Support use of clang::lifetime_capture_by in constructors

[usx95]

https://godbolt.org/z/KqM3znMhT

struct S{
    S(const int& s [[clang::lifetimebound]]);
};
struct T{
    T(const int& t [[clang::lifetime_capture_by(this)]]);
};

void foo() {
    S s(1); // Error. Good.
    T t(1); // No diagnostic here. Bad.
}

[hokein]

This case is already supported. The warning is not shown because this warning is disabled by default, see https://godbolt.org/z/rqKhW9qEW.

[usx95]

Hmm. I expected it to not work.
I think the current code path should not be taken for constructors and we should use the initialisation path. For example, we have a false positive:
https://godbolt.org/z/rzzM16dGc

struct T{
    T(const int& t [[clang::lifetime_capture_by(this)]]);
};

int foo(const T& t);

void foo() {
    auto x = foo(T(1)); // Error. False positive.
}

[hokein]

I see.

I think this is a case (I hit it in #117690 as well) where both the container and temporary object are destroyed at the end of the full expression, and the lifetimebound doesn't trigger on this case.

Do we want to diagnose on the following case? I think probably not (in order to reduce noise)

void foo() {
    foo(T(1)); 
}

[usx95]

struct T{
    T(const int& t [[clang::lifetime_capture_by(this)]]);
};

int foo(const T& t);
int bar(const T& t [[clang::lifetimebound]]);

void foo() {
    auto x = foo(T(1)); // Should be ok.
    auto y = bar(T(1)); // error is expected.
}

I would lean towards treating this as we treat lifetimebound for ctors here:

llvm-project/clang/lib/Sema/CheckExprLifetime.cpp

Lines 624 to 630 in 4a7b56e

	if (CheckCoroCall || Callee->getParamDecl(I)->hasAttr<LifetimeBoundAttr>())
	VisitLifetimeBoundArg(Callee->getParamDecl(I), Arg);
	else if (EnableGSLAnalysis && I == 0) {
	// Perform GSL analysis for the first argument
	if (shouldTrackFirstArgument(Callee)) {
	VisitGSLPointerArg(Callee, Arg);
	} else if (auto *Ctor = dyn_cast<CXXConstructExpr>(Call);

And not trigger checkCaptureByLifetime for constructors