Ensure manage_exception_state uses correct TLS

[zhykzhykzhyk]

Solves #323

[CreoValis]

Wouldn't access through a volatile * be a better solution here? The "no inline" spec probably works around the issue, but that just seems to push the dtor body to the point where gcc couldn't "see into it" to optimize it away.
I think using a volatile pointer cast would forbid it from even optimizing it away.

Like so:

    manage_exception_state() {
        exception_state_ = *(volatile __cxa_eh_globals *)__cxa_get_globals();
    }
    ~manage_exception_state() {
        *(volatile __cxa_eh_globals *)__cxa_get_globals() = exception_state_;
    }
`

[olk]

ty

[pdimov]

This merge causes build failures on Linux:

gcc.compile.c++ bin.v2/libs/fiber/build/gcc-11/release/x86_64/threading-multi/visibility-hidden/algo/round_robin.o
In file included from ./boost/context/fiber.hpp:12,
                 from ./boost/fiber/context.hpp:28,
                 from ./boost/fiber/algo/round_robin.hpp:16,
                 from /home/runner/work/boost/boost/libs/fiber/src/algo/round_robin.cpp:7:
./boost/context/fiber_fcontext.hpp: In constructor ‘__cxxabiv1::manage_exception_state::manage_exception_state()’:
./boost/context/fiber_fcontext.hpp:91:76: error: ambiguous overload for ‘operator=’ (operand types are ‘__cxxabiv1::__cxa_eh_globals’ and ‘volatile __cxxabiv1::__cxa_eh_globals’)
   91 |         exception_state_ = *(volatile __cxa_eh_globals *)__cxa_get_globals();
      |                                                                            ^
./boost/context/fiber_fcontext.hpp:83:8: note: candidate: ‘constexpr __cxxabiv1::__cxa_eh_globals& __cxxabiv1::__cxa_eh_globals::operator=(const __cxxabiv1::__cxa_eh_globals&)’ (near match)
   83 | struct __cxa_eh_globals {
      |        ^~~~~~~~~~~~~~~~
./boost/context/fiber_fcontext.hpp:83:8: note:   conversion of argument 1 would be ill-formed:
./boost/context/fiber_fcontext.hpp:91:28: error: binding reference of type ‘const __cxxabiv1::__cxa_eh_globals&’ to ‘volatile __cxxabiv1::__cxa_eh_globals’ discards qualifiers
   91 |         exception_state_ = *(volatile __cxa_eh_globals *)__cxa_get_globals();
      |                            ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
./boost/context/fiber_fcontext.hpp:83:8: note: candidate: ‘constexpr __cxxabiv1::__cxa_eh_globals& __cxxabiv1::__cxa_eh_globals::operator=(__cxxabiv1::__cxa_eh_globals&&)’ (near match)
   83 | struct __cxa_eh_globals {
      |        ^~~~~~~~~~~~~~~~
./boost/context/fiber_fcontext.hpp:83:8: note:   conversion of argument 1 would be ill-formed:
./boost/context/fiber_fcontext.hpp:91:28: error: cannot bind rvalue reference of type ‘__cxxabiv1::__cxa_eh_globals&&’ to lvalue of type ‘volatile __cxxabiv1::__cxa_eh_globals’
   91 |         exception_state_ = *(volatile __cxa_eh_globals *)__cxa_get_globals();
      |                            ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
./boost/context/fiber_fcontext.hpp: In destructor ‘__cxxabiv1::manage_exception_state::~manage_exception_state()’:
./boost/context/fiber_fcontext.hpp:94:61: error: passing ‘volatile __cxxabiv1::__cxa_eh_globals’ as ‘this’ argument discards qualifiers [-fpermissive]
   94 |         *(volatile __cxa_eh_globals *)__cxa_get_globals() = exception_state_;
      |                                                             ^~~~~~~~~~~~~~~~
./boost/context/fiber_fcontext.hpp:83:8: note:   in call to ‘constexpr __cxxabiv1::__cxa_eh_globals& __cxxabiv1::__cxa_eh_globals::operator=(const __cxxabiv1::__cxa_eh_globals&)’
   83 | struct __cxa_eh_globals {
      |        ^~~~~~~~~~~~~~~~

https://github.com/boostorg/boost/actions/runs/22891458318/job/66415405269

[pdimov]

Also here https://github.com/boostorg/context/actions/runs/22295861288/job/64492292871

[grisumbras]

And on Mac with GCC: https://github.com/grisumbras/boost-conan-index/actions/runs/22950381877/job/66613532089#step:6:19760

[zhykzhykzhyk]

Wouldn't access through a volatile * be a better solution here? The "no inline" spec probably works around the issue, but that just seems to push the dtor body to the point where gcc couldn't "see into it" to optimize it away.
I think using a volatile pointer cast would forbid it from even optimizing it away.

Like so:

    manage_exception_state() {
        exception_state_ = *(volatile __cxa_eh_globals *)__cxa_get_globals();
    }
    ~manage_exception_state() {
        *(volatile __cxa_eh_globals *)__cxa_get_globals() = exception_state_;
    }
`

The problem is coming from the result of __cxa_get_globals() is being cached, casting the pointer to volatile doesn't prevent compiler to cache the pointer it self.

[zhykzhykzhyk]

@olk The volatile change breaks the code, can you merge this PR as-is?