[interpreter] Add explicit OSR polling bytecode.

src/interpreter/interpreter.cc

Index: src/interpreter/interpreter.cc
diff --git a/src/interpreter/interpreter.cc b/src/interpreter/interpreter.cc
index 8ff3c159dfeb79b604823d43ccb7d0206509b9ed..d2afe03d41e286c7741bbdfb9d6dafcd4e566f8d 100644
--- a/src/interpreter/interpreter.cc
+++ b/src/interpreter/interpreter.cc
@@ -1793,6 +1793,33 @@ void Interpreter::DoStackCheck(InterpreterAssembler* assembler) {
   }
 }
 
+// OsrPoll <loop_depth>
+//
+// Performs a loop nesting check and potentially triggers OSR.
+void Interpreter::DoOsrPoll(InterpreterAssembler* assembler) {
+  Node* loop_depth = __ BytecodeOperandFlag(0);
+  Node* osr_level = __ LoadOSRNestingLevel();
+
+  // Check if OSR points at the given {loop_depth} are armed by comparing it to
+  // the current {osr_level} loaded from the header of the BytecodeArray.
+  Label ok(assembler), osr_armed(assembler, Label::kDeferred);
+  Node* condition = __ Int32GreaterThanOrEqual(loop_depth, osr_level);

[rmcilroy, 2016/07/25 12:40:24]

It looks like OSR nesting level initialized to 0 in the bytecode array, and
OSRPoll always passes level zero currently. Doesn't that mean we always end up
trying to OSR?

[Michael Starzinger, 2016/07/25 13:40:48]

The {OsrPoll} handler will fall-through (i.e. the 'ok' label) when {loop_depth
>= bytecode->ost_nesting_level()} and only trigger (i.e. the 'osr_armed' label)
when this condition does not hold. Hence the initialization to 0 should not
cause it to trigger.

Also, I plan to make the loop depth tracking zero based, so top-level loops
within a function will contain a OsrPoll(0) as well.

[rmcilroy, 2016/07/26 09:26:44]

You are right, sorry I got the labels on the branch mixed up.
Right, sounds fine now I got the condition the right way around.

+  __ Branch(condition, &ok, &osr_armed);
+
+  __ Bind(&ok);
+  __ Dispatch();
+
+  __ Bind(&osr_armed);
+  {
+    Callable callable = CodeFactory::InterpreterOnStackReplacement(isolate_);
+    Node* target = __ HeapConstant(callable.code());
+    Node* context = __ GetContext();
+    // TODO(4764): We don't have a CallStub method with zero arguments, add one!

[rmcilroy, 2016/07/25 12:40:24]

Could you do that with this CL please?

[Michael Starzinger, 2016/07/25 13:40:48]

Done.

+    __ CallStub(callable.descriptor(), target, context, context);
+    __ Dispatch();
+  }
+}
+
 // Throw
 //
 // Throws the exception in the accumulator.