Jakub Jelinek wrote:

> So firstprivate clause handling remaps them then if declare target indirect
> is used? If so, the patch looks reasonable to me.

[I have now updated the patch to turn the testcase to ensure
that is also keeps works at runtime.]

OpenMP leaves it a bit open when the remapping has to happen,
but one can construct cases – in particular with unified-shared memory –
where it is not possible to do this upon entry to a target region.

Thus, it has to be done when the function is invoked, e.g.

i = (*g) ();

is turned (in the target region but only on the device side) into

i = (*GOMP_target_map_indirect_ptr (g)) ();

Thus, as long as the host pointer value is transferred to the device,
it works – as the lookup is done on the device side. Directly using a
device address (remap when mapping to the target) will also not shorten
the lookup, i.e. there is no need for it.

Does it still look reasonable to you?

Tobias

PS: The current OpenMP specification, it is listed mainly described via
the glossary (newest change is the addition of dummy procedure):

"indirect device invocation – An indirect call to the _device_ version of 
a _procedure_ on a _device_ other than the _host-device_, through a 
function pointer (C/C++), a pointer to a member function (C++), a dummy 
procedure (Fortran), or a procedure pointer (Fortran) that refers to the 
host version of the _procedure_."