Hello
On 26/07/2021 8:56 pm, Jakub Jelinek wrote:
On Mon, Jul 26, 2021 at 08:28:16PM +0100, Kwok Cheung Yeung wrote:
In Section 1.2.2 of the OpenMP TR10 spec, 'target variant' is defined as:
A version of a device routine that can only be executed as part of a target
region.
Yes, that is a target variant, but I'm pretty sure we've decided that
the target construct added for declare target is actually not a dynamic
property. So basically mostly return to the 5.0 wording with clarifications
for Fortran. See
https://github.com/OpenMP/spec/issues/2612#issuecomment-849742988
for details.
Making the target in construct dynamic would pretty much force all the
scoring to be dynamic as well.
In that comment, Deepak says:
So, we decided to keep the target trait static, requiring that the declare
target directive must be explicit and that the function version must be
different from the version of the function that may be called outside of a
target region (with the additional clarification that whether it differs or not
will be implementation defined).
"the function version must be different from the version of the function that
may be called outside of a target region": This is what we do not have in GCC at
the moment - the function versions called within and outside target regions are
the same on the host.
"whether it differs or not will be implementation defined": So whether a
function with 'declare target' and a metadirective involving a 'target'
construct behaves the same or not when called from both inside and outside of a
target region is implementation defined?
I will leave the treatment of target constructs in the selector as it is then,
with both calls going to the same function with the metadirective resolving to
the 'target' variant. I will try to address your other concerns later.
Thanks
Kwok
