[turbofan] Frame elision for code stubs.

Removed Frame::needs_frame and the function-wide logic using it in
favor of FrameAccessState::has_frame, which can be set on a more
granular level, and driving it block by block.

BUG= v8:4533
LOG=N

Review URL: https://codereview.chromium.org/1775323002

Cr-Commit-Position: refs/heads/master@{#35139}

src/compiler/arm/code-generator-arm.cc

@@ -230,7 +230,8 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
         value_(value),
         scratch0_(scratch0),
         scratch1_(scratch1),
-        mode_(mode) {}
+        mode_(mode),
+        must_save_lr_(!gen->frame_access_state()->has_frame()) {}
 
   void Generate() final {
     if (mode_ > RecordWriteMode::kValueIsPointer) {
@@ -244,7 +245,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
                                              : OMIT_REMEMBERED_SET;
     SaveFPRegsMode const save_fp_mode =
         frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       // We need to save and restore lr if the frame was elided.
       __ Push(lr);
     }
@@ -257,7 +258,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
       __ add(scratch1_, object_, Operand(index_));
     }
     __ CallStub(&stub);
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       __ Pop(lr);
     }
   }
@@ -270,6 +271,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
   Register const scratch0_;
   Register const scratch1_;
   RecordWriteMode const mode_;
+  bool must_save_lr_;
 };
 
 
@@ -386,6 +388,11 @@ Condition FlagsConditionToCondition(FlagsCondition condition) {
     DCHECK_EQ(LeaveCC, i.OutputSBit());           \
   } while (0)
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  __ LeaveFrame(StackFrame::MANUAL);
+}
+
+void CodeGenerator::AssembleSetupStackPointer() {}
 
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
@@ -402,7 +409,7 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ sub(sp, sp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (FLAG_enable_embedded_constant_pool) {
       __ ldr(cp, MemOperand(fp, StandardFrameConstants::kConstantPoolOffset));
     }
@@ -580,7 +587,7 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->has_frame()) {
         __ ldr(i.OutputRegister(), MemOperand(fp, 0));
       } else {
         __ mov(i.OutputRegister(), fp);
@@ -1271,7 +1278,7 @@ void CodeGenerator::AssembleDeoptimizerCall(
 
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (descriptor->IsCFunctionCall()) {
       if (FLAG_enable_embedded_constant_pool) {
         __ Push(lr, fp, pp);
@@ -1286,10 +1293,7 @@ void CodeGenerator::AssemblePrologue() {
     } else {
       __ StubPrologue(info()->GetOutputStackFrameType());
     }
-  } else {
-    frame()->SetElidedFrameSizeInSlots(0);
   }
-  frame_access_state()->SetFrameAccessToDefault();
 
   int stack_shrink_slots = frame()->GetSpillSlotCount();
   if (info()->is_osr()) {
@@ -1359,15 +1363,15 @@ void CodeGenerator::AssembleReturn() {
   }
 
   if (descriptor->IsCFunctionCall()) {
-    __ LeaveFrame(StackFrame::MANUAL);
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrame();
+  } else if (frame_access_state()->has_frame()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
       __ b(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
-      __ LeaveFrame(StackFrame::MANUAL);
+      AssembleDeconstructFrame();
     }
   }
   __ Ret(pop_count);

src/compiler/arm64/code-generator-arm64.cc

@@ -279,7 +279,8 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
         value_(value),
         scratch0_(scratch0),
         scratch1_(scratch1),
-        mode_(mode) {}
+        mode_(mode),
+        must_save_lr_(!gen->frame_access_state()->has_frame()) {}
 
   void Generate() final {
     if (mode_ > RecordWriteMode::kValueIsPointer) {
@@ -293,7 +294,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
                                              : OMIT_REMEMBERED_SET;
     SaveFPRegsMode const save_fp_mode =
         frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       // We need to save and restore lr if the frame was elided.
       __ Push(lr);
     }
@@ -301,7 +302,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
                          remembered_set_action, save_fp_mode);
     __ Add(scratch1_, object_, index_);
     __ CallStub(&stub);
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       __ Pop(lr);
     }
   }
@@ -313,6 +314,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
   Register const scratch0_;
   Register const scratch1_;
   RecordWriteMode const mode_;
+  bool must_save_lr_;
 };
 
 
@@ -466,6 +468,15 @@ Condition FlagsConditionToCondition(FlagsCondition condition) {
     }                                                                       \
   } while (0)
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  const CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
+  if (descriptor->IsCFunctionCall() || descriptor->UseNativeStack()) {
+    __ Mov(csp, fp);
+  } else {
+    __ Mov(jssp, fp);
+  }
+  __ Pop(fp, lr);
+}
 
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
@@ -482,7 +493,7 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ Claim(-sp_slot_delta);
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     __ Ldr(lr, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
     __ Ldr(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
   }
@@ -672,7 +683,7 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       __ mov(i.OutputRegister(), fp);
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->has_frame()) {
         __ ldr(i.OutputRegister(), MemOperand(fp, 0));
       } else {
         __ mov(i.OutputRegister(), fp);
@@ -1492,6 +1503,14 @@ void CodeGenerator::AssembleDeoptimizerCall(
   __ Call(deopt_entry, RelocInfo::RUNTIME_ENTRY);
 }
 
+void CodeGenerator::AssembleSetupStackPointer() {
+  const CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
+  if (descriptor->UseNativeStack() || descriptor->IsCFunctionCall()) {
+    __ SetStackPointer(csp);
+  } else {
+    __ SetStackPointer(jssp);
+  }
+}
 
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
@@ -1499,19 +1518,12 @@ void CodeGenerator::AssemblePrologue() {
     __ AssertCspAligned();
   }
 
-  frame()->AlignFrame(16);
   int stack_shrink_slots = frame()->GetSpillSlotCount();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (descriptor->IsJSFunctionCall()) {
       DCHECK(!descriptor->UseNativeStack());
-      __ SetStackPointer(jssp);
       __ Prologue(this->info()->GeneratePreagedPrologue());
     } else {
-      if (descriptor->UseNativeStack() || descriptor->IsCFunctionCall()) {
-        __ SetStackPointer(csp);
-      } else {
-        __ SetStackPointer(jssp);
-      }
       if (descriptor->IsCFunctionCall()) {
         __ Push(lr, fp);
         __ Mov(fp, masm_.StackPointer());
@@ -1521,15 +1533,8 @@ void CodeGenerator::AssemblePrologue() {
                         frame()->GetTotalFrameSlotCount());
       }
     }
-  } else {
-    if (descriptor->UseNativeStack()) {
-      __ SetStackPointer(csp);
-    } else {
-      __ SetStackPointer(jssp);
-    }
-    frame()->SetElidedFrameSizeInSlots(0);
   }
-  frame_access_state()->SetFrameAccessToDefault();
+
   if (info()->is_osr()) {
     // TurboFan OSR-compiled functions cannot be entered directly.
     __ Abort(kShouldNotDirectlyEnterOsrFunction);
@@ -1590,22 +1595,18 @@ void CodeGenerator::AssembleReturn() {
 
   int pop_count = static_cast<int>(descriptor->StackParameterCount());
   if (descriptor->IsCFunctionCall()) {
-    __ Mov(csp, fp);
-    __ Pop(fp, lr);
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrame();
+  } else if (frame_access_state()->has_frame()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
       __ B(&return_label_);
       return;
     } else {
       __ Bind(&return_label_);
+      AssembleDeconstructFrame();
       if (descriptor->UseNativeStack()) {
-        __ Mov(csp, fp);
         pop_count += (pop_count & 1);  // align
-      } else {
-        __ Mov(jssp, fp);
       }
-      __ Pop(fp, lr);
     }
   } else if (descriptor->UseNativeStack()) {
     pop_count += (pop_count & 1);  // align

src/compiler/code-generator.cc

@@ -56,12 +56,8 @@ CodeGenerator::CodeGenerator(Frame* frame, Linkage* linkage,
   for (int i = 0; i < code->InstructionBlockCount(); ++i) {
     new (&labels_[i]) Label;
   }
-  if (code->ContainsCall()) {
-    frame->MarkNeedsFrame();
-  }
 }
 
-
 Handle<Code> CodeGenerator::GenerateCode() {
   CompilationInfo* info = this->info();
 
@@ -80,10 +76,6 @@ Handle<Code> CodeGenerator::GenerateCode() {
   }
   // Architecture-specific, linkage-specific prologue.
   info->set_prologue_offset(masm()->pc_offset());
-  AssemblePrologue();
-  if (linkage()->GetIncomingDescriptor()->InitializeRootRegister()) {
-    masm()->InitializeRootRegister();
-  }
 
   // Define deoptimization literals for all inlined functions.
   DCHECK_EQ(0u, deoptimization_literals_.size());
@@ -104,6 +96,9 @@ Handle<Code> CodeGenerator::GenerateCode() {
     }
   }
 
+  // Finish the Frame
+  frame()->AlignFrame(kFrameAlignmentInBytes);
+  AssembleSetupStackPointer();
   // Assemble all non-deferred blocks, followed by deferred ones.
   for (int deferred = 0; deferred < 2; ++deferred) {
     for (const InstructionBlock* block : code()->instruction_blocks()) {
@@ -143,9 +138,24 @@ Handle<Code> CodeGenerator::GenerateCode() {
         SNPrintF(buffer, " --");
         masm()->RecordComment(buffer_start);
       }
+
+      frame_access_state()->MarkHasFrame(block->needs_frame());
+
       masm()->bind(GetLabel(current_block_));
+      if (block->must_construct_frame()) {
+        AssemblePrologue();
+        // We need to setup the root register after we assemble the prologue, to
+        // avoid clobbering callee saved registers in case of C linkage and
+        // using the roots.
+        // TODO(mtrofin): investigate how we can avoid doing this repeatedly.
+        if (linkage()->GetIncomingDescriptor()->InitializeRootRegister()) {
+          masm()->InitializeRootRegister();
+        }
+      }
+
       for (int i = block->code_start(); i < block->code_end(); ++i) {
-        AssembleInstruction(code()->InstructionAt(i));
+        Instruction* instr = code()->InstructionAt(i);
+        AssembleInstruction(instr, block);
       }
     }
   }
@@ -290,9 +300,12 @@ bool CodeGenerator::IsMaterializableFromRoot(
   return false;
 }
 
-
-void CodeGenerator::AssembleInstruction(Instruction* instr) {
+void CodeGenerator::AssembleInstruction(Instruction* instr,
+                                        const InstructionBlock* block) {
   AssembleGaps(instr);
+  if (instr->IsJump() && block->must_deconstruct_frame()) {
+    AssembleDeconstructFrame();
+  }
   AssembleSourcePosition(instr);
   // Assemble architecture-specific code for the instruction.
   AssembleArchInstruction(instr);
@@ -761,13 +774,11 @@ DeoptimizationExit* CodeGenerator::AddDeoptimizationExit(
 }
 
 int CodeGenerator::TailCallFrameStackSlotDelta(int stack_param_delta) {
-  CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
-  int spill_slots = frame()->GetSpillSlotCount();
-  bool has_frame = descriptor->IsJSFunctionCall() || spill_slots > 0;
   // Leave the PC on the stack on platforms that have that as part of their ABI
   int pc_slots = V8_TARGET_ARCH_STORES_RETURN_ADDRESS_ON_STACK ? 1 : 0;
-  int sp_slot_delta =
-      has_frame ? (frame()->GetTotalFrameSlotCount() - pc_slots) : 0;
+  int sp_slot_delta = frame_access_state()->has_frame()
+                          ? (frame()->GetTotalFrameSlotCount() - pc_slots)
+                          : 0;
   // Discard only slots that won't be used by new parameters.
   sp_slot_delta += stack_param_delta;
   return sp_slot_delta;

src/compiler/code-generator.h

@@ -85,7 +85,7 @@ class CodeGenerator final : public GapResolver::Assembler {
                                 Heap::RootListIndex* index_return);
 
   // Assemble code for the specified instruction.
-  void AssembleInstruction(Instruction* instr);
+  void AssembleInstruction(Instruction* instr, const InstructionBlock* block);
   void AssembleSourcePosition(Instruction* instr);
   void AssembleGaps(Instruction* instr);
 
@@ -106,6 +106,9 @@ class CodeGenerator final : public GapResolver::Assembler {
   // Generates an architecture-specific, descriptor-specific prologue
   // to set up a stack frame.
   void AssemblePrologue();
+
+  void AssembleSetupStackPointer();
+
   // Generates an architecture-specific, descriptor-specific return sequence
   // to tear down a stack frame.
   void AssembleReturn();
@@ -113,6 +116,8 @@ class CodeGenerator final : public GapResolver::Assembler {
   // Generates code to deconstruct a the caller's frame, including arguments.
   void AssembleDeconstructActivationRecord(int stack_param_delta);
 
+  void AssembleDeconstructFrame();
+
   // Generates code to manipulate the stack in preparation for a tail call.
   void AssemblePrepareTailCall(int stack_param_delta);
 

src/compiler/code-stub-assembler.cc

@@ -734,7 +734,8 @@ Node* CodeStubAssembler::ChangeFloat64ToTagged(Node* value) {
     } else {
       Node* pair = Int32AddWithOverflow(value32, value32);
       Node* overflow = Projection(1, pair);
-      Label if_overflow(this, Label::kDeferred), if_notoverflow(this);
+      Label if_overflow(this, Label::kDeferred), if_notoverflow(this),
+          if_join(this);
       Branch(overflow, &if_overflow, &if_notoverflow);
       Bind(&if_overflow);
       Goto(&if_valueisheapnumber);

src/compiler/frame-elider.cc

@@ -22,8 +22,9 @@ void FrameElider::MarkBlocks() {
   for (InstructionBlock* block : instruction_blocks()) {
     if (block->needs_frame()) continue;
     for (int i = block->code_start(); i < block->code_end(); ++i) {
-      if (InstructionAt(i)->IsCall() ||
-          InstructionAt(i)->opcode() == ArchOpcode::kArchDeoptimize) {
+      const Instruction* instr = InstructionAt(i);
+      if (instr->IsCall() || instr->IsDeoptimizeCall() ||
+          instr->arch_opcode() == ArchOpcode::kArchStackPointer) {
         block->mark_needs_frame();
         break;
       }
@@ -50,6 +51,16 @@ void FrameElider::MarkDeConstruction() {
       for (RpoNumber& succ : block->successors()) {
         if (!InstructionBlockAt(succ)->needs_frame()) {
           DCHECK_EQ(1U, block->SuccessorCount());
+          const Instruction* last =
+              InstructionAt(block->last_instruction_index());
+          if (last->IsThrow() || last->IsTailCall() ||
+              last->IsDeoptimizeCall()) {
+            // We need to keep the frame if we exit the block through any
+            // of these.
+            continue;
+          }
+          // The only cases when we need to deconstruct are ret and jump.
+          DCHECK(last->IsRet() || last->IsJump());
           block->mark_must_deconstruct_frame();
         }
       }

src/compiler/frame.cc

@@ -13,9 +13,7 @@ namespace internal {
 namespace compiler {
 
 Frame::Frame(int fixed_frame_size_in_slots, const CallDescriptor* descriptor)
-    : needs_frame_((descriptor != nullptr) &&
-                   descriptor->RequiresFrameAsIncoming()),
-      frame_slot_count_(fixed_frame_size_in_slots),
+    : frame_slot_count_(fixed_frame_size_in_slots),
       callee_saved_slot_count_(0),
       spill_slot_count_(0),
       allocated_registers_(nullptr),
@@ -34,8 +32,13 @@ int Frame::AlignFrame(int alignment) {
   return delta;
 }
 
+void FrameAccessState::MarkHasFrame(bool state) {
+  has_frame_ = state;
+  SetFrameAccessToDefault();
+}
+
 void FrameAccessState::SetFrameAccessToDefault() {
-  if (frame()->needs_frame() && !FLAG_turbo_sp_frame_access) {
+  if (has_frame() && !FLAG_turbo_sp_frame_access) {
     SetFrameAccessToFP();
   } else {
     SetFrameAccessToSP();
@@ -46,7 +49,6 @@ void FrameAccessState::SetFrameAccessToDefault() {
 FrameOffset FrameAccessState::GetFrameOffset(int spill_slot) const {
   const int frame_offset = FrameSlotToFPOffset(spill_slot);
   if (access_frame_with_fp()) {
-    DCHECK(frame()->needs_frame());
     return FrameOffset::FromFramePointer(frame_offset);
   } else {
     // No frame. Retrieve all parameters relative to stack pointer.

src/compiler/frame.h

@@ -81,26 +81,13 @@ class Frame : public ZoneObject {
   explicit Frame(int fixed_frame_size_in_slots,
                  const CallDescriptor* descriptor);
 
-  inline bool needs_frame() const { return needs_frame_; }
-  inline void MarkNeedsFrame() { needs_frame_ = true; }
-
   inline int GetTotalFrameSlotCount() const { return frame_slot_count_; }
 
-  inline int GetSPToFPSlotCount() const {
-    return GetTotalFrameSlotCount() -
-           StandardFrameConstants::kFixedSlotCountAboveFp;
-  }
   inline int GetSavedCalleeRegisterSlotCount() const {
     return callee_saved_slot_count_;
   }
   inline int GetSpillSlotCount() const { return spill_slot_count_; }
 
-  inline void SetElidedFrameSizeInSlots(int slots) {
-    DCHECK_EQ(0, callee_saved_slot_count_);
-    DCHECK_EQ(0, spill_slot_count_);
-    frame_slot_count_ = slots;
-  }
-
   void SetAllocatedRegisters(BitVector* regs) {
     DCHECK(allocated_registers_ == nullptr);
     allocated_registers_ = regs;
@@ -120,21 +107,18 @@ class Frame : public ZoneObject {
     int alignment_slots = alignment / kPointerSize;
     int delta = alignment_slots - (frame_slot_count_ & (alignment_slots - 1));
     if (delta != alignment_slots) {
-      DCHECK(needs_frame_);
       frame_slot_count_ += delta;
     }
     return delta;
   }
 
   void AllocateSavedCalleeRegisterSlots(int count) {
-    needs_frame_ = true;
     frame_slot_count_ += count;
     callee_saved_slot_count_ += count;
   }
 
   int AllocateSpillSlot(int width) {
     DCHECK_EQ(0, callee_saved_slot_count_);
-    needs_frame_ = true;
     int frame_slot_count_before = frame_slot_count_;
     int slot = AllocateAlignedFrameSlot(width);
     spill_slot_count_ += (frame_slot_count_ - frame_slot_count_before);
@@ -146,7 +130,6 @@ class Frame : public ZoneObject {
   int ReserveSpillSlots(size_t slot_count) {
     DCHECK_EQ(0, callee_saved_slot_count_);
     DCHECK_EQ(0, spill_slot_count_);
-    needs_frame_ = true;
     spill_slot_count_ += static_cast<int>(slot_count);
     frame_slot_count_ += static_cast<int>(slot_count);
     return frame_slot_count_ - 1;
@@ -168,7 +151,6 @@ class Frame : public ZoneObject {
   }
 
  private:
-  bool needs_frame_;
   int frame_slot_count_;
   int callee_saved_slot_count_;
   int spill_slot_count_;
@@ -210,23 +192,33 @@ class FrameOffset {
 class FrameAccessState : public ZoneObject {
  public:
   explicit FrameAccessState(Frame* const frame)
-      : frame_(frame), access_frame_with_fp_(false), sp_delta_(0) {
-    SetFrameAccessToDefault();
-  }
+      : frame_(frame),
+        access_frame_with_fp_(false),
+        sp_delta_(0),
+        has_frame_(false) {}
 
   Frame* frame() const { return frame_; }
+  void MarkHasFrame(bool state);
 
   int sp_delta() const { return sp_delta_; }
   void ClearSPDelta() { sp_delta_ = 0; }
   void IncreaseSPDelta(int amount) { sp_delta_ += amount; }
 
   bool access_frame_with_fp() const { return access_frame_with_fp_; }
+
+  // Regardless of how we access slots on the stack - using sp or fp - do we
+  // have a frame, at the current stage in code generation.
+  bool has_frame() const { return has_frame_; }
+
   void SetFrameAccessToDefault();
   void SetFrameAccessToFP() { access_frame_with_fp_ = true; }
   void SetFrameAccessToSP() { access_frame_with_fp_ = false; }
 
   int GetSPToFPSlotCount() const {
-    return frame_->GetSPToFPSlotCount() + sp_delta();
+    int frame_slot_count =
+        (has_frame() ? frame()->GetTotalFrameSlotCount() : kElidedFrameSlots) -
+        StandardFrameConstants::kFixedSlotCountAboveFp;
+    return frame_slot_count + sp_delta();
   }
   int GetSPToFPOffset() const { return GetSPToFPSlotCount() * kPointerSize; }
 
@@ -240,6 +232,7 @@ class FrameAccessState : public ZoneObject {
   Frame* const frame_;
   bool access_frame_with_fp_;
   int sp_delta_;
+  bool has_frame_;
 };
 }  // namespace compiler
 }  // namespace internal

src/compiler/ia32/code-generator-ia32.cc

@@ -330,6 +330,12 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
     __ bind(&done);                                          \
   } while (false)
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  __ mov(esp, ebp);
+  __ pop(ebp);
+}
+
+void CodeGenerator::AssembleSetupStackPointer() {}
 
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
@@ -346,7 +352,7 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ sub(esp, Immediate(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     __ mov(ebp, MemOperand(ebp, 0));
   }
   frame_access_state()->SetFrameAccessToSP();
@@ -514,7 +520,7 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       __ mov(i.OutputRegister(), ebp);
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->has_frame()) {
         __ mov(i.OutputRegister(), Operand(ebp, 0));
       } else {
         __ mov(i.OutputRegister(), ebp);
@@ -1607,7 +1613,7 @@ void CodeGenerator::AssembleDeoptimizerCall(
 
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (descriptor->IsCFunctionCall()) {
       __ push(ebp);
       __ mov(ebp, esp);
@@ -1616,11 +1622,7 @@ void CodeGenerator::AssemblePrologue() {
     } else {
       __ StubPrologue(info()->GetOutputStackFrameType());
     }
-  } else {
-    frame()->SetElidedFrameSizeInSlots(kPCOnStackSize / kPointerSize);
   }
-  frame_access_state()->SetFrameAccessToDefault();
-
   int stack_shrink_slots = frame()->GetSpillSlotCount();
   if (info()->is_osr()) {
     // TurboFan OSR-compiled functions cannot be entered directly.
@@ -1666,17 +1668,15 @@ void CodeGenerator::AssembleReturn() {
   }
 
   if (descriptor->IsCFunctionCall()) {
-    __ mov(esp, ebp);  // Move stack pointer back to frame pointer.
-    __ pop(ebp);       // Pop caller's frame pointer.
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrame();
+  } else if (frame_access_state()->has_frame()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
       __ jmp(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
-      __ mov(esp, ebp);  // Move stack pointer back to frame pointer.
-      __ pop(ebp);       // Pop caller's frame pointer.
+      AssembleDeconstructFrame();
     }
   }
   size_t pop_size = descriptor->StackParameterCount() * kPointerSize;

src/compiler/instruction-selector.cc

@@ -1190,7 +1190,6 @@ void InstructionSelector::VisitLoadStackPointer(Node* node) {
 
 void InstructionSelector::VisitLoadFramePointer(Node* node) {
   OperandGenerator g(this);
-  frame_->MarkNeedsFrame();
   Emit(kArchFramePointer, g.DefineAsRegister(node));
 }
 

src/compiler/instruction.h

@@ -812,6 +812,23 @@ class Instruction final {
            OutputCount() == 0 && TempCount() == 0;
   }
 
+  bool IsDeoptimizeCall() const {
+    return arch_opcode() == ArchOpcode::kArchDeoptimize ||
+           FlagsModeField::decode(opcode()) == kFlags_deoptimize;
+  }
+
+  bool IsJump() const { return arch_opcode() == ArchOpcode::kArchJmp; }
+  bool IsRet() const { return arch_opcode() == ArchOpcode::kArchRet; }
+  bool IsTailCall() const {
+    return arch_opcode() == ArchOpcode::kArchTailCallCodeObject ||
+           arch_opcode() == ArchOpcode::kArchTailCallCodeObjectFromJSFunction ||
+           arch_opcode() == ArchOpcode::kArchTailCallJSFunction ||
+           arch_opcode() == ArchOpcode::kArchTailCallJSFunctionFromJSFunction;
+  }
+  bool IsThrow() const {
+    return arch_opcode() == ArchOpcode::kArchThrowTerminator;
+  }
+
   enum GapPosition {
     START,
     END,

src/compiler/jump-threading.cc

@@ -96,7 +96,9 @@ bool JumpThreading::ComputeForwarding(Zone* local_zone,
           // the frame at start. So we should move the decision of whether
           // to build a frame or not in the register allocator, and trickle it
           // here and to the code generator.
-          if (frame_at_start || !block->must_deconstruct_frame()) {
+          if (frame_at_start ||
+              !(block->must_deconstruct_frame() ||
+                block->must_construct_frame())) {
             fw = code->InputRpo(instr, 0);
           }
           fallthru = false;

src/compiler/mips/code-generator-mips.cc

@@ -224,7 +224,8 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
         value_(value),
         scratch0_(scratch0),
         scratch1_(scratch1),
-        mode_(mode) {}
+        mode_(mode),
+        must_save_lr_(!gen->frame_access_state()->has_frame()) {}
 
   void Generate() final {
     if (mode_ > RecordWriteMode::kValueIsPointer) {
@@ -238,7 +239,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
                                              : OMIT_REMEMBERED_SET;
     SaveFPRegsMode const save_fp_mode =
         frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       // We need to save and restore ra if the frame was elided.
       __ Push(ra);
     }
@@ -246,7 +247,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
                          remembered_set_action, save_fp_mode);
     __ Addu(scratch1_, object_, index_);
     __ CallStub(&stub);
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       __ Pop(ra);
     }
   }
@@ -258,6 +259,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
   Register const scratch0_;
   Register const scratch1_;
   RecordWriteMode const mode_;
+  bool must_save_lr_;
 };
 
 
@@ -470,6 +472,13 @@ FPUCondition FlagsConditionToConditionCmpFPU(bool& predicate,
     __ bind(&done);                                                           \
   }
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  __ mov(sp, fp);
+  __ Pop(ra, fp);
+}
+
+void CodeGenerator::AssembleSetupStackPointer() {}
+
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
   if (sp_slot_delta > 0) {
@@ -485,7 +494,7 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ Subu(sp, sp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     __ lw(ra, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
     __ lw(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
   }
@@ -645,7 +654,7 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       __ mov(i.OutputRegister(), fp);
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->has_frame()) {
         __ lw(i.OutputRegister(), MemOperand(fp, 0));
       } else {
         __ mov(i.OutputRegister(), fp);
@@ -1564,7 +1573,7 @@ void CodeGenerator::AssembleDeoptimizerCall(
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
   int stack_shrink_slots = frame()->GetSpillSlotCount();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (descriptor->IsCFunctionCall()) {
       __ Push(ra, fp);
       __ mov(fp, sp);
@@ -1573,10 +1582,7 @@ void CodeGenerator::AssemblePrologue() {
     } else {
       __ StubPrologue(info()->GetOutputStackFrameType());
     }
-  } else {
-    frame()->SetElidedFrameSizeInSlots(0);
   }
-  frame_access_state()->SetFrameAccessToDefault();
 
   if (info()->is_osr()) {
     // TurboFan OSR-compiled functions cannot be entered directly.
@@ -1638,17 +1644,15 @@ void CodeGenerator::AssembleReturn() {
   }
 
   if (descriptor->IsCFunctionCall()) {
-    __ mov(sp, fp);
-    __ Pop(ra, fp);
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrame();
+  } else if (frame_access_state()->has_frame()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
       __ Branch(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
-      __ mov(sp, fp);
-      __ Pop(ra, fp);
+      AssembleDeconstructFrame();
     }
   }
   if (pop_count != 0) {

src/compiler/mips64/code-generator-mips64.cc

@@ -224,7 +224,8 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
         value_(value),
         scratch0_(scratch0),
         scratch1_(scratch1),
-        mode_(mode) {}
+        mode_(mode),
+        must_save_lr_(gen->frame_access_state()->has_frame()) {}
 
   void Generate() final {
     if (mode_ > RecordWriteMode::kValueIsPointer) {
@@ -238,7 +239,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
                                              : OMIT_REMEMBERED_SET;
     SaveFPRegsMode const save_fp_mode =
         frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       // We need to save and restore ra if the frame was elided.
       __ Push(ra);
     }
@@ -246,7 +247,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
                          remembered_set_action, save_fp_mode);
     __ Daddu(scratch1_, object_, index_);
     __ CallStub(&stub);
-    if (!frame()->needs_frame()) {
+    if (must_save_lr_) {
       __ Pop(ra);
     }
   }
@@ -258,6 +259,7 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
   Register const scratch0_;
   Register const scratch1_;
   RecordWriteMode const mode_;
+  bool must_save_lr_;
 };
 
 
@@ -482,6 +484,13 @@ FPUCondition FlagsConditionToConditionCmpFPU(bool& predicate,
     __ bind(&done);                                                           \
   }
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  __ mov(sp, fp);
+  __ Pop(ra, fp);
+}
+
+void CodeGenerator::AssembleSetupStackPointer() {}
+
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
   if (sp_slot_delta > 0) {
@@ -497,7 +506,7 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ Dsubu(sp, sp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     __ ld(ra, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
     __ ld(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
   }
@@ -655,7 +664,7 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       __ mov(i.OutputRegister(), fp);
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->has_frame()) {
         __ ld(i.OutputRegister(), MemOperand(fp, 0));
       } else {
         __ mov(i.OutputRegister(), fp);
@@ -1870,7 +1879,7 @@ void CodeGenerator::AssembleDeoptimizerCall(
 
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (descriptor->IsCFunctionCall()) {
       __ Push(ra, fp);
       __ mov(fp, sp);
@@ -1879,10 +1888,7 @@ void CodeGenerator::AssemblePrologue() {
     } else {
       __ StubPrologue(info()->GetOutputStackFrameType());
     }
-  } else {
-    frame()->SetElidedFrameSizeInSlots(0);
   }
-  frame_access_state()->SetFrameAccessToDefault();
 
   int stack_shrink_slots = frame()->GetSpillSlotCount();
   if (info()->is_osr()) {
@@ -1941,17 +1947,15 @@ void CodeGenerator::AssembleReturn() {
   }
 
   if (descriptor->IsCFunctionCall()) {
-    __ mov(sp, fp);
-    __ Pop(ra, fp);
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrame();
+  } else if (frame_access_state()->has_frame()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
       __ Branch(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
-      __ mov(sp, fp);
-      __ Pop(ra, fp);
+      AssembleDeconstructFrame();
     }
   }
   int pop_count = static_cast<int>(descriptor->StackParameterCount());

src/compiler/pipeline.cc

@@ -276,6 +276,9 @@ class PipelineData {
         info()->isolate(), instruction_zone(), instruction_blocks);
     if (descriptor && descriptor->RequiresFrameAsIncoming()) {
       sequence_->instruction_blocks()[0]->mark_needs_frame();
+    } else {
+      DCHECK_EQ(0, descriptor->CalleeSavedFPRegisters());
+      DCHECK_EQ(0, descriptor->CalleeSavedRegisters());
     }
   }
 
@@ -1370,6 +1373,7 @@ Handle<Code> Pipeline::ScheduleAndGenerateCode(
   AllocateRegisters(
       RegisterConfiguration::ArchDefault(RegisterConfiguration::TURBOFAN),
       call_descriptor, run_verifier);
+  Run<FrameElisionPhase>();
   if (data->compilation_failed()) {
     info()->AbortOptimization(kNotEnoughVirtualRegistersRegalloc);
     return Handle<Code>();
@@ -1378,11 +1382,7 @@ Handle<Code> Pipeline::ScheduleAndGenerateCode(
   BeginPhaseKind("code generation");
   // TODO(mtrofin): move this off to the register allocator.
   bool generate_frame_at_start =
-      !FLAG_turbo_frame_elision || !data_->info()->IsStub() ||
-      !data_->frame()->needs_frame() ||
-      data_->sequence()->instruction_blocks().front()->needs_frame() ||
-      linkage.GetIncomingDescriptor()->CalleeSavedFPRegisters() != 0 ||
-      linkage.GetIncomingDescriptor()->CalleeSavedRegisters() != 0;
+      data_->sequence()->instruction_blocks().front()->must_construct_frame();
   // Optimimize jumps.
   if (FLAG_turbo_jt) {
     Run<JumpThreadingPhase>(generate_frame_at_start);
@@ -1502,7 +1502,6 @@ void Pipeline::AllocateRegisters(const RegisterConfiguration* config,
   }
 
   Run<LocateSpillSlotsPhase>();
-  Run<FrameElisionPhase>();
 
   if (FLAG_trace_turbo_graph) {
     OFStream os(stdout);

src/compiler/x64/code-generator-x64.cc

@@ -602,6 +602,12 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
     }                                                            \
   } while (false)
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  __ movq(rsp, rbp);
+  __ popq(rbp);
+}
+
+void CodeGenerator::AssembleSetupStackPointer() {}
 
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
@@ -618,7 +624,7 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ subq(rsp, Immediate(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     __ movq(rbp, MemOperand(rbp, 0));
   }
   frame_access_state()->SetFrameAccessToSP();
@@ -775,7 +781,7 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       __ movq(i.OutputRegister(), rbp);
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->has_frame()) {
         __ movq(i.OutputRegister(), Operand(rbp, 0));
       } else {
         __ movq(i.OutputRegister(), rbp);
@@ -1929,7 +1935,7 @@ static const int kQuadWordSize = 16;
 
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (descriptor->IsCFunctionCall()) {
       __ pushq(rbp);
       __ movq(rbp, rsp);
@@ -1938,11 +1944,7 @@ void CodeGenerator::AssemblePrologue() {
     } else {
       __ StubPrologue(info()->GetOutputStackFrameType());
     }
-  } else {
-    frame()->SetElidedFrameSizeInSlots(kPCOnStackSize / kPointerSize);
   }
-  frame_access_state()->SetFrameAccessToDefault();
-
   int stack_shrink_slots = frame()->GetSpillSlotCount();
   if (info()->is_osr()) {
     // TurboFan OSR-compiled functions cannot be entered directly.
@@ -2022,17 +2024,15 @@ void CodeGenerator::AssembleReturn() {
   }
 
   if (descriptor->IsCFunctionCall()) {
-    __ movq(rsp, rbp);  // Move stack pointer back to frame pointer.
-    __ popq(rbp);       // Pop caller's frame pointer.
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrame();
+  } else if (frame_access_state()->has_frame()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
       __ jmp(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
-      __ movq(rsp, rbp);  // Move stack pointer back to frame pointer.
-      __ popq(rbp);       // Pop caller's frame pointer.
+      AssembleDeconstructFrame();
     }
   }
   size_t pop_size = descriptor->StackParameterCount() * kPointerSize;

src/globals.h

@@ -124,6 +124,11 @@ const int kFloatSize     = sizeof(float);     // NOLINT
 const int kDoubleSize    = sizeof(double);    // NOLINT
 const int kIntptrSize    = sizeof(intptr_t);  // NOLINT
 const int kPointerSize   = sizeof(void*);     // NOLINT
+#if V8_TARGET_ARCH_ARM64
+const int kFrameAlignmentInBytes = 2 * kPointerSize;
+#else
+const int kFrameAlignmentInBytes = kPointerSize;
+#endif
 #if V8_TARGET_ARCH_X64 && V8_TARGET_ARCH_32_BIT
 const int kRegisterSize  = kPointerSize + kPointerSize;
 #else
@@ -132,6 +137,12 @@ const int kRegisterSize  = kPointerSize;
 const int kPCOnStackSize = kRegisterSize;
 const int kFPOnStackSize = kRegisterSize;
 
+#if V8_TARGET_ARCH_X64 || V8_TARGET_ARCH_IA32
+const int kElidedFrameSlots = kPCOnStackSize / kPointerSize;
+#else
+const int kElidedFrameSlots = 0;
+#endif
+
 const int kDoubleSizeLog2 = 3;
 
 #if V8_HOST_ARCH_64_BIT