js-inlining-heuristic.cc 32.4 KB
Newer Older
1 2 3 4 5 6
// Copyright 2015 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#include "src/compiler/js-inlining-heuristic.h"

7
#include "src/codegen/optimized-compilation-info.h"
8
#include "src/compiler/common-operator.h"
9
#include "src/compiler/compiler-source-position-table.h"
10
#include "src/compiler/js-heap-broker.h"
11
#include "src/compiler/node-matchers.h"
12
#include "src/compiler/simplified-operator.h"
13
#include "src/objects/objects-inl.h"
14 15 16 17 18

namespace v8 {
namespace internal {
namespace compiler {

19 20 21
#define TRACE(...)                                                             \
  do {                                                                         \
    if (FLAG_trace_turbo_inlining) StdoutStream{} << __VA_ARGS__ << std::endl; \
22
  } while (false)
23

24
namespace {
25 26
bool IsSmall(int const size) {
  return size <= FLAG_max_inlined_bytecode_size_small;
27
}
28 29 30 31

bool CanConsiderForInlining(JSHeapBroker* broker,
                            SharedFunctionInfoRef const& shared,
                            FeedbackVectorRef const& feedback_vector) {
32 33 34 35 36 37 38
  SharedFunctionInfo::Inlineability inlineability = shared.GetInlineability();
  if (inlineability != SharedFunctionInfo::kIsInlineable) {
    TRACE("Cannot consider "
          << shared << " for inlining (reason: " << inlineability << ")");
    return false;
  }

39
  DCHECK(shared.HasBytecodeArray());
40
  if (!broker->IsSerializedForCompilation(shared, feedback_vector)) {
41 42
    TRACE_BROKER_MISSING(
        broker, "data for " << shared << " (not serialized for compilation)");
43 44
    TRACE("Cannot consider " << shared << " for inlining with "
                             << feedback_vector << " (missing data)");
45 46
    return false;
  }
47 48

  TRACE("Considering " << shared << " for inlining with " << feedback_vector);
49 50 51 52 53
  return true;
}

bool CanConsiderForInlining(JSHeapBroker* broker,
                            JSFunctionRef const& function) {
54 55 56 57 58 59
  if (!function.has_feedback_vector()) {
    TRACE("Cannot consider " << function
                             << " for inlining (no feedback vector)");
    return false;
  }

60 61 62
  if (!function.serialized()) {
    TRACE_BROKER_MISSING(
        broker, "data for " << function << " (cannot consider for inlining)");
63
    TRACE("Cannot consider " << function << " for inlining (missing data)");
64 65 66 67 68 69
    return false;
  }
  return CanConsiderForInlining(broker, function.shared(),
                                function.feedback_vector());
}

70 71 72 73
}  // namespace

JSInliningHeuristic::Candidate JSInliningHeuristic::CollectFunctions(
    Node* node, int functions_size) {
74
  DCHECK_NE(0, functions_size);
75 76 77 78 79
  Node* callee = node->InputAt(0);
  Candidate out;
  out.node = node;

  HeapObjectMatcher m(callee);
80
  if (m.HasResolvedValue() && m.Ref(broker()).IsJSFunction()) {
81 82
    out.functions[0] = m.Ref(broker()).AsJSFunction();
    JSFunctionRef function = out.functions[0].value();
83
    if (CanConsiderForInlining(broker(), function)) {
84
      out.bytecode[0] = function.shared().GetBytecodeArray();
85 86
      out.num_functions = 1;
      return out;
87
    }
88
  }
89 90
  if (m.IsPhi()) {
    int const value_input_count = m.node()->op()->ValueInputCount();
91 92 93 94
    if (value_input_count > functions_size) {
      out.num_functions = 0;
      return out;
    }
95
    for (int n = 0; n < value_input_count; ++n) {
96
      HeapObjectMatcher m(callee->InputAt(n));
97
      if (!m.HasResolvedValue() || !m.Ref(broker()).IsJSFunction()) {
98 99 100 101 102 103
        out.num_functions = 0;
        return out;
      }

      out.functions[n] = m.Ref(broker()).AsJSFunction();
      JSFunctionRef function = out.functions[n].value();
104
      if (CanConsiderForInlining(broker(), function)) {
105
        out.bytecode[n] = function.shared().GetBytecodeArray();
106
      }
107
    }
108 109
    out.num_functions = value_input_count;
    return out;
110
  }
111 112 113 114 115 116 117 118 119 120 121 122 123 124
  if (m.IsCheckClosure()) {
    DCHECK(!out.functions[0].has_value());
    FeedbackCellRef feedback_cell(broker(), FeedbackCellOf(m.op()));
    SharedFunctionInfoRef shared_info =
        feedback_cell.shared_function_info().value();
    out.shared_info = shared_info;
    if (feedback_cell.value().IsFeedbackVector() &&
        CanConsiderForInlining(broker(), shared_info,
                               feedback_cell.value().AsFeedbackVector())) {
      out.bytecode[0] = shared_info.GetBytecodeArray();
    }
    out.num_functions = 1;
    return out;
  }
125
  if (m.IsJSCreateClosure()) {
126
    DCHECK(!out.functions[0].has_value());
127 128 129
    JSCreateClosureNode n(callee);
    CreateClosureParameters const& p = n.Parameters();
    FeedbackCellRef feedback_cell = n.GetFeedbackCellRefChecked(broker());
130 131 132 133 134
    SharedFunctionInfoRef shared_info(broker(), p.shared_info());
    out.shared_info = shared_info;
    if (feedback_cell.value().IsFeedbackVector() &&
        CanConsiderForInlining(broker(), shared_info,
                               feedback_cell.value().AsFeedbackVector())) {
135
      out.bytecode[0] = shared_info.GetBytecodeArray();
136
    }
137 138
    out.num_functions = 1;
    return out;
139
  }
140 141
  out.num_functions = 0;
  return out;
142
}
143

144 145 146
Reduction JSInliningHeuristic::Reduce(Node* node) {
  if (!IrOpcode::IsInlineeOpcode(node->opcode())) return NoChange();

147
  if (total_inlined_bytecode_size_ >= FLAG_max_inlined_bytecode_size_absolute) {
148 149 150
    return NoChange();
  }

151 152 153 154 155
  // Check if we already saw that {node} before, and if so, just skip it.
  if (seen_.find(node->id()) != seen_.end()) return NoChange();
  seen_.insert(node->id());

  // Check if the {node} is an appropriate candidate for inlining.
156
  Candidate candidate = CollectFunctions(node, kMaxCallPolymorphism);
157 158 159
  if (candidate.num_functions == 0) {
    return NoChange();
  } else if (candidate.num_functions > 1 && !FLAG_polymorphic_inlining) {
160 161 162
    TRACE("Not considering call site #"
          << node->id() << ":" << node->op()->mnemonic()
          << ", because polymorphic inlining is disabled");
163 164 165
    return NoChange();
  }

166
  bool can_inline_candidate = false, candidate_is_small = true;
167
  candidate.total_size = 0;
168
  Node* frame_state = NodeProperties::GetFrameStateInput(node);
169
  FrameStateInfo const& frame_info = FrameStateInfoOf(frame_state->op());
170
  Handle<SharedFunctionInfo> frame_shared_info;
171
  for (int i = 0; i < candidate.num_functions; ++i) {
172 173 174 175 176 177 178 179
    if (!candidate.bytecode[i].has_value()) {
      candidate.can_inline_function[i] = false;
      continue;
    }

    SharedFunctionInfoRef shared = candidate.functions[i].has_value()
                                       ? candidate.functions[i].value().shared()
                                       : candidate.shared_info.value();
180 181
    candidate.can_inline_function[i] = candidate.bytecode[i].has_value();
    CHECK_IMPLIES(candidate.can_inline_function[i], shared.IsInlineable());
182 183 184 185 186 187 188 189 190 191
    // Do not allow direct recursion i.e. f() -> f(). We still allow indirect
    // recurion like f() -> g() -> f(). The indirect recursion is helpful in
    // cases where f() is a small dispatch function that calls the appropriate
    // function. In the case of direct recursion, we only have some static
    // information for the first level of inlining and it may not be that useful
    // to just inline one level in recursive calls. In some cases like tail
    // recursion we may benefit from recursive inlining, if we have additional
    // analysis that converts them to iterative implementations. Though it is
    // not obvious if such an anlysis is needed.
    if (frame_info.shared_info().ToHandle(&frame_shared_info) &&
192
        frame_shared_info.equals(shared.object())) {
193 194 195
      TRACE("Not considering call site #" << node->id() << ":"
                                          << node->op()->mnemonic()
                                          << ", because of recursive inlining");
196 197
      candidate.can_inline_function[i] = false;
    }
198 199
    if (candidate.can_inline_function[i]) {
      can_inline_candidate = true;
200
      BytecodeArrayRef bytecode = candidate.bytecode[i].value();
201
      candidate.total_size += bytecode.length();
202 203 204
      unsigned inlined_bytecode_size = 0;
      if (candidate.functions[i].has_value()) {
        JSFunctionRef function = candidate.functions[i].value();
205
        if (function.HasAttachedOptimizedCode()) {
206 207 208 209 210 211
          inlined_bytecode_size = function.code().inlined_bytecode_size();
          candidate.total_size += inlined_bytecode_size;
        }
      }
      candidate_is_small = candidate_is_small &&
                           IsSmall(bytecode.length() + inlined_bytecode_size);
212 213
    }
  }
214
  if (!can_inline_candidate) return NoChange();
215

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
  // Gather feedback on how often this call site has been hit before.
  if (node->opcode() == IrOpcode::kJSCall) {
    CallParameters const p = CallParametersOf(node->op());
    candidate.frequency = p.frequency();
  } else {
    ConstructParameters const p = ConstructParametersOf(node->op());
    candidate.frequency = p.frequency();
  }

  // Don't consider a {candidate} whose frequency is below the
  // threshold, i.e. a call site that is only hit once every N
  // invocations of the caller.
  if (candidate.frequency.IsKnown() &&
      candidate.frequency.value() < FLAG_min_inlining_frequency) {
    return NoChange();
  }

233
  // Forcibly inline small functions here. In the case of polymorphic inlining
234 235
  // candidate_is_small is set only when all functions are small.
  if (candidate_is_small) {
236 237
    TRACE("Inlining small function(s) at call site #"
          << node->id() << ":" << node->op()->mnemonic());
238
    return InlineCandidate(candidate, true);
239 240
  }

241
  // In the general case we remember the candidate for later.
242
  candidates_.insert(candidate);
243 244 245
  return NoChange();
}

246
void JSInliningHeuristic::Finalize() {
247
  if (candidates_.empty()) return;  // Nothing to do without candidates.
248 249
  if (FLAG_trace_turbo_inlining) PrintCandidates();

250 251 252
  // We inline at most one candidate in every iteration of the fixpoint.
  // This is to ensure that we don't consume the full inlining budget
  // on things that aren't called very often.
253 254 255 256 257
  // TODO(bmeurer): Use std::priority_queue instead of std::set here.
  while (!candidates_.empty()) {
    auto i = candidates_.begin();
    Candidate candidate = *i;
    candidates_.erase(i);
258

259 260 261
    // Ignore this candidate if it's no longer valid.
    if (!IrOpcode::IsInlineeOpcode(candidate.node->opcode())) continue;
    if (candidate.node->IsDead()) continue;
262

263 264 265 266
    // Make sure we have some extra budget left, so that any small functions
    // exposed by this function would be given a chance to inline.
    double size_of_candidate =
        candidate.total_size * FLAG_reserve_inline_budget_scale_factor;
267 268
    int total_size =
        total_inlined_bytecode_size_ + static_cast<int>(size_of_candidate);
269 270 271 272 273
    if (total_size > FLAG_max_inlined_bytecode_size_cumulative) {
      // Try if any smaller functions are available to inline.
      continue;
    }

274 275
    Reduction const reduction = InlineCandidate(candidate, false);
    if (reduction.Changed()) return;
276
  }
277 278
}

279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471
namespace {

struct NodeAndIndex {
  Node* node;
  int index;
};

bool CollectStateValuesOwnedUses(Node* node, Node* state_values,
                                 NodeAndIndex* uses_buffer, size_t* use_count,
                                 size_t max_uses) {
  // Only accumulate states that are not shared with other users.
  if (state_values->UseCount() > 1) return true;
  for (int i = 0; i < state_values->InputCount(); i++) {
    Node* input = state_values->InputAt(i);
    if (input->opcode() == IrOpcode::kStateValues) {
      if (!CollectStateValuesOwnedUses(node, input, uses_buffer, use_count,
                                       max_uses)) {
        return false;
      }
    } else if (input == node) {
      if (*use_count >= max_uses) return false;
      uses_buffer[*use_count] = {state_values, i};
      (*use_count)++;
    }
  }
  return true;
}

}  // namespace

Node* JSInliningHeuristic::DuplicateStateValuesAndRename(Node* state_values,
                                                         Node* from, Node* to,
                                                         StateCloneMode mode) {
  // Only rename in states that are not shared with other users. This needs to
  // be in sync with the condition in {CollectStateValuesOwnedUses}.
  if (state_values->UseCount() > 1) return state_values;
  Node* copy = mode == kChangeInPlace ? state_values : nullptr;
  for (int i = 0; i < state_values->InputCount(); i++) {
    Node* input = state_values->InputAt(i);
    Node* processed;
    if (input->opcode() == IrOpcode::kStateValues) {
      processed = DuplicateStateValuesAndRename(input, from, to, mode);
    } else if (input == from) {
      processed = to;
    } else {
      processed = input;
    }
    if (processed != input) {
      if (!copy) {
        copy = graph()->CloneNode(state_values);
      }
      copy->ReplaceInput(i, processed);
    }
  }
  return copy ? copy : state_values;
}

namespace {

bool CollectFrameStateUniqueUses(Node* node, Node* frame_state,
                                 NodeAndIndex* uses_buffer, size_t* use_count,
                                 size_t max_uses) {
  // Only accumulate states that are not shared with other users.
  if (frame_state->UseCount() > 1) return true;
  if (frame_state->InputAt(kFrameStateStackInput) == node) {
    if (*use_count >= max_uses) return false;
    uses_buffer[*use_count] = {frame_state, kFrameStateStackInput};
    (*use_count)++;
  }
  if (!CollectStateValuesOwnedUses(node,
                                   frame_state->InputAt(kFrameStateLocalsInput),
                                   uses_buffer, use_count, max_uses)) {
    return false;
  }
  return true;
}

}  // namespace

Node* JSInliningHeuristic::DuplicateFrameStateAndRename(Node* frame_state,
                                                        Node* from, Node* to,
                                                        StateCloneMode mode) {
  // Only rename in states that are not shared with other users. This needs to
  // be in sync with the condition in {DuplicateFrameStateAndRename}.
  if (frame_state->UseCount() > 1) return frame_state;
  Node* copy = mode == kChangeInPlace ? frame_state : nullptr;
  if (frame_state->InputAt(kFrameStateStackInput) == from) {
    if (!copy) {
      copy = graph()->CloneNode(frame_state);
    }
    copy->ReplaceInput(kFrameStateStackInput, to);
  }
  Node* locals = frame_state->InputAt(kFrameStateLocalsInput);
  Node* new_locals = DuplicateStateValuesAndRename(locals, from, to, mode);
  if (new_locals != locals) {
    if (!copy) {
      copy = graph()->CloneNode(frame_state);
    }
    copy->ReplaceInput(kFrameStateLocalsInput, new_locals);
  }
  return copy ? copy : frame_state;
}

bool JSInliningHeuristic::TryReuseDispatch(Node* node, Node* callee,
                                           Node** if_successes, Node** calls,
                                           Node** inputs, int input_count) {
  // We will try to reuse the control flow branch created for computing
  // the {callee} target of the call. We only reuse the branch if there
  // is no side-effect between the call and the branch, and if the callee is
  // only used as the target (and possibly also in the related frame states).

  // We are trying to match the following pattern:
  //
  //         C1     C2
  //          .     .
  //          |     |
  //         Merge(merge)  <-----------------+
  //           ^    ^                        |
  //  V1  V2   |    |         E1  E2         |
  //   .  .    |    +----+     .  .          |
  //   |  |    |         |     |  |          |
  //  Phi(callee)      EffectPhi(effect_phi) |
  //     ^                    ^              |
  //     |                    |              |
  //     +----+               |              |
  //     |    |               |              |
  //     |   StateValues      |              |
  //     |       ^            |              |
  //     +----+  |            |              |
  //     |    |  |            |              |
  //     |    FrameState      |              |
  //     |           ^        |              |
  //     |           |        |          +---+
  //     |           |        |          |   |
  //     +----+     Checkpoint(checkpoint)   |
  //     |    |           ^                  |
  //     |    StateValues |    +-------------+
  //     |        |       |    |
  //     +-----+  |       |    |
  //     |     |  |       |    |
  //     |     FrameState |    |
  //     |             ^  |    |
  //     +-----------+ |  |    |
  //                  Call(node)
  //                   |
  //                   |
  //                   .
  //
  // The {callee} here is a phi that merges the possible call targets, {node}
  // is the actual call that we will try to duplicate and connect to the
  // control that comes into {merge}. There can be a {checkpoint} between
  // the call and the calle phi.
  //
  // The idea is to get rid of the merge, effect phi and phi, then duplicate
  // the call (with all the frame states and such), and connect the duplicated
  // calls and states directly to the inputs of the ex-phi, ex-effect-phi and
  // ex-merge. The tricky part is to make sure that there is no interference
  // from the outside. In particular, there should not be any unaccounted uses
  // of the  phi, effect-phi and merge because we will remove them from
  // the graph.
  //
  //     V1              E1   C1  V2   E2               C2
  //     .                .    .  .    .                .
  //     |                |    |  |    |                |
  //     +----+           |    |  +----+                |
  //     |    |           |    |  |    |                |
  //     |   StateValues  |    |  |   StateValues       |
  //     |       ^        |    |  |       ^             |
  //     +----+  |        |    |  +----+  |             |
  //     |    |  |        |    |  |    |  |             |
  //     |    FrameState  |    |  |    FrameState       |
  //     |           ^    |    |  |           ^         |
  //     |           |    |    |  |           |         |
  //     |           |    |    |  |           |         |
  //     +----+     Checkpoint |  +----+     Checkpoint |
  //     |    |           ^    |  |    |           ^    |
  //     |    StateValues |    |  |    StateValues |    |
  //     |        |       |    |  |        |       |    |
  //     +-----+  |       |    |  +-----+  |       |    |
  //     |     |  |       |    |  |     |  |       |    |
  //     |     FrameState |    |  |     FrameState |    |
  //     |              ^ |    |  |              ^ |    |
  //     +-------------+| |    |  +-------------+| |    |
  //                   Call----+                Call----+
  //                     |                       |
  //                     +-------+  +------------+
  //                             |  |
  //                             Merge
  //                             EffectPhi
  //                             Phi
  //                              |
  //                             ...

472 473 474 475 476
  // Bailout if the call is not polymorphic anymore (other reducers might
  // have replaced the callee phi with a constant).
  if (callee->opcode() != IrOpcode::kPhi) return false;
  int const num_calls = callee->op()->ValueInputCount();

477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604
  // If there is a control node between the callee computation
  // and the call, bail out.
  Node* merge = NodeProperties::GetControlInput(callee);
  if (NodeProperties::GetControlInput(node) != merge) return false;

  // If there is a non-checkpoint effect node between the callee computation
  // and the call, bail out. We will drop any checkpoint between the call and
  // the callee phi because the callee computation should have its own
  // checkpoint that the call can fall back to.
  Node* checkpoint = nullptr;
  Node* effect = NodeProperties::GetEffectInput(node);
  if (effect->opcode() == IrOpcode::kCheckpoint) {
    checkpoint = effect;
    if (NodeProperties::GetControlInput(checkpoint) != merge) return false;
    effect = NodeProperties::GetEffectInput(effect);
  }
  if (effect->opcode() != IrOpcode::kEffectPhi) return false;
  if (NodeProperties::GetControlInput(effect) != merge) return false;
  Node* effect_phi = effect;

  // The effect phi, the callee, the call and the checkpoint must be the only
  // users of the merge.
  for (Node* merge_use : merge->uses()) {
    if (merge_use != effect_phi && merge_use != callee && merge_use != node &&
        merge_use != checkpoint) {
      return false;
    }
  }

  // The effect phi must be only used by the checkpoint or the call.
  for (Node* effect_phi_use : effect_phi->uses()) {
    if (effect_phi_use != node && effect_phi_use != checkpoint) return false;
  }

  // We must replace the callee phi with the appropriate constant in
  // the entire subgraph reachable by inputs from the call (terminating
  // at phis and merges). Since we do not want to walk (and later duplicate)
  // the subgraph here, we limit the possible uses to this set:
  //
  // 1. In the call (as a target).
  // 2. The checkpoint between the call and the callee computation merge.
  // 3. The lazy deoptimization frame state.
  //
  // This corresponds to the most common pattern, where the function is
  // called with only local variables or constants as arguments.
  //
  // To check the uses, we first collect all the occurrences of callee in 1, 2
  // and 3, and then we check that all uses of callee are in the collected
  // occurrences. If there is an unaccounted use, we do not try to rewire
  // the control flow.
  //
  // Note: With CFG, this would be much easier and more robust - we would just
  // duplicate all the nodes between the merge and the call, replacing all
  // occurrences of the {callee} phi with the appropriate constant.

  // First compute the set of uses that are only reachable from 2 and 3.
  const size_t kMaxUses = 8;
  NodeAndIndex replaceable_uses[kMaxUses];
  size_t replaceable_uses_count = 0;

  // Collect the uses to check case 2.
  Node* checkpoint_state = nullptr;
  if (checkpoint) {
    checkpoint_state = checkpoint->InputAt(0);
    if (!CollectFrameStateUniqueUses(callee, checkpoint_state, replaceable_uses,
                                     &replaceable_uses_count, kMaxUses)) {
      return false;
    }
  }

  // Collect the uses to check case 3.
  Node* frame_state = NodeProperties::GetFrameStateInput(node);
  if (!CollectFrameStateUniqueUses(callee, frame_state, replaceable_uses,
                                   &replaceable_uses_count, kMaxUses)) {
    return false;
  }

  // Bail out if there is a use of {callee} that is not reachable from 1, 2
  // and 3.
  for (Edge edge : callee->use_edges()) {
    // Case 1 (use by the call as a target).
    if (edge.from() == node && edge.index() == 0) continue;
    // Case 2 and 3 - used in checkpoint and/or lazy deopt frame states.
    bool found = false;
    for (size_t i = 0; i < replaceable_uses_count; i++) {
      if (replaceable_uses[i].node == edge.from() &&
          replaceable_uses[i].index == edge.index()) {
        found = true;
        break;
      }
    }
    if (!found) return false;
  }

  // Clone the call and the framestate, including the uniquely reachable
  // state values, making sure that we replace the phi with the constant.
  for (int i = 0; i < num_calls; ++i) {
    // Clone the calls for each branch.
    // We need to specialize the calls to the correct target, effect, and
    // control. We also need to duplicate the checkpoint and the lazy
    // frame state, and change all the uses of the callee to the constant
    // callee.
    Node* target = callee->InputAt(i);
    Node* effect = effect_phi->InputAt(i);
    Node* control = merge->InputAt(i);

    if (checkpoint) {
      // Duplicate the checkpoint.
      Node* new_checkpoint_state = DuplicateFrameStateAndRename(
          checkpoint_state, callee, target,
          (i == num_calls - 1) ? kChangeInPlace : kCloneState);
      effect = graph()->NewNode(checkpoint->op(), new_checkpoint_state, effect,
                                control);
    }

    // Duplicate the call.
    Node* new_lazy_frame_state = DuplicateFrameStateAndRename(
        frame_state, callee, target,
        (i == num_calls - 1) ? kChangeInPlace : kCloneState);
    inputs[0] = target;
    inputs[input_count - 3] = new_lazy_frame_state;
    inputs[input_count - 2] = effect;
    inputs[input_count - 1] = control;
    calls[i] = if_successes[i] =
        graph()->NewNode(node->op(), input_count, inputs);
  }

  // Mark the control inputs dead, so that we can kill the merge.
605 606 607
  node->ReplaceInput(input_count - 1, jsgraph()->Dead());
  callee->ReplaceInput(num_calls, jsgraph()->Dead());
  effect_phi->ReplaceInput(num_calls, jsgraph()->Dead());
608
  if (checkpoint) {
609
    checkpoint->ReplaceInput(2, jsgraph()->Dead());
610 611 612 613 614 615 616 617 618 619 620
  }

  merge->Kill();
  return true;
}

void JSInliningHeuristic::CreateOrReuseDispatch(Node* node, Node* callee,
                                                Candidate const& candidate,
                                                Node** if_successes,
                                                Node** calls, Node** inputs,
                                                int input_count) {
621 622
  SourcePositionTable::Scope position(
      source_positions_, source_positions_->GetSourcePosition(node));
623
  if (TryReuseDispatch(node, callee, if_successes, calls, inputs,
624 625 626 627
                       input_count)) {
    return;
  }

628 629
  STATIC_ASSERT(JSCallOrConstructNode::kHaveIdenticalLayouts);

630 631 632 633 634 635 636
  Node* fallthrough_control = NodeProperties::GetControlInput(node);
  int const num_calls = candidate.num_functions;

  // Create the appropriate control flow to dispatch to the cloned calls.
  for (int i = 0; i < num_calls; ++i) {
    // TODO(2206): Make comparison be based on underlying SharedFunctionInfo
    // instead of the target JSFunction reference directly.
637
    Node* target = jsgraph()->Constant(candidate.functions[i].value());
638 639 640 641 642 643 644 645 646 647 648 649 650 651
    if (i != (num_calls - 1)) {
      Node* check =
          graph()->NewNode(simplified()->ReferenceEqual(), callee, target);
      Node* branch =
          graph()->NewNode(common()->Branch(), check, fallthrough_control);
      fallthrough_control = graph()->NewNode(common()->IfFalse(), branch);
      if_successes[i] = graph()->NewNode(common()->IfTrue(), branch);
    } else {
      if_successes[i] = fallthrough_control;
    }

    // Clone the calls for each branch.
    // The first input to the call is the actual target (which we specialize
    // to the known {target}); the last input is the control dependency.
652 653 654
    // We also specialize the new.target of JSConstruct {node}s if it refers
    // to the same node as the {node}'s target input, so that we can later
    // properly inline the JSCreate operations.
655
    if (node->opcode() == IrOpcode::kJSConstruct) {
656 657 658 659 660
      // TODO(jgruber, v8:10675): This branch seems unreachable.
      JSConstructNode n(node);
      if (inputs[n.TargetIndex()] == inputs[n.NewTargetIndex()]) {
        inputs[n.NewTargetIndex()] = target;
      }
661
    }
662
    inputs[JSCallOrConstructNode::TargetIndex()] = target;
663 664 665 666 667 668
    inputs[input_count - 1] = if_successes[i];
    calls[i] = if_successes[i] =
        graph()->NewNode(node->op(), input_count, inputs);
  }
}

669
Reduction JSInliningHeuristic::InlineCandidate(Candidate const& candidate,
670
                                               bool small_function) {
671 672 673
  int const num_calls = candidate.num_functions;
  Node* const node = candidate.node;
  if (num_calls == 1) {
674
    Reduction const reduction = inliner_.ReduceJSCall(node);
675
    if (reduction.Changed()) {
676
      total_inlined_bytecode_size_ += candidate.bytecode[0].value().length();
677 678 679 680
    }
    return reduction;
  }

681
  // Expand the JSCall/JSConstruct node to a subgraph first if
682 683 684 685 686 687 688 689 690 691 692 693 694 695
  // we have multiple known target functions.
  DCHECK_LT(1, num_calls);
  Node* calls[kMaxCallPolymorphism + 1];
  Node* if_successes[kMaxCallPolymorphism];
  Node* callee = NodeProperties::GetValueInput(node, 0);

  // Setup the inputs for the cloned call nodes.
  int const input_count = node->InputCount();
  Node** inputs = graph()->zone()->NewArray<Node*>(input_count);
  for (int i = 0; i < input_count; ++i) {
    inputs[i] = node->InputAt(i);
  }

  // Create the appropriate control flow to dispatch to the cloned calls.
696 697
  CreateOrReuseDispatch(node, callee, candidate, if_successes, calls, inputs,
                        input_count);
698 699 700

  // Check if we have an exception projection for the call {node}.
  Node* if_exception = nullptr;
701
  if (NodeProperties::IsExceptionalCall(node, &if_exception)) {
702 703
    Node* if_exceptions[kMaxCallPolymorphism + 1];
    for (int i = 0; i < num_calls; ++i) {
704
      if_successes[i] = graph()->NewNode(common()->IfSuccess(), calls[i]);
705 706 707
      if_exceptions[i] =
          graph()->NewNode(common()->IfException(), calls[i], calls[i]);
    }
708 709

    // Morph the {if_exception} projection into a join.
710
    Node* exception_control =
711
        graph()->NewNode(common()->Merge(num_calls), num_calls, if_exceptions);
712 713 714 715
    if_exceptions[num_calls] = exception_control;
    Node* exception_effect = graph()->NewNode(common()->EffectPhi(num_calls),
                                              num_calls + 1, if_exceptions);
    Node* exception_value = graph()->NewNode(
716 717
        common()->Phi(MachineRepresentation::kTagged, num_calls), num_calls + 1,
        if_exceptions);
718 719
    ReplaceWithValue(if_exception, exception_value, exception_effect,
                     exception_control);
720 721
  }

722
  // Morph the original call site into a join of the dispatched call sites.
723
  Node* control =
724 725 726 727 728 729 730 731 732 733
      graph()->NewNode(common()->Merge(num_calls), num_calls, if_successes);
  calls[num_calls] = control;
  Node* effect =
      graph()->NewNode(common()->EffectPhi(num_calls), num_calls + 1, calls);
  Node* value =
      graph()->NewNode(common()->Phi(MachineRepresentation::kTagged, num_calls),
                       num_calls + 1, calls);
  ReplaceWithValue(node, value, effect, control);

  // Inline the individual, cloned call sites.
734 735 736
  for (int i = 0; i < num_calls && total_inlined_bytecode_size_ <
                                       FLAG_max_inlined_bytecode_size_absolute;
       ++i) {
737
    if (candidate.can_inline_function[i] &&
738 739 740
        (small_function || total_inlined_bytecode_size_ <
                               FLAG_max_inlined_bytecode_size_cumulative)) {
      Node* node = calls[i];
741 742
      Reduction const reduction = inliner_.ReduceJSCall(node);
      if (reduction.Changed()) {
743
        total_inlined_bytecode_size_ += candidate.bytecode[i]->length();
744 745 746 747
        // Killing the call node is not strictly necessary, but it is safer to
        // make sure we do not resurrect the node.
        node->Kill();
      }
748 749 750 751 752
    }
  }

  return Replace(value);
}
753

754 755
bool JSInliningHeuristic::CandidateCompare::operator()(
    const Candidate& left, const Candidate& right) const {
756
  if (right.frequency.IsUnknown()) {
757 758 759 760 761 762
    if (left.frequency.IsUnknown()) {
      // If left and right are both unknown then the ordering is indeterminate,
      // which breaks strict weak ordering requirements, so we fall back to the
      // node id as a tie breaker.
      return left.node->id() > right.node->id();
    }
763
    return true;
764 765 766 767 768
  } else if (left.frequency.IsUnknown()) {
    return false;
  } else if (left.frequency.value() > right.frequency.value()) {
    return true;
  } else if (left.frequency.value() < right.frequency.value()) {
769 770 771
    return false;
  } else {
    return left.node->id() > right.node->id();
772
  }
773 774 775
}

void JSInliningHeuristic::PrintCandidates() {
776
  StdoutStream os;
777
  os << candidates_.size() << " candidate(s) for inlining:" << std::endl;
778
  for (const Candidate& candidate : candidates_) {
779 780 781
    os << "- candidate: " << candidate.node->op()->mnemonic() << " node #"
       << candidate.node->id() << " with frequency " << candidate.frequency
       << ", " << candidate.num_functions << " target(s):" << std::endl;
782
    for (int i = 0; i < candidate.num_functions; ++i) {
783 784 785 786
      SharedFunctionInfoRef shared = candidate.functions[i].has_value()
                                         ? candidate.functions[i]->shared()
                                         : candidate.shared_info.value();
      os << "  - target: " << shared;
787
      if (candidate.bytecode[i].has_value()) {
788
        os << ", bytecode size: " << candidate.bytecode[i]->length();
789 790
        if (candidate.functions[i].has_value()) {
          JSFunctionRef function = candidate.functions[i].value();
791
          if (function.HasAttachedOptimizedCode()) {
792 793 794 795
            os << ", existing opt code's inlined bytecode size: "
               << function.code().inlined_bytecode_size();
          }
        }
796
      } else {
797
        os << ", no bytecode";
798
      }
799
      os << std::endl;
800
    }
801
  }
802 803
}

804 805 806 807 808 809 810 811 812 813
Graph* JSInliningHeuristic::graph() const { return jsgraph()->graph(); }

CommonOperatorBuilder* JSInliningHeuristic::common() const {
  return jsgraph()->common();
}

SimplifiedOperatorBuilder* JSInliningHeuristic::simplified() const {
  return jsgraph()->simplified();
}

814 815
#undef TRACE

816 817 818
}  // namespace compiler
}  // namespace internal
}  // namespace v8