• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===-- HexagonISelDAGToDAG.cpp - A dag to dag inst selector for Hexagon --===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines an instruction selector for the Hexagon target.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "Hexagon.h"
15 #include "HexagonISelDAGToDAG.h"
16 #include "HexagonISelLowering.h"
17 #include "HexagonMachineFunctionInfo.h"
18 #include "HexagonTargetMachine.h"
19 #include "llvm/CodeGen/FunctionLoweringInfo.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/SelectionDAGISel.h"
22 #include "llvm/IR/Intrinsics.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 /// createHexagonISelDag - This pass converts a legalized DAG into a
58 /// Hexagon-specific DAG, ready for instruction scheduling.
59 ///
60 namespace llvm {
createHexagonISelDag(HexagonTargetMachine & TM,CodeGenOpt::Level OptLevel)61 FunctionPass *createHexagonISelDag(HexagonTargetMachine &TM,
62                                    CodeGenOpt::Level OptLevel) {
63   return new HexagonDAGToDAGISel(TM, OptLevel);
64 }
65 }
66 
SelectIndexedLoad(LoadSDNode * LD,const SDLoc & dl)67 void HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, const SDLoc &dl) {
68   SDValue Chain = LD->getChain();
69   SDValue Base = LD->getBasePtr();
70   SDValue Offset = LD->getOffset();
71   int32_t Inc = cast<ConstantSDNode>(Offset.getNode())->getSExtValue();
72   EVT LoadedVT = LD->getMemoryVT();
73   unsigned Opcode = 0;
74 
75   // Check for zero extended loads. Treat any-extend loads as zero extended
76   // loads.
77   ISD::LoadExtType ExtType = LD->getExtensionType();
78   bool IsZeroExt = (ExtType == ISD::ZEXTLOAD || ExtType == ISD::EXTLOAD);
79   bool IsValidInc = HII->isValidAutoIncImm(LoadedVT, Inc);
80 
81   assert(LoadedVT.isSimple());
82   switch (LoadedVT.getSimpleVT().SimpleTy) {
83   case MVT::i8:
84     if (IsZeroExt)
85       Opcode = IsValidInc ? Hexagon::L2_loadrub_pi : Hexagon::L2_loadrub_io;
86     else
87       Opcode = IsValidInc ? Hexagon::L2_loadrb_pi : Hexagon::L2_loadrb_io;
88     break;
89   case MVT::i16:
90     if (IsZeroExt)
91       Opcode = IsValidInc ? Hexagon::L2_loadruh_pi : Hexagon::L2_loadruh_io;
92     else
93       Opcode = IsValidInc ? Hexagon::L2_loadrh_pi : Hexagon::L2_loadrh_io;
94     break;
95   case MVT::i32:
96   case MVT::f32:
97   case MVT::v2i16:
98   case MVT::v4i8:
99     Opcode = IsValidInc ? Hexagon::L2_loadri_pi : Hexagon::L2_loadri_io;
100     break;
101   case MVT::i64:
102   case MVT::f64:
103   case MVT::v2i32:
104   case MVT::v4i16:
105   case MVT::v8i8:
106     Opcode = IsValidInc ? Hexagon::L2_loadrd_pi : Hexagon::L2_loadrd_io;
107     break;
108   case MVT::v64i8:
109   case MVT::v32i16:
110   case MVT::v16i32:
111   case MVT::v8i64:
112   case MVT::v128i8:
113   case MVT::v64i16:
114   case MVT::v32i32:
115   case MVT::v16i64:
116     if (isAlignedMemNode(LD)) {
117       if (LD->isNonTemporal())
118         Opcode = IsValidInc ? Hexagon::V6_vL32b_nt_pi : Hexagon::V6_vL32b_nt_ai;
119       else
120         Opcode = IsValidInc ? Hexagon::V6_vL32b_pi : Hexagon::V6_vL32b_ai;
121     } else {
122       Opcode = IsValidInc ? Hexagon::V6_vL32Ub_pi : Hexagon::V6_vL32Ub_ai;
123     }
124     break;
125   default:
126     llvm_unreachable("Unexpected memory type in indexed load");
127   }
128 
129   SDValue IncV = CurDAG->getTargetConstant(Inc, dl, MVT::i32);
130   MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
131   MemOp[0] = LD->getMemOperand();
132 
133   auto getExt64 = [this,ExtType] (MachineSDNode *N, const SDLoc &dl)
134         -> MachineSDNode* {
135     if (ExtType == ISD::ZEXTLOAD || ExtType == ISD::EXTLOAD) {
136       SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
137       return CurDAG->getMachineNode(Hexagon::A4_combineir, dl, MVT::i64,
138                                     Zero, SDValue(N, 0));
139     }
140     if (ExtType == ISD::SEXTLOAD)
141       return CurDAG->getMachineNode(Hexagon::A2_sxtw, dl, MVT::i64,
142                                     SDValue(N, 0));
143     return N;
144   };
145 
146   //                  Loaded value   Next address   Chain
147   SDValue From[3] = { SDValue(LD,0), SDValue(LD,1), SDValue(LD,2) };
148   SDValue To[3];
149 
150   EVT ValueVT = LD->getValueType(0);
151   if (ValueVT == MVT::i64 && ExtType != ISD::NON_EXTLOAD) {
152     // A load extending to i64 will actually produce i32, which will then
153     // need to be extended to i64.
154     assert(LoadedVT.getSizeInBits() <= 32);
155     ValueVT = MVT::i32;
156   }
157 
158   if (IsValidInc) {
159     MachineSDNode *L = CurDAG->getMachineNode(Opcode, dl, ValueVT,
160                                               MVT::i32, MVT::Other, Base,
161                                               IncV, Chain);
162     L->setMemRefs(MemOp, MemOp+1);
163     To[1] = SDValue(L, 1); // Next address.
164     To[2] = SDValue(L, 2); // Chain.
165     // Handle special case for extension to i64.
166     if (LD->getValueType(0) == MVT::i64)
167       L = getExt64(L, dl);
168     To[0] = SDValue(L, 0); // Loaded (extended) value.
169   } else {
170     SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
171     MachineSDNode *L = CurDAG->getMachineNode(Opcode, dl, ValueVT, MVT::Other,
172                                               Base, Zero, Chain);
173     L->setMemRefs(MemOp, MemOp+1);
174     To[2] = SDValue(L, 1); // Chain.
175     MachineSDNode *A = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
176                                               Base, IncV);
177     To[1] = SDValue(A, 0); // Next address.
178     // Handle special case for extension to i64.
179     if (LD->getValueType(0) == MVT::i64)
180       L = getExt64(L, dl);
181     To[0] = SDValue(L, 0); // Loaded (extended) value.
182   }
183   ReplaceUses(From, To, 3);
184   CurDAG->RemoveDeadNode(LD);
185 }
186 
LoadInstrForLoadIntrinsic(SDNode * IntN)187 MachineSDNode *HexagonDAGToDAGISel::LoadInstrForLoadIntrinsic(SDNode *IntN) {
188   if (IntN->getOpcode() != ISD::INTRINSIC_W_CHAIN)
189     return nullptr;
190 
191   SDLoc dl(IntN);
192   unsigned IntNo = cast<ConstantSDNode>(IntN->getOperand(1))->getZExtValue();
193 
194   static std::map<unsigned,unsigned> LoadPciMap = {
195     { Intrinsic::hexagon_circ_ldb,  Hexagon::L2_loadrb_pci  },
196     { Intrinsic::hexagon_circ_ldub, Hexagon::L2_loadrub_pci },
197     { Intrinsic::hexagon_circ_ldh,  Hexagon::L2_loadrh_pci  },
198     { Intrinsic::hexagon_circ_lduh, Hexagon::L2_loadruh_pci },
199     { Intrinsic::hexagon_circ_ldw,  Hexagon::L2_loadri_pci  },
200     { Intrinsic::hexagon_circ_ldd,  Hexagon::L2_loadrd_pci  },
201   };
202   auto FLC = LoadPciMap.find(IntNo);
203   if (FLC != LoadPciMap.end()) {
204     EVT ValTy = (IntNo == Intrinsic::hexagon_circ_ldd) ? MVT::i64 : MVT::i32;
205     EVT RTys[] = { ValTy, MVT::i32, MVT::Other };
206     // Operands: { Base, Increment, Modifier, Chain }
207     auto Inc = cast<ConstantSDNode>(IntN->getOperand(5));
208     SDValue I = CurDAG->getTargetConstant(Inc->getSExtValue(), dl, MVT::i32);
209     MachineSDNode *Res = CurDAG->getMachineNode(FLC->second, dl, RTys,
210           { IntN->getOperand(2), I, IntN->getOperand(4),
211             IntN->getOperand(0) });
212     return Res;
213   }
214 
215   return nullptr;
216 }
217 
StoreInstrForLoadIntrinsic(MachineSDNode * LoadN,SDNode * IntN)218 SDNode *HexagonDAGToDAGISel::StoreInstrForLoadIntrinsic(MachineSDNode *LoadN,
219       SDNode *IntN) {
220   // The "LoadN" is just a machine load instruction. The intrinsic also
221   // involves storing it. Generate an appropriate store to the location
222   // given in the intrinsic's operand(3).
223   uint64_t F = HII->get(LoadN->getMachineOpcode()).TSFlags;
224   unsigned SizeBits = (F >> HexagonII::MemAccessSizePos) &
225                       HexagonII::MemAccesSizeMask;
226   unsigned Size = 1U << (SizeBits-1);
227 
228   SDLoc dl(IntN);
229   MachinePointerInfo PI;
230   SDValue TS;
231   SDValue Loc = IntN->getOperand(3);
232 
233   if (Size >= 4)
234     TS = CurDAG->getStore(SDValue(LoadN, 2), dl, SDValue(LoadN, 0), Loc, PI,
235                           Size);
236   else
237     TS = CurDAG->getTruncStore(SDValue(LoadN, 2), dl, SDValue(LoadN, 0), Loc,
238                                PI, MVT::getIntegerVT(Size * 8), Size);
239 
240   SDNode *StoreN;
241   {
242     HandleSDNode Handle(TS);
243     SelectStore(TS.getNode());
244     StoreN = Handle.getValue().getNode();
245   }
246 
247   // Load's results are { Loaded value, Updated pointer, Chain }
248   ReplaceUses(SDValue(IntN, 0), SDValue(LoadN, 1));
249   ReplaceUses(SDValue(IntN, 1), SDValue(StoreN, 0));
250   return StoreN;
251 }
252 
tryLoadOfLoadIntrinsic(LoadSDNode * N)253 bool HexagonDAGToDAGISel::tryLoadOfLoadIntrinsic(LoadSDNode *N) {
254   // The intrinsics for load circ/brev perform two operations:
255   // 1. Load a value V from the specified location, using the addressing
256   //    mode corresponding to the intrinsic.
257   // 2. Store V into a specified location. This location is typically a
258   //    local, temporary object.
259   // In many cases, the program using these intrinsics will immediately
260   // load V again from the local object. In those cases, when certain
261   // conditions are met, the last load can be removed.
262   // This function identifies and optimizes this pattern. If the pattern
263   // cannot be optimized, it returns nullptr, which will cause the load
264   // to be selected separately from the intrinsic (which will be handled
265   // in SelectIntrinsicWChain).
266 
267   SDValue Ch = N->getOperand(0);
268   SDValue Loc = N->getOperand(1);
269 
270   // Assume that the load and the intrinsic are connected directly with a
271   // chain:
272   //   t1: i32,ch = int.load ..., ..., ..., Loc, ...    // <-- C
273   //   t2: i32,ch = load t1:1, Loc, ...
274   SDNode *C = Ch.getNode();
275 
276   if (C->getOpcode() != ISD::INTRINSIC_W_CHAIN)
277     return false;
278 
279   // The second load can only be eliminated if its extension type matches
280   // that of the load instruction corresponding to the intrinsic. The user
281   // can provide an address of an unsigned variable to store the result of
282   // a sign-extending intrinsic into (or the other way around).
283   ISD::LoadExtType IntExt;
284   switch (cast<ConstantSDNode>(C->getOperand(1))->getZExtValue()) {
285     case Intrinsic::hexagon_circ_ldub:
286     case Intrinsic::hexagon_circ_lduh:
287       IntExt = ISD::ZEXTLOAD;
288       break;
289     case Intrinsic::hexagon_circ_ldw:
290     case Intrinsic::hexagon_circ_ldd:
291       IntExt = ISD::NON_EXTLOAD;
292       break;
293     default:
294       IntExt = ISD::SEXTLOAD;
295       break;
296   }
297   if (N->getExtensionType() != IntExt)
298     return false;
299 
300   // Make sure the target location for the loaded value in the load intrinsic
301   // is the location from which LD (or N) is loading.
302   if (C->getNumOperands() < 4 || Loc.getNode() != C->getOperand(3).getNode())
303     return false;
304 
305   if (MachineSDNode *L = LoadInstrForLoadIntrinsic(C)) {
306     SDNode *S = StoreInstrForLoadIntrinsic(L, C);
307     SDValue F[] = { SDValue(N,0), SDValue(N,1), SDValue(C,0), SDValue(C,1) };
308     SDValue T[] = { SDValue(L,0), SDValue(S,0), SDValue(L,1), SDValue(S,0) };
309     ReplaceUses(F, T, array_lengthof(T));
310     // This transformation will leave the intrinsic dead. If it remains in
311     // the DAG, the selection code will see it again, but without the load,
312     // and it will generate a store that is normally required for it.
313     CurDAG->RemoveDeadNode(C);
314     return true;
315   }
316   return false;
317 }
318 
319 // Convert the bit-reverse load intrinsic to appropriate target instruction.
SelectBrevLdIntrinsic(SDNode * IntN)320 bool HexagonDAGToDAGISel::SelectBrevLdIntrinsic(SDNode *IntN) {
321   if (IntN->getOpcode() != ISD::INTRINSIC_W_CHAIN)
322     return false;
323 
324   const SDLoc &dl(IntN);
325   unsigned IntNo = cast<ConstantSDNode>(IntN->getOperand(1))->getZExtValue();
326 
327   static const std::map<unsigned, unsigned> LoadBrevMap = {
328     { Intrinsic::hexagon_L2_loadrb_pbr, Hexagon::L2_loadrb_pbr },
329     { Intrinsic::hexagon_L2_loadrub_pbr, Hexagon::L2_loadrub_pbr },
330     { Intrinsic::hexagon_L2_loadrh_pbr, Hexagon::L2_loadrh_pbr },
331     { Intrinsic::hexagon_L2_loadruh_pbr, Hexagon::L2_loadruh_pbr },
332     { Intrinsic::hexagon_L2_loadri_pbr, Hexagon::L2_loadri_pbr },
333     { Intrinsic::hexagon_L2_loadrd_pbr, Hexagon::L2_loadrd_pbr }
334   };
335   auto FLI = LoadBrevMap.find(IntNo);
336   if (FLI != LoadBrevMap.end()) {
337     EVT ValTy =
338         (IntNo == Intrinsic::hexagon_L2_loadrd_pbr) ? MVT::i64 : MVT::i32;
339     EVT RTys[] = { ValTy, MVT::i32, MVT::Other };
340     // Operands of Intrinsic: {chain, enum ID of intrinsic, baseptr,
341     // modifier}.
342     // Operands of target instruction: { Base, Modifier, Chain }.
343     MachineSDNode *Res = CurDAG->getMachineNode(
344         FLI->second, dl, RTys,
345         {IntN->getOperand(2), IntN->getOperand(3), IntN->getOperand(0)});
346 
347     MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
348     MemOp[0] = cast<MemIntrinsicSDNode>(IntN)->getMemOperand();
349     Res->setMemRefs(MemOp, MemOp + 1);
350 
351     ReplaceUses(SDValue(IntN, 0), SDValue(Res, 0));
352     ReplaceUses(SDValue(IntN, 1), SDValue(Res, 1));
353     ReplaceUses(SDValue(IntN, 2), SDValue(Res, 2));
354     CurDAG->RemoveDeadNode(IntN);
355     return true;
356   }
357   return false;
358 }
359 
360 /// Generate a machine instruction node for the new circlar buffer intrinsics.
361 /// The new versions use a CSx register instead of the K field.
SelectNewCircIntrinsic(SDNode * IntN)362 bool HexagonDAGToDAGISel::SelectNewCircIntrinsic(SDNode *IntN) {
363   if (IntN->getOpcode() != ISD::INTRINSIC_W_CHAIN)
364     return false;
365 
366   SDLoc DL(IntN);
367   unsigned IntNo = cast<ConstantSDNode>(IntN->getOperand(1))->getZExtValue();
368   SmallVector<SDValue, 7> Ops;
369 
370   static std::map<unsigned,unsigned> LoadNPcMap = {
371     { Intrinsic::hexagon_L2_loadrub_pci, Hexagon::PS_loadrub_pci },
372     { Intrinsic::hexagon_L2_loadrb_pci, Hexagon::PS_loadrb_pci },
373     { Intrinsic::hexagon_L2_loadruh_pci, Hexagon::PS_loadruh_pci },
374     { Intrinsic::hexagon_L2_loadrh_pci, Hexagon::PS_loadrh_pci },
375     { Intrinsic::hexagon_L2_loadri_pci, Hexagon::PS_loadri_pci },
376     { Intrinsic::hexagon_L2_loadrd_pci, Hexagon::PS_loadrd_pci },
377     { Intrinsic::hexagon_L2_loadrub_pcr, Hexagon::PS_loadrub_pcr },
378     { Intrinsic::hexagon_L2_loadrb_pcr, Hexagon::PS_loadrb_pcr },
379     { Intrinsic::hexagon_L2_loadruh_pcr, Hexagon::PS_loadruh_pcr },
380     { Intrinsic::hexagon_L2_loadrh_pcr, Hexagon::PS_loadrh_pcr },
381     { Intrinsic::hexagon_L2_loadri_pcr, Hexagon::PS_loadri_pcr },
382     { Intrinsic::hexagon_L2_loadrd_pcr, Hexagon::PS_loadrd_pcr }
383   };
384   auto FLI = LoadNPcMap.find (IntNo);
385   if (FLI != LoadNPcMap.end()) {
386     EVT ValTy = MVT::i32;
387     if (IntNo == Intrinsic::hexagon_L2_loadrd_pci ||
388         IntNo == Intrinsic::hexagon_L2_loadrd_pcr)
389       ValTy = MVT::i64;
390     EVT RTys[] = { ValTy, MVT::i32, MVT::Other };
391     // Handle load.*_pci case which has 6 operands.
392     if (IntN->getNumOperands() == 6) {
393       auto Inc = cast<ConstantSDNode>(IntN->getOperand(3));
394       SDValue I = CurDAG->getTargetConstant(Inc->getSExtValue(), DL, MVT::i32);
395       // Operands: { Base, Increment, Modifier, Start, Chain }.
396       Ops = { IntN->getOperand(2), I, IntN->getOperand(4), IntN->getOperand(5),
397               IntN->getOperand(0) };
398     } else
399       // Handle load.*_pcr case which has 5 operands.
400       // Operands: { Base, Modifier, Start, Chain }.
401       Ops = { IntN->getOperand(2), IntN->getOperand(3), IntN->getOperand(4),
402               IntN->getOperand(0) };
403     MachineSDNode *Res = CurDAG->getMachineNode(FLI->second, DL, RTys, Ops);
404     ReplaceUses(SDValue(IntN, 0), SDValue(Res, 0));
405     ReplaceUses(SDValue(IntN, 1), SDValue(Res, 1));
406     ReplaceUses(SDValue(IntN, 2), SDValue(Res, 2));
407     CurDAG->RemoveDeadNode(IntN);
408     return true;
409   }
410 
411   static std::map<unsigned,unsigned> StoreNPcMap = {
412     { Intrinsic::hexagon_S2_storerb_pci, Hexagon::PS_storerb_pci },
413     { Intrinsic::hexagon_S2_storerh_pci, Hexagon::PS_storerh_pci },
414     { Intrinsic::hexagon_S2_storerf_pci, Hexagon::PS_storerf_pci },
415     { Intrinsic::hexagon_S2_storeri_pci, Hexagon::PS_storeri_pci },
416     { Intrinsic::hexagon_S2_storerd_pci, Hexagon::PS_storerd_pci },
417     { Intrinsic::hexagon_S2_storerb_pcr, Hexagon::PS_storerb_pcr },
418     { Intrinsic::hexagon_S2_storerh_pcr, Hexagon::PS_storerh_pcr },
419     { Intrinsic::hexagon_S2_storerf_pcr, Hexagon::PS_storerf_pcr },
420     { Intrinsic::hexagon_S2_storeri_pcr, Hexagon::PS_storeri_pcr },
421     { Intrinsic::hexagon_S2_storerd_pcr, Hexagon::PS_storerd_pcr }
422   };
423   auto FSI = StoreNPcMap.find (IntNo);
424   if (FSI != StoreNPcMap.end()) {
425     EVT RTys[] = { MVT::i32, MVT::Other };
426     // Handle store.*_pci case which has 7 operands.
427     if (IntN->getNumOperands() == 7) {
428       auto Inc = cast<ConstantSDNode>(IntN->getOperand(3));
429       SDValue I = CurDAG->getTargetConstant(Inc->getSExtValue(), DL, MVT::i32);
430       // Operands: { Base, Increment, Modifier, Value, Start, Chain }.
431       Ops = { IntN->getOperand(2), I, IntN->getOperand(4), IntN->getOperand(5),
432               IntN->getOperand(6), IntN->getOperand(0) };
433     } else
434       // Handle store.*_pcr case which has 6 operands.
435       // Operands: { Base, Modifier, Value, Start, Chain }.
436       Ops = { IntN->getOperand(2), IntN->getOperand(3), IntN->getOperand(4),
437               IntN->getOperand(5), IntN->getOperand(0) };
438     MachineSDNode *Res = CurDAG->getMachineNode(FSI->second, DL, RTys, Ops);
439     ReplaceUses(SDValue(IntN, 0), SDValue(Res, 0));
440     ReplaceUses(SDValue(IntN, 1), SDValue(Res, 1));
441     CurDAG->RemoveDeadNode(IntN);
442     return true;
443   }
444 
445   return false;
446 }
447 
SelectLoad(SDNode * N)448 void HexagonDAGToDAGISel::SelectLoad(SDNode *N) {
449   SDLoc dl(N);
450   LoadSDNode *LD = cast<LoadSDNode>(N);
451 
452   // Handle indexed loads.
453   ISD::MemIndexedMode AM = LD->getAddressingMode();
454   if (AM != ISD::UNINDEXED) {
455     SelectIndexedLoad(LD, dl);
456     return;
457   }
458 
459   // Handle patterns using circ/brev load intrinsics.
460   if (tryLoadOfLoadIntrinsic(LD))
461     return;
462 
463   SelectCode(LD);
464 }
465 
SelectIndexedStore(StoreSDNode * ST,const SDLoc & dl)466 void HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, const SDLoc &dl) {
467   SDValue Chain = ST->getChain();
468   SDValue Base = ST->getBasePtr();
469   SDValue Offset = ST->getOffset();
470   SDValue Value = ST->getValue();
471   // Get the constant value.
472   int32_t Inc = cast<ConstantSDNode>(Offset.getNode())->getSExtValue();
473   EVT StoredVT = ST->getMemoryVT();
474   EVT ValueVT = Value.getValueType();
475 
476   bool IsValidInc = HII->isValidAutoIncImm(StoredVT, Inc);
477   unsigned Opcode = 0;
478 
479   assert(StoredVT.isSimple());
480   switch (StoredVT.getSimpleVT().SimpleTy) {
481   case MVT::i8:
482     Opcode = IsValidInc ? Hexagon::S2_storerb_pi : Hexagon::S2_storerb_io;
483     break;
484   case MVT::i16:
485     Opcode = IsValidInc ? Hexagon::S2_storerh_pi : Hexagon::S2_storerh_io;
486     break;
487   case MVT::i32:
488   case MVT::f32:
489   case MVT::v2i16:
490   case MVT::v4i8:
491     Opcode = IsValidInc ? Hexagon::S2_storeri_pi : Hexagon::S2_storeri_io;
492     break;
493   case MVT::i64:
494   case MVT::f64:
495   case MVT::v2i32:
496   case MVT::v4i16:
497   case MVT::v8i8:
498     Opcode = IsValidInc ? Hexagon::S2_storerd_pi : Hexagon::S2_storerd_io;
499     break;
500   case MVT::v64i8:
501   case MVT::v32i16:
502   case MVT::v16i32:
503   case MVT::v8i64:
504   case MVT::v128i8:
505   case MVT::v64i16:
506   case MVT::v32i32:
507   case MVT::v16i64:
508     if (isAlignedMemNode(ST)) {
509       if (ST->isNonTemporal())
510         Opcode = IsValidInc ? Hexagon::V6_vS32b_nt_pi : Hexagon::V6_vS32b_nt_ai;
511       else
512         Opcode = IsValidInc ? Hexagon::V6_vS32b_pi : Hexagon::V6_vS32b_ai;
513     } else {
514       Opcode = IsValidInc ? Hexagon::V6_vS32Ub_pi : Hexagon::V6_vS32Ub_ai;
515     }
516     break;
517   default:
518     llvm_unreachable("Unexpected memory type in indexed store");
519   }
520 
521   if (ST->isTruncatingStore() && ValueVT.getSizeInBits() == 64) {
522     assert(StoredVT.getSizeInBits() < 64 && "Not a truncating store");
523     Value = CurDAG->getTargetExtractSubreg(Hexagon::isub_lo,
524                                            dl, MVT::i32, Value);
525   }
526 
527   SDValue IncV = CurDAG->getTargetConstant(Inc, dl, MVT::i32);
528   MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
529   MemOp[0] = ST->getMemOperand();
530 
531   //                  Next address   Chain
532   SDValue From[2] = { SDValue(ST,0), SDValue(ST,1) };
533   SDValue To[2];
534 
535   if (IsValidInc) {
536     // Build post increment store.
537     SDValue Ops[] = { Base, IncV, Value, Chain };
538     MachineSDNode *S = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::Other,
539                                               Ops);
540     S->setMemRefs(MemOp, MemOp + 1);
541     To[0] = SDValue(S, 0);
542     To[1] = SDValue(S, 1);
543   } else {
544     SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
545     SDValue Ops[] = { Base, Zero, Value, Chain };
546     MachineSDNode *S = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops);
547     S->setMemRefs(MemOp, MemOp + 1);
548     To[1] = SDValue(S, 0);
549     MachineSDNode *A = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
550                                               Base, IncV);
551     To[0] = SDValue(A, 0);
552   }
553 
554   ReplaceUses(From, To, 2);
555   CurDAG->RemoveDeadNode(ST);
556 }
557 
SelectStore(SDNode * N)558 void HexagonDAGToDAGISel::SelectStore(SDNode *N) {
559   SDLoc dl(N);
560   StoreSDNode *ST = cast<StoreSDNode>(N);
561 
562   // Handle indexed stores.
563   ISD::MemIndexedMode AM = ST->getAddressingMode();
564   if (AM != ISD::UNINDEXED) {
565     SelectIndexedStore(ST, dl);
566     return;
567   }
568 
569   SelectCode(ST);
570 }
571 
SelectSHL(SDNode * N)572 void HexagonDAGToDAGISel::SelectSHL(SDNode *N) {
573   SDLoc dl(N);
574   SDValue Shl_0 = N->getOperand(0);
575   SDValue Shl_1 = N->getOperand(1);
576 
577   auto Default = [this,N] () -> void { SelectCode(N); };
578 
579   if (N->getValueType(0) != MVT::i32 || Shl_1.getOpcode() != ISD::Constant)
580     return Default();
581 
582   // RHS is const.
583   int32_t ShlConst = cast<ConstantSDNode>(Shl_1)->getSExtValue();
584 
585   if (Shl_0.getOpcode() == ISD::MUL) {
586     SDValue Mul_0 = Shl_0.getOperand(0); // Val
587     SDValue Mul_1 = Shl_0.getOperand(1); // Const
588     // RHS of mul is const.
589     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mul_1)) {
590       int32_t ValConst = C->getSExtValue() << ShlConst;
591       if (isInt<9>(ValConst)) {
592         SDValue Val = CurDAG->getTargetConstant(ValConst, dl, MVT::i32);
593         SDNode *Result = CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl,
594                                                 MVT::i32, Mul_0, Val);
595         ReplaceNode(N, Result);
596         return;
597       }
598     }
599     return Default();
600   }
601 
602   if (Shl_0.getOpcode() == ISD::SUB) {
603     SDValue Sub_0 = Shl_0.getOperand(0); // Const 0
604     SDValue Sub_1 = Shl_0.getOperand(1); // Val
605     if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(Sub_0)) {
606       if (C1->getSExtValue() != 0 || Sub_1.getOpcode() != ISD::SHL)
607         return Default();
608       SDValue Shl2_0 = Sub_1.getOperand(0); // Val
609       SDValue Shl2_1 = Sub_1.getOperand(1); // Const
610       if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(Shl2_1)) {
611         int32_t ValConst = 1 << (ShlConst + C2->getSExtValue());
612         if (isInt<9>(-ValConst)) {
613           SDValue Val = CurDAG->getTargetConstant(-ValConst, dl, MVT::i32);
614           SDNode *Result = CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl,
615                                                   MVT::i32, Shl2_0, Val);
616           ReplaceNode(N, Result);
617           return;
618         }
619       }
620     }
621   }
622 
623   return Default();
624 }
625 
626 //
627 // Handling intrinsics for circular load and bitreverse load.
628 //
SelectIntrinsicWChain(SDNode * N)629 void HexagonDAGToDAGISel::SelectIntrinsicWChain(SDNode *N) {
630   if (MachineSDNode *L = LoadInstrForLoadIntrinsic(N)) {
631     StoreInstrForLoadIntrinsic(L, N);
632     CurDAG->RemoveDeadNode(N);
633     return;
634   }
635 
636   // Handle bit-reverse load intrinsics.
637   if (SelectBrevLdIntrinsic(N))
638     return;
639 
640   if (SelectNewCircIntrinsic(N))
641     return;
642 
643   unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
644   if (IntNo == Intrinsic::hexagon_V6_vgathermw ||
645       IntNo == Intrinsic::hexagon_V6_vgathermw_128B ||
646       IntNo == Intrinsic::hexagon_V6_vgathermh ||
647       IntNo == Intrinsic::hexagon_V6_vgathermh_128B ||
648       IntNo == Intrinsic::hexagon_V6_vgathermhw ||
649       IntNo == Intrinsic::hexagon_V6_vgathermhw_128B) {
650     SelectV65Gather(N);
651     return;
652   }
653   if (IntNo == Intrinsic::hexagon_V6_vgathermwq ||
654       IntNo == Intrinsic::hexagon_V6_vgathermwq_128B ||
655       IntNo == Intrinsic::hexagon_V6_vgathermhq ||
656       IntNo == Intrinsic::hexagon_V6_vgathermhq_128B ||
657       IntNo == Intrinsic::hexagon_V6_vgathermhwq ||
658       IntNo == Intrinsic::hexagon_V6_vgathermhwq_128B) {
659     SelectV65GatherPred(N);
660     return;
661   }
662 
663   SelectCode(N);
664 }
665 
SelectIntrinsicWOChain(SDNode * N)666 void HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) {
667   unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
668   unsigned Bits;
669   switch (IID) {
670   case Intrinsic::hexagon_S2_vsplatrb:
671     Bits = 8;
672     break;
673   case Intrinsic::hexagon_S2_vsplatrh:
674     Bits = 16;
675     break;
676   case Intrinsic::hexagon_V6_vaddcarry:
677   case Intrinsic::hexagon_V6_vaddcarry_128B:
678   case Intrinsic::hexagon_V6_vsubcarry:
679   case Intrinsic::hexagon_V6_vsubcarry_128B:
680     SelectHVXDualOutput(N);
681     return;
682   default:
683     SelectCode(N);
684     return;
685   }
686 
687   SDValue V = N->getOperand(1);
688   SDValue U;
689   if (keepsLowBits(V, Bits, U)) {
690     SDValue R = CurDAG->getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
691                                 N->getOperand(0), U);
692     ReplaceNode(N, R.getNode());
693     SelectCode(R.getNode());
694     return;
695   }
696   SelectCode(N);
697 }
698 
699 //
700 // Map floating point constant values.
701 //
SelectConstantFP(SDNode * N)702 void HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) {
703   SDLoc dl(N);
704   ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N);
705   APInt A = CN->getValueAPF().bitcastToAPInt();
706   if (N->getValueType(0) == MVT::f32) {
707     SDValue V = CurDAG->getTargetConstant(A.getZExtValue(), dl, MVT::i32);
708     ReplaceNode(N, CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl, MVT::f32, V));
709     return;
710   }
711   if (N->getValueType(0) == MVT::f64) {
712     SDValue V = CurDAG->getTargetConstant(A.getZExtValue(), dl, MVT::i64);
713     ReplaceNode(N, CurDAG->getMachineNode(Hexagon::CONST64, dl, MVT::f64, V));
714     return;
715   }
716 
717   SelectCode(N);
718 }
719 
720 //
721 // Map boolean values.
722 //
SelectConstant(SDNode * N)723 void HexagonDAGToDAGISel::SelectConstant(SDNode *N) {
724   if (N->getValueType(0) == MVT::i1) {
725     assert(!(cast<ConstantSDNode>(N)->getZExtValue() >> 1));
726     unsigned Opc = (cast<ConstantSDNode>(N)->getSExtValue() != 0)
727                       ? Hexagon::PS_true
728                       : Hexagon::PS_false;
729     ReplaceNode(N, CurDAG->getMachineNode(Opc, SDLoc(N), MVT::i1));
730     return;
731   }
732 
733   SelectCode(N);
734 }
735 
SelectFrameIndex(SDNode * N)736 void HexagonDAGToDAGISel::SelectFrameIndex(SDNode *N) {
737   MachineFrameInfo &MFI = MF->getFrameInfo();
738   const HexagonFrameLowering *HFI = HST->getFrameLowering();
739   int FX = cast<FrameIndexSDNode>(N)->getIndex();
740   unsigned StkA = HFI->getStackAlignment();
741   unsigned MaxA = MFI.getMaxAlignment();
742   SDValue FI = CurDAG->getTargetFrameIndex(FX, MVT::i32);
743   SDLoc DL(N);
744   SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
745   SDNode *R = nullptr;
746 
747   // Use PS_fi when:
748   // - the object is fixed, or
749   // - there are no objects with higher-than-default alignment, or
750   // - there are no dynamically allocated objects.
751   // Otherwise, use PS_fia.
752   if (FX < 0 || MaxA <= StkA || !MFI.hasVarSizedObjects()) {
753     R = CurDAG->getMachineNode(Hexagon::PS_fi, DL, MVT::i32, FI, Zero);
754   } else {
755     auto &HMFI = *MF->getInfo<HexagonMachineFunctionInfo>();
756     unsigned AR = HMFI.getStackAlignBaseVReg();
757     SDValue CH = CurDAG->getEntryNode();
758     SDValue Ops[] = { CurDAG->getCopyFromReg(CH, DL, AR, MVT::i32), FI, Zero };
759     R = CurDAG->getMachineNode(Hexagon::PS_fia, DL, MVT::i32, Ops);
760   }
761 
762   ReplaceNode(N, R);
763 }
764 
SelectAddSubCarry(SDNode * N)765 void HexagonDAGToDAGISel::SelectAddSubCarry(SDNode *N) {
766   unsigned OpcCarry = N->getOpcode() == HexagonISD::ADDC ? Hexagon::A4_addp_c
767                                                          : Hexagon::A4_subp_c;
768   SDNode *C = CurDAG->getMachineNode(OpcCarry, SDLoc(N), N->getVTList(),
769                                      { N->getOperand(0), N->getOperand(1),
770                                        N->getOperand(2) });
771   ReplaceNode(N, C);
772 }
773 
SelectVAlign(SDNode * N)774 void HexagonDAGToDAGISel::SelectVAlign(SDNode *N) {
775   MVT ResTy = N->getValueType(0).getSimpleVT();
776   if (HST->isHVXVectorType(ResTy, true))
777     return SelectHvxVAlign(N);
778 
779   const SDLoc &dl(N);
780   unsigned VecLen = ResTy.getSizeInBits();
781   if (VecLen == 32) {
782     SDValue Ops[] = {
783       CurDAG->getTargetConstant(Hexagon::DoubleRegsRegClassID, dl, MVT::i32),
784       N->getOperand(0),
785       CurDAG->getTargetConstant(Hexagon::isub_hi, dl, MVT::i32),
786       N->getOperand(1),
787       CurDAG->getTargetConstant(Hexagon::isub_lo, dl, MVT::i32)
788     };
789     SDNode *R = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl,
790                                        MVT::i64, Ops);
791 
792     // Shift right by "(Addr & 0x3) * 8" bytes.
793     SDValue M0 = CurDAG->getTargetConstant(0x18, dl, MVT::i32);
794     SDValue M1 = CurDAG->getTargetConstant(0x03, dl, MVT::i32);
795     SDNode *C = CurDAG->getMachineNode(Hexagon::S4_andi_asl_ri, dl, MVT::i32,
796                                        M0, N->getOperand(2), M1);
797     SDNode *S = CurDAG->getMachineNode(Hexagon::S2_lsr_r_p, dl, MVT::i64,
798                                        SDValue(R, 0), SDValue(C, 0));
799     SDValue E = CurDAG->getTargetExtractSubreg(Hexagon::isub_lo, dl, ResTy,
800                                                SDValue(S, 0));
801     ReplaceNode(N, E.getNode());
802   } else {
803     assert(VecLen == 64);
804     SDNode *Pu = CurDAG->getMachineNode(Hexagon::C2_tfrrp, dl, MVT::v8i1,
805                                         N->getOperand(2));
806     SDNode *VA = CurDAG->getMachineNode(Hexagon::S2_valignrb, dl, ResTy,
807                                         N->getOperand(0), N->getOperand(1),
808                                         SDValue(Pu,0));
809     ReplaceNode(N, VA);
810   }
811 }
812 
SelectVAlignAddr(SDNode * N)813 void HexagonDAGToDAGISel::SelectVAlignAddr(SDNode *N) {
814   const SDLoc &dl(N);
815   SDValue A = N->getOperand(1);
816   int Mask = -cast<ConstantSDNode>(A.getNode())->getSExtValue();
817   assert(isPowerOf2_32(-Mask));
818 
819   SDValue M = CurDAG->getTargetConstant(Mask, dl, MVT::i32);
820   SDNode *AA = CurDAG->getMachineNode(Hexagon::A2_andir, dl, MVT::i32,
821                                       N->getOperand(0), M);
822   ReplaceNode(N, AA);
823 }
824 
825 // Handle these nodes here to avoid having to write patterns for all
826 // combinations of input/output types. In all cases, the resulting
827 // instruction is the same.
SelectTypecast(SDNode * N)828 void HexagonDAGToDAGISel::SelectTypecast(SDNode *N) {
829   SDValue Op = N->getOperand(0);
830   MVT OpTy = Op.getValueType().getSimpleVT();
831   SDNode *T = CurDAG->MorphNodeTo(N, N->getOpcode(),
832                                   CurDAG->getVTList(OpTy), {Op});
833   ReplaceNode(T, Op.getNode());
834 }
835 
SelectP2D(SDNode * N)836 void HexagonDAGToDAGISel::SelectP2D(SDNode *N) {
837   MVT ResTy = N->getValueType(0).getSimpleVT();
838   SDNode *T = CurDAG->getMachineNode(Hexagon::C2_mask, SDLoc(N), ResTy,
839                                      N->getOperand(0));
840   ReplaceNode(N, T);
841 }
842 
SelectD2P(SDNode * N)843 void HexagonDAGToDAGISel::SelectD2P(SDNode *N) {
844   const SDLoc &dl(N);
845   MVT ResTy = N->getValueType(0).getSimpleVT();
846   SDValue Zero = CurDAG->getTargetConstant(0, dl, MVT::i32);
847   SDNode *T = CurDAG->getMachineNode(Hexagon::A4_vcmpbgtui, dl, ResTy,
848                                      N->getOperand(0), Zero);
849   ReplaceNode(N, T);
850 }
851 
SelectV2Q(SDNode * N)852 void HexagonDAGToDAGISel::SelectV2Q(SDNode *N) {
853   const SDLoc &dl(N);
854   MVT ResTy = N->getValueType(0).getSimpleVT();
855 
856   SDValue C = CurDAG->getTargetConstant(-1, dl, MVT::i32);
857   SDNode *R = CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl, MVT::i32, C);
858   SDNode *T = CurDAG->getMachineNode(Hexagon::V6_vandvrt, dl, ResTy,
859                                      N->getOperand(0), SDValue(R,0));
860   ReplaceNode(N, T);
861 }
862 
SelectQ2V(SDNode * N)863 void HexagonDAGToDAGISel::SelectQ2V(SDNode *N) {
864   const SDLoc &dl(N);
865   MVT ResTy = N->getValueType(0).getSimpleVT();
866 
867   SDValue C = CurDAG->getTargetConstant(-1, dl, MVT::i32);
868   SDNode *R = CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl, MVT::i32, C);
869   SDNode *T = CurDAG->getMachineNode(Hexagon::V6_vandqrt, dl, ResTy,
870                                      N->getOperand(0), SDValue(R,0));
871   ReplaceNode(N, T);
872 }
873 
Select(SDNode * N)874 void HexagonDAGToDAGISel::Select(SDNode *N) {
875   if (N->isMachineOpcode())
876     return N->setNodeId(-1);  // Already selected.
877 
878   switch (N->getOpcode()) {
879   case ISD::Constant:             return SelectConstant(N);
880   case ISD::ConstantFP:           return SelectConstantFP(N);
881   case ISD::FrameIndex:           return SelectFrameIndex(N);
882   case ISD::SHL:                  return SelectSHL(N);
883   case ISD::LOAD:                 return SelectLoad(N);
884   case ISD::STORE:                return SelectStore(N);
885   case ISD::INTRINSIC_W_CHAIN:    return SelectIntrinsicWChain(N);
886   case ISD::INTRINSIC_WO_CHAIN:   return SelectIntrinsicWOChain(N);
887 
888   case HexagonISD::ADDC:
889   case HexagonISD::SUBC:          return SelectAddSubCarry(N);
890   case HexagonISD::VALIGN:        return SelectVAlign(N);
891   case HexagonISD::VALIGNADDR:    return SelectVAlignAddr(N);
892   case HexagonISD::TYPECAST:      return SelectTypecast(N);
893   case HexagonISD::P2D:           return SelectP2D(N);
894   case HexagonISD::D2P:           return SelectD2P(N);
895   case HexagonISD::Q2V:           return SelectQ2V(N);
896   case HexagonISD::V2Q:           return SelectV2Q(N);
897   }
898 
899   if (HST->useHVXOps()) {
900     switch (N->getOpcode()) {
901     case ISD::VECTOR_SHUFFLE:     return SelectHvxShuffle(N);
902     case HexagonISD::VROR:        return SelectHvxRor(N);
903     }
904   }
905 
906   SelectCode(N);
907 }
908 
909 bool HexagonDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue & Op,unsigned ConstraintID,std::vector<SDValue> & OutOps)910 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
911                              std::vector<SDValue> &OutOps) {
912   SDValue Inp = Op, Res;
913 
914   switch (ConstraintID) {
915   default:
916     return true;
917   case InlineAsm::Constraint_i:
918   case InlineAsm::Constraint_o: // Offsetable.
919   case InlineAsm::Constraint_v: // Not offsetable.
920   case InlineAsm::Constraint_m: // Memory.
921     if (SelectAddrFI(Inp, Res))
922       OutOps.push_back(Res);
923     else
924       OutOps.push_back(Inp);
925     break;
926   }
927 
928   OutOps.push_back(CurDAG->getTargetConstant(0, SDLoc(Op), MVT::i32));
929   return false;
930 }
931 
932 
isMemOPCandidate(SDNode * I,SDNode * U)933 static bool isMemOPCandidate(SDNode *I, SDNode *U) {
934   // I is an operand of U. Check if U is an arithmetic (binary) operation
935   // usable in a memop, where the other operand is a loaded value, and the
936   // result of U is stored in the same location.
937 
938   if (!U->hasOneUse())
939     return false;
940   unsigned Opc = U->getOpcode();
941   switch (Opc) {
942     case ISD::ADD:
943     case ISD::SUB:
944     case ISD::AND:
945     case ISD::OR:
946       break;
947     default:
948       return false;
949   }
950 
951   SDValue S0 = U->getOperand(0);
952   SDValue S1 = U->getOperand(1);
953   SDValue SY = (S0.getNode() == I) ? S1 : S0;
954 
955   SDNode *UUse = *U->use_begin();
956   if (UUse->getNumValues() != 1)
957     return false;
958 
959   // Check if one of the inputs to U is a load instruction and the output
960   // is used by a store instruction. If so and they also have the same
961   // base pointer, then don't preoprocess this node sequence as it
962   // can be matched to a memop.
963   SDNode *SYNode = SY.getNode();
964   if (UUse->getOpcode() == ISD::STORE && SYNode->getOpcode() == ISD::LOAD) {
965     SDValue LDBasePtr = cast<MemSDNode>(SYNode)->getBasePtr();
966     SDValue STBasePtr = cast<MemSDNode>(UUse)->getBasePtr();
967     if (LDBasePtr == STBasePtr)
968       return true;
969   }
970   return false;
971 }
972 
973 
974 // Transform: (or (select c x 0) z)  ->  (select c (or x z) z)
975 //            (or (select c 0 y) z)  ->  (select c z (or y z))
ppSimplifyOrSelect0(std::vector<SDNode * > && Nodes)976 void HexagonDAGToDAGISel::ppSimplifyOrSelect0(std::vector<SDNode*> &&Nodes) {
977   SelectionDAG &DAG = *CurDAG;
978 
979   for (auto I : Nodes) {
980     if (I->getOpcode() != ISD::OR)
981       continue;
982 
983     auto IsZero = [] (const SDValue &V) -> bool {
984       if (ConstantSDNode *SC = dyn_cast<ConstantSDNode>(V.getNode()))
985         return SC->isNullValue();
986       return false;
987     };
988     auto IsSelect0 = [IsZero] (const SDValue &Op) -> bool {
989       if (Op.getOpcode() != ISD::SELECT)
990         return false;
991       return IsZero(Op.getOperand(1)) || IsZero(Op.getOperand(2));
992     };
993 
994     SDValue N0 = I->getOperand(0), N1 = I->getOperand(1);
995     EVT VT = I->getValueType(0);
996     bool SelN0 = IsSelect0(N0);
997     SDValue SOp = SelN0 ? N0 : N1;
998     SDValue VOp = SelN0 ? N1 : N0;
999 
1000     if (SOp.getOpcode() == ISD::SELECT && SOp.getNode()->hasOneUse()) {
1001       SDValue SC = SOp.getOperand(0);
1002       SDValue SX = SOp.getOperand(1);
1003       SDValue SY = SOp.getOperand(2);
1004       SDLoc DLS = SOp;
1005       if (IsZero(SY)) {
1006         SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SX, VOp);
1007         SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, NewOr, VOp);
1008         DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1009       } else if (IsZero(SX)) {
1010         SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SY, VOp);
1011         SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, VOp, NewOr);
1012         DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1013       }
1014     }
1015   }
1016 }
1017 
1018 // Transform: (store ch val (add x (add (shl y c) e)))
1019 //        to: (store ch val (add x (shl (add y d) c))),
1020 // where e = (shl d c) for some integer d.
1021 // The purpose of this is to enable generation of loads/stores with
1022 // shifted addressing mode, i.e. mem(x+y<<#c). For that, the shift
1023 // value c must be 0, 1 or 2.
ppAddrReorderAddShl(std::vector<SDNode * > && Nodes)1024 void HexagonDAGToDAGISel::ppAddrReorderAddShl(std::vector<SDNode*> &&Nodes) {
1025   SelectionDAG &DAG = *CurDAG;
1026 
1027   for (auto I : Nodes) {
1028     if (I->getOpcode() != ISD::STORE)
1029       continue;
1030 
1031     // I matched: (store ch val Off)
1032     SDValue Off = I->getOperand(2);
1033     // Off needs to match: (add x (add (shl y c) (shl d c))))
1034     if (Off.getOpcode() != ISD::ADD)
1035       continue;
1036     // Off matched: (add x T0)
1037     SDValue T0 = Off.getOperand(1);
1038     // T0 needs to match: (add T1 T2):
1039     if (T0.getOpcode() != ISD::ADD)
1040       continue;
1041     // T0 matched: (add T1 T2)
1042     SDValue T1 = T0.getOperand(0);
1043     SDValue T2 = T0.getOperand(1);
1044     // T1 needs to match: (shl y c)
1045     if (T1.getOpcode() != ISD::SHL)
1046       continue;
1047     SDValue C = T1.getOperand(1);
1048     ConstantSDNode *CN = dyn_cast<ConstantSDNode>(C.getNode());
1049     if (CN == nullptr)
1050       continue;
1051     unsigned CV = CN->getZExtValue();
1052     if (CV > 2)
1053       continue;
1054     // T2 needs to match e, where e = (shl d c) for some d.
1055     ConstantSDNode *EN = dyn_cast<ConstantSDNode>(T2.getNode());
1056     if (EN == nullptr)
1057       continue;
1058     unsigned EV = EN->getZExtValue();
1059     if (EV % (1 << CV) != 0)
1060       continue;
1061     unsigned DV = EV / (1 << CV);
1062 
1063     // Replace T0 with: (shl (add y d) c)
1064     SDLoc DL = SDLoc(I);
1065     EVT VT = T0.getValueType();
1066     SDValue D = DAG.getConstant(DV, DL, VT);
1067     // NewAdd = (add y d)
1068     SDValue NewAdd = DAG.getNode(ISD::ADD, DL, VT, T1.getOperand(0), D);
1069     // NewShl = (shl NewAdd c)
1070     SDValue NewShl = DAG.getNode(ISD::SHL, DL, VT, NewAdd, C);
1071     ReplaceNode(T0.getNode(), NewShl.getNode());
1072   }
1073 }
1074 
1075 // Transform: (load ch (add x (and (srl y c) Mask)))
1076 //        to: (load ch (add x (shl (srl y d) d-c)))
1077 // where
1078 // Mask = 00..0 111..1 0.0
1079 //          |     |     +-- d-c 0s, and d-c is 0, 1 or 2.
1080 //          |     +-------- 1s
1081 //          +-------------- at most c 0s
1082 // Motivating example:
1083 // DAG combiner optimizes (add x (shl (srl y 5) 2))
1084 //                     to (add x (and (srl y 3) 1FFFFFFC))
1085 // which results in a constant-extended and(##...,lsr). This transformation
1086 // undoes this simplification for cases where the shl can be folded into
1087 // an addressing mode.
ppAddrRewriteAndSrl(std::vector<SDNode * > && Nodes)1088 void HexagonDAGToDAGISel::ppAddrRewriteAndSrl(std::vector<SDNode*> &&Nodes) {
1089   SelectionDAG &DAG = *CurDAG;
1090 
1091   for (SDNode *N : Nodes) {
1092     unsigned Opc = N->getOpcode();
1093     if (Opc != ISD::LOAD && Opc != ISD::STORE)
1094       continue;
1095     SDValue Addr = Opc == ISD::LOAD ? N->getOperand(1) : N->getOperand(2);
1096     // Addr must match: (add x T0)
1097     if (Addr.getOpcode() != ISD::ADD)
1098       continue;
1099     SDValue T0 = Addr.getOperand(1);
1100     // T0 must match: (and T1 Mask)
1101     if (T0.getOpcode() != ISD::AND)
1102       continue;
1103 
1104     // We have an AND.
1105     //
1106     // Check the first operand. It must be: (srl y c).
1107     SDValue S = T0.getOperand(0);
1108     if (S.getOpcode() != ISD::SRL)
1109       continue;
1110     ConstantSDNode *SN = dyn_cast<ConstantSDNode>(S.getOperand(1).getNode());
1111     if (SN == nullptr)
1112       continue;
1113     if (SN->getAPIntValue().getBitWidth() != 32)
1114       continue;
1115     uint32_t CV = SN->getZExtValue();
1116 
1117     // Check the second operand: the supposed mask.
1118     ConstantSDNode *MN = dyn_cast<ConstantSDNode>(T0.getOperand(1).getNode());
1119     if (MN == nullptr)
1120       continue;
1121     if (MN->getAPIntValue().getBitWidth() != 32)
1122       continue;
1123     uint32_t Mask = MN->getZExtValue();
1124     // Examine the mask.
1125     uint32_t TZ = countTrailingZeros(Mask);
1126     uint32_t M1 = countTrailingOnes(Mask >> TZ);
1127     uint32_t LZ = countLeadingZeros(Mask);
1128     // Trailing zeros + middle ones + leading zeros must equal the width.
1129     if (TZ + M1 + LZ != 32)
1130       continue;
1131     // The number of trailing zeros will be encoded in the addressing mode.
1132     if (TZ > 2)
1133       continue;
1134     // The number of leading zeros must be at most c.
1135     if (LZ > CV)
1136       continue;
1137 
1138     // All looks good.
1139     SDValue Y = S.getOperand(0);
1140     EVT VT = Addr.getValueType();
1141     SDLoc dl(S);
1142     // TZ = D-C, so D = TZ+C.
1143     SDValue D = DAG.getConstant(TZ+CV, dl, VT);
1144     SDValue DC = DAG.getConstant(TZ, dl, VT);
1145     SDValue NewSrl = DAG.getNode(ISD::SRL, dl, VT, Y, D);
1146     SDValue NewShl = DAG.getNode(ISD::SHL, dl, VT, NewSrl, DC);
1147     ReplaceNode(T0.getNode(), NewShl.getNode());
1148   }
1149 }
1150 
1151 // Transform: (op ... (zext i1 c) ...) -> (select c (op ... 0 ...)
1152 //                                                  (op ... 1 ...))
ppHoistZextI1(std::vector<SDNode * > && Nodes)1153 void HexagonDAGToDAGISel::ppHoistZextI1(std::vector<SDNode*> &&Nodes) {
1154   SelectionDAG &DAG = *CurDAG;
1155 
1156   for (SDNode *N : Nodes) {
1157     unsigned Opc = N->getOpcode();
1158     if (Opc != ISD::ZERO_EXTEND)
1159       continue;
1160     SDValue OpI1 = N->getOperand(0);
1161     EVT OpVT = OpI1.getValueType();
1162     if (!OpVT.isSimple() || OpVT.getSimpleVT() != MVT::i1)
1163       continue;
1164     for (auto I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1165       SDNode *U = *I;
1166       if (U->getNumValues() != 1)
1167         continue;
1168       EVT UVT = U->getValueType(0);
1169       if (!UVT.isSimple() || !UVT.isInteger() || UVT.getSimpleVT() == MVT::i1)
1170         continue;
1171       if (isMemOPCandidate(N, U))
1172         continue;
1173 
1174       // Potentially simplifiable operation.
1175       unsigned I1N = I.getOperandNo();
1176       SmallVector<SDValue,2> Ops(U->getNumOperands());
1177       for (unsigned i = 0, n = U->getNumOperands(); i != n; ++i)
1178         Ops[i] = U->getOperand(i);
1179       EVT BVT = Ops[I1N].getValueType();
1180 
1181       SDLoc dl(U);
1182       SDValue C0 = DAG.getConstant(0, dl, BVT);
1183       SDValue C1 = DAG.getConstant(1, dl, BVT);
1184       SDValue If0, If1;
1185 
1186       if (isa<MachineSDNode>(U)) {
1187         unsigned UseOpc = U->getMachineOpcode();
1188         Ops[I1N] = C0;
1189         If0 = SDValue(DAG.getMachineNode(UseOpc, dl, UVT, Ops), 0);
1190         Ops[I1N] = C1;
1191         If1 = SDValue(DAG.getMachineNode(UseOpc, dl, UVT, Ops), 0);
1192       } else {
1193         unsigned UseOpc = U->getOpcode();
1194         Ops[I1N] = C0;
1195         If0 = DAG.getNode(UseOpc, dl, UVT, Ops);
1196         Ops[I1N] = C1;
1197         If1 = DAG.getNode(UseOpc, dl, UVT, Ops);
1198       }
1199       SDValue Sel = DAG.getNode(ISD::SELECT, dl, UVT, OpI1, If1, If0);
1200       DAG.ReplaceAllUsesWith(U, Sel.getNode());
1201     }
1202   }
1203 }
1204 
PreprocessISelDAG()1205 void HexagonDAGToDAGISel::PreprocessISelDAG() {
1206   // Repack all nodes before calling each preprocessing function,
1207   // because each of them can modify the set of nodes.
1208   auto getNodes = [this] () -> std::vector<SDNode*> {
1209     std::vector<SDNode*> T;
1210     T.reserve(CurDAG->allnodes_size());
1211     for (SDNode &N : CurDAG->allnodes())
1212       T.push_back(&N);
1213     return T;
1214   };
1215 
1216   // Transform: (or (select c x 0) z)  ->  (select c (or x z) z)
1217   //            (or (select c 0 y) z)  ->  (select c z (or y z))
1218   ppSimplifyOrSelect0(getNodes());
1219 
1220   // Transform: (store ch val (add x (add (shl y c) e)))
1221   //        to: (store ch val (add x (shl (add y d) c))),
1222   // where e = (shl d c) for some integer d.
1223   // The purpose of this is to enable generation of loads/stores with
1224   // shifted addressing mode, i.e. mem(x+y<<#c). For that, the shift
1225   // value c must be 0, 1 or 2.
1226   ppAddrReorderAddShl(getNodes());
1227 
1228   // Transform: (load ch (add x (and (srl y c) Mask)))
1229   //        to: (load ch (add x (shl (srl y d) d-c)))
1230   // where
1231   // Mask = 00..0 111..1 0.0
1232   //          |     |     +-- d-c 0s, and d-c is 0, 1 or 2.
1233   //          |     +-------- 1s
1234   //          +-------------- at most c 0s
1235   // Motivating example:
1236   // DAG combiner optimizes (add x (shl (srl y 5) 2))
1237   //                     to (add x (and (srl y 3) 1FFFFFFC))
1238   // which results in a constant-extended and(##...,lsr). This transformation
1239   // undoes this simplification for cases where the shl can be folded into
1240   // an addressing mode.
1241   ppAddrRewriteAndSrl(getNodes());
1242 
1243   // Transform: (op ... (zext i1 c) ...) -> (select c (op ... 0 ...)
1244   //                                                  (op ... 1 ...))
1245   ppHoistZextI1(getNodes());
1246 
1247   DEBUG_WITH_TYPE("isel", {
1248     dbgs() << "Preprocessed (Hexagon) selection DAG:";
1249     CurDAG->dump();
1250   });
1251 
1252   if (EnableAddressRebalancing) {
1253     rebalanceAddressTrees();
1254 
1255     DEBUG_WITH_TYPE("isel", {
1256       dbgs() << "Address tree balanced selection DAG:";
1257       CurDAG->dump();
1258     });
1259   }
1260 }
1261 
EmitFunctionEntryCode()1262 void HexagonDAGToDAGISel::EmitFunctionEntryCode() {
1263   auto &HST = static_cast<const HexagonSubtarget&>(MF->getSubtarget());
1264   auto &HFI = *HST.getFrameLowering();
1265   if (!HFI.needsAligna(*MF))
1266     return;
1267 
1268   MachineFrameInfo &MFI = MF->getFrameInfo();
1269   MachineBasicBlock *EntryBB = &MF->front();
1270   unsigned AR = FuncInfo->CreateReg(MVT::i32);
1271   unsigned MaxA = MFI.getMaxAlignment();
1272   BuildMI(EntryBB, DebugLoc(), HII->get(Hexagon::PS_aligna), AR)
1273       .addImm(MaxA);
1274   MF->getInfo<HexagonMachineFunctionInfo>()->setStackAlignBaseVReg(AR);
1275 }
1276 
1277 // Match a frame index that can be used in an addressing mode.
SelectAddrFI(SDValue & N,SDValue & R)1278 bool HexagonDAGToDAGISel::SelectAddrFI(SDValue &N, SDValue &R) {
1279   if (N.getOpcode() != ISD::FrameIndex)
1280     return false;
1281   auto &HFI = *HST->getFrameLowering();
1282   MachineFrameInfo &MFI = MF->getFrameInfo();
1283   int FX = cast<FrameIndexSDNode>(N)->getIndex();
1284   if (!MFI.isFixedObjectIndex(FX) && HFI.needsAligna(*MF))
1285     return false;
1286   R = CurDAG->getTargetFrameIndex(FX, MVT::i32);
1287   return true;
1288 }
1289 
SelectAddrGA(SDValue & N,SDValue & R)1290 inline bool HexagonDAGToDAGISel::SelectAddrGA(SDValue &N, SDValue &R) {
1291   return SelectGlobalAddress(N, R, false, 0);
1292 }
1293 
SelectAddrGP(SDValue & N,SDValue & R)1294 inline bool HexagonDAGToDAGISel::SelectAddrGP(SDValue &N, SDValue &R) {
1295   return SelectGlobalAddress(N, R, true, 0);
1296 }
1297 
SelectAnyImm(SDValue & N,SDValue & R)1298 inline bool HexagonDAGToDAGISel::SelectAnyImm(SDValue &N, SDValue &R) {
1299   return SelectAnyImmediate(N, R, 0);
1300 }
1301 
SelectAnyImm0(SDValue & N,SDValue & R)1302 inline bool HexagonDAGToDAGISel::SelectAnyImm0(SDValue &N, SDValue &R) {
1303   return SelectAnyImmediate(N, R, 0);
1304 }
SelectAnyImm1(SDValue & N,SDValue & R)1305 inline bool HexagonDAGToDAGISel::SelectAnyImm1(SDValue &N, SDValue &R) {
1306   return SelectAnyImmediate(N, R, 1);
1307 }
SelectAnyImm2(SDValue & N,SDValue & R)1308 inline bool HexagonDAGToDAGISel::SelectAnyImm2(SDValue &N, SDValue &R) {
1309   return SelectAnyImmediate(N, R, 2);
1310 }
SelectAnyImm3(SDValue & N,SDValue & R)1311 inline bool HexagonDAGToDAGISel::SelectAnyImm3(SDValue &N, SDValue &R) {
1312   return SelectAnyImmediate(N, R, 3);
1313 }
1314 
SelectAnyInt(SDValue & N,SDValue & R)1315 inline bool HexagonDAGToDAGISel::SelectAnyInt(SDValue &N, SDValue &R) {
1316   EVT T = N.getValueType();
1317   if (!T.isInteger() || T.getSizeInBits() != 32 || !isa<ConstantSDNode>(N))
1318     return false;
1319   R = N;
1320   return true;
1321 }
1322 
SelectAnyImmediate(SDValue & N,SDValue & R,uint32_t LogAlign)1323 bool HexagonDAGToDAGISel::SelectAnyImmediate(SDValue &N, SDValue &R,
1324                                              uint32_t LogAlign) {
1325   auto IsAligned = [LogAlign] (uint64_t V) -> bool {
1326     return alignTo(V, (uint64_t)1 << LogAlign) == V;
1327   };
1328 
1329   switch (N.getOpcode()) {
1330   case ISD::Constant: {
1331     if (N.getValueType() != MVT::i32)
1332       return false;
1333     int32_t V = cast<const ConstantSDNode>(N)->getZExtValue();
1334     if (!IsAligned(V))
1335       return false;
1336     R = CurDAG->getTargetConstant(V, SDLoc(N), N.getValueType());
1337     return true;
1338   }
1339   case HexagonISD::JT:
1340   case HexagonISD::CP:
1341     // These are assumed to always be aligned at least 8-byte boundary.
1342     if (LogAlign > 3)
1343       return false;
1344     R = N.getOperand(0);
1345     return true;
1346   case ISD::ExternalSymbol:
1347     // Symbols may be aligned at any boundary.
1348     if (LogAlign > 0)
1349       return false;
1350     R = N;
1351     return true;
1352   case ISD::BlockAddress:
1353     // Block address is always aligned at least 4-byte boundary.
1354     if (LogAlign > 2 || !IsAligned(cast<BlockAddressSDNode>(N)->getOffset()))
1355       return false;
1356     R = N;
1357     return true;
1358   }
1359 
1360   if (SelectGlobalAddress(N, R, false, LogAlign) ||
1361       SelectGlobalAddress(N, R, true, LogAlign))
1362     return true;
1363 
1364   return false;
1365 }
1366 
SelectGlobalAddress(SDValue & N,SDValue & R,bool UseGP,uint32_t LogAlign)1367 bool HexagonDAGToDAGISel::SelectGlobalAddress(SDValue &N, SDValue &R,
1368                                               bool UseGP, uint32_t LogAlign) {
1369   auto IsAligned = [LogAlign] (uint64_t V) -> bool {
1370     return alignTo(V, (uint64_t)1 << LogAlign) == V;
1371   };
1372 
1373   switch (N.getOpcode()) {
1374   case ISD::ADD: {
1375     SDValue N0 = N.getOperand(0);
1376     SDValue N1 = N.getOperand(1);
1377     unsigned GAOpc = N0.getOpcode();
1378     if (UseGP && GAOpc != HexagonISD::CONST32_GP)
1379       return false;
1380     if (!UseGP && GAOpc != HexagonISD::CONST32)
1381       return false;
1382     if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N1)) {
1383       SDValue Addr = N0.getOperand(0);
1384       // For the purpose of alignment, sextvalue and zextvalue are the same.
1385       if (!IsAligned(Const->getZExtValue()))
1386         return false;
1387       if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Addr)) {
1388         if (GA->getOpcode() == ISD::TargetGlobalAddress) {
1389           uint64_t NewOff = GA->getOffset() + (uint64_t)Const->getSExtValue();
1390           R = CurDAG->getTargetGlobalAddress(GA->getGlobal(), SDLoc(Const),
1391                                              N.getValueType(), NewOff);
1392           return true;
1393         }
1394       }
1395     }
1396     break;
1397   }
1398   case HexagonISD::CP:
1399   case HexagonISD::JT:
1400   case HexagonISD::CONST32:
1401     // The operand(0) of CONST32 is TargetGlobalAddress, which is what we
1402     // want in the instruction.
1403     if (!UseGP)
1404       R = N.getOperand(0);
1405     return !UseGP;
1406   case HexagonISD::CONST32_GP:
1407     if (UseGP)
1408       R = N.getOperand(0);
1409     return UseGP;
1410   default:
1411     return false;
1412   }
1413 
1414   return false;
1415 }
1416 
DetectUseSxtw(SDValue & N,SDValue & R)1417 bool HexagonDAGToDAGISel::DetectUseSxtw(SDValue &N, SDValue &R) {
1418   // This (complex pattern) function is meant to detect a sign-extension
1419   // i32->i64 on a per-operand basis. This would allow writing single
1420   // patterns that would cover a number of combinations of different ways
1421   // a sign-extensions could be written. For example:
1422   //   (mul (DetectUseSxtw x) (DetectUseSxtw y)) -> (M2_dpmpyss_s0 x y)
1423   // could match either one of these:
1424   //   (mul (sext x) (sext_inreg y))
1425   //   (mul (sext-load *p) (sext_inreg y))
1426   //   (mul (sext_inreg x) (sext y))
1427   // etc.
1428   //
1429   // The returned value will have type i64 and its low word will
1430   // contain the value being extended. The high bits are not specified.
1431   // The returned type is i64 because the original type of N was i64,
1432   // but the users of this function should only use the low-word of the
1433   // result, e.g.
1434   //  (mul sxtw:x, sxtw:y) -> (M2_dpmpyss_s0 (LoReg sxtw:x), (LoReg sxtw:y))
1435 
1436   if (N.getValueType() != MVT::i64)
1437     return false;
1438   unsigned Opc = N.getOpcode();
1439   switch (Opc) {
1440     case ISD::SIGN_EXTEND:
1441     case ISD::SIGN_EXTEND_INREG: {
1442       // sext_inreg has the source type as a separate operand.
1443       EVT T = Opc == ISD::SIGN_EXTEND
1444                 ? N.getOperand(0).getValueType()
1445                 : cast<VTSDNode>(N.getOperand(1))->getVT();
1446       unsigned SW = T.getSizeInBits();
1447       if (SW == 32)
1448         R = N.getOperand(0);
1449       else if (SW < 32)
1450         R = N;
1451       else
1452         return false;
1453       break;
1454     }
1455     case ISD::LOAD: {
1456       LoadSDNode *L = cast<LoadSDNode>(N);
1457       if (L->getExtensionType() != ISD::SEXTLOAD)
1458         return false;
1459       // All extending loads extend to i32, so even if the value in
1460       // memory is shorter than 32 bits, it will be i32 after the load.
1461       if (L->getMemoryVT().getSizeInBits() > 32)
1462         return false;
1463       R = N;
1464       break;
1465     }
1466     case ISD::SRA: {
1467       auto *S = dyn_cast<ConstantSDNode>(N.getOperand(1));
1468       if (!S || S->getZExtValue() != 32)
1469         return false;
1470       R = N;
1471       break;
1472     }
1473     default:
1474       return false;
1475   }
1476   EVT RT = R.getValueType();
1477   if (RT == MVT::i64)
1478     return true;
1479   assert(RT == MVT::i32);
1480   // This is only to produce a value of type i64. Do not rely on the
1481   // high bits produced by this.
1482   const SDLoc &dl(N);
1483   SDValue Ops[] = {
1484     CurDAG->getTargetConstant(Hexagon::DoubleRegsRegClassID, dl, MVT::i32),
1485     R, CurDAG->getTargetConstant(Hexagon::isub_hi, dl, MVT::i32),
1486     R, CurDAG->getTargetConstant(Hexagon::isub_lo, dl, MVT::i32)
1487   };
1488   SDNode *T = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl,
1489                                      MVT::i64, Ops);
1490   R = SDValue(T, 0);
1491   return true;
1492 }
1493 
keepsLowBits(const SDValue & Val,unsigned NumBits,SDValue & Src)1494 bool HexagonDAGToDAGISel::keepsLowBits(const SDValue &Val, unsigned NumBits,
1495       SDValue &Src) {
1496   unsigned Opc = Val.getOpcode();
1497   switch (Opc) {
1498   case ISD::SIGN_EXTEND:
1499   case ISD::ZERO_EXTEND:
1500   case ISD::ANY_EXTEND: {
1501     const SDValue &Op0 = Val.getOperand(0);
1502     EVT T = Op0.getValueType();
1503     if (T.isInteger() && T.getSizeInBits() == NumBits) {
1504       Src = Op0;
1505       return true;
1506     }
1507     break;
1508   }
1509   case ISD::SIGN_EXTEND_INREG:
1510   case ISD::AssertSext:
1511   case ISD::AssertZext:
1512     if (Val.getOperand(0).getValueType().isInteger()) {
1513       VTSDNode *T = cast<VTSDNode>(Val.getOperand(1));
1514       if (T->getVT().getSizeInBits() == NumBits) {
1515         Src = Val.getOperand(0);
1516         return true;
1517       }
1518     }
1519     break;
1520   case ISD::AND: {
1521     // Check if this is an AND with NumBits of lower bits set to 1.
1522     uint64_t Mask = (1 << NumBits) - 1;
1523     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1524       if (C->getZExtValue() == Mask) {
1525         Src = Val.getOperand(1);
1526         return true;
1527       }
1528     }
1529     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1530       if (C->getZExtValue() == Mask) {
1531         Src = Val.getOperand(0);
1532         return true;
1533       }
1534     }
1535     break;
1536   }
1537   case ISD::OR:
1538   case ISD::XOR: {
1539     // OR/XOR with the lower NumBits bits set to 0.
1540     uint64_t Mask = (1 << NumBits) - 1;
1541     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1542       if ((C->getZExtValue() & Mask) == 0) {
1543         Src = Val.getOperand(1);
1544         return true;
1545       }
1546     }
1547     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1548       if ((C->getZExtValue() & Mask) == 0) {
1549         Src = Val.getOperand(0);
1550         return true;
1551       }
1552     }
1553     break;
1554   }
1555   default:
1556     break;
1557   }
1558   return false;
1559 }
1560 
isAlignedMemNode(const MemSDNode * N) const1561 bool HexagonDAGToDAGISel::isAlignedMemNode(const MemSDNode *N) const {
1562   return N->getAlignment() >= N->getMemoryVT().getStoreSize();
1563 }
1564 
isSmallStackStore(const StoreSDNode * N) const1565 bool HexagonDAGToDAGISel::isSmallStackStore(const StoreSDNode *N) const {
1566   unsigned StackSize = MF->getFrameInfo().estimateStackSize(*MF);
1567   switch (N->getMemoryVT().getStoreSize()) {
1568     case 1:
1569       return StackSize <= 56;   // 1*2^6 - 8
1570     case 2:
1571       return StackSize <= 120;  // 2*2^6 - 8
1572     case 4:
1573       return StackSize <= 248;  // 4*2^6 - 8
1574     default:
1575       return false;
1576   }
1577 }
1578 
1579 // Return true when the given node fits in a positive half word.
isPositiveHalfWord(const SDNode * N) const1580 bool HexagonDAGToDAGISel::isPositiveHalfWord(const SDNode *N) const {
1581   if (const ConstantSDNode *CN = dyn_cast<const ConstantSDNode>(N)) {
1582     int64_t V = CN->getSExtValue();
1583     return V > 0 && isInt<16>(V);
1584   }
1585   if (N->getOpcode() == ISD::SIGN_EXTEND_INREG) {
1586     const VTSDNode *VN = dyn_cast<const VTSDNode>(N->getOperand(1));
1587     return VN->getVT().getSizeInBits() <= 16;
1588   }
1589   return false;
1590 }
1591 
hasOneUse(const SDNode * N) const1592 bool HexagonDAGToDAGISel::hasOneUse(const SDNode *N) const {
1593   return !CheckSingleUse || N->hasOneUse();
1594 }
1595 
1596 ////////////////////////////////////////////////////////////////////////////////
1597 // Rebalancing of address calculation trees
1598 
isOpcodeHandled(const SDNode * N)1599 static bool isOpcodeHandled(const SDNode *N) {
1600   switch (N->getOpcode()) {
1601     case ISD::ADD:
1602     case ISD::MUL:
1603       return true;
1604     case ISD::SHL:
1605       // We only handle constant shifts because these can be easily flattened
1606       // into multiplications by 2^Op1.
1607       return isa<ConstantSDNode>(N->getOperand(1).getNode());
1608     default:
1609       return false;
1610   }
1611 }
1612 
1613 /// Return the weight of an SDNode
getWeight(SDNode * N)1614 int HexagonDAGToDAGISel::getWeight(SDNode *N) {
1615   if (!isOpcodeHandled(N))
1616     return 1;
1617   assert(RootWeights.count(N) && "Cannot get weight of unseen root!");
1618   assert(RootWeights[N] != -1 && "Cannot get weight of unvisited root!");
1619   assert(RootWeights[N] != -2 && "Cannot get weight of RAWU'd root!");
1620   return RootWeights[N];
1621 }
1622 
getHeight(SDNode * N)1623 int HexagonDAGToDAGISel::getHeight(SDNode *N) {
1624   if (!isOpcodeHandled(N))
1625     return 0;
1626   assert(RootWeights.count(N) && RootWeights[N] >= 0 &&
1627       "Cannot query height of unvisited/RAUW'd node!");
1628   return RootHeights[N];
1629 }
1630 
1631 namespace {
1632 struct WeightedLeaf {
1633   SDValue Value;
1634   int Weight;
1635   int InsertionOrder;
1636 
WeightedLeaf__anon24b34d940811::WeightedLeaf1637   WeightedLeaf() : Value(SDValue()) { }
1638 
WeightedLeaf__anon24b34d940811::WeightedLeaf1639   WeightedLeaf(SDValue Value, int Weight, int InsertionOrder) :
1640     Value(Value), Weight(Weight), InsertionOrder(InsertionOrder) {
1641     assert(Weight >= 0 && "Weight must be >= 0");
1642   }
1643 
Compare__anon24b34d940811::WeightedLeaf1644   static bool Compare(const WeightedLeaf &A, const WeightedLeaf &B) {
1645     assert(A.Value.getNode() && B.Value.getNode());
1646     return A.Weight == B.Weight ?
1647             (A.InsertionOrder > B.InsertionOrder) :
1648             (A.Weight > B.Weight);
1649   }
1650 };
1651 
1652 /// A specialized priority queue for WeigthedLeaves. It automatically folds
1653 /// constants and allows removal of non-top elements while maintaining the
1654 /// priority order.
1655 class LeafPrioQueue {
1656   SmallVector<WeightedLeaf, 8> Q;
1657   bool HaveConst;
1658   WeightedLeaf ConstElt;
1659   unsigned Opcode;
1660 
1661 public:
empty()1662   bool empty() {
1663     return (!HaveConst && Q.empty());
1664   }
1665 
size()1666   size_t size() {
1667     return Q.size() + HaveConst;
1668   }
1669 
hasConst()1670   bool hasConst() {
1671     return HaveConst;
1672   }
1673 
top()1674   const WeightedLeaf &top() {
1675     if (HaveConst)
1676       return ConstElt;
1677     return Q.front();
1678   }
1679 
pop()1680   WeightedLeaf pop() {
1681     if (HaveConst) {
1682       HaveConst = false;
1683       return ConstElt;
1684     }
1685     std::pop_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1686     return Q.pop_back_val();
1687   }
1688 
push(WeightedLeaf L,bool SeparateConst=true)1689   void push(WeightedLeaf L, bool SeparateConst=true) {
1690     if (!HaveConst && SeparateConst && isa<ConstantSDNode>(L.Value)) {
1691       if (Opcode == ISD::MUL &&
1692           cast<ConstantSDNode>(L.Value)->getSExtValue() == 1)
1693         return;
1694       if (Opcode == ISD::ADD &&
1695           cast<ConstantSDNode>(L.Value)->getSExtValue() == 0)
1696         return;
1697 
1698       HaveConst = true;
1699       ConstElt = L;
1700     } else {
1701       Q.push_back(L);
1702       std::push_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1703     }
1704   }
1705 
1706   /// Push L to the bottom of the queue regardless of its weight. If L is
1707   /// constant, it will not be folded with other constants in the queue.
pushToBottom(WeightedLeaf L)1708   void pushToBottom(WeightedLeaf L) {
1709     L.Weight = 1000;
1710     push(L, false);
1711   }
1712 
1713   /// Search for a SHL(x, [<=MaxAmount]) subtree in the queue, return the one of
1714   /// lowest weight and remove it from the queue.
1715   WeightedLeaf findSHL(uint64_t MaxAmount);
1716 
1717   WeightedLeaf findMULbyConst();
1718 
LeafPrioQueue(unsigned Opcode)1719   LeafPrioQueue(unsigned Opcode) :
1720     HaveConst(false), Opcode(Opcode) { }
1721 };
1722 } // end anonymous namespace
1723 
findSHL(uint64_t MaxAmount)1724 WeightedLeaf LeafPrioQueue::findSHL(uint64_t MaxAmount) {
1725   int ResultPos;
1726   WeightedLeaf Result;
1727 
1728   for (int Pos = 0, End = Q.size(); Pos != End; ++Pos) {
1729     const WeightedLeaf &L = Q[Pos];
1730     const SDValue &Val = L.Value;
1731     if (Val.getOpcode() != ISD::SHL ||
1732         !isa<ConstantSDNode>(Val.getOperand(1)) ||
1733         Val.getConstantOperandVal(1) > MaxAmount)
1734       continue;
1735     if (!Result.Value.getNode() || Result.Weight > L.Weight ||
1736         (Result.Weight == L.Weight && Result.InsertionOrder > L.InsertionOrder))
1737     {
1738       Result = L;
1739       ResultPos = Pos;
1740     }
1741   }
1742 
1743   if (Result.Value.getNode()) {
1744     Q.erase(&Q[ResultPos]);
1745     std::make_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1746   }
1747 
1748   return Result;
1749 }
1750 
findMULbyConst()1751 WeightedLeaf LeafPrioQueue::findMULbyConst() {
1752   int ResultPos;
1753   WeightedLeaf Result;
1754 
1755   for (int Pos = 0, End = Q.size(); Pos != End; ++Pos) {
1756     const WeightedLeaf &L = Q[Pos];
1757     const SDValue &Val = L.Value;
1758     if (Val.getOpcode() != ISD::MUL ||
1759         !isa<ConstantSDNode>(Val.getOperand(1)) ||
1760         Val.getConstantOperandVal(1) > 127)
1761       continue;
1762     if (!Result.Value.getNode() || Result.Weight > L.Weight ||
1763         (Result.Weight == L.Weight && Result.InsertionOrder > L.InsertionOrder))
1764     {
1765       Result = L;
1766       ResultPos = Pos;
1767     }
1768   }
1769 
1770   if (Result.Value.getNode()) {
1771     Q.erase(&Q[ResultPos]);
1772     std::make_heap(Q.begin(), Q.end(), WeightedLeaf::Compare);
1773   }
1774 
1775   return Result;
1776 }
1777 
getMultiplierForSHL(SDNode * N)1778 SDValue HexagonDAGToDAGISel::getMultiplierForSHL(SDNode *N) {
1779   uint64_t MulFactor = 1ull << N->getConstantOperandVal(1);
1780   return CurDAG->getConstant(MulFactor, SDLoc(N),
1781                              N->getOperand(1).getValueType());
1782 }
1783 
1784 /// @returns the value x for which 2^x is a factor of Val
getPowerOf2Factor(SDValue Val)1785 static unsigned getPowerOf2Factor(SDValue Val) {
1786   if (Val.getOpcode() == ISD::MUL) {
1787     unsigned MaxFactor = 0;
1788     for (int i = 0; i < 2; ++i) {
1789       ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(i));
1790       if (!C)
1791         continue;
1792       const APInt &CInt = C->getAPIntValue();
1793       if (CInt.getBoolValue())
1794         MaxFactor = CInt.countTrailingZeros();
1795     }
1796     return MaxFactor;
1797   }
1798   if (Val.getOpcode() == ISD::SHL) {
1799     if (!isa<ConstantSDNode>(Val.getOperand(1).getNode()))
1800       return 0;
1801     return (unsigned) Val.getConstantOperandVal(1);
1802   }
1803 
1804   return 0;
1805 }
1806 
1807 /// @returns true if V>>Amount will eliminate V's operation on its child
willShiftRightEliminate(SDValue V,unsigned Amount)1808 static bool willShiftRightEliminate(SDValue V, unsigned Amount) {
1809   if (V.getOpcode() == ISD::MUL) {
1810     SDValue Ops[] = { V.getOperand(0), V.getOperand(1) };
1811     for (int i = 0; i < 2; ++i)
1812       if (isa<ConstantSDNode>(Ops[i].getNode()) &&
1813           V.getConstantOperandVal(i) % (1ULL << Amount) == 0) {
1814         uint64_t NewConst = V.getConstantOperandVal(i) >> Amount;
1815         return (NewConst == 1);
1816       }
1817   } else if (V.getOpcode() == ISD::SHL) {
1818     return (Amount == V.getConstantOperandVal(1));
1819   }
1820 
1821   return false;
1822 }
1823 
factorOutPowerOf2(SDValue V,unsigned Power)1824 SDValue HexagonDAGToDAGISel::factorOutPowerOf2(SDValue V, unsigned Power) {
1825   SDValue Ops[] = { V.getOperand(0), V.getOperand(1) };
1826   if (V.getOpcode() == ISD::MUL) {
1827     for (int i=0; i < 2; ++i) {
1828       if (isa<ConstantSDNode>(Ops[i].getNode()) &&
1829           V.getConstantOperandVal(i) % ((uint64_t)1 << Power) == 0) {
1830         uint64_t NewConst = V.getConstantOperandVal(i) >> Power;
1831         if (NewConst == 1)
1832           return Ops[!i];
1833         Ops[i] = CurDAG->getConstant(NewConst,
1834                                      SDLoc(V), V.getValueType());
1835         break;
1836       }
1837     }
1838   } else if (V.getOpcode() == ISD::SHL) {
1839     uint64_t ShiftAmount = V.getConstantOperandVal(1);
1840     if (ShiftAmount == Power)
1841       return Ops[0];
1842     Ops[1] = CurDAG->getConstant(ShiftAmount - Power,
1843                                  SDLoc(V), V.getValueType());
1844   }
1845 
1846   return CurDAG->getNode(V.getOpcode(), SDLoc(V), V.getValueType(), Ops);
1847 }
1848 
isTargetConstant(const SDValue & V)1849 static bool isTargetConstant(const SDValue &V) {
1850   return V.getOpcode() == HexagonISD::CONST32 ||
1851          V.getOpcode() == HexagonISD::CONST32_GP;
1852 }
1853 
getUsesInFunction(const Value * V)1854 unsigned HexagonDAGToDAGISel::getUsesInFunction(const Value *V) {
1855   if (GAUsesInFunction.count(V))
1856     return GAUsesInFunction[V];
1857 
1858   unsigned Result = 0;
1859   const Function &CurF = CurDAG->getMachineFunction().getFunction();
1860   for (const User *U : V->users()) {
1861     if (isa<Instruction>(U) &&
1862         cast<Instruction>(U)->getParent()->getParent() == &CurF)
1863       ++Result;
1864   }
1865 
1866   GAUsesInFunction[V] = Result;
1867 
1868   return Result;
1869 }
1870 
1871 /// Note - After calling this, N may be dead. It may have been replaced by a
1872 /// new node, so always use the returned value in place of N.
1873 ///
1874 /// @returns The SDValue taking the place of N (which could be N if it is
1875 /// unchanged)
balanceSubTree(SDNode * N,bool TopLevel)1876 SDValue HexagonDAGToDAGISel::balanceSubTree(SDNode *N, bool TopLevel) {
1877   assert(RootWeights.count(N) && "Cannot balance non-root node.");
1878   assert(RootWeights[N] != -2 && "This node was RAUW'd!");
1879   assert(!TopLevel || N->getOpcode() == ISD::ADD);
1880 
1881   // Return early if this node was already visited
1882   if (RootWeights[N] != -1)
1883     return SDValue(N, 0);
1884 
1885   assert(isOpcodeHandled(N));
1886 
1887   SDValue Op0 = N->getOperand(0);
1888   SDValue Op1 = N->getOperand(1);
1889 
1890   // Return early if the operands will remain unchanged or are all roots
1891   if ((!isOpcodeHandled(Op0.getNode()) || RootWeights.count(Op0.getNode())) &&
1892       (!isOpcodeHandled(Op1.getNode()) || RootWeights.count(Op1.getNode()))) {
1893     SDNode *Op0N = Op0.getNode();
1894     int Weight;
1895     if (isOpcodeHandled(Op0N) && RootWeights[Op0N] == -1) {
1896       Weight = getWeight(balanceSubTree(Op0N).getNode());
1897       // Weight = calculateWeight(Op0N);
1898     } else
1899       Weight = getWeight(Op0N);
1900 
1901     SDNode *Op1N = N->getOperand(1).getNode(); // Op1 may have been RAUWd
1902     if (isOpcodeHandled(Op1N) && RootWeights[Op1N] == -1) {
1903       Weight += getWeight(balanceSubTree(Op1N).getNode());
1904       // Weight += calculateWeight(Op1N);
1905     } else
1906       Weight += getWeight(Op1N);
1907 
1908     RootWeights[N] = Weight;
1909     RootHeights[N] = std::max(getHeight(N->getOperand(0).getNode()),
1910                               getHeight(N->getOperand(1).getNode())) + 1;
1911 
1912     LLVM_DEBUG(dbgs() << "--> No need to balance root (Weight=" << Weight
1913                       << " Height=" << RootHeights[N] << "): ");
1914     LLVM_DEBUG(N->dump(CurDAG));
1915 
1916     return SDValue(N, 0);
1917   }
1918 
1919   LLVM_DEBUG(dbgs() << "** Balancing root node: ");
1920   LLVM_DEBUG(N->dump(CurDAG));
1921 
1922   unsigned NOpcode = N->getOpcode();
1923 
1924   LeafPrioQueue Leaves(NOpcode);
1925   SmallVector<SDValue, 4> Worklist;
1926   Worklist.push_back(SDValue(N, 0));
1927 
1928   // SHL nodes will be converted to MUL nodes
1929   if (NOpcode == ISD::SHL)
1930     NOpcode = ISD::MUL;
1931 
1932   bool CanFactorize = false;
1933   WeightedLeaf Mul1, Mul2;
1934   unsigned MaxPowerOf2 = 0;
1935   WeightedLeaf GA;
1936 
1937   // Do not try to factor out a shift if there is already a shift at the tip of
1938   // the tree.
1939   bool HaveTopLevelShift = false;
1940   if (TopLevel &&
1941       ((isOpcodeHandled(Op0.getNode()) && Op0.getOpcode() == ISD::SHL &&
1942                         Op0.getConstantOperandVal(1) < 4) ||
1943        (isOpcodeHandled(Op1.getNode()) && Op1.getOpcode() == ISD::SHL &&
1944                         Op1.getConstantOperandVal(1) < 4)))
1945     HaveTopLevelShift = true;
1946 
1947   // Flatten the subtree into an ordered list of leaves; at the same time
1948   // determine whether the tree is already balanced.
1949   int InsertionOrder = 0;
1950   SmallDenseMap<SDValue, int> NodeHeights;
1951   bool Imbalanced = false;
1952   int CurrentWeight = 0;
1953   while (!Worklist.empty()) {
1954     SDValue Child = Worklist.pop_back_val();
1955 
1956     if (Child.getNode() != N && RootWeights.count(Child.getNode())) {
1957       // CASE 1: Child is a root note
1958 
1959       int Weight = RootWeights[Child.getNode()];
1960       if (Weight == -1) {
1961         Child = balanceSubTree(Child.getNode());
1962         // calculateWeight(Child.getNode());
1963         Weight = getWeight(Child.getNode());
1964       } else if (Weight == -2) {
1965         // Whoops, this node was RAUWd by one of the balanceSubTree calls we
1966         // made. Our worklist isn't up to date anymore.
1967         // Restart the whole process.
1968         LLVM_DEBUG(dbgs() << "--> Subtree was RAUWd. Restarting...\n");
1969         return balanceSubTree(N, TopLevel);
1970       }
1971 
1972       NodeHeights[Child] = 1;
1973       CurrentWeight += Weight;
1974 
1975       unsigned PowerOf2;
1976       if (TopLevel && !CanFactorize && !HaveTopLevelShift &&
1977           (Child.getOpcode() == ISD::MUL || Child.getOpcode() == ISD::SHL) &&
1978           Child.hasOneUse() && (PowerOf2 = getPowerOf2Factor(Child))) {
1979         // Try to identify two factorizable MUL/SHL children greedily. Leave
1980         // them out of the priority queue for now so we can deal with them
1981         // after.
1982         if (!Mul1.Value.getNode()) {
1983           Mul1 = WeightedLeaf(Child, Weight, InsertionOrder++);
1984           MaxPowerOf2 = PowerOf2;
1985         } else {
1986           Mul2 = WeightedLeaf(Child, Weight, InsertionOrder++);
1987           MaxPowerOf2 = std::min(MaxPowerOf2, PowerOf2);
1988 
1989           // Our addressing modes can only shift by a maximum of 3
1990           if (MaxPowerOf2 > 3)
1991             MaxPowerOf2 = 3;
1992 
1993           CanFactorize = true;
1994         }
1995       } else
1996         Leaves.push(WeightedLeaf(Child, Weight, InsertionOrder++));
1997     } else if (!isOpcodeHandled(Child.getNode())) {
1998       // CASE 2: Child is an unhandled kind of node (e.g. constant)
1999       int Weight = getWeight(Child.getNode());
2000 
2001       NodeHeights[Child] = getHeight(Child.getNode());
2002       CurrentWeight += Weight;
2003 
2004       if (isTargetConstant(Child) && !GA.Value.getNode())
2005         GA = WeightedLeaf(Child, Weight, InsertionOrder++);
2006       else
2007         Leaves.push(WeightedLeaf(Child, Weight, InsertionOrder++));
2008     } else {
2009       // CASE 3: Child is a subtree of same opcode
2010       // Visit children first, then flatten.
2011       unsigned ChildOpcode = Child.getOpcode();
2012       assert(ChildOpcode == NOpcode ||
2013              (NOpcode == ISD::MUL && ChildOpcode == ISD::SHL));
2014 
2015       // Convert SHL to MUL
2016       SDValue Op1;
2017       if (ChildOpcode == ISD::SHL)
2018         Op1 = getMultiplierForSHL(Child.getNode());
2019       else
2020         Op1 = Child->getOperand(1);
2021 
2022       if (!NodeHeights.count(Op1) || !NodeHeights.count(Child->getOperand(0))) {
2023         assert(!NodeHeights.count(Child) && "Parent visited before children?");
2024         // Visit children first, then re-visit this node
2025         Worklist.push_back(Child);
2026         Worklist.push_back(Op1);
2027         Worklist.push_back(Child->getOperand(0));
2028       } else {
2029         // Back at this node after visiting the children
2030         if (std::abs(NodeHeights[Op1] - NodeHeights[Child->getOperand(0)]) > 1)
2031           Imbalanced = true;
2032 
2033         NodeHeights[Child] = std::max(NodeHeights[Op1],
2034                                       NodeHeights[Child->getOperand(0)]) + 1;
2035       }
2036     }
2037   }
2038 
2039   LLVM_DEBUG(dbgs() << "--> Current height=" << NodeHeights[SDValue(N, 0)]
2040                     << " weight=" << CurrentWeight
2041                     << " imbalanced=" << Imbalanced << "\n");
2042 
2043   // Transform MUL(x, C * 2^Y) + SHL(z, Y) -> SHL(ADD(MUL(x, C), z), Y)
2044   //  This factors out a shift in order to match memw(a<<Y+b).
2045   if (CanFactorize && (willShiftRightEliminate(Mul1.Value, MaxPowerOf2) ||
2046                        willShiftRightEliminate(Mul2.Value, MaxPowerOf2))) {
2047     LLVM_DEBUG(dbgs() << "--> Found common factor for two MUL children!\n");
2048     int Weight = Mul1.Weight + Mul2.Weight;
2049     int Height = std::max(NodeHeights[Mul1.Value], NodeHeights[Mul2.Value]) + 1;
2050     SDValue Mul1Factored = factorOutPowerOf2(Mul1.Value, MaxPowerOf2);
2051     SDValue Mul2Factored = factorOutPowerOf2(Mul2.Value, MaxPowerOf2);
2052     SDValue Sum = CurDAG->getNode(ISD::ADD, SDLoc(N), Mul1.Value.getValueType(),
2053                                   Mul1Factored, Mul2Factored);
2054     SDValue Const = CurDAG->getConstant(MaxPowerOf2, SDLoc(N),
2055                                         Mul1.Value.getValueType());
2056     SDValue New = CurDAG->getNode(ISD::SHL, SDLoc(N), Mul1.Value.getValueType(),
2057                                   Sum, Const);
2058     NodeHeights[New] = Height;
2059     Leaves.push(WeightedLeaf(New, Weight, Mul1.InsertionOrder));
2060   } else if (Mul1.Value.getNode()) {
2061     // We failed to factorize two MULs, so now the Muls are left outside the
2062     // queue... add them back.
2063     Leaves.push(Mul1);
2064     if (Mul2.Value.getNode())
2065       Leaves.push(Mul2);
2066     CanFactorize = false;
2067   }
2068 
2069   // Combine GA + Constant -> GA+Offset, but only if GA is not used elsewhere
2070   // and the root node itself is not used more than twice. This reduces the
2071   // amount of additional constant extenders introduced by this optimization.
2072   bool CombinedGA = false;
2073   if (NOpcode == ISD::ADD && GA.Value.getNode() && Leaves.hasConst() &&
2074       GA.Value.hasOneUse() && N->use_size() < 3) {
2075     GlobalAddressSDNode *GANode =
2076       cast<GlobalAddressSDNode>(GA.Value.getOperand(0));
2077     ConstantSDNode *Offset = cast<ConstantSDNode>(Leaves.top().Value);
2078 
2079     if (getUsesInFunction(GANode->getGlobal()) == 1 && Offset->hasOneUse() &&
2080         getTargetLowering()->isOffsetFoldingLegal(GANode)) {
2081       LLVM_DEBUG(dbgs() << "--> Combining GA and offset ("
2082                         << Offset->getSExtValue() << "): ");
2083       LLVM_DEBUG(GANode->dump(CurDAG));
2084 
2085       SDValue NewTGA =
2086         CurDAG->getTargetGlobalAddress(GANode->getGlobal(), SDLoc(GA.Value),
2087             GANode->getValueType(0),
2088             GANode->getOffset() + (uint64_t)Offset->getSExtValue());
2089       GA.Value = CurDAG->getNode(GA.Value.getOpcode(), SDLoc(GA.Value),
2090           GA.Value.getValueType(), NewTGA);
2091       GA.Weight += Leaves.top().Weight;
2092 
2093       NodeHeights[GA.Value] = getHeight(GA.Value.getNode());
2094       CombinedGA = true;
2095 
2096       Leaves.pop(); // Remove the offset constant from the queue
2097     }
2098   }
2099 
2100   if ((RebalanceOnlyForOptimizations && !CanFactorize && !CombinedGA) ||
2101       (RebalanceOnlyImbalancedTrees && !Imbalanced)) {
2102     RootWeights[N] = CurrentWeight;
2103     RootHeights[N] = NodeHeights[SDValue(N, 0)];
2104 
2105     return SDValue(N, 0);
2106   }
2107 
2108   // Combine GA + SHL(x, C<=31) so we will match Rx=add(#u8,asl(Rx,#U5))
2109   if (NOpcode == ISD::ADD && GA.Value.getNode()) {
2110     WeightedLeaf SHL = Leaves.findSHL(31);
2111     if (SHL.Value.getNode()) {
2112       int Height = std::max(NodeHeights[GA.Value], NodeHeights[SHL.Value]) + 1;
2113       GA.Value = CurDAG->getNode(ISD::ADD, SDLoc(GA.Value),
2114                                  GA.Value.getValueType(),
2115                                  GA.Value, SHL.Value);
2116       GA.Weight = SHL.Weight; // Specifically ignore the GA weight here
2117       NodeHeights[GA.Value] = Height;
2118     }
2119   }
2120 
2121   if (GA.Value.getNode())
2122     Leaves.push(GA);
2123 
2124   // If this is the top level and we haven't factored out a shift, we should try
2125   // to move a constant to the bottom to match addressing modes like memw(rX+C)
2126   if (TopLevel && !CanFactorize && Leaves.hasConst()) {
2127     LLVM_DEBUG(dbgs() << "--> Pushing constant to tip of tree.");
2128     Leaves.pushToBottom(Leaves.pop());
2129   }
2130 
2131   const DataLayout &DL = CurDAG->getDataLayout();
2132   const TargetLowering &TLI = *getTargetLowering();
2133 
2134   // Rebuild the tree using Huffman's algorithm
2135   while (Leaves.size() > 1) {
2136     WeightedLeaf L0 = Leaves.pop();
2137 
2138     // See whether we can grab a MUL to form an add(Rx,mpyi(Ry,#u6)),
2139     // otherwise just get the next leaf
2140     WeightedLeaf L1 = Leaves.findMULbyConst();
2141     if (!L1.Value.getNode())
2142       L1 = Leaves.pop();
2143 
2144     assert(L0.Weight <= L1.Weight && "Priority queue is broken!");
2145 
2146     SDValue V0 = L0.Value;
2147     int V0Weight = L0.Weight;
2148     SDValue V1 = L1.Value;
2149     int V1Weight = L1.Weight;
2150 
2151     // Make sure that none of these nodes have been RAUW'd
2152     if ((RootWeights.count(V0.getNode()) && RootWeights[V0.getNode()] == -2) ||
2153         (RootWeights.count(V1.getNode()) && RootWeights[V1.getNode()] == -2)) {
2154       LLVM_DEBUG(dbgs() << "--> Subtree was RAUWd. Restarting...\n");
2155       return balanceSubTree(N, TopLevel);
2156     }
2157 
2158     ConstantSDNode *V0C = dyn_cast<ConstantSDNode>(V0);
2159     ConstantSDNode *V1C = dyn_cast<ConstantSDNode>(V1);
2160     EVT VT = N->getValueType(0);
2161     SDValue NewNode;
2162 
2163     if (V0C && !V1C) {
2164       std::swap(V0, V1);
2165       std::swap(V0C, V1C);
2166     }
2167 
2168     // Calculate height of this node
2169     assert(NodeHeights.count(V0) && NodeHeights.count(V1) &&
2170            "Children must have been visited before re-combining them!");
2171     int Height = std::max(NodeHeights[V0], NodeHeights[V1]) + 1;
2172 
2173     // Rebuild this node (and restore SHL from MUL if needed)
2174     if (V1C && NOpcode == ISD::MUL && V1C->getAPIntValue().isPowerOf2())
2175       NewNode = CurDAG->getNode(
2176           ISD::SHL, SDLoc(V0), VT, V0,
2177           CurDAG->getConstant(
2178               V1C->getAPIntValue().logBase2(), SDLoc(N),
2179               TLI.getScalarShiftAmountTy(DL, V0.getValueType())));
2180     else
2181       NewNode = CurDAG->getNode(NOpcode, SDLoc(N), VT, V0, V1);
2182 
2183     NodeHeights[NewNode] = Height;
2184 
2185     int Weight = V0Weight + V1Weight;
2186     Leaves.push(WeightedLeaf(NewNode, Weight, L0.InsertionOrder));
2187 
2188     LLVM_DEBUG(dbgs() << "--> Built new node (Weight=" << Weight
2189                       << ",Height=" << Height << "):\n");
2190     LLVM_DEBUG(NewNode.dump());
2191   }
2192 
2193   assert(Leaves.size() == 1);
2194   SDValue NewRoot = Leaves.top().Value;
2195 
2196   assert(NodeHeights.count(NewRoot));
2197   int Height = NodeHeights[NewRoot];
2198 
2199   // Restore SHL if we earlier converted it to a MUL
2200   if (NewRoot.getOpcode() == ISD::MUL) {
2201     ConstantSDNode *V1C = dyn_cast<ConstantSDNode>(NewRoot.getOperand(1));
2202     if (V1C && V1C->getAPIntValue().isPowerOf2()) {
2203       EVT VT = NewRoot.getValueType();
2204       SDValue V0 = NewRoot.getOperand(0);
2205       NewRoot = CurDAG->getNode(
2206           ISD::SHL, SDLoc(NewRoot), VT, V0,
2207           CurDAG->getConstant(
2208               V1C->getAPIntValue().logBase2(), SDLoc(NewRoot),
2209               TLI.getScalarShiftAmountTy(DL, V0.getValueType())));
2210     }
2211   }
2212 
2213   if (N != NewRoot.getNode()) {
2214     LLVM_DEBUG(dbgs() << "--> Root is now: ");
2215     LLVM_DEBUG(NewRoot.dump());
2216 
2217     // Replace all uses of old root by new root
2218     CurDAG->ReplaceAllUsesWith(N, NewRoot.getNode());
2219     // Mark that we have RAUW'd N
2220     RootWeights[N] = -2;
2221   } else {
2222     LLVM_DEBUG(dbgs() << "--> Root unchanged.\n");
2223   }
2224 
2225   RootWeights[NewRoot.getNode()] = Leaves.top().Weight;
2226   RootHeights[NewRoot.getNode()] = Height;
2227 
2228   return NewRoot;
2229 }
2230 
rebalanceAddressTrees()2231 void HexagonDAGToDAGISel::rebalanceAddressTrees() {
2232   for (auto I = CurDAG->allnodes_begin(), E = CurDAG->allnodes_end(); I != E;) {
2233     SDNode *N = &*I++;
2234     if (N->getOpcode() != ISD::LOAD && N->getOpcode() != ISD::STORE)
2235       continue;
2236 
2237     SDValue BasePtr = cast<MemSDNode>(N)->getBasePtr();
2238     if (BasePtr.getOpcode() != ISD::ADD)
2239       continue;
2240 
2241     // We've already processed this node
2242     if (RootWeights.count(BasePtr.getNode()))
2243       continue;
2244 
2245     LLVM_DEBUG(dbgs() << "** Rebalancing address calculation in node: ");
2246     LLVM_DEBUG(N->dump(CurDAG));
2247 
2248     // FindRoots
2249     SmallVector<SDNode *, 4> Worklist;
2250 
2251     Worklist.push_back(BasePtr.getOperand(0).getNode());
2252     Worklist.push_back(BasePtr.getOperand(1).getNode());
2253 
2254     while (!Worklist.empty()) {
2255       SDNode *N = Worklist.pop_back_val();
2256       unsigned Opcode = N->getOpcode();
2257 
2258       if (!isOpcodeHandled(N))
2259         continue;
2260 
2261       Worklist.push_back(N->getOperand(0).getNode());
2262       Worklist.push_back(N->getOperand(1).getNode());
2263 
2264       // Not a root if it has only one use and same opcode as its parent
2265       if (N->hasOneUse() && Opcode == N->use_begin()->getOpcode())
2266         continue;
2267 
2268       // This root node has already been processed
2269       if (RootWeights.count(N))
2270         continue;
2271 
2272       RootWeights[N] = -1;
2273     }
2274 
2275     // Balance node itself
2276     RootWeights[BasePtr.getNode()] = -1;
2277     SDValue NewBasePtr = balanceSubTree(BasePtr.getNode(), /*TopLevel=*/ true);
2278 
2279     if (N->getOpcode() == ISD::LOAD)
2280       N = CurDAG->UpdateNodeOperands(N, N->getOperand(0),
2281             NewBasePtr, N->getOperand(2));
2282     else
2283       N = CurDAG->UpdateNodeOperands(N, N->getOperand(0), N->getOperand(1),
2284             NewBasePtr, N->getOperand(3));
2285 
2286     LLVM_DEBUG(dbgs() << "--> Final node: ");
2287     LLVM_DEBUG(N->dump(CurDAG));
2288   }
2289 
2290   CurDAG->RemoveDeadNodes();
2291   GAUsesInFunction.clear();
2292   RootHeights.clear();
2293   RootWeights.clear();
2294 }
2295