Adding upstream version 115.7.0esr.upstream/115.7.0esrupstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

1 files changed, 3086 insertions, 0 deletions

diff --git a/js/src/jit/ScalarReplacement.cpp b/js/src/jit/ScalarReplacement.cpp
new file mode 100644
index 0000000000..0ded3601aa
--- /dev/null
+++ b/js/src/jit/ScalarReplacement.cpp
@@ -0,0 +1,3086 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*-
+ * vim: set ts=8 sts=2 et sw=2 tw=80:
+ * This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "jit/ScalarReplacement.h"
+
+#include "jit/IonAnalysis.h"
+#include "jit/JitSpewer.h"
+#include "jit/MIR.h"
+#include "jit/MIRGenerator.h"
+#include "jit/MIRGraph.h"
+#include "jit/WarpBuilderShared.h"
+#include "js/Vector.h"
+#include "vm/ArgumentsObject.h"
+
+#include "gc/ObjectKind-inl.h"
+
+namespace js {
+namespace jit {
+
+template <typename MemoryView>
+class EmulateStateOf {
+ private:
+  using BlockState = typename MemoryView::BlockState;
+
+  MIRGenerator* mir_;
+  MIRGraph& graph_;
+
+  // Block state at the entrance of all basic blocks.
+  Vector<BlockState*, 8, SystemAllocPolicy> states_;
+
+ public:
+  EmulateStateOf(MIRGenerator* mir, MIRGraph& graph)
+      : mir_(mir), graph_(graph) {}
+
+  bool run(MemoryView& view);
+};
+
+template <typename MemoryView>
+bool EmulateStateOf<MemoryView>::run(MemoryView& view) {
+  // Initialize the current block state of each block to an unknown state.
+  if (!states_.appendN(nullptr, graph_.numBlocks())) {
+    return false;
+  }
+
+  // Initialize the first block which needs to be traversed in RPO.
+  MBasicBlock* startBlock = view.startingBlock();
+  if (!view.initStartingState(&states_[startBlock->id()])) {
+    return false;
+  }
+
+  // Iterate over each basic block which has a valid entry state, and merge
+  // the state in the successor blocks.
+  for (ReversePostorderIterator block = graph_.rpoBegin(startBlock);
+       block != graph_.rpoEnd(); block++) {
+    if (mir_->shouldCancel(MemoryView::phaseName)) {
+      return false;
+    }
+
+    // Get the block state as the result of the merge of all predecessors
+    // which have already been visited in RPO.  This means that backedges
+    // are not yet merged into the loop.
+    BlockState* state = states_[block->id()];
+    if (!state) {
+      continue;
+    }
+    view.setEntryBlockState(state);
+
+    // Iterates over resume points, phi and instructions.
+    for (MNodeIterator iter(*block); iter;) {
+      // Increment the iterator before visiting the instruction, as the
+      // visit function might discard itself from the basic block.
+      MNode* ins = *iter++;
+      if (ins->isDefinition()) {
+        MDefinition* def = ins->toDefinition();
+        switch (def->op()) {
+#define MIR_OP(op)                 \
+  case MDefinition::Opcode::op:    \
+    view.visit##op(def->to##op()); \
+    break;
+          MIR_OPCODE_LIST(MIR_OP)
+#undef MIR_OP
+        }
+      } else {
+        view.visitResumePoint(ins->toResumePoint());
+      }
+      if (!graph_.alloc().ensureBallast()) {
+        return false;
+      }
+      if (view.oom()) {
+        return false;
+      }
+    }
+
+    // For each successor, merge the current state into the state of the
+    // successors.
+    for (size_t s = 0; s < block->numSuccessors(); s++) {
+      MBasicBlock* succ = block->getSuccessor(s);
+      if (!view.mergeIntoSuccessorState(*block, succ, &states_[succ->id()])) {
+        return false;
+      }
+    }
+  }
+
+  states_.clear();
+  return true;
+}
+
+static inline bool IsOptimizableObjectInstruction(MInstruction* ins) {
+  return ins->isNewObject() || ins->isNewPlainObject() ||
+         ins->isNewCallObject() || ins->isNewIterator();
+}
+
+static bool PhiOperandEqualTo(MDefinition* operand, MInstruction* newObject) {
+  if (operand == newObject) {
+    return true;
+  }
+
+  switch (operand->op()) {
+    case MDefinition::Opcode::GuardShape:
+      return PhiOperandEqualTo(operand->toGuardShape()->input(), newObject);
+
+    case MDefinition::Opcode::GuardToClass:
+      return PhiOperandEqualTo(operand->toGuardToClass()->input(), newObject);
+
+    case MDefinition::Opcode::CheckIsObj:
+      return PhiOperandEqualTo(operand->toCheckIsObj()->input(), newObject);
+
+    case MDefinition::Opcode::Unbox:
+      return PhiOperandEqualTo(operand->toUnbox()->input(), newObject);
+
+    default:
+      return false;
+  }
+}
+
+// Return true if all phi operands are equal to |newObject|.
+static bool PhiOperandsEqualTo(MPhi* phi, MInstruction* newObject) {
+  MOZ_ASSERT(IsOptimizableObjectInstruction(newObject));
+
+  for (size_t i = 0, e = phi->numOperands(); i < e; i++) {
+    if (!PhiOperandEqualTo(phi->getOperand(i), newObject)) {
+      return false;
+    }
+  }
+  return true;
+}
+
+static bool IsObjectEscaped(MDefinition* ins, MInstruction* newObject,
+                            const Shape* shapeDefault = nullptr);
+
+// Returns False if the lambda is not escaped and if it is optimizable by
+// ScalarReplacementOfObject.
+static bool IsLambdaEscaped(MInstruction* ins, MInstruction* lambda,
+                            MInstruction* newObject, const Shape* shape) {
+  MOZ_ASSERT(lambda->isLambda() || lambda->isFunctionWithProto());
+  MOZ_ASSERT(IsOptimizableObjectInstruction(newObject));
+  JitSpewDef(JitSpew_Escape, "Check lambda\n", ins);
+  JitSpewIndent spewIndent(JitSpew_Escape);
+
+  // The scope chain is not escaped if none of the Lambdas which are
+  // capturing it are escaped.
+  for (MUseIterator i(ins->usesBegin()); i != ins->usesEnd(); i++) {
+    MNode* consumer = (*i)->consumer();
+    if (!consumer->isDefinition()) {
+      // Cannot optimize if it is observable from fun.arguments or others.
+      if (!consumer->toResumePoint()->isRecoverableOperand(*i)) {
+        JitSpew(JitSpew_Escape, "Observable lambda cannot be recovered");
+        return true;
+      }
+      continue;
+    }
+
+    MDefinition* def = consumer->toDefinition();
+    switch (def->op()) {
+      case MDefinition::Opcode::GuardToFunction: {
+        auto* guard = def->toGuardToFunction();
+        if (IsLambdaEscaped(guard, lambda, newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::GuardFunctionScript: {
+        auto* guard = def->toGuardFunctionScript();
+        BaseScript* actual;
+        if (lambda->isLambda()) {
+          actual = lambda->toLambda()->templateFunction()->baseScript();
+        } else {
+          actual = lambda->toFunctionWithProto()->function()->baseScript();
+        }
+        if (actual != guard->expected()) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching script guard\n",
+                     guard);
+          return true;
+        }
+        if (IsLambdaEscaped(guard, lambda, newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::FunctionEnvironment: {
+        if (IsObjectEscaped(def->toFunctionEnvironment(), newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      default:
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", def);
+        return true;
+    }
+  }
+  JitSpew(JitSpew_Escape, "Lambda is not escaped");
+  return false;
+}
+
+static bool IsLambdaEscaped(MInstruction* lambda, MInstruction* newObject,
+                            const Shape* shape) {
+  return IsLambdaEscaped(lambda, lambda, newObject, shape);
+}
+
+// Returns False if the object is not escaped and if it is optimizable by
+// ScalarReplacementOfObject.
+//
+// For the moment, this code is dumb as it only supports objects which are not
+// changing shape.
+static bool IsObjectEscaped(MDefinition* ins, MInstruction* newObject,
+                            const Shape* shapeDefault) {
+  MOZ_ASSERT(ins->type() == MIRType::Object || ins->isPhi());
+  MOZ_ASSERT(IsOptimizableObjectInstruction(newObject));
+
+  JitSpewDef(JitSpew_Escape, "Check object\n", ins);
+  JitSpewIndent spewIndent(JitSpew_Escape);
+
+  const Shape* shape = shapeDefault;
+  if (!shape) {
+    if (ins->isNewPlainObject()) {
+      shape = ins->toNewPlainObject()->shape();
+    } else if (JSObject* templateObj = MObjectState::templateObjectOf(ins)) {
+      shape = templateObj->shape();
+    }
+  }
+
+  if (!shape) {
+    JitSpew(JitSpew_Escape, "No shape defined.");
+    return true;
+  }
+
+  // Check if the object is escaped. If the object is not the first argument
+  // of either a known Store / Load, then we consider it as escaped. This is a
+  // cheap and conservative escape analysis.
+  for (MUseIterator i(ins->usesBegin()); i != ins->usesEnd(); i++) {
+    MNode* consumer = (*i)->consumer();
+    if (!consumer->isDefinition()) {
+      // Cannot optimize if it is observable from fun.arguments or others.
+      if (!consumer->toResumePoint()->isRecoverableOperand(*i)) {
+        JitSpew(JitSpew_Escape, "Observable object cannot be recovered");
+        return true;
+      }
+      continue;
+    }
+
+    MDefinition* def = consumer->toDefinition();
+    switch (def->op()) {
+      case MDefinition::Opcode::StoreFixedSlot:
+      case MDefinition::Opcode::LoadFixedSlot:
+        // Not escaped if it is the first argument.
+        if (def->indexOf(*i) == 0) {
+          break;
+        }
+
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", def);
+        return true;
+
+      case MDefinition::Opcode::PostWriteBarrier:
+        break;
+
+      case MDefinition::Opcode::Slots: {
+#ifdef DEBUG
+        // Assert that MSlots are only used by MStoreDynamicSlot and
+        // MLoadDynamicSlot.
+        MSlots* ins = def->toSlots();
+        MOZ_ASSERT(ins->object() != 0);
+        for (MUseIterator i(ins->usesBegin()); i != ins->usesEnd(); i++) {
+          // toDefinition should normally never fail, since they don't get
+          // captured by resume points.
+          MDefinition* def = (*i)->consumer()->toDefinition();
+          MOZ_ASSERT(def->op() == MDefinition::Opcode::StoreDynamicSlot ||
+                     def->op() == MDefinition::Opcode::LoadDynamicSlot);
+        }
+#endif
+        break;
+      }
+
+      case MDefinition::Opcode::GuardShape: {
+        MGuardShape* guard = def->toGuardShape();
+        MOZ_ASSERT(!ins->isGuardShape());
+        if (shape != guard->shape()) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching guard shape\n", guard);
+          return true;
+        }
+        if (IsObjectEscaped(def->toInstruction(), newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::GuardToClass: {
+        MGuardToClass* guard = def->toGuardToClass();
+        if (!shape || shape->getObjectClass() != guard->getClass()) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching class guard\n", guard);
+          return true;
+        }
+        if (IsObjectEscaped(def->toInstruction(), newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::CheckIsObj: {
+        if (IsObjectEscaped(def->toInstruction(), newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::Unbox: {
+        if (def->type() != MIRType::Object) {
+          JitSpewDef(JitSpew_Escape, "has an invalid unbox\n", def);
+          return true;
+        }
+        if (IsObjectEscaped(def->toInstruction(), newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::Lambda:
+      case MDefinition::Opcode::FunctionWithProto: {
+        if (IsLambdaEscaped(def->toInstruction(), newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::Phi: {
+        auto* phi = def->toPhi();
+        if (!PhiOperandsEqualTo(phi, newObject)) {
+          JitSpewDef(JitSpew_Escape, "has different phi operands\n", def);
+          return true;
+        }
+        if (IsObjectEscaped(phi, newObject, shape)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::Compare: {
+        bool canFold;
+        if (!def->toCompare()->tryFold(&canFold)) {
+          JitSpewDef(JitSpew_Escape, "has an unsupported compare\n", def);
+          return true;
+        }
+        break;
+      }
+
+      // Doesn't escape the object.
+      case MDefinition::Opcode::IsObject:
+        break;
+
+      // This instruction is a no-op used to verify that scalar replacement
+      // is working as expected in jit-test.
+      case MDefinition::Opcode::AssertRecoveredOnBailout:
+        break;
+
+      // This is just a special flavor of constant which lets us optimize
+      // out some guards in certain circumstances. We'll turn this into a
+      // regular constant later.
+      case MDefinition::Opcode::ConstantProto:
+        break;
+
+      // We definitely don't need barriers for objects that don't exist.
+      case MDefinition::Opcode::AssertCanElidePostWriteBarrier:
+        break;
+
+      default:
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", def);
+        return true;
+    }
+  }
+
+  JitSpew(JitSpew_Escape, "Object is not escaped");
+  return false;
+}
+
+class ObjectMemoryView : public MDefinitionVisitorDefaultNoop {
+ public:
+  using BlockState = MObjectState;
+  static const char phaseName[];
+
+ private:
+  TempAllocator& alloc_;
+  MConstant* undefinedVal_;
+  MInstruction* obj_;
+  MBasicBlock* startBlock_;
+  BlockState* state_;
+
+  // Used to improve the memory usage by sharing common modification.
+  const MResumePoint* lastResumePoint_;
+
+  bool oom_;
+
+ public:
+  ObjectMemoryView(TempAllocator& alloc, MInstruction* obj);
+
+  MBasicBlock* startingBlock();
+  bool initStartingState(BlockState** pState);
+
+  void setEntryBlockState(BlockState* state);
+  bool mergeIntoSuccessorState(MBasicBlock* curr, MBasicBlock* succ,
+                               BlockState** pSuccState);
+
+#ifdef DEBUG
+  void assertSuccess();
+#else
+  void assertSuccess() {}
+#endif
+
+  bool oom() const { return oom_; }
+
+ private:
+  MDefinition* functionForCallObject(MDefinition* ins);
+
+ public:
+  void visitResumePoint(MResumePoint* rp);
+  void visitObjectState(MObjectState* ins);
+  void visitStoreFixedSlot(MStoreFixedSlot* ins);
+  void visitLoadFixedSlot(MLoadFixedSlot* ins);
+  void visitPostWriteBarrier(MPostWriteBarrier* ins);
+  void visitStoreDynamicSlot(MStoreDynamicSlot* ins);
+  void visitLoadDynamicSlot(MLoadDynamicSlot* ins);
+  void visitGuardShape(MGuardShape* ins);
+  void visitGuardToClass(MGuardToClass* ins);
+  void visitCheckIsObj(MCheckIsObj* ins);
+  void visitUnbox(MUnbox* ins);
+  void visitFunctionEnvironment(MFunctionEnvironment* ins);
+  void visitGuardToFunction(MGuardToFunction* ins);
+  void visitGuardFunctionScript(MGuardFunctionScript* ins);
+  void visitLambda(MLambda* ins);
+  void visitFunctionWithProto(MFunctionWithProto* ins);
+  void visitPhi(MPhi* ins);
+  void visitCompare(MCompare* ins);
+  void visitConstantProto(MConstantProto* ins);
+  void visitIsObject(MIsObject* ins);
+  void visitAssertCanElidePostWriteBarrier(
+      MAssertCanElidePostWriteBarrier* ins);
+};
+
+/* static */ const char ObjectMemoryView::phaseName[] =
+    "Scalar Replacement of Object";
+
+ObjectMemoryView::ObjectMemoryView(TempAllocator& alloc, MInstruction* obj)
+    : alloc_(alloc),
+      undefinedVal_(nullptr),
+      obj_(obj),
+      startBlock_(obj->block()),
+      state_(nullptr),
+      lastResumePoint_(nullptr),
+      oom_(false) {
+  // Annotate snapshots RValue such that we recover the store first.
+  obj_->setIncompleteObject();
+
+  // Annotate the instruction such that we do not replace it by a
+  // Magic(JS_OPTIMIZED_OUT) in case of removed uses.
+  obj_->setImplicitlyUsedUnchecked();
+}
+
+MBasicBlock* ObjectMemoryView::startingBlock() { return startBlock_; }
+
+bool ObjectMemoryView::initStartingState(BlockState** pState) {
+  // Uninitialized slots have an "undefined" value.
+  undefinedVal_ = MConstant::New(alloc_, UndefinedValue());
+  startBlock_->insertBefore(obj_, undefinedVal_);
+
+  // Create a new block state and insert at it at the location of the new
+  // object.
+  BlockState* state = BlockState::New(alloc_, obj_);
+  if (!state) {
+    return false;
+  }
+
+  startBlock_->insertAfter(obj_, state);
+
+  // Initialize the properties of the object state.
+  state->initFromTemplateObject(alloc_, undefinedVal_);
+
+  // Hold out of resume point until it is visited.
+  state->setInWorklist();
+
+  *pState = state;
+  return true;
+}
+
+void ObjectMemoryView::setEntryBlockState(BlockState* state) { state_ = state; }
+
+bool ObjectMemoryView::mergeIntoSuccessorState(MBasicBlock* curr,
+                                               MBasicBlock* succ,
+                                               BlockState** pSuccState) {
+  BlockState* succState = *pSuccState;
+
+  // When a block has no state yet, create an empty one for the
+  // successor.
+  if (!succState) {
+    // If the successor is not dominated then the object cannot flow
+    // in this basic block without a Phi.  We know that no Phi exist
+    // in non-dominated successors as the conservative escaped
+    // analysis fails otherwise.  Such condition can succeed if the
+    // successor is a join at the end of a if-block and the object
+    // only exists within the branch.
+    if (!startBlock_->dominates(succ)) {
+      return true;
+    }
+
+    // If there is only one predecessor, carry over the last state of the
+    // block to the successor.  As the block state is immutable, if the
+    // current block has multiple successors, they will share the same entry
+    // state.
+    if (succ->numPredecessors() <= 1 || !state_->numSlots()) {
+      *pSuccState = state_;
+      return true;
+    }
+
+    // If we have multiple predecessors, then we allocate one Phi node for
+    // each predecessor, and create a new block state which only has phi
+    // nodes.  These would later be removed by the removal of redundant phi
+    // nodes.
+    succState = BlockState::Copy(alloc_, state_);
+    if (!succState) {
+      return false;
+    }
+
+    size_t numPreds = succ->numPredecessors();
+    for (size_t slot = 0; slot < state_->numSlots(); slot++) {
+      MPhi* phi = MPhi::New(alloc_.fallible());
+      if (!phi || !phi->reserveLength(numPreds)) {
+        return false;
+      }
+
+      // Fill the input of the successors Phi with undefined
+      // values, and each block later fills the Phi inputs.
+      for (size_t p = 0; p < numPreds; p++) {
+        phi->addInput(undefinedVal_);
+      }
+
+      // Add Phi in the list of Phis of the basic block.
+      succ->addPhi(phi);
+      succState->setSlot(slot, phi);
+    }
+
+    // Insert the newly created block state instruction at the beginning
+    // of the successor block, after all the phi nodes.  Note that it
+    // would be captured by the entry resume point of the successor
+    // block.
+    succ->insertBefore(succ->safeInsertTop(), succState);
+    *pSuccState = succState;
+  }
+
+  MOZ_ASSERT_IF(succ == startBlock_, startBlock_->isLoopHeader());
+  if (succ->numPredecessors() > 1 && succState->numSlots() &&
+      succ != startBlock_) {
+    // We need to re-compute successorWithPhis as the previous EliminatePhis
+    // phase might have removed all the Phis from the successor block.
+    size_t currIndex;
+    MOZ_ASSERT(!succ->phisEmpty());
+    if (curr->successorWithPhis()) {
+      MOZ_ASSERT(curr->successorWithPhis() == succ);
+      currIndex = curr->positionInPhiSuccessor();
+    } else {
+      currIndex = succ->indexForPredecessor(curr);
+      curr->setSuccessorWithPhis(succ, currIndex);
+    }
+    MOZ_ASSERT(succ->getPredecessor(currIndex) == curr);
+
+    // Copy the current slot states to the index of current block in all the
+    // Phi created during the first visit of the successor.
+    for (size_t slot = 0; slot < state_->numSlots(); slot++) {
+      MPhi* phi = succState->getSlot(slot)->toPhi();
+      phi->replaceOperand(currIndex, state_->getSlot(slot));
+    }
+  }
+
+  return true;
+}
+
+#ifdef DEBUG
+void ObjectMemoryView::assertSuccess() {
+  for (MUseIterator i(obj_->usesBegin()); i != obj_->usesEnd(); i++) {
+    MNode* ins = (*i)->consumer();
+    MDefinition* def = nullptr;
+
+    // Resume points have been replaced by the object state.
+    if (ins->isResumePoint() ||
+        (def = ins->toDefinition())->isRecoveredOnBailout()) {
+      MOZ_ASSERT(obj_->isIncompleteObject());
+      continue;
+    }
+
+    // The only remaining uses would be removed by DCE, which will also
+    // recover the object on bailouts.
+    MOZ_ASSERT(def->isSlots() || def->isLambda() || def->isFunctionWithProto());
+    MOZ_ASSERT(!def->hasDefUses());
+  }
+}
+#endif
+
+void ObjectMemoryView::visitResumePoint(MResumePoint* rp) {
+  // As long as the MObjectState is not yet seen next to the allocation, we do
+  // not patch the resume point to recover the side effects.
+  if (!state_->isInWorklist()) {
+    rp->addStore(alloc_, state_, lastResumePoint_);
+    lastResumePoint_ = rp;
+  }
+}
+
+void ObjectMemoryView::visitObjectState(MObjectState* ins) {
+  if (ins->isInWorklist()) {
+    ins->setNotInWorklist();
+  }
+}
+
+void ObjectMemoryView::visitStoreFixedSlot(MStoreFixedSlot* ins) {
+  // Skip stores made on other objects.
+  if (ins->object() != obj_) {
+    return;
+  }
+
+  // Clone the state and update the slot value.
+  if (state_->hasFixedSlot(ins->slot())) {
+    state_ = BlockState::Copy(alloc_, state_);
+    if (!state_) {
+      oom_ = true;
+      return;
+    }
+
+    state_->setFixedSlot(ins->slot(), ins->value());
+    ins->block()->insertBefore(ins->toInstruction(), state_);
+  } else {
+    // UnsafeSetReserveSlot can access baked-in slots which are guarded by
+    // conditions, which are not seen by the escape analysis.
+    MBail* bailout = MBail::New(alloc_, BailoutKind::Inevitable);
+    ins->block()->insertBefore(ins, bailout);
+  }
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitLoadFixedSlot(MLoadFixedSlot* ins) {
+  // Skip loads made on other objects.
+  if (ins->object() != obj_) {
+    return;
+  }
+
+  // Replace load by the slot value.
+  if (state_->hasFixedSlot(ins->slot())) {
+    ins->replaceAllUsesWith(state_->getFixedSlot(ins->slot()));
+  } else {
+    // UnsafeGetReserveSlot can access baked-in slots which are guarded by
+    // conditions, which are not seen by the escape analysis.
+    MBail* bailout = MBail::New(alloc_, BailoutKind::Inevitable);
+    ins->block()->insertBefore(ins, bailout);
+    ins->replaceAllUsesWith(undefinedVal_);
+  }
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitPostWriteBarrier(MPostWriteBarrier* ins) {
+  // Skip loads made on other objects.
+  if (ins->object() != obj_) {
+    return;
+  }
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitStoreDynamicSlot(MStoreDynamicSlot* ins) {
+  // Skip stores made on other objects.
+  MSlots* slots = ins->slots()->toSlots();
+  if (slots->object() != obj_) {
+    // Guard objects are replaced when they are visited.
+    MOZ_ASSERT(!slots->object()->isGuardShape() ||
+               slots->object()->toGuardShape()->object() != obj_);
+    return;
+  }
+
+  // Clone the state and update the slot value.
+  if (state_->hasDynamicSlot(ins->slot())) {
+    state_ = BlockState::Copy(alloc_, state_);
+    if (!state_) {
+      oom_ = true;
+      return;
+    }
+
+    state_->setDynamicSlot(ins->slot(), ins->value());
+    ins->block()->insertBefore(ins->toInstruction(), state_);
+  } else {
+    // UnsafeSetReserveSlot can access baked-in slots which are guarded by
+    // conditions, which are not seen by the escape analysis.
+    MBail* bailout = MBail::New(alloc_, BailoutKind::Inevitable);
+    ins->block()->insertBefore(ins, bailout);
+  }
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitLoadDynamicSlot(MLoadDynamicSlot* ins) {
+  // Skip loads made on other objects.
+  MSlots* slots = ins->slots()->toSlots();
+  if (slots->object() != obj_) {
+    // Guard objects are replaced when they are visited.
+    MOZ_ASSERT(!slots->object()->isGuardShape() ||
+               slots->object()->toGuardShape()->object() != obj_);
+    return;
+  }
+
+  // Replace load by the slot value.
+  if (state_->hasDynamicSlot(ins->slot())) {
+    ins->replaceAllUsesWith(state_->getDynamicSlot(ins->slot()));
+  } else {
+    // UnsafeGetReserveSlot can access baked-in slots which are guarded by
+    // conditions, which are not seen by the escape analysis.
+    MBail* bailout = MBail::New(alloc_, BailoutKind::Inevitable);
+    ins->block()->insertBefore(ins, bailout);
+    ins->replaceAllUsesWith(undefinedVal_);
+  }
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitGuardShape(MGuardShape* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != obj_) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(obj_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitGuardToClass(MGuardToClass* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != obj_) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(obj_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitCheckIsObj(MCheckIsObj* ins) {
+  // Skip checks on other objects.
+  if (ins->input() != obj_) {
+    return;
+  }
+
+  // Replace the check by its object.
+  ins->replaceAllUsesWith(obj_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitUnbox(MUnbox* ins) {
+  // Skip unrelated unboxes.
+  if (ins->input() != obj_) {
+    return;
+  }
+  MOZ_ASSERT(ins->type() == MIRType::Object);
+
+  // Replace the unbox with the object.
+  ins->replaceAllUsesWith(obj_);
+
+  // Remove the unbox.
+  ins->block()->discard(ins);
+}
+
+MDefinition* ObjectMemoryView::functionForCallObject(MDefinition* ins) {
+  // Return early when we don't replace MNewCallObject.
+  if (!obj_->isNewCallObject()) {
+    return nullptr;
+  }
+
+  // Unwrap instructions until we found either MLambda or MFunctionWithProto.
+  // Return the function instruction if their environment chain matches the
+  // MNewCallObject we're about to replace.
+  while (true) {
+    switch (ins->op()) {
+      case MDefinition::Opcode::Lambda: {
+        if (ins->toLambda()->environmentChain() == obj_) {
+          return ins;
+        }
+        return nullptr;
+      }
+      case MDefinition::Opcode::FunctionWithProto: {
+        if (ins->toFunctionWithProto()->environmentChain() == obj_) {
+          return ins;
+        }
+        return nullptr;
+      }
+      case MDefinition::Opcode::FunctionEnvironment:
+        ins = ins->toFunctionEnvironment()->function();
+        break;
+      case MDefinition::Opcode::GuardToFunction:
+        ins = ins->toGuardToFunction()->object();
+        break;
+      case MDefinition::Opcode::GuardFunctionScript:
+        ins = ins->toGuardFunctionScript()->function();
+        break;
+      default:
+        return nullptr;
+    }
+  }
+}
+
+void ObjectMemoryView::visitFunctionEnvironment(MFunctionEnvironment* ins) {
+  // Skip function environment which are not aliases of the NewCallObject.
+  if (!functionForCallObject(ins)) {
+    return;
+  }
+
+  // Replace the function environment by the scope chain of the lambda.
+  ins->replaceAllUsesWith(obj_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitGuardToFunction(MGuardToFunction* ins) {
+  // Skip guards on other objects.
+  auto* function = functionForCallObject(ins);
+  if (!function) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(function);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitGuardFunctionScript(MGuardFunctionScript* ins) {
+  // Skip guards on other objects.
+  auto* function = functionForCallObject(ins);
+  if (!function) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(function);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitLambda(MLambda* ins) {
+  if (ins->environmentChain() != obj_) {
+    return;
+  }
+
+  // In order to recover the lambda we need to recover the scope chain, as the
+  // lambda is holding it.
+  ins->setIncompleteObject();
+}
+
+void ObjectMemoryView::visitFunctionWithProto(MFunctionWithProto* ins) {
+  if (ins->environmentChain() != obj_) {
+    return;
+  }
+
+  ins->setIncompleteObject();
+}
+
+void ObjectMemoryView::visitPhi(MPhi* ins) {
+  // Skip phis on other objects.
+  if (!PhiOperandsEqualTo(ins, obj_)) {
+    return;
+  }
+
+  // Replace the phi by its object.
+  ins->replaceAllUsesWith(obj_);
+
+  // Remove original instruction.
+  ins->block()->discardPhi(ins);
+}
+
+void ObjectMemoryView::visitCompare(MCompare* ins) {
+  // Skip unrelated comparisons.
+  if (ins->lhs() != obj_ && ins->rhs() != obj_) {
+    return;
+  }
+
+  bool folded;
+  MOZ_ALWAYS_TRUE(ins->tryFold(&folded));
+
+  auto* cst = MConstant::New(alloc_, BooleanValue(folded));
+  ins->block()->insertBefore(ins, cst);
+
+  // Replace the comparison with a constant.
+  ins->replaceAllUsesWith(cst);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitConstantProto(MConstantProto* ins) {
+  if (ins->getReceiverObject() != obj_) {
+    return;
+  }
+
+  auto* cst = ins->protoObject();
+  ins->replaceAllUsesWith(cst);
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitIsObject(MIsObject* ins) {
+  // Skip unrelated tests.
+  if (ins->input() != obj_) {
+    return;
+  }
+
+  auto* cst = MConstant::New(alloc_, BooleanValue(true));
+  ins->block()->insertBefore(ins, cst);
+
+  // Replace the test with a constant.
+  ins->replaceAllUsesWith(cst);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ObjectMemoryView::visitAssertCanElidePostWriteBarrier(
+    MAssertCanElidePostWriteBarrier* ins) {
+  if (ins->object() != obj_) {
+    return;
+  }
+
+  ins->block()->discard(ins);
+}
+
+static bool IndexOf(MDefinition* ins, int32_t* res) {
+  MOZ_ASSERT(ins->isLoadElement() || ins->isStoreElement());
+  MDefinition* indexDef = ins->getOperand(1);  // ins->index();
+  if (indexDef->isSpectreMaskIndex()) {
+    indexDef = indexDef->toSpectreMaskIndex()->index();
+  }
+  if (indexDef->isBoundsCheck()) {
+    indexDef = indexDef->toBoundsCheck()->index();
+  }
+  if (indexDef->isToNumberInt32()) {
+    indexDef = indexDef->toToNumberInt32()->getOperand(0);
+  }
+  MConstant* indexDefConst = indexDef->maybeConstantValue();
+  if (!indexDefConst || indexDefConst->type() != MIRType::Int32) {
+    return false;
+  }
+  *res = indexDefConst->toInt32();
+  return true;
+}
+
+static inline bool IsOptimizableArrayInstruction(MInstruction* ins) {
+  return ins->isNewArray() || ins->isNewArrayObject();
+}
+
+// We don't support storing holes when doing scalar replacement, so any
+// optimizable MNewArrayObject instruction is guaranteed to be packed.
+static inline bool IsPackedArray(MInstruction* ins) {
+  return ins->isNewArrayObject();
+}
+
+// Returns False if the elements is not escaped and if it is optimizable by
+// ScalarReplacementOfArray.
+static bool IsElementEscaped(MDefinition* def, MInstruction* newArray,
+                             uint32_t arraySize) {
+  MOZ_ASSERT(def->isElements());
+  MOZ_ASSERT(IsOptimizableArrayInstruction(newArray));
+
+  JitSpewDef(JitSpew_Escape, "Check elements\n", def);
+  JitSpewIndent spewIndent(JitSpew_Escape);
+
+  for (MUseIterator i(def->usesBegin()); i != def->usesEnd(); i++) {
+    // The MIRType::Elements cannot be captured in a resume point as
+    // it does not represent a value allocation.
+    MDefinition* access = (*i)->consumer()->toDefinition();
+
+    switch (access->op()) {
+      case MDefinition::Opcode::LoadElement: {
+        MOZ_ASSERT(access->toLoadElement()->elements() == def);
+
+        // If the index is not a constant then this index can alias
+        // all others. We do not handle this case.
+        int32_t index;
+        if (!IndexOf(access, &index)) {
+          JitSpewDef(JitSpew_Escape,
+                     "has a load element with a non-trivial index\n", access);
+          return true;
+        }
+        if (index < 0 || arraySize <= uint32_t(index)) {
+          JitSpewDef(JitSpew_Escape,
+                     "has a load element with an out-of-bound index\n", access);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::StoreElement: {
+        MStoreElement* storeElem = access->toStoreElement();
+        MOZ_ASSERT(storeElem->elements() == def);
+
+        // StoreElement must bail out if it stores to a hole, in case
+        // there is a setter on the prototype chain. If this StoreElement
+        // might store to a hole, we can't scalar-replace it.
+        if (storeElem->needsHoleCheck()) {
+          JitSpewDef(JitSpew_Escape, "has a store element with a hole check\n",
+                     storeElem);
+          return true;
+        }
+
+        // If the index is not a constant then this index can alias
+        // all others. We do not handle this case.
+        int32_t index;
+        if (!IndexOf(storeElem, &index)) {
+          JitSpewDef(JitSpew_Escape,
+                     "has a store element with a non-trivial index\n",
+                     storeElem);
+          return true;
+        }
+        if (index < 0 || arraySize <= uint32_t(index)) {
+          JitSpewDef(JitSpew_Escape,
+                     "has a store element with an out-of-bound index\n",
+                     storeElem);
+          return true;
+        }
+
+        // Dense element holes are written using MStoreHoleValueElement instead
+        // of MStoreElement.
+        MOZ_ASSERT(storeElem->value()->type() != MIRType::MagicHole);
+        break;
+      }
+
+      case MDefinition::Opcode::SetInitializedLength:
+        MOZ_ASSERT(access->toSetInitializedLength()->elements() == def);
+        break;
+
+      case MDefinition::Opcode::InitializedLength:
+        MOZ_ASSERT(access->toInitializedLength()->elements() == def);
+        break;
+
+      case MDefinition::Opcode::ArrayLength:
+        MOZ_ASSERT(access->toArrayLength()->elements() == def);
+        break;
+
+      case MDefinition::Opcode::ApplyArray:
+        MOZ_ASSERT(access->toApplyArray()->getElements() == def);
+        if (!IsPackedArray(newArray)) {
+          JitSpewDef(JitSpew_Escape, "is not guaranteed to be packed\n",
+                     access);
+          return true;
+        }
+        break;
+
+      case MDefinition::Opcode::ConstructArray:
+        MOZ_ASSERT(access->toConstructArray()->getElements() == def);
+        if (!IsPackedArray(newArray)) {
+          JitSpewDef(JitSpew_Escape, "is not guaranteed to be packed\n",
+                     access);
+          return true;
+        }
+        break;
+
+      default:
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", access);
+        return true;
+    }
+  }
+  JitSpew(JitSpew_Escape, "Elements is not escaped");
+  return false;
+}
+
+// Returns False if the array is not escaped and if it is optimizable by
+// ScalarReplacementOfArray.
+//
+// For the moment, this code is dumb as it only supports arrays which are not
+// changing length, with only access with known constants.
+static bool IsArrayEscaped(MInstruction* ins, MInstruction* newArray) {
+  MOZ_ASSERT(ins->type() == MIRType::Object);
+  MOZ_ASSERT(IsOptimizableArrayInstruction(newArray));
+
+  JitSpewDef(JitSpew_Escape, "Check array\n", ins);
+  JitSpewIndent spewIndent(JitSpew_Escape);
+
+  const Shape* shape;
+  uint32_t length;
+  if (newArray->isNewArrayObject()) {
+    length = newArray->toNewArrayObject()->length();
+    shape = newArray->toNewArrayObject()->shape();
+  } else {
+    length = newArray->toNewArray()->length();
+    JSObject* templateObject = newArray->toNewArray()->templateObject();
+    if (!templateObject) {
+      JitSpew(JitSpew_Escape, "No template object defined.");
+      return true;
+    }
+    shape = templateObject->shape();
+  }
+
+  if (length >= 16) {
+    JitSpew(JitSpew_Escape, "Array has too many elements");
+    return true;
+  }
+
+  // Check if the object is escaped. If the object is not the first argument
+  // of either a known Store / Load, then we consider it as escaped. This is a
+  // cheap and conservative escape analysis.
+  for (MUseIterator i(ins->usesBegin()); i != ins->usesEnd(); i++) {
+    MNode* consumer = (*i)->consumer();
+    if (!consumer->isDefinition()) {
+      // Cannot optimize if it is observable from fun.arguments or others.
+      if (!consumer->toResumePoint()->isRecoverableOperand(*i)) {
+        JitSpew(JitSpew_Escape, "Observable array cannot be recovered");
+        return true;
+      }
+      continue;
+    }
+
+    MDefinition* def = consumer->toDefinition();
+    switch (def->op()) {
+      case MDefinition::Opcode::Elements: {
+        MElements* elem = def->toElements();
+        MOZ_ASSERT(elem->object() == ins);
+        if (IsElementEscaped(elem, newArray, length)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", elem);
+          return true;
+        }
+
+        break;
+      }
+
+      case MDefinition::Opcode::GuardShape: {
+        MGuardShape* guard = def->toGuardShape();
+        if (shape != guard->shape()) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching guard shape\n", guard);
+          return true;
+        }
+        if (IsArrayEscaped(guard, newArray)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+
+        break;
+      }
+
+      case MDefinition::Opcode::GuardToClass: {
+        MGuardToClass* guard = def->toGuardToClass();
+        if (shape->getObjectClass() != guard->getClass()) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching class guard\n", guard);
+          return true;
+        }
+        if (IsArrayEscaped(guard, newArray)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+
+        break;
+      }
+
+      case MDefinition::Opcode::GuardArrayIsPacked: {
+        auto* guard = def->toGuardArrayIsPacked();
+        if (!IsPackedArray(newArray)) {
+          JitSpewDef(JitSpew_Escape, "is not guaranteed to be packed\n", def);
+          return true;
+        }
+        if (IsArrayEscaped(guard, newArray)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::Unbox: {
+        if (def->type() != MIRType::Object) {
+          JitSpewDef(JitSpew_Escape, "has an invalid unbox\n", def);
+          return true;
+        }
+        if (IsArrayEscaped(def->toInstruction(), newArray)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      // This instruction is supported for |JSOp::OptimizeSpreadCall|.
+      case MDefinition::Opcode::Compare: {
+        bool canFold;
+        if (!def->toCompare()->tryFold(&canFold)) {
+          JitSpewDef(JitSpew_Escape, "has an unsupported compare\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::PostWriteBarrier:
+      case MDefinition::Opcode::PostWriteElementBarrier:
+        break;
+
+      // This instruction is a no-op used to verify that scalar replacement
+      // is working as expected in jit-test.
+      case MDefinition::Opcode::AssertRecoveredOnBailout:
+        break;
+
+      default:
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", def);
+        return true;
+    }
+  }
+
+  JitSpew(JitSpew_Escape, "Array is not escaped");
+  return false;
+}
+
+// This class replaces every MStoreElement and MSetInitializedLength by an
+// MArrayState which emulates the content of the array. All MLoadElement,
+// MInitializedLength and MArrayLength are replaced by the corresponding value.
+//
+// In order to restore the value of the array correctly in case of bailouts, we
+// replace all reference of the allocation by the MArrayState definition.
+class ArrayMemoryView : public MDefinitionVisitorDefaultNoop {
+ public:
+  using BlockState = MArrayState;
+  static const char* phaseName;
+
+ private:
+  TempAllocator& alloc_;
+  MConstant* undefinedVal_;
+  MConstant* length_;
+  MInstruction* arr_;
+  MBasicBlock* startBlock_;
+  BlockState* state_;
+
+  // Used to improve the memory usage by sharing common modification.
+  const MResumePoint* lastResumePoint_;
+
+  bool oom_;
+
+ public:
+  ArrayMemoryView(TempAllocator& alloc, MInstruction* arr);
+
+  MBasicBlock* startingBlock();
+  bool initStartingState(BlockState** pState);
+
+  void setEntryBlockState(BlockState* state);
+  bool mergeIntoSuccessorState(MBasicBlock* curr, MBasicBlock* succ,
+                               BlockState** pSuccState);
+
+#ifdef DEBUG
+  void assertSuccess();
+#else
+  void assertSuccess() {}
+#endif
+
+  bool oom() const { return oom_; }
+
+ private:
+  bool isArrayStateElements(MDefinition* elements);
+  void discardInstruction(MInstruction* ins, MDefinition* elements);
+
+ public:
+  void visitResumePoint(MResumePoint* rp);
+  void visitArrayState(MArrayState* ins);
+  void visitStoreElement(MStoreElement* ins);
+  void visitLoadElement(MLoadElement* ins);
+  void visitSetInitializedLength(MSetInitializedLength* ins);
+  void visitInitializedLength(MInitializedLength* ins);
+  void visitArrayLength(MArrayLength* ins);
+  void visitPostWriteBarrier(MPostWriteBarrier* ins);
+  void visitPostWriteElementBarrier(MPostWriteElementBarrier* ins);
+  void visitGuardShape(MGuardShape* ins);
+  void visitGuardToClass(MGuardToClass* ins);
+  void visitGuardArrayIsPacked(MGuardArrayIsPacked* ins);
+  void visitUnbox(MUnbox* ins);
+  void visitCompare(MCompare* ins);
+  void visitApplyArray(MApplyArray* ins);
+  void visitConstructArray(MConstructArray* ins);
+};
+
+const char* ArrayMemoryView::phaseName = "Scalar Replacement of Array";
+
+ArrayMemoryView::ArrayMemoryView(TempAllocator& alloc, MInstruction* arr)
+    : alloc_(alloc),
+      undefinedVal_(nullptr),
+      length_(nullptr),
+      arr_(arr),
+      startBlock_(arr->block()),
+      state_(nullptr),
+      lastResumePoint_(nullptr),
+      oom_(false) {
+  // Annotate snapshots RValue such that we recover the store first.
+  arr_->setIncompleteObject();
+
+  // Annotate the instruction such that we do not replace it by a
+  // Magic(JS_OPTIMIZED_OUT) in case of removed uses.
+  arr_->setImplicitlyUsedUnchecked();
+}
+
+MBasicBlock* ArrayMemoryView::startingBlock() { return startBlock_; }
+
+bool ArrayMemoryView::initStartingState(BlockState** pState) {
+  // Uninitialized elements have an "undefined" value.
+  undefinedVal_ = MConstant::New(alloc_, UndefinedValue());
+  MConstant* initLength = MConstant::New(alloc_, Int32Value(0));
+  arr_->block()->insertBefore(arr_, undefinedVal_);
+  arr_->block()->insertBefore(arr_, initLength);
+
+  // Create a new block state and insert at it at the location of the new array.
+  BlockState* state = BlockState::New(alloc_, arr_, initLength);
+  if (!state) {
+    return false;
+  }
+
+  startBlock_->insertAfter(arr_, state);
+
+  // Initialize the elements of the array state.
+  state->initFromTemplateObject(alloc_, undefinedVal_);
+
+  // Hold out of resume point until it is visited.
+  state->setInWorklist();
+
+  *pState = state;
+  return true;
+}
+
+void ArrayMemoryView::setEntryBlockState(BlockState* state) { state_ = state; }
+
+bool ArrayMemoryView::mergeIntoSuccessorState(MBasicBlock* curr,
+                                              MBasicBlock* succ,
+                                              BlockState** pSuccState) {
+  BlockState* succState = *pSuccState;
+
+  // When a block has no state yet, create an empty one for the
+  // successor.
+  if (!succState) {
+    // If the successor is not dominated then the array cannot flow
+    // in this basic block without a Phi.  We know that no Phi exist
+    // in non-dominated successors as the conservative escaped
+    // analysis fails otherwise.  Such condition can succeed if the
+    // successor is a join at the end of a if-block and the array
+    // only exists within the branch.
+    if (!startBlock_->dominates(succ)) {
+      return true;
+    }
+
+    // If there is only one predecessor, carry over the last state of the
+    // block to the successor.  As the block state is immutable, if the
+    // current block has multiple successors, they will share the same entry
+    // state.
+    if (succ->numPredecessors() <= 1 || !state_->numElements()) {
+      *pSuccState = state_;
+      return true;
+    }
+
+    // If we have multiple predecessors, then we allocate one Phi node for
+    // each predecessor, and create a new block state which only has phi
+    // nodes.  These would later be removed by the removal of redundant phi
+    // nodes.
+    succState = BlockState::Copy(alloc_, state_);
+    if (!succState) {
+      return false;
+    }
+
+    size_t numPreds = succ->numPredecessors();
+    for (size_t index = 0; index < state_->numElements(); index++) {
+      MPhi* phi = MPhi::New(alloc_.fallible());
+      if (!phi || !phi->reserveLength(numPreds)) {
+        return false;
+      }
+
+      // Fill the input of the successors Phi with undefined
+      // values, and each block later fills the Phi inputs.
+      for (size_t p = 0; p < numPreds; p++) {
+        phi->addInput(undefinedVal_);
+      }
+
+      // Add Phi in the list of Phis of the basic block.
+      succ->addPhi(phi);
+      succState->setElement(index, phi);
+    }
+
+    // Insert the newly created block state instruction at the beginning
+    // of the successor block, after all the phi nodes.  Note that it
+    // would be captured by the entry resume point of the successor
+    // block.
+    succ->insertBefore(succ->safeInsertTop(), succState);
+    *pSuccState = succState;
+  }
+
+  MOZ_ASSERT_IF(succ == startBlock_, startBlock_->isLoopHeader());
+  if (succ->numPredecessors() > 1 && succState->numElements() &&
+      succ != startBlock_) {
+    // We need to re-compute successorWithPhis as the previous EliminatePhis
+    // phase might have removed all the Phis from the successor block.
+    size_t currIndex;
+    MOZ_ASSERT(!succ->phisEmpty());
+    if (curr->successorWithPhis()) {
+      MOZ_ASSERT(curr->successorWithPhis() == succ);
+      currIndex = curr->positionInPhiSuccessor();
+    } else {
+      currIndex = succ->indexForPredecessor(curr);
+      curr->setSuccessorWithPhis(succ, currIndex);
+    }
+    MOZ_ASSERT(succ->getPredecessor(currIndex) == curr);
+
+    // Copy the current element states to the index of current block in all
+    // the Phi created during the first visit of the successor.
+    for (size_t index = 0; index < state_->numElements(); index++) {
+      MPhi* phi = succState->getElement(index)->toPhi();
+      phi->replaceOperand(currIndex, state_->getElement(index));
+    }
+  }
+
+  return true;
+}
+
+#ifdef DEBUG
+void ArrayMemoryView::assertSuccess() { MOZ_ASSERT(!arr_->hasLiveDefUses()); }
+#endif
+
+void ArrayMemoryView::visitResumePoint(MResumePoint* rp) {
+  // As long as the MArrayState is not yet seen next to the allocation, we do
+  // not patch the resume point to recover the side effects.
+  if (!state_->isInWorklist()) {
+    rp->addStore(alloc_, state_, lastResumePoint_);
+    lastResumePoint_ = rp;
+  }
+}
+
+void ArrayMemoryView::visitArrayState(MArrayState* ins) {
+  if (ins->isInWorklist()) {
+    ins->setNotInWorklist();
+  }
+}
+
+bool ArrayMemoryView::isArrayStateElements(MDefinition* elements) {
+  return elements->isElements() && elements->toElements()->object() == arr_;
+}
+
+void ArrayMemoryView::discardInstruction(MInstruction* ins,
+                                         MDefinition* elements) {
+  MOZ_ASSERT(elements->isElements());
+  ins->block()->discard(ins);
+  if (!elements->hasLiveDefUses()) {
+    elements->block()->discard(elements->toInstruction());
+  }
+}
+
+void ArrayMemoryView::visitStoreElement(MStoreElement* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->elements();
+  if (!isArrayStateElements(elements)) {
+    return;
+  }
+
+  // Register value of the setter in the state.
+  int32_t index;
+  MOZ_ALWAYS_TRUE(IndexOf(ins, &index));
+  state_ = BlockState::Copy(alloc_, state_);
+  if (!state_) {
+    oom_ = true;
+    return;
+  }
+
+  state_->setElement(index, ins->value());
+  ins->block()->insertBefore(ins, state_);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void ArrayMemoryView::visitLoadElement(MLoadElement* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->elements();
+  if (!isArrayStateElements(elements)) {
+    return;
+  }
+
+  // Replace by the value contained at the index.
+  int32_t index;
+  MOZ_ALWAYS_TRUE(IndexOf(ins, &index));
+
+  // The only way to store a hole value in a new array is with
+  // StoreHoleValueElement, which IsElementEscaped does not allow.
+  // Therefore, we do not have to do a hole check.
+  MDefinition* element = state_->getElement(index);
+  MOZ_ASSERT(element->type() != MIRType::MagicHole);
+
+  ins->replaceAllUsesWith(element);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void ArrayMemoryView::visitSetInitializedLength(MSetInitializedLength* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->elements();
+  if (!isArrayStateElements(elements)) {
+    return;
+  }
+
+  // Replace by the new initialized length.  Note that the argument of
+  // MSetInitializedLength is the last index and not the initialized length.
+  // To obtain the length, we need to add 1 to it, and thus we need to create
+  // a new constant that we register in the ArrayState.
+  state_ = BlockState::Copy(alloc_, state_);
+  if (!state_) {
+    oom_ = true;
+    return;
+  }
+
+  int32_t initLengthValue = ins->index()->maybeConstantValue()->toInt32() + 1;
+  MConstant* initLength = MConstant::New(alloc_, Int32Value(initLengthValue));
+  ins->block()->insertBefore(ins, initLength);
+  ins->block()->insertBefore(ins, state_);
+  state_->setInitializedLength(initLength);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void ArrayMemoryView::visitInitializedLength(MInitializedLength* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->elements();
+  if (!isArrayStateElements(elements)) {
+    return;
+  }
+
+  // Replace by the value of the length.
+  ins->replaceAllUsesWith(state_->initializedLength());
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void ArrayMemoryView::visitArrayLength(MArrayLength* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->elements();
+  if (!isArrayStateElements(elements)) {
+    return;
+  }
+
+  // Replace by the value of the length.
+  if (!length_) {
+    length_ = MConstant::New(alloc_, Int32Value(state_->numElements()));
+    arr_->block()->insertBefore(arr_, length_);
+  }
+  ins->replaceAllUsesWith(length_);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void ArrayMemoryView::visitPostWriteBarrier(MPostWriteBarrier* ins) {
+  // Skip barriers on other objects.
+  if (ins->object() != arr_) {
+    return;
+  }
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArrayMemoryView::visitPostWriteElementBarrier(
+    MPostWriteElementBarrier* ins) {
+  // Skip barriers on other objects.
+  if (ins->object() != arr_) {
+    return;
+  }
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArrayMemoryView::visitGuardShape(MGuardShape* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != arr_) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(arr_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArrayMemoryView::visitGuardToClass(MGuardToClass* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != arr_) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(arr_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArrayMemoryView::visitGuardArrayIsPacked(MGuardArrayIsPacked* ins) {
+  // Skip guards on other objects.
+  if (ins->array() != arr_) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(arr_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArrayMemoryView::visitUnbox(MUnbox* ins) {
+  // Skip unrelated unboxes.
+  if (ins->getOperand(0) != arr_) {
+    return;
+  }
+  MOZ_ASSERT(ins->type() == MIRType::Object);
+
+  // Replace the unbox with the array object.
+  ins->replaceAllUsesWith(arr_);
+
+  // Remove the unbox.
+  ins->block()->discard(ins);
+}
+
+void ArrayMemoryView::visitCompare(MCompare* ins) {
+  // Skip unrelated comparisons.
+  if (ins->lhs() != arr_ && ins->rhs() != arr_) {
+    return;
+  }
+
+  bool folded;
+  MOZ_ALWAYS_TRUE(ins->tryFold(&folded));
+
+  auto* cst = MConstant::New(alloc_, BooleanValue(folded));
+  ins->block()->insertBefore(ins, cst);
+
+  // Replace the comparison with a constant.
+  ins->replaceAllUsesWith(cst);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArrayMemoryView::visitApplyArray(MApplyArray* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->getElements();
+  if (!isArrayStateElements(elements)) {
+    return;
+  }
+
+  uint32_t numElements = state_->numElements();
+
+  CallInfo callInfo(alloc_, /*constructing=*/false, ins->ignoresReturnValue());
+  if (!callInfo.initForApplyArray(ins->getFunction(), ins->getThis(),
+                                  numElements)) {
+    oom_ = true;
+    return;
+  }
+
+  for (uint32_t i = 0; i < numElements; i++) {
+    auto* element = state_->getElement(i);
+    MOZ_ASSERT(element->type() != MIRType::MagicHole);
+
+    callInfo.initArg(i, element);
+  }
+
+  auto addUndefined = [this]() { return undefinedVal_; };
+
+  bool needsThisCheck = false;
+  bool isDOMCall = false;
+  auto* call = MakeCall(alloc_, addUndefined, callInfo, needsThisCheck,
+                        ins->getSingleTarget(), isDOMCall);
+  if (!call) {
+    oom_ = true;
+    return;
+  }
+  if (!ins->maybeCrossRealm()) {
+    call->setNotCrossRealm();
+  }
+
+  ins->block()->insertBefore(ins, call);
+  ins->replaceAllUsesWith(call);
+
+  call->stealResumePoint(ins);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void ArrayMemoryView::visitConstructArray(MConstructArray* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->getElements();
+  if (!isArrayStateElements(elements)) {
+    return;
+  }
+
+  uint32_t numElements = state_->numElements();
+
+  CallInfo callInfo(alloc_, /*constructing=*/true, ins->ignoresReturnValue());
+  if (!callInfo.initForConstructArray(ins->getFunction(), ins->getThis(),
+                                      ins->getNewTarget(), numElements)) {
+    oom_ = true;
+    return;
+  }
+
+  for (uint32_t i = 0; i < numElements; i++) {
+    auto* element = state_->getElement(i);
+    MOZ_ASSERT(element->type() != MIRType::MagicHole);
+
+    callInfo.initArg(i, element);
+  }
+
+  auto addUndefined = [this]() { return undefinedVal_; };
+
+  bool needsThisCheck = ins->needsThisCheck();
+  bool isDOMCall = false;
+  auto* call = MakeCall(alloc_, addUndefined, callInfo, needsThisCheck,
+                        ins->getSingleTarget(), isDOMCall);
+  if (!call) {
+    oom_ = true;
+    return;
+  }
+  if (!ins->maybeCrossRealm()) {
+    call->setNotCrossRealm();
+  }
+
+  ins->block()->insertBefore(ins, call);
+  ins->replaceAllUsesWith(call);
+
+  call->stealResumePoint(ins);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+static inline bool IsOptimizableArgumentsInstruction(MInstruction* ins) {
+  return ins->isCreateArgumentsObject() ||
+         ins->isCreateInlinedArgumentsObject();
+}
+
+class ArgumentsReplacer : public MDefinitionVisitorDefaultNoop {
+ private:
+  MIRGenerator* mir_;
+  MIRGraph& graph_;
+  MInstruction* args_;
+
+  bool oom_ = false;
+
+  TempAllocator& alloc() { return graph_.alloc(); }
+
+  bool isInlinedArguments() const {
+    return args_->isCreateInlinedArgumentsObject();
+  }
+
+  MNewArrayObject* inlineArgsArray(MInstruction* ins, Shape* shape,
+                                   uint32_t begin, uint32_t count);
+
+  void visitGuardToClass(MGuardToClass* ins);
+  void visitGuardProto(MGuardProto* ins);
+  void visitGuardArgumentsObjectFlags(MGuardArgumentsObjectFlags* ins);
+  void visitUnbox(MUnbox* ins);
+  void visitGetArgumentsObjectArg(MGetArgumentsObjectArg* ins);
+  void visitLoadArgumentsObjectArg(MLoadArgumentsObjectArg* ins);
+  void visitLoadArgumentsObjectArgHole(MLoadArgumentsObjectArgHole* ins);
+  void visitInArgumentsObjectArg(MInArgumentsObjectArg* ins);
+  void visitArgumentsObjectLength(MArgumentsObjectLength* ins);
+  void visitApplyArgsObj(MApplyArgsObj* ins);
+  void visitArrayFromArgumentsObject(MArrayFromArgumentsObject* ins);
+  void visitArgumentsSlice(MArgumentsSlice* ins);
+  void visitLoadFixedSlot(MLoadFixedSlot* ins);
+
+  bool oom() const { return oom_; }
+
+ public:
+  ArgumentsReplacer(MIRGenerator* mir, MIRGraph& graph, MInstruction* args)
+      : mir_(mir), graph_(graph), args_(args) {
+    MOZ_ASSERT(IsOptimizableArgumentsInstruction(args_));
+  }
+
+  bool escapes(MInstruction* ins, bool guardedForMapped = false);
+  bool run();
+  void assertSuccess();
+};
+
+// Returns false if the arguments object does not escape.
+bool ArgumentsReplacer::escapes(MInstruction* ins, bool guardedForMapped) {
+  MOZ_ASSERT(ins->type() == MIRType::Object);
+
+  JitSpewDef(JitSpew_Escape, "Check arguments object\n", ins);
+  JitSpewIndent spewIndent(JitSpew_Escape);
+
+  // We can replace inlined arguments in scripts with OSR entries, but
+  // the outermost arguments object has already been allocated before
+  // we enter via OSR and can't be replaced.
+  if (ins->isCreateArgumentsObject() && graph_.osrBlock()) {
+    JitSpew(JitSpew_Escape, "Can't replace outermost OSR arguments");
+    return true;
+  }
+
+  // Check all uses to see whether they can be supported without
+  // allocating an ArgumentsObject.
+  for (MUseIterator i(ins->usesBegin()); i != ins->usesEnd(); i++) {
+    MNode* consumer = (*i)->consumer();
+
+    // If a resume point can observe this instruction, we can only optimize
+    // if it is recoverable.
+    if (consumer->isResumePoint()) {
+      if (!consumer->toResumePoint()->isRecoverableOperand(*i)) {
+        JitSpew(JitSpew_Escape, "Observable args object cannot be recovered");
+        return true;
+      }
+      continue;
+    }
+
+    MDefinition* def = consumer->toDefinition();
+    switch (def->op()) {
+      case MDefinition::Opcode::GuardToClass: {
+        MGuardToClass* guard = def->toGuardToClass();
+        if (!guard->isArgumentsObjectClass()) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching class guard\n", guard);
+          return true;
+        }
+        bool isMapped = guard->getClass() == &MappedArgumentsObject::class_;
+        if (escapes(guard, isMapped)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::GuardProto: {
+        if (escapes(def->toInstruction(), guardedForMapped)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::GuardArgumentsObjectFlags: {
+        if (escapes(def->toInstruction(), guardedForMapped)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::Unbox: {
+        if (def->type() != MIRType::Object) {
+          JitSpewDef(JitSpew_Escape, "has an invalid unbox\n", def);
+          return true;
+        }
+        if (escapes(def->toInstruction())) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::LoadFixedSlot: {
+        MLoadFixedSlot* load = def->toLoadFixedSlot();
+
+        // We can replace arguments.callee.
+        if (load->slot() == ArgumentsObject::CALLEE_SLOT) {
+          MOZ_ASSERT(guardedForMapped);
+          continue;
+        }
+        JitSpew(JitSpew_Escape, "is escaped by unsupported LoadFixedSlot\n");
+        return true;
+      }
+
+      case MDefinition::Opcode::ApplyArgsObj: {
+        if (ins == def->toApplyArgsObj()->getThis()) {
+          JitSpew(JitSpew_Escape, "is escaped as |this| arg of ApplyArgsObj\n");
+          return true;
+        }
+        MOZ_ASSERT(ins == def->toApplyArgsObj()->getArgsObj());
+        break;
+      }
+
+      // This is a replaceable consumer.
+      case MDefinition::Opcode::ArgumentsObjectLength:
+      case MDefinition::Opcode::GetArgumentsObjectArg:
+      case MDefinition::Opcode::LoadArgumentsObjectArg:
+      case MDefinition::Opcode::LoadArgumentsObjectArgHole:
+      case MDefinition::Opcode::InArgumentsObjectArg:
+      case MDefinition::Opcode::ArrayFromArgumentsObject:
+      case MDefinition::Opcode::ArgumentsSlice:
+        break;
+
+      // This instruction is a no-op used to test that scalar replacement
+      // is working as expected.
+      case MDefinition::Opcode::AssertRecoveredOnBailout:
+        break;
+
+      default:
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", def);
+        return true;
+    }
+  }
+
+  JitSpew(JitSpew_Escape, "ArgumentsObject is not escaped");
+  return false;
+}
+
+// Replacing the arguments object is simpler than replacing an object
+// or array, because the arguments object does not change state.
+bool ArgumentsReplacer::run() {
+  MBasicBlock* startBlock = args_->block();
+
+  // Iterate over each basic block.
+  for (ReversePostorderIterator block = graph_.rpoBegin(startBlock);
+       block != graph_.rpoEnd(); block++) {
+    if (mir_->shouldCancel("Scalar replacement of Arguments Object")) {
+      return false;
+    }
+
+    // Iterates over phis and instructions.
+    // We do not have to visit resume points. Any resume points that capture
+    // the argument object will be handled by the Sink pass.
+    for (MDefinitionIterator iter(*block); iter;) {
+      // Increment the iterator before visiting the instruction, as the
+      // visit function might discard itself from the basic block.
+      MDefinition* def = *iter++;
+      switch (def->op()) {
+#define MIR_OP(op)              \
+  case MDefinition::Opcode::op: \
+    visit##op(def->to##op());   \
+    break;
+        MIR_OPCODE_LIST(MIR_OP)
+#undef MIR_OP
+      }
+      if (!graph_.alloc().ensureBallast()) {
+        return false;
+      }
+      if (oom()) {
+        return false;
+      }
+    }
+  }
+
+  assertSuccess();
+  return true;
+}
+
+void ArgumentsReplacer::assertSuccess() {
+  MOZ_ASSERT(args_->canRecoverOnBailout());
+  MOZ_ASSERT(!args_->hasLiveDefUses());
+}
+
+void ArgumentsReplacer::visitGuardToClass(MGuardToClass* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != args_) {
+    return;
+  }
+  MOZ_ASSERT(ins->isArgumentsObjectClass());
+
+  // Replace the guard with the args object.
+  ins->replaceAllUsesWith(args_);
+
+  // Remove the guard.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitGuardProto(MGuardProto* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != args_) {
+    return;
+  }
+
+  // The prototype can only be changed through explicit operations, for example
+  // by calling |Reflect.setPrototype|. We have already determined that the args
+  // object doesn't escape, so its prototype can't be mutated.
+
+  // Replace the guard with the args object.
+  ins->replaceAllUsesWith(args_);
+
+  // Remove the guard.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitGuardArgumentsObjectFlags(
+    MGuardArgumentsObjectFlags* ins) {
+  // Skip other arguments objects.
+  if (ins->argsObject() != args_) {
+    return;
+  }
+
+#ifdef DEBUG
+  // Each *_OVERRIDDEN_BIT can only be set by setting or deleting a
+  // property of the args object. We have already determined that the
+  // args object doesn't escape, so its properties can't be mutated.
+  //
+  // FORWARDED_ARGUMENTS_BIT is set if any mapped argument is closed
+  // over, which is an immutable property of the script. Because we
+  // are replacing the args object for a known script, we can check
+  // the flag once, which is done when we first attach the CacheIR,
+  // and rely on it.  (Note that this wouldn't be true if we didn't
+  // know the origin of args_, because it could be passed in from
+  // another function.)
+  uint32_t supportedBits = ArgumentsObject::LENGTH_OVERRIDDEN_BIT |
+                           ArgumentsObject::ITERATOR_OVERRIDDEN_BIT |
+                           ArgumentsObject::ELEMENT_OVERRIDDEN_BIT |
+                           ArgumentsObject::CALLEE_OVERRIDDEN_BIT |
+                           ArgumentsObject::FORWARDED_ARGUMENTS_BIT;
+
+  MOZ_ASSERT((ins->flags() & ~supportedBits) == 0);
+  MOZ_ASSERT_IF(ins->flags() & ArgumentsObject::FORWARDED_ARGUMENTS_BIT,
+                !args_->block()->info().anyFormalIsForwarded());
+#endif
+
+  // Replace the guard with the args object.
+  ins->replaceAllUsesWith(args_);
+
+  // Remove the guard.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitUnbox(MUnbox* ins) {
+  // Skip unrelated unboxes.
+  if (ins->getOperand(0) != args_) {
+    return;
+  }
+  MOZ_ASSERT(ins->type() == MIRType::Object);
+
+  // Replace the unbox with the args object.
+  ins->replaceAllUsesWith(args_);
+
+  // Remove the unbox.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitGetArgumentsObjectArg(
+    MGetArgumentsObjectArg* ins) {
+  // Skip other arguments objects.
+  if (ins->argsObject() != args_) {
+    return;
+  }
+
+  // We don't support setting arguments in ArgumentsReplacer::escapes,
+  // so we can load the initial value of the argument without worrying
+  // about it being stale.
+  MDefinition* getArg;
+  if (isInlinedArguments()) {
+    // Inlined frames have direct access to the actual arguments.
+    auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+    if (ins->argno() < actualArgs->numActuals()) {
+      getArg = actualArgs->getArg(ins->argno());
+    } else {
+      // Omitted arguments are not mapped to the arguments object, and
+      // will always be undefined.
+      auto* undef = MConstant::New(alloc(), UndefinedValue());
+      ins->block()->insertBefore(ins, undef);
+      getArg = undef;
+    }
+  } else {
+    // Load the argument from the frame.
+    auto* index = MConstant::New(alloc(), Int32Value(ins->argno()));
+    ins->block()->insertBefore(ins, index);
+
+    auto* loadArg = MGetFrameArgument::New(alloc(), index);
+    ins->block()->insertBefore(ins, loadArg);
+    getArg = loadArg;
+  }
+  ins->replaceAllUsesWith(getArg);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitLoadArgumentsObjectArg(
+    MLoadArgumentsObjectArg* ins) {
+  // Skip other arguments objects.
+  if (ins->argsObject() != args_) {
+    return;
+  }
+
+  MDefinition* index = ins->index();
+
+  MInstruction* loadArg;
+  if (isInlinedArguments()) {
+    auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+
+    // Insert bounds check.
+    auto* length =
+        MConstant::New(alloc(), Int32Value(actualArgs->numActuals()));
+    ins->block()->insertBefore(ins, length);
+
+    MInstruction* check = MBoundsCheck::New(alloc(), index, length);
+    check->setBailoutKind(ins->bailoutKind());
+    ins->block()->insertBefore(ins, check);
+
+    if (mir_->outerInfo().hadBoundsCheckBailout()) {
+      check->setNotMovable();
+    }
+
+    loadArg = MGetInlinedArgument::New(alloc(), check, actualArgs);
+    if (!loadArg) {
+      oom_ = true;
+      return;
+    }
+  } else {
+    // Insert bounds check.
+    auto* length = MArgumentsLength::New(alloc());
+    ins->block()->insertBefore(ins, length);
+
+    MInstruction* check = MBoundsCheck::New(alloc(), index, length);
+    check->setBailoutKind(ins->bailoutKind());
+    ins->block()->insertBefore(ins, check);
+
+    if (mir_->outerInfo().hadBoundsCheckBailout()) {
+      check->setNotMovable();
+    }
+
+    if (JitOptions.spectreIndexMasking) {
+      check = MSpectreMaskIndex::New(alloc(), check, length);
+      ins->block()->insertBefore(ins, check);
+    }
+
+    loadArg = MGetFrameArgument::New(alloc(), check);
+  }
+  ins->block()->insertBefore(ins, loadArg);
+  ins->replaceAllUsesWith(loadArg);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitLoadArgumentsObjectArgHole(
+    MLoadArgumentsObjectArgHole* ins) {
+  // Skip other arguments objects.
+  if (ins->argsObject() != args_) {
+    return;
+  }
+
+  MDefinition* index = ins->index();
+
+  MInstruction* loadArg;
+  if (isInlinedArguments()) {
+    auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+
+    loadArg = MGetInlinedArgumentHole::New(alloc(), index, actualArgs);
+    if (!loadArg) {
+      oom_ = true;
+      return;
+    }
+  } else {
+    auto* length = MArgumentsLength::New(alloc());
+    ins->block()->insertBefore(ins, length);
+
+    loadArg = MGetFrameArgumentHole::New(alloc(), index, length);
+  }
+  loadArg->setBailoutKind(ins->bailoutKind());
+  ins->block()->insertBefore(ins, loadArg);
+  ins->replaceAllUsesWith(loadArg);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitInArgumentsObjectArg(MInArgumentsObjectArg* ins) {
+  // Skip other arguments objects.
+  if (ins->argsObject() != args_) {
+    return;
+  }
+
+  MDefinition* index = ins->index();
+
+  // Ensure the index is non-negative.
+  auto* guardedIndex = MGuardInt32IsNonNegative::New(alloc(), index);
+  guardedIndex->setBailoutKind(ins->bailoutKind());
+  ins->block()->insertBefore(ins, guardedIndex);
+
+  MInstruction* length;
+  if (isInlinedArguments()) {
+    uint32_t argc = args_->toCreateInlinedArgumentsObject()->numActuals();
+    length = MConstant::New(alloc(), Int32Value(argc));
+  } else {
+    length = MArgumentsLength::New(alloc());
+  }
+  ins->block()->insertBefore(ins, length);
+
+  auto* compare = MCompare::New(alloc(), guardedIndex, length, JSOp::Lt,
+                                MCompare::Compare_Int32);
+  ins->block()->insertBefore(ins, compare);
+  ins->replaceAllUsesWith(compare);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitArgumentsObjectLength(
+    MArgumentsObjectLength* ins) {
+  // Skip other arguments objects.
+  if (ins->argsObject() != args_) {
+    return;
+  }
+
+  MInstruction* length;
+  if (isInlinedArguments()) {
+    uint32_t argc = args_->toCreateInlinedArgumentsObject()->numActuals();
+    length = MConstant::New(alloc(), Int32Value(argc));
+  } else {
+    length = MArgumentsLength::New(alloc());
+  }
+  ins->block()->insertBefore(ins, length);
+  ins->replaceAllUsesWith(length);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitApplyArgsObj(MApplyArgsObj* ins) {
+  // Skip other arguments objects.
+  if (ins->getArgsObj() != args_) {
+    return;
+  }
+
+  MInstruction* newIns;
+  if (isInlinedArguments()) {
+    auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+    CallInfo callInfo(alloc(), /*constructing=*/false,
+                      ins->ignoresReturnValue());
+
+    callInfo.initForApplyInlinedArgs(ins->getFunction(), ins->getThis(),
+                                     actualArgs->numActuals());
+    for (uint32_t i = 0; i < actualArgs->numActuals(); i++) {
+      callInfo.initArg(i, actualArgs->getArg(i));
+    }
+
+    auto addUndefined = [this, &ins]() -> MConstant* {
+      MConstant* undef = MConstant::New(alloc(), UndefinedValue());
+      ins->block()->insertBefore(ins, undef);
+      return undef;
+    };
+
+    bool needsThisCheck = false;
+    bool isDOMCall = false;
+    auto* call = MakeCall(alloc(), addUndefined, callInfo, needsThisCheck,
+                          ins->getSingleTarget(), isDOMCall);
+    if (!call) {
+      oom_ = true;
+      return;
+    }
+    if (!ins->maybeCrossRealm()) {
+      call->setNotCrossRealm();
+    }
+    newIns = call;
+  } else {
+    auto* numArgs = MArgumentsLength::New(alloc());
+    ins->block()->insertBefore(ins, numArgs);
+
+    // TODO: Should we rename MApplyArgs?
+    auto* apply = MApplyArgs::New(alloc(), ins->getSingleTarget(),
+                                  ins->getFunction(), numArgs, ins->getThis());
+    apply->setBailoutKind(ins->bailoutKind());
+    if (!ins->maybeCrossRealm()) {
+      apply->setNotCrossRealm();
+    }
+    if (ins->ignoresReturnValue()) {
+      apply->setIgnoresReturnValue();
+    }
+    newIns = apply;
+  }
+
+  ins->block()->insertBefore(ins, newIns);
+  ins->replaceAllUsesWith(newIns);
+
+  newIns->stealResumePoint(ins);
+  ins->block()->discard(ins);
+}
+
+MNewArrayObject* ArgumentsReplacer::inlineArgsArray(MInstruction* ins,
+                                                    Shape* shape,
+                                                    uint32_t begin,
+                                                    uint32_t count) {
+  auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+
+  // Contrary to |WarpBuilder::build_Rest()|, we can always create
+  // MNewArrayObject, because we're guaranteed to have a shape and all
+  // arguments can be stored into fixed elements.
+  static_assert(
+      gc::CanUseFixedElementsForArray(ArgumentsObject::MaxInlinedArgs));
+
+  gc::Heap heap = gc::Heap::Default;
+
+  // Allocate an array of the correct size.
+  auto* shapeConstant = MConstant::NewShape(alloc(), shape);
+  ins->block()->insertBefore(ins, shapeConstant);
+
+  auto* newArray = MNewArrayObject::New(alloc(), shapeConstant, count, heap);
+  ins->block()->insertBefore(ins, newArray);
+
+  if (count) {
+    auto* elements = MElements::New(alloc(), newArray);
+    ins->block()->insertBefore(ins, elements);
+
+    MConstant* index = nullptr;
+    for (uint32_t i = 0; i < count; i++) {
+      index = MConstant::New(alloc(), Int32Value(i));
+      ins->block()->insertBefore(ins, index);
+
+      MDefinition* arg = actualArgs->getArg(begin + i);
+      auto* store = MStoreElement::NewUnbarriered(alloc(), elements, index, arg,
+                                                  /* needsHoleCheck = */ false);
+      ins->block()->insertBefore(ins, store);
+
+      auto* barrier = MPostWriteBarrier::New(alloc(), newArray, arg);
+      ins->block()->insertBefore(ins, barrier);
+    }
+
+    auto* initLength = MSetInitializedLength::New(alloc(), elements, index);
+    ins->block()->insertBefore(ins, initLength);
+  }
+
+  return newArray;
+}
+
+void ArgumentsReplacer::visitArrayFromArgumentsObject(
+    MArrayFromArgumentsObject* ins) {
+  // Skip other arguments objects.
+  if (ins->argsObject() != args_) {
+    return;
+  }
+
+  // We can only replace `arguments` because we've verified that the `arguments`
+  // object hasn't been modified in any way. This implies that the arguments
+  // stored in the stack frame haven't been changed either.
+  //
+  // The idea to replace `arguments` in spread calls `f(...arguments)` is now as
+  // follows:
+  // We replace |MArrayFromArgumentsObject| with the identical instructions we
+  // emit when building a rest-array object, cf. |WarpBuilder::build_Rest()|. In
+  // a next step, scalar replacement will then replace these new instructions
+  // themselves.
+
+  Shape* shape = ins->shape();
+  MOZ_ASSERT(shape);
+
+  MDefinition* replacement;
+  if (isInlinedArguments()) {
+    auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+    uint32_t numActuals = actualArgs->numActuals();
+    MOZ_ASSERT(numActuals <= ArgumentsObject::MaxInlinedArgs);
+
+    replacement = inlineArgsArray(ins, shape, 0, numActuals);
+  } else {
+    // We can use |MRest| to read all arguments, because we've guaranteed that
+    // the arguments stored in the stack frame haven't changed; see the comment
+    // at the start of this method.
+
+    auto* numActuals = MArgumentsLength::New(alloc());
+    ins->block()->insertBefore(ins, numActuals);
+
+    // Set |numFormals| to zero to read all arguments, including any formals.
+    uint32_t numFormals = 0;
+
+    auto* rest = MRest::New(alloc(), numActuals, numFormals, shape);
+    ins->block()->insertBefore(ins, rest);
+
+    replacement = rest;
+  }
+
+  ins->replaceAllUsesWith(replacement);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+static uint32_t NormalizeSlice(MDefinition* def, uint32_t length) {
+  int32_t value = def->toConstant()->toInt32();
+  if (value < 0) {
+    return std::max(int32_t(uint32_t(value) + length), 0);
+  }
+  return std::min(uint32_t(value), length);
+}
+
+void ArgumentsReplacer::visitArgumentsSlice(MArgumentsSlice* ins) {
+  // Skip other arguments objects.
+  if (ins->object() != args_) {
+    return;
+  }
+
+  // Optimise the common pattern |Array.prototype.slice.call(arguments, begin)|,
+  // where |begin| is a non-negative, constant int32.
+  //
+  // An absent end-index is replaced by |arguments.length|, so we try to match
+  // |Array.prototype.slice.call(arguments, begin, arguments.length)|.
+  if (isInlinedArguments()) {
+    // When this is an inlined arguments, |arguments.length| has been replaced
+    // by a constant.
+    if (ins->begin()->isConstant() && ins->end()->isConstant()) {
+      auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+      uint32_t numActuals = actualArgs->numActuals();
+      MOZ_ASSERT(numActuals <= ArgumentsObject::MaxInlinedArgs);
+
+      uint32_t begin = NormalizeSlice(ins->begin(), numActuals);
+      uint32_t end = NormalizeSlice(ins->end(), numActuals);
+      uint32_t count = end > begin ? end - begin : 0;
+      MOZ_ASSERT(count <= numActuals);
+
+      Shape* shape = ins->templateObj()->shape();
+      auto* newArray = inlineArgsArray(ins, shape, begin, count);
+
+      ins->replaceAllUsesWith(newArray);
+
+      // Remove original instruction.
+      ins->block()->discard(ins);
+      return;
+    }
+  } else {
+    // Otherwise |arguments.length| is emitted as MArgumentsLength.
+    if (ins->begin()->isConstant() && ins->end()->isArgumentsLength()) {
+      int32_t begin = ins->begin()->toConstant()->toInt32();
+      if (begin >= 0) {
+        auto* numActuals = MArgumentsLength::New(alloc());
+        ins->block()->insertBefore(ins, numActuals);
+
+        // Set |numFormals| to read all arguments starting at |begin|.
+        uint32_t numFormals = begin;
+
+        Shape* shape = ins->templateObj()->shape();
+
+        // Use MRest because it can be scalar replaced, which enables further
+        // optimizations.
+        auto* rest = MRest::New(alloc(), numActuals, numFormals, shape);
+        ins->block()->insertBefore(ins, rest);
+
+        ins->replaceAllUsesWith(rest);
+
+        // Remove original instruction.
+        ins->block()->discard(ins);
+        return;
+      }
+    }
+  }
+
+  MInstruction* numArgs;
+  if (isInlinedArguments()) {
+    uint32_t argc = args_->toCreateInlinedArgumentsObject()->numActuals();
+    numArgs = MConstant::New(alloc(), Int32Value(argc));
+  } else {
+    numArgs = MArgumentsLength::New(alloc());
+  }
+  ins->block()->insertBefore(ins, numArgs);
+
+  auto* begin = MNormalizeSliceTerm::New(alloc(), ins->begin(), numArgs);
+  ins->block()->insertBefore(ins, begin);
+
+  auto* end = MNormalizeSliceTerm::New(alloc(), ins->end(), numArgs);
+  ins->block()->insertBefore(ins, end);
+
+  bool isMax = false;
+  auto* beginMin = MMinMax::New(alloc(), begin, end, MIRType::Int32, isMax);
+  ins->block()->insertBefore(ins, beginMin);
+
+  // Safe to truncate because both operands are positive and end >= beginMin.
+  auto* count = MSub::New(alloc(), end, beginMin, MIRType::Int32);
+  count->setTruncateKind(TruncateKind::Truncate);
+  ins->block()->insertBefore(ins, count);
+
+  MInstruction* replacement;
+  if (isInlinedArguments()) {
+    auto* actualArgs = args_->toCreateInlinedArgumentsObject();
+    replacement =
+        MInlineArgumentsSlice::New(alloc(), beginMin, count, actualArgs,
+                                   ins->templateObj(), ins->initialHeap());
+  } else {
+    replacement = MFrameArgumentsSlice::New(
+        alloc(), beginMin, count, ins->templateObj(), ins->initialHeap());
+  }
+  ins->block()->insertBefore(ins, replacement);
+
+  ins->replaceAllUsesWith(replacement);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void ArgumentsReplacer::visitLoadFixedSlot(MLoadFixedSlot* ins) {
+  // Skip other arguments objects.
+  if (ins->object() != args_) {
+    return;
+  }
+
+  MOZ_ASSERT(ins->slot() == ArgumentsObject::CALLEE_SLOT);
+
+  MDefinition* replacement;
+  if (isInlinedArguments()) {
+    replacement = args_->toCreateInlinedArgumentsObject()->getCallee();
+  } else {
+    auto* callee = MCallee::New(alloc());
+    ins->block()->insertBefore(ins, callee);
+    replacement = callee;
+  }
+  ins->replaceAllUsesWith(replacement);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+static inline bool IsOptimizableRestInstruction(MInstruction* ins) {
+  return ins->isRest();
+}
+
+class RestReplacer : public MDefinitionVisitorDefaultNoop {
+ private:
+  MIRGenerator* mir_;
+  MIRGraph& graph_;
+  MInstruction* rest_;
+
+  TempAllocator& alloc() { return graph_.alloc(); }
+  MRest* rest() const { return rest_->toRest(); }
+
+  bool isRestElements(MDefinition* elements);
+  void discardInstruction(MInstruction* ins, MDefinition* elements);
+  MDefinition* restLength(MInstruction* ins);
+  void visitLength(MInstruction* ins, MDefinition* elements);
+
+  void visitGuardToClass(MGuardToClass* ins);
+  void visitGuardShape(MGuardShape* ins);
+  void visitGuardArrayIsPacked(MGuardArrayIsPacked* ins);
+  void visitUnbox(MUnbox* ins);
+  void visitCompare(MCompare* ins);
+  void visitLoadElement(MLoadElement* ins);
+  void visitArrayLength(MArrayLength* ins);
+  void visitInitializedLength(MInitializedLength* ins);
+  void visitApplyArray(MApplyArray* ins);
+  void visitConstructArray(MConstructArray* ins);
+
+  bool escapes(MElements* ins);
+
+ public:
+  RestReplacer(MIRGenerator* mir, MIRGraph& graph, MInstruction* rest)
+      : mir_(mir), graph_(graph), rest_(rest) {
+    MOZ_ASSERT(IsOptimizableRestInstruction(rest_));
+  }
+
+  bool escapes(MInstruction* ins);
+  bool run();
+  void assertSuccess();
+};
+
+// Returns false if the rest array object does not escape.
+bool RestReplacer::escapes(MInstruction* ins) {
+  MOZ_ASSERT(ins->type() == MIRType::Object);
+
+  JitSpewDef(JitSpew_Escape, "Check rest array\n", ins);
+  JitSpewIndent spewIndent(JitSpew_Escape);
+
+  // We can replace rest arrays in scripts with OSR entries, but the outermost
+  // rest object has already been allocated before we enter via OSR and can't be
+  // replaced.
+  // See also the same restriction when replacing |arguments|.
+  if (graph_.osrBlock()) {
+    JitSpew(JitSpew_Escape, "Can't replace outermost OSR rest array");
+    return true;
+  }
+
+  // Check all uses to see whether they can be supported without allocating an
+  // ArrayObject for the rest parameter.
+  for (MUseIterator i(ins->usesBegin()); i != ins->usesEnd(); i++) {
+    MNode* consumer = (*i)->consumer();
+
+    // If a resume point can observe this instruction, we can only optimize
+    // if it is recoverable.
+    if (consumer->isResumePoint()) {
+      if (!consumer->toResumePoint()->isRecoverableOperand(*i)) {
+        JitSpew(JitSpew_Escape, "Observable rest array cannot be recovered");
+        return true;
+      }
+      continue;
+    }
+
+    MDefinition* def = consumer->toDefinition();
+    switch (def->op()) {
+      case MDefinition::Opcode::Elements: {
+        auto* elem = def->toElements();
+        MOZ_ASSERT(elem->object() == ins);
+        if (escapes(elem)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::GuardShape: {
+        const Shape* shape = rest()->shape();
+        if (!shape) {
+          JitSpew(JitSpew_Escape, "No shape defined.");
+          return true;
+        }
+
+        auto* guard = def->toGuardShape();
+        if (shape != guard->shape()) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching guard shape\n", def);
+          return true;
+        }
+        if (escapes(guard)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::GuardToClass: {
+        auto* guard = def->toGuardToClass();
+        if (guard->getClass() != &ArrayObject::class_) {
+          JitSpewDef(JitSpew_Escape, "has a non-matching class guard\n", def);
+          return true;
+        }
+        if (escapes(guard)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::GuardArrayIsPacked: {
+        // Rest arrays are always packed as long as they aren't modified.
+        auto* guard = def->toGuardArrayIsPacked();
+        if (escapes(guard)) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      case MDefinition::Opcode::Unbox: {
+        if (def->type() != MIRType::Object) {
+          JitSpewDef(JitSpew_Escape, "has an invalid unbox\n", def);
+          return true;
+        }
+        if (escapes(def->toInstruction())) {
+          JitSpewDef(JitSpew_Escape, "is indirectly escaped by\n", def);
+          return true;
+        }
+        break;
+      }
+
+      // This instruction is supported for |JSOp::OptimizeSpreadCall|.
+      case MDefinition::Opcode::Compare: {
+        bool canFold;
+        if (!def->toCompare()->tryFold(&canFold)) {
+          JitSpewDef(JitSpew_Escape, "has an unsupported compare\n", def);
+          return true;
+        }
+        break;
+      }
+
+      // This instruction is a no-op used to test that scalar replacement is
+      // working as expected.
+      case MDefinition::Opcode::AssertRecoveredOnBailout:
+        break;
+
+      default:
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", def);
+        return true;
+    }
+  }
+
+  JitSpew(JitSpew_Escape, "Rest array object is not escaped");
+  return false;
+}
+
+bool RestReplacer::escapes(MElements* ins) {
+  JitSpewDef(JitSpew_Escape, "Check rest array elements\n", ins);
+  JitSpewIndent spewIndent(JitSpew_Escape);
+
+  for (MUseIterator i(ins->usesBegin()); i != ins->usesEnd(); i++) {
+    // The MIRType::Elements cannot be captured in a resume point as it does not
+    // represent a value allocation.
+    MDefinition* def = (*i)->consumer()->toDefinition();
+
+    switch (def->op()) {
+      case MDefinition::Opcode::LoadElement:
+        MOZ_ASSERT(def->toLoadElement()->elements() == ins);
+        break;
+
+      case MDefinition::Opcode::ArrayLength:
+        MOZ_ASSERT(def->toArrayLength()->elements() == ins);
+        break;
+
+      case MDefinition::Opcode::InitializedLength:
+        MOZ_ASSERT(def->toInitializedLength()->elements() == ins);
+        break;
+
+      case MDefinition::Opcode::ApplyArray:
+        MOZ_ASSERT(def->toApplyArray()->getElements() == ins);
+        break;
+
+      case MDefinition::Opcode::ConstructArray:
+        MOZ_ASSERT(def->toConstructArray()->getElements() == ins);
+        break;
+
+      default:
+        JitSpewDef(JitSpew_Escape, "is escaped by\n", def);
+        return true;
+    }
+  }
+
+  JitSpew(JitSpew_Escape, "Rest array object is not escaped");
+  return false;
+}
+
+// Replacing the rest array object is simpler than replacing an object or array,
+// because the rest array object does not change state.
+bool RestReplacer::run() {
+  MBasicBlock* startBlock = rest_->block();
+
+  // Iterate over each basic block.
+  for (ReversePostorderIterator block = graph_.rpoBegin(startBlock);
+       block != graph_.rpoEnd(); block++) {
+    if (mir_->shouldCancel("Scalar replacement of rest array object")) {
+      return false;
+    }
+
+    // Iterates over phis and instructions.
+    // We do not have to visit resume points. Any resume points that capture the
+    // rest array object will be handled by the Sink pass.
+    for (MDefinitionIterator iter(*block); iter;) {
+      // Increment the iterator before visiting the instruction, as the visit
+      // function might discard itself from the basic block.
+      MDefinition* def = *iter++;
+      switch (def->op()) {
+#define MIR_OP(op)              \
+  case MDefinition::Opcode::op: \
+    visit##op(def->to##op());   \
+    break;
+        MIR_OPCODE_LIST(MIR_OP)
+#undef MIR_OP
+      }
+      if (!graph_.alloc().ensureBallast()) {
+        return false;
+      }
+    }
+  }
+
+  assertSuccess();
+  return true;
+}
+
+void RestReplacer::assertSuccess() {
+  MOZ_ASSERT(rest_->canRecoverOnBailout());
+  MOZ_ASSERT(!rest_->hasLiveDefUses());
+}
+
+bool RestReplacer::isRestElements(MDefinition* elements) {
+  return elements->isElements() && elements->toElements()->object() == rest_;
+}
+
+void RestReplacer::discardInstruction(MInstruction* ins,
+                                      MDefinition* elements) {
+  MOZ_ASSERT(elements->isElements());
+  ins->block()->discard(ins);
+  if (!elements->hasLiveDefUses()) {
+    elements->block()->discard(elements->toInstruction());
+  }
+}
+
+void RestReplacer::visitGuardToClass(MGuardToClass* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != rest_) {
+    return;
+  }
+  MOZ_ASSERT(ins->getClass() == &ArrayObject::class_);
+
+  // Replace the guard with the array object.
+  ins->replaceAllUsesWith(rest_);
+
+  // Remove the guard.
+  ins->block()->discard(ins);
+}
+
+void RestReplacer::visitGuardShape(MGuardShape* ins) {
+  // Skip guards on other objects.
+  if (ins->object() != rest_) {
+    return;
+  }
+
+  // Replace the guard with the array object.
+  ins->replaceAllUsesWith(rest_);
+
+  // Remove the guard.
+  ins->block()->discard(ins);
+}
+
+void RestReplacer::visitGuardArrayIsPacked(MGuardArrayIsPacked* ins) {
+  // Skip guards on other objects.
+  if (ins->array() != rest_) {
+    return;
+  }
+
+  // Replace the guard by its object.
+  ins->replaceAllUsesWith(rest_);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void RestReplacer::visitUnbox(MUnbox* ins) {
+  // Skip unrelated unboxes.
+  if (ins->input() != rest_) {
+    return;
+  }
+  MOZ_ASSERT(ins->type() == MIRType::Object);
+
+  // Replace the unbox with the array object.
+  ins->replaceAllUsesWith(rest_);
+
+  // Remove the unbox.
+  ins->block()->discard(ins);
+}
+
+void RestReplacer::visitCompare(MCompare* ins) {
+  // Skip unrelated comparisons.
+  if (ins->lhs() != rest_ && ins->rhs() != rest_) {
+    return;
+  }
+
+  bool folded;
+  MOZ_ALWAYS_TRUE(ins->tryFold(&folded));
+
+  auto* cst = MConstant::New(alloc(), BooleanValue(folded));
+  ins->block()->insertBefore(ins, cst);
+
+  // Replace the comparison with a constant.
+  ins->replaceAllUsesWith(cst);
+
+  // Remove original instruction.
+  ins->block()->discard(ins);
+}
+
+void RestReplacer::visitLoadElement(MLoadElement* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->elements();
+  if (!isRestElements(elements)) {
+    return;
+  }
+
+  MDefinition* index = ins->index();
+
+  // Adjust the index to skip any extra formals.
+  if (uint32_t formals = rest()->numFormals()) {
+    auto* numFormals = MConstant::New(alloc(), Int32Value(formals));
+    ins->block()->insertBefore(ins, numFormals);
+
+    auto* add = MAdd::New(alloc(), index, numFormals, TruncateKind::Truncate);
+    ins->block()->insertBefore(ins, add);
+
+    index = add;
+  }
+
+  auto* loadArg = MGetFrameArgument::New(alloc(), index);
+
+  ins->block()->insertBefore(ins, loadArg);
+  ins->replaceAllUsesWith(loadArg);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+MDefinition* RestReplacer::restLength(MInstruction* ins) {
+  // Compute |Math.max(numActuals - numFormals, 0)| for the rest array length.
+
+  auto* numActuals = rest()->numActuals();
+
+  if (uint32_t formals = rest()->numFormals()) {
+    auto* numFormals = MConstant::New(alloc(), Int32Value(formals));
+    ins->block()->insertBefore(ins, numFormals);
+
+    auto* length = MSub::New(alloc(), numActuals, numFormals, MIRType::Int32);
+    length->setTruncateKind(TruncateKind::Truncate);
+    ins->block()->insertBefore(ins, length);
+
+    auto* zero = MConstant::New(alloc(), Int32Value(0));
+    ins->block()->insertBefore(ins, zero);
+
+    bool isMax = true;
+    auto* minmax = MMinMax::New(alloc(), length, zero, MIRType::Int32, isMax);
+    ins->block()->insertBefore(ins, minmax);
+
+    return minmax;
+  }
+
+  return numActuals;
+}
+
+void RestReplacer::visitLength(MInstruction* ins, MDefinition* elements) {
+  MOZ_ASSERT(ins->isArrayLength() || ins->isInitializedLength());
+
+  // Skip other array objects.
+  if (!isRestElements(elements)) {
+    return;
+  }
+
+  MDefinition* replacement = restLength(ins);
+
+  ins->replaceAllUsesWith(replacement);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void RestReplacer::visitArrayLength(MArrayLength* ins) {
+  visitLength(ins, ins->elements());
+}
+
+void RestReplacer::visitInitializedLength(MInitializedLength* ins) {
+  // The initialized length of a rest array is equal to its length.
+  visitLength(ins, ins->elements());
+}
+
+void RestReplacer::visitApplyArray(MApplyArray* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->getElements();
+  if (!isRestElements(elements)) {
+    return;
+  }
+
+  auto* numActuals = restLength(ins);
+
+  auto* apply =
+      MApplyArgs::New(alloc(), ins->getSingleTarget(), ins->getFunction(),
+                      numActuals, ins->getThis(), rest()->numFormals());
+  apply->setBailoutKind(ins->bailoutKind());
+  if (!ins->maybeCrossRealm()) {
+    apply->setNotCrossRealm();
+  }
+  if (ins->ignoresReturnValue()) {
+    apply->setIgnoresReturnValue();
+  }
+  ins->block()->insertBefore(ins, apply);
+
+  ins->replaceAllUsesWith(apply);
+
+  apply->stealResumePoint(ins);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+void RestReplacer::visitConstructArray(MConstructArray* ins) {
+  // Skip other array objects.
+  MDefinition* elements = ins->getElements();
+  if (!isRestElements(elements)) {
+    return;
+  }
+
+  auto* numActuals = restLength(ins);
+
+  auto* construct = MConstructArgs::New(
+      alloc(), ins->getSingleTarget(), ins->getFunction(), numActuals,
+      ins->getThis(), ins->getNewTarget(), rest()->numFormals());
+  construct->setBailoutKind(ins->bailoutKind());
+  if (!ins->maybeCrossRealm()) {
+    construct->setNotCrossRealm();
+  }
+
+  ins->block()->insertBefore(ins, construct);
+  ins->replaceAllUsesWith(construct);
+
+  construct->stealResumePoint(ins);
+
+  // Remove original instruction.
+  discardInstruction(ins, elements);
+}
+
+bool ScalarReplacement(MIRGenerator* mir, MIRGraph& graph) {
+  JitSpew(JitSpew_Escape, "Begin (ScalarReplacement)");
+
+  EmulateStateOf<ObjectMemoryView> replaceObject(mir, graph);
+  EmulateStateOf<ArrayMemoryView> replaceArray(mir, graph);
+  bool addedPhi = false;
+
+  for (ReversePostorderIterator block = graph.rpoBegin();
+       block != graph.rpoEnd(); block++) {
+    if (mir->shouldCancel("Scalar Replacement (main loop)")) {
+      return false;
+    }
+
+    for (MInstructionIterator ins = block->begin(); ins != block->end();
+         ins++) {
+      if (IsOptimizableObjectInstruction(*ins) &&
+          !IsObjectEscaped(*ins, *ins)) {
+        ObjectMemoryView view(graph.alloc(), *ins);
+        if (!replaceObject.run(view)) {
+          return false;
+        }
+        view.assertSuccess();
+        addedPhi = true;
+        continue;
+      }
+
+      if (IsOptimizableArrayInstruction(*ins) && !IsArrayEscaped(*ins, *ins)) {
+        ArrayMemoryView view(graph.alloc(), *ins);
+        if (!replaceArray.run(view)) {
+          return false;
+        }
+        view.assertSuccess();
+        addedPhi = true;
+        continue;
+      }
+
+      if (IsOptimizableArgumentsInstruction(*ins)) {
+        ArgumentsReplacer replacer(mir, graph, *ins);
+        if (replacer.escapes(*ins)) {
+          continue;
+        }
+        if (!replacer.run()) {
+          return false;
+        }
+        continue;
+      }
+
+      if (IsOptimizableRestInstruction(*ins)) {
+        RestReplacer replacer(mir, graph, *ins);
+        if (replacer.escapes(*ins)) {
+          continue;
+        }
+        if (!replacer.run()) {
+          return false;
+        }
+        continue;
+      }
+    }
+  }
+
+  if (addedPhi) {
+    // Phis added by Scalar Replacement are only redundant Phis which are
+    // not directly captured by any resume point but only by the MDefinition
+    // state. The conservative observability only focuses on Phis which are
+    // not used as resume points operands.
+    AssertExtendedGraphCoherency(graph);
+    if (!EliminatePhis(mir, graph, ConservativeObservability)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+} /* namespace jit */
+} /* namespace js */