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
Coloring__anona824ea740111::Coloring113 Coloring(ArrayRef<Node> Ord) : Order(Ord) {
114 build();
115 if (!color())
116 Colors.clear();
117 }
118
colors__anona824ea740111::Coloring119 const MapType &colors() const {
120 return Colors;
121 }
122
other__anona824ea740111::Coloring123 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
conj__anona824ea740111::Coloring139 Node conj(Node Pos) {
140 Node Num = Order.size();
141 return (Pos < Num/2) ? Pos + Num/2 : Pos - Num/2;
142 }
143
getColor__anona824ea740111::Coloring144 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
getUniqueColor(const NodeSet & Nodes)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
build()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
color()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)
dump() const276 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
PermNetwork__anona824ea740411::PermNetwork337 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
getControls__anona824ea740411::PermNetwork350 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
ctl__anona824ea740411::PermNetwork367 uint8_t ctl(ElemType Pos, unsigned Step) const {
368 return Table[Pos][Step];
369 }
size__anona824ea740411::PermNetwork370 unsigned size() const {
371 return Order.size();
372 }
steps__anona824ea740411::PermNetwork373 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 {
ForwardDeltaNetwork__anona824ea740411::ForwardDeltaNetwork385 ForwardDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
386
run__anona824ea740411::ForwardDeltaNetwork387 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 {
ReverseDeltaNetwork__anona824ea740411::ReverseDeltaNetwork399 ReverseDeltaNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord) {}
400
run__anona824ea740411::ReverseDeltaNetwork401 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 {
BenesNetwork__anona824ea740411::BenesNetwork413 BenesNetwork(ArrayRef<ElemType> Ord) : PermNetwork(Ord, 2) {}
414
run__anona824ea740411::BenesNetwork415 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
route(ElemType * P,RowType * T,unsigned Size,unsigned Step)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
route(ElemType * P,RowType * T,unsigned Size,unsigned Step)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
route(ElemType * P,RowType * T,unsigned Size,unsigned Step)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 {
OpRef__anona824ea740711::OpRef623 OpRef(SDValue V) : OpV(V) {}
isValue__anona824ea740711::OpRef624 bool isValue() const { return OpV.getNode() != nullptr; }
isValid__anona824ea740711::OpRef625 bool isValid() const { return isValue() || !(OpN & Invalid); }
isUndef__anona824ea740711::OpRef626 bool isUndef() const { return OpN & Undef; }
res__anona824ea740711::OpRef627 static OpRef res(int N) { return OpRef(Whole | (N & Index)); }
fail__anona824ea740711::OpRef628 static OpRef fail() { return OpRef(Invalid); }
629
lo__anona824ea740711::OpRef630 static OpRef lo(const OpRef &R) {
631 assert(!R.isValue());
632 return OpRef(R.OpN & (Undef | Index | LoHalf));
633 }
hi__anona824ea740711::OpRef634 static OpRef hi(const OpRef &R) {
635 assert(!R.isValue());
636 return OpRef(R.OpN & (Undef | Index | HiHalf));
637 }
undef__anona824ea740711::OpRef638 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:
OpRef__anona824ea740711::OpRef664 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 {
ResultStack__anona824ea740711::ResultStack677 ResultStack(SDNode *Inp)
678 : InpNode(Inp), InpTy(Inp->getValueType(0).getSimpleVT()) {}
679 SDNode *InpNode;
680 MVT InpTy;
push__anona824ea740711::ResultStack681 unsigned push(const NodeTemplate &Res) {
682 List.push_back(Res);
683 return List.size()-1;
684 }
push__anona824ea740711::ResultStack685 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 }
empty__anona824ea740711::ResultStack692 bool empty() const { return List.empty(); }
size__anona824ea740711::ResultStack693 unsigned size() const { return List.size(); }
top__anona824ea740711::ResultStack694 unsigned top() const { return size()-1; }
operator []__anona824ea740711::ResultStack695 const NodeTemplate &operator[](unsigned I) const { return List[I]; }
reset__anona824ea740711::ResultStack696 unsigned reset(unsigned NewTop) {
697 List.resize(NewTop+1);
698 return NewTop;
699 }
700
701 using BaseType = std::vector<NodeTemplate>;
begin__anona824ea740711::ResultStack702 BaseType::iterator begin() { return List.begin(); }
end__anona824ea740711::ResultStack703 BaseType::iterator end() { return List.end(); }
begin__anona824ea740711::ResultStack704 BaseType::const_iterator begin() const { return List.begin(); }
end__anona824ea740711::ResultStack705 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)
print(raw_ostream & OS,const SelectionDAG & G) const715 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
print(raw_ostream & OS,const SelectionDAG & G) const738 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
print(raw_ostream & OS,const SelectionDAG & G) const752 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 {
ShuffleMask__anona824ea740911::ShuffleMask768 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
lo__anona824ea740911::ShuffleMask780 ShuffleMask lo() const {
781 size_t H = Mask.size()/2;
782 return ShuffleMask(Mask.take_front(H));
783 }
hi__anona824ea740911::ShuffleMask784 ShuffleMask hi() const {
785 size_t H = Mask.size()/2;
786 return ShuffleMask(Mask.take_back(H));
787 }
788
print__anona824ea740911::ShuffleMask789 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
operator <<(raw_ostream & OS,const ShuffleMask & SM)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
lo(ArrayRef<int> Vuu)813 ArrayRef<int> lo(ArrayRef<int> Vuu) { return Vuu.take_front(Vuu.size() / 2); }
hi(ArrayRef<int> Vuu)814 ArrayRef<int> hi(ArrayRef<int> Vuu) { return Vuu.take_back(Vuu.size() / 2); }
815
vshuffvdd(ArrayRef<int> Vu,ArrayRef<int> Vv,unsigned Rt)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
vdealvdd(ArrayRef<int> Vu,ArrayRef<int> Vv,unsigned Rt)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
vpack(ArrayRef<int> Vu,ArrayRef<int> Vv,unsigned Size,bool TakeOdd)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
vshuff(ArrayRef<int> Vu,ArrayRef<int> Vv,unsigned Size,bool TakeOdd)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
vdeal(ArrayRef<int> Vu,ArrayRef<int> Vv,unsigned Size,bool TakeOdd)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
vdealb4w(ArrayRef<int> Vu,ArrayRef<int> Vv)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>
mask(ShuffFunc S,unsigned Length,OptArgs...args)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
getHexagonLowering(SelectionDAG & G)915 static const HexagonTargetLowering &getHexagonLowering(SelectionDAG &G) {
916 return static_cast<const HexagonTargetLowering&>(G.getTargetLoweringInfo());
917 }
getHexagonSubtarget(SelectionDAG & G)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
HvxSelectorllvm::HvxSelector930 HvxSelector(HexagonDAGToDAGISel &HS, SelectionDAG &G)
931 : Lower(getHexagonLowering(G)), ISel(HS), DAG(G),
932 HST(getHexagonSubtarget(G)), HwLen(HST.getVectorLength()) {}
933
getSingleVTllvm::HvxSelector934 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
getPairVTllvm::HvxSelector940 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
getBoolVTllvm::HvxSelector946 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
splitMask(ArrayRef<int> Mask,MutableArrayRef<int> MaskL,MutableArrayRef<int> MaskR)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
findStrip(ArrayRef<int> A,int Inc,unsigned MaxLen)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
isUndef(ArrayRef<int> Mask)1032 static bool isUndef(ArrayRef<int> Mask) {
1033 for (int Idx : Mask)
1034 if (Idx != -1)
1035 return false;
1036 return true;
1037 }
1038
isIdentity(ArrayRef<int> Mask)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
isLowHalfOnly(ArrayRef<int> Mask)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
getInputSegmentList(ShuffleMask SM,unsigned SegLen)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
getOutputSegmentMap(ShuffleMask SM,unsigned SegLen)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
packSegmentMask(ArrayRef<int> Mask,ArrayRef<unsigned> OutSegMap,unsigned SegLen,MutableArrayRef<int> PackedMask)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
selectVectorConstants(SDNode * N)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
materialize(const ResultStack & Results)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
concats(OpRef Lo,OpRef Hi,ResultStack & Results)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
funnels(OpRef Va,OpRef Vb,int Amount,ResultStack & Results)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.
packs(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results,MutableArrayRef<int> NewMask,unsigned Options)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 (int I = 0, E = SegMap.size(); I != E; ++I) {
1341 unsigned X = SegMap[I];
1342 if (X == ~0u)
1343 continue;
1344 if (Seg0 == ~0u)
1345 Seg0 = X;
1346 else if (Seg1 != ~0u)
1347 break;
1348 if (X == ~1u || X != Seg0)
1349 Seg1 = X;
1350 }
1351
1352 if (SegCount == 1) {
1353 unsigned SrcOp = SegList[0] / 2;
1354 for (int I = 0; I != static_cast<int>(VecLen); ++I) {
1355 int M = SM.Mask[I];
1356 if (M >= 0) {
1357 M -= SrcOp * HwLen;
1358 assert(M >= 0);
1359 }
1360 NewMask[I] = M;
1361 }
1362 return Inp[SrcOp];
1363 }
1364
1365 if (SegCount == 2) {
1366 // Seg0 should not be undef here: this would imply a SegList
1367 // with <= 1 elements, which was checked earlier.
1368 assert(Seg0 != ~0u);
1369
1370 // If Seg0 or Seg1 are "multi-defined", pick them from the input
1371 // segment list instead.
1372 if (Seg0 == ~1u || Seg1 == ~1u) {
1373 if (Seg0 == Seg1) {
1374 Seg0 = SegList[0];
1375 Seg1 = SegList[1];
1376 } else if (Seg0 == ~1u) {
1377 Seg0 = SegList[0] != Seg1 ? SegList[0] : SegList[1];
1378 } else {
1379 assert(Seg1 == ~1u);
1380 Seg1 = SegList[0] != Seg0 ? SegList[0] : SegList[1];
1381 }
1382 }
1383 assert(Seg0 != ~1u && Seg1 != ~1u);
1384
1385 assert(Seg0 != Seg1 && "Expecting different segments");
1386 const SDLoc &dl(Results.InpNode);
1387 Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(SegLen, dl)});
1388 OpRef HL = OpRef::res(Results.top());
1389
1390 // Va = AB, Vb = CD
1391
1392 if (Seg0 / 2 == Seg1 / 2) {
1393 // Same input vector.
1394 Va = Inp[Seg0 / 2];
1395 if (Seg0 > Seg1) {
1396 // Swap halves.
1397 Results.push(Hexagon::V6_vror, Ty, {Inp[Seg0 / 2], HL});
1398 Va = OpRef::res(Results.top());
1399 }
1400 packSegmentMask(SM.Mask, {Seg0, Seg1}, SegLen, MaskH);
1401 } else if (Seg0 % 2 == Seg1 % 2) {
1402 // Picking AC, BD, CA, or DB.
1403 // vshuff(CD,AB,HL) -> BD:AC
1404 // vshuff(AB,CD,HL) -> DB:CA
1405 auto Vs = (Seg0 == 0 || Seg0 == 1) ? std::make_pair(Vb, Va) // AC or BD
1406 : std::make_pair(Va, Vb); // CA or DB
1407 Results.push(Hexagon::V6_vshuffvdd, PairTy, {Vs.first, Vs.second, HL});
1408 OpRef P = OpRef::res(Results.top());
1409 Va = (Seg0 == 0 || Seg0 == 2) ? OpRef::lo(P) : OpRef::hi(P);
1410 packSegmentMask(SM.Mask, {Seg0, Seg1}, SegLen, MaskH);
1411 } else {
1412 // Picking AD, BC, CB, or DA.
1413 if ((Seg0 == 0 && Seg1 == 3) || (Seg0 == 2 && Seg1 == 1)) {
1414 // AD or BC: this can be done using vmux.
1415 // Q = V6_pred_scalar2 SegLen
1416 // V = V6_vmux Q, (Va, Vb) or (Vb, Va)
1417 Results.push(Hexagon::V6_pred_scalar2, getBoolVT(), {HL});
1418 OpRef Qt = OpRef::res(Results.top());
1419 auto Vs = (Seg0 == 0) ? std::make_pair(Va, Vb) // AD
1420 : std::make_pair(Vb, Va); // CB
1421 Results.push(Hexagon::V6_vmux, Ty, {Qt, Vs.first, Vs.second});
1422 Va = OpRef::res(Results.top());
1423 packSegmentMask(SM.Mask, {Seg0, Seg1}, SegLen, MaskH);
1424 } else {
1425 // BC or DA: this could be done via valign by SegLen.
1426 // Do nothing here, because valign (if possible) will be generated
1427 // later on (make sure the Seg0 values are as expected).
1428 assert(Seg0 == 1 || Seg0 == 3);
1429 }
1430 }
1431 }
1432
1433 // Check if the arguments can be packed by valign(Va,Vb) or valign(Vb,Va).
1434 // FIXME: maybe remove this?
1435 ShuffleMask SMH(MaskH);
1436 assert(SMH.Mask.size() == VecLen);
1437 shuffles::MaskT MaskA(SMH.Mask);
1438
1439 if (SMH.MaxSrc - SMH.MinSrc >= static_cast<int>(HwLen)) {
1440 // valign(Lo=Va,Hi=Vb) won't work. Try swapping Va/Vb.
1441 shuffles::MaskT Swapped(SMH.Mask);
1442 ShuffleVectorSDNode::commuteMask(Swapped);
1443 ShuffleMask SW(Swapped);
1444 if (SW.MaxSrc - SW.MinSrc < static_cast<int>(HwLen)) {
1445 MaskA.assign(SW.Mask.begin(), SW.Mask.end());
1446 std::swap(Va, Vb);
1447 }
1448 }
1449 ShuffleMask SMA(MaskA);
1450 assert(SMA.Mask.size() == VecLen);
1451
1452 if (SMA.MaxSrc - SMA.MinSrc < static_cast<int>(HwLen)) {
1453 int ShiftR = SMA.MinSrc;
1454 if (ShiftR >= static_cast<int>(HwLen)) {
1455 Va = Vb;
1456 Vb = OpRef::undef(Ty);
1457 ShiftR -= HwLen;
1458 }
1459 OpRef RetVal = valign(Va, Vb, ShiftR, Ty, Results);
1460
1461 for (int I = 0; I != static_cast<int>(VecLen); ++I) {
1462 int M = SMA.Mask[I];
1463 if (M != -1)
1464 M -= SMA.MinSrc;
1465 NewMask[I] = M;
1466 }
1467 return RetVal;
1468 }
1469
1470 // By here, packing by segment (half-vector) shuffling, and vector alignment
1471 // failed. Try vmux.
1472 // Note: since this is using the original mask, Va and Vb must not have been
1473 // modified.
1474
1475 if (Options & PackMux) {
1476 // If elements picked from Va and Vb have all different (source) indexes
1477 // (relative to the start of the argument), do a mux, and update the mask.
1478 BitVector Picked(HwLen);
1479 SmallVector<uint8_t,128> MuxBytes(HwLen);
1480 bool CanMux = true;
1481 for (int I = 0; I != static_cast<int>(VecLen); ++I) {
1482 int M = SM.Mask[I];
1483 if (M == -1)
1484 continue;
1485 if (M >= static_cast<int>(HwLen))
1486 M -= HwLen;
1487 else
1488 MuxBytes[M] = 0xFF;
1489 if (Picked[M]) {
1490 CanMux = false;
1491 break;
1492 }
1493 NewMask[I] = M;
1494 }
1495 if (CanMux)
1496 return vmuxs(MuxBytes, Va, Vb, Results);
1497 }
1498 return OpRef::fail();
1499 }
1500
1501 // Va, Vb are vector pairs. If SM only uses two single vectors from Va/Vb,
1502 // pack these vectors into a pair, and remap SM into NewMask to use the
1503 // new pair instead.
packp(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results,MutableArrayRef<int> NewMask)1504 OpRef HvxSelector::packp(ShuffleMask SM, OpRef Va, OpRef Vb,
1505 ResultStack &Results, MutableArrayRef<int> NewMask) {
1506 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1507 SmallVector<unsigned, 4> SegList = getInputSegmentList(SM.Mask, HwLen);
1508 if (SegList.empty())
1509 return OpRef::undef(getPairVT(MVT::i8));
1510
1511 // If more than two halves are used, bail.
1512 // TODO: be more aggressive here?
1513 unsigned SegCount = SegList.size();
1514 if (SegCount > 2)
1515 return OpRef::fail();
1516
1517 MVT HalfTy = getSingleVT(MVT::i8);
1518
1519 OpRef Inp[2] = { Va, Vb };
1520 OpRef Out[2] = { OpRef::undef(HalfTy), OpRef::undef(HalfTy) };
1521
1522 // Really make sure we have at most 2 vectors used in the mask.
1523 assert(SegCount <= 2);
1524
1525 for (int I = 0, E = SegList.size(); I != E; ++I) {
1526 unsigned S = SegList[I];
1527 OpRef Op = Inp[S / 2];
1528 Out[I] = (S & 1) ? OpRef::hi(Op) : OpRef::lo(Op);
1529 }
1530
1531 // NOTE: Using SegList as the packing map here (not SegMap). This works,
1532 // because we're not concerned here about the order of the segments (i.e.
1533 // single vectors) in the output pair. Changing the order of vectors is
1534 // free (as opposed to changing the order of vector halves as in packs),
1535 // and so there is no extra cost added in case the order needs to be
1536 // changed later.
1537 packSegmentMask(SM.Mask, SegList, HwLen, NewMask);
1538 return concats(Out[0], Out[1], Results);
1539 }
1540
vmuxs(ArrayRef<uint8_t> Bytes,OpRef Va,OpRef Vb,ResultStack & Results)1541 OpRef HvxSelector::vmuxs(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1542 ResultStack &Results) {
1543 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1544 MVT ByteTy = getSingleVT(MVT::i8);
1545 MVT BoolTy = MVT::getVectorVT(MVT::i1, HwLen);
1546 const SDLoc &dl(Results.InpNode);
1547 SDValue B = getVectorConstant(Bytes, dl);
1548 Results.push(Hexagon::V6_vd0, ByteTy, {});
1549 Results.push(Hexagon::V6_veqb, BoolTy, {OpRef(B), OpRef::res(-1)});
1550 Results.push(Hexagon::V6_vmux, ByteTy, {OpRef::res(-1), Vb, Va});
1551 return OpRef::res(Results.top());
1552 }
1553
vmuxp(ArrayRef<uint8_t> Bytes,OpRef Va,OpRef Vb,ResultStack & Results)1554 OpRef HvxSelector::vmuxp(ArrayRef<uint8_t> Bytes, OpRef Va, OpRef Vb,
1555 ResultStack &Results) {
1556 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1557 size_t S = Bytes.size() / 2;
1558 OpRef L = vmuxs(Bytes.take_front(S), OpRef::lo(Va), OpRef::lo(Vb), Results);
1559 OpRef H = vmuxs(Bytes.drop_front(S), OpRef::hi(Va), OpRef::hi(Vb), Results);
1560 return concats(L, H, Results);
1561 }
1562
shuffs1(ShuffleMask SM,OpRef Va,ResultStack & Results)1563 OpRef HvxSelector::shuffs1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1564 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1565 unsigned VecLen = SM.Mask.size();
1566 assert(HwLen == VecLen);
1567 (void)VecLen;
1568 assert(all_of(SM.Mask, [this](int M) { return M == -1 || M < int(HwLen); }));
1569
1570 if (isIdentity(SM.Mask))
1571 return Va;
1572 if (isUndef(SM.Mask))
1573 return OpRef::undef(getSingleVT(MVT::i8));
1574
1575 // First, check for rotations.
1576 if (auto Dist = rotationDistance(SM, VecLen)) {
1577 OpRef Rotate = funnels(Va, Va, *Dist, Results);
1578 if (Rotate.isValid())
1579 return Rotate;
1580 }
1581 unsigned HalfLen = HwLen / 2;
1582 assert(isPowerOf2_32(HalfLen));
1583
1584 // Handle special case where the output is the same half of the input
1585 // repeated twice, i.e. if Va = AB, then handle the output of AA or BB.
1586 std::pair<int, unsigned> Strip1 = findStrip(SM.Mask, 1, HalfLen);
1587 if ((Strip1.first & ~HalfLen) == 0 && Strip1.second == HalfLen) {
1588 std::pair<int, unsigned> Strip2 =
1589 findStrip(SM.Mask.drop_front(HalfLen), 1, HalfLen);
1590 if (Strip1 == Strip2) {
1591 const SDLoc &dl(Results.InpNode);
1592 Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(HalfLen, dl)});
1593 Results.push(Hexagon::V6_vshuffvdd, getPairVT(MVT::i8),
1594 {Va, Va, OpRef::res(Results.top())});
1595 OpRef S = OpRef::res(Results.top());
1596 return (Strip1.first == 0) ? OpRef::lo(S) : OpRef::hi(S);
1597 }
1598 }
1599
1600 OpRef P = perfect(SM, Va, Results);
1601 if (P.isValid())
1602 return P;
1603 return butterfly(SM, Va, Results);
1604 }
1605
shuffs2(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results)1606 OpRef HvxSelector::shuffs2(ShuffleMask SM, OpRef Va, OpRef Vb,
1607 ResultStack &Results) {
1608 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1609 if (isUndef(SM.Mask))
1610 return OpRef::undef(getSingleVT(MVT::i8));
1611
1612 OpRef C = contracting(SM, Va, Vb, Results);
1613 if (C.isValid())
1614 return C;
1615
1616 int VecLen = SM.Mask.size();
1617 shuffles::MaskT PackedMask(VecLen);
1618 OpRef P = packs(SM, Va, Vb, Results, PackedMask);
1619 if (P.isValid())
1620 return shuffs1(ShuffleMask(PackedMask), P, Results);
1621
1622 // TODO: Before we split the mask, try perfect shuffle on concatenated
1623 // operands.
1624
1625 shuffles::MaskT MaskL(VecLen), MaskR(VecLen);
1626 splitMask(SM.Mask, MaskL, MaskR);
1627
1628 OpRef L = shuffs1(ShuffleMask(MaskL), Va, Results);
1629 OpRef R = shuffs1(ShuffleMask(MaskR), Vb, Results);
1630 if (!L.isValid() || !R.isValid())
1631 return OpRef::fail();
1632
1633 SmallVector<uint8_t, 128> Bytes(VecLen);
1634 for (int I = 0; I != VecLen; ++I) {
1635 if (MaskL[I] != -1)
1636 Bytes[I] = 0xFF;
1637 }
1638 return vmuxs(Bytes, L, R, Results);
1639 }
1640
shuffp1(ShuffleMask SM,OpRef Va,ResultStack & Results)1641 OpRef HvxSelector::shuffp1(ShuffleMask SM, OpRef Va, ResultStack &Results) {
1642 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1643 int VecLen = SM.Mask.size();
1644
1645 if (isIdentity(SM.Mask))
1646 return Va;
1647 if (isUndef(SM.Mask))
1648 return OpRef::undef(getPairVT(MVT::i8));
1649
1650 shuffles::MaskT PackedMask(VecLen);
1651 OpRef P = packs(SM, OpRef::lo(Va), OpRef::hi(Va), Results, PackedMask);
1652 if (P.isValid()) {
1653 ShuffleMask PM(PackedMask);
1654 OpRef E = expanding(PM, P, Results);
1655 if (E.isValid())
1656 return E;
1657
1658 OpRef L = shuffs1(PM.lo(), P, Results);
1659 OpRef H = shuffs1(PM.hi(), P, Results);
1660 if (L.isValid() && H.isValid())
1661 return concats(L, H, Results);
1662 }
1663
1664 if (!isLowHalfOnly(SM.Mask)) {
1665 // Doing a perfect shuffle on a low-half mask (i.e. where the upper half
1666 // is all-undef) may produce a perfect shuffle that generates legitimate
1667 // upper half. This isn't wrong, but if the perfect shuffle was possible,
1668 // then there is a good chance that a shorter (contracting) code may be
1669 // used as well (e.g. V6_vshuffeb, etc).
1670 OpRef R = perfect(SM, Va, Results);
1671 if (R.isValid())
1672 return R;
1673 // TODO commute the mask and try the opposite order of the halves.
1674 }
1675
1676 OpRef L = shuffs2(SM.lo(), OpRef::lo(Va), OpRef::hi(Va), Results);
1677 OpRef H = shuffs2(SM.hi(), OpRef::lo(Va), OpRef::hi(Va), Results);
1678 if (L.isValid() && H.isValid())
1679 return concats(L, H, Results);
1680
1681 return OpRef::fail();
1682 }
1683
shuffp2(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results)1684 OpRef HvxSelector::shuffp2(ShuffleMask SM, OpRef Va, OpRef Vb,
1685 ResultStack &Results) {
1686 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1687 if (isUndef(SM.Mask))
1688 return OpRef::undef(getPairVT(MVT::i8));
1689
1690 int VecLen = SM.Mask.size();
1691 SmallVector<int,256> PackedMask(VecLen);
1692 OpRef P = packp(SM, Va, Vb, Results, PackedMask);
1693 if (P.isValid())
1694 return shuffp1(ShuffleMask(PackedMask), P, Results);
1695
1696 SmallVector<int,256> MaskL(VecLen), MaskR(VecLen);
1697 splitMask(SM.Mask, MaskL, MaskR);
1698
1699 OpRef L = shuffp1(ShuffleMask(MaskL), Va, Results);
1700 OpRef R = shuffp1(ShuffleMask(MaskR), Vb, Results);
1701 if (!L.isValid() || !R.isValid())
1702 return OpRef::fail();
1703
1704 // Mux the results.
1705 SmallVector<uint8_t,256> Bytes(VecLen);
1706 for (int I = 0; I != VecLen; ++I) {
1707 if (MaskL[I] != -1)
1708 Bytes[I] = 0xFF;
1709 }
1710 return vmuxp(Bytes, L, R, Results);
1711 }
1712
1713 namespace {
1714 struct Deleter : public SelectionDAG::DAGNodeDeletedListener {
1715 template <typename T>
Deleter__anona824ea740e11::Deleter1716 Deleter(SelectionDAG &D, T &C)
1717 : SelectionDAG::DAGNodeDeletedListener(D, [&C] (SDNode *N, SDNode *E) {
1718 C.erase(N);
1719 }) {}
1720 };
1721
1722 template <typename T>
1723 struct NullifyingVector : public T {
1724 DenseMap<SDNode*, SDNode**> Refs;
NullifyingVector__anona824ea740e11::NullifyingVector1725 NullifyingVector(T &&V) : T(V) {
1726 for (unsigned i = 0, e = T::size(); i != e; ++i) {
1727 SDNode *&N = T::operator[](i);
1728 Refs[N] = &N;
1729 }
1730 }
erase__anona824ea740e11::NullifyingVector1731 void erase(SDNode *N) {
1732 auto F = Refs.find(N);
1733 if (F != Refs.end())
1734 *F->second = nullptr;
1735 }
1736 };
1737 }
1738
select(SDNode * ISelN)1739 void HvxSelector::select(SDNode *ISelN) {
1740 // What's important here is to select the right set of nodes. The main
1741 // selection algorithm loops over nodes in a topological order, i.e. users
1742 // are visited before their operands.
1743 //
1744 // It is an error to have an unselected node with a selected operand, and
1745 // there is an assertion in the main selector code to enforce that.
1746 //
1747 // Such a situation could occur if we selected a node, which is both a
1748 // subnode of ISelN, and a subnode of an unrelated (and yet unselected)
1749 // node in the DAG.
1750 assert(ISelN->getOpcode() == HexagonISD::ISEL);
1751 SDNode *N0 = ISelN->getOperand(0).getNode();
1752 if (N0->isMachineOpcode()) {
1753 ISel.ReplaceNode(ISelN, N0);
1754 return;
1755 }
1756
1757 // There could have been nodes created (i.e. inserted into the DAG)
1758 // that are now dead. Remove them, in case they use any of the nodes
1759 // to select (and make them look shared).
1760 DAG.RemoveDeadNodes();
1761
1762 SetVector<SDNode*> SubNodes, TmpQ;
1763 std::map<SDNode*,unsigned> NumOps;
1764
1765 // Don't want to select N0 if it's shared with another node, except if
1766 // it's shared with other ISELs.
1767 auto IsISelN = [](SDNode *T) { return T->getOpcode() == HexagonISD::ISEL; };
1768 if (llvm::all_of(N0->uses(), IsISelN))
1769 SubNodes.insert(N0);
1770
1771 auto InSubNodes = [&SubNodes](SDNode *T) { return SubNodes.count(T); };
1772 for (unsigned I = 0; I != SubNodes.size(); ++I) {
1773 SDNode *S = SubNodes[I];
1774 unsigned OpN = 0;
1775 // Only add subnodes that are only reachable from N0.
1776 for (SDValue Op : S->ops()) {
1777 SDNode *O = Op.getNode();
1778 if (llvm::all_of(O->uses(), InSubNodes)) {
1779 SubNodes.insert(O);
1780 ++OpN;
1781 }
1782 }
1783 NumOps.insert({S, OpN});
1784 if (OpN == 0)
1785 TmpQ.insert(S);
1786 }
1787
1788 for (unsigned I = 0; I != TmpQ.size(); ++I) {
1789 SDNode *S = TmpQ[I];
1790 for (SDNode *U : S->uses()) {
1791 if (U == ISelN)
1792 continue;
1793 auto F = NumOps.find(U);
1794 assert(F != NumOps.end());
1795 if (F->second > 0 && !--F->second)
1796 TmpQ.insert(F->first);
1797 }
1798 }
1799
1800 // Remove the marker.
1801 ISel.ReplaceNode(ISelN, N0);
1802
1803 assert(SubNodes.size() == TmpQ.size());
1804 NullifyingVector<decltype(TmpQ)::vector_type> Queue(TmpQ.takeVector());
1805
1806 Deleter DUQ(DAG, Queue);
1807 for (SDNode *S : reverse(Queue)) {
1808 if (S == nullptr)
1809 continue;
1810 DEBUG_WITH_TYPE("isel", {dbgs() << "HVX selecting: "; S->dump(&DAG);});
1811 ISel.Select(S);
1812 }
1813 }
1814
scalarizeShuffle(ArrayRef<int> Mask,const SDLoc & dl,MVT ResTy,SDValue Va,SDValue Vb,SDNode * N)1815 bool HvxSelector::scalarizeShuffle(ArrayRef<int> Mask, const SDLoc &dl,
1816 MVT ResTy, SDValue Va, SDValue Vb,
1817 SDNode *N) {
1818 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
1819 MVT ElemTy = ResTy.getVectorElementType();
1820 assert(ElemTy == MVT::i8);
1821 unsigned VecLen = Mask.size();
1822 bool HavePairs = (2*HwLen == VecLen);
1823 MVT SingleTy = getSingleVT(MVT::i8);
1824
1825 // The prior attempts to handle this shuffle may have left a bunch of
1826 // dead nodes in the DAG (such as constants). These nodes will be added
1827 // at the end of DAG's node list, which at that point had already been
1828 // sorted topologically. In the main selection loop, the node list is
1829 // traversed backwards from the root node, which means that any new
1830 // nodes (from the end of the list) will not be visited.
1831 // Scalarization will replace the shuffle node with the scalarized
1832 // expression, and if that expression reused any if the leftoever (dead)
1833 // nodes, these nodes would not be selected (since the "local" selection
1834 // only visits nodes that are not in AllNodes).
1835 // To avoid this issue, remove all dead nodes from the DAG now.
1836 // DAG.RemoveDeadNodes();
1837
1838 SmallVector<SDValue,128> Ops;
1839 LLVMContext &Ctx = *DAG.getContext();
1840 MVT LegalTy = Lower.getTypeToTransformTo(Ctx, ElemTy).getSimpleVT();
1841 for (int I : Mask) {
1842 if (I < 0) {
1843 Ops.push_back(ISel.selectUndef(dl, LegalTy));
1844 continue;
1845 }
1846 SDValue Vec;
1847 unsigned M = I;
1848 if (M < VecLen) {
1849 Vec = Va;
1850 } else {
1851 Vec = Vb;
1852 M -= VecLen;
1853 }
1854 if (HavePairs) {
1855 if (M < HwLen) {
1856 Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, SingleTy, Vec);
1857 } else {
1858 Vec = DAG.getTargetExtractSubreg(Hexagon::vsub_hi, dl, SingleTy, Vec);
1859 M -= HwLen;
1860 }
1861 }
1862 SDValue Idx = DAG.getConstant(M, dl, MVT::i32);
1863 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, LegalTy, {Vec, Idx});
1864 SDValue L = Lower.LowerOperation(Ex, DAG);
1865 assert(L.getNode());
1866 Ops.push_back(L);
1867 }
1868
1869 SDValue LV;
1870 if (2*HwLen == VecLen) {
1871 SDValue B0 = DAG.getBuildVector(SingleTy, dl, {Ops.data(), HwLen});
1872 SDValue L0 = Lower.LowerOperation(B0, DAG);
1873 SDValue B1 = DAG.getBuildVector(SingleTy, dl, {Ops.data()+HwLen, HwLen});
1874 SDValue L1 = Lower.LowerOperation(B1, DAG);
1875 // XXX CONCAT_VECTORS is legal for HVX vectors. Legalizing (lowering)
1876 // functions may expect to be called only for illegal operations, so
1877 // make sure that they are not called for legal ones. Develop a better
1878 // mechanism for dealing with this.
1879 LV = DAG.getNode(ISD::CONCAT_VECTORS, dl, ResTy, {L0, L1});
1880 } else {
1881 SDValue BV = DAG.getBuildVector(ResTy, dl, Ops);
1882 LV = Lower.LowerOperation(BV, DAG);
1883 }
1884
1885 assert(!N->use_empty());
1886 SDValue IS = DAG.getNode(HexagonISD::ISEL, dl, ResTy, LV);
1887 ISel.ReplaceNode(N, IS.getNode());
1888 select(IS.getNode());
1889 DAG.RemoveDeadNodes();
1890 return true;
1891 }
1892
getPerfectCompletions(ShuffleMask SM,unsigned Width)1893 SmallVector<uint32_t, 8> HvxSelector::getPerfectCompletions(ShuffleMask SM,
1894 unsigned Width) {
1895 auto possibilities = [](ArrayRef<uint8_t> Bs, unsigned Width) -> uint32_t {
1896 unsigned Impossible = ~(1u << Width) + 1;
1897 for (unsigned I = 0, E = Bs.size(); I != E; ++I) {
1898 uint8_t B = Bs[I];
1899 if (B == 0xff)
1900 continue;
1901 if (~Impossible == 0)
1902 break;
1903 for (unsigned Log = 0; Log != Width; ++Log) {
1904 if (Impossible & (1u << Log))
1905 continue;
1906 unsigned Expected = (I >> Log) % 2;
1907 if (B != Expected)
1908 Impossible |= (1u << Log);
1909 }
1910 }
1911 return ~Impossible;
1912 };
1913
1914 SmallVector<uint32_t, 8> Worklist(Width);
1915
1916 for (unsigned BitIdx = 0; BitIdx != Width; ++BitIdx) {
1917 SmallVector<uint8_t> BitValues(SM.Mask.size());
1918 for (int i = 0, e = SM.Mask.size(); i != e; ++i) {
1919 int M = SM.Mask[i];
1920 if (M < 0)
1921 BitValues[i] = 0xff;
1922 else
1923 BitValues[i] = (M & (1u << BitIdx)) != 0;
1924 }
1925 Worklist[BitIdx] = possibilities(BitValues, Width);
1926 }
1927
1928 // If there is a word P in Worklist that matches multiple possibilities,
1929 // then if any other word Q matches any of the possibilities matched by P,
1930 // then Q matches all the possibilities matched by P. In fact, P == Q.
1931 // In other words, for each words P, Q, the sets of possibilities matched
1932 // by P and Q are either equal or disjoint (no partial overlap).
1933 //
1934 // Illustration: For 4-bit values there are 4 complete sequences:
1935 // a: 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1
1936 // b: 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1
1937 // c: 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1
1938 // d: 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1
1939 //
1940 // Words containing unknown bits that match two of the complete
1941 // sequences:
1942 // ab: 0 u u 1 0 u u 1 0 u u 1 0 u u 1
1943 // ac: 0 u 0 u u 1 u 1 0 u 0 u u 1 u 1
1944 // ad: 0 u 0 u 0 u 0 u u 1 u 1 u 1 u 1
1945 // bc: 0 0 u u u u 1 1 0 0 u u u u 1 1
1946 // bd: 0 0 u u 0 0 u u u u 1 1 u u 1 1
1947 // cd: 0 0 0 0 u u u u u u u u 1 1 1 1
1948 //
1949 // Proof of the claim above:
1950 // Let P be a word that matches s0 and s1. For that to happen, all known
1951 // bits in P must match s0 and s1 exactly.
1952 // Assume there is Q that matches s1. Note that since P and Q came from
1953 // the same shuffle mask, the positions of unknown bits in P and Q match
1954 // exactly, which makes the indices of known bits be exactly the same
1955 // between P and Q. Since P matches s0 and s1, the known bits of P much
1956 // match both s0 and s1. Also, since Q matches s1, the known bits in Q
1957 // are exactly the same as in s1, which means that they are exactly the
1958 // same as in P. This implies that P == Q.
1959
1960 // There can be a situation where there are more entries with the same
1961 // bits set than there are set bits (e.g. value 9 occuring more than 2
1962 // times). In such cases it will be impossible to complete this to a
1963 // perfect shuffle.
1964 SmallVector<uint32_t, 8> Sorted(Worklist);
1965 llvm::sort(Sorted.begin(), Sorted.end());
1966
1967 for (unsigned I = 0, E = Sorted.size(); I != E;) {
1968 unsigned P = Sorted[I], Count = 1;
1969 while (++I != E && P == Sorted[I])
1970 ++Count;
1971 if ((unsigned)llvm::popcount(P) < Count) {
1972 // Reset all occurences of P, if there are more occurrences of P
1973 // than there are bits in P.
1974 for_each(Worklist, [P](unsigned &Q) {
1975 if (Q == P)
1976 Q = 0;
1977 });
1978 }
1979 }
1980
1981 return Worklist;
1982 }
1983
1984 SmallVector<uint32_t, 8>
completeToPerfect(ArrayRef<uint32_t> Completions,unsigned Width)1985 HvxSelector::completeToPerfect(ArrayRef<uint32_t> Completions, unsigned Width) {
1986 // Pick a completion if there are multiple possibilities. For now just
1987 // select any valid completion.
1988 SmallVector<uint32_t, 8> Comps(Completions);
1989
1990 for (unsigned I = 0; I != Width; ++I) {
1991 uint32_t P = Comps[I];
1992 assert(P != 0);
1993 if (isPowerOf2_32(P))
1994 continue;
1995 // T = least significant bit of P.
1996 uint32_t T = P ^ ((P - 1) & P);
1997 // Clear T in all remaining words matching P.
1998 for (unsigned J = I + 1; J != Width; ++J) {
1999 if (Comps[J] == P)
2000 Comps[J] ^= T;
2001 }
2002 Comps[I] = T;
2003 }
2004
2005 #ifndef NDEBUG
2006 // Check that we have generated a valid completion.
2007 uint32_t OrAll = 0;
2008 for (unsigned I = 0, E = Comps.size(); I != E; ++I) {
2009 uint32_t C = Comps[I];
2010 assert(isPowerOf2_32(C));
2011 OrAll |= C;
2012 }
2013 assert(OrAll == (1u << Width) -1);
2014 #endif
2015
2016 return Comps;
2017 }
2018
rotationDistance(ShuffleMask SM,unsigned WrapAt)2019 std::optional<int> HvxSelector::rotationDistance(ShuffleMask SM,
2020 unsigned WrapAt) {
2021 std::optional<int> Dist;
2022 for (int I = 0, E = SM.Mask.size(); I != E; ++I) {
2023 int M = SM.Mask[I];
2024 if (M < 0)
2025 continue;
2026 if (Dist) {
2027 if ((I + *Dist) % static_cast<int>(WrapAt) != M)
2028 return std::nullopt;
2029 } else {
2030 // Integer a%b operator assumes rounding towards zero by /, so it
2031 // "misbehaves" when a crosses 0 (the remainder also changes sign).
2032 // Add WrapAt in an attempt to keep I+Dist non-negative.
2033 Dist = M - I;
2034 if (Dist < 0)
2035 Dist = *Dist + WrapAt;
2036 }
2037 }
2038 return Dist;
2039 }
2040
contracting(ShuffleMask SM,OpRef Va,OpRef Vb,ResultStack & Results)2041 OpRef HvxSelector::contracting(ShuffleMask SM, OpRef Va, OpRef Vb,
2042 ResultStack &Results) {
2043 DEBUG_WITH_TYPE("isel", { dbgs() << __func__ << '\n'; });
2044 if (!Va.isValid() || !Vb.isValid())
2045 return OpRef::fail();
2046
2047 // Contracting shuffles, i.e. instructions that always discard some bytes
2048 // from the operand vectors.
2049 //
2050 // Funnel shifts
2051 // V6_vshuff{e,o}b
2052 // V6_vshuf{e,o}h
2053 // V6_vdealb4w
2054 // V6_vpack{e,o}{b,h}
2055
2056 int VecLen = SM.Mask.size();
2057
2058 // First, check for funnel shifts.
2059 if (auto Dist = rotationDistance(SM, 2 * VecLen)) {
2060 OpRef Funnel = funnels(Va, Vb, *Dist, Results);
2061 if (Funnel.isValid())
2062 return Funnel;
2063 }
2064
2065 MVT SingleTy = getSingleVT(MVT::i8);
2066 MVT PairTy = getPairVT(MVT::i8);
2067
2068 auto same = [](ArrayRef<int> Mask1, ArrayRef<int> Mask2) -> bool {
2069 return Mask1 == Mask2;
2070 };
2071
2072 using PackConfig = std::pair<unsigned, bool>;
2073 PackConfig Packs[] = {
2074 {1, false}, // byte, even
2075 {1, true}, // byte, odd
2076 {2, false}, // half, even
2077 {2, true}, // half, odd
2078 };
2079
2080 { // Check vpack
2081 unsigned Opcodes[] = {
2082 Hexagon::V6_vpackeb,
2083 Hexagon::V6_vpackob,
2084 Hexagon::V6_vpackeh,
2085 Hexagon::V6_vpackoh,
2086 };
2087 for (int i = 0, e = std::size(Opcodes); i != e; ++i) {
2088 auto [Size, Odd] = Packs[i];
2089 if (same(SM.Mask, shuffles::mask(shuffles::vpack, HwLen, Size, Odd))) {
2090 Results.push(Opcodes[i], SingleTy, {Vb, Va});
2091 return OpRef::res(Results.top());
2092 }
2093 }
2094 }
2095
2096 { // Check vshuff
2097 unsigned Opcodes[] = {
2098 Hexagon::V6_vshuffeb,
2099 Hexagon::V6_vshuffob,
2100 Hexagon::V6_vshufeh,
2101 Hexagon::V6_vshufoh,
2102 };
2103 for (int i = 0, e = std::size(Opcodes); i != e; ++i) {
2104 auto [Size, Odd] = Packs[i];
2105 if (same(SM.Mask, shuffles::mask(shuffles::vshuff, HwLen, Size, Odd))) {
2106 Results.push(Opcodes[i], SingleTy, {Vb, Va});
2107 return OpRef::res(Results.top());
2108 }
2109 }
2110 }
2111
2112 { // Check vdeal
2113 // There is no "V6_vdealeb", etc, but the supposed behavior of vdealeb
2114 // is equivalent to "(V6_vpackeb (V6_vdealvdd Vu, Vv, -2))". Other such
2115 // variants of "deal" can be done similarly.
2116 unsigned Opcodes[] = {
2117 Hexagon::V6_vpackeb,
2118 Hexagon::V6_vpackob,
2119 Hexagon::V6_vpackeh,
2120 Hexagon::V6_vpackoh,
2121 };
2122 const SDLoc &dl(Results.InpNode);
2123
2124 for (int i = 0, e = std::size(Opcodes); i != e; ++i) {
2125 auto [Size, Odd] = Packs[i];
2126 if (same(SM.Mask, shuffles::mask(shuffles::vdeal, HwLen, Size, Odd))) {
2127 Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(-2 * Size, dl)});
2128 Results.push(Hexagon::V6_vdealvdd, PairTy, {Vb, Va, OpRef::res(-1)});
2129 auto vdeal = OpRef::res(Results.top());
2130 Results.push(Opcodes[i], SingleTy,
2131 {OpRef::hi(vdeal), OpRef::lo(vdeal)});
2132 return OpRef::res(Results.top());
2133 }
2134 }
2135 }
2136
2137 if (same(SM.Mask, shuffles::mask(shuffles::vdealb4w, HwLen))) {
2138 Results.push(Hexagon::V6_vdealb4w, SingleTy, {Vb, Va});
2139 return OpRef::res(Results.top());
2140 }
2141
2142 return OpRef::fail();
2143 }
2144
expanding(ShuffleMask SM,OpRef Va,ResultStack & Results)2145 OpRef HvxSelector::expanding(ShuffleMask SM, OpRef Va, ResultStack &Results) {
2146 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
2147 // Expanding shuffles (using all elements and inserting into larger vector):
2148 //
2149 // V6_vunpacku{b,h} [*]
2150 //
2151 // [*] Only if the upper elements (filled with 0s) are "don't care" in Mask.
2152 //
2153 // Note: V6_vunpacko{b,h} are or-ing the high byte/half in the result, so
2154 // they are not shuffles.
2155 //
2156 // The argument is a single vector.
2157
2158 int VecLen = SM.Mask.size();
2159 assert(2*HwLen == unsigned(VecLen) && "Expecting vector-pair type");
2160
2161 std::pair<int,unsigned> Strip = findStrip(SM.Mask, 1, VecLen);
2162
2163 // The patterns for the unpacks, in terms of the starting offsets of the
2164 // consecutive strips (L = length of the strip, N = VecLen):
2165 //
2166 // vunpacku: 0, -1, L, -1, 2L, -1 ...
2167
2168 if (Strip.first != 0)
2169 return OpRef::fail();
2170
2171 // The vunpackus only handle byte and half-word.
2172 if (Strip.second != 1 && Strip.second != 2)
2173 return OpRef::fail();
2174
2175 int N = VecLen;
2176 int L = Strip.second;
2177
2178 // First, check the non-ignored strips.
2179 for (int I = 2*L; I < N; I += 2*L) {
2180 auto S = findStrip(SM.Mask.drop_front(I), 1, N-I);
2181 if (S.second != unsigned(L))
2182 return OpRef::fail();
2183 if (2*S.first != I)
2184 return OpRef::fail();
2185 }
2186 // Check the -1s.
2187 for (int I = L; I < N; I += 2*L) {
2188 auto S = findStrip(SM.Mask.drop_front(I), 0, N-I);
2189 if (S.first != -1 || S.second != unsigned(L))
2190 return OpRef::fail();
2191 }
2192
2193 unsigned Opc = Strip.second == 1 ? Hexagon::V6_vunpackub
2194 : Hexagon::V6_vunpackuh;
2195 Results.push(Opc, getPairVT(MVT::i8), {Va});
2196 return OpRef::res(Results.top());
2197 }
2198
perfect(ShuffleMask SM,OpRef Va,ResultStack & Results)2199 OpRef HvxSelector::perfect(ShuffleMask SM, OpRef Va, ResultStack &Results) {
2200 DEBUG_WITH_TYPE("isel", { dbgs() << __func__ << '\n'; });
2201 // V6_vdeal{b,h}
2202 // V6_vshuff{b,h}
2203
2204 // V6_vshufoe{b,h} those are quivalent to vshuffvdd(..,{1,2})
2205 // V6_vshuffvdd (V6_vshuff)
2206 // V6_dealvdd (V6_vdeal)
2207
2208 int VecLen = SM.Mask.size();
2209 assert(isPowerOf2_32(VecLen) && Log2_32(VecLen) <= 8);
2210 unsigned LogLen = Log2_32(VecLen);
2211 unsigned HwLog = Log2_32(HwLen);
2212 // The result length must be the same as the length of a single vector,
2213 // or a vector pair.
2214 assert(LogLen == HwLog || LogLen == HwLog + 1);
2215 bool HavePairs = LogLen == HwLog + 1;
2216
2217 SmallVector<unsigned, 8> Perm(LogLen);
2218
2219 // Check if this could be a perfect shuffle, or a combination of perfect
2220 // shuffles.
2221 //
2222 // Consider this permutation (using hex digits to make the ASCII diagrams
2223 // easier to read):
2224 // { 0, 8, 1, 9, 2, A, 3, B, 4, C, 5, D, 6, E, 7, F }.
2225 // This is a "deal" operation: divide the input into two halves, and
2226 // create the output by picking elements by alternating between these two
2227 // halves:
2228 // 0 1 2 3 4 5 6 7 --> 0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F [*]
2229 // 8 9 A B C D E F
2230 //
2231 // Aside from a few special explicit cases (V6_vdealb, etc.), HVX provides
2232 // a somwehat different mechanism that could be used to perform shuffle/
2233 // deal operations: a 2x2 transpose.
2234 // Consider the halves of inputs again, they can be interpreted as a 2x8
2235 // matrix. A 2x8 matrix can be looked at four 2x2 matrices concatenated
2236 // together. Now, when considering 2 elements at a time, it will be a 2x4
2237 // matrix (with elements 01, 23, 45, etc.), or two 2x2 matrices:
2238 // 01 23 45 67
2239 // 89 AB CD EF
2240 // With groups of 4, this will become a single 2x2 matrix, and so on.
2241 //
2242 // The 2x2 transpose instruction works by transposing each of the 2x2
2243 // matrices (or "sub-matrices"), given a specific group size. For example,
2244 // if the group size is 1 (i.e. each element is its own group), there
2245 // will be four transposes of the four 2x2 matrices that form the 2x8.
2246 // For example, with the inputs as above, the result will be:
2247 // 0 8 2 A 4 C 6 E
2248 // 1 9 3 B 5 D 7 F
2249 // Now, this result can be tranposed again, but with the group size of 2:
2250 // 08 19 4C 5D
2251 // 2A 3B 6E 7F
2252 // If we then transpose that result, but with the group size of 4, we get:
2253 // 0819 2A3B
2254 // 4C5D 6E7F
2255 // If we concatenate these two rows, it will be
2256 // 0 8 1 9 2 A 3 B 4 C 5 D 6 E 7 F
2257 // which is the same as the "deal" [*] above.
2258 //
2259 // In general, a "deal" of individual elements is a series of 2x2 transposes,
2260 // with changing group size. HVX has two instructions:
2261 // Vdd = V6_vdealvdd Vu, Vv, Rt
2262 // Vdd = V6_shufvdd Vu, Vv, Rt
2263 // that perform exactly that. The register Rt controls which transposes are
2264 // going to happen: a bit at position n (counting from 0) indicates that a
2265 // transpose with a group size of 2^n will take place. If multiple bits are
2266 // set, multiple transposes will happen: vdealvdd will perform them starting
2267 // with the largest group size, vshuffvdd will do them in the reverse order.
2268 //
2269 // The main observation is that each 2x2 transpose corresponds to swapping
2270 // columns of bits in the binary representation of the values.
2271 //
2272 // The numbers {3,2,1,0} and the log2 of the number of contiguous 1 bits
2273 // in a given column. The * denote the columns that will be swapped.
2274 // The transpose with the group size 2^n corresponds to swapping columns
2275 // 3 (the highest log) and log2(n):
2276 //
2277 // 3 2 1 0 0 2 1 3 0 2 3 1
2278 // * * * * * *
2279 // 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2280 // 1 0 0 0 1 8 1 0 0 0 8 1 0 0 0 8 1 0 0 0
2281 // 2 0 0 1 0 2 0 0 1 0 1 0 0 0 1 1 0 0 0 1
2282 // 3 0 0 1 1 A 1 0 1 0 9 1 0 0 1 9 1 0 0 1
2283 // 4 0 1 0 0 4 0 1 0 0 4 0 1 0 0 2 0 0 1 0
2284 // 5 0 1 0 1 C 1 1 0 0 C 1 1 0 0 A 1 0 1 0
2285 // 6 0 1 1 0 6 0 1 1 0 5 0 1 0 1 3 0 0 1 1
2286 // 7 0 1 1 1 E 1 1 1 0 D 1 1 0 1 B 1 0 1 1
2287 // 8 1 0 0 0 1 0 0 0 1 2 0 0 1 0 4 0 1 0 0
2288 // 9 1 0 0 1 9 1 0 0 1 A 1 0 1 0 C 1 1 0 0
2289 // A 1 0 1 0 3 0 0 1 1 3 0 0 1 1 5 0 1 0 1
2290 // B 1 0 1 1 B 1 0 1 1 B 1 0 1 1 D 1 1 0 1
2291 // C 1 1 0 0 5 0 1 0 1 6 0 1 1 0 6 0 1 1 0
2292 // D 1 1 0 1 D 1 1 0 1 E 1 1 1 0 E 1 1 1 0
2293 // E 1 1 1 0 7 0 1 1 1 7 0 1 1 1 7 0 1 1 1
2294 // F 1 1 1 1 F 1 1 1 1 F 1 1 1 1 F 1 1 1 1
2295
2296 // There is one special case that is not a perfect shuffle, but can be
2297 // turned into one easily: when the shuffle operates on a vector pair,
2298 // but the two vectors in the pair are swapped. The code that identifies
2299 // perfect shuffles will reject it, unless the order is reversed.
2300 shuffles::MaskT MaskStorage(SM.Mask);
2301 bool InvertedPair = false;
2302 if (HavePairs && SM.Mask[0] >= int(HwLen)) {
2303 for (int i = 0, e = SM.Mask.size(); i != e; ++i) {
2304 int M = SM.Mask[i];
2305 MaskStorage[i] = M >= int(HwLen) ? M - HwLen : M + HwLen;
2306 }
2307 InvertedPair = true;
2308 SM = ShuffleMask(MaskStorage);
2309 }
2310
2311 auto Comps = getPerfectCompletions(SM, LogLen);
2312 if (llvm::any_of(Comps, [](uint32_t P) { return P == 0; }))
2313 return OpRef::fail();
2314
2315 auto Pick = completeToPerfect(Comps, LogLen);
2316 for (unsigned I = 0; I != LogLen; ++I)
2317 Perm[I] = Log2_32(Pick[I]);
2318
2319 // Once we have Perm, represent it as cycles. Denote the maximum log2
2320 // (equal to log2(VecLen)-1) as M. The cycle containing M can then be
2321 // written as (M a1 a2 a3 ... an). That cycle can be broken up into
2322 // simple swaps as (M a1)(M a2)(M a3)...(M an), with the composition
2323 // order being from left to right. Any (contiguous) segment where the
2324 // values ai, ai+1...aj are either all increasing or all decreasing,
2325 // can be implemented via a single vshuffvdd/vdealvdd respectively.
2326 //
2327 // If there is a cycle (a1 a2 ... an) that does not involve M, it can
2328 // be written as (M an)(a1 a2 ... an)(M a1). The first two cycles can
2329 // then be folded to get (M a1 a2 ... an)(M a1), and the above procedure
2330 // can be used to generate a sequence of vshuffvdd/vdealvdd.
2331 //
2332 // Example:
2333 // Assume M = 4 and consider a permutation (0 1)(2 3). It can be written
2334 // as (4 0 1)(4 0) composed with (4 2 3)(4 2), or simply
2335 // (4 0 1)(4 0)(4 2 3)(4 2).
2336 // It can then be expanded into swaps as
2337 // (4 0)(4 1)(4 0)(4 2)(4 3)(4 2),
2338 // and broken up into "increasing" segments as
2339 // [(4 0)(4 1)] [(4 0)(4 2)(4 3)] [(4 2)].
2340 // This is equivalent to
2341 // (4 0 1)(4 0 2 3)(4 2),
2342 // which can be implemented as 3 vshufvdd instructions.
2343
2344 using CycleType = SmallVector<unsigned, 8>;
2345 std::set<CycleType> Cycles;
2346 std::set<unsigned> All;
2347
2348 for (unsigned I : Perm)
2349 All.insert(I);
2350
2351 // If the cycle contains LogLen-1, move it to the front of the cycle.
2352 // Otherwise, return the cycle unchanged.
2353 auto canonicalize = [LogLen](const CycleType &C) -> CycleType {
2354 unsigned LogPos, N = C.size();
2355 for (LogPos = 0; LogPos != N; ++LogPos)
2356 if (C[LogPos] == LogLen - 1)
2357 break;
2358 if (LogPos == N)
2359 return C;
2360
2361 CycleType NewC(C.begin() + LogPos, C.end());
2362 NewC.append(C.begin(), C.begin() + LogPos);
2363 return NewC;
2364 };
2365
2366 auto pfs = [](const std::set<CycleType> &Cs, unsigned Len) {
2367 // Ordering: shuff: 5 0 1 2 3 4, deal: 5 4 3 2 1 0 (for Log=6),
2368 // for bytes zero is included, for halfwords is not.
2369 if (Cs.size() != 1)
2370 return 0u;
2371 const CycleType &C = *Cs.begin();
2372 if (C[0] != Len - 1)
2373 return 0u;
2374 int D = Len - C.size();
2375 if (D != 0 && D != 1)
2376 return 0u;
2377
2378 bool IsDeal = true, IsShuff = true;
2379 for (unsigned I = 1; I != Len - D; ++I) {
2380 if (C[I] != Len - 1 - I)
2381 IsDeal = false;
2382 if (C[I] != I - (1 - D)) // I-1, I
2383 IsShuff = false;
2384 }
2385 // At most one, IsDeal or IsShuff, can be non-zero.
2386 assert(!(IsDeal || IsShuff) || IsDeal != IsShuff);
2387 static unsigned Deals[] = {Hexagon::V6_vdealb, Hexagon::V6_vdealh};
2388 static unsigned Shufs[] = {Hexagon::V6_vshuffb, Hexagon::V6_vshuffh};
2389 return IsDeal ? Deals[D] : (IsShuff ? Shufs[D] : 0);
2390 };
2391
2392 while (!All.empty()) {
2393 unsigned A = *All.begin();
2394 All.erase(A);
2395 CycleType C;
2396 C.push_back(A);
2397 for (unsigned B = Perm[A]; B != A; B = Perm[B]) {
2398 C.push_back(B);
2399 All.erase(B);
2400 }
2401 if (C.size() <= 1)
2402 continue;
2403 Cycles.insert(canonicalize(C));
2404 }
2405
2406 MVT SingleTy = getSingleVT(MVT::i8);
2407 MVT PairTy = getPairVT(MVT::i8);
2408
2409 // Recognize patterns for V6_vdeal{b,h} and V6_vshuff{b,h}.
2410 if (unsigned(VecLen) == HwLen) {
2411 if (unsigned SingleOpc = pfs(Cycles, LogLen)) {
2412 Results.push(SingleOpc, SingleTy, {Va});
2413 return OpRef::res(Results.top());
2414 }
2415 }
2416
2417 // From the cycles, construct the sequence of values that will
2418 // then form the control values for vdealvdd/vshuffvdd, i.e.
2419 // (M a1 a2)(M a3 a4 a5)... -> a1 a2 a3 a4 a5
2420 // This essentially strips the M value from the cycles where
2421 // it's present, and performs the insertion of M (then stripping)
2422 // for cycles without M (as described in an earlier comment).
2423 SmallVector<unsigned, 8> SwapElems;
2424 // When the input is extended (i.e. single vector becomes a pair),
2425 // this is done by using an "undef" vector as the second input.
2426 // However, then we get
2427 // input 1: GOODBITS
2428 // input 2: ........
2429 // but we need
2430 // input 1: ....BITS
2431 // input 2: ....GOOD
2432 // Then at the end, this needs to be undone. To accomplish this,
2433 // artificially add "LogLen-1" at both ends of the sequence.
2434 if (!HavePairs)
2435 SwapElems.push_back(LogLen - 1);
2436 for (const CycleType &C : Cycles) {
2437 // Do the transformation: (a1..an) -> (M a1..an)(M a1).
2438 unsigned First = (C[0] == LogLen - 1) ? 1 : 0;
2439 SwapElems.append(C.begin() + First, C.end());
2440 if (First == 0)
2441 SwapElems.push_back(C[0]);
2442 }
2443 if (!HavePairs)
2444 SwapElems.push_back(LogLen - 1);
2445
2446 const SDLoc &dl(Results.InpNode);
2447 OpRef Arg = HavePairs ? Va : concats(Va, OpRef::undef(SingleTy), Results);
2448 if (InvertedPair)
2449 Arg = concats(OpRef::hi(Arg), OpRef::lo(Arg), Results);
2450
2451 for (unsigned I = 0, E = SwapElems.size(); I != E;) {
2452 bool IsInc = I == E - 1 || SwapElems[I] < SwapElems[I + 1];
2453 unsigned S = (1u << SwapElems[I]);
2454 if (I < E - 1) {
2455 while (++I < E - 1 && IsInc == (SwapElems[I] < SwapElems[I + 1]))
2456 S |= 1u << SwapElems[I];
2457 // The above loop will not add a bit for the final SwapElems[I+1],
2458 // so add it here.
2459 S |= 1u << SwapElems[I];
2460 }
2461 ++I;
2462
2463 NodeTemplate Res;
2464 Results.push(Hexagon::A2_tfrsi, MVT::i32, {getConst32(S, dl)});
2465 Res.Opc = IsInc ? Hexagon::V6_vshuffvdd : Hexagon::V6_vdealvdd;
2466 Res.Ty = PairTy;
2467 Res.Ops = {OpRef::hi(Arg), OpRef::lo(Arg), OpRef::res(-1)};
2468 Results.push(Res);
2469 Arg = OpRef::res(Results.top());
2470 }
2471
2472 return HavePairs ? Arg : OpRef::lo(Arg);
2473 }
2474
butterfly(ShuffleMask SM,OpRef Va,ResultStack & Results)2475 OpRef HvxSelector::butterfly(ShuffleMask SM, OpRef Va, ResultStack &Results) {
2476 DEBUG_WITH_TYPE("isel", {dbgs() << __func__ << '\n';});
2477 // Butterfly shuffles.
2478 //
2479 // V6_vdelta
2480 // V6_vrdelta
2481 // V6_vror
2482
2483 // The assumption here is that all elements picked by Mask are in the
2484 // first operand to the vector_shuffle. This assumption is enforced
2485 // by the caller.
2486
2487 MVT ResTy = getSingleVT(MVT::i8);
2488 PermNetwork::Controls FC, RC;
2489 const SDLoc &dl(Results.InpNode);
2490 int VecLen = SM.Mask.size();
2491
2492 for (int M : SM.Mask) {
2493 if (M != -1 && M >= VecLen)
2494 return OpRef::fail();
2495 }
2496
2497 // Try the deltas/benes for both single vectors and vector pairs.
2498 ForwardDeltaNetwork FN(SM.Mask);
2499 if (FN.run(FC)) {
2500 SDValue Ctl = getVectorConstant(FC, dl);
2501 Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(Ctl)});
2502 return OpRef::res(Results.top());
2503 }
2504
2505 // Try reverse delta.
2506 ReverseDeltaNetwork RN(SM.Mask);
2507 if (RN.run(RC)) {
2508 SDValue Ctl = getVectorConstant(RC, dl);
2509 Results.push(Hexagon::V6_vrdelta, ResTy, {Va, OpRef(Ctl)});
2510 return OpRef::res(Results.top());
2511 }
2512
2513 // Do Benes.
2514 BenesNetwork BN(SM.Mask);
2515 if (BN.run(FC, RC)) {
2516 SDValue CtlF = getVectorConstant(FC, dl);
2517 SDValue CtlR = getVectorConstant(RC, dl);
2518 Results.push(Hexagon::V6_vdelta, ResTy, {Va, OpRef(CtlF)});
2519 Results.push(Hexagon::V6_vrdelta, ResTy,
2520 {OpRef::res(-1), OpRef(CtlR)});
2521 return OpRef::res(Results.top());
2522 }
2523
2524 return OpRef::fail();
2525 }
2526
getConst32(int Val,const SDLoc & dl)2527 SDValue HvxSelector::getConst32(int Val, const SDLoc &dl) {
2528 return DAG.getTargetConstant(Val, dl, MVT::i32);
2529 }
2530
getVectorConstant(ArrayRef<uint8_t> Data,const SDLoc & dl)2531 SDValue HvxSelector::getVectorConstant(ArrayRef<uint8_t> Data,
2532 const SDLoc &dl) {
2533 SmallVector<SDValue, 128> Elems;
2534 for (uint8_t C : Data)
2535 Elems.push_back(DAG.getConstant(C, dl, MVT::i8));
2536 MVT VecTy = MVT::getVectorVT(MVT::i8, Data.size());
2537 SDValue BV = DAG.getBuildVector(VecTy, dl, Elems);
2538 SDValue LV = Lower.LowerOperation(BV, DAG);
2539 DAG.RemoveDeadNode(BV.getNode());
2540 return DAG.getNode(HexagonISD::ISEL, dl, VecTy, LV);
2541 }
2542
selectExtractSubvector(SDNode * N)2543 void HvxSelector::selectExtractSubvector(SDNode *N) {
2544 SDValue Inp = N->getOperand(0);
2545 MVT ResTy = N->getValueType(0).getSimpleVT();
2546 auto IdxN = cast<ConstantSDNode>(N->getOperand(1));
2547 unsigned Idx = IdxN->getZExtValue();
2548
2549 [[maybe_unused]] MVT InpTy = Inp.getValueType().getSimpleVT();
2550 [[maybe_unused]] unsigned ResLen = ResTy.getVectorNumElements();
2551 assert(InpTy.getVectorElementType() == ResTy.getVectorElementType());
2552 assert(2 * ResLen == InpTy.getVectorNumElements());
2553 assert(Idx == 0 || Idx == ResLen);
2554
2555 unsigned SubReg = Idx == 0 ? Hexagon::vsub_lo : Hexagon::vsub_hi;
2556 SDValue Ext = DAG.getTargetExtractSubreg(SubReg, SDLoc(N), ResTy, Inp);
2557
2558 ISel.ReplaceNode(N, Ext.getNode());
2559 }
2560
selectShuffle(SDNode * N)2561 void HvxSelector::selectShuffle(SDNode *N) {
2562 DEBUG_WITH_TYPE("isel", {
2563 dbgs() << "Starting " << __func__ << " on node:\n";
2564 N->dump(&DAG);
2565 });
2566 MVT ResTy = N->getValueType(0).getSimpleVT();
2567 // Assume that vector shuffles operate on vectors of bytes.
2568 assert(ResTy.isVector() && ResTy.getVectorElementType() == MVT::i8);
2569
2570 auto *SN = cast<ShuffleVectorSDNode>(N);
2571 std::vector<int> Mask(SN->getMask().begin(), SN->getMask().end());
2572 // This shouldn't really be necessary. Is it?
2573 for (int &Idx : Mask)
2574 if (Idx != -1 && Idx < 0)
2575 Idx = -1;
2576
2577 unsigned VecLen = Mask.size();
2578 bool HavePairs = (2*HwLen == VecLen);
2579 assert(ResTy.getSizeInBits() / 8 == VecLen);
2580
2581 // Vd = vector_shuffle Va, Vb, Mask
2582 //
2583
2584 bool UseLeft = false, UseRight = false;
2585 for (unsigned I = 0; I != VecLen; ++I) {
2586 if (Mask[I] == -1)
2587 continue;
2588 unsigned Idx = Mask[I];
2589 assert(Idx < 2*VecLen);
2590 if (Idx < VecLen)
2591 UseLeft = true;
2592 else
2593 UseRight = true;
2594 }
2595
2596 DEBUG_WITH_TYPE("isel", {
2597 dbgs() << "VecLen=" << VecLen << " HwLen=" << HwLen << " UseLeft="
2598 << UseLeft << " UseRight=" << UseRight << " HavePairs="
2599 << HavePairs << '\n';
2600 });
2601 // If the mask is all -1's, generate "undef".
2602 if (!UseLeft && !UseRight) {
2603 ISel.ReplaceNode(N, ISel.selectUndef(SDLoc(SN), ResTy).getNode());
2604 return;
2605 }
2606
2607 SDValue Vec0 = N->getOperand(0);
2608 SDValue Vec1 = N->getOperand(1);
2609 assert(Vec0.getValueType() == ResTy && Vec1.getValueType() == ResTy);
2610
2611 ResultStack Results(SN);
2612 OpRef Va = OpRef::undef(ResTy);
2613 OpRef Vb = OpRef::undef(ResTy);
2614
2615 if (!Vec0.isUndef()) {
2616 Results.push(TargetOpcode::COPY, ResTy, {Vec0});
2617 Va = OpRef::OpRef::res(Results.top());
2618 }
2619 if (!Vec1.isUndef()) {
2620 Results.push(TargetOpcode::COPY, ResTy, {Vec1});
2621 Vb = OpRef::res(Results.top());
2622 }
2623
2624 OpRef Res = !HavePairs ? shuffs2(ShuffleMask(Mask), Va, Vb, Results)
2625 : shuffp2(ShuffleMask(Mask), Va, Vb, Results);
2626
2627 bool Done = Res.isValid();
2628 if (Done) {
2629 // Make sure that Res is on the stack before materializing.
2630 Results.push(TargetOpcode::COPY, ResTy, {Res});
2631 materialize(Results);
2632 } else {
2633 Done = scalarizeShuffle(Mask, SDLoc(N), ResTy, Vec0, Vec1, N);
2634 }
2635
2636 if (!Done) {
2637 #ifndef NDEBUG
2638 dbgs() << "Unhandled shuffle:\n";
2639 SN->dumpr(&DAG);
2640 #endif
2641 llvm_unreachable("Failed to select vector shuffle");
2642 }
2643 }
2644
selectRor(SDNode * N)2645 void HvxSelector::selectRor(SDNode *N) {
2646 // If this is a rotation by less than 8, use V6_valignbi.
2647 MVT Ty = N->getValueType(0).getSimpleVT();
2648 const SDLoc &dl(N);
2649 SDValue VecV = N->getOperand(0);
2650 SDValue RotV = N->getOperand(1);
2651 SDNode *NewN = nullptr;
2652
2653 if (auto *CN = dyn_cast<ConstantSDNode>(RotV.getNode())) {
2654 unsigned S = CN->getZExtValue() % HST.getVectorLength();
2655 if (S == 0) {
2656 NewN = VecV.getNode();
2657 } else if (isUInt<3>(S)) {
2658 NewN = DAG.getMachineNode(Hexagon::V6_valignbi, dl, Ty,
2659 {VecV, VecV, getConst32(S, dl)});
2660 }
2661 }
2662
2663 if (!NewN)
2664 NewN = DAG.getMachineNode(Hexagon::V6_vror, dl, Ty, {VecV, RotV});
2665
2666 ISel.ReplaceNode(N, NewN);
2667 }
2668
selectVAlign(SDNode * N)2669 void HvxSelector::selectVAlign(SDNode *N) {
2670 SDValue Vv = N->getOperand(0);
2671 SDValue Vu = N->getOperand(1);
2672 SDValue Rt = N->getOperand(2);
2673 SDNode *NewN = DAG.getMachineNode(Hexagon::V6_valignb, SDLoc(N),
2674 N->getValueType(0), {Vv, Vu, Rt});
2675 ISel.ReplaceNode(N, NewN);
2676 DAG.RemoveDeadNode(N);
2677 }
2678
PreprocessHvxISelDAG()2679 void HexagonDAGToDAGISel::PreprocessHvxISelDAG() {
2680 auto getNodes = [this]() -> std::vector<SDNode *> {
2681 std::vector<SDNode *> T;
2682 T.reserve(CurDAG->allnodes_size());
2683 for (SDNode &N : CurDAG->allnodes())
2684 T.push_back(&N);
2685 return T;
2686 };
2687
2688 ppHvxShuffleOfShuffle(getNodes());
2689 }
2690
2691 template <> struct std::hash<SDValue> {
operator ()std::hash2692 std::size_t operator()(SDValue V) const {
2693 return std::hash<const void *>()(V.getNode()) +
2694 std::hash<unsigned>()(V.getResNo());
2695 };
2696 };
2697
ppHvxShuffleOfShuffle(std::vector<SDNode * > && Nodes)2698 void HexagonDAGToDAGISel::ppHvxShuffleOfShuffle(std::vector<SDNode *> &&Nodes) {
2699 // Motivating case:
2700 // t10: v64i32 = ...
2701 // t46: v128i8 = vector_shuffle<...> t44, t45
2702 // t48: v128i8 = vector_shuffle<...> t44, t45
2703 // t42: v128i8 = vector_shuffle<...> t46, t48
2704 // t12: v32i32 = extract_subvector t10, Constant:i32<0>
2705 // t44: v128i8 = bitcast t12
2706 // t15: v32i32 = extract_subvector t10, Constant:i32<32>
2707 // t45: v128i8 = bitcast t15
2708 SelectionDAG &DAG = *CurDAG;
2709 unsigned HwLen = HST->getVectorLength();
2710
2711 struct SubVectorInfo {
2712 SubVectorInfo(SDValue S, unsigned H) : Src(S), HalfIdx(H) {}
2713 SDValue Src;
2714 unsigned HalfIdx;
2715 };
2716
2717 using MapType = std::unordered_map<SDValue, unsigned>;
2718
2719 auto getMaskElt = [&](unsigned Idx, ShuffleVectorSDNode *Shuff0,
2720 ShuffleVectorSDNode *Shuff1,
2721 const MapType &OpMap) -> int {
2722 // Treat Shuff0 and Shuff1 as operands to another vector shuffle, and
2723 // Idx as a (non-undef) element of the top level shuffle's mask, that
2724 // is, index into concat(Shuff0, Shuff1).
2725 // Assuming that Shuff0 and Shuff1 both operate on subvectors of the
2726 // same source vector (as described by OpMap), return the index of
2727 // that source vector corresponding to Idx.
2728 ShuffleVectorSDNode *OpShuff = Idx < HwLen ? Shuff0 : Shuff1;
2729 if (Idx >= HwLen)
2730 Idx -= HwLen;
2731
2732 // Get the mask index that M points at in the corresponding operand.
2733 int MaybeN = OpShuff->getMaskElt(Idx);
2734 if (MaybeN < 0)
2735 return -1;
2736
2737 auto N = static_cast<unsigned>(MaybeN);
2738 unsigned SrcBase = N < HwLen ? OpMap.at(OpShuff->getOperand(0))
2739 : OpMap.at(OpShuff->getOperand(1));
2740 if (N >= HwLen)
2741 N -= HwLen;
2742
2743 return N + SrcBase;
2744 };
2745
2746 auto fold3 = [&](SDValue TopShuff, SDValue Inp, MapType &&OpMap) -> SDValue {
2747 // Fold all 3 shuffles into a single one.
2748 auto *This = cast<ShuffleVectorSDNode>(TopShuff);
2749 auto *S0 = cast<ShuffleVectorSDNode>(TopShuff.getOperand(0));
2750 auto *S1 = cast<ShuffleVectorSDNode>(TopShuff.getOperand(1));
2751 ArrayRef<int> TopMask = This->getMask();
2752 // This should be guaranteed by type checks in the caller, and the fact
2753 // that all shuffles should have been promoted to operate on MVT::i8.
2754 assert(TopMask.size() == S0->getMask().size() &&
2755 TopMask.size() == S1->getMask().size());
2756 assert(TopMask.size() == HwLen);
2757
2758 SmallVector<int, 256> FoldedMask(2 * HwLen);
2759 for (unsigned I = 0; I != HwLen; ++I) {
2760 int MaybeM = TopMask[I];
2761 if (MaybeM >= 0) {
2762 FoldedMask[I] =
2763 getMaskElt(static_cast<unsigned>(MaybeM), S0, S1, OpMap);
2764 } else {
2765 FoldedMask[I] = -1;
2766 }
2767 }
2768 // The second half of the result will be all-undef.
2769 std::fill(FoldedMask.begin() + HwLen, FoldedMask.end(), -1);
2770
2771 // Return
2772 // FoldedShuffle = (Shuffle Inp, undef, FoldedMask)
2773 // (LoHalf FoldedShuffle)
2774 const SDLoc &dl(TopShuff);
2775 MVT SingleTy = MVT::getVectorVT(MVT::i8, HwLen);
2776 MVT PairTy = MVT::getVectorVT(MVT::i8, 2 * HwLen);
2777 SDValue FoldedShuff =
2778 DAG.getVectorShuffle(PairTy, dl, DAG.getBitcast(PairTy, Inp),
2779 DAG.getUNDEF(PairTy), FoldedMask);
2780 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SingleTy, FoldedShuff,
2781 DAG.getConstant(0, dl, MVT::i32));
2782 };
2783
2784 auto getSourceInfo = [](SDValue V) -> std::optional<SubVectorInfo> {
2785 while (V.getOpcode() == ISD::BITCAST)
2786 V = V.getOperand(0);
2787 if (V.getOpcode() != ISD::EXTRACT_SUBVECTOR)
2788 return std::nullopt;
2789 return SubVectorInfo(V.getOperand(0),
2790 !cast<ConstantSDNode>(V.getOperand(1))->isZero());
2791 };
2792
2793 for (SDNode *N : Nodes) {
2794 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
2795 continue;
2796 EVT ResTy = N->getValueType(0);
2797 if (ResTy.getVectorElementType() != MVT::i8)
2798 continue;
2799 if (ResTy.getVectorNumElements() != HwLen)
2800 continue;
2801
2802 SDValue V0 = N->getOperand(0);
2803 SDValue V1 = N->getOperand(1);
2804 if (V0.getOpcode() != ISD::VECTOR_SHUFFLE)
2805 continue;
2806 if (V1.getOpcode() != ISD::VECTOR_SHUFFLE)
2807 continue;
2808 if (V0.getValueType() != ResTy || V1.getValueType() != ResTy)
2809 continue;
2810
2811 // Check if all operands of the two operand shuffles are extract_subvectors
2812 // from the same vector pair.
2813 auto V0A = getSourceInfo(V0.getOperand(0));
2814 if (!V0A.has_value())
2815 continue;
2816 auto V0B = getSourceInfo(V0.getOperand(1));
2817 if (!V0B.has_value() || V0B->Src != V0A->Src)
2818 continue;
2819 auto V1A = getSourceInfo(V1.getOperand(0));
2820 if (!V1A.has_value() || V1A->Src != V0A->Src)
2821 continue;
2822 auto V1B = getSourceInfo(V1.getOperand(1));
2823 if (!V1B.has_value() || V1B->Src != V0A->Src)
2824 continue;
2825
2826 // The source must be a pair. This should be guaranteed here,
2827 // but check just in case.
2828 assert(V0A->Src.getValueType().getSizeInBits() == 16 * HwLen);
2829
2830 MapType OpMap = {
2831 {V0.getOperand(0), V0A->HalfIdx * HwLen},
2832 {V0.getOperand(1), V0B->HalfIdx * HwLen},
2833 {V1.getOperand(0), V1A->HalfIdx * HwLen},
2834 {V1.getOperand(1), V1B->HalfIdx * HwLen},
2835 };
2836 SDValue NewS = fold3(SDValue(N, 0), V0A->Src, std::move(OpMap));
2837 ReplaceNode(N, NewS.getNode());
2838 }
2839 }
2840
SelectHvxExtractSubvector(SDNode * N)2841 void HexagonDAGToDAGISel::SelectHvxExtractSubvector(SDNode *N) {
2842 HvxSelector(*this, *CurDAG).selectExtractSubvector(N);
2843 }
2844
SelectHvxShuffle(SDNode * N)2845 void HexagonDAGToDAGISel::SelectHvxShuffle(SDNode *N) {
2846 HvxSelector(*this, *CurDAG).selectShuffle(N);
2847 }
2848
SelectHvxRor(SDNode * N)2849 void HexagonDAGToDAGISel::SelectHvxRor(SDNode *N) {
2850 HvxSelector(*this, *CurDAG).selectRor(N);
2851 }
2852
SelectHvxVAlign(SDNode * N)2853 void HexagonDAGToDAGISel::SelectHvxVAlign(SDNode *N) {
2854 HvxSelector(*this, *CurDAG).selectVAlign(N);
2855 }
2856
SelectV65GatherPred(SDNode * N)2857 void HexagonDAGToDAGISel::SelectV65GatherPred(SDNode *N) {
2858 const SDLoc &dl(N);
2859 SDValue Chain = N->getOperand(0);
2860 SDValue Address = N->getOperand(2);
2861 SDValue Predicate = N->getOperand(3);
2862 SDValue Base = N->getOperand(4);
2863 SDValue Modifier = N->getOperand(5);
2864 SDValue Offset = N->getOperand(6);
2865 SDValue ImmOperand = CurDAG->getTargetConstant(0, dl, MVT::i32);
2866
2867 unsigned Opcode;
2868 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2869 switch (IntNo) {
2870 default:
2871 llvm_unreachable("Unexpected HVX gather intrinsic.");
2872 case Intrinsic::hexagon_V6_vgathermhq:
2873 case Intrinsic::hexagon_V6_vgathermhq_128B:
2874 Opcode = Hexagon::V6_vgathermhq_pseudo;
2875 break;
2876 case Intrinsic::hexagon_V6_vgathermwq:
2877 case Intrinsic::hexagon_V6_vgathermwq_128B:
2878 Opcode = Hexagon::V6_vgathermwq_pseudo;
2879 break;
2880 case Intrinsic::hexagon_V6_vgathermhwq:
2881 case Intrinsic::hexagon_V6_vgathermhwq_128B:
2882 Opcode = Hexagon::V6_vgathermhwq_pseudo;
2883 break;
2884 }
2885
2886 SDVTList VTs = CurDAG->getVTList(MVT::Other);
2887 SDValue Ops[] = { Address, ImmOperand,
2888 Predicate, Base, Modifier, Offset, Chain };
2889 SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2890
2891 MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2892 CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2893
2894 ReplaceNode(N, Result);
2895 }
2896
SelectV65Gather(SDNode * N)2897 void HexagonDAGToDAGISel::SelectV65Gather(SDNode *N) {
2898 const SDLoc &dl(N);
2899 SDValue Chain = N->getOperand(0);
2900 SDValue Address = N->getOperand(2);
2901 SDValue Base = N->getOperand(3);
2902 SDValue Modifier = N->getOperand(4);
2903 SDValue Offset = N->getOperand(5);
2904 SDValue ImmOperand = CurDAG->getTargetConstant(0, dl, MVT::i32);
2905
2906 unsigned Opcode;
2907 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2908 switch (IntNo) {
2909 default:
2910 llvm_unreachable("Unexpected HVX gather intrinsic.");
2911 case Intrinsic::hexagon_V6_vgathermh:
2912 case Intrinsic::hexagon_V6_vgathermh_128B:
2913 Opcode = Hexagon::V6_vgathermh_pseudo;
2914 break;
2915 case Intrinsic::hexagon_V6_vgathermw:
2916 case Intrinsic::hexagon_V6_vgathermw_128B:
2917 Opcode = Hexagon::V6_vgathermw_pseudo;
2918 break;
2919 case Intrinsic::hexagon_V6_vgathermhw:
2920 case Intrinsic::hexagon_V6_vgathermhw_128B:
2921 Opcode = Hexagon::V6_vgathermhw_pseudo;
2922 break;
2923 }
2924
2925 SDVTList VTs = CurDAG->getVTList(MVT::Other);
2926 SDValue Ops[] = { Address, ImmOperand, Base, Modifier, Offset, Chain };
2927 SDNode *Result = CurDAG->getMachineNode(Opcode, dl, VTs, Ops);
2928
2929 MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2930 CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
2931
2932 ReplaceNode(N, Result);
2933 }
2934
SelectHVXDualOutput(SDNode * N)2935 void HexagonDAGToDAGISel::SelectHVXDualOutput(SDNode *N) {
2936 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2937 SDNode *Result;
2938 switch (IID) {
2939 case Intrinsic::hexagon_V6_vaddcarry: {
2940 std::array<SDValue, 3> Ops = {
2941 {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2942 SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v64i1);
2943 Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2944 break;
2945 }
2946 case Intrinsic::hexagon_V6_vaddcarry_128B: {
2947 std::array<SDValue, 3> Ops = {
2948 {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2949 SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v128i1);
2950 Result = CurDAG->getMachineNode(Hexagon::V6_vaddcarry, SDLoc(N), VTs, Ops);
2951 break;
2952 }
2953 case Intrinsic::hexagon_V6_vsubcarry: {
2954 std::array<SDValue, 3> Ops = {
2955 {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2956 SDVTList VTs = CurDAG->getVTList(MVT::v16i32, MVT::v64i1);
2957 Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2958 break;
2959 }
2960 case Intrinsic::hexagon_V6_vsubcarry_128B: {
2961 std::array<SDValue, 3> Ops = {
2962 {N->getOperand(1), N->getOperand(2), N->getOperand(3)}};
2963 SDVTList VTs = CurDAG->getVTList(MVT::v32i32, MVT::v128i1);
2964 Result = CurDAG->getMachineNode(Hexagon::V6_vsubcarry, SDLoc(N), VTs, Ops);
2965 break;
2966 }
2967 default:
2968 llvm_unreachable("Unexpected HVX dual output intrinsic.");
2969 }
2970 ReplaceUses(N, Result);
2971 ReplaceUses(SDValue(N, 0), SDValue(Result, 0));
2972 ReplaceUses(SDValue(N, 1), SDValue(Result, 1));
2973 CurDAG->RemoveDeadNode(N);
2974 }
2975