1 //===-- XCoreISelLowering.cpp - XCore 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 implements the XCoreTargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "XCoreISelLowering.h"
16 #include "XCoreMachineFunctionInfo.h"
17 #include "XCoreSubtarget.h"
18 #include "XCoreTargetMachine.h"
19 #include "XCoreTargetObjectFile.h"
20 #include "llvm/CodeGen/CallingConvLower.h"
21 #include "llvm/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineJumpTableInfo.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAGISel.h"
27 #include "llvm/CodeGen/ValueTypes.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/GlobalAlias.h"
33 #include "llvm/IR/GlobalVariable.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/KnownBits.h"
38 #include "llvm/Support/raw_ostream.h"
43 #define DEBUG_TYPE "xcore-lower"
45 const char *XCoreTargetLowering::
46 getTargetNodeName(unsigned Opcode) const
48 switch ((XCoreISD::NodeType)Opcode)
50 case XCoreISD::FIRST_NUMBER : break;
51 case XCoreISD::BL : return "XCoreISD::BL";
52 case XCoreISD::PCRelativeWrapper : return "XCoreISD::PCRelativeWrapper";
53 case XCoreISD::DPRelativeWrapper : return "XCoreISD::DPRelativeWrapper";
54 case XCoreISD::CPRelativeWrapper : return "XCoreISD::CPRelativeWrapper";
55 case XCoreISD::LDWSP : return "XCoreISD::LDWSP";
56 case XCoreISD::STWSP : return "XCoreISD::STWSP";
57 case XCoreISD::RETSP : return "XCoreISD::RETSP";
58 case XCoreISD::LADD : return "XCoreISD::LADD";
59 case XCoreISD::LSUB : return "XCoreISD::LSUB";
60 case XCoreISD::LMUL : return "XCoreISD::LMUL";
61 case XCoreISD::MACCU : return "XCoreISD::MACCU";
62 case XCoreISD::MACCS : return "XCoreISD::MACCS";
63 case XCoreISD::CRC8 : return "XCoreISD::CRC8";
64 case XCoreISD::BR_JT : return "XCoreISD::BR_JT";
65 case XCoreISD::BR_JT32 : return "XCoreISD::BR_JT32";
66 case XCoreISD::FRAME_TO_ARGS_OFFSET : return "XCoreISD::FRAME_TO_ARGS_OFFSET";
67 case XCoreISD::EH_RETURN : return "XCoreISD::EH_RETURN";
68 case XCoreISD::MEMBARRIER : return "XCoreISD::MEMBARRIER";
73 XCoreTargetLowering::XCoreTargetLowering(const TargetMachine &TM,
74 const XCoreSubtarget &Subtarget)
75 : TargetLowering(TM), TM(TM), Subtarget(Subtarget) {
77 // Set up the register classes.
78 addRegisterClass(MVT::i32, &XCore::GRRegsRegClass);
80 // Compute derived properties from the register classes
81 computeRegisterProperties(Subtarget.getRegisterInfo());
83 setStackPointerRegisterToSaveRestore(XCore::SP);
85 setSchedulingPreference(Sched::Source);
87 // Use i32 for setcc operations results (slt, sgt, ...).
88 setBooleanContents(ZeroOrOneBooleanContent);
89 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
91 // XCore does not have the NodeTypes below.
92 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
93 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
96 setOperationAction(ISD::ADD, MVT::i64, Custom);
97 setOperationAction(ISD::SUB, MVT::i64, Custom);
98 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Custom);
99 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Custom);
100 setOperationAction(ISD::MULHS, MVT::i32, Expand);
101 setOperationAction(ISD::MULHU, MVT::i32, Expand);
102 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
103 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
104 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
107 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
108 setOperationAction(ISD::ROTL , MVT::i32, Expand);
109 setOperationAction(ISD::ROTR , MVT::i32, Expand);
111 setOperationAction(ISD::TRAP, MVT::Other, Legal);
114 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
116 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
117 setOperationAction(ISD::BlockAddress, MVT::i32 , Custom);
119 // Conversion of i64 -> double produces constantpool nodes
120 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
123 for (MVT VT : MVT::integer_valuetypes()) {
124 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
125 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
126 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
128 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Expand);
129 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Expand);
132 // Custom expand misaligned loads / stores.
133 setOperationAction(ISD::LOAD, MVT::i32, Custom);
134 setOperationAction(ISD::STORE, MVT::i32, Custom);
137 setOperationAction(ISD::VAEND, MVT::Other, Expand);
138 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
139 setOperationAction(ISD::VAARG, MVT::Other, Custom);
140 setOperationAction(ISD::VASTART, MVT::Other, Custom);
143 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
144 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
145 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
147 // Exception handling
148 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
149 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
152 // We request a fence for ATOMIC_* instructions, to reduce them to Monotonic.
153 // As we are always Sequential Consistent, an ATOMIC_FENCE becomes a no OP.
154 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
155 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
156 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
158 // TRAMPOLINE is custom lowered.
159 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
160 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
162 // We want to custom lower some of our intrinsics.
163 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
165 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 4;
166 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize
167 = MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 2;
169 // We have target-specific dag combine patterns for the following nodes:
170 setTargetDAGCombine(ISD::STORE);
171 setTargetDAGCombine(ISD::ADD);
172 setTargetDAGCombine(ISD::INTRINSIC_VOID);
173 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
175 setMinFunctionAlignment(1);
176 setPrefFunctionAlignment(2);
179 bool XCoreTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
180 if (Val.getOpcode() != ISD::LOAD)
183 EVT VT1 = Val.getValueType();
184 if (!VT1.isSimple() || !VT1.isInteger() ||
185 !VT2.isSimple() || !VT2.isInteger())
188 switch (VT1.getSimpleVT().SimpleTy) {
197 SDValue XCoreTargetLowering::
198 LowerOperation(SDValue Op, SelectionDAG &DAG) const {
199 switch (Op.getOpcode())
201 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
202 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
203 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
204 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
205 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
206 case ISD::LOAD: return LowerLOAD(Op, DAG);
207 case ISD::STORE: return LowerSTORE(Op, DAG);
208 case ISD::VAARG: return LowerVAARG(Op, DAG);
209 case ISD::VASTART: return LowerVASTART(Op, DAG);
210 case ISD::SMUL_LOHI: return LowerSMUL_LOHI(Op, DAG);
211 case ISD::UMUL_LOHI: return LowerUMUL_LOHI(Op, DAG);
212 // FIXME: Remove these when LegalizeDAGTypes lands.
214 case ISD::SUB: return ExpandADDSUB(Op.getNode(), DAG);
215 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
216 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
217 case ISD::FRAME_TO_ARGS_OFFSET: return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
218 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
219 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
220 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
221 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
222 case ISD::ATOMIC_LOAD: return LowerATOMIC_LOAD(Op, DAG);
223 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op, DAG);
225 llvm_unreachable("unimplemented operand");
229 /// ReplaceNodeResults - Replace the results of node with an illegal result
230 /// type with new values built out of custom code.
231 void XCoreTargetLowering::ReplaceNodeResults(SDNode *N,
232 SmallVectorImpl<SDValue>&Results,
233 SelectionDAG &DAG) const {
234 switch (N->getOpcode()) {
236 llvm_unreachable("Don't know how to custom expand this!");
239 Results.push_back(ExpandADDSUB(N, DAG));
244 //===----------------------------------------------------------------------===//
245 // Misc Lower Operation implementation
246 //===----------------------------------------------------------------------===//
248 SDValue XCoreTargetLowering::getGlobalAddressWrapper(SDValue GA,
249 const GlobalValue *GV,
250 SelectionDAG &DAG) const {
251 // FIXME there is no actual debug info here
254 if (GV->getValueType()->isFunctionTy())
255 return DAG.getNode(XCoreISD::PCRelativeWrapper, dl, MVT::i32, GA);
257 const auto *GVar = dyn_cast<GlobalVariable>(GV);
258 if ((GV->hasSection() && GV->getSection().startswith(".cp.")) ||
259 (GVar && GVar->isConstant() && GV->hasLocalLinkage()))
260 return DAG.getNode(XCoreISD::CPRelativeWrapper, dl, MVT::i32, GA);
262 return DAG.getNode(XCoreISD::DPRelativeWrapper, dl, MVT::i32, GA);
265 static bool IsSmallObject(const GlobalValue *GV, const XCoreTargetLowering &XTL) {
266 if (XTL.getTargetMachine().getCodeModel() == CodeModel::Small)
269 Type *ObjType = GV->getValueType();
270 if (!ObjType->isSized())
273 auto &DL = GV->getParent()->getDataLayout();
274 unsigned ObjSize = DL.getTypeAllocSize(ObjType);
275 return ObjSize < CodeModelLargeSize && ObjSize != 0;
278 SDValue XCoreTargetLowering::
279 LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const
281 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
282 const GlobalValue *GV = GN->getGlobal();
284 int64_t Offset = GN->getOffset();
285 if (IsSmallObject(GV, *this)) {
286 // We can only fold positive offsets that are a multiple of the word size.
287 int64_t FoldedOffset = std::max(Offset & ~3, (int64_t)0);
288 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, FoldedOffset);
289 GA = getGlobalAddressWrapper(GA, GV, DAG);
290 // Handle the rest of the offset.
291 if (Offset != FoldedOffset) {
292 SDValue Remaining = DAG.getConstant(Offset - FoldedOffset, DL, MVT::i32);
293 GA = DAG.getNode(ISD::ADD, DL, MVT::i32, GA, Remaining);
297 // Ideally we would not fold in offset with an index <= 11.
298 Type *Ty = Type::getInt8PtrTy(*DAG.getContext());
299 Constant *GA = ConstantExpr::getBitCast(const_cast<GlobalValue*>(GV), Ty);
300 Ty = Type::getInt32Ty(*DAG.getContext());
301 Constant *Idx = ConstantInt::get(Ty, Offset);
302 Constant *GAI = ConstantExpr::getGetElementPtr(
303 Type::getInt8Ty(*DAG.getContext()), GA, Idx);
304 SDValue CP = DAG.getConstantPool(GAI, MVT::i32);
305 return DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL,
306 DAG.getEntryNode(), CP, MachinePointerInfo());
310 SDValue XCoreTargetLowering::
311 LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const
314 auto PtrVT = getPointerTy(DAG.getDataLayout());
315 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
316 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT);
318 return DAG.getNode(XCoreISD::PCRelativeWrapper, DL, PtrVT, Result);
321 SDValue XCoreTargetLowering::
322 LowerConstantPool(SDValue Op, SelectionDAG &DAG) const
324 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
325 // FIXME there isn't really debug info here
327 EVT PtrVT = Op.getValueType();
329 if (CP->isMachineConstantPoolEntry()) {
330 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
331 CP->getAlignment(), CP->getOffset());
333 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
334 CP->getAlignment(), CP->getOffset());
336 return DAG.getNode(XCoreISD::CPRelativeWrapper, dl, MVT::i32, Res);
339 unsigned XCoreTargetLowering::getJumpTableEncoding() const {
340 return MachineJumpTableInfo::EK_Inline;
343 SDValue XCoreTargetLowering::
344 LowerBR_JT(SDValue Op, SelectionDAG &DAG) const
346 SDValue Chain = Op.getOperand(0);
347 SDValue Table = Op.getOperand(1);
348 SDValue Index = Op.getOperand(2);
350 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
351 unsigned JTI = JT->getIndex();
352 MachineFunction &MF = DAG.getMachineFunction();
353 const MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
354 SDValue TargetJT = DAG.getTargetJumpTable(JT->getIndex(), MVT::i32);
356 unsigned NumEntries = MJTI->getJumpTables()[JTI].MBBs.size();
357 if (NumEntries <= 32) {
358 return DAG.getNode(XCoreISD::BR_JT, dl, MVT::Other, Chain, TargetJT, Index);
360 assert((NumEntries >> 31) == 0);
361 SDValue ScaledIndex = DAG.getNode(ISD::SHL, dl, MVT::i32, Index,
362 DAG.getConstant(1, dl, MVT::i32));
363 return DAG.getNode(XCoreISD::BR_JT32, dl, MVT::Other, Chain, TargetJT,
367 SDValue XCoreTargetLowering::lowerLoadWordFromAlignedBasePlusOffset(
368 const SDLoc &DL, SDValue Chain, SDValue Base, int64_t Offset,
369 SelectionDAG &DAG) const {
370 auto PtrVT = getPointerTy(DAG.getDataLayout());
371 if ((Offset & 0x3) == 0) {
372 return DAG.getLoad(PtrVT, DL, Chain, Base, MachinePointerInfo());
374 // Lower to pair of consecutive word aligned loads plus some bit shifting.
375 int32_t HighOffset = alignTo(Offset, 4);
376 int32_t LowOffset = HighOffset - 4;
377 SDValue LowAddr, HighAddr;
378 if (GlobalAddressSDNode *GASD =
379 dyn_cast<GlobalAddressSDNode>(Base.getNode())) {
380 LowAddr = DAG.getGlobalAddress(GASD->getGlobal(), DL, Base.getValueType(),
382 HighAddr = DAG.getGlobalAddress(GASD->getGlobal(), DL, Base.getValueType(),
385 LowAddr = DAG.getNode(ISD::ADD, DL, MVT::i32, Base,
386 DAG.getConstant(LowOffset, DL, MVT::i32));
387 HighAddr = DAG.getNode(ISD::ADD, DL, MVT::i32, Base,
388 DAG.getConstant(HighOffset, DL, MVT::i32));
390 SDValue LowShift = DAG.getConstant((Offset - LowOffset) * 8, DL, MVT::i32);
391 SDValue HighShift = DAG.getConstant((HighOffset - Offset) * 8, DL, MVT::i32);
393 SDValue Low = DAG.getLoad(PtrVT, DL, Chain, LowAddr, MachinePointerInfo());
394 SDValue High = DAG.getLoad(PtrVT, DL, Chain, HighAddr, MachinePointerInfo());
395 SDValue LowShifted = DAG.getNode(ISD::SRL, DL, MVT::i32, Low, LowShift);
396 SDValue HighShifted = DAG.getNode(ISD::SHL, DL, MVT::i32, High, HighShift);
397 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, LowShifted, HighShifted);
398 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Low.getValue(1),
400 SDValue Ops[] = { Result, Chain };
401 return DAG.getMergeValues(Ops, DL);
404 static bool isWordAligned(SDValue Value, SelectionDAG &DAG)
406 KnownBits Known = DAG.computeKnownBits(Value);
407 return Known.countMinTrailingZeros() >= 2;
410 SDValue XCoreTargetLowering::
411 LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
412 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
413 LoadSDNode *LD = cast<LoadSDNode>(Op);
414 assert(LD->getExtensionType() == ISD::NON_EXTLOAD &&
415 "Unexpected extension type");
416 assert(LD->getMemoryVT() == MVT::i32 && "Unexpected load EVT");
417 if (allowsMisalignedMemoryAccesses(LD->getMemoryVT(),
418 LD->getAddressSpace(),
422 auto &TD = DAG.getDataLayout();
423 unsigned ABIAlignment = TD.getABITypeAlignment(
424 LD->getMemoryVT().getTypeForEVT(*DAG.getContext()));
425 // Leave aligned load alone.
426 if (LD->getAlignment() >= ABIAlignment)
429 SDValue Chain = LD->getChain();
430 SDValue BasePtr = LD->getBasePtr();
433 if (!LD->isVolatile()) {
434 const GlobalValue *GV;
436 if (DAG.isBaseWithConstantOffset(BasePtr) &&
437 isWordAligned(BasePtr->getOperand(0), DAG)) {
438 SDValue NewBasePtr = BasePtr->getOperand(0);
439 Offset = cast<ConstantSDNode>(BasePtr->getOperand(1))->getSExtValue();
440 return lowerLoadWordFromAlignedBasePlusOffset(DL, Chain, NewBasePtr,
443 if (TLI.isGAPlusOffset(BasePtr.getNode(), GV, Offset) &&
444 MinAlign(GV->getAlignment(), 4) == 4) {
445 SDValue NewBasePtr = DAG.getGlobalAddress(GV, DL,
446 BasePtr->getValueType(0));
447 return lowerLoadWordFromAlignedBasePlusOffset(DL, Chain, NewBasePtr,
452 if (LD->getAlignment() == 2) {
454 DAG.getExtLoad(ISD::ZEXTLOAD, DL, MVT::i32, Chain, BasePtr,
455 LD->getPointerInfo(), MVT::i16,
456 /* Alignment = */ 2, LD->getMemOperand()->getFlags());
457 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
458 DAG.getConstant(2, DL, MVT::i32));
460 DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain, HighAddr,
461 LD->getPointerInfo().getWithOffset(2), MVT::i16,
462 /* Alignment = */ 2, LD->getMemOperand()->getFlags());
463 SDValue HighShifted = DAG.getNode(ISD::SHL, DL, MVT::i32, High,
464 DAG.getConstant(16, DL, MVT::i32));
465 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, Low, HighShifted);
466 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Low.getValue(1),
468 SDValue Ops[] = { Result, Chain };
469 return DAG.getMergeValues(Ops, DL);
472 // Lower to a call to __misaligned_load(BasePtr).
473 Type *IntPtrTy = TD.getIntPtrType(*DAG.getContext());
474 TargetLowering::ArgListTy Args;
475 TargetLowering::ArgListEntry Entry;
478 Entry.Node = BasePtr;
479 Args.push_back(Entry);
481 TargetLowering::CallLoweringInfo CLI(DAG);
482 CLI.setDebugLoc(DL).setChain(Chain).setLibCallee(
483 CallingConv::C, IntPtrTy,
484 DAG.getExternalSymbol("__misaligned_load",
485 getPointerTy(DAG.getDataLayout())),
488 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
489 SDValue Ops[] = { CallResult.first, CallResult.second };
490 return DAG.getMergeValues(Ops, DL);
493 SDValue XCoreTargetLowering::
494 LowerSTORE(SDValue Op, SelectionDAG &DAG) const
496 StoreSDNode *ST = cast<StoreSDNode>(Op);
497 assert(!ST->isTruncatingStore() && "Unexpected store type");
498 assert(ST->getMemoryVT() == MVT::i32 && "Unexpected store EVT");
499 if (allowsMisalignedMemoryAccesses(ST->getMemoryVT(),
500 ST->getAddressSpace(),
501 ST->getAlignment())) {
504 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(
505 ST->getMemoryVT().getTypeForEVT(*DAG.getContext()));
506 // Leave aligned store alone.
507 if (ST->getAlignment() >= ABIAlignment) {
510 SDValue Chain = ST->getChain();
511 SDValue BasePtr = ST->getBasePtr();
512 SDValue Value = ST->getValue();
515 if (ST->getAlignment() == 2) {
517 SDValue High = DAG.getNode(ISD::SRL, dl, MVT::i32, Value,
518 DAG.getConstant(16, dl, MVT::i32));
519 SDValue StoreLow = DAG.getTruncStore(
520 Chain, dl, Low, BasePtr, ST->getPointerInfo(), MVT::i16,
521 /* Alignment = */ 2, ST->getMemOperand()->getFlags());
522 SDValue HighAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, BasePtr,
523 DAG.getConstant(2, dl, MVT::i32));
524 SDValue StoreHigh = DAG.getTruncStore(
525 Chain, dl, High, HighAddr, ST->getPointerInfo().getWithOffset(2),
526 MVT::i16, /* Alignment = */ 2, ST->getMemOperand()->getFlags());
527 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, StoreLow, StoreHigh);
530 // Lower to a call to __misaligned_store(BasePtr, Value).
531 Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
532 TargetLowering::ArgListTy Args;
533 TargetLowering::ArgListEntry Entry;
536 Entry.Node = BasePtr;
537 Args.push_back(Entry);
540 Args.push_back(Entry);
542 TargetLowering::CallLoweringInfo CLI(DAG);
543 CLI.setDebugLoc(dl).setChain(Chain).setCallee(
544 CallingConv::C, Type::getVoidTy(*DAG.getContext()),
545 DAG.getExternalSymbol("__misaligned_store",
546 getPointerTy(DAG.getDataLayout())),
549 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
550 return CallResult.second;
553 SDValue XCoreTargetLowering::
554 LowerSMUL_LOHI(SDValue Op, SelectionDAG &DAG) const
556 assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::SMUL_LOHI &&
557 "Unexpected operand to lower!");
559 SDValue LHS = Op.getOperand(0);
560 SDValue RHS = Op.getOperand(1);
561 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
562 SDValue Hi = DAG.getNode(XCoreISD::MACCS, dl,
563 DAG.getVTList(MVT::i32, MVT::i32), Zero, Zero,
565 SDValue Lo(Hi.getNode(), 1);
566 SDValue Ops[] = { Lo, Hi };
567 return DAG.getMergeValues(Ops, dl);
570 SDValue XCoreTargetLowering::
571 LowerUMUL_LOHI(SDValue Op, SelectionDAG &DAG) const
573 assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::UMUL_LOHI &&
574 "Unexpected operand to lower!");
576 SDValue LHS = Op.getOperand(0);
577 SDValue RHS = Op.getOperand(1);
578 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
579 SDValue Hi = DAG.getNode(XCoreISD::LMUL, dl,
580 DAG.getVTList(MVT::i32, MVT::i32), LHS, RHS,
582 SDValue Lo(Hi.getNode(), 1);
583 SDValue Ops[] = { Lo, Hi };
584 return DAG.getMergeValues(Ops, dl);
587 /// isADDADDMUL - Return whether Op is in a form that is equivalent to
588 /// add(add(mul(x,y),a),b). If requireIntermediatesHaveOneUse is true then
589 /// each intermediate result in the calculation must also have a single use.
590 /// If the Op is in the correct form the constituent parts are written to Mul0,
591 /// Mul1, Addend0 and Addend1.
593 isADDADDMUL(SDValue Op, SDValue &Mul0, SDValue &Mul1, SDValue &Addend0,
594 SDValue &Addend1, bool requireIntermediatesHaveOneUse)
596 if (Op.getOpcode() != ISD::ADD)
598 SDValue N0 = Op.getOperand(0);
599 SDValue N1 = Op.getOperand(1);
602 if (N0.getOpcode() == ISD::ADD) {
605 } else if (N1.getOpcode() == ISD::ADD) {
611 if (requireIntermediatesHaveOneUse && !AddOp.hasOneUse())
613 if (OtherOp.getOpcode() == ISD::MUL) {
614 // add(add(a,b),mul(x,y))
615 if (requireIntermediatesHaveOneUse && !OtherOp.hasOneUse())
617 Mul0 = OtherOp.getOperand(0);
618 Mul1 = OtherOp.getOperand(1);
619 Addend0 = AddOp.getOperand(0);
620 Addend1 = AddOp.getOperand(1);
623 if (AddOp.getOperand(0).getOpcode() == ISD::MUL) {
624 // add(add(mul(x,y),a),b)
625 if (requireIntermediatesHaveOneUse && !AddOp.getOperand(0).hasOneUse())
627 Mul0 = AddOp.getOperand(0).getOperand(0);
628 Mul1 = AddOp.getOperand(0).getOperand(1);
629 Addend0 = AddOp.getOperand(1);
633 if (AddOp.getOperand(1).getOpcode() == ISD::MUL) {
634 // add(add(a,mul(x,y)),b)
635 if (requireIntermediatesHaveOneUse && !AddOp.getOperand(1).hasOneUse())
637 Mul0 = AddOp.getOperand(1).getOperand(0);
638 Mul1 = AddOp.getOperand(1).getOperand(1);
639 Addend0 = AddOp.getOperand(0);
646 SDValue XCoreTargetLowering::
647 TryExpandADDWithMul(SDNode *N, SelectionDAG &DAG) const
651 if (N->getOperand(0).getOpcode() == ISD::MUL) {
652 Mul = N->getOperand(0);
653 Other = N->getOperand(1);
654 } else if (N->getOperand(1).getOpcode() == ISD::MUL) {
655 Mul = N->getOperand(1);
656 Other = N->getOperand(0);
661 SDValue LL, RL, AddendL, AddendH;
662 LL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
663 Mul.getOperand(0), DAG.getConstant(0, dl, MVT::i32));
664 RL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
665 Mul.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
666 AddendL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
667 Other, DAG.getConstant(0, dl, MVT::i32));
668 AddendH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
669 Other, DAG.getConstant(1, dl, MVT::i32));
670 APInt HighMask = APInt::getHighBitsSet(64, 32);
671 unsigned LHSSB = DAG.ComputeNumSignBits(Mul.getOperand(0));
672 unsigned RHSSB = DAG.ComputeNumSignBits(Mul.getOperand(1));
673 if (DAG.MaskedValueIsZero(Mul.getOperand(0), HighMask) &&
674 DAG.MaskedValueIsZero(Mul.getOperand(1), HighMask)) {
675 // The inputs are both zero-extended.
676 SDValue Hi = DAG.getNode(XCoreISD::MACCU, dl,
677 DAG.getVTList(MVT::i32, MVT::i32), AddendH,
679 SDValue Lo(Hi.getNode(), 1);
680 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
682 if (LHSSB > 32 && RHSSB > 32) {
683 // The inputs are both sign-extended.
684 SDValue Hi = DAG.getNode(XCoreISD::MACCS, dl,
685 DAG.getVTList(MVT::i32, MVT::i32), AddendH,
687 SDValue Lo(Hi.getNode(), 1);
688 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
691 LH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
692 Mul.getOperand(0), DAG.getConstant(1, dl, MVT::i32));
693 RH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
694 Mul.getOperand(1), DAG.getConstant(1, dl, MVT::i32));
695 SDValue Hi = DAG.getNode(XCoreISD::MACCU, dl,
696 DAG.getVTList(MVT::i32, MVT::i32), AddendH,
698 SDValue Lo(Hi.getNode(), 1);
699 RH = DAG.getNode(ISD::MUL, dl, MVT::i32, LL, RH);
700 LH = DAG.getNode(ISD::MUL, dl, MVT::i32, LH, RL);
701 Hi = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, RH);
702 Hi = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, LH);
703 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
706 SDValue XCoreTargetLowering::
707 ExpandADDSUB(SDNode *N, SelectionDAG &DAG) const
709 assert(N->getValueType(0) == MVT::i64 &&
710 (N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) &&
711 "Unknown operand to lower!");
713 if (N->getOpcode() == ISD::ADD)
714 if (SDValue Result = TryExpandADDWithMul(N, DAG))
719 // Extract components
720 SDValue LHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
722 DAG.getConstant(0, dl, MVT::i32));
723 SDValue LHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
725 DAG.getConstant(1, dl, MVT::i32));
726 SDValue RHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
728 DAG.getConstant(0, dl, MVT::i32));
729 SDValue RHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
731 DAG.getConstant(1, dl, MVT::i32));
734 unsigned Opcode = (N->getOpcode() == ISD::ADD) ? XCoreISD::LADD :
736 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
737 SDValue Lo = DAG.getNode(Opcode, dl, DAG.getVTList(MVT::i32, MVT::i32),
739 SDValue Carry(Lo.getNode(), 1);
741 SDValue Hi = DAG.getNode(Opcode, dl, DAG.getVTList(MVT::i32, MVT::i32),
743 SDValue Ignored(Hi.getNode(), 1);
745 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
748 SDValue XCoreTargetLowering::
749 LowerVAARG(SDValue Op, SelectionDAG &DAG) const
751 // Whist llvm does not support aggregate varargs we can ignore
752 // the possibility of the ValueType being an implicit byVal vararg.
753 SDNode *Node = Op.getNode();
754 EVT VT = Node->getValueType(0); // not an aggregate
755 SDValue InChain = Node->getOperand(0);
756 SDValue VAListPtr = Node->getOperand(1);
757 EVT PtrVT = VAListPtr.getValueType();
758 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
761 DAG.getLoad(PtrVT, dl, InChain, VAListPtr, MachinePointerInfo(SV));
762 // Increment the pointer, VAList, to the next vararg
763 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAList,
764 DAG.getIntPtrConstant(VT.getSizeInBits() / 8,
766 // Store the incremented VAList to the legalized pointer
767 InChain = DAG.getStore(VAList.getValue(1), dl, nextPtr, VAListPtr,
768 MachinePointerInfo(SV));
769 // Load the actual argument out of the pointer VAList
770 return DAG.getLoad(VT, dl, InChain, VAList, MachinePointerInfo());
773 SDValue XCoreTargetLowering::
774 LowerVASTART(SDValue Op, SelectionDAG &DAG) const
777 // vastart stores the address of the VarArgsFrameIndex slot into the
778 // memory location argument
779 MachineFunction &MF = DAG.getMachineFunction();
780 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
781 SDValue Addr = DAG.getFrameIndex(XFI->getVarArgsFrameIndex(), MVT::i32);
782 return DAG.getStore(Op.getOperand(0), dl, Addr, Op.getOperand(1),
783 MachinePointerInfo());
786 SDValue XCoreTargetLowering::LowerFRAMEADDR(SDValue Op,
787 SelectionDAG &DAG) const {
788 // This nodes represent llvm.frameaddress on the DAG.
789 // It takes one operand, the index of the frame address to return.
790 // An index of zero corresponds to the current function's frame address.
791 // An index of one to the parent's frame address, and so on.
792 // Depths > 0 not supported yet!
793 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
796 MachineFunction &MF = DAG.getMachineFunction();
797 const TargetRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
798 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op),
799 RegInfo->getFrameRegister(MF), MVT::i32);
802 SDValue XCoreTargetLowering::
803 LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const {
804 // This nodes represent llvm.returnaddress on the DAG.
805 // It takes one operand, the index of the return address to return.
806 // An index of zero corresponds to the current function's return address.
807 // An index of one to the parent's return address, and so on.
808 // Depths > 0 not supported yet!
809 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
812 MachineFunction &MF = DAG.getMachineFunction();
813 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
814 int FI = XFI->createLRSpillSlot(MF);
815 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
816 return DAG.getLoad(getPointerTy(DAG.getDataLayout()), SDLoc(Op),
817 DAG.getEntryNode(), FIN,
818 MachinePointerInfo::getFixedStack(MF, FI));
821 SDValue XCoreTargetLowering::
822 LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const {
823 // This node represents offset from frame pointer to first on-stack argument.
824 // This is needed for correct stack adjustment during unwind.
825 // However, we don't know the offset until after the frame has be finalised.
826 // This is done during the XCoreFTAOElim pass.
827 return DAG.getNode(XCoreISD::FRAME_TO_ARGS_OFFSET, SDLoc(Op), MVT::i32);
830 SDValue XCoreTargetLowering::
831 LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
832 // OUTCHAIN = EH_RETURN(INCHAIN, OFFSET, HANDLER)
833 // This node represents 'eh_return' gcc dwarf builtin, which is used to
834 // return from exception. The general meaning is: adjust stack by OFFSET and
835 // pass execution to HANDLER.
836 MachineFunction &MF = DAG.getMachineFunction();
837 SDValue Chain = Op.getOperand(0);
838 SDValue Offset = Op.getOperand(1);
839 SDValue Handler = Op.getOperand(2);
842 // Absolute SP = (FP + FrameToArgs) + Offset
843 const TargetRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
844 SDValue Stack = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
845 RegInfo->getFrameRegister(MF), MVT::i32);
846 SDValue FrameToArgs = DAG.getNode(XCoreISD::FRAME_TO_ARGS_OFFSET, dl,
848 Stack = DAG.getNode(ISD::ADD, dl, MVT::i32, Stack, FrameToArgs);
849 Stack = DAG.getNode(ISD::ADD, dl, MVT::i32, Stack, Offset);
851 // R0=ExceptionPointerRegister R1=ExceptionSelectorRegister
852 // which leaves 2 caller saved registers, R2 & R3 for us to use.
853 unsigned StackReg = XCore::R2;
854 unsigned HandlerReg = XCore::R3;
856 SDValue OutChains[] = {
857 DAG.getCopyToReg(Chain, dl, StackReg, Stack),
858 DAG.getCopyToReg(Chain, dl, HandlerReg, Handler)
861 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
863 return DAG.getNode(XCoreISD::EH_RETURN, dl, MVT::Other, Chain,
864 DAG.getRegister(StackReg, MVT::i32),
865 DAG.getRegister(HandlerReg, MVT::i32));
869 SDValue XCoreTargetLowering::
870 LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const {
871 return Op.getOperand(0);
874 SDValue XCoreTargetLowering::
875 LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const {
876 SDValue Chain = Op.getOperand(0);
877 SDValue Trmp = Op.getOperand(1); // trampoline
878 SDValue FPtr = Op.getOperand(2); // nested function
879 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
881 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
884 // LDAPF_u10 r11, nest
885 // LDW_2rus r11, r11[0]
886 // STWSP_ru6 r11, sp[0]
887 // LDAPF_u10 r11, fptr
888 // LDW_2rus r11, r11[0]
894 SDValue OutChains[5];
900 DAG.getStore(Chain, dl, DAG.getConstant(0x0a3cd805, dl, MVT::i32), Addr,
901 MachinePointerInfo(TrmpAddr));
903 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
904 DAG.getConstant(4, dl, MVT::i32));
906 DAG.getStore(Chain, dl, DAG.getConstant(0xd80456c0, dl, MVT::i32), Addr,
907 MachinePointerInfo(TrmpAddr, 4));
909 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
910 DAG.getConstant(8, dl, MVT::i32));
912 DAG.getStore(Chain, dl, DAG.getConstant(0x27fb0a3c, dl, MVT::i32), Addr,
913 MachinePointerInfo(TrmpAddr, 8));
915 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
916 DAG.getConstant(12, dl, MVT::i32));
918 DAG.getStore(Chain, dl, Nest, Addr, MachinePointerInfo(TrmpAddr, 12));
920 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
921 DAG.getConstant(16, dl, MVT::i32));
923 DAG.getStore(Chain, dl, FPtr, Addr, MachinePointerInfo(TrmpAddr, 16));
925 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
928 SDValue XCoreTargetLowering::
929 LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
931 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
933 case Intrinsic::xcore_crc8:
934 EVT VT = Op.getValueType();
936 DAG.getNode(XCoreISD::CRC8, DL, DAG.getVTList(VT, VT),
937 Op.getOperand(1), Op.getOperand(2) , Op.getOperand(3));
938 SDValue Crc(Data.getNode(), 1);
939 SDValue Results[] = { Crc, Data };
940 return DAG.getMergeValues(Results, DL);
945 SDValue XCoreTargetLowering::
946 LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const {
948 return DAG.getNode(XCoreISD::MEMBARRIER, DL, MVT::Other, Op.getOperand(0));
951 SDValue XCoreTargetLowering::
952 LowerATOMIC_LOAD(SDValue Op, SelectionDAG &DAG) const {
953 AtomicSDNode *N = cast<AtomicSDNode>(Op);
954 assert(N->getOpcode() == ISD::ATOMIC_LOAD && "Bad Atomic OP");
955 assert((N->getOrdering() == AtomicOrdering::Unordered ||
956 N->getOrdering() == AtomicOrdering::Monotonic) &&
957 "setInsertFencesForAtomic(true) expects unordered / monotonic");
958 if (N->getMemoryVT() == MVT::i32) {
959 if (N->getAlignment() < 4)
960 report_fatal_error("atomic load must be aligned");
961 return DAG.getLoad(getPointerTy(DAG.getDataLayout()), SDLoc(Op),
962 N->getChain(), N->getBasePtr(), N->getPointerInfo(),
963 N->getAlignment(), N->getMemOperand()->getFlags(),
964 N->getAAInfo(), N->getRanges());
966 if (N->getMemoryVT() == MVT::i16) {
967 if (N->getAlignment() < 2)
968 report_fatal_error("atomic load must be aligned");
969 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), MVT::i32, N->getChain(),
970 N->getBasePtr(), N->getPointerInfo(), MVT::i16,
971 N->getAlignment(), N->getMemOperand()->getFlags(),
974 if (N->getMemoryVT() == MVT::i8)
975 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), MVT::i32, N->getChain(),
976 N->getBasePtr(), N->getPointerInfo(), MVT::i8,
977 N->getAlignment(), N->getMemOperand()->getFlags(),
982 SDValue XCoreTargetLowering::
983 LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) const {
984 AtomicSDNode *N = cast<AtomicSDNode>(Op);
985 assert(N->getOpcode() == ISD::ATOMIC_STORE && "Bad Atomic OP");
986 assert((N->getOrdering() == AtomicOrdering::Unordered ||
987 N->getOrdering() == AtomicOrdering::Monotonic) &&
988 "setInsertFencesForAtomic(true) expects unordered / monotonic");
989 if (N->getMemoryVT() == MVT::i32) {
990 if (N->getAlignment() < 4)
991 report_fatal_error("atomic store must be aligned");
992 return DAG.getStore(N->getChain(), SDLoc(Op), N->getVal(), N->getBasePtr(),
993 N->getPointerInfo(), N->getAlignment(),
994 N->getMemOperand()->getFlags(), N->getAAInfo());
996 if (N->getMemoryVT() == MVT::i16) {
997 if (N->getAlignment() < 2)
998 report_fatal_error("atomic store must be aligned");
999 return DAG.getTruncStore(N->getChain(), SDLoc(Op), N->getVal(),
1000 N->getBasePtr(), N->getPointerInfo(), MVT::i16,
1001 N->getAlignment(), N->getMemOperand()->getFlags(),
1004 if (N->getMemoryVT() == MVT::i8)
1005 return DAG.getTruncStore(N->getChain(), SDLoc(Op), N->getVal(),
1006 N->getBasePtr(), N->getPointerInfo(), MVT::i8,
1007 N->getAlignment(), N->getMemOperand()->getFlags(),
1012 //===----------------------------------------------------------------------===//
1013 // Calling Convention Implementation
1014 //===----------------------------------------------------------------------===//
1016 #include "XCoreGenCallingConv.inc"
1018 //===----------------------------------------------------------------------===//
1019 // Call Calling Convention Implementation
1020 //===----------------------------------------------------------------------===//
1022 /// XCore call implementation
1024 XCoreTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1025 SmallVectorImpl<SDValue> &InVals) const {
1026 SelectionDAG &DAG = CLI.DAG;
1028 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1029 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1030 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1031 SDValue Chain = CLI.Chain;
1032 SDValue Callee = CLI.Callee;
1033 bool &isTailCall = CLI.IsTailCall;
1034 CallingConv::ID CallConv = CLI.CallConv;
1035 bool isVarArg = CLI.IsVarArg;
1037 // XCore target does not yet support tail call optimization.
1040 // For now, only CallingConv::C implemented
1044 report_fatal_error("Unsupported calling convention");
1045 case CallingConv::Fast:
1046 case CallingConv::C:
1047 return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall,
1048 Outs, OutVals, Ins, dl, DAG, InVals);
1052 /// LowerCallResult - Lower the result values of a call into the
1053 /// appropriate copies out of appropriate physical registers / memory locations.
1054 static SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1055 const SmallVectorImpl<CCValAssign> &RVLocs,
1056 const SDLoc &dl, SelectionDAG &DAG,
1057 SmallVectorImpl<SDValue> &InVals) {
1058 SmallVector<std::pair<int, unsigned>, 4> ResultMemLocs;
1059 // Copy results out of physical registers.
1060 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1061 const CCValAssign &VA = RVLocs[i];
1062 if (VA.isRegLoc()) {
1063 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getValVT(),
1064 InFlag).getValue(1);
1065 InFlag = Chain.getValue(2);
1066 InVals.push_back(Chain.getValue(0));
1068 assert(VA.isMemLoc());
1069 ResultMemLocs.push_back(std::make_pair(VA.getLocMemOffset(),
1071 // Reserve space for this result.
1072 InVals.push_back(SDValue());
1076 // Copy results out of memory.
1077 SmallVector<SDValue, 4> MemOpChains;
1078 for (unsigned i = 0, e = ResultMemLocs.size(); i != e; ++i) {
1079 int offset = ResultMemLocs[i].first;
1080 unsigned index = ResultMemLocs[i].second;
1081 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other);
1082 SDValue Ops[] = { Chain, DAG.getConstant(offset / 4, dl, MVT::i32) };
1083 SDValue load = DAG.getNode(XCoreISD::LDWSP, dl, VTs, Ops);
1084 InVals[index] = load;
1085 MemOpChains.push_back(load.getValue(1));
1088 // Transform all loads nodes into one single node because
1089 // all load nodes are independent of each other.
1090 if (!MemOpChains.empty())
1091 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1096 /// LowerCCCCallTo - functions arguments are copied from virtual
1097 /// regs to (physical regs)/(stack frame), CALLSEQ_START and
1098 /// CALLSEQ_END are emitted.
1099 /// TODO: isTailCall, sret.
1100 SDValue XCoreTargetLowering::LowerCCCCallTo(
1101 SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg,
1102 bool isTailCall, const SmallVectorImpl<ISD::OutputArg> &Outs,
1103 const SmallVectorImpl<SDValue> &OutVals,
1104 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1105 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1107 // Analyze operands of the call, assigning locations to each operand.
1108 SmallVector<CCValAssign, 16> ArgLocs;
1109 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1112 // The ABI dictates there should be one stack slot available to the callee
1113 // on function entry (for saving lr).
1114 CCInfo.AllocateStack(4, 4);
1116 CCInfo.AnalyzeCallOperands(Outs, CC_XCore);
1118 SmallVector<CCValAssign, 16> RVLocs;
1119 // Analyze return values to determine the number of bytes of stack required.
1120 CCState RetCCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1122 RetCCInfo.AllocateStack(CCInfo.getNextStackOffset(), 4);
1123 RetCCInfo.AnalyzeCallResult(Ins, RetCC_XCore);
1125 // Get a count of how many bytes are to be pushed on the stack.
1126 unsigned NumBytes = RetCCInfo.getNextStackOffset();
1127 auto PtrVT = getPointerTy(DAG.getDataLayout());
1129 Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl);
1131 SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass;
1132 SmallVector<SDValue, 12> MemOpChains;
1134 // Walk the register/memloc assignments, inserting copies/loads.
1135 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1136 CCValAssign &VA = ArgLocs[i];
1137 SDValue Arg = OutVals[i];
1139 // Promote the value if needed.
1140 switch (VA.getLocInfo()) {
1141 default: llvm_unreachable("Unknown loc info!");
1142 case CCValAssign::Full: break;
1143 case CCValAssign::SExt:
1144 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1146 case CCValAssign::ZExt:
1147 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1149 case CCValAssign::AExt:
1150 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1154 // Arguments that can be passed on register must be kept at
1155 // RegsToPass vector
1156 if (VA.isRegLoc()) {
1157 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1159 assert(VA.isMemLoc());
1161 int Offset = VA.getLocMemOffset();
1163 MemOpChains.push_back(DAG.getNode(XCoreISD::STWSP, dl, MVT::Other,
1165 DAG.getConstant(Offset/4, dl,
1170 // Transform all store nodes into one single node because
1171 // all store nodes are independent of each other.
1172 if (!MemOpChains.empty())
1173 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1175 // Build a sequence of copy-to-reg nodes chained together with token
1176 // chain and flag operands which copy the outgoing args into registers.
1177 // The InFlag in necessary since all emitted instructions must be
1180 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1181 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1182 RegsToPass[i].second, InFlag);
1183 InFlag = Chain.getValue(1);
1186 // If the callee is a GlobalAddress node (quite common, every direct call is)
1187 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1188 // Likewise ExternalSymbol -> TargetExternalSymbol.
1189 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1190 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, MVT::i32);
1191 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee))
1192 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), MVT::i32);
1194 // XCoreBranchLink = #chain, #target_address, #opt_in_flags...
1195 // = Chain, Callee, Reg#1, Reg#2, ...
1197 // Returns a chain & a flag for retval copy to use.
1198 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1199 SmallVector<SDValue, 8> Ops;
1200 Ops.push_back(Chain);
1201 Ops.push_back(Callee);
1203 // Add argument registers to the end of the list so that they are
1204 // known live into the call.
1205 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1206 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1207 RegsToPass[i].second.getValueType()));
1209 if (InFlag.getNode())
1210 Ops.push_back(InFlag);
1212 Chain = DAG.getNode(XCoreISD::BL, dl, NodeTys, Ops);
1213 InFlag = Chain.getValue(1);
1215 // Create the CALLSEQ_END node.
1216 Chain = DAG.getCALLSEQ_END(Chain, DAG.getConstant(NumBytes, dl, PtrVT, true),
1217 DAG.getConstant(0, dl, PtrVT, true), InFlag, dl);
1218 InFlag = Chain.getValue(1);
1220 // Handle result values, copying them out of physregs into vregs that we
1222 return LowerCallResult(Chain, InFlag, RVLocs, dl, DAG, InVals);
1225 //===----------------------------------------------------------------------===//
1226 // Formal Arguments Calling Convention Implementation
1227 //===----------------------------------------------------------------------===//
1230 struct ArgDataPair { SDValue SDV; ISD::ArgFlagsTy Flags; };
1233 /// XCore formal arguments implementation
1234 SDValue XCoreTargetLowering::LowerFormalArguments(
1235 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1236 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1237 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1241 report_fatal_error("Unsupported calling convention");
1242 case CallingConv::C:
1243 case CallingConv::Fast:
1244 return LowerCCCArguments(Chain, CallConv, isVarArg,
1245 Ins, dl, DAG, InVals);
1249 /// LowerCCCArguments - transform physical registers into
1250 /// virtual registers and generate load operations for
1251 /// arguments places on the stack.
1253 SDValue XCoreTargetLowering::LowerCCCArguments(
1254 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1255 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1256 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1257 MachineFunction &MF = DAG.getMachineFunction();
1258 MachineFrameInfo &MFI = MF.getFrameInfo();
1259 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1260 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
1262 // Assign locations to all of the incoming arguments.
1263 SmallVector<CCValAssign, 16> ArgLocs;
1264 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1267 CCInfo.AnalyzeFormalArguments(Ins, CC_XCore);
1269 unsigned StackSlotSize = XCoreFrameLowering::stackSlotSize();
1271 unsigned LRSaveSize = StackSlotSize;
1274 XFI->setReturnStackOffset(CCInfo.getNextStackOffset() + LRSaveSize);
1276 // All getCopyFromReg ops must precede any getMemcpys to prevent the
1277 // scheduler clobbering a register before it has been copied.
1279 // 1. CopyFromReg (and load) arg & vararg registers.
1280 // 2. Chain CopyFromReg nodes into a TokenFactor.
1281 // 3. Memcpy 'byVal' args & push final InVals.
1282 // 4. Chain mem ops nodes into a TokenFactor.
1283 SmallVector<SDValue, 4> CFRegNode;
1284 SmallVector<ArgDataPair, 4> ArgData;
1285 SmallVector<SDValue, 4> MemOps;
1287 // 1a. CopyFromReg (and load) arg registers.
1288 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1290 CCValAssign &VA = ArgLocs[i];
1293 if (VA.isRegLoc()) {
1294 // Arguments passed in registers
1295 EVT RegVT = VA.getLocVT();
1296 switch (RegVT.getSimpleVT().SimpleTy) {
1300 errs() << "LowerFormalArguments Unhandled argument type: "
1301 << RegVT.getEVTString() << "\n";
1303 llvm_unreachable(nullptr);
1306 unsigned VReg = RegInfo.createVirtualRegister(&XCore::GRRegsRegClass);
1307 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1308 ArgIn = DAG.getCopyFromReg(Chain, dl, VReg, RegVT);
1309 CFRegNode.push_back(ArgIn.getValue(ArgIn->getNumValues() - 1));
1313 assert(VA.isMemLoc());
1314 // Load the argument to a virtual register
1315 unsigned ObjSize = VA.getLocVT().getSizeInBits()/8;
1316 if (ObjSize > StackSlotSize) {
1317 errs() << "LowerFormalArguments Unhandled argument type: "
1318 << EVT(VA.getLocVT()).getEVTString()
1321 // Create the frame index object for this incoming parameter...
1322 int FI = MFI.CreateFixedObject(ObjSize,
1323 LRSaveSize + VA.getLocMemOffset(),
1326 // Create the SelectionDAG nodes corresponding to a load
1327 //from this parameter
1328 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1329 ArgIn = DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
1330 MachinePointerInfo::getFixedStack(MF, FI));
1332 const ArgDataPair ADP = { ArgIn, Ins[i].Flags };
1333 ArgData.push_back(ADP);
1336 // 1b. CopyFromReg vararg registers.
1338 // Argument registers
1339 static const MCPhysReg ArgRegs[] = {
1340 XCore::R0, XCore::R1, XCore::R2, XCore::R3
1342 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
1343 unsigned FirstVAReg = CCInfo.getFirstUnallocated(ArgRegs);
1344 if (FirstVAReg < array_lengthof(ArgRegs)) {
1346 // Save remaining registers, storing higher register numbers at a higher
1348 for (int i = array_lengthof(ArgRegs) - 1; i >= (int)FirstVAReg; --i) {
1349 // Create a stack slot
1350 int FI = MFI.CreateFixedObject(4, offset, true);
1351 if (i == (int)FirstVAReg) {
1352 XFI->setVarArgsFrameIndex(FI);
1354 offset -= StackSlotSize;
1355 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1356 // Move argument from phys reg -> virt reg
1357 unsigned VReg = RegInfo.createVirtualRegister(&XCore::GRRegsRegClass);
1358 RegInfo.addLiveIn(ArgRegs[i], VReg);
1359 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1360 CFRegNode.push_back(Val.getValue(Val->getNumValues() - 1));
1361 // Move argument from virt reg -> stack
1363 DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
1364 MemOps.push_back(Store);
1367 // This will point to the next argument passed via stack.
1368 XFI->setVarArgsFrameIndex(
1369 MFI.CreateFixedObject(4, LRSaveSize + CCInfo.getNextStackOffset(),
1374 // 2. chain CopyFromReg nodes into a TokenFactor.
1375 if (!CFRegNode.empty())
1376 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, CFRegNode);
1378 // 3. Memcpy 'byVal' args & push final InVals.
1379 // Aggregates passed "byVal" need to be copied by the callee.
1380 // The callee will use a pointer to this copy, rather than the original
1382 for (SmallVectorImpl<ArgDataPair>::const_iterator ArgDI = ArgData.begin(),
1383 ArgDE = ArgData.end();
1384 ArgDI != ArgDE; ++ArgDI) {
1385 if (ArgDI->Flags.isByVal() && ArgDI->Flags.getByValSize()) {
1386 unsigned Size = ArgDI->Flags.getByValSize();
1387 unsigned Align = std::max(StackSlotSize, ArgDI->Flags.getByValAlign());
1388 // Create a new object on the stack and copy the pointee into it.
1389 int FI = MFI.CreateStackObject(Size, Align, false);
1390 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1391 InVals.push_back(FIN);
1392 MemOps.push_back(DAG.getMemcpy(Chain, dl, FIN, ArgDI->SDV,
1393 DAG.getConstant(Size, dl, MVT::i32),
1394 Align, false, false, false,
1395 MachinePointerInfo(),
1396 MachinePointerInfo()));
1398 InVals.push_back(ArgDI->SDV);
1402 // 4, chain mem ops nodes into a TokenFactor.
1403 if (!MemOps.empty()) {
1404 MemOps.push_back(Chain);
1405 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
1411 //===----------------------------------------------------------------------===//
1412 // Return Value Calling Convention Implementation
1413 //===----------------------------------------------------------------------===//
1415 bool XCoreTargetLowering::
1416 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1418 const SmallVectorImpl<ISD::OutputArg> &Outs,
1419 LLVMContext &Context) const {
1420 SmallVector<CCValAssign, 16> RVLocs;
1421 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1422 if (!CCInfo.CheckReturn(Outs, RetCC_XCore))
1424 if (CCInfo.getNextStackOffset() != 0 && isVarArg)
1430 XCoreTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
1432 const SmallVectorImpl<ISD::OutputArg> &Outs,
1433 const SmallVectorImpl<SDValue> &OutVals,
1434 const SDLoc &dl, SelectionDAG &DAG) const {
1436 XCoreFunctionInfo *XFI =
1437 DAG.getMachineFunction().getInfo<XCoreFunctionInfo>();
1438 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
1440 // CCValAssign - represent the assignment of
1441 // the return value to a location
1442 SmallVector<CCValAssign, 16> RVLocs;
1444 // CCState - Info about the registers and stack slot.
1445 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1448 // Analyze return values.
1450 CCInfo.AllocateStack(XFI->getReturnStackOffset(), 4);
1452 CCInfo.AnalyzeReturn(Outs, RetCC_XCore);
1455 SmallVector<SDValue, 4> RetOps(1, Chain);
1457 // Return on XCore is always a "retsp 0"
1458 RetOps.push_back(DAG.getConstant(0, dl, MVT::i32));
1460 SmallVector<SDValue, 4> MemOpChains;
1461 // Handle return values that must be copied to memory.
1462 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1463 CCValAssign &VA = RVLocs[i];
1466 assert(VA.isMemLoc());
1468 report_fatal_error("Can't return value from vararg function in memory");
1471 int Offset = VA.getLocMemOffset();
1472 unsigned ObjSize = VA.getLocVT().getSizeInBits() / 8;
1473 // Create the frame index object for the memory location.
1474 int FI = MFI.CreateFixedObject(ObjSize, Offset, false);
1476 // Create a SelectionDAG node corresponding to a store
1477 // to this memory location.
1478 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1479 MemOpChains.push_back(DAG.getStore(
1480 Chain, dl, OutVals[i], FIN,
1481 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)));
1484 // Transform all store nodes into one single node because
1485 // all stores are independent of each other.
1486 if (!MemOpChains.empty())
1487 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1489 // Now handle return values copied to registers.
1490 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1491 CCValAssign &VA = RVLocs[i];
1494 // Copy the result values into the output registers.
1495 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
1497 // guarantee that all emitted copies are
1498 // stuck together, avoiding something bad
1499 Flag = Chain.getValue(1);
1500 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1503 RetOps[0] = Chain; // Update chain.
1505 // Add the flag if we have it.
1507 RetOps.push_back(Flag);
1509 return DAG.getNode(XCoreISD::RETSP, dl, MVT::Other, RetOps);
1512 //===----------------------------------------------------------------------===//
1513 // Other Lowering Code
1514 //===----------------------------------------------------------------------===//
1517 XCoreTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
1518 MachineBasicBlock *BB) const {
1519 const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
1520 DebugLoc dl = MI.getDebugLoc();
1521 assert((MI.getOpcode() == XCore::SELECT_CC) &&
1522 "Unexpected instr type to insert");
1524 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
1525 // control-flow pattern. The incoming instruction knows the destination vreg
1526 // to set, the condition code register to branch on, the true/false values to
1527 // select between, and a branch opcode to use.
1528 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1529 MachineFunction::iterator It = ++BB->getIterator();
1534 // cmpTY ccX, r1, r2
1536 // fallthrough --> copy0MBB
1537 MachineBasicBlock *thisMBB = BB;
1538 MachineFunction *F = BB->getParent();
1539 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1540 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1541 F->insert(It, copy0MBB);
1542 F->insert(It, sinkMBB);
1544 // Transfer the remainder of BB and its successor edges to sinkMBB.
1545 sinkMBB->splice(sinkMBB->begin(), BB,
1546 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1547 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1549 // Next, add the true and fallthrough blocks as its successors.
1550 BB->addSuccessor(copy0MBB);
1551 BB->addSuccessor(sinkMBB);
1553 BuildMI(BB, dl, TII.get(XCore::BRFT_lru6))
1554 .addReg(MI.getOperand(1).getReg())
1558 // %FalseValue = ...
1559 // # fallthrough to sinkMBB
1562 // Update machine-CFG edges
1563 BB->addSuccessor(sinkMBB);
1566 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1569 BuildMI(*BB, BB->begin(), dl, TII.get(XCore::PHI), MI.getOperand(0).getReg())
1570 .addReg(MI.getOperand(3).getReg())
1572 .addReg(MI.getOperand(2).getReg())
1575 MI.eraseFromParent(); // The pseudo instruction is gone now.
1579 //===----------------------------------------------------------------------===//
1580 // Target Optimization Hooks
1581 //===----------------------------------------------------------------------===//
1583 SDValue XCoreTargetLowering::PerformDAGCombine(SDNode *N,
1584 DAGCombinerInfo &DCI) const {
1585 SelectionDAG &DAG = DCI.DAG;
1587 switch (N->getOpcode()) {
1589 case ISD::INTRINSIC_VOID:
1590 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
1591 case Intrinsic::xcore_outt:
1592 case Intrinsic::xcore_outct:
1593 case Intrinsic::xcore_chkct: {
1594 SDValue OutVal = N->getOperand(3);
1595 // These instructions ignore the high bits.
1596 if (OutVal.hasOneUse()) {
1597 unsigned BitWidth = OutVal.getValueSizeInBits();
1598 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, 8);
1600 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1601 !DCI.isBeforeLegalizeOps());
1602 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1603 if (TLI.ShrinkDemandedConstant(OutVal, DemandedMask, TLO) ||
1604 TLI.SimplifyDemandedBits(OutVal, DemandedMask, Known, TLO))
1605 DCI.CommitTargetLoweringOpt(TLO);
1609 case Intrinsic::xcore_setpt: {
1610 SDValue Time = N->getOperand(3);
1611 // This instruction ignores the high bits.
1612 if (Time.hasOneUse()) {
1613 unsigned BitWidth = Time.getValueSizeInBits();
1614 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, 16);
1616 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1617 !DCI.isBeforeLegalizeOps());
1618 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1619 if (TLI.ShrinkDemandedConstant(Time, DemandedMask, TLO) ||
1620 TLI.SimplifyDemandedBits(Time, DemandedMask, Known, TLO))
1621 DCI.CommitTargetLoweringOpt(TLO);
1627 case XCoreISD::LADD: {
1628 SDValue N0 = N->getOperand(0);
1629 SDValue N1 = N->getOperand(1);
1630 SDValue N2 = N->getOperand(2);
1631 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1632 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1633 EVT VT = N0.getValueType();
1635 // canonicalize constant to RHS
1637 return DAG.getNode(XCoreISD::LADD, dl, DAG.getVTList(VT, VT), N1, N0, N2);
1639 // fold (ladd 0, 0, x) -> 0, x & 1
1640 if (N0C && N0C->isNullValue() && N1C && N1C->isNullValue()) {
1641 SDValue Carry = DAG.getConstant(0, dl, VT);
1642 SDValue Result = DAG.getNode(ISD::AND, dl, VT, N2,
1643 DAG.getConstant(1, dl, VT));
1644 SDValue Ops[] = { Result, Carry };
1645 return DAG.getMergeValues(Ops, dl);
1648 // fold (ladd x, 0, y) -> 0, add x, y iff carry is unused and y has only the
1650 if (N1C && N1C->isNullValue() && N->hasNUsesOfValue(0, 1)) {
1651 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
1652 VT.getSizeInBits() - 1);
1653 KnownBits Known = DAG.computeKnownBits(N2);
1654 if ((Known.Zero & Mask) == Mask) {
1655 SDValue Carry = DAG.getConstant(0, dl, VT);
1656 SDValue Result = DAG.getNode(ISD::ADD, dl, VT, N0, N2);
1657 SDValue Ops[] = { Result, Carry };
1658 return DAG.getMergeValues(Ops, dl);
1663 case XCoreISD::LSUB: {
1664 SDValue N0 = N->getOperand(0);
1665 SDValue N1 = N->getOperand(1);
1666 SDValue N2 = N->getOperand(2);
1667 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1668 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1669 EVT VT = N0.getValueType();
1671 // fold (lsub 0, 0, x) -> x, -x iff x has only the low bit set
1672 if (N0C && N0C->isNullValue() && N1C && N1C->isNullValue()) {
1673 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
1674 VT.getSizeInBits() - 1);
1675 KnownBits Known = DAG.computeKnownBits(N2);
1676 if ((Known.Zero & Mask) == Mask) {
1677 SDValue Borrow = N2;
1678 SDValue Result = DAG.getNode(ISD::SUB, dl, VT,
1679 DAG.getConstant(0, dl, VT), N2);
1680 SDValue Ops[] = { Result, Borrow };
1681 return DAG.getMergeValues(Ops, dl);
1685 // fold (lsub x, 0, y) -> 0, sub x, y iff borrow is unused and y has only the
1687 if (N1C && N1C->isNullValue() && N->hasNUsesOfValue(0, 1)) {
1688 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
1689 VT.getSizeInBits() - 1);
1690 KnownBits Known = DAG.computeKnownBits(N2);
1691 if ((Known.Zero & Mask) == Mask) {
1692 SDValue Borrow = DAG.getConstant(0, dl, VT);
1693 SDValue Result = DAG.getNode(ISD::SUB, dl, VT, N0, N2);
1694 SDValue Ops[] = { Result, Borrow };
1695 return DAG.getMergeValues(Ops, dl);
1700 case XCoreISD::LMUL: {
1701 SDValue N0 = N->getOperand(0);
1702 SDValue N1 = N->getOperand(1);
1703 SDValue N2 = N->getOperand(2);
1704 SDValue N3 = N->getOperand(3);
1705 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1706 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1707 EVT VT = N0.getValueType();
1708 // Canonicalize multiplicative constant to RHS. If both multiplicative
1709 // operands are constant canonicalize smallest to RHS.
1710 if ((N0C && !N1C) ||
1711 (N0C && N1C && N0C->getZExtValue() < N1C->getZExtValue()))
1712 return DAG.getNode(XCoreISD::LMUL, dl, DAG.getVTList(VT, VT),
1716 if (N1C && N1C->isNullValue()) {
1717 // If the high result is unused fold to add(a, b)
1718 if (N->hasNUsesOfValue(0, 0)) {
1719 SDValue Lo = DAG.getNode(ISD::ADD, dl, VT, N2, N3);
1720 SDValue Ops[] = { Lo, Lo };
1721 return DAG.getMergeValues(Ops, dl);
1723 // Otherwise fold to ladd(a, b, 0)
1725 DAG.getNode(XCoreISD::LADD, dl, DAG.getVTList(VT, VT), N2, N3, N1);
1726 SDValue Carry(Result.getNode(), 1);
1727 SDValue Ops[] = { Carry, Result };
1728 return DAG.getMergeValues(Ops, dl);
1733 // Fold 32 bit expressions such as add(add(mul(x,y),a),b) ->
1734 // lmul(x, y, a, b). The high result of lmul will be ignored.
1735 // This is only profitable if the intermediate results are unused
1737 SDValue Mul0, Mul1, Addend0, Addend1;
1738 if (N->getValueType(0) == MVT::i32 &&
1739 isADDADDMUL(SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, true)) {
1740 SDValue Ignored = DAG.getNode(XCoreISD::LMUL, dl,
1741 DAG.getVTList(MVT::i32, MVT::i32), Mul0,
1742 Mul1, Addend0, Addend1);
1743 SDValue Result(Ignored.getNode(), 1);
1746 APInt HighMask = APInt::getHighBitsSet(64, 32);
1747 // Fold 64 bit expression such as add(add(mul(x,y),a),b) ->
1748 // lmul(x, y, a, b) if all operands are zero-extended. We do this
1749 // before type legalization as it is messy to match the operands after
1751 if (N->getValueType(0) == MVT::i64 &&
1752 isADDADDMUL(SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, false) &&
1753 DAG.MaskedValueIsZero(Mul0, HighMask) &&
1754 DAG.MaskedValueIsZero(Mul1, HighMask) &&
1755 DAG.MaskedValueIsZero(Addend0, HighMask) &&
1756 DAG.MaskedValueIsZero(Addend1, HighMask)) {
1757 SDValue Mul0L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1758 Mul0, DAG.getConstant(0, dl, MVT::i32));
1759 SDValue Mul1L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1760 Mul1, DAG.getConstant(0, dl, MVT::i32));
1761 SDValue Addend0L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1762 Addend0, DAG.getConstant(0, dl, MVT::i32));
1763 SDValue Addend1L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1764 Addend1, DAG.getConstant(0, dl, MVT::i32));
1765 SDValue Hi = DAG.getNode(XCoreISD::LMUL, dl,
1766 DAG.getVTList(MVT::i32, MVT::i32), Mul0L, Mul1L,
1767 Addend0L, Addend1L);
1768 SDValue Lo(Hi.getNode(), 1);
1769 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
1774 // Replace unaligned store of unaligned load with memmove.
1775 StoreSDNode *ST = cast<StoreSDNode>(N);
1776 if (!DCI.isBeforeLegalize() ||
1777 allowsMisalignedMemoryAccesses(ST->getMemoryVT(),
1778 ST->getAddressSpace(),
1779 ST->getAlignment()) ||
1780 ST->isVolatile() || ST->isIndexed()) {
1783 SDValue Chain = ST->getChain();
1785 unsigned StoreBits = ST->getMemoryVT().getStoreSizeInBits();
1786 assert((StoreBits % 8) == 0 &&
1787 "Store size in bits must be a multiple of 8");
1788 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(
1789 ST->getMemoryVT().getTypeForEVT(*DCI.DAG.getContext()));
1790 unsigned Alignment = ST->getAlignment();
1791 if (Alignment >= ABIAlignment) {
1795 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(ST->getValue())) {
1796 if (LD->hasNUsesOfValue(1, 0) && ST->getMemoryVT() == LD->getMemoryVT() &&
1797 LD->getAlignment() == Alignment &&
1798 !LD->isVolatile() && !LD->isIndexed() &&
1799 Chain.reachesChainWithoutSideEffects(SDValue(LD, 1))) {
1800 bool isTail = isInTailCallPosition(DAG, ST, Chain);
1801 return DAG.getMemmove(Chain, dl, ST->getBasePtr(),
1803 DAG.getConstant(StoreBits/8, dl, MVT::i32),
1804 Alignment, false, isTail, ST->getPointerInfo(),
1805 LD->getPointerInfo());
1814 void XCoreTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
1816 const APInt &DemandedElts,
1817 const SelectionDAG &DAG,
1818 unsigned Depth) const {
1820 switch (Op.getOpcode()) {
1822 case XCoreISD::LADD:
1823 case XCoreISD::LSUB:
1824 if (Op.getResNo() == 1) {
1825 // Top bits of carry / borrow are clear.
1826 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1827 Known.getBitWidth() - 1);
1830 case ISD::INTRINSIC_W_CHAIN:
1832 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1834 case Intrinsic::xcore_getts:
1835 // High bits are known to be zero.
1836 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1837 Known.getBitWidth() - 16);
1839 case Intrinsic::xcore_int:
1840 case Intrinsic::xcore_inct:
1841 // High bits are known to be zero.
1842 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1843 Known.getBitWidth() - 8);
1845 case Intrinsic::xcore_testct:
1846 // Result is either 0 or 1.
1847 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1848 Known.getBitWidth() - 1);
1850 case Intrinsic::xcore_testwct:
1851 // Result is in the range 0 - 4.
1852 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1853 Known.getBitWidth() - 3);
1861 //===----------------------------------------------------------------------===//
1862 // Addressing mode description hooks
1863 //===----------------------------------------------------------------------===//
1865 static inline bool isImmUs(int64_t val)
1867 return (val >= 0 && val <= 11);
1870 static inline bool isImmUs2(int64_t val)
1872 return (val%2 == 0 && isImmUs(val/2));
1875 static inline bool isImmUs4(int64_t val)
1877 return (val%4 == 0 && isImmUs(val/4));
1880 /// isLegalAddressingMode - Return true if the addressing mode represented
1881 /// by AM is legal for this target, for a load/store of the specified type.
1882 bool XCoreTargetLowering::isLegalAddressingMode(const DataLayout &DL,
1883 const AddrMode &AM, Type *Ty,
1885 Instruction *I) const {
1886 if (Ty->getTypeID() == Type::VoidTyID)
1887 return AM.Scale == 0 && isImmUs(AM.BaseOffs) && isImmUs4(AM.BaseOffs);
1889 unsigned Size = DL.getTypeAllocSize(Ty);
1891 return Size >= 4 && !AM.HasBaseReg && AM.Scale == 0 &&
1898 if (AM.Scale == 0) {
1899 return isImmUs(AM.BaseOffs);
1902 return AM.Scale == 1 && AM.BaseOffs == 0;
1906 if (AM.Scale == 0) {
1907 return isImmUs2(AM.BaseOffs);
1910 return AM.Scale == 2 && AM.BaseOffs == 0;
1913 if (AM.Scale == 0) {
1914 return isImmUs4(AM.BaseOffs);
1917 return AM.Scale == 4 && AM.BaseOffs == 0;
1921 //===----------------------------------------------------------------------===//
1922 // XCore Inline Assembly Support
1923 //===----------------------------------------------------------------------===//
1925 std::pair<unsigned, const TargetRegisterClass *>
1926 XCoreTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
1927 StringRef Constraint,
1929 if (Constraint.size() == 1) {
1930 switch (Constraint[0]) {
1933 return std::make_pair(0U, &XCore::GRRegsRegClass);
1936 // Use the default implementation in TargetLowering to convert the register
1937 // constraint into a member of a register class.
1938 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);