xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARC/ARCISelLowering.cpp (revision 7899f917b1c0ea178f1d2be0cfb452086d079d23)
1 //===- ARCISelLowering.cpp - ARC DAG Lowering Impl --------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the ARCTargetLowering class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "ARCISelLowering.h"
14 #include "ARC.h"
15 #include "ARCMachineFunctionInfo.h"
16 #include "ARCSubtarget.h"
17 #include "ARCTargetMachine.h"
18 #include "MCTargetDesc/ARCInfo.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineJumpTableInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/ValueTypes.h"
26 #include "llvm/IR/CallingConv.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/Support/Debug.h"
29 #include <algorithm>
30 
31 #define DEBUG_TYPE "arc-lower"
32 
33 using namespace llvm;
34 
35 static SDValue lowerCallResult(SDValue Chain, SDValue InGlue,
36                                const SmallVectorImpl<CCValAssign> &RVLocs,
37                                SDLoc dl, SelectionDAG &DAG,
38                                SmallVectorImpl<SDValue> &InVals);
39 
40 static ARCCC::CondCode ISDCCtoARCCC(ISD::CondCode isdCC) {
41   switch (isdCC) {
42   case ISD::SETUEQ:
43     return ARCCC::EQ;
44   case ISD::SETUGT:
45     return ARCCC::HI;
46   case ISD::SETUGE:
47     return ARCCC::HS;
48   case ISD::SETULT:
49     return ARCCC::LO;
50   case ISD::SETULE:
51     return ARCCC::LS;
52   case ISD::SETUNE:
53     return ARCCC::NE;
54   case ISD::SETEQ:
55     return ARCCC::EQ;
56   case ISD::SETGT:
57     return ARCCC::GT;
58   case ISD::SETGE:
59     return ARCCC::GE;
60   case ISD::SETLT:
61     return ARCCC::LT;
62   case ISD::SETLE:
63     return ARCCC::LE;
64   case ISD::SETNE:
65     return ARCCC::NE;
66   default:
67     llvm_unreachable("Unhandled ISDCC code.");
68   }
69 }
70 
71 void ARCTargetLowering::ReplaceNodeResults(SDNode *N,
72                                            SmallVectorImpl<SDValue> &Results,
73                                            SelectionDAG &DAG) const {
74   LLVM_DEBUG(dbgs() << "[ARC-ISEL] ReplaceNodeResults ");
75   LLVM_DEBUG(N->dump(&DAG));
76   LLVM_DEBUG(dbgs() << "; use_count=" << N->use_size() << "\n");
77 
78   switch (N->getOpcode()) {
79   case ISD::READCYCLECOUNTER:
80     if (N->getValueType(0) == MVT::i64) {
81       // We read the TIMER0 and zero-extend it to 64-bits as the intrinsic
82       // requires.
83       SDValue V =
84           DAG.getNode(ISD::READCYCLECOUNTER, SDLoc(N),
85                       DAG.getVTList(MVT::i32, MVT::Other), N->getOperand(0));
86       SDValue Op = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), MVT::i64, V);
87       Results.push_back(Op);
88       Results.push_back(V.getValue(1));
89     }
90     break;
91   default:
92     break;
93   }
94 }
95 
96 ARCTargetLowering::ARCTargetLowering(const TargetMachine &TM,
97                                      const ARCSubtarget &Subtarget)
98     : TargetLowering(TM), Subtarget(Subtarget) {
99   // Set up the register classes.
100   addRegisterClass(MVT::i32, &ARC::GPR32RegClass);
101 
102   // Compute derived properties from the register classes
103   computeRegisterProperties(Subtarget.getRegisterInfo());
104 
105   setStackPointerRegisterToSaveRestore(ARC::SP);
106 
107   setSchedulingPreference(Sched::Source);
108 
109   // Use i32 for setcc operations results (slt, sgt, ...).
110   setBooleanContents(ZeroOrOneBooleanContent);
111   setBooleanVectorContents(ZeroOrOneBooleanContent);
112 
113   for (unsigned Opc = 0; Opc < ISD::BUILTIN_OP_END; ++Opc)
114     setOperationAction(Opc, MVT::i32, Expand);
115 
116   // Operations to get us off of the ground.
117   // Basic.
118   setOperationAction(ISD::ADD, MVT::i32, Legal);
119   setOperationAction(ISD::SUB, MVT::i32, Legal);
120   setOperationAction(ISD::AND, MVT::i32, Legal);
121   setOperationAction(ISD::SMAX, MVT::i32, Legal);
122   setOperationAction(ISD::SMIN, MVT::i32, Legal);
123 
124   setOperationAction(ISD::ADDC, MVT::i32, Legal);
125   setOperationAction(ISD::ADDE, MVT::i32, Legal);
126   setOperationAction(ISD::SUBC, MVT::i32, Legal);
127   setOperationAction(ISD::SUBE, MVT::i32, Legal);
128 
129   // Need barrel shifter.
130   setOperationAction(ISD::SHL, MVT::i32, Legal);
131   setOperationAction(ISD::SRA, MVT::i32, Legal);
132   setOperationAction(ISD::SRL, MVT::i32, Legal);
133   setOperationAction(ISD::ROTR, MVT::i32, Legal);
134 
135   setOperationAction(ISD::Constant, MVT::i32, Legal);
136   setOperationAction(ISD::UNDEF, MVT::i32, Legal);
137 
138   // Need multiplier
139   setOperationAction(ISD::MUL, MVT::i32, Legal);
140   setOperationAction(ISD::MULHS, MVT::i32, Legal);
141   setOperationAction(ISD::MULHU, MVT::i32, Legal);
142   setOperationAction(ISD::LOAD, MVT::i32, Legal);
143   setOperationAction(ISD::STORE, MVT::i32, Legal);
144 
145   setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
146   setOperationAction(ISD::BR_CC, MVT::i32, Custom);
147   setOperationAction(ISD::BRCOND, MVT::Other, Expand);
148   setOperationAction(ISD::BR_JT, MVT::Other, Expand);
149   setOperationAction(ISD::JumpTable, MVT::i32, Custom);
150 
151   // Have pseudo instruction for frame addresses.
152   setOperationAction(ISD::FRAMEADDR, MVT::i32, Legal);
153   // Custom lower global addresses.
154   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
155 
156   // Expand var-args ops.
157   setOperationAction(ISD::VASTART, MVT::Other, Custom);
158   setOperationAction(ISD::VAEND, MVT::Other, Expand);
159   setOperationAction(ISD::VAARG, MVT::Other, Expand);
160   setOperationAction(ISD::VACOPY, MVT::Other, Expand);
161 
162   // Other expansions
163   setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
164   setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
165 
166   // Sign extend inreg
167   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Custom);
168 
169   // TODO: Predicate these with `options.hasBitScan() ? Legal : Expand`
170   //       when the HasBitScan predicate is available.
171   setOperationAction(ISD::CTLZ, MVT::i32, Legal);
172   setOperationAction(ISD::CTTZ, MVT::i32, Legal);
173 
174   setOperationAction(ISD::READCYCLECOUNTER, MVT::i32, Legal);
175   setOperationAction(ISD::READCYCLECOUNTER, MVT::i64,
176                      isTypeLegal(MVT::i64) ? Legal : Custom);
177 
178   setMaxAtomicSizeInBitsSupported(0);
179 }
180 
181 const char *ARCTargetLowering::getTargetNodeName(unsigned Opcode) const {
182   switch (Opcode) {
183   case ARCISD::BL:
184     return "ARCISD::BL";
185   case ARCISD::CMOV:
186     return "ARCISD::CMOV";
187   case ARCISD::CMP:
188     return "ARCISD::CMP";
189   case ARCISD::BRcc:
190     return "ARCISD::BRcc";
191   case ARCISD::RET:
192     return "ARCISD::RET";
193   case ARCISD::GAWRAPPER:
194     return "ARCISD::GAWRAPPER";
195   }
196   return nullptr;
197 }
198 
199 //===----------------------------------------------------------------------===//
200 //  Misc Lower Operation implementation
201 //===----------------------------------------------------------------------===//
202 
203 SDValue ARCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
204   SDValue LHS = Op.getOperand(0);
205   SDValue RHS = Op.getOperand(1);
206   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
207   SDValue TVal = Op.getOperand(2);
208   SDValue FVal = Op.getOperand(3);
209   SDLoc dl(Op);
210   ARCCC::CondCode ArcCC = ISDCCtoARCCC(CC);
211   assert(LHS.getValueType() == MVT::i32 && "Only know how to SELECT_CC i32");
212   SDValue Cmp = DAG.getNode(ARCISD::CMP, dl, MVT::Glue, LHS, RHS);
213   return DAG.getNode(ARCISD::CMOV, dl, TVal.getValueType(), TVal, FVal,
214                      DAG.getConstant(ArcCC, dl, MVT::i32), Cmp);
215 }
216 
217 SDValue ARCTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
218                                                   SelectionDAG &DAG) const {
219   SDValue Op0 = Op.getOperand(0);
220   SDLoc dl(Op);
221   assert(Op.getValueType() == MVT::i32 &&
222          "Unhandled target sign_extend_inreg.");
223   // These are legal
224   unsigned Width = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits();
225   if (Width == 16 || Width == 8)
226     return Op;
227   if (Width >= 32) {
228     return {};
229   }
230   SDValue LS = DAG.getNode(ISD::SHL, dl, MVT::i32, Op0,
231                            DAG.getConstant(32 - Width, dl, MVT::i32));
232   SDValue SR = DAG.getNode(ISD::SRA, dl, MVT::i32, LS,
233                            DAG.getConstant(32 - Width, dl, MVT::i32));
234   return SR;
235 }
236 
237 SDValue ARCTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
238   SDValue Chain = Op.getOperand(0);
239   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
240   SDValue LHS = Op.getOperand(2);
241   SDValue RHS = Op.getOperand(3);
242   SDValue Dest = Op.getOperand(4);
243   SDLoc dl(Op);
244   ARCCC::CondCode arcCC = ISDCCtoARCCC(CC);
245   assert(LHS.getValueType() == MVT::i32 && "Only know how to BR_CC i32");
246   return DAG.getNode(ARCISD::BRcc, dl, MVT::Other, Chain, Dest, LHS, RHS,
247                      DAG.getConstant(arcCC, dl, MVT::i32));
248 }
249 
250 SDValue ARCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
251   auto *N = cast<JumpTableSDNode>(Op);
252   SDValue GA = DAG.getTargetJumpTable(N->getIndex(), MVT::i32);
253   return DAG.getNode(ARCISD::GAWRAPPER, SDLoc(N), MVT::i32, GA);
254 }
255 
256 #include "ARCGenCallingConv.inc"
257 
258 //===----------------------------------------------------------------------===//
259 //                  Call Calling Convention Implementation
260 //===----------------------------------------------------------------------===//
261 
262 /// ARC call implementation
263 SDValue ARCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
264                                      SmallVectorImpl<SDValue> &InVals) const {
265   SelectionDAG &DAG = CLI.DAG;
266   SDLoc &dl = CLI.DL;
267   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
268   SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
269   SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
270   SDValue Chain = CLI.Chain;
271   SDValue Callee = CLI.Callee;
272   CallingConv::ID CallConv = CLI.CallConv;
273   bool IsVarArg = CLI.IsVarArg;
274   bool &IsTailCall = CLI.IsTailCall;
275 
276   IsTailCall = false; // Do not support tail calls yet.
277 
278   SmallVector<CCValAssign, 16> ArgLocs;
279   CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
280                  *DAG.getContext());
281 
282   CCInfo.AnalyzeCallOperands(Outs, CC_ARC);
283 
284   SmallVector<CCValAssign, 16> RVLocs;
285   // Analyze return values to determine the number of bytes of stack required.
286   CCState RetCCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
287                     *DAG.getContext());
288   RetCCInfo.AllocateStack(CCInfo.getStackSize(), Align(4));
289   RetCCInfo.AnalyzeCallResult(Ins, RetCC_ARC);
290 
291   // Get a count of how many bytes are to be pushed on the stack.
292   unsigned NumBytes = RetCCInfo.getStackSize();
293 
294   Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl);
295 
296   SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass;
297   SmallVector<SDValue, 12> MemOpChains;
298 
299   SDValue StackPtr;
300   // Walk the register/memloc assignments, inserting copies/loads.
301   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
302     CCValAssign &VA = ArgLocs[i];
303     SDValue Arg = OutVals[i];
304 
305     // Promote the value if needed.
306     switch (VA.getLocInfo()) {
307     default:
308       llvm_unreachable("Unknown loc info!");
309     case CCValAssign::Full:
310       break;
311     case CCValAssign::SExt:
312       Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
313       break;
314     case CCValAssign::ZExt:
315       Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
316       break;
317     case CCValAssign::AExt:
318       Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
319       break;
320     }
321 
322     // Arguments that can be passed on register must be kept at
323     // RegsToPass vector
324     if (VA.isRegLoc()) {
325       RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
326     } else {
327       assert(VA.isMemLoc() && "Must be register or memory argument.");
328       if (!StackPtr.getNode())
329         StackPtr = DAG.getCopyFromReg(Chain, dl, ARC::SP,
330                                       getPointerTy(DAG.getDataLayout()));
331       // Calculate the stack position.
332       SDValue SOffset = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
333       SDValue PtrOff = DAG.getNode(
334           ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), StackPtr, SOffset);
335 
336       SDValue Store =
337           DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo());
338       MemOpChains.push_back(Store);
339       IsTailCall = false;
340     }
341   }
342 
343   // Transform all store nodes into one single node because
344   // all store nodes are independent of each other.
345   if (!MemOpChains.empty())
346     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
347 
348   // Build a sequence of copy-to-reg nodes chained together with token
349   // chain and flag operands which copy the outgoing args into registers.
350   // The Glue in necessary since all emitted instructions must be
351   // stuck together.
352   SDValue Glue;
353   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
354     Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
355                              RegsToPass[i].second, Glue);
356     Glue = Chain.getValue(1);
357   }
358 
359   // If the callee is a GlobalAddress node (quite common, every direct call is)
360   // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
361   // Likewise ExternalSymbol -> TargetExternalSymbol.
362   bool IsDirect = true;
363   if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee))
364     Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, MVT::i32);
365   else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee))
366     Callee = DAG.getTargetExternalSymbol(E->getSymbol(), MVT::i32);
367   else
368     IsDirect = false;
369   // Branch + Link = #chain, #target_address, #opt_in_flags...
370   //             = Chain, Callee, Reg#1, Reg#2, ...
371   //
372   // Returns a chain & a glue for retval copy to use.
373   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
374   SmallVector<SDValue, 8> Ops;
375   Ops.push_back(Chain);
376   Ops.push_back(Callee);
377 
378   for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
379     Ops.push_back(DAG.getRegister(RegsToPass[i].first,
380                                   RegsToPass[i].second.getValueType()));
381 
382   // Add a register mask operand representing the call-preserved registers.
383   const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
384   const uint32_t *Mask =
385       TRI->getCallPreservedMask(DAG.getMachineFunction(), CallConv);
386   assert(Mask && "Missing call preserved mask for calling convention");
387   Ops.push_back(DAG.getRegisterMask(Mask));
388 
389   if (Glue.getNode())
390     Ops.push_back(Glue);
391 
392   Chain = DAG.getNode(IsDirect ? ARCISD::BL : ARCISD::JL, dl, NodeTys, Ops);
393   Glue = Chain.getValue(1);
394 
395   // Create the CALLSEQ_END node.
396   Chain = DAG.getCALLSEQ_END(Chain, NumBytes, 0, Glue, dl);
397   Glue = Chain.getValue(1);
398 
399   // Handle result values, copying them out of physregs into vregs that we
400   // return.
401   if (IsTailCall)
402     return Chain;
403   return lowerCallResult(Chain, Glue, RVLocs, dl, DAG, InVals);
404 }
405 
406 /// Lower the result values of a call into the appropriate copies out of
407 /// physical registers / memory locations.
408 static SDValue lowerCallResult(SDValue Chain, SDValue Glue,
409                                const SmallVectorImpl<CCValAssign> &RVLocs,
410                                SDLoc dl, SelectionDAG &DAG,
411                                SmallVectorImpl<SDValue> &InVals) {
412   SmallVector<std::pair<int, unsigned>, 4> ResultMemLocs;
413   // Copy results out of physical registers.
414   for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
415     const CCValAssign &VA = RVLocs[i];
416     if (VA.isRegLoc()) {
417       SDValue RetValue;
418       RetValue =
419           DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getValVT(), Glue);
420       Chain = RetValue.getValue(1);
421       Glue = RetValue.getValue(2);
422       InVals.push_back(RetValue);
423     } else {
424       assert(VA.isMemLoc() && "Must be memory location.");
425       ResultMemLocs.push_back(
426           std::make_pair(VA.getLocMemOffset(), InVals.size()));
427 
428       // Reserve space for this result.
429       InVals.push_back(SDValue());
430     }
431   }
432 
433   // Copy results out of memory.
434   SmallVector<SDValue, 4> MemOpChains;
435   for (unsigned i = 0, e = ResultMemLocs.size(); i != e; ++i) {
436     int Offset = ResultMemLocs[i].first;
437     unsigned Index = ResultMemLocs[i].second;
438     SDValue StackPtr = DAG.getRegister(ARC::SP, MVT::i32);
439     SDValue SpLoc = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr,
440                                 DAG.getConstant(Offset, dl, MVT::i32));
441     SDValue Load =
442         DAG.getLoad(MVT::i32, dl, Chain, SpLoc, MachinePointerInfo());
443     InVals[Index] = Load;
444     MemOpChains.push_back(Load.getValue(1));
445   }
446 
447   // Transform all loads nodes into one single node because
448   // all load nodes are independent of each other.
449   if (!MemOpChains.empty())
450     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
451 
452   return Chain;
453 }
454 
455 //===----------------------------------------------------------------------===//
456 //             Formal Arguments Calling Convention Implementation
457 //===----------------------------------------------------------------------===//
458 
459 namespace {
460 
461 struct ArgDataPair {
462   SDValue SDV;
463   ISD::ArgFlagsTy Flags;
464 };
465 
466 } // end anonymous namespace
467 
468 /// ARC formal arguments implementation
469 SDValue ARCTargetLowering::LowerFormalArguments(
470     SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
471     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
472     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
473   switch (CallConv) {
474   default:
475     llvm_unreachable("Unsupported calling convention");
476   case CallingConv::C:
477   case CallingConv::Fast:
478     return LowerCallArguments(Chain, CallConv, IsVarArg, Ins, dl, DAG, InVals);
479   }
480 }
481 
482 /// Transform physical registers into virtual registers, and generate load
483 /// operations for argument places on the stack.
484 SDValue ARCTargetLowering::LowerCallArguments(
485     SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
486     const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
487     SmallVectorImpl<SDValue> &InVals) const {
488   MachineFunction &MF = DAG.getMachineFunction();
489   MachineFrameInfo &MFI = MF.getFrameInfo();
490   MachineRegisterInfo &RegInfo = MF.getRegInfo();
491   auto *AFI = MF.getInfo<ARCFunctionInfo>();
492 
493   // Assign locations to all of the incoming arguments.
494   SmallVector<CCValAssign, 16> ArgLocs;
495   CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
496                  *DAG.getContext());
497 
498   CCInfo.AnalyzeFormalArguments(Ins, CC_ARC);
499 
500   unsigned StackSlotSize = 4;
501 
502   if (!IsVarArg)
503     AFI->setReturnStackOffset(CCInfo.getStackSize());
504 
505   // All getCopyFromReg ops must precede any getMemcpys to prevent the
506   // scheduler clobbering a register before it has been copied.
507   // The stages are:
508   // 1. CopyFromReg (and load) arg & vararg registers.
509   // 2. Chain CopyFromReg nodes into a TokenFactor.
510   // 3. Memcpy 'byVal' args & push final InVals.
511   // 4. Chain mem ops nodes into a TokenFactor.
512   SmallVector<SDValue, 4> CFRegNode;
513   SmallVector<ArgDataPair, 4> ArgData;
514   SmallVector<SDValue, 4> MemOps;
515 
516   // 1a. CopyFromReg (and load) arg registers.
517   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
518     CCValAssign &VA = ArgLocs[i];
519     SDValue ArgIn;
520 
521     if (VA.isRegLoc()) {
522       // Arguments passed in registers
523       EVT RegVT = VA.getLocVT();
524       switch (RegVT.getSimpleVT().SimpleTy) {
525       default: {
526         LLVM_DEBUG(errs() << "LowerFormalArguments Unhandled argument type: "
527                           << (unsigned)RegVT.getSimpleVT().SimpleTy << "\n");
528         llvm_unreachable("Unhandled LowerFormalArguments type.");
529       }
530       case MVT::i32:
531         unsigned VReg = RegInfo.createVirtualRegister(&ARC::GPR32RegClass);
532         RegInfo.addLiveIn(VA.getLocReg(), VReg);
533         ArgIn = DAG.getCopyFromReg(Chain, dl, VReg, RegVT);
534         CFRegNode.push_back(ArgIn.getValue(ArgIn->getNumValues() - 1));
535       }
536     } else {
537       // Only arguments passed on the stack should make it here.
538       assert(VA.isMemLoc());
539       // Load the argument to a virtual register
540       unsigned ObjSize = VA.getLocVT().getStoreSize();
541       assert((ObjSize <= StackSlotSize) && "Unhandled argument");
542 
543       // Create the frame index object for this incoming parameter...
544       int FI = MFI.CreateFixedObject(ObjSize, VA.getLocMemOffset(), true);
545 
546       // Create the SelectionDAG nodes corresponding to a load
547       // from this parameter
548       SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
549       ArgIn = DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
550                           MachinePointerInfo::getFixedStack(MF, FI));
551     }
552     const ArgDataPair ADP = {ArgIn, Ins[i].Flags};
553     ArgData.push_back(ADP);
554   }
555 
556   // 1b. CopyFromReg vararg registers.
557   if (IsVarArg) {
558     // Argument registers
559     static const MCPhysReg ArgRegs[] = {ARC::R0, ARC::R1, ARC::R2, ARC::R3,
560                                         ARC::R4, ARC::R5, ARC::R6, ARC::R7};
561     auto *AFI = MF.getInfo<ARCFunctionInfo>();
562     unsigned FirstVAReg = CCInfo.getFirstUnallocated(ArgRegs);
563     if (FirstVAReg < std::size(ArgRegs)) {
564       int Offset = 0;
565       // Save remaining registers, storing higher register numbers at a higher
566       // address
567       // There are (std::size(ArgRegs) - FirstVAReg) registers which
568       // need to be saved.
569       int VarFI = MFI.CreateFixedObject((std::size(ArgRegs) - FirstVAReg) * 4,
570                                         CCInfo.getStackSize(), true);
571       AFI->setVarArgsFrameIndex(VarFI);
572       SDValue FIN = DAG.getFrameIndex(VarFI, MVT::i32);
573       for (unsigned i = FirstVAReg; i < std::size(ArgRegs); i++) {
574         // Move argument from phys reg -> virt reg
575         unsigned VReg = RegInfo.createVirtualRegister(&ARC::GPR32RegClass);
576         RegInfo.addLiveIn(ArgRegs[i], VReg);
577         SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
578         CFRegNode.push_back(Val.getValue(Val->getNumValues() - 1));
579         SDValue VAObj = DAG.getNode(ISD::ADD, dl, MVT::i32, FIN,
580                                     DAG.getConstant(Offset, dl, MVT::i32));
581         // Move argument from virt reg -> stack
582         SDValue Store =
583             DAG.getStore(Val.getValue(1), dl, Val, VAObj, MachinePointerInfo());
584         MemOps.push_back(Store);
585         Offset += 4;
586       }
587     } else {
588       llvm_unreachable("Too many var args parameters.");
589     }
590   }
591 
592   // 2. Chain CopyFromReg nodes into a TokenFactor.
593   if (!CFRegNode.empty())
594     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, CFRegNode);
595 
596   // 3. Memcpy 'byVal' args & push final InVals.
597   // Aggregates passed "byVal" need to be copied by the callee.
598   // The callee will use a pointer to this copy, rather than the original
599   // pointer.
600   for (const auto &ArgDI : ArgData) {
601     if (ArgDI.Flags.isByVal() && ArgDI.Flags.getByValSize()) {
602       unsigned Size = ArgDI.Flags.getByValSize();
603       Align Alignment =
604           std::max(Align(StackSlotSize), ArgDI.Flags.getNonZeroByValAlign());
605       // Create a new object on the stack and copy the pointee into it.
606       int FI = MFI.CreateStackObject(Size, Alignment, false);
607       SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
608       InVals.push_back(FIN);
609       MemOps.push_back(DAG.getMemcpy(
610           Chain, dl, FIN, ArgDI.SDV, DAG.getConstant(Size, dl, MVT::i32),
611           Alignment, false, false, false, MachinePointerInfo(),
612           MachinePointerInfo()));
613     } else {
614       InVals.push_back(ArgDI.SDV);
615     }
616   }
617 
618   // 4. Chain mem ops nodes into a TokenFactor.
619   if (!MemOps.empty()) {
620     MemOps.push_back(Chain);
621     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
622   }
623 
624   return Chain;
625 }
626 
627 //===----------------------------------------------------------------------===//
628 //               Return Value Calling Convention Implementation
629 //===----------------------------------------------------------------------===//
630 
631 bool ARCTargetLowering::CanLowerReturn(
632     CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg,
633     const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
634   SmallVector<CCValAssign, 16> RVLocs;
635   CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
636   if (!CCInfo.CheckReturn(Outs, RetCC_ARC))
637     return false;
638   if (CCInfo.getStackSize() != 0 && IsVarArg)
639     return false;
640   return true;
641 }
642 
643 SDValue
644 ARCTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
645                                bool IsVarArg,
646                                const SmallVectorImpl<ISD::OutputArg> &Outs,
647                                const SmallVectorImpl<SDValue> &OutVals,
648                                const SDLoc &dl, SelectionDAG &DAG) const {
649   auto *AFI = DAG.getMachineFunction().getInfo<ARCFunctionInfo>();
650   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
651 
652   // CCValAssign - represent the assignment of
653   // the return value to a location
654   SmallVector<CCValAssign, 16> RVLocs;
655 
656   // CCState - Info about the registers and stack slot.
657   CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
658                  *DAG.getContext());
659 
660   // Analyze return values.
661   if (!IsVarArg)
662     CCInfo.AllocateStack(AFI->getReturnStackOffset(), Align(4));
663 
664   CCInfo.AnalyzeReturn(Outs, RetCC_ARC);
665 
666   SDValue Glue;
667   SmallVector<SDValue, 4> RetOps(1, Chain);
668   SmallVector<SDValue, 4> MemOpChains;
669   // Handle return values that must be copied to memory.
670   for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
671     CCValAssign &VA = RVLocs[i];
672     if (VA.isRegLoc())
673       continue;
674     assert(VA.isMemLoc());
675     if (IsVarArg) {
676       report_fatal_error("Can't return value from vararg function in memory");
677     }
678 
679     int Offset = VA.getLocMemOffset();
680     unsigned ObjSize = VA.getLocVT().getStoreSize();
681     // Create the frame index object for the memory location.
682     int FI = MFI.CreateFixedObject(ObjSize, Offset, false);
683 
684     // Create a SelectionDAG node corresponding to a store
685     // to this memory location.
686     SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
687     MemOpChains.push_back(DAG.getStore(
688         Chain, dl, OutVals[i], FIN,
689         MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)));
690   }
691 
692   // Transform all store nodes into one single node because
693   // all stores are independent of each other.
694   if (!MemOpChains.empty())
695     Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
696 
697   // Now handle return values copied to registers.
698   for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
699     CCValAssign &VA = RVLocs[i];
700     if (!VA.isRegLoc())
701       continue;
702     // Copy the result values into the output registers.
703     Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Glue);
704 
705     // guarantee that all emitted copies are
706     // stuck together, avoiding something bad
707     Glue = Chain.getValue(1);
708     RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
709   }
710 
711   RetOps[0] = Chain; // Update chain.
712 
713   // Add the glue if we have it.
714   if (Glue.getNode())
715     RetOps.push_back(Glue);
716 
717   // What to do with the RetOps?
718   return DAG.getNode(ARCISD::RET, dl, MVT::Other, RetOps);
719 }
720 
721 //===----------------------------------------------------------------------===//
722 // Target Optimization Hooks
723 //===----------------------------------------------------------------------===//
724 
725 SDValue ARCTargetLowering::PerformDAGCombine(SDNode *N,
726                                              DAGCombinerInfo &DCI) const {
727   return {};
728 }
729 
730 //===----------------------------------------------------------------------===//
731 //  Addressing mode description hooks
732 //===----------------------------------------------------------------------===//
733 
734 /// Return true if the addressing mode represented by AM is legal for this
735 /// target, for a load/store of the specified type.
736 bool ARCTargetLowering::isLegalAddressingMode(const DataLayout &DL,
737                                               const AddrMode &AM, Type *Ty,
738                                               unsigned AS,
739                                               Instruction *I) const {
740   return AM.Scale == 0;
741 }
742 
743 // Don't emit tail calls for the time being.
744 bool ARCTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
745   return false;
746 }
747 
748 SDValue ARCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
749   const ARCRegisterInfo &ARI = *Subtarget.getRegisterInfo();
750   MachineFunction &MF = DAG.getMachineFunction();
751   MachineFrameInfo &MFI = MF.getFrameInfo();
752   MFI.setFrameAddressIsTaken(true);
753 
754   EVT VT = Op.getValueType();
755   SDLoc dl(Op);
756   assert(Op.getConstantOperandVal(0) == 0 &&
757          "Only support lowering frame addr of current frame.");
758   Register FrameReg = ARI.getFrameRegister(MF);
759   return DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
760 }
761 
762 SDValue ARCTargetLowering::LowerGlobalAddress(SDValue Op,
763                                               SelectionDAG &DAG) const {
764   const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
765   const GlobalValue *GV = GN->getGlobal();
766   SDLoc dl(GN);
767   int64_t Offset = GN->getOffset();
768   SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, Offset);
769   return DAG.getNode(ARCISD::GAWRAPPER, dl, MVT::i32, GA);
770 }
771 
772 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
773   MachineFunction &MF = DAG.getMachineFunction();
774   auto *FuncInfo = MF.getInfo<ARCFunctionInfo>();
775 
776   // vastart just stores the address of the VarArgsFrameIndex slot into the
777   // memory location argument.
778   SDLoc dl(Op);
779   EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
780   SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
781   const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
782   return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
783                       MachinePointerInfo(SV));
784 }
785 
786 SDValue ARCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
787   switch (Op.getOpcode()) {
788   case ISD::GlobalAddress:
789     return LowerGlobalAddress(Op, DAG);
790   case ISD::FRAMEADDR:
791     return LowerFRAMEADDR(Op, DAG);
792   case ISD::SELECT_CC:
793     return LowerSELECT_CC(Op, DAG);
794   case ISD::BR_CC:
795     return LowerBR_CC(Op, DAG);
796   case ISD::SIGN_EXTEND_INREG:
797     return LowerSIGN_EXTEND_INREG(Op, DAG);
798   case ISD::JumpTable:
799     return LowerJumpTable(Op, DAG);
800   case ISD::VASTART:
801     return LowerVASTART(Op, DAG);
802   case ISD::READCYCLECOUNTER:
803     // As of LLVM 3.8, the lowering code insists that we customize it even
804     // though we've declared the i32 version as legal. This is because it only
805     // thinks i64 is the truly supported version. We've already converted the
806     // i64 version to a widened i32.
807     assert(Op.getSimpleValueType() == MVT::i32);
808     return Op;
809   default:
810     llvm_unreachable("unimplemented operand");
811   }
812 }
813