[sparkplug][deoptimizer] Deoptimize to baseline.

If we have baseline code, deoptimize to baseline instead of the
interpreter. The process is similar to deopting to the interpreter.
We just use different builtins
(BaselineEnterAtBytecode/BaselineEnterAtNextBytecode) instead of
InterpreterEnterBytecodeDispatch/InterpreterEnterBytecodeAdvance, that
patch an interpreter frame to a baseline frame and continue execution in
baseline code (based on the deopt type, at the current or next
bytecode).

Bug: v8:11420
Change-Id: Iabaefb36c05155a435c7b380906a86d9b9d549fa
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2695591
Commit-Queue: Patrick Thier <pthier@chromium.org>
Reviewed-by: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#73609}

src/baseline/arm/baseline-assembler-arm-inl.h

@@ -92,6 +92,10 @@ void BaselineAssembler::Bind(Label* label) {
   __ bind(label);
 }
 
+void BaselineAssembler::JumpTarget() {
+  // NOP on arm.
+}
+
 void BaselineAssembler::Jump(Label* target, Label::Distance distance) {
   __ b(target);
 }

src/baseline/arm64/baseline-assembler-arm64-inl.h

@@ -87,6 +87,8 @@ void BaselineAssembler::Bind(Label* label) {
   __ BindJumpTarget(label);
 }
 
+void BaselineAssembler::JumpTarget() { __ JumpTarget(); }
+
 void BaselineAssembler::Jump(Label* target, Label::Distance distance) {
   __ B(target);
 }

src/baseline/baseline-assembler.h

@@ -40,6 +40,9 @@ class BaselineAssembler {
   inline void DebugBreak();
 
   inline void Bind(Label* label);
+  // Marks the current position as a valid jump target on CFI enabled
+  // architectures.
+  inline void JumpTarget();
   inline void JumpIf(Condition cc, Label* target,
                      Label::Distance distance = Label::kFar);
   inline void Jump(Label* target, Label::Distance distance = Label::kFar);

src/baseline/baseline-compiler.cc

@@ -426,6 +426,7 @@ void BaselineCompiler::PreVisitSingleBytecode() {
 
 void BaselineCompiler::VisitSingleBytecode() {
   int offset = iterator().current_offset();
+  bool is_marked_as_jump_target = false;
   if (labels_[offset]) {
     // Bind labels for this offset that have already been linked to a
     // jump (i.e. forward jumps, excluding jump tables).
@@ -436,15 +437,23 @@ void BaselineCompiler::VisitSingleBytecode() {
     labels_[offset]->linked.Clear();
 #endif
     __ Bind(&labels_[offset]->unlinked);
+    is_marked_as_jump_target = true;
   }
 
   // Record positions of exception handlers.
   if (iterator().current_offset() == *next_handler_offset_) {
     __ ExceptionHandler();
     next_handler_offset_++;
+    is_marked_as_jump_target = true;
   }
   DCHECK_LT(iterator().current_offset(), *next_handler_offset_);
 
+  // Mark position as valid jump target, if it isn't one already.
+  // This is required for the deoptimizer, when CFI is enabled.
+  if (!is_marked_as_jump_target) {
+    __ JumpTarget();
+  }
+
   if (FLAG_code_comments) {
     std::ostringstream str;
     str << "[ ";

src/baseline/ia32/baseline-assembler-ia32-inl.h

@@ -90,6 +90,10 @@ MemOperand BaselineAssembler::FeedbackVectorOperand() {
 
 void BaselineAssembler::Bind(Label* label) { __ bind(label); }
 
+void BaselineAssembler::JumpTarget() {
+  // NOP on ia32.
+}
+
 void BaselineAssembler::Jump(Label* target, Label::Distance distance) {
   __ jmp(target, distance);
 }

src/baseline/x64/baseline-assembler-x64-inl.h

@@ -93,6 +93,10 @@ MemOperand BaselineAssembler::FeedbackVectorOperand() {
 
 void BaselineAssembler::Bind(Label* label) { __ bind(label); }
 
+void BaselineAssembler::JumpTarget() {
+  // NOP on x64.
+}
+
 void BaselineAssembler::Jump(Label* target, Label::Distance distance) {
   __ jmp(target, distance);
 }

src/builtins/arm/builtins-arm.cc

@@ -3462,6 +3462,109 @@ void Builtins::Generate_DeoptimizationEntry_Lazy(MacroAssembler* masm) {
   Generate_DeoptimizationEntry(masm, DeoptimizeKind::kLazy);
 }
 
+namespace {
+
+// Converts an interpreter frame into a baseline frame and continues execution
+// in baseline code (baseline code has to exist on the shared function info),
+// either at the start or the end of the current bytecode.
+void Generate_BaselineEntry(MacroAssembler* masm, bool next_bytecode) {
+  // Get bytecode array and bytecode offset from the stack frame.
+  __ ldr(kInterpreterBytecodeArrayRegister,
+         MemOperand(fp, InterpreterFrameConstants::kBytecodeArrayFromFp));
+  __ ldr(kInterpreterBytecodeOffsetRegister,
+         MemOperand(fp, InterpreterFrameConstants::kBytecodeOffsetFromFp));
+  __ SmiUntag(kInterpreterBytecodeOffsetRegister);
+
+  // Get function from the frame.
+  Register closure = r1;
+  __ ldr(closure, MemOperand(fp, StandardFrameConstants::kFunctionOffset));
+
+  // Replace BytecodeOffset with the feedback vector.
+  Register feedback_vector = r2;
+  __ ldr(feedback_vector,
+         FieldMemOperand(closure, JSFunction::kFeedbackCellOffset));
+  __ ldr(feedback_vector, FieldMemOperand(feedback_vector, Cell::kValueOffset));
+  if (__ emit_debug_code()) {
+    Register scratch = r3;
+    __ CompareObjectType(feedback_vector, scratch, scratch,
+                         FEEDBACK_VECTOR_TYPE);
+    __ Assert(eq, AbortReason::kExpectedFeedbackVector);
+  }
+  __ str(feedback_vector,
+         MemOperand(fp, InterpreterFrameConstants::kBytecodeOffsetFromFp));
+  feedback_vector = no_reg;
+
+  // Get the Code object from the shared function info.
+  Register code_obj = r4;
+  __ ldr(code_obj,
+         FieldMemOperand(closure, JSFunction::kSharedFunctionInfoOffset));
+  __ ldr(code_obj,
+         FieldMemOperand(code_obj, SharedFunctionInfo::kFunctionDataOffset));
+  __ ldr(code_obj,
+         FieldMemOperand(code_obj, BaselineData::kBaselineCodeOffset));
+  closure = no_reg;
+
+  // Compute baseline pc for bytecode offset.
+  __ Push(kInterpreterAccumulatorRegister);
+  ExternalReference get_baseline_pc_extref =
+      next_bytecode
+          ? ExternalReference::baseline_end_pc_for_bytecode_offset()
+          : ExternalReference::baseline_start_pc_for_bytecode_offset();
+  Register get_baseline_pc = r3;
+  __ Move(get_baseline_pc, get_baseline_pc_extref);
+
+  // If the code deoptimizes during the implicit function entry stack interrupt
+  // check, it will have a bailout ID of kFunctionEntryBytecodeOffset, which is
+  // not a valid bytecode offset.
+  // TODO(pthier): Investigate if it is feasible to handle this special case
+  // in TurboFan instead of here.
+  Label valid_bytecode_offset, function_entry_bytecode;
+  __ cmp(kInterpreterBytecodeOffsetRegister,
+         Operand(BytecodeArray::kHeaderSize - kHeapObjectTag +
+                 kFunctionEntryBytecodeOffset));
+  __ b(eq, &function_entry_bytecode);
+
+  __ sub(kInterpreterBytecodeOffsetRegister, kInterpreterBytecodeOffsetRegister,
+         Operand(BytecodeArray::kHeaderSize - kHeapObjectTag));
+
+  __ bind(&valid_bytecode_offset);
+  {
+    FrameScope scope(masm, StackFrame::INTERNAL);
+    __ PrepareCallCFunction(3, 0, r0);
+    Register arg_reg_1 = r0;
+    Register arg_reg_2 = r1;
+    Register arg_reg_3 = r2;
+    __ mov(arg_reg_1, code_obj);
+    __ mov(arg_reg_2, kInterpreterBytecodeOffsetRegister);
+    __ mov(arg_reg_3, kInterpreterBytecodeArrayRegister);
+    __ CallCFunction(get_baseline_pc, 3, 0);
+  }
+  __ add(code_obj, code_obj, Operand(Code::kHeaderSize - kHeapObjectTag));
+  __ add(code_obj, code_obj, kReturnRegister0);
+  __ Pop(kInterpreterAccumulatorRegister);
+
+  __ Jump(code_obj);
+  __ Trap();  // Unreachable.
+
+  __ bind(&function_entry_bytecode);
+  // If the bytecode offset is kFunctionEntryOffset, get the start address of
+  // the first bytecode.
+  __ mov(kInterpreterBytecodeOffsetRegister, Operand(0));
+  __ Move(get_baseline_pc,
+          ExternalReference::baseline_start_pc_for_bytecode_offset());
+  __ b(&valid_bytecode_offset);
+}
+
+}  // namespace
+
+void Builtins::Generate_BaselineEnterAtBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, false);
+}
+
+void Builtins::Generate_BaselineEnterAtNextBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, true);
+}
+
 void Builtins::Generate_DynamicCheckMapsTrampoline(MacroAssembler* masm) {
   FrameScope scope(masm, StackFrame::MANUAL);
   __ EnterFrame(StackFrame::INTERNAL);

src/builtins/arm64/builtins-arm64.cc

@@ -3969,6 +3969,111 @@ void Builtins::Generate_DeoptimizationEntry_Lazy(MacroAssembler* masm) {
   Generate_DeoptimizationEntry(masm, DeoptimizeKind::kLazy);
 }
 
+namespace {
+
+// Converts an interpreter frame into a baseline frame and continues execution
+// in baseline code (baseline code has to exist on the shared function info),
+// either at the start or the end of the current bytecode.
+void Generate_BaselineEntry(MacroAssembler* masm, bool next_bytecode) {
+  // Get bytecode array and bytecode offset from the stack frame.
+  __ Ldr(kInterpreterBytecodeArrayRegister,
+         MemOperand(fp, InterpreterFrameConstants::kBytecodeArrayFromFp));
+  __ SmiUntag(kInterpreterBytecodeOffsetRegister,
+              MemOperand(fp, InterpreterFrameConstants::kBytecodeOffsetFromFp));
+
+  // Get function from the frame.
+  Register closure = x1;
+  __ Ldr(closure, MemOperand(fp, StandardFrameConstants::kFunctionOffset));
+
+  // Replace BytecodeOffset with the feedback vector.
+  Register feedback_vector = x2;
+  __ LoadTaggedPointerField(
+      feedback_vector,
+      FieldMemOperand(closure, JSFunction::kFeedbackCellOffset));
+  __ LoadTaggedPointerField(
+      feedback_vector, FieldMemOperand(feedback_vector, Cell::kValueOffset));
+  if (__ emit_debug_code()) {
+    Register scratch = x3;
+    __ CompareObjectType(feedback_vector, scratch, scratch,
+                         FEEDBACK_VECTOR_TYPE);
+    __ Assert(eq, AbortReason::kExpectedFeedbackVector);
+  }
+  __ Str(feedback_vector,
+         MemOperand(fp, InterpreterFrameConstants::kBytecodeOffsetFromFp));
+  feedback_vector = no_reg;
+
+  // Get the Code object from the shared function info.
+  Register code_obj = x22;
+  __ LoadTaggedPointerField(
+      code_obj,
+      FieldMemOperand(closure, JSFunction::kSharedFunctionInfoOffset));
+  __ LoadTaggedPointerField(
+      code_obj,
+      FieldMemOperand(code_obj, SharedFunctionInfo::kFunctionDataOffset));
+  __ LoadTaggedPointerField(
+      code_obj, FieldMemOperand(code_obj, BaselineData::kBaselineCodeOffset));
+  closure = no_reg;
+
+  // Compute baseline pc for bytecode offset.
+  __ Push(padreg, kInterpreterAccumulatorRegister);
+  ExternalReference get_baseline_pc_extref =
+      next_bytecode
+          ? ExternalReference::baseline_end_pc_for_bytecode_offset()
+          : ExternalReference::baseline_start_pc_for_bytecode_offset();
+  Register get_baseline_pc = x3;
+  __ Mov(get_baseline_pc, get_baseline_pc_extref);
+
+  // If the code deoptimizes during the implicit function entry stack interrupt
+  // check, it will have a bailout ID of kFunctionEntryBytecodeOffset, which is
+  // not a valid bytecode offset.
+  // TODO(pthier): Investigate if it is feasible to handle this special case
+  // in TurboFan instead of here.
+  Label valid_bytecode_offset, function_entry_bytecode;
+  __ cmp(kInterpreterBytecodeOffsetRegister,
+         Operand(BytecodeArray::kHeaderSize - kHeapObjectTag +
+                 kFunctionEntryBytecodeOffset));
+  __ B(eq, &function_entry_bytecode);
+
+  __ Sub(kInterpreterBytecodeOffsetRegister, kInterpreterBytecodeOffsetRegister,
+         (BytecodeArray::kHeaderSize - kHeapObjectTag));
+
+  __ bind(&valid_bytecode_offset);
+  {
+    FrameScope scope(masm, StackFrame::INTERNAL);
+    Register arg_reg_1 = x0;
+    Register arg_reg_2 = x1;
+    Register arg_reg_3 = x2;
+    __ Mov(arg_reg_1, code_obj);
+    __ Mov(arg_reg_2, kInterpreterBytecodeOffsetRegister);
+    __ Mov(arg_reg_3, kInterpreterBytecodeArrayRegister);
+    __ CallCFunction(get_baseline_pc, 3, 0);
+  }
+  __ Add(code_obj, code_obj, Code::kHeaderSize - kHeapObjectTag);
+  __ Add(code_obj, code_obj, kReturnRegister0);
+  __ Pop(kInterpreterAccumulatorRegister, padreg);
+
+  __ Jump(code_obj);
+  __ Trap();  // Unreachable.
+
+  __ bind(&function_entry_bytecode);
+  // If the bytecode offset is kFunctionEntryOffset, get the start address of
+  // the first bytecode.
+  __ Mov(kInterpreterBytecodeOffsetRegister, Operand(0));
+  __ Mov(get_baseline_pc,
+         ExternalReference::baseline_start_pc_for_bytecode_offset());
+  __ B(&valid_bytecode_offset);
+}
+
+}  // namespace
+
+void Builtins::Generate_BaselineEnterAtBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, false);
+}
+
+void Builtins::Generate_BaselineEnterAtNextBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, true);
+}
+
 void Builtins::Generate_DynamicCheckMapsTrampoline(MacroAssembler* masm) {
   FrameScope scope(masm, StackFrame::MANUAL);
   __ EnterFrame(StackFrame::INTERNAL);

src/builtins/builtins-definitions.h

@@ -141,6 +141,8 @@ namespace internal {
   ASM(BaselineOutOfLinePrologue, BaselineOutOfLinePrologue)                    \
   ASM(BaselineOnStackReplacement, ContextOnly)                                 \
   ASM(BaselineLeaveFrame, BaselineLeaveFrame)                                  \
+  ASM(BaselineEnterAtBytecode, Void)                                           \
+  ASM(BaselineEnterAtNextBytecode, Void)                                       \
                                                                                \
   /* Code life-cycle */                                                        \
   TFC(CompileLazy, JSTrampoline)                                               \

src/builtins/ia32/builtins-ia32.cc

@@ -4073,6 +4073,108 @@ void Builtins::Generate_DeoptimizationEntry_Lazy(MacroAssembler* masm) {
   Generate_DeoptimizationEntry(masm, DeoptimizeKind::kLazy);
 }
 
+namespace {
+
+// Converts an interpreter frame into a baseline frame and continues execution
+// in baseline code (baseline code has to exist on the shared function info),
+// either at the start or the end of the current bytecode.
+void Generate_BaselineEntry(MacroAssembler* masm, bool next_bytecode) {
+  // Get bytecode array and bytecode offset from the stack frame.
+  __ mov(kInterpreterBytecodeArrayRegister,
+         MemOperand(ebp, InterpreterFrameConstants::kBytecodeArrayFromFp));
+  __ mov(kInterpreterBytecodeOffsetRegister,
+         MemOperand(ebp, InterpreterFrameConstants::kBytecodeOffsetFromFp));
+  __ SmiUntag(kInterpreterBytecodeOffsetRegister);
+
+  // Get function from the frame.
+  Register closure = esi;
+  __ mov(closure, MemOperand(ebp, StandardFrameConstants::kFunctionOffset));
+
+  // Replace BytecodeOffset with the feedback vector.
+  Register feedback_vector = ecx;
+  __ mov(feedback_vector,
+         FieldOperand(closure, JSFunction::kFeedbackCellOffset));
+  __ mov(feedback_vector, FieldOperand(feedback_vector, Cell::kValueOffset));
+  if (__ emit_debug_code()) {
+    __ push(kInterpreterAccumulatorRegister);
+    Register scratch = kInterpreterAccumulatorRegister;
+    __ CmpObjectType(feedback_vector, FEEDBACK_VECTOR_TYPE, scratch);
+    __ Assert(equal, AbortReason::kExpectedFeedbackVector);
+    scratch = no_reg;
+    __ pop(kInterpreterAccumulatorRegister);
+  }
+  __ mov(MemOperand(ebp, InterpreterFrameConstants::kBytecodeOffsetFromFp),
+         feedback_vector);
+  feedback_vector = no_reg;
+
+  // Get the Code object from the shared function info.
+  Register code_obj = closure;
+  __ mov(code_obj,
+         FieldOperand(closure, JSFunction::kSharedFunctionInfoOffset));
+  closure = no_reg;
+  __ mov(code_obj,
+         FieldOperand(code_obj, SharedFunctionInfo::kFunctionDataOffset));
+  __ mov(code_obj, FieldOperand(code_obj, BaselineData::kBaselineCodeOffset));
+
+  // Compute baseline pc for bytecode offset.
+  __ push(kInterpreterAccumulatorRegister);
+  ExternalReference get_baseline_pc_extref =
+      next_bytecode
+          ? ExternalReference::baseline_end_pc_for_bytecode_offset()
+          : ExternalReference::baseline_start_pc_for_bytecode_offset();
+  Register get_baseline_pc = ecx;
+  __ LoadAddress(get_baseline_pc, get_baseline_pc_extref);
+  // If the code deoptimizes during the implicit function entry stack interrupt
+  // check, it will have a bailout ID of kFunctionEntryBytecodeOffset, which is
+  // not a valid bytecode offset.
+  // TODO(pthier): Investigate if it is feasible to handle this special case
+  // in TurboFan instead of here.
+  Label valid_bytecode_offset, function_entry_bytecode;
+  __ cmp(kInterpreterBytecodeOffsetRegister,
+         Immediate(BytecodeArray::kHeaderSize - kHeapObjectTag +
+                   kFunctionEntryBytecodeOffset));
+  __ j(equal, &function_entry_bytecode);
+
+  __ sub(kInterpreterBytecodeOffsetRegister,
+         Immediate(BytecodeArray::kHeaderSize - kHeapObjectTag));
+
+  __ bind(&valid_bytecode_offset);
+  {
+    FrameScope scope(masm, StackFrame::INTERNAL);
+    __ PrepareCallCFunction(3, eax);
+    __ mov(Operand(esp, 0 * kSystemPointerSize), code_obj);
+    __ mov(Operand(esp, 1 * kSystemPointerSize),
+           kInterpreterBytecodeOffsetRegister);
+    __ mov(Operand(esp, 2 * kSystemPointerSize),
+           kInterpreterBytecodeArrayRegister);
+    __ CallCFunction(get_baseline_pc, 3);
+  }
+  __ lea(code_obj,
+         FieldOperand(code_obj, kReturnRegister0, times_1, Code::kHeaderSize));
+  __ pop(kInterpreterAccumulatorRegister);
+
+  __ jmp(code_obj);
+  __ Trap();  // Unreachable.
+
+  __ bind(&function_entry_bytecode);
+  // If the bytecode offset is kFunctionEntryOffset, get the start address of
+  // the first bytecode.
+  __ mov(kInterpreterBytecodeOffsetRegister, Immediate(0));
+  __ LoadAddress(get_baseline_pc,
+                 ExternalReference::baseline_start_pc_for_bytecode_offset());
+  __ jmp(&valid_bytecode_offset);
+}
+
+}  // namespace
+
+void Builtins::Generate_BaselineEnterAtBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, false);
+}
+
+void Builtins::Generate_BaselineEnterAtNextBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, true);
+}
+
 void Builtins::Generate_DynamicCheckMapsTrampoline(MacroAssembler* masm) {
   FrameScope scope(masm, StackFrame::MANUAL);
   __ EnterFrame(StackFrame::INTERNAL);

src/builtins/x64/builtins-x64.cc

@@ -4343,6 +4343,106 @@ void Builtins::Generate_DeoptimizationEntry_Lazy(MacroAssembler* masm) {
   Generate_DeoptimizationEntry(masm, DeoptimizeKind::kLazy);
 }
 
+namespace {
+
+// Converts an interpreter frame into a baseline frame and continues execution
+// in baseline code (baseline code has to exist on the shared function info),
+// either at the start or the end of the current bytecode.
+void Generate_BaselineEntry(MacroAssembler* masm, bool next_bytecode) {
+  // Get bytecode array and bytecode offset from the stack frame.
+  __ movq(kInterpreterBytecodeArrayRegister,
+          MemOperand(rbp, InterpreterFrameConstants::kBytecodeArrayFromFp));
+  __ SmiUntag(
+      kInterpreterBytecodeOffsetRegister,
+      MemOperand(rbp, InterpreterFrameConstants::kBytecodeOffsetFromFp));
+
+  // Get function from the frame.
+  Register closure = rdi;
+  __ movq(closure, MemOperand(rbp, StandardFrameConstants::kFunctionOffset));
+
+  // Replace BytecodeOffset with the feedback vector.
+  Register feedback_vector = rbx;
+  __ LoadTaggedPointerField(
+      feedback_vector, FieldOperand(closure, JSFunction::kFeedbackCellOffset));
+  __ LoadTaggedPointerField(feedback_vector,
+                            FieldOperand(feedback_vector, Cell::kValueOffset));
+  if (__ emit_debug_code()) {
+    __ CmpObjectType(feedback_vector, FEEDBACK_VECTOR_TYPE, kScratchRegister);
+    __ Assert(equal, AbortReason::kExpectedFeedbackVector);
+  }
+  __ movq(MemOperand(rbp, InterpreterFrameConstants::kBytecodeOffsetFromFp),
+          feedback_vector);
+  feedback_vector = no_reg;
+
+  // Get the Code object from the shared function info.
+  Register code_obj = rbx;
+  __ LoadTaggedPointerField(
+      code_obj, FieldOperand(closure, JSFunction::kSharedFunctionInfoOffset));
+  __ LoadTaggedPointerField(
+      code_obj,
+      FieldOperand(code_obj, SharedFunctionInfo::kFunctionDataOffset));
+  __ LoadTaggedPointerField(
+      code_obj, FieldOperand(code_obj, BaselineData::kBaselineCodeOffset));
+  closure = no_reg;
+
+  // Compute baseline pc for bytecode offset.
+  __ pushq(kInterpreterAccumulatorRegister);
+  ExternalReference get_baseline_pc_extref =
+      next_bytecode
+          ? ExternalReference::baseline_end_pc_for_bytecode_offset()
+          : ExternalReference::baseline_start_pc_for_bytecode_offset();
+  Register get_baseline_pc = rax;
+  __ LoadAddress(get_baseline_pc, get_baseline_pc_extref);
+
+  // If the code deoptimizes during the implicit function entry stack interrupt
+  // check, it will have a bailout ID of kFunctionEntryBytecodeOffset, which is
+  // not a valid bytecode offset.
+  // TODO(pthier): Investigate if it is feasible to handle this special case
+  // in TurboFan instead of here.
+  Label valid_bytecode_offset, function_entry_bytecode;
+  __ cmpq(kInterpreterBytecodeOffsetRegister,
+          Immediate(BytecodeArray::kHeaderSize - kHeapObjectTag +
+                    kFunctionEntryBytecodeOffset));
+  __ j(equal, &function_entry_bytecode);
+
+  __ subq(kInterpreterBytecodeOffsetRegister,
+          Immediate(BytecodeArray::kHeaderSize - kHeapObjectTag));
+
+  __ bind(&valid_bytecode_offset);
+  {
+    FrameScope scope(masm, StackFrame::INTERNAL);
+    __ PrepareCallCFunction(3);
+    __ movq(arg_reg_1, code_obj);
+    __ movq(arg_reg_2, kInterpreterBytecodeOffsetRegister);
+    __ movq(arg_reg_3, kInterpreterBytecodeArrayRegister);
+    __ CallCFunction(get_baseline_pc, 3);
+  }
+  __ leaq(code_obj,
+          FieldOperand(code_obj, kReturnRegister0, times_1, Code::kHeaderSize));
+  __ popq(kInterpreterAccumulatorRegister);
+
+  __ jmp(code_obj);
+  __ Trap();  // Unreachable.
+
+  __ bind(&function_entry_bytecode);
+  // If the bytecode offset is kFunctionEntryOffset, get the start address of
+  // the first bytecode.
+  __ movq(kInterpreterBytecodeOffsetRegister, Immediate(0));
+  __ LoadAddress(get_baseline_pc,
+                 ExternalReference::baseline_start_pc_for_bytecode_offset());
+  __ jmp(&valid_bytecode_offset);
+}
+
+}  // namespace
+
+void Builtins::Generate_BaselineEnterAtBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, false);
+}
+
+void Builtins::Generate_BaselineEnterAtNextBytecode(MacroAssembler* masm) {
+  Generate_BaselineEntry(masm, true);
+}
+
 void Builtins::Generate_DynamicCheckMapsTrampoline(MacroAssembler* masm) {
   FrameScope scope(masm, StackFrame::MANUAL);
   __ EnterFrame(StackFrame::INTERNAL);

src/codegen/external-reference.cc

@@ -621,6 +621,35 @@ ExternalReference::address_of_enable_experimental_regexp_engine() {
   return ExternalReference(&FLAG_enable_experimental_regexp_engine);
 }
 
+namespace {
+
+static uintptr_t BaselineStartPCForBytecodeOffset(Address raw_code_obj,
+                                                  int bytecode_offset,
+                                                  Address raw_bytecode_array) {
+  Code code_obj = Code::cast(Object(raw_code_obj));
+  BytecodeArray bytecode_array =
+      BytecodeArray::cast(Object(raw_bytecode_array));
+  return code_obj.GetBaselineStartPCForBytecodeOffset(bytecode_offset,
+                                                      bytecode_array);
+}
+
+static uintptr_t BaselineEndPCForBytecodeOffset(Address raw_code_obj,
+                                                int bytecode_offset,
+                                                Address raw_bytecode_array) {
+  Code code_obj = Code::cast(Object(raw_code_obj));
+  BytecodeArray bytecode_array =
+      BytecodeArray::cast(Object(raw_bytecode_array));
+  return code_obj.GetBaselineEndPCForBytecodeOffset(bytecode_offset,
+                                                    bytecode_array);
+}
+
+}  // namespace
+
+FUNCTION_REFERENCE(baseline_end_pc_for_bytecode_offset,
+                   BaselineEndPCForBytecodeOffset)
+FUNCTION_REFERENCE(baseline_start_pc_for_bytecode_offset,
+                   BaselineStartPCForBytecodeOffset)
+
 ExternalReference ExternalReference::thread_in_wasm_flag_address_address(
     Isolate* isolate) {
   return ExternalReference(isolate->thread_in_wasm_flag_address_address());

src/codegen/external-reference.h

@@ -124,6 +124,8 @@ class StatsCounter;
   V(address_of_wasm_i8x16_splat_0x33, "wasm_i8x16_splat_0x33")                 \
   V(address_of_wasm_i8x16_splat_0x55, "wasm_i8x16_splat_0x55")                 \
   V(address_of_wasm_i16x8_splat_0x0001, "wasm_16x8_splat_0x0001")              \
+  V(baseline_start_pc_for_bytecode_offset, "BaselineStartPCForBytecodeOffset") \
+  V(baseline_end_pc_for_bytecode_offset, "BaselineEndPCForBytecodeOffset")     \
   V(bytecode_size_table_address, "Bytecodes::bytecode_size_table_address")     \
   V(check_object_type, "check_object_type")                                    \
   V(compute_integer_hash, "ComputeSeededHash")                                 \

src/deoptimizer/deoptimizer-cfi-builtins.cc

@@ -14,6 +14,8 @@ void Builtins_ContinueToJavaScriptBuiltinWithResult();
 void Builtins_ContinueToJavaScriptBuiltin();
 void construct_stub_create_deopt_addr();
 void construct_stub_invoke_deopt_addr();
+void Builtins_BaselineEnterAtBytecode();
+void Builtins_BaselineEnterAtNextBytecode();
 typedef void (*function_ptr)();
 }
 
@@ -30,6 +32,8 @@ constexpr function_ptr builtins[] = {
     &Builtins_ContinueToJavaScriptBuiltin,
     &construct_stub_create_deopt_addr,
     &construct_stub_invoke_deopt_addr,
+    &Builtins_BaselineEnterAtBytecode,
+    &Builtins_BaselineEnterAtNextBytecode,
 };
 
 bool Deoptimizer::IsValidReturnAddress(Address address) {

src/deoptimizer/deoptimizer.cc

@@ -991,11 +991,25 @@ void Deoptimizer::DoComputeOutputFrames() {
       stack_guard->real_jslimit() - kStackLimitSlackForDeoptimizationInBytes);
 }
 
+namespace {
+
+// Get the dispatch builtin for unoptimized frames.
+Builtins::Name DispatchBuiltinFor(bool is_baseline, bool advance_bc) {
+  if (is_baseline) {
+    return advance_bc ? Builtins::kBaselineEnterAtNextBytecode
+                      : Builtins::kBaselineEnterAtBytecode;
+  } else {
+    return advance_bc ? Builtins::kInterpreterEnterBytecodeAdvance
+                      : Builtins::kInterpreterEnterBytecodeDispatch;
+  }
+}
+
+}  // namespace
+
 void Deoptimizer::DoComputeUnoptimizedFrame(TranslatedFrame* translated_frame,
                                             int frame_index,
                                             bool goto_catch_handler) {
   SharedFunctionInfo shared = translated_frame->raw_shared_info();
-
   TranslatedFrame::iterator value_iterator = translated_frame->begin();
   const bool is_bottommost = (0 == frame_index);
   const bool is_topmost = (output_count_ - 1 == frame_index);
@@ -1020,15 +1034,10 @@ void Deoptimizer::DoComputeUnoptimizedFrame(TranslatedFrame* translated_frame,
   const uint32_t output_frame_size = frame_info.frame_size_in_bytes();
 
   TranslatedFrame::iterator function_iterator = value_iterator++;
-  if (verbose_tracing_enabled()) {
-    PrintF(trace_scope()->file(), "  translating unoptimized frame ");
-    std::unique_ptr<char[]> name = shared.DebugNameCStr();
-    PrintF(trace_scope()->file(), "%s", name.get());
-    PrintF(trace_scope()->file(),
-           " => bytecode_offset=%d, variable_frame_size=%d, frame_size=%d%s\n",
-           real_bytecode_offset, frame_info.frame_size_in_bytes_without_fixed(),
-           output_frame_size, goto_catch_handler ? " (throw)" : "");
-  }
+
+  BytecodeArray bytecode_array =
+      shared.HasBreakInfo() ? shared.GetDebugInfo().DebugBytecodeArray()
+                            : shared.GetBytecodeArray(isolate());
 
   // Allocate and store the output frame description.
   FrameDescription* output_frame = new (output_frame_size)
@@ -1039,6 +1048,34 @@ void Deoptimizer::DoComputeUnoptimizedFrame(TranslatedFrame* translated_frame,
   CHECK_NULL(output_[frame_index]);
   output_[frame_index] = output_frame;
 
+  // Compute this frame's PC and state.
+  // For interpreted frames, the PC will be a special builtin that
+  // continues the bytecode dispatch. Note that non-topmost and lazy-style
+  // bailout handlers also advance the bytecode offset before dispatch, hence
+  // simulating what normal handlers do upon completion of the operation.
+  // For baseline frames, the PC will be a builtin to convert the interpreter
+  // frame to a baseline frame before continuing execution of baseline code.
+  // We can't directly continue into baseline code, because of CFI.
+  Builtins* builtins = isolate_->builtins();
+  const bool advance_bc =
+      (!is_topmost || (deopt_kind_ == DeoptimizeKind::kLazy)) &&
+      !goto_catch_handler;
+  const bool is_baseline = shared.HasBaselineData();
+  Code dispatch_builtin =
+      builtins->builtin(DispatchBuiltinFor(is_baseline, advance_bc));
+
+  if (verbose_tracing_enabled()) {
+    PrintF(trace_scope()->file(), "  translating %s frame ",
+           is_baseline ? "baseline" : "interpreted");
+    std::unique_ptr<char[]> name = shared.DebugNameCStr();
+    PrintF(trace_scope()->file(), "%s", name.get());
+    PrintF(trace_scope()->file(), " => bytecode_offset=%d, ",
+           real_bytecode_offset);
+    PrintF(trace_scope()->file(), "variable_frame_size=%d, frame_size=%d%s\n",
+           frame_info.frame_size_in_bytes_without_fixed(), output_frame_size,
+           goto_catch_handler ? " (throw)" : "");
+  }
+
   // The top address of the frame is computed from the previous frame's top and
   // this frame's size.
   const intptr_t top_address =
@@ -1145,9 +1182,6 @@ void Deoptimizer::DoComputeUnoptimizedFrame(TranslatedFrame* translated_frame,
   frame_writer.PushRawValue(argc, "actual argument count\n");
 
   // Set the bytecode array pointer.
-  Object bytecode_array = shared.HasBreakInfo()
-                              ? shared.GetDebugInfo().DebugBytecodeArray()
-                              : shared.GetBytecodeArray(isolate());
   frame_writer.PushRawObject(bytecode_array, "bytecode array\n");
 
   // The bytecode offset was mentioned explicitly in the BEGIN_FRAME.
@@ -1237,26 +1271,16 @@ void Deoptimizer::DoComputeUnoptimizedFrame(TranslatedFrame* translated_frame,
   CHECK_EQ(translated_frame->end(), value_iterator);
   CHECK_EQ(0u, frame_writer.top_offset());
 
-  // Compute this frame's PC and state. The PC will be a special builtin that
-  // continues the bytecode dispatch. Note that non-topmost and lazy-style
-  // bailout handlers also advance the bytecode offset before dispatch, hence
-  // simulating what normal handlers do upon completion of the operation.
-  Builtins* builtins = isolate_->builtins();
-  Code dispatch_builtin =
-      (!is_topmost || (deopt_kind_ == DeoptimizeKind::kLazy)) &&
-              !goto_catch_handler
-          ? builtins->builtin(Builtins::kInterpreterEnterBytecodeAdvance)
-          : builtins->builtin(Builtins::kInterpreterEnterBytecodeDispatch);
+  const intptr_t pc =
+      static_cast<intptr_t>(dispatch_builtin.InstructionStart());
   if (is_topmost) {
     // Only the pc of the topmost frame needs to be signed since it is
     // authenticated at the end of the DeoptimizationEntry builtin.
     const intptr_t top_most_pc = PointerAuthentication::SignAndCheckPC(
-        static_cast<intptr_t>(dispatch_builtin.InstructionStart()),
-        frame_writer.frame()->GetTop());
+        pc, frame_writer.frame()->GetTop());
     output_frame->SetPc(top_most_pc);
   } else {
-    output_frame->SetPc(
-        static_cast<intptr_t>(dispatch_builtin.InstructionStart()));
+    output_frame->SetPc(pc);
   }
 
   // Update constant pool.

src/execution/frames.cc

@@ -225,16 +225,13 @@ namespace {
 
 bool IsInterpreterFramePc(Isolate* isolate, Address pc,
                           StackFrame::State* state) {
-  Code interpreter_entry_trampoline =
-      isolate->builtins()->builtin(Builtins::kInterpreterEntryTrampoline);
-  Code interpreter_bytecode_advance =
-      isolate->builtins()->builtin(Builtins::kInterpreterEnterBytecodeAdvance);
-  Code interpreter_bytecode_dispatch =
-      isolate->builtins()->builtin(Builtins::kInterpreterEnterBytecodeDispatch);
-
-  if (interpreter_entry_trampoline.contains(isolate, pc) ||
-      interpreter_bytecode_advance.contains(isolate, pc) ||
-      interpreter_bytecode_dispatch.contains(isolate, pc)) {
+  Builtins::Name builtin_index = InstructionStream::TryLookupCode(isolate, pc);
+  if (builtin_index != Builtins::kNoBuiltinId &&
+      (builtin_index == Builtins::kInterpreterEntryTrampoline ||
+       builtin_index == Builtins::kInterpreterEnterBytecodeAdvance ||
+       builtin_index == Builtins::kInterpreterEnterBytecodeDispatch ||
+       builtin_index == Builtins::kBaselineEnterAtBytecode ||
+       builtin_index == Builtins::kBaselineEnterAtNextBytecode)) {
     return true;
   } else if (FLAG_interpreted_frames_native_stack) {
     intptr_t marker = Memory<intptr_t>(
@@ -251,7 +248,7 @@ bool IsInterpreterFramePc(Isolate* isolate, Address pc,
     } else if (!isolate->heap()->InSpaceSlow(pc, CODE_SPACE)) {
       return false;
     }
-    interpreter_entry_trampoline =
+    Code interpreter_entry_trampoline =
         isolate->heap()->GcSafeFindCodeForInnerPointer(pc);
     return interpreter_entry_trampoline.is_interpreter_trampoline_builtin();
   } else {
@@ -595,7 +592,10 @@ StackFrame::Type StackFrame::ComputeType(const StackFrameIteratorBase* iterator,
       switch (code_obj.kind()) {
         case CodeKind::BUILTIN:
           if (StackFrame::IsTypeMarker(marker)) break;
-          if (code_obj.is_interpreter_trampoline_builtin()) {
+          if (code_obj.is_interpreter_trampoline_builtin() ||
+              // Frames for baseline entry trampolines on the stack are still
+              // interpreted frames.
+              code_obj.is_baseline_trampoline_builtin()) {
             return INTERPRETED;
           }
           if (code_obj.is_baseline_leave_frame_builtin()) {
@@ -1838,8 +1838,8 @@ int BaselineFrame::GetBytecodeOffset() const {
 }
 
 intptr_t BaselineFrame::GetPCForBytecodeOffset(int bytecode_offset) const {
-  return LookupCode().GetBaselinePCForBytecodeOffset(bytecode_offset,
-                                                     GetBytecodeArray());
+  return LookupCode().GetBaselineStartPCForBytecodeOffset(bytecode_offset,
+                                                          GetBytecodeArray());
 }
 
 void BaselineFrame::PatchContext(Context value) {

src/objects/code-inl.h

@@ -371,7 +371,7 @@ CodeKind Code::kind() const {
 int Code::GetBytecodeOffsetForBaselinePC(Address baseline_pc,
                                          BytecodeArray bytecodes) {
   DisallowGarbageCollection no_gc;
-  CHECK(!is_baseline_prologue_builtin());
+  CHECK(!is_baseline_trampoline_builtin());
   if (is_baseline_leave_frame_builtin()) return kFunctionExitBytecodeOffset;
   CHECK_EQ(kind(), CodeKind::BASELINE);
   baseline::BytecodeOffsetIterator offset_iterator(
@@ -382,13 +382,33 @@ int Code::GetBytecodeOffsetForBaselinePC(Address baseline_pc,
 }
 
 uintptr_t Code::GetBaselinePCForBytecodeOffset(int bytecode_offset,
+                                               BytecodeToPCPosition position,
                                                BytecodeArray bytecodes) {
   DisallowGarbageCollection no_gc;
   CHECK_EQ(kind(), CodeKind::BASELINE);
   baseline::BytecodeOffsetIterator offset_iterator(
       ByteArray::cast(bytecode_offset_table()), bytecodes);
   offset_iterator.AdvanceToBytecodeOffset(bytecode_offset);
-  return offset_iterator.current_pc_start_offset();
+  uintptr_t pc = 0;
+  if (position == kPcAtStartOfBytecode) {
+    pc = offset_iterator.current_pc_start_offset();
+  } else {
+    DCHECK_EQ(position, kPcAtEndOfBytecode);
+    pc = offset_iterator.current_pc_end_offset();
+  }
+  return pc;
+}
+
+uintptr_t Code::GetBaselineStartPCForBytecodeOffset(int bytecode_offset,
+                                                    BytecodeArray bytecodes) {
+  return GetBaselinePCForBytecodeOffset(bytecode_offset, kPcAtStartOfBytecode,
+                                        bytecodes);
+}
+
+uintptr_t Code::GetBaselineEndPCForBytecodeOffset(int bytecode_offset,
+                                                  BytecodeArray bytecodes) {
+  return GetBaselinePCForBytecodeOffset(bytecode_offset, kPcAtEndOfBytecode,
+                                        bytecodes);
 }
 
 void Code::initialize_flags(CodeKind kind, bool is_turbofanned, int stack_slots,
@@ -414,12 +434,16 @@ inline bool Code::is_interpreter_trampoline_builtin() const {
           index == Builtins::kInterpreterEnterBytecodeDispatch);
 }
 
-inline bool Code::is_baseline_leave_frame_builtin() const {
-  return builtin_index() == Builtins::kBaselineLeaveFrame;
+inline bool Code::is_baseline_trampoline_builtin() const {
+  const int index = builtin_index();
+  return index != Builtins::kNoBuiltinId &&
+         (index == Builtins::kBaselineOutOfLinePrologue ||
+          index == Builtins::kBaselineEnterAtBytecode ||
+          index == Builtins::kBaselineEnterAtNextBytecode);
 }
 
-inline bool Code::is_baseline_prologue_builtin() const {
-  return builtin_index() == Builtins::kBaselineOutOfLinePrologue;
+inline bool Code::is_baseline_leave_frame_builtin() const {
+  return builtin_index() == Builtins::kBaselineLeaveFrame;
 }
 
 inline bool Code::checks_optimization_marker() const {

src/objects/code.h

@@ -276,7 +276,7 @@ class Code : public HeapObject {
   inline bool is_interpreter_trampoline_builtin() const;
 
   // Testers for baseline builtins.
-  inline bool is_baseline_prologue_builtin() const;
+  inline bool is_baseline_trampoline_builtin() const;
   inline bool is_baseline_leave_frame_builtin() const;
 
   // Tells whether the code checks the optimization marker in the function's
@@ -405,8 +405,12 @@ class Code : public HeapObject {
   static inline void CopyRelocInfoToByteArray(ByteArray dest,
                                               const CodeDesc& desc);
 
-  inline uintptr_t GetBaselinePCForBytecodeOffset(int bytecode_offset,
-                                                  BytecodeArray bytecodes);
+  inline uintptr_t GetBaselineStartPCForBytecodeOffset(int bytecode_offset,
+                                                       BytecodeArray bytecodes);
+
+  inline uintptr_t GetBaselineEndPCForBytecodeOffset(int bytecode_offset,
+                                                     BytecodeArray bytecodes);
+
   inline int GetBytecodeOffsetForBaselinePC(Address baseline_pc,
                                             BytecodeArray bytecodes);
 
@@ -553,6 +557,17 @@ class Code : public HeapObject {
   bool is_promise_rejection() const;
   bool is_exception_caught() const;
 
+  enum BytecodeToPCPosition {
+    kPcAtStartOfBytecode,
+    // End of bytecode equals the start of the next bytecode.
+    // We need it when we deoptimize to the next bytecode (lazy deopt or deopt
+    // of non-topmost frame).
+    kPcAtEndOfBytecode
+  };
+  inline uintptr_t GetBaselinePCForBytecodeOffset(int bytecode_offset,
+                                                  BytecodeToPCPosition position,
+                                                  BytecodeArray bytecodes);
+
   OBJECT_CONSTRUCTORS(Code, HeapObject);
 };