xref: /freebsd/contrib/llvm-project/llvm/lib/Target/Hexagon/HexagonISelDAGToDAGHVX.cpp (revision 7a6dacaca14b62ca4b74406814becb87a3fefac0)
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/BitVector.h"
14 #include "llvm/ADT/SetVector.h"
15 #include "llvm/CodeGen/MachineInstrBuilder.h"
16 #include "llvm/CodeGen/SelectionDAGISel.h"
17 #include "llvm/IR/Intrinsics.h"
18 #include "llvm/IR/IntrinsicsHexagon.h"
19 #include "llvm/Support/CommandLine.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Support/MathExtras.h"
22 
23 #include <algorithm>
24 #include <cmath>
25 #include <deque>
26 #include <functional>
27 #include <map>
28 #include <optional>
29 #include <set>
30 #include <unordered_map>
31 #include <utility>
32 #include <vector>
33 
34 #define DEBUG_TYPE "hexagon-isel"
35 using namespace llvm;
36 
37 namespace {
38 
39 // --------------------------------------------------------------------
40 // Implementation of permutation networks.
41 
42 // Implementation of the node routing through butterfly networks:
43 // - Forward delta.
44 // - Reverse delta.
45 // - Benes.
46 //
47 //
48 // Forward delta network consists of log(N) steps, where N is the number
49 // of inputs. In each step, an input can stay in place, or it can get
50 // routed to another position[1]. The step after that consists of two
51 // networks, each half in size in terms of the number of nodes. In those
52 // terms, in the given step, an input can go to either the upper or the
53 // lower network in the next step.
54 //
55 // [1] Hexagon's vdelta/vrdelta allow an element to be routed to both
56 // positions as long as there is no conflict.
57 
58 // Here's a delta network for 8 inputs, only the switching routes are
59 // shown:
60 //
61 //         Steps:
62 //         |- 1 ---------------|- 2 -----|- 3 -|
63 //
64 // Inp[0] ***                 ***       ***   *** Out[0]
65 //           \               /   \     /   \ /
66 //            \             /     \   /     X
67 //             \           /       \ /     / \
68 // Inp[1] ***   \         /   ***   X   ***   *** Out[1]
69 //           \   \       /   /   \ / \ /
70 //            \   \     /   /     X   X
71 //             \   \   /   /     / \ / \
72 // Inp[2] ***   \   \ /   /   ***   X   ***   *** Out[2]
73 //           \   \   X   /   /     / \     \ /
74 //            \   \ / \ /   /     /   \     X
75 //             \   X   X   /     /     \   / \
76 // Inp[3] ***   \ / \ / \ /   ***       ***   *** Out[3]
77 //           \   X   X   X   /
78 //            \ / \ / \ / \ /
79 //             X   X   X   X
80 //            / \ / \ / \ / \
81 //           /   X   X   X   \
82 // Inp[4] ***   / \ / \ / \   ***       ***   *** Out[4]
83 //             /   X   X   \     \     /   \ /
84 //            /   / \ / \   \     \   /     X
85 //           /   /   X   \   \     \ /     / \
86 // Inp[5] ***   /   / \   \   ***   X   ***   *** Out[5]
87 //             /   /   \   \     \ / \ /
88 //            /   /     \   \     X   X
89 //           /   /       \   \   / \ / \
90 // Inp[6] ***   /         \   ***   X   ***   *** Out[6]
91 //             /           \       / \     \ /
92 //            /             \     /   \     X
93 //           /               \   /     \   / \
94 // Inp[7] ***                 ***       ***   *** Out[7]
95 //
96 //
97 // Reverse delta network is same as delta network, with the steps in
98 // the opposite order.
99 //
100 //
101 // Benes network is a forward delta network immediately followed by
102 // a reverse delta network.
103 
104 enum class ColorKind { None, Red, Black };
105 
106 // Graph coloring utility used to partition nodes into two groups:
107 // they will correspond to nodes routed to the upper and lower networks.
108 struct Coloring {
109   using Node = int;
110   using MapType = std::map<Node, ColorKind>;
111   static constexpr Node Ignore = Node(-1);
112 
113   Coloring(ArrayRef<Node> Ord) : Order(Ord) {
114     build();
115     if (!color())
116       Colors.clear();
117   }
118 
119   const MapType &colors() const {
120     return Colors;
121   }
122 
123   ColorKind other(ColorKind Color) {
124     if (Color == ColorKind::None)
125       return ColorKind::Red;
126     return Color == ColorKind::Red ? ColorKind::Black : ColorKind::Red;
127   }
128 
129   LLVM_DUMP_METHOD void dump() const;
130 
131 private:
132   ArrayRef<Node> Order;
133   MapType Colors;
134   std::set<Node> Needed;
135 
136   using NodeSet = std::set<Node>;
137   std::map<Node,NodeSet> Edges;
138 
139   Node conj(Node Pos) {
140     Node Num = Order.size();
141     return (Pos < Num/2) ? Pos + Num/2 : Pos - Num/2;
142   }
143 
144   ColorKind getColor(Node N) {
145     auto F = Colors.find(N);
146     return F != Colors.end() ? F->second : ColorKind::None;
147   }
148 
149   std::pair<bool, ColorKind> getUniqueColor(const NodeSet &Nodes);
150 
151   void build();
152   bool color();
153 };
154 } // namespace
155 
156 std::pair<bool, ColorKind> Coloring::getUniqueColor(const NodeSet &Nodes) {
157   auto Color = ColorKind::None;
158   for (Node N : Nodes) {
159     ColorKind ColorN = getColor(N);
160     if (ColorN == ColorKind::None)
161       continue;
162     if (Color == ColorKind::None)
163       Color = ColorN;
164     else if (Color != ColorKind::None && Color != ColorN)
165       return { false, ColorKind::None };
166   }
167   return { true, Color };
168 }
169 
170 void Coloring::build() {
171   // Add Order[P] and Order[conj(P)] to Edges.
172   for (unsigned P = 0; P != Order.size(); ++P) {
173     Node I = Order[P];
174     if (I != Ignore) {
175       Needed.insert(I);
176       Node PC = Order[conj(P)];
177       if (PC != Ignore && PC != I)
178         Edges[I].insert(PC);
179     }
180   }
181   // Add I and conj(I) to Edges.
182   for (unsigned I = 0; I != Order.size(); ++I) {
183     if (!Needed.count(I))
184       continue;
185     Node C = conj(I);
186     // This will create an entry in the edge table, even if I is not
187     // connected to any other node. This is necessary, because it still
188     // needs to be colored.
189     NodeSet &Is = Edges[I];
190     if (Needed.count(C))
191       Is.insert(C);
192   }
193 }
194 
195 bool Coloring::color() {
196   SetVector<Node> FirstQ;
197   auto Enqueue = [this,&FirstQ] (Node N) {
198     SetVector<Node> Q;
199     Q.insert(N);
200     for (unsigned I = 0; I != Q.size(); ++I) {
201       NodeSet &Ns = Edges[Q[I]];
202       Q.insert(Ns.begin(), Ns.end());
203     }
204     FirstQ.insert(Q.begin(), Q.end());
205   };
206   for (Node N : Needed)
207     Enqueue(N);
208 
209   for (Node N : FirstQ) {
210     if (Colors.count(N))
211       continue;
212     NodeSet &Ns = Edges[N];
213     auto P = getUniqueColor(Ns);
214     if (!P.first)
215       return false;
216     Colors[N] = other(P.second);
217   }
218 
219   // First, color nodes that don't have any dups.
220   for (auto E : Edges) {
221     Node N = E.first;
222     if (!Needed.count(conj(N)) || Colors.count(N))
223       continue;
224     auto P = getUniqueColor(E.second);
225     if (!P.first)
226       return false;
227     Colors[N] = other(P.second);
228   }
229 
230   // Now, nodes that are still uncolored. Since the graph can be modified
231   // in this step, create a work queue.
232   std::vector<Node> WorkQ;
233   for (auto E : Edges) {
234     Node N = E.first;
235     if (!Colors.count(N))
236       WorkQ.push_back(N);
237   }
238 
239   for (Node N : WorkQ) {
240     NodeSet &Ns = Edges[N];
241     auto P = getUniqueColor(Ns);
242     if (P.first) {
243       Colors[N] = other(P.second);
244       continue;
245     }
246 
247     // Coloring failed. Split this node.
248     Node C = conj(N);
249     ColorKind ColorN = other(ColorKind::None);
250     ColorKind ColorC = other(ColorN);
251     NodeSet &Cs = Edges[C];
252     NodeSet CopyNs = Ns;
253     for (Node M : CopyNs) {
254       ColorKind ColorM = getColor(M);
255       if (ColorM == ColorC) {
256         // Connect M with C, disconnect M from N.
257         Cs.insert(M);
258         Edges[M].insert(C);
259         Ns.erase(M);
260         Edges[M].erase(N);
261       }
262     }
263     Colors[N] = ColorN;
264     Colors[C] = ColorC;
265   }
266 
267   // Explicitly assign "None" to all uncolored nodes.
268   for (unsigned I = 0; I != Order.size(); ++I)
269     if (Colors.count(I) == 0)
270       Colors[I] = ColorKind::None;
271 
272   return true;
273 }
274 
275 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
276 void Coloring::dump() const {
277   dbgs() << "{ Order:   {";
278   for (Node P : Order) {
279     if (P != Ignore)
280       dbgs() << ' ' << P;
281     else
282       dbgs() << " -";
283   }
284   dbgs() << " }\n";
285   dbgs() << "  Needed: {";
286   for (Node N : Needed)
287     dbgs() << ' ' << N;
288   dbgs() << " }\n";
289 
290   dbgs() << "  Edges: {\n";
291   for (auto E : Edges) {
292     dbgs() << "    " << E.first << " -> {";
293     for (auto N : E.second)
294       dbgs() << ' ' << N;
295     dbgs() << " }\n";
296   }
297   dbgs() << "  }\n";
298 
299   auto ColorKindToName = [](ColorKind C) {
300     switch (C) {
301     case ColorKind::None:
302       return "None";
303     case ColorKind::Red:
304       return "Red";
305     case ColorKind::Black:
306       return "Black";
307     }
308     llvm_unreachable("all ColorKinds should be handled by the switch above");
309   };
310 
311   dbgs() << "  Colors: {\n";
312   for (auto C : Colors)
313     dbgs() << "    " << C.first << " -> " << ColorKindToName(C.second) << "\n";
314   dbgs() << "  }\n}\n";
315 }
316 #endif
317 
318 namespace {
319 // Base class of for reordering networks. They don't strictly need to be
320 // permutations, as outputs with repeated occurrences of an input element
321 // are allowed.
322 struct PermNetwork {
323   using Controls = std::vector<uint8_t>;
324   using ElemType = int;
325   static constexpr ElemType Ignore = ElemType(-1);
326 
327   enum : uint8_t {
328     None,
329     Pass,
330     Switch
331   };
332   enum : uint8_t {
333     Forward,
334     Reverse
335   };
336 
337   PermNetwork(ArrayRef<ElemType> Ord, unsigned Mult = 1) {
338     Order.assign(Ord.data(), Ord.data()+Ord.size());
339     Log = 0;
340 
341     unsigned S = Order.size();
342     while (S >>= 1)
343       ++Log;
344 
345     Table.resize(Order.size());
346     for (RowType &Row : Table)
347       Row.resize(Mult*Log, None);
348   }
349 
350   void getControls(Controls &V, unsigned StartAt, uint8_t Dir) const {
351     unsigned Size = Order.size();
352     V.resize(Size);
353     for (unsigned I = 0; I != Size; ++I) {
354       unsigned W = 0;
355       for (unsigned L = 0; L != Log; ++L) {
356         unsigned C = ctl(I, StartAt+L) == Switch;
357         if (Dir == Forward)
358           W |= C << (Log-1-L);
359         else
360           W |= C << L;
361       }
362       assert(isUInt<8>(W));
363       V[I] = uint8_t(W);
364     }
365   }
366 
367   uint8_t ctl(ElemType Pos, unsigned Step) const {
368     return Table[Pos][Step];
369   }
370   unsigned size() const {
371     return Order.size();
372   }
373   unsigned steps() const {
374     return Log;
375   }
376 
377 protected:
378   unsigned Log;
379   std::vector<ElemType> Order;
380   using RowType = std::vector<uint8_t>;
381   std::vector<RowType> Table;
382 };
383 
384 struct ForwardDeltaNetwork : public PermNetwork {
385   ForwardDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
386 
387   bool run(Controls &V) {
388     if (!route(Order.data(), Table.data(), size(), 0))
389       return false;
390     getControls(V, 0, Forward);
391     return true;
392   }
393 
394 private:
395   bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
396 };
397 
398 struct ReverseDeltaNetwork : public PermNetwork {
399   ReverseDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
400 
401   bool run(Controls &V) {
402     if (!route(Order.data(), Table.data(), size(), 0))
403       return false;
404     getControls(V, 0, Reverse);
405     return true;
406   }
407 
408 private:
409   bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
410 };
411 
412 struct BenesNetwork : public PermNetwork {
413   BenesNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord, 2) {}
414 
415   bool run(Controls &F, Controls &R) {
416     if (!route(Order.data(), Table.data(), size(), 0))
417       return false;
418 
419     getControls(F, 0, Forward);
420     getControls(R, Log, Reverse);
421     return true;
422   }
423 
424 private:
425   bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
426 };
427 } // namespace
428 
429 bool ForwardDeltaNetwork::route(ElemType *P, RowType *T, unsigned Size,
430                                 unsigned Step) {
431   bool UseUp = false, UseDown = false;
432   ElemType Num = Size;
433 
434   // Cannot use coloring here, because coloring is used to determine
435   // the "big" switch, i.e. the one that changes halves, and in a forward
436   // network, a color can be simultaneously routed to both halves in the
437   // step we're working on.
438   for (ElemType J = 0; J != Num; ++J) {
439     ElemType I = P[J];
440     // I is the position in the input,
441     // J is the position in the output.
442     if (I == Ignore)
443       continue;
444     uint8_t S;
445     if (I < Num/2)
446       S = (J < Num/2) ? Pass : Switch;
447     else
448       S = (J < Num/2) ? Switch : Pass;
449 
450     // U is the element in the table that needs to be updated.
451     ElemType U = (S == Pass) ? I : (I < Num/2 ? I+Num/2 : I-Num/2);
452     if (U < Num/2)
453       UseUp = true;
454     else
455       UseDown = true;
456     if (T[U][Step] != S && T[U][Step] != None)
457       return false;
458     T[U][Step] = S;
459   }
460 
461   for (ElemType J = 0; J != Num; ++J)
462     if (P[J] != Ignore && P[J] >= Num/2)
463       P[J] -= Num/2;
464 
465   if (Step+1 < Log) {
466     if (UseUp   && !route(P,        T,        Size/2, Step+1))
467       return false;
468     if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
469       return false;
470   }
471   return true;
472 }
473 
474 bool ReverseDeltaNetwork::route(ElemType *P, RowType *T, unsigned Size,
475                                 unsigned Step) {
476   unsigned Pets = Log-1 - Step;
477   bool UseUp = false, UseDown = false;
478   ElemType Num = Size;
479 
480   // In this step half-switching occurs, so coloring can be used.
481   Coloring G({P,Size});
482   const Coloring::MapType &M = G.colors();
483   if (M.empty())
484     return false;
485 
486   ColorKind ColorUp = ColorKind::None;
487   for (ElemType J = 0; J != Num; ++J) {
488     ElemType I = P[J];
489     // I is the position in the input,
490     // J is the position in the output.
491     if (I == Ignore)
492       continue;
493     ColorKind C = M.at(I);
494     if (C == ColorKind::None)
495       continue;
496     // During "Step", inputs cannot switch halves, so if the "up" color
497     // is still unknown, make sure that it is selected in such a way that
498     // "I" will stay in the same half.
499     bool InpUp = I < Num/2;
500     if (ColorUp == ColorKind::None)
501       ColorUp = InpUp ? C : G.other(C);
502     if ((C == ColorUp) != InpUp) {
503       // If I should go to a different half than where is it now, give up.
504       return false;
505     }
506 
507     uint8_t S;
508     if (InpUp) {
509       S = (J < Num/2) ? Pass : Switch;
510       UseUp = true;
511     } else {
512       S = (J < Num/2) ? Switch : Pass;
513       UseDown = true;
514     }
515     T[J][Pets] = S;
516   }
517 
518   // Reorder the working permutation according to the computed switch table
519   // for the last step (i.e. Pets).
520   for (ElemType J = 0, E = Size / 2; J != E; ++J) {
521     ElemType PJ = P[J];         // Current values of P[J]
522     ElemType PC = P[J+Size/2];  // and P[conj(J)]
523     ElemType QJ = PJ;           // New values of P[J]
524     ElemType QC = PC;           // and P[conj(J)]
525     if (T[J][Pets] == Switch)
526       QC = PJ;
527     if (T[J+Size/2][Pets] == Switch)
528       QJ = PC;
529     P[J] = QJ;
530     P[J+Size/2] = QC;
531   }
532 
533   for (ElemType J = 0; J != Num; ++J)
534     if (P[J] != Ignore && P[J] >= Num/2)
535       P[J] -= Num/2;
536 
537   if (Step+1 < Log) {
538     if (UseUp && !route(P, T, Size/2, Step+1))
539       return false;
540     if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
541       return false;
542   }
543   return true;
544 }
545 
546 bool BenesNetwork::route(ElemType *P, RowType *T, unsigned Size,
547                          unsigned Step) {
548   Coloring G({P,Size});
549   const Coloring::MapType &M = G.colors();
550   if (M.empty())
551     return false;
552   ElemType Num = Size;
553 
554   unsigned Pets = 2*Log-1 - Step;
555   bool UseUp = false, UseDown = false;
556 
557   // Both assignments, i.e. Red->Up and Red->Down are valid, but they will
558   // result in different controls. Let's pick the one where the first
559   // control will be "Pass".
560   ColorKind ColorUp = ColorKind::None;
561   for (ElemType J = 0; J != Num; ++J) {
562     ElemType I = P[J];
563     if (I == Ignore)
564       continue;
565     ColorKind C = M.at(I);
566     if (C == ColorKind::None)
567       continue;
568     if (ColorUp == ColorKind::None) {
569       ColorUp = (I < Num / 2) ? ColorKind::Red : ColorKind::Black;
570     }
571     unsigned CI = (I < Num/2) ? I+Num/2 : I-Num/2;
572     if (C == ColorUp) {
573       if (I < Num/2)
574         T[I][Step] = Pass;
575       else
576         T[CI][Step] = Switch;
577       T[J][Pets] = (J < Num/2) ? Pass : Switch;
578       UseUp = true;
579     } else { // Down
580       if (I < Num/2)
581         T[CI][Step] = Switch;
582       else
583         T[I][Step] = Pass;
584       T[J][Pets] = (J < Num/2) ? Switch : Pass;
585       UseDown = true;
586     }
587   }
588 
589   // Reorder the working permutation according to the computed switch table
590   // for the last step (i.e. Pets).
591   for (ElemType J = 0; J != Num/2; ++J) {
592     ElemType PJ = P[J];         // Current values of P[J]
593     ElemType PC = P[J+Num/2];   // and P[conj(J)]
594     ElemType QJ = PJ;           // New values of P[J]
595     ElemType QC = PC;           // and P[conj(J)]
596     if (T[J][Pets] == Switch)
597       QC = PJ;
598     if (T[J+Num/2][Pets] == Switch)
599       QJ = PC;
600     P[J] = QJ;
601     P[J+Num/2] = QC;
602   }
603 
604   for (ElemType J = 0; J != Num; ++J)
605     if (P[J] != Ignore && P[J] >= Num/2)
606       P[J] -= Num/2;
607 
608   if (Step+1 < Log) {
609     if (UseUp && !route(P, T, Size/2, Step+1))
610       return false;
611     if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
612       return false;
613   }
614   return true;
615 }
616 
617 // --------------------------------------------------------------------
618 // Support for building selection results (output instructions that are
619 // parts of the final selection).
620 
621 namespace {
622 struct OpRef {
623   OpRef(SDValue V) : OpV(V) {}
624   bool isValue() const { return OpV.getNode() != nullptr; }
625   bool isValid() const { return isValue() || !(OpN & Invalid); }
626   bool isUndef() const { return OpN & Undef; }
627   static OpRef res(int N) { return OpRef(Whole | (N & Index)); }
628   static OpRef fail() { return OpRef(Invalid); }
629 
630   static OpRef lo(const OpRef &R) {
631     assert(!R.isValue());
632     return OpRef(R.OpN & (Undef | Index | LoHalf));
633   }
634   static OpRef hi(const OpRef &R) {
635     assert(!R.isValue());
636     return OpRef(R.OpN & (Undef | Index | HiHalf));
637   }
638   static OpRef undef(MVT Ty) { return OpRef(Undef | Ty.SimpleTy); }
639 
640   // Direct value.
641   SDValue OpV = SDValue();
642 
643   // Reference to the operand of the input node:
644   // If the 31st bit is 1, it's undef, otherwise, bits 28..0 are the
645   // operand index:
646   // If bit 30 is set, it's the high half of the operand.
647   // If bit 29 is set, it's the low half of the operand.
648   unsigned OpN = 0;
649 
650   enum : unsigned {
651     Invalid = 0x10000000,
652     LoHalf  = 0x20000000,
653     HiHalf  = 0x40000000,
654     Whole   = LoHalf | HiHalf,
655     Undef   = 0x80000000,
656     Index   = 0x0FFFFFFF,  // Mask of the index value.
657     IndexBits = 28,
658   };
659 
660   LLVM_DUMP_METHOD
661   void print(raw_ostream &OS, const SelectionDAG &G) const;
662 
663 private:
664   OpRef(unsigned N) : OpN(N) {}
665 };
666 
667 struct NodeTemplate {
668   NodeTemplate() = default;
669   unsigned Opc = 0;
670   MVT Ty = MVT::Other;
671   std::vector<OpRef> Ops;
672 
673   LLVM_DUMP_METHOD void print(raw_ostream &OS, const SelectionDAG &G) const;
674 };
675 
676 struct ResultStack {
677   ResultStack(SDNode *Inp)
678     : InpNode(Inp), InpTy(Inp->getValueType(0).getSimpleVT()) {}
679   SDNode *InpNode;
680   MVT InpTy;
681   unsigned push(const NodeTemplate &Res) {
682     List.push_back(Res);
683     return List.size()-1;
684   }
685   unsigned push(unsigned Opc, MVT Ty, std::vector<OpRef> &&Ops) {
686     NodeTemplate Res;
687     Res.Opc = Opc;
688     Res.Ty = Ty;
689     Res.Ops = Ops;
690     return push(Res);
691   }
692   bool empty() const { return List.empty(); }
693   unsigned size() const { return List.size(); }
694   unsigned top() const { return size()-1; }
695   const NodeTemplate &operator[](unsigned I) const { return List[I]; }
696   unsigned reset(unsigned NewTop) {
697     List.resize(NewTop+1);
698     return NewTop;
699   }
700 
701   using BaseType = std::vector<NodeTemplate>;
702   BaseType::iterator begin() { return List.begin(); }
703   BaseType::iterator end()   { return List.end(); }
704   BaseType::const_iterator begin() const { return List.begin(); }
705   BaseType::const_iterator end() const   { return List.end(); }
706 
707   BaseType List;
708 
709   LLVM_DUMP_METHOD
710   void print(raw_ostream &OS, const SelectionDAG &G) const;
711 };
712 } // namespace
713 
714 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
715 void OpRef::print(raw_ostream &OS, const SelectionDAG &G) const {
716   if (isValue()) {
717     OpV.getNode()->print(OS, &G);
718     return;
719   }
720   if (OpN & Invalid) {
721     OS << "invalid";
722     return;
723   }
724   if (OpN & Undef) {
725     OS << "undef";
726     return;
727   }
728   if ((OpN & Whole) != Whole) {
729     assert((OpN & Whole) == LoHalf || (OpN & Whole) == HiHalf);
730     if (OpN & LoHalf)
731       OS << "lo ";
732     else
733       OS << "hi ";
734   }
735   OS << '#' << SignExtend32(OpN & Index, IndexBits);
736 }
737 
738 void NodeTemplate::print(raw_ostream &OS, const SelectionDAG &G) const {
739   const TargetInstrInfo &TII = *G.getSubtarget().getInstrInfo();
740   OS << format("%8s", EVT(Ty).getEVTString().c_str()) << "  "
741      << TII.getName(Opc);
742   bool Comma = false;
743   for (const auto &R : Ops) {
744     if (Comma)
745       OS << ',';
746     Comma = true;
747     OS << ' ';
748     R.print(OS, G);
749   }
750 }
751 
752 void ResultStack::print(raw_ostream &OS, const SelectionDAG &G) const {
753   OS << "Input node:\n";
754 #ifndef NDEBUG
755   InpNode->dumpr(&G);
756 #endif
757   OS << "Result templates:\n";
758   for (unsigned I = 0, E = List.size(); I != E; ++I) {
759     OS << '[' << I << "] ";
760     List[I].print(OS, G);
761     OS << '\n';
762   }
763 }
764 #endif
765 
766 namespace {
767 struct ShuffleMask {
768   ShuffleMask(ArrayRef<int> M) : Mask(M) {
769     for (int M : Mask) {
770       if (M == -1)
771         continue;
772       MinSrc = (MinSrc == -1) ? M : std::min(MinSrc, M);
773       MaxSrc = (MaxSrc == -1) ? M : std::max(MaxSrc, M);
774     }
775   }
776 
777   ArrayRef<int> Mask;
778   int MinSrc = -1, MaxSrc = -1;
779 
780   ShuffleMask lo() const {
781     size_t H = Mask.size()/2;
782     return ShuffleMask(Mask.take_front(H));
783   }
784   ShuffleMask hi() const {
785     size_t H = Mask.size()/2;
786     return ShuffleMask(Mask.take_back(H));
787   }
788 
789   void print(raw_ostream &OS) const {
790     OS << "MinSrc:" << MinSrc << ", MaxSrc:" << MaxSrc << " {";
791     for (int M : Mask)
792       OS << ' ' << M;
793     OS << " }";
794   }
795 };
796 
797 LLVM_ATTRIBUTE_UNUSED
798 raw_ostream &operator<<(raw_ostream &OS, const ShuffleMask &SM) {
799   SM.print(OS);
800   return OS;
801 }
802 } // namespace
803 
804 namespace shuffles {
805 using MaskT = SmallVector<int, 128>;
806 // Vdd = vshuffvdd(Vu, Vv, Rt)
807 // Vdd = vdealvdd(Vu, Vv, Rt)
808 // Vd  = vpack(Vu, Vv, Size, TakeOdd)
809 // Vd  = vshuff(Vu, Vv, Size, TakeOdd)
810 // Vd  = vdeal(Vu, Vv, Size, TakeOdd)
811 // Vd  = vdealb4w(Vu, Vv)
812 
813 ArrayRef<int> lo(ArrayRef<int> Vuu) { return Vuu.take_front(Vuu.size() / 2); }
814 ArrayRef<int> hi(ArrayRef<int> Vuu) { return Vuu.take_back(Vuu.size() / 2); }
815 
816 MaskT vshuffvdd(ArrayRef<int> Vu, ArrayRef<int> Vv, unsigned Rt) {
817   int Len = Vu.size();
818   MaskT Vdd(2 * Len);
819   std::copy(Vv.begin(), Vv.end(), Vdd.begin());
820   std::copy(Vu.begin(), Vu.end(), Vdd.begin() + Len);
821 
822   auto Vd0 = MutableArrayRef<int>(Vdd).take_front(Len);
823   auto Vd1 = MutableArrayRef<int>(Vdd).take_back(Len);
824 
825   for (int Offset = 1; Offset < Len; Offset *= 2) {
826     if ((Rt & Offset) == 0)
827       continue;
828     for (int i = 0; i != Len; ++i) {
829       if ((i & Offset) == 0)
830         std::swap(Vd1[i], Vd0[i + Offset]);
831     }
832   }
833   return Vdd;
834 }
835 
836 MaskT vdealvdd(ArrayRef<int> Vu, ArrayRef<int> Vv, unsigned Rt) {
837   int Len = Vu.size();
838   MaskT Vdd(2 * Len);
839   std::copy(Vv.begin(), Vv.end(), Vdd.begin());
840   std::copy(Vu.begin(), Vu.end(), Vdd.begin() + Len);
841 
842   auto Vd0 = MutableArrayRef<int>(Vdd).take_front(Len);
843   auto Vd1 = MutableArrayRef<int>(Vdd).take_back(Len);
844 
845   for (int Offset = Len / 2; Offset > 0; Offset /= 2) {
846     if ((Rt & Offset) == 0)
847       continue;
848     for (int i = 0; i != Len; ++i) {
849       if ((i & Offset) == 0)
850         std::swap(Vd1[i], Vd0[i + Offset]);
851     }
852   }
853   return Vdd;
854 }
855 
856 MaskT vpack(ArrayRef<int> Vu, ArrayRef<int> Vv, unsigned Size, bool TakeOdd) {
857   int Len = Vu.size();
858   MaskT Vd(Len);
859   auto Odd = static_cast<int>(TakeOdd);
860   for (int i = 0, e = Len / (2 * Size); i != e; ++i) {
861     for (int b = 0; b != static_cast<int>(Size); ++b) {
862       // clang-format off
863       Vd[i * Size + b]           = Vv[(2 * i + Odd) * Size + b];
864       Vd[i * Size + b + Len / 2] = Vu[(2 * i + Odd) * Size + b];
865       // clang-format on
866     }
867   }
868   return Vd;
869 }
870 
871 MaskT vshuff(ArrayRef<int> Vu, ArrayRef<int> Vv, unsigned Size, bool TakeOdd) {
872   int Len = Vu.size();
873   MaskT Vd(Len);
874   auto Odd = static_cast<int>(TakeOdd);
875   for (int i = 0, e = Len / (2 * Size); i != e; ++i) {
876     for (int b = 0; b != static_cast<int>(Size); ++b) {
877       Vd[(2 * i + 0) * Size + b] = Vv[(2 * i + Odd) * Size + b];
878       Vd[(2 * i + 1) * Size + b] = Vu[(2 * i + Odd) * Size + b];
879     }
880   }
881   return Vd;
882 }
883 
884 MaskT vdeal(ArrayRef<int> Vu, ArrayRef<int> Vv, unsigned Size, bool TakeOdd) {
885   int Len = Vu.size();
886   MaskT T = vdealvdd(Vu, Vv, Len - 2 * Size);
887   return vpack(hi(T), lo(T), Size, TakeOdd);
888 }
889 
890 MaskT vdealb4w(ArrayRef<int> Vu, ArrayRef<int> Vv) {
891   int Len = Vu.size();
892   MaskT Vd(Len);
893   for (int i = 0, e = Len / 4; i != e; ++i) {
894     Vd[0 * (Len / 4) + i] = Vv[4 * i + 0];
895     Vd[1 * (Len / 4) + i] = Vv[4 * i + 2];
896     Vd[2 * (Len / 4) + i] = Vu[4 * i + 0];
897     Vd[3 * (Len / 4) + i] = Vu[4 * i + 2];
898   }
899   return Vd;
900 }
901 
902 template <typename ShuffFunc, typename... OptArgs>
903 auto mask(ShuffFunc S, unsigned Length, OptArgs... args) -> MaskT {
904   MaskT Vu(Length), Vv(Length);
905   std::iota(Vu.begin(), Vu.end(), Length); // High
906   std::iota(Vv.begin(), Vv.end(), 0);      // Low
907   return S(Vu, Vv, args...);
908 }
909 
910 } // namespace shuffles
911 
912 // --------------------------------------------------------------------
913 // The HvxSelector class.
914 
915 static const HexagonTargetLowering &getHexagonLowering(SelectionDAG &G) {
916   return static_cast<const HexagonTargetLowering&>(G.getTargetLoweringInfo());
917 }
918 static const HexagonSubtarget &getHexagonSubtarget(SelectionDAG &G) {
919   return G.getSubtarget<HexagonSubtarget>();
920 }
921 
922 namespace llvm {
923   struct HvxSelector {
924     const HexagonTargetLowering &Lower;
925     HexagonDAGToDAGISel &ISel;
926     SelectionDAG &DAG;
927     const HexagonSubtarget &HST;
928     const unsigned HwLen;
929 
930     HvxSelector(HexagonDAGToDAGISel &HS, SelectionDAG &G)
931         : Lower(getHexagonLowering(G)), ISel(HS), DAG(G),
932           HST(getHexagonSubtarget(G)), HwLen(HST.getVectorLength()) {}
933 
934     MVT getSingleVT(MVT ElemTy) const {
935       assert(ElemTy != MVT::i1 && "Use getBoolVT for predicates");
936       unsigned NumElems = HwLen / (ElemTy.getSizeInBits() / 8);
937       return MVT::getVectorVT(ElemTy, NumElems);
938     }
939 
940     MVT getPairVT(MVT ElemTy) const {
941       assert(ElemTy != MVT::i1); // Suspicious: there are no predicate pairs.
942       unsigned NumElems = (2 * HwLen) / (ElemTy.getSizeInBits() / 8);
943       return MVT::getVectorVT(ElemTy, NumElems);
944     }
945 
946     MVT getBoolVT() const {
947       // Return HwLen x i1.
948       return MVT::getVectorVT(MVT::i1, HwLen);
949     }
950 
951     void selectExtractSubvector(SDNode *N);
952     void selectShuffle(SDNode *N);
953     void selectRor(SDNode *N);
954     void selectVAlign(SDNode *N);
955 
956     static SmallVector<uint32_t, 8> getPerfectCompletions(ShuffleMask SM,
957                                                           unsigned Width);
958     static SmallVector<uint32_t, 8> completeToPerfect(
959         ArrayRef<uint32_t> Completions, unsigned Width);
960     static std::optional<int> rotationDistance(ShuffleMask SM, unsigned WrapAt);
961 
962   private:
963     void select(SDNode *ISelN);
964     void materialize(const ResultStack &Results);
965 
966     SDValue getConst32(int Val, const SDLoc &dl);
967     SDValue getVectorConstant(ArrayRef<uint8_t> Data, const SDLoc &dl);
968 
969     enum : unsigned {
970       None,
971       PackMux,
972     };
973     OpRef concats(OpRef Va, OpRef Vb, ResultStack &Results);
974     OpRef funnels(OpRef Va, OpRef Vb, int Amount, ResultStack &Results);
975 
976     OpRef packs(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results,
977                 MutableArrayRef<int> NewMask, unsigned Options = None);
978     OpRef packp(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results,
979                 MutableArrayRef<int> NewMask);
980     OpRef vmuxs(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
981                 ResultStack &Results);
982     OpRef vmuxp(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
983                 ResultStack &Results);
984 
985     OpRef shuffs1(ShuffleMask SM, OpRef Va, ResultStack &Results);
986     OpRef shuffs2(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
987     OpRef shuffp1(ShuffleMask SM, OpRef Va, ResultStack &Results);
988     OpRef shuffp2(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
989 
990     OpRef butterfly(ShuffleMask SM, OpRef Va, ResultStack &Results);
991     OpRef contracting(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
992     OpRef expanding(ShuffleMask SM, OpRef Va, ResultStack &Results);
993     OpRef perfect(ShuffleMask SM, OpRef Va, ResultStack &Results);
994 
995     bool selectVectorConstants(SDNode *N);
996     bool scalarizeShuffle(ArrayRef<int> Mask, const SDLoc &dl, MVT ResTy,
997                           SDValue Va, SDValue Vb, SDNode *N);
998   };
999 } // namespace llvm
1000 
1001 static void splitMask(ArrayRef<int> Mask, MutableArrayRef<int> MaskL,
1002                       MutableArrayRef<int> MaskR) {
1003   unsigned VecLen = Mask.size();
1004   assert(MaskL.size() == VecLen && MaskR.size() == VecLen);
1005   for (unsigned I = 0; I != VecLen; ++I) {
1006     int M = Mask[I];
1007     if (M < 0) {
1008       MaskL[I] = MaskR[I] = -1;
1009     } else if (unsigned(M) < VecLen) {
1010       MaskL[I] = M;
1011       MaskR[I] = -1;
1012     } else {
1013       MaskL[I] = -1;
1014       MaskR[I] = M-VecLen;
1015     }
1016   }
1017 }
1018 
1019 static std::pair<int,unsigned> findStrip(ArrayRef<int> A, int Inc,
1020                                          unsigned MaxLen) {
1021   assert(A.size() > 0 && A.size() >= MaxLen);
1022   int F = A[0];
1023   int E = F;
1024   for (unsigned I = 1; I != MaxLen; ++I) {
1025     if (A[I] - E != Inc)
1026       return { F, I };
1027     E = A[I];
1028   }
1029   return { F, MaxLen };
1030 }
1031 
1032 static bool isUndef(ArrayRef<int> Mask) {
1033   for (int Idx : Mask)
1034     if (Idx != -1)
1035       return false;
1036   return true;
1037 }
1038 
1039 static bool isIdentity(ArrayRef<int> Mask) {
1040   for (int I = 0, E = Mask.size(); I != E; ++I) {
1041     int M = Mask[I];
1042     if (M >= 0 && M != I)
1043       return false;
1044   }
1045   return true;
1046 }
1047 
1048 static bool isLowHalfOnly(ArrayRef<int> Mask) {
1049   int L = Mask.size();
1050   assert(L % 2 == 0);
1051   // Check if the second half of the mask is all-undef.
1052   return llvm::all_of(Mask.drop_front(L / 2), [](int M) { return M < 0; });
1053 }
1054 
1055 static SmallVector<unsigned, 4> getInputSegmentList(ShuffleMask SM,
1056                                                     unsigned SegLen) {
1057   assert(isPowerOf2_32(SegLen));
1058   SmallVector<unsigned, 4> SegList;
1059   if (SM.MaxSrc == -1)
1060     return SegList;
1061 
1062   unsigned Shift = Log2_32(SegLen);
1063   BitVector Segs(alignTo(SM.MaxSrc + 1, SegLen) >> Shift);
1064 
1065   for (int M : SM.Mask) {
1066     if (M >= 0)
1067       Segs.set(M >> Shift);
1068   }
1069 
1070   for (unsigned B : Segs.set_bits())
1071     SegList.push_back(B);
1072   return SegList;
1073 }
1074 
1075 static SmallVector<unsigned, 4> getOutputSegmentMap(ShuffleMask SM,
1076                                                     unsigned SegLen) {
1077   // Calculate the layout of the output segments in terms of the input
1078   // segments.
1079   // For example [1,3,1,0] means that the output consists of 4 output
1080   // segments, where the first output segment has only elements of the
1081   // input segment at index 1. The next output segment only has elements
1082   // of the input segment 3, etc.
1083   // If an output segment only has undef elements, the value will be ~0u.
1084   // If an output segment has elements from more than one input segment,
1085   // the corresponding value will be ~1u.
1086   unsigned MaskLen = SM.Mask.size();
1087   assert(MaskLen % SegLen == 0);
1088   SmallVector<unsigned, 4> Map(MaskLen / SegLen);
1089 
1090   for (int S = 0, E = Map.size(); S != E; ++S) {
1091     unsigned Idx = ~0u;
1092     for (int I = 0; I != static_cast<int>(SegLen); ++I) {
1093       int M = SM.Mask[S*SegLen + I];
1094       if (M < 0)
1095         continue;
1096       unsigned G = M / SegLen; // Input segment of this element.
1097       if (Idx == ~0u) {
1098         Idx = G;
1099       } else if (Idx != G) {
1100         Idx = ~1u;
1101         break;
1102       }
1103     }
1104     Map[S] = Idx;
1105   }
1106 
1107   return Map;
1108 }
1109 
1110 static void packSegmentMask(ArrayRef<int> Mask, ArrayRef<unsigned> OutSegMap,
1111                             unsigned SegLen, MutableArrayRef<int> PackedMask) {
1112   SmallVector<unsigned, 4> InvMap;
1113   for (int I = OutSegMap.size() - 1; I >= 0; --I) {
1114     unsigned S = OutSegMap[I];
1115     assert(S != ~0u && "Unexpected undef");
1116     assert(S != ~1u && "Unexpected multi");
1117     if (InvMap.size() <= S)
1118       InvMap.resize(S+1);
1119     InvMap[S] = I;
1120   }
1121 
1122   unsigned Shift = Log2_32(SegLen);
1123   for (int I = 0, E = Mask.size(); I != E; ++I) {
1124     int M = Mask[I];
1125     if (M >= 0) {
1126       int OutIdx = InvMap[M >> Shift];
1127       M = (M & (SegLen-1)) + SegLen*OutIdx;
1128     }
1129     PackedMask[I] = M;
1130   }
1131 }
1132 
1133 bool HvxSelector::selectVectorConstants(SDNode *N) {
1134   // Constant vectors are generated as loads from constant pools or as
1135   // splats of a constant value. Since they are generated during the
1136   // selection process, the main selection algorithm is not aware of them.
1137   // Select them directly here.
1138   SmallVector<SDNode*,4> Nodes;
1139   SetVector<SDNode*> WorkQ;
1140 
1141   // The DAG can change (due to CSE) during selection, so cache all the
1142   // unselected nodes first to avoid traversing a mutating DAG.
1143   WorkQ.insert(N);
1144   for (unsigned i = 0; i != WorkQ.size(); ++i) {
1145     SDNode *W = WorkQ[i];
1146     if (!W->isMachineOpcode() && W->getOpcode() == HexagonISD::ISEL)
1147       Nodes.push_back(W);
1148     for (unsigned j = 0, f = W->getNumOperands(); j != f; ++j)
1149       WorkQ.insert(W->getOperand(j).getNode());
1150   }
1151 
1152   for (SDNode *L : Nodes)
1153     select(L);
1154 
1155   return !Nodes.empty();
1156 }
1157 
1158 void HvxSelector::materialize(const ResultStack &Results) {
1159   DEBUG_WITH_TYPE("isel", {
1160     dbgs() << "Materializing\n";
1161     Results.print(dbgs(), DAG);
1162   });
1163   if (Results.empty())
1164     return;
1165   const SDLoc &dl(Results.InpNode);
1166   std::vector<SDValue> Output;
1167 
1168   for (unsigned I = 0, E = Results.size(); I != E; ++I) {
1169     const NodeTemplate &Node = Results[I];
1170     std::vector<SDValue> Ops;
1171     for (const OpRef &R : Node.Ops) {
1172       assert(R.isValid());
1173       if (R.isValue()) {
1174         Ops.push_back(R.OpV);
1175         continue;
1176       }
1177       if (R.OpN & OpRef::Undef) {
1178         MVT::SimpleValueType SVT = MVT::SimpleValueType(R.OpN & OpRef::Index);
1179         Ops.push_back(ISel.selectUndef(dl, MVT(SVT)));
1180         continue;
1181       }
1182       // R is an index of a result.
1183       unsigned Part = R.OpN & OpRef::Whole;
1184       int Idx = SignExtend32(R.OpN & OpRef::Index, OpRef::IndexBits);
1185       if (Idx < 0)
1186         Idx += I;
1187       assert(Idx >= 0 && unsigned(Idx) < Output.size());
1188       SDValue Op = Output[Idx];
1189       MVT OpTy = Op.getValueType().getSimpleVT();
1190       if (Part != OpRef::Whole) {
1191         assert(Part == OpRef::LoHalf || Part == OpRef::HiHalf);
1192         MVT HalfTy = MVT::getVectorVT(OpTy.getVectorElementType(),
1193                                       OpTy.getVectorNumElements()/2);
1194         unsigned Sub = (Part == OpRef::LoHalf) ? Hexagon::vsub_lo
1195                                                : Hexagon::vsub_hi;
1196         Op = DAG.getTargetExtractSubreg(Sub, dl, HalfTy, Op);
1197       }
1198       Ops.push_back(Op);
1199     } // for (Node : Results)
1200 
1201     assert(Node.Ty != MVT::Other);
1202     SDNode *ResN = (Node.Opc == TargetOpcode::COPY)
1203                       ? Ops.front().getNode()
1204                       : DAG.getMachineNode(Node.Opc, dl, Node.Ty, Ops);
1205     Output.push_back(SDValue(ResN, 0));
1206   }
1207 
1208   SDNode *OutN = Output.back().getNode();
1209   SDNode *InpN = Results.InpNode;
1210   DEBUG_WITH_TYPE("isel", {
1211     dbgs() << "Generated node:\n";
1212     OutN->dumpr(&DAG);
1213   });
1214 
1215   ISel.ReplaceNode(InpN, OutN);
1216   selectVectorConstants(OutN);
1217   DAG.RemoveDeadNodes();
1218 }
1219 
1220 OpRef HvxSelector::concats(OpRef Lo, OpRef Hi, ResultStack &Results) {
1221   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1222   const SDLoc &dl(Results.InpNode);
1223   Results.push(TargetOpcode::REG_SEQUENCE, getPairVT(MVT::i8), {
1224     getConst32(Hexagon::HvxWRRegClassID, dl),
1225     Lo, getConst32(Hexagon::vsub_lo, dl),
1226     Hi, getConst32(Hexagon::vsub_hi, dl),
1227   });
1228   return OpRef::res(Results.top());
1229 }
1230 
1231 OpRef HvxSelector::funnels(OpRef Va, OpRef Vb, int Amount,
1232                            ResultStack &Results) {
1233   // Do a funnel shift towards the low end (shift right) by Amount bytes.
1234   // If Amount < 0, treat it as shift left, i.e. do a shift right by
1235   // Amount + HwLen.
1236   auto VecLen = static_cast<int>(HwLen);
1237 
1238   if (Amount == 0)
1239     return Va;
1240   if (Amount == VecLen)
1241     return Vb;
1242 
1243   MVT Ty = getSingleVT(MVT::i8);
1244   const SDLoc &dl(Results.InpNode);
1245 
1246   if (Amount < 0)
1247     Amount += VecLen;
1248   if (Amount > VecLen) {
1249     Amount -= VecLen;
1250     std::swap(Va, Vb);
1251   }
1252 
1253   if (isUInt<3>(Amount)) {
1254     SDValue A = getConst32(Amount, dl);
1255     Results.push(Hexagon::V6_valignbi, Ty, {Vb, Va, A});
1256   } else if (isUInt<3>(VecLen - Amount)) {
1257     SDValue A = getConst32(VecLen - Amount, dl);
1258     Results.push(Hexagon::V6_vlalignbi, Ty, {Vb, Va, A});
1259   } else {
1260     SDValue A = getConst32(Amount, dl);
1261     Results.push(Hexagon::A2_tfrsi, Ty, {A});
1262     Results.push(Hexagon::V6_valignb, Ty, {Vb, Va, OpRef::res(-1)});
1263   }
1264   return OpRef::res(Results.top());
1265 }
1266 
1267 // Va, Vb are single vectors. If SM only uses two vector halves from Va/Vb,
1268 // pack these halves into a single vector, and remap SM into NewMask to use
1269 // the new vector instead.
1270 OpRef HvxSelector::packs(ShuffleMask SM, OpRef Va, OpRef Vb,
1271                          ResultStack &Results, MutableArrayRef<int> NewMask,
1272                          unsigned Options) {
1273   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1274   if (!Va.isValid() || !Vb.isValid())
1275     return OpRef::fail();
1276 
1277   if (Vb.isUndef()) {
1278     std::copy(SM.Mask.begin(), SM.Mask.end(), NewMask.begin());
1279     return Va;
1280   }
1281   if (Va.isUndef()) {
1282     std::copy(SM.Mask.begin(), SM.Mask.end(), NewMask.begin());
1283     ShuffleVectorSDNode::commuteMask(NewMask);
1284     return Vb;
1285   }
1286 
1287   MVT Ty = getSingleVT(MVT::i8);
1288   MVT PairTy = getPairVT(MVT::i8);
1289   OpRef Inp[2] = {Va, Vb};
1290   unsigned VecLen = SM.Mask.size();
1291 
1292   auto valign = [this](OpRef Lo, OpRef Hi, unsigned Amt, MVT Ty,
1293                        ResultStack &Results) {
1294     if (Amt == 0)
1295       return Lo;
1296     const SDLoc &dl(Results.InpNode);
1297     if (isUInt<3>(Amt) || isUInt<3>(HwLen - Amt)) {
1298       bool IsRight = isUInt<3>(Amt); // Right align.
1299       SDValue S = getConst32(IsRight ? Amt : HwLen - Amt, dl);
1300       unsigned Opc = IsRight ? Hexagon::V6_valignbi : Hexagon::V6_vlalignbi;
1301       Results.push(Opc, Ty, {Hi, Lo, S});
1302       return OpRef::res(Results.top());
1303     }
1304     Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(Amt, dl)});
1305     OpRef A = OpRef::res(Results.top());
1306     Results.push(Hexagon::V6_valignb, Ty, {Hi, Lo, A});
1307     return OpRef::res(Results.top());
1308   };
1309 
1310   // Segment is a vector half.
1311   unsigned SegLen = HwLen / 2;
1312 
1313   // Check if we can shuffle vector halves around to get the used elements
1314   // into a single vector.
1315   shuffles::MaskT MaskH(SM.Mask);
1316   SmallVector<unsigned, 4> SegList = getInputSegmentList(SM.Mask, SegLen);
1317   unsigned SegCount = SegList.size();
1318   SmallVector<unsigned, 4> SegMap = getOutputSegmentMap(SM.Mask, SegLen);
1319 
1320   if (SegList.empty())
1321     return OpRef::undef(Ty);
1322 
1323   // NOTE:
1324   // In the following part of the function, where the segments are rearranged,
1325   // the shuffle mask SM can be of any length that is a multiple of a vector
1326   // (i.e. a multiple of 2*SegLen), and non-zero.
1327   // The output segment map is computed, and it may have any even number of
1328   // entries, but the rearrangement of input segments will be done based only
1329   // on the first two (non-undef) entries in the segment map.
1330   // For example, if the output map is 3, 1, 1, 3 (it can have at most two
1331   // distinct entries!), the segments 1 and 3 of Va/Vb will be packaged into
1332   // a single vector V = 3:1. The output mask will then be updated to use
1333   // seg(0,V), seg(1,V), seg(1,V), seg(0,V).
1334   //
1335   // Picking the segments based on the output map is an optimization. For
1336   // correctness it is only necessary that Seg0 and Seg1 are the two input
1337   // segments that are used in the output.
1338 
1339   unsigned Seg0 = ~0u, Seg1 = ~0u;
1340   for (unsigned X : SegMap) {
1341     if (X == ~0u)
1342       continue;
1343     if (Seg0 == ~0u)
1344       Seg0 = X;
1345     else if (Seg1 != ~0u)
1346       break;
1347     if (X == ~1u || X != Seg0)
1348       Seg1 = X;
1349   }
1350 
1351   if (SegCount == 1) {
1352     unsigned SrcOp = SegList[0] / 2;
1353     for (int I = 0; I != static_cast<int>(VecLen); ++I) {
1354       int M = SM.Mask[I];
1355       if (M >= 0) {
1356         M -= SrcOp * HwLen;
1357         assert(M >= 0);
1358       }
1359       NewMask[I] = M;
1360     }
1361     return Inp[SrcOp];
1362   }
1363 
1364   if (SegCount == 2) {
1365     // Seg0 should not be undef here: this would imply a SegList
1366     // with <= 1 elements, which was checked earlier.
1367     assert(Seg0 != ~0u);
1368 
1369     // If Seg0 or Seg1 are "multi-defined", pick them from the input
1370     // segment list instead.
1371     if (Seg0 == ~1u || Seg1 == ~1u) {
1372       if (Seg0 == Seg1) {
1373         Seg0 = SegList[0];
1374         Seg1 = SegList[1];
1375       } else if (Seg0 == ~1u) {
1376         Seg0 = SegList[0] != Seg1 ? SegList[0] : SegList[1];
1377       } else {
1378         assert(Seg1 == ~1u);
1379         Seg1 = SegList[0] != Seg0 ? SegList[0] : SegList[1];
1380       }
1381     }
1382     assert(Seg0 != ~1u && Seg1 != ~1u);
1383 
1384     assert(Seg0 != Seg1 && "Expecting different segments");
1385     const SDLoc &dl(Results.InpNode);
1386     Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(SegLen, dl)});
1387     OpRef HL = OpRef::res(Results.top());
1388 
1389     // Va = AB, Vb = CD
1390 
1391     if (Seg0 / 2 == Seg1 / 2) {
1392       // Same input vector.
1393       Va = Inp[Seg0 / 2];
1394       if (Seg0 > Seg1) {
1395         // Swap halves.
1396         Results.push(Hexagon::V6_vror, Ty, {Inp[Seg0 / 2], HL});
1397         Va = OpRef::res(Results.top());
1398       }
1399       packSegmentMask(SM.Mask, {Seg0, Seg1}, SegLen, MaskH);
1400     } else if (Seg0 % 2 == Seg1 % 2) {
1401       // Picking AC, BD, CA, or DB.
1402       // vshuff(CD,AB,HL) -> BD:AC
1403       // vshuff(AB,CD,HL) -> DB:CA
1404       auto Vs = (Seg0 == 0 || Seg0 == 1) ? std::make_pair(Vb, Va)  // AC or BD
1405                                          : std::make_pair(Va, Vb); // CA or DB
1406       Results.push(Hexagon::V6_vshuffvdd, PairTy, {Vs.first, Vs.second, HL});
1407       OpRef P = OpRef::res(Results.top());
1408       Va = (Seg0 == 0 || Seg0 == 2) ? OpRef::lo(P) : OpRef::hi(P);
1409       packSegmentMask(SM.Mask, {Seg0, Seg1}, SegLen, MaskH);
1410     } else {
1411       // Picking AD, BC, CB, or DA.
1412       if ((Seg0 == 0 && Seg1 == 3) || (Seg0 == 2 && Seg1 == 1)) {
1413         // AD or BC: this can be done using vmux.
1414         // Q = V6_pred_scalar2 SegLen
1415         // V = V6_vmux Q, (Va, Vb) or (Vb, Va)
1416         Results.push(Hexagon::V6_pred_scalar2, getBoolVT(), {HL});
1417         OpRef Qt = OpRef::res(Results.top());
1418         auto Vs = (Seg0 == 0) ? std::make_pair(Va, Vb)  // AD
1419                               : std::make_pair(Vb, Va); // CB
1420         Results.push(Hexagon::V6_vmux, Ty, {Qt, Vs.first, Vs.second});
1421         Va = OpRef::res(Results.top());
1422         packSegmentMask(SM.Mask, {Seg0, Seg1}, SegLen, MaskH);
1423       } else {
1424         // BC or DA: this could be done via valign by SegLen.
1425         // Do nothing here, because valign (if possible) will be generated
1426         // later on (make sure the Seg0 values are as expected).
1427         assert(Seg0 == 1 || Seg0 == 3);
1428       }
1429     }
1430   }
1431 
1432   // Check if the arguments can be packed by valign(Va,Vb) or valign(Vb,Va).
1433   // FIXME: maybe remove this?
1434   ShuffleMask SMH(MaskH);
1435   assert(SMH.Mask.size() == VecLen);
1436   shuffles::MaskT MaskA(SMH.Mask);
1437 
1438   if (SMH.MaxSrc - SMH.MinSrc >= static_cast<int>(HwLen)) {
1439     // valign(Lo=Va,Hi=Vb) won't work. Try swapping Va/Vb.
1440     shuffles::MaskT Swapped(SMH.Mask);
1441     ShuffleVectorSDNode::commuteMask(Swapped);
1442     ShuffleMask SW(Swapped);
1443     if (SW.MaxSrc - SW.MinSrc < static_cast<int>(HwLen)) {
1444       MaskA.assign(SW.Mask.begin(), SW.Mask.end());
1445       std::swap(Va, Vb);
1446     }
1447   }
1448   ShuffleMask SMA(MaskA);
1449   assert(SMA.Mask.size() == VecLen);
1450 
1451   if (SMA.MaxSrc - SMA.MinSrc < static_cast<int>(HwLen)) {
1452     int ShiftR = SMA.MinSrc;
1453     if (ShiftR >= static_cast<int>(HwLen)) {
1454       Va = Vb;
1455       Vb = OpRef::undef(Ty);
1456       ShiftR -= HwLen;
1457     }
1458     OpRef RetVal = valign(Va, Vb, ShiftR, Ty, Results);
1459 
1460     for (int I = 0; I != static_cast<int>(VecLen); ++I) {
1461       int M = SMA.Mask[I];
1462       if (M != -1)
1463         M -= SMA.MinSrc;
1464       NewMask[I] = M;
1465     }
1466     return RetVal;
1467   }
1468 
1469   // By here, packing by segment (half-vector) shuffling, and vector alignment
1470   // failed. Try vmux.
1471   // Note: since this is using the original mask, Va and Vb must not have been
1472   // modified.
1473 
1474   if (Options & PackMux) {
1475     // If elements picked from Va and Vb have all different (source) indexes
1476     // (relative to the start of the argument), do a mux, and update the mask.
1477     BitVector Picked(HwLen);
1478     SmallVector<uint8_t,128> MuxBytes(HwLen);
1479     bool CanMux = true;
1480     for (int I = 0; I != static_cast<int>(VecLen); ++I) {
1481       int M = SM.Mask[I];
1482       if (M == -1)
1483         continue;
1484       if (M >= static_cast<int>(HwLen))
1485         M -= HwLen;
1486       else
1487         MuxBytes[M] = 0xFF;
1488       if (Picked[M]) {
1489         CanMux = false;
1490         break;
1491       }
1492       NewMask[I] = M;
1493     }
1494     if (CanMux)
1495       return vmuxs(MuxBytes, Va, Vb, Results);
1496   }
1497   return OpRef::fail();
1498 }
1499 
1500 // Va, Vb are vector pairs. If SM only uses two single vectors from Va/Vb,
1501 // pack these vectors into a pair, and remap SM into NewMask to use the
1502 // new pair instead.
1503 OpRef HvxSelector::packp(ShuffleMask SM, OpRef Va, OpRef Vb,
1504                          ResultStack &Results, MutableArrayRef<int> NewMask) {
1505   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1506   SmallVector<unsigned, 4> SegList = getInputSegmentList(SM.Mask, HwLen);
1507   if (SegList.empty())
1508     return OpRef::undef(getPairVT(MVT::i8));
1509 
1510   // If more than two halves are used, bail.
1511   // TODO: be more aggressive here?
1512   unsigned SegCount = SegList.size();
1513   if (SegCount > 2)
1514     return OpRef::fail();
1515 
1516   MVT HalfTy = getSingleVT(MVT::i8);
1517 
1518   OpRef Inp[2] = { Va, Vb };
1519   OpRef Out[2] = { OpRef::undef(HalfTy), OpRef::undef(HalfTy) };
1520 
1521   // Really make sure we have at most 2 vectors used in the mask.
1522   assert(SegCount <= 2);
1523 
1524   for (int I = 0, E = SegList.size(); I != E; ++I) {
1525     unsigned S = SegList[I];
1526     OpRef Op = Inp[S / 2];
1527     Out[I] = (S & 1) ? OpRef::hi(Op) : OpRef::lo(Op);
1528   }
1529 
1530   // NOTE: Using SegList as the packing map here (not SegMap). This works,
1531   // because we're not concerned here about the order of the segments (i.e.
1532   // single vectors) in the output pair. Changing the order of vectors is
1533   // free (as opposed to changing the order of vector halves as in packs),
1534   // and so there is no extra cost added in case the order needs to be
1535   // changed later.
1536   packSegmentMask(SM.Mask, SegList, HwLen, NewMask);
1537   return concats(Out[0], Out[1], Results);
1538 }
1539 
1540 OpRef HvxSelector::vmuxs(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1541                          ResultStack &Results) {
1542   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1543   MVT ByteTy = getSingleVT(MVT::i8);
1544   MVT BoolTy = MVT::getVectorVT(MVT::i1, HwLen);
1545   const SDLoc &dl(Results.InpNode);
1546   SDValue B = getVectorConstant(Bytes, dl);
1547   Results.push(Hexagon::V6_vd0, ByteTy, {});
1548   Results.push(Hexagon::V6_veqb, BoolTy, {OpRef(B), OpRef::res(-1)});
1549   Results.push(Hexagon::V6_vmux, ByteTy, {OpRef::res(-1), Vb, Va});
1550   return OpRef::res(Results.top());
1551 }
1552 
1553 OpRef HvxSelector::vmuxp(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1554                          ResultStack &Results) {
1555   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1556   size_t S = Bytes.size() / 2;
1557   OpRef L = vmuxs(Bytes.take_front(S), OpRef::lo(Va), OpRef::lo(Vb), Results);
1558   OpRef H = vmuxs(Bytes.drop_front(S), OpRef::hi(Va), OpRef::hi(Vb), Results);
1559   return concats(L, H, Results);
1560 }
1561 
1562 OpRef HvxSelector::shuffs1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1563   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1564   unsigned VecLen = SM.Mask.size();
1565   assert(HwLen == VecLen);
1566   (void)VecLen;
1567   assert(all_of(SM.Mask, [this](int M) { return M == -1 || M < int(HwLen); }));
1568 
1569   if (isIdentity(SM.Mask))
1570     return Va;
1571   if (isUndef(SM.Mask))
1572     return OpRef::undef(getSingleVT(MVT::i8));
1573 
1574   // First, check for rotations.
1575   if (auto Dist = rotationDistance(SM, VecLen)) {
1576     OpRef Rotate = funnels(Va, Va, *Dist, Results);
1577     if (Rotate.isValid())
1578       return Rotate;
1579   }
1580   unsigned HalfLen = HwLen / 2;
1581   assert(isPowerOf2_32(HalfLen));
1582 
1583   // Handle special case where the output is the same half of the input
1584   // repeated twice, i.e. if Va = AB, then handle the output of AA or BB.
1585   std::pair<int, unsigned> Strip1 = findStrip(SM.Mask, 1, HalfLen);
1586   if ((Strip1.first & ~HalfLen) == 0 && Strip1.second == HalfLen) {
1587     std::pair<int, unsigned> Strip2 =
1588         findStrip(SM.Mask.drop_front(HalfLen), 1, HalfLen);
1589     if (Strip1 == Strip2) {
1590       const SDLoc &dl(Results.InpNode);
1591       Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(HalfLen, dl)});
1592       Results.push(Hexagon::V6_vshuffvdd, getPairVT(MVT::i8),
1593                    {Va, Va, OpRef::res(Results.top())});
1594       OpRef S = OpRef::res(Results.top());
1595       return (Strip1.first == 0) ? OpRef::lo(S) : OpRef::hi(S);
1596     }
1597   }
1598 
1599   OpRef P = perfect(SM, Va, Results);
1600   if (P.isValid())
1601     return P;
1602   return butterfly(SM, Va, Results);
1603 }
1604 
1605 OpRef HvxSelector::shuffs2(ShuffleMask SM, OpRef Va, OpRef Vb,
1606                            ResultStack &Results) {
1607   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1608   if (isUndef(SM.Mask))
1609     return OpRef::undef(getSingleVT(MVT::i8));
1610 
1611   OpRef C = contracting(SM, Va, Vb, Results);
1612   if (C.isValid())
1613     return C;
1614 
1615   int VecLen = SM.Mask.size();
1616   shuffles::MaskT PackedMask(VecLen);
1617   OpRef P = packs(SM, Va, Vb, Results, PackedMask);
1618   if (P.isValid())
1619     return shuffs1(ShuffleMask(PackedMask), P, Results);
1620 
1621   // TODO: Before we split the mask, try perfect shuffle on concatenated
1622   // operands.
1623 
1624   shuffles::MaskT MaskL(VecLen), MaskR(VecLen);
1625   splitMask(SM.Mask, MaskL, MaskR);
1626 
1627   OpRef L = shuffs1(ShuffleMask(MaskL), Va, Results);
1628   OpRef R = shuffs1(ShuffleMask(MaskR), Vb, Results);
1629   if (!L.isValid() || !R.isValid())
1630     return OpRef::fail();
1631 
1632   SmallVector<uint8_t, 128> Bytes(VecLen);
1633   for (int I = 0; I != VecLen; ++I) {
1634     if (MaskL[I] != -1)
1635       Bytes[I] = 0xFF;
1636   }
1637   return vmuxs(Bytes, L, R, Results);
1638 }
1639 
1640 OpRef HvxSelector::shuffp1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1641   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1642   int VecLen = SM.Mask.size();
1643 
1644   if (isIdentity(SM.Mask))
1645     return Va;
1646   if (isUndef(SM.Mask))
1647     return OpRef::undef(getPairVT(MVT::i8));
1648 
1649   shuffles::MaskT PackedMask(VecLen);
1650   OpRef P = packs(SM, OpRef::lo(Va), OpRef::hi(Va), Results, PackedMask);
1651   if (P.isValid()) {
1652     ShuffleMask PM(PackedMask);
1653     OpRef E = expanding(PM, P, Results);
1654     if (E.isValid())
1655       return E;
1656 
1657     OpRef L = shuffs1(PM.lo(), P, Results);
1658     OpRef H = shuffs1(PM.hi(), P, Results);
1659     if (L.isValid() && H.isValid())
1660       return concats(L, H, Results);
1661   }
1662 
1663   if (!isLowHalfOnly(SM.Mask)) {
1664     // Doing a perfect shuffle on a low-half mask (i.e. where the upper half
1665     // is all-undef) may produce a perfect shuffle that generates legitimate
1666     // upper half. This isn't wrong, but if the perfect shuffle was possible,
1667     // then there is a good chance that a shorter (contracting) code may be
1668     // used as well (e.g. V6_vshuffeb, etc).
1669     OpRef R = perfect(SM, Va, Results);
1670     if (R.isValid())
1671       return R;
1672     // TODO commute the mask and try the opposite order of the halves.
1673   }
1674 
1675   OpRef L = shuffs2(SM.lo(), OpRef::lo(Va), OpRef::hi(Va), Results);
1676   OpRef H = shuffs2(SM.hi(), OpRef::lo(Va), OpRef::hi(Va), Results);
1677   if (L.isValid() && H.isValid())
1678     return concats(L, H, Results);
1679 
1680   return OpRef::fail();
1681 }
1682 
1683 OpRef HvxSelector::shuffp2(ShuffleMask SM, OpRef Va, OpRef Vb,
1684                            ResultStack &Results) {
1685   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1686   if (isUndef(SM.Mask))
1687     return OpRef::undef(getPairVT(MVT::i8));
1688 
1689   int VecLen = SM.Mask.size();
1690   SmallVector<int,256> PackedMask(VecLen);
1691   OpRef P = packp(SM, Va, Vb, Results, PackedMask);
1692   if (P.isValid())
1693     return shuffp1(ShuffleMask(PackedMask), P, Results);
1694 
1695   SmallVector<int,256> MaskL(VecLen), MaskR(VecLen);
1696   splitMask(SM.Mask, MaskL, MaskR);
1697 
1698   OpRef L = shuffp1(ShuffleMask(MaskL), Va, Results);
1699   OpRef R = shuffp1(ShuffleMask(MaskR), Vb, Results);
1700   if (!L.isValid() || !R.isValid())
1701     return OpRef::fail();
1702 
1703   // Mux the results.
1704   SmallVector<uint8_t,256> Bytes(VecLen);
1705   for (int I = 0; I != VecLen; ++I) {
1706     if (MaskL[I] != -1)
1707       Bytes[I] = 0xFF;
1708   }
1709   return vmuxp(Bytes, L, R, Results);
1710 }
1711 
1712 namespace {
1713   struct Deleter : public SelectionDAG::DAGNodeDeletedListener {
1714     template <typename T>
1715     Deleter(SelectionDAG &D, T &C)
1716       : SelectionDAG::DAGNodeDeletedListener(D, [&C] (SDNode *N, SDNode *E) {
1717                                                   C.erase(N);
1718                                                 }) {}
1719   };
1720 
1721   template <typename T>
1722   struct NullifyingVector : public T {
1723     DenseMap<SDNode*, SDNode**> Refs;
1724     NullifyingVector(T &&V) : T(V) {
1725       for (unsigned i = 0, e = T::size(); i != e; ++i) {
1726         SDNode *&N = T::operator[](i);
1727         Refs[N] = &N;
1728       }
1729     }
1730     void erase(SDNode *N) {
1731       auto F = Refs.find(N);
1732       if (F != Refs.end())
1733         *F->second = nullptr;
1734     }
1735   };
1736 }
1737 
1738 void HvxSelector::select(SDNode *ISelN) {
1739   // What's important here is to select the right set of nodes. The main
1740   // selection algorithm loops over nodes in a topological order, i.e. users
1741   // are visited before their operands.
1742   //
1743   // It is an error to have an unselected node with a selected operand, and
1744   // there is an assertion in the main selector code to enforce that.
1745   //
1746   // Such a situation could occur if we selected a node, which is both a
1747   // subnode of ISelN, and a subnode of an unrelated (and yet unselected)
1748   // node in the DAG.
1749   assert(ISelN->getOpcode() == HexagonISD::ISEL);
1750   SDNode *N0 = ISelN->getOperand(0).getNode();
1751 
1752   // There could have been nodes created (i.e. inserted into the DAG)
1753   // that are now dead. Remove them, in case they use any of the nodes
1754   // to select (and make them look shared).
1755   DAG.RemoveDeadNodes();
1756 
1757   SetVector<SDNode *> SubNodes;
1758 
1759   if (!N0->isMachineOpcode()) {
1760     // Don't want to select N0 if it's shared with another node, except if
1761     // it's shared with other ISELs.
1762     auto IsISelN = [](SDNode *T) { return T->getOpcode() == HexagonISD::ISEL; };
1763     if (llvm::all_of(N0->uses(), IsISelN))
1764       SubNodes.insert(N0);
1765   }
1766   if (SubNodes.empty()) {
1767     ISel.ReplaceNode(ISelN, N0);
1768     return;
1769   }
1770 
1771   // Need to manually select the nodes that are dominated by the ISEL. Other
1772   // nodes are reachable from the rest of the DAG, and so will be selected
1773   // by the DAG selection routine.
1774   SetVector<SDNode*> Dom, NonDom;
1775   Dom.insert(N0);
1776 
1777   auto IsDomRec = [&Dom, &NonDom] (SDNode *T, auto Rec) -> bool {
1778     if (Dom.count(T))
1779       return true;
1780     if (T->use_empty() || NonDom.count(T))
1781       return false;
1782     for (SDNode *U : T->uses()) {
1783       // If T is reachable from a known non-dominated node, then T itself
1784       // is non-dominated.
1785       if (!Rec(U, Rec)) {
1786         NonDom.insert(T);
1787         return false;
1788       }
1789     }
1790     Dom.insert(T);
1791     return true;
1792   };
1793 
1794   auto IsDom = [&IsDomRec] (SDNode *T) { return IsDomRec(T, IsDomRec); };
1795 
1796   // Add the rest of nodes dominated by ISEL to SubNodes.
1797   for (unsigned I = 0; I != SubNodes.size(); ++I) {
1798     for (SDValue Op : SubNodes[I]->ops()) {
1799       SDNode *O = Op.getNode();
1800       if (IsDom(O))
1801         SubNodes.insert(O);
1802     }
1803   }
1804 
1805   // Do a topological sort of nodes from Dom.
1806   SetVector<SDNode*> TmpQ;
1807 
1808   std::map<SDNode *, unsigned> OpCount;
1809   for (SDNode *T : Dom) {
1810     unsigned NumDomOps = llvm::count_if(T->ops(), [&Dom](const SDUse &U) {
1811                              return Dom.count(U.getNode());
1812                            });
1813 
1814     OpCount.insert({T, NumDomOps});
1815     if (NumDomOps == 0)
1816       TmpQ.insert(T);
1817   }
1818 
1819   for (unsigned I = 0; I != TmpQ.size(); ++I) {
1820     SDNode *S = TmpQ[I];
1821     for (SDNode *U : S->uses()) {
1822       if (U == ISelN)
1823         continue;
1824       auto F = OpCount.find(U);
1825       assert(F != OpCount.end());
1826       if (F->second > 0 && !--F->second)
1827         TmpQ.insert(F->first);
1828     }
1829   }
1830 
1831   // Remove the marker.
1832   ISel.ReplaceNode(ISelN, N0);
1833 
1834   assert(SubNodes.size() == TmpQ.size());
1835   NullifyingVector<decltype(TmpQ)::vector_type> Queue(TmpQ.takeVector());
1836 
1837   Deleter DUQ(DAG, Queue);
1838   for (SDNode *S : reverse(Queue)) {
1839     if (S == nullptr)
1840       continue;
1841     DEBUG_WITH_TYPE("isel", {dbgs() << "HVX selecting: "; S->dump(&DAG);});
1842     ISel.Select(S);
1843   }
1844 }
1845 
1846 bool HvxSelector::scalarizeShuffle(ArrayRef<int> Mask, const SDLoc &dl,
1847                                    MVT ResTy, SDValue Va, SDValue Vb,
1848                                    SDNode *N) {
1849   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1850   MVT ElemTy = ResTy.getVectorElementType();
1851   assert(ElemTy == MVT::i8);
1852   unsigned VecLen = Mask.size();
1853   bool HavePairs = (2*HwLen == VecLen);
1854   MVT SingleTy = getSingleVT(MVT::i8);
1855 
1856   // The prior attempts to handle this shuffle may have left a bunch of
1857   // dead nodes in the DAG (such as constants). These nodes will be added
1858   // at the end of DAG's node list, which at that point had already been
1859   // sorted topologically. In the main selection loop, the node list is
1860   // traversed backwards from the root node, which means that any new
1861   // nodes (from the end of the list) will not be visited.
1862   // Scalarization will replace the shuffle node with the scalarized
1863   // expression, and if that expression reused any if the leftoever (dead)
1864   // nodes, these nodes would not be selected (since the "local" selection
1865   // only visits nodes that are not in AllNodes).
1866   // To avoid this issue, remove all dead nodes from the DAG now.
1867 //  DAG.RemoveDeadNodes();
1868 
1869   SmallVector<SDValue,128> Ops;
1870   LLVMContext &Ctx = *DAG.getContext();
1871   MVT LegalTy = Lower.getTypeToTransformTo(Ctx, ElemTy).getSimpleVT();
1872   for (int I : Mask) {
1873     if (I < 0) {
1874       Ops.push_back(ISel.selectUndef(dl, LegalTy));
1875       continue;
1876     }
1877     SDValue Vec;
1878     unsigned M = I;
1879     if (M < VecLen) {
1880       Vec = Va;
1881     } else {
1882       Vec = Vb;
1883       M -= VecLen;
1884     }
1885     if (HavePairs) {
1886       if (M < HwLen) {
1887         Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, SingleTy, Vec);
1888       } else {
1889         Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_hi, dl, SingleTy, Vec);
1890         M -= HwLen;
1891       }
1892     }
1893     SDValue Idx = DAG.getConstant(M, dl, MVT::i32);
1894     SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, LegalTy, {Vec, Idx});
1895     SDValue L = Lower.LowerOperation(Ex, DAG);
1896     assert(L.getNode());
1897     Ops.push_back(L);
1898   }
1899 
1900   SDValue LV;
1901   if (2*HwLen == VecLen) {
1902     SDValue B0 = DAG.getBuildVector(SingleTy, dl, {Ops.data(), HwLen});
1903     SDValue L0 = Lower.LowerOperation(B0, DAG);
1904     SDValue B1 = DAG.getBuildVector(SingleTy, dl, {Ops.data()+HwLen, HwLen});
1905     SDValue L1 = Lower.LowerOperation(B1, DAG);
1906     // XXX CONCAT_VECTORS is legal for HVX vectors. Legalizing (lowering)
1907     // functions may expect to be called only for illegal operations, so
1908     // make sure that they are not called for legal ones. Develop a better
1909     // mechanism for dealing with this.
1910     LV = DAG.getNode(ISD::CONCAT_VECTORS, dl, ResTy, {L0, L1});
1911   } else {
1912     SDValue BV = DAG.getBuildVector(ResTy, dl, Ops);
1913     LV = Lower.LowerOperation(BV, DAG);
1914   }
1915 
1916   assert(!N->use_empty());
1917   SDValue IS = DAG.getNode(HexagonISD::ISEL, dl, ResTy, LV);
1918   ISel.ReplaceNode(N, IS.getNode());
1919   select(IS.getNode());
1920   DAG.RemoveDeadNodes();
1921   return true;
1922 }
1923 
1924 SmallVector<uint32_t, 8> HvxSelector::getPerfectCompletions(ShuffleMask SM,
1925                                                             unsigned Width) {
1926   auto possibilities = [](ArrayRef<uint8_t> Bs, unsigned Width) -> uint32_t {
1927     unsigned Impossible = ~(1u << Width) + 1;
1928     for (unsigned I = 0, E = Bs.size(); I != E; ++I) {
1929       uint8_t B = Bs[I];
1930       if (B == 0xff)
1931         continue;
1932       if (~Impossible == 0)
1933         break;
1934       for (unsigned Log = 0; Log != Width; ++Log) {
1935         if (Impossible & (1u << Log))
1936           continue;
1937         unsigned Expected = (I >> Log) % 2;
1938         if (B != Expected)
1939           Impossible |= (1u << Log);
1940       }
1941     }
1942     return ~Impossible;
1943   };
1944 
1945   SmallVector<uint32_t, 8> Worklist(Width);
1946 
1947   for (unsigned BitIdx = 0; BitIdx != Width; ++BitIdx) {
1948     SmallVector<uint8_t> BitValues(SM.Mask.size());
1949     for (int i = 0, e = SM.Mask.size(); i != e; ++i) {
1950       int M = SM.Mask[i];
1951       if (M < 0)
1952         BitValues[i] = 0xff;
1953       else
1954         BitValues[i] = (M & (1u << BitIdx)) != 0;
1955     }
1956     Worklist[BitIdx] = possibilities(BitValues, Width);
1957   }
1958 
1959   // If there is a word P in Worklist that matches multiple possibilities,
1960   // then if any other word Q matches any of the possibilities matched by P,
1961   // then Q matches all the possibilities matched by P. In fact, P == Q.
1962   // In other words, for each words P, Q, the sets of possibilities matched
1963   // by P and Q are either equal or disjoint (no partial overlap).
1964   //
1965   // Illustration: For 4-bit values there are 4 complete sequences:
1966   // a:  0 1 0 1  0 1 0 1  0 1 0 1  0 1 0 1
1967   // b:  0 0 1 1  0 0 1 1  0 0 1 1  0 0 1 1
1968   // c:  0 0 0 0  1 1 1 1  0 0 0 0  1 1 1 1
1969   // d:  0 0 0 0  0 0 0 0  1 1 1 1  1 1 1 1
1970   //
1971   // Words containing unknown bits that match two of the complete
1972   // sequences:
1973   // ab: 0 u u 1  0 u u 1  0 u u 1  0 u u 1
1974   // ac: 0 u 0 u  u 1 u 1  0 u 0 u  u 1 u 1
1975   // ad: 0 u 0 u  0 u 0 u  u 1 u 1  u 1 u 1
1976   // bc: 0 0 u u  u u 1 1  0 0 u u  u u 1 1
1977   // bd: 0 0 u u  0 0 u u  u u 1 1  u u 1 1
1978   // cd: 0 0 0 0  u u u u  u u u u  1 1 1 1
1979   //
1980   // Proof of the claim above:
1981   // Let P be a word that matches s0 and s1. For that to happen, all known
1982   // bits in P must match s0 and s1 exactly.
1983   // Assume there is Q that matches s1. Note that since P and Q came from
1984   // the same shuffle mask, the positions of unknown bits in P and Q match
1985   // exactly, which makes the indices of known bits be exactly the same
1986   // between P and Q. Since P matches s0 and s1, the known bits of P much
1987   // match both s0 and s1. Also, since Q matches s1, the known bits in Q
1988   // are exactly the same as in s1, which means that they are exactly the
1989   // same as in P. This implies that P == Q.
1990 
1991   // There can be a situation where there are more entries with the same
1992   // bits set than there are set bits (e.g. value 9 occuring more than 2
1993   // times). In such cases it will be impossible to complete this to a
1994   // perfect shuffle.
1995   SmallVector<uint32_t, 8> Sorted(Worklist);
1996   llvm::sort(Sorted.begin(), Sorted.end());
1997 
1998   for (unsigned I = 0, E = Sorted.size(); I != E;) {
1999     unsigned P = Sorted[I], Count = 1;
2000     while (++I != E && P == Sorted[I])
2001       ++Count;
2002     if ((unsigned)llvm::popcount(P) < Count) {
2003       // Reset all occurences of P, if there are more occurrences of P
2004       // than there are bits in P.
2005       for (unsigned &Q : Worklist) {
2006         if (Q == P)
2007           Q = 0;
2008       }
2009     }
2010   }
2011 
2012   return Worklist;
2013 }
2014 
2015 SmallVector<uint32_t, 8>
2016 HvxSelector::completeToPerfect(ArrayRef<uint32_t> Completions, unsigned Width) {
2017   // Pick a completion if there are multiple possibilities. For now just
2018   // select any valid completion.
2019   SmallVector<uint32_t, 8> Comps(Completions);
2020 
2021   for (unsigned I = 0; I != Width; ++I) {
2022     uint32_t P = Comps[I];
2023     assert(P != 0);
2024     if (isPowerOf2_32(P))
2025       continue;
2026     // T = least significant bit of P.
2027     uint32_t T = P ^ ((P - 1) & P);
2028     // Clear T in all remaining words matching P.
2029     for (unsigned J = I + 1; J != Width; ++J) {
2030       if (Comps[J] == P)
2031         Comps[J] ^= T;
2032     }
2033     Comps[I] = T;
2034   }
2035 
2036 #ifndef NDEBUG
2037   // Check that we have generated a valid completion.
2038   uint32_t OrAll = 0;
2039   for (uint32_t C : Comps) {
2040     assert(isPowerOf2_32(C));
2041     OrAll |= C;
2042   }
2043   assert(OrAll == (1u << Width) -1);
2044 #endif
2045 
2046   return Comps;
2047 }
2048 
2049 std::optional<int> HvxSelector::rotationDistance(ShuffleMask SM,
2050                                                  unsigned WrapAt) {
2051   std::optional<int> Dist;
2052   for (int I = 0, E = SM.Mask.size(); I != E; ++I) {
2053     int M = SM.Mask[I];
2054     if (M < 0)
2055       continue;
2056     if (Dist) {
2057       if ((I + *Dist) % static_cast<int>(WrapAt) != M)
2058         return std::nullopt;
2059     } else {
2060       // Integer a%b operator assumes rounding towards zero by /, so it
2061       // "misbehaves" when a crosses 0 (the remainder also changes sign).
2062       // Add WrapAt in an attempt to keep I+Dist non-negative.
2063       Dist = M - I;
2064       if (Dist < 0)
2065         Dist = *Dist + WrapAt;
2066     }
2067   }
2068   return Dist;
2069 }
2070 
2071 OpRef HvxSelector::contracting(ShuffleMask SM, OpRef Va, OpRef Vb,
2072                                ResultStack &Results) {
2073   DEBUG_WITH_TYPE("isel", { dbgs() << __func__ << '\n'; });
2074   if (!Va.isValid() || !Vb.isValid())
2075     return OpRef::fail();
2076 
2077   // Contracting shuffles, i.e. instructions that always discard some bytes
2078   // from the operand vectors.
2079   //
2080   // Funnel shifts
2081   // V6_vshuff{e,o}b
2082   // V6_vshuf{e,o}h
2083   // V6_vdealb4w
2084   // V6_vpack{e,o}{b,h}
2085 
2086   int VecLen = SM.Mask.size();
2087 
2088   // First, check for funnel shifts.
2089   if (auto Dist = rotationDistance(SM, 2 * VecLen)) {
2090     OpRef Funnel = funnels(Va, Vb, *Dist, Results);
2091     if (Funnel.isValid())
2092       return Funnel;
2093   }
2094 
2095   MVT SingleTy = getSingleVT(MVT::i8);
2096   MVT PairTy = getPairVT(MVT::i8);
2097 
2098   auto same = [](ArrayRef<int> Mask1, ArrayRef<int> Mask2) -> bool {
2099     return Mask1 == Mask2;
2100   };
2101 
2102   using PackConfig = std::pair<unsigned, bool>;
2103   PackConfig Packs[] = {
2104       {1, false}, // byte, even
2105       {1, true},  // byte, odd
2106       {2, false}, // half, even
2107       {2, true},  // half, odd
2108   };
2109 
2110   { // Check vpack
2111     unsigned Opcodes[] = {
2112         Hexagon::V6_vpackeb,
2113         Hexagon::V6_vpackob,
2114         Hexagon::V6_vpackeh,
2115         Hexagon::V6_vpackoh,
2116     };
2117     for (int i = 0, e = std::size(Opcodes); i != e; ++i) {
2118       auto [Size, Odd] = Packs[i];
2119       if (same(SM.Mask, shuffles::mask(shuffles::vpack, HwLen, Size, Odd))) {
2120         Results.push(Opcodes[i], SingleTy, {Vb, Va});
2121         return OpRef::res(Results.top());
2122       }
2123     }
2124   }
2125 
2126   { // Check vshuff
2127     unsigned Opcodes[] = {
2128         Hexagon::V6_vshuffeb,
2129         Hexagon::V6_vshuffob,
2130         Hexagon::V6_vshufeh,
2131         Hexagon::V6_vshufoh,
2132     };
2133     for (int i = 0, e = std::size(Opcodes); i != e; ++i) {
2134       auto [Size, Odd] = Packs[i];
2135       if (same(SM.Mask, shuffles::mask(shuffles::vshuff, HwLen, Size, Odd))) {
2136         Results.push(Opcodes[i], SingleTy, {Vb, Va});
2137         return OpRef::res(Results.top());
2138       }
2139     }
2140   }
2141 
2142   { // Check vdeal
2143     // There is no "V6_vdealeb", etc, but the supposed behavior of vdealeb
2144     // is equivalent to "(V6_vpackeb (V6_vdealvdd Vu, Vv, -2))". Other such
2145     // variants of "deal" can be done similarly.
2146     unsigned Opcodes[] = {
2147         Hexagon::V6_vpackeb,
2148         Hexagon::V6_vpackob,
2149         Hexagon::V6_vpackeh,
2150         Hexagon::V6_vpackoh,
2151     };
2152     const SDLoc &dl(Results.InpNode);
2153 
2154     for (int i = 0, e = std::size(Opcodes); i != e; ++i) {
2155       auto [Size, Odd] = Packs[i];
2156       if (same(SM.Mask, shuffles::mask(shuffles::vdeal, HwLen, Size, Odd))) {
2157         Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(-2 * Size, dl)});
2158         Results.push(Hexagon::V6_vdealvdd, PairTy, {Vb, Va, OpRef::res(-1)});
2159         auto vdeal = OpRef::res(Results.top());
2160         Results.push(Opcodes[i], SingleTy,
2161                      {OpRef::hi(vdeal), OpRef::lo(vdeal)});
2162         return OpRef::res(Results.top());
2163       }
2164     }
2165   }
2166 
2167   if (same(SM.Mask, shuffles::mask(shuffles::vdealb4w, HwLen))) {
2168     Results.push(Hexagon::V6_vdealb4w, SingleTy, {Vb, Va});
2169     return OpRef::res(Results.top());
2170   }
2171 
2172   return OpRef::fail();
2173 }
2174 
2175 OpRef HvxSelector::expanding(ShuffleMask SM, OpRef Va, ResultStack &Results) {
2176   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
2177   // Expanding shuffles (using all elements and inserting into larger vector):
2178   //
2179   // V6_vunpacku{b,h} [*]
2180   //
2181   // [*] Only if the upper elements (filled with 0s) are "don't care" in Mask.
2182   //
2183   // Note: V6_vunpacko{b,h} are or-ing the high byte/half in the result, so
2184   // they are not shuffles.
2185   //
2186   // The argument is a single vector.
2187 
2188   int VecLen = SM.Mask.size();
2189   assert(2*HwLen == unsigned(VecLen) && "Expecting vector-pair type");
2190 
2191   std::pair<int,unsigned> Strip = findStrip(SM.Mask, 1, VecLen);
2192 
2193   // The patterns for the unpacks, in terms of the starting offsets of the
2194   // consecutive strips (L = length of the strip, N = VecLen):
2195   //
2196   // vunpacku:  0, -1, L, -1, 2L, -1 ...
2197 
2198   if (Strip.first != 0)
2199     return OpRef::fail();
2200 
2201   // The vunpackus only handle byte and half-word.
2202   if (Strip.second != 1 && Strip.second != 2)
2203     return OpRef::fail();
2204 
2205   int N = VecLen;
2206   int L = Strip.second;
2207 
2208   // First, check the non-ignored strips.
2209   for (int I = 2*L; I < N; I += 2*L) {
2210     auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
2211     if (S.second != unsigned(L))
2212       return OpRef::fail();
2213     if (2*S.first != I)
2214       return OpRef::fail();
2215   }
2216   // Check the -1s.
2217   for (int I = L; I < N; I += 2*L) {
2218     auto S = findStrip(SM.Mask.drop_front(I), 0, N-I);
2219     if (S.first != -1 || S.second != unsigned(L))
2220       return OpRef::fail();
2221   }
2222 
2223   unsigned Opc = Strip.second == 1 ? Hexagon::V6_vunpackub
2224                                    : Hexagon::V6_vunpackuh;
2225   Results.push(Opc, getPairVT(MVT::i8), {Va});
2226   return OpRef::res(Results.top());
2227 }
2228 
2229 OpRef HvxSelector::perfect(ShuffleMask SM, OpRef Va, ResultStack &Results) {
2230   DEBUG_WITH_TYPE("isel", { dbgs() << __func__ << '\n'; });
2231   // V6_vdeal{b,h}
2232   // V6_vshuff{b,h}
2233 
2234   // V6_vshufoe{b,h}  those are quivalent to vshuffvdd(..,{1,2})
2235   // V6_vshuffvdd (V6_vshuff)
2236   // V6_dealvdd (V6_vdeal)
2237 
2238   int VecLen = SM.Mask.size();
2239   assert(isPowerOf2_32(VecLen) && Log2_32(VecLen) <= 8);
2240   unsigned LogLen = Log2_32(VecLen);
2241   unsigned HwLog = Log2_32(HwLen);
2242   // The result length must be the same as the length of a single vector,
2243   // or a vector pair.
2244   assert(LogLen == HwLog || LogLen == HwLog + 1);
2245   bool HavePairs = LogLen == HwLog + 1;
2246 
2247   SmallVector<unsigned, 8> Perm(LogLen);
2248 
2249   // Check if this could be a perfect shuffle, or a combination of perfect
2250   // shuffles.
2251   //
2252   // Consider this permutation (using hex digits to make the ASCII diagrams
2253   // easier to read):
2254   //   { 0, 8, 1, 9, 2, A, 3, B, 4, C, 5, D, 6, E, 7, F }.
2255   // This is a "deal" operation: divide the input into two halves, and
2256   // create the output by picking elements by alternating between these two
2257   // halves:
2258   //   0 1 2 3 4 5 6 7    -->    0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F  [*]
2259   //   8 9 A B C D E F
2260   //
2261   // Aside from a few special explicit cases (V6_vdealb, etc.), HVX provides
2262   // a somwehat different mechanism that could be used to perform shuffle/
2263   // deal operations: a 2x2 transpose.
2264   // Consider the halves of inputs again, they can be interpreted as a 2x8
2265   // matrix. A 2x8 matrix can be looked at four 2x2 matrices concatenated
2266   // together. Now, when considering 2 elements at a time, it will be a 2x4
2267   // matrix (with elements 01, 23, 45, etc.), or two 2x2 matrices:
2268   //   01 23  45 67
2269   //   89 AB  CD EF
2270   // With groups of 4, this will become a single 2x2 matrix, and so on.
2271   //
2272   // The 2x2 transpose instruction works by transposing each of the 2x2
2273   // matrices (or "sub-matrices"), given a specific group size. For example,
2274   // if the group size is 1 (i.e. each element is its own group), there
2275   // will be four transposes of the four 2x2 matrices that form the 2x8.
2276   // For example, with the inputs as above, the result will be:
2277   //   0 8  2 A  4 C  6 E
2278   //   1 9  3 B  5 D  7 F
2279   // Now, this result can be tranposed again, but with the group size of 2:
2280   //   08 19  4C 5D
2281   //   2A 3B  6E 7F
2282   // If we then transpose that result, but with the group size of 4, we get:
2283   //   0819 2A3B
2284   //   4C5D 6E7F
2285   // If we concatenate these two rows, it will be
2286   //   0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F
2287   // which is the same as the "deal" [*] above.
2288   //
2289   // In general, a "deal" of individual elements is a series of 2x2 transposes,
2290   // with changing group size. HVX has two instructions:
2291   //   Vdd = V6_vdealvdd Vu, Vv, Rt
2292   //   Vdd = V6_shufvdd  Vu, Vv, Rt
2293   // that perform exactly that. The register Rt controls which transposes are
2294   // going to happen: a bit at position n (counting from 0) indicates that a
2295   // transpose with a group size of 2^n will take place. If multiple bits are
2296   // set, multiple transposes will happen: vdealvdd will perform them starting
2297   // with the largest group size, vshuffvdd will do them in the reverse order.
2298   //
2299   // The main observation is that each 2x2 transpose corresponds to swapping
2300   // columns of bits in the binary representation of the values.
2301   //
2302   // The numbers {3,2,1,0} and the log2 of the number of contiguous 1 bits
2303   // in a given column. The * denote the columns that will be swapped.
2304   // The transpose with the group size 2^n corresponds to swapping columns
2305   // 3 (the highest log) and log2(n):
2306   //
2307   //     3 2 1 0         0 2 1 3         0 2 3 1
2308   //     *     *             * *           * *
2309   //  0  0 0 0 0      0  0 0 0 0      0  0 0 0 0      0  0 0 0 0
2310   //  1  0 0 0 1      8  1 0 0 0      8  1 0 0 0      8  1 0 0 0
2311   //  2  0 0 1 0      2  0 0 1 0      1  0 0 0 1      1  0 0 0 1
2312   //  3  0 0 1 1      A  1 0 1 0      9  1 0 0 1      9  1 0 0 1
2313   //  4  0 1 0 0      4  0 1 0 0      4  0 1 0 0      2  0 0 1 0
2314   //  5  0 1 0 1      C  1 1 0 0      C  1 1 0 0      A  1 0 1 0
2315   //  6  0 1 1 0      6  0 1 1 0      5  0 1 0 1      3  0 0 1 1
2316   //  7  0 1 1 1      E  1 1 1 0      D  1 1 0 1      B  1 0 1 1
2317   //  8  1 0 0 0      1  0 0 0 1      2  0 0 1 0      4  0 1 0 0
2318   //  9  1 0 0 1      9  1 0 0 1      A  1 0 1 0      C  1 1 0 0
2319   //  A  1 0 1 0      3  0 0 1 1      3  0 0 1 1      5  0 1 0 1
2320   //  B  1 0 1 1      B  1 0 1 1      B  1 0 1 1      D  1 1 0 1
2321   //  C  1 1 0 0      5  0 1 0 1      6  0 1 1 0      6  0 1 1 0
2322   //  D  1 1 0 1      D  1 1 0 1      E  1 1 1 0      E  1 1 1 0
2323   //  E  1 1 1 0      7  0 1 1 1      7  0 1 1 1      7  0 1 1 1
2324   //  F  1 1 1 1      F  1 1 1 1      F  1 1 1 1      F  1 1 1 1
2325 
2326   // There is one special case that is not a perfect shuffle, but can be
2327   // turned into one easily: when the shuffle operates on a vector pair,
2328   // but the two vectors in the pair are swapped. The code that identifies
2329   // perfect shuffles will reject it, unless the order is reversed.
2330   shuffles::MaskT MaskStorage(SM.Mask);
2331   bool InvertedPair = false;
2332   if (HavePairs && SM.Mask[0] >= int(HwLen)) {
2333     for (int i = 0, e = SM.Mask.size(); i != e; ++i) {
2334       int M = SM.Mask[i];
2335       MaskStorage[i] = M >= int(HwLen) ? M - HwLen : M + HwLen;
2336     }
2337     InvertedPair = true;
2338     SM = ShuffleMask(MaskStorage);
2339   }
2340 
2341   auto Comps = getPerfectCompletions(SM, LogLen);
2342   if (llvm::is_contained(Comps, 0))
2343     return OpRef::fail();
2344 
2345   auto Pick = completeToPerfect(Comps, LogLen);
2346   for (unsigned I = 0; I != LogLen; ++I)
2347     Perm[I] = Log2_32(Pick[I]);
2348 
2349   // Once we have Perm, represent it as cycles. Denote the maximum log2
2350   // (equal to log2(VecLen)-1) as M. The cycle containing M can then be
2351   // written as (M a1 a2 a3 ... an). That cycle can be broken up into
2352   // simple swaps as (M a1)(M a2)(M a3)...(M an), with the composition
2353   // order being from left to right. Any (contiguous) segment where the
2354   // values ai, ai+1...aj are either all increasing or all decreasing,
2355   // can be implemented via a single vshuffvdd/vdealvdd respectively.
2356   //
2357   // If there is a cycle (a1 a2 ... an) that does not involve M, it can
2358   // be written as (M an)(a1 a2 ... an)(M a1). The first two cycles can
2359   // then be folded to get (M a1 a2 ... an)(M a1), and the above procedure
2360   // can be used to generate a sequence of vshuffvdd/vdealvdd.
2361   //
2362   // Example:
2363   // Assume M = 4 and consider a permutation (0 1)(2 3). It can be written
2364   // as (4 0 1)(4 0) composed with (4 2 3)(4 2), or simply
2365   //   (4 0 1)(4 0)(4 2 3)(4 2).
2366   // It can then be expanded into swaps as
2367   //   (4 0)(4 1)(4 0)(4 2)(4 3)(4 2),
2368   // and broken up into "increasing" segments as
2369   //   [(4 0)(4 1)] [(4 0)(4 2)(4 3)] [(4 2)].
2370   // This is equivalent to
2371   //   (4 0 1)(4 0 2 3)(4 2),
2372   // which can be implemented as 3 vshufvdd instructions.
2373 
2374   using CycleType = SmallVector<unsigned, 8>;
2375   std::set<CycleType> Cycles;
2376   std::set<unsigned> All;
2377 
2378   for (unsigned I : Perm)
2379     All.insert(I);
2380 
2381   // If the cycle contains LogLen-1, move it to the front of the cycle.
2382   // Otherwise, return the cycle unchanged.
2383   auto canonicalize = [LogLen](const CycleType &C) -> CycleType {
2384     unsigned LogPos, N = C.size();
2385     for (LogPos = 0; LogPos != N; ++LogPos)
2386       if (C[LogPos] == LogLen - 1)
2387         break;
2388     if (LogPos == N)
2389       return C;
2390 
2391     CycleType NewC(C.begin() + LogPos, C.end());
2392     NewC.append(C.begin(), C.begin() + LogPos);
2393     return NewC;
2394   };
2395 
2396   auto pfs = [](const std::set<CycleType> &Cs, unsigned Len) {
2397     // Ordering: shuff: 5 0 1 2 3 4, deal: 5 4 3 2 1 0 (for Log=6),
2398     // for bytes zero is included, for halfwords is not.
2399     if (Cs.size() != 1)
2400       return 0u;
2401     const CycleType &C = *Cs.begin();
2402     if (C[0] != Len - 1)
2403       return 0u;
2404     int D = Len - C.size();
2405     if (D != 0 && D != 1)
2406       return 0u;
2407 
2408     bool IsDeal = true, IsShuff = true;
2409     for (unsigned I = 1; I != Len - D; ++I) {
2410       if (C[I] != Len - 1 - I)
2411         IsDeal = false;
2412       if (C[I] != I - (1 - D)) // I-1, I
2413         IsShuff = false;
2414     }
2415     // At most one, IsDeal or IsShuff, can be non-zero.
2416     assert(!(IsDeal || IsShuff) || IsDeal != IsShuff);
2417     static unsigned Deals[] = {Hexagon::V6_vdealb, Hexagon::V6_vdealh};
2418     static unsigned Shufs[] = {Hexagon::V6_vshuffb, Hexagon::V6_vshuffh};
2419     return IsDeal ? Deals[D] : (IsShuff ? Shufs[D] : 0);
2420   };
2421 
2422   while (!All.empty()) {
2423     unsigned A = *All.begin();
2424     All.erase(A);
2425     CycleType C;
2426     C.push_back(A);
2427     for (unsigned B = Perm[A]; B != A; B = Perm[B]) {
2428       C.push_back(B);
2429       All.erase(B);
2430     }
2431     if (C.size() <= 1)
2432       continue;
2433     Cycles.insert(canonicalize(C));
2434   }
2435 
2436   MVT SingleTy = getSingleVT(MVT::i8);
2437   MVT PairTy = getPairVT(MVT::i8);
2438 
2439   // Recognize patterns for V6_vdeal{b,h} and V6_vshuff{b,h}.
2440   if (unsigned(VecLen) == HwLen) {
2441     if (unsigned SingleOpc = pfs(Cycles, LogLen)) {
2442       Results.push(SingleOpc, SingleTy, {Va});
2443       return OpRef::res(Results.top());
2444     }
2445   }
2446 
2447   // From the cycles, construct the sequence of values that will
2448   // then form the control values for vdealvdd/vshuffvdd, i.e.
2449   // (M a1 a2)(M a3 a4 a5)... -> a1 a2 a3 a4 a5
2450   // This essentially strips the M value from the cycles where
2451   // it's present, and performs the insertion of M (then stripping)
2452   // for cycles without M (as described in an earlier comment).
2453   SmallVector<unsigned, 8> SwapElems;
2454   // When the input is extended (i.e. single vector becomes a pair),
2455   // this is done by using an "undef" vector as the second input.
2456   // However, then we get
2457   //   input 1: GOODBITS
2458   //   input 2: ........
2459   // but we need
2460   //   input 1: ....BITS
2461   //   input 2: ....GOOD
2462   // Then at the end, this needs to be undone. To accomplish this,
2463   // artificially add "LogLen-1" at both ends of the sequence.
2464   if (!HavePairs)
2465     SwapElems.push_back(LogLen - 1);
2466   for (const CycleType &C : Cycles) {
2467     // Do the transformation: (a1..an) -> (M a1..an)(M a1).
2468     unsigned First = (C[0] == LogLen - 1) ? 1 : 0;
2469     SwapElems.append(C.begin() + First, C.end());
2470     if (First == 0)
2471       SwapElems.push_back(C[0]);
2472   }
2473   if (!HavePairs)
2474     SwapElems.push_back(LogLen - 1);
2475 
2476   const SDLoc &dl(Results.InpNode);
2477   OpRef Arg = HavePairs ? Va : concats(Va, OpRef::undef(SingleTy), Results);
2478   if (InvertedPair)
2479     Arg = concats(OpRef::hi(Arg), OpRef::lo(Arg), Results);
2480 
2481   for (unsigned I = 0, E = SwapElems.size(); I != E;) {
2482     bool IsInc = I == E - 1 || SwapElems[I] < SwapElems[I + 1];
2483     unsigned S = (1u << SwapElems[I]);
2484     if (I < E - 1) {
2485       while (++I < E - 1 && IsInc == (SwapElems[I] < SwapElems[I + 1]))
2486         S |= 1u << SwapElems[I];
2487       // The above loop will not add a bit for the final SwapElems[I+1],
2488       // so add it here.
2489       S |= 1u << SwapElems[I];
2490     }
2491     ++I;
2492 
2493     NodeTemplate Res;
2494     Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(S, dl)});
2495     Res.Opc = IsInc ? Hexagon::V6_vshuffvdd : Hexagon::V6_vdealvdd;
2496     Res.Ty = PairTy;
2497     Res.Ops = {OpRef::hi(Arg), OpRef::lo(Arg), OpRef::res(-1)};
2498     Results.push(Res);
2499     Arg = OpRef::res(Results.top());
2500   }
2501 
2502   return HavePairs ? Arg : OpRef::lo(Arg);
2503 }
2504 
2505 OpRef HvxSelector::butterfly(ShuffleMask SM, OpRef Va, ResultStack &Results) {
2506   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
2507   // Butterfly shuffles.
2508   //
2509   // V6_vdelta
2510   // V6_vrdelta
2511   // V6_vror
2512 
2513   // The assumption here is that all elements picked by Mask are in the
2514   // first operand to the vector_shuffle. This assumption is enforced
2515   // by the caller.
2516 
2517   MVT ResTy = getSingleVT(MVT::i8);
2518   PermNetwork::Controls FC, RC;
2519   const SDLoc &dl(Results.InpNode);
2520   int VecLen = SM.Mask.size();
2521 
2522   for (int M : SM.Mask) {
2523     if (M != -1 && M >= VecLen)
2524       return OpRef::fail();
2525   }
2526 
2527   // Try the deltas/benes for both single vectors and vector pairs.
2528   ForwardDeltaNetwork FN(SM.Mask);
2529   if (FN.run(FC)) {
2530     SDValue Ctl = getVectorConstant(FC, dl);
2531     Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(Ctl)});
2532     return OpRef::res(Results.top());
2533   }
2534 
2535   // Try reverse delta.
2536   ReverseDeltaNetwork RN(SM.Mask);
2537   if (RN.run(RC)) {
2538     SDValue Ctl = getVectorConstant(RC, dl);
2539     Results.push(Hexagon::V6_vrdelta, ResTy, {Va, OpRef(Ctl)});
2540     return OpRef::res(Results.top());
2541   }
2542 
2543   // Do Benes.
2544   BenesNetwork BN(SM.Mask);
2545   if (BN.run(FC, RC)) {
2546     SDValue CtlF = getVectorConstant(FC, dl);
2547     SDValue CtlR = getVectorConstant(RC, dl);
2548     Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(CtlF)});
2549     Results.push(Hexagon::V6_vrdelta, ResTy,
2550                  {OpRef::res(-1), OpRef(CtlR)});
2551     return OpRef::res(Results.top());
2552   }
2553 
2554   return OpRef::fail();
2555 }
2556 
2557 SDValue HvxSelector::getConst32(int Val, const SDLoc &dl) {
2558   return DAG.getTargetConstant(Val, dl, MVT::i32);
2559 }
2560 
2561 SDValue HvxSelector::getVectorConstant(ArrayRef<uint8_t> Data,
2562                                        const SDLoc &dl) {
2563   SmallVector<SDValue, 128> Elems;
2564   for (uint8_t C : Data)
2565     Elems.push_back(DAG.getConstant(C, dl, MVT::i8));
2566   MVT VecTy = MVT::getVectorVT(MVT::i8, Data.size());
2567   SDValue BV = DAG.getBuildVector(VecTy, dl, Elems);
2568   SDValue LV = Lower.LowerOperation(BV, DAG);
2569   DAG.RemoveDeadNode(BV.getNode());
2570   return DAG.getNode(HexagonISD::ISEL, dl, VecTy, LV);
2571 }
2572 
2573 void HvxSelector::selectExtractSubvector(SDNode *N) {
2574   SDValue Inp = N->getOperand(0);
2575   MVT ResTy = N->getValueType(0).getSimpleVT();
2576   unsigned Idx = N->getConstantOperandVal(1);
2577 
2578   [[maybe_unused]] MVT InpTy = Inp.getValueType().getSimpleVT();
2579   [[maybe_unused]] unsigned ResLen = ResTy.getVectorNumElements();
2580   assert(InpTy.getVectorElementType() == ResTy.getVectorElementType());
2581   assert(2 * ResLen == InpTy.getVectorNumElements());
2582   assert(Idx == 0 || Idx == ResLen);
2583 
2584   unsigned SubReg = Idx == 0 ? Hexagon::vsub_lo : Hexagon::vsub_hi;
2585   SDValue Ext = DAG.getTargetExtractSubreg(SubReg, SDLoc(N), ResTy, Inp);
2586 
2587   ISel.ReplaceNode(N, Ext.getNode());
2588 }
2589 
2590 void HvxSelector::selectShuffle(SDNode *N) {
2591   DEBUG_WITH_TYPE("isel", {
2592     dbgs() << "Starting " << __func__ << " on node:\n";
2593     N->dump(&DAG);
2594   });
2595   MVT ResTy = N->getValueType(0).getSimpleVT();
2596   // Assume that vector shuffles operate on vectors of bytes.
2597   assert(ResTy.isVector() && ResTy.getVectorElementType() == MVT::i8);
2598 
2599   auto *SN = cast<ShuffleVectorSDNode>(N);
2600   std::vector<int> Mask(SN->getMask().begin(), SN->getMask().end());
2601   // This shouldn't really be necessary. Is it?
2602   for (int &Idx : Mask)
2603     if (Idx != -1 && Idx < 0)
2604       Idx = -1;
2605 
2606   unsigned VecLen = Mask.size();
2607   bool HavePairs = (2*HwLen == VecLen);
2608   assert(ResTy.getSizeInBits() / 8 == VecLen);
2609 
2610   // Vd = vector_shuffle Va, Vb, Mask
2611   //
2612 
2613   bool UseLeft = false, UseRight = false;
2614   for (unsigned I = 0; I != VecLen; ++I) {
2615     if (Mask[I] == -1)
2616       continue;
2617     unsigned Idx = Mask[I];
2618     assert(Idx < 2*VecLen);
2619     if (Idx < VecLen)
2620       UseLeft = true;
2621     else
2622       UseRight = true;
2623   }
2624 
2625   DEBUG_WITH_TYPE("isel", {
2626     dbgs() << "VecLen=" << VecLen << " HwLen=" << HwLen << " UseLeft="
2627            << UseLeft << " UseRight=" << UseRight << " HavePairs="
2628            << HavePairs << '\n';
2629   });
2630   // If the mask is all -1's, generate "undef".
2631   if (!UseLeft && !UseRight) {
2632     ISel.ReplaceNode(N, ISel.selectUndef(SDLoc(SN), ResTy).getNode());
2633     return;
2634   }
2635 
2636   SDValue Vec0 = N->getOperand(0);
2637   SDValue Vec1 = N->getOperand(1);
2638   assert(Vec0.getValueType() == ResTy && Vec1.getValueType() == ResTy);
2639 
2640   ResultStack Results(SN);
2641   OpRef Va = OpRef::undef(ResTy);
2642   OpRef Vb = OpRef::undef(ResTy);
2643 
2644   if (!Vec0.isUndef()) {
2645     Results.push(TargetOpcode::COPY, ResTy, {Vec0});
2646     Va = OpRef::OpRef::res(Results.top());
2647   }
2648   if (!Vec1.isUndef()) {
2649     Results.push(TargetOpcode::COPY, ResTy, {Vec1});
2650     Vb = OpRef::res(Results.top());
2651   }
2652 
2653   OpRef Res = !HavePairs ? shuffs2(ShuffleMask(Mask), Va, Vb, Results)
2654                          : shuffp2(ShuffleMask(Mask), Va, Vb, Results);
2655 
2656   bool Done = Res.isValid();
2657   if (Done) {
2658     // Make sure that Res is on the stack before materializing.
2659     Results.push(TargetOpcode::COPY, ResTy, {Res});
2660     materialize(Results);
2661   } else {
2662     Done = scalarizeShuffle(Mask, SDLoc(N), ResTy, Vec0, Vec1, N);
2663   }
2664 
2665   if (!Done) {
2666 #ifndef NDEBUG
2667     dbgs() << "Unhandled shuffle:\n";
2668     SN->dumpr(&DAG);
2669 #endif
2670     llvm_unreachable("Failed to select vector shuffle");
2671   }
2672 }
2673 
2674 void HvxSelector::selectRor(SDNode *N) {
2675   // If this is a rotation by less than 8, use V6_valignbi.
2676   MVT Ty = N->getValueType(0).getSimpleVT();
2677   const SDLoc &dl(N);
2678   SDValue VecV = N->getOperand(0);
2679   SDValue RotV = N->getOperand(1);
2680   SDNode *NewN = nullptr;
2681 
2682   if (auto *CN = dyn_cast<ConstantSDNode>(RotV.getNode())) {
2683     unsigned S = CN->getZExtValue() % HST.getVectorLength();
2684     if (S == 0) {
2685       NewN = VecV.getNode();
2686     } else if (isUInt<3>(S)) {
2687       NewN = DAG.getMachineNode(Hexagon::V6_valignbi, dl, Ty,
2688                                 {VecV, VecV, getConst32(S, dl)});
2689     }
2690   }
2691 
2692   if (!NewN)
2693     NewN = DAG.getMachineNode(Hexagon::V6_vror, dl, Ty, {VecV, RotV});
2694 
2695   ISel.ReplaceNode(N, NewN);
2696 }
2697 
2698 void HvxSelector::selectVAlign(SDNode *N) {
2699   SDValue Vv = N->getOperand(0);
2700   SDValue Vu = N->getOperand(1);
2701   SDValue Rt = N->getOperand(2);
2702   SDNode *NewN = DAG.getMachineNode(Hexagon::V6_valignb, SDLoc(N),
2703                                     N->getValueType(0), {Vv, Vu, Rt});
2704   ISel.ReplaceNode(N, NewN);
2705   DAG.RemoveDeadNode(N);
2706 }
2707 
2708 void HexagonDAGToDAGISel::PreprocessHvxISelDAG() {
2709   auto getNodes = [this]() -> std::vector<SDNode *> {
2710     std::vector<SDNode *> T;
2711     T.reserve(CurDAG->allnodes_size());
2712     for (SDNode &N : CurDAG->allnodes())
2713       T.push_back(&N);
2714     return T;
2715   };
2716 
2717   ppHvxShuffleOfShuffle(getNodes());
2718 }
2719 
2720 template <> struct std::hash<SDValue> {
2721   std::size_t operator()(SDValue V) const {
2722     return std::hash<const void *>()(V.getNode()) +
2723            std::hash<unsigned>()(V.getResNo());
2724   };
2725 };
2726 
2727 void HexagonDAGToDAGISel::ppHvxShuffleOfShuffle(std::vector<SDNode *> &&Nodes) {
2728   // Motivating case:
2729   //   t10: v64i32 = ...
2730   //         t46: v128i8 = vector_shuffle<...> t44, t45
2731   //         t48: v128i8 = vector_shuffle<...> t44, t45
2732   //       t42: v128i8 = vector_shuffle<...> t46, t48
2733   //     t12: v32i32 = extract_subvector t10, Constant:i32<0>
2734   //   t44: v128i8 = bitcast t12
2735   //     t15: v32i32 = extract_subvector t10, Constant:i32<32>
2736   //   t45: v128i8 = bitcast t15
2737   SelectionDAG &DAG = *CurDAG;
2738   unsigned HwLen = HST->getVectorLength();
2739 
2740   struct SubVectorInfo {
2741     SubVectorInfo(SDValue S, unsigned H) : Src(S), HalfIdx(H) {}
2742     SDValue Src;
2743     unsigned HalfIdx;
2744   };
2745 
2746   using MapType = std::unordered_map<SDValue, unsigned>;
2747 
2748   auto getMaskElt = [&](unsigned Idx, ShuffleVectorSDNode *Shuff0,
2749                         ShuffleVectorSDNode *Shuff1,
2750                         const MapType &OpMap) -> int {
2751     // Treat Shuff0 and Shuff1 as operands to another vector shuffle, and
2752     // Idx as a (non-undef) element of the top level shuffle's mask, that
2753     // is, index into concat(Shuff0, Shuff1).
2754     // Assuming that Shuff0 and Shuff1 both operate on subvectors of the
2755     // same source vector (as described by OpMap), return the index of
2756     // that source vector corresponding to Idx.
2757     ShuffleVectorSDNode *OpShuff = Idx < HwLen ? Shuff0 : Shuff1;
2758     if (Idx >= HwLen)
2759       Idx -= HwLen;
2760 
2761     // Get the mask index that M points at in the corresponding operand.
2762     int MaybeN = OpShuff->getMaskElt(Idx);
2763     if (MaybeN < 0)
2764       return -1;
2765 
2766     auto N = static_cast<unsigned>(MaybeN);
2767     unsigned SrcBase = N < HwLen ? OpMap.at(OpShuff->getOperand(0))
2768                                  : OpMap.at(OpShuff->getOperand(1));
2769     if (N >= HwLen)
2770       N -= HwLen;
2771 
2772     return N + SrcBase;
2773   };
2774 
2775   auto fold3 = [&](SDValue TopShuff, SDValue Inp, MapType &&OpMap) -> SDValue {
2776     // Fold all 3 shuffles into a single one.
2777     auto *This = cast<ShuffleVectorSDNode>(TopShuff);
2778     auto *S0 = cast<ShuffleVectorSDNode>(TopShuff.getOperand(0));
2779     auto *S1 = cast<ShuffleVectorSDNode>(TopShuff.getOperand(1));
2780     ArrayRef<int> TopMask = This->getMask();
2781     // This should be guaranteed by type checks in the caller, and the fact
2782     // that all shuffles should have been promoted to operate on MVT::i8.
2783     assert(TopMask.size() == S0->getMask().size() &&
2784            TopMask.size() == S1->getMask().size());
2785     assert(TopMask.size() == HwLen);
2786 
2787     SmallVector<int, 256> FoldedMask(2 * HwLen);
2788     for (unsigned I = 0; I != HwLen; ++I) {
2789       int MaybeM = TopMask[I];
2790       if (MaybeM >= 0) {
2791         FoldedMask[I] =
2792             getMaskElt(static_cast<unsigned>(MaybeM), S0, S1, OpMap);
2793       } else {
2794         FoldedMask[I] = -1;
2795       }
2796     }
2797     // The second half of the result will be all-undef.
2798     std::fill(FoldedMask.begin() + HwLen, FoldedMask.end(), -1);
2799 
2800     // Return
2801     //   FoldedShuffle = (Shuffle Inp, undef, FoldedMask)
2802     //   (LoHalf FoldedShuffle)
2803     const SDLoc &dl(TopShuff);
2804     MVT SingleTy = MVT::getVectorVT(MVT::i8, HwLen);
2805     MVT PairTy = MVT::getVectorVT(MVT::i8, 2 * HwLen);
2806     SDValue FoldedShuff =
2807         DAG.getVectorShuffle(PairTy, dl, DAG.getBitcast(PairTy, Inp),
2808                              DAG.getUNDEF(PairTy), FoldedMask);
2809     return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SingleTy, FoldedShuff,
2810                        DAG.getConstant(0, dl, MVT::i32));
2811   };
2812 
2813   auto getSourceInfo = [](SDValue V) -> std::optional<SubVectorInfo> {
2814     while (V.getOpcode() == ISD::BITCAST)
2815       V = V.getOperand(0);
2816     if (V.getOpcode() != ISD::EXTRACT_SUBVECTOR)
2817       return std::nullopt;
2818     return SubVectorInfo(V.getOperand(0),
2819                          !cast<ConstantSDNode>(V.getOperand(1))->isZero());
2820   };
2821 
2822   for (SDNode *N : Nodes) {
2823     if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2824       continue;
2825     EVT ResTy = N->getValueType(0);
2826     if (ResTy.getVectorElementType() != MVT::i8)
2827       continue;
2828     if (ResTy.getVectorNumElements() != HwLen)
2829       continue;
2830 
2831     SDValue V0 = N->getOperand(0);
2832     SDValue V1 = N->getOperand(1);
2833     if (V0.getOpcode() != ISD::VECTOR_SHUFFLE)
2834       continue;
2835     if (V1.getOpcode() != ISD::VECTOR_SHUFFLE)
2836       continue;
2837     if (V0.getValueType() != ResTy || V1.getValueType() != ResTy)
2838       continue;
2839 
2840     // Check if all operands of the two operand shuffles are extract_subvectors
2841     // from the same vector pair.
2842     auto V0A = getSourceInfo(V0.getOperand(0));
2843     if (!V0A.has_value())
2844       continue;
2845     auto V0B = getSourceInfo(V0.getOperand(1));
2846     if (!V0B.has_value() || V0B->Src != V0A->Src)
2847       continue;
2848     auto V1A = getSourceInfo(V1.getOperand(0));
2849     if (!V1A.has_value() || V1A->Src != V0A->Src)
2850       continue;
2851     auto V1B = getSourceInfo(V1.getOperand(1));
2852     if (!V1B.has_value() || V1B->Src != V0A->Src)
2853       continue;
2854 
2855     // The source must be a pair. This should be guaranteed here,
2856     // but check just in case.
2857     assert(V0A->Src.getValueType().getSizeInBits() == 16 * HwLen);
2858 
2859     MapType OpMap = {
2860         {V0.getOperand(0), V0A->HalfIdx * HwLen},
2861         {V0.getOperand(1), V0B->HalfIdx * HwLen},
2862         {V1.getOperand(0), V1A->HalfIdx * HwLen},
2863         {V1.getOperand(1), V1B->HalfIdx * HwLen},
2864     };
2865     SDValue NewS = fold3(SDValue(N, 0), V0A->Src, std::move(OpMap));
2866     ReplaceNode(N, NewS.getNode());
2867   }
2868 }
2869 
2870 void HexagonDAGToDAGISel::SelectHvxExtractSubvector(SDNode *N) {
2871   HvxSelector(*this, *CurDAG).selectExtractSubvector(N);
2872 }
2873 
2874 void HexagonDAGToDAGISel::SelectHvxShuffle(SDNode *N) {
2875   HvxSelector(*this, *CurDAG).selectShuffle(N);
2876 }
2877 
2878 void HexagonDAGToDAGISel::SelectHvxRor(SDNode *N) {
2879   HvxSelector(*this, *CurDAG).selectRor(N);
2880 }
2881 
2882 void HexagonDAGToDAGISel::SelectHvxVAlign(SDNode *N) {
2883   HvxSelector(*this, *CurDAG).selectVAlign(N);
2884 }
2885 
2886 void HexagonDAGToDAGISel::SelectV65GatherPred(SDNode *N) {
2887   const SDLoc &dl(N);
2888   SDValue Chain = N->getOperand(0);
2889   SDValue Address = N->getOperand(2);
2890   SDValue Predicate = N->getOperand(3);
2891   SDValue Base = N->getOperand(4);
2892   SDValue Modifier = N->getOperand(5);
2893   SDValue Offset = N->getOperand(6);
2894   SDValue ImmOperand = CurDAG->getTargetConstant(0, dl, MVT::i32);
2895 
2896   unsigned Opcode;
2897   unsigned IntNo = N->getConstantOperandVal(1);
2898   switch (IntNo) {
2899   default:
2900     llvm_unreachable("Unexpected HVX gather intrinsic.");
2901   case Intrinsic::hexagon_V6_vgathermhq:
2902   case Intrinsic::hexagon_V6_vgathermhq_128B:
2903     Opcode = Hexagon::V6_vgathermhq_pseudo;
2904     break;
2905   case Intrinsic::hexagon_V6_vgathermwq:
2906   case Intrinsic::hexagon_V6_vgathermwq_128B:
2907     Opcode = Hexagon::V6_vgathermwq_pseudo;
2908     break;
2909   case Intrinsic::hexagon_V6_vgathermhwq:
2910   case Intrinsic::hexagon_V6_vgathermhwq_128B:
2911     Opcode = Hexagon::V6_vgathermhwq_pseudo;
2912     break;
2913   }
2914 
2915   SDVTList VTs = CurDAG->getVTList(MVT::Other);
2916   SDValue Ops[] = { Address, ImmOperand,
2917                     Predicate, Base, Modifier, Offset, Chain };
2918   SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2919 
2920   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2921   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2922 
2923   ReplaceNode(N, Result);
2924 }
2925 
2926 void HexagonDAGToDAGISel::SelectV65Gather(SDNode *N) {
2927   const SDLoc &dl(N);
2928   SDValue Chain = N->getOperand(0);
2929   SDValue Address = N->getOperand(2);
2930   SDValue Base = N->getOperand(3);
2931   SDValue Modifier = N->getOperand(4);
2932   SDValue Offset = N->getOperand(5);
2933   SDValue ImmOperand = CurDAG->getTargetConstant(0, dl, MVT::i32);
2934 
2935   unsigned Opcode;
2936   unsigned IntNo = N->getConstantOperandVal(1);
2937   switch (IntNo) {
2938   default:
2939     llvm_unreachable("Unexpected HVX gather intrinsic.");
2940   case Intrinsic::hexagon_V6_vgathermh:
2941   case Intrinsic::hexagon_V6_vgathermh_128B:
2942     Opcode = Hexagon::V6_vgathermh_pseudo;
2943     break;
2944   case Intrinsic::hexagon_V6_vgathermw:
2945   case Intrinsic::hexagon_V6_vgathermw_128B:
2946     Opcode = Hexagon::V6_vgathermw_pseudo;
2947     break;
2948   case Intrinsic::hexagon_V6_vgathermhw:
2949   case Intrinsic::hexagon_V6_vgathermhw_128B:
2950     Opcode = Hexagon::V6_vgathermhw_pseudo;
2951     break;
2952   }
2953 
2954   SDVTList VTs = CurDAG->getVTList(MVT::Other);
2955   SDValue Ops[] = { Address, ImmOperand, Base, Modifier, Offset, Chain };
2956   SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2957 
2958   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2959   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2960 
2961   ReplaceNode(N, Result);
2962 }
2963 
2964 void HexagonDAGToDAGISel::SelectHVXDualOutput(SDNode *N) {
2965   unsigned IID = N->getConstantOperandVal(0);
2966   SDNode *Result;
2967   switch (IID) {
2968   case Intrinsic::hexagon_V6_vaddcarry: {
2969     std::array<SDValue, 3> Ops = {
2970         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2971     SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v64i1);
2972     Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2973     break;
2974   }
2975   case Intrinsic::hexagon_V6_vaddcarry_128B: {
2976     std::array<SDValue, 3> Ops = {
2977         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2978     SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v128i1);
2979     Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2980     break;
2981   }
2982   case Intrinsic::hexagon_V6_vsubcarry: {
2983     std::array<SDValue, 3> Ops = {
2984         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2985     SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v64i1);
2986     Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2987     break;
2988   }
2989   case Intrinsic::hexagon_V6_vsubcarry_128B: {
2990     std::array<SDValue, 3> Ops = {
2991         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2992     SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v128i1);
2993     Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2994     break;
2995   }
2996   default:
2997     llvm_unreachable("Unexpected HVX dual output intrinsic.");
2998   }
2999   ReplaceUses(N, Result);
3000   ReplaceUses(SDValue(N, 0), SDValue(Result, 0));
3001   ReplaceUses(SDValue(N, 1), SDValue(Result, 1));
3002   CurDAG->RemoveDeadNode(N);
3003 }
3004