xref: /freebsd/contrib/llvm-project/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp (revision d56accc7c3dcc897489b6a07834763a03b9f3d68)
1 //===-- HexagonISelDAGToDAG.cpp - A dag to dag inst selector for Hexagon --===//
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 defines an instruction selector for the Hexagon target.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "HexagonISelDAGToDAG.h"
14 #include "Hexagon.h"
15 #include "HexagonISelLowering.h"
16 #include "HexagonMachineFunctionInfo.h"
17 #include "HexagonTargetMachine.h"
18 #include "llvm/CodeGen/FunctionLoweringInfo.h"
19 #include "llvm/CodeGen/MachineInstrBuilder.h"
20 #include "llvm/CodeGen/SelectionDAGISel.h"
21 #include "llvm/IR/Intrinsics.h"
22 #include "llvm/IR/IntrinsicsHexagon.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/Debug.h"
25 using namespace llvm;
26 
27 #define DEBUG_TYPE "hexagon-isel"
28 
29 static
30 cl::opt<bool>
31 EnableAddressRebalancing("isel-rebalance-addr", cl::Hidden, cl::init(true),
32   cl::desc("Rebalance address calculation trees to improve "
33           "instruction selection"));
34 
35 // Rebalance only if this allows e.g. combining a GA with an offset or
36 // factoring out a shift.
37 static
38 cl::opt<bool>
39 RebalanceOnlyForOptimizations("rebalance-only-opt", cl::Hidden, cl::init(false),
40   cl::desc("Rebalance address tree only if this allows optimizations"));
41 
42 static
43 cl::opt<bool>
44 RebalanceOnlyImbalancedTrees("rebalance-only-imbal", cl::Hidden,
45   cl::init(false), cl::desc("Rebalance address tree only if it is imbalanced"));
46 
47 static cl::opt<bool> CheckSingleUse("hexagon-isel-su", cl::Hidden,
48   cl::init(true), cl::desc("Enable checking of SDNode's single-use status"));
49 
50 //===----------------------------------------------------------------------===//
51 // Instruction Selector Implementation
52 //===----------------------------------------------------------------------===//
53 
54 #define GET_DAGISEL_BODY HexagonDAGToDAGISel
55 #include "HexagonGenDAGISel.inc"
56 
57 namespace llvm {
58 /// createHexagonISelDag - This pass converts a legalized DAG into a
59 /// Hexagon-specific DAG, ready for instruction scheduling.
60 FunctionPass *createHexagonISelDag(HexagonTargetMachine &TM,
61                                    CodeGenOpt::Level OptLevel) {
62   return new HexagonDAGToDAGISel(TM, OptLevel);
63 }
64 }
65 
66 void HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, const SDLoc &dl) {
67   SDValue Chain = LD->getChain();
68   SDValue Base = LD->getBasePtr();
69   SDValue Offset = LD->getOffset();
70   int32_t Inc = cast<ConstantSDNode>(Offset.getNode())->getSExtValue();
71   EVT LoadedVT = LD->getMemoryVT();
72   unsigned Opcode = 0;
73 
74   // Check for zero extended loads. Treat any-extend loads as zero extended
75   // loads.
76   ISD::LoadExtType ExtType = LD->getExtensionType();
77   bool IsZeroExt = (ExtType == ISD::ZEXTLOAD || ExtType == ISD::EXTLOAD);
78   bool IsValidInc = HII->isValidAutoIncImm(LoadedVT, Inc);
79 
80   assert(LoadedVT.isSimple());
81   switch (LoadedVT.getSimpleVT().SimpleTy) {
82   case MVT::i8:
83     if (IsZeroExt)
84       Opcode = IsValidInc ? Hexagon::L2_loadrub_pi : Hexagon::L2_loadrub_io;
85     else
86       Opcode = IsValidInc ? Hexagon::L2_loadrb_pi : Hexagon::L2_loadrb_io;
87     break;
88   case MVT::i16:
89     if (IsZeroExt)
90       Opcode = IsValidInc ? Hexagon::L2_loadruh_pi : Hexagon::L2_loadruh_io;
91     else
92       Opcode = IsValidInc ? Hexagon::L2_loadrh_pi : Hexagon::L2_loadrh_io;
93     break;
94   case MVT::i32:
95   case MVT::f32:
96   case MVT::v2i16:
97   case MVT::v4i8:
98     Opcode = IsValidInc ? Hexagon::L2_loadri_pi : Hexagon::L2_loadri_io;
99     break;
100   case MVT::i64:
101   case MVT::f64:
102   case MVT::v2i32:
103   case MVT::v4i16:
104   case MVT::v8i8:
105     Opcode = IsValidInc ? Hexagon::L2_loadrd_pi : Hexagon::L2_loadrd_io;
106     break;
107   case MVT::v64i8:
108   case MVT::v32i16:
109   case MVT::v16i32:
110   case MVT::v8i64:
111   case MVT::v128i8:
112   case MVT::v64i16:
113   case MVT::v32i32:
114   case MVT::v16i64:
115     if (isAlignedMemNode(LD)) {
116       if (LD->isNonTemporal())
117         Opcode = IsValidInc ? Hexagon::V6_vL32b_nt_pi : Hexagon::V6_vL32b_nt_ai;
118       else
119         Opcode = IsValidInc ? Hexagon::V6_vL32b_pi : Hexagon::V6_vL32b_ai;
120     } else {
121       Opcode = IsValidInc ? Hexagon::V6_vL32Ub_pi : Hexagon::V6_vL32Ub_ai;
122     }
123     break;
124   default:
125     llvm_unreachable("Unexpected memory type in indexed load");
126   }
127 
128   SDValue IncV = CurDAG->getTargetConstant(Inc, dl, MVT::i32);
129   MachineMemOperand *MemOp = LD->getMemOperand();
130 
131   auto getExt64 = [this,ExtType] (MachineSDNode *N, const SDLoc &dl)
132         -> MachineSDNode* {
133     if (ExtType == ISD::ZEXTLOAD || ExtType == ISD::EXTLOAD) {
134       SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
135       return CurDAG->getMachineNode(Hexagon::A4_combineir, dl, MVT::i64,
136                                     Zero, SDValue(N, 0));
137     }
138     if (ExtType == ISD::SEXTLOAD)
139       return CurDAG->getMachineNode(Hexagon::A2_sxtw, dl, MVT::i64,
140                                     SDValue(N, 0));
141     return N;
142   };
143 
144   //                  Loaded value   Next address   Chain
145   SDValue From[3] = { SDValue(LD,0), SDValue(LD,1), SDValue(LD,2) };
146   SDValue To[3];
147 
148   EVT ValueVT = LD->getValueType(0);
149   if (ValueVT == MVT::i64 && ExtType != ISD::NON_EXTLOAD) {
150     // A load extending to i64 will actually produce i32, which will then
151     // need to be extended to i64.
152     assert(LoadedVT.getSizeInBits() <= 32);
153     ValueVT = MVT::i32;
154   }
155 
156   if (IsValidInc) {
157     MachineSDNode *L = CurDAG->getMachineNode(Opcode, dl, ValueVT,
158                                               MVT::i32, MVT::Other, Base,
159                                               IncV, Chain);
160     CurDAG->setNodeMemRefs(L, {MemOp});
161     To[1] = SDValue(L, 1); // Next address.
162     To[2] = SDValue(L, 2); // Chain.
163     // Handle special case for extension to i64.
164     if (LD->getValueType(0) == MVT::i64)
165       L = getExt64(L, dl);
166     To[0] = SDValue(L, 0); // Loaded (extended) value.
167   } else {
168     SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
169     MachineSDNode *L = CurDAG->getMachineNode(Opcode, dl, ValueVT, MVT::Other,
170                                               Base, Zero, Chain);
171     CurDAG->setNodeMemRefs(L, {MemOp});
172     To[2] = SDValue(L, 1); // Chain.
173     MachineSDNode *A = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
174                                               Base, IncV);
175     To[1] = SDValue(A, 0); // Next address.
176     // Handle special case for extension to i64.
177     if (LD->getValueType(0) == MVT::i64)
178       L = getExt64(L, dl);
179     To[0] = SDValue(L, 0); // Loaded (extended) value.
180   }
181   ReplaceUses(From, To, 3);
182   CurDAG->RemoveDeadNode(LD);
183 }
184 
185 MachineSDNode *HexagonDAGToDAGISel::LoadInstrForLoadIntrinsic(SDNode *IntN) {
186   if (IntN->getOpcode() != ISD::INTRINSIC_W_CHAIN)
187     return nullptr;
188 
189   SDLoc dl(IntN);
190   unsigned IntNo = cast<ConstantSDNode>(IntN->getOperand(1))->getZExtValue();
191 
192   static std::map<unsigned,unsigned> LoadPciMap = {
193     { Intrinsic::hexagon_circ_ldb,  Hexagon::L2_loadrb_pci  },
194     { Intrinsic::hexagon_circ_ldub, Hexagon::L2_loadrub_pci },
195     { Intrinsic::hexagon_circ_ldh,  Hexagon::L2_loadrh_pci  },
196     { Intrinsic::hexagon_circ_lduh, Hexagon::L2_loadruh_pci },
197     { Intrinsic::hexagon_circ_ldw,  Hexagon::L2_loadri_pci  },
198     { Intrinsic::hexagon_circ_ldd,  Hexagon::L2_loadrd_pci  },
199   };
200   auto FLC = LoadPciMap.find(IntNo);
201   if (FLC != LoadPciMap.end()) {
202     EVT ValTy = (IntNo == Intrinsic::hexagon_circ_ldd) ? MVT::i64 : MVT::i32;
203     EVT RTys[] = { ValTy, MVT::i32, MVT::Other };
204     // Operands: { Base, Increment, Modifier, Chain }
205     auto Inc = cast<ConstantSDNode>(IntN->getOperand(5));
206     SDValue I = CurDAG->getTargetConstant(Inc->getSExtValue(), dl, MVT::i32);
207     MachineSDNode *Res = CurDAG->getMachineNode(FLC->second, dl, RTys,
208           { IntN->getOperand(2), I, IntN->getOperand(4),
209             IntN->getOperand(0) });
210     return Res;
211   }
212 
213   return nullptr;
214 }
215 
216 SDNode *HexagonDAGToDAGISel::StoreInstrForLoadIntrinsic(MachineSDNode *LoadN,
217       SDNode *IntN) {
218   // The "LoadN" is just a machine load instruction. The intrinsic also
219   // involves storing it. Generate an appropriate store to the location
220   // given in the intrinsic's operand(3).
221   uint64_t F = HII->get(LoadN->getMachineOpcode()).TSFlags;
222   unsigned SizeBits = (F >> HexagonII::MemAccessSizePos) &
223                       HexagonII::MemAccesSizeMask;
224   unsigned Size = 1U << (SizeBits-1);
225 
226   SDLoc dl(IntN);
227   MachinePointerInfo PI;
228   SDValue TS;
229   SDValue Loc = IntN->getOperand(3);
230 
231   if (Size >= 4)
232     TS = CurDAG->getStore(SDValue(LoadN, 2), dl, SDValue(LoadN, 0), Loc, PI,
233                           Align(Size));
234   else
235     TS = CurDAG->getTruncStore(SDValue(LoadN, 2), dl, SDValue(LoadN, 0), Loc,
236                                PI, MVT::getIntegerVT(Size * 8), Align(Size));
237 
238   SDNode *StoreN;
239   {
240     HandleSDNode Handle(TS);
241     SelectStore(TS.getNode());
242     StoreN = Handle.getValue().getNode();
243   }
244 
245   // Load's results are { Loaded value, Updated pointer, Chain }
246   ReplaceUses(SDValue(IntN, 0), SDValue(LoadN, 1));
247   ReplaceUses(SDValue(IntN, 1), SDValue(StoreN, 0));
248   return StoreN;
249 }
250 
251 bool HexagonDAGToDAGISel::tryLoadOfLoadIntrinsic(LoadSDNode *N) {
252   // The intrinsics for load circ/brev perform two operations:
253   // 1. Load a value V from the specified location, using the addressing
254   //    mode corresponding to the intrinsic.
255   // 2. Store V into a specified location. This location is typically a
256   //    local, temporary object.
257   // In many cases, the program using these intrinsics will immediately
258   // load V again from the local object. In those cases, when certain
259   // conditions are met, the last load can be removed.
260   // This function identifies and optimizes this pattern. If the pattern
261   // cannot be optimized, it returns nullptr, which will cause the load
262   // to be selected separately from the intrinsic (which will be handled
263   // in SelectIntrinsicWChain).
264 
265   SDValue Ch = N->getOperand(0);
266   SDValue Loc = N->getOperand(1);
267 
268   // Assume that the load and the intrinsic are connected directly with a
269   // chain:
270   //   t1: i32,ch = int.load ..., ..., ..., Loc, ...    // <-- C
271   //   t2: i32,ch = load t1:1, Loc, ...
272   SDNode *C = Ch.getNode();
273 
274   if (C->getOpcode() != ISD::INTRINSIC_W_CHAIN)
275     return false;
276 
277   // The second load can only be eliminated if its extension type matches
278   // that of the load instruction corresponding to the intrinsic. The user
279   // can provide an address of an unsigned variable to store the result of
280   // a sign-extending intrinsic into (or the other way around).
281   ISD::LoadExtType IntExt;
282   switch (cast<ConstantSDNode>(C->getOperand(1))->getZExtValue()) {
283     case Intrinsic::hexagon_circ_ldub:
284     case Intrinsic::hexagon_circ_lduh:
285       IntExt = ISD::ZEXTLOAD;
286       break;
287     case Intrinsic::hexagon_circ_ldw:
288     case Intrinsic::hexagon_circ_ldd:
289       IntExt = ISD::NON_EXTLOAD;
290       break;
291     default:
292       IntExt = ISD::SEXTLOAD;
293       break;
294   }
295   if (N->getExtensionType() != IntExt)
296     return false;
297 
298   // Make sure the target location for the loaded value in the load intrinsic
299   // is the location from which LD (or N) is loading.
300   if (C->getNumOperands() < 4 || Loc.getNode() != C->getOperand(3).getNode())
301     return false;
302 
303   if (MachineSDNode *L = LoadInstrForLoadIntrinsic(C)) {
304     SDNode *S = StoreInstrForLoadIntrinsic(L, C);
305     SDValue F[] = { SDValue(N,0), SDValue(N,1), SDValue(C,0), SDValue(C,1) };
306     SDValue T[] = { SDValue(L,0), SDValue(S,0), SDValue(L,1), SDValue(S,0) };
307     ReplaceUses(F, T, array_lengthof(T));
308     // This transformation will leave the intrinsic dead. If it remains in
309     // the DAG, the selection code will see it again, but without the load,
310     // and it will generate a store that is normally required for it.
311     CurDAG->RemoveDeadNode(C);
312     return true;
313   }
314   return false;
315 }
316 
317 // Convert the bit-reverse load intrinsic to appropriate target instruction.
318 bool HexagonDAGToDAGISel::SelectBrevLdIntrinsic(SDNode *IntN) {
319   if (IntN->getOpcode() != ISD::INTRINSIC_W_CHAIN)
320     return false;
321 
322   const SDLoc &dl(IntN);
323   unsigned IntNo = cast<ConstantSDNode>(IntN->getOperand(1))->getZExtValue();
324 
325   static const std::map<unsigned, unsigned> LoadBrevMap = {
326     { Intrinsic::hexagon_L2_loadrb_pbr, Hexagon::L2_loadrb_pbr },
327     { Intrinsic::hexagon_L2_loadrub_pbr, Hexagon::L2_loadrub_pbr },
328     { Intrinsic::hexagon_L2_loadrh_pbr, Hexagon::L2_loadrh_pbr },
329     { Intrinsic::hexagon_L2_loadruh_pbr, Hexagon::L2_loadruh_pbr },
330     { Intrinsic::hexagon_L2_loadri_pbr, Hexagon::L2_loadri_pbr },
331     { Intrinsic::hexagon_L2_loadrd_pbr, Hexagon::L2_loadrd_pbr }
332   };
333   auto FLI = LoadBrevMap.find(IntNo);
334   if (FLI != LoadBrevMap.end()) {
335     EVT ValTy =
336         (IntNo == Intrinsic::hexagon_L2_loadrd_pbr) ? MVT::i64 : MVT::i32;
337     EVT RTys[] = { ValTy, MVT::i32, MVT::Other };
338     // Operands of Intrinsic: {chain, enum ID of intrinsic, baseptr,
339     // modifier}.
340     // Operands of target instruction: { Base, Modifier, Chain }.
341     MachineSDNode *Res = CurDAG->getMachineNode(
342         FLI->second, dl, RTys,
343         {IntN->getOperand(2), IntN->getOperand(3), IntN->getOperand(0)});
344 
345     MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(IntN)->getMemOperand();
346     CurDAG->setNodeMemRefs(Res, {MemOp});
347 
348     ReplaceUses(SDValue(IntN, 0), SDValue(Res, 0));
349     ReplaceUses(SDValue(IntN, 1), SDValue(Res, 1));
350     ReplaceUses(SDValue(IntN, 2), SDValue(Res, 2));
351     CurDAG->RemoveDeadNode(IntN);
352     return true;
353   }
354   return false;
355 }
356 
357 /// Generate a machine instruction node for the new circlar buffer intrinsics.
358 /// The new versions use a CSx register instead of the K field.
359 bool HexagonDAGToDAGISel::SelectNewCircIntrinsic(SDNode *IntN) {
360   if (IntN->getOpcode() != ISD::INTRINSIC_W_CHAIN)
361     return false;
362 
363   SDLoc DL(IntN);
364   unsigned IntNo = cast<ConstantSDNode>(IntN->getOperand(1))->getZExtValue();
365   SmallVector<SDValue, 7> Ops;
366 
367   static std::map<unsigned,unsigned> LoadNPcMap = {
368     { Intrinsic::hexagon_L2_loadrub_pci, Hexagon::PS_loadrub_pci },
369     { Intrinsic::hexagon_L2_loadrb_pci, Hexagon::PS_loadrb_pci },
370     { Intrinsic::hexagon_L2_loadruh_pci, Hexagon::PS_loadruh_pci },
371     { Intrinsic::hexagon_L2_loadrh_pci, Hexagon::PS_loadrh_pci },
372     { Intrinsic::hexagon_L2_loadri_pci, Hexagon::PS_loadri_pci },
373     { Intrinsic::hexagon_L2_loadrd_pci, Hexagon::PS_loadrd_pci },
374     { Intrinsic::hexagon_L2_loadrub_pcr, Hexagon::PS_loadrub_pcr },
375     { Intrinsic::hexagon_L2_loadrb_pcr, Hexagon::PS_loadrb_pcr },
376     { Intrinsic::hexagon_L2_loadruh_pcr, Hexagon::PS_loadruh_pcr },
377     { Intrinsic::hexagon_L2_loadrh_pcr, Hexagon::PS_loadrh_pcr },
378     { Intrinsic::hexagon_L2_loadri_pcr, Hexagon::PS_loadri_pcr },
379     { Intrinsic::hexagon_L2_loadrd_pcr, Hexagon::PS_loadrd_pcr }
380   };
381   auto FLI = LoadNPcMap.find (IntNo);
382   if (FLI != LoadNPcMap.end()) {
383     EVT ValTy = MVT::i32;
384     if (IntNo == Intrinsic::hexagon_L2_loadrd_pci ||
385         IntNo == Intrinsic::hexagon_L2_loadrd_pcr)
386       ValTy = MVT::i64;
387     EVT RTys[] = { ValTy, MVT::i32, MVT::Other };
388     // Handle load.*_pci case which has 6 operands.
389     if (IntN->getNumOperands() == 6) {
390       auto Inc = cast<ConstantSDNode>(IntN->getOperand(3));
391       SDValue I = CurDAG->getTargetConstant(Inc->getSExtValue(), DL, MVT::i32);
392       // Operands: { Base, Increment, Modifier, Start, Chain }.
393       Ops = { IntN->getOperand(2), I, IntN->getOperand(4), IntN->getOperand(5),
394               IntN->getOperand(0) };
395     } else
396       // Handle load.*_pcr case which has 5 operands.
397       // Operands: { Base, Modifier, Start, Chain }.
398       Ops = { IntN->getOperand(2), IntN->getOperand(3), IntN->getOperand(4),
399               IntN->getOperand(0) };
400     MachineSDNode *Res = CurDAG->getMachineNode(FLI->second, DL, RTys, Ops);
401     ReplaceUses(SDValue(IntN, 0), SDValue(Res, 0));
402     ReplaceUses(SDValue(IntN, 1), SDValue(Res, 1));
403     ReplaceUses(SDValue(IntN, 2), SDValue(Res, 2));
404     CurDAG->RemoveDeadNode(IntN);
405     return true;
406   }
407 
408   static std::map<unsigned,unsigned> StoreNPcMap = {
409     { Intrinsic::hexagon_S2_storerb_pci, Hexagon::PS_storerb_pci },
410     { Intrinsic::hexagon_S2_storerh_pci, Hexagon::PS_storerh_pci },
411     { Intrinsic::hexagon_S2_storerf_pci, Hexagon::PS_storerf_pci },
412     { Intrinsic::hexagon_S2_storeri_pci, Hexagon::PS_storeri_pci },
413     { Intrinsic::hexagon_S2_storerd_pci, Hexagon::PS_storerd_pci },
414     { Intrinsic::hexagon_S2_storerb_pcr, Hexagon::PS_storerb_pcr },
415     { Intrinsic::hexagon_S2_storerh_pcr, Hexagon::PS_storerh_pcr },
416     { Intrinsic::hexagon_S2_storerf_pcr, Hexagon::PS_storerf_pcr },
417     { Intrinsic::hexagon_S2_storeri_pcr, Hexagon::PS_storeri_pcr },
418     { Intrinsic::hexagon_S2_storerd_pcr, Hexagon::PS_storerd_pcr }
419   };
420   auto FSI = StoreNPcMap.find (IntNo);
421   if (FSI != StoreNPcMap.end()) {
422     EVT RTys[] = { MVT::i32, MVT::Other };
423     // Handle store.*_pci case which has 7 operands.
424     if (IntN->getNumOperands() == 7) {
425       auto Inc = cast<ConstantSDNode>(IntN->getOperand(3));
426       SDValue I = CurDAG->getTargetConstant(Inc->getSExtValue(), DL, MVT::i32);
427       // Operands: { Base, Increment, Modifier, Value, Start, Chain }.
428       Ops = { IntN->getOperand(2), I, IntN->getOperand(4), IntN->getOperand(5),
429               IntN->getOperand(6), IntN->getOperand(0) };
430     } else
431       // Handle store.*_pcr case which has 6 operands.
432       // Operands: { Base, Modifier, Value, Start, Chain }.
433       Ops = { IntN->getOperand(2), IntN->getOperand(3), IntN->getOperand(4),
434               IntN->getOperand(5), IntN->getOperand(0) };
435     MachineSDNode *Res = CurDAG->getMachineNode(FSI->second, DL, RTys, Ops);
436     ReplaceUses(SDValue(IntN, 0), SDValue(Res, 0));
437     ReplaceUses(SDValue(IntN, 1), SDValue(Res, 1));
438     CurDAG->RemoveDeadNode(IntN);
439     return true;
440   }
441 
442   return false;
443 }
444 
445 void HexagonDAGToDAGISel::SelectLoad(SDNode *N) {
446   SDLoc dl(N);
447   LoadSDNode *LD = cast<LoadSDNode>(N);
448 
449   // Handle indexed loads.
450   ISD::MemIndexedMode AM = LD->getAddressingMode();
451   if (AM != ISD::UNINDEXED) {
452     SelectIndexedLoad(LD, dl);
453     return;
454   }
455 
456   // Handle patterns using circ/brev load intrinsics.
457   if (tryLoadOfLoadIntrinsic(LD))
458     return;
459 
460   SelectCode(LD);
461 }
462 
463 void HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, const SDLoc &dl) {
464   SDValue Chain = ST->getChain();
465   SDValue Base = ST->getBasePtr();
466   SDValue Offset = ST->getOffset();
467   SDValue Value = ST->getValue();
468   // Get the constant value.
469   int32_t Inc = cast<ConstantSDNode>(Offset.getNode())->getSExtValue();
470   EVT StoredVT = ST->getMemoryVT();
471   EVT ValueVT = Value.getValueType();
472 
473   bool IsValidInc = HII->isValidAutoIncImm(StoredVT, Inc);
474   unsigned Opcode = 0;
475 
476   assert(StoredVT.isSimple());
477   switch (StoredVT.getSimpleVT().SimpleTy) {
478   case MVT::i8:
479     Opcode = IsValidInc ? Hexagon::S2_storerb_pi : Hexagon::S2_storerb_io;
480     break;
481   case MVT::i16:
482     Opcode = IsValidInc ? Hexagon::S2_storerh_pi : Hexagon::S2_storerh_io;
483     break;
484   case MVT::i32:
485   case MVT::f32:
486   case MVT::v2i16:
487   case MVT::v4i8:
488     Opcode = IsValidInc ? Hexagon::S2_storeri_pi : Hexagon::S2_storeri_io;
489     break;
490   case MVT::i64:
491   case MVT::f64:
492   case MVT::v2i32:
493   case MVT::v4i16:
494   case MVT::v8i8:
495     Opcode = IsValidInc ? Hexagon::S2_storerd_pi : Hexagon::S2_storerd_io;
496     break;
497   case MVT::v64i8:
498   case MVT::v32i16:
499   case MVT::v16i32:
500   case MVT::v8i64:
501   case MVT::v128i8:
502   case MVT::v64i16:
503   case MVT::v32i32:
504   case MVT::v16i64:
505     if (isAlignedMemNode(ST)) {
506       if (ST->isNonTemporal())
507         Opcode = IsValidInc ? Hexagon::V6_vS32b_nt_pi : Hexagon::V6_vS32b_nt_ai;
508       else
509         Opcode = IsValidInc ? Hexagon::V6_vS32b_pi : Hexagon::V6_vS32b_ai;
510     } else {
511       Opcode = IsValidInc ? Hexagon::V6_vS32Ub_pi : Hexagon::V6_vS32Ub_ai;
512     }
513     break;
514   default:
515     llvm_unreachable("Unexpected memory type in indexed store");
516   }
517 
518   if (ST->isTruncatingStore() && ValueVT.getSizeInBits() == 64) {
519     assert(StoredVT.getSizeInBits() < 64 && "Not a truncating store");
520     Value = CurDAG->getTargetExtractSubreg(Hexagon::isub_lo,
521                                            dl, MVT::i32, Value);
522   }
523 
524   SDValue IncV = CurDAG->getTargetConstant(Inc, dl, MVT::i32);
525   MachineMemOperand *MemOp = ST->getMemOperand();
526 
527   //                  Next address   Chain
528   SDValue From[2] = { SDValue(ST,0), SDValue(ST,1) };
529   SDValue To[2];
530 
531   if (IsValidInc) {
532     // Build post increment store.
533     SDValue Ops[] = { Base, IncV, Value, Chain };
534     MachineSDNode *S = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::Other,
535                                               Ops);
536     CurDAG->setNodeMemRefs(S, {MemOp});
537     To[0] = SDValue(S, 0);
538     To[1] = SDValue(S, 1);
539   } else {
540     SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
541     SDValue Ops[] = { Base, Zero, Value, Chain };
542     MachineSDNode *S = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops);
543     CurDAG->setNodeMemRefs(S, {MemOp});
544     To[1] = SDValue(S, 0);
545     MachineSDNode *A = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
546                                               Base, IncV);
547     To[0] = SDValue(A, 0);
548   }
549 
550   ReplaceUses(From, To, 2);
551   CurDAG->RemoveDeadNode(ST);
552 }
553 
554 void HexagonDAGToDAGISel::SelectStore(SDNode *N) {
555   SDLoc dl(N);
556   StoreSDNode *ST = cast<StoreSDNode>(N);
557 
558   // Handle indexed stores.
559   ISD::MemIndexedMode AM = ST->getAddressingMode();
560   if (AM != ISD::UNINDEXED) {
561     SelectIndexedStore(ST, dl);
562     return;
563   }
564 
565   SelectCode(ST);
566 }
567 
568 void HexagonDAGToDAGISel::SelectSHL(SDNode *N) {
569   SDLoc dl(N);
570   SDValue Shl_0 = N->getOperand(0);
571   SDValue Shl_1 = N->getOperand(1);
572 
573   auto Default = [this,N] () -> void { SelectCode(N); };
574 
575   if (N->getValueType(0) != MVT::i32 || Shl_1.getOpcode() != ISD::Constant)
576     return Default();
577 
578   // RHS is const.
579   int32_t ShlConst = cast<ConstantSDNode>(Shl_1)->getSExtValue();
580 
581   if (Shl_0.getOpcode() == ISD::MUL) {
582     SDValue Mul_0 = Shl_0.getOperand(0); // Val
583     SDValue Mul_1 = Shl_0.getOperand(1); // Const
584     // RHS of mul is const.
585     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mul_1)) {
586       int32_t ValConst = C->getSExtValue() << ShlConst;
587       if (isInt<9>(ValConst)) {
588         SDValue Val = CurDAG->getTargetConstant(ValConst, dl, MVT::i32);
589         SDNode *Result = CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl,
590                                                 MVT::i32, Mul_0, Val);
591         ReplaceNode(N, Result);
592         return;
593       }
594     }
595     return Default();
596   }
597 
598   if (Shl_0.getOpcode() == ISD::SUB) {
599     SDValue Sub_0 = Shl_0.getOperand(0); // Const 0
600     SDValue Sub_1 = Shl_0.getOperand(1); // Val
601     if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Sub_0)) {
602       if (C1->getSExtValue() != 0 || Sub_1.getOpcode() != ISD::SHL)
603         return Default();
604       SDValue Shl2_0 = Sub_1.getOperand(0); // Val
605       SDValue Shl2_1 = Sub_1.getOperand(1); // Const
606       if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(Shl2_1)) {
607         int32_t ValConst = 1 << (ShlConst + C2->getSExtValue());
608         if (isInt<9>(-ValConst)) {
609           SDValue Val = CurDAG->getTargetConstant(-ValConst, dl, MVT::i32);
610           SDNode *Result = CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl,
611                                                   MVT::i32, Shl2_0, Val);
612           ReplaceNode(N, Result);
613           return;
614         }
615       }
616     }
617   }
618 
619   return Default();
620 }
621 
622 //
623 // Handling intrinsics for circular load and bitreverse load.
624 //
625 void HexagonDAGToDAGISel::SelectIntrinsicWChain(SDNode *N) {
626   if (MachineSDNode *L = LoadInstrForLoadIntrinsic(N)) {
627     StoreInstrForLoadIntrinsic(L, N);
628     CurDAG->RemoveDeadNode(N);
629     return;
630   }
631 
632   // Handle bit-reverse load intrinsics.
633   if (SelectBrevLdIntrinsic(N))
634     return;
635 
636   if (SelectNewCircIntrinsic(N))
637     return;
638 
639   unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
640   if (IntNo == Intrinsic::hexagon_V6_vgathermw ||
641       IntNo == Intrinsic::hexagon_V6_vgathermw_128B ||
642       IntNo == Intrinsic::hexagon_V6_vgathermh ||
643       IntNo == Intrinsic::hexagon_V6_vgathermh_128B ||
644       IntNo == Intrinsic::hexagon_V6_vgathermhw ||
645       IntNo == Intrinsic::hexagon_V6_vgathermhw_128B) {
646     SelectV65Gather(N);
647     return;
648   }
649   if (IntNo == Intrinsic::hexagon_V6_vgathermwq ||
650       IntNo == Intrinsic::hexagon_V6_vgathermwq_128B ||
651       IntNo == Intrinsic::hexagon_V6_vgathermhq ||
652       IntNo == Intrinsic::hexagon_V6_vgathermhq_128B ||
653       IntNo == Intrinsic::hexagon_V6_vgathermhwq ||
654       IntNo == Intrinsic::hexagon_V6_vgathermhwq_128B) {
655     SelectV65GatherPred(N);
656     return;
657   }
658 
659   SelectCode(N);
660 }
661 
662 void HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) {
663   unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
664   unsigned Bits;
665   switch (IID) {
666   case Intrinsic::hexagon_S2_vsplatrb:
667     Bits = 8;
668     break;
669   case Intrinsic::hexagon_S2_vsplatrh:
670     Bits = 16;
671     break;
672   case Intrinsic::hexagon_V6_vaddcarry:
673   case Intrinsic::hexagon_V6_vaddcarry_128B:
674   case Intrinsic::hexagon_V6_vsubcarry:
675   case Intrinsic::hexagon_V6_vsubcarry_128B:
676     SelectHVXDualOutput(N);
677     return;
678   default:
679     SelectCode(N);
680     return;
681   }
682 
683   SDValue V = N->getOperand(1);
684   SDValue U;
685   if (keepsLowBits(V, Bits, U)) {
686     SDValue R = CurDAG->getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
687                                 N->getOperand(0), U);
688     ReplaceNode(N, R.getNode());
689     SelectCode(R.getNode());
690     return;
691   }
692   SelectCode(N);
693 }
694 
695 //
696 // Map floating point constant values.
697 //
698 void HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) {
699   SDLoc dl(N);
700   auto *CN = cast<ConstantFPSDNode>(N);
701   APInt A = CN->getValueAPF().bitcastToAPInt();
702   if (N->getValueType(0) == MVT::f32) {
703     SDValue V = CurDAG->getTargetConstant(A.getZExtValue(), dl, MVT::i32);
704     ReplaceNode(N, CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl, MVT::f32, V));
705     return;
706   }
707   if (N->getValueType(0) == MVT::f64) {
708     SDValue V = CurDAG->getTargetConstant(A.getZExtValue(), dl, MVT::i64);
709     ReplaceNode(N, CurDAG->getMachineNode(Hexagon::CONST64, dl, MVT::f64, V));
710     return;
711   }
712 
713   SelectCode(N);
714 }
715 
716 //
717 // Map boolean values.
718 //
719 void HexagonDAGToDAGISel::SelectConstant(SDNode *N) {
720   if (N->getValueType(0) == MVT::i1) {
721     assert(!(cast<ConstantSDNode>(N)->getZExtValue() >> 1));
722     unsigned Opc = (cast<ConstantSDNode>(N)->getSExtValue() != 0)
723                       ? Hexagon::PS_true
724                       : Hexagon::PS_false;
725     ReplaceNode(N, CurDAG->getMachineNode(Opc, SDLoc(N), MVT::i1));
726     return;
727   }
728 
729   SelectCode(N);
730 }
731 
732 void HexagonDAGToDAGISel::SelectFrameIndex(SDNode *N) {
733   MachineFrameInfo &MFI = MF->getFrameInfo();
734   const HexagonFrameLowering *HFI = HST->getFrameLowering();
735   int FX = cast<FrameIndexSDNode>(N)->getIndex();
736   Align StkA = HFI->getStackAlign();
737   Align MaxA = MFI.getMaxAlign();
738   SDValue FI = CurDAG->getTargetFrameIndex(FX, MVT::i32);
739   SDLoc DL(N);
740   SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
741   SDNode *R = nullptr;
742 
743   // Use PS_fi when:
744   // - the object is fixed, or
745   // - there are no objects with higher-than-default alignment, or
746   // - there are no dynamically allocated objects.
747   // Otherwise, use PS_fia.
748   if (FX < 0 || MaxA <= StkA || !MFI.hasVarSizedObjects()) {
749     R = CurDAG->getMachineNode(Hexagon::PS_fi, DL, MVT::i32, FI, Zero);
750   } else {
751     auto &HMFI = *MF->getInfo<HexagonMachineFunctionInfo>();
752     unsigned AR = HMFI.getStackAlignBaseVReg();
753     SDValue CH = CurDAG->getEntryNode();
754     SDValue Ops[] = { CurDAG->getCopyFromReg(CH, DL, AR, MVT::i32), FI, Zero };
755     R = CurDAG->getMachineNode(Hexagon::PS_fia, DL, MVT::i32, Ops);
756   }
757 
758   ReplaceNode(N, R);
759 }
760 
761 void HexagonDAGToDAGISel::SelectAddSubCarry(SDNode *N) {
762   unsigned OpcCarry = N->getOpcode() == HexagonISD::ADDC ? Hexagon::A4_addp_c
763                                                          : Hexagon::A4_subp_c;
764   SDNode *C = CurDAG->getMachineNode(OpcCarry, SDLoc(N), N->getVTList(),
765                                      { N->getOperand(0), N->getOperand(1),
766                                        N->getOperand(2) });
767   ReplaceNode(N, C);
768 }
769 
770 void HexagonDAGToDAGISel::SelectVAlign(SDNode *N) {
771   MVT ResTy = N->getValueType(0).getSimpleVT();
772   if (HST->isHVXVectorType(ResTy, true))
773     return SelectHvxVAlign(N);
774 
775   const SDLoc &dl(N);
776   unsigned VecLen = ResTy.getSizeInBits();
777   if (VecLen == 32) {
778     SDValue Ops[] = {
779       CurDAG->getTargetConstant(Hexagon::DoubleRegsRegClassID, dl, MVT::i32),
780       N->getOperand(0),
781       CurDAG->getTargetConstant(Hexagon::isub_hi, dl, MVT::i32),
782       N->getOperand(1),
783       CurDAG->getTargetConstant(Hexagon::isub_lo, dl, MVT::i32)
784     };
785     SDNode *R = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl,
786                                        MVT::i64, Ops);
787 
788     // Shift right by "(Addr & 0x3) * 8" bytes.
789     SDNode *C;
790     SDValue M0 = CurDAG->getTargetConstant(0x18, dl, MVT::i32);
791     SDValue M1 = CurDAG->getTargetConstant(0x03, dl, MVT::i32);
792     if (HST->useCompound()) {
793       C = CurDAG->getMachineNode(Hexagon::S4_andi_asl_ri, dl, MVT::i32,
794                                  M0, N->getOperand(2), M1);
795     } else {
796       SDNode *T = CurDAG->getMachineNode(Hexagon::S2_asl_i_r, dl, MVT::i32,
797                                          N->getOperand(2), M1);
798       C = CurDAG->getMachineNode(Hexagon::A2_andir, dl, MVT::i32,
799                                  SDValue(T, 0), M0);
800     }
801     SDNode *S = CurDAG->getMachineNode(Hexagon::S2_lsr_r_p, dl, MVT::i64,
802                                        SDValue(R, 0), SDValue(C, 0));
803     SDValue E = CurDAG->getTargetExtractSubreg(Hexagon::isub_lo, dl, ResTy,
804                                                SDValue(S, 0));
805     ReplaceNode(N, E.getNode());
806   } else {
807     assert(VecLen == 64);
808     SDNode *Pu = CurDAG->getMachineNode(Hexagon::C2_tfrrp, dl, MVT::v8i1,
809                                         N->getOperand(2));
810     SDNode *VA = CurDAG->getMachineNode(Hexagon::S2_valignrb, dl, ResTy,
811                                         N->getOperand(0), N->getOperand(1),
812                                         SDValue(Pu,0));
813     ReplaceNode(N, VA);
814   }
815 }
816 
817 void HexagonDAGToDAGISel::SelectVAlignAddr(SDNode *N) {
818   const SDLoc &dl(N);
819   SDValue A = N->getOperand(1);
820   int Mask = -cast<ConstantSDNode>(A.getNode())->getSExtValue();
821   assert(isPowerOf2_32(-Mask));
822 
823   SDValue M = CurDAG->getTargetConstant(Mask, dl, MVT::i32);
824   SDNode *AA = CurDAG->getMachineNode(Hexagon::A2_andir, dl, MVT::i32,
825                                       N->getOperand(0), M);
826   ReplaceNode(N, AA);
827 }
828 
829 // Handle these nodes here to avoid having to write patterns for all
830 // combinations of input/output types. In all cases, the resulting
831 // instruction is the same.
832 void HexagonDAGToDAGISel::SelectTypecast(SDNode *N) {
833   SDValue Op = N->getOperand(0);
834   MVT OpTy = Op.getValueType().getSimpleVT();
835   SDNode *T = CurDAG->MorphNodeTo(N, N->getOpcode(),
836                                   CurDAG->getVTList(OpTy), {Op});
837   ReplaceNode(T, Op.getNode());
838 }
839 
840 void HexagonDAGToDAGISel::SelectP2D(SDNode *N) {
841   MVT ResTy = N->getValueType(0).getSimpleVT();
842   SDNode *T = CurDAG->getMachineNode(Hexagon::C2_mask, SDLoc(N), ResTy,
843                                      N->getOperand(0));
844   ReplaceNode(N, T);
845 }
846 
847 void HexagonDAGToDAGISel::SelectD2P(SDNode *N) {
848   const SDLoc &dl(N);
849   MVT ResTy = N->getValueType(0).getSimpleVT();
850   SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
851   SDNode *T = CurDAG->getMachineNode(Hexagon::A4_vcmpbgtui, dl, ResTy,
852                                      N->getOperand(0), Zero);
853   ReplaceNode(N, T);
854 }
855 
856 void HexagonDAGToDAGISel::SelectV2Q(SDNode *N) {
857   const SDLoc &dl(N);
858   MVT ResTy = N->getValueType(0).getSimpleVT();
859   // The argument to V2Q should be a single vector.
860   MVT OpTy = N->getOperand(0).getValueType().getSimpleVT(); (void)OpTy;
861   assert(HST->getVectorLength() * 8 == OpTy.getSizeInBits());
862 
863   SDValue C = CurDAG->getTargetConstant(-1, dl, MVT::i32);
864   SDNode *R = CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl, MVT::i32, C);
865   SDNode *T = CurDAG->getMachineNode(Hexagon::V6_vandvrt, dl, ResTy,
866                                      N->getOperand(0), SDValue(R,0));
867   ReplaceNode(N, T);
868 }
869 
870 void HexagonDAGToDAGISel::SelectQ2V(SDNode *N) {
871   const SDLoc &dl(N);
872   MVT ResTy = N->getValueType(0).getSimpleVT();
873   // The result of V2Q should be a single vector.
874   assert(HST->getVectorLength() * 8 == ResTy.getSizeInBits());
875 
876   SDValue C = CurDAG->getTargetConstant(-1, dl, MVT::i32);
877   SDNode *R = CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl, MVT::i32, C);
878   SDNode *T = CurDAG->getMachineNode(Hexagon::V6_vandqrt, dl, ResTy,
879                                      N->getOperand(0), SDValue(R,0));
880   ReplaceNode(N, T);
881 }
882 
883 void HexagonDAGToDAGISel::Select(SDNode *N) {
884   if (N->isMachineOpcode())
885     return N->setNodeId(-1);  // Already selected.
886 
887   switch (N->getOpcode()) {
888   case ISD::Constant:             return SelectConstant(N);
889   case ISD::ConstantFP:           return SelectConstantFP(N);
890   case ISD::FrameIndex:           return SelectFrameIndex(N);
891   case ISD::SHL:                  return SelectSHL(N);
892   case ISD::LOAD:                 return SelectLoad(N);
893   case ISD::STORE:                return SelectStore(N);
894   case ISD::INTRINSIC_W_CHAIN:    return SelectIntrinsicWChain(N);
895   case ISD::INTRINSIC_WO_CHAIN:   return SelectIntrinsicWOChain(N);
896 
897   case HexagonISD::ADDC:
898   case HexagonISD::SUBC:          return SelectAddSubCarry(N);
899   case HexagonISD::VALIGN:        return SelectVAlign(N);
900   case HexagonISD::VALIGNADDR:    return SelectVAlignAddr(N);
901   case HexagonISD::TYPECAST:      return SelectTypecast(N);
902   case HexagonISD::P2D:           return SelectP2D(N);
903   case HexagonISD::D2P:           return SelectD2P(N);
904   case HexagonISD::Q2V:           return SelectQ2V(N);
905   case HexagonISD::V2Q:           return SelectV2Q(N);
906   }
907 
908   if (HST->useHVXOps()) {
909     switch (N->getOpcode()) {
910     case ISD::VECTOR_SHUFFLE:     return SelectHvxShuffle(N);
911     case HexagonISD::VROR:        return SelectHvxRor(N);
912     }
913   }
914 
915   SelectCode(N);
916 }
917 
918 bool HexagonDAGToDAGISel::
919 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
920                              std::vector<SDValue> &OutOps) {
921   SDValue Inp = Op, Res;
922 
923   switch (ConstraintID) {
924   default:
925     return true;
926   case InlineAsm::Constraint_o: // Offsetable.
927   case InlineAsm::Constraint_v: // Not offsetable.
928   case InlineAsm::Constraint_m: // Memory.
929     if (SelectAddrFI(Inp, Res))
930       OutOps.push_back(Res);
931     else
932       OutOps.push_back(Inp);
933     break;
934   }
935 
936   OutOps.push_back(CurDAG->getTargetConstant(0, SDLoc(Op), MVT::i32));
937   return false;
938 }
939 
940 
941 static bool isMemOPCandidate(SDNode *I, SDNode *U) {
942   // I is an operand of U. Check if U is an arithmetic (binary) operation
943   // usable in a memop, where the other operand is a loaded value, and the
944   // result of U is stored in the same location.
945 
946   if (!U->hasOneUse())
947     return false;
948   unsigned Opc = U->getOpcode();
949   switch (Opc) {
950     case ISD::ADD:
951     case ISD::SUB:
952     case ISD::AND:
953     case ISD::OR:
954       break;
955     default:
956       return false;
957   }
958 
959   SDValue S0 = U->getOperand(0);
960   SDValue S1 = U->getOperand(1);
961   SDValue SY = (S0.getNode() == I) ? S1 : S0;
962 
963   SDNode *UUse = *U->use_begin();
964   if (UUse->getNumValues() != 1)
965     return false;
966 
967   // Check if one of the inputs to U is a load instruction and the output
968   // is used by a store instruction. If so and they also have the same
969   // base pointer, then don't preoprocess this node sequence as it
970   // can be matched to a memop.
971   SDNode *SYNode = SY.getNode();
972   if (UUse->getOpcode() == ISD::STORE && SYNode->getOpcode() == ISD::LOAD) {
973     SDValue LDBasePtr = cast<MemSDNode>(SYNode)->getBasePtr();
974     SDValue STBasePtr = cast<MemSDNode>(UUse)->getBasePtr();
975     if (LDBasePtr == STBasePtr)
976       return true;
977   }
978   return false;
979 }
980 
981 
982 // Transform: (or (select c x 0) z)  ->  (select c (or x z) z)
983 //            (or (select c 0 y) z)  ->  (select c z (or y z))
984 void HexagonDAGToDAGISel::ppSimplifyOrSelect0(std::vector<SDNode*> &&Nodes) {
985   SelectionDAG &DAG = *CurDAG;
986 
987   for (auto I : Nodes) {
988     if (I->getOpcode() != ISD::OR)
989       continue;
990 
991     auto IsZero = [] (const SDValue &V) -> bool {
992       if (ConstantSDNode *SC = dyn_cast<ConstantSDNode>(V.getNode()))
993         return SC->isZero();
994       return false;
995     };
996     auto IsSelect0 = [IsZero] (const SDValue &Op) -> bool {
997       if (Op.getOpcode() != ISD::SELECT)
998         return false;
999       return IsZero(Op.getOperand(1)) || IsZero(Op.getOperand(2));
1000     };
1001 
1002     SDValue N0 = I->getOperand(0), N1 = I->getOperand(1);
1003     EVT VT = I->getValueType(0);
1004     bool SelN0 = IsSelect0(N0);
1005     SDValue SOp = SelN0 ? N0 : N1;
1006     SDValue VOp = SelN0 ? N1 : N0;
1007 
1008     if (SOp.getOpcode() == ISD::SELECT && SOp.getNode()->hasOneUse()) {
1009       SDValue SC = SOp.getOperand(0);
1010       SDValue SX = SOp.getOperand(1);
1011       SDValue SY = SOp.getOperand(2);
1012       SDLoc DLS = SOp;
1013       if (IsZero(SY)) {
1014         SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SX, VOp);
1015         SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, NewOr, VOp);
1016         DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1017       } else if (IsZero(SX)) {
1018         SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SY, VOp);
1019         SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, VOp, NewOr);
1020         DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1021       }
1022     }
1023   }
1024 }
1025 
1026 // Transform: (store ch val (add x (add (shl y c) e)))
1027 //        to: (store ch val (add x (shl (add y d) c))),
1028 // where e = (shl d c) for some integer d.
1029 // The purpose of this is to enable generation of loads/stores with
1030 // shifted addressing mode, i.e. mem(x+y<<#c). For that, the shift
1031 // value c must be 0, 1 or 2.
1032 void HexagonDAGToDAGISel::ppAddrReorderAddShl(std::vector<SDNode*> &&Nodes) {
1033   SelectionDAG &DAG = *CurDAG;
1034 
1035   for (auto I : Nodes) {
1036     if (I->getOpcode() != ISD::STORE)
1037       continue;
1038 
1039     // I matched: (store ch val Off)
1040     SDValue Off = I->getOperand(2);
1041     // Off needs to match: (add x (add (shl y c) (shl d c))))
1042     if (Off.getOpcode() != ISD::ADD)
1043       continue;
1044     // Off matched: (add x T0)
1045     SDValue T0 = Off.getOperand(1);
1046     // T0 needs to match: (add T1 T2):
1047     if (T0.getOpcode() != ISD::ADD)
1048       continue;
1049     // T0 matched: (add T1 T2)
1050     SDValue T1 = T0.getOperand(0);
1051     SDValue T2 = T0.getOperand(1);
1052     // T1 needs to match: (shl y c)
1053     if (T1.getOpcode() != ISD::SHL)
1054       continue;
1055     SDValue C = T1.getOperand(1);
1056     ConstantSDNode *CN = dyn_cast<ConstantSDNode>(C.getNode());
1057     if (CN == nullptr)
1058       continue;
1059     unsigned CV = CN->getZExtValue();
1060     if (CV > 2)
1061       continue;
1062     // T2 needs to match e, where e = (shl d c) for some d.
1063     ConstantSDNode *EN = dyn_cast<ConstantSDNode>(T2.getNode());
1064     if (EN == nullptr)
1065       continue;
1066     unsigned EV = EN->getZExtValue();
1067     if (EV % (1 << CV) != 0)
1068       continue;
1069     unsigned DV = EV / (1 << CV);
1070 
1071     // Replace T0 with: (shl (add y d) c)
1072     SDLoc DL = SDLoc(I);
1073     EVT VT = T0.getValueType();
1074     SDValue D = DAG.getConstant(DV, DL, VT);
1075     // NewAdd = (add y d)
1076     SDValue NewAdd = DAG.getNode(ISD::ADD, DL, VT, T1.getOperand(0), D);
1077     // NewShl = (shl NewAdd c)
1078     SDValue NewShl = DAG.getNode(ISD::SHL, DL, VT, NewAdd, C);
1079     ReplaceNode(T0.getNode(), NewShl.getNode());
1080   }
1081 }
1082 
1083 // Transform: (load ch (add x (and (srl y c) Mask)))
1084 //        to: (load ch (add x (shl (srl y d) d-c)))
1085 // where
1086 // Mask = 00..0 111..1 0.0
1087 //          |     |     +-- d-c 0s, and d-c is 0, 1 or 2.
1088 //          |     +-------- 1s
1089 //          +-------------- at most c 0s
1090 // Motivating example:
1091 // DAG combiner optimizes (add x (shl (srl y 5) 2))
1092 //                     to (add x (and (srl y 3) 1FFFFFFC))
1093 // which results in a constant-extended and(##...,lsr). This transformation
1094 // undoes this simplification for cases where the shl can be folded into
1095 // an addressing mode.
1096 void HexagonDAGToDAGISel::ppAddrRewriteAndSrl(std::vector<SDNode*> &&Nodes) {
1097   SelectionDAG &DAG = *CurDAG;
1098 
1099   for (SDNode *N : Nodes) {
1100     unsigned Opc = N->getOpcode();
1101     if (Opc != ISD::LOAD && Opc != ISD::STORE)
1102       continue;
1103     SDValue Addr = Opc == ISD::LOAD ? N->getOperand(1) : N->getOperand(2);
1104     // Addr must match: (add x T0)
1105     if (Addr.getOpcode() != ISD::ADD)
1106       continue;
1107     SDValue T0 = Addr.getOperand(1);
1108     // T0 must match: (and T1 Mask)
1109     if (T0.getOpcode() != ISD::AND)
1110       continue;
1111 
1112     // We have an AND.
1113     //
1114     // Check the first operand. It must be: (srl y c).
1115     SDValue S = T0.getOperand(0);
1116     if (S.getOpcode() != ISD::SRL)
1117       continue;
1118     ConstantSDNode *SN = dyn_cast<ConstantSDNode>(S.getOperand(1).getNode());
1119     if (SN == nullptr)
1120       continue;
1121     if (SN->getAPIntValue().getBitWidth() != 32)
1122       continue;
1123     uint32_t CV = SN->getZExtValue();
1124 
1125     // Check the second operand: the supposed mask.
1126     ConstantSDNode *MN = dyn_cast<ConstantSDNode>(T0.getOperand(1).getNode());
1127     if (MN == nullptr)
1128       continue;
1129     if (MN->getAPIntValue().getBitWidth() != 32)
1130       continue;
1131     uint32_t Mask = MN->getZExtValue();
1132     // Examine the mask.
1133     uint32_t TZ = countTrailingZeros(Mask);
1134     uint32_t M1 = countTrailingOnes(Mask >> TZ);
1135     uint32_t LZ = countLeadingZeros(Mask);
1136     // Trailing zeros + middle ones + leading zeros must equal the width.
1137     if (TZ + M1 + LZ != 32)
1138       continue;
1139     // The number of trailing zeros will be encoded in the addressing mode.
1140     if (TZ > 2)
1141       continue;
1142     // The number of leading zeros must be at most c.
1143     if (LZ > CV)
1144       continue;
1145 
1146     // All looks good.
1147     SDValue Y = S.getOperand(0);
1148     EVT VT = Addr.getValueType();
1149     SDLoc dl(S);
1150     // TZ = D-C, so D = TZ+C.
1151     SDValue D = DAG.getConstant(TZ+CV, dl, VT);
1152     SDValue DC = DAG.getConstant(TZ, dl, VT);
1153     SDValue NewSrl = DAG.getNode(ISD::SRL, dl, VT, Y, D);
1154     SDValue NewShl = DAG.getNode(ISD::SHL, dl, VT, NewSrl, DC);
1155     ReplaceNode(T0.getNode(), NewShl.getNode());
1156   }
1157 }
1158 
1159 // Transform: (op ... (zext i1 c) ...) -> (select c (op ... 0 ...)
1160 //                                                  (op ... 1 ...))
1161 void HexagonDAGToDAGISel::ppHoistZextI1(std::vector<SDNode*> &&Nodes) {
1162   SelectionDAG &DAG = *CurDAG;
1163 
1164   for (SDNode *N : Nodes) {
1165     unsigned Opc = N->getOpcode();
1166     if (Opc != ISD::ZERO_EXTEND)
1167       continue;
1168     SDValue OpI1 = N->getOperand(0);
1169     EVT OpVT = OpI1.getValueType();
1170     if (!OpVT.isSimple() || OpVT.getSimpleVT() != MVT::i1)
1171       continue;
1172     for (auto I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1173       SDNode *U = *I;
1174       if (U->getNumValues() != 1)
1175         continue;
1176       EVT UVT = U->getValueType(0);
1177       if (!UVT.isSimple() || !UVT.isInteger() || UVT.getSimpleVT() == MVT::i1)
1178         continue;
1179       // Do not generate select for all i1 vector type.
1180       if (UVT.isVector() && UVT.getVectorElementType() == MVT::i1)
1181         continue;
1182       if (isMemOPCandidate(N, U))
1183         continue;
1184 
1185       // Potentially simplifiable operation.
1186       unsigned I1N = I.getOperandNo();
1187       SmallVector<SDValue,2> Ops(U->getNumOperands());
1188       for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i)
1189         Ops[i] = U->getOperand(i);
1190       EVT BVT = Ops[I1N].getValueType();
1191 
1192       const SDLoc &dl(U);
1193       SDValue C0 = DAG.getConstant(0, dl, BVT);
1194       SDValue C1 = DAG.getConstant(1, dl, BVT);
1195       SDValue If0, If1;
1196 
1197       if (isa<MachineSDNode>(U)) {
1198         unsigned UseOpc = U->getMachineOpcode();
1199         Ops[I1N] = C0;
1200         If0 = SDValue(DAG.getMachineNode(UseOpc, dl, UVT, Ops), 0);
1201         Ops[I1N] = C1;
1202         If1 = SDValue(DAG.getMachineNode(UseOpc, dl, UVT, Ops), 0);
1203       } else {
1204         unsigned UseOpc = U->getOpcode();
1205         Ops[I1N] = C0;
1206         If0 = DAG.getNode(UseOpc, dl, UVT, Ops);
1207         Ops[I1N] = C1;
1208         If1 = DAG.getNode(UseOpc, dl, UVT, Ops);
1209       }
1210       // We're generating a SELECT way after legalization, so keep the types
1211       // simple.
1212       unsigned UW = UVT.getSizeInBits();
1213       EVT SVT = (UW == 32 || UW == 64) ? MVT::getIntegerVT(UW) : UVT;
1214       SDValue Sel = DAG.getNode(ISD::SELECT, dl, SVT, OpI1,
1215                                 DAG.getBitcast(SVT, If1),
1216                                 DAG.getBitcast(SVT, If0));
1217       SDValue Ret = DAG.getBitcast(UVT, Sel);
1218       DAG.ReplaceAllUsesWith(U, Ret.getNode());
1219     }
1220   }
1221 }
1222 
1223 void HexagonDAGToDAGISel::PreprocessISelDAG() {
1224   // Repack all nodes before calling each preprocessing function,
1225   // because each of them can modify the set of nodes.
1226   auto getNodes = [this] () -> std::vector<SDNode*> {
1227     std::vector<SDNode*> T;
1228     T.reserve(CurDAG->allnodes_size());
1229     for (SDNode &N : CurDAG->allnodes())
1230       T.push_back(&N);
1231     return T;
1232   };
1233 
1234   // Transform: (or (select c x 0) z)  ->  (select c (or x z) z)
1235   //            (or (select c 0 y) z)  ->  (select c z (or y z))
1236   ppSimplifyOrSelect0(getNodes());
1237 
1238   // Transform: (store ch val (add x (add (shl y c) e)))
1239   //        to: (store ch val (add x (shl (add y d) c))),
1240   // where e = (shl d c) for some integer d.
1241   // The purpose of this is to enable generation of loads/stores with
1242   // shifted addressing mode, i.e. mem(x+y<<#c). For that, the shift
1243   // value c must be 0, 1 or 2.
1244   ppAddrReorderAddShl(getNodes());
1245 
1246   // Transform: (load ch (add x (and (srl y c) Mask)))
1247   //        to: (load ch (add x (shl (srl y d) d-c)))
1248   // where
1249   // Mask = 00..0 111..1 0.0
1250   //          |     |     +-- d-c 0s, and d-c is 0, 1 or 2.
1251   //          |     +-------- 1s
1252   //          +-------------- at most c 0s
1253   // Motivating example:
1254   // DAG combiner optimizes (add x (shl (srl y 5) 2))
1255   //                     to (add x (and (srl y 3) 1FFFFFFC))
1256   // which results in a constant-extended and(##...,lsr). This transformation
1257   // undoes this simplification for cases where the shl can be folded into
1258   // an addressing mode.
1259   ppAddrRewriteAndSrl(getNodes());
1260 
1261   // Transform: (op ... (zext i1 c) ...) -> (select c (op ... 0 ...)
1262   //                                                  (op ... 1 ...))
1263   ppHoistZextI1(getNodes());
1264 
1265   DEBUG_WITH_TYPE("isel", {
1266     dbgs() << "Preprocessed (Hexagon) selection DAG:";
1267     CurDAG->dump();
1268   });
1269 
1270   if (EnableAddressRebalancing) {
1271     rebalanceAddressTrees();
1272 
1273     DEBUG_WITH_TYPE("isel", {
1274       dbgs() << "Address tree balanced selection DAG:";
1275       CurDAG->dump();
1276     });
1277   }
1278 }
1279 
1280 void HexagonDAGToDAGISel::emitFunctionEntryCode() {
1281   auto &HST = MF->getSubtarget<HexagonSubtarget>();
1282   auto &HFI = *HST.getFrameLowering();
1283   if (!HFI.needsAligna(*MF))
1284     return;
1285 
1286   MachineFrameInfo &MFI = MF->getFrameInfo();
1287   MachineBasicBlock *EntryBB = &MF->front();
1288   Register AR = FuncInfo->CreateReg(MVT::i32);
1289   Align EntryMaxA = MFI.getMaxAlign();
1290   BuildMI(EntryBB, DebugLoc(), HII->get(Hexagon::PS_aligna), AR)
1291       .addImm(EntryMaxA.value());
1292   MF->getInfo<HexagonMachineFunctionInfo>()->setStackAlignBaseVReg(AR);
1293 }
1294 
1295 void HexagonDAGToDAGISel::updateAligna() {
1296   auto &HFI = *MF->getSubtarget<HexagonSubtarget>().getFrameLowering();
1297   if (!HFI.needsAligna(*MF))
1298     return;
1299   auto *AlignaI = const_cast<MachineInstr*>(HFI.getAlignaInstr(*MF));
1300   assert(AlignaI != nullptr);
1301   unsigned MaxA = MF->getFrameInfo().getMaxAlign().value();
1302   if (AlignaI->getOperand(1).getImm() < MaxA)
1303     AlignaI->getOperand(1).setImm(MaxA);
1304 }
1305 
1306 // Match a frame index that can be used in an addressing mode.
1307 bool HexagonDAGToDAGISel::SelectAddrFI(SDValue &N, SDValue &R) {
1308   if (N.getOpcode() != ISD::FrameIndex)
1309     return false;
1310   auto &HFI = *HST->getFrameLowering();
1311   MachineFrameInfo &MFI = MF->getFrameInfo();
1312   int FX = cast<FrameIndexSDNode>(N)->getIndex();
1313   if (!MFI.isFixedObjectIndex(FX) && HFI.needsAligna(*MF))
1314     return false;
1315   R = CurDAG->getTargetFrameIndex(FX, MVT::i32);
1316   return true;
1317 }
1318 
1319 inline bool HexagonDAGToDAGISel::SelectAddrGA(SDValue &N, SDValue &R) {
1320   return SelectGlobalAddress(N, R, false, Align(1));
1321 }
1322 
1323 inline bool HexagonDAGToDAGISel::SelectAddrGP(SDValue &N, SDValue &R) {
1324   return SelectGlobalAddress(N, R, true, Align(1));
1325 }
1326 
1327 inline bool HexagonDAGToDAGISel::SelectAnyImm(SDValue &N, SDValue &R) {
1328   return SelectAnyImmediate(N, R, Align(1));
1329 }
1330 
1331 inline bool HexagonDAGToDAGISel::SelectAnyImm0(SDValue &N, SDValue &R) {
1332   return SelectAnyImmediate(N, R, Align(1));
1333 }
1334 inline bool HexagonDAGToDAGISel::SelectAnyImm1(SDValue &N, SDValue &R) {
1335   return SelectAnyImmediate(N, R, Align(2));
1336 }
1337 inline bool HexagonDAGToDAGISel::SelectAnyImm2(SDValue &N, SDValue &R) {
1338   return SelectAnyImmediate(N, R, Align(4));
1339 }
1340 inline bool HexagonDAGToDAGISel::SelectAnyImm3(SDValue &N, SDValue &R) {
1341   return SelectAnyImmediate(N, R, Align(8));
1342 }
1343 
1344 inline bool HexagonDAGToDAGISel::SelectAnyInt(SDValue &N, SDValue &R) {
1345   EVT T = N.getValueType();
1346   if (!T.isInteger() || T.getSizeInBits() != 32 || !isa<ConstantSDNode>(N))
1347     return false;
1348   R = N;
1349   return true;
1350 }
1351 
1352 bool HexagonDAGToDAGISel::SelectAnyImmediate(SDValue &N, SDValue &R,
1353                                              Align Alignment) {
1354   switch (N.getOpcode()) {
1355   case ISD::Constant: {
1356     if (N.getValueType() != MVT::i32)
1357       return false;
1358     int32_t V = cast<const ConstantSDNode>(N)->getZExtValue();
1359     if (!isAligned(Alignment, V))
1360       return false;
1361     R = CurDAG->getTargetConstant(V, SDLoc(N), N.getValueType());
1362     return true;
1363   }
1364   case HexagonISD::JT:
1365   case HexagonISD::CP:
1366     // These are assumed to always be aligned at least 8-byte boundary.
1367     if (Alignment > Align(8))
1368       return false;
1369     R = N.getOperand(0);
1370     return true;
1371   case ISD::ExternalSymbol:
1372     // Symbols may be aligned at any boundary.
1373     if (Alignment > Align(1))
1374       return false;
1375     R = N;
1376     return true;
1377   case ISD::BlockAddress:
1378     // Block address is always aligned at least 4-byte boundary.
1379     if (Alignment > Align(4) ||
1380         !isAligned(Alignment, cast<BlockAddressSDNode>(N)->getOffset()))
1381       return false;
1382     R = N;
1383     return true;
1384   }
1385 
1386   if (SelectGlobalAddress(N, R, false, Alignment) ||
1387       SelectGlobalAddress(N, R, true, Alignment))
1388     return true;
1389 
1390   return false;
1391 }
1392 
1393 bool HexagonDAGToDAGISel::SelectGlobalAddress(SDValue &N, SDValue &R,
1394                                               bool UseGP, Align Alignment) {
1395   switch (N.getOpcode()) {
1396   case ISD::ADD: {
1397     SDValue N0 = N.getOperand(0);
1398     SDValue N1 = N.getOperand(1);
1399     unsigned GAOpc = N0.getOpcode();
1400     if (UseGP && GAOpc != HexagonISD::CONST32_GP)
1401       return false;
1402     if (!UseGP && GAOpc != HexagonISD::CONST32)
1403       return false;
1404     if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N1)) {
1405       if (!isAligned(Alignment, Const->getZExtValue()))
1406         return false;
1407       SDValue Addr = N0.getOperand(0);
1408       if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Addr)) {
1409         if (GA->getOpcode() == ISD::TargetGlobalAddress) {
1410           uint64_t NewOff = GA->getOffset() + (uint64_t)Const->getSExtValue();
1411           R = CurDAG->getTargetGlobalAddress(GA->getGlobal(), SDLoc(Const),
1412                                              N.getValueType(), NewOff);
1413           return true;
1414         }
1415       }
1416     }
1417     break;
1418   }
1419   case HexagonISD::CP:
1420   case HexagonISD::JT:
1421   case HexagonISD::CONST32:
1422     // The operand(0) of CONST32 is TargetGlobalAddress, which is what we
1423     // want in the instruction.
1424     if (!UseGP)
1425       R = N.getOperand(0);
1426     return !UseGP;
1427   case HexagonISD::CONST32_GP:
1428     if (UseGP)
1429       R = N.getOperand(0);
1430     return UseGP;
1431   default:
1432     return false;
1433   }
1434 
1435   return false;
1436 }
1437 
1438 bool HexagonDAGToDAGISel::DetectUseSxtw(SDValue &N, SDValue &R) {
1439   // This (complex pattern) function is meant to detect a sign-extension
1440   // i32->i64 on a per-operand basis. This would allow writing single
1441   // patterns that would cover a number of combinations of different ways
1442   // a sign-extensions could be written. For example:
1443   //   (mul (DetectUseSxtw x) (DetectUseSxtw y)) -> (M2_dpmpyss_s0 x y)
1444   // could match either one of these:
1445   //   (mul (sext x) (sext_inreg y))
1446   //   (mul (sext-load *p) (sext_inreg y))
1447   //   (mul (sext_inreg x) (sext y))
1448   // etc.
1449   //
1450   // The returned value will have type i64 and its low word will
1451   // contain the value being extended. The high bits are not specified.
1452   // The returned type is i64 because the original type of N was i64,
1453   // but the users of this function should only use the low-word of the
1454   // result, e.g.
1455   //  (mul sxtw:x, sxtw:y) -> (M2_dpmpyss_s0 (LoReg sxtw:x), (LoReg sxtw:y))
1456 
1457   if (N.getValueType() != MVT::i64)
1458     return false;
1459   unsigned Opc = N.getOpcode();
1460   switch (Opc) {
1461     case ISD::SIGN_EXTEND:
1462     case ISD::SIGN_EXTEND_INREG: {
1463       // sext_inreg has the source type as a separate operand.
1464       EVT T = Opc == ISD::SIGN_EXTEND
1465                 ? N.getOperand(0).getValueType()
1466                 : cast<VTSDNode>(N.getOperand(1))->getVT();
1467       unsigned SW = T.getSizeInBits();
1468       if (SW == 32)
1469         R = N.getOperand(0);
1470       else if (SW < 32)
1471         R = N;
1472       else
1473         return false;
1474       break;
1475     }
1476     case ISD::LOAD: {
1477       LoadSDNode *L = cast<LoadSDNode>(N);
1478       if (L->getExtensionType() != ISD::SEXTLOAD)
1479         return false;
1480       // All extending loads extend to i32, so even if the value in
1481       // memory is shorter than 32 bits, it will be i32 after the load.
1482       if (L->getMemoryVT().getSizeInBits() > 32)
1483         return false;
1484       R = N;
1485       break;
1486     }
1487     case ISD::SRA: {
1488       auto *S = dyn_cast<ConstantSDNode>(N.getOperand(1));
1489       if (!S || S->getZExtValue() != 32)
1490         return false;
1491       R = N;
1492       break;
1493     }
1494     default:
1495       return false;
1496   }
1497   EVT RT = R.getValueType();
1498   if (RT == MVT::i64)
1499     return true;
1500   assert(RT == MVT::i32);
1501   // This is only to produce a value of type i64. Do not rely on the
1502   // high bits produced by this.
1503   const SDLoc &dl(N);
1504   SDValue Ops[] = {
1505     CurDAG->getTargetConstant(Hexagon::DoubleRegsRegClassID, dl, MVT::i32),
1506     R, CurDAG->getTargetConstant(Hexagon::isub_hi, dl, MVT::i32),
1507     R, CurDAG->getTargetConstant(Hexagon::isub_lo, dl, MVT::i32)
1508   };
1509   SDNode *T = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl,
1510                                      MVT::i64, Ops);
1511   R = SDValue(T, 0);
1512   return true;
1513 }
1514 
1515 bool HexagonDAGToDAGISel::keepsLowBits(const SDValue &Val, unsigned NumBits,
1516       SDValue &Src) {
1517   unsigned Opc = Val.getOpcode();
1518   switch (Opc) {
1519   case ISD::SIGN_EXTEND:
1520   case ISD::ZERO_EXTEND:
1521   case ISD::ANY_EXTEND: {
1522     const SDValue &Op0 = Val.getOperand(0);
1523     EVT T = Op0.getValueType();
1524     if (T.isInteger() && T.getSizeInBits() == NumBits) {
1525       Src = Op0;
1526       return true;
1527     }
1528     break;
1529   }
1530   case ISD::SIGN_EXTEND_INREG:
1531   case ISD::AssertSext:
1532   case ISD::AssertZext:
1533     if (Val.getOperand(0).getValueType().isInteger()) {
1534       VTSDNode *T = cast<VTSDNode>(Val.getOperand(1));
1535       if (T->getVT().getSizeInBits() == NumBits) {
1536         Src = Val.getOperand(0);
1537         return true;
1538       }
1539     }
1540     break;
1541   case ISD::AND: {
1542     // Check if this is an AND with NumBits of lower bits set to 1.
1543     uint64_t Mask = (1 << NumBits) - 1;
1544     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1545       if (C->getZExtValue() == Mask) {
1546         Src = Val.getOperand(1);
1547         return true;
1548       }
1549     }
1550     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1551       if (C->getZExtValue() == Mask) {
1552         Src = Val.getOperand(0);
1553         return true;
1554       }
1555     }
1556     break;
1557   }
1558   case ISD::OR:
1559   case ISD::XOR: {
1560     // OR/XOR with the lower NumBits bits set to 0.
1561     uint64_t Mask = (1 << NumBits) - 1;
1562     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1563       if ((C->getZExtValue() & Mask) == 0) {
1564         Src = Val.getOperand(1);
1565         return true;
1566       }
1567     }
1568     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1569       if ((C->getZExtValue() & Mask) == 0) {
1570         Src = Val.getOperand(0);
1571         return true;
1572       }
1573     }
1574     break;
1575   }
1576   default:
1577     break;
1578   }
1579   return false;
1580 }
1581 
1582 bool HexagonDAGToDAGISel::isAlignedMemNode(const MemSDNode *N) const {
1583   return N->getAlignment() >= N->getMemoryVT().getStoreSize();
1584 }
1585 
1586 bool HexagonDAGToDAGISel::isSmallStackStore(const StoreSDNode *N) const {
1587   unsigned StackSize = MF->getFrameInfo().estimateStackSize(*MF);
1588   switch (N->getMemoryVT().getStoreSize()) {
1589     case 1:
1590       return StackSize <= 56;   // 1*2^6 - 8
1591     case 2:
1592       return StackSize <= 120;  // 2*2^6 - 8
1593     case 4:
1594       return StackSize <= 248;  // 4*2^6 - 8
1595     default:
1596       return false;
1597   }
1598 }
1599 
1600 // Return true when the given node fits in a positive half word.
1601 bool HexagonDAGToDAGISel::isPositiveHalfWord(const SDNode *N) const {
1602   if (const ConstantSDNode *CN = dyn_cast<const ConstantSDNode>(N)) {
1603     int64_t V = CN->getSExtValue();
1604     return V > 0 && isInt<16>(V);
1605   }
1606   if (N->getOpcode() == ISD::SIGN_EXTEND_INREG) {
1607     const VTSDNode *VN = dyn_cast<const VTSDNode>(N->getOperand(1));
1608     return VN->getVT().getSizeInBits() <= 16;
1609   }
1610   return false;
1611 }
1612 
1613 bool HexagonDAGToDAGISel::hasOneUse(const SDNode *N) const {
1614   return !CheckSingleUse || N->hasOneUse();
1615 }
1616 
1617 ////////////////////////////////////////////////////////////////////////////////
1618 // Rebalancing of address calculation trees
1619 
1620 static bool isOpcodeHandled(const SDNode *N) {
1621   switch (N->getOpcode()) {
1622     case ISD::ADD:
1623     case ISD::MUL:
1624       return true;
1625     case ISD::SHL:
1626       // We only handle constant shifts because these can be easily flattened
1627       // into multiplications by 2^Op1.
1628       return isa<ConstantSDNode>(N->getOperand(1).getNode());
1629     default:
1630       return false;
1631   }
1632 }
1633 
1634 /// Return the weight of an SDNode
1635 int HexagonDAGToDAGISel::getWeight(SDNode *N) {
1636   if (!isOpcodeHandled(N))
1637     return 1;
1638   assert(RootWeights.count(N) && "Cannot get weight of unseen root!");
1639   assert(RootWeights[N] != -1 && "Cannot get weight of unvisited root!");
1640   assert(RootWeights[N] != -2 && "Cannot get weight of RAWU'd root!");
1641   return RootWeights[N];
1642 }
1643 
1644 int HexagonDAGToDAGISel::getHeight(SDNode *N) {
1645   if (!isOpcodeHandled(N))
1646     return 0;
1647   assert(RootWeights.count(N) && RootWeights[N] >= 0 &&
1648       "Cannot query height of unvisited/RAUW'd node!");
1649   return RootHeights[N];
1650 }
1651 
1652 namespace {
1653 struct WeightedLeaf {
1654   SDValue Value;
1655   int Weight;
1656   int InsertionOrder;
1657 
1658   WeightedLeaf() : Value(SDValue()) { }
1659 
1660   WeightedLeaf(SDValue Value, int Weight, int InsertionOrder) :
1661     Value(Value), Weight(Weight), InsertionOrder(InsertionOrder) {
1662     assert(Weight >= 0 && "Weight must be >= 0");
1663   }
1664 
1665   static bool Compare(const WeightedLeaf &A, const WeightedLeaf &B) {
1666     assert(A.Value.getNode() && B.Value.getNode());
1667     return A.Weight == B.Weight ?
1668             (A.InsertionOrder > B.InsertionOrder) :
1669             (A.Weight > B.Weight);
1670   }
1671 };
1672 
1673 /// A specialized priority queue for WeigthedLeaves. It automatically folds
1674 /// constants and allows removal of non-top elements while maintaining the
1675 /// priority order.
1676 class LeafPrioQueue {
1677   SmallVector<WeightedLeaf, 8> Q;
1678   bool HaveConst;
1679   WeightedLeaf ConstElt;
1680   unsigned Opcode;
1681 
1682 public:
1683   bool empty() {
1684     return (!HaveConst && Q.empty());
1685   }
1686 
1687   size_t size() {
1688     return Q.size() + HaveConst;
1689   }
1690 
1691   bool hasConst() {
1692     return HaveConst;
1693   }
1694 
1695   const WeightedLeaf &top() {
1696     if (HaveConst)
1697       return ConstElt;
1698     return Q.front();
1699   }
1700 
1701   WeightedLeaf pop() {
1702     if (HaveConst) {
1703       HaveConst = false;
1704       return ConstElt;
1705     }
1706     std::pop_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1707     return Q.pop_back_val();
1708   }
1709 
1710   void push(WeightedLeaf L, bool SeparateConst=true) {
1711     if (!HaveConst && SeparateConst && isa<ConstantSDNode>(L.Value)) {
1712       if (Opcode == ISD::MUL &&
1713           cast<ConstantSDNode>(L.Value)->getSExtValue() == 1)
1714         return;
1715       if (Opcode == ISD::ADD &&
1716           cast<ConstantSDNode>(L.Value)->getSExtValue() == 0)
1717         return;
1718 
1719       HaveConst = true;
1720       ConstElt = L;
1721     } else {
1722       Q.push_back(L);
1723       std::push_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1724     }
1725   }
1726 
1727   /// Push L to the bottom of the queue regardless of its weight. If L is
1728   /// constant, it will not be folded with other constants in the queue.
1729   void pushToBottom(WeightedLeaf L) {
1730     L.Weight = 1000;
1731     push(L, false);
1732   }
1733 
1734   /// Search for a SHL(x, [<=MaxAmount]) subtree in the queue, return the one of
1735   /// lowest weight and remove it from the queue.
1736   WeightedLeaf findSHL(uint64_t MaxAmount);
1737 
1738   WeightedLeaf findMULbyConst();
1739 
1740   LeafPrioQueue(unsigned Opcode) :
1741     HaveConst(false), Opcode(Opcode) { }
1742 };
1743 } // end anonymous namespace
1744 
1745 WeightedLeaf LeafPrioQueue::findSHL(uint64_t MaxAmount) {
1746   int ResultPos;
1747   WeightedLeaf Result;
1748 
1749   for (int Pos = 0, End = Q.size(); Pos != End; ++Pos) {
1750     const WeightedLeaf &L = Q[Pos];
1751     const SDValue &Val = L.Value;
1752     if (Val.getOpcode() != ISD::SHL ||
1753         !isa<ConstantSDNode>(Val.getOperand(1)) ||
1754         Val.getConstantOperandVal(1) > MaxAmount)
1755       continue;
1756     if (!Result.Value.getNode() || Result.Weight > L.Weight ||
1757         (Result.Weight == L.Weight && Result.InsertionOrder > L.InsertionOrder))
1758     {
1759       Result = L;
1760       ResultPos = Pos;
1761     }
1762   }
1763 
1764   if (Result.Value.getNode()) {
1765     Q.erase(&Q[ResultPos]);
1766     std::make_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1767   }
1768 
1769   return Result;
1770 }
1771 
1772 WeightedLeaf LeafPrioQueue::findMULbyConst() {
1773   int ResultPos;
1774   WeightedLeaf Result;
1775 
1776   for (int Pos = 0, End = Q.size(); Pos != End; ++Pos) {
1777     const WeightedLeaf &L = Q[Pos];
1778     const SDValue &Val = L.Value;
1779     if (Val.getOpcode() != ISD::MUL ||
1780         !isa<ConstantSDNode>(Val.getOperand(1)) ||
1781         Val.getConstantOperandVal(1) > 127)
1782       continue;
1783     if (!Result.Value.getNode() || Result.Weight > L.Weight ||
1784         (Result.Weight == L.Weight && Result.InsertionOrder > L.InsertionOrder))
1785     {
1786       Result = L;
1787       ResultPos = Pos;
1788     }
1789   }
1790 
1791   if (Result.Value.getNode()) {
1792     Q.erase(&Q[ResultPos]);
1793     std::make_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1794   }
1795 
1796   return Result;
1797 }
1798 
1799 SDValue HexagonDAGToDAGISel::getMultiplierForSHL(SDNode *N) {
1800   uint64_t MulFactor = 1ull << N->getConstantOperandVal(1);
1801   return CurDAG->getConstant(MulFactor, SDLoc(N),
1802                              N->getOperand(1).getValueType());
1803 }
1804 
1805 /// @returns the value x for which 2^x is a factor of Val
1806 static unsigned getPowerOf2Factor(SDValue Val) {
1807   if (Val.getOpcode() == ISD::MUL) {
1808     unsigned MaxFactor = 0;
1809     for (int i = 0; i < 2; ++i) {
1810       ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(i));
1811       if (!C)
1812         continue;
1813       const APInt &CInt = C->getAPIntValue();
1814       if (CInt.getBoolValue())
1815         MaxFactor = CInt.countTrailingZeros();
1816     }
1817     return MaxFactor;
1818   }
1819   if (Val.getOpcode() == ISD::SHL) {
1820     if (!isa<ConstantSDNode>(Val.getOperand(1).getNode()))
1821       return 0;
1822     return (unsigned) Val.getConstantOperandVal(1);
1823   }
1824 
1825   return 0;
1826 }
1827 
1828 /// @returns true if V>>Amount will eliminate V's operation on its child
1829 static bool willShiftRightEliminate(SDValue V, unsigned Amount) {
1830   if (V.getOpcode() == ISD::MUL) {
1831     SDValue Ops[] = { V.getOperand(0), V.getOperand(1) };
1832     for (int i = 0; i < 2; ++i)
1833       if (isa<ConstantSDNode>(Ops[i].getNode()) &&
1834           V.getConstantOperandVal(i) % (1ULL << Amount) == 0) {
1835         uint64_t NewConst = V.getConstantOperandVal(i) >> Amount;
1836         return (NewConst == 1);
1837       }
1838   } else if (V.getOpcode() == ISD::SHL) {
1839     return (Amount == V.getConstantOperandVal(1));
1840   }
1841 
1842   return false;
1843 }
1844 
1845 SDValue HexagonDAGToDAGISel::factorOutPowerOf2(SDValue V, unsigned Power) {
1846   SDValue Ops[] = { V.getOperand(0), V.getOperand(1) };
1847   if (V.getOpcode() == ISD::MUL) {
1848     for (int i=0; i < 2; ++i) {
1849       if (isa<ConstantSDNode>(Ops[i].getNode()) &&
1850           V.getConstantOperandVal(i) % ((uint64_t)1 << Power) == 0) {
1851         uint64_t NewConst = V.getConstantOperandVal(i) >> Power;
1852         if (NewConst == 1)
1853           return Ops[!i];
1854         Ops[i] = CurDAG->getConstant(NewConst,
1855                                      SDLoc(V), V.getValueType());
1856         break;
1857       }
1858     }
1859   } else if (V.getOpcode() == ISD::SHL) {
1860     uint64_t ShiftAmount = V.getConstantOperandVal(1);
1861     if (ShiftAmount == Power)
1862       return Ops[0];
1863     Ops[1] = CurDAG->getConstant(ShiftAmount - Power,
1864                                  SDLoc(V), V.getValueType());
1865   }
1866 
1867   return CurDAG->getNode(V.getOpcode(), SDLoc(V), V.getValueType(), Ops);
1868 }
1869 
1870 static bool isTargetConstant(const SDValue &V) {
1871   return V.getOpcode() == HexagonISD::CONST32 ||
1872          V.getOpcode() == HexagonISD::CONST32_GP;
1873 }
1874 
1875 unsigned HexagonDAGToDAGISel::getUsesInFunction(const Value *V) {
1876   if (GAUsesInFunction.count(V))
1877     return GAUsesInFunction[V];
1878 
1879   unsigned Result = 0;
1880   const Function &CurF = CurDAG->getMachineFunction().getFunction();
1881   for (const User *U : V->users()) {
1882     if (isa<Instruction>(U) &&
1883         cast<Instruction>(U)->getParent()->getParent() == &CurF)
1884       ++Result;
1885   }
1886 
1887   GAUsesInFunction[V] = Result;
1888 
1889   return Result;
1890 }
1891 
1892 /// Note - After calling this, N may be dead. It may have been replaced by a
1893 /// new node, so always use the returned value in place of N.
1894 ///
1895 /// @returns The SDValue taking the place of N (which could be N if it is
1896 /// unchanged)
1897 SDValue HexagonDAGToDAGISel::balanceSubTree(SDNode *N, bool TopLevel) {
1898   assert(RootWeights.count(N) && "Cannot balance non-root node.");
1899   assert(RootWeights[N] != -2 && "This node was RAUW'd!");
1900   assert(!TopLevel || N->getOpcode() == ISD::ADD);
1901 
1902   // Return early if this node was already visited
1903   if (RootWeights[N] != -1)
1904     return SDValue(N, 0);
1905 
1906   assert(isOpcodeHandled(N));
1907 
1908   SDValue Op0 = N->getOperand(0);
1909   SDValue Op1 = N->getOperand(1);
1910 
1911   // Return early if the operands will remain unchanged or are all roots
1912   if ((!isOpcodeHandled(Op0.getNode()) || RootWeights.count(Op0.getNode())) &&
1913       (!isOpcodeHandled(Op1.getNode()) || RootWeights.count(Op1.getNode()))) {
1914     SDNode *Op0N = Op0.getNode();
1915     int Weight;
1916     if (isOpcodeHandled(Op0N) && RootWeights[Op0N] == -1) {
1917       Weight = getWeight(balanceSubTree(Op0N).getNode());
1918       // Weight = calculateWeight(Op0N);
1919     } else
1920       Weight = getWeight(Op0N);
1921 
1922     SDNode *Op1N = N->getOperand(1).getNode(); // Op1 may have been RAUWd
1923     if (isOpcodeHandled(Op1N) && RootWeights[Op1N] == -1) {
1924       Weight += getWeight(balanceSubTree(Op1N).getNode());
1925       // Weight += calculateWeight(Op1N);
1926     } else
1927       Weight += getWeight(Op1N);
1928 
1929     RootWeights[N] = Weight;
1930     RootHeights[N] = std::max(getHeight(N->getOperand(0).getNode()),
1931                               getHeight(N->getOperand(1).getNode())) + 1;
1932 
1933     LLVM_DEBUG(dbgs() << "--> No need to balance root (Weight=" << Weight
1934                       << " Height=" << RootHeights[N] << "): ");
1935     LLVM_DEBUG(N->dump(CurDAG));
1936 
1937     return SDValue(N, 0);
1938   }
1939 
1940   LLVM_DEBUG(dbgs() << "** Balancing root node: ");
1941   LLVM_DEBUG(N->dump(CurDAG));
1942 
1943   unsigned NOpcode = N->getOpcode();
1944 
1945   LeafPrioQueue Leaves(NOpcode);
1946   SmallVector<SDValue, 4> Worklist;
1947   Worklist.push_back(SDValue(N, 0));
1948 
1949   // SHL nodes will be converted to MUL nodes
1950   if (NOpcode == ISD::SHL)
1951     NOpcode = ISD::MUL;
1952 
1953   bool CanFactorize = false;
1954   WeightedLeaf Mul1, Mul2;
1955   unsigned MaxPowerOf2 = 0;
1956   WeightedLeaf GA;
1957 
1958   // Do not try to factor out a shift if there is already a shift at the tip of
1959   // the tree.
1960   bool HaveTopLevelShift = false;
1961   if (TopLevel &&
1962       ((isOpcodeHandled(Op0.getNode()) && Op0.getOpcode() == ISD::SHL &&
1963                         Op0.getConstantOperandVal(1) < 4) ||
1964        (isOpcodeHandled(Op1.getNode()) && Op1.getOpcode() == ISD::SHL &&
1965                         Op1.getConstantOperandVal(1) < 4)))
1966     HaveTopLevelShift = true;
1967 
1968   // Flatten the subtree into an ordered list of leaves; at the same time
1969   // determine whether the tree is already balanced.
1970   int InsertionOrder = 0;
1971   SmallDenseMap<SDValue, int> NodeHeights;
1972   bool Imbalanced = false;
1973   int CurrentWeight = 0;
1974   while (!Worklist.empty()) {
1975     SDValue Child = Worklist.pop_back_val();
1976 
1977     if (Child.getNode() != N && RootWeights.count(Child.getNode())) {
1978       // CASE 1: Child is a root note
1979 
1980       int Weight = RootWeights[Child.getNode()];
1981       if (Weight == -1) {
1982         Child = balanceSubTree(Child.getNode());
1983         // calculateWeight(Child.getNode());
1984         Weight = getWeight(Child.getNode());
1985       } else if (Weight == -2) {
1986         // Whoops, this node was RAUWd by one of the balanceSubTree calls we
1987         // made. Our worklist isn't up to date anymore.
1988         // Restart the whole process.
1989         LLVM_DEBUG(dbgs() << "--> Subtree was RAUWd. Restarting...\n");
1990         return balanceSubTree(N, TopLevel);
1991       }
1992 
1993       NodeHeights[Child] = 1;
1994       CurrentWeight += Weight;
1995 
1996       unsigned PowerOf2;
1997       if (TopLevel && !CanFactorize && !HaveTopLevelShift &&
1998           (Child.getOpcode() == ISD::MUL || Child.getOpcode() == ISD::SHL) &&
1999           Child.hasOneUse() && (PowerOf2 = getPowerOf2Factor(Child))) {
2000         // Try to identify two factorizable MUL/SHL children greedily. Leave
2001         // them out of the priority queue for now so we can deal with them
2002         // after.
2003         if (!Mul1.Value.getNode()) {
2004           Mul1 = WeightedLeaf(Child, Weight, InsertionOrder++);
2005           MaxPowerOf2 = PowerOf2;
2006         } else {
2007           Mul2 = WeightedLeaf(Child, Weight, InsertionOrder++);
2008           MaxPowerOf2 = std::min(MaxPowerOf2, PowerOf2);
2009 
2010           // Our addressing modes can only shift by a maximum of 3
2011           if (MaxPowerOf2 > 3)
2012             MaxPowerOf2 = 3;
2013 
2014           CanFactorize = true;
2015         }
2016       } else
2017         Leaves.push(WeightedLeaf(Child, Weight, InsertionOrder++));
2018     } else if (!isOpcodeHandled(Child.getNode())) {
2019       // CASE 2: Child is an unhandled kind of node (e.g. constant)
2020       int Weight = getWeight(Child.getNode());
2021 
2022       NodeHeights[Child] = getHeight(Child.getNode());
2023       CurrentWeight += Weight;
2024 
2025       if (isTargetConstant(Child) && !GA.Value.getNode())
2026         GA = WeightedLeaf(Child, Weight, InsertionOrder++);
2027       else
2028         Leaves.push(WeightedLeaf(Child, Weight, InsertionOrder++));
2029     } else {
2030       // CASE 3: Child is a subtree of same opcode
2031       // Visit children first, then flatten.
2032       unsigned ChildOpcode = Child.getOpcode();
2033       assert(ChildOpcode == NOpcode ||
2034              (NOpcode == ISD::MUL && ChildOpcode == ISD::SHL));
2035 
2036       // Convert SHL to MUL
2037       SDValue Op1;
2038       if (ChildOpcode == ISD::SHL)
2039         Op1 = getMultiplierForSHL(Child.getNode());
2040       else
2041         Op1 = Child->getOperand(1);
2042 
2043       if (!NodeHeights.count(Op1) || !NodeHeights.count(Child->getOperand(0))) {
2044         assert(!NodeHeights.count(Child) && "Parent visited before children?");
2045         // Visit children first, then re-visit this node
2046         Worklist.push_back(Child);
2047         Worklist.push_back(Op1);
2048         Worklist.push_back(Child->getOperand(0));
2049       } else {
2050         // Back at this node after visiting the children
2051         if (std::abs(NodeHeights[Op1] - NodeHeights[Child->getOperand(0)]) > 1)
2052           Imbalanced = true;
2053 
2054         NodeHeights[Child] = std::max(NodeHeights[Op1],
2055                                       NodeHeights[Child->getOperand(0)]) + 1;
2056       }
2057     }
2058   }
2059 
2060   LLVM_DEBUG(dbgs() << "--> Current height=" << NodeHeights[SDValue(N, 0)]
2061                     << " weight=" << CurrentWeight
2062                     << " imbalanced=" << Imbalanced << "\n");
2063 
2064   // Transform MUL(x, C * 2^Y) + SHL(z, Y) -> SHL(ADD(MUL(x, C), z), Y)
2065   //  This factors out a shift in order to match memw(a<<Y+b).
2066   if (CanFactorize && (willShiftRightEliminate(Mul1.Value, MaxPowerOf2) ||
2067                        willShiftRightEliminate(Mul2.Value, MaxPowerOf2))) {
2068     LLVM_DEBUG(dbgs() << "--> Found common factor for two MUL children!\n");
2069     int Weight = Mul1.Weight + Mul2.Weight;
2070     int Height = std::max(NodeHeights[Mul1.Value], NodeHeights[Mul2.Value]) + 1;
2071     SDValue Mul1Factored = factorOutPowerOf2(Mul1.Value, MaxPowerOf2);
2072     SDValue Mul2Factored = factorOutPowerOf2(Mul2.Value, MaxPowerOf2);
2073     SDValue Sum = CurDAG->getNode(ISD::ADD, SDLoc(N), Mul1.Value.getValueType(),
2074                                   Mul1Factored, Mul2Factored);
2075     SDValue Const = CurDAG->getConstant(MaxPowerOf2, SDLoc(N),
2076                                         Mul1.Value.getValueType());
2077     SDValue New = CurDAG->getNode(ISD::SHL, SDLoc(N), Mul1.Value.getValueType(),
2078                                   Sum, Const);
2079     NodeHeights[New] = Height;
2080     Leaves.push(WeightedLeaf(New, Weight, Mul1.InsertionOrder));
2081   } else if (Mul1.Value.getNode()) {
2082     // We failed to factorize two MULs, so now the Muls are left outside the
2083     // queue... add them back.
2084     Leaves.push(Mul1);
2085     if (Mul2.Value.getNode())
2086       Leaves.push(Mul2);
2087     CanFactorize = false;
2088   }
2089 
2090   // Combine GA + Constant -> GA+Offset, but only if GA is not used elsewhere
2091   // and the root node itself is not used more than twice. This reduces the
2092   // amount of additional constant extenders introduced by this optimization.
2093   bool CombinedGA = false;
2094   if (NOpcode == ISD::ADD && GA.Value.getNode() && Leaves.hasConst() &&
2095       GA.Value.hasOneUse() && N->use_size() < 3) {
2096     GlobalAddressSDNode *GANode =
2097       cast<GlobalAddressSDNode>(GA.Value.getOperand(0));
2098     ConstantSDNode *Offset = cast<ConstantSDNode>(Leaves.top().Value);
2099 
2100     if (getUsesInFunction(GANode->getGlobal()) == 1 && Offset->hasOneUse() &&
2101         getTargetLowering()->isOffsetFoldingLegal(GANode)) {
2102       LLVM_DEBUG(dbgs() << "--> Combining GA and offset ("
2103                         << Offset->getSExtValue() << "): ");
2104       LLVM_DEBUG(GANode->dump(CurDAG));
2105 
2106       SDValue NewTGA =
2107         CurDAG->getTargetGlobalAddress(GANode->getGlobal(), SDLoc(GA.Value),
2108             GANode->getValueType(0),
2109             GANode->getOffset() + (uint64_t)Offset->getSExtValue());
2110       GA.Value = CurDAG->getNode(GA.Value.getOpcode(), SDLoc(GA.Value),
2111           GA.Value.getValueType(), NewTGA);
2112       GA.Weight += Leaves.top().Weight;
2113 
2114       NodeHeights[GA.Value] = getHeight(GA.Value.getNode());
2115       CombinedGA = true;
2116 
2117       Leaves.pop(); // Remove the offset constant from the queue
2118     }
2119   }
2120 
2121   if ((RebalanceOnlyForOptimizations && !CanFactorize && !CombinedGA) ||
2122       (RebalanceOnlyImbalancedTrees && !Imbalanced)) {
2123     RootWeights[N] = CurrentWeight;
2124     RootHeights[N] = NodeHeights[SDValue(N, 0)];
2125 
2126     return SDValue(N, 0);
2127   }
2128 
2129   // Combine GA + SHL(x, C<=31) so we will match Rx=add(#u8,asl(Rx,#U5))
2130   if (NOpcode == ISD::ADD && GA.Value.getNode()) {
2131     WeightedLeaf SHL = Leaves.findSHL(31);
2132     if (SHL.Value.getNode()) {
2133       int Height = std::max(NodeHeights[GA.Value], NodeHeights[SHL.Value]) + 1;
2134       GA.Value = CurDAG->getNode(ISD::ADD, SDLoc(GA.Value),
2135                                  GA.Value.getValueType(),
2136                                  GA.Value, SHL.Value);
2137       GA.Weight = SHL.Weight; // Specifically ignore the GA weight here
2138       NodeHeights[GA.Value] = Height;
2139     }
2140   }
2141 
2142   if (GA.Value.getNode())
2143     Leaves.push(GA);
2144 
2145   // If this is the top level and we haven't factored out a shift, we should try
2146   // to move a constant to the bottom to match addressing modes like memw(rX+C)
2147   if (TopLevel && !CanFactorize && Leaves.hasConst()) {
2148     LLVM_DEBUG(dbgs() << "--> Pushing constant to tip of tree.");
2149     Leaves.pushToBottom(Leaves.pop());
2150   }
2151 
2152   const DataLayout &DL = CurDAG->getDataLayout();
2153   const TargetLowering &TLI = *getTargetLowering();
2154 
2155   // Rebuild the tree using Huffman's algorithm
2156   while (Leaves.size() > 1) {
2157     WeightedLeaf L0 = Leaves.pop();
2158 
2159     // See whether we can grab a MUL to form an add(Rx,mpyi(Ry,#u6)),
2160     // otherwise just get the next leaf
2161     WeightedLeaf L1 = Leaves.findMULbyConst();
2162     if (!L1.Value.getNode())
2163       L1 = Leaves.pop();
2164 
2165     assert(L0.Weight <= L1.Weight && "Priority queue is broken!");
2166 
2167     SDValue V0 = L0.Value;
2168     int V0Weight = L0.Weight;
2169     SDValue V1 = L1.Value;
2170     int V1Weight = L1.Weight;
2171 
2172     // Make sure that none of these nodes have been RAUW'd
2173     if ((RootWeights.count(V0.getNode()) && RootWeights[V0.getNode()] == -2) ||
2174         (RootWeights.count(V1.getNode()) && RootWeights[V1.getNode()] == -2)) {
2175       LLVM_DEBUG(dbgs() << "--> Subtree was RAUWd. Restarting...\n");
2176       return balanceSubTree(N, TopLevel);
2177     }
2178 
2179     ConstantSDNode *V0C = dyn_cast<ConstantSDNode>(V0);
2180     ConstantSDNode *V1C = dyn_cast<ConstantSDNode>(V1);
2181     EVT VT = N->getValueType(0);
2182     SDValue NewNode;
2183 
2184     if (V0C && !V1C) {
2185       std::swap(V0, V1);
2186       std::swap(V0C, V1C);
2187     }
2188 
2189     // Calculate height of this node
2190     assert(NodeHeights.count(V0) && NodeHeights.count(V1) &&
2191            "Children must have been visited before re-combining them!");
2192     int Height = std::max(NodeHeights[V0], NodeHeights[V1]) + 1;
2193 
2194     // Rebuild this node (and restore SHL from MUL if needed)
2195     if (V1C && NOpcode == ISD::MUL && V1C->getAPIntValue().isPowerOf2())
2196       NewNode = CurDAG->getNode(
2197           ISD::SHL, SDLoc(V0), VT, V0,
2198           CurDAG->getConstant(
2199               V1C->getAPIntValue().logBase2(), SDLoc(N),
2200               TLI.getScalarShiftAmountTy(DL, V0.getValueType())));
2201     else
2202       NewNode = CurDAG->getNode(NOpcode, SDLoc(N), VT, V0, V1);
2203 
2204     NodeHeights[NewNode] = Height;
2205 
2206     int Weight = V0Weight + V1Weight;
2207     Leaves.push(WeightedLeaf(NewNode, Weight, L0.InsertionOrder));
2208 
2209     LLVM_DEBUG(dbgs() << "--> Built new node (Weight=" << Weight
2210                       << ",Height=" << Height << "):\n");
2211     LLVM_DEBUG(NewNode.dump());
2212   }
2213 
2214   assert(Leaves.size() == 1);
2215   SDValue NewRoot = Leaves.top().Value;
2216 
2217   assert(NodeHeights.count(NewRoot));
2218   int Height = NodeHeights[NewRoot];
2219 
2220   // Restore SHL if we earlier converted it to a MUL
2221   if (NewRoot.getOpcode() == ISD::MUL) {
2222     ConstantSDNode *V1C = dyn_cast<ConstantSDNode>(NewRoot.getOperand(1));
2223     if (V1C && V1C->getAPIntValue().isPowerOf2()) {
2224       EVT VT = NewRoot.getValueType();
2225       SDValue V0 = NewRoot.getOperand(0);
2226       NewRoot = CurDAG->getNode(
2227           ISD::SHL, SDLoc(NewRoot), VT, V0,
2228           CurDAG->getConstant(
2229               V1C->getAPIntValue().logBase2(), SDLoc(NewRoot),
2230               TLI.getScalarShiftAmountTy(DL, V0.getValueType())));
2231     }
2232   }
2233 
2234   if (N != NewRoot.getNode()) {
2235     LLVM_DEBUG(dbgs() << "--> Root is now: ");
2236     LLVM_DEBUG(NewRoot.dump());
2237 
2238     // Replace all uses of old root by new root
2239     CurDAG->ReplaceAllUsesWith(N, NewRoot.getNode());
2240     // Mark that we have RAUW'd N
2241     RootWeights[N] = -2;
2242   } else {
2243     LLVM_DEBUG(dbgs() << "--> Root unchanged.\n");
2244   }
2245 
2246   RootWeights[NewRoot.getNode()] = Leaves.top().Weight;
2247   RootHeights[NewRoot.getNode()] = Height;
2248 
2249   return NewRoot;
2250 }
2251 
2252 void HexagonDAGToDAGISel::rebalanceAddressTrees() {
2253   for (SDNode &Node : llvm::make_early_inc_range(CurDAG->allnodes())) {
2254     SDNode *N = &Node;
2255     if (N->getOpcode() != ISD::LOAD && N->getOpcode() != ISD::STORE)
2256       continue;
2257 
2258     SDValue BasePtr = cast<MemSDNode>(N)->getBasePtr();
2259     if (BasePtr.getOpcode() != ISD::ADD)
2260       continue;
2261 
2262     // We've already processed this node
2263     if (RootWeights.count(BasePtr.getNode()))
2264       continue;
2265 
2266     LLVM_DEBUG(dbgs() << "** Rebalancing address calculation in node: ");
2267     LLVM_DEBUG(N->dump(CurDAG));
2268 
2269     // FindRoots
2270     SmallVector<SDNode *, 4> Worklist;
2271 
2272     Worklist.push_back(BasePtr.getOperand(0).getNode());
2273     Worklist.push_back(BasePtr.getOperand(1).getNode());
2274 
2275     while (!Worklist.empty()) {
2276       SDNode *N = Worklist.pop_back_val();
2277       unsigned Opcode = N->getOpcode();
2278 
2279       if (!isOpcodeHandled(N))
2280         continue;
2281 
2282       Worklist.push_back(N->getOperand(0).getNode());
2283       Worklist.push_back(N->getOperand(1).getNode());
2284 
2285       // Not a root if it has only one use and same opcode as its parent
2286       if (N->hasOneUse() && Opcode == N->use_begin()->getOpcode())
2287         continue;
2288 
2289       // This root node has already been processed
2290       if (RootWeights.count(N))
2291         continue;
2292 
2293       RootWeights[N] = -1;
2294     }
2295 
2296     // Balance node itself
2297     RootWeights[BasePtr.getNode()] = -1;
2298     SDValue NewBasePtr = balanceSubTree(BasePtr.getNode(), /*TopLevel=*/ true);
2299 
2300     if (N->getOpcode() == ISD::LOAD)
2301       N = CurDAG->UpdateNodeOperands(N, N->getOperand(0),
2302             NewBasePtr, N->getOperand(2));
2303     else
2304       N = CurDAG->UpdateNodeOperands(N, N->getOperand(0), N->getOperand(1),
2305             NewBasePtr, N->getOperand(3));
2306 
2307     LLVM_DEBUG(dbgs() << "--> Final node: ");
2308     LLVM_DEBUG(N->dump(CurDAG));
2309   }
2310 
2311   CurDAG->RemoveDeadNodes();
2312   GAUsesInFunction.clear();
2313   RootHeights.clear();
2314   RootWeights.clear();
2315 }
2316