RFR(XS): 8227384: C2 compilation fails with "graph should be schedulable" when running with -XX:-EliminateLocks

[Vladimir Kozlov]

On 9/26/19 8:37 AM, Roland Westrelin wrote:
> 
> Hi Vladimir,
> 
>> My concern is that your loading from Klass::prototype_header_offset (or make constant) instead of
>> loading from object's mark word in LoadNod Ideal() and Value() methods bypassing Store into mark
>> word or possible access to object from other thread before this load. It is okay if
>> object/allocation is not escaping but not in general case I think.
> 
> LoadNode::is_new_object_mark_load() checks the memory edge of the load:

Douh! Sorry, I missed this obvious check. You are right. Fix is good.

Thanks,
Vladimir

> 
> 1565     if (alloc != NULL && mem->is_Proj() &&
> 1566         mem->in(0) != NULL &&
> 1567         mem->in(0) == alloc->initialization() &&
> 
> So it can't bypass a store unless the store was captured but stores to
> headers are not captured (see below).
> 
> Roland.
> 
> int InitializeNode::captured_store_insertion_point(intptr_t start,
>                                                     int size_in_bytes,
>                                                     PhaseTransform* phase) {
>    const int FAIL = 0, MAX_STORE = BytesPerLong;
> 
>    if (is_complete())
>      return FAIL;                // arraycopy got here first; punt
> 
>    assert(allocation() != NULL, "must be present");
> 
>    // no negatives, no header fields:
>    if (start < (intptr_t) allocation()->minimum_header_size())  return FAIL;
>