On Wed, 28 Oct 2015, Bernd Schmidt wrote:
> On 10/28/2015 05:54 PM, Alexander Monakov wrote: > > --- a/gcc/config/nvptx/nvptx.c > > +++ b/gcc/config/nvptx/nvptx.c > > @@ -1893,7 +1893,7 @@ init_output_initializer (FILE *file, const char *name, > > const_tree type, > > && TREE_CODE (type) != REAL_TYPE) > > | | sz < 0 > > | | sz > HOST_BITS_PER_WIDE_INT) > > - type = ptr_type_node; > > + type = char_type_node; > > decl_chunk_size = int_size_in_bytes (type); > > decl_chunk_mode = int_mode_for_mode (TYPE_MODE (type)); > > decl_offset = 0; > > The idea here was that if you have a struct with a pointer field, and an > initialization of it that uses a symbolic address, you'd be able to output the > initializer. I don't quite see how that would still work after your patch. > > You say "with no libgomp testsuite regressions", did you run any other tests? I didn't; a simple test exercising pointer fields appears to fail, indeed. So what's the way forward here? Unless packed, a structure with pointer fields will have size divisible by sizeof(void*), so can we simply pick the largest PTX type evenly dividing the size of the original type? Thanks. Alexander
