[RFA] better alpha_register_virtual_type

[RFA] better alpha_register_virtual_type

[Richard Henderson]

Ok?


r~


	* alpha-tdep.c (alpha_register_virtual_type): Use void_data_ptr
	for SP, GP; void_func_ptr for PC; non-language-specific types
	for all others.
	* alpha-tdep.h (ALPHA_GP_REGNUM): New.

Index: alpha-tdep.c
===================================================================
RCS file: /cvs/src/src/gdb/alpha-tdep.c,v
retrieving revision 1.95
diff -c -p -d -u -p -r1.95 alpha-tdep.c
--- alpha-tdep.c	2 Jun 2003 04:32:19 -0000	1.95
+++ alpha-tdep.c	2 Jun 2003 04:46:18 -0000
@@ -89,8 +89,17 @@ alpha_register_convertible (int regno)
 static struct type *
 alpha_register_virtual_type (int regno)
 {
-  return ((regno >= FP0_REGNUM && regno < (FP0_REGNUM+31))
-	  ? builtin_type_double : builtin_type_long);
+  if (regno == ALPHA_SP_REGNUM || regno == ALPHA_GP_REGNUM)
+    return builtin_type_void_data_ptr;
+  if (regno == ALPHA_PC_REGNUM)
+    return builtin_type_void_func_ptr;
+
+  /* Don't need to worry about little vs big endian until 
+     some jerk tries to port to alpha-unicosmk.  */
+  if (regno >= FP0_REGNUM && regno < FP0_REGNUM + 31)
+    return builtin_type_ieee_double_little;
+
+  return builtin_type_int64;
 }
 
 /* Is REGNUM a member of REGGROUP?  */
Index: alpha-tdep.h
===================================================================
RCS file: /cvs/src/src/gdb/alpha-tdep.h,v
retrieving revision 1.15
diff -c -p -d -u -p -r1.15 alpha-tdep.h
--- alpha-tdep.h	1 Jun 2003 21:46:37 -0000	1.15
+++ alpha-tdep.h	2 Jun 2003 04:46:19 -0000
@@ -43,9 +43,10 @@
 #define ALPHA_GCC_FP_REGNUM 15  /* Used by gcc as frame register */
 #define ALPHA_A0_REGNUM     16  /* Loc of first arg during a subr call */
 #define ALPHA_T9_REGNUM     23  /* Return address register for OSF/1 __div* */
+#define ALPHA_RA_REGNUM     26  /* Contains return address value */
 #define ALPHA_T12_REGNUM    27  /* Contains start addr of current proc */
+#define ALPHA_GP_REGNUM     29  /* Contains the global pointer */
 #define ALPHA_SP_REGNUM     30  /* Contains address of top of stack */
-#define ALPHA_RA_REGNUM     26  /* Contains return address value */
 #define ALPHA_ZERO_REGNUM   31  /* Read-only register, always 0 */
 #define ALPHA_FP0_REGNUM    32  /* Floating point register 0 */
 #define ALPHA_FPA0_REGNUM   48  /* First float arg during a subr call */

Re: [RFA] better alpha_register_virtual_type

[Daniel Jacobowitz]

On Sun, Jun 01, 2003 at 10:03:58PM -0700, Richard Henderson wrote:
> Ok?

Yes.  I was a little surprised by the void_data_ptr/void_func_ptr bit,
but I see that d10v does the same thing, so it must be right :)

Hmm, I'm not sure.  Any particular reason for this patch?  The
ieee_double_little and int64 parts are definitely correct, but 

> 
> 
> r~
> 
> 
> 	* alpha-tdep.c (alpha_register_virtual_type): Use void_data_ptr
> 	for SP, GP; void_func_ptr for PC; non-language-specific types
> 	for all others.
> 	* alpha-tdep.h (ALPHA_GP_REGNUM): New.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: [RFA] better alpha_register_virtual_type

[Richard Henderson]

On Mon, Jun 02, 2003 at 12:30:36PM -0400, Daniel Jacobowitz wrote:
> Yes.  I was a little surprised by the void_data_ptr/void_func_ptr bit,
> but I see that d10v does the same thing, so it must be right :)
>
> Hmm, I'm not sure.  Any particular reason for this patch?

Well, the void_func_ptr bit is nice because "info r" yields

	pc             0x12000053c      0x12000053c <main+16>

The void_data_ptr bits I think just document which registers
are ABI mandated to contain pointer values all of the time.

Actually, I have a related question here.  Something that I
didn't notice earlier is that d10v is using register_type,
not register_virtual_type.  Looking at the guts of regcache.c,
it would appear that the later is deprecated, since not 
having a register_type hook (among other things) results in
legacy_p being set.

I thought it obvious to rename my existing hook, but that changes
the behaviour of "info r" -- I no longer get the pc decoded, and
indeed "ptype $pc" once again yields int64_t.

Is this a bug elsewhere in gdb, or what?

Oh, and wrt regcache's legacy_p, it seems to want you to implement
the pseudo_register_{read,write} hooks, even if the target doesn't
have any.  But nevertheless d10v doesn't implement the hooks.
Perhaps the predicates should be modified to notice that there are
no pseudos defined?


r~

Re: [RFA] better alpha_register_virtual_type

[Daniel Jacobowitz]

On Mon, Jun 02, 2003 at 09:43:05AM -0700, Richard Henderson wrote:
> On Mon, Jun 02, 2003 at 12:30:36PM -0400, Daniel Jacobowitz wrote:
> > Yes.  I was a little surprised by the void_data_ptr/void_func_ptr bit,
> > but I see that d10v does the same thing, so it must be right :)
> >
> > Hmm, I'm not sure.  Any particular reason for this patch?
> 
> Well, the void_func_ptr bit is nice because "info r" yields
> 
> 	pc             0x12000053c      0x12000053c <main+16>
> 
> The void_data_ptr bits I think just document which registers
> are ABI mandated to contain pointer values all of the time.

Hmm, that's pretty nice.  Sure.

> Actually, I have a related question here.  Something that I
> didn't notice earlier is that d10v is using register_type,
> not register_virtual_type.  Looking at the guts of regcache.c,
> it would appear that the later is deprecated, since not 
> having a register_type hook (among other things) results in
> legacy_p being set.
> 
> I thought it obvious to rename my existing hook, but that changes
> the behaviour of "info r" -- I no longer get the pc decoded, and
> indeed "ptype $pc" once again yields int64_t.
> 
> Is this a bug elsewhere in gdb, or what?

Hum.  That seems strange if you look at init_regcache_descr, since
gdbarch_register_type and REGISTER_VIRTUAL_TYPE are used similarly.  I
can't see how legacy_p would affect this.

What does ptype $pc say - does it show up as an int64_t or a code
pointer?

> Oh, and wrt regcache's legacy_p, it seems to want you to implement
> the pseudo_register_{read,write} hooks, even if the target doesn't
> have any.  But nevertheless d10v doesn't implement the hooks.
> Perhaps the predicates should be modified to notice that there are
> no pseudos defined?

I don't see what you mean; it's:
  if (!gdbarch_pseudo_register_read_p (gdbarch)
      && !gdbarch_pseudo_register_write_p (gdbarch)
      && !gdbarch_register_type_p (gdbarch))

but gdbarch_register_type_p should be true.

"maint print registers" might be handy here.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: [RFA] better alpha_register_virtual_type

[Andrew Cagney]

> On Mon, Jun 02, 2003 at 12:30:36PM -0400, Daniel Jacobowitz wrote:
> 
>> Yes.  I was a little surprised by the void_data_ptr/void_func_ptr bit,
>> but I see that d10v does the same thing, so it must be right :)

For the d10v it is 100% right.  The d10v's PC needs to be a code pointer 
so that it it is automatically converted into a CORE_ADDR when used in 
equations vis:

(gdb) ptype $pc
type = void (*)()
(gdb) x/i $pc
0x10140b8 <main+20>:    ld      r0, @r11        ||      nop
(gdb) print/x $pc
$1 = 0x10140b8
(gdb) print/x (int)$pc
$2 = 0x502e

>> Hmm, I'm not sure.  Any particular reason for this patch?
> 
> 
> Well, the void_func_ptr bit is nice because "info r" yields
> 
> 	pc             0x12000053c      0x12000053c <main+16>
> 
> The void_data_ptr bits I think just document which registers
> are ABI mandated to contain pointer values all of the time.
> 
> Actually, I have a related question here.  Something that I
> didn't notice earlier is that d10v is using register_type,
> not register_virtual_type.  Looking at the guts of regcache.c,
> it would appear that the later is deprecated, since not 
> having a register_type hook (among other things) results in
> legacy_p being set.
> 
> I thought it obvious to rename my existing hook, but that changes
> the behaviour of "info r" -- I no longer get the pc decoded, and
> indeed "ptype $pc" once again yields int64_t.

Something weird is happening.  The only thing notable about the d10v is 
that it doesn't set PC_REGNUM (per above, the d10v works).  Perhaphs the 
alpha definition is confusing std-regs.c / builtin-regs.c (but that 
doesn't make much sense).

> Is this a bug elsewhere in gdb, or what?

Perhaps.

Andrew

Re: [RFA] better alpha_register_virtual_type

[Richard Henderson]

On Mon, Jun 02, 2003 at 03:06:19PM -0400, Andrew Cagney wrote:
> >I thought it obvious to rename my existing hook, but that changes
> >the behaviour of "info r" -- I no longer get the pc decoded, and
> >indeed "ptype $pc" once again yields int64_t.
> 
> Something weird is happening.  The only thing notable about the d10v is 
> that it doesn't set PC_REGNUM (per above, the d10v works).  Perhaphs the 
> alpha definition is confusing std-regs.c / builtin-regs.c (but that 
> doesn't make much sense).

I have no idea.  I can't replicate this now.  I've made the
change to _register_type again and things seem to work as
expected.  Perhaps I did something stupid before.   *shrug*


r~