rfc: auto-casted vector types

rfc: auto-casted vector types

[Aldy Hernandez]

The PPC E500 API specifies that vector types of type __ev64_opaque__ 
(V2SImode) can interconvert with other vector types without a cast.

For example:

	#define __vector __attribute__((vector_size(8)))
	typedef int                     __vector __ev64_opaque__;
	typedef float                   __vector __ev64_fs__;
	typedef int						__vector __ev64_int__;

	__ev64_fs__ f;
	__ev64_int__ i;
	__ev64_opaque__ op;

	/* Do not require a cast.  */
	op = f;
	f = op;
	op = i;
	i = op;

I had originally kludged a solution in comptypes() and 
convert_for_assignment() because I thought nobody would be interested 
(or would agree), but Nick has since cleaned up my patches a bit, and 
they don't look that bad: three lines of actual generic code.

I'd like to get input on:

	(a) the feasibility of including these patches
	(b) alternate solutions that would be accepted
	(c) including the patch below in contrib/

As you can probably guess, I didn't design the API, so no flames please 
;-).

Comments?

2003-02-13  Nick Clifton  <nickc@redhat.com>
	    Aldy Hernandez  <aldyh@redhat.com>

	* doc/tm.texi: Document TARGET_VECTOR_TYPES_COMPATIBLE.

	* c-typeck.c (comptypes): Take into account
	TARGET_VECTOR_TYPES_COMPATIBLE.
	(convert_for_assignment): Same.

	* config/rs6000/rs6000-protos.h
	(rs6000_spe_vector_types_compatible): Prototype.

	* config/rs6000/rs6000.c (is_ev64_opaque_type): New.
	(rs6000_spe_vector_types_compatible): New.

	* config/rs6000/rs6000.h (TARGET_VECTOR_TYPES_COMPATIBLE): Define.

Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/gcc/gcc/doc/tm.texi,v
retrieving revision 1.198
diff -c -p -r1.198 tm.texi
*** doc/tm.texi	31 Jan 2003 23:34:17 -0000	1.198
--- doc/tm.texi	13 Feb 2003 20:50:20 -0000
*************** Define this macro to be nonzero if the p
*** 1302,1307 ****
--- 1302,1312 ----
   involving vector mode @var{mode}.  At the very least, it must have 
move
   patterns for this mode.

+ @findex TARGET_VECTOR_TYPES_COMPATIBLE
+ @item TARGET_VECTOR_TYPES_COMPATIBLE(@var{tree1}, @var{tree2})
+ Define this macro to return nonzero if no cast is needed when copying
+ a value of type @var{tree1} to a value of type @var{tree2}.
+
   @findex STACK_SAVEAREA_MODE
   @item STACK_SAVEAREA_MODE (@var{save_level})
   If defined, an expression of type @code{enum machine_mode} that
Index: c-typeck.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/c-typeck.c,v
retrieving revision 1.218
diff -c -p -r1.218 c-typeck.c
*** c-typeck.c	1 Feb 2003 13:09:39 -0000	1.218
--- c-typeck.c	13 Feb 2003 20:50:35 -0000
*************** comptypes (type1, type2)
*** 574,579 ****
--- 574,586 ----
   	val = 1;
         break;

+     case VECTOR_TYPE:
+ #ifdef TARGET_VECTOR_TYPES_COMPATIBLE
+       /* The target might allow certain vector types to be 
compatible.  */
+       val = TARGET_VECTOR_TYPES_COMPATIBLE (t1, t2);
+ #endif
+       break;
+
       default:
         break;
       }
*************** convert_for_assignment (type, rhs, errty
*** 4061,4066 ****
--- 4068,4079 ----
         rhs = build1 (NOP_EXPR, type, rhs);
         return rhs;
       }
+ #ifdef TARGET_VECTOR_TYPES_COMPATIBLE
+   /* Some types can interconvert without explicit casts.  */
+   else if (codel == VECTOR_TYPE && coder == VECTOR_TYPE
+ 	   && TARGET_VECTOR_TYPES_COMPATIBLE (type, rhstype))
+     return convert (type, rhs);
+ #endif
     /* Arithmetic types all interconvert, and enum is treated like int. 
  */
     else if ((codel == INTEGER_TYPE || codel == REAL_TYPE
   	    || codel == ENUMERAL_TYPE || codel == COMPLEX_TYPE
Index: config/rs6000/rs6000.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/rs6000/rs6000.c,v
retrieving revision 1.422
diff -c -p -r1.422 rs6000.c
*** config/rs6000/rs6000.c	10 Feb 2003 18:42:19 -0000	1.422
--- config/rs6000/rs6000.c	13 Feb 2003 20:51:03 -0000
*************** static void is_altivec_return_reg PARAMS
*** 265,270 ****
--- 265,271 ----
   static rtx generate_set_vrsave PARAMS ((rtx, rs6000_stack_t *, int));
   static void altivec_frame_fixup PARAMS ((rtx, rtx, HOST_WIDE_INT));
   static int easy_vector_constant PARAMS ((rtx));
+ static int is_ev64_opaque_type PARAMS ((tree));

   /* Hash table stuff for keeping track of TOC entries.  */

*************** rs6000_memory_move_cost (mode, class, in
*** 13522,13527 ****
--- 13523,13556 ----
       return 4 * HARD_REGNO_NREGS (FIRST_ALTIVEC_REGNO, mode);
     else
       return 4 + rs6000_register_move_cost (mode, class, GENERAL_REGS);
+ }
+
+ static int
+ is_ev64_opaque_type (type)
+      tree type;
+ {
+   return (TYPE_NAME (type)
+ 	  && TREE_CODE (TYPE_NAME (type)) == TYPE_DECL
+ 	  && DECL_NAME (TYPE_NAME (type))
+ 	  && strcmp (IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (type))),
+ 		     "__ev64_opaque__") == 0);
+ }
+
+ /* Return true iff vector type1 can be converted into vector type2.  
*/
+
+ int
+ rs6000_spe_vector_types_compatible (t1, t2)
+      tree t1;
+      tree t2;
+ {
+   if (TREE_CODE (t1) != VECTOR_TYPE || TREE_CODE (t2) != VECTOR_TYPE)
+     return 0;
+
+   if (TYPE_NAME (t1) || TYPE_NAME (t2))
+     return is_ev64_opaque_type (t1) || is_ev64_opaque_type (t2);
+
+   /* The V2SI type is the basis of the ev64_opaque type.  */
+   return t1 == V2SI_type_node || t2 == V2SI_type_node;
   }

   #include "gt-rs6000.h"
Index: config/rs6000/rs6000.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/rs6000/rs6000.h,v
retrieving revision 1.251
diff -c -p -r1.251 rs6000.h
*** config/rs6000/rs6000.h	8 Feb 2003 03:59:40 -0000	1.251
--- config/rs6000/rs6000.h	13 Feb 2003 20:51:11 -0000
*************** extern int rs6000_default_long_calls;
*** 892,897 ****
--- 892,901 ----
            || (MODE) == V1DImode          \
            || (MODE) == V2SImode)

+ /* Allow SPE vector types to be inter-converted.  */
+ #define TARGET_VECTOR_TYPES_COMPATIBLE(t1,t2) \
+    (TARGET_SPE ? rs6000_spe_vector_types_compatible (t1, t2) : 0)
+
   /* Define this macro to be nonzero if the port is prepared to handle
      insns involving vector mode MODE.  At the very least, it must have
      move patterns for this mode.  */

Re: rfc: auto-casted vector types

[Aldy Hernandez]

ping.

could i get some comments on this, or the patches look sensible enough 
to submit for inclusion?

aldy

On Thursday, February 13, 2003, at 01:01  PM, Aldy Hernandez wrote:

> The PPC E500 API specifies that vector types of type __ev64_opaque__ 
> (V2SImode) can interconvert with other vector types without a cast.
>
> For example:
>
> 	#define __vector __attribute__((vector_size(8)))
> 	typedef int                     __vector __ev64_opaque__;
> 	typedef float                   __vector __ev64_fs__;
> 	typedef int						__vector __ev64_int__;
>
> 	__ev64_fs__ f;
> 	__ev64_int__ i;
> 	__ev64_opaque__ op;
>
> 	/* Do not require a cast.  */
> 	op = f;
> 	f = op;
> 	op = i;
> 	i = op;
>
> I had originally kludged a solution in comptypes() and 
> convert_for_assignment() because I thought nobody would be interested 
> (or would agree), but Nick has since cleaned up my patches a bit, and 
> they don't look that bad: three lines of actual generic code.
>
> I'd like to get input on:
>
> 	(a) the feasibility of including these patches
> 	(b) alternate solutions that would be accepted
> 	(c) including the patch below in contrib/
>
> As you can probably guess, I didn't design the API, so no flames 
> please ;-).
>
> Comments?
>
> 2003-02-13  Nick Clifton  <nickc@redhat.com>
> 	    Aldy Hernandez  <aldyh@redhat.com>
>
> 	* doc/tm.texi: Document TARGET_VECTOR_TYPES_COMPATIBLE.
>
> 	* c-typeck.c (comptypes): Take into account
> 	TARGET_VECTOR_TYPES_COMPATIBLE.
> 	(convert_for_assignment): Same.
>
> 	* config/rs6000/rs6000-protos.h
> 	(rs6000_spe_vector_types_compatible): Prototype.
>
> 	* config/rs6000/rs6000.c (is_ev64_opaque_type): New.
> 	(rs6000_spe_vector_types_compatible): New.
>
> 	* config/rs6000/rs6000.h (TARGET_VECTOR_TYPES_COMPATIBLE): Define.
>
> Index: doc/tm.texi
> ===================================================================
> RCS file: /cvs/gcc/gcc/gcc/doc/tm.texi,v
> retrieving revision 1.198
> diff -c -p -r1.198 tm.texi
> *** doc/tm.texi	31 Jan 2003 23:34:17 -0000	1.198
> --- doc/tm.texi	13 Feb 2003 20:50:20 -0000
> *************** Define this macro to be nonzero if the p
> *** 1302,1307 ****
> --- 1302,1312 ----
>   involving vector mode @var{mode}.  At the very least, it must have 
> move
>   patterns for this mode.
>
> + @findex TARGET_VECTOR_TYPES_COMPATIBLE
> + @item TARGET_VECTOR_TYPES_COMPATIBLE(@var{tree1}, @var{tree2})
> + Define this macro to return nonzero if no cast is needed when copying
> + a value of type @var{tree1} to a value of type @var{tree2}.
> +
>   @findex STACK_SAVEAREA_MODE
>   @item STACK_SAVEAREA_MODE (@var{save_level})
>   If defined, an expression of type @code{enum machine_mode} that
> Index: c-typeck.c
> ===================================================================
> RCS file: /cvs/gcc/gcc/gcc/c-typeck.c,v
> retrieving revision 1.218
> diff -c -p -r1.218 c-typeck.c
> *** c-typeck.c	1 Feb 2003 13:09:39 -0000	1.218
> --- c-typeck.c	13 Feb 2003 20:50:35 -0000
> *************** comptypes (type1, type2)
> *** 574,579 ****
> --- 574,586 ----
>   	val = 1;
>         break;
>
> +     case VECTOR_TYPE:
> + #ifdef TARGET_VECTOR_TYPES_COMPATIBLE
> +       /* The target might allow certain vector types to be 
> compatible.  */
> +       val = TARGET_VECTOR_TYPES_COMPATIBLE (t1, t2);
> + #endif
> +       break;
> +
>       default:
>         break;
>       }
> *************** convert_for_assignment (type, rhs, errty
> *** 4061,4066 ****
> --- 4068,4079 ----
>         rhs = build1 (NOP_EXPR, type, rhs);
>         return rhs;
>       }
> + #ifdef TARGET_VECTOR_TYPES_COMPATIBLE
> +   /* Some types can interconvert without explicit casts.  */
> +   else if (codel == VECTOR_TYPE && coder == VECTOR_TYPE
> + 	   && TARGET_VECTOR_TYPES_COMPATIBLE (type, rhstype))
> +     return convert (type, rhs);
> + #endif
>     /* Arithmetic types all interconvert, and enum is treated like 
> int.  */
>     else if ((codel == INTEGER_TYPE || codel == REAL_TYPE
>   	    || codel == ENUMERAL_TYPE || codel == COMPLEX_TYPE
> Index: config/rs6000/rs6000.c
> ===================================================================
> RCS file: /cvs/gcc/gcc/gcc/config/rs6000/rs6000.c,v
> retrieving revision 1.422
> diff -c -p -r1.422 rs6000.c
> *** config/rs6000/rs6000.c	10 Feb 2003 18:42:19 -0000	1.422
> --- config/rs6000/rs6000.c	13 Feb 2003 20:51:03 -0000
> *************** static void is_altivec_return_reg PARAMS
> *** 265,270 ****
> --- 265,271 ----
>   static rtx generate_set_vrsave PARAMS ((rtx, rs6000_stack_t *, int));
>   static void altivec_frame_fixup PARAMS ((rtx, rtx, HOST_WIDE_INT));
>   static int easy_vector_constant PARAMS ((rtx));
> + static int is_ev64_opaque_type PARAMS ((tree));
>
>   /* Hash table stuff for keeping track of TOC entries.  */
>
> *************** rs6000_memory_move_cost (mode, class, in
> *** 13522,13527 ****
> --- 13523,13556 ----
>       return 4 * HARD_REGNO_NREGS (FIRST_ALTIVEC_REGNO, mode);
>     else
>       return 4 + rs6000_register_move_cost (mode, class, GENERAL_REGS);
> + }
> +
> + static int
> + is_ev64_opaque_type (type)
> +      tree type;
> + {
> +   return (TYPE_NAME (type)
> + 	  && TREE_CODE (TYPE_NAME (type)) == TYPE_DECL
> + 	  && DECL_NAME (TYPE_NAME (type))
> + 	  && strcmp (IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (type))),
> + 		     "__ev64_opaque__") == 0);
> + }
> +
> + /* Return true iff vector type1 can be converted into vector type2.  
> */
> +
> + int
> + rs6000_spe_vector_types_compatible (t1, t2)
> +      tree t1;
> +      tree t2;
> + {
> +   if (TREE_CODE (t1) != VECTOR_TYPE || TREE_CODE (t2) != VECTOR_TYPE)
> +     return 0;
> +
> +   if (TYPE_NAME (t1) || TYPE_NAME (t2))
> +     return is_ev64_opaque_type (t1) || is_ev64_opaque_type (t2);
> +
> +   /* The V2SI type is the basis of the ev64_opaque type.  */
> +   return t1 == V2SI_type_node || t2 == V2SI_type_node;
>   }
>
>   #include "gt-rs6000.h"
> Index: config/rs6000/rs6000.h
> ===================================================================
> RCS file: /cvs/gcc/gcc/gcc/config/rs6000/rs6000.h,v
> retrieving revision 1.251
> diff -c -p -r1.251 rs6000.h
> *** config/rs6000/rs6000.h	8 Feb 2003 03:59:40 -0000	1.251
> --- config/rs6000/rs6000.h	13 Feb 2003 20:51:11 -0000
> *************** extern int rs6000_default_long_calls;
> *** 892,897 ****
> --- 892,901 ----
>            || (MODE) == V1DImode          \
>            || (MODE) == V2SImode)
>
> + /* Allow SPE vector types to be inter-converted.  */
> + #define TARGET_VECTOR_TYPES_COMPATIBLE(t1,t2) \
> +    (TARGET_SPE ? rs6000_spe_vector_types_compatible (t1, t2) : 0)
> +
>   /* Define this macro to be nonzero if the port is prepared to handle
>      insns involving vector mode MODE.  At the very least, it must have
>      move patterns for this mode.  */
>

Re: rfc: auto-casted vector types

[Zack Weinberg]

Aldy Hernandez <aldy@andrews.edu> writes:

> ping.
>
> could i get some comments on this, or the patches look sensible enough
> to submit for inclusion?

It looks decent to me.  I would suggest that here

>> + @findex TARGET_VECTOR_TYPES_COMPATIBLE
>> + @item TARGET_VECTOR_TYPES_COMPATIBLE(@var{tree1}, @var{tree2})
>> + Define this macro to return nonzero if no cast is needed when copying
>> + a value of type @var{tree1} to a value of type @var{tree2}.

you say that this works only for vector types (I realize this is
implied by the name, but it should be said explicitly anyway); and
rather than putting #ifdefs in c-typeck.c, make this a target hook
whose default is 'always return 0'.

zw

Re: rfc: auto-casted vector types

[Jim Wilson]

Aldy Hernandez <aldyh@redhat.com> writes:

> The PPC E500 API specifies that vector types of type __ev64_opaque__
> (V2SImode) can interconvert with other vector types without a cast.

Is there any publicly available documentation?  My copy of the draft docs
say they are Motorola confidential.  If there is documentation available,
it would be useful to point to it somewhere.

I tried searching the Motorola web site.  I found Altivec PIM documentation,
but no SPE PIM documentation.

> 	* doc/tm.texi: Document TARGET_VECTOR_TYPES_COMPATIBLE.
> 	* c-typeck.c (comptypes): Take into account
> 	TARGET_VECTOR_TYPES_COMPATIBLE.
> 	(convert_for_assignment): Same.
>       ...

The patch looks pretty reasonable to me also.  I agree with Zack's suggestions.

It might be useful to note in the comments in the rs6000 file that the
current implementation has a known flaw.  It pretends that V2SI is the opaque
vector type.  This means that it accidentally allows conversions to/from the
V2SI vector type without explicit casts.  In order to fix this, we need to
use a different otherwise unused vector type for the opaque vector type.
Nick started on a patch to do this, but didn't have enough time to finish it.

Jim

Re: rfc: auto-casted vector types

[Kumar Gala]

The SPE PIM docs are not currently publicly available.

- kumar

On Monday, February 17, 2003, at 04:25 PM, Jim Wilson wrote:

> Aldy Hernandez <aldyh@redhat.com> writes:
>
>> The PPC E500 API specifies that vector types of type __ev64_opaque__
>> (V2SImode) can interconvert with other vector types without a cast.
>
> Is there any publicly available documentation?  My copy of the draft 
> docs
> say they are Motorola confidential.  If there is documentation 
> available,
> it would be useful to point to it somewhere.
>
> I tried searching the Motorola web site.  I found Altivec PIM 
> documentation,
> but no SPE PIM documentation.
>
>> 	* doc/tm.texi: Document TARGET_VECTOR_TYPES_COMPATIBLE.
>> 	* c-typeck.c (comptypes): Take into account
>> 	TARGET_VECTOR_TYPES_COMPATIBLE.
>> 	(convert_for_assignment): Same.
>>       ...
>
> The patch looks pretty reasonable to me also.  I agree with Zack's 
> suggestions.
>
> It might be useful to note in the comments in the rs6000 file that the
> current implementation has a known flaw.  It pretends that V2SI is the 
> opaque
> vector type.  This means that it accidentally allows conversions 
> to/from the
> V2SI vector type without explicit casts.  In order to fix this, we 
> need to
> use a different otherwise unused vector type for the opaque vector 
> type.
> Nick started on a patch to do this, but didn't have enough time to 
> finish it.
>
> Jim

Re: rfc: auto-casted vector types

[Kumar Gala]

> The patch looks pretty reasonable to me also.  I agree with Zack's 
> suggestions.
>
> It might be useful to note in the comments in the rs6000 file that the
> current implementation has a known flaw.  It pretends that V2SI is the 
> opaque
> vector type.  This means that it accidentally allows conversions 
> to/from the
> V2SI vector type without explicit casts.  In order to fix this, we 
> need to
> use a different otherwise unused vector type for the opaque vector 
> type.
> Nick started on a patch to do this, but didn't have enough time to 
> finish it.

If this is not successful can we make the opaque type map to a 64-bit 
data type(V1SI?) instead of the V2SI.  I think this would be more 
natural if we have to provide something for opaque.

eg

__ev64_opaque__ x = { 0, 3 }; // should fail
__ev64_opaque__ y = { 3 }; // should be ok

- kumar

Re: rfc: auto-casted vector types

[Geoff Keating]

Jim Wilson <wilson@redhat.com> writes:

> It might be useful to note in the comments in the rs6000 file that the
> current implementation has a known flaw.  It pretends that V2SI is the opaque
> vector type.  This means that it accidentally allows conversions to/from the
> V2SI vector type without explicit casts.  In order to fix this, we need to
> use a different otherwise unused vector type for the opaque vector type.
> Nick started on a patch to do this, but didn't have enough time to finish it.

This shouldn't be done at the RTL level at all; the right thing to do
is to create a special '__ev64_opaque__' type which is different from
'vector int' and just happens to also have V2SImode (or DImode or
whatever is convenient---I'd vote for DImode since that's what GCC
uses for all-purpose "blob of data" types).

-- 
- Geoffrey Keating <geoffk@geoffk.org>

patch: Re: rfc: auto-casted vector types

[Aldy Hernandez]

> Is there any publicly available documentation?  My copy of the draft 
> docs
> say they are Motorola confidential.  If there is documentation 
> available,
> it would be useful to point to it somewhere.

According to Kumar, no.

> The patch looks pretty reasonable to me also.  I agree with Zack's 
> suggestions.

Done.

> It might be useful to note in the comments in the rs6000 file that the
> current implementation has a known flaw.  It pretends that V2SI is the 
> opaque
> vector type.  This means that it accidentally allows conversions 
> to/from the
> V2SI vector type without explicit casts.  In order to fix this, we 
> need to

I added a comment.

> use a different otherwise unused vector type for the opaque vector 
> type.
> Nick started on a patch to do this, but didn't have enough time to 
> finish it.

As I recall he used V8QI, but that had other problems.  V1DI, as Kumar 
suggests,
is probably a better approach, but will require heavily massaging 
spe.h/spe.md.  I'll try that later, since that will probably break 
other things I don't have time to fix right now ;-).

OK, I'm formally submitting the patch now.  I've built an spe toolchain 
with this, and a generic powerpc-linux bootstrap/regtest is in progress.

OK to install when tests finish?

Aldy

2003-02-17  Nick Clifton  <nickc@redhat.com>
	    Aldy Hernandez  <aldyh@redhat.com>

	* defaults.h (TARGET_VECTOR_TYPES_COMPATIBLE): Define.

	* doc/tm.texi: Document TARGET_VECTOR_TYPES_COMPATIBLE.

	* c-typeck.c (comptypes): Take into account
	TARGET_VECTOR_TYPES_COMPATIBLE.
	(convert_for_assignment): Same.
	(really_start_incremental_init): Disallow initialization of
	ev64_opaque types.

	* config/rs6000/rs6000-protos.h
	(rs6000_spe_vector_types_compatible): Prototype.

	* config/rs6000/rs6000.c (is_ev64_opaque_type): New.
	(rs6000_spe_vector_types_compatible): New.

	* config/rs6000/rs6000.h (TARGET_VECTOR_TYPES_COMPATIBLE): Define.

Index: defaults.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/defaults.h,v
retrieving revision 1.101
diff -c -p -r1.101 defaults.h
*** defaults.h	24 Jan 2003 22:07:00 -0000	1.101
--- defaults.h	18 Feb 2003 00:09:35 -0000
*************** You Lose!  You must define PREFERRED_DEB
*** 595,600 ****
--- 595,607 ----
   #define VECTOR_MODE_SUPPORTED_P(MODE) 0
   #endif

+ /*  Define this macro to return nonzero if no cast is needed when
+     copying a vector value of type TYPE1 to a vector value of
+     type TYPE2.  */
+ #ifndef TARGET_VECTOR_TYPES_COMPATIBLE
+ #define TARGET_VECTOR_TYPES_COMPATIBLE(TYPE1, TYPE2) 0
+ #endif
+
   /* Determine whether __cxa_atexit, rather than atexit, is used to
      register C++ destructors for local statics and global objects.  */
   #ifndef DEFAULT_USE_CXA_ATEXIT
Index: c-typeck.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/c-typeck.c,v
retrieving revision 1.218
diff -c -p -r1.218 c-typeck.c
*** c-typeck.c	1 Feb 2003 13:09:39 -0000	1.218
--- c-typeck.c	18 Feb 2003 00:09:41 -0000
*************** comptypes (type1, type2)
*** 574,579 ****
--- 574,584 ----
   	val = 1;
         break;

+     case VECTOR_TYPE:
+       /* The target might allow certain vector types to be 
compatible.  */
+       val = TARGET_VECTOR_TYPES_COMPATIBLE (t1, t2);
+       break;
+
       default:
         break;
       }
*************** convert_for_assignment (type, rhs, errty
*** 4061,4066 ****
--- 4066,4075 ----
         rhs = build1 (NOP_EXPR, type, rhs);
         return rhs;
       }
+   /* Some types can interconvert without explicit casts.  */
+   else if (codel == VECTOR_TYPE && coder == VECTOR_TYPE
+ 	   && TARGET_VECTOR_TYPES_COMPATIBLE (type, rhstype))
+     return convert (type, rhs);
     /* Arithmetic types all interconvert, and enum is treated like int. 
  */
     else if ((codel == INTEGER_TYPE || codel == REAL_TYPE
   	    || codel == ENUMERAL_TYPE || codel == COMPLEX_TYPE
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/gcc/gcc/doc/tm.texi,v
retrieving revision 1.198
diff -c -p -r1.198 tm.texi
*** doc/tm.texi	31 Jan 2003 23:34:17 -0000	1.198
--- doc/tm.texi	18 Feb 2003 00:10:20 -0000
*************** Define this macro to be nonzero if the p
*** 1302,1307 ****
--- 1302,1313 ----
   involving vector mode @var{mode}.  At the very least, it must have 
move
   patterns for this mode.

+ @findex TARGET_VECTOR_TYPES_COMPATIBLE
+ @item TARGET_VECTOR_TYPES_COMPATIBLE(@var{tree1}, @var{tree2})
+ Define this macro to return nonzero if no cast is needed when copying
+ a vector value of type @var{tree1} to a vector value of type
+ @var{tree2}.
+
   @findex STACK_SAVEAREA_MODE
   @item STACK_SAVEAREA_MODE (@var{save_level})
   If defined, an expression of type @code{enum machine_mode} that
Index: config/rs6000/rs6000.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/rs6000/rs6000.c,v
retrieving revision 1.422
diff -c -p -r1.422 rs6000.c
*** config/rs6000/rs6000.c	10 Feb 2003 18:42:19 -0000	1.422
--- config/rs6000/rs6000.c	18 Feb 2003 00:10:29 -0000
*************** static void is_altivec_return_reg PARAMS
*** 265,270 ****
--- 265,271 ----
   static rtx generate_set_vrsave PARAMS ((rtx, rs6000_stack_t *, int));
   static void altivec_frame_fixup PARAMS ((rtx, rtx, HOST_WIDE_INT));
   static int easy_vector_constant PARAMS ((rtx));
+ static int is_ev64_opaque_type PARAMS ((tree));

   /* Hash table stuff for keeping track of TOC entries.  */

*************** rs6000_memory_move_cost (mode, class, in
*** 13522,13527 ****
--- 13523,13558 ----
       return 4 * HARD_REGNO_NREGS (FIRST_ALTIVEC_REGNO, mode);
     else
       return 4 + rs6000_register_move_cost (mode, class, GENERAL_REGS);
+ }
+
+ static int
+ is_ev64_opaque_type (type)
+      tree type;
+ {
+   return (TYPE_NAME (type)
+ 	  && TREE_CODE (TYPE_NAME (type)) == TYPE_DECL
+ 	  && DECL_NAME (TYPE_NAME (type))
+ 	  && strcmp (IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (type))),
+ 		     "__ev64_opaque__") == 0);
+ }
+
+ /* Return true iff vector type1 can be converted into vector type2.  
*/
+
+ int
+ rs6000_spe_vector_types_compatible (t1, t2)
+      tree t1;
+      tree t2;
+ {
+   if (TREE_CODE (t1) != VECTOR_TYPE || TREE_CODE (t2) != VECTOR_TYPE)
+     return 0;
+
+   if (TYPE_NAME (t1) || TYPE_NAME (t2))
+     return is_ev64_opaque_type (t1) || is_ev64_opaque_type (t2);
+
+   /* FIXME: We assume V2SI is the opaque type, so we accidentally
+      allow inter conversion to and from V2SI modes.  We could use
+      V1D1, and rewrite <spe.h> accordingly.  */
+   return t1 == V2SI_type_node || t2 == V2SI_type_node;
   }

   #include "gt-rs6000.h"
Index: config/rs6000/rs6000.h
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/rs6000/rs6000.h,v
retrieving revision 1.251
diff -c -p -r1.251 rs6000.h
*** config/rs6000/rs6000.h	8 Feb 2003 03:59:40 -0000	1.251
--- config/rs6000/rs6000.h	18 Feb 2003 00:10:41 -0000
*************** extern int rs6000_default_long_calls;
*** 892,897 ****
--- 892,901 ----
            || (MODE) == V1DImode          \
            || (MODE) == V2SImode)

+ /* Allow SPE vector types to be inter-converted.  */
+ #define TARGET_VECTOR_TYPES_COMPATIBLE(t1,t2) \
+    (TARGET_SPE ? rs6000_spe_vector_types_compatible (t1, t2) : 0)
+
   /* Define this macro to be nonzero if the port is prepared to handle
      insns involving vector mode MODE.  At the very least, it must have
      move patterns for this mode.  */

Re: rfc: auto-casted vector types

[Mark Mitchell]

--On Monday, February 17, 2003 11:16:34 AM -0800 Aldy Hernandez 
<aldy@andrews.edu> wrote:

> ping.
>
> could i get some comments on this, or the patches look sensible enough to
> submit for inclusion?

To me, this approach looks plausible.  I'd rather we use #if than #ifdef; 
can't that TARGET_* stuff go in the target hooks bits?  Also, you need some 
comments for the new functions.

I don't object to the patch in principle, though.

-- 
Mark Mitchell                mark@codesourcery.com
CodeSourcery, LLC            http://www.codesourcery.com

Re: rfc: auto-casted vector types

[Aldy Hernandez]

>>>>> "Geoff" == Geoff Keating <geoffk@geoffk.org> writes:

 > Jim Wilson <wilson@redhat.com> writes:
 >> It might be useful to note in the comments in the rs6000 file that the
 >> current implementation has a known flaw.  It pretends that V2SI is the opaque
 >> vector type.  This means that it accidentally allows conversions to/from the
 >> V2SI vector type without explicit casts.  In order to fix this, we need to
 >> use a different otherwise unused vector type for the opaque vector type.
 >> Nick started on a patch to do this, but didn't have enough time to finish it.

 > This shouldn't be done at the RTL level at all; the right thing to do
 > is to create a special '__ev64_opaque__' type which is different from
 > 'vector int' and just happens to also have V2SImode (or DImode or
 > whatever is convenient---I'd vote for DImode since that's what GCC
 > uses for all-purpose "blob of data" types).

Actually, I agree with this because I just ran into a similar problem
this morning with an upcoming patch that I solved by looking at the
type instead.

I'll put it on my TODO list.

Aldy