for loop condition variable destroyed twice by 'continue'

[Quuxplusone]

Bugzilla Link	PR49585
Status	CONFIRMED
Importance	P enhancement
Reported by	Aleksander Miera (ammiera@hotmail.com)
Reported on	2021-03-14 06:38:41 -0700
Last modified on	2021-03-17 14:22:37 -0700
Version	11.0
Hardware	PC Linux
CC	htmldeveloper@gmail.com, llvm-bugs@lists.llvm.org, neeilans@live.com, richard-llvm@metafoo.co.uk
Fixed by commit(s)
Attachments	test.cc (1493 bytes, text/x-c++src)
Blocks
Blocked by
See also

Created attachment 24646
example program

Exactly as the title states, assuming that getData(i) returns a shared_ptr<foo>
that is stored in an underlying vector, el gets destroyed, effectively leaving
the vector in a corrupted state.
This occurs only in the presence of the continue, without it (e.g. when
printing inside the condition), everything works fine. Same if el in not
created in the condition check.
Reporting against v.11, but when fiddling with godbolt the bug seems to have
been there since forever and is replicable for C++11 and any newer standard.

void testLoop()
{
    std::cout << __FUNCTION__ << " start" << '\n';
    Holder h;
    for (auto i = 0u; auto el = h.getData(i); ++i) {
        if (i %2)
            continue;
        h.getData(i)->print();
    }
    std::cout << __FUNCTION__ << " end" << '\n';
}

It has initially been found for MSVC, if it is related:
https://stackoverflow.com/questions/66382186/possible-msvc-compiler-bug#
https://godbolt.org/z/T43n4e

[Quuxplusone]

Attached test.cc (1493 bytes, text/x-c++src): example program

[Quuxplusone]

Somewhat reduced: https://godbolt.org/z/veEzx9

Wow. So, we're emitting code for 'continue;' that branches through the cleanup for the condition variable to the increment expression. But the increment expression also branches through that cleanup to the start of the loop. So we run the destructor twice.

[Quuxplusone]

https://reviews.llvm.org/D98816