// gcc -O2 -Wall -S uninit-anon-bitfield.c

struct foo {
  unsigned int a:16;
  unsigned int b:11;
  unsigned int :5;
};

extern struct foo bar;

void foo(int a, int b)
{
  struct foo tmp;

  tmp.a = a;
  tmp.b = b;

  bar = tmp;
}

warns while it should not, the anon bitfield is just a padding, can't be
assigned to.  There is no warning with -fno-tree-sra.
The bug is in SRA IMHO, when scalarizing a struct assignment, it should certainly
ignore the anon bitfields (TREE_CODE (x->element) == FIELD_DECL && DECL_NAME (x->element) == NULL && TREE_CODE (TREE_TYPE (x->element)) == BIT_FIELD_TYPE ?).
Both in generate_element_copy, generate_copy_inout and perhaps in other places
as well.  Or not instantiate the anon bitfields at all.

There is another bug, e.g. for
struct foo {
  unsigned int a:16;
  unsigned int b:11;
  unsigned int :5;
};

extern struct foo bar;

void foo(struct foo a)
{
  struct foo tmp;

  tmp = a;

  bar = tmp;
}
SRA certainly should not have decided to use element copies at all, that makes
many times worse code.