On Wed, 8 Sep 2021 11:34:21 GMT, Leo Korinth <[email protected]> wrote:
>> src/hotspot/share/memory/allocation.hpp line 439:
>>
>>> 437: void* operator new(size_t size, const std::nothrow_t&
>>> nothrow_constant) throw() {
>>> 438: address res = (address)resource_allocate_bytes(size,
>>> AllocFailStrategy::RETURN_NULL);
>>> 439: DEBUG_ONLY(if (res != NULL) _thread_last_allocated =
>>> RESOURCE_AREA;)
>>
>> Maybe we should also guard against the possibility of nested allocations,
>> which may trash `_thread_last_allocated`?
>>
>>
>> #define PUSH_RESOURCE_OBJ_ALLOC_TYPE(t) \
>> assert(_thread_last_allocated == STACK_OR_EMBEDDED, "must not be nested");
>> \
>> DEBUG_ONLY(_thread_last_allocated = t); \
>>
>> ...
>> if (res != NULL) {
>> PUSH_RESOURCE_OBJ_ALLOC_TYPE(RESOURCE_AREA);
>> }
>>
>>
>> Similarly, the `ResourceObj` constructor should use a corresponding
>> `POP_RESOURCE_OBJ_ALLOC_TYPE` macro.
>
> I added a `set_type` method that ensures that the `_thread_last_allocated`
> always transition over a `STACK_OR_EMBEDDED`. I did *not* create a PUSH/POP
> macro pair because i believe it would give the false impression that we are
> doing a stack operation.
>
> Other than that I also made `allocation_type` use a `u8` as backing type. I
> also removed the now unused `allocation_mask` and the now unimportant detail
> that `STACK_OR_EMBEDDED = 0`.
Hmm, u8 was not what I was thinking, I will change that to a uint8_t in the
next update...
-------------
PR: https://git.openjdk.java.net/jdk/pull/5387
