This series of patches adds support for OpenMP iterators in the 'map'
clause of the 'target' construct (and it's derivatives such as 'target
enter data'), and the 'to' and 'from' constructs of the 'target update'
construct, currently for C and C++ only.
The approach in this patch differs from Tobias' WFC patch
(https://gcc.gnu.org/pipermail/gcc-patches/2021-December/586237.html) in
that it does not rely on generating a callback function - instead,
during Gimplification it generates loop(s) to evaluate every iteration
of the iterator expression, and the results (i.e. addresses, as the
expression should be an lvalue) are placed into a new array. This array
is then used as the 'hostaddrs' entry for that particular map. Libgomp
detects this (the corresponding size entry is set to SIZE_MAX, which
shouldn't normally occur) and inserts the contents of the array into the
map information before continuing on as normal.
Caveats:
- In section 2.21.7.1 of the OpenMP 5.1 standard, it states that 'If an
expression that is used to form a list item in a map clause contains an
iterator identifier, the list item instances that would result from
different values of the iterator must not have the same containing array
and must not have base pointers that share original storage' - this is
currently not enforced (it would prohibit something like map
(iterator(i=0:10), to: x[i]) while x is an int[]). As the expression in
the iterator is more-or-less unbound, it would be very difficult to
determine this at compile time. At runtime in libgomp, I suppose we
could check every iterator-derived mapping to ensure that they all
access unique entries in mem_map?
- The clause finishing currently generates spurious firstprivate maps -
the patch currently just ignores them when in iterator clauses, but is
there a better way of doing this?
- Clause reordering does not work too well with iterators. I believe the
current approach to reordering on trunk is a bit buggy in the first
place, so I just added enough to get the clauses through the pass
without ICEing.
The GCC gomp tests and all the libgomp tests have been run without
regressions on an x86-64 host with NVPTX offloading. Testing on AMD GCN
to follow.
Kwok
- [PATCH 0/3] openmp: Add support for iterators in OpenMP ... Kwok Cheung Yeung
-
