Jan,

I suppose things could have changed since 2005, but taking "Haskell on a Shared-Memory Multiprocessor" (Harris, Marlow, Peyton Jones 2005) as the reference point, the figure in section 3.3 makes clear that a thunk's payload (free variables) remains valid until update with an indirection.

The paper is also clear that there is only one header word, i.e., that a copy of the original header word isn't stashed in the header and that this header word is replaced with a blackhole (code pointer or info pointer).  They do not state outright that such a copy isn't stashed in the payload together with the free variables, but if that were the case it would seem an odd omission given the level of detail otherwise.

Kei

On 09/16/2015 10:48 AM, Jan-Willem Maessen wrote:

I *believe* the answer is "Black holing happens as / after the thunk
contents are copied to the stack, so the contents of the thunk are now
dead and the data is live on a stack somewhere."

-Jan

--
Kei Davis
Applied Computer Science Group CCS-7, Mail Stop B287
Los Alamos National Laboratory
Los Alamos, NM 87545, U.S.A.