CreateFixedObject(NumWords * 4, VA.getLocMemOffset(),
2579 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2580 InVals.push_back(FIN);
2581 ReadByValArg(MF, Chain, dl, OutChains, DAG, NumWords, FIN, VA, Flags);
2583 LastFI = CopyMips64ByValRegs(MF, Chain, dl, OutChains, DAG, VA, Flags,
2584 MFI, IsRegLoc, InVals, MipsFI,
2589 // Arguments stored on registers
2591 EVT RegVT = VA.getLocVT();
2592 unsigned ArgReg = VA.getLocReg();
2593 TargetRegisterClass *RC = 0;
2595 if (RegVT == MVT::i32)
2596 RC = Mips::CPURegsRegisterClass;
2597 else if (RegVT == MVT::i64)
2598 RC = Mips::CPU64RegsRegisterClass;
2599 else if (RegVT == MVT::f32)
2600 RC = Mips::FGR32RegisterClass;
2601 else if (RegVT == MVT::f64)
2602 RC = HasMips64 ? Mips::FGR64RegisterClass : Mips::AFGR64RegisterClass;
2604 llvm_unreachable("RegVT not supported by FormalArguments Lowering");
2606 // Transform the arguments stored on
2607 // physical registers into virtual ones
2608 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgReg, RC);
2609 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2611 // If this is an 8 or 16-bit value, it has been passed promoted
2612 // to 32 bits. Insert an assert[sz]ext to capture this, then
2613 // truncate to the right size.
2614 if (VA.getLocInfo() != CCValAssign::Full) {
2615 unsigned Opcode = 0;
2616 if (VA.getLocInfo() == CCValAssign::SExt)
2617 Opcode = ISD::AssertSext;
2618 else if (VA.getLocInfo() == CCValAssign::ZExt)
2619 Opcode = ISD::AssertZext;
2621 ArgValue = DAG.getNode(Opcode, dl, RegVT, ArgValue,
2622 DAG.getValueType(ValVT));
2623 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue);
2626 // Handle floating point arguments passed in integer registers.
2627 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
2628 (RegVT == MVT::i64 && ValVT == MVT::f64))
2629 ArgValue = DAG.getNode(ISD::BITCAST, dl, ValVT, ArgValue);
2630 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) {
2631 unsigned Reg2 = AddLiveIn(DAG.getMachineFunction(),
2632 getNextIntArgReg(ArgReg), RC);
2633 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl, Reg2, RegVT);
2634 if (!Subtarget->isLittle())
2635 std::swap(ArgValue, ArgValue2);
2636 ArgValue = DAG.getNode(MipsISD::BuildPairF64, dl, MVT::f64,
2637 ArgValue, ArgValue2);
2640 InVals.push_back(ArgValue);
2641 } else { // VA.isRegLoc()
2644 assert(VA.isMemLoc());
2646 // The stack pointer offset is relative to the caller stack frame.
2647 LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2648 VA.getLocMemOffset(), true);
2650 // Create load nodes to retrieve arguments from the stack
2651 SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy());
2652 InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN,
2653 MachinePointerInfo::getFixedStack(LastFI),
2654 false, false, false, 0));
2658 // The mips ABIs for returning structs by value requires that we copy
2659 // the sret argument into $v0 for the return. Save the argument into
2660 // a virtual register so that we can access it from the return points.
2661 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2662 unsigned Reg = MipsFI->getSRetReturnReg();
2664 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i32));
2665 MipsFI->setSRetReturnReg(Reg);
2667 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2668 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2672 unsigned NumOfRegs = IsO32 ? 4 : 8;
2673 const unsigned *ArgRegs = IsO32 ? O32IntRegs : Mips64IntRegs;
2674 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumOfRegs);
2675 int FirstRegSlotOffset = IsO32 ? 0 : -64 ; // offset of $a0's slot.
2676 TargetRegisterClass *RC
2677 = IsO32 ? Mips::CPURegsRegisterClass : Mips::CPU64RegsRegisterClass;
2678 unsigned RegSize = RC->getSize();
2679 int RegSlotOffset = FirstRegSlotOffset + Idx * RegSize;
2681 // Offset of the first variable argument from stack pointer.
2682 int FirstVaArgOffset;
2684 if (IsO32 || (Idx == NumOfRegs)) {
2686 (CCInfo.getNextStackOffset() + RegSize - 1) / RegSize * RegSize;
2688 FirstVaArgOffset = RegSlotOffset;
2690 // Record the frame index of the first variable argument
2691 // which is a value necessary to VASTART.
2692 LastFI = MFI->CreateFixedObject(RegSize, FirstVaArgOffset, true);
2693 MipsFI->setVarArgsFrameIndex(LastFI);
2695 // Copy the integer registers that have not been used for argument passing
2696 // to the argument register save area. For O32, the save area is allocated
2697 // in the caller's stack frame, while for N32/64, it is allocated in the
2698 // callee's stack frame.
2699 for (int StackOffset = RegSlotOffset;
2700 Idx < NumOfRegs; ++Idx, StackOffset += RegSize) {
2701 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), ArgRegs[Idx], RC);
2702 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg,
2703 MVT::getIntegerVT(RegSize * 8));
2704 LastFI = MFI->CreateFixedObject(RegSize, StackOffset, true);
2705 SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy());
2706 OutChains.push_back(DAG.getStore(Chain, dl, ArgValue, PtrOff,
2707 MachinePointerInfo(), false, false, 0));
2711 MipsFI->setLastInArgFI(LastFI);
2713 // All stores are grouped in one node to allow the matching between
2714 // the size of Ins and InVals. This only happens when on varg functions
2715 if (!OutChains.empty()) {
2716 OutChains.push_back(Chain);
2717 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2718 &OutChains[0], OutChains.size());
2724 //===----------------------------------------------------------------------===//
2725 // Return Value Calling Convention Implementation
2726 //===----------------------------------------------------------------------===//
2729 MipsTargetLowering::LowerReturn(SDValue Chain,
2730 CallingConv::ID CallConv, bool isVarArg,
2731 const SmallVectorImpl<ISD::OutputArg> &Outs,
2732 const SmallVectorImpl<SDValue> &OutVals,
2733 DebugLoc dl, SelectionDAG &DAG) const {
2735 // CCValAssign - represent the assignment of
2736 // the return value to a location
2737 SmallVector<CCValAssign, 16> RVLocs;
2739 // CCState - Info about the registers and stack slot.
2740 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2741 getTargetMachine(), RVLocs, *DAG.getContext());
2743 // Analize return values.
2744 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
2746 // If this is the first return lowered for this function, add
2747 // the regs to the liveout set for the function.
2748 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2749 for (unsigned i = 0; i != RVLocs.size(); ++i)
2750 if (RVLocs[i].isRegLoc())
2751 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2756 // Copy the result values into the output registers.
2757 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2758 CCValAssign &VA = RVLocs[i];
2759 assert(VA.isRegLoc() && "Can only return in registers!");
2761 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
2763 // guarantee that all emitted copies are
2764 // stuck together, avoiding something bad
2765 Flag = Chain.getValue(1);
2768 // The mips ABIs for returning structs by value requires that we copy
2769 // the sret argument into $v0 for the return. We saved the argument into
2770 // a virtual register in the entry block, so now we copy the value out
2772 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
2773 MachineFunction &MF = DAG.getMachineFunction();
2774 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2775 unsigned Reg = MipsFI->getSRetReturnReg();
2778 llvm_unreachable("sret virtual register not created in the entry block");
2779 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2781 Chain = DAG.getCopyToReg(Chain, dl, Mips::V0, Val, Flag);
2782 Flag = Chain.getValue(1);
2785 // Return on Mips is always a "jr $ra"
2787 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2788 Chain, DAG.getRegister(Mips::RA, MVT::i32), Flag);
2790 return DAG.getNode(MipsISD::Ret, dl, MVT::Other,
2791 Chain, DAG.getRegister(Mips::RA, MVT::i32));
2794 //===----------------------------------------------------------------------===//
2795 // Mips Inline Assembly Support
2796 //===----------------------------------------------------------------------===//
2798 /// getConstraintType - Given a constraint letter, return the type of
2799 /// constraint it is for this target.
2800 MipsTargetLowering::ConstraintType MipsTargetLowering::
2801 getConstraintType(const std::string &Constraint) const
2803 // Mips specific constrainy
2804 // GCC config/mips/constraints.md
2806 // 'd' : An address register. Equivalent to r
2807 // unless generating MIPS16 code.
2808 // 'y' : Equivalent to r; retained for
2809 // backwards compatibility.
2810 // 'f' : Floating Point registers.
2811 if (Constraint.size() == 1) {
2812 switch (Constraint[0]) {
2817 return C_RegisterClass;
2821 return TargetLowering::getConstraintType(Constraint);
2824 /// Examine constraint type and operand type and determine a weight value.
2825 /// This object must already have been set up with the operand type
2826 /// and the current alternative constraint selected.
2827 TargetLowering::ConstraintWeight
2828 MipsTargetLowering::getSingleConstraintMatchWeight(
2829 AsmOperandInfo &info, const char *constraint) const {
2830 ConstraintWeight weight = CW_Invalid;
2831 Value *CallOperandVal = info.CallOperandVal;
2832 // If we don't have a value, we can't do a match,
2833 // but allow it at the lowest weight.
2834 if (CallOperandVal == NULL)
2836 Type *type = CallOperandVal->getType();
2837 // Look at the constraint type.
2838 switch (*constraint) {
2840 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
2844 if (type->isIntegerTy())
2845 weight = CW_Register;
2848 if (type->isFloatTy())
2849 weight = CW_Register;
2855 /// Given a register class constraint, like 'r', if this corresponds directly
2856 /// to an LLVM register class, return a register of 0 and the register class
2858 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering::
2859 getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const
2861 if (Constraint.size() == 1) {
2862 switch (Constraint[0]) {
2863 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
2864 case 'y': // Same as 'r'. Exists for compatibility.
2866 return std::make_pair(0U, Mips::CPURegsRegisterClass);
2869 return std::make_pair(0U, Mips::FGR32RegisterClass);
2871 if ((!Subtarget->isSingleFloat()) && (!Subtarget->isFP64bit()))
2872 return std::make_pair(0U, Mips::AFGR64RegisterClass);
2876 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2880 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
2881 // The Mips target isn't yet aware of offsets.
2885 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2886 if (VT != MVT::f32 && VT != MVT::f64)
2888 if (Imm.isNegZero())
2890 return Imm.isZero();