Re: [patch rfc] push_dummy_call return a frame ID; eliminate needfor read_fp

[Andrew Cagney <ac131313 at redhat dot com>]

Hmm,

Patch withdrawn.  I've just figured out a more straight forward way of 
doing this.  The `else' comment:

> +	   re-identify the dummy breakpoint.  Code using
> +	   push_dummy_call() will have previously computed the dummy
> +	   frame's ID (and hence set "code_addr" to something
> +	   non-zero.  */
> +	id = frame_id_build (read_fp (), sal.pc);
> +      /* else ... is the sal.pc the frame ID's code_addr?  */
> +

gives the hint.

The frame ID's code_addr is always be the address of the breakpoint, so 
instead of having push_dummy_call() return the frame ID, it can return 
the ID's stack_addr, and then the hand function call code can explicitly 
construct the frame ID from that and the sap.pc.

I'll just need to check the Arm.

Andrew


> Now that the inferior function call code is all in a single function, more, and significant, cleanups become possible.
>
> This patch modifies the architecture method push_dummy_call() so that it returns a frame ID, and not just a stack pointer.  That frame ID is then used:
>
> - to create the breakpoint (the ID identifies the breakpoint's frame)
>
> - to identify the dummy (somewhat indirectly at present as it still goes via SAVE_DUMMY_FRAME_TOS() and generic_save_call_dummy_addr())
>
> This means that:
>
> - push_dummy_call(regcache, ...) creates the dummy frame's ID
>
> - unwind_dummy_id(frame, ...) is then expected to return that same value
>
> so the need for one to match the other is hopefully both clear and easy ...
>
> Future changes will involve:
>
> - deprecating read_fp / TARGET_READ_FP, as core GDB no longer contains any legitimate calls
>
> - merging SAVE_DUMMY_FRAME_TOS and generic_save_call_dummy_addr() into a single method that receives the full frame ID, and can not be overridden by the architecture.
>
> Someone studying this carefully will probably realise that instead of having push_dummy_call() return the ID, the code could simply use:
>
>     flush cached frames;
>     id = get_frame_id (get_current_frame ());
>
> but that would incure additional overhead.
>
> Mark,
> I think this patch resolves the i386's FP/SP frame ID problem.  It makes push_dummy_call responsible for computing the ID, and then just passes it through.
>
> Richard,
> For the ARM, I've pulled a quick hack.  I'm not sure what would happen if the ARM started returning a full frame ID.
>
> I'll leave this one sit for a week.
>
> Andrew
>
>
>
> 2003-04-22  Andrew Cagney  <cagney at redhat dot com>
>
> 	* d10v-tdep.c (d10v_push_dummy_call): Return a frame ID.
> 	(d10v_gdbarch_init): Do not set read_fp.
> 	(d10v_read_fp): Delete function.
> 	* arm-tdep.c (arm_push_dummy_call): Return the SP in a frame ID.
> 	* infcall.c (call_function_by_hand): New variable "id".
> 	Initialize using gdbarch_push_dummy_call when possible.  For
> 	legacy code, use the the frame ID's "stack_addr" instead of "sp"
> 	to track the stack's inner most address.
> 	* gdbarch.sh (PUSH_DUMMY_CALL): Return the frame ID instead of the
> 	stack's inner most address.
> 	* gdbarch.h, gdbarch.c: Re-generate.
> 	
> Index: doc/ChangeLog
> 2003-04-22  Andrew Cagney  <cagney at redhat dot com>
>
> 	* gdbint.texinfo (Target Architecture Definition): Update
> 	"push_dummy_call", returns a frame ID.  Cross reference
> 	"push_dummy_call" and "unwind_dummy_id".  Drop reference to
> 	SAVE_DUMMY_FRAME_TOS.