[Bug middle-end/92088] aggregates with VLAs and nested functions are broken

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92088

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 47041
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47041&action=edit
patch for the inliner

void foo(int n)
{
  struct X { int a[n]; } y;

  inline struct X __attribute__((always_inline)) baz (struct X x)
    {
      x.a[0] = 1;
      return x;
    }

  y.a[0] = 0;
  y = baz(y);
  if (y.a[0] != 1)
    __builtin_abort ();
}

doesn't require any other middle-end fixes but straight runs into the inliner
deficiency.  Note that we don't use DECL_BY_REFERENCE here so the ABI for
passing this might be a bit odd (or not properly defined at all - we're
dealing with nested functions anyway).

Anyhow - attached is a patc^Whack for the inliner.  We then inline this as

  struct X xD.1985 [value-expr: *_52];
...
  _52 = __builtin_alloca_with_align (_48, 32);
  _53 = __builtin_memcpy (_52, y.1_46, _48);

but we obviously lack building a stack_save/restore pair around inlinings
of alloca-calling functions (or such param setup).