Hi Kwok, Tobias! On 2019-07-29T21:55:52+0100, Kwok Cheung Yeung <k...@codesourcery.com> wrote: > > if (omp_is_reference (new_var) > > && TREE_CODE (TREE_TYPE (new_var)) != POINTER_TYPE) > > As is, this code in lower_omp_target will always reject optional arguments, > so > it must be changed. This was introduced in commit r261025 (SVN trunk) as a > fix > for PR 85879, but it also breaks support for arrays in firstprivate, which > was > probably an unintended side-effect.
Do we have sufficient testsuite coverage for "arrays in firstprivate", in
all languages?
Grüße
Thomas
> I have found that allowing POINTER_TYPEs
> that are also DECL_BY_REFERENCE through in this condition allows both
> optional
> arguments and arrays to work without regressing the tests in r261025.
> * omp-low.c (lower_omp_target): Create temporary for received value
> and take the address for new_var if the original variable was a
> DECL_BY_REFERENCE. [...]
> --- a/gcc/omp-low.c
> +++ b/gcc/omp-low.c
> @@ -11173,7 +11173,8 @@ lower_omp_target (gimple_stmt_iterator *gsi_p,
> omp_context *ctx)
> {
> gcc_assert (is_gimple_omp_oacc (ctx->stmt));
> if (omp_is_reference (new_var)
> - && TREE_CODE (TREE_TYPE (new_var)) != POINTER_TYPE)
> + && (TREE_CODE (TREE_TYPE (new_var)) != POINTER_TYPE
> + || DECL_BY_REFERENCE (var)))
> {
> /* Create a local object to hold the instance
> value. */
Grüße
Thomas
signature.asc
Description: PGP signature
