VM: Fix initialization of statics in presence of exceptions.

VM: Fix initialization of statics in presence of exceptions.

In a static- or top-level variable v

var v = expr
static v = expr

If the initializer expression expr throws, the VM throws a
CyclicInitializationError when loading v afterwards, even though
this has nothing to do with cyclic initialization.

I'm changing the way implicit static getters and static initializers are
compiled. Static initializers are invoked from static getters.

They look as follows:

get:v {
  if (field.value === transition_sentinel) {
    field.value = null;
    throw new CyclicInitializationError();
  }
  if (field.value === sentinel) {
    field.value = transition_sentinel;
    init:v();
  }
  return field.value;
}


init:v {
  try {
    field.value = expr;
  } catch {
    if (field.value === transition_sentinel) {
      field.value = null;
    }
    rethrow;
  }
}

BUG=http://dartbug.com/5802
TEST=language/lazy_static3_test,
language/throwing_lazy_variable_test,
co19/Language/12_Expressions/30_Identifier_Reference_A08_t02

R=hausner@google.com

Committed: https://code.google.com/p/dart/source/detail?r=29797

[hausner, 2013-10-31 21:56:44]

LGTM, with caveat.

Thanks for fixing the bug. As I pointed out in your previous changelist, I think
we need to change the approach you take with creating init:x functions for each
initialized static field.

[Florian Schneider, 2013-11-01 12:23:06]

On 2013/10/31 21:56:44, hausner wrote:
> LGTM, with caveat.
> 
> Thanks for fixing the bug. As I pointed out in your previous changelist, I
think
> we need to change the approach you take with creating init:x functions for
each
> initialized static field.

I'm also not too happy with the extremely verbose code for creating a synthetic
try-catch in
the parser. I'll look into how a runtime function can simplify it.

[Florian Schneider, 2013-11-01 12:23:08]

On 2013/10/31 21:56:44, hausner wrote:
> LGTM, with caveat.
> 
> Thanks for fixing the bug. As I pointed out in your previous changelist, I
think
> we need to change the approach you take with creating init:x functions for
each
> initialized static field.

I'm also not too happy with the extremely verbose code for creating a synthetic
try-catch in
the parser. I'll look into how a runtime function can simplify it.

[Florian Schneider, 2013-11-01 12:43:48]

https://codereview.chromium.org/54713003/diff/100001/runtime/vm/class_finaliz...
File runtime/vm/class_finalizer.cc (right):

https://codereview.chromium.org/54713003/diff/100001/runtime/vm/class_finaliz...
runtime/vm/class_finalizer.cc:1225: Function::NewStaticInitializer(field));
This addresses your comment to my previous CL: I now store the token position of
the field with the initializer function. When parsing the initializer function,
I simply skip ahead over the field identifier and the = token to the initializer
expression. No idea why I didn't do that in the first place.

[fschneider, 2013-11-04 10:09:01]

Committed patchset #4 manually as r29797 (presubmit successful).