diff --git a/compiler/optimizing/builder.cc b/compiler/optimizing/builder.cc
index f858f82..365599f 100644
--- a/compiler/optimizing/builder.cc
+++ b/compiler/optimizing/builder.cc
@@ -652,8 +652,8 @@
     DCHECK((optimized_invoke_type == invoke_type) || (optimized_invoke_type != kDirect)
            || compiler_driver_->GetCompilerOptions().GetCompilePic());
     bool is_recursive =
-        (target_method.dex_method_index == dex_compilation_unit_->GetDexMethodIndex());
-    DCHECK(!is_recursive || (target_method.dex_file == dex_compilation_unit_->GetDexFile()));
+        (target_method.dex_method_index == outer_compilation_unit_->GetDexMethodIndex())
+        && (target_method.dex_file == outer_compilation_unit_->GetDexFile());
 
     if (optimized_invoke_type == kStatic) {
       ScopedObjectAccess soa(Thread::Current());
diff --git a/compiler/optimizing/graph_visualizer.cc b/compiler/optimizing/graph_visualizer.cc
index 7da4f2d..fd2e4e8 100644
--- a/compiler/optimizing/graph_visualizer.cc
+++ b/compiler/optimizing/graph_visualizer.cc
@@ -280,6 +280,13 @@
         << instance_of->MustDoNullCheck() << std::noboolalpha;
   }
 
+  void VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) OVERRIDE {
+    StartAttributeStream("dex_file_index") << invoke->GetDexMethodIndex();
+    StartAttributeStream("recursive") << std::boolalpha
+                                      << invoke->IsRecursive()
+                                      << std::noboolalpha;
+  }
+
   bool IsPass(const char* name) {
     return strcmp(pass_name_, name) == 0;
   }
diff --git a/compiler/optimizing/inliner.cc b/compiler/optimizing/inliner.cc
index 997f980..15f3deb 100644
--- a/compiler/optimizing/inliner.cc
+++ b/compiler/optimizing/inliner.cc
@@ -36,8 +36,8 @@
 
 namespace art {
 
-static constexpr int kMaxInlineCodeUnits = 100;
-static constexpr int kDepthLimit = 5;
+static constexpr int kMaxInlineCodeUnits = 18;
+static constexpr int kDepthLimit = 3;
 
 void HInliner::Run() {
   if (graph_->IsDebuggable()) {
@@ -46,8 +46,15 @@
     return;
   }
   const GrowableArray<HBasicBlock*>& blocks = graph_->GetReversePostOrder();
+  HBasicBlock* next_block = blocks.Get(0);
   for (size_t i = 0; i < blocks.Size(); ++i) {
-    HBasicBlock* block = blocks.Get(i);
+    // Because we are changing the graph when inlining, we need to remember the next block.
+    // This avoids doing the inlining work again on the inlined blocks.
+    if (blocks.Get(i) != next_block) {
+      continue;
+    }
+    HBasicBlock* block = next_block;
+    next_block = (i == blocks.Size() - 1) ? nullptr : blocks.Get(i + 1);
     for (HInstruction* instruction = block->GetFirstInstruction(); instruction != nullptr;) {
       HInstruction* next = instruction->GetNext();
       HInvokeStaticOrDirect* call = instruction->AsInvokeStaticOrDirect();
@@ -90,10 +97,10 @@
     return false;
   }
 
-  bool can_use_dex_cache = true;
+  bool same_dex_file = true;
   const DexFile& outer_dex_file = *outer_compilation_unit_.GetDexFile();
   if (resolved_method->GetDexFile()->GetLocation().compare(outer_dex_file.GetLocation()) != 0) {
-    can_use_dex_cache = false;
+    same_dex_file = false;
   }
 
   const DexFile::CodeItem* code_item = resolved_method->GetCodeItem();
@@ -140,7 +147,7 @@
     return false;
   }
 
-  if (!TryBuildAndInline(resolved_method, invoke_instruction, method_index, can_use_dex_cache)) {
+  if (!TryBuildAndInline(resolved_method, invoke_instruction, method_index, same_dex_file)) {
     return false;
   }
 
@@ -152,7 +159,7 @@
 bool HInliner::TryBuildAndInline(Handle<mirror::ArtMethod> resolved_method,
                                  HInvoke* invoke_instruction,
                                  uint32_t method_index,
-                                 bool can_use_dex_cache) const {
+                                 bool same_dex_file) const {
   ScopedObjectAccess soa(Thread::Current());
   const DexFile::CodeItem* code_item = resolved_method->GetCodeItem();
   const DexFile& caller_dex_file = *caller_compilation_unit_.GetDexFile();
@@ -254,6 +261,31 @@
     inliner.Run();
   }
 
+  // TODO: We should abort only if all predecessors throw. However,
+  // HGraph::InlineInto currently does not handle an exit block with
+  // a throw predecessor.
+  HBasicBlock* exit_block = callee_graph->GetExitBlock();
+  if (exit_block == nullptr) {
+    VLOG(compiler) << "Method " << PrettyMethod(method_index, caller_dex_file)
+                   << " could not be inlined because it has an infinite loop";
+    resolved_method->SetShouldNotInline();
+    return false;
+  }
+
+  bool has_throw_predecessor = false;
+  for (size_t i = 0, e = exit_block->GetPredecessors().Size(); i < e; ++i) {
+    if (exit_block->GetPredecessors().Get(i)->GetLastInstruction()->IsThrow()) {
+      has_throw_predecessor = true;
+      break;
+    }
+  }
+  if (has_throw_predecessor) {
+    VLOG(compiler) << "Method " << PrettyMethod(method_index, caller_dex_file)
+                   << " could not be inlined because one branch always throws";
+    resolved_method->SetShouldNotInline();
+    return false;
+  }
+
   HReversePostOrderIterator it(*callee_graph);
   it.Advance();  // Past the entry block, it does not contain instructions that prevent inlining.
   for (; !it.Done(); it.Advance()) {
@@ -269,27 +301,24 @@
          !instr_it.Done();
          instr_it.Advance()) {
       HInstruction* current = instr_it.Current();
-      if (current->IsSuspendCheck()) {
-        continue;
-      }
 
-      if (current->CanThrow()) {
+      if (current->IsInvokeInterface()) {
+        // Disable inlining of interface calls. The cost in case of entering the
+        // resolution conflict is currently too high.
         VLOG(compiler) << "Method " << PrettyMethod(method_index, caller_dex_file)
-                       << " could not be inlined because " << current->DebugName()
-                       << " can throw";
+                       << " could not be inlined because it has an interface call.";
         resolved_method->SetShouldNotInline();
         return false;
       }
 
-      if (current->NeedsEnvironment()) {
+      if (!same_dex_file && current->NeedsEnvironment()) {
         VLOG(compiler) << "Method " << PrettyMethod(method_index, caller_dex_file)
                        << " could not be inlined because " << current->DebugName()
-                       << " needs an environment";
-        resolved_method->SetShouldNotInline();
+                       << " needs an environment and is in a different dex file";
         return false;
       }
 
-      if (!can_use_dex_cache && current->NeedsDexCache()) {
+      if (!same_dex_file && current->NeedsDexCache()) {
         VLOG(compiler) << "Method " << PrettyMethod(method_index, caller_dex_file)
                        << " could not be inlined because " << current->DebugName()
                        << " it is in a different dex file and requires access to the dex cache";
@@ -302,10 +331,6 @@
 
   callee_graph->InlineInto(graph_, invoke_instruction);
 
-  if (callee_graph->HasBoundsChecks()) {
-    graph_->SetHasBoundsChecks(true);
-  }
-
   return true;
 }
 
diff --git a/compiler/optimizing/inliner.h b/compiler/optimizing/inliner.h
index 1dbc7d3..09a36c6 100644
--- a/compiler/optimizing/inliner.h
+++ b/compiler/optimizing/inliner.h
@@ -51,7 +51,7 @@
   bool TryBuildAndInline(Handle<mirror::ArtMethod> resolved_method,
                          HInvoke* invoke_instruction,
                          uint32_t method_index,
-                         bool can_use_dex_cache) const;
+                         bool same_dex_file) const;
 
   const DexCompilationUnit& outer_compilation_unit_;
   const DexCompilationUnit& caller_compilation_unit_;
diff --git a/compiler/optimizing/nodes.cc b/compiler/optimizing/nodes.cc
index 80d4b4a..06f6a7f 100644
--- a/compiler/optimizing/nodes.cc
+++ b/compiler/optimizing/nodes.cc
@@ -1314,6 +1314,29 @@
 
 void HGraph::InlineInto(HGraph* outer_graph, HInvoke* invoke) {
   DCHECK(HasExitBlock()) << "Unimplemented scenario";
+  // Update the environments in this graph to have the invoke's environment
+  // as parent.
+  {
+    HReversePostOrderIterator it(*this);
+    it.Advance();  // Skip the entry block, we do not need to update the entry's suspend check.
+    for (; !it.Done(); it.Advance()) {
+      HBasicBlock* block = it.Current();
+      for (HInstructionIterator instr_it(block->GetInstructions());
+           !instr_it.Done();
+           instr_it.Advance()) {
+        HInstruction* current = instr_it.Current();
+        if (current->NeedsEnvironment()) {
+          current->GetEnvironment()->SetAndCopyParentChain(
+              outer_graph->GetArena(), invoke->GetEnvironment());
+        }
+      }
+    }
+  }
+  outer_graph->UpdateMaximumNumberOfOutVRegs(GetMaximumNumberOfOutVRegs());
+  if (HasBoundsChecks()) {
+    outer_graph->SetHasBoundsChecks(true);
+  }
+
   if (GetBlocks().Size() == 3) {
     // Simple case of an entry block, a body block, and an exit block.
     // Put the body block's instruction into `invoke`'s block.
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 3144c5c..005d50e 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -143,6 +143,7 @@
         dex_file_(dex_file),
         method_idx_(method_idx),
         invoke_type_(invoke_type),
+        in_ssa_form_(false),
         should_generate_constructor_barrier_(should_generate_constructor_barrier),
         cached_null_constant_(nullptr),
         cached_int_constants_(std::less<int32_t>(), arena->Adapter()),
@@ -174,6 +175,7 @@
     // users remaining when being visited.
     if (!AnalyzeNaturalLoops()) return false;
     TransformToSsa();
+    in_ssa_form_ = true;
     return true;
   }
 
@@ -216,11 +218,16 @@
     maximum_number_of_out_vregs_ = new_value;
   }
 
+  void UpdateMaximumNumberOfOutVRegs(uint16_t other_value) {
+    maximum_number_of_out_vregs_ = std::max(maximum_number_of_out_vregs_, other_value);
+  }
+
   void UpdateTemporariesVRegSlots(size_t slots) {
     temporaries_vreg_slots_ = std::max(slots, temporaries_vreg_slots_);
   }
 
   size_t GetTemporariesVRegSlots() const {
+    DCHECK(!in_ssa_form_);
     return temporaries_vreg_slots_;
   }
 
@@ -229,6 +236,7 @@
   }
 
   uint16_t GetNumberOfVRegs() const {
+    DCHECK(!in_ssa_form_);
     return number_of_vregs_;
   }
 
@@ -237,6 +245,7 @@
   }
 
   uint16_t GetNumberOfLocalVRegs() const {
+    DCHECK(!in_ssa_form_);
     return number_of_vregs_ - number_of_in_vregs_;
   }
 
@@ -381,6 +390,11 @@
   // If inlined, this encodes how the callee is being invoked.
   const InvokeType invoke_type_;
 
+  // Whether the graph has been transformed to SSA form. Only used
+  // in debug mode to ensure we are not using properties only valid
+  // for non-SSA form (like the number of temporaries).
+  bool in_ssa_form_;
+
   const bool should_generate_constructor_barrier_;
 
   // Cached constants.
@@ -1121,14 +1135,16 @@
                const DexFile& dex_file,
                uint32_t method_idx,
                uint32_t dex_pc,
-               InvokeType invoke_type)
+               InvokeType invoke_type,
+               HInstruction* holder)
      : vregs_(arena, number_of_vregs),
        locations_(arena, number_of_vregs),
        parent_(nullptr),
        dex_file_(dex_file),
        method_idx_(method_idx),
        dex_pc_(dex_pc),
-       invoke_type_(invoke_type) {
+       invoke_type_(invoke_type),
+       holder_(holder) {
     vregs_.SetSize(number_of_vregs);
     for (size_t i = 0; i < number_of_vregs; i++) {
       vregs_.Put(i, HUserRecord<HEnvironment*>());
@@ -1140,19 +1156,24 @@
     }
   }
 
-  HEnvironment(ArenaAllocator* arena, const HEnvironment& to_copy)
+  HEnvironment(ArenaAllocator* arena, const HEnvironment& to_copy, HInstruction* holder)
       : HEnvironment(arena,
                      to_copy.Size(),
                      to_copy.GetDexFile(),
                      to_copy.GetMethodIdx(),
                      to_copy.GetDexPc(),
-                     to_copy.GetInvokeType()) {}
+                     to_copy.GetInvokeType(),
+                     holder) {}
 
   void SetAndCopyParentChain(ArenaAllocator* allocator, HEnvironment* parent) {
-    parent_ = new (allocator) HEnvironment(allocator, *parent);
-    parent_->CopyFrom(parent);
-    if (parent->GetParent() != nullptr) {
-      parent_->SetAndCopyParentChain(allocator, parent->GetParent());
+    if (parent_ != nullptr) {
+      parent_->SetAndCopyParentChain(allocator, parent);
+    } else {
+      parent_ = new (allocator) HEnvironment(allocator, *parent, holder_);
+      parent_->CopyFrom(parent);
+      if (parent->GetParent() != nullptr) {
+        parent_->SetAndCopyParentChain(allocator, parent->GetParent());
+      }
     }
   }
 
@@ -1202,6 +1223,10 @@
     return dex_file_;
   }
 
+  HInstruction* GetHolder() const {
+    return holder_;
+  }
+
  private:
   // Record instructions' use entries of this environment for constant-time removal.
   // It should only be called by HInstruction when a new environment use is added.
@@ -1219,6 +1244,10 @@
   const uint32_t dex_pc_;
   const InvokeType invoke_type_;
 
+  // The instruction that holds this environment. Only used in debug mode
+  // to ensure the graph is consistent.
+  HInstruction* const holder_;
+
   friend class HInstruction;
 
   DISALLOW_COPY_AND_ASSIGN(HEnvironment);
@@ -1425,13 +1454,18 @@
   HEnvironment* GetEnvironment() const { return environment_; }
   // Set the `environment_` field. Raw because this method does not
   // update the uses lists.
-  void SetRawEnvironment(HEnvironment* environment) { environment_ = environment; }
+  void SetRawEnvironment(HEnvironment* environment) {
+    DCHECK(environment_ == nullptr);
+    DCHECK_EQ(environment->GetHolder(), this);
+    environment_ = environment;
+  }
 
   // Set the environment of this instruction, copying it from `environment`. While
   // copying, the uses lists are being updated.
   void CopyEnvironmentFrom(HEnvironment* environment) {
+    DCHECK(environment_ == nullptr);
     ArenaAllocator* allocator = GetBlock()->GetGraph()->GetArena();
-    environment_ = new (allocator) HEnvironment(allocator, *environment);
+    environment_ = new (allocator) HEnvironment(allocator, *environment, this);
     environment_->CopyFrom(environment);
     if (environment->GetParent() != nullptr) {
       environment_->SetAndCopyParentChain(allocator, environment->GetParent());
@@ -1440,8 +1474,9 @@
 
   void CopyEnvironmentFromWithLoopPhiAdjustment(HEnvironment* environment,
                                                 HBasicBlock* block) {
+    DCHECK(environment_ == nullptr);
     ArenaAllocator* allocator = GetBlock()->GetGraph()->GetArena();
-    environment_ = new (allocator) HEnvironment(allocator, *environment);
+    environment_ = new (allocator) HEnvironment(allocator, *environment, this);
     environment_->CopyFromWithLoopPhiAdjustment(environment, block);
     if (environment->GetParent() != nullptr) {
       environment_->SetAndCopyParentChain(allocator, environment->GetParent());
@@ -2420,6 +2455,12 @@
     intrinsic_ = intrinsic;
   }
 
+  bool IsInlined() const {
+    return GetEnvironment()->GetParent() != nullptr;
+  }
+
+  bool CanThrow() const OVERRIDE { return true; }
+
   DECLARE_INSTRUCTION(Invoke);
 
  protected:
diff --git a/compiler/optimizing/nodes_test.cc b/compiler/optimizing/nodes_test.cc
index 782cde4..fef77aa 100644
--- a/compiler/optimizing/nodes_test.cc
+++ b/compiler/optimizing/nodes_test.cc
@@ -51,7 +51,7 @@
   exit_block->AddInstruction(new (&allocator) HExit());
 
   HEnvironment* environment = new (&allocator) HEnvironment(
-      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic);
+      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic, null_check);
   null_check->SetRawEnvironment(environment);
   environment->SetRawEnvAt(0, parameter);
   parameter->AddEnvUseAt(null_check->GetEnvironment(), 0);
@@ -132,7 +132,7 @@
   ASSERT_TRUE(parameter1->GetUses().HasOnlyOneUse());
 
   HEnvironment* environment = new (&allocator) HEnvironment(
-      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic);
+      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic, with_environment);
   GrowableArray<HInstruction*> array(&allocator, 1);
   array.Add(parameter1);
 
@@ -143,13 +143,13 @@
   ASSERT_TRUE(parameter1->GetEnvUses().HasOnlyOneUse());
 
   HEnvironment* parent1 = new (&allocator) HEnvironment(
-      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic);
+      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic, nullptr);
   parent1->CopyFrom(array);
 
   ASSERT_EQ(parameter1->GetEnvUses().SizeSlow(), 2u);
 
   HEnvironment* parent2 = new (&allocator) HEnvironment(
-      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic);
+      &allocator, 1, graph->GetDexFile(), graph->GetMethodIdx(), 0, kStatic, nullptr);
   parent2->CopyFrom(array);
   parent1->SetAndCopyParentChain(&allocator, parent2);
 
diff --git a/compiler/optimizing/prepare_for_register_allocation.cc b/compiler/optimizing/prepare_for_register_allocation.cc
index 538736b..a249aa9 100644
--- a/compiler/optimizing/prepare_for_register_allocation.cc
+++ b/compiler/optimizing/prepare_for_register_allocation.cc
@@ -88,7 +88,11 @@
 
     // The static call will initialize the class so there's no need for a clinit check if
     // it's the first user.
-    if (last_input == invoke->GetPrevious()) {
+    // There is one special case where we still need the clinit check, when inlining. Because
+    // currently the callee is responsible for reporting parameters to the GC, the code
+    // that walks the stack during `artQuickResolutionTrampoline` cannot be interrupted for GC.
+    // Therefore we cannot allocate any object in that code, including loading a new class.
+    if (last_input == invoke->GetPrevious() && !invoke->IsInlined()) {
       last_input->SetMustGenerateClinitCheck(false);
     }
 
@@ -102,7 +106,7 @@
 
     // If the load class instruction is no longer used, remove it from
     // the graph.
-    if (!last_input->HasUses()) {
+    if (!last_input->HasUses() && !(last_input->MustGenerateClinitCheck() && invoke->IsInlined())) {
       last_input->GetBlock()->RemoveInstruction(last_input);
     }
   }
diff --git a/compiler/optimizing/register_allocator.cc b/compiler/optimizing/register_allocator.cc
index d4ff4d8..9a859bf 100644
--- a/compiler/optimizing/register_allocator.cc
+++ b/compiler/optimizing/register_allocator.cc
@@ -1583,7 +1583,7 @@
       while (env_use != nullptr && env_use->GetPosition() <= range->GetEnd()) {
         DCHECK(current->CoversSlow(env_use->GetPosition())
                || (env_use->GetPosition() == range->GetEnd()));
-        HEnvironment* environment = env_use->GetUser()->GetEnvironment();
+        HEnvironment* environment = env_use->GetEnvironment();
         environment->SetLocationAt(env_use->GetInputIndex(), source);
         env_use = env_use->GetNext();
       }
diff --git a/compiler/optimizing/ssa_builder.cc b/compiler/optimizing/ssa_builder.cc
index c51d248..c4612af 100644
--- a/compiler/optimizing/ssa_builder.cc
+++ b/compiler/optimizing/ssa_builder.cc
@@ -548,7 +548,8 @@
       GetGraph()->GetDexFile(),
       GetGraph()->GetMethodIdx(),
       instruction->GetDexPc(),
-      GetGraph()->GetInvokeType());
+      GetGraph()->GetInvokeType(),
+      instruction);
   environment->CopyFrom(*current_locals_);
   instruction->SetRawEnvironment(environment);
 }
diff --git a/compiler/optimizing/ssa_liveness_analysis.h b/compiler/optimizing/ssa_liveness_analysis.h
index 4b19c5b..4cbe29a 100644
--- a/compiler/optimizing/ssa_liveness_analysis.h
+++ b/compiler/optimizing/ssa_liveness_analysis.h
@@ -117,6 +117,7 @@
         || user->IsPhi()
         || (GetPosition() == user->GetLifetimePosition() + 1)
         || (GetPosition() == user->GetLifetimePosition()));
+    DCHECK(environment == nullptr || user == nullptr);
     DCHECK(next_ == nullptr || next->GetPosition() >= GetPosition());
   }
 
@@ -128,6 +129,7 @@
   void SetNext(UsePosition* next) { next_ = next; }
 
   HInstruction* GetUser() const { return user_; }
+  HEnvironment* GetEnvironment() const { return environment_; }
 
   bool GetIsEnvironment() const { return environment_ != nullptr; }
   bool IsSynthesized() const { return user_ == nullptr; }
@@ -280,7 +282,7 @@
       }
       DCHECK(first_use_->GetPosition() + 1 == position);
       UsePosition* new_use = new (allocator_) UsePosition(
-          instruction, environment, input_index, position, cursor->GetNext());
+          instruction, nullptr /* environment */, input_index, position, cursor->GetNext());
       cursor->SetNext(new_use);
       if (first_range_->GetEnd() == first_use_->GetPosition()) {
         first_range_->end_ = position;
@@ -290,10 +292,10 @@
 
     if (is_environment) {
       first_env_use_ = new (allocator_) UsePosition(
-          instruction, environment, input_index, position, first_env_use_);
+          nullptr /* instruction */, environment, input_index, position, first_env_use_);
     } else {
       first_use_ = new (allocator_) UsePosition(
-          instruction, environment, input_index, position, first_use_);
+          instruction, nullptr /* environment */, input_index, position, first_use_);
     }
 
     if (is_environment && !keep_alive) {