Re: SH: double pseudo registers little-endian/big-endian

[Thomas Schwinge <thomas@codesourcery.com>]

[-- Attachment #1: Type: text/plain, Size: 6750 bytes --]

Hi!

Ping.


On Thu, 30 Aug 2012 17:52:36 +0200, I wrote:
> Little-endian SH GDB:
> 
>     (gdb) set $dr0 = -1
>     (gdb) info registers dr0 fr0 fr1
>     dr0            -1       (raw 0xbff0000000000000)
>     fr0            -1.875   (raw 0xbff00000)
>     fr1            0        (raw 0x00000000)
> 
> This is correct.  Big-endian SH GDB:
> 
>     (gdb) set $dr0 = -1
>     (gdb) info registers dr0 fr0 fr1
>     dr0            -1       (raw 0xbff0000000000000)
>     fr0            8.63634257e-41   (raw 0x0000f0bf)
>     fr1            0        (raw 0x00000000)
> 
> Here, the value stored in fr0 is bogus.  This comes from the
> sh_register_convert_to_virtual conversion: as the commentary says, this
> should be applied in the little-endian case only.  The reasons that dr0
> is displayed correctly nevertheless is due to sh_register_convert_to_raw
> (erroneously) doing the same conversion backwards, too.
> 
> A similar problem exists when displaying double values loaded as part of
> the program text.
> 
> Here is a patch (basically copying code from sh64-tdep.c which got this
> right), and also fixes some commentary.  OK to commit?
> 
> gdb/
> 	sh-tdep.c (sh_register_convert_to_virtual)
> 	(sh_register_convert_to_raw): Add a gdbarch parameter.  Update
> 	all callers.  Just do a memcpy if not the little-endian case.
> 
> Index: sh-tdep.c
> ===================================================================
> RCS file: /cvs/src/src/gdb/sh-tdep.c,v
> retrieving revision 1.246
> diff -u -p -r1.246 sh-tdep.c
> --- sh-tdep.c	22 Jul 2012 16:52:41 -0000	1.246
> +++ sh-tdep.c	30 Aug 2012 15:43:24 -0000
> @@ -342,15 +342,16 @@ sh_sh4_register_name (struct gdbarch *gd
>      "r0b0", "r1b0", "r2b0", "r3b0", "r4b0", "r5b0", "r6b0", "r7b0",
>      /* bank 1 51 - 58 */
>      "r0b1", "r1b1", "r2b1", "r3b1", "r4b1", "r5b1", "r6b1", "r7b1",
> +    /* 59 - 66 */
>      "", "", "", "", "", "", "", "",
>      /* pseudo bank register.  */
>      "",
> -    /* double precision (pseudo) 59 - 66 */
> +    /* double precision (pseudo) 68 - 75 */
>      "dr0", "dr2", "dr4", "dr6", "dr8", "dr10", "dr12", "dr14",
> -    /* vectors (pseudo) 67 - 70 */
> +    /* vectors (pseudo) 76 - 79 */
>      "fv0", "fv4", "fv8", "fv12",
> -    /* FIXME: missing XF 71 - 86 */
> -    /* FIXME: missing XD 87 - 94 */
> +    /* FIXME: missing XF */
> +    /* FIXME: missing XD */
>    };
>    if (reg_nr < 0)
>      return NULL;
> @@ -379,12 +380,13 @@ sh_sh4_nofpu_register_name (struct gdbar
>      "r0b0", "r1b0", "r2b0", "r3b0", "r4b0", "r5b0", "r6b0", "r7b0",
>      /* bank 1 51 - 58 */
>      "r0b1", "r1b1", "r2b1", "r3b1", "r4b1", "r5b1", "r6b1", "r7b1",
> +    /* 59 - 66 */
>      "", "", "", "", "", "", "", "",
>      /* pseudo bank register.  */
>      "",
> -    /* double precision (pseudo) 59 - 66 -- not for nofpu target */
> +    /* double precision (pseudo) 68 - 75 -- not for nofpu target */
>      "", "", "", "", "", "", "", "",
> -    /* vectors (pseudo) 67 - 70 -- not for nofpu target */
> +    /* vectors (pseudo) 76 - 79 -- not for nofpu target */
>      "", "", "", "",
>    };
>    if (reg_nr < 0)
> @@ -1521,16 +1523,16 @@ sh_register_reggroup_p (struct gdbarch *
>  
>  /* On the sh4, the DRi pseudo registers are problematic if the target
>     is little endian.  When the user writes one of those registers, for
> -   instance with 'ser var $dr0=1', we want the double to be stored
> +   instance with 'set var $dr0=1', we want the double to be stored
>     like this: 
> -   fr0 = 0x00 0x00 0x00 0x00 0x00 0xf0 0x3f 
> -   fr1 = 0x00 0x00 0x00 0x00 0x00 0x00 0x00 
> +   fr0 = 0x00 0x00 0xf0 0x3f 
> +   fr1 = 0x00 0x00 0x00 0x00 
>  
>     This corresponds to little endian byte order & big endian word
>     order.  However if we let gdb write the register w/o conversion, it
>     will write fr0 and fr1 this way:
> -   fr0 = 0x00 0x00 0x00 0x00 0x00 0x00 0x00
> -   fr1 = 0x00 0x00 0x00 0x00 0x00 0xf0 0x3f
> +   fr0 = 0x00 0x00 0x00 0x00
> +   fr1 = 0x00 0x00 0xf0 0x3f
>     because it will consider fr0 and fr1 as a single LE stretch of memory.
>     
>     To achieve what we want we must force gdb to store things in
> @@ -1539,16 +1541,23 @@ sh_register_reggroup_p (struct gdbarch *
>  
>     In case the target is big endian, there is no problem, the
>     raw bytes will look like:
> -   fr0 = 0x3f 0xf0 0x00 0x00 0x00 0x00 0x00
> -   fr1 = 0x00 0x00 0x00 0x00 0x00 0x00 0x00 
> +   fr0 = 0x3f 0xf0 0x00 0x00
> +   fr1 = 0x00 0x00 0x00 0x00
>  
>     The other pseudo registers (the FVs) also don't pose a problem
>     because they are stored as 4 individual FP elements.  */
>  
>  static void
> -sh_register_convert_to_virtual (int regnum, struct type *type,
> -				char *from, char *to)
> +sh_register_convert_to_virtual (struct gdbarch *gdbarch, int regnum,
> +				struct type *type, char *from, char *to)
>  {
> +  if (gdbarch_byte_order (gdbarch) != BFD_ENDIAN_LITTLE)
> +    {
> +      /* It is a no-op.  */
> +      memcpy (to, from, register_size (gdbarch, regnum));
> +      return;
> +    }
> +
>    if (regnum >= DR0_REGNUM && regnum <= DR_LAST_REGNUM)
>      {
>        DOUBLEST val;
> @@ -1562,9 +1571,16 @@ sh_register_convert_to_virtual (int regn
>  }
>  
>  static void
> -sh_register_convert_to_raw (struct type *type, int regnum,
> -			    const void *from, void *to)
> +sh_register_convert_to_raw (struct gdbarch *gdbarch, struct type *type,
> +			    int regnum, const void *from, void *to)
>  {
> +  if (gdbarch_byte_order (gdbarch) != BFD_ENDIAN_LITTLE)
> +    {
> +      /* It is a no-op.  */
> +      memcpy (to, from, register_size (gdbarch, regnum));
> +      return;
> +    }
> +
>    if (regnum >= DR0_REGNUM && regnum <= DR_LAST_REGNUM)
>      {
>        DOUBLEST val = extract_typed_floating (from, type);
> @@ -1643,7 +1659,7 @@ sh_pseudo_register_read (struct gdbarch 
>        if (status == REG_VALID)
>  	{
>  	  /* We must pay attention to the endiannes. */
> -	  sh_register_convert_to_virtual (reg_nr,
> +	  sh_register_convert_to_virtual (gdbarch, reg_nr,
>  					  register_type (gdbarch, reg_nr),
>  					  temp_buffer, buffer);
>  	}
> @@ -1685,7 +1701,7 @@ sh_pseudo_register_write (struct gdbarch
>        base_regnum = dr_reg_base_num (gdbarch, reg_nr);
>  
>        /* We must pay attention to the endiannes.  */
> -      sh_register_convert_to_raw (register_type (gdbarch, reg_nr),
> +      sh_register_convert_to_raw (gdbarch, register_type (gdbarch, reg_nr),
>  				  reg_nr, buffer, temp_buffer);
>  
>        /* Write the real regs for which this one is an alias.  */


Grüße,
 Thomas

[-- Attachment #2: Type: application/pgp-signature, Size: 489 bytes --]