https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93553
Tobias Burnus <burnus at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |burnus at gcc dot gnu.org
--- Comment #6 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Slightly simplified one can use a scalar 'c' instead of an array. One gets
(-fdump-tree-gimple):
#pragma omp parallel reduction(+:x) shared(c)
{
integer(kind=8) D.3933;
integer(kind=8) D.3934;
#pragma omp for nowait private(i)
for (i = 1; i <= 8; i = i + 1)
{
D.3933 = (integer(kind=8)) i;
D.3934 = D.3933 + -1;
(*x)[D.3934] = c;
During omp-lowering, one has:
…
D.3924 = S.0 + -1;
D.3940 = x;
(*D.3940)[D.3924] = 0;
The "D.3940 = x;" is processed and the "D.3940" ends up as "tp" argument to
"scan_omp_1_op" which calls then in turn:
tree repl = remap_decl (t, &ctx->cb)
The latter then calls
omp_copy_decl
and returns error_mark_node → this gives then the gcc_assert in scan_omp1_op.
By contrast, "S.0" appears in ctx->cb->decl_map and, hence, omp_copy_decl is
not called for it.
