[turbofan] Fix tail call optimization.

src/compiler/arm/instruction-selector-arm.cc

Index: src/compiler/arm/instruction-selector-arm.cc
diff --git a/src/compiler/arm/instruction-selector-arm.cc b/src/compiler/arm/instruction-selector-arm.cc
index af55951f64e51831a4a119dcc1300527f8e374b9..76a0acb410f0b8fcec908285b8935c94df6e8751 100644
--- a/src/compiler/arm/instruction-selector-arm.cc
+++ b/src/compiler/arm/instruction-selector-arm.cc
@@ -1080,12 +1080,11 @@ void InstructionSelector::VisitFloat64RoundTiesAway(Node* node) {
 }
 
 
-void InstructionSelector::VisitCall(Node* node, BasicBlock* handler,
-                                    CallMode call_mode) {
+void InstructionSelector::VisitCall(Node* node, BasicBlock* handler) {
   ArmOperandGenerator g(this);
   const CallDescriptor* descriptor = OpParameter<const CallDescriptor*>(node);
 
-  FrameStateDescriptor* frame_state_descriptor = NULL;
+  FrameStateDescriptor* frame_state_descriptor = nullptr;
   if (descriptor->NeedsFrameState()) {
     frame_state_descriptor =
         GetFrameStateDescriptor(node->InputAt(descriptor->InputCount()));
@@ -1094,12 +1093,11 @@ void InstructionSelector::VisitCall(Node* node, BasicBlock* handler,
   CallBuffer buffer(zone(), descriptor, frame_state_descriptor);
 
   // Compute InstructionOperands for inputs and outputs.
-  // TODO(turbofan): on ARM64 it's probably better to use the code object in a
+  // TODO(turbofan): on ARM it's probably better to use the code object in a
   // register if there are multiple uses of it. Improve constant pool and the
   // heuristics in the register allocator for where to emit constants.
   InitializeCallBuffer(node, &buffer, true, false);
 
-  // TODO(dcarney): might be possible to use claim/poke instead
   // Push any stack arguments.
   for (Node* node : base::Reversed(buffer.pushed_nodes)) {
     Emit(kArmPush, g.NoOutput(), g.UseRegister(node));
@@ -1107,21 +1105,20 @@ void InstructionSelector::VisitCall(Node* node, BasicBlock* handler,
 
   // Pass label of exception handler block.
   CallDescriptor::Flags flags = descriptor->flags();
-  if (handler != nullptr) {
+  if (handler) {
     flags |= CallDescriptor::kHasExceptionHandler;
     buffer.instruction_args.push_back(g.Label(handler));
   }
 
   // Select the appropriate opcode based on the call type.
-  bool is_tail_call = call_mode == TAIL_CALL;
   InstructionCode opcode;
   switch (descriptor->kind()) {
     case CallDescriptor::kCallCodeObject: {
-      opcode = is_tail_call ? kArchTailCallCodeObject : kArchCallCodeObject;
+      opcode = kArchCallCodeObject;
       break;
     }
     case CallDescriptor::kCallJSFunction:
-      opcode = is_tail_call ? kArchTailCallJSFunction : kArchCallJSFunction;
+      opcode = kArchCallJSFunction;
       break;
     default:
       UNREACHABLE();
@@ -1130,13 +1127,95 @@ void InstructionSelector::VisitCall(Node* node, BasicBlock* handler,
   opcode |= MiscField::encode(flags);
 
   // Emit the call instruction.
-  size_t size = is_tail_call ? 0 : buffer.outputs.size();
-  InstructionOperand* first_output =
-      size > 0 ? &buffer.outputs.front() : nullptr;
-  Instruction* call_instr =
-      Emit(opcode, size, first_output, buffer.instruction_args.size(),
-           &buffer.instruction_args.front());
-  call_instr->MarkAsCall();
+  size_t const output_count = buffer.outputs.size();
+  auto* outputs = output_count ? &buffer.outputs.front() : nullptr;
+  Emit(opcode, output_count, outputs, buffer.instruction_args.size(),
+       &buffer.instruction_args.front())->MarkAsCall();
+}
+
+
+void InstructionSelector::VisitTailCall(Node* node) {
+  ArmOperandGenerator g(this);
+  CallDescriptor const* descriptor = OpParameter<CallDescriptor const*>(node);
+  DCHECK_NE(0, descriptor->flags() & CallDescriptor::kSupportsTailCalls);
+  DCHECK_EQ(0, descriptor->flags() & CallDescriptor::kPatchableCallSite);
+  DCHECK_EQ(0, descriptor->flags() & CallDescriptor::kNeedsNopAfterCall);
+
+  // TODO(turbofan): Relax restriction for stack parameters.
+  if (descriptor->UsesOnlyRegisters() &&
+      descriptor->HasSameReturnLocationsAs(
+          linkage()->GetIncomingDescriptor())) {
+    CallBuffer buffer(zone(), descriptor, nullptr);
+
+    // Compute InstructionOperands for inputs and outputs.
+    // TODO(turbofan): on ARM it's probably better to use the code object in a
+    // register if there are multiple uses of it. Improve constant pool and the
+    // heuristics in the register allocator for where to emit constants.
+    InitializeCallBuffer(node, &buffer, true, false);
+
+    DCHECK_EQ(0u, buffer.pushed_nodes.size());
+
+    // Select the appropriate opcode based on the call type.
+    InstructionCode opcode;
+    switch (descriptor->kind()) {
+      case CallDescriptor::kCallCodeObject:
+        opcode = kArchTailCallCodeObject;
+        break;
+      case CallDescriptor::kCallJSFunction:
+        opcode = kArchTailCallJSFunction;
+        break;
+      default:
+        UNREACHABLE();
+        return;
+    }
+    opcode |= MiscField::encode(descriptor->flags());
+
+    // Emit the tailcall instruction.
+    Emit(opcode, 0, nullptr, buffer.instruction_args.size(),
+         &buffer.instruction_args.front());
+  } else {
+    FrameStateDescriptor* frame_state_descriptor =

[Jarin, 2015/05/05 09:15:24]

There seems to be some serious overlap with the VisitCall method in this branch.
Could you reuse that code?

[Benedikt Meurer, 2015/05/05 09:40:21]

As said, we will unify VisitCall and VisitTailCall, once we have an idea of what
it should look like.

+        descriptor->NeedsFrameState()
+            ? GetFrameStateDescriptor(
+                  node->InputAt(static_cast<int>(descriptor->InputCount())))
+            : nullptr;
+
+    CallBuffer buffer(zone(), descriptor, frame_state_descriptor);
+
+    // Compute InstructionOperands for inputs and outputs.
+    // TODO(turbofan): on ARM it's probably better to use the code object in a
+    // register if there are multiple uses of it. Improve constant pool and the
+    // heuristics in the register allocator for where to emit constants.
+    InitializeCallBuffer(node, &buffer, true, false);
+
+    // Push any stack arguments.
+    for (Node* node : base::Reversed(buffer.pushed_nodes)) {
+      Emit(kArmPush, g.NoOutput(), g.UseRegister(node));
+    }
+
+    // Select the appropriate opcode based on the call type.
+    InstructionCode opcode;
+    switch (descriptor->kind()) {
+      case CallDescriptor::kCallCodeObject: {
+        opcode = kArchCallCodeObject;
+        break;
+      }
+      case CallDescriptor::kCallJSFunction:
+        opcode = kArchCallJSFunction;
+        break;
+      default:
+        UNREACHABLE();
+        return;
+    }
+    opcode |= MiscField::encode(descriptor->flags());
+
+    // Emit the call instruction.
+    size_t const output_count = buffer.outputs.size();
+    auto* outputs = output_count ? &buffer.outputs.front() : nullptr;
+    Emit(opcode, output_count, outputs, buffer.instruction_args.size(),
+         &buffer.instruction_args.front())->MarkAsCall();
+    Emit(kArchRet, 0, nullptr, output_count, outputs);
+  }
 }