1 // Copyright 2014 the V8 project authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4 
5 #include "src/compiler/backend/gap-resolver.h"
6 
7 #include <algorithm>
8 #include <set>
9 
10 #include "src/base/enum-set.h"
11 #include "src/codegen/register-configuration.h"
12 
13 namespace v8 {
14 namespace internal {
15 namespace compiler {
16 
17 namespace {
18 
19 // Splits a FP move between two location operands into the equivalent series of
20 // moves between smaller sub-operands, e.g. a double move to two single moves.
21 // This helps reduce the number of cycles that would normally occur under FP
22 // aliasing, and makes swaps much easier to implement.
Split(MoveOperands * move,MachineRepresentation smaller_rep,ParallelMove * moves)23 MoveOperands* Split(MoveOperands* move, MachineRepresentation smaller_rep,
24                     ParallelMove* moves) {
25   DCHECK(!kSimpleFPAliasing);
26   // Splitting is only possible when the slot size is the same as float size.
27   DCHECK_EQ(kSystemPointerSize, kFloatSize);
28   const LocationOperand& src_loc = LocationOperand::cast(move->source());
29   const LocationOperand& dst_loc = LocationOperand::cast(move->destination());
30   MachineRepresentation dst_rep = dst_loc.representation();
31   DCHECK_NE(smaller_rep, dst_rep);
32   auto src_kind = src_loc.location_kind();
33   auto dst_kind = dst_loc.location_kind();
34 
35   int aliases =
36       1 << (ElementSizeLog2Of(dst_rep) - ElementSizeLog2Of(smaller_rep));
37   int base = -1;
38   USE(base);
39   DCHECK_EQ(aliases, RegisterConfiguration::Default()->GetAliases(
40                          dst_rep, 0, smaller_rep, &base));
41 
42   int src_index = -1;
43   int slot_size = (1 << ElementSizeLog2Of(smaller_rep)) / kSystemPointerSize;
44   int src_step = 1;
45   if (src_kind == LocationOperand::REGISTER) {
46     src_index = src_loc.register_code() * aliases;
47   } else {
48     src_index = src_loc.index();
49     // For operands that occupy multiple slots, the index refers to the last
50     // slot. On little-endian architectures, we start at the high slot and use a
51     // negative step so that register-to-slot moves are in the correct order.
52     src_step = -slot_size;
53   }
54   int dst_index = -1;
55   int dst_step = 1;
56   if (dst_kind == LocationOperand::REGISTER) {
57     dst_index = dst_loc.register_code() * aliases;
58   } else {
59     dst_index = dst_loc.index();
60     dst_step = -slot_size;
61   }
62 
63   // Reuse 'move' for the first fragment. It is not pending.
64   move->set_source(AllocatedOperand(src_kind, smaller_rep, src_index));
65   move->set_destination(AllocatedOperand(dst_kind, smaller_rep, dst_index));
66   // Add the remaining fragment moves.
67   for (int i = 1; i < aliases; ++i) {
68     src_index += src_step;
69     dst_index += dst_step;
70     moves->AddMove(AllocatedOperand(src_kind, smaller_rep, src_index),
71                    AllocatedOperand(dst_kind, smaller_rep, dst_index));
72   }
73   // Return the first fragment.
74   return move;
75 }
76 
77 enum MoveOperandKind : uint8_t { kConstant, kGpReg, kFpReg, kStack };
78 
GetKind(const InstructionOperand & move)79 MoveOperandKind GetKind(const InstructionOperand& move) {
80   if (move.IsConstant()) return kConstant;
81   LocationOperand loc_op = LocationOperand::cast(move);
82   if (loc_op.location_kind() != LocationOperand::REGISTER) return kStack;
83   return IsFloatingPoint(loc_op.representation()) ? kFpReg : kGpReg;
84 }
85 
86 }  // namespace
87 
Resolve(ParallelMove * moves)88 void GapResolver::Resolve(ParallelMove* moves) {
89   base::EnumSet<MoveOperandKind, uint8_t> source_kinds;
90   base::EnumSet<MoveOperandKind, uint8_t> destination_kinds;
91 
92   // Remove redundant moves, collect source kinds and destination kinds to
93   // detect simple non-overlapping moves, and collect FP move representations if
94   // aliasing is non-simple.
95   int fp_reps = 0;
96   size_t nmoves = moves->size();
97   for (size_t i = 0; i < nmoves;) {
98     MoveOperands* move = (*moves)[i];
99     if (move->IsRedundant()) {
100       nmoves--;
101       if (i < nmoves) (*moves)[i] = (*moves)[nmoves];
102       continue;
103     }
104     i++;
105     source_kinds.Add(GetKind(move->source()));
106     destination_kinds.Add(GetKind(move->destination()));
107     if (!kSimpleFPAliasing && move->destination().IsFPRegister()) {
108       fp_reps |= RepresentationBit(
109           LocationOperand::cast(move->destination()).representation());
110     }
111   }
112   if (nmoves != moves->size()) moves->resize(nmoves);
113 
114   if ((source_kinds & destination_kinds).empty() || moves->size() < 2) {
115     // Fast path for non-conflicting parallel moves.
116     for (MoveOperands* move : *moves) {
117       assembler_->AssembleMove(&move->source(), &move->destination());
118     }
119     return;
120   }
121 
122   if (!kSimpleFPAliasing) {
123     if (fp_reps && !base::bits::IsPowerOfTwo(fp_reps)) {
124       // Start with the smallest FP moves, so we never encounter smaller moves
125       // in the middle of a cycle of larger moves.
126       if ((fp_reps & RepresentationBit(MachineRepresentation::kFloat32)) != 0) {
127         split_rep_ = MachineRepresentation::kFloat32;
128         for (size_t i = 0; i < moves->size(); ++i) {
129           auto move = (*moves)[i];
130           if (!move->IsEliminated() && move->destination().IsFloatRegister())
131             PerformMove(moves, move);
132         }
133       }
134       if ((fp_reps & RepresentationBit(MachineRepresentation::kFloat64)) != 0) {
135         split_rep_ = MachineRepresentation::kFloat64;
136         for (size_t i = 0; i < moves->size(); ++i) {
137           auto move = (*moves)[i];
138           if (!move->IsEliminated() && move->destination().IsDoubleRegister())
139             PerformMove(moves, move);
140         }
141       }
142     }
143     split_rep_ = MachineRepresentation::kSimd128;
144   }
145 
146   for (size_t i = 0; i < moves->size(); ++i) {
147     auto move = (*moves)[i];
148     if (!move->IsEliminated()) PerformMove(moves, move);
149   }
150 }
151 
PerformMove(ParallelMove * moves,MoveOperands * move)152 void GapResolver::PerformMove(ParallelMove* moves, MoveOperands* move) {
153   // Each call to this function performs a move and deletes it from the move
154   // graph.  We first recursively perform any move blocking this one.  We mark a
155   // move as "pending" on entry to PerformMove in order to detect cycles in the
156   // move graph.  We use operand swaps to resolve cycles, which means that a
157   // call to PerformMove could change any source operand in the move graph.
158   DCHECK(!move->IsPending());
159   DCHECK(!move->IsRedundant());
160 
161   // Clear this move's destination to indicate a pending move.  The actual
162   // destination is saved on the side.
163   InstructionOperand source = move->source();
164   DCHECK(!source.IsInvalid());  // Or else it will look eliminated.
165   InstructionOperand destination = move->destination();
166   move->SetPending();
167 
168   // We may need to split moves between FP locations differently.
169   const bool is_fp_loc_move =
170       !kSimpleFPAliasing && destination.IsFPLocationOperand();
171 
172   // Perform a depth-first traversal of the move graph to resolve dependencies.
173   // Any unperformed, unpending move with a source the same as this one's
174   // destination blocks this one so recursively perform all such moves.
175   for (size_t i = 0; i < moves->size(); ++i) {
176     auto other = (*moves)[i];
177     if (other->IsEliminated()) continue;
178     if (other->IsPending()) continue;
179     if (other->source().InterferesWith(destination)) {
180       if (is_fp_loc_move &&
181           LocationOperand::cast(other->source()).representation() >
182               split_rep_) {
183         // 'other' must also be an FP location move. Break it into fragments
184         // of the same size as 'move'. 'other' is set to one of the fragments,
185         // and the rest are appended to 'moves'.
186         other = Split(other, split_rep_, moves);
187         // 'other' may not block destination now.
188         if (!other->source().InterferesWith(destination)) continue;
189       }
190       // Though PerformMove can change any source operand in the move graph,
191       // this call cannot create a blocking move via a swap (this loop does not
192       // miss any).  Assume there is a non-blocking move with source A and this
193       // move is blocked on source B and there is a swap of A and B.  Then A and
194       // B must be involved in the same cycle (or they would not be swapped).
195       // Since this move's destination is B and there is only a single incoming
196       // edge to an operand, this move must also be involved in the same cycle.
197       // In that case, the blocking move will be created but will be "pending"
198       // when we return from PerformMove.
199       PerformMove(moves, other);
200     }
201   }
202 
203   // This move's source may have changed due to swaps to resolve cycles and so
204   // it may now be the last move in the cycle.  If so remove it.
205   source = move->source();
206   if (source.EqualsCanonicalized(destination)) {
207     move->Eliminate();
208     return;
209   }
210 
211   // We are about to resolve this move and don't need it marked as pending, so
212   // restore its destination.
213   move->set_destination(destination);
214 
215   // The move may be blocked on a (at most one) pending move, in which case we
216   // have a cycle.  Search for such a blocking move and perform a swap to
217   // resolve it.
218   auto blocker =
219       std::find_if(moves->begin(), moves->end(), [&](MoveOperands* move) {
220         return !move->IsEliminated() &&
221                move->source().InterferesWith(destination);
222       });
223   if (blocker == moves->end()) {
224     // The easy case: This move is not blocked.
225     assembler_->AssembleMove(&source, &destination);
226     move->Eliminate();
227     return;
228   }
229 
230   // Ensure source is a register or both are stack slots, to limit swap cases.
231   if (source.IsStackSlot() || source.IsFPStackSlot()) {
232     std::swap(source, destination);
233   }
234   assembler_->AssembleSwap(&source, &destination);
235   move->Eliminate();
236 
237   // Update outstanding moves whose source may now have been moved.
238   if (is_fp_loc_move) {
239     // We may have to split larger moves.
240     for (size_t i = 0; i < moves->size(); ++i) {
241       auto other = (*moves)[i];
242       if (other->IsEliminated()) continue;
243       if (source.InterferesWith(other->source())) {
244         if (LocationOperand::cast(other->source()).representation() >
245             split_rep_) {
246           other = Split(other, split_rep_, moves);
247           if (!source.InterferesWith(other->source())) continue;
248         }
249         other->set_source(destination);
250       } else if (destination.InterferesWith(other->source())) {
251         if (LocationOperand::cast(other->source()).representation() >
252             split_rep_) {
253           other = Split(other, split_rep_, moves);
254           if (!destination.InterferesWith(other->source())) continue;
255         }
256         other->set_source(source);
257       }
258     }
259   } else {
260     for (auto other : *moves) {
261       if (other->IsEliminated()) continue;
262       if (source.EqualsCanonicalized(other->source())) {
263         other->set_source(destination);
264       } else if (destination.EqualsCanonicalized(other->source())) {
265         other->set_source(source);
266       }
267     }
268   }
269 }
270 }  // namespace compiler
271 }  // namespace internal
272 }  // namespace v8
273