Hi,
we've been encountering an ICE for OpenACC host_data sections, which
has a use_device() clause similar to OpenMP use_device_ptr.
The ICE happens in make_decl_rtl() for scan-created variables, which IIUC,
should not be entered at all for automatic variables.
I believe the problem is, unlike other variable creation cases where the
code is split out into an offloaded child function, a host_data section
is actually host side code, so the child function local variable processing
doesn't apply here; the use_device() referenced variable has to be added
to the current host function.
So here is the quite small fix. This fixed the ICE for OpenACC on trunk
and gomp4. However when I tested it for OpenMP using the case that Julian
provided here[1], the same ICE appeared to be already fixed. I'm not sure
if some other interim change covered it up for OpenMP.
This patch was tested on trunk without regressions. Okay for trunk?
[1] https://gcc.gnu.org/ml/gcc-patches/2015-11/msg00104.html
Thanks,
Chung-Lin
* omp-low.c (scan_sharing_clauses): Call add_local_decl() for
use_device/use_device_ptr variables.
Index: omp-low.c
===================================================================
--- omp-low.c (revision 232047)
+++ omp-low.c (working copy)
@@ -1972,7 +1972,10 @@ scan_sharing_clauses (tree clauses, omp_context *c
gcc_assert (DECL_P (decl2));
install_var_local (decl2, ctx);
}
- install_var_local (decl, ctx);
+ decl = install_var_local (decl, ctx);
+ /* use_device/use_device_ptr items are actually host side variables,
+ not on the offloaded target; add to current function here. */
+ add_local_decl (cfun, decl);
break;
case OMP_CLAUSE_IS_DEVICE_PTR:
