Re: [RFC v2 1/4] tls: add macros for coroutine-safe TLS variables

[Peter Maydell]

On Wed, 1 Dec 2021 at 17:19, Stefan Hajnoczi <stefanha@redhat.com> wrote:
>
> Compiler optimizations can cache TLS values across coroutine yield
> points, resulting in stale values from the previous thread when a
> coroutine is re-entered by a new thread.
>
> Serge Guelton developed an __attribute__((noinline)) wrapper and tested
> it with clang and gcc. I formatted his idea according to QEMU's coding
> style and wrote documentation.

> +#ifdef QEMU_CO_TLS_ADDR
> +#define QEMU_DEFINE_STATIC_CO_TLS(type, var)                    \
> +    __thread type co_tls_##var;  \
> +    static inline type get_##var(void)                          \
> +    { type *p; QEMU_CO_TLS_ADDR(p, co_tls_##var); return *p; }  \
> +    static inline void set_##var(type v)                        \
> +    { type *p; QEMU_CO_TLS_ADDR(p, co_tls_##var); *p = v; }     \
> +    static inline type *get_ptr_##var(void)                     \
> +    { type *p; QEMU_CO_TLS_ADDR(p, co_tls_##var); return p; }
> +#else
> +#define QEMU_DEFINE_STATIC_CO_TLS(type, var)                    \
> +    static __thread type co_tls_##var;                          \
> +    static __attribute__((noinline, unused)) type get_##var(void)       \
> +    { return co_tls_##var; }                                    \
> +    static __attribute__((noinline, unused)) void set_##var(type v)     \
> +    { co_tls_##var = v; }                                       \
> +    static __attribute__((noinline, unused)) type *get_ptr_##var(void)  \
> +    { return &co_tls_##var; }
> +#endif

My compiler-developer colleagues present the following case where
'noinline' is not sufficient for the compiler to definitely
use different values of the address-of-the-TLS-var across an
intervening function call:

  __thread int i;

  __attribute__((noinline)) long get_ptr_i()
  {
    return (long)&i;
  }

  void switcher();

  int g()
  {
    long a = get_ptr_i();
    switcher();
    return a == get_ptr_i();
  }

Trunk clang optimizes the comparison of the two pointers down
to "must be always true" even though they might not be if the
switcher() function has resulted in a change-of-thread:
  https://godbolt.org/z/hd67zh6jW

The 'optnone' attribute (clang-specific) seems to be able
to suppress this attribute. The equivalent on gcc may
(or may not) be 'noipa'.

-- PMM