[builtins] support exception handling in TFJ builtins

src/isolate.cc

Index: src/isolate.cc
diff --git a/src/isolate.cc b/src/isolate.cc
index b4b3c38289dcf1a3fc92f7eff04156c284f271a9..51978d82bb08c88e3afb074760d73295bb516ef9 100644
--- a/src/isolate.cc
+++ b/src/isolate.cc
@@ -1280,6 +1280,28 @@ Object* Isolate::UnwindAndFindHandler() {
       }
     }
 
+    if (frame->is_stub() && catchable_by_js) {
+      StubFrame* stub_frame = StubFrame::cast(frame);
+      int stack_depth = 0;
+      int code_offset =
+          stub_frame->LookupExceptionHandlerInTable(&stack_depth, nullptr);
+      if (code_offset >= 0) {
+        Address return_sp = frame->fp() -
+                            StandardFrameConstants::kFixedFrameSizeFromFp -
+                            stack_depth * kPointerSize;
+        STATIC_ASSERT(TryBlockConstant::kElementCount == 1);
+        Object* maybe_context = Memory::Object_at(return_sp - kPointerSize);
+        context = maybe_context->IsContext() ? Context::cast(maybe_context)
+                                             : raw_native_context();

[caitp, 2016/08/18 03:08:57]

I'm not sure about this:

For TFJ builtins, it looks like the frame should be more or less identical to
normal JS frames, however `maybe_context` is never a Context in the cctests
provided. I haven't checked if this would behave differently with proper TFJ
builtins or not yet.

Since stubs are basically internal, I think always using the currently entered
context is _probably_ fine though, but maybe not?

+
+        code = frame->LookupCode();
+        handler_sp = return_sp;
+        handler_fp = frame->fp();
+        offset = code_offset;

[caitp, 2016/08/18 03:08:57]

Without setting the offset, the stub is restarted from the beginning, and it
gets ugly and breaks things. I'm not actually 100% sure this does work, but it
does seem to get to the right spot, at least (in the two test cases).

+        break;
+      }
+    }
+
     RemoveMaterializedObjectsOnUnwind(frame);
   }