1//===-- HexagonISelDAGToDAGHVX.cpp ----------------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9#include "Hexagon.h"
10#include "HexagonISelDAGToDAG.h"
11#include "HexagonISelLowering.h"
12#include "HexagonTargetMachine.h"
13#include "llvm/ADT/SetVector.h"
14#include "llvm/CodeGen/MachineInstrBuilder.h"
15#include "llvm/CodeGen/SelectionDAGISel.h"
16#include "llvm/IR/Intrinsics.h"
17#include "llvm/IR/IntrinsicsHexagon.h"
18#include "llvm/Support/CommandLine.h"
19#include "llvm/Support/Debug.h"
20
21#include <deque>
22#include <map>
23#include <set>
24#include <utility>
25#include <vector>
26
27#define DEBUG_TYPE "hexagon-isel"
28
29using namespace llvm;
30
31namespace {
32
33// --------------------------------------------------------------------
34// Implementation of permutation networks.
35
36// Implementation of the node routing through butterfly networks:
37// - Forward delta.
38// - Reverse delta.
39// - Benes.
40//
41//
42// Forward delta network consists of log(N) steps, where N is the number
43// of inputs. In each step, an input can stay in place, or it can get
44// routed to another position[1]. The step after that consists of two
45// networks, each half in size in terms of the number of nodes. In those
46// terms, in the given step, an input can go to either the upper or the
47// lower network in the next step.
48//
49// [1] Hexagon's vdelta/vrdelta allow an element to be routed to both
50// positions as long as there is no conflict.
51
52// Here's a delta network for 8 inputs, only the switching routes are
53// shown:
54//
55//         Steps:
56//         |- 1 ---------------|- 2 -----|- 3 -|
57//
58// Inp[0] ***                 ***       ***   *** Out[0]
59//           \               /   \     /   \ /
60//            \             /     \   /     X
61//             \           /       \ /     / \
62// Inp[1] ***   \         /   ***   X   ***   *** Out[1]
63//           \   \       /   /   \ / \ /
64//            \   \     /   /     X   X
65//             \   \   /   /     / \ / \
66// Inp[2] ***   \   \ /   /   ***   X   ***   *** Out[2]
67//           \   \   X   /   /     / \     \ /
68//            \   \ / \ /   /     /   \     X
69//             \   X   X   /     /     \   / \
70// Inp[3] ***   \ / \ / \ /   ***       ***   *** Out[3]
71//           \   X   X   X   /
72//            \ / \ / \ / \ /
73//             X   X   X   X
74//            / \ / \ / \ / \
75//           /   X   X   X   \
76// Inp[4] ***   / \ / \ / \   ***       ***   *** Out[4]
77//             /   X   X   \     \     /   \ /
78//            /   / \ / \   \     \   /     X
79//           /   /   X   \   \     \ /     / \
80// Inp[5] ***   /   / \   \   ***   X   ***   *** Out[5]
81//             /   /   \   \     \ / \ /
82//            /   /     \   \     X   X
83//           /   /       \   \   / \ / \
84// Inp[6] ***   /         \   ***   X   ***   *** Out[6]
85//             /           \       / \     \ /
86//            /             \     /   \     X
87//           /               \   /     \   / \
88// Inp[7] ***                 ***       ***   *** Out[7]
89//
90//
91// Reverse delta network is same as delta network, with the steps in
92// the opposite order.
93//
94//
95// Benes network is a forward delta network immediately followed by
96// a reverse delta network.
97
98enum class ColorKind { None, Red, Black };
99
100// Graph coloring utility used to partition nodes into two groups:
101// they will correspond to nodes routed to the upper and lower networks.
102struct Coloring {
103  using Node = int;
104  using MapType = std::map<Node, ColorKind>;
105  static constexpr Node Ignore = Node(-1);
106
107  Coloring(ArrayRef<Node> Ord) : Order(Ord) {
108    build();
109    if (!color())
110      Colors.clear();
111  }
112
113  const MapType &colors() const {
114    return Colors;
115  }
116
117  ColorKind other(ColorKind Color) {
118    if (Color == ColorKind::None)
119      return ColorKind::Red;
120    return Color == ColorKind::Red ? ColorKind::Black : ColorKind::Red;
121  }
122
123  LLVM_DUMP_METHOD void dump() const;
124
125private:
126  ArrayRef<Node> Order;
127  MapType Colors;
128  std::set<Node> Needed;
129
130  using NodeSet = std::set<Node>;
131  std::map<Node,NodeSet> Edges;
132
133  Node conj(Node Pos) {
134    Node Num = Order.size();
135    return (Pos < Num/2) ? Pos + Num/2 : Pos - Num/2;
136  }
137
138  ColorKind getColor(Node N) {
139    auto F = Colors.find(N);
140    return F != Colors.end() ? F->second : ColorKind::None;
141  }
142
143  std::pair<bool, ColorKind> getUniqueColor(const NodeSet &Nodes);
144
145  void build();
146  bool color();
147};
148} // namespace
149
150std::pair<bool, ColorKind> Coloring::getUniqueColor(const NodeSet &Nodes) {
151  auto Color = ColorKind::None;
152  for (Node N : Nodes) {
153    ColorKind ColorN = getColor(N);
154    if (ColorN == ColorKind::None)
155      continue;
156    if (Color == ColorKind::None)
157      Color = ColorN;
158    else if (Color != ColorKind::None && Color != ColorN)
159      return { false, ColorKind::None };
160  }
161  return { true, Color };
162}
163
164void Coloring::build() {
165  // Add Order[P] and Order[conj(P)] to Edges.
166  for (unsigned P = 0; P != Order.size(); ++P) {
167    Node I = Order[P];
168    if (I != Ignore) {
169      Needed.insert(I);
170      Node PC = Order[conj(P)];
171      if (PC != Ignore && PC != I)
172        Edges[I].insert(PC);
173    }
174  }
175  // Add I and conj(I) to Edges.
176  for (unsigned I = 0; I != Order.size(); ++I) {
177    if (!Needed.count(I))
178      continue;
179    Node C = conj(I);
180    // This will create an entry in the edge table, even if I is not
181    // connected to any other node. This is necessary, because it still
182    // needs to be colored.
183    NodeSet &Is = Edges[I];
184    if (Needed.count(C))
185      Is.insert(C);
186  }
187}
188
189bool Coloring::color() {
190  SetVector<Node> FirstQ;
191  auto Enqueue = [this,&FirstQ] (Node N) {
192    SetVector<Node> Q;
193    Q.insert(N);
194    for (unsigned I = 0; I != Q.size(); ++I) {
195      NodeSet &Ns = Edges[Q[I]];
196      Q.insert(Ns.begin(), Ns.end());
197    }
198    FirstQ.insert(Q.begin(), Q.end());
199  };
200  for (Node N : Needed)
201    Enqueue(N);
202
203  for (Node N : FirstQ) {
204    if (Colors.count(N))
205      continue;
206    NodeSet &Ns = Edges[N];
207    auto P = getUniqueColor(Ns);
208    if (!P.first)
209      return false;
210    Colors[N] = other(P.second);
211  }
212
213  // First, color nodes that don't have any dups.
214  for (auto E : Edges) {
215    Node N = E.first;
216    if (!Needed.count(conj(N)) || Colors.count(N))
217      continue;
218    auto P = getUniqueColor(E.second);
219    if (!P.first)
220      return false;
221    Colors[N] = other(P.second);
222  }
223
224  // Now, nodes that are still uncolored. Since the graph can be modified
225  // in this step, create a work queue.
226  std::vector<Node> WorkQ;
227  for (auto E : Edges) {
228    Node N = E.first;
229    if (!Colors.count(N))
230      WorkQ.push_back(N);
231  }
232
233  for (unsigned I = 0; I < WorkQ.size(); ++I) {
234    Node N = WorkQ[I];
235    NodeSet &Ns = Edges[N];
236    auto P = getUniqueColor(Ns);
237    if (P.first) {
238      Colors[N] = other(P.second);
239      continue;
240    }
241
242    // Coloring failed. Split this node.
243    Node C = conj(N);
244    ColorKind ColorN = other(ColorKind::None);
245    ColorKind ColorC = other(ColorN);
246    NodeSet &Cs = Edges[C];
247    NodeSet CopyNs = Ns;
248    for (Node M : CopyNs) {
249      ColorKind ColorM = getColor(M);
250      if (ColorM == ColorC) {
251        // Connect M with C, disconnect M from N.
252        Cs.insert(M);
253        Edges[M].insert(C);
254        Ns.erase(M);
255        Edges[M].erase(N);
256      }
257    }
258    Colors[N] = ColorN;
259    Colors[C] = ColorC;
260  }
261
262  // Explicitly assign "None" to all uncolored nodes.
263  for (unsigned I = 0; I != Order.size(); ++I)
264    if (Colors.count(I) == 0)
265      Colors[I] = ColorKind::None;
266
267  return true;
268}
269
270#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
271void Coloring::dump() const {
272  dbgs() << "{ Order:   {";
273  for (unsigned I = 0; I != Order.size(); ++I) {
274    Node P = Order[I];
275    if (P != Ignore)
276      dbgs() << ' ' << P;
277    else
278      dbgs() << " -";
279  }
280  dbgs() << " }\n";
281  dbgs() << "  Needed: {";
282  for (Node N : Needed)
283    dbgs() << ' ' << N;
284  dbgs() << " }\n";
285
286  dbgs() << "  Edges: {\n";
287  for (auto E : Edges) {
288    dbgs() << "    " << E.first << " -> {";
289    for (auto N : E.second)
290      dbgs() << ' ' << N;
291    dbgs() << " }\n";
292  }
293  dbgs() << "  }\n";
294
295  auto ColorKindToName = [](ColorKind C) {
296    switch (C) {
297    case ColorKind::None:
298      return "None";
299    case ColorKind::Red:
300      return "Red";
301    case ColorKind::Black:
302      return "Black";
303    }
304    llvm_unreachable("all ColorKinds should be handled by the switch above");
305  };
306
307  dbgs() << "  Colors: {\n";
308  for (auto C : Colors)
309    dbgs() << "    " << C.first << " -> " << ColorKindToName(C.second) << "\n";
310  dbgs() << "  }\n}\n";
311}
312#endif
313
314namespace {
315// Base class of for reordering networks. They don't strictly need to be
316// permutations, as outputs with repeated occurrences of an input element
317// are allowed.
318struct PermNetwork {
319  using Controls = std::vector<uint8_t>;
320  using ElemType = int;
321  static constexpr ElemType Ignore = ElemType(-1);
322
323  enum : uint8_t {
324    None,
325    Pass,
326    Switch
327  };
328  enum : uint8_t {
329    Forward,
330    Reverse
331  };
332
333  PermNetwork(ArrayRef<ElemType> Ord, unsigned Mult = 1) {
334    Order.assign(Ord.data(), Ord.data()+Ord.size());
335    Log = 0;
336
337    unsigned S = Order.size();
338    while (S >>= 1)
339      ++Log;
340
341    Table.resize(Order.size());
342    for (RowType &Row : Table)
343      Row.resize(Mult*Log, None);
344  }
345
346  void getControls(Controls &V, unsigned StartAt, uint8_t Dir) const {
347    unsigned Size = Order.size();
348    V.resize(Size);
349    for (unsigned I = 0; I != Size; ++I) {
350      unsigned W = 0;
351      for (unsigned L = 0; L != Log; ++L) {
352        unsigned C = ctl(I, StartAt+L) == Switch;
353        if (Dir == Forward)
354          W |= C << (Log-1-L);
355        else
356          W |= C << L;
357      }
358      assert(isUInt<8>(W));
359      V[I] = uint8_t(W);
360    }
361  }
362
363  uint8_t ctl(ElemType Pos, unsigned Step) const {
364    return Table[Pos][Step];
365  }
366  unsigned size() const {
367    return Order.size();
368  }
369  unsigned steps() const {
370    return Log;
371  }
372
373protected:
374  unsigned Log;
375  std::vector<ElemType> Order;
376  using RowType = std::vector<uint8_t>;
377  std::vector<RowType> Table;
378};
379
380struct ForwardDeltaNetwork : public PermNetwork {
381  ForwardDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
382
383  bool run(Controls &V) {
384    if (!route(Order.data(), Table.data(), size(), 0))
385      return false;
386    getControls(V, 0, Forward);
387    return true;
388  }
389
390private:
391  bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
392};
393
394struct ReverseDeltaNetwork : public PermNetwork {
395  ReverseDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
396
397  bool run(Controls &V) {
398    if (!route(Order.data(), Table.data(), size(), 0))
399      return false;
400    getControls(V, 0, Reverse);
401    return true;
402  }
403
404private:
405  bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
406};
407
408struct BenesNetwork : public PermNetwork {
409  BenesNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord, 2) {}
410
411  bool run(Controls &F, Controls &R) {
412    if (!route(Order.data(), Table.data(), size(), 0))
413      return false;
414
415    getControls(F, 0, Forward);
416    getControls(R, Log, Reverse);
417    return true;
418  }
419
420private:
421  bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
422};
423} // namespace
424
425bool ForwardDeltaNetwork::route(ElemType *P, RowType *T, unsigned Size,
426                                unsigned Step) {
427  bool UseUp = false, UseDown = false;
428  ElemType Num = Size;
429
430  // Cannot use coloring here, because coloring is used to determine
431  // the "big" switch, i.e. the one that changes halves, and in a forward
432  // network, a color can be simultaneously routed to both halves in the
433  // step we're working on.
434  for (ElemType J = 0; J != Num; ++J) {
435    ElemType I = P[J];
436    // I is the position in the input,
437    // J is the position in the output.
438    if (I == Ignore)
439      continue;
440    uint8_t S;
441    if (I < Num/2)
442      S = (J < Num/2) ? Pass : Switch;
443    else
444      S = (J < Num/2) ? Switch : Pass;
445
446    // U is the element in the table that needs to be updated.
447    ElemType U = (S == Pass) ? I : (I < Num/2 ? I+Num/2 : I-Num/2);
448    if (U < Num/2)
449      UseUp = true;
450    else
451      UseDown = true;
452    if (T[U][Step] != S && T[U][Step] != None)
453      return false;
454    T[U][Step] = S;
455  }
456
457  for (ElemType J = 0; J != Num; ++J)
458    if (P[J] != Ignore && P[J] >= Num/2)
459      P[J] -= Num/2;
460
461  if (Step+1 < Log) {
462    if (UseUp   && !route(P,        T,        Size/2, Step+1))
463      return false;
464    if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
465      return false;
466  }
467  return true;
468}
469
470bool ReverseDeltaNetwork::route(ElemType *P, RowType *T, unsigned Size,
471                                unsigned Step) {
472  unsigned Pets = Log-1 - Step;
473  bool UseUp = false, UseDown = false;
474  ElemType Num = Size;
475
476  // In this step half-switching occurs, so coloring can be used.
477  Coloring G({P,Size});
478  const Coloring::MapType &M = G.colors();
479  if (M.empty())
480    return false;
481
482  ColorKind ColorUp = ColorKind::None;
483  for (ElemType J = 0; J != Num; ++J) {
484    ElemType I = P[J];
485    // I is the position in the input,
486    // J is the position in the output.
487    if (I == Ignore)
488      continue;
489    ColorKind C = M.at(I);
490    if (C == ColorKind::None)
491      continue;
492    // During "Step", inputs cannot switch halves, so if the "up" color
493    // is still unknown, make sure that it is selected in such a way that
494    // "I" will stay in the same half.
495    bool InpUp = I < Num/2;
496    if (ColorUp == ColorKind::None)
497      ColorUp = InpUp ? C : G.other(C);
498    if ((C == ColorUp) != InpUp) {
499      // If I should go to a different half than where is it now, give up.
500      return false;
501    }
502
503    uint8_t S;
504    if (InpUp) {
505      S = (J < Num/2) ? Pass : Switch;
506      UseUp = true;
507    } else {
508      S = (J < Num/2) ? Switch : Pass;
509      UseDown = true;
510    }
511    T[J][Pets] = S;
512  }
513
514  // Reorder the working permutation according to the computed switch table
515  // for the last step (i.e. Pets).
516  for (ElemType J = 0, E = Size / 2; J != E; ++J) {
517    ElemType PJ = P[J];         // Current values of P[J]
518    ElemType PC = P[J+Size/2];  // and P[conj(J)]
519    ElemType QJ = PJ;           // New values of P[J]
520    ElemType QC = PC;           // and P[conj(J)]
521    if (T[J][Pets] == Switch)
522      QC = PJ;
523    if (T[J+Size/2][Pets] == Switch)
524      QJ = PC;
525    P[J] = QJ;
526    P[J+Size/2] = QC;
527  }
528
529  for (ElemType J = 0; J != Num; ++J)
530    if (P[J] != Ignore && P[J] >= Num/2)
531      P[J] -= Num/2;
532
533  if (Step+1 < Log) {
534    if (UseUp && !route(P, T, Size/2, Step+1))
535      return false;
536    if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
537      return false;
538  }
539  return true;
540}
541
542bool BenesNetwork::route(ElemType *P, RowType *T, unsigned Size,
543                         unsigned Step) {
544  Coloring G({P,Size});
545  const Coloring::MapType &M = G.colors();
546  if (M.empty())
547    return false;
548  ElemType Num = Size;
549
550  unsigned Pets = 2*Log-1 - Step;
551  bool UseUp = false, UseDown = false;
552
553  // Both assignments, i.e. Red->Up and Red->Down are valid, but they will
554  // result in different controls. Let's pick the one where the first
555  // control will be "Pass".
556  ColorKind ColorUp = ColorKind::None;
557  for (ElemType J = 0; J != Num; ++J) {
558    ElemType I = P[J];
559    if (I == Ignore)
560      continue;
561    ColorKind C = M.at(I);
562    if (C == ColorKind::None)
563      continue;
564    if (ColorUp == ColorKind::None) {
565      ColorUp = (I < Num / 2) ? ColorKind::Red : ColorKind::Black;
566    }
567    unsigned CI = (I < Num/2) ? I+Num/2 : I-Num/2;
568    if (C == ColorUp) {
569      if (I < Num/2)
570        T[I][Step] = Pass;
571      else
572        T[CI][Step] = Switch;
573      T[J][Pets] = (J < Num/2) ? Pass : Switch;
574      UseUp = true;
575    } else { // Down
576      if (I < Num/2)
577        T[CI][Step] = Switch;
578      else
579        T[I][Step] = Pass;
580      T[J][Pets] = (J < Num/2) ? Switch : Pass;
581      UseDown = true;
582    }
583  }
584
585  // Reorder the working permutation according to the computed switch table
586  // for the last step (i.e. Pets).
587  for (ElemType J = 0; J != Num/2; ++J) {
588    ElemType PJ = P[J];         // Current values of P[J]
589    ElemType PC = P[J+Num/2];   // and P[conj(J)]
590    ElemType QJ = PJ;           // New values of P[J]
591    ElemType QC = PC;           // and P[conj(J)]
592    if (T[J][Pets] == Switch)
593      QC = PJ;
594    if (T[J+Num/2][Pets] == Switch)
595      QJ = PC;
596    P[J] = QJ;
597    P[J+Num/2] = QC;
598  }
599
600  for (ElemType J = 0; J != Num; ++J)
601    if (P[J] != Ignore && P[J] >= Num/2)
602      P[J] -= Num/2;
603
604  if (Step+1 < Log) {
605    if (UseUp && !route(P, T, Size/2, Step+1))
606      return false;
607    if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
608      return false;
609  }
610  return true;
611}
612
613// --------------------------------------------------------------------
614// Support for building selection results (output instructions that are
615// parts of the final selection).
616
617namespace {
618struct OpRef {
619  OpRef(SDValue V) : OpV(V) {}
620  bool isValue() const { return OpV.getNode() != nullptr; }
621  bool isValid() const { return isValue() || !(OpN & Invalid); }
622  static OpRef res(int N) { return OpRef(Whole | (N & Index)); }
623  static OpRef fail() { return OpRef(Invalid); }
624
625  static OpRef lo(const OpRef &R) {
626    assert(!R.isValue());
627    return OpRef(R.OpN & (Undef | Index | LoHalf));
628  }
629  static OpRef hi(const OpRef &R) {
630    assert(!R.isValue());
631    return OpRef(R.OpN & (Undef | Index | HiHalf));
632  }
633  static OpRef undef(MVT Ty) { return OpRef(Undef | Ty.SimpleTy); }
634
635  // Direct value.
636  SDValue OpV = SDValue();
637
638  // Reference to the operand of the input node:
639  // If the 31st bit is 1, it's undef, otherwise, bits 28..0 are the
640  // operand index:
641  // If bit 30 is set, it's the high half of the operand.
642  // If bit 29 is set, it's the low half of the operand.
643  unsigned OpN = 0;
644
645  enum : unsigned {
646    Invalid = 0x10000000,
647    LoHalf  = 0x20000000,
648    HiHalf  = 0x40000000,
649    Whole   = LoHalf | HiHalf,
650    Undef   = 0x80000000,
651    Index   = 0x0FFFFFFF,  // Mask of the index value.
652    IndexBits = 28,
653  };
654
655  LLVM_DUMP_METHOD
656  void print(raw_ostream &OS, const SelectionDAG &G) const;
657
658private:
659  OpRef(unsigned N) : OpN(N) {}
660};
661
662struct NodeTemplate {
663  NodeTemplate() = default;
664  unsigned Opc = 0;
665  MVT Ty = MVT::Other;
666  std::vector<OpRef> Ops;
667
668  LLVM_DUMP_METHOD void print(raw_ostream &OS, const SelectionDAG &G) const;
669};
670
671struct ResultStack {
672  ResultStack(SDNode *Inp)
673    : InpNode(Inp), InpTy(Inp->getValueType(0).getSimpleVT()) {}
674  SDNode *InpNode;
675  MVT InpTy;
676  unsigned push(const NodeTemplate &Res) {
677    List.push_back(Res);
678    return List.size()-1;
679  }
680  unsigned push(unsigned Opc, MVT Ty, std::vector<OpRef> &&Ops) {
681    NodeTemplate Res;
682    Res.Opc = Opc;
683    Res.Ty = Ty;
684    Res.Ops = Ops;
685    return push(Res);
686  }
687  bool empty() const { return List.empty(); }
688  unsigned size() const { return List.size(); }
689  unsigned top() const { return size()-1; }
690  const NodeTemplate &operator[](unsigned I) const { return List[I]; }
691  unsigned reset(unsigned NewTop) {
692    List.resize(NewTop+1);
693    return NewTop;
694  }
695
696  using BaseType = std::vector<NodeTemplate>;
697  BaseType::iterator begin() { return List.begin(); }
698  BaseType::iterator end()   { return List.end(); }
699  BaseType::const_iterator begin() const { return List.begin(); }
700  BaseType::const_iterator end() const   { return List.end(); }
701
702  BaseType List;
703
704  LLVM_DUMP_METHOD
705  void print(raw_ostream &OS, const SelectionDAG &G) const;
706};
707} // namespace
708
709#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
710void OpRef::print(raw_ostream &OS, const SelectionDAG &G) const {
711  if (isValue()) {
712    OpV.getNode()->print(OS, &G);
713    return;
714  }
715  if (OpN & Invalid) {
716    OS << "invalid";
717    return;
718  }
719  if (OpN & Undef) {
720    OS << "undef";
721    return;
722  }
723  if ((OpN & Whole) != Whole) {
724    assert((OpN & Whole) == LoHalf || (OpN & Whole) == HiHalf);
725    if (OpN & LoHalf)
726      OS << "lo ";
727    else
728      OS << "hi ";
729  }
730  OS << '#' << SignExtend32(OpN & Index, IndexBits);
731}
732
733void NodeTemplate::print(raw_ostream &OS, const SelectionDAG &G) const {
734  const TargetInstrInfo &TII = *G.getSubtarget().getInstrInfo();
735  OS << format("%8s", EVT(Ty).getEVTString().c_str()) << "  "
736     << TII.getName(Opc);
737  bool Comma = false;
738  for (const auto &R : Ops) {
739    if (Comma)
740      OS << ',';
741    Comma = true;
742    OS << ' ';
743    R.print(OS, G);
744  }
745}
746
747void ResultStack::print(raw_ostream &OS, const SelectionDAG &G) const {
748  OS << "Input node:\n";
749#ifndef NDEBUG
750  InpNode->dumpr(&G);
751#endif
752  OS << "Result templates:\n";
753  for (unsigned I = 0, E = List.size(); I != E; ++I) {
754    OS << '[' << I << "] ";
755    List[I].print(OS, G);
756    OS << '\n';
757  }
758}
759#endif
760
761namespace {
762struct ShuffleMask {
763  ShuffleMask(ArrayRef<int> M) : Mask(M) {
764    for (unsigned I = 0, E = Mask.size(); I != E; ++I) {
765      int M = Mask[I];
766      if (M == -1)
767        continue;
768      MinSrc = (MinSrc == -1) ? M : std::min(MinSrc, M);
769      MaxSrc = (MaxSrc == -1) ? M : std::max(MaxSrc, M);
770    }
771  }
772
773  ArrayRef<int> Mask;
774  int MinSrc = -1, MaxSrc = -1;
775
776  ShuffleMask lo() const {
777    size_t H = Mask.size()/2;
778    return ShuffleMask(Mask.take_front(H));
779  }
780  ShuffleMask hi() const {
781    size_t H = Mask.size()/2;
782    return ShuffleMask(Mask.take_back(H));
783  }
784
785  void print(raw_ostream &OS) const {
786    OS << "MinSrc:" << MinSrc << ", MaxSrc:" << MaxSrc << " {";
787    for (int M : Mask)
788      OS << ' ' << M;
789    OS << " }";
790  }
791};
792} // namespace
793
794// --------------------------------------------------------------------
795// The HvxSelector class.
796
797static const HexagonTargetLowering &getHexagonLowering(SelectionDAG &G) {
798  return static_cast<const HexagonTargetLowering&>(G.getTargetLoweringInfo());
799}
800static const HexagonSubtarget &getHexagonSubtarget(SelectionDAG &G) {
801  return static_cast<const HexagonSubtarget&>(G.getSubtarget());
802}
803
804namespace llvm {
805  struct HvxSelector {
806    const HexagonTargetLowering &Lower;
807    HexagonDAGToDAGISel &ISel;
808    SelectionDAG &DAG;
809    const HexagonSubtarget &HST;
810    const unsigned HwLen;
811
812    HvxSelector(HexagonDAGToDAGISel &HS, SelectionDAG &G)
813      : Lower(getHexagonLowering(G)),  ISel(HS), DAG(G),
814        HST(getHexagonSubtarget(G)), HwLen(HST.getVectorLength()) {}
815
816    MVT getSingleVT(MVT ElemTy) const {
817      unsigned NumElems = HwLen / (ElemTy.getSizeInBits()/8);
818      return MVT::getVectorVT(ElemTy, NumElems);
819    }
820
821    MVT getPairVT(MVT ElemTy) const {
822      unsigned NumElems = (2*HwLen) / (ElemTy.getSizeInBits()/8);
823      return MVT::getVectorVT(ElemTy, NumElems);
824    }
825
826    void selectShuffle(SDNode *N);
827    void selectRor(SDNode *N);
828    void selectVAlign(SDNode *N);
829
830  private:
831    void materialize(const ResultStack &Results);
832
833    SDValue getVectorConstant(ArrayRef<uint8_t> Data, const SDLoc &dl);
834
835    enum : unsigned {
836      None,
837      PackMux,
838    };
839    OpRef concat(OpRef Va, OpRef Vb, ResultStack &Results);
840    OpRef packs(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results,
841                MutableArrayRef<int> NewMask, unsigned Options = None);
842    OpRef packp(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results,
843                MutableArrayRef<int> NewMask);
844    OpRef vmuxs(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
845                ResultStack &Results);
846    OpRef vmuxp(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
847                ResultStack &Results);
848
849    OpRef shuffs1(ShuffleMask SM, OpRef Va, ResultStack &Results);
850    OpRef shuffs2(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
851    OpRef shuffp1(ShuffleMask SM, OpRef Va, ResultStack &Results);
852    OpRef shuffp2(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
853
854    OpRef butterfly(ShuffleMask SM, OpRef Va, ResultStack &Results);
855    OpRef contracting(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
856    OpRef expanding(ShuffleMask SM, OpRef Va, ResultStack &Results);
857    OpRef perfect(ShuffleMask SM, OpRef Va, ResultStack &Results);
858
859    bool selectVectorConstants(SDNode *N);
860    bool scalarizeShuffle(ArrayRef<int> Mask, const SDLoc &dl, MVT ResTy,
861                          SDValue Va, SDValue Vb, SDNode *N);
862
863  };
864}
865
866static void splitMask(ArrayRef<int> Mask, MutableArrayRef<int> MaskL,
867                      MutableArrayRef<int> MaskR) {
868  unsigned VecLen = Mask.size();
869  assert(MaskL.size() == VecLen && MaskR.size() == VecLen);
870  for (unsigned I = 0; I != VecLen; ++I) {
871    int M = Mask[I];
872    if (M < 0) {
873      MaskL[I] = MaskR[I] = -1;
874    } else if (unsigned(M) < VecLen) {
875      MaskL[I] = M;
876      MaskR[I] = -1;
877    } else {
878      MaskL[I] = -1;
879      MaskR[I] = M-VecLen;
880    }
881  }
882}
883
884static std::pair<int,unsigned> findStrip(ArrayRef<int> A, int Inc,
885                                         unsigned MaxLen) {
886  assert(A.size() > 0 && A.size() >= MaxLen);
887  int F = A[0];
888  int E = F;
889  for (unsigned I = 1; I != MaxLen; ++I) {
890    if (A[I] - E != Inc)
891      return { F, I };
892    E = A[I];
893  }
894  return { F, MaxLen };
895}
896
897static bool isUndef(ArrayRef<int> Mask) {
898  for (int Idx : Mask)
899    if (Idx != -1)
900      return false;
901  return true;
902}
903
904static bool isIdentity(ArrayRef<int> Mask) {
905  for (int I = 0, E = Mask.size(); I != E; ++I) {
906    int M = Mask[I];
907    if (M >= 0 && M != I)
908      return false;
909  }
910  return true;
911}
912
913static bool isPermutation(ArrayRef<int> Mask) {
914  // Check by adding all numbers only works if there is no overflow.
915  assert(Mask.size() < 0x00007FFF && "Sanity failure");
916  int Sum = 0;
917  for (int Idx : Mask) {
918    if (Idx == -1)
919      return false;
920    Sum += Idx;
921  }
922  int N = Mask.size();
923  return 2*Sum == N*(N-1);
924}
925
926bool HvxSelector::selectVectorConstants(SDNode *N) {
927  // Constant vectors are generated as loads from constant pools or as
928  // splats of a constant value. Since they are generated during the
929  // selection process, the main selection algorithm is not aware of them.
930  // Select them directly here.
931  SmallVector<SDNode*,4> Nodes;
932  SetVector<SDNode*> WorkQ;
933
934  // The one-use test for VSPLATW's operand may fail due to dead nodes
935  // left over in the DAG.
936  DAG.RemoveDeadNodes();
937
938  // The DAG can change (due to CSE) during selection, so cache all the
939  // unselected nodes first to avoid traversing a mutating DAG.
940
941  auto IsNodeToSelect = [] (SDNode *N) {
942    if (N->isMachineOpcode())
943      return false;
944    switch (N->getOpcode()) {
945      case HexagonISD::VZERO:
946      case HexagonISD::VSPLATW:
947        return true;
948      case ISD::LOAD: {
949        SDValue Addr = cast<LoadSDNode>(N)->getBasePtr();
950        unsigned AddrOpc = Addr.getOpcode();
951        if (AddrOpc == HexagonISD::AT_PCREL || AddrOpc == HexagonISD::CP)
952          if (Addr.getOperand(0).getOpcode() == ISD::TargetConstantPool)
953            return true;
954      }
955      break;
956    }
957    // Make sure to select the operand of VSPLATW.
958    bool IsSplatOp = N->hasOneUse() &&
959                     N->use_begin()->getOpcode() == HexagonISD::VSPLATW;
960    return IsSplatOp;
961  };
962
963  WorkQ.insert(N);
964  for (unsigned i = 0; i != WorkQ.size(); ++i) {
965    SDNode *W = WorkQ[i];
966    if (IsNodeToSelect(W))
967      Nodes.push_back(W);
968    for (unsigned j = 0, f = W->getNumOperands(); j != f; ++j)
969      WorkQ.insert(W->getOperand(j).getNode());
970  }
971
972  for (SDNode *L : Nodes)
973    ISel.Select(L);
974
975  return !Nodes.empty();
976}
977
978void HvxSelector::materialize(const ResultStack &Results) {
979  DEBUG_WITH_TYPE("isel", {
980    dbgs() << "Materializing\n";
981    Results.print(dbgs(), DAG);
982  });
983  if (Results.empty())
984    return;
985  const SDLoc &dl(Results.InpNode);
986  std::vector<SDValue> Output;
987
988  for (unsigned I = 0, E = Results.size(); I != E; ++I) {
989    const NodeTemplate &Node = Results[I];
990    std::vector<SDValue> Ops;
991    for (const OpRef &R : Node.Ops) {
992      assert(R.isValid());
993      if (R.isValue()) {
994        Ops.push_back(R.OpV);
995        continue;
996      }
997      if (R.OpN & OpRef::Undef) {
998        MVT::SimpleValueType SVT = MVT::SimpleValueType(R.OpN & OpRef::Index);
999        Ops.push_back(ISel.selectUndef(dl, MVT(SVT)));
1000        continue;
1001      }
1002      // R is an index of a result.
1003      unsigned Part = R.OpN & OpRef::Whole;
1004      int Idx = SignExtend32(R.OpN & OpRef::Index, OpRef::IndexBits);
1005      if (Idx < 0)
1006        Idx += I;
1007      assert(Idx >= 0 && unsigned(Idx) < Output.size());
1008      SDValue Op = Output[Idx];
1009      MVT OpTy = Op.getValueType().getSimpleVT();
1010      if (Part != OpRef::Whole) {
1011        assert(Part == OpRef::LoHalf || Part == OpRef::HiHalf);
1012        MVT HalfTy = MVT::getVectorVT(OpTy.getVectorElementType(),
1013                                      OpTy.getVectorNumElements()/2);
1014        unsigned Sub = (Part == OpRef::LoHalf) ? Hexagon::vsub_lo
1015                                               : Hexagon::vsub_hi;
1016        Op = DAG.getTargetExtractSubreg(Sub, dl, HalfTy, Op);
1017      }
1018      Ops.push_back(Op);
1019    } // for (Node : Results)
1020
1021    assert(Node.Ty != MVT::Other);
1022    SDNode *ResN = (Node.Opc == TargetOpcode::COPY)
1023                      ? Ops.front().getNode()
1024                      : DAG.getMachineNode(Node.Opc, dl, Node.Ty, Ops);
1025    Output.push_back(SDValue(ResN, 0));
1026  }
1027
1028  SDNode *OutN = Output.back().getNode();
1029  SDNode *InpN = Results.InpNode;
1030  DEBUG_WITH_TYPE("isel", {
1031    dbgs() << "Generated node:\n";
1032    OutN->dumpr(&DAG);
1033  });
1034
1035  ISel.ReplaceNode(InpN, OutN);
1036  selectVectorConstants(OutN);
1037  DAG.RemoveDeadNodes();
1038}
1039
1040OpRef HvxSelector::concat(OpRef Lo, OpRef Hi, ResultStack &Results) {
1041  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1042  const SDLoc &dl(Results.InpNode);
1043  Results.push(TargetOpcode::REG_SEQUENCE, getPairVT(MVT::i8), {
1044    DAG.getTargetConstant(Hexagon::HvxWRRegClassID, dl, MVT::i32),
1045    Lo, DAG.getTargetConstant(Hexagon::vsub_lo, dl, MVT::i32),
1046    Hi, DAG.getTargetConstant(Hexagon::vsub_hi, dl, MVT::i32),
1047  });
1048  return OpRef::res(Results.top());
1049}
1050
1051// Va, Vb are single vectors, SM can be arbitrarily long.
1052OpRef HvxSelector::packs(ShuffleMask SM, OpRef Va, OpRef Vb,
1053                         ResultStack &Results, MutableArrayRef<int> NewMask,
1054                         unsigned Options) {
1055  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1056  if (!Va.isValid() || !Vb.isValid())
1057    return OpRef::fail();
1058
1059  int VecLen = SM.Mask.size();
1060  MVT Ty = getSingleVT(MVT::i8);
1061
1062  auto IsExtSubvector = [] (ShuffleMask M) {
1063    assert(M.MinSrc >= 0 && M.MaxSrc >= 0);
1064    for (int I = 0, E = M.Mask.size(); I != E; ++I) {
1065      if (M.Mask[I] >= 0 && M.Mask[I]-I != M.MinSrc)
1066        return false;
1067    }
1068    return true;
1069  };
1070
1071  if (SM.MaxSrc - SM.MinSrc < int(HwLen)) {
1072    if (SM.MinSrc == 0 || SM.MinSrc == int(HwLen) || !IsExtSubvector(SM)) {
1073      // If the mask picks elements from only one of the operands, return
1074      // that operand, and update the mask to use index 0 to refer to the
1075      // first element of that operand.
1076      // If the mask extracts a subvector, it will be handled below, so
1077      // skip it here.
1078      if (SM.MaxSrc < int(HwLen)) {
1079        memcpy(NewMask.data(), SM.Mask.data(), sizeof(int)*VecLen);
1080        return Va;
1081      }
1082      if (SM.MinSrc >= int(HwLen)) {
1083        for (int I = 0; I != VecLen; ++I) {
1084          int M = SM.Mask[I];
1085          if (M != -1)
1086            M -= HwLen;
1087          NewMask[I] = M;
1088        }
1089        return Vb;
1090      }
1091    }
1092    int MinSrc = SM.MinSrc;
1093    if (SM.MaxSrc < int(HwLen)) {
1094      Vb = Va;
1095    } else if (SM.MinSrc > int(HwLen)) {
1096      Va = Vb;
1097      MinSrc = SM.MinSrc - HwLen;
1098    }
1099    const SDLoc &dl(Results.InpNode);
1100    if (isUInt<3>(MinSrc) || isUInt<3>(HwLen-MinSrc)) {
1101      bool IsRight = isUInt<3>(MinSrc); // Right align.
1102      SDValue S = DAG.getTargetConstant(IsRight ? MinSrc : HwLen-MinSrc,
1103                                        dl, MVT::i32);
1104      unsigned Opc = IsRight ? Hexagon::V6_valignbi
1105                             : Hexagon::V6_vlalignbi;
1106      Results.push(Opc, Ty, {Vb, Va, S});
1107    } else {
1108      SDValue S = DAG.getTargetConstant(MinSrc, dl, MVT::i32);
1109      Results.push(Hexagon::A2_tfrsi, MVT::i32, {S});
1110      unsigned Top = Results.top();
1111      Results.push(Hexagon::V6_valignb, Ty, {Vb, Va, OpRef::res(Top)});
1112    }
1113    for (int I = 0; I != VecLen; ++I) {
1114      int M = SM.Mask[I];
1115      if (M != -1)
1116        M -= SM.MinSrc;
1117      NewMask[I] = M;
1118    }
1119    return OpRef::res(Results.top());
1120  }
1121
1122  if (Options & PackMux) {
1123    // If elements picked from Va and Vb have all different (source) indexes
1124    // (relative to the start of the argument), do a mux, and update the mask.
1125    BitVector Picked(HwLen);
1126    SmallVector<uint8_t,128> MuxBytes(HwLen);
1127    bool CanMux = true;
1128    for (int I = 0; I != VecLen; ++I) {
1129      int M = SM.Mask[I];
1130      if (M == -1)
1131        continue;
1132      if (M >= int(HwLen))
1133        M -= HwLen;
1134      else
1135        MuxBytes[M] = 0xFF;
1136      if (Picked[M]) {
1137        CanMux = false;
1138        break;
1139      }
1140      NewMask[I] = M;
1141    }
1142    if (CanMux)
1143      return vmuxs(MuxBytes, Va, Vb, Results);
1144  }
1145
1146  return OpRef::fail();
1147}
1148
1149OpRef HvxSelector::packp(ShuffleMask SM, OpRef Va, OpRef Vb,
1150                         ResultStack &Results, MutableArrayRef<int> NewMask) {
1151  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1152  unsigned HalfMask = 0;
1153  unsigned LogHw = Log2_32(HwLen);
1154  for (int M : SM.Mask) {
1155    if (M == -1)
1156      continue;
1157    HalfMask |= (1u << (M >> LogHw));
1158  }
1159
1160  if (HalfMask == 0)
1161    return OpRef::undef(getPairVT(MVT::i8));
1162
1163  // If more than two halves are used, bail.
1164  // TODO: be more aggressive here?
1165  if (countPopulation(HalfMask) > 2)
1166    return OpRef::fail();
1167
1168  MVT HalfTy = getSingleVT(MVT::i8);
1169
1170  OpRef Inp[2] = { Va, Vb };
1171  OpRef Out[2] = { OpRef::undef(HalfTy), OpRef::undef(HalfTy) };
1172
1173  uint8_t HalfIdx[4] = { 0xFF, 0xFF, 0xFF, 0xFF };
1174  unsigned Idx = 0;
1175  for (unsigned I = 0; I != 4; ++I) {
1176    if ((HalfMask & (1u << I)) == 0)
1177      continue;
1178    assert(Idx < 2);
1179    OpRef Op = Inp[I/2];
1180    Out[Idx] = (I & 1) ? OpRef::hi(Op) : OpRef::lo(Op);
1181    HalfIdx[I] = Idx++;
1182  }
1183
1184  int VecLen = SM.Mask.size();
1185  for (int I = 0; I != VecLen; ++I) {
1186    int M = SM.Mask[I];
1187    if (M >= 0) {
1188      uint8_t Idx = HalfIdx[M >> LogHw];
1189      assert(Idx == 0 || Idx == 1);
1190      M = (M & (HwLen-1)) + HwLen*Idx;
1191    }
1192    NewMask[I] = M;
1193  }
1194
1195  return concat(Out[0], Out[1], Results);
1196}
1197
1198OpRef HvxSelector::vmuxs(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1199                         ResultStack &Results) {
1200  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1201  MVT ByteTy = getSingleVT(MVT::i8);
1202  MVT BoolTy = MVT::getVectorVT(MVT::i1, 8*HwLen); // XXX
1203  const SDLoc &dl(Results.InpNode);
1204  SDValue B = getVectorConstant(Bytes, dl);
1205  Results.push(Hexagon::V6_vd0, ByteTy, {});
1206  Results.push(Hexagon::V6_veqb, BoolTy, {OpRef(B), OpRef::res(-1)});
1207  Results.push(Hexagon::V6_vmux, ByteTy, {OpRef::res(-1), Vb, Va});
1208  return OpRef::res(Results.top());
1209}
1210
1211OpRef HvxSelector::vmuxp(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1212                         ResultStack &Results) {
1213  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1214  size_t S = Bytes.size() / 2;
1215  OpRef L = vmuxs(Bytes.take_front(S), OpRef::lo(Va), OpRef::lo(Vb), Results);
1216  OpRef H = vmuxs(Bytes.drop_front(S), OpRef::hi(Va), OpRef::hi(Vb), Results);
1217  return concat(L, H, Results);
1218}
1219
1220OpRef HvxSelector::shuffs1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1221  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1222  unsigned VecLen = SM.Mask.size();
1223  assert(HwLen == VecLen);
1224  (void)VecLen;
1225  assert(all_of(SM.Mask, [this](int M) { return M == -1 || M < int(HwLen); }));
1226
1227  if (isIdentity(SM.Mask))
1228    return Va;
1229  if (isUndef(SM.Mask))
1230    return OpRef::undef(getSingleVT(MVT::i8));
1231
1232  OpRef P = perfect(SM, Va, Results);
1233  if (P.isValid())
1234    return P;
1235  return butterfly(SM, Va, Results);
1236}
1237
1238OpRef HvxSelector::shuffs2(ShuffleMask SM, OpRef Va, OpRef Vb,
1239                           ResultStack &Results) {
1240  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1241  if (isUndef(SM.Mask))
1242    return OpRef::undef(getSingleVT(MVT::i8));
1243
1244  OpRef C = contracting(SM, Va, Vb, Results);
1245  if (C.isValid())
1246    return C;
1247
1248  int VecLen = SM.Mask.size();
1249  SmallVector<int,128> NewMask(VecLen);
1250  OpRef P = packs(SM, Va, Vb, Results, NewMask);
1251  if (P.isValid())
1252    return shuffs1(ShuffleMask(NewMask), P, Results);
1253
1254  SmallVector<int,128> MaskL(VecLen), MaskR(VecLen);
1255  splitMask(SM.Mask, MaskL, MaskR);
1256
1257  OpRef L = shuffs1(ShuffleMask(MaskL), Va, Results);
1258  OpRef R = shuffs1(ShuffleMask(MaskR), Vb, Results);
1259  if (!L.isValid() || !R.isValid())
1260    return OpRef::fail();
1261
1262  SmallVector<uint8_t,128> Bytes(VecLen);
1263  for (int I = 0; I != VecLen; ++I) {
1264    if (MaskL[I] != -1)
1265      Bytes[I] = 0xFF;
1266  }
1267  return vmuxs(Bytes, L, R, Results);
1268}
1269
1270OpRef HvxSelector::shuffp1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1271  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1272  int VecLen = SM.Mask.size();
1273
1274  if (isIdentity(SM.Mask))
1275    return Va;
1276  if (isUndef(SM.Mask))
1277    return OpRef::undef(getPairVT(MVT::i8));
1278
1279  SmallVector<int,128> PackedMask(VecLen);
1280  OpRef P = packs(SM, OpRef::lo(Va), OpRef::hi(Va), Results, PackedMask);
1281  if (P.isValid()) {
1282    ShuffleMask PM(PackedMask);
1283    OpRef E = expanding(PM, P, Results);
1284    if (E.isValid())
1285      return E;
1286
1287    OpRef L = shuffs1(PM.lo(), P, Results);
1288    OpRef H = shuffs1(PM.hi(), P, Results);
1289    if (L.isValid() && H.isValid())
1290      return concat(L, H, Results);
1291  }
1292
1293  OpRef R = perfect(SM, Va, Results);
1294  if (R.isValid())
1295    return R;
1296  // TODO commute the mask and try the opposite order of the halves.
1297
1298  OpRef L = shuffs2(SM.lo(), OpRef::lo(Va), OpRef::hi(Va), Results);
1299  OpRef H = shuffs2(SM.hi(), OpRef::lo(Va), OpRef::hi(Va), Results);
1300  if (L.isValid() && H.isValid())
1301    return concat(L, H, Results);
1302
1303  return OpRef::fail();
1304}
1305
1306OpRef HvxSelector::shuffp2(ShuffleMask SM, OpRef Va, OpRef Vb,
1307                           ResultStack &Results) {
1308  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1309  if (isUndef(SM.Mask))
1310    return OpRef::undef(getPairVT(MVT::i8));
1311
1312  int VecLen = SM.Mask.size();
1313  SmallVector<int,256> PackedMask(VecLen);
1314  OpRef P = packp(SM, Va, Vb, Results, PackedMask);
1315  if (P.isValid())
1316    return shuffp1(ShuffleMask(PackedMask), P, Results);
1317
1318  SmallVector<int,256> MaskL(VecLen), MaskR(VecLen);
1319  splitMask(SM.Mask, MaskL, MaskR);
1320
1321  OpRef L = shuffp1(ShuffleMask(MaskL), Va, Results);
1322  OpRef R = shuffp1(ShuffleMask(MaskR), Vb, Results);
1323  if (!L.isValid() || !R.isValid())
1324    return OpRef::fail();
1325
1326  // Mux the results.
1327  SmallVector<uint8_t,256> Bytes(VecLen);
1328  for (int I = 0; I != VecLen; ++I) {
1329    if (MaskL[I] != -1)
1330      Bytes[I] = 0xFF;
1331  }
1332  return vmuxp(Bytes, L, R, Results);
1333}
1334
1335namespace {
1336  struct Deleter : public SelectionDAG::DAGNodeDeletedListener {
1337    template <typename T>
1338    Deleter(SelectionDAG &D, T &C)
1339      : SelectionDAG::DAGNodeDeletedListener(D, [&C] (SDNode *N, SDNode *E) {
1340                                                  C.erase(N);
1341                                                }) {}
1342  };
1343
1344  template <typename T>
1345  struct NullifyingVector : public T {
1346    DenseMap<SDNode*, SDNode**> Refs;
1347    NullifyingVector(T &&V) : T(V) {
1348      for (unsigned i = 0, e = T::size(); i != e; ++i) {
1349        SDNode *&N = T::operator[](i);
1350        Refs[N] = &N;
1351      }
1352    }
1353    void erase(SDNode *N) {
1354      auto F = Refs.find(N);
1355      if (F != Refs.end())
1356        *F->second = nullptr;
1357    }
1358  };
1359}
1360
1361bool HvxSelector::scalarizeShuffle(ArrayRef<int> Mask, const SDLoc &dl,
1362                                   MVT ResTy, SDValue Va, SDValue Vb,
1363                                   SDNode *N) {
1364  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1365  MVT ElemTy = ResTy.getVectorElementType();
1366  assert(ElemTy == MVT::i8);
1367  unsigned VecLen = Mask.size();
1368  bool HavePairs = (2*HwLen == VecLen);
1369  MVT SingleTy = getSingleVT(MVT::i8);
1370
1371  // The prior attempts to handle this shuffle may have left a bunch of
1372  // dead nodes in the DAG (such as constants). These nodes will be added
1373  // at the end of DAG's node list, which at that point had already been
1374  // sorted topologically. In the main selection loop, the node list is
1375  // traversed backwards from the root node, which means that any new
1376  // nodes (from the end of the list) will not be visited.
1377  // Scalarization will replace the shuffle node with the scalarized
1378  // expression, and if that expression reused any if the leftoever (dead)
1379  // nodes, these nodes would not be selected (since the "local" selection
1380  // only visits nodes that are not in AllNodes).
1381  // To avoid this issue, remove all dead nodes from the DAG now.
1382  DAG.RemoveDeadNodes();
1383  DenseSet<SDNode*> AllNodes;
1384  for (SDNode &S : DAG.allnodes())
1385    AllNodes.insert(&S);
1386
1387  Deleter DUA(DAG, AllNodes);
1388
1389  SmallVector<SDValue,128> Ops;
1390  LLVMContext &Ctx = *DAG.getContext();
1391  MVT LegalTy = Lower.getTypeToTransformTo(Ctx, ElemTy).getSimpleVT();
1392  for (int I : Mask) {
1393    if (I < 0) {
1394      Ops.push_back(ISel.selectUndef(dl, LegalTy));
1395      continue;
1396    }
1397    SDValue Vec;
1398    unsigned M = I;
1399    if (M < VecLen) {
1400      Vec = Va;
1401    } else {
1402      Vec = Vb;
1403      M -= VecLen;
1404    }
1405    if (HavePairs) {
1406      if (M < HwLen) {
1407        Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, SingleTy, Vec);
1408      } else {
1409        Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_hi, dl, SingleTy, Vec);
1410        M -= HwLen;
1411      }
1412    }
1413    SDValue Idx = DAG.getConstant(M, dl, MVT::i32);
1414    SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, LegalTy, {Vec, Idx});
1415    SDValue L = Lower.LowerOperation(Ex, DAG);
1416    assert(L.getNode());
1417    Ops.push_back(L);
1418  }
1419
1420  SDValue LV;
1421  if (2*HwLen == VecLen) {
1422    SDValue B0 = DAG.getBuildVector(SingleTy, dl, {Ops.data(), HwLen});
1423    SDValue L0 = Lower.LowerOperation(B0, DAG);
1424    SDValue B1 = DAG.getBuildVector(SingleTy, dl, {Ops.data()+HwLen, HwLen});
1425    SDValue L1 = Lower.LowerOperation(B1, DAG);
1426    // XXX CONCAT_VECTORS is legal for HVX vectors. Legalizing (lowering)
1427    // functions may expect to be called only for illegal operations, so
1428    // make sure that they are not called for legal ones. Develop a better
1429    // mechanism for dealing with this.
1430    LV = DAG.getNode(ISD::CONCAT_VECTORS, dl, ResTy, {L0, L1});
1431  } else {
1432    SDValue BV = DAG.getBuildVector(ResTy, dl, Ops);
1433    LV = Lower.LowerOperation(BV, DAG);
1434  }
1435
1436  assert(!N->use_empty());
1437  ISel.ReplaceNode(N, LV.getNode());
1438
1439  if (AllNodes.count(LV.getNode())) {
1440    DAG.RemoveDeadNodes();
1441    return true;
1442  }
1443
1444  // The lowered build-vector node will now need to be selected. It needs
1445  // to be done here because this node and its submodes are not included
1446  // in the main selection loop.
1447  // Implement essentially the same topological ordering algorithm as is
1448  // used in SelectionDAGISel.
1449
1450  SetVector<SDNode*> SubNodes, TmpQ;
1451  std::map<SDNode*,unsigned> NumOps;
1452
1453  SubNodes.insert(LV.getNode());
1454  for (unsigned I = 0; I != SubNodes.size(); ++I) {
1455    unsigned OpN = 0;
1456    SDNode *S = SubNodes[I];
1457    for (SDValue Op : S->ops()) {
1458      if (AllNodes.count(Op.getNode()))
1459        continue;
1460      SubNodes.insert(Op.getNode());
1461      ++OpN;
1462    }
1463    NumOps.insert({S, OpN});
1464    if (OpN == 0)
1465      TmpQ.insert(S);
1466  }
1467
1468  for (unsigned I = 0; I != TmpQ.size(); ++I) {
1469    SDNode *S = TmpQ[I];
1470    for (SDNode *U : S->uses()) {
1471      if (!SubNodes.count(U))
1472        continue;
1473      auto F = NumOps.find(U);
1474      assert(F != NumOps.end());
1475      assert(F->second > 0);
1476      if (!--F->second)
1477        TmpQ.insert(F->first);
1478    }
1479  }
1480  assert(SubNodes.size() == TmpQ.size());
1481  NullifyingVector<decltype(TmpQ)::vector_type> Queue(TmpQ.takeVector());
1482
1483  Deleter DUQ(DAG, Queue);
1484  for (SDNode *S : reverse(Queue))
1485    if (S != nullptr)
1486      ISel.Select(S);
1487
1488  DAG.RemoveDeadNodes();
1489  return true;
1490}
1491
1492OpRef HvxSelector::contracting(ShuffleMask SM, OpRef Va, OpRef Vb,
1493                               ResultStack &Results) {
1494  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1495  if (!Va.isValid() || !Vb.isValid())
1496    return OpRef::fail();
1497
1498  // Contracting shuffles, i.e. instructions that always discard some bytes
1499  // from the operand vectors.
1500  //
1501  // V6_vshuff{e,o}b
1502  // V6_vdealb4w
1503  // V6_vpack{e,o}{b,h}
1504
1505  int VecLen = SM.Mask.size();
1506  std::pair<int,unsigned> Strip = findStrip(SM.Mask, 1, VecLen);
1507  MVT ResTy = getSingleVT(MVT::i8);
1508
1509  // The following shuffles only work for bytes and halfwords. This requires
1510  // the strip length to be 1 or 2.
1511  if (Strip.second != 1 && Strip.second != 2)
1512    return OpRef::fail();
1513
1514  // The patterns for the shuffles, in terms of the starting offsets of the
1515  // consecutive strips (L = length of the strip, N = VecLen):
1516  //
1517  // vpacke:    0, 2L, 4L ... N+0, N+2L, N+4L ...      L = 1 or 2
1518  // vpacko:    L, 3L, 5L ... N+L, N+3L, N+5L ...      L = 1 or 2
1519  //
1520  // vshuffe:   0, N+0, 2L, N+2L, 4L ...               L = 1 or 2
1521  // vshuffo:   L, N+L, 3L, N+3L, 5L ...               L = 1 or 2
1522  //
1523  // vdealb4w:  0, 4, 8 ... 2, 6, 10 ... N+0, N+4, N+8 ... N+2, N+6, N+10 ...
1524
1525  // The value of the element in the mask following the strip will decide
1526  // what kind of a shuffle this can be.
1527  int NextInMask = SM.Mask[Strip.second];
1528
1529  // Check if NextInMask could be 2L, 3L or 4, i.e. if it could be a mask
1530  // for vpack or vdealb4w. VecLen > 4, so NextInMask for vdealb4w would
1531  // satisfy this.
1532  if (NextInMask < VecLen) {
1533    // vpack{e,o} or vdealb4w
1534    if (Strip.first == 0 && Strip.second == 1 && NextInMask == 4) {
1535      int N = VecLen;
1536      // Check if this is vdealb4w (L=1).
1537      for (int I = 0; I != N/4; ++I)
1538        if (SM.Mask[I] != 4*I)
1539          return OpRef::fail();
1540      for (int I = 0; I != N/4; ++I)
1541        if (SM.Mask[I+N/4] != 2 + 4*I)
1542          return OpRef::fail();
1543      for (int I = 0; I != N/4; ++I)
1544        if (SM.Mask[I+N/2] != N + 4*I)
1545          return OpRef::fail();
1546      for (int I = 0; I != N/4; ++I)
1547        if (SM.Mask[I+3*N/4] != N+2 + 4*I)
1548          return OpRef::fail();
1549      // Matched mask for vdealb4w.
1550      Results.push(Hexagon::V6_vdealb4w, ResTy, {Vb, Va});
1551      return OpRef::res(Results.top());
1552    }
1553
1554    // Check if this is vpack{e,o}.
1555    int N = VecLen;
1556    int L = Strip.second;
1557    // Check if the first strip starts at 0 or at L.
1558    if (Strip.first != 0 && Strip.first != L)
1559      return OpRef::fail();
1560    // Examine the rest of the mask.
1561    for (int I = L; I < N; I += L) {
1562      auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
1563      // Check whether the mask element at the beginning of each strip
1564      // increases by 2L each time.
1565      if (S.first - Strip.first != 2*I)
1566        return OpRef::fail();
1567      // Check whether each strip is of the same length.
1568      if (S.second != unsigned(L))
1569        return OpRef::fail();
1570    }
1571
1572    // Strip.first == 0  =>  vpacke
1573    // Strip.first == L  =>  vpacko
1574    assert(Strip.first == 0 || Strip.first == L);
1575    using namespace Hexagon;
1576    NodeTemplate Res;
1577    Res.Opc = Strip.second == 1 // Number of bytes.
1578                  ? (Strip.first == 0 ? V6_vpackeb : V6_vpackob)
1579                  : (Strip.first == 0 ? V6_vpackeh : V6_vpackoh);
1580    Res.Ty = ResTy;
1581    Res.Ops = { Vb, Va };
1582    Results.push(Res);
1583    return OpRef::res(Results.top());
1584  }
1585
1586  // Check if this is vshuff{e,o}.
1587  int N = VecLen;
1588  int L = Strip.second;
1589  std::pair<int,unsigned> PrevS = Strip;
1590  bool Flip = false;
1591  for (int I = L; I < N; I += L) {
1592    auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
1593    if (S.second != PrevS.second)
1594      return OpRef::fail();
1595    int Diff = Flip ? PrevS.first - S.first + 2*L
1596                    : S.first - PrevS.first;
1597    if (Diff != N)
1598      return OpRef::fail();
1599    Flip ^= true;
1600    PrevS = S;
1601  }
1602  // Strip.first == 0  =>  vshuffe
1603  // Strip.first == L  =>  vshuffo
1604  assert(Strip.first == 0 || Strip.first == L);
1605  using namespace Hexagon;
1606  NodeTemplate Res;
1607  Res.Opc = Strip.second == 1 // Number of bytes.
1608                ? (Strip.first == 0 ? V6_vshuffeb : V6_vshuffob)
1609                : (Strip.first == 0 ?  V6_vshufeh :  V6_vshufoh);
1610  Res.Ty = ResTy;
1611  Res.Ops = { Vb, Va };
1612  Results.push(Res);
1613  return OpRef::res(Results.top());
1614}
1615
1616OpRef HvxSelector::expanding(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1617  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1618  // Expanding shuffles (using all elements and inserting into larger vector):
1619  //
1620  // V6_vunpacku{b,h} [*]
1621  //
1622  // [*] Only if the upper elements (filled with 0s) are "don't care" in Mask.
1623  //
1624  // Note: V6_vunpacko{b,h} are or-ing the high byte/half in the result, so
1625  // they are not shuffles.
1626  //
1627  // The argument is a single vector.
1628
1629  int VecLen = SM.Mask.size();
1630  assert(2*HwLen == unsigned(VecLen) && "Expecting vector-pair type");
1631
1632  std::pair<int,unsigned> Strip = findStrip(SM.Mask, 1, VecLen);
1633
1634  // The patterns for the unpacks, in terms of the starting offsets of the
1635  // consecutive strips (L = length of the strip, N = VecLen):
1636  //
1637  // vunpacku:  0, -1, L, -1, 2L, -1 ...
1638
1639  if (Strip.first != 0)
1640    return OpRef::fail();
1641
1642  // The vunpackus only handle byte and half-word.
1643  if (Strip.second != 1 && Strip.second != 2)
1644    return OpRef::fail();
1645
1646  int N = VecLen;
1647  int L = Strip.second;
1648
1649  // First, check the non-ignored strips.
1650  for (int I = 2*L; I < 2*N; I += 2*L) {
1651    auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
1652    if (S.second != unsigned(L))
1653      return OpRef::fail();
1654    if (2*S.first != I)
1655      return OpRef::fail();
1656  }
1657  // Check the -1s.
1658  for (int I = L; I < 2*N; I += 2*L) {
1659    auto S = findStrip(SM.Mask.drop_front(I), 0, N-I);
1660    if (S.first != -1 || S.second != unsigned(L))
1661      return OpRef::fail();
1662  }
1663
1664  unsigned Opc = Strip.second == 1 ? Hexagon::V6_vunpackub
1665                                   : Hexagon::V6_vunpackuh;
1666  Results.push(Opc, getPairVT(MVT::i8), {Va});
1667  return OpRef::res(Results.top());
1668}
1669
1670OpRef HvxSelector::perfect(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1671  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1672  // V6_vdeal{b,h}
1673  // V6_vshuff{b,h}
1674
1675  // V6_vshufoe{b,h}  those are quivalent to vshuffvdd(..,{1,2})
1676  // V6_vshuffvdd (V6_vshuff)
1677  // V6_dealvdd (V6_vdeal)
1678
1679  int VecLen = SM.Mask.size();
1680  assert(isPowerOf2_32(VecLen) && Log2_32(VecLen) <= 8);
1681  unsigned LogLen = Log2_32(VecLen);
1682  unsigned HwLog = Log2_32(HwLen);
1683  // The result length must be the same as the length of a single vector,
1684  // or a vector pair.
1685  assert(LogLen == HwLog || LogLen == HwLog+1);
1686  bool Extend = (LogLen == HwLog);
1687
1688  if (!isPermutation(SM.Mask))
1689    return OpRef::fail();
1690
1691  SmallVector<unsigned,8> Perm(LogLen);
1692
1693  // Check if this could be a perfect shuffle, or a combination of perfect
1694  // shuffles.
1695  //
1696  // Consider this permutation (using hex digits to make the ASCII diagrams
1697  // easier to read):
1698  //   { 0, 8, 1, 9, 2, A, 3, B, 4, C, 5, D, 6, E, 7, F }.
1699  // This is a "deal" operation: divide the input into two halves, and
1700  // create the output by picking elements by alternating between these two
1701  // halves:
1702  //   0 1 2 3 4 5 6 7    -->    0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F  [*]
1703  //   8 9 A B C D E F
1704  //
1705  // Aside from a few special explicit cases (V6_vdealb, etc.), HVX provides
1706  // a somwehat different mechanism that could be used to perform shuffle/
1707  // deal operations: a 2x2 transpose.
1708  // Consider the halves of inputs again, they can be interpreted as a 2x8
1709  // matrix. A 2x8 matrix can be looked at four 2x2 matrices concatenated
1710  // together. Now, when considering 2 elements at a time, it will be a 2x4
1711  // matrix (with elements 01, 23, 45, etc.), or two 2x2 matrices:
1712  //   01 23  45 67
1713  //   89 AB  CD EF
1714  // With groups of 4, this will become a single 2x2 matrix, and so on.
1715  //
1716  // The 2x2 transpose instruction works by transposing each of the 2x2
1717  // matrices (or "sub-matrices"), given a specific group size. For example,
1718  // if the group size is 1 (i.e. each element is its own group), there
1719  // will be four transposes of the four 2x2 matrices that form the 2x8.
1720  // For example, with the inputs as above, the result will be:
1721  //   0 8  2 A  4 C  6 E
1722  //   1 9  3 B  5 D  7 F
1723  // Now, this result can be tranposed again, but with the group size of 2:
1724  //   08 19  4C 5D
1725  //   2A 3B  6E 7F
1726  // If we then transpose that result, but with the group size of 4, we get:
1727  //   0819 2A3B
1728  //   4C5D 6E7F
1729  // If we concatenate these two rows, it will be
1730  //   0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F
1731  // which is the same as the "deal" [*] above.
1732  //
1733  // In general, a "deal" of individual elements is a series of 2x2 transposes,
1734  // with changing group size. HVX has two instructions:
1735  //   Vdd = V6_vdealvdd Vu, Vv, Rt
1736  //   Vdd = V6_shufvdd  Vu, Vv, Rt
1737  // that perform exactly that. The register Rt controls which transposes are
1738  // going to happen: a bit at position n (counting from 0) indicates that a
1739  // transpose with a group size of 2^n will take place. If multiple bits are
1740  // set, multiple transposes will happen: vdealvdd will perform them starting
1741  // with the largest group size, vshuffvdd will do them in the reverse order.
1742  //
1743  // The main observation is that each 2x2 transpose corresponds to swapping
1744  // columns of bits in the binary representation of the values.
1745  //
1746  // The numbers {3,2,1,0} and the log2 of the number of contiguous 1 bits
1747  // in a given column. The * denote the columns that will be swapped.
1748  // The transpose with the group size 2^n corresponds to swapping columns
1749  // 3 (the highest log) and log2(n):
1750  //
1751  //     3 2 1 0         0 2 1 3         0 2 3 1
1752  //     *     *             * *           * *
1753  //  0  0 0 0 0      0  0 0 0 0      0  0 0 0 0      0  0 0 0 0
1754  //  1  0 0 0 1      8  1 0 0 0      8  1 0 0 0      8  1 0 0 0
1755  //  2  0 0 1 0      2  0 0 1 0      1  0 0 0 1      1  0 0 0 1
1756  //  3  0 0 1 1      A  1 0 1 0      9  1 0 0 1      9  1 0 0 1
1757  //  4  0 1 0 0      4  0 1 0 0      4  0 1 0 0      2  0 0 1 0
1758  //  5  0 1 0 1      C  1 1 0 0      C  1 1 0 0      A  1 0 1 0
1759  //  6  0 1 1 0      6  0 1 1 0      5  0 1 0 1      3  0 0 1 1
1760  //  7  0 1 1 1      E  1 1 1 0      D  1 1 0 1      B  1 0 1 1
1761  //  8  1 0 0 0      1  0 0 0 1      2  0 0 1 0      4  0 1 0 0
1762  //  9  1 0 0 1      9  1 0 0 1      A  1 0 1 0      C  1 1 0 0
1763  //  A  1 0 1 0      3  0 0 1 1      3  0 0 1 1      5  0 1 0 1
1764  //  B  1 0 1 1      B  1 0 1 1      B  1 0 1 1      D  1 1 0 1
1765  //  C  1 1 0 0      5  0 1 0 1      6  0 1 1 0      6  0 1 1 0
1766  //  D  1 1 0 1      D  1 1 0 1      E  1 1 1 0      E  1 1 1 0
1767  //  E  1 1 1 0      7  0 1 1 1      7  0 1 1 1      7  0 1 1 1
1768  //  F  1 1 1 1      F  1 1 1 1      F  1 1 1 1      F  1 1 1 1
1769
1770  auto XorPow2 = [] (ArrayRef<int> Mask, unsigned Num) {
1771    unsigned X = Mask[0] ^ Mask[Num/2];
1772    // Check that the first half has the X's bits clear.
1773    if ((Mask[0] & X) != 0)
1774      return 0u;
1775    for (unsigned I = 1; I != Num/2; ++I) {
1776      if (unsigned(Mask[I] ^ Mask[I+Num/2]) != X)
1777        return 0u;
1778      if ((Mask[I] & X) != 0)
1779        return 0u;
1780    }
1781    return X;
1782  };
1783
1784  // Create a vector of log2's for each column: Perm[i] corresponds to
1785  // the i-th bit (lsb is 0).
1786  assert(VecLen > 2);
1787  for (unsigned I = VecLen; I >= 2; I >>= 1) {
1788    // Examine the initial segment of Mask of size I.
1789    unsigned X = XorPow2(SM.Mask, I);
1790    if (!isPowerOf2_32(X))
1791      return OpRef::fail();
1792    // Check the other segments of Mask.
1793    for (int J = I; J < VecLen; J += I) {
1794      if (XorPow2(SM.Mask.slice(J, I), I) != X)
1795        return OpRef::fail();
1796    }
1797    Perm[Log2_32(X)] = Log2_32(I)-1;
1798  }
1799
1800  // Once we have Perm, represent it as cycles. Denote the maximum log2
1801  // (equal to log2(VecLen)-1) as M. The cycle containing M can then be
1802  // written as (M a1 a2 a3 ... an). That cycle can be broken up into
1803  // simple swaps as (M a1)(M a2)(M a3)...(M an), with the composition
1804  // order being from left to right. Any (contiguous) segment where the
1805  // values ai, ai+1...aj are either all increasing or all decreasing,
1806  // can be implemented via a single vshuffvdd/vdealvdd respectively.
1807  //
1808  // If there is a cycle (a1 a2 ... an) that does not involve M, it can
1809  // be written as (M an)(a1 a2 ... an)(M a1). The first two cycles can
1810  // then be folded to get (M a1 a2 ... an)(M a1), and the above procedure
1811  // can be used to generate a sequence of vshuffvdd/vdealvdd.
1812  //
1813  // Example:
1814  // Assume M = 4 and consider a permutation (0 1)(2 3). It can be written
1815  // as (4 0 1)(4 0) composed with (4 2 3)(4 2), or simply
1816  //   (4 0 1)(4 0)(4 2 3)(4 2).
1817  // It can then be expanded into swaps as
1818  //   (4 0)(4 1)(4 0)(4 2)(4 3)(4 2),
1819  // and broken up into "increasing" segments as
1820  //   [(4 0)(4 1)] [(4 0)(4 2)(4 3)] [(4 2)].
1821  // This is equivalent to
1822  //   (4 0 1)(4 0 2 3)(4 2),
1823  // which can be implemented as 3 vshufvdd instructions.
1824
1825  using CycleType = SmallVector<unsigned,8>;
1826  std::set<CycleType> Cycles;
1827  std::set<unsigned> All;
1828
1829  for (unsigned I : Perm)
1830    All.insert(I);
1831
1832  // If the cycle contains LogLen-1, move it to the front of the cycle.
1833  // Otherwise, return the cycle unchanged.
1834  auto canonicalize = [LogLen](const CycleType &C) -> CycleType {
1835    unsigned LogPos, N = C.size();
1836    for (LogPos = 0; LogPos != N; ++LogPos)
1837      if (C[LogPos] == LogLen-1)
1838        break;
1839    if (LogPos == N)
1840      return C;
1841
1842    CycleType NewC(C.begin()+LogPos, C.end());
1843    NewC.append(C.begin(), C.begin()+LogPos);
1844    return NewC;
1845  };
1846
1847  auto pfs = [](const std::set<CycleType> &Cs, unsigned Len) {
1848    // Ordering: shuff: 5 0 1 2 3 4, deal: 5 4 3 2 1 0 (for Log=6),
1849    // for bytes zero is included, for halfwords is not.
1850    if (Cs.size() != 1)
1851      return 0u;
1852    const CycleType &C = *Cs.begin();
1853    if (C[0] != Len-1)
1854      return 0u;
1855    int D = Len - C.size();
1856    if (D != 0 && D != 1)
1857      return 0u;
1858
1859    bool IsDeal = true, IsShuff = true;
1860    for (unsigned I = 1; I != Len-D; ++I) {
1861      if (C[I] != Len-1-I)
1862        IsDeal = false;
1863      if (C[I] != I-(1-D))  // I-1, I
1864        IsShuff = false;
1865    }
1866    // At most one, IsDeal or IsShuff, can be non-zero.
1867    assert(!(IsDeal || IsShuff) || IsDeal != IsShuff);
1868    static unsigned Deals[] = { Hexagon::V6_vdealb, Hexagon::V6_vdealh };
1869    static unsigned Shufs[] = { Hexagon::V6_vshuffb, Hexagon::V6_vshuffh };
1870    return IsDeal ? Deals[D] : (IsShuff ? Shufs[D] : 0);
1871  };
1872
1873  while (!All.empty()) {
1874    unsigned A = *All.begin();
1875    All.erase(A);
1876    CycleType C;
1877    C.push_back(A);
1878    for (unsigned B = Perm[A]; B != A; B = Perm[B]) {
1879      C.push_back(B);
1880      All.erase(B);
1881    }
1882    if (C.size() <= 1)
1883      continue;
1884    Cycles.insert(canonicalize(C));
1885  }
1886
1887  MVT SingleTy = getSingleVT(MVT::i8);
1888  MVT PairTy = getPairVT(MVT::i8);
1889
1890  // Recognize patterns for V6_vdeal{b,h} and V6_vshuff{b,h}.
1891  if (unsigned(VecLen) == HwLen) {
1892    if (unsigned SingleOpc = pfs(Cycles, LogLen)) {
1893      Results.push(SingleOpc, SingleTy, {Va});
1894      return OpRef::res(Results.top());
1895    }
1896  }
1897
1898  SmallVector<unsigned,8> SwapElems;
1899  if (HwLen == unsigned(VecLen))
1900    SwapElems.push_back(LogLen-1);
1901
1902  for (const CycleType &C : Cycles) {
1903    unsigned First = (C[0] == LogLen-1) ? 1 : 0;
1904    SwapElems.append(C.begin()+First, C.end());
1905    if (First == 0)
1906      SwapElems.push_back(C[0]);
1907  }
1908
1909  const SDLoc &dl(Results.InpNode);
1910  OpRef Arg = !Extend ? Va
1911                      : concat(Va, OpRef::undef(SingleTy), Results);
1912
1913  for (unsigned I = 0, E = SwapElems.size(); I != E; ) {
1914    bool IsInc = I == E-1 || SwapElems[I] < SwapElems[I+1];
1915    unsigned S = (1u << SwapElems[I]);
1916    if (I < E-1) {
1917      while (++I < E-1 && IsInc == (SwapElems[I] < SwapElems[I+1]))
1918        S |= 1u << SwapElems[I];
1919      // The above loop will not add a bit for the final SwapElems[I+1],
1920      // so add it here.
1921      S |= 1u << SwapElems[I];
1922    }
1923    ++I;
1924
1925    NodeTemplate Res;
1926    Results.push(Hexagon::A2_tfrsi, MVT::i32,
1927                 { DAG.getTargetConstant(S, dl, MVT::i32) });
1928    Res.Opc = IsInc ? Hexagon::V6_vshuffvdd : Hexagon::V6_vdealvdd;
1929    Res.Ty = PairTy;
1930    Res.Ops = { OpRef::hi(Arg), OpRef::lo(Arg), OpRef::res(-1) };
1931    Results.push(Res);
1932    Arg = OpRef::res(Results.top());
1933  }
1934
1935  return !Extend ? Arg : OpRef::lo(Arg);
1936}
1937
1938OpRef HvxSelector::butterfly(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1939  DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1940  // Butterfly shuffles.
1941  //
1942  // V6_vdelta
1943  // V6_vrdelta
1944  // V6_vror
1945
1946  // The assumption here is that all elements picked by Mask are in the
1947  // first operand to the vector_shuffle. This assumption is enforced
1948  // by the caller.
1949
1950  MVT ResTy = getSingleVT(MVT::i8);
1951  PermNetwork::Controls FC, RC;
1952  const SDLoc &dl(Results.InpNode);
1953  int VecLen = SM.Mask.size();
1954
1955  for (int M : SM.Mask) {
1956    if (M != -1 && M >= VecLen)
1957      return OpRef::fail();
1958  }
1959
1960  // Try the deltas/benes for both single vectors and vector pairs.
1961  ForwardDeltaNetwork FN(SM.Mask);
1962  if (FN.run(FC)) {
1963    SDValue Ctl = getVectorConstant(FC, dl);
1964    Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(Ctl)});
1965    return OpRef::res(Results.top());
1966  }
1967
1968  // Try reverse delta.
1969  ReverseDeltaNetwork RN(SM.Mask);
1970  if (RN.run(RC)) {
1971    SDValue Ctl = getVectorConstant(RC, dl);
1972    Results.push(Hexagon::V6_vrdelta, ResTy, {Va, OpRef(Ctl)});
1973    return OpRef::res(Results.top());
1974  }
1975
1976  // Do Benes.
1977  BenesNetwork BN(SM.Mask);
1978  if (BN.run(FC, RC)) {
1979    SDValue CtlF = getVectorConstant(FC, dl);
1980    SDValue CtlR = getVectorConstant(RC, dl);
1981    Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(CtlF)});
1982    Results.push(Hexagon::V6_vrdelta, ResTy,
1983                 {OpRef::res(-1), OpRef(CtlR)});
1984    return OpRef::res(Results.top());
1985  }
1986
1987  return OpRef::fail();
1988}
1989
1990SDValue HvxSelector::getVectorConstant(ArrayRef<uint8_t> Data,
1991                                       const SDLoc &dl) {
1992  SmallVector<SDValue, 128> Elems;
1993  for (uint8_t C : Data)
1994    Elems.push_back(DAG.getConstant(C, dl, MVT::i8));
1995  MVT VecTy = MVT::getVectorVT(MVT::i8, Data.size());
1996  SDValue BV = DAG.getBuildVector(VecTy, dl, Elems);
1997  SDValue LV = Lower.LowerOperation(BV, DAG);
1998  DAG.RemoveDeadNode(BV.getNode());
1999  return LV;
2000}
2001
2002void HvxSelector::selectShuffle(SDNode *N) {
2003  DEBUG_WITH_TYPE("isel", {
2004    dbgs() << "Starting " << __func__ << " on node:\n";
2005    N->dump(&DAG);
2006  });
2007  MVT ResTy = N->getValueType(0).getSimpleVT();
2008  // Assume that vector shuffles operate on vectors of bytes.
2009  assert(ResTy.isVector() && ResTy.getVectorElementType() == MVT::i8);
2010
2011  auto *SN = cast<ShuffleVectorSDNode>(N);
2012  std::vector<int> Mask(SN->getMask().begin(), SN->getMask().end());
2013  // This shouldn't really be necessary. Is it?
2014  for (int &Idx : Mask)
2015    if (Idx != -1 && Idx < 0)
2016      Idx = -1;
2017
2018  unsigned VecLen = Mask.size();
2019  bool HavePairs = (2*HwLen == VecLen);
2020  assert(ResTy.getSizeInBits() / 8 == VecLen);
2021
2022  // Vd = vector_shuffle Va, Vb, Mask
2023  //
2024
2025  bool UseLeft = false, UseRight = false;
2026  for (unsigned I = 0; I != VecLen; ++I) {
2027    if (Mask[I] == -1)
2028      continue;
2029    unsigned Idx = Mask[I];
2030    assert(Idx < 2*VecLen);
2031    if (Idx < VecLen)
2032      UseLeft = true;
2033    else
2034      UseRight = true;
2035  }
2036
2037  DEBUG_WITH_TYPE("isel", {
2038    dbgs() << "VecLen=" << VecLen << " HwLen=" << HwLen << " UseLeft="
2039           << UseLeft << " UseRight=" << UseRight << " HavePairs="
2040           << HavePairs << '\n';
2041  });
2042  // If the mask is all -1's, generate "undef".
2043  if (!UseLeft && !UseRight) {
2044    ISel.ReplaceNode(N, ISel.selectUndef(SDLoc(SN), ResTy).getNode());
2045    return;
2046  }
2047
2048  SDValue Vec0 = N->getOperand(0);
2049  SDValue Vec1 = N->getOperand(1);
2050  ResultStack Results(SN);
2051  Results.push(TargetOpcode::COPY, ResTy, {Vec0});
2052  Results.push(TargetOpcode::COPY, ResTy, {Vec1});
2053  OpRef Va = OpRef::res(Results.top()-1);
2054  OpRef Vb = OpRef::res(Results.top());
2055
2056  OpRef Res = !HavePairs ? shuffs2(ShuffleMask(Mask), Va, Vb, Results)
2057                         : shuffp2(ShuffleMask(Mask), Va, Vb, Results);
2058
2059  bool Done = Res.isValid();
2060  if (Done) {
2061    // Make sure that Res is on the stack before materializing.
2062    Results.push(TargetOpcode::COPY, ResTy, {Res});
2063    materialize(Results);
2064  } else {
2065    Done = scalarizeShuffle(Mask, SDLoc(N), ResTy, Vec0, Vec1, N);
2066  }
2067
2068  if (!Done) {
2069#ifndef NDEBUG
2070    dbgs() << "Unhandled shuffle:\n";
2071    SN->dumpr(&DAG);
2072#endif
2073    llvm_unreachable("Failed to select vector shuffle");
2074  }
2075}
2076
2077void HvxSelector::selectRor(SDNode *N) {
2078  // If this is a rotation by less than 8, use V6_valignbi.
2079  MVT Ty = N->getValueType(0).getSimpleVT();
2080  const SDLoc &dl(N);
2081  SDValue VecV = N->getOperand(0);
2082  SDValue RotV = N->getOperand(1);
2083  SDNode *NewN = nullptr;
2084
2085  if (auto *CN = dyn_cast<ConstantSDNode>(RotV.getNode())) {
2086    unsigned S = CN->getZExtValue() % HST.getVectorLength();
2087    if (S == 0) {
2088      NewN = VecV.getNode();
2089    } else if (isUInt<3>(S)) {
2090      SDValue C = DAG.getTargetConstant(S, dl, MVT::i32);
2091      NewN = DAG.getMachineNode(Hexagon::V6_valignbi, dl, Ty,
2092                                {VecV, VecV, C});
2093    }
2094  }
2095
2096  if (!NewN)
2097    NewN = DAG.getMachineNode(Hexagon::V6_vror, dl, Ty, {VecV, RotV});
2098
2099  ISel.ReplaceNode(N, NewN);
2100}
2101
2102void HvxSelector::selectVAlign(SDNode *N) {
2103  SDValue Vv = N->getOperand(0);
2104  SDValue Vu = N->getOperand(1);
2105  SDValue Rt = N->getOperand(2);
2106  SDNode *NewN = DAG.getMachineNode(Hexagon::V6_valignb, SDLoc(N),
2107                                    N->getValueType(0), {Vv, Vu, Rt});
2108  ISel.ReplaceNode(N, NewN);
2109  DAG.RemoveDeadNode(N);
2110}
2111
2112void HexagonDAGToDAGISel::SelectHvxShuffle(SDNode *N) {
2113  HvxSelector(*this, *CurDAG).selectShuffle(N);
2114}
2115
2116void HexagonDAGToDAGISel::SelectHvxRor(SDNode *N) {
2117  HvxSelector(*this, *CurDAG).selectRor(N);
2118}
2119
2120void HexagonDAGToDAGISel::SelectHvxVAlign(SDNode *N) {
2121  HvxSelector(*this, *CurDAG).selectVAlign(N);
2122}
2123
2124void HexagonDAGToDAGISel::SelectV65GatherPred(SDNode *N) {
2125  const SDLoc &dl(N);
2126  SDValue Chain = N->getOperand(0);
2127  SDValue Address = N->getOperand(2);
2128  SDValue Predicate = N->getOperand(3);
2129  SDValue Base = N->getOperand(4);
2130  SDValue Modifier = N->getOperand(5);
2131  SDValue Offset = N->getOperand(6);
2132
2133  unsigned Opcode;
2134  unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2135  switch (IntNo) {
2136  default:
2137    llvm_unreachable("Unexpected HVX gather intrinsic.");
2138  case Intrinsic::hexagon_V6_vgathermhq:
2139  case Intrinsic::hexagon_V6_vgathermhq_128B:
2140    Opcode = Hexagon::V6_vgathermhq_pseudo;
2141    break;
2142  case Intrinsic::hexagon_V6_vgathermwq:
2143  case Intrinsic::hexagon_V6_vgathermwq_128B:
2144    Opcode = Hexagon::V6_vgathermwq_pseudo;
2145    break;
2146  case Intrinsic::hexagon_V6_vgathermhwq:
2147  case Intrinsic::hexagon_V6_vgathermhwq_128B:
2148    Opcode = Hexagon::V6_vgathermhwq_pseudo;
2149    break;
2150  }
2151
2152  SDVTList VTs = CurDAG->getVTList(MVT::Other);
2153  SDValue Ops[] = { Address, Predicate, Base, Modifier, Offset, Chain };
2154  SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2155
2156  MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2157  CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2158
2159  ReplaceNode(N, Result);
2160}
2161
2162void HexagonDAGToDAGISel::SelectV65Gather(SDNode *N) {
2163  const SDLoc &dl(N);
2164  SDValue Chain = N->getOperand(0);
2165  SDValue Address = N->getOperand(2);
2166  SDValue Base = N->getOperand(3);
2167  SDValue Modifier = N->getOperand(4);
2168  SDValue Offset = N->getOperand(5);
2169
2170  unsigned Opcode;
2171  unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2172  switch (IntNo) {
2173  default:
2174    llvm_unreachable("Unexpected HVX gather intrinsic.");
2175  case Intrinsic::hexagon_V6_vgathermh:
2176  case Intrinsic::hexagon_V6_vgathermh_128B:
2177    Opcode = Hexagon::V6_vgathermh_pseudo;
2178    break;
2179  case Intrinsic::hexagon_V6_vgathermw:
2180  case Intrinsic::hexagon_V6_vgathermw_128B:
2181    Opcode = Hexagon::V6_vgathermw_pseudo;
2182    break;
2183  case Intrinsic::hexagon_V6_vgathermhw:
2184  case Intrinsic::hexagon_V6_vgathermhw_128B:
2185    Opcode = Hexagon::V6_vgathermhw_pseudo;
2186    break;
2187  }
2188
2189  SDVTList VTs = CurDAG->getVTList(MVT::Other);
2190  SDValue Ops[] = { Address, Base, Modifier, Offset, Chain };
2191  SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2192
2193  MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2194  CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2195
2196  ReplaceNode(N, Result);
2197}
2198
2199void HexagonDAGToDAGISel::SelectHVXDualOutput(SDNode *N) {
2200  unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2201  SDNode *Result;
2202  switch (IID) {
2203  case Intrinsic::hexagon_V6_vaddcarry: {
2204    SmallVector<SDValue, 3> Ops = { N->getOperand(1), N->getOperand(2),
2205                                    N->getOperand(3) };
2206    SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v512i1);
2207    Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2208    break;
2209  }
2210  case Intrinsic::hexagon_V6_vaddcarry_128B: {
2211    SmallVector<SDValue, 3> Ops = { N->getOperand(1), N->getOperand(2),
2212                                    N->getOperand(3) };
2213    SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v1024i1);
2214    Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2215    break;
2216  }
2217  case Intrinsic::hexagon_V6_vsubcarry: {
2218    SmallVector<SDValue, 3> Ops = { N->getOperand(1), N->getOperand(2),
2219                                    N->getOperand(3) };
2220    SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v512i1);
2221    Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2222    break;
2223  }
2224  case Intrinsic::hexagon_V6_vsubcarry_128B: {
2225    SmallVector<SDValue, 3> Ops = { N->getOperand(1), N->getOperand(2),
2226                                    N->getOperand(3) };
2227    SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v1024i1);
2228    Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2229    break;
2230  }
2231  default:
2232    llvm_unreachable("Unexpected HVX dual output intrinsic.");
2233  }
2234  ReplaceUses(N, Result);
2235  ReplaceUses(SDValue(N, 0), SDValue(Result, 0));
2236  ReplaceUses(SDValue(N, 1), SDValue(Result, 1));
2237  CurDAG->RemoveDeadNode(N);
2238}
2239