X87: [crankshaft] Support ES6 tail call elimination.

  port 22938040 (r34542)

  original commit message:
  HInvokeFunction and HApplyArguments instructions now support tail calling.

  Inlining of calls at tail position is not supported yet and therefore still disabled.

  The tail-call-megatest was modified so that the usages of "arguments" object do not disable Crankshaft.

BUG=

Review URL: https://codereview.chromium.org/1767343003

Cr-Commit-Position: refs/heads/master@{#34590}

src/crankshaft/x87/lithium-codegen-x87.cc

@@ -3207,13 +3207,24 @@ void LCodeGen::DoApplyArguments(LApplyArguments* instr) {
 
   // Invoke the function.
   __ bind(&invoke);
+
+  InvokeFlag flag = CALL_FUNCTION;
+  if (instr->hydrogen()->tail_call_mode() == TailCallMode::kAllow) {
+    // TODO(ishell): drop current frame before pushing arguments to the stack.
+    flag = JUMP_FUNCTION;
+    ParameterCount actual(eax);
+    // It is safe to use ebx, ecx and edx as scratch registers here given that
+    // 1) we are not going to return to caller function anyway,
+    // 2) ebx (expected arguments count) and edx (new.target) will be
+    //    initialized below.
+    PrepareForTailCall(actual, ebx, ecx, edx);
+  }
+
   DCHECK(instr->HasPointerMap());
   LPointerMap* pointers = instr->pointer_map();
-  SafepointGenerator safepoint_generator(
-      this, pointers, Safepoint::kLazyDeopt);
+  SafepointGenerator safepoint_generator(this, pointers, Safepoint::kLazyDeopt);
   ParameterCount actual(eax);
-  __ InvokeFunction(function, no_reg, actual, CALL_FUNCTION,
-                    safepoint_generator);
+  __ InvokeFunction(function, no_reg, actual, flag, safepoint_generator);
 }
 
 
@@ -3257,10 +3268,9 @@ void LCodeGen::DoDeclareGlobals(LDeclareGlobals* instr) {
   CallRuntime(Runtime::kDeclareGlobals, instr);
 }
 
-
 void LCodeGen::CallKnownFunction(Handle<JSFunction> function,
                                  int formal_parameter_count, int arity,
-                                 LInstruction* instr) {
+                                 bool is_tail_call, LInstruction* instr) {
   bool dont_adapt_arguments =
       formal_parameter_count == SharedFunctionInfo::kDontAdaptArgumentsSentinel;
   bool can_invoke_directly =
@@ -3276,21 +3286,38 @@ void LCodeGen::CallKnownFunction(Handle<JSFunction> function,
     __ mov(edx, factory()->undefined_value());
     __ mov(eax, arity);
 
+    bool is_self_call = function.is_identical_to(info()->closure());
+
     // Invoke function directly.
-    if (function.is_identical_to(info()->closure())) {
-      __ CallSelf();
+    if (is_self_call) {
+      Handle<Code> self(reinterpret_cast<Code**>(__ CodeObject().location()));
+      if (is_tail_call) {
+        __ Jump(self, RelocInfo::CODE_TARGET);
+      } else {
+        __ Call(self, RelocInfo::CODE_TARGET);
+      }
     } else {
-      __ call(FieldOperand(function_reg, JSFunction::kCodeEntryOffset));
+      Operand target = FieldOperand(function_reg, JSFunction::kCodeEntryOffset);
+      if (is_tail_call) {
+        __ jmp(target);
+      } else {
+        __ call(target);
+      }
+    }
+
+    if (!is_tail_call) {
+      // Set up deoptimization.
+      RecordSafepointWithLazyDeopt(instr, RECORD_SIMPLE_SAFEPOINT);
     }
-    RecordSafepointWithLazyDeopt(instr, RECORD_SIMPLE_SAFEPOINT);
   } else {
     // We need to adapt arguments.
     LPointerMap* pointers = instr->pointer_map();
     SafepointGenerator generator(
         this, pointers, Safepoint::kLazyDeopt);
-    ParameterCount count(arity);
+    ParameterCount actual(arity);
     ParameterCount expected(formal_parameter_count);
-    __ InvokeFunction(function_reg, expected, count, CALL_FUNCTION, generator);
+    InvokeFlag flag = is_tail_call ? JUMP_FUNCTION : CALL_FUNCTION;
+    __ InvokeFunction(function_reg, expected, actual, flag, generator);
   }
 }
 
@@ -3764,23 +3791,77 @@ void LCodeGen::DoMathExp(LMathExp* instr) {
   X87CommitWrite(result_reg);
 }
 
+void LCodeGen::PrepareForTailCall(const ParameterCount& actual,
+                                  Register scratch1, Register scratch2,
+                                  Register scratch3) {
+#if DEBUG
+  if (actual.is_reg()) {
+    DCHECK(!AreAliased(actual.reg(), scratch1, scratch2, scratch3));
+  } else {
+    DCHECK(!AreAliased(scratch1, scratch2, scratch3));
+  }
+#endif
+  if (FLAG_code_comments) {
+    if (actual.is_reg()) {
+      Comment(";;; PrepareForTailCall, actual: %s {", actual.reg().ToString());
+    } else {
+      Comment(";;; PrepareForTailCall, actual: %d {", actual.immediate());
+    }
+  }
+
+  // Check if next frame is an arguments adaptor frame.
+  Register caller_args_count_reg = scratch1;
+  Label no_arguments_adaptor, formal_parameter_count_loaded;
+  __ mov(scratch2, Operand(ebp, StandardFrameConstants::kCallerFPOffset));
+  __ cmp(Operand(scratch2, StandardFrameConstants::kContextOffset),
+         Immediate(Smi::FromInt(StackFrame::ARGUMENTS_ADAPTOR)));
+  __ j(not_equal, &no_arguments_adaptor, Label::kNear);
+
+  // Drop current frame and load arguments count from arguments adaptor frame.
+  __ mov(ebp, scratch2);
+  __ mov(caller_args_count_reg,
+         Operand(ebp, ArgumentsAdaptorFrameConstants::kLengthOffset));
+  __ SmiUntag(caller_args_count_reg);
+  __ jmp(&formal_parameter_count_loaded, Label::kNear);
+
+  __ bind(&no_arguments_adaptor);
+  // Load caller's formal parameter count.
+  __ mov(caller_args_count_reg,
+         Immediate(info()->literal()->parameter_count()));
+
+  __ bind(&formal_parameter_count_loaded);
+  __ PrepareForTailCall(actual, caller_args_count_reg, scratch2, scratch3,
+                        ReturnAddressState::kNotOnStack);
+  Comment(";;; }");
+}
 
 void LCodeGen::DoInvokeFunction(LInvokeFunction* instr) {
+  HInvokeFunction* hinstr = instr->hydrogen();
   DCHECK(ToRegister(instr->context()).is(esi));
   DCHECK(ToRegister(instr->function()).is(edi));
   DCHECK(instr->HasPointerMap());
 
-  Handle<JSFunction> known_function = instr->hydrogen()->known_function();
+  bool is_tail_call = hinstr->tail_call_mode() == TailCallMode::kAllow;
+
+  if (is_tail_call) {
+    ParameterCount actual(instr->arity());
+    // It is safe to use ebx, ecx and edx as scratch registers here given that
+    // 1) we are not going to return to caller function anyway,
+    // 2) ebx (expected arguments count) and edx (new.target) will be
+    //    initialized below.
+    PrepareForTailCall(actual, ebx, ecx, edx);
+  }
+
+  Handle<JSFunction> known_function = hinstr->known_function();
   if (known_function.is_null()) {
     LPointerMap* pointers = instr->pointer_map();
-    SafepointGenerator generator(
-        this, pointers, Safepoint::kLazyDeopt);
-    ParameterCount count(instr->arity());
-    __ InvokeFunction(edi, no_reg, count, CALL_FUNCTION, generator);
+    SafepointGenerator generator(this, pointers, Safepoint::kLazyDeopt);
+    ParameterCount actual(instr->arity());
+    InvokeFlag flag = is_tail_call ? JUMP_FUNCTION : CALL_FUNCTION;
+    __ InvokeFunction(edi, no_reg, actual, flag, generator);
   } else {
-    CallKnownFunction(known_function,
-                      instr->hydrogen()->formal_parameter_count(),
-                      instr->arity(), instr);
+    CallKnownFunction(known_function, hinstr->formal_parameter_count(),
+                      instr->arity(), is_tail_call, instr);
   }
 }
 

src/crankshaft/x87/lithium-codegen-x87.h

@@ -219,11 +219,14 @@ class LCodeGen: public LCodeGenBase {
 
   void LoadContextFromDeferred(LOperand* context);
 
-  // Generate a direct call to a known function.  Expects the function
+  void PrepareForTailCall(const ParameterCount& actual, Register scratch1,
+                          Register scratch2, Register scratch3);
+
+  // Generate a direct call to a known function. Expects the function
   // to be in edi.
   void CallKnownFunction(Handle<JSFunction> function,
                          int formal_parameter_count, int arity,
-                         LInstruction* instr);
+                         bool is_tail_call, LInstruction* instr);
 
   void RecordSafepointWithLazyDeopt(LInstruction* instr,
                                     SafepointMode safepoint_mode);

src/crankshaft/x87/lithium-x87.h

@@ -553,6 +553,7 @@ class LApplyArguments final : public LTemplateInstruction<1, 4, 0> {
   LOperand* elements() { return inputs_[3]; }
 
   DECLARE_CONCRETE_INSTRUCTION(ApplyArguments, "apply-arguments")
+  DECLARE_HYDROGEN_ACCESSOR(ApplyArguments)
 };
 
 

src/x87/builtins-x87.cc

@@ -1907,16 +1907,18 @@ void PrepareForTailCall(MacroAssembler* masm, Register args_reg,
   }
 
   // Check if next frame is an arguments adaptor frame.
+  Register caller_args_count_reg = scratch1;
   Label no_arguments_adaptor, formal_parameter_count_loaded;
   __ mov(scratch2, Operand(ebp, StandardFrameConstants::kCallerFPOffset));
   __ cmp(Operand(scratch2, StandardFrameConstants::kContextOffset),
          Immediate(Smi::FromInt(StackFrame::ARGUMENTS_ADAPTOR)));
   __ j(not_equal, &no_arguments_adaptor, Label::kNear);
 
-  // Drop arguments adaptor frame and load arguments count.
+  // Drop current frame and load arguments count from arguments adaptor frame.
   __ mov(ebp, scratch2);
-  __ mov(scratch1, Operand(ebp, ArgumentsAdaptorFrameConstants::kLengthOffset));
-  __ SmiUntag(scratch1);
+  __ mov(caller_args_count_reg,
+         Operand(ebp, ArgumentsAdaptorFrameConstants::kLengthOffset));
+  __ SmiUntag(caller_args_count_reg);
   __ jmp(&formal_parameter_count_loaded, Label::kNear);
 
   __ bind(&no_arguments_adaptor);
@@ -1925,57 +1927,15 @@ void PrepareForTailCall(MacroAssembler* masm, Register args_reg,
   __ mov(scratch1,
          FieldOperand(scratch1, JSFunction::kSharedFunctionInfoOffset));
   __ mov(
-      scratch1,
+      caller_args_count_reg,
       FieldOperand(scratch1, SharedFunctionInfo::kFormalParameterCountOffset));
-  __ SmiUntag(scratch1);
+  __ SmiUntag(caller_args_count_reg);
 
   __ bind(&formal_parameter_count_loaded);
 
-  // Calculate the destination address where we will put the return address
-  // after we drop current frame.
-  Register new_sp_reg = scratch2;
-  __ sub(scratch1, args_reg);
-  __ lea(new_sp_reg, Operand(ebp, scratch1, times_pointer_size,
-                             StandardFrameConstants::kCallerPCOffset));
-
-  if (FLAG_debug_code) {
-    __ cmp(esp, new_sp_reg);
-    __ Check(below, kStackAccessBelowStackPointer);
-  }
-
-  // Copy receiver and return address as well.
-  Register count_reg = scratch1;
-  __ lea(count_reg, Operand(args_reg, 2));
-
-  // Copy return address from caller's frame to current frame's return address
-  // to avoid its trashing and let the following loop copy it to the right
-  // place.
-  Register tmp_reg = scratch3;
-  __ mov(tmp_reg, Operand(ebp, StandardFrameConstants::kCallerPCOffset));
-  __ mov(Operand(esp, 0), tmp_reg);
-
-  // Restore caller's frame pointer now as it could be overwritten by
-  // the copying loop.
-  __ mov(ebp, Operand(ebp, StandardFrameConstants::kCallerFPOffset));
-
-  Operand src(esp, count_reg, times_pointer_size, 0);
-  Operand dst(new_sp_reg, count_reg, times_pointer_size, 0);
-
-  // Now copy callee arguments to the caller frame going backwards to avoid
-  // callee arguments corruption (source and destination areas could overlap).
-  Label loop, entry;
-  __ jmp(&entry, Label::kNear);
-  __ bind(&loop);
-  __ dec(count_reg);
-  __ mov(tmp_reg, src);
-  __ mov(dst, tmp_reg);
-  __ bind(&entry);
-  __ cmp(count_reg, Immediate(0));
-  __ j(not_equal, &loop, Label::kNear);
-
-  // Leave current frame.
-  __ mov(esp, new_sp_reg);
-
+  ParameterCount callee_args_count(args_reg);
+  __ PrepareForTailCall(callee_args_count, caller_args_count_reg, scratch2,
+                        scratch3, ReturnAddressState::kOnStack);
   __ bind(&done);
 }
 }  // namespace

src/x87/macro-assembler-x87.cc

@@ -2024,6 +2024,77 @@ void MacroAssembler::JumpToExternalReference(const ExternalReference& ext) {
   jmp(ces.GetCode(), RelocInfo::CODE_TARGET);
 }
 
+void MacroAssembler::PrepareForTailCall(const ParameterCount& callee_args_count,
+                                        Register caller_args_count_reg,
+                                        Register scratch0, Register scratch1,
+                                        ReturnAddressState ra_state) {
+#if DEBUG
+  if (callee_args_count.is_reg()) {
+    DCHECK(!AreAliased(callee_args_count.reg(), caller_args_count_reg, scratch0,
+                       scratch1));
+  } else {
+    DCHECK(!AreAliased(caller_args_count_reg, scratch0, scratch1));
+  }
+#endif
+
+  // Calculate the destination address where we will put the return address
+  // after we drop current frame.
+  Register new_sp_reg = scratch0;
+  if (callee_args_count.is_reg()) {
+    sub(caller_args_count_reg, callee_args_count.reg());
+    lea(new_sp_reg, Operand(ebp, caller_args_count_reg, times_pointer_size,
+                            StandardFrameConstants::kCallerPCOffset));
+  } else {
+    lea(new_sp_reg, Operand(ebp, caller_args_count_reg, times_pointer_size,
+                            StandardFrameConstants::kCallerPCOffset -
+                                callee_args_count.immediate() * kPointerSize));
+  }
+
+  if (FLAG_debug_code) {
+    cmp(esp, new_sp_reg);
+    Check(below, kStackAccessBelowStackPointer);
+  }
+
+  // Copy return address from caller's frame to current frame's return address
+  // to avoid its trashing and let the following loop copy it to the right
+  // place.
+  Register tmp_reg = scratch1;
+  if (ra_state == ReturnAddressState::kOnStack) {
+    mov(tmp_reg, Operand(ebp, StandardFrameConstants::kCallerPCOffset));
+    mov(Operand(esp, 0), tmp_reg);
+  } else {
+    DCHECK(ReturnAddressState::kNotOnStack == ra_state);
+    Push(Operand(ebp, StandardFrameConstants::kCallerPCOffset));
+  }
+
+  // Restore caller's frame pointer now as it could be overwritten by
+  // the copying loop.
+  mov(ebp, Operand(ebp, StandardFrameConstants::kCallerFPOffset));
+
+  // +2 here is to copy both receiver and return address.
+  Register count_reg = caller_args_count_reg;
+  if (callee_args_count.is_reg()) {
+    lea(count_reg, Operand(callee_args_count.reg(), 2));
+  } else {
+    mov(count_reg, Immediate(callee_args_count.immediate() + 2));
+    // TODO(ishell): Unroll copying loop for small immediate values.
+  }
+
+  // Now copy callee arguments to the caller frame going backwards to avoid
+  // callee arguments corruption (source and destination areas could overlap).
+  Label loop, entry;
+  jmp(&entry, Label::kNear);
+  bind(&loop);
+  dec(count_reg);
+  mov(tmp_reg, Operand(esp, count_reg, times_pointer_size, 0));
+  mov(Operand(new_sp_reg, count_reg, times_pointer_size, 0), tmp_reg);
+  bind(&entry);
+  cmp(count_reg, Immediate(0));
+  j(not_equal, &loop, Label::kNear);
+
+  // Leave current frame.
+  mov(esp, new_sp_reg);
+}
 
 void MacroAssembler::InvokePrologue(const ParameterCount& expected,
                                     const ParameterCount& actual,

src/x87/macro-assembler-x87.h

@@ -44,6 +44,8 @@ enum PointersToHereCheck {
 
 enum RegisterValueType { REGISTER_VALUE_IS_SMI, REGISTER_VALUE_IS_INT32 };
 
+enum class ReturnAddressState { kOnStack, kNotOnStack };
+
 #ifdef DEBUG
 bool AreAliased(Register reg1, Register reg2, Register reg3 = no_reg,
                 Register reg4 = no_reg, Register reg5 = no_reg,
@@ -318,6 +320,16 @@ class MacroAssembler: public Assembler {
   // ---------------------------------------------------------------------------
   // JavaScript invokes
 
+  // Removes current frame and its arguments from the stack preserving
+  // the arguments and a return address pushed to the stack for the next call.
+  // |ra_state| defines whether return address is already pushed to stack or
+  // not. Both |callee_args_count| and |caller_args_count_reg| do not include
+  // receiver. |callee_args_count| is not modified, |caller_args_count_reg|
+  // is trashed.
+  void PrepareForTailCall(const ParameterCount& callee_args_count,
+                          Register caller_args_count_reg, Register scratch0,
+                          Register scratch1, ReturnAddressState ra_state);
+
   // Invoke the JavaScript function code by either calling or jumping.
 
   void InvokeFunctionCode(Register function, Register new_target,
@@ -757,12 +769,6 @@ class MacroAssembler: public Assembler {
   void Popcnt(Register dst, Register src) { Popcnt(dst, Operand(src)); }
   void Popcnt(Register dst, const Operand& src);
 
-  // Emit call to the code we are currently generating.
-  void CallSelf() {
-    Handle<Code> self(reinterpret_cast<Code**>(CodeObject().location()));
-    call(self, RelocInfo::CODE_TARGET);
-  }
-
   // Move if the registers are not identical.
   void Move(Register target, Register source);