1 //===-- HexagonISelDAGToDAGHVX.cpp ----------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "Hexagon.h"
10 #include "HexagonISelDAGToDAG.h"
11 #include "HexagonISelLowering.h"
12 #include "HexagonTargetMachine.h"
13 #include "llvm/ADT/SetVector.h"
14 #include "llvm/CodeGen/MachineInstrBuilder.h"
15 #include "llvm/CodeGen/SelectionDAGISel.h"
16 #include "llvm/IR/Intrinsics.h"
17 #include "llvm/IR/IntrinsicsHexagon.h"
18 #include "llvm/Support/CommandLine.h"
19 #include "llvm/Support/Debug.h"
20 
21 #include <deque>
22 #include <map>
23 #include <set>
24 #include <utility>
25 #include <vector>
26 
27 #define DEBUG_TYPE "hexagon-isel"
28 
29 using namespace llvm;
30 
31 namespace {
32 
33 // --------------------------------------------------------------------
34 // Implementation of permutation networks.
35 
36 // Implementation of the node routing through butterfly networks:
37 // - Forward delta.
38 // - Reverse delta.
39 // - Benes.
40 //
41 //
42 // Forward delta network consists of log(N) steps, where N is the number
43 // of inputs. In each step, an input can stay in place, or it can get
44 // routed to another position[1]. The step after that consists of two
45 // networks, each half in size in terms of the number of nodes. In those
46 // terms, in the given step, an input can go to either the upper or the
47 // lower network in the next step.
48 //
49 // [1] Hexagon's vdelta/vrdelta allow an element to be routed to both
50 // positions as long as there is no conflict.
51 
52 // Here's a delta network for 8 inputs, only the switching routes are
53 // shown:
54 //
55 //         Steps:
56 //         |- 1 ---------------|- 2 -----|- 3 -|
57 //
58 // Inp[0] ***                 ***       ***   *** Out[0]
59 //           \               /   \     /   \ /
60 //            \             /     \   /     X
61 //             \           /       \ /     / \
62 // Inp[1] ***   \         /   ***   X   ***   *** Out[1]
63 //           \   \       /   /   \ / \ /
64 //            \   \     /   /     X   X
65 //             \   \   /   /     / \ / \
66 // Inp[2] ***   \   \ /   /   ***   X   ***   *** Out[2]
67 //           \   \   X   /   /     / \     \ /
68 //            \   \ / \ /   /     /   \     X
69 //             \   X   X   /     /     \   / \
70 // Inp[3] ***   \ / \ / \ /   ***       ***   *** Out[3]
71 //           \   X   X   X   /
72 //            \ / \ / \ / \ /
73 //             X   X   X   X
74 //            / \ / \ / \ / \
75 //           /   X   X   X   \
76 // Inp[4] ***   / \ / \ / \   ***       ***   *** Out[4]
77 //             /   X   X   \     \     /   \ /
78 //            /   / \ / \   \     \   /     X
79 //           /   /   X   \   \     \ /     / \
80 // Inp[5] ***   /   / \   \   ***   X   ***   *** Out[5]
81 //             /   /   \   \     \ / \ /
82 //            /   /     \   \     X   X
83 //           /   /       \   \   / \ / \
84 // Inp[6] ***   /         \   ***   X   ***   *** Out[6]
85 //             /           \       / \     \ /
86 //            /             \     /   \     X
87 //           /               \   /     \   / \
88 // Inp[7] ***                 ***       ***   *** Out[7]
89 //
90 //
91 // Reverse delta network is same as delta network, with the steps in
92 // the opposite order.
93 //
94 //
95 // Benes network is a forward delta network immediately followed by
96 // a reverse delta network.
97 
98 enum class ColorKind { None, Red, Black };
99 
100 // Graph coloring utility used to partition nodes into two groups:
101 // they will correspond to nodes routed to the upper and lower networks.
102 struct Coloring {
103   using Node = int;
104   using MapType = std::map<Node, ColorKind>;
105   static constexpr Node Ignore = Node(-1);
106 
Coloring__anonb73ef63f0111::Coloring107   Coloring(ArrayRef<Node> Ord) : Order(Ord) {
108     build();
109     if (!color())
110       Colors.clear();
111   }
112 
colors__anonb73ef63f0111::Coloring113   const MapType &colors() const {
114     return Colors;
115   }
116 
other__anonb73ef63f0111::Coloring117   ColorKind other(ColorKind Color) {
118     if (Color == ColorKind::None)
119       return ColorKind::Red;
120     return Color == ColorKind::Red ? ColorKind::Black : ColorKind::Red;
121   }
122 
123   LLVM_DUMP_METHOD void dump() const;
124 
125 private:
126   ArrayRef<Node> Order;
127   MapType Colors;
128   std::set<Node> Needed;
129 
130   using NodeSet = std::set<Node>;
131   std::map<Node,NodeSet> Edges;
132 
conj__anonb73ef63f0111::Coloring133   Node conj(Node Pos) {
134     Node Num = Order.size();
135     return (Pos < Num/2) ? Pos + Num/2 : Pos - Num/2;
136   }
137 
getColor__anonb73ef63f0111::Coloring138   ColorKind getColor(Node N) {
139     auto F = Colors.find(N);
140     return F != Colors.end() ? F->second : ColorKind::None;
141   }
142 
143   std::pair<bool, ColorKind> getUniqueColor(const NodeSet &Nodes);
144 
145   void build();
146   bool color();
147 };
148 } // namespace
149 
getUniqueColor(const NodeSet & Nodes)150 std::pair<bool, ColorKind> Coloring::getUniqueColor(const NodeSet &Nodes) {
151   auto Color = ColorKind::None;
152   for (Node N : Nodes) {
153     ColorKind ColorN = getColor(N);
154     if (ColorN == ColorKind::None)
155       continue;
156     if (Color == ColorKind::None)
157       Color = ColorN;
158     else if (Color != ColorKind::None && Color != ColorN)
159       return { false, ColorKind::None };
160   }
161   return { true, Color };
162 }
163 
build()164 void Coloring::build() {
165   // Add Order[P] and Order[conj(P)] to Edges.
166   for (unsigned P = 0; P != Order.size(); ++P) {
167     Node I = Order[P];
168     if (I != Ignore) {
169       Needed.insert(I);
170       Node PC = Order[conj(P)];
171       if (PC != Ignore && PC != I)
172         Edges[I].insert(PC);
173     }
174   }
175   // Add I and conj(I) to Edges.
176   for (unsigned I = 0; I != Order.size(); ++I) {
177     if (!Needed.count(I))
178       continue;
179     Node C = conj(I);
180     // This will create an entry in the edge table, even if I is not
181     // connected to any other node. This is necessary, because it still
182     // needs to be colored.
183     NodeSet &Is = Edges[I];
184     if (Needed.count(C))
185       Is.insert(C);
186   }
187 }
188 
color()189 bool Coloring::color() {
190   SetVector<Node> FirstQ;
191   auto Enqueue = [this,&FirstQ] (Node N) {
192     SetVector<Node> Q;
193     Q.insert(N);
194     for (unsigned I = 0; I != Q.size(); ++I) {
195       NodeSet &Ns = Edges[Q[I]];
196       Q.insert(Ns.begin(), Ns.end());
197     }
198     FirstQ.insert(Q.begin(), Q.end());
199   };
200   for (Node N : Needed)
201     Enqueue(N);
202 
203   for (Node N : FirstQ) {
204     if (Colors.count(N))
205       continue;
206     NodeSet &Ns = Edges[N];
207     auto P = getUniqueColor(Ns);
208     if (!P.first)
209       return false;
210     Colors[N] = other(P.second);
211   }
212 
213   // First, color nodes that don't have any dups.
214   for (auto E : Edges) {
215     Node N = E.first;
216     if (!Needed.count(conj(N)) || Colors.count(N))
217       continue;
218     auto P = getUniqueColor(E.second);
219     if (!P.first)
220       return false;
221     Colors[N] = other(P.second);
222   }
223 
224   // Now, nodes that are still uncolored. Since the graph can be modified
225   // in this step, create a work queue.
226   std::vector<Node> WorkQ;
227   for (auto E : Edges) {
228     Node N = E.first;
229     if (!Colors.count(N))
230       WorkQ.push_back(N);
231   }
232 
233   for (unsigned I = 0; I < WorkQ.size(); ++I) {
234     Node N = WorkQ[I];
235     NodeSet &Ns = Edges[N];
236     auto P = getUniqueColor(Ns);
237     if (P.first) {
238       Colors[N] = other(P.second);
239       continue;
240     }
241 
242     // Coloring failed. Split this node.
243     Node C = conj(N);
244     ColorKind ColorN = other(ColorKind::None);
245     ColorKind ColorC = other(ColorN);
246     NodeSet &Cs = Edges[C];
247     NodeSet CopyNs = Ns;
248     for (Node M : CopyNs) {
249       ColorKind ColorM = getColor(M);
250       if (ColorM == ColorC) {
251         // Connect M with C, disconnect M from N.
252         Cs.insert(M);
253         Edges[M].insert(C);
254         Ns.erase(M);
255         Edges[M].erase(N);
256       }
257     }
258     Colors[N] = ColorN;
259     Colors[C] = ColorC;
260   }
261 
262   // Explicitly assign "None" to all uncolored nodes.
263   for (unsigned I = 0; I != Order.size(); ++I)
264     if (Colors.count(I) == 0)
265       Colors[I] = ColorKind::None;
266 
267   return true;
268 }
269 
270 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump() const271 void Coloring::dump() const {
272   dbgs() << "{ Order:   {";
273   for (unsigned I = 0; I != Order.size(); ++I) {
274     Node P = Order[I];
275     if (P != Ignore)
276       dbgs() << ' ' << P;
277     else
278       dbgs() << " -";
279   }
280   dbgs() << " }\n";
281   dbgs() << "  Needed: {";
282   for (Node N : Needed)
283     dbgs() << ' ' << N;
284   dbgs() << " }\n";
285 
286   dbgs() << "  Edges: {\n";
287   for (auto E : Edges) {
288     dbgs() << "    " << E.first << " -> {";
289     for (auto N : E.second)
290       dbgs() << ' ' << N;
291     dbgs() << " }\n";
292   }
293   dbgs() << "  }\n";
294 
295   auto ColorKindToName = [](ColorKind C) {
296     switch (C) {
297     case ColorKind::None:
298       return "None";
299     case ColorKind::Red:
300       return "Red";
301     case ColorKind::Black:
302       return "Black";
303     }
304     llvm_unreachable("all ColorKinds should be handled by the switch above");
305   };
306 
307   dbgs() << "  Colors: {\n";
308   for (auto C : Colors)
309     dbgs() << "    " << C.first << " -> " << ColorKindToName(C.second) << "\n";
310   dbgs() << "  }\n}\n";
311 }
312 #endif
313 
314 namespace {
315 // Base class of for reordering networks. They don't strictly need to be
316 // permutations, as outputs with repeated occurrences of an input element
317 // are allowed.
318 struct PermNetwork {
319   using Controls = std::vector<uint8_t>;
320   using ElemType = int;
321   static constexpr ElemType Ignore = ElemType(-1);
322 
323   enum : uint8_t {
324     None,
325     Pass,
326     Switch
327   };
328   enum : uint8_t {
329     Forward,
330     Reverse
331   };
332 
PermNetwork__anonb73ef63f0411::PermNetwork333   PermNetwork(ArrayRef<ElemType> Ord, unsigned Mult = 1) {
334     Order.assign(Ord.data(), Ord.data()+Ord.size());
335     Log = 0;
336 
337     unsigned S = Order.size();
338     while (S >>= 1)
339       ++Log;
340 
341     Table.resize(Order.size());
342     for (RowType &Row : Table)
343       Row.resize(Mult*Log, None);
344   }
345 
getControls__anonb73ef63f0411::PermNetwork346   void getControls(Controls &V, unsigned StartAt, uint8_t Dir) const {
347     unsigned Size = Order.size();
348     V.resize(Size);
349     for (unsigned I = 0; I != Size; ++I) {
350       unsigned W = 0;
351       for (unsigned L = 0; L != Log; ++L) {
352         unsigned C = ctl(I, StartAt+L) == Switch;
353         if (Dir == Forward)
354           W |= C << (Log-1-L);
355         else
356           W |= C << L;
357       }
358       assert(isUInt<8>(W));
359       V[I] = uint8_t(W);
360     }
361   }
362 
ctl__anonb73ef63f0411::PermNetwork363   uint8_t ctl(ElemType Pos, unsigned Step) const {
364     return Table[Pos][Step];
365   }
size__anonb73ef63f0411::PermNetwork366   unsigned size() const {
367     return Order.size();
368   }
steps__anonb73ef63f0411::PermNetwork369   unsigned steps() const {
370     return Log;
371   }
372 
373 protected:
374   unsigned Log;
375   std::vector<ElemType> Order;
376   using RowType = std::vector<uint8_t>;
377   std::vector<RowType> Table;
378 };
379 
380 struct ForwardDeltaNetwork : public PermNetwork {
ForwardDeltaNetwork__anonb73ef63f0411::ForwardDeltaNetwork381   ForwardDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
382 
run__anonb73ef63f0411::ForwardDeltaNetwork383   bool run(Controls &V) {
384     if (!route(Order.data(), Table.data(), size(), 0))
385       return false;
386     getControls(V, 0, Forward);
387     return true;
388   }
389 
390 private:
391   bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
392 };
393 
394 struct ReverseDeltaNetwork : public PermNetwork {
ReverseDeltaNetwork__anonb73ef63f0411::ReverseDeltaNetwork395   ReverseDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
396 
run__anonb73ef63f0411::ReverseDeltaNetwork397   bool run(Controls &V) {
398     if (!route(Order.data(), Table.data(), size(), 0))
399       return false;
400     getControls(V, 0, Reverse);
401     return true;
402   }
403 
404 private:
405   bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
406 };
407 
408 struct BenesNetwork : public PermNetwork {
BenesNetwork__anonb73ef63f0411::BenesNetwork409   BenesNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord, 2) {}
410 
run__anonb73ef63f0411::BenesNetwork411   bool run(Controls &F, Controls &R) {
412     if (!route(Order.data(), Table.data(), size(), 0))
413       return false;
414 
415     getControls(F, 0, Forward);
416     getControls(R, Log, Reverse);
417     return true;
418   }
419 
420 private:
421   bool route(ElemType *P, RowType *T, unsigned Size, unsigned Step);
422 };
423 } // namespace
424 
route(ElemType * P,RowType * T,unsigned Size,unsigned Step)425 bool ForwardDeltaNetwork::route(ElemType *P, RowType *T, unsigned Size,
426                                 unsigned Step) {
427   bool UseUp = false, UseDown = false;
428   ElemType Num = Size;
429 
430   // Cannot use coloring here, because coloring is used to determine
431   // the "big" switch, i.e. the one that changes halves, and in a forward
432   // network, a color can be simultaneously routed to both halves in the
433   // step we're working on.
434   for (ElemType J = 0; J != Num; ++J) {
435     ElemType I = P[J];
436     // I is the position in the input,
437     // J is the position in the output.
438     if (I == Ignore)
439       continue;
440     uint8_t S;
441     if (I < Num/2)
442       S = (J < Num/2) ? Pass : Switch;
443     else
444       S = (J < Num/2) ? Switch : Pass;
445 
446     // U is the element in the table that needs to be updated.
447     ElemType U = (S == Pass) ? I : (I < Num/2 ? I+Num/2 : I-Num/2);
448     if (U < Num/2)
449       UseUp = true;
450     else
451       UseDown = true;
452     if (T[U][Step] != S && T[U][Step] != None)
453       return false;
454     T[U][Step] = S;
455   }
456 
457   for (ElemType J = 0; J != Num; ++J)
458     if (P[J] != Ignore && P[J] >= Num/2)
459       P[J] -= Num/2;
460 
461   if (Step+1 < Log) {
462     if (UseUp   && !route(P,        T,        Size/2, Step+1))
463       return false;
464     if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
465       return false;
466   }
467   return true;
468 }
469 
route(ElemType * P,RowType * T,unsigned Size,unsigned Step)470 bool ReverseDeltaNetwork::route(ElemType *P, RowType *T, unsigned Size,
471                                 unsigned Step) {
472   unsigned Pets = Log-1 - Step;
473   bool UseUp = false, UseDown = false;
474   ElemType Num = Size;
475 
476   // In this step half-switching occurs, so coloring can be used.
477   Coloring G({P,Size});
478   const Coloring::MapType &M = G.colors();
479   if (M.empty())
480     return false;
481 
482   ColorKind ColorUp = ColorKind::None;
483   for (ElemType J = 0; J != Num; ++J) {
484     ElemType I = P[J];
485     // I is the position in the input,
486     // J is the position in the output.
487     if (I == Ignore)
488       continue;
489     ColorKind C = M.at(I);
490     if (C == ColorKind::None)
491       continue;
492     // During "Step", inputs cannot switch halves, so if the "up" color
493     // is still unknown, make sure that it is selected in such a way that
494     // "I" will stay in the same half.
495     bool InpUp = I < Num/2;
496     if (ColorUp == ColorKind::None)
497       ColorUp = InpUp ? C : G.other(C);
498     if ((C == ColorUp) != InpUp) {
499       // If I should go to a different half than where is it now, give up.
500       return false;
501     }
502 
503     uint8_t S;
504     if (InpUp) {
505       S = (J < Num/2) ? Pass : Switch;
506       UseUp = true;
507     } else {
508       S = (J < Num/2) ? Switch : Pass;
509       UseDown = true;
510     }
511     T[J][Pets] = S;
512   }
513 
514   // Reorder the working permutation according to the computed switch table
515   // for the last step (i.e. Pets).
516   for (ElemType J = 0, E = Size / 2; J != E; ++J) {
517     ElemType PJ = P[J];         // Current values of P[J]
518     ElemType PC = P[J+Size/2];  // and P[conj(J)]
519     ElemType QJ = PJ;           // New values of P[J]
520     ElemType QC = PC;           // and P[conj(J)]
521     if (T[J][Pets] == Switch)
522       QC = PJ;
523     if (T[J+Size/2][Pets] == Switch)
524       QJ = PC;
525     P[J] = QJ;
526     P[J+Size/2] = QC;
527   }
528 
529   for (ElemType J = 0; J != Num; ++J)
530     if (P[J] != Ignore && P[J] >= Num/2)
531       P[J] -= Num/2;
532 
533   if (Step+1 < Log) {
534     if (UseUp && !route(P, T, Size/2, Step+1))
535       return false;
536     if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
537       return false;
538   }
539   return true;
540 }
541 
route(ElemType * P,RowType * T,unsigned Size,unsigned Step)542 bool BenesNetwork::route(ElemType *P, RowType *T, unsigned Size,
543                          unsigned Step) {
544   Coloring G({P,Size});
545   const Coloring::MapType &M = G.colors();
546   if (M.empty())
547     return false;
548   ElemType Num = Size;
549 
550   unsigned Pets = 2*Log-1 - Step;
551   bool UseUp = false, UseDown = false;
552 
553   // Both assignments, i.e. Red->Up and Red->Down are valid, but they will
554   // result in different controls. Let's pick the one where the first
555   // control will be "Pass".
556   ColorKind ColorUp = ColorKind::None;
557   for (ElemType J = 0; J != Num; ++J) {
558     ElemType I = P[J];
559     if (I == Ignore)
560       continue;
561     ColorKind C = M.at(I);
562     if (C == ColorKind::None)
563       continue;
564     if (ColorUp == ColorKind::None) {
565       ColorUp = (I < Num / 2) ? ColorKind::Red : ColorKind::Black;
566     }
567     unsigned CI = (I < Num/2) ? I+Num/2 : I-Num/2;
568     if (C == ColorUp) {
569       if (I < Num/2)
570         T[I][Step] = Pass;
571       else
572         T[CI][Step] = Switch;
573       T[J][Pets] = (J < Num/2) ? Pass : Switch;
574       UseUp = true;
575     } else { // Down
576       if (I < Num/2)
577         T[CI][Step] = Switch;
578       else
579         T[I][Step] = Pass;
580       T[J][Pets] = (J < Num/2) ? Switch : Pass;
581       UseDown = true;
582     }
583   }
584 
585   // Reorder the working permutation according to the computed switch table
586   // for the last step (i.e. Pets).
587   for (ElemType J = 0; J != Num/2; ++J) {
588     ElemType PJ = P[J];         // Current values of P[J]
589     ElemType PC = P[J+Num/2];   // and P[conj(J)]
590     ElemType QJ = PJ;           // New values of P[J]
591     ElemType QC = PC;           // and P[conj(J)]
592     if (T[J][Pets] == Switch)
593       QC = PJ;
594     if (T[J+Num/2][Pets] == Switch)
595       QJ = PC;
596     P[J] = QJ;
597     P[J+Num/2] = QC;
598   }
599 
600   for (ElemType J = 0; J != Num; ++J)
601     if (P[J] != Ignore && P[J] >= Num/2)
602       P[J] -= Num/2;
603 
604   if (Step+1 < Log) {
605     if (UseUp && !route(P, T, Size/2, Step+1))
606       return false;
607     if (UseDown && !route(P+Size/2, T+Size/2, Size/2, Step+1))
608       return false;
609   }
610   return true;
611 }
612 
613 // --------------------------------------------------------------------
614 // Support for building selection results (output instructions that are
615 // parts of the final selection).
616 
617 namespace {
618 struct OpRef {
OpRef__anonb73ef63f0711::OpRef619   OpRef(SDValue V) : OpV(V) {}
isValue__anonb73ef63f0711::OpRef620   bool isValue() const { return OpV.getNode() != nullptr; }
isValid__anonb73ef63f0711::OpRef621   bool isValid() const { return isValue() || !(OpN & Invalid); }
res__anonb73ef63f0711::OpRef622   static OpRef res(int N) { return OpRef(Whole | (N & Index)); }
fail__anonb73ef63f0711::OpRef623   static OpRef fail() { return OpRef(Invalid); }
624 
lo__anonb73ef63f0711::OpRef625   static OpRef lo(const OpRef &R) {
626     assert(!R.isValue());
627     return OpRef(R.OpN & (Undef | Index | LoHalf));
628   }
hi__anonb73ef63f0711::OpRef629   static OpRef hi(const OpRef &R) {
630     assert(!R.isValue());
631     return OpRef(R.OpN & (Undef | Index | HiHalf));
632   }
undef__anonb73ef63f0711::OpRef633   static OpRef undef(MVT Ty) { return OpRef(Undef | Ty.SimpleTy); }
634 
635   // Direct value.
636   SDValue OpV = SDValue();
637 
638   // Reference to the operand of the input node:
639   // If the 31st bit is 1, it's undef, otherwise, bits 28..0 are the
640   // operand index:
641   // If bit 30 is set, it's the high half of the operand.
642   // If bit 29 is set, it's the low half of the operand.
643   unsigned OpN = 0;
644 
645   enum : unsigned {
646     Invalid = 0x10000000,
647     LoHalf  = 0x20000000,
648     HiHalf  = 0x40000000,
649     Whole   = LoHalf | HiHalf,
650     Undef   = 0x80000000,
651     Index   = 0x0FFFFFFF,  // Mask of the index value.
652     IndexBits = 28,
653   };
654 
655   LLVM_DUMP_METHOD
656   void print(raw_ostream &OS, const SelectionDAG &G) const;
657 
658 private:
OpRef__anonb73ef63f0711::OpRef659   OpRef(unsigned N) : OpN(N) {}
660 };
661 
662 struct NodeTemplate {
663   NodeTemplate() = default;
664   unsigned Opc = 0;
665   MVT Ty = MVT::Other;
666   std::vector<OpRef> Ops;
667 
668   LLVM_DUMP_METHOD void print(raw_ostream &OS, const SelectionDAG &G) const;
669 };
670 
671 struct ResultStack {
ResultStack__anonb73ef63f0711::ResultStack672   ResultStack(SDNode *Inp)
673     : InpNode(Inp), InpTy(Inp->getValueType(0).getSimpleVT()) {}
674   SDNode *InpNode;
675   MVT InpTy;
push__anonb73ef63f0711::ResultStack676   unsigned push(const NodeTemplate &Res) {
677     List.push_back(Res);
678     return List.size()-1;
679   }
push__anonb73ef63f0711::ResultStack680   unsigned push(unsigned Opc, MVT Ty, std::vector<OpRef> &&Ops) {
681     NodeTemplate Res;
682     Res.Opc = Opc;
683     Res.Ty = Ty;
684     Res.Ops = Ops;
685     return push(Res);
686   }
empty__anonb73ef63f0711::ResultStack687   bool empty() const { return List.empty(); }
size__anonb73ef63f0711::ResultStack688   unsigned size() const { return List.size(); }
top__anonb73ef63f0711::ResultStack689   unsigned top() const { return size()-1; }
operator []__anonb73ef63f0711::ResultStack690   const NodeTemplate &operator[](unsigned I) const { return List[I]; }
reset__anonb73ef63f0711::ResultStack691   unsigned reset(unsigned NewTop) {
692     List.resize(NewTop+1);
693     return NewTop;
694   }
695 
696   using BaseType = std::vector<NodeTemplate>;
begin__anonb73ef63f0711::ResultStack697   BaseType::iterator begin() { return List.begin(); }
end__anonb73ef63f0711::ResultStack698   BaseType::iterator end()   { return List.end(); }
begin__anonb73ef63f0711::ResultStack699   BaseType::const_iterator begin() const { return List.begin(); }
end__anonb73ef63f0711::ResultStack700   BaseType::const_iterator end() const   { return List.end(); }
701 
702   BaseType List;
703 
704   LLVM_DUMP_METHOD
705   void print(raw_ostream &OS, const SelectionDAG &G) const;
706 };
707 } // namespace
708 
709 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS,const SelectionDAG & G) const710 void OpRef::print(raw_ostream &OS, const SelectionDAG &G) const {
711   if (isValue()) {
712     OpV.getNode()->print(OS, &G);
713     return;
714   }
715   if (OpN & Invalid) {
716     OS << "invalid";
717     return;
718   }
719   if (OpN & Undef) {
720     OS << "undef";
721     return;
722   }
723   if ((OpN & Whole) != Whole) {
724     assert((OpN & Whole) == LoHalf || (OpN & Whole) == HiHalf);
725     if (OpN & LoHalf)
726       OS << "lo ";
727     else
728       OS << "hi ";
729   }
730   OS << '#' << SignExtend32(OpN & Index, IndexBits);
731 }
732 
print(raw_ostream & OS,const SelectionDAG & G) const733 void NodeTemplate::print(raw_ostream &OS, const SelectionDAG &G) const {
734   const TargetInstrInfo &TII = *G.getSubtarget().getInstrInfo();
735   OS << format("%8s", EVT(Ty).getEVTString().c_str()) << "  "
736      << TII.getName(Opc);
737   bool Comma = false;
738   for (const auto &R : Ops) {
739     if (Comma)
740       OS << ',';
741     Comma = true;
742     OS << ' ';
743     R.print(OS, G);
744   }
745 }
746 
print(raw_ostream & OS,const SelectionDAG & G) const747 void ResultStack::print(raw_ostream &OS, const SelectionDAG &G) const {
748   OS << "Input node:\n";
749 #ifndef NDEBUG
750   InpNode->dumpr(&G);
751 #endif
752   OS << "Result templates:\n";
753   for (unsigned I = 0, E = List.size(); I != E; ++I) {
754     OS << '[' << I << "] ";
755     List[I].print(OS, G);
756     OS << '\n';
757   }
758 }
759 #endif
760 
761 namespace {
762 struct ShuffleMask {
ShuffleMask__anonb73ef63f0911::ShuffleMask763   ShuffleMask(ArrayRef<int> M) : Mask(M) {
764     for (unsigned I = 0, E = Mask.size(); I != E; ++I) {
765       int M = Mask[I];
766       if (M == -1)
767         continue;
768       MinSrc = (MinSrc == -1) ? M : std::min(MinSrc, M);
769       MaxSrc = (MaxSrc == -1) ? M : std::max(MaxSrc, M);
770     }
771   }
772 
773   ArrayRef<int> Mask;
774   int MinSrc = -1, MaxSrc = -1;
775 
lo__anonb73ef63f0911::ShuffleMask776   ShuffleMask lo() const {
777     size_t H = Mask.size()/2;
778     return ShuffleMask(Mask.take_front(H));
779   }
hi__anonb73ef63f0911::ShuffleMask780   ShuffleMask hi() const {
781     size_t H = Mask.size()/2;
782     return ShuffleMask(Mask.take_back(H));
783   }
784 
print__anonb73ef63f0911::ShuffleMask785   void print(raw_ostream &OS) const {
786     OS << "MinSrc:" << MinSrc << ", MaxSrc:" << MaxSrc << " {";
787     for (int M : Mask)
788       OS << ' ' << M;
789     OS << " }";
790   }
791 };
792 
793 LLVM_ATTRIBUTE_UNUSED
operator <<(raw_ostream & OS,const ShuffleMask & SM)794 raw_ostream &operator<<(raw_ostream &OS, const ShuffleMask &SM) {
795   SM.print(OS);
796   return OS;
797 }
798 } // namespace
799 
800 // --------------------------------------------------------------------
801 // The HvxSelector class.
802 
getHexagonLowering(SelectionDAG & G)803 static const HexagonTargetLowering &getHexagonLowering(SelectionDAG &G) {
804   return static_cast<const HexagonTargetLowering&>(G.getTargetLoweringInfo());
805 }
getHexagonSubtarget(SelectionDAG & G)806 static const HexagonSubtarget &getHexagonSubtarget(SelectionDAG &G) {
807   return static_cast<const HexagonSubtarget&>(G.getSubtarget());
808 }
809 
810 namespace llvm {
811   struct HvxSelector {
812     const HexagonTargetLowering &Lower;
813     HexagonDAGToDAGISel &ISel;
814     SelectionDAG &DAG;
815     const HexagonSubtarget &HST;
816     const unsigned HwLen;
817 
HvxSelectorllvm::HvxSelector818     HvxSelector(HexagonDAGToDAGISel &HS, SelectionDAG &G)
819       : Lower(getHexagonLowering(G)),  ISel(HS), DAG(G),
820         HST(getHexagonSubtarget(G)), HwLen(HST.getVectorLength()) {}
821 
getSingleVTllvm::HvxSelector822     MVT getSingleVT(MVT ElemTy) const {
823       unsigned NumElems = HwLen / (ElemTy.getSizeInBits()/8);
824       return MVT::getVectorVT(ElemTy, NumElems);
825     }
826 
getPairVTllvm::HvxSelector827     MVT getPairVT(MVT ElemTy) const {
828       unsigned NumElems = (2*HwLen) / (ElemTy.getSizeInBits()/8);
829       return MVT::getVectorVT(ElemTy, NumElems);
830     }
831 
832     void selectShuffle(SDNode *N);
833     void selectRor(SDNode *N);
834     void selectVAlign(SDNode *N);
835 
836   private:
837     void select(SDNode *ISelN);
838     void materialize(const ResultStack &Results);
839 
840     SDValue getVectorConstant(ArrayRef<uint8_t> Data, const SDLoc &dl);
841 
842     enum : unsigned {
843       None,
844       PackMux,
845     };
846     OpRef concat(OpRef Va, OpRef Vb, ResultStack &Results);
847     OpRef packs(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results,
848                 MutableArrayRef<int> NewMask, unsigned Options = None);
849     OpRef packp(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results,
850                 MutableArrayRef<int> NewMask);
851     OpRef vmuxs(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
852                 ResultStack &Results);
853     OpRef vmuxp(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
854                 ResultStack &Results);
855 
856     OpRef shuffs1(ShuffleMask SM, OpRef Va, ResultStack &Results);
857     OpRef shuffs2(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
858     OpRef shuffp1(ShuffleMask SM, OpRef Va, ResultStack &Results);
859     OpRef shuffp2(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
860 
861     OpRef butterfly(ShuffleMask SM, OpRef Va, ResultStack &Results);
862     OpRef contracting(ShuffleMask SM, OpRef Va, OpRef Vb, ResultStack &Results);
863     OpRef expanding(ShuffleMask SM, OpRef Va, ResultStack &Results);
864     OpRef perfect(ShuffleMask SM, OpRef Va, ResultStack &Results);
865 
866     bool selectVectorConstants(SDNode *N);
867     bool scalarizeShuffle(ArrayRef<int> Mask, const SDLoc &dl, MVT ResTy,
868                           SDValue Va, SDValue Vb, SDNode *N);
869 
870   };
871 }
872 
splitMask(ArrayRef<int> Mask,MutableArrayRef<int> MaskL,MutableArrayRef<int> MaskR)873 static void splitMask(ArrayRef<int> Mask, MutableArrayRef<int> MaskL,
874                       MutableArrayRef<int> MaskR) {
875   unsigned VecLen = Mask.size();
876   assert(MaskL.size() == VecLen && MaskR.size() == VecLen);
877   for (unsigned I = 0; I != VecLen; ++I) {
878     int M = Mask[I];
879     if (M < 0) {
880       MaskL[I] = MaskR[I] = -1;
881     } else if (unsigned(M) < VecLen) {
882       MaskL[I] = M;
883       MaskR[I] = -1;
884     } else {
885       MaskL[I] = -1;
886       MaskR[I] = M-VecLen;
887     }
888   }
889 }
890 
findStrip(ArrayRef<int> A,int Inc,unsigned MaxLen)891 static std::pair<int,unsigned> findStrip(ArrayRef<int> A, int Inc,
892                                          unsigned MaxLen) {
893   assert(A.size() > 0 && A.size() >= MaxLen);
894   int F = A[0];
895   int E = F;
896   for (unsigned I = 1; I != MaxLen; ++I) {
897     if (A[I] - E != Inc)
898       return { F, I };
899     E = A[I];
900   }
901   return { F, MaxLen };
902 }
903 
isUndef(ArrayRef<int> Mask)904 static bool isUndef(ArrayRef<int> Mask) {
905   for (int Idx : Mask)
906     if (Idx != -1)
907       return false;
908   return true;
909 }
910 
isIdentity(ArrayRef<int> Mask)911 static bool isIdentity(ArrayRef<int> Mask) {
912   for (int I = 0, E = Mask.size(); I != E; ++I) {
913     int M = Mask[I];
914     if (M >= 0 && M != I)
915       return false;
916   }
917   return true;
918 }
919 
isPermutation(ArrayRef<int> Mask)920 static bool isPermutation(ArrayRef<int> Mask) {
921   // Check by adding all numbers only works if there is no overflow.
922   assert(Mask.size() < 0x00007FFF && "Sanity failure");
923   int Sum = 0;
924   for (int Idx : Mask) {
925     if (Idx == -1)
926       return false;
927     Sum += Idx;
928   }
929   int N = Mask.size();
930   return 2*Sum == N*(N-1);
931 }
932 
selectVectorConstants(SDNode * N)933 bool HvxSelector::selectVectorConstants(SDNode *N) {
934   // Constant vectors are generated as loads from constant pools or as
935   // splats of a constant value. Since they are generated during the
936   // selection process, the main selection algorithm is not aware of them.
937   // Select them directly here.
938   SmallVector<SDNode*,4> Nodes;
939   SetVector<SDNode*> WorkQ;
940 
941   // The DAG can change (due to CSE) during selection, so cache all the
942   // unselected nodes first to avoid traversing a mutating DAG.
943   WorkQ.insert(N);
944   for (unsigned i = 0; i != WorkQ.size(); ++i) {
945     SDNode *W = WorkQ[i];
946     if (!W->isMachineOpcode() && W->getOpcode() == HexagonISD::ISEL)
947       Nodes.push_back(W);
948     for (unsigned j = 0, f = W->getNumOperands(); j != f; ++j)
949       WorkQ.insert(W->getOperand(j).getNode());
950   }
951 
952   for (SDNode *L : Nodes)
953     select(L);
954 
955   return !Nodes.empty();
956 }
957 
materialize(const ResultStack & Results)958 void HvxSelector::materialize(const ResultStack &Results) {
959   DEBUG_WITH_TYPE("isel", {
960     dbgs() << "Materializing\n";
961     Results.print(dbgs(), DAG);
962   });
963   if (Results.empty())
964     return;
965   const SDLoc &dl(Results.InpNode);
966   std::vector<SDValue> Output;
967 
968   for (unsigned I = 0, E = Results.size(); I != E; ++I) {
969     const NodeTemplate &Node = Results[I];
970     std::vector<SDValue> Ops;
971     for (const OpRef &R : Node.Ops) {
972       assert(R.isValid());
973       if (R.isValue()) {
974         Ops.push_back(R.OpV);
975         continue;
976       }
977       if (R.OpN & OpRef::Undef) {
978         MVT::SimpleValueType SVT = MVT::SimpleValueType(R.OpN & OpRef::Index);
979         Ops.push_back(ISel.selectUndef(dl, MVT(SVT)));
980         continue;
981       }
982       // R is an index of a result.
983       unsigned Part = R.OpN & OpRef::Whole;
984       int Idx = SignExtend32(R.OpN & OpRef::Index, OpRef::IndexBits);
985       if (Idx < 0)
986         Idx += I;
987       assert(Idx >= 0 && unsigned(Idx) < Output.size());
988       SDValue Op = Output[Idx];
989       MVT OpTy = Op.getValueType().getSimpleVT();
990       if (Part != OpRef::Whole) {
991         assert(Part == OpRef::LoHalf || Part == OpRef::HiHalf);
992         MVT HalfTy = MVT::getVectorVT(OpTy.getVectorElementType(),
993                                       OpTy.getVectorNumElements()/2);
994         unsigned Sub = (Part == OpRef::LoHalf) ? Hexagon::vsub_lo
995                                                : Hexagon::vsub_hi;
996         Op = DAG.getTargetExtractSubreg(Sub, dl, HalfTy, Op);
997       }
998       Ops.push_back(Op);
999     } // for (Node : Results)
1000 
1001     assert(Node.Ty != MVT::Other);
1002     SDNode *ResN = (Node.Opc == TargetOpcode::COPY)
1003                       ? Ops.front().getNode()
1004                       : DAG.getMachineNode(Node.Opc, dl, Node.Ty, Ops);
1005     Output.push_back(SDValue(ResN, 0));
1006   }
1007 
1008   SDNode *OutN = Output.back().getNode();
1009   SDNode *InpN = Results.InpNode;
1010   DEBUG_WITH_TYPE("isel", {
1011     dbgs() << "Generated node:\n";
1012     OutN->dumpr(&DAG);
1013   });
1014 
1015   ISel.ReplaceNode(InpN, OutN);
1016   selectVectorConstants(OutN);
1017   DAG.RemoveDeadNodes();
1018 }
1019 
concat(OpRef Lo,OpRef Hi,ResultStack & Results)1020 OpRef HvxSelector::concat(OpRef Lo, OpRef Hi, ResultStack &Results) {
1021   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1022   const SDLoc &dl(Results.InpNode);
1023   Results.push(TargetOpcode::REG_SEQUENCE, getPairVT(MVT::i8), {
1024     DAG.getTargetConstant(Hexagon::HvxWRRegClassID, dl, MVT::i32),
1025     Lo, DAG.getTargetConstant(Hexagon::vsub_lo, dl, MVT::i32),
1026     Hi, DAG.getTargetConstant(Hexagon::vsub_hi, dl, MVT::i32),
1027   });
1028   return OpRef::res(Results.top());
1029 }
1030 
1031 // Va, Vb are single vectors, SM can be arbitrarily long.
packs(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results,MutableArrayRef<int> NewMask,unsigned Options)1032 OpRef HvxSelector::packs(ShuffleMask SM, OpRef Va, OpRef Vb,
1033                          ResultStack &Results, MutableArrayRef<int> NewMask,
1034                          unsigned Options) {
1035   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1036   if (!Va.isValid() || !Vb.isValid())
1037     return OpRef::fail();
1038 
1039   int VecLen = SM.Mask.size();
1040   MVT Ty = getSingleVT(MVT::i8);
1041 
1042   auto IsExtSubvector = [] (ShuffleMask M) {
1043     assert(M.MinSrc >= 0 && M.MaxSrc >= 0);
1044     for (int I = 0, E = M.Mask.size(); I != E; ++I) {
1045       if (M.Mask[I] >= 0 && M.Mask[I]-I != M.MinSrc)
1046         return false;
1047     }
1048     return true;
1049   };
1050 
1051   if (SM.MaxSrc - SM.MinSrc < int(HwLen)) {
1052     if (SM.MinSrc == 0 || SM.MinSrc == int(HwLen) || !IsExtSubvector(SM)) {
1053       // If the mask picks elements from only one of the operands, return
1054       // that operand, and update the mask to use index 0 to refer to the
1055       // first element of that operand.
1056       // If the mask extracts a subvector, it will be handled below, so
1057       // skip it here.
1058       if (SM.MaxSrc < int(HwLen)) {
1059         memcpy(NewMask.data(), SM.Mask.data(), sizeof(int)*VecLen);
1060         return Va;
1061       }
1062       if (SM.MinSrc >= int(HwLen)) {
1063         for (int I = 0; I != VecLen; ++I) {
1064           int M = SM.Mask[I];
1065           if (M != -1)
1066             M -= HwLen;
1067           NewMask[I] = M;
1068         }
1069         return Vb;
1070       }
1071     }
1072     int MinSrc = SM.MinSrc;
1073     if (SM.MaxSrc < int(HwLen)) {
1074       Vb = Va;
1075     } else if (SM.MinSrc > int(HwLen)) {
1076       Va = Vb;
1077       MinSrc = SM.MinSrc - HwLen;
1078     }
1079     const SDLoc &dl(Results.InpNode);
1080     if (isUInt<3>(MinSrc) || isUInt<3>(HwLen-MinSrc)) {
1081       bool IsRight = isUInt<3>(MinSrc); // Right align.
1082       SDValue S = DAG.getTargetConstant(IsRight ? MinSrc : HwLen-MinSrc,
1083                                         dl, MVT::i32);
1084       unsigned Opc = IsRight ? Hexagon::V6_valignbi
1085                              : Hexagon::V6_vlalignbi;
1086       Results.push(Opc, Ty, {Vb, Va, S});
1087     } else {
1088       SDValue S = DAG.getTargetConstant(MinSrc, dl, MVT::i32);
1089       Results.push(Hexagon::A2_tfrsi, MVT::i32, {S});
1090       unsigned Top = Results.top();
1091       Results.push(Hexagon::V6_valignb, Ty, {Vb, Va, OpRef::res(Top)});
1092     }
1093     for (int I = 0; I != VecLen; ++I) {
1094       int M = SM.Mask[I];
1095       if (M != -1)
1096         M -= SM.MinSrc;
1097       NewMask[I] = M;
1098     }
1099     return OpRef::res(Results.top());
1100   }
1101 
1102   if (Options & PackMux) {
1103     // If elements picked from Va and Vb have all different (source) indexes
1104     // (relative to the start of the argument), do a mux, and update the mask.
1105     BitVector Picked(HwLen);
1106     SmallVector<uint8_t,128> MuxBytes(HwLen);
1107     bool CanMux = true;
1108     for (int I = 0; I != VecLen; ++I) {
1109       int M = SM.Mask[I];
1110       if (M == -1)
1111         continue;
1112       if (M >= int(HwLen))
1113         M -= HwLen;
1114       else
1115         MuxBytes[M] = 0xFF;
1116       if (Picked[M]) {
1117         CanMux = false;
1118         break;
1119       }
1120       NewMask[I] = M;
1121     }
1122     if (CanMux)
1123       return vmuxs(MuxBytes, Va, Vb, Results);
1124   }
1125 
1126   return OpRef::fail();
1127 }
1128 
packp(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results,MutableArrayRef<int> NewMask)1129 OpRef HvxSelector::packp(ShuffleMask SM, OpRef Va, OpRef Vb,
1130                          ResultStack &Results, MutableArrayRef<int> NewMask) {
1131   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1132   unsigned HalfMask = 0;
1133   unsigned LogHw = Log2_32(HwLen);
1134   for (int M : SM.Mask) {
1135     if (M == -1)
1136       continue;
1137     HalfMask |= (1u << (M >> LogHw));
1138   }
1139 
1140   if (HalfMask == 0)
1141     return OpRef::undef(getPairVT(MVT::i8));
1142 
1143   // If more than two halves are used, bail.
1144   // TODO: be more aggressive here?
1145   if (countPopulation(HalfMask) > 2)
1146     return OpRef::fail();
1147 
1148   MVT HalfTy = getSingleVT(MVT::i8);
1149 
1150   OpRef Inp[2] = { Va, Vb };
1151   OpRef Out[2] = { OpRef::undef(HalfTy), OpRef::undef(HalfTy) };
1152 
1153   uint8_t HalfIdx[4] = { 0xFF, 0xFF, 0xFF, 0xFF };
1154   unsigned Idx = 0;
1155   for (unsigned I = 0; I != 4; ++I) {
1156     if ((HalfMask & (1u << I)) == 0)
1157       continue;
1158     assert(Idx < 2);
1159     OpRef Op = Inp[I/2];
1160     Out[Idx] = (I & 1) ? OpRef::hi(Op) : OpRef::lo(Op);
1161     HalfIdx[I] = Idx++;
1162   }
1163 
1164   int VecLen = SM.Mask.size();
1165   for (int I = 0; I != VecLen; ++I) {
1166     int M = SM.Mask[I];
1167     if (M >= 0) {
1168       uint8_t Idx = HalfIdx[M >> LogHw];
1169       assert(Idx == 0 || Idx == 1);
1170       M = (M & (HwLen-1)) + HwLen*Idx;
1171     }
1172     NewMask[I] = M;
1173   }
1174 
1175   return concat(Out[0], Out[1], Results);
1176 }
1177 
vmuxs(ArrayRef<uint8_t> Bytes,OpRef Va,OpRef Vb,ResultStack & Results)1178 OpRef HvxSelector::vmuxs(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1179                          ResultStack &Results) {
1180   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1181   MVT ByteTy = getSingleVT(MVT::i8);
1182   MVT BoolTy = MVT::getVectorVT(MVT::i1, HwLen);
1183   const SDLoc &dl(Results.InpNode);
1184   SDValue B = getVectorConstant(Bytes, dl);
1185   Results.push(Hexagon::V6_vd0, ByteTy, {});
1186   Results.push(Hexagon::V6_veqb, BoolTy, {OpRef(B), OpRef::res(-1)});
1187   Results.push(Hexagon::V6_vmux, ByteTy, {OpRef::res(-1), Vb, Va});
1188   return OpRef::res(Results.top());
1189 }
1190 
vmuxp(ArrayRef<uint8_t> Bytes,OpRef Va,OpRef Vb,ResultStack & Results)1191 OpRef HvxSelector::vmuxp(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1192                          ResultStack &Results) {
1193   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1194   size_t S = Bytes.size() / 2;
1195   OpRef L = vmuxs(Bytes.take_front(S), OpRef::lo(Va), OpRef::lo(Vb), Results);
1196   OpRef H = vmuxs(Bytes.drop_front(S), OpRef::hi(Va), OpRef::hi(Vb), Results);
1197   return concat(L, H, Results);
1198 }
1199 
shuffs1(ShuffleMask SM,OpRef Va,ResultStack & Results)1200 OpRef HvxSelector::shuffs1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1201   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1202   unsigned VecLen = SM.Mask.size();
1203   assert(HwLen == VecLen);
1204   (void)VecLen;
1205   assert(all_of(SM.Mask, [this](int M) { return M == -1 || M < int(HwLen); }));
1206 
1207   if (isIdentity(SM.Mask))
1208     return Va;
1209   if (isUndef(SM.Mask))
1210     return OpRef::undef(getSingleVT(MVT::i8));
1211 
1212   OpRef P = perfect(SM, Va, Results);
1213   if (P.isValid())
1214     return P;
1215   return butterfly(SM, Va, Results);
1216 }
1217 
shuffs2(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results)1218 OpRef HvxSelector::shuffs2(ShuffleMask SM, OpRef Va, OpRef Vb,
1219                            ResultStack &Results) {
1220   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1221   if (isUndef(SM.Mask))
1222     return OpRef::undef(getSingleVT(MVT::i8));
1223 
1224   OpRef C = contracting(SM, Va, Vb, Results);
1225   if (C.isValid())
1226     return C;
1227 
1228   int VecLen = SM.Mask.size();
1229   SmallVector<int,128> NewMask(VecLen);
1230   OpRef P = packs(SM, Va, Vb, Results, NewMask);
1231   if (P.isValid())
1232     return shuffs1(ShuffleMask(NewMask), P, Results);
1233 
1234   SmallVector<int,128> MaskL(VecLen), MaskR(VecLen);
1235   splitMask(SM.Mask, MaskL, MaskR);
1236 
1237   OpRef L = shuffs1(ShuffleMask(MaskL), Va, Results);
1238   OpRef R = shuffs1(ShuffleMask(MaskR), Vb, Results);
1239   if (!L.isValid() || !R.isValid())
1240     return OpRef::fail();
1241 
1242   SmallVector<uint8_t,128> Bytes(VecLen);
1243   for (int I = 0; I != VecLen; ++I) {
1244     if (MaskL[I] != -1)
1245       Bytes[I] = 0xFF;
1246   }
1247   return vmuxs(Bytes, L, R, Results);
1248 }
1249 
shuffp1(ShuffleMask SM,OpRef Va,ResultStack & Results)1250 OpRef HvxSelector::shuffp1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1251   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1252   int VecLen = SM.Mask.size();
1253 
1254   if (isIdentity(SM.Mask))
1255     return Va;
1256   if (isUndef(SM.Mask))
1257     return OpRef::undef(getPairVT(MVT::i8));
1258 
1259   SmallVector<int,128> PackedMask(VecLen);
1260   OpRef P = packs(SM, OpRef::lo(Va), OpRef::hi(Va), Results, PackedMask);
1261   if (P.isValid()) {
1262     ShuffleMask PM(PackedMask);
1263     OpRef E = expanding(PM, P, Results);
1264     if (E.isValid())
1265       return E;
1266 
1267     OpRef L = shuffs1(PM.lo(), P, Results);
1268     OpRef H = shuffs1(PM.hi(), P, Results);
1269     if (L.isValid() && H.isValid())
1270       return concat(L, H, Results);
1271   }
1272 
1273   OpRef R = perfect(SM, Va, Results);
1274   if (R.isValid())
1275     return R;
1276   // TODO commute the mask and try the opposite order of the halves.
1277 
1278   OpRef L = shuffs2(SM.lo(), OpRef::lo(Va), OpRef::hi(Va), Results);
1279   OpRef H = shuffs2(SM.hi(), OpRef::lo(Va), OpRef::hi(Va), Results);
1280   if (L.isValid() && H.isValid())
1281     return concat(L, H, Results);
1282 
1283   return OpRef::fail();
1284 }
1285 
shuffp2(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results)1286 OpRef HvxSelector::shuffp2(ShuffleMask SM, OpRef Va, OpRef Vb,
1287                            ResultStack &Results) {
1288   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1289   if (isUndef(SM.Mask))
1290     return OpRef::undef(getPairVT(MVT::i8));
1291 
1292   int VecLen = SM.Mask.size();
1293   SmallVector<int,256> PackedMask(VecLen);
1294   OpRef P = packp(SM, Va, Vb, Results, PackedMask);
1295   if (P.isValid())
1296     return shuffp1(ShuffleMask(PackedMask), P, Results);
1297 
1298   SmallVector<int,256> MaskL(VecLen), MaskR(VecLen);
1299   splitMask(SM.Mask, MaskL, MaskR);
1300 
1301   OpRef L = shuffp1(ShuffleMask(MaskL), Va, Results);
1302   OpRef R = shuffp1(ShuffleMask(MaskR), Vb, Results);
1303   if (!L.isValid() || !R.isValid())
1304     return OpRef::fail();
1305 
1306   // Mux the results.
1307   SmallVector<uint8_t,256> Bytes(VecLen);
1308   for (int I = 0; I != VecLen; ++I) {
1309     if (MaskL[I] != -1)
1310       Bytes[I] = 0xFF;
1311   }
1312   return vmuxp(Bytes, L, R, Results);
1313 }
1314 
1315 namespace {
1316   struct Deleter : public SelectionDAG::DAGNodeDeletedListener {
1317     template <typename T>
Deleter__anonb73ef63f0d11::Deleter1318     Deleter(SelectionDAG &D, T &C)
1319       : SelectionDAG::DAGNodeDeletedListener(D, [&C] (SDNode *N, SDNode *E) {
1320                                                   C.erase(N);
1321                                                 }) {}
1322   };
1323 
1324   template <typename T>
1325   struct NullifyingVector : public T {
1326     DenseMap<SDNode*, SDNode**> Refs;
NullifyingVector__anonb73ef63f0d11::NullifyingVector1327     NullifyingVector(T &&V) : T(V) {
1328       for (unsigned i = 0, e = T::size(); i != e; ++i) {
1329         SDNode *&N = T::operator[](i);
1330         Refs[N] = &N;
1331       }
1332     }
erase__anonb73ef63f0d11::NullifyingVector1333     void erase(SDNode *N) {
1334       auto F = Refs.find(N);
1335       if (F != Refs.end())
1336         *F->second = nullptr;
1337     }
1338   };
1339 }
1340 
select(SDNode * ISelN)1341 void HvxSelector::select(SDNode *ISelN) {
1342   // What's important here is to select the right set of nodes. The main
1343   // selection algorithm loops over nodes in a topological order, i.e. users
1344   // are visited before their operands.
1345   //
1346   // It is an error to have an unselected node with a selected operand, and
1347   // there is an assertion in the main selector code to enforce that.
1348   //
1349   // Such a situation could occur if we selected a node, which is both a
1350   // subnode of ISelN, and a subnode of an unrelated (and yet unselected)
1351   // node in the DAG.
1352   assert(ISelN->getOpcode() == HexagonISD::ISEL);
1353   SDNode *N0 = ISelN->getOperand(0).getNode();
1354   if (N0->isMachineOpcode()) {
1355     ISel.ReplaceNode(ISelN, N0);
1356     return;
1357   }
1358 
1359   // There could have been nodes created (i.e. inserted into the DAG)
1360   // that are now dead. Remove them, in case they use any of the nodes
1361   // to select (and make them look shared).
1362   DAG.RemoveDeadNodes();
1363 
1364   SetVector<SDNode*> SubNodes, TmpQ;
1365   std::map<SDNode*,unsigned> NumOps;
1366 
1367   // Don't want to select N0 if it's shared with another node, except if
1368   // it's shared with other ISELs.
1369   auto IsISelN = [](SDNode *T) { return T->getOpcode() == HexagonISD::ISEL; };
1370   if (llvm::all_of(N0->uses(), IsISelN))
1371     SubNodes.insert(N0);
1372 
1373   auto InSubNodes = [&SubNodes](SDNode *T) { return SubNodes.count(T); };
1374   for (unsigned I = 0; I != SubNodes.size(); ++I) {
1375     SDNode *S = SubNodes[I];
1376     unsigned OpN = 0;
1377     // Only add subnodes that are only reachable from N0.
1378     for (SDValue Op : S->ops()) {
1379       SDNode *O = Op.getNode();
1380       if (llvm::all_of(O->uses(), InSubNodes)) {
1381         SubNodes.insert(O);
1382         ++OpN;
1383       }
1384     }
1385     NumOps.insert({S, OpN});
1386     if (OpN == 0)
1387       TmpQ.insert(S);
1388   }
1389 
1390   for (unsigned I = 0; I != TmpQ.size(); ++I) {
1391     SDNode *S = TmpQ[I];
1392     for (SDNode *U : S->uses()) {
1393       if (U == ISelN)
1394         continue;
1395       auto F = NumOps.find(U);
1396       assert(F != NumOps.end());
1397       if (F->second > 0 && !--F->second)
1398         TmpQ.insert(F->first);
1399     }
1400   }
1401 
1402   // Remove the marker.
1403   ISel.ReplaceNode(ISelN, N0);
1404 
1405   assert(SubNodes.size() == TmpQ.size());
1406   NullifyingVector<decltype(TmpQ)::vector_type> Queue(TmpQ.takeVector());
1407 
1408   Deleter DUQ(DAG, Queue);
1409   for (SDNode *S : reverse(Queue)) {
1410     if (S == nullptr)
1411       continue;
1412     DEBUG_WITH_TYPE("isel", {dbgs() << "HVX selecting: "; S->dump(&DAG);});
1413     ISel.Select(S);
1414   }
1415 }
1416 
scalarizeShuffle(ArrayRef<int> Mask,const SDLoc & dl,MVT ResTy,SDValue Va,SDValue Vb,SDNode * N)1417 bool HvxSelector::scalarizeShuffle(ArrayRef<int> Mask, const SDLoc &dl,
1418                                    MVT ResTy, SDValue Va, SDValue Vb,
1419                                    SDNode *N) {
1420   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1421   MVT ElemTy = ResTy.getVectorElementType();
1422   assert(ElemTy == MVT::i8);
1423   unsigned VecLen = Mask.size();
1424   bool HavePairs = (2*HwLen == VecLen);
1425   MVT SingleTy = getSingleVT(MVT::i8);
1426 
1427   // The prior attempts to handle this shuffle may have left a bunch of
1428   // dead nodes in the DAG (such as constants). These nodes will be added
1429   // at the end of DAG's node list, which at that point had already been
1430   // sorted topologically. In the main selection loop, the node list is
1431   // traversed backwards from the root node, which means that any new
1432   // nodes (from the end of the list) will not be visited.
1433   // Scalarization will replace the shuffle node with the scalarized
1434   // expression, and if that expression reused any if the leftoever (dead)
1435   // nodes, these nodes would not be selected (since the "local" selection
1436   // only visits nodes that are not in AllNodes).
1437   // To avoid this issue, remove all dead nodes from the DAG now.
1438 //  DAG.RemoveDeadNodes();
1439 
1440   SmallVector<SDValue,128> Ops;
1441   LLVMContext &Ctx = *DAG.getContext();
1442   MVT LegalTy = Lower.getTypeToTransformTo(Ctx, ElemTy).getSimpleVT();
1443   for (int I : Mask) {
1444     if (I < 0) {
1445       Ops.push_back(ISel.selectUndef(dl, LegalTy));
1446       continue;
1447     }
1448     SDValue Vec;
1449     unsigned M = I;
1450     if (M < VecLen) {
1451       Vec = Va;
1452     } else {
1453       Vec = Vb;
1454       M -= VecLen;
1455     }
1456     if (HavePairs) {
1457       if (M < HwLen) {
1458         Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, SingleTy, Vec);
1459       } else {
1460         Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_hi, dl, SingleTy, Vec);
1461         M -= HwLen;
1462       }
1463     }
1464     SDValue Idx = DAG.getConstant(M, dl, MVT::i32);
1465     SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, LegalTy, {Vec, Idx});
1466     SDValue L = Lower.LowerOperation(Ex, DAG);
1467     assert(L.getNode());
1468     Ops.push_back(L);
1469   }
1470 
1471   SDValue LV;
1472   if (2*HwLen == VecLen) {
1473     SDValue B0 = DAG.getBuildVector(SingleTy, dl, {Ops.data(), HwLen});
1474     SDValue L0 = Lower.LowerOperation(B0, DAG);
1475     SDValue B1 = DAG.getBuildVector(SingleTy, dl, {Ops.data()+HwLen, HwLen});
1476     SDValue L1 = Lower.LowerOperation(B1, DAG);
1477     // XXX CONCAT_VECTORS is legal for HVX vectors. Legalizing (lowering)
1478     // functions may expect to be called only for illegal operations, so
1479     // make sure that they are not called for legal ones. Develop a better
1480     // mechanism for dealing with this.
1481     LV = DAG.getNode(ISD::CONCAT_VECTORS, dl, ResTy, {L0, L1});
1482   } else {
1483     SDValue BV = DAG.getBuildVector(ResTy, dl, Ops);
1484     LV = Lower.LowerOperation(BV, DAG);
1485   }
1486 
1487   assert(!N->use_empty());
1488   SDValue IS = DAG.getNode(HexagonISD::ISEL, dl, ResTy, LV);
1489   ISel.ReplaceNode(N, IS.getNode());
1490   select(IS.getNode());
1491   DAG.RemoveDeadNodes();
1492   return true;
1493 }
1494 
contracting(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results)1495 OpRef HvxSelector::contracting(ShuffleMask SM, OpRef Va, OpRef Vb,
1496                                ResultStack &Results) {
1497   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1498   if (!Va.isValid() || !Vb.isValid())
1499     return OpRef::fail();
1500 
1501   // Contracting shuffles, i.e. instructions that always discard some bytes
1502   // from the operand vectors.
1503   //
1504   // V6_vshuff{e,o}b
1505   // V6_vdealb4w
1506   // V6_vpack{e,o}{b,h}
1507 
1508   int VecLen = SM.Mask.size();
1509   std::pair<int,unsigned> Strip = findStrip(SM.Mask, 1, VecLen);
1510   MVT ResTy = getSingleVT(MVT::i8);
1511 
1512   // The following shuffles only work for bytes and halfwords. This requires
1513   // the strip length to be 1 or 2.
1514   if (Strip.second != 1 && Strip.second != 2)
1515     return OpRef::fail();
1516 
1517   // The patterns for the shuffles, in terms of the starting offsets of the
1518   // consecutive strips (L = length of the strip, N = VecLen):
1519   //
1520   // vpacke:    0, 2L, 4L ... N+0, N+2L, N+4L ...      L = 1 or 2
1521   // vpacko:    L, 3L, 5L ... N+L, N+3L, N+5L ...      L = 1 or 2
1522   //
1523   // vshuffe:   0, N+0, 2L, N+2L, 4L ...               L = 1 or 2
1524   // vshuffo:   L, N+L, 3L, N+3L, 5L ...               L = 1 or 2
1525   //
1526   // vdealb4w:  0, 4, 8 ... 2, 6, 10 ... N+0, N+4, N+8 ... N+2, N+6, N+10 ...
1527 
1528   // The value of the element in the mask following the strip will decide
1529   // what kind of a shuffle this can be.
1530   int NextInMask = SM.Mask[Strip.second];
1531 
1532   // Check if NextInMask could be 2L, 3L or 4, i.e. if it could be a mask
1533   // for vpack or vdealb4w. VecLen > 4, so NextInMask for vdealb4w would
1534   // satisfy this.
1535   if (NextInMask < VecLen) {
1536     // vpack{e,o} or vdealb4w
1537     if (Strip.first == 0 && Strip.second == 1 && NextInMask == 4) {
1538       int N = VecLen;
1539       // Check if this is vdealb4w (L=1).
1540       for (int I = 0; I != N/4; ++I)
1541         if (SM.Mask[I] != 4*I)
1542           return OpRef::fail();
1543       for (int I = 0; I != N/4; ++I)
1544         if (SM.Mask[I+N/4] != 2 + 4*I)
1545           return OpRef::fail();
1546       for (int I = 0; I != N/4; ++I)
1547         if (SM.Mask[I+N/2] != N + 4*I)
1548           return OpRef::fail();
1549       for (int I = 0; I != N/4; ++I)
1550         if (SM.Mask[I+3*N/4] != N+2 + 4*I)
1551           return OpRef::fail();
1552       // Matched mask for vdealb4w.
1553       Results.push(Hexagon::V6_vdealb4w, ResTy, {Vb, Va});
1554       return OpRef::res(Results.top());
1555     }
1556 
1557     // Check if this is vpack{e,o}.
1558     int N = VecLen;
1559     int L = Strip.second;
1560     // Check if the first strip starts at 0 or at L.
1561     if (Strip.first != 0 && Strip.first != L)
1562       return OpRef::fail();
1563     // Examine the rest of the mask.
1564     for (int I = L; I < N; I += L) {
1565       auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
1566       // Check whether the mask element at the beginning of each strip
1567       // increases by 2L each time.
1568       if (S.first - Strip.first != 2*I)
1569         return OpRef::fail();
1570       // Check whether each strip is of the same length.
1571       if (S.second != unsigned(L))
1572         return OpRef::fail();
1573     }
1574 
1575     // Strip.first == 0  =>  vpacke
1576     // Strip.first == L  =>  vpacko
1577     assert(Strip.first == 0 || Strip.first == L);
1578     using namespace Hexagon;
1579     NodeTemplate Res;
1580     Res.Opc = Strip.second == 1 // Number of bytes.
1581                   ? (Strip.first == 0 ? V6_vpackeb : V6_vpackob)
1582                   : (Strip.first == 0 ? V6_vpackeh : V6_vpackoh);
1583     Res.Ty = ResTy;
1584     Res.Ops = { Vb, Va };
1585     Results.push(Res);
1586     return OpRef::res(Results.top());
1587   }
1588 
1589   // Check if this is vshuff{e,o}.
1590   int N = VecLen;
1591   int L = Strip.second;
1592   std::pair<int,unsigned> PrevS = Strip;
1593   bool Flip = false;
1594   for (int I = L; I < N; I += L) {
1595     auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
1596     if (S.second != PrevS.second)
1597       return OpRef::fail();
1598     int Diff = Flip ? PrevS.first - S.first + 2*L
1599                     : S.first - PrevS.first;
1600     if (Diff != N)
1601       return OpRef::fail();
1602     Flip ^= true;
1603     PrevS = S;
1604   }
1605   // Strip.first == 0  =>  vshuffe
1606   // Strip.first == L  =>  vshuffo
1607   assert(Strip.first == 0 || Strip.first == L);
1608   using namespace Hexagon;
1609   NodeTemplate Res;
1610   Res.Opc = Strip.second == 1 // Number of bytes.
1611                 ? (Strip.first == 0 ? V6_vshuffeb : V6_vshuffob)
1612                 : (Strip.first == 0 ?  V6_vshufeh :  V6_vshufoh);
1613   Res.Ty = ResTy;
1614   Res.Ops = { Vb, Va };
1615   Results.push(Res);
1616   return OpRef::res(Results.top());
1617 }
1618 
expanding(ShuffleMask SM,OpRef Va,ResultStack & Results)1619 OpRef HvxSelector::expanding(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1620   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1621   // Expanding shuffles (using all elements and inserting into larger vector):
1622   //
1623   // V6_vunpacku{b,h} [*]
1624   //
1625   // [*] Only if the upper elements (filled with 0s) are "don't care" in Mask.
1626   //
1627   // Note: V6_vunpacko{b,h} are or-ing the high byte/half in the result, so
1628   // they are not shuffles.
1629   //
1630   // The argument is a single vector.
1631 
1632   int VecLen = SM.Mask.size();
1633   assert(2*HwLen == unsigned(VecLen) && "Expecting vector-pair type");
1634 
1635   std::pair<int,unsigned> Strip = findStrip(SM.Mask, 1, VecLen);
1636 
1637   // The patterns for the unpacks, in terms of the starting offsets of the
1638   // consecutive strips (L = length of the strip, N = VecLen):
1639   //
1640   // vunpacku:  0, -1, L, -1, 2L, -1 ...
1641 
1642   if (Strip.first != 0)
1643     return OpRef::fail();
1644 
1645   // The vunpackus only handle byte and half-word.
1646   if (Strip.second != 1 && Strip.second != 2)
1647     return OpRef::fail();
1648 
1649   int N = VecLen;
1650   int L = Strip.second;
1651 
1652   // First, check the non-ignored strips.
1653   for (int I = 2*L; I < 2*N; I += 2*L) {
1654     auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
1655     if (S.second != unsigned(L))
1656       return OpRef::fail();
1657     if (2*S.first != I)
1658       return OpRef::fail();
1659   }
1660   // Check the -1s.
1661   for (int I = L; I < 2*N; I += 2*L) {
1662     auto S = findStrip(SM.Mask.drop_front(I), 0, N-I);
1663     if (S.first != -1 || S.second != unsigned(L))
1664       return OpRef::fail();
1665   }
1666 
1667   unsigned Opc = Strip.second == 1 ? Hexagon::V6_vunpackub
1668                                    : Hexagon::V6_vunpackuh;
1669   Results.push(Opc, getPairVT(MVT::i8), {Va});
1670   return OpRef::res(Results.top());
1671 }
1672 
perfect(ShuffleMask SM,OpRef Va,ResultStack & Results)1673 OpRef HvxSelector::perfect(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1674   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1675   // V6_vdeal{b,h}
1676   // V6_vshuff{b,h}
1677 
1678   // V6_vshufoe{b,h}  those are quivalent to vshuffvdd(..,{1,2})
1679   // V6_vshuffvdd (V6_vshuff)
1680   // V6_dealvdd (V6_vdeal)
1681 
1682   int VecLen = SM.Mask.size();
1683   assert(isPowerOf2_32(VecLen) && Log2_32(VecLen) <= 8);
1684   unsigned LogLen = Log2_32(VecLen);
1685   unsigned HwLog = Log2_32(HwLen);
1686   // The result length must be the same as the length of a single vector,
1687   // or a vector pair.
1688   assert(LogLen == HwLog || LogLen == HwLog+1);
1689   bool HavePairs = LogLen == HwLog+1;
1690 
1691   if (!isPermutation(SM.Mask))
1692     return OpRef::fail();
1693 
1694   SmallVector<unsigned,8> Perm(LogLen);
1695 
1696   // Check if this could be a perfect shuffle, or a combination of perfect
1697   // shuffles.
1698   //
1699   // Consider this permutation (using hex digits to make the ASCII diagrams
1700   // easier to read):
1701   //   { 0, 8, 1, 9, 2, A, 3, B, 4, C, 5, D, 6, E, 7, F }.
1702   // This is a "deal" operation: divide the input into two halves, and
1703   // create the output by picking elements by alternating between these two
1704   // halves:
1705   //   0 1 2 3 4 5 6 7    -->    0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F  [*]
1706   //   8 9 A B C D E F
1707   //
1708   // Aside from a few special explicit cases (V6_vdealb, etc.), HVX provides
1709   // a somwehat different mechanism that could be used to perform shuffle/
1710   // deal operations: a 2x2 transpose.
1711   // Consider the halves of inputs again, they can be interpreted as a 2x8
1712   // matrix. A 2x8 matrix can be looked at four 2x2 matrices concatenated
1713   // together. Now, when considering 2 elements at a time, it will be a 2x4
1714   // matrix (with elements 01, 23, 45, etc.), or two 2x2 matrices:
1715   //   01 23  45 67
1716   //   89 AB  CD EF
1717   // With groups of 4, this will become a single 2x2 matrix, and so on.
1718   //
1719   // The 2x2 transpose instruction works by transposing each of the 2x2
1720   // matrices (or "sub-matrices"), given a specific group size. For example,
1721   // if the group size is 1 (i.e. each element is its own group), there
1722   // will be four transposes of the four 2x2 matrices that form the 2x8.
1723   // For example, with the inputs as above, the result will be:
1724   //   0 8  2 A  4 C  6 E
1725   //   1 9  3 B  5 D  7 F
1726   // Now, this result can be tranposed again, but with the group size of 2:
1727   //   08 19  4C 5D
1728   //   2A 3B  6E 7F
1729   // If we then transpose that result, but with the group size of 4, we get:
1730   //   0819 2A3B
1731   //   4C5D 6E7F
1732   // If we concatenate these two rows, it will be
1733   //   0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F
1734   // which is the same as the "deal" [*] above.
1735   //
1736   // In general, a "deal" of individual elements is a series of 2x2 transposes,
1737   // with changing group size. HVX has two instructions:
1738   //   Vdd = V6_vdealvdd Vu, Vv, Rt
1739   //   Vdd = V6_shufvdd  Vu, Vv, Rt
1740   // that perform exactly that. The register Rt controls which transposes are
1741   // going to happen: a bit at position n (counting from 0) indicates that a
1742   // transpose with a group size of 2^n will take place. If multiple bits are
1743   // set, multiple transposes will happen: vdealvdd will perform them starting
1744   // with the largest group size, vshuffvdd will do them in the reverse order.
1745   //
1746   // The main observation is that each 2x2 transpose corresponds to swapping
1747   // columns of bits in the binary representation of the values.
1748   //
1749   // The numbers {3,2,1,0} and the log2 of the number of contiguous 1 bits
1750   // in a given column. The * denote the columns that will be swapped.
1751   // The transpose with the group size 2^n corresponds to swapping columns
1752   // 3 (the highest log) and log2(n):
1753   //
1754   //     3 2 1 0         0 2 1 3         0 2 3 1
1755   //     *     *             * *           * *
1756   //  0  0 0 0 0      0  0 0 0 0      0  0 0 0 0      0  0 0 0 0
1757   //  1  0 0 0 1      8  1 0 0 0      8  1 0 0 0      8  1 0 0 0
1758   //  2  0 0 1 0      2  0 0 1 0      1  0 0 0 1      1  0 0 0 1
1759   //  3  0 0 1 1      A  1 0 1 0      9  1 0 0 1      9  1 0 0 1
1760   //  4  0 1 0 0      4  0 1 0 0      4  0 1 0 0      2  0 0 1 0
1761   //  5  0 1 0 1      C  1 1 0 0      C  1 1 0 0      A  1 0 1 0
1762   //  6  0 1 1 0      6  0 1 1 0      5  0 1 0 1      3  0 0 1 1
1763   //  7  0 1 1 1      E  1 1 1 0      D  1 1 0 1      B  1 0 1 1
1764   //  8  1 0 0 0      1  0 0 0 1      2  0 0 1 0      4  0 1 0 0
1765   //  9  1 0 0 1      9  1 0 0 1      A  1 0 1 0      C  1 1 0 0
1766   //  A  1 0 1 0      3  0 0 1 1      3  0 0 1 1      5  0 1 0 1
1767   //  B  1 0 1 1      B  1 0 1 1      B  1 0 1 1      D  1 1 0 1
1768   //  C  1 1 0 0      5  0 1 0 1      6  0 1 1 0      6  0 1 1 0
1769   //  D  1 1 0 1      D  1 1 0 1      E  1 1 1 0      E  1 1 1 0
1770   //  E  1 1 1 0      7  0 1 1 1      7  0 1 1 1      7  0 1 1 1
1771   //  F  1 1 1 1      F  1 1 1 1      F  1 1 1 1      F  1 1 1 1
1772 
1773   // There is one special case that is not a perfect shuffle, but
1774   // can be turned into one easily: when the shuffle operates on
1775   // a vector pair, but the two vectors in the pair are swapped.
1776   // The code below that identifies perfect shuffles will reject
1777   // it, unless the order is reversed.
1778   SmallVector<int,128> MaskStorage(SM.Mask.begin(), SM.Mask.end());
1779   bool InvertedPair = false;
1780   if (HavePairs && SM.Mask[0] >= int(HwLen)) {
1781     for (int i = 0, e = SM.Mask.size(); i != e; ++i) {
1782       int M = SM.Mask[i];
1783       MaskStorage[i] = M >= int(HwLen) ? M-HwLen : M+HwLen;
1784     }
1785     InvertedPair = true;
1786   }
1787   ArrayRef<int> LocalMask(MaskStorage);
1788 
1789   auto XorPow2 = [] (ArrayRef<int> Mask, unsigned Num) {
1790     unsigned X = Mask[0] ^ Mask[Num/2];
1791     // Check that the first half has the X's bits clear.
1792     if ((Mask[0] & X) != 0)
1793       return 0u;
1794     for (unsigned I = 1; I != Num/2; ++I) {
1795       if (unsigned(Mask[I] ^ Mask[I+Num/2]) != X)
1796         return 0u;
1797       if ((Mask[I] & X) != 0)
1798         return 0u;
1799     }
1800     return X;
1801   };
1802 
1803   // Create a vector of log2's for each column: Perm[i] corresponds to
1804   // the i-th bit (lsb is 0).
1805   assert(VecLen > 2);
1806   for (unsigned I = VecLen; I >= 2; I >>= 1) {
1807     // Examine the initial segment of Mask of size I.
1808     unsigned X = XorPow2(LocalMask, I);
1809     if (!isPowerOf2_32(X))
1810       return OpRef::fail();
1811     // Check the other segments of Mask.
1812     for (int J = I; J < VecLen; J += I) {
1813       if (XorPow2(LocalMask.slice(J, I), I) != X)
1814         return OpRef::fail();
1815     }
1816     Perm[Log2_32(X)] = Log2_32(I)-1;
1817   }
1818 
1819   // Once we have Perm, represent it as cycles. Denote the maximum log2
1820   // (equal to log2(VecLen)-1) as M. The cycle containing M can then be
1821   // written as (M a1 a2 a3 ... an). That cycle can be broken up into
1822   // simple swaps as (M a1)(M a2)(M a3)...(M an), with the composition
1823   // order being from left to right. Any (contiguous) segment where the
1824   // values ai, ai+1...aj are either all increasing or all decreasing,
1825   // can be implemented via a single vshuffvdd/vdealvdd respectively.
1826   //
1827   // If there is a cycle (a1 a2 ... an) that does not involve M, it can
1828   // be written as (M an)(a1 a2 ... an)(M a1). The first two cycles can
1829   // then be folded to get (M a1 a2 ... an)(M a1), and the above procedure
1830   // can be used to generate a sequence of vshuffvdd/vdealvdd.
1831   //
1832   // Example:
1833   // Assume M = 4 and consider a permutation (0 1)(2 3). It can be written
1834   // as (4 0 1)(4 0) composed with (4 2 3)(4 2), or simply
1835   //   (4 0 1)(4 0)(4 2 3)(4 2).
1836   // It can then be expanded into swaps as
1837   //   (4 0)(4 1)(4 0)(4 2)(4 3)(4 2),
1838   // and broken up into "increasing" segments as
1839   //   [(4 0)(4 1)] [(4 0)(4 2)(4 3)] [(4 2)].
1840   // This is equivalent to
1841   //   (4 0 1)(4 0 2 3)(4 2),
1842   // which can be implemented as 3 vshufvdd instructions.
1843 
1844   using CycleType = SmallVector<unsigned,8>;
1845   std::set<CycleType> Cycles;
1846   std::set<unsigned> All;
1847 
1848   for (unsigned I : Perm)
1849     All.insert(I);
1850 
1851   // If the cycle contains LogLen-1, move it to the front of the cycle.
1852   // Otherwise, return the cycle unchanged.
1853   auto canonicalize = [LogLen](const CycleType &C) -> CycleType {
1854     unsigned LogPos, N = C.size();
1855     for (LogPos = 0; LogPos != N; ++LogPos)
1856       if (C[LogPos] == LogLen-1)
1857         break;
1858     if (LogPos == N)
1859       return C;
1860 
1861     CycleType NewC(C.begin()+LogPos, C.end());
1862     NewC.append(C.begin(), C.begin()+LogPos);
1863     return NewC;
1864   };
1865 
1866   auto pfs = [](const std::set<CycleType> &Cs, unsigned Len) {
1867     // Ordering: shuff: 5 0 1 2 3 4, deal: 5 4 3 2 1 0 (for Log=6),
1868     // for bytes zero is included, for halfwords is not.
1869     if (Cs.size() != 1)
1870       return 0u;
1871     const CycleType &C = *Cs.begin();
1872     if (C[0] != Len-1)
1873       return 0u;
1874     int D = Len - C.size();
1875     if (D != 0 && D != 1)
1876       return 0u;
1877 
1878     bool IsDeal = true, IsShuff = true;
1879     for (unsigned I = 1; I != Len-D; ++I) {
1880       if (C[I] != Len-1-I)
1881         IsDeal = false;
1882       if (C[I] != I-(1-D))  // I-1, I
1883         IsShuff = false;
1884     }
1885     // At most one, IsDeal or IsShuff, can be non-zero.
1886     assert(!(IsDeal || IsShuff) || IsDeal != IsShuff);
1887     static unsigned Deals[] = { Hexagon::V6_vdealb, Hexagon::V6_vdealh };
1888     static unsigned Shufs[] = { Hexagon::V6_vshuffb, Hexagon::V6_vshuffh };
1889     return IsDeal ? Deals[D] : (IsShuff ? Shufs[D] : 0);
1890   };
1891 
1892   while (!All.empty()) {
1893     unsigned A = *All.begin();
1894     All.erase(A);
1895     CycleType C;
1896     C.push_back(A);
1897     for (unsigned B = Perm[A]; B != A; B = Perm[B]) {
1898       C.push_back(B);
1899       All.erase(B);
1900     }
1901     if (C.size() <= 1)
1902       continue;
1903     Cycles.insert(canonicalize(C));
1904   }
1905 
1906   MVT SingleTy = getSingleVT(MVT::i8);
1907   MVT PairTy = getPairVT(MVT::i8);
1908 
1909   // Recognize patterns for V6_vdeal{b,h} and V6_vshuff{b,h}.
1910   if (unsigned(VecLen) == HwLen) {
1911     if (unsigned SingleOpc = pfs(Cycles, LogLen)) {
1912       Results.push(SingleOpc, SingleTy, {Va});
1913       return OpRef::res(Results.top());
1914     }
1915   }
1916 
1917   // From the cycles, construct the sequence of values that will
1918   // then form the control values for vdealvdd/vshuffvdd, i.e.
1919   // (M a1 a2)(M a3 a4 a5)... -> a1 a2 a3 a4 a5
1920   // This essentially strips the M value from the cycles where
1921   // it's present, and performs the insertion of M (then stripping)
1922   // for cycles without M (as described in an earlier comment).
1923   SmallVector<unsigned,8> SwapElems;
1924   // When the input is extended (i.e. single vector becomes a pair),
1925   // this is done by using an "undef" vector as the second input.
1926   // However, then we get
1927   //   input 1: GOODBITS
1928   //   input 2: ........
1929   // but we need
1930   //   input 1: ....BITS
1931   //   input 2: ....GOOD
1932   // Then at the end, this needs to be undone. To accomplish this,
1933   // artificially add "LogLen-1" at both ends of the sequence.
1934   if (!HavePairs)
1935     SwapElems.push_back(LogLen-1);
1936   for (const CycleType &C : Cycles) {
1937     // Do the transformation: (a1..an) -> (M a1..an)(M a1).
1938     unsigned First = (C[0] == LogLen-1) ? 1 : 0;
1939     SwapElems.append(C.begin()+First, C.end());
1940     if (First == 0)
1941       SwapElems.push_back(C[0]);
1942   }
1943   if (!HavePairs)
1944     SwapElems.push_back(LogLen-1);
1945 
1946   const SDLoc &dl(Results.InpNode);
1947   OpRef Arg = HavePairs ? Va
1948                         : concat(Va, OpRef::undef(SingleTy), Results);
1949   if (InvertedPair)
1950     Arg = concat(OpRef::hi(Arg), OpRef::lo(Arg), Results);
1951 
1952   for (unsigned I = 0, E = SwapElems.size(); I != E; ) {
1953     bool IsInc = I == E-1 || SwapElems[I] < SwapElems[I+1];
1954     unsigned S = (1u << SwapElems[I]);
1955     if (I < E-1) {
1956       while (++I < E-1 && IsInc == (SwapElems[I] < SwapElems[I+1]))
1957         S |= 1u << SwapElems[I];
1958       // The above loop will not add a bit for the final SwapElems[I+1],
1959       // so add it here.
1960       S |= 1u << SwapElems[I];
1961     }
1962     ++I;
1963 
1964     NodeTemplate Res;
1965     Results.push(Hexagon::A2_tfrsi, MVT::i32,
1966                  { DAG.getTargetConstant(S, dl, MVT::i32) });
1967     Res.Opc = IsInc ? Hexagon::V6_vshuffvdd : Hexagon::V6_vdealvdd;
1968     Res.Ty = PairTy;
1969     Res.Ops = { OpRef::hi(Arg), OpRef::lo(Arg), OpRef::res(-1) };
1970     Results.push(Res);
1971     Arg = OpRef::res(Results.top());
1972   }
1973 
1974   return HavePairs ? Arg : OpRef::lo(Arg);
1975 }
1976 
butterfly(ShuffleMask SM,OpRef Va,ResultStack & Results)1977 OpRef HvxSelector::butterfly(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1978   DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1979   // Butterfly shuffles.
1980   //
1981   // V6_vdelta
1982   // V6_vrdelta
1983   // V6_vror
1984 
1985   // The assumption here is that all elements picked by Mask are in the
1986   // first operand to the vector_shuffle. This assumption is enforced
1987   // by the caller.
1988 
1989   MVT ResTy = getSingleVT(MVT::i8);
1990   PermNetwork::Controls FC, RC;
1991   const SDLoc &dl(Results.InpNode);
1992   int VecLen = SM.Mask.size();
1993 
1994   for (int M : SM.Mask) {
1995     if (M != -1 && M >= VecLen)
1996       return OpRef::fail();
1997   }
1998 
1999   // Try the deltas/benes for both single vectors and vector pairs.
2000   ForwardDeltaNetwork FN(SM.Mask);
2001   if (FN.run(FC)) {
2002     SDValue Ctl = getVectorConstant(FC, dl);
2003     Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(Ctl)});
2004     return OpRef::res(Results.top());
2005   }
2006 
2007   // Try reverse delta.
2008   ReverseDeltaNetwork RN(SM.Mask);
2009   if (RN.run(RC)) {
2010     SDValue Ctl = getVectorConstant(RC, dl);
2011     Results.push(Hexagon::V6_vrdelta, ResTy, {Va, OpRef(Ctl)});
2012     return OpRef::res(Results.top());
2013   }
2014 
2015   // Do Benes.
2016   BenesNetwork BN(SM.Mask);
2017   if (BN.run(FC, RC)) {
2018     SDValue CtlF = getVectorConstant(FC, dl);
2019     SDValue CtlR = getVectorConstant(RC, dl);
2020     Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(CtlF)});
2021     Results.push(Hexagon::V6_vrdelta, ResTy,
2022                  {OpRef::res(-1), OpRef(CtlR)});
2023     return OpRef::res(Results.top());
2024   }
2025 
2026   return OpRef::fail();
2027 }
2028 
getVectorConstant(ArrayRef<uint8_t> Data,const SDLoc & dl)2029 SDValue HvxSelector::getVectorConstant(ArrayRef<uint8_t> Data,
2030                                        const SDLoc &dl) {
2031   SmallVector<SDValue, 128> Elems;
2032   for (uint8_t C : Data)
2033     Elems.push_back(DAG.getConstant(C, dl, MVT::i8));
2034   MVT VecTy = MVT::getVectorVT(MVT::i8, Data.size());
2035   SDValue BV = DAG.getBuildVector(VecTy, dl, Elems);
2036   SDValue LV = Lower.LowerOperation(BV, DAG);
2037   DAG.RemoveDeadNode(BV.getNode());
2038   return DAG.getNode(HexagonISD::ISEL, dl, VecTy, LV);
2039 }
2040 
selectShuffle(SDNode * N)2041 void HvxSelector::selectShuffle(SDNode *N) {
2042   DEBUG_WITH_TYPE("isel", {
2043     dbgs() << "Starting " << __func__ << " on node:\n";
2044     N->dump(&DAG);
2045   });
2046   MVT ResTy = N->getValueType(0).getSimpleVT();
2047   // Assume that vector shuffles operate on vectors of bytes.
2048   assert(ResTy.isVector() && ResTy.getVectorElementType() == MVT::i8);
2049 
2050   auto *SN = cast<ShuffleVectorSDNode>(N);
2051   std::vector<int> Mask(SN->getMask().begin(), SN->getMask().end());
2052   // This shouldn't really be necessary. Is it?
2053   for (int &Idx : Mask)
2054     if (Idx != -1 && Idx < 0)
2055       Idx = -1;
2056 
2057   unsigned VecLen = Mask.size();
2058   bool HavePairs = (2*HwLen == VecLen);
2059   assert(ResTy.getSizeInBits() / 8 == VecLen);
2060 
2061   // Vd = vector_shuffle Va, Vb, Mask
2062   //
2063 
2064   bool UseLeft = false, UseRight = false;
2065   for (unsigned I = 0; I != VecLen; ++I) {
2066     if (Mask[I] == -1)
2067       continue;
2068     unsigned Idx = Mask[I];
2069     assert(Idx < 2*VecLen);
2070     if (Idx < VecLen)
2071       UseLeft = true;
2072     else
2073       UseRight = true;
2074   }
2075 
2076   DEBUG_WITH_TYPE("isel", {
2077     dbgs() << "VecLen=" << VecLen << " HwLen=" << HwLen << " UseLeft="
2078            << UseLeft << " UseRight=" << UseRight << " HavePairs="
2079            << HavePairs << '\n';
2080   });
2081   // If the mask is all -1's, generate "undef".
2082   if (!UseLeft && !UseRight) {
2083     ISel.ReplaceNode(N, ISel.selectUndef(SDLoc(SN), ResTy).getNode());
2084     return;
2085   }
2086 
2087   SDValue Vec0 = N->getOperand(0);
2088   SDValue Vec1 = N->getOperand(1);
2089   ResultStack Results(SN);
2090   Results.push(TargetOpcode::COPY, ResTy, {Vec0});
2091   Results.push(TargetOpcode::COPY, ResTy, {Vec1});
2092   OpRef Va = OpRef::res(Results.top()-1);
2093   OpRef Vb = OpRef::res(Results.top());
2094 
2095   OpRef Res = !HavePairs ? shuffs2(ShuffleMask(Mask), Va, Vb, Results)
2096                          : shuffp2(ShuffleMask(Mask), Va, Vb, Results);
2097 
2098   bool Done = Res.isValid();
2099   if (Done) {
2100     // Make sure that Res is on the stack before materializing.
2101     Results.push(TargetOpcode::COPY, ResTy, {Res});
2102     materialize(Results);
2103   } else {
2104     Done = scalarizeShuffle(Mask, SDLoc(N), ResTy, Vec0, Vec1, N);
2105   }
2106 
2107   if (!Done) {
2108 #ifndef NDEBUG
2109     dbgs() << "Unhandled shuffle:\n";
2110     SN->dumpr(&DAG);
2111 #endif
2112     llvm_unreachable("Failed to select vector shuffle");
2113   }
2114 }
2115 
selectRor(SDNode * N)2116 void HvxSelector::selectRor(SDNode *N) {
2117   // If this is a rotation by less than 8, use V6_valignbi.
2118   MVT Ty = N->getValueType(0).getSimpleVT();
2119   const SDLoc &dl(N);
2120   SDValue VecV = N->getOperand(0);
2121   SDValue RotV = N->getOperand(1);
2122   SDNode *NewN = nullptr;
2123 
2124   if (auto *CN = dyn_cast<ConstantSDNode>(RotV.getNode())) {
2125     unsigned S = CN->getZExtValue() % HST.getVectorLength();
2126     if (S == 0) {
2127       NewN = VecV.getNode();
2128     } else if (isUInt<3>(S)) {
2129       SDValue C = DAG.getTargetConstant(S, dl, MVT::i32);
2130       NewN = DAG.getMachineNode(Hexagon::V6_valignbi, dl, Ty,
2131                                 {VecV, VecV, C});
2132     }
2133   }
2134 
2135   if (!NewN)
2136     NewN = DAG.getMachineNode(Hexagon::V6_vror, dl, Ty, {VecV, RotV});
2137 
2138   ISel.ReplaceNode(N, NewN);
2139 }
2140 
selectVAlign(SDNode * N)2141 void HvxSelector::selectVAlign(SDNode *N) {
2142   SDValue Vv = N->getOperand(0);
2143   SDValue Vu = N->getOperand(1);
2144   SDValue Rt = N->getOperand(2);
2145   SDNode *NewN = DAG.getMachineNode(Hexagon::V6_valignb, SDLoc(N),
2146                                     N->getValueType(0), {Vv, Vu, Rt});
2147   ISel.ReplaceNode(N, NewN);
2148   DAG.RemoveDeadNode(N);
2149 }
2150 
SelectHvxShuffle(SDNode * N)2151 void HexagonDAGToDAGISel::SelectHvxShuffle(SDNode *N) {
2152   HvxSelector(*this, *CurDAG).selectShuffle(N);
2153 }
2154 
SelectHvxRor(SDNode * N)2155 void HexagonDAGToDAGISel::SelectHvxRor(SDNode *N) {
2156   HvxSelector(*this, *CurDAG).selectRor(N);
2157 }
2158 
SelectHvxVAlign(SDNode * N)2159 void HexagonDAGToDAGISel::SelectHvxVAlign(SDNode *N) {
2160   HvxSelector(*this, *CurDAG).selectVAlign(N);
2161 }
2162 
SelectV65GatherPred(SDNode * N)2163 void HexagonDAGToDAGISel::SelectV65GatherPred(SDNode *N) {
2164   const SDLoc &dl(N);
2165   SDValue Chain = N->getOperand(0);
2166   SDValue Address = N->getOperand(2);
2167   SDValue Predicate = N->getOperand(3);
2168   SDValue Base = N->getOperand(4);
2169   SDValue Modifier = N->getOperand(5);
2170   SDValue Offset = N->getOperand(6);
2171 
2172   unsigned Opcode;
2173   unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2174   switch (IntNo) {
2175   default:
2176     llvm_unreachable("Unexpected HVX gather intrinsic.");
2177   case Intrinsic::hexagon_V6_vgathermhq:
2178   case Intrinsic::hexagon_V6_vgathermhq_128B:
2179     Opcode = Hexagon::V6_vgathermhq_pseudo;
2180     break;
2181   case Intrinsic::hexagon_V6_vgathermwq:
2182   case Intrinsic::hexagon_V6_vgathermwq_128B:
2183     Opcode = Hexagon::V6_vgathermwq_pseudo;
2184     break;
2185   case Intrinsic::hexagon_V6_vgathermhwq:
2186   case Intrinsic::hexagon_V6_vgathermhwq_128B:
2187     Opcode = Hexagon::V6_vgathermhwq_pseudo;
2188     break;
2189   }
2190 
2191   SDVTList VTs = CurDAG->getVTList(MVT::Other);
2192   SDValue Ops[] = { Address, Predicate, Base, Modifier, Offset, Chain };
2193   SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2194 
2195   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2196   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2197 
2198   ReplaceNode(N, Result);
2199 }
2200 
SelectV65Gather(SDNode * N)2201 void HexagonDAGToDAGISel::SelectV65Gather(SDNode *N) {
2202   const SDLoc &dl(N);
2203   SDValue Chain = N->getOperand(0);
2204   SDValue Address = N->getOperand(2);
2205   SDValue Base = N->getOperand(3);
2206   SDValue Modifier = N->getOperand(4);
2207   SDValue Offset = N->getOperand(5);
2208 
2209   unsigned Opcode;
2210   unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2211   switch (IntNo) {
2212   default:
2213     llvm_unreachable("Unexpected HVX gather intrinsic.");
2214   case Intrinsic::hexagon_V6_vgathermh:
2215   case Intrinsic::hexagon_V6_vgathermh_128B:
2216     Opcode = Hexagon::V6_vgathermh_pseudo;
2217     break;
2218   case Intrinsic::hexagon_V6_vgathermw:
2219   case Intrinsic::hexagon_V6_vgathermw_128B:
2220     Opcode = Hexagon::V6_vgathermw_pseudo;
2221     break;
2222   case Intrinsic::hexagon_V6_vgathermhw:
2223   case Intrinsic::hexagon_V6_vgathermhw_128B:
2224     Opcode = Hexagon::V6_vgathermhw_pseudo;
2225     break;
2226   }
2227 
2228   SDVTList VTs = CurDAG->getVTList(MVT::Other);
2229   SDValue Ops[] = { Address, Base, Modifier, Offset, Chain };
2230   SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2231 
2232   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2233   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2234 
2235   ReplaceNode(N, Result);
2236 }
2237 
SelectHVXDualOutput(SDNode * N)2238 void HexagonDAGToDAGISel::SelectHVXDualOutput(SDNode *N) {
2239   unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2240   SDNode *Result;
2241   switch (IID) {
2242   case Intrinsic::hexagon_V6_vaddcarry: {
2243     std::array<SDValue, 3> Ops = {
2244         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2245     SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v64i1);
2246     Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2247     break;
2248   }
2249   case Intrinsic::hexagon_V6_vaddcarry_128B: {
2250     std::array<SDValue, 3> Ops = {
2251         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2252     SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v128i1);
2253     Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2254     break;
2255   }
2256   case Intrinsic::hexagon_V6_vsubcarry: {
2257     std::array<SDValue, 3> Ops = {
2258         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2259     SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v64i1);
2260     Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2261     break;
2262   }
2263   case Intrinsic::hexagon_V6_vsubcarry_128B: {
2264     std::array<SDValue, 3> Ops = {
2265         {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2266     SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v128i1);
2267     Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2268     break;
2269   }
2270   default:
2271     llvm_unreachable("Unexpected HVX dual output intrinsic.");
2272   }
2273   ReplaceUses(N, Result);
2274   ReplaceUses(SDValue(N, 0), SDValue(Result, 0));
2275   ReplaceUses(SDValue(N, 1), SDValue(Result, 1));
2276   CurDAG->RemoveDeadNode(N);
2277 }
2278