[Bug middle-end/66820] [5/6 Regression] internal compiler error: in get_expr_operands, at tree-ssa-operands.c:910

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66820

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, the problem is that the fold_stmt folding added in that revision creates
new decls that it doesn't put into the current gimplification context vars, but
instead creates whole function temporaries.

So we have in *.gimple:
      #pragma omp parallel num_threads(num_nthreads) shared(filenames)
        {
          char * D.2570;
          long unsigned int D.2571;

          {
            char filename[508];
            int i;

            try
              {
                i = 0;
                D.2570 = filenames[i];
                strcpy (&filename, D.2570);
                D.2571 = __builtin_strlen (&filename);
                D.2572 = &filename + D.2571;
                __builtin_memcpy (D.2572, "[data]", 7);
              }
            finally
              {
                filename = {CLOBBER};
              }
          }
where the D.2572 temporary has been created by fold_stmt, but it hasn't been
added into the parallel region (nor has private(D.2572) clause which would work
too).
Right now in gimplify.c maybe_fold_stmt has a hack to avoid folding anything
inside of ORT_TARGET regions, perhaps we should extend that also to
(ctx->region_type & (ORT_PARALLEL | ORT_TASK)) != 0 too (and adjust lower_omp
to fold_stmt accordingly even when taskreg_nesting_level is non-zero).
Or get rid of fold_stmt during gimplification altogether, though that is
supposedly not suitable for gcc 5 (and keep doing it only in forwprop and
later)?