On 10 Jun 15:52, Bernd Schmidt wrote:
> On 04/17/2014 08:33 PM, Ilya Verbin wrote:
> >+{
> >+ /* Collect all omp-target global variables to offload_vars, if they have
> >not
> >+ been gathered earlier by input_offload_tables. */
> >+ if (vec_safe_is_empty (offload_vars))
>
> What if a variable was entered into the table by something other
> than input_offload_tables? We'll skip this code entirely, which
> doesn't seem right. Can we even get here after input_offload_tables
> has been called, and if so, maybe this step of collecting variables
> belongs elsewhere?
>
> Also, the previous code did the same for functions, and I can't find
> anything corresponding to that after the patch. Is this intentional?
I'll try to explain with an example bellow:
Suppose there are 2 source files: test1.c and test2.c.
1. During the compilation of test1.c:
1.1. In expand_omp_target gcc adds new target functions into offload_funcs;
1.2. In output_offload_tables gcc adds all target variables into offload_vars;
1.3. In output_offload_tables gcc streams offload_funcs/vars into TARGET
LTO_section_offload_table.
And if there is -flto, it also streams them into the HOST
LTO_section_offload_table;
1.4. In omp_finish_file gcc writes addresses from offload_funcs/vars into
test1.o.
2. The same steps happen for test2.c.
3a. If there is no -flto, ld will join raw tables from test1.o and test2.o.
And accel compiler will join tables from target
LTO_section_offload_table.
For now this mode isn't implemented, to run accel compiler we need -flto.
3b. If there is -flto (let's consider WHOPR mode, since LTO mode is
simpler), there are 2 stages:
3.1. WPA:
3.1.1. In input_offload_tables gcc reads host LTO_section_offload_table from
test1.o and test2.o;
3.1.2. In output_offload_tables gcc streams the joined tables into
LTO_section_offload_table in the new partition xxx.ltrans0.o;
3.2. LTRANS:
3.2.1. In input_offload_tables gcc reads host LTO_section_offload_table from
xxx.ltrans0.o;
3.2.2. In omp_finish_file gcc writes addresses from offload_funcs/vars into the
final xxx.ltrans0.ltrans.o.
So, the question is what is the right place for collecting decls into
offload_funcs/vars?
I collect offload_funcs in expand_omp_target where they're created.
But for offload_vars I couldn't find a place better than output_offload_tables.
That's why I added "if (vec_safe_is_empty (offload_vars))".
If the var decls have been read by input_offload_tables on the step 3.1.1,
there is no need to
collect them from FOR_EACH_DEFINED_VARIABLE on the step 3.1.2, because that
order might be incorrect.
Thanks,
-- Ilya
