diff mbox series

[RFA/RFC] C++/OpenMP: Supporting (first)private for member variables [PR110347] - or VALUE_EXPR and gimplify

Message ID 4d72cb5b-a0d5-4a38-adf1-754a069c4a2c@baylibre.com
State New
Headers show
Series [RFA/RFC] C++/OpenMP: Supporting (first)private for member variables [PR110347] - or VALUE_EXPR and gimplify | expand

Commit Message

Tobias Burnus Feb. 16, 2024, 3:40 p.m. UTC 
The following works with PARALLEL but not with TARGET.

OpenMP states the following is supposed to work:

    A = 5;  // == this->A
    B = 6;  // == this->B
    C[44] = 7; // == this->C; assume 'int C[100]'

    #pragma <parallel|target> firstprivate(A,C) private(B)
    {
      A += 5;  // Now: A is 10.
      B = 7;
      C[44] += 7; // Now C is 14
      // It is unspecified what value this->{A,B,C} has
    }
    // {A,B,C[44]} == this->{A,B,C[44]} are still {5,6,7}

* * *

In the C++ FE, that's handled by creating a temporary variable:
       v = create_temporary_var (TREE_TYPE (m));
with
       SET_DECL_VALUE_EXPR (v, m);
       DECL_OMP_PRIVATIZED_MEMBER(v)
where 'm' is, e.g., 'this->A' - and a bunch of
   'if (DECL_OMP_PRIVATIZED_MEMBER(decl))'
in the g++ FE, only.

For PARALLEL, the VALUE_EXPR survives until omp-low.cc, which handles 
this for (first)privatizing.

But for TARGET, in gimplify.cc, after the following call in 
gimplify_omp_workshare

16813  gimple *g = gimplify_and_return_first (OMP_BODY (expr), &body);

the 'A' in the body will be turned into 'this->A'.

* * *

Thus, while there is after omplower the expected

   #pragma omp target ... firstprivate(A)

and also

    D.3081 = .omp_data_i->A; A= ...;

what actually gets used is

    D.3084 = .omp_data_i->D.3046;
    this = D.3084;
    D.2996 = this->A;

which unsurprisingly breaks.

* * *

This can be "fixed" by using the following patch.

With that patch, the -fdump-tree-omplower looks fine. But it does then 
fail with:

  during RTL pass: expand
  g2.cpp:11:7: internal compiler error: in make_decl_rtl, at varasm.cc:1443

for the 'A' with 'B = A' (where B is a non-member var) and 'A' is still 
as the value expr 'this->A'.

  }


* * *

Any idea / suggestion how to handle this best?

One way I see would be to add a lang-hook here to check for 
DECL_OMP_PRIVATIZED_MEMBER, similar to the hack above. And
then ensure that the DECL_VALUE_EXPR points to the var decl
in the target region (i.e. some hacking in omp-low.cc).

I have no idea whether that would - nor whether that would be
the way forward. - Thoughts?

Tobias
diff mbox series

Patch

--- a/gcc/gimplify.cc
+++ b/gcc/gimplify.cc
@@ -3285,12 +3285,15 @@  gimplify_var_or_parm_decl (tree *expr_p)
    if (gimplify_omp_ctxp && omp_notice_variable (gimplify_omp_ctxp, 
decl, true))
      return GS_ALL_DONE;

+ if (!flag_openmp) // Assume: C++'s DECL_OMP_PRIVATIZED_MEMBER (decl)
+ {
    /* If the decl is an alias for another expression, substitute it. */
    if (DECL_HAS_VALUE_EXPR_P (decl))
      {
        *expr_p = unshare_expr (DECL_VALUE_EXPR (decl));
        return GS_OK;
      }
+ }

    return GS_ALL_DONE;