Re: [PATCH] Disallow pseudo-registers in agent expression.

["Jim Blandy" <jimb at red-bean dot com>]

On Jan 28, 2008 3:51 PM, andrzej zaborowski <balrogg@gmail.com> wrote:
> On 27/01/2008, Jim Blandy <jimb@red-bean.com> wrote:
> > On some targets, all the user-visible registers are pseudo-registers.
> > What we need is a gdbarch method (optional) that the tracepoint code
> > could call, passing it a struct agent_expr and a pseudo-register
> > number, to have the architecture code append the appropriate bytecodes
> > to access that register.
>
> That's ofcourse doable but sounds tough and bug-prone. The mapping of
> numbers to raw registers is to some extent obvious and unlikely to
> change, while the pseudo-registers may change with gdb version so the
> target has more gdb version dependence.

I'm not sure I expressed myself well.  I was imagining something like this:

*** ax-gdb.c	07 Jan 2008 08:31:13 -0800	1.40
--- ax-gdb.c	28 Jan 2008 17:21:28 -0800	
***************
*** 1607,1615 ****
  	if (reg == -1)
  	  internal_error (__FILE__, __LINE__,
  			  _("Register $%s not available"), name);
! 	value->kind = axs_lvalue_register;
! 	value->u.reg = reg;
! 	value->type = register_type (current_gdbarch, reg);
        }
        break;

--- 1607,1627 ----
  	if (reg == -1)
  	  internal_error (__FILE__, __LINE__,
  			  _("Register $%s not available"), name);
!         if (reg < gdbarch_num_regs (current_gdbarch))
!           {
!             value->kind = axs_lvalue_register;
!             value->u.reg = reg;
!             value->type = register_type (current_gdbarch, reg);
!           }
!         else
!           {
!             /* It's a pseudo-register reference; ask the architecture
!                to generate appropriate bytecode for it.  */
!             if (! gdbarch_gen_pseudo_reg_ax (current_gdbarch, ax, value, reg))
!               error (_("Couldn't generate agent expression"
!                        " to refer to register '%s'"),
!                      gdbarch_register_name (current_gdbarch, reg));
!           }
        }
        break;

This seems pretty reasonable to me: only the architecture knows how
pseudo-register values are computed from the raw register values --- a
pseudo-register might be two raw registers spliced together, for
example --- so only the architecture knows the right agent bytecode to
generate.