Re: RFC/Patch for PR fortran/17675

[ja2morri at csclub dot uwaterloo dot ca (James A. Morrison)]

Steve Ellcey <sje@cup.hp.com> writes:

> Here is a patch I created to fix PR fortran/17675.  I have some
> questions about it and I imagine some changes will be desired before it
> is checked in.  On IA64 HP-UX this patch reduced my fortran test suite
> failures from 95 to 31.

 Cool, I've got a build going on sparc-linux.
 
> This code is only activated for machines with STRICT_ALIGNMENT set,
> should we try to do anything for machines where SLOW_UNALIGNED_ACCESS
> is set?  Should we add padding for all machines?  If alignment doesn't
> matter on a machine then I assume the alignment for all types would
> be the same and this code would never add any padding.
> 
> What is the best way to determine the correct padding to use.  I simply
> found the type with the largest required alignment, added sufficient
> padding to align that field, and let the rest fall where they had to.
> Is there a better way to ensure that the maximum amount of data is
> properly aligned?
> 
> Another question I had was that this change is going to result in a
> union tree type being built by create_common with no field at offset
> zero.  I don't know if that is OK or not.  Is a dummy field needed or is
> it OK to have a hole in the tree that is created to represent a common
> block.
> 
> Comments on the patch?

 Since you have everything aligned already, you could put an assert
in build_field like:
        gcc_assert (known_align >= desired_align || !STRICT_ALIGNMENT);
 
> Steve Ellcey
> sje@cup.hp.com
> 
> 2004-12-20  Steve Ellcey  <sje@cup.hp.com>
> 
> 	* gcc/fortran/trans-common.c (finish_equivalences):  Add padding
> 	to front of common to try and align contents.

> 
> *** gcc.orig/gcc/fortran/trans-common.c	Mon Dec 20 16:03:31 2004
> --- gcc/gcc/fortran/trans-common.c	Mon Dec 20 16:31:23 2004
> *************** finish_equivalences (gfc_namespace *ns)
> *** 753,758 ****
> --- 753,789 ----
>           for (v = current_segment; v; v = v->next)
>   	  v->offset -= min_offset;
>   
> + 	if (STRICT_ALIGNMENT)
> + 	  {
> + 	    /* Find the type with the strictest alignment requirements and
> + 	       ensure it is aligned.  Hope that makes most of the others
> + 	       fall into place.  */
> + 
> + 	    unsigned HOST_WIDE_INT max_alignment = 0;
> +   	    HOST_WIDE_INT max_align_offset = 0;
> + 	    HOST_WIDE_INT padding = 0;
> + 	    for (v = current_segment; v; v = v->next)
> + 	      {
> + 	        if (TYPE_ALIGN (v->field) > max_alignment)
> + 		  {
> + 		    max_alignment = TYPE_ALIGN (v->field);
> + 		    max_align_offset = v->offset;
> + 		  }
> + 	      }
> + 
> + 	    /* If the field with the largest needed alignment isn't already
> + 	       aligned, add padding to front of common (increase everyone's
> + 	       offset) to align that field and hope everything else falls into
> + 	       place.  */
> + 
> + 	    if ((max_align_offset % max_alignment) != 0)
> + 	      {
> + 	        padding = max_alignment - (max_align_offset % max_alignment);
> + 	        for (v = current_segment; v; v = v->next)
> + 		  v->offset += padding;
> + 	      }
> + 	  }
> + 
>           current_common = current_segment;
>           create_common (NULL);
>           break;
> 


-- 
Thanks,
Jim

http://www.student.cs.uwaterloo.ca/~ja2morri/
http://phython.blogspot.com
http://open.nit.ca/wiki/?page=jim