Patchwork [Fortran] FINAL (prep patches 2/5): Add internal STRIDE intrinsic

login
register
mail settings
Submitter Tobias Burnus
Date Dec. 31, 2012, 12:17 p.m.
Message ID <50E18240.8040806@net-b.de>
Download mbox | patch
Permalink /patch/208850/
State New
Headers show

Comments

Tobias Burnus - Dec. 31, 2012, 12:17 p.m.
The attached patch adds a new - internal only - intrinsic, STRIDE, which 
returns the stride of an array with array descriptor; for an "integer :: 
array(5,5)", the stride of dim=2 is 5.

This new intrinsic is only internally available and will be used by the 
finalization wrapper to handle noncontiguous arrays.

Build on x86-64-gnu-linux.
OK for the trunk?

Tobias
Steve Kargl - Dec. 31, 2012, 6:50 p.m.
On Mon, Dec 31, 2012 at 01:17:04PM +0100, Tobias Burnus wrote:
> The attached patch adds a new - internal only - intrinsic, STRIDE, which 
> returns the stride of an array with array descriptor; for an "integer :: 
> array(5,5)", the stride of dim=2 is 5.
> 
> This new intrinsic is only internally available and will be used by the 
> finalization wrapper to handle noncontiguous arrays.
> 

What is the plan for when J3 decides to add a
STRIDE intrinsic to Fortran20xy, which may
have different semantics to what your function
does?

At a minimum, please consider adding a leading
underscore to the function.  This will clearly
never conflict with a future Fortran standard
(unless J3 looses it collected mind).
Tobias Burnus - Dec. 31, 2012, 7:19 p.m.
Hi Steve,

Steve Kargl:
> On Mon, Dec 31, 2012 at 01:17:04PM +0100, Tobias Burnus wrote:
>> This new intrinsic is only internally available and will be used by the
>> finalization wrapper to handle noncontiguous arrays.
> What is the plan for when J3 decides to add a
> STRIDE intrinsic

Well, we then add it. As written, this intrinsic is *only* *internally* 
available.

Actually, I do hope that J3 will add a better access to Fortran's array 
descriptor. Currently, one can do a lot of manipulations from C (since 
ISO/IEC TS 29113:2012) but not from Fortran. The better access from 
Fortran also matches the wish of the UK's BSI Fortran Panel: 
ftp://ftp.nag.co.uk/sc22wg5/N1901-N1950/N1923.pdf (4th non-indent 
paragraph).

> At a minimum, please consider adding a leading underscore to the function.

If you had read the patch carefully, you had seen that the intrinsic is 
defined as GFC_PREFIX ("stride"), which expands on most systems to 
"_F._stride".

However, even if I had named it "stride" it were not available as it is 
marked as "make_from_module();". The latter implies that 
gfc_find_function (name)  won't find it. You have to use a symbol which 
has  "sym->intmod_sym_id" set - which is not possible from the user code.

The  "make_from_module();" technique has been added for procedures like 
"c_sizeof" which should be handled via the normal intrinsic handling 
(intrinsic.c, check.c, iresolve.c, trans-intrinsic.c), but where the 
procedure should only be available after using "USE ISO_C_BINDING".

By the way, the symbol also never appears in the tree / assembler as it 
is converted into the respective component access of the array 
descriptor, e.g.  "variable.dim[i].stride".

Tobias
Steve Kargl - Dec. 31, 2012, 7:38 p.m.
On Mon, Dec 31, 2012 at 08:19:49PM +0100, Tobias Burnus wrote:
> Hi Steve,
> 
> Steve Kargl:
> > On Mon, Dec 31, 2012 at 01:17:04PM +0100, Tobias Burnus wrote:
> >> This new intrinsic is only internally available and will be used by the
> >> finalization wrapper to handle noncontiguous arrays.
> > What is the plan for when J3 decides to add a
> > STRIDE intrinsic
> 
> Well, we then add it. As written, this intrinsic is *only* *internally* 
> available.
> 
> Actually, I do hope that J3 will add a better access to Fortran's array 
> descriptor. Currently, one can do a lot of manipulations from C (since 
> ISO/IEC TS 29113:2012) but not from Fortran. The better access from 
> Fortran also matches the wish of the UK's BSI Fortran Panel: 
> ftp://ftp.nag.co.uk/sc22wg5/N1901-N1950/N1923.pdf (4th non-indent 
> paragraph).
> 
> > At a minimum, please consider adding a leading underscore to the function.
> 
> If you had read the patch carefully, you had seen that the intrinsic is 
> defined as GFC_PREFIX ("stride"), which expands on most systems to 
> "_F._stride".
> 

Ah, I did miss the GFC_PREFIX declaration.  I was going off
your (mis)description of what you were adding to the list
of intrinsics.  Namely, you stated you were adding a STRIDE
intrinsic when in fact you are adding a GFC_PREFIX("stride")
intrinsic.
Paul Richard Thomas - Jan. 4, 2013, 9:10 p.m.
Dear Tobias,

This one is also OK for trunk.  It does cross my mind, however, that
we should offer STRIDE as a gcc extension, in anticipation of F201x.

Cheers

Paul

On 31 December 2012 13:17, Tobias Burnus <burnus@net-b.de> wrote:
> The attached patch adds a new - internal only - intrinsic, STRIDE, which
> returns the stride of an array with array descriptor; for an "integer ::
> array(5,5)", the stride of dim=2 is 5.
>
> This new intrinsic is only internally available and will be used by the
> finalization wrapper to handle noncontiguous arrays.
>
> Build on x86-64-gnu-linux.
> OK for the trunk?
>
> Tobias

Patch

2012-12-31  Tobias Burnus  <burnus@net-b.de>

	* intrinsic.c (add_functions): New internal intrinsic
	function GFC_PREFIX ("stride").
	* gfortran.h (gfc_isym_id): Add GFC_ISYM_STRIDE.
	* intrinsic.h (gfc_resolve_stride): New prototypes.
	* iresolve.c (gfc_resolve_stride): New function.
	* trans-intrinsic.c (conv_intrinsic_stride): New static
	function.
	(gfc_conv_intrinsic_function): Use it.

diff --git a/gcc/fortran/gfortran.h b/gcc/fortran/gfortran.h
index a419af3..027cab6 100644
--- a/gcc/fortran/gfortran.h
+++ b/gcc/fortran/gfortran.h
@@ -521,6 +521,7 @@  enum gfc_isym_id
   GFC_ISYM_SR_KIND,
   GFC_ISYM_STAT,
   GFC_ISYM_STORAGE_SIZE,
+  GFC_ISYM_STRIDE,
   GFC_ISYM_SUM,
   GFC_ISYM_SYMLINK,
   GFC_ISYM_SYMLNK,
diff --git a/gcc/fortran/intrinsic.c b/gcc/fortran/intrinsic.c
index 274c921..8328b9b 100644
--- a/gcc/fortran/intrinsic.c
+++ b/gcc/fortran/intrinsic.c
@@ -2640,6 +2640,14 @@  add_functions (void)
 
   make_generic ("size", GFC_ISYM_SIZE, GFC_STD_F95);
 
+  /* Obtain the stride for a given dimensions; to be used only internally.
+     "make_from_module" makes inaccessible for external users.  */
+  add_sym_2 (GFC_PREFIX ("stride"), GFC_ISYM_STRIDE, CLASS_INQUIRY, ACTUAL_NO,
+	     BT_INTEGER, gfc_index_integer_kind, GFC_STD_GNU,
+	     NULL, NULL, gfc_resolve_stride,
+	     ar, BT_REAL, dr, REQUIRED, dm, BT_INTEGER, ii, OPTIONAL);
+  make_from_module();
+
   add_sym_1 ("sizeof", GFC_ISYM_SIZEOF, CLASS_IMPURE, ACTUAL_NO, BT_INTEGER, ii,
 	     GFC_STD_GNU, gfc_check_sizeof, NULL, NULL,
 	     x, BT_UNKNOWN, 0, REQUIRED);
diff --git a/gcc/fortran/intrinsic.h b/gcc/fortran/intrinsic.h
index 2635ba6..4540ad8 100644
--- a/gcc/fortran/intrinsic.h
+++ b/gcc/fortran/intrinsic.h
@@ -546,6 +546,7 @@  void gfc_resolve_signal (gfc_expr *, gfc_expr *, gfc_expr *);
 void gfc_resolve_sin (gfc_expr *, gfc_expr *);
 void gfc_resolve_sinh (gfc_expr *, gfc_expr *);
 void gfc_resolve_size (gfc_expr *, gfc_expr *, gfc_expr *, gfc_expr *);
+void gfc_resolve_stride (gfc_expr *, gfc_expr *, gfc_expr *);
 void gfc_resolve_spacing (gfc_expr *, gfc_expr *);
 void gfc_resolve_spread (gfc_expr *, gfc_expr *, gfc_expr *, gfc_expr *);
 void gfc_resolve_sqrt (gfc_expr *, gfc_expr *);
diff --git a/gcc/fortran/iresolve.c b/gcc/fortran/iresolve.c
index 3f981d8..c884fc1 100644
--- a/gcc/fortran/iresolve.c
+++ b/gcc/fortran/iresolve.c
@@ -2314,6 +2314,15 @@  gfc_resolve_size (gfc_expr *f, gfc_expr *array ATTRIBUTE_UNUSED,
 
 
 void
+gfc_resolve_stride (gfc_expr *f, gfc_expr *array ATTRIBUTE_UNUSED,
+		  gfc_expr *dim ATTRIBUTE_UNUSED)
+{
+  f->ts.type = BT_INTEGER;
+  f->ts.kind = gfc_index_integer_kind;
+}
+
+
+void
 gfc_resolve_spacing (gfc_expr *f, gfc_expr *x)
 {
   f->ts = x->ts;
diff --git a/gcc/fortran/trans-intrinsic.c b/gcc/fortran/trans-intrinsic.c
index 5a89be1..e0b5f11 100644
--- a/gcc/fortran/trans-intrinsic.c
+++ b/gcc/fortran/trans-intrinsic.c
@@ -1657,6 +1657,35 @@  conv_intrinsic_cobound (gfc_se * se, gfc_expr * expr)
 
 
 static void
+conv_intrinsic_stride (gfc_se * se, gfc_expr * expr)
+{
+  gfc_actual_arglist *array_arg;
+  gfc_actual_arglist *dim_arg;
+  gfc_se argse;
+  tree desc, tmp;
+
+  array_arg = expr->value.function.actual;
+  dim_arg = array_arg->next;
+
+  gcc_assert (array_arg->expr->expr_type == EXPR_VARIABLE);
+
+  gfc_init_se (&argse, NULL);
+  gfc_conv_expr_descriptor (&argse, array_arg->expr);
+  gfc_add_block_to_block (&se->pre, &argse.pre);
+  gfc_add_block_to_block (&se->post, &argse.post);
+  desc = argse.expr;
+
+  gcc_assert (dim_arg->expr);
+  gfc_init_se (&argse, NULL);
+  gfc_conv_expr_type (&argse, dim_arg->expr, gfc_array_index_type);
+  gfc_add_block_to_block (&se->pre, &argse.pre);
+  tmp = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type,
+			 argse.expr, gfc_index_one_node);
+  se->expr = gfc_conv_descriptor_stride_get (desc, tmp);
+}
+
+
+static void
 gfc_conv_intrinsic_abs (gfc_se * se, gfc_expr * expr)
 {
   tree arg, cabs;
@@ -6806,6 +6835,10 @@  gfc_conv_intrinsic_function (gfc_se * se, gfc_expr * expr)
       gfc_conv_intrinsic_spacing (se, expr);
       break;
 
+    case GFC_ISYM_STRIDE:
+      conv_intrinsic_stride (se, expr);
+      break;
+
     case GFC_ISYM_SUM:
       gfc_conv_intrinsic_arith (se, expr, PLUS_EXPR, false);
       break;