[3/6] OpenMP 4.0 C++ FE support

On Wed, Oct 09, 2013 at 08:17:55PM -0400, Jason Merrill wrote:
> Please add a comment explaining why an invisiref decl is handled specially.

Ok.

> >@@ -1345,6 +1384,30 @@ cplus_decl_attributes (tree *decl, tree
> >       || *decl == error_mark_node)
> >     return;
> >
> >+  /* Add implicit "omp declare target" attribute if requested.  */
> >+  if (current_omp_declare_target_attribute
> 
> This can be called during template instantiation, so we don't want
> to use the value of this from the current parse context.  Either
> move the variable into cp_parser and tack on the attribute in the
> parser as well, or move the variable into saved_scope so that it's
> handled by push_to_top_level.

Chose the latter.

> >+  /* #pragma omp declare simd attribute needs to be always finalized.  */
> >+  if (flag_openmp
> >+      && is_attribute_p ("omp declare simd", name))
> >+    return true;
> 
> The comment doesn't seem to match the code here; the code is always
> deferring the attribute until instantiation time.  Maybe just
> s/finalized/deferred/?

Ok.

> >+      if (TREE_VALUE (attr) == NULL_TREE
> >+	  || TREE_CODE (TREE_VALUE (attr)) != TREE_LIST)
> >+	continue;
> 
> You're breaking the convention that the TREE_VALUE of an attribute
> is always a list?  That seems like a bad idea.

Changed, though it will affect Aldy's changes a little bit (two liner I
guess).

> >+cp_omp_mappable_type (tree type)
> 
> Do you want to check TYPE_OBJ_P here?  Can a reference be mappable?

References and pointers are mappable, for map clause they references are
treated specially (what the reference points to is actually mapped, plus
pointer (well, reference) adjustment, and mappable type of what it
references is checked).

> >+	  && RECORD_OR_UNION_CODE_P (TREE_CODE (CP_DECL_CONTEXT (*decl))))
> 
> DECL_CLASS_SCOPE_P (*decl)
> 
> >+	       && (TREE_CODE (CP_DECL_CONTEXT (*decl)) == FUNCTION_DECL
> 
> DECL_FUNCTION_SCOPE_P (*decl)
> 
> >+		    "for pointer type length expression is not optional");
> 
> I'd say "must be specified".

Ok.

> >+		  omp_out = build_fold_addr_expr (omp_out);
> >+		  omp_out = build_static_cast (ptype, omp_out,
> >+					       tf_warning_or_error);
> >+		  omp_in = build_fold_addr_expr (omp_in);
> >+		  omp_in = build_static_cast (ptype, omp_in,
> >+					      tf_warning_or_error);
> >+		  if (omp_out == error_mark_node || omp_in == error_mark_node)
> >+		    return true;
> >+		  omp_out = build1 (INDIRECT_REF, atype, omp_out);
> >+		  omp_in = build1 (INDIRECT_REF, atype, omp_in);
> 
> Casting to a reference type would be shorter.

Changed.

> >+	      error ("linear clause applied to non-integral non-pointer");
> 
> This error should include the offending type.

Ok.

Here is the patch then, does this look right?

2013-10-10  Jakub Jelinek  <jakub@redhat.com>

gcc/c/
	* c-typeck.c (handle_omp_array_sections_1): Change wording
	of errors requesting specification of length.
	(c_finish_omp_clauses): For error about non-integral non-pointer
	linear clause include type in the diagnostic message.
	* c-parser.c (c_finish_omp_declare_simd): If
	c_omp_declare_simd_clauses_to_numbers returned non-NULL, store
	the clauses into TREE_VALUE of an extra TREE_LIST pointed out
	by TREE_VALUE of the attribute.
gcc/cp/
	* decl.c (grokfndecl): If omp declare simd has any clauses,
	store them into TREE_VALUE of an extra TREE_LIST pointed out
	by TREE_VALUE of the attribute.
	* pt.c (apply_late_template_attributes): Likewise.
	* parser.c (cp_parser_late_parsing_omp_declare_simd): Likewise.
	(cp_parser_omp_declare_target, cp_parser_omp_end_declare_target):
	Use scope_chain->omp_declare_target_attribute instead of
	current_omp_declare_target_attribute.
	* cp-tree.h (struct saved_scope): Add omp_declare_target_attribute
	field.
	(current_omp_declare_target_attribute): Remove.
	* cp-gimplify.c (cp_genericize_r) <case OMP_CLAUSE_REDUCTION>: Add
	comment.
	* decl2.c (current_omp_declare_target_attribute): Remove.
	(is_late_template_attribute): Fix up comment about "omp declare simd"
	attribute.
	(cp_check_const_attributes): Don't expect TREE_VALUEs other
	than NULL_TREE or TREE_LIST.
	(cplus_decl_attributes): Use scope_chain->omp_declare_target_attribute
	instead of current_omp_declare_target_attribute.  Use
	DECL_CLASS_SCOPE_P and DECL_FUNCTION_SCOPE_P macros.
	* semantics.c (handle_omp_array_sections_1): Change wording
	of errors requesting specification of length.
	(finish_omp_reduction_clause): Static cast omp_{out,in,priv,orig}
	to reference type and convert_from_reference, instead of
	taking address, static cast to pointer type and INDIRECT_REF.
	(finish_omp_clauses): For error about non-integral non-pointer
	linear clause include type in the diagnostic message.
gcc/
	* tree.c (omp_remove_redundant_declare_simd_attrs): Adjust
	for the clauses being stored into TREE_VALUE of an extra TREE_LIST
	pointed out by TREE_VALUE of the attribute.
gcc/testsuite/
	* g++.dg/gomp/depend-1.C (foo): Adjust expected error messages.
	* g++.dg/gomp/depend-2.C (foo): Likewise.
	* c-c++-common/gomp/depend-1.c (foo): Likewise.
	* c-c++-common/gomp/map-1.c (foo): Likewise.

	Jakub

[3/6] OpenMP 4.0 C++ FE support

Commit Message

Comments

Patch