[patch] Fix fortran access to special register on SPU arch

[patch] Fix fortran access to special register on SPU arch

[Markus Deuling]

[-- Attachment #1: Type: text/plain, Size: 1502 bytes --]

Hi,

SPU architecture has 128 special register:

(gdb) ptype $r0
type = union __spu_builtin_type_vec128 {
    int128_t uint128;
    int64_t v2_int64[2];
    int32_t v4_int32[4];
    int16_t v8_int16[8];
    int8_t v16_int8[16];
    double v2_double[2];
    float v4_float[4];
}
(gdb) 

when debugging a fortran binary GDB cannot access single elements of the above 
elements regarded as arrays of int64, int32 ...

(gdb) p $r0%v2_int64
$1 = (635655159808, 0)
(gdb) p $r0%v2_int64(0)
Attempt to take address of value not located in memory.
(gdb) set $r0%v2_int64(0)=2
Attempt to take address of value not located in memory.
(gdb) 

Access to fortran subranges is possible:

(gdb) p $r0%v2_int64(0:0)
$2 = (635655159808)
(gdb) p $r0%v2_int64(0:1)
$3 = (635655159808, 0)
(gdb) p $r0%v2_int64     
$4 = (635655159808, 0)
(gdb) 

This patch invokes value_subscript for the special case that a multi_f77_subscript
is in a register instead of memory.

(gdb) p $r0%uint128
$2 = 0x00000094000000000000000000000000
(gdb) p $r0%v2_int64
$3 = (635655159808, 0)
(gdb) p $r0%v2_int64(0)
$4 = 635655159808

(gdb) set $r0%v2_int64(1)=5
(gdb) p $r0%v2_int64(1)
$6 = 5
(gdb) 

the gdb.fortran testsuite showed no regression on SPU.
If this patch is ok it would be great to have it in gdb 6.8. Ok ?

ChangeLog:

	* eval.c (evaluate_subexp_standard): Call value_subscript for
	accessing fortran array elements in registers.

Regards,
Markus

-- 
  Markus Deuling
  GNU Toolchain for Linux on Cell BE
  deuling@de.ibm.com


[-- Attachment #2: diff-fortran --]
[-- Type: text/plain, Size: 461 bytes --]

diff -urpN src/gdb/eval.c dev/gdb/eval.c
--- src/gdb/eval.c	2008-02-11 05:48:36.000000000 +0100
+++ dev/gdb/eval.c	2008-03-03 19:23:37.000000000 +0100
@@ -1720,6 +1720,8 @@ evaluate_subexp_standard (struct type *e
 	   returns the correct type value */
 
 	deprecated_set_value_type (arg1, tmp_type);
+	if (VALUE_LVAL (arg1) == lval_register)
+	  return value_subscript (arg1, arg2);
 	return value_ind (value_add (value_coerce_array (arg1), arg2));
       }
 

Re: [patch] Fix fortran access to special register on SPU arch

[Michael Snyder]

I understand the failure mode (not located in memory), but
if value_subscript works in the register case, why wouldn't
it work in general, rather than value_ind?

On Mon, 2008-03-03 at 20:06 +0100, Markus Deuling wrote:
> Hi,
> 
> SPU architecture has 128 special register:
> 
> (gdb) ptype $r0
> type = union __spu_builtin_type_vec128 {
>     int128_t uint128;
>     int64_t v2_int64[2];
>     int32_t v4_int32[4];
>     int16_t v8_int16[8];
>     int8_t v16_int8[16];
>     double v2_double[2];
>     float v4_float[4];
> }
> (gdb) 
> 
> when debugging a fortran binary GDB cannot access single elements of the above 
> elements regarded as arrays of int64, int32 ...
> 
> (gdb) p $r0%v2_int64
> $1 = (635655159808, 0)
> (gdb) p $r0%v2_int64(0)
> Attempt to take address of value not located in memory.
> (gdb) set $r0%v2_int64(0)=2
> Attempt to take address of value not located in memory.
> (gdb) 
> 
> Access to fortran subranges is possible:
> 
> (gdb) p $r0%v2_int64(0:0)
> $2 = (635655159808)
> (gdb) p $r0%v2_int64(0:1)
> $3 = (635655159808, 0)
> (gdb) p $r0%v2_int64     
> $4 = (635655159808, 0)
> (gdb) 
> 
> This patch invokes value_subscript for the special case that a multi_f77_subscript
> is in a register instead of memory.
> 
> (gdb) p $r0%uint128
> $2 = 0x00000094000000000000000000000000
> (gdb) p $r0%v2_int64
> $3 = (635655159808, 0)
> (gdb) p $r0%v2_int64(0)
> $4 = 635655159808
> 
> (gdb) set $r0%v2_int64(1)=5
> (gdb) p $r0%v2_int64(1)
> $6 = 5
> (gdb) 
> 
> the gdb.fortran testsuite showed no regression on SPU.
> If this patch is ok it would be great to have it in gdb 6.8. Ok ?
> 
> ChangeLog:
> 
> 	* eval.c (evaluate_subexp_standard): Call value_subscript for
> 	accessing fortran array elements in registers.
> 
> Regards,
> Markus
> 
> plain text document attachment (diff-fortran)
> diff -urpN src/gdb/eval.c dev/gdb/eval.c
> --- src/gdb/eval.c	2008-02-11 05:48:36.000000000 +0100
> +++ dev/gdb/eval.c	2008-03-03 19:23:37.000000000 +0100
> @@ -1720,6 +1720,8 @@ evaluate_subexp_standard (struct type *e
>  	   returns the correct type value */
>  
>  	deprecated_set_value_type (arg1, tmp_type);
> +	if (VALUE_LVAL (arg1) == lval_register)
> +	  return value_subscript (arg1, arg2);
>  	return value_ind (value_add (value_coerce_array (arg1), arg2));
>        }
>  

Re: [patch] Fix fortran access to special register on SPU arch

[Markus Deuling]

[-- Attachment #1: Type: text/plain, Size: 333 bytes --]

Hi Michael,

Michael Snyder schrieb:
> I understand the failure mode (not located in memory), but
> if value_subscript works in the register case, why wouldn't
> it work in general, rather than value_ind?
> 

it does. Attached patch works fine. Ok ?



-- 
  Markus Deuling
  GNU Toolchain for Linux on Cell BE
  deuling@de.ibm.com


[-- Attachment #2: diff-fortran --]
[-- Type: text/plain, Size: 446 bytes --]

diff -urpN src/gdb/eval.c dev/gdb/eval.c
--- src/gdb/eval.c	2008-02-11 05:48:36.000000000 +0100
+++ dev/gdb/eval.c	2008-03-03 21:19:32.000000000 +0100
@@ -1720,7 +1720,7 @@ evaluate_subexp_standard (struct type *e
 	   returns the correct type value */
 
 	deprecated_set_value_type (arg1, tmp_type);
-	return value_ind (value_add (value_coerce_array (arg1), arg2));
+	return value_subscript (arg1, arg2);
       }
 
     case BINOP_LOGICAL_AND: