[PATCH 2/2] bpf: CO-RE consumer of deferred constants

[David Faust]

On 10/31/26 7:54 PM, York Jasper Niebuhr wrote:
> Use deferred constants to preserve semantic information required for BPF
> CO-RE relocations.
> 
> Mark enum types for deferred enum values and record types carrying
> "preserve_access_index" for deferred field offsets.  Deferred constants
> encountered during BPF lowering are either turned into CO-RE relocations
> or resolved to their ordinary compile-time values when no relocation is
> requested.
> 
> Add a "preserve_enum_values" attribute for enum types.  Enumerators of a
> type carrying this attribute automatically generate
> BPF_RELO_ENUMVAL_VALUE relocations.
> 
> Remove the bpf_enum_mappings workaround in favor of direct enum identity
> recovery from deferred cst arguments to __builtin_preserve_enum_value.
> 
> For record types carrying "preserve_access_index", deferred field
> offsets originating from offsetof expressions are converted into
> BPF_RELO_FIELD_BYTE_OFFSET relocations.
> 
> BPF CO-RE relocations can not currently be emitted for static data.  Add
> an IPA pass which resolves deferred constants in static initializers
> before the generic deferred-constant barrier.  Initializers which would
> require CO-RE relocations are diagnosed.  Support for such initializers
> could be added in the future by generating relocated runtime
> initialization code, for cases such as:
> 
>   struct task_struct {
>     ...
>     int x;
>     ...
>   } __attribute__((preserve_access_index));
> 
>   static size_t OFF_X = offsetof (struct task_struct, x);
> 
So I gave this a try to see if the above static initializer thing
is an issue for the kernel selftests.

Instead I encountered another problem when using an enumerator to index
into an array, e.g.:

  enum bpf_map_type {
    BPF_MAP_TYPE_START,
    BPF_MAP_TYPE_RINGBUF,
    __BPF_MAP_TYPE_MAX
  };
  int a[BPF_MAP_TYPE_RINGBUF];

$ ./xgcc -B. ~/playpen/bpf/defer.c -c
/home/dfaust/playpen/bpf/defer.c:8:5: error: variably modified ‘a’ at file scope
    8 | int a[BPF_MAP_TYPE_RINGBUF];
      |     ^

It's coming from this spot in c-decl.cc grokdeclarator:7444

		else if ((decl_context == NORMAL || decl_context == FIELD)
			 && current_scope == file_scope)
		  {
		    error_at (loc, "variably modified %qE at file scope", name);
		    size = integer_one_node;
		  }

Not sure whether this indicates a bug in the C/middle-end portion,
or due to how it is being used from the BPF backend, e.g. if the
plugin_finish_type hook needs to take more care rather than just
marking any enumeral type.