[turbofan] Fix tail call optimization.

src/compiler/arm64/instruction-selector-arm64.cc

Index: src/compiler/arm64/instruction-selector-arm64.cc
diff --git a/src/compiler/arm64/instruction-selector-arm64.cc b/src/compiler/arm64/instruction-selector-arm64.cc
index 874e93cb551fc8c1997a9a755af2d1d1629e50c1..f653520c36e95c6cd3524fd267af9affea43b26d 100644
--- a/src/compiler/arm64/instruction-selector-arm64.cc
+++ b/src/compiler/arm64/instruction-selector-arm64.cc
@@ -1204,12 +1204,11 @@ void InstructionSelector::VisitFloat64RoundTiesAway(Node* node) {
 }
 
 
-void InstructionSelector::VisitCall(Node* node, BasicBlock* handler,
-                                    CallMode call_mode) {
+void InstructionSelector::VisitCall(Node* node, BasicBlock* handler) {
   Arm64OperandGenerator 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()));
@@ -1261,15 +1260,14 @@ void InstructionSelector::VisitCall(Node* node, BasicBlock* 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();
@@ -1278,13 +1276,120 @@ 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) {
+  Arm64OperandGenerator g(this);
+  const CallDescriptor* descriptor = OpParameter<const CallDescriptor*>(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 ARM64 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 = nullptr;
+    if (descriptor->NeedsFrameState()) {
+      frame_state_descriptor =
+          GetFrameStateDescriptor(node->InputAt(descriptor->InputCount()));
+    }
+
+    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
+    // 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 the arguments to the stack.
+    bool pushed_count_uneven = buffer.pushed_nodes.size() & 1;
+    int aligned_push_count = buffer.pushed_nodes.size();
+    // TODO(dcarney): claim and poke probably take small immediates,
+    //                loop here or whatever.
+    // Bump the stack pointer(s).
+    if (aligned_push_count > 0) {
+      // TODO(dcarney): it would be better to bump the csp here only
+      //                and emit paired stores with increment for non c frames.
+      Emit(kArm64Claim, g.NoOutput(), g.TempImmediate(aligned_push_count));
+    }
+    // Move arguments to the stack.
+    {
+      int slot = buffer.pushed_nodes.size() - 1;
+      // Emit the uneven pushes.
+      if (pushed_count_uneven) {
+        Node* input = buffer.pushed_nodes[slot];
+        Emit(kArm64Poke, g.NoOutput(), g.UseRegister(input),
+             g.TempImmediate(slot));
+        slot--;
+      }
+      // Now all pushes can be done in pairs.
+      for (; slot >= 0; slot -= 2) {
+        Emit(kArm64PokePair, g.NoOutput(),
+             g.UseRegister(buffer.pushed_nodes[slot]),
+             g.UseRegister(buffer.pushed_nodes[slot - 1]),
+             g.TempImmediate(slot));
+      }
+    }
+
+    // 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);
+  }
 }