 No.21605
holub Tester
Jul 1, 2023, 19:59
|
chat log related to the topic: og> template<int Width, int AddrShift> typename emu::detail::handler_entry_size<Width>::uX handler_entry_read_tap<Width, AddrShift>::read(offs_t offset, uX mem_mask) const { this->ref(); uX data = this->m_next->read(offset, mem_mask); m_tap(offset, data, mem_mask); this->unref(); return data; } for reference ref/unref is incrementing/decrementing a counter and testing for zero. Sadly it's required because the access may delete the tap and if the tap is deleted you get a use-after-free on the m_tap(...) call (that's the delegate call) smf> could the tap delete be defered? og> yes, it is, by incrementing its reference counter at the start of the method and decrementing it afterwards unref deletes the object if the refcount reaches 0 smf> I mean without refcounting in the method, like when it gets removed then put it in a list of things to delete at the next vblank that way you only have a cost when deleting, not on every method call og> could be something to look at, yes |
|---|---|
|
No.23167
holub Tester
3 days ago
|
I did couple quick experiments. Looks like moving free from unref doesn't change situation. Even if 100% times it's false, that doesn't make too much difference. Only removing if completely make runtime a bit faster. |
|
No.23168
galibert Developer
3 days ago
|
Yup, it's not about the free, which almost never happens, it's all about reading the refcount and testing it. Reads are costly, conditional branches are costly. Writes, comparatively, aren't. |