[Bug libstdc++/66017] New: Undefined behaviour in std::set<long long>

[public at hansmi dot ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66017

            Bug ID: 66017
           Summary: Undefined behaviour in std::set<long long>
           Product: gcc
           Version: 5.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libstdc++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: public at hansmi dot ch
  Target Milestone: ---

Created attachment 35463
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35463&action=edit
Test program and output after building with Clang 3.6 (using GCC toolchain
5.1.0)

When building for Linux x86 (Debian 8 (Jessie), 32 bit) using `long long' as
std::set's value type causes UBSan as included in Clang 3.5 and 3.6 to report
an downcast/upcast of a misaligned address at runtime and ASan to report
undefined behaviour, all of them in _Rb_tree.

The simplest example I could find:

---
#include <set>

int main(int, char **)
{
  std::set<long long> foo {1LL};
}
---

std::set::begin, std::set::end, set::set::empty cause reports too.

This is not reproducible when compiling with GCC 5.1.0 (with the same options
sans those specific to Clang) and neither when building for x86-64 with either
compiler.

Reproduced using:

- Clang 3.5 w/ GCC toolchain 4.9
- Clang 3.6 w/ GCC toolchain 4.9
- Clang 3.6 w/ GCC toolchain 5.1.0

Shorter value types for std::set, e.g. `long' or `char', work. Packaging the
`long long' in another type, e.g. a struct, works too. The issue does not occur
with libc++.

Bug 60734 reported something similar, though there seem to be more issues. I'm
uncertain as to whether it's an issue in _Rb_tree, __aligned_buffer or another
place altogether.

Original report at LLVM/Clang: https://llvm.org/bugs/show_bug.cgi?id=23413