[webkit-dev] Set once pointer types and static analysis
Ryosuke Niwa
rniwa at apple.com
Mon Oct 28 11:48:38 PDT 2024
Hi all,
In WebKit, it’s fairly common to write a member variable as RefPtr or std::unique_ptr that later gets lazily initialized to some value but never unset or assigned of a different value after that.
e.g.
class Foo {
Bar& bar() {
if (!m_bar)
m_bar = Bar::create();
return *m_bar;
}
Ref<Bar> protectedBar() { return bar(); }
RefPtr<Bar> m_bar;
}
Assuming there is no other code modifying m_bar, foo->bar()->method() is always safe to call even if method wasn’t a trivial function. Right now, static analyzer doesn’t recognize this pattern so we’d be forced to write code like this: foo->protectedBar()->method() where ensureProtectedBar is a wrapper around ensureBar which returns Ref<Bar>.
A suggestion was made that static analyzer can recognize this patten. Specifically, if we introduced a new smart pointer types that only allow setting the value once, static analyzer can allow foo->bar()->method()and avoid ref-churn in some cases:
e.g.
class Foo {
Bar& bar() {
if (!m_bar)
m_bar = Bar::create();
return *m_bar;
}
SetOnceRefPtr<Bar> m_bar;
}
SetOnceRefPtr::operator=(T*) release asserts that m_ptr isn’t set, and doesn’t have a move constructor, operator=(nullptr_t), leakRef(), releaseNonNull(), etc… which can override the value of m_ptr after setting the value via operator= or in constructor.
We could create various variants: SetOnceRef, SetOnceUniquePtr, SetOnceUniqueRef.
Do you think this will be useful?
- R. Niwa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-dev/attachments/20241028/0e8a731e/attachment.htm>
More information about the webkit-dev
mailing list