re PR tree-optimization/68128 (A huge regression in Parboil v2.5 OpenMP CUTCP test (2.5 times lower performance))

	PR tree-optimization/68128
	* tree.h (OMP_CLAUSE_SHARED_READONLY): Define.
	* gimplify.c: Include gimple-walk.h.
	(enum gimplify_omp_var_data): Add GOVD_WRITTEN.
	(omp_notice_variable): Set flags to n->value if n already
	exists in target region, but we need to jump to do_outer.
	(omp_shared_to_firstprivate_optimizable_decl_p,
	omp_mark_stores, omp_find_stores_op, omp_find_stores_stmt): New
	functions.
	(gimplify_adjust_omp_clauses_1): Set OMP_CLAUSE_SHARED_READONLY
	on OMP_CLAUSE_SHARED if it is a scalar non-addressable that is
	not modified in the body.  Call omp_mark_stores for outer
	contexts on OMP_CLAUSE_SHARED clauses if they could be written
	in the body or on OMP_CLAUSE_LASTPRIVATE.
	(gimplify_adjust_omp_clauses): Add body argument, call
	omp_find_stores_{stmt,op} on the body through walk_gimple_seq.
	Set OMP_CLAUSE_SHARED_READONLY
	on OMP_CLAUSE_SHARED if it is a scalar non-addressable that is
	not modified in the body.  Call omp_mark_stores for outer
	contexts on OMP_CLAUSE_SHARED clauses if they could be written
	in the body or on OMP_CLAUSE_LASTPRIVATE or on OMP_CLAUSE_LINEAR
	without OMP_CLAUSE_LINEAR_NO_COPYOUT or on OMP_CLAUSE_REDUCTION.
	(gimplify_oacc_cache, gimplify_omp_parallel, gimplify_omp_task,
	gimplify_omp_for, gimplify_omp_workshare, gimplify_omp_target_update,
	gimplify_expr): Adjust gimplify_adjust_omp_clauses callers.
	* tree-nested.c (convert_nonlocal_omp_clauses,
	convert_local_omp_clauses): Clear OMP_CLAUSE_SHARED_READONLY on
	non-local vars or local vars referenced from nested routines.
	* omp-low.c (scan_sharing_clauses): For OMP_CLAUSE_SHARED_READONLY
	attempt to optimize it into OMP_CLAUSE_FIRSTPRIVATE.  Even for
	TREE_READONLY, don't call use_pointer_for_field with non-NULL
	second argument until we are sure we are keeping OMP_CLAUSE_SHARED.

	* gcc.dg/gomp/pr68128-1.c: New test.
	* gcc.dg/gomp/pr68128-2.c: New test.

From-SVN: r230932