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)
407 DAG.computeKnownBits(Value, Known);
408 return Known.countMinTrailingZeros() >= 2;
411 SDValue XCoreTargetLowering::
412 LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
413 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
414 LoadSDNode *LD = cast<LoadSDNode>(Op);
415 assert(LD->getExtensionType() == ISD::NON_EXTLOAD &&
416 "Unexpected extension type");
417 assert(LD->getMemoryVT() == MVT::i32 && "Unexpected load EVT");
418 if (allowsMisalignedMemoryAccesses(LD->getMemoryVT(),
419 LD->getAddressSpace(),
423 auto &TD = DAG.getDataLayout();
424 unsigned ABIAlignment = TD.getABITypeAlignment(
425 LD->getMemoryVT().getTypeForEVT(*DAG.getContext()));
426 // Leave aligned load alone.
427 if (LD->getAlignment() >= ABIAlignment)
430 SDValue Chain = LD->getChain();
431 SDValue BasePtr = LD->getBasePtr();
434 if (!LD->isVolatile()) {
435 const GlobalValue *GV;
437 if (DAG.isBaseWithConstantOffset(BasePtr) &&
438 isWordAligned(BasePtr->getOperand(0), DAG)) {
439 SDValue NewBasePtr = BasePtr->getOperand(0);
440 Offset = cast<ConstantSDNode>(BasePtr->getOperand(1))->getSExtValue();
441 return lowerLoadWordFromAlignedBasePlusOffset(DL, Chain, NewBasePtr,
444 if (TLI.isGAPlusOffset(BasePtr.getNode(), GV, Offset) &&
445 MinAlign(GV->getAlignment(), 4) == 4) {
446 SDValue NewBasePtr = DAG.getGlobalAddress(GV, DL,
447 BasePtr->getValueType(0));
448 return lowerLoadWordFromAlignedBasePlusOffset(DL, Chain, NewBasePtr,
453 if (LD->getAlignment() == 2) {
455 DAG.getExtLoad(ISD::ZEXTLOAD, DL, MVT::i32, Chain, BasePtr,
456 LD->getPointerInfo(), MVT::i16,
457 /* Alignment = */ 2, LD->getMemOperand()->getFlags());
458 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
459 DAG.getConstant(2, DL, MVT::i32));
461 DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain, HighAddr,
462 LD->getPointerInfo().getWithOffset(2), MVT::i16,
463 /* Alignment = */ 2, LD->getMemOperand()->getFlags());
464 SDValue HighShifted = DAG.getNode(ISD::SHL, DL, MVT::i32, High,
465 DAG.getConstant(16, DL, MVT::i32));
466 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, Low, HighShifted);
467 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Low.getValue(1),
469 SDValue Ops[] = { Result, Chain };
470 return DAG.getMergeValues(Ops, DL);
473 // Lower to a call to __misaligned_load(BasePtr).
474 Type *IntPtrTy = TD.getIntPtrType(*DAG.getContext());
475 TargetLowering::ArgListTy Args;
476 TargetLowering::ArgListEntry Entry;
479 Entry.Node = BasePtr;
480 Args.push_back(Entry);
482 TargetLowering::CallLoweringInfo CLI(DAG);
483 CLI.setDebugLoc(DL).setChain(Chain).setLibCallee(
484 CallingConv::C, IntPtrTy,
485 DAG.getExternalSymbol("__misaligned_load",
486 getPointerTy(DAG.getDataLayout())),
489 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
490 SDValue Ops[] = { CallResult.first, CallResult.second };
491 return DAG.getMergeValues(Ops, DL);
494 SDValue XCoreTargetLowering::
495 LowerSTORE(SDValue Op, SelectionDAG &DAG) const
497 StoreSDNode *ST = cast<StoreSDNode>(Op);
498 assert(!ST->isTruncatingStore() && "Unexpected store type");
499 assert(ST->getMemoryVT() == MVT::i32 && "Unexpected store EVT");
500 if (allowsMisalignedMemoryAccesses(ST->getMemoryVT(),
501 ST->getAddressSpace(),
502 ST->getAlignment())) {
505 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(
506 ST->getMemoryVT().getTypeForEVT(*DAG.getContext()));
507 // Leave aligned store alone.
508 if (ST->getAlignment() >= ABIAlignment) {
511 SDValue Chain = ST->getChain();
512 SDValue BasePtr = ST->getBasePtr();
513 SDValue Value = ST->getValue();
516 if (ST->getAlignment() == 2) {
518 SDValue High = DAG.getNode(ISD::SRL, dl, MVT::i32, Value,
519 DAG.getConstant(16, dl, MVT::i32));
520 SDValue StoreLow = DAG.getTruncStore(
521 Chain, dl, Low, BasePtr, ST->getPointerInfo(), MVT::i16,
522 /* Alignment = */ 2, ST->getMemOperand()->getFlags());
523 SDValue HighAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, BasePtr,
524 DAG.getConstant(2, dl, MVT::i32));
525 SDValue StoreHigh = DAG.getTruncStore(
526 Chain, dl, High, HighAddr, ST->getPointerInfo().getWithOffset(2),
527 MVT::i16, /* Alignment = */ 2, ST->getMemOperand()->getFlags());
528 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, StoreLow, StoreHigh);
531 // Lower to a call to __misaligned_store(BasePtr, Value).
532 Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
533 TargetLowering::ArgListTy Args;
534 TargetLowering::ArgListEntry Entry;
537 Entry.Node = BasePtr;
538 Args.push_back(Entry);
541 Args.push_back(Entry);
543 TargetLowering::CallLoweringInfo CLI(DAG);
544 CLI.setDebugLoc(dl).setChain(Chain).setCallee(
545 CallingConv::C, Type::getVoidTy(*DAG.getContext()),
546 DAG.getExternalSymbol("__misaligned_store",
547 getPointerTy(DAG.getDataLayout())),
550 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
551 return CallResult.second;
554 SDValue XCoreTargetLowering::
555 LowerSMUL_LOHI(SDValue Op, SelectionDAG &DAG) const
557 assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::SMUL_LOHI &&
558 "Unexpected operand to lower!");
560 SDValue LHS = Op.getOperand(0);
561 SDValue RHS = Op.getOperand(1);
562 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
563 SDValue Hi = DAG.getNode(XCoreISD::MACCS, dl,
564 DAG.getVTList(MVT::i32, MVT::i32), Zero, Zero,
566 SDValue Lo(Hi.getNode(), 1);
567 SDValue Ops[] = { Lo, Hi };
568 return DAG.getMergeValues(Ops, dl);
571 SDValue XCoreTargetLowering::
572 LowerUMUL_LOHI(SDValue Op, SelectionDAG &DAG) const
574 assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::UMUL_LOHI &&
575 "Unexpected operand to lower!");
577 SDValue LHS = Op.getOperand(0);
578 SDValue RHS = Op.getOperand(1);
579 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
580 SDValue Hi = DAG.getNode(XCoreISD::LMUL, dl,
581 DAG.getVTList(MVT::i32, MVT::i32), LHS, RHS,
583 SDValue Lo(Hi.getNode(), 1);
584 SDValue Ops[] = { Lo, Hi };
585 return DAG.getMergeValues(Ops, dl);
588 /// isADDADDMUL - Return whether Op is in a form that is equivalent to
589 /// add(add(mul(x,y),a),b). If requireIntermediatesHaveOneUse is true then
590 /// each intermediate result in the calculation must also have a single use.
591 /// If the Op is in the correct form the constituent parts are written to Mul0,
592 /// Mul1, Addend0 and Addend1.
594 isADDADDMUL(SDValue Op, SDValue &Mul0, SDValue &Mul1, SDValue &Addend0,
595 SDValue &Addend1, bool requireIntermediatesHaveOneUse)
597 if (Op.getOpcode() != ISD::ADD)
599 SDValue N0 = Op.getOperand(0);
600 SDValue N1 = Op.getOperand(1);
603 if (N0.getOpcode() == ISD::ADD) {
606 } else if (N1.getOpcode() == ISD::ADD) {
612 if (requireIntermediatesHaveOneUse && !AddOp.hasOneUse())
614 if (OtherOp.getOpcode() == ISD::MUL) {
615 // add(add(a,b),mul(x,y))
616 if (requireIntermediatesHaveOneUse && !OtherOp.hasOneUse())
618 Mul0 = OtherOp.getOperand(0);
619 Mul1 = OtherOp.getOperand(1);
620 Addend0 = AddOp.getOperand(0);
621 Addend1 = AddOp.getOperand(1);
624 if (AddOp.getOperand(0).getOpcode() == ISD::MUL) {
625 // add(add(mul(x,y),a),b)
626 if (requireIntermediatesHaveOneUse && !AddOp.getOperand(0).hasOneUse())
628 Mul0 = AddOp.getOperand(0).getOperand(0);
629 Mul1 = AddOp.getOperand(0).getOperand(1);
630 Addend0 = AddOp.getOperand(1);
634 if (AddOp.getOperand(1).getOpcode() == ISD::MUL) {
635 // add(add(a,mul(x,y)),b)
636 if (requireIntermediatesHaveOneUse && !AddOp.getOperand(1).hasOneUse())
638 Mul0 = AddOp.getOperand(1).getOperand(0);
639 Mul1 = AddOp.getOperand(1).getOperand(1);
640 Addend0 = AddOp.getOperand(0);
647 SDValue XCoreTargetLowering::
648 TryExpandADDWithMul(SDNode *N, SelectionDAG &DAG) const
652 if (N->getOperand(0).getOpcode() == ISD::MUL) {
653 Mul = N->getOperand(0);
654 Other = N->getOperand(1);
655 } else if (N->getOperand(1).getOpcode() == ISD::MUL) {
656 Mul = N->getOperand(1);
657 Other = N->getOperand(0);
662 SDValue LL, RL, AddendL, AddendH;
663 LL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
664 Mul.getOperand(0), DAG.getConstant(0, dl, MVT::i32));
665 RL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
666 Mul.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
667 AddendL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
668 Other, DAG.getConstant(0, dl, MVT::i32));
669 AddendH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
670 Other, DAG.getConstant(1, dl, MVT::i32));
671 APInt HighMask = APInt::getHighBitsSet(64, 32);
672 unsigned LHSSB = DAG.ComputeNumSignBits(Mul.getOperand(0));
673 unsigned RHSSB = DAG.ComputeNumSignBits(Mul.getOperand(1));
674 if (DAG.MaskedValueIsZero(Mul.getOperand(0), HighMask) &&
675 DAG.MaskedValueIsZero(Mul.getOperand(1), HighMask)) {
676 // The inputs are both zero-extended.
677 SDValue Hi = DAG.getNode(XCoreISD::MACCU, dl,
678 DAG.getVTList(MVT::i32, MVT::i32), AddendH,
680 SDValue Lo(Hi.getNode(), 1);
681 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
683 if (LHSSB > 32 && RHSSB > 32) {
684 // The inputs are both sign-extended.
685 SDValue Hi = DAG.getNode(XCoreISD::MACCS, dl,
686 DAG.getVTList(MVT::i32, MVT::i32), AddendH,
688 SDValue Lo(Hi.getNode(), 1);
689 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
692 LH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
693 Mul.getOperand(0), DAG.getConstant(1, dl, MVT::i32));
694 RH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
695 Mul.getOperand(1), DAG.getConstant(1, dl, MVT::i32));
696 SDValue Hi = DAG.getNode(XCoreISD::MACCU, dl,
697 DAG.getVTList(MVT::i32, MVT::i32), AddendH,
699 SDValue Lo(Hi.getNode(), 1);
700 RH = DAG.getNode(ISD::MUL, dl, MVT::i32, LL, RH);
701 LH = DAG.getNode(ISD::MUL, dl, MVT::i32, LH, RL);
702 Hi = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, RH);
703 Hi = DAG.getNode(ISD::ADD, dl, MVT::i32, Hi, LH);
704 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
707 SDValue XCoreTargetLowering::
708 ExpandADDSUB(SDNode *N, SelectionDAG &DAG) const
710 assert(N->getValueType(0) == MVT::i64 &&
711 (N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) &&
712 "Unknown operand to lower!");
714 if (N->getOpcode() == ISD::ADD)
715 if (SDValue Result = TryExpandADDWithMul(N, DAG))
720 // Extract components
721 SDValue LHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
723 DAG.getConstant(0, dl, MVT::i32));
724 SDValue LHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
726 DAG.getConstant(1, dl, MVT::i32));
727 SDValue RHSL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
729 DAG.getConstant(0, dl, MVT::i32));
730 SDValue RHSH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
732 DAG.getConstant(1, dl, MVT::i32));
735 unsigned Opcode = (N->getOpcode() == ISD::ADD) ? XCoreISD::LADD :
737 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
738 SDValue Lo = DAG.getNode(Opcode, dl, DAG.getVTList(MVT::i32, MVT::i32),
740 SDValue Carry(Lo.getNode(), 1);
742 SDValue Hi = DAG.getNode(Opcode, dl, DAG.getVTList(MVT::i32, MVT::i32),
744 SDValue Ignored(Hi.getNode(), 1);
746 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
749 SDValue XCoreTargetLowering::
750 LowerVAARG(SDValue Op, SelectionDAG &DAG) const
752 // Whist llvm does not support aggregate varargs we can ignore
753 // the possibility of the ValueType being an implicit byVal vararg.
754 SDNode *Node = Op.getNode();
755 EVT VT = Node->getValueType(0); // not an aggregate
756 SDValue InChain = Node->getOperand(0);
757 SDValue VAListPtr = Node->getOperand(1);
758 EVT PtrVT = VAListPtr.getValueType();
759 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
762 DAG.getLoad(PtrVT, dl, InChain, VAListPtr, MachinePointerInfo(SV));
763 // Increment the pointer, VAList, to the next vararg
764 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAList,
765 DAG.getIntPtrConstant(VT.getSizeInBits() / 8,
767 // Store the incremented VAList to the legalized pointer
768 InChain = DAG.getStore(VAList.getValue(1), dl, nextPtr, VAListPtr,
769 MachinePointerInfo(SV));
770 // Load the actual argument out of the pointer VAList
771 return DAG.getLoad(VT, dl, InChain, VAList, MachinePointerInfo());
774 SDValue XCoreTargetLowering::
775 LowerVASTART(SDValue Op, SelectionDAG &DAG) const
778 // vastart stores the address of the VarArgsFrameIndex slot into the
779 // memory location argument
780 MachineFunction &MF = DAG.getMachineFunction();
781 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
782 SDValue Addr = DAG.getFrameIndex(XFI->getVarArgsFrameIndex(), MVT::i32);
783 return DAG.getStore(Op.getOperand(0), dl, Addr, Op.getOperand(1),
784 MachinePointerInfo());
787 SDValue XCoreTargetLowering::LowerFRAMEADDR(SDValue Op,
788 SelectionDAG &DAG) const {
789 // This nodes represent llvm.frameaddress on the DAG.
790 // It takes one operand, the index of the frame address to return.
791 // An index of zero corresponds to the current function's frame address.
792 // An index of one to the parent's frame address, and so on.
793 // Depths > 0 not supported yet!
794 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
797 MachineFunction &MF = DAG.getMachineFunction();
798 const TargetRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
799 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op),
800 RegInfo->getFrameRegister(MF), MVT::i32);
803 SDValue XCoreTargetLowering::
804 LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const {
805 // This nodes represent llvm.returnaddress on the DAG.
806 // It takes one operand, the index of the return address to return.
807 // An index of zero corresponds to the current function's return address.
808 // An index of one to the parent's return address, and so on.
809 // Depths > 0 not supported yet!
810 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
813 MachineFunction &MF = DAG.getMachineFunction();
814 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
815 int FI = XFI->createLRSpillSlot(MF);
816 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
817 return DAG.getLoad(getPointerTy(DAG.getDataLayout()), SDLoc(Op),
818 DAG.getEntryNode(), FIN,
819 MachinePointerInfo::getFixedStack(MF, FI));
822 SDValue XCoreTargetLowering::
823 LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const {
824 // This node represents offset from frame pointer to first on-stack argument.
825 // This is needed for correct stack adjustment during unwind.
826 // However, we don't know the offset until after the frame has be finalised.
827 // This is done during the XCoreFTAOElim pass.
828 return DAG.getNode(XCoreISD::FRAME_TO_ARGS_OFFSET, SDLoc(Op), MVT::i32);
831 SDValue XCoreTargetLowering::
832 LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
833 // OUTCHAIN = EH_RETURN(INCHAIN, OFFSET, HANDLER)
834 // This node represents 'eh_return' gcc dwarf builtin, which is used to
835 // return from exception. The general meaning is: adjust stack by OFFSET and
836 // pass execution to HANDLER.
837 MachineFunction &MF = DAG.getMachineFunction();
838 SDValue Chain = Op.getOperand(0);
839 SDValue Offset = Op.getOperand(1);
840 SDValue Handler = Op.getOperand(2);
843 // Absolute SP = (FP + FrameToArgs) + Offset
844 const TargetRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
845 SDValue Stack = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
846 RegInfo->getFrameRegister(MF), MVT::i32);
847 SDValue FrameToArgs = DAG.getNode(XCoreISD::FRAME_TO_ARGS_OFFSET, dl,
849 Stack = DAG.getNode(ISD::ADD, dl, MVT::i32, Stack, FrameToArgs);
850 Stack = DAG.getNode(ISD::ADD, dl, MVT::i32, Stack, Offset);
852 // R0=ExceptionPointerRegister R1=ExceptionSelectorRegister
853 // which leaves 2 caller saved registers, R2 & R3 for us to use.
854 unsigned StackReg = XCore::R2;
855 unsigned HandlerReg = XCore::R3;
857 SDValue OutChains[] = {
858 DAG.getCopyToReg(Chain, dl, StackReg, Stack),
859 DAG.getCopyToReg(Chain, dl, HandlerReg, Handler)
862 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
864 return DAG.getNode(XCoreISD::EH_RETURN, dl, MVT::Other, Chain,
865 DAG.getRegister(StackReg, MVT::i32),
866 DAG.getRegister(HandlerReg, MVT::i32));
870 SDValue XCoreTargetLowering::
871 LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const {
872 return Op.getOperand(0);
875 SDValue XCoreTargetLowering::
876 LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const {
877 SDValue Chain = Op.getOperand(0);
878 SDValue Trmp = Op.getOperand(1); // trampoline
879 SDValue FPtr = Op.getOperand(2); // nested function
880 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
882 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
885 // LDAPF_u10 r11, nest
886 // LDW_2rus r11, r11[0]
887 // STWSP_ru6 r11, sp[0]
888 // LDAPF_u10 r11, fptr
889 // LDW_2rus r11, r11[0]
895 SDValue OutChains[5];
901 DAG.getStore(Chain, dl, DAG.getConstant(0x0a3cd805, dl, MVT::i32), Addr,
902 MachinePointerInfo(TrmpAddr));
904 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
905 DAG.getConstant(4, dl, MVT::i32));
907 DAG.getStore(Chain, dl, DAG.getConstant(0xd80456c0, dl, MVT::i32), Addr,
908 MachinePointerInfo(TrmpAddr, 4));
910 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
911 DAG.getConstant(8, dl, MVT::i32));
913 DAG.getStore(Chain, dl, DAG.getConstant(0x27fb0a3c, dl, MVT::i32), Addr,
914 MachinePointerInfo(TrmpAddr, 8));
916 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
917 DAG.getConstant(12, dl, MVT::i32));
919 DAG.getStore(Chain, dl, Nest, Addr, MachinePointerInfo(TrmpAddr, 12));
921 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
922 DAG.getConstant(16, dl, MVT::i32));
924 DAG.getStore(Chain, dl, FPtr, Addr, MachinePointerInfo(TrmpAddr, 16));
926 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
929 SDValue XCoreTargetLowering::
930 LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
932 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
934 case Intrinsic::xcore_crc8:
935 EVT VT = Op.getValueType();
937 DAG.getNode(XCoreISD::CRC8, DL, DAG.getVTList(VT, VT),
938 Op.getOperand(1), Op.getOperand(2) , Op.getOperand(3));
939 SDValue Crc(Data.getNode(), 1);
940 SDValue Results[] = { Crc, Data };
941 return DAG.getMergeValues(Results, DL);
946 SDValue XCoreTargetLowering::
947 LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const {
949 return DAG.getNode(XCoreISD::MEMBARRIER, DL, MVT::Other, Op.getOperand(0));
952 SDValue XCoreTargetLowering::
953 LowerATOMIC_LOAD(SDValue Op, SelectionDAG &DAG) const {
954 AtomicSDNode *N = cast<AtomicSDNode>(Op);
955 assert(N->getOpcode() == ISD::ATOMIC_LOAD && "Bad Atomic OP");
956 assert((N->getOrdering() == AtomicOrdering::Unordered ||
957 N->getOrdering() == AtomicOrdering::Monotonic) &&
958 "setInsertFencesForAtomic(true) expects unordered / monotonic");
959 if (N->getMemoryVT() == MVT::i32) {
960 if (N->getAlignment() < 4)
961 report_fatal_error("atomic load must be aligned");
962 return DAG.getLoad(getPointerTy(DAG.getDataLayout()), SDLoc(Op),
963 N->getChain(), N->getBasePtr(), N->getPointerInfo(),
964 N->getAlignment(), N->getMemOperand()->getFlags(),
965 N->getAAInfo(), N->getRanges());
967 if (N->getMemoryVT() == MVT::i16) {
968 if (N->getAlignment() < 2)
969 report_fatal_error("atomic load must be aligned");
970 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), MVT::i32, N->getChain(),
971 N->getBasePtr(), N->getPointerInfo(), MVT::i16,
972 N->getAlignment(), N->getMemOperand()->getFlags(),
975 if (N->getMemoryVT() == MVT::i8)
976 return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), MVT::i32, N->getChain(),
977 N->getBasePtr(), N->getPointerInfo(), MVT::i8,
978 N->getAlignment(), N->getMemOperand()->getFlags(),
983 SDValue XCoreTargetLowering::
984 LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) const {
985 AtomicSDNode *N = cast<AtomicSDNode>(Op);
986 assert(N->getOpcode() == ISD::ATOMIC_STORE && "Bad Atomic OP");
987 assert((N->getOrdering() == AtomicOrdering::Unordered ||
988 N->getOrdering() == AtomicOrdering::Monotonic) &&
989 "setInsertFencesForAtomic(true) expects unordered / monotonic");
990 if (N->getMemoryVT() == MVT::i32) {
991 if (N->getAlignment() < 4)
992 report_fatal_error("atomic store must be aligned");
993 return DAG.getStore(N->getChain(), SDLoc(Op), N->getVal(), N->getBasePtr(),
994 N->getPointerInfo(), N->getAlignment(),
995 N->getMemOperand()->getFlags(), N->getAAInfo());
997 if (N->getMemoryVT() == MVT::i16) {
998 if (N->getAlignment() < 2)
999 report_fatal_error("atomic store must be aligned");
1000 return DAG.getTruncStore(N->getChain(), SDLoc(Op), N->getVal(),
1001 N->getBasePtr(), N->getPointerInfo(), MVT::i16,
1002 N->getAlignment(), N->getMemOperand()->getFlags(),
1005 if (N->getMemoryVT() == MVT::i8)
1006 return DAG.getTruncStore(N->getChain(), SDLoc(Op), N->getVal(),
1007 N->getBasePtr(), N->getPointerInfo(), MVT::i8,
1008 N->getAlignment(), N->getMemOperand()->getFlags(),
1013 //===----------------------------------------------------------------------===//
1014 // Calling Convention Implementation
1015 //===----------------------------------------------------------------------===//
1017 #include "XCoreGenCallingConv.inc"
1019 //===----------------------------------------------------------------------===//
1020 // Call Calling Convention Implementation
1021 //===----------------------------------------------------------------------===//
1023 /// XCore call implementation
1025 XCoreTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1026 SmallVectorImpl<SDValue> &InVals) const {
1027 SelectionDAG &DAG = CLI.DAG;
1029 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1030 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1031 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1032 SDValue Chain = CLI.Chain;
1033 SDValue Callee = CLI.Callee;
1034 bool &isTailCall = CLI.IsTailCall;
1035 CallingConv::ID CallConv = CLI.CallConv;
1036 bool isVarArg = CLI.IsVarArg;
1038 // XCore target does not yet support tail call optimization.
1041 // For now, only CallingConv::C implemented
1045 report_fatal_error("Unsupported calling convention");
1046 case CallingConv::Fast:
1047 case CallingConv::C:
1048 return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall,
1049 Outs, OutVals, Ins, dl, DAG, InVals);
1053 /// LowerCallResult - Lower the result values of a call into the
1054 /// appropriate copies out of appropriate physical registers / memory locations.
1055 static SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
1056 const SmallVectorImpl<CCValAssign> &RVLocs,
1057 const SDLoc &dl, SelectionDAG &DAG,
1058 SmallVectorImpl<SDValue> &InVals) {
1059 SmallVector<std::pair<int, unsigned>, 4> ResultMemLocs;
1060 // Copy results out of physical registers.
1061 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1062 const CCValAssign &VA = RVLocs[i];
1063 if (VA.isRegLoc()) {
1064 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getValVT(),
1065 InFlag).getValue(1);
1066 InFlag = Chain.getValue(2);
1067 InVals.push_back(Chain.getValue(0));
1069 assert(VA.isMemLoc());
1070 ResultMemLocs.push_back(std::make_pair(VA.getLocMemOffset(),
1072 // Reserve space for this result.
1073 InVals.push_back(SDValue());
1077 // Copy results out of memory.
1078 SmallVector<SDValue, 4> MemOpChains;
1079 for (unsigned i = 0, e = ResultMemLocs.size(); i != e; ++i) {
1080 int offset = ResultMemLocs[i].first;
1081 unsigned index = ResultMemLocs[i].second;
1082 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other);
1083 SDValue Ops[] = { Chain, DAG.getConstant(offset / 4, dl, MVT::i32) };
1084 SDValue load = DAG.getNode(XCoreISD::LDWSP, dl, VTs, Ops);
1085 InVals[index] = load;
1086 MemOpChains.push_back(load.getValue(1));
1089 // Transform all loads nodes into one single node because
1090 // all load nodes are independent of each other.
1091 if (!MemOpChains.empty())
1092 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1097 /// LowerCCCCallTo - functions arguments are copied from virtual
1098 /// regs to (physical regs)/(stack frame), CALLSEQ_START and
1099 /// CALLSEQ_END are emitted.
1100 /// TODO: isTailCall, sret.
1101 SDValue XCoreTargetLowering::LowerCCCCallTo(
1102 SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg,
1103 bool isTailCall, const SmallVectorImpl<ISD::OutputArg> &Outs,
1104 const SmallVectorImpl<SDValue> &OutVals,
1105 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1106 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1108 // Analyze operands of the call, assigning locations to each operand.
1109 SmallVector<CCValAssign, 16> ArgLocs;
1110 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1113 // The ABI dictates there should be one stack slot available to the callee
1114 // on function entry (for saving lr).
1115 CCInfo.AllocateStack(4, 4);
1117 CCInfo.AnalyzeCallOperands(Outs, CC_XCore);
1119 SmallVector<CCValAssign, 16> RVLocs;
1120 // Analyze return values to determine the number of bytes of stack required.
1121 CCState RetCCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1123 RetCCInfo.AllocateStack(CCInfo.getNextStackOffset(), 4);
1124 RetCCInfo.AnalyzeCallResult(Ins, RetCC_XCore);
1126 // Get a count of how many bytes are to be pushed on the stack.
1127 unsigned NumBytes = RetCCInfo.getNextStackOffset();
1128 auto PtrVT = getPointerTy(DAG.getDataLayout());
1130 Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl);
1132 SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass;
1133 SmallVector<SDValue, 12> MemOpChains;
1135 // Walk the register/memloc assignments, inserting copies/loads.
1136 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1137 CCValAssign &VA = ArgLocs[i];
1138 SDValue Arg = OutVals[i];
1140 // Promote the value if needed.
1141 switch (VA.getLocInfo()) {
1142 default: llvm_unreachable("Unknown loc info!");
1143 case CCValAssign::Full: break;
1144 case CCValAssign::SExt:
1145 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1147 case CCValAssign::ZExt:
1148 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1150 case CCValAssign::AExt:
1151 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1155 // Arguments that can be passed on register must be kept at
1156 // RegsToPass vector
1157 if (VA.isRegLoc()) {
1158 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1160 assert(VA.isMemLoc());
1162 int Offset = VA.getLocMemOffset();
1164 MemOpChains.push_back(DAG.getNode(XCoreISD::STWSP, dl, MVT::Other,
1166 DAG.getConstant(Offset/4, dl,
1171 // Transform all store nodes into one single node because
1172 // all store nodes are independent of each other.
1173 if (!MemOpChains.empty())
1174 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1176 // Build a sequence of copy-to-reg nodes chained together with token
1177 // chain and flag operands which copy the outgoing args into registers.
1178 // The InFlag in necessary since all emitted instructions must be
1181 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1182 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1183 RegsToPass[i].second, InFlag);
1184 InFlag = Chain.getValue(1);
1187 // If the callee is a GlobalAddress node (quite common, every direct call is)
1188 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1189 // Likewise ExternalSymbol -> TargetExternalSymbol.
1190 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1191 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, MVT::i32);
1192 else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee))
1193 Callee = DAG.getTargetExternalSymbol(E->getSymbol(), MVT::i32);
1195 // XCoreBranchLink = #chain, #target_address, #opt_in_flags...
1196 // = Chain, Callee, Reg#1, Reg#2, ...
1198 // Returns a chain & a flag for retval copy to use.
1199 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1200 SmallVector<SDValue, 8> Ops;
1201 Ops.push_back(Chain);
1202 Ops.push_back(Callee);
1204 // Add argument registers to the end of the list so that they are
1205 // known live into the call.
1206 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1207 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1208 RegsToPass[i].second.getValueType()));
1210 if (InFlag.getNode())
1211 Ops.push_back(InFlag);
1213 Chain = DAG.getNode(XCoreISD::BL, dl, NodeTys, Ops);
1214 InFlag = Chain.getValue(1);
1216 // Create the CALLSEQ_END node.
1217 Chain = DAG.getCALLSEQ_END(Chain, DAG.getConstant(NumBytes, dl, PtrVT, true),
1218 DAG.getConstant(0, dl, PtrVT, true), InFlag, dl);
1219 InFlag = Chain.getValue(1);
1221 // Handle result values, copying them out of physregs into vregs that we
1223 return LowerCallResult(Chain, InFlag, RVLocs, dl, DAG, InVals);
1226 //===----------------------------------------------------------------------===//
1227 // Formal Arguments Calling Convention Implementation
1228 //===----------------------------------------------------------------------===//
1231 struct ArgDataPair { SDValue SDV; ISD::ArgFlagsTy Flags; };
1234 /// XCore formal arguments implementation
1235 SDValue XCoreTargetLowering::LowerFormalArguments(
1236 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1237 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1238 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1242 report_fatal_error("Unsupported calling convention");
1243 case CallingConv::C:
1244 case CallingConv::Fast:
1245 return LowerCCCArguments(Chain, CallConv, isVarArg,
1246 Ins, dl, DAG, InVals);
1250 /// LowerCCCArguments - transform physical registers into
1251 /// virtual registers and generate load operations for
1252 /// arguments places on the stack.
1254 SDValue XCoreTargetLowering::LowerCCCArguments(
1255 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1256 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1257 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1258 MachineFunction &MF = DAG.getMachineFunction();
1259 MachineFrameInfo &MFI = MF.getFrameInfo();
1260 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1261 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
1263 // Assign locations to all of the incoming arguments.
1264 SmallVector<CCValAssign, 16> ArgLocs;
1265 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1268 CCInfo.AnalyzeFormalArguments(Ins, CC_XCore);
1270 unsigned StackSlotSize = XCoreFrameLowering::stackSlotSize();
1272 unsigned LRSaveSize = StackSlotSize;
1275 XFI->setReturnStackOffset(CCInfo.getNextStackOffset() + LRSaveSize);
1277 // All getCopyFromReg ops must precede any getMemcpys to prevent the
1278 // scheduler clobbering a register before it has been copied.
1280 // 1. CopyFromReg (and load) arg & vararg registers.
1281 // 2. Chain CopyFromReg nodes into a TokenFactor.
1282 // 3. Memcpy 'byVal' args & push final InVals.
1283 // 4. Chain mem ops nodes into a TokenFactor.
1284 SmallVector<SDValue, 4> CFRegNode;
1285 SmallVector<ArgDataPair, 4> ArgData;
1286 SmallVector<SDValue, 4> MemOps;
1288 // 1a. CopyFromReg (and load) arg registers.
1289 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1291 CCValAssign &VA = ArgLocs[i];
1294 if (VA.isRegLoc()) {
1295 // Arguments passed in registers
1296 EVT RegVT = VA.getLocVT();
1297 switch (RegVT.getSimpleVT().SimpleTy) {
1301 errs() << "LowerFormalArguments Unhandled argument type: "
1302 << RegVT.getEVTString() << "\n";
1304 llvm_unreachable(nullptr);
1307 unsigned VReg = RegInfo.createVirtualRegister(&XCore::GRRegsRegClass);
1308 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1309 ArgIn = DAG.getCopyFromReg(Chain, dl, VReg, RegVT);
1310 CFRegNode.push_back(ArgIn.getValue(ArgIn->getNumValues() - 1));
1314 assert(VA.isMemLoc());
1315 // Load the argument to a virtual register
1316 unsigned ObjSize = VA.getLocVT().getSizeInBits()/8;
1317 if (ObjSize > StackSlotSize) {
1318 errs() << "LowerFormalArguments Unhandled argument type: "
1319 << EVT(VA.getLocVT()).getEVTString()
1322 // Create the frame index object for this incoming parameter...
1323 int FI = MFI.CreateFixedObject(ObjSize,
1324 LRSaveSize + VA.getLocMemOffset(),
1327 // Create the SelectionDAG nodes corresponding to a load
1328 //from this parameter
1329 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1330 ArgIn = DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
1331 MachinePointerInfo::getFixedStack(MF, FI));
1333 const ArgDataPair ADP = { ArgIn, Ins[i].Flags };
1334 ArgData.push_back(ADP);
1337 // 1b. CopyFromReg vararg registers.
1339 // Argument registers
1340 static const MCPhysReg ArgRegs[] = {
1341 XCore::R0, XCore::R1, XCore::R2, XCore::R3
1343 XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>();
1344 unsigned FirstVAReg = CCInfo.getFirstUnallocated(ArgRegs);
1345 if (FirstVAReg < array_lengthof(ArgRegs)) {
1347 // Save remaining registers, storing higher register numbers at a higher
1349 for (int i = array_lengthof(ArgRegs) - 1; i >= (int)FirstVAReg; --i) {
1350 // Create a stack slot
1351 int FI = MFI.CreateFixedObject(4, offset, true);
1352 if (i == (int)FirstVAReg) {
1353 XFI->setVarArgsFrameIndex(FI);
1355 offset -= StackSlotSize;
1356 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1357 // Move argument from phys reg -> virt reg
1358 unsigned VReg = RegInfo.createVirtualRegister(&XCore::GRRegsRegClass);
1359 RegInfo.addLiveIn(ArgRegs[i], VReg);
1360 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1361 CFRegNode.push_back(Val.getValue(Val->getNumValues() - 1));
1362 // Move argument from virt reg -> stack
1364 DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
1365 MemOps.push_back(Store);
1368 // This will point to the next argument passed via stack.
1369 XFI->setVarArgsFrameIndex(
1370 MFI.CreateFixedObject(4, LRSaveSize + CCInfo.getNextStackOffset(),
1375 // 2. chain CopyFromReg nodes into a TokenFactor.
1376 if (!CFRegNode.empty())
1377 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, CFRegNode);
1379 // 3. Memcpy 'byVal' args & push final InVals.
1380 // Aggregates passed "byVal" need to be copied by the callee.
1381 // The callee will use a pointer to this copy, rather than the original
1383 for (SmallVectorImpl<ArgDataPair>::const_iterator ArgDI = ArgData.begin(),
1384 ArgDE = ArgData.end();
1385 ArgDI != ArgDE; ++ArgDI) {
1386 if (ArgDI->Flags.isByVal() && ArgDI->Flags.getByValSize()) {
1387 unsigned Size = ArgDI->Flags.getByValSize();
1388 unsigned Align = std::max(StackSlotSize, ArgDI->Flags.getByValAlign());
1389 // Create a new object on the stack and copy the pointee into it.
1390 int FI = MFI.CreateStackObject(Size, Align, false);
1391 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1392 InVals.push_back(FIN);
1393 MemOps.push_back(DAG.getMemcpy(Chain, dl, FIN, ArgDI->SDV,
1394 DAG.getConstant(Size, dl, MVT::i32),
1395 Align, false, false, false,
1396 MachinePointerInfo(),
1397 MachinePointerInfo()));
1399 InVals.push_back(ArgDI->SDV);
1403 // 4, chain mem ops nodes into a TokenFactor.
1404 if (!MemOps.empty()) {
1405 MemOps.push_back(Chain);
1406 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
1412 //===----------------------------------------------------------------------===//
1413 // Return Value Calling Convention Implementation
1414 //===----------------------------------------------------------------------===//
1416 bool XCoreTargetLowering::
1417 CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
1419 const SmallVectorImpl<ISD::OutputArg> &Outs,
1420 LLVMContext &Context) const {
1421 SmallVector<CCValAssign, 16> RVLocs;
1422 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1423 if (!CCInfo.CheckReturn(Outs, RetCC_XCore))
1425 if (CCInfo.getNextStackOffset() != 0 && isVarArg)
1431 XCoreTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
1433 const SmallVectorImpl<ISD::OutputArg> &Outs,
1434 const SmallVectorImpl<SDValue> &OutVals,
1435 const SDLoc &dl, SelectionDAG &DAG) const {
1437 XCoreFunctionInfo *XFI =
1438 DAG.getMachineFunction().getInfo<XCoreFunctionInfo>();
1439 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
1441 // CCValAssign - represent the assignment of
1442 // the return value to a location
1443 SmallVector<CCValAssign, 16> RVLocs;
1445 // CCState - Info about the registers and stack slot.
1446 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1449 // Analyze return values.
1451 CCInfo.AllocateStack(XFI->getReturnStackOffset(), 4);
1453 CCInfo.AnalyzeReturn(Outs, RetCC_XCore);
1456 SmallVector<SDValue, 4> RetOps(1, Chain);
1458 // Return on XCore is always a "retsp 0"
1459 RetOps.push_back(DAG.getConstant(0, dl, MVT::i32));
1461 SmallVector<SDValue, 4> MemOpChains;
1462 // Handle return values that must be copied to memory.
1463 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1464 CCValAssign &VA = RVLocs[i];
1467 assert(VA.isMemLoc());
1469 report_fatal_error("Can't return value from vararg function in memory");
1472 int Offset = VA.getLocMemOffset();
1473 unsigned ObjSize = VA.getLocVT().getSizeInBits() / 8;
1474 // Create the frame index object for the memory location.
1475 int FI = MFI.CreateFixedObject(ObjSize, Offset, false);
1477 // Create a SelectionDAG node corresponding to a store
1478 // to this memory location.
1479 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1480 MemOpChains.push_back(DAG.getStore(
1481 Chain, dl, OutVals[i], FIN,
1482 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)));
1485 // Transform all store nodes into one single node because
1486 // all stores are independent of each other.
1487 if (!MemOpChains.empty())
1488 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1490 // Now handle return values copied to registers.
1491 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1492 CCValAssign &VA = RVLocs[i];
1495 // Copy the result values into the output registers.
1496 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
1498 // guarantee that all emitted copies are
1499 // stuck together, avoiding something bad
1500 Flag = Chain.getValue(1);
1501 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1504 RetOps[0] = Chain; // Update chain.
1506 // Add the flag if we have it.
1508 RetOps.push_back(Flag);
1510 return DAG.getNode(XCoreISD::RETSP, dl, MVT::Other, RetOps);
1513 //===----------------------------------------------------------------------===//
1514 // Other Lowering Code
1515 //===----------------------------------------------------------------------===//
1518 XCoreTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
1519 MachineBasicBlock *BB) const {
1520 const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
1521 DebugLoc dl = MI.getDebugLoc();
1522 assert((MI.getOpcode() == XCore::SELECT_CC) &&
1523 "Unexpected instr type to insert");
1525 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
1526 // control-flow pattern. The incoming instruction knows the destination vreg
1527 // to set, the condition code register to branch on, the true/false values to
1528 // select between, and a branch opcode to use.
1529 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1530 MachineFunction::iterator It = ++BB->getIterator();
1535 // cmpTY ccX, r1, r2
1537 // fallthrough --> copy0MBB
1538 MachineBasicBlock *thisMBB = BB;
1539 MachineFunction *F = BB->getParent();
1540 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1541 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1542 F->insert(It, copy0MBB);
1543 F->insert(It, sinkMBB);
1545 // Transfer the remainder of BB and its successor edges to sinkMBB.
1546 sinkMBB->splice(sinkMBB->begin(), BB,
1547 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1548 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1550 // Next, add the true and fallthrough blocks as its successors.
1551 BB->addSuccessor(copy0MBB);
1552 BB->addSuccessor(sinkMBB);
1554 BuildMI(BB, dl, TII.get(XCore::BRFT_lru6))
1555 .addReg(MI.getOperand(1).getReg())
1559 // %FalseValue = ...
1560 // # fallthrough to sinkMBB
1563 // Update machine-CFG edges
1564 BB->addSuccessor(sinkMBB);
1567 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1570 BuildMI(*BB, BB->begin(), dl, TII.get(XCore::PHI), MI.getOperand(0).getReg())
1571 .addReg(MI.getOperand(3).getReg())
1573 .addReg(MI.getOperand(2).getReg())
1576 MI.eraseFromParent(); // The pseudo instruction is gone now.
1580 //===----------------------------------------------------------------------===//
1581 // Target Optimization Hooks
1582 //===----------------------------------------------------------------------===//
1584 SDValue XCoreTargetLowering::PerformDAGCombine(SDNode *N,
1585 DAGCombinerInfo &DCI) const {
1586 SelectionDAG &DAG = DCI.DAG;
1588 switch (N->getOpcode()) {
1590 case ISD::INTRINSIC_VOID:
1591 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
1592 case Intrinsic::xcore_outt:
1593 case Intrinsic::xcore_outct:
1594 case Intrinsic::xcore_chkct: {
1595 SDValue OutVal = N->getOperand(3);
1596 // These instructions ignore the high bits.
1597 if (OutVal.hasOneUse()) {
1598 unsigned BitWidth = OutVal.getValueSizeInBits();
1599 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, 8);
1601 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1602 !DCI.isBeforeLegalizeOps());
1603 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1604 if (TLI.ShrinkDemandedConstant(OutVal, DemandedMask, TLO) ||
1605 TLI.SimplifyDemandedBits(OutVal, DemandedMask, Known, TLO))
1606 DCI.CommitTargetLoweringOpt(TLO);
1610 case Intrinsic::xcore_setpt: {
1611 SDValue Time = N->getOperand(3);
1612 // This instruction ignores the high bits.
1613 if (Time.hasOneUse()) {
1614 unsigned BitWidth = Time.getValueSizeInBits();
1615 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, 16);
1617 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1618 !DCI.isBeforeLegalizeOps());
1619 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1620 if (TLI.ShrinkDemandedConstant(Time, DemandedMask, TLO) ||
1621 TLI.SimplifyDemandedBits(Time, DemandedMask, Known, TLO))
1622 DCI.CommitTargetLoweringOpt(TLO);
1628 case XCoreISD::LADD: {
1629 SDValue N0 = N->getOperand(0);
1630 SDValue N1 = N->getOperand(1);
1631 SDValue N2 = N->getOperand(2);
1632 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1633 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1634 EVT VT = N0.getValueType();
1636 // canonicalize constant to RHS
1638 return DAG.getNode(XCoreISD::LADD, dl, DAG.getVTList(VT, VT), N1, N0, N2);
1640 // fold (ladd 0, 0, x) -> 0, x & 1
1641 if (N0C && N0C->isNullValue() && N1C && N1C->isNullValue()) {
1642 SDValue Carry = DAG.getConstant(0, dl, VT);
1643 SDValue Result = DAG.getNode(ISD::AND, dl, VT, N2,
1644 DAG.getConstant(1, dl, VT));
1645 SDValue Ops[] = { Result, Carry };
1646 return DAG.getMergeValues(Ops, dl);
1649 // fold (ladd x, 0, y) -> 0, add x, y iff carry is unused and y has only the
1651 if (N1C && N1C->isNullValue() && N->hasNUsesOfValue(0, 1)) {
1653 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
1654 VT.getSizeInBits() - 1);
1655 DAG.computeKnownBits(N2, Known);
1656 if ((Known.Zero & Mask) == Mask) {
1657 SDValue Carry = DAG.getConstant(0, dl, VT);
1658 SDValue Result = DAG.getNode(ISD::ADD, dl, VT, N0, N2);
1659 SDValue Ops[] = { Result, Carry };
1660 return DAG.getMergeValues(Ops, dl);
1665 case XCoreISD::LSUB: {
1666 SDValue N0 = N->getOperand(0);
1667 SDValue N1 = N->getOperand(1);
1668 SDValue N2 = N->getOperand(2);
1669 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1670 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1671 EVT VT = N0.getValueType();
1673 // fold (lsub 0, 0, x) -> x, -x iff x has only the low bit set
1674 if (N0C && N0C->isNullValue() && N1C && N1C->isNullValue()) {
1676 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
1677 VT.getSizeInBits() - 1);
1678 DAG.computeKnownBits(N2, Known);
1679 if ((Known.Zero & Mask) == Mask) {
1680 SDValue Borrow = N2;
1681 SDValue Result = DAG.getNode(ISD::SUB, dl, VT,
1682 DAG.getConstant(0, dl, VT), N2);
1683 SDValue Ops[] = { Result, Borrow };
1684 return DAG.getMergeValues(Ops, dl);
1688 // fold (lsub x, 0, y) -> 0, sub x, y iff borrow is unused and y has only the
1690 if (N1C && N1C->isNullValue() && N->hasNUsesOfValue(0, 1)) {
1692 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
1693 VT.getSizeInBits() - 1);
1694 DAG.computeKnownBits(N2, Known);
1695 if ((Known.Zero & Mask) == Mask) {
1696 SDValue Borrow = DAG.getConstant(0, dl, VT);
1697 SDValue Result = DAG.getNode(ISD::SUB, dl, VT, N0, N2);
1698 SDValue Ops[] = { Result, Borrow };
1699 return DAG.getMergeValues(Ops, dl);
1704 case XCoreISD::LMUL: {
1705 SDValue N0 = N->getOperand(0);
1706 SDValue N1 = N->getOperand(1);
1707 SDValue N2 = N->getOperand(2);
1708 SDValue N3 = N->getOperand(3);
1709 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1710 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1711 EVT VT = N0.getValueType();
1712 // Canonicalize multiplicative constant to RHS. If both multiplicative
1713 // operands are constant canonicalize smallest to RHS.
1714 if ((N0C && !N1C) ||
1715 (N0C && N1C && N0C->getZExtValue() < N1C->getZExtValue()))
1716 return DAG.getNode(XCoreISD::LMUL, dl, DAG.getVTList(VT, VT),
1720 if (N1C && N1C->isNullValue()) {
1721 // If the high result is unused fold to add(a, b)
1722 if (N->hasNUsesOfValue(0, 0)) {
1723 SDValue Lo = DAG.getNode(ISD::ADD, dl, VT, N2, N3);
1724 SDValue Ops[] = { Lo, Lo };
1725 return DAG.getMergeValues(Ops, dl);
1727 // Otherwise fold to ladd(a, b, 0)
1729 DAG.getNode(XCoreISD::LADD, dl, DAG.getVTList(VT, VT), N2, N3, N1);
1730 SDValue Carry(Result.getNode(), 1);
1731 SDValue Ops[] = { Carry, Result };
1732 return DAG.getMergeValues(Ops, dl);
1737 // Fold 32 bit expressions such as add(add(mul(x,y),a),b) ->
1738 // lmul(x, y, a, b). The high result of lmul will be ignored.
1739 // This is only profitable if the intermediate results are unused
1741 SDValue Mul0, Mul1, Addend0, Addend1;
1742 if (N->getValueType(0) == MVT::i32 &&
1743 isADDADDMUL(SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, true)) {
1744 SDValue Ignored = DAG.getNode(XCoreISD::LMUL, dl,
1745 DAG.getVTList(MVT::i32, MVT::i32), Mul0,
1746 Mul1, Addend0, Addend1);
1747 SDValue Result(Ignored.getNode(), 1);
1750 APInt HighMask = APInt::getHighBitsSet(64, 32);
1751 // Fold 64 bit expression such as add(add(mul(x,y),a),b) ->
1752 // lmul(x, y, a, b) if all operands are zero-extended. We do this
1753 // before type legalization as it is messy to match the operands after
1755 if (N->getValueType(0) == MVT::i64 &&
1756 isADDADDMUL(SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, false) &&
1757 DAG.MaskedValueIsZero(Mul0, HighMask) &&
1758 DAG.MaskedValueIsZero(Mul1, HighMask) &&
1759 DAG.MaskedValueIsZero(Addend0, HighMask) &&
1760 DAG.MaskedValueIsZero(Addend1, HighMask)) {
1761 SDValue Mul0L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1762 Mul0, DAG.getConstant(0, dl, MVT::i32));
1763 SDValue Mul1L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1764 Mul1, DAG.getConstant(0, dl, MVT::i32));
1765 SDValue Addend0L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1766 Addend0, DAG.getConstant(0, dl, MVT::i32));
1767 SDValue Addend1L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
1768 Addend1, DAG.getConstant(0, dl, MVT::i32));
1769 SDValue Hi = DAG.getNode(XCoreISD::LMUL, dl,
1770 DAG.getVTList(MVT::i32, MVT::i32), Mul0L, Mul1L,
1771 Addend0L, Addend1L);
1772 SDValue Lo(Hi.getNode(), 1);
1773 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
1778 // Replace unaligned store of unaligned load with memmove.
1779 StoreSDNode *ST = cast<StoreSDNode>(N);
1780 if (!DCI.isBeforeLegalize() ||
1781 allowsMisalignedMemoryAccesses(ST->getMemoryVT(),
1782 ST->getAddressSpace(),
1783 ST->getAlignment()) ||
1784 ST->isVolatile() || ST->isIndexed()) {
1787 SDValue Chain = ST->getChain();
1789 unsigned StoreBits = ST->getMemoryVT().getStoreSizeInBits();
1790 assert((StoreBits % 8) == 0 &&
1791 "Store size in bits must be a multiple of 8");
1792 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(
1793 ST->getMemoryVT().getTypeForEVT(*DCI.DAG.getContext()));
1794 unsigned Alignment = ST->getAlignment();
1795 if (Alignment >= ABIAlignment) {
1799 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(ST->getValue())) {
1800 if (LD->hasNUsesOfValue(1, 0) && ST->getMemoryVT() == LD->getMemoryVT() &&
1801 LD->getAlignment() == Alignment &&
1802 !LD->isVolatile() && !LD->isIndexed() &&
1803 Chain.reachesChainWithoutSideEffects(SDValue(LD, 1))) {
1804 bool isTail = isInTailCallPosition(DAG, ST, Chain);
1805 return DAG.getMemmove(Chain, dl, ST->getBasePtr(),
1807 DAG.getConstant(StoreBits/8, dl, MVT::i32),
1808 Alignment, false, isTail, ST->getPointerInfo(),
1809 LD->getPointerInfo());
1818 void XCoreTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
1820 const APInt &DemandedElts,
1821 const SelectionDAG &DAG,
1822 unsigned Depth) const {
1824 switch (Op.getOpcode()) {
1826 case XCoreISD::LADD:
1827 case XCoreISD::LSUB:
1828 if (Op.getResNo() == 1) {
1829 // Top bits of carry / borrow are clear.
1830 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1831 Known.getBitWidth() - 1);
1834 case ISD::INTRINSIC_W_CHAIN:
1836 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1838 case Intrinsic::xcore_getts:
1839 // High bits are known to be zero.
1840 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1841 Known.getBitWidth() - 16);
1843 case Intrinsic::xcore_int:
1844 case Intrinsic::xcore_inct:
1845 // High bits are known to be zero.
1846 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1847 Known.getBitWidth() - 8);
1849 case Intrinsic::xcore_testct:
1850 // Result is either 0 or 1.
1851 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1852 Known.getBitWidth() - 1);
1854 case Intrinsic::xcore_testwct:
1855 // Result is in the range 0 - 4.
1856 Known.Zero = APInt::getHighBitsSet(Known.getBitWidth(),
1857 Known.getBitWidth() - 3);
1865 //===----------------------------------------------------------------------===//
1866 // Addressing mode description hooks
1867 //===----------------------------------------------------------------------===//
1869 static inline bool isImmUs(int64_t val)
1871 return (val >= 0 && val <= 11);
1874 static inline bool isImmUs2(int64_t val)
1876 return (val%2 == 0 && isImmUs(val/2));
1879 static inline bool isImmUs4(int64_t val)
1881 return (val%4 == 0 && isImmUs(val/4));
1884 /// isLegalAddressingMode - Return true if the addressing mode represented
1885 /// by AM is legal for this target, for a load/store of the specified type.
1886 bool XCoreTargetLowering::isLegalAddressingMode(const DataLayout &DL,
1887 const AddrMode &AM, Type *Ty,
1889 Instruction *I) const {
1890 if (Ty->getTypeID() == Type::VoidTyID)
1891 return AM.Scale == 0 && isImmUs(AM.BaseOffs) && isImmUs4(AM.BaseOffs);
1893 unsigned Size = DL.getTypeAllocSize(Ty);
1895 return Size >= 4 && !AM.HasBaseReg && AM.Scale == 0 &&
1902 if (AM.Scale == 0) {
1903 return isImmUs(AM.BaseOffs);
1906 return AM.Scale == 1 && AM.BaseOffs == 0;
1910 if (AM.Scale == 0) {
1911 return isImmUs2(AM.BaseOffs);
1914 return AM.Scale == 2 && AM.BaseOffs == 0;
1917 if (AM.Scale == 0) {
1918 return isImmUs4(AM.BaseOffs);
1921 return AM.Scale == 4 && AM.BaseOffs == 0;
1925 //===----------------------------------------------------------------------===//
1926 // XCore Inline Assembly Support
1927 //===----------------------------------------------------------------------===//
1929 std::pair<unsigned, const TargetRegisterClass *>
1930 XCoreTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
1931 StringRef Constraint,
1933 if (Constraint.size() == 1) {
1934 switch (Constraint[0]) {
1937 return std::make_pair(0U, &XCore::GRRegsRegClass);
1940 // Use the default implementation in TargetLowering to convert the register
1941 // constraint into a member of a register class.
1942 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);