Patchwork [fortran] PR fortran/53732

login
register
mail settings
Submitter Mikael Morin
Date July 3, 2012, 7:56 p.m.
Message ID <4FF34E86.7020006@sfr.fr>
Download mbox | patch
Permalink /patch/168853/
State New
Headers show

Comments

Mikael Morin - July 3, 2012, 7:56 p.m.
Hello,

Revision 180898:
http://gcc.gnu.org/viewcvs?view=revision&revision=180898
introduced at the end of gfc_add_loop_ss_code some self recursive calls 
to handle reductions' loops (like in sum for example).
However, as gfc_add_loop_ss_code already calls itself to handle scalar 
and vector subscripts, the new calls had to be disabled if there was 
such a subscript, as in that case the reduction(s) would have been 
handled from within the subscript call.
Back then, I didn't pay attention, however, to the fact that the 
subscript flag was set to true in all the subscript calls, so the 
reductions are then handled with subscript set to true, which introduces 
an unwanted conversion to gfc_array_index_type and the failure leading 
to the PR.

The patch below (which is the one posted on bugzilla) disables reduction 
handling if subscript is true, so that the reductions are handled from 
the outer call instead of from the subscript call.
I have finally convinced myself that it is always right, as if there is 
a reduction in a subscript, gfc_add_loop_ss_code will call 
gfc_conv_expr_descriptor which will create a temporary, and the 
reduction will be handled separatedly in a new loop initializing the 
temporary.

Regression tested on x86_64-unknown-freebsd9.0.
OK for trunk/4.7?

Mikael
2012-07-15  Mikael Morin  <mikael@gcc.gnu.org>

	PR fortran/53732
	* trans-array.c (gfc_add_loop_ss_code): Disable self recursive calls
	handling nested loop(s) if the subscript flag is true.
2012-07-03  Mikael Morin  <mikael@gcc.gnu.org>

	PR fortran/53732
	* gfortran.dg/inline_sum_4.f90: New test.
! { dg-do compile }
!
! PR fortran/53732
! this was leading to an internal "mismatching comparison operand types" error
!
! Contributed by minzastro <minzastro@googlemail.com>

program test
implicit none

real*8 arr(4, 4, 4, 4)

arr(:,:,:,:) = 1d0

arr(1, :, :, = sum(arr, dim=1, mask=(arr(:,:,:,:) > 0d0))

end program test
Tobias Burnus - July 4, 2012, 7:29 a.m.
Hello Mikael,

On 07/03/2012 09:56 PM, Mikael Morin wrote:
> The patch below (which is the one posted on bugzilla) disables 
> reduction handling if subscript is true, so that the reductions are 
> handled from the outer call instead of from the subscript call.
> I have finally convinced myself that it is always right, as if there 
> is a reduction in a subscript, gfc_add_loop_ss_code will call 
> gfc_conv_expr_descriptor which will create a temporary, and the 
> reduction will be handled separatedly in a new loop initializing the 
> temporary.
>
> Regression tested on x86_64-unknown-freebsd9.0.
> OK for trunk/4.7?

OK. Thanks for the patch. (I believe the patch is correct; I am not 100% 
sure that it is as the scalarizer tends to confuse me.)

Tobias

Patch

diff --git a/trans-array.c b/trans-array.c
index f135af1..d289ac3 100644
--- a/trans-array.c
+++ b/trans-array.c
@@ -2398,7 +2398,6 @@  gfc_add_loop_ss_code (gfc_loopinfo * loop, gfc_ss * ss, bool subscript,
   gfc_ss_info *ss_info;
   gfc_array_info *info;
   gfc_expr *expr;
-  bool skip_nested = false;
   int n;
 
   /* Don't evaluate the arguments for realloc_lhs_loop_for_fcn_call; otherwise,
@@ -2487,12 +2486,7 @@  gfc_add_loop_ss_code (gfc_loopinfo * loop, gfc_ss * ss, bool subscript,
 	  /* Add the expressions for scalar and vector subscripts.  */
 	  for (n = 0; n < GFC_MAX_DIMENSIONS; n++)
 	    if (info->subscript[n])
-	      {
-		gfc_add_loop_ss_code (loop, info->subscript[n], true, where);
-		/* The recursive call will have taken care of the nested loops.
-		   No need to do it twice.  */
-		skip_nested = true;
-	      }
+	      gfc_add_loop_ss_code (loop, info->subscript[n], true, where);
 
 	  set_vector_loop_bounds (ss);
 	  break;
@@ -2548,7 +2542,7 @@  gfc_add_loop_ss_code (gfc_loopinfo * loop, gfc_ss * ss, bool subscript,
 	}
     }
 
-  if (!skip_nested)
+  if (!subscript)
     for (nested_loop = loop->nested; nested_loop;
 	 nested_loop = nested_loop->next)
       gfc_add_loop_ss_code (nested_loop, nested_loop->ss, subscript, where);