Re: Bug with GCC's handling of lifetimes of implicit-lifetime types

[Richard Biener <richard.guenther@gmail.com>]

On Sat, Dec 10, 2022 at 7:45 PM Andrew Pinski via Gcc <gcc@gcc.gnu.org> wrote:
>
> On Sat, Dec 10, 2022 at 10:36 AM Jonathan Wakely via Gcc
> <gcc@gcc.gnu.org> wrote:
> >
> > On Sat, 10 Dec 2022 at 17:42, Gavin Ray via Gcc <gcc@gcc.gnu.org> wrote:
> > >
> > > This came up when I was asking around about what the proper way was to:
> > >
> > > - Allocate aligned storage for a buffer pool/page cache
> > > - Then create pointers to "Page" structs inside of the storage memory area
> > >
> > > I thought something like this might do:
> > >
> > > struct buffer_pool
> > > {
> > >   alignas(PAGE_SIZE) std::byte storage[NUM_PAGES * PAGE_SIZE];
> > >   page* pages = new (storage) page[NUM_PAGES];
> > > }
> > >
> > > Someone told me that this was a valid solution but not to do it, because it
> > > wouldn't function properly on GCC
> > > They gave this as a reproduction:
> > >
> > > https://godbolt.org/z/EhzM37Gzh
> > >
> > > I'm not experienced enough with C++ to grok the connection between this
> > > repro and my code,
> >
> > Me neither. I don't think there is any connection, because I don't
> > think the repro shows what they think it shows.
> >
> > > but I figured
> > > I'd post it on the mailing list in case it was useful for others/might get
> > > fixed in the future =)
> > >
> > > They said it had to do with "handling of lifetimes of implicit-lifetime
> > > types"
> >
> > I don't think that code is a valid implementation of
> > start_lifetime_as. Without a proper implementation of
> > start_lifetime_as (which GCC doesn't provide yet), GCC does not allow
> > you to read the bytes of a float as an int, and doesn't give you the
> > bytes of 1.0f, it gives you 0.
> >
> > https://godbolt.org/z/dvncY9Pea works for GCC. But this has nothing to
> > do your code above, as far as I can see.
>
> See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107115#c10 for what
> is going wrong.
> Basically GCC does not have a way to express this in the IR currently
> and there are proposals there on how to do it.

I wouldn't call them "proposals" - basically the C++ language providing
holes into the TBAA system is a misdesign, it will be incredibly difficult
to implement this "hole" without sacrifying optimization which means
people will complain endlessly why std::start_lifetime_as isn't a way
to circumvent TBAA without losing optimization.

But yes, I think std::start_lifetime_as needs to be implemented in the
C++ frontend and not in the library.  As a band-aid that works you can
use

template <typename T>
auto start_lifetime_as(void* p) noexcept -> T* {
  __asm__ volatile ("" : : "g" (p) : "memory");
  return std::launder((T*)p);
}

that will a) force what is pointed to be by 'p' addressable and
b) make all aliased memory considered clobbered (including *p).

Have fun with that.

In the end we'd need something like this, for less optimization
effect maybe with a way to specify that only *(T *)p is clobbered.

template <typename T>
auto start_lifetime_as(void* p) noexcept -> T* {
  typedef T Tp __attribute__((may_alias));
  __asm__ volatile ("" : "=m" (*(Tp *)p) : "g" (p), "m" (*(Tp *)p));
  return std::launder((T*)p);
}

might work, but the typedef/casting/dereferencing might be
problematic for some 'T'.  Note on the GIMPLE level asms
with memory inputs/outputs are hard barriers for everything.

Richard.

> Thanks,
> Andrew Pinski