[Bulk] [OpenACC 0/7] host_data construct

[Julian Brown]

On Mon, 26 Oct 2015 19:34:22 +0100
Jakub Jelinek <jakub@redhat.com> wrote:

> Your use_device sounds very similar to use_device_ptr clause in
> OpenMP, which is allowed on #pragma omp target data construct and is
> implemented quite a bit differently from this; it is unclear if the
> OpenACC standard requires this kind of implementation, or you just
> chose to implement it this way.  In particular, the GOMP_target_data
> call puts the variables mentioned in the use_device_ptr clauses into
> the mapping structures (similarly how map clause appears) and the
> corresponding vars are privatized within the target data region
> (which is a host region, basically a fancy { } braces), where the
> private variables contain the offloading device's pointers.

As the author of the original patch, I have to say using the mapping
structures seems like a far better approach, but I've hit some trouble
with the details of adapting OpenACC to use that method.

Firstly, on trunk at least, use_device_ptr variables are restricted to
pointer or array types: that restriction doesn't exist in OpenACC, nor
actually could I find it in the OpenMP 4.1 document (my guess is the
standards are supposed to match in this regard). I think that a program
such as this should work:

void target_fn (int *targ_data);

int
main (int argc, char *argv[])
{
  char out;
  int myvar;
#pragma omp target enter data map(to: myvar)

#pragma omp target data use_device_ptr(myvar) map(from:out)
  {
    target_fn (&myvar);
    out = 5;
  }

  return 0;
}

"myvar" would have its address taken in the use_device_ptr region, and
places where the corresponding mapped variable has its address taken
would be replaced by a direct use of the mapped pointer. (Or is that
not a well-formed thing to do, in general?). This fails with "error:
'use_device_ptr' variable is neither a pointer nor an array".

Secondly, attempts to use use_device_ptr on (e.g.
dynamically-allocated) arrays accessed through a pointer cause an ICE
with the existing trunk OpenMP code:

#include <stdlib.h>

void target_fn (char *targ_data);

int
main (int argc, char *argv[])
{
  char *myarr, out;

  myarr = malloc (1024);

#pragma omp target data map(to: myarr[0:1024])
  {
#pragma omp target data use_device_ptr(myarr) map(from:out)
    {
      target_fn (myarr);
      out = 5;
    }
  }

  return 0;
}

udp3.c: In function 'main':
udp3.c:6:1: internal compiler error: in make_decl_rtl, at varasm.c:1298
 main (int argc, char *argv[])
 ^
0x111256b make_decl_rtl(tree_node*)
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/varasm.c:1294
0x9ea005 expand_expr_real_1(tree_node*, rtx_def*, machine_mode, expand_modifier, rtx_def**, bool)
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/expr.c:9559
0x9e31c2 expand_expr_real(tree_node*, rtx_def*, machine_mode, expand_modifier, rtx_def**, bool)
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/expr.c:7892
0x9cb4ae expand_expr
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/expr.h:255
0x9d907d expand_assignment(tree_node*, tree_node*, bool)
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/expr.c:5089
0x89e219 expand_gimple_stmt_1
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/cfgexpand.c:3576
0x89e60d expand_gimple_stmt
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/cfgexpand.c:3672
0x8a5773 expand_gimple_basic_block
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/cfgexpand.c:5676
0x8a72d4 execute
        /scratch/jbrown/openacc-trunk/src/gcc-mainline/gcc/cfgexpand.c:6288

Furthermore, this looks strange to me (006t.omplower):

  .omp_data_arr.5.out = &out;
  myarr.8 = myarr;
  .omp_data_arr.5.myarr = myarr.8;
  #pragma omp target data map(from:out [len: 1]) use_device_ptr(myarr)
    {
      D.2436 = .omp_data_arr.5.myarr;
      myarr = D.2436;

That's clobbering the original myarr variable, right?

Any clues on these two? The omp-low.c code is rather opaque to me...

Thanks,

Julian