After the first libgfortran memory allocator preparation patch, this is the
actual patch that organizes unified_shared_memory allocation into libgfortran.

In the current OpenMP requires implementation, the requires_mask is collected
through offload LTO processing, and presented to libgomp when registering
offload images through GOMP_offload_register_ver() (called by the mkoffload generated
constructor linked into the program binary)

This means that the only reliable place to access omp_requires_mask is in
GOMP_offload_register_ver, however since it is called through an ELF constructor
in the *main program*, this runs later than libgfortran/runtime/main.c:init() constructor,
and because some libgfortran init actions there start allocating memory, this can cause
more deallocation errors later.

Another issue is that CUDA appears to be registering some cleanup actions using atexit(),
which forces libgomp to register gomp_target_fini() using atexit as well (to properly run
before the underlying CUDA stuff disappears). This happens to us here as well.

So to summarize we need to: (1) order libgfortran init actions after omp_requires_mask
processing is done, and (2) order libgfortran cleanup actions before gomp_target_fini,
to properly deallocate stuff without crashing.

The above explanation is for why there's a little new set of definitions, as well as
callback registering functions exported from libgomp to libgfortran, basically to register
libgfortran init/fini actions into libgomp to run.

Inside GOMP_offload_register_ver, after omp_requires_mask processing is done, we call into
libgfortran through a new _gfortran_mem_allocators_init function to insert the omp_free/alloc/etc.
based allocators into the Fortran runtime, when GOMP_REQUIRES_UNIFIED_SHARED_MEMORY is set.

All symbol references between libgfortran/libgomp are defined with weak symbols. Test of the
weak symbols are also used to determine if the other library exists in this program.

A final issue is: the case where we have an OpenMP program that does NOT have offloading.
We cannot passively determine in libgomp/libgfortran whether offloading exists or not, only the
main program itself can, by seeing if the hidden __OFFLOAD_TABLE__ exists.

When we do init/fini libgomp callback registering for OpenMP programs, those with no offloading
will not have those callback properly run (because of no offload image loading)
Therefore the solution here is a constructor added into the crtoffloadend.o fragment that does
a "null" call of GOMP_offload_register_ver, solely for triggering the post-offload_register callbacks
when __OFFLOAD_TABLE__ is NULL. (and because of this, the crtoffloadend.o Makefile rule is adjusted
to compile with PIC)

I know this is a big pile of yarn wrt how the main program/libgomp/libgfortran interacts, but it's
finally working. Again tested without regressions. Preparing to commit to devel/omp/gcc-12, and seeking
approval for mainline when the requires patches are in.

Thanks,
Chung-Lin

2022-08-15  Chung-Lin Tang  <cltang@codesourcery.com>

libgcc/
	* Makefile.in (crtoffloadend$(objext)): Add $(PICFLAG) to compile rule.
	* offloadstuff.c (GOMP_offload_register_ver): Add declaration of weak
	symbol.
	(__OFFLOAD_TABLE__): Likewise.
	(init_non_offload): New function.

libgfortran/

	* gfortran.map (GFORTRAN_13): New namespace.
	(_gfortran_mem_allocators_init): New name inside GFORTRAN_13.
	* libgfortran.h (mem_allocators_init): New exported declaration.
	* runtime/main.c (do_init): Rename from init, add run-once guard code.
	(cleanup): Add run-once guard code.
	(GOMP_post_offload_register_callback): Declare weak symbol.
	(GOMP_pre_gomp_target_fini_callback): Likewise.
	(init): New constructor to register offload callbacks, or call do_init
	when not OpenMP.
	* runtime/memory.c (gfortran_malloc): New pointer variable.
	(gfortran_calloc): Likewise.
	(gfortran_realloc): Likewise.
	(gfortran_free): Likewise.
	(mem_allocators_init): New function.
	(xmalloc): Use gfortran_malloc.
	(xmallocarray): Use gfortran_malloc.
	(xcalloc): Use gfortran_calloc.
	(xrealloc): Use gfortran_realloc.
	(xfree): Use gfortran_free.

libgomp/

	* libgomp.map (GOMP_5.1.2): New version namespace.
	(GOMP_post_offload_register_callback): New name inside GOMP_5.1.2.
	(GOMP_pre_gomp_target_fini_callback): Likewise.
	(GOMP_DEFINE_CALLBACK_SET): Macro to define callback set.
	(post_offload_register): Define callback set for after offload image
	register.
	(pre_gomp_target_fini): Define callback set for before gomp_target_fini
	is called.
	(libgfortran_malloc_usm): New function.
	(libgfortran_calloc_usm): Likewise
	(libgfortran_realloc_usm): Likewise
	(libgfortran_free_usm): Likewise.
	(_gfortran_mem_allocators_init): Declare weak symbol.
	(gomp_libgfortran_omp_allocators_init): New function.
	(GOMP_offload_register_ver): Add handling of host_table == NULL, calling
	into libgfortran to set unified_shared_memory allocators, and execution
	of post_offload_register callbacks.
	(gomp_target_init): Register all pre_gomp_target_fini callbacks to run
	at end of main using atexit().