OpenMP: Noncontiguous "target update" for Fortran

(Note: On OG14 branch, this was tacked on to "OpenMP: Array shaping
operator and strided "target update" for C", losing its original
commit message from OG13; I've restored it as a separate patch for
OG15, and merged with "Strided/rectangular 'target update'
out-of-bounds array lookup fix" and the Fortran part of "Dimension
ordering for array-shaping operator for C and C++".)

This patch implements noncontiguous "target update" for Fortran.
The existing middle end/runtime bits relating to C++ support are reused,
with some small adjustments, e.g.:

  1. The node used to map the OMP "array descriptor" (from omp-low.cc
     onwards) now uses the OMP_CLAUSE_SIZE field as a bias (the difference
     between the "virtual origin" element with zero indices in each
     dimension and the first element actually stored in memory).

  2. The OMP_CLAUSE_SIZE field of a GOMP_MAP_DIM_STRIDE node may now be
     used to store a "span", which is the distance in bytes between
     two adjacent elements in an array (with unit stride) when that is
     different from the element size, as it can be in Fortran.

The implementation goes to some effort to massage Fortran array metadata
(array descriptors) into a form that can ultimately be consumed by
omp_target_memcpy_rect_worker. The method for doing this is described
in comments in the patch body.

2023-07-03  Julian Brown  <julian@codesourcery.com>

gcc/ChangeLog
	* gimplify.cc (gimplify_adjust_omp_clauses): Don't gimplify
	VIEW_CONVERT_EXPR away in GOMP_MAP_TO_GRID/GOMP_MAP_FROM_GRID clauses.
	* omp-low.cc (omp_noncontig_descriptor_type): Add SPAN field.
	(scan_sharing_clauses): Don't store descriptor size in its
	OMP_CLAUSE_SIZE field.
	(lower_omp_target): Add missing OMP_CLAUSE_MAP check.  Add special-case
	string handling.  Handle span and bias.  Use low bound instead of zero
	as index for trailing full dimensions.

gcc/fortran/ChangeLog
	* trans-openmp.cc (gfc_omp_deep_map_kind_p): Handle
	GOMP_MAP_{TO,FROM}_GRID, GOMP_MAP_GRID_{DIM,STRIDE}.
	(gfc_trans_omp_arrayshape_type, gfc_omp_calculate_gcd,
	gfc_desc_to_omp_noncontig_array, gfc_omp_contiguous_update_p): New
	functions.
	(gfc_trans_omp_clauses): Handle noncontiguous to/from clauses for OMP
	"target update" directives.

gcc/testsuite/ChangeLog
	* gfortran.dg/gomp/noncontig-updates-1.f90: New test.
	* gfortran.dg/gomp/noncontig-updates-2.f90: New test.
	* gfortran.dg/gomp/noncontig-updates-3.f90: New test.
	* gfortran.dg/gomp/noncontig-updates-4.f90: New test.

libgomp/ChangeLog
	* libgomp.h (omp_noncontig_array_desc): Add span field.
	* target.c (omp_target_memcpy_rect_worker): Add span parameter. Update
	forward declaration. Handle span != element_size.
	(gomp_update): Handle bias in descriptor's size slot.  Update calls to
	omp_target_memcpy_rect_worker.
	* testsuite/libgomp.fortran/noncontig-updates-1.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-2.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-3.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-4.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-5.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-6.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-7.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-8.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-9.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-10.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-11.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-12.f90: New test.
	* testsuite/libgomp.fortran/noncontig-updates-13.f90: New test.

Co-Authored-By: Sandra Loosemore <sloosemore@baylibre.com>

7 files changed, 659 insertions, 11 deletions

diff --git a/gcc/fortran/trans-openmp.cc b/gcc/fortran/trans-openmp.cc
index 66709dd..946bd8f 100644
--- a/gcc/fortran/trans-openmp.cc
+++ b/gcc/fortran/trans-openmp.cc
@@ -2372,6 +2372,10 @@ gfc_omp_deep_map_kind_p (tree clause)
     case GOMP_MAP_FIRSTPRIVATE_POINTER:
     case GOMP_MAP_FIRSTPRIVATE_REFERENCE:
     case GOMP_MAP_ATTACH_DETACH:
+    case GOMP_MAP_TO_GRID:
+    case GOMP_MAP_FROM_GRID:
+    case GOMP_MAP_GRID_DIM:
+    case GOMP_MAP_GRID_STRIDE:
       break;
     default:
       gcc_unreachable ();
@@ -3633,6 +3637,346 @@ get_symbol_rooted_namelist (hash_map<gfc_symbol *,
   return NULL;
 }
 
+/* We build an "un-Fortrannish" array-of-arrays here to pass our calculated
+   array bounds to the middle end for strided/rectangular OpenMP
+   "target update" operations.  */
+
+static tree
+gfc_trans_omp_arrayshape_type (tree type, vec<tree> *dims)
+{
+  gcc_assert (dims->length () > 0);
+
+  for (unsigned i = 0; i < dims->length (); i++)
+    {
+      tree dim = fold_convert (sizetype, (*dims)[i]);
+      /* We need the index of the last element, not the array size.  */
+      dim = size_binop (MINUS_EXPR, dim, size_one_node);
+      tree idxtype = build_index_type (dim);
+      type = build_array_type (type, idxtype);
+    }
+
+  return type;
+}
+
+/* Emit code to find the greatest common divisor of two (gfc_array_index_type)
+   trees to BLOCK.  This is Euclid's algorithm:
+
+     int
+     gcd (int a, int b)
+     {
+       int tmp;
+       while (b != 0)
+	 {
+	   tmp = b;
+	   b = a % b;
+	   a = tmp;
+	 }
+       return a;
+     }
+*/
+
+static void
+gfc_omp_calculate_gcd (stmtblock_t *block, tree dst, tree a, tree b)
+{
+  tree tmp = gfc_create_var (gfc_array_index_type, "tmp");
+  tree avar = gfc_create_var (gfc_array_index_type, "a");
+  tree bvar = gfc_create_var (gfc_array_index_type, "b");
+
+  /* Avoid clobbering the inputs.  */
+  gfc_add_modify (block, avar, a);
+  gfc_add_modify (block, bvar, b);
+
+  tree label_cond = gfc_build_label_decl (NULL_TREE);
+  tree label_loop = gfc_build_label_decl (NULL_TREE);
+  TREE_USED (label_cond) = 1;
+  TREE_USED (label_loop) = 1;
+
+  gfc_add_expr_to_block (block, build1_v (GOTO_EXPR, label_cond));
+  gfc_add_expr_to_block (block, build1_v (LABEL_EXPR, label_loop));
+
+  gfc_add_modify (block, tmp, bvar);
+  gfc_add_modify (block, bvar,
+		  fold_build2_loc (input_location, TRUNC_MOD_EXPR,
+				   gfc_array_index_type, avar, bvar));
+  gfc_add_modify (block, avar, tmp);
+
+  gfc_add_expr_to_block (block, build1_v (LABEL_EXPR, label_cond));
+
+  tmp = fold_build2_loc (input_location, NE_EXPR, boolean_type_node, bvar,
+			 gfc_index_zero_node);
+  tmp = build3_v (COND_EXPR, tmp, build1_v (GOTO_EXPR, label_loop),
+		  build_empty_stmt (input_location));
+  gfc_add_expr_to_block (block, tmp);
+
+  gfc_add_modify (block, dst, avar);
+}
+
+/* Convert a gfortran array descriptor -- specifically the per-dimension
+   strides -- into a form that can be easily translated to a noncontiguous
+   OpenMP "target update" operation.  We emit a specialized version of a
+   function like this inline:
+
+     void
+     gfc_desc_to_omp_noncontig_array (int *dims, int *strides, int ndims,
+				      int *fstrides, int *flo, int *fhi)
+     {
+       dims[ndims - 1] = (fhi[ndims - 1] - flo[ndims - 1] + 1);
+       strides[0] = fstrides[0];
+       if (ndims > 1)
+	 strides[ndims - 1] = 1;
+       if (ndims == 2)
+	 dims[0] = fstrides[1];
+       else if (ndims > 2)
+	 {
+	   int grains[ndims - 2];
+
+	   int bigger_grain = fstrides[ndims - 1];
+	   for (int i = ndims - 2; i > 0; i--)
+	     {
+	       grains[i - 1] = gcd (fstrides[i], bigger_grain);
+	       bigger_grain = grains[i - 1];
+	     }
+
+	   int volume = 1;
+	   for (int i = 0; i < ndims - 2; i++)
+	     {
+	       int g = grains[i];
+	       dims[i] = g / volume;
+	       strides[i + 1] = fstrides[i + 1] / g;
+	       volume = volume * dims[i];
+	     }
+	   dims[ndims - 2] = fstrides[ndims - 1] / volume;
+	 }
+     }
+
+   where "fstrides", "flo" and "fhi" represent the stride, low bound and upper
+   bound of each dimension in the Fortran array descriptor.
+
+   (Note that most of the complexity only applies to arrays with more than two
+   dimensions, and the final stanza won't be emitted at all for lower-ranked
+   arrays.)
+
+   The output of the algorithm is a set of dimensions dims[] = { D, C, B, A }
+   "as if" the array was declared like this (in C):
+
+     type arr[A][B][C][D];
+
+   i.e. with the innermost dimension first, and a set of strides (in terms of
+   the step size along each dimension, without previous dimensions multiplied
+   in).
+
+   As an example, if we have an array:
+
+     allocate (arr(18,19,20,21,22))
+
+   and an update operation:
+
+     !$omp target update to(arr(1:3:2,1:4:3,1:5:4,1:6:5,1:7:6))
+
+   the strides we see in the Fortran array descriptor will be:
+
+     2 54 1368 34200 861840
+
+   as given by:
+
+     2 = stride0
+     54 = dim0 * stride1
+     1368 = dim0 * dim1 * stride2
+     34200 = dim0 * dim1 * dim2 * stride3
+     861840 = dim0 * dim1 * dim2 * dim3 * stride4
+
+   where "dimN" are the extents of each dimension (18,19,20,21,22), and
+   "strideN" are the strides in terms of step length along each dimension
+   (2,3,4,5,6).
+
+   We'd like to figure out what the original dimN, strideN were from the
+   Fortran array descriptor, but that's in general impossible.  Furthermore,
+   if we naively divide a stride by the preceding stride, the result isn't
+   necessarily an integer, as for e.g.:
+
+     861840/34200 = 25.2
+
+   What we can do though is figure out the greatest common divisor of
+   each stride and the preceding one, from the largest down, and use those as
+   units of granularity, i.e. the size of the corresponding dimension we pass
+   to the middle-end/runtime.  The stepwise stride is then the number of
+   times each "grain" fits into the Fortran array descriptor stride.
+
+   The output of the algorithm will be:
+
+     dims  strides
+     18    2
+     76    3
+     5     1
+     126   5
+     9     1
+
+   These numbers work fine for libgomp target.c:omp_target_memcpy_rect_worker.
+   Multiplying them through also gives the same numbers as the source Fortran
+   array strides, i.e. dim0*dim1*dim2*stride3 (18*76*5*5) = 34200.  */
+
+static void
+gfc_desc_to_omp_noncontig_array (stmtblock_t *block, tree *ompdimsp,
+				 tree *ompstridesp, tree desc, int ndims)
+{
+  tree lastdim = build_int_cst (gfc_array_index_type, ndims - 1);
+  tree dimrange = build_index_type (lastdim);
+  tree ndimarrtype = build_array_type (gfc_array_index_type, dimrange);
+  tree ompdims = gfc_create_var (ndimarrtype, "dims");
+  tree ompstrides = gfc_create_var (ndimarrtype, "strides");
+
+  *ompdimsp = ompdims;
+  *ompstridesp = ompstrides;
+
+  /* dims[ndims - 1] = (fhi[ndims - 1] - flo[ndims - 1] + 1);  */
+  tree lastlbound = gfc_conv_array_lbound (desc, ndims - 1);
+  tree lastubound = gfc_conv_array_ubound (desc, ndims - 1);
+  tree lastrange = fold_build2_loc (input_location, MINUS_EXPR,
+				    gfc_array_index_type, lastubound,
+				    lastlbound);
+  lastrange = fold_build2_loc (input_location, PLUS_EXPR, gfc_array_index_type,
+			       lastrange, gfc_index_one_node);
+
+  gfc_add_modify (block,
+		  gfc_build_array_ref (ompdims, lastdim, NULL_TREE, true),
+		  lastrange);
+
+  /* strides[0] = fstrides[0];  */
+  tree stride0 = gfc_conv_array_stride (desc, 0);
+  gfc_add_modify (block,
+		  gfc_build_array_ref (ompstrides, gfc_index_zero_node,
+				       NULL_TREE, true),
+		  stride0);
+
+  if (ndims > 1)
+    /* strides[ndims - 1] = 1;  */
+    gfc_add_modify (block,
+		    gfc_build_array_ref (ompstrides, lastdim, NULL_TREE, true),
+		    gfc_index_one_node);
+
+  if (ndims == 2)
+    /* dims[0] = fstrides[1];  */
+    gfc_add_modify (block,
+		    gfc_build_array_ref (ompdims, gfc_index_zero_node,
+					 NULL_TREE, true),
+		    gfc_conv_array_stride (desc, 1));
+  else if (ndims > 2)
+    {
+      /* int grains[ndims - 2];  */
+      tree lastgrain = build_int_cst (gfc_array_index_type, ndims - 3);
+      tree grainrange = build_index_type (lastgrain);
+      tree grainarrtype = build_array_type (gfc_array_index_type, grainrange);
+      tree grains = gfc_create_var (grainarrtype, "grains");
+
+      /* int bigger_grain = fstrides[ndims - 1];  */
+      tree bigger_grain = gfc_create_var (gfc_array_index_type, "bigger_grain");
+      tree fstridem1 = gfc_conv_array_stride (desc, ndims - 1);
+      gfc_add_modify (block, bigger_grain, fstridem1);
+
+      /*
+	for (int i = ndims - 2; i > 0; i--)
+	  {
+	    grains[i - 1] = gcd (fstrides[i], bigger_grain);
+	    bigger_grain = grains[i - 1];
+	  }
+      */
+      stmtblock_t loop_body;
+      gfc_init_block (&loop_body);
+
+      tree idx = gfc_create_var (gfc_array_index_type, "idx");
+
+      tree gcdtmp = gfc_create_var (gfc_array_index_type, "tmp");
+      gfc_omp_calculate_gcd (&loop_body, gcdtmp,
+			     gfc_conv_descriptor_stride_get (desc, idx),
+			     bigger_grain);
+      tree idxm1 = fold_build2_loc (input_location, MINUS_EXPR,
+				    gfc_array_index_type, idx,
+				    gfc_index_one_node);
+      gfc_add_modify (&loop_body,
+		      gfc_build_array_ref (grains, idxm1, NULL_TREE, true),
+		      gcdtmp);
+      gfc_add_modify (&loop_body, bigger_grain, gcdtmp);
+
+      gfc_simple_for_loop (block, idx,
+			   build_int_cst (gfc_array_index_type, ndims - 2),
+			   gfc_index_zero_node, GT_EXPR,
+			   build_int_cst (gfc_array_index_type, -1),
+			   gfc_finish_block (&loop_body));
+      /*
+	 int volume = 1;
+	 for (int i = 0; i < ndims - 2; i++)
+	   {
+	     int g = grains[i];
+	     dims[i] = g / volume;
+	     strides[i + 1] = fstrides[i + 1] / g;
+	     volume = volume * dims[i];
+	   }
+      */
+      tree volume = gfc_create_var (gfc_array_index_type, "volume");
+      gfc_add_modify (block, volume, gfc_index_one_node);
+
+      gfc_init_block (&loop_body);
+      tree grain = gfc_create_var (gfc_array_index_type, "grain");
+      gfc_add_modify (&loop_body, grain,
+		      gfc_build_array_ref (grains, idx, NULL_TREE, true));
+      tree dims_i = gfc_build_array_ref (ompdims, idx, NULL_TREE, true);
+      gfc_add_modify (&loop_body, dims_i,
+		      fold_build2_loc (input_location, TRUNC_DIV_EXPR,
+				       gfc_array_index_type, grain, volume));
+      tree nidx = fold_build2_loc (input_location, PLUS_EXPR,
+				   gfc_array_index_type, idx,
+				   gfc_index_one_node);
+      tree strides_ni = gfc_build_array_ref (ompstrides, nidx, NULL_TREE, true);
+      tree fstrides_ni = gfc_conv_descriptor_stride_get (desc, nidx);
+      gfc_add_modify (&loop_body, strides_ni,
+		      fold_build2_loc (input_location, TRUNC_DIV_EXPR,
+				       gfc_array_index_type, fstrides_ni,
+				       grain));
+      gfc_add_modify (&loop_body, volume,
+		      fold_build2_loc (input_location, MULT_EXPR,
+				       gfc_array_index_type, volume, dims_i));
+
+      gfc_simple_for_loop (block, idx, gfc_index_zero_node,
+			   build_int_cst (gfc_array_index_type, ndims - 2),
+			   LT_EXPR, gfc_index_one_node,
+			   gfc_finish_block (&loop_body));
+
+      /* dims[ndims - 2] = fstrides[ndims - 1] / volume;  */
+      tree dimsm2
+	= gfc_build_array_ref (ompdims,
+			       build_int_cst (gfc_array_index_type, ndims - 2),
+			       NULL_TREE, true);
+      gfc_add_modify (block, dimsm2,
+		      fold_build2_loc (input_location, TRUNC_DIV_EXPR,
+				       gfc_array_index_type, fstridem1,
+				       volume));
+    }
+}
+
+/* Return TRUE if update for N can definitely be done with a single contiguous
+   transfer.  If no or if we can't tell, return FALSE.  */
+
+static bool
+gfc_omp_contiguous_update_p (gfc_omp_namelist *n)
+{
+  gfc_expr *contig_expr = n->expr;
+
+  if (!n->expr)
+    {
+      if (n->sym->attr.contiguous)
+	return true;
+
+      tree desc = gfc_trans_omp_variable (n->sym, false);
+      tree type = TREE_TYPE (desc);
+      if (!GFC_ARRAY_TYPE_P (type) && !GFC_DESCRIPTOR_TYPE_P (type))
+	return true;
+
+      contig_expr = gfc_lval_expr_from_sym (n->sym);
+    }
+
+  return gfc_is_simply_contiguous (contig_expr, false, true);
+}
+
 static tree
 gfc_trans_omp_clauses (stmtblock_t *block, gfc_omp_clauses *clauses,
 		       locus where, toc_directive cd = TOC_OPENMP)
@@ -5262,6 +5606,162 @@ gfc_trans_omp_clauses (stmtblock_t *block, gfc_omp_clauses *clauses,
 		default:
 		  gcc_unreachable ();
 		}
+
+	      if ((list == OMP_LIST_TO || list == OMP_LIST_FROM)
+		  && (!n->expr
+		       || (n->expr
+			   && n->expr->ref
+			   && n->expr->ref->type == REF_ARRAY))
+		  && !gfc_omp_contiguous_update_p (n))
+		{
+		  int ndims;
+		  gfc_se se;
+		  gfc_init_se (&se, NULL);
+
+		  tree desc, span = NULL_TREE;
+
+		  if (n->expr)
+		    {
+		      if (n->expr->rank)
+			gfc_conv_expr_descriptor (&se, n->expr);
+		      else
+			gfc_conv_expr (&se, n->expr);
+
+		      desc = se.expr;
+		      /* The span is the distance between two array elements
+			 along the innermost dimension (there may be padding
+			 or other data between elements, e.g. of a derived-type
+			 array).  */
+		      span = gfc_get_array_span (desc, n->expr);
+		      ndims = n->expr->ref->u.ar.dimen;
+		    }
+		  else
+		    {
+		      desc = gfc_trans_omp_variable (n->sym, false);
+		      tree type = TREE_TYPE (desc);
+		      if (GFC_DESCRIPTOR_TYPE_P (type))
+			span = gfc_conv_descriptor_span_get (desc);
+		      ndims = GFC_TYPE_ARRAY_RANK (type);
+		    }
+
+		  gfc_add_block_to_block (block, &se.pre);
+
+		  tree ompdims, ompstrides;
+
+		  gfc_desc_to_omp_noncontig_array (block, &ompdims,
+						   &ompstrides, desc, ndims);
+
+		  tree type = TREE_TYPE (desc);
+		  tree etype = gfc_get_element_type (type);
+		  tree elsize = fold_convert (gfc_array_index_type,
+					      size_in_bytes (etype));
+
+		  tree ptr = gfc_conv_array_data (desc);
+		  tree offset = gfc_conv_array_offset (desc);
+
+		  if (!span)
+		    /* The span is the element size.  */
+		    span = elsize;
+
+		  tree node = build_omp_clause (input_location, OMP_CLAUSE_MAP);
+
+		  switch (list)
+		    {
+		    case OMP_LIST_TO:
+		      OMP_CLAUSE_SET_MAP_KIND (node, GOMP_MAP_TO_GRID);
+		      break;
+		    case OMP_LIST_FROM:
+		      OMP_CLAUSE_SET_MAP_KIND (node, GOMP_MAP_FROM_GRID);
+		      break;
+		    default:
+		      gcc_unreachable ();
+		    }
+
+		  gcc_assert (POINTER_TYPE_P (TREE_TYPE (ptr)));
+		  tree byte_offset = fold_convert (sizetype, offset);
+		  byte_offset = size_binop (MULT_EXPR, byte_offset,
+					    fold_convert (sizetype, span));
+		  tree origin
+		    = fold_build2_loc (input_location, POINTER_PLUS_EXPR,
+				       TREE_TYPE (ptr), ptr, byte_offset);
+
+		  OMP_CLAUSE_SIZE (node) = elsize;
+
+		  omp_clauses = gfc_trans_add_clause (node, omp_clauses);
+
+		  auto_vec<tree, 5> dims;
+
+		  for (int r = 0; r < ndims; r++)
+		    {
+		      tree d
+			= gfc_build_array_ref (ompdims,
+					       build_int_cst
+						 (gfc_array_index_type, r),
+					       NULL_TREE, true);
+		      d = gfc_evaluate_now (d, block);
+		      dims.safe_push (d);
+		    }
+
+		  for (int r = ndims - 1; r >= 0; r--)
+		    {
+		      tree stride_r, len_r, lowbound_r;
+
+		      tree rcst = build_int_cst (gfc_array_index_type, r);
+
+		      stride_r = gfc_build_array_ref (ompstrides, rcst,
+						      NULL_TREE, true);
+		      lowbound_r = gfc_conv_array_lbound (desc, r);
+		      len_r
+			= fold_build2_loc (input_location, MINUS_EXPR,
+					   gfc_array_index_type,
+					   gfc_conv_array_ubound (desc, r),
+					   lowbound_r);
+		      len_r
+			= fold_build2_loc (input_location, PLUS_EXPR,
+					   gfc_array_index_type,
+					   len_r, gfc_index_one_node);
+
+		      lowbound_r
+			= fold_build2_loc (input_location, MULT_EXPR,
+					   gfc_array_index_type, lowbound_r,
+					   stride_r);
+
+		      stride_r = gfc_evaluate_now (stride_r, block);
+		      lowbound_r = gfc_evaluate_now (lowbound_r, block);
+		      len_r = gfc_evaluate_now (len_r, block);
+
+		      tree dim = build_omp_clause (input_location,
+						   OMP_CLAUSE_MAP);
+		      OMP_CLAUSE_SET_MAP_KIND (dim, GOMP_MAP_GRID_DIM);
+		      OMP_CLAUSE_DECL (dim) = lowbound_r;
+		      OMP_CLAUSE_SIZE (dim) = len_r;
+
+		      omp_clauses = gfc_trans_add_clause (dim, omp_clauses);
+
+		      if (!integer_onep (stride_r)
+			  || (r == 0 && !operand_equal_p (span, elsize)))
+			{
+			  tree snode = build_omp_clause (input_location,
+							 OMP_CLAUSE_MAP);
+			  OMP_CLAUSE_SET_MAP_KIND (snode,
+						   GOMP_MAP_GRID_STRIDE);
+			  OMP_CLAUSE_DECL (snode) = stride_r;
+			  if (r == 0 && !operand_equal_p (span, elsize))
+			    OMP_CLAUSE_SIZE (snode) = span;
+			  omp_clauses = gfc_trans_add_clause (snode,
+							      omp_clauses);
+			}
+		    }
+		  origin = build_fold_indirect_ref (origin);
+		  tree eltype = gfc_get_element_type (TREE_TYPE (desc));
+		  tree arrtype
+		    = gfc_trans_omp_arrayshape_type (eltype, &dims);
+		  OMP_CLAUSE_DECL (node)
+		    = build1_loc (input_location, VIEW_CONVERT_EXPR,
+				  arrtype, origin);
+		  continue;
+		}
+
 	      tree node = build_omp_clause (gfc_get_location (&n->where),
 					    clause_code);
 	      if (n->expr == NULL
diff --git a/gcc/gimplify.cc b/gcc/gimplify.cc
index f52e362..aa1dc91 100644
--- a/gcc/gimplify.cc
+++ b/gcc/gimplify.cc
@@ -16035,6 +16035,16 @@ gimplify_adjust_omp_clauses (gimple_seq *pre_p, gimple_seq body, tree *list_p,
 	      if (OMP_CLAUSE_MAP_KIND (c) == GOMP_MAP_STRUCT)
 		break;
 
+	      /* If we have a non-contiguous (strided/rectangular) update
+		 operation with a VIEW_CONVERT_EXPR, we need to be careful not
+		 to gimplify the conversion away, because we need it during
+		 omp-low.cc in order to retrieve the array's dimensions.  Just
+		 gimplify partially instead.  */
+	      if ((OMP_CLAUSE_MAP_KIND (c) == GOMP_MAP_TO_GRID
+		   || OMP_CLAUSE_MAP_KIND (c) == GOMP_MAP_FROM_GRID)
+		  && TREE_CODE (*pd) == VIEW_CONVERT_EXPR)
+		pd = &TREE_OPERAND (*pd, 0);
+
 	      /* We've already partly gimplified this in
 		 gimplify_scan_omp_clauses.  Don't do any more.  */
 	      if (code == OMP_TARGET && OMP_CLAUSE_MAP_IN_REDUCTION (c))
diff --git a/gcc/omp-low.cc b/gcc/omp-low.cc
index 5a150ee..7689d62 100644
--- a/gcc/omp-low.cc
+++ b/gcc/omp-low.cc
@@ -1293,6 +1293,11 @@ omp_noncontig_descriptor_type (location_t loc)
   TREE_CHAIN (field) = fields;
   fields = field;
 
+  field = build_decl (loc, FIELD_DECL, get_identifier ("__span"),
+		      size_type_node);
+  TREE_CHAIN (field) = fields;
+  fields = field;
+
   tree ptr_size_type = build_pointer_type (size_type_node);
 
   field = build_decl (loc, FIELD_DECL, get_identifier ("__dim"), ptr_size_type);
@@ -1957,7 +1962,6 @@ scan_sharing_clauses (tree clauses, omp_context *ctx)
 					  OMP_CLAUSE_MAP);
 	      OMP_CLAUSE_SET_MAP_KIND (dn, GOMP_MAP_TO_PSET);
 	      OMP_CLAUSE_DECL (dn) = desc;
-	      OMP_CLAUSE_SIZE (dn) = TYPE_SIZE_UNIT (desc_type);
 
 	      OMP_CLAUSE_CHAIN (dn) = OMP_CLAUSE_CHAIN (c);
 	      OMP_CLAUSE_CHAIN (c) = dn;
@@ -13411,6 +13415,7 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 			&& OMP_CLAUSE_MAP_KIND (nc) == GOMP_MAP_TO_PSET);
 	    c = nc;
 	    while ((nc = OMP_CLAUSE_CHAIN (c))
+		   && OMP_CLAUSE_CODE (nc) == OMP_CLAUSE_MAP
 		   && (OMP_CLAUSE_MAP_KIND (nc) == GOMP_MAP_GRID_DIM
 		       || OMP_CLAUSE_MAP_KIND (nc) == GOMP_MAP_GRID_STRIDE))
 	      c = nc;
@@ -13851,7 +13856,7 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 		int i, dims = 0;
 		auto_vec<tree> tdims;
 		bool pointer_based = false, handled_pointer_section = false;
-		tree arrsize = fold_convert (sizetype, elsize);
+		tree arrsize = size_one_node;
 
 		/* Allow a single (maybe strided) array section if we have a
 		   pointer base.  */
@@ -13863,8 +13868,12 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 		    dims = 1;
 		  }
 		else
+		  /* NOTE: Don't treat (e.g. Fortran, fixed-length) strings as
+		     array types here; array section syntax isn't applicable to
+		     strings.  */
 		  for (tree itype = type;
-		       TREE_CODE (itype) == ARRAY_TYPE;
+		       TREE_CODE (itype) == ARRAY_TYPE
+		       && !TYPE_STRING_FLAG (itype);
 		       itype = TREE_TYPE (itype))
 		    {
 		      tdims.safe_push (itype);
@@ -13905,13 +13914,16 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 		oc = c;
 		c = dn;
 
+		tree span = NULL_TREE;
+
 		for (i = 0; i < dims; i++)
 		  {
 		    nc = OMP_CLAUSE_CHAIN (c);
 		    tree dim = NULL_TREE, index = NULL_TREE, len = NULL_TREE,
 			 stride = size_one_node;
 
-		    if (OMP_CLAUSE_CODE (nc) == OMP_CLAUSE_MAP
+		    if (nc
+			&& OMP_CLAUSE_CODE (nc) == OMP_CLAUSE_MAP
 			&& OMP_CLAUSE_MAP_KIND (nc) == GOMP_MAP_GRID_DIM)
 		      {
 			index = OMP_CLAUSE_DECL (nc);
@@ -13928,6 +13940,18 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 			  {
 			    stride = OMP_CLAUSE_DECL (nc2);
 			    stride = fold_convert (sizetype, stride);
+			    if (OMP_CLAUSE_SIZE (nc2))
+			      {
+				/* If the element size is not the same as the
+				   distance between two adjacent array
+				   elements (in the innermost dimension),
+				   retrieve the latter value ("span") from the
+				   size field of the stride.  We only expect to
+				   see one such field per array.  */
+				gcc_assert (!span);
+				span = OMP_CLAUSE_SIZE (nc2);
+				span = fold_convert (sizetype, span);
+			      }
 			    nc = nc2;
 			  }
 
@@ -13985,7 +14009,8 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 			dim = size_binop (MINUS_EXPR, maxval, minval);
 			dim = size_binop (PLUS_EXPR, dim, size_one_node);
 			len = dim;
-			index = size_zero_node;
+			index = minval;
+			nc = c;
 		      }
 
 		    if (TREE_CODE (dim) != INTEGER_CST)
@@ -14007,10 +14032,55 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 		    CONSTRUCTOR_APPEND_ELT (vstride, cidx, stride);
 		  }
 
-		/* The size of the whole array -- to make sure we find any
-		   part of the array via splay-tree lookup that might be
-		   mapped on the target at runtime.  */
-		OMP_CLAUSE_SIZE (oc) = arrsize;
+		tree bias = size_zero_node;
+		tree volume = size_one_node;
+		tree enclosure = size_one_node;
+		for (i = dims - 1; i >= 0; i--)
+		  {
+		    tree dim = (*vdim)[i].value;
+		    tree index = (*vindex)[i].value;
+		    tree stride = (*vstride)[i].value;
+		    tree len = (*vlen)[i].value;
+
+		    /* For the bias we want, e.g.:
+
+			   index[0] * stride[0] * dim[1] * dim[2]
+			 + index[1] * stride[1] * dim[2]
+			 + index[2] * stride[2]
+
+		       All multiplied by "span" (or "elsize").  */
+
+		    tree index_stride = size_binop (MULT_EXPR, index, stride);
+		    bias = size_binop (PLUS_EXPR, bias,
+				       size_binop (MULT_EXPR, volume,
+						   index_stride));
+		    volume = size_binop (MULT_EXPR, volume, dim);
+
+		    if (i == 0)
+		      {
+			tree elems_covered = size_binop (MINUS_EXPR, len,
+							 size_one_node);
+			elems_covered = size_binop (MULT_EXPR, elems_covered,
+						    stride);
+			elems_covered = size_binop (PLUS_EXPR, elems_covered,
+						    size_one_node);
+			enclosure = size_binop (MULT_EXPR, enclosure,
+						elems_covered);
+		      }
+		    else
+		      enclosure = volume;
+		  }
+
+		/* If we don't have a separate span size, use the element size
+		   instead.  */
+		if (!span)
+		  span = fold_convert (sizetype, elsize);
+
+		/* The size of a volume enclosing the elements to be
+		   transferred.  */
+		OMP_CLAUSE_SIZE (oc) = size_binop (MULT_EXPR, enclosure, span);
+		/* And the bias of the first element we will update.  */
+		OMP_CLAUSE_SIZE (dn) = size_binop (MULT_EXPR, bias, span);
 
 		tree cdim = build_constructor (size_arr_type, vdim);
 		tree cindex = build_constructor (size_arr_type, vindex);
@@ -14041,13 +14111,14 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 
 		tree ndims_field = TYPE_FIELDS (desc_type);
 		tree elemsize_field = DECL_CHAIN (ndims_field);
-		tree dim_field = DECL_CHAIN (elemsize_field);
+		tree span_field = DECL_CHAIN (elemsize_field);
+		tree dim_field = DECL_CHAIN (span_field);
 		tree index_field = DECL_CHAIN (dim_field);
 		tree len_field = DECL_CHAIN (index_field);
 		tree stride_field = DECL_CHAIN (len_field);
 
 		vec<constructor_elt, va_gc> *v;
-		vec_alloc (v, 6);
+		vec_alloc (v, 7);
 
 		bool all_static = (TREE_STATIC (dim_tmp)
 				   && TREE_STATIC (index_tmp)
@@ -14077,6 +14148,7 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
 
 		CONSTRUCTOR_APPEND_ELT (v, ndims_field, ndims);
 		CONSTRUCTOR_APPEND_ELT (v, elemsize_field, elsize);
+		CONSTRUCTOR_APPEND_ELT (v, span_field, span);
 		CONSTRUCTOR_APPEND_ELT (v, dim_field, dim_tmp);
 		CONSTRUCTOR_APPEND_ELT (v, index_field, index_tmp);
 		CONSTRUCTOR_APPEND_ELT (v, len_field, len_tmp);
diff --git a/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-1.f90 b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-1.f90
new file mode 100644
index 0000000..5c60f5c
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-1.f90
@@ -0,0 +1,19 @@
+! { dg-additional-options "-fdump-tree-original" }
+
+integer :: basicarray(100)
+integer, allocatable :: allocarray(:)
+
+allocate(allocarray(1:20))
+
+!$omp target update to(basicarray)
+
+!$omp target update from(basicarray(:))
+
+!$omp target update to(allocarray)
+
+!$omp target update from(allocarray(:))
+
+end
+
+! { dg-final { scan-tree-dump-times {omp target update from\(} 2 "original" } }
+! { dg-final { scan-tree-dump-times {omp target update to\(} 2 "original" } }
diff --git a/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-2.f90 b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-2.f90
new file mode 100644
index 0000000..f5a5273
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-2.f90
@@ -0,0 +1,16 @@
+! { dg-additional-options "-fdump-tree-original" }
+
+integer, allocatable :: allocarray(:)
+integer, allocatable :: allocarray2(:,:)
+
+allocate(allocarray(1:20))
+allocate(allocarray2(1:20,1:20))
+
+! This one must be noncontiguous
+!$omp target update to(allocarray(::2))
+! { dg-final { scan-tree-dump {omp target update map\(to_grid:} "original" } }
+
+!$omp target update from(allocarray2(:,5:15))
+! { dg-final { scan-tree-dump {omp target update from\(} "original" } }
+
+end
diff --git a/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-3.f90 b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-3.f90
new file mode 100644
index 0000000..5cbfe7c
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-3.f90
@@ -0,0 +1,16 @@
+! { dg-additional-options "-fdump-tree-original" }
+
+integer, allocatable :: allocarray(:,:)
+
+allocate(allocarray(1:20,1:20))
+
+! This one could possibly be handled as a contiguous update - but isn't,
+! for now.
+!$omp target update to(allocarray(1:20,5:15))
+! { dg-final { scan-tree-dump {omp target update map\(to_grid:} "original" } }
+
+!$omp target update from(allocarray(:,5:15:2))
+! { dg-final { scan-tree-dump {omp target update map\(from_grid:} "original" } }
+
+end
+
diff --git a/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-4.f90 b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-4.f90
new file mode 100644
index 0000000..53152aa
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/gomp/noncontig-updates-4.f90
@@ -0,0 +1,15 @@
+! { dg-additional-options "-fdump-tree-original" }
+
+integer, target :: tgtarray(20)
+integer, pointer, contiguous :: arrayptr(:)
+
+arrayptr => tgtarray
+
+!$omp target update from(arrayptr)
+! { dg-final { scan-tree-dump {omp target update from\(} "original" } }
+
+!$omp target update to(arrayptr(::2))
+! { dg-final { scan-tree-dump {omp target update map\(to_grid:} "original" } }
+
+end
+