[Fortran-dev][Patch, Fortran] C_F_Pointer cleanup

[Fortran-dev][Patch, Fortran] C_F_Pointer cleanup

[Tobias Burnus]

[-- Attachment #1: Type: text/plain, Size: 1185 bytes --]

Dear all,

that's a follow up cleanup to the patch, which has just been merged. 
Most parts should be really obvious and nice, however, the offset part 
isn't. As the offset is not part of the descriptor defined at DTS 
29113:2012's "8.2 C descriptors" (p. 17), we will have to drop it at 
some point. Hence, I used the less readable name "tmp" and an integer 
divide, instead of multiplying "stride" by "tmp" when setting the "sm" 
and changing "tmp" to, e.g., "size".

Bootstrapped (with C) and regtested on x86-64-linux.
OK for the branch?

Currently failing are the following 14 (13) test cases:

gfortran.dg/optional_dim_3.f90 (I have a patch)

gfortran.dg/associated_2.f90
gfortran.dg/auto_char_dummy_array_1.f90
gfortran.dg/auto_char_len_3.f90
gfortran.dg/class_array_1.f03
gfortran.dg/class_array_2.f03
gfortran.dg/class_array_3.f03
gfortran.dg/class_to_type_1.f03
gfortran.dg/proc_decl_23.f90
gfortran.dg/select_type_26.f03
gfortran.dg/select_type_27.f03
gfortran.dg/read_eof_all.f90
gfortran.dg/transfer_intrinsic_3.f90
gfortran.dg/subref_array_pointer_2.f90

(Plus gfortran.dg/lto/pr45586 and gfortran.dg/realloc_on_assign_5.f03, 
but those also fail on the trunk.)

Tobias

[-- Attachment #2: ftn-dev-c_f_ptr.diff --]
[-- Type: text/x-patch, Size: 4209 bytes --]

2012-07-15  Tobias Burnus  <burnus@net-b.de>

	* trans-expr.c (conv_isocbinding_procedure): For C_F_Pointer,
	directly set extent and sm instead of using ubound and stride.

2012-07-15  Tobias Burnus  <burnus@net-b.de>

	* gfortran.dg/c_f_pointer_tests_3.f90: Update scan-tree-dump
	pattern.

Index: gcc/fortran/trans-expr.c
===================================================================
--- gcc/fortran/trans-expr.c	(Revision 189481)
+++ gcc/fortran/trans-expr.c	(Arbeitskopie)
@@ -3315,7 +3315,7 @@ conv_isocbinding_procedure (gfc_se * se, gfc_symbo
       gfc_se fptrse;
       gfc_se shapese;
       gfc_ss *ss, *shape_ss;
-      tree desc, dim, tmp, stride, offset;
+      tree desc, dim, tmp, sm, offset;
       stmtblock_t body, block;
       gfc_loopinfo loop;
 
@@ -3378,9 +3378,10 @@ conv_isocbinding_procedure (gfc_se * se, gfc_symbo
       gfc_copy_loopinfo_to_se (&shapese, &loop);
       shapese.ss = shape_ss;
 
-      stride = gfc_create_var (gfc_array_index_type, "stride");
+      sm = gfc_create_var (gfc_array_index_type, "sm");
       offset = gfc_create_var (gfc_array_index_type, "offset");
-      gfc_add_modify (&block, stride, gfc_index_one_node);
+      tmp = size_in_bytes (gfc_get_element_type (TREE_TYPE (desc)));
+      gfc_add_modify (&block, sm, fold_convert (TREE_TYPE (sm), tmp));
       gfc_add_modify (&block, offset, gfc_index_zero_node);
 
       /* Loop body.  */
@@ -3389,23 +3390,27 @@ conv_isocbinding_procedure (gfc_se * se, gfc_symbo
       dim = fold_build2_loc (input_location, MINUS_EXPR, gfc_array_index_type,
 			     loop.loopvar[0], loop.from[0]);
 
-      /* Set bounds and stride. */
+      /* Set bounds and stride multiplier. */
       gfc_conv_descriptor_lbound_set (&body, desc, dim, gfc_index_one_node);
-      gfc_conv_descriptor_stride_set (&body, desc, dim, stride);
+      gfc_conv_descriptor_sm_set (&body, desc, dim, sm);
 
       gfc_conv_expr (&shapese, arg->next->next->expr);
       gfc_add_block_to_block (&body, &shapese.pre);
-      gfc_conv_descriptor_ubound_set (&body, desc, dim, shapese.expr);
+      gfc_conv_descriptor_extent_set (&body, desc, dim, shapese.expr);
       gfc_add_block_to_block (&body, &shapese.post);
 
-      /* Calculate offset. */
+      /* Calculate offset. Change from the stride multiplier back to the
+	 stride.  */
+      tmp = fold_build2_loc (input_location, TRUNC_DIV_EXPR,
+			     gfc_array_index_type, sm,
+			     fold_convert (TREE_TYPE (sm), tmp));
       gfc_add_modify (&body, offset,
 		      fold_build2_loc (input_location, PLUS_EXPR,
-				       gfc_array_index_type, offset, stride));
+				       gfc_array_index_type, offset, tmp));
       /* Update stride.  */
-      gfc_add_modify (&body, stride,
+      gfc_add_modify (&body, sm,
 		      fold_build2_loc (input_location, MULT_EXPR,
-				       gfc_array_index_type, stride,
+				       gfc_array_index_type, sm,
 				       fold_convert (gfc_array_index_type,
 						     shapese.expr)));
       /* Finish scalarization loop.  */ 
Index: gcc/testsuite/gfortran.dg/c_f_pointer_tests_3.f90
===================================================================
--- gcc/testsuite/gfortran.dg/c_f_pointer_tests_3.f90	(Revision 189481)
+++ gcc/testsuite/gfortran.dg/c_f_pointer_tests_3.f90	(Arbeitskopie)
@@ -31,10 +31,10 @@ end program test
 !
 ! Array c_f_pointer:
 !
-! { dg-final { scan-tree-dump-times " fptr_array.data = cptr;" 1 "original" } }
-! { dg-final { scan-tree-dump-times " fptr_array.dim\\\[S..\\\].lbound = 1;" 1 "original" } }
-! { dg-final { scan-tree-dump-times " fptr_array.dim\\\[S..\\\].ubound = " 1 "original" } }
-! { dg-final { scan-tree-dump-times " fptr_array.dim\\\[S..\\\].stride = " 1 "original" } }
+! { dg-final { scan-tree-dump-times " fptr_array.base_addr = cptr;" 1 "original" } }
+! { dg-final { scan-tree-dump-times " fptr_array.dim\\\[S..\\\].lower_bound = 1;" 1 "original" } }
+! { dg-final { scan-tree-dump-times " fptr_array.dim\\\[S..\\\].extent = " 1 "original" } }
+! { dg-final { scan-tree-dump-times " fptr_array.dim\\\[S..\\\].sm = " 1 "original" } }
 !
 ! Check c_f_procpointer
 ! { dg-final { scan-tree-dump-times "  fprocptr = .integer.kind=4. .\\*<.*>. ... cfunptr;" 1 "original" } }

Re: [Fortran-dev][Patch, Fortran] C_F_Pointer cleanup

[Mikael Morin]

On 15/07/2012 08:23, Tobias Burnus wrote:
> Dear all,
> 
> that's a follow up cleanup to the patch, which has just been merged.
> Most parts should be really obvious and nice, however, the offset part
> isn't. As the offset is not part of the descriptor defined at DTS
> 29113:2012's "8.2 C descriptors" (p. 17), we will have to drop it at
> some point. Hence, I used the less readable name "tmp" and an integer
> divide, instead of multiplying "stride" by "tmp" when setting the "sm"
> and changing "tmp" to, e.g., "size".
> 
> Bootstrapped (with C) and regtested on x86-64-linux.
> OK for the branch?
> 
OK.
I guess the division will be dropped too (the middle-end does the
reverse multiplication implicitly in the pointer arithmetic).
I wonder which one is faster (the casts in the second version are
certainly not nice for the middle-end):

ptr + (offset_in_bytes / elt_size_in_bytes)
VS
(sometype *) (((char *)ptr) + offset_in_bytes)

Mikael

Re: [Fortran-dev][Patch, Fortran] C_F_Pointer cleanup

[Tobias Burnus]

On 07/20/2012 11:52 AM, Mikael Morin wrote
> OK.

Thanks for the review.

> I guess the division will be dropped too (the middle-end does the
> reverse multiplication implicitly in the pointer arithmetic).
> I wonder which one is faster (the casts in the second version are
> certainly not nice for the middle-end):
>
> ptr + (offset_in_bytes / elt_size_in_bytes)
> VS
> (sometype *) (((char *)ptr) + offset_in_bytes)

As we have to support strides which aren't multiples of the element 
size, we have to use the second form. We currently do so for BT_CLASS 
but not for BT_TYPE. Actually, half of the remaining failures on the 
branch seem to be due to the wrong usage of stride- vs. byte-based 
offsets. For instance in SELECT TYPE .. TYPE IS, where one has a BT_TYPE 
and which fails.

I think the conversion to byte offsets will be the next step. 
Afterwards, we should get rid of the offset.

Tobias

Re: [Fortran-dev][Patch, Fortran] C_F_Pointer cleanup

[Richard Guenther]

On Fri, Jul 20, 2012 at 12:17 PM, Tobias Burnus <burnus@net-b.de> wrote:
> On 07/20/2012 11:52 AM, Mikael Morin wrote
>>
>> OK.
>
>
> Thanks for the review.
>
>
>> I guess the division will be dropped too (the middle-end does the
>> reverse multiplication implicitly in the pointer arithmetic).
>> I wonder which one is faster (the casts in the second version are
>> certainly not nice for the middle-end):
>>
>> ptr + (offset_in_bytes / elt_size_in_bytes)
>> VS
>> (sometype *) (((char *)ptr) + offset_in_bytes)
>
>
> As we have to support strides which aren't multiples of the element size, we
> have to use the second form. We currently do so for BT_CLASS but not for
> BT_TYPE. Actually, half of the remaining failures on the branch seem to be
> due to the wrong usage of stride- vs. byte-based offsets. For instance in
> SELECT TYPE .. TYPE IS, where one has a BT_TYPE and which fails.
>
> I think the conversion to byte offsets will be the next step. Afterwards, we
> should get rid of the offset.

The middle-end POINTER_PLUS_EXPR requires byte offsets for all
pointer types, so you can avoid the conversion to char * if ptr is already
a pointer type.

Richard.

> Tobias