Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Ulrich Weigand]

Mark Kettenis wrote:

+/* Return a default for the architecture-specific operations.  */
+
+static void *
+dwarf2_frame_init (struct gdbarch *gdbarch)
+{
+  struct dwarf2_frame_ops *ops;
+  
+  ops = GDBARCH_OBSTACK_ZALLOC (gdbarch, struct dwarf2_frame_ops);
+  ops->init_reg = dwarf2_frame_default_init_reg;
+  return ops;
+}
+
+/* Set the architecture-specific register state initialization
+   function for GDBARCH to INIT_REG.  */
+
+void
+dwarf2_frame_set_init_reg (struct gdbarch *gdbarch,
+                          void (*init_reg) (struct gdbarch *, int,
+                                            struct dwarf2_frame_state_reg *))
+{
+  struct dwarf2_frame_ops *ops;
+
+  ops = gdbarch_data (gdbarch, dwarf2_frame_data);
+  ops->init_reg = init_reg;
+}

Unfortunately this now crashes on s390, because the 
dwarf2_frame_init routine is called from within
_initialize_dwarf2_frame, while the s390 backend calls
dwarf2_frame_set_init_reg from within _initialize_s390_tdep.

Now, in the generated initialize_all_files routine, the
s390_tdep init routine is called *before* the dwarf2_frame
one, and hence dwarf2_frame_set_init_reg gets called before
dwarf2_frame_init.

Thus, the gdbarch_data call returns NULL, and the assignment
to ops->init_reg crashes.

Is there some usual way to solve this sort of init-order
issues with the gdb init sequence?

Bye,
Ulrich

-- 
  Dr. Ulrich Weigand
  weigand@informatik.uni-erlangen.de

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Daniel Jacobowitz]

On Mon, Feb 16, 2004 at 02:27:24AM +0100, Ulrich Weigand wrote:
> +void
> +dwarf2_frame_set_init_reg (struct gdbarch *gdbarch,
> +                          void (*init_reg) (struct gdbarch *, int,
> +                                            struct dwarf2_frame_state_reg *))

> Unfortunately this now crashes on s390, because the 
> dwarf2_frame_init routine is called from within
> _initialize_dwarf2_frame, while the s390 backend calls
> dwarf2_frame_set_init_reg from within _initialize_s390_tdep.

Er, how?  You shouldn't have a gdbarch in _initialize_s390_tdep.  You
don't have one until s390_gdbarch_init.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Ulrich Weigand]

Daniel Jacobowitz wrote:

> On Mon, Feb 16, 2004 at 02:27:24AM +0100, Ulrich Weigand wrote:
> > +void
> > +dwarf2_frame_set_init_reg (struct gdbarch *gdbarch,
> > +                          void (*init_reg) (struct gdbarch *, int,
> > +                                            struct dwarf2_frame_state_reg *))
> 
> > Unfortunately this now crashes on s390, because the 
> > dwarf2_frame_init routine is called from within
> > _initialize_dwarf2_frame, while the s390 backend calls
> > dwarf2_frame_set_init_reg from within _initialize_s390_tdep.
> 
> Er, how?  You shouldn't have a gdbarch in _initialize_s390_tdep.  You
> don't have one until s390_gdbarch_init.

Sorry, my analysis wasn't completely correct.  What actually happens
is an init-order problem during the initialization of the gdbarch
in initialize_current_architecture.  The problem is that I call
dwarf2_frame_set_init_reg from within the s390_gdbarch_init routine
(which I gather is the right place?), and at this point in time,
the gdbarch_data call in dwarf2_frame_set_init_reg returns NULL
because the gdbarch hasn't finished initialization yet:

gdbarch_data (struct gdbarch *gdbarch, struct gdbarch_data *data)
{
  gdb_assert (data->index < gdbarch->nr_data);
  /* The data-pointer isn't initialized, call init() to get a value but
     only if the architecture initializaiton has completed.  Otherwise
     punt - hope that the caller knows what they are doing.  */
  if (gdbarch->data[data->index] == NULL
      && gdbarch->initialized_p)

Bye,
Ulrich

-- 
  Dr. Ulrich Weigand
  weigand@informatik.uni-erlangen.de

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Daniel Jacobowitz]

On Mon, Feb 16, 2004 at 02:01:29PM +0100, Ulrich Weigand wrote:
> Daniel Jacobowitz wrote:
> 
> > On Mon, Feb 16, 2004 at 02:27:24AM +0100, Ulrich Weigand wrote:
> > > +void
> > > +dwarf2_frame_set_init_reg (struct gdbarch *gdbarch,
> > > +                          void (*init_reg) (struct gdbarch *, int,
> > > +                                            struct dwarf2_frame_state_reg *))
> > 
> > > Unfortunately this now crashes on s390, because the 
> > > dwarf2_frame_init routine is called from within
> > > _initialize_dwarf2_frame, while the s390 backend calls
> > > dwarf2_frame_set_init_reg from within _initialize_s390_tdep.
> > 
> > Er, how?  You shouldn't have a gdbarch in _initialize_s390_tdep.  You
> > don't have one until s390_gdbarch_init.
> 
> Sorry, my analysis wasn't completely correct.  What actually happens
> is an init-order problem during the initialization of the gdbarch
> in initialize_current_architecture.  The problem is that I call
> dwarf2_frame_set_init_reg from within the s390_gdbarch_init routine
> (which I gather is the right place?), and at this point in time,
> the gdbarch_data call in dwarf2_frame_set_init_reg returns NULL
> because the gdbarch hasn't finished initialization yet:
> 
> gdbarch_data (struct gdbarch *gdbarch, struct gdbarch_data *data)
> {
>   gdb_assert (data->index < gdbarch->nr_data);
>   /* The data-pointer isn't initialized, call init() to get a value but
>      only if the architecture initializaiton has completed.  Otherwise
>      punt - hope that the caller knows what they are doing.  */
>   if (gdbarch->data[data->index] == NULL
>       && gdbarch->initialized_p)

Oopsie.  Yes, now I can see what's going on - I'm not sure what should
change here.  Andrew, do you recall if the initialized_p check was for
any specific problem, or might we be able to remove it?

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Andrew Cagney]

>> Sorry, my analysis wasn't completely correct.  What actually happens
>> is an init-order problem during the initialization of the gdbarch
>> in initialize_current_architecture.  The problem is that I call
>> dwarf2_frame_set_init_reg from within the s390_gdbarch_init routine
>> (which I gather is the right place?), and at this point in time,
>> the gdbarch_data call in dwarf2_frame_set_init_reg returns NULL
>> because the gdbarch hasn't finished initialization yet:

I've committed the attatched.  It follows the convention described in 
that revised doco patch I posted:
http://sources.redhat.com/ml/gdb-patches/2004-02/msg00381.html

Andrew

>> gdbarch_data (struct gdbarch *gdbarch, struct gdbarch_data *data)
>> {
>>   gdb_assert (data->index < gdbarch->nr_data);
>>   /* The data-pointer isn't initialized, call init() to get a value but
>>      only if the architecture initializaiton has completed.  Otherwise
>>      punt - hope that the caller knows what they are doing.  */
>>   if (gdbarch->data[data->index] == NULL
>>       && gdbarch->initialized_p)
> 
> 
> Oopsie.  Yes, now I can see what's going on - I'm not sure what should
> change here.  Andrew, do you recall if the initialized_p check was for
> any specific problem, or might we be able to remove it?
> 
> -- Daniel Jacobowitz MontaVista Software Debian GNU/Linux Developer 

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Andrew Cagney]

[-- Attachment #1: Type: text/plain, Size: 673 bytes --]

> 
> Sorry, my analysis wasn't completely correct.  What actually happens
> is an init-order problem during the initialization of the gdbarch
> in initialize_current_architecture.  The problem is that I call
> dwarf2_frame_set_init_reg from within the s390_gdbarch_init routine
> (which I gather is the right place?), and at this point in time,
> the gdbarch_data call in dwarf2_frame_set_init_reg returns NULL
> because the gdbarch hasn't finished initialization yet:
> 
> I've committed the attatched.  It follows the convention described in that revised doco patch I posted:
> http://sources.redhat.com/ml/gdb-patches/2004-02/msg00381.html
> 
> Andrew 

with patch ...


[-- Attachment #2: diffs --]
[-- Type: text/plain, Size: 1408 bytes --]

2004-02-16  Andrew Cagney  <cagney@redhat.com>

	* dwarf2-frame.c (dwarf2_frame_ops): New function.
	(dwarf2_frame_set_init_reg): Use, instead of gdbarch_data.
	(dwarf2_frame_init_reg): Ditto.

Index: dwarf2-frame.c
===================================================================
RCS file: /cvs/src/src/gdb/dwarf2-frame.c,v
retrieving revision 1.30
diff -u -r1.30 dwarf2-frame.c
--- dwarf2-frame.c	15 Feb 2004 21:29:26 -0000	1.30
+++ dwarf2-frame.c	16 Feb 2004 20:28:01 -0000
@@ -518,6 +518,20 @@
   return ops;
 }
 
+static struct dwarf2_frame_ops *
+dwarf2_frame_ops (struct gdbarch *gdbarch)
+{
+  struct dwarf2_frame_ops *ops = gdbarch_data (gdbarch, dwarf2_frame_data);
+  if (ops == NULL)
+    {
+      /* ULGH, called during architecture initialization.  Patch
+         things up.  */
+      ops = dwarf2_frame_init (gdbarch);
+      set_gdbarch_data (gdbarch, dwarf2_frame_data, ops);
+    }
+  return ops;
+}
+
 /* Set the architecture-specific register state initialization
    function for GDBARCH to INIT_REG.  */
 
@@ -528,7 +542,7 @@
 {
   struct dwarf2_frame_ops *ops;
 
-  ops = gdbarch_data (gdbarch, dwarf2_frame_data);
+  ops = dwarf2_frame_ops (gdbarch);
   ops->init_reg = init_reg;
 }
 
@@ -540,7 +554,7 @@
 {
   struct dwarf2_frame_ops *ops;
 
-  ops = gdbarch_data (gdbarch, dwarf2_frame_data);
+  ops = dwarf2_frame_ops (gdbarch);
   ops->init_reg (gdbarch, regnum, reg);
 }
 \f

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Mark Kettenis]

   Date: Mon, 16 Feb 2004 15:55:52 -0500
   From: Andrew Cagney <cagney@gnu.org>

   > I've committed the attatched.  It follows the convention
   > described in that revised doco patch I posted:
   > http://sources.redhat.com/ml/gdb-patches/2004-02/msg00381.html
   > 
   > Andrew 

   with patch ...

   2004-02-16  Andrew Cagney  <cagney@redhat.com>

	   * dwarf2-frame.c (dwarf2_frame_ops): New function.
	   (dwarf2_frame_set_init_reg): Use, instead of gdbarch_data.
	   (dwarf2_frame_init_reg): Ditto.

Thanks Andrew.  Can't say I'm too happy with yet another function, but
if that's the it's supposed to work...  Perhaps we should revisit the
way these per-architecture keys work someday.

Mark

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Andrew Cagney]

> Thanks Andrew.  Can't say I'm too happy with yet another function, but
> if that's the it's supposed to work...  Perhaps we should revisit the
> way these per-architecture keys work someday.

It's how it works.  Revising how it is supposed to work at some stage 
would be useful..

Andrew

Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization

[Ulrich Weigand]

Andrew Cagney wrote:

> I've committed the attatched.  It follows the convention described in 
> that revised doco patch I posted:
> http://sources.redhat.com/ml/gdb-patches/2004-02/msg00381.html

This fixes the problem I was seeing.  Using DWARF-2 frames now works
on s390 without any changes to common code. :-)

Thanks!
Ulrich

-- 
  Dr. Ulrich Weigand
  weigand@informatik.uni-erlangen.de