Re: [PATCH 1/2] Use gdbarch obstack to allocate types in alloc_type_arch

[Doug Evans <xdje42 at gmail dot com>]

On Sat, Aug 29, 2015 at 2:26 PM, Patrick Palka <patrick@parcs.ath.cx> wrote:
>...
> Some background for this change: The TYPE_OBJFILE_OWNED macro tells us
> who owns a given type, and according to the macro's documentation a
> type is always owned either by an objfile or by a gdbarch.  Given this
> binary encoding of ownership it doesn't seem to make much sense for
> _any_ type to be allocated by malloc.  All types should be allocated
> on an objfile obstack or on a gdbarch obstack.

I can imagine types being allocated by malloc and tracked via whatever
object "owns" them. All that matters is that when the owner is freed
all its owned objects are freed too. Whether those objects are
malloc'd or in an obstack is just an implementation detail.

I know that that's not how arch-owned types worked prior to your patch.
I'm just saying they *could* have worked that way.
Moving them to an obstack obviates the need for keeping a copy of
the pointer so it can later be freed.

> To support allocating a type by malloc, I think the type ownership
> scheme should have three possible cases: that a type is owned by an
> objfile, or that it's owned by a gdbarch, or that it's owned by the
> caller.  This new last case could correspond to a type that's
> allocated by malloc instead of on an obstack.
>
> Does this make sense, or maybe I am misunderstanding what "owning" means here?

I think you've got it right.

There's still, technically, an open issue of "What happens if an arch
is freed and there's still a value in the history that uses that type?
Would we then have to preserve that type again? (bleah)"

I don't have too strong an opinion on what's right here.
If someone wants to allow for arches being freed and thus
"preserved" types have to be "owned" by something else, ok,
but I don't see it as an urgent need. We *could* just say that
arches are never freed for now.