[RFA,5/6] Add validity bits for psymtab high and low fields

  Right now some psymtab code checks whether a psymtab's textlow or
texthigh fields are valid by comparing against 0.

I imagine this is mildly wrong in the current environment, but once
psymtabs are relocated dynamically, it will no longer be correct,
because it will be much more normal to see a psymtab with a textlow of
zero -- this will just mean it appears at the start of the text
section.

This patch introduces validity bits to handle this situation more
nicely, and changes users of the code to follow.

ChangeLog
2018-05-03  Tom Tromey  <tromey@redhat.com>

	* dbxread.c (end_psymtab): Use texthigh_valid and textlow_valid.
	* mdebugread.c (parse_partial_symbols): Use textlow_valid.
	(psymtab_to_symtab_1): Use texthigh_valid and textlow_valid.
	* psympriv.h (struct partial_symtab) <textlow_, texthigh_>: Update
	comment.
	<textlow_valid, texthigh_valid>: New fields.
	(set_psymtab_textlow, set_psymtab_texthigh): New inline functions.
	(SET_PSYMTAB_TEXTLOW, SET_PSYMTAB_TEXTHIGH): Redefine.
	* xcoffread.c (scan_xcoff_symtab): Use textlow_valid.
---
 gdb/ChangeLog    | 12 ++++++++++++
 gdb/dbxread.c    |  6 ++----
 gdb/mdebugread.c |  8 ++++----
 gdb/psympriv.h   | 33 ++++++++++++++++++++++++++++++---
 gdb/xcoffread.c  |  2 +-
 5 files changed, 49 insertions(+), 12 deletions(-)
  

On 05/03/2018 03:36 PM, Tom Tromey wrote:
> Right now some psymtab code checks whether a psymtab's textlow or
> texthigh fields are valid by comparing against 0.
> 
> I imagine this is mildly wrong in the current environment, but once
> psymtabs are relocated dynamically, it will no longer be correct,
> because it will be much more normal to see a psymtab with a textlow of
> zero -- this will just mean it appears at the start of the text
> section.
> 

Also wrong if 0 is a valid address! [as a recent patch* serves to remind]

> This patch introduces validity bits to handle this situation more
> nicely, and changes users of the code to follow.
> 
> ChangeLog
> 2018-05-03  Tom Tromey  <tromey@redhat.com>
> 
> 	* dbxread.c (end_psymtab): Use texthigh_valid and textlow_valid.

Does this line not also qualify for texthigh_valid:

  if (PSYMTAB_TEXTHIGH (pst) == 0 && last_function_name
      && gdbarch_sofun_address_maybe_missing (gdbarch))

?

There's also this line in dbx_read_symtab that might qualify (from case N_SCOPE):

	      valu = nlist.n_value + last_function_start;
	      if (PSYMTAB_TEXTHIGH (pst) == 0 || valu > PSYMTAB_TEXTHIGH (pst))
		SET_PSYMTAB_TEXTHIGH (pst, valu);
	      break;

Also in this function, can textlow_not_set be replaced by textlow_valid? [The same with the textlow_not_set parameter to dbx_end_psymtab?]

> 	* mdebugread.c (parse_partial_symbols): Use textlow_valid.
> 	(psymtab_to_symtab_1): Use texthigh_valid and textlow_valid.
> 	* psympriv.h (struct partial_symtab) <textlow_, texthigh_>: Update
> 	comment.
> 	<textlow_valid, texthigh_valid>: New fields.
> 	(set_psymtab_textlow, set_psymtab_texthigh): New inline functions.
> 	(SET_PSYMTAB_TEXTLOW, SET_PSYMTAB_TEXTHIGH): Redefine.
> 	* xcoffread.c (scan_xcoff_symtab): Use textlow_valid.

Otherwise, this all LGTM (IANAM).

Keith

* https://sourceware.org/ml/gdb-patches/2018-05/msg00325.html
  

>>>>> "Keith" == Keith Seitz <keiths@redhat.com> writes:

>> * dbxread.c (end_psymtab): Use texthigh_valid and textlow_valid.

Keith> Does this line not also qualify for texthigh_valid:
Keith>   if (PSYMTAB_TEXTHIGH (pst) == 0 && last_function_name
Keith>       && gdbarch_sofun_address_maybe_missing (gdbarch))
Keith> ?

This is a bit obscure, and I'm not totally sure it is correct.  The idea
is that this only happens in a pathological case, which we want to
detect.  There is a comment there:

  /* Under Solaris, the N_SO symbols always have a value of 0,
     instead of the usual address of the .o file.  Therefore,
     we have to do some tricks to fill in texthigh and textlow.
..


Also the final patch in this series changes this line to:

  if (PSYMTAB_RAW_TEXTHIGH (pst) == 0 && last_function_name
      && gdbarch_sofun_address_maybe_missing (gdbarch))


Keith> There's also this line in dbx_read_symtab that might qualify (from case N_SCOPE):

Keith> 	      valu = nlist.n_value + last_function_start;
Keith> 	      if (PSYMTAB_TEXTHIGH (pst) == 0 || valu > PSYMTAB_TEXTHIGH (pst))
Keith> 		SET_PSYMTAB_TEXTHIGH (pst, valu);
Keith> 	      break;

This also gets changed to use the RAW_ forms.

Keith> Also in this function, can textlow_not_set be replaced by
Keith> textlow_valid? [The same with the textlow_not_set parameter to
Keith> dbx_end_psymtab?]

I think it is a good idea, but there is one spot setting textlow_not_set
without invoking SET_PSYMTAB_TEXTLOW.  From the N_SO case:

	    prev_textlow_not_set = textlow_not_set;

	    /* A zero value is probably an indication for the SunPRO 3.0
	       compiler.  dbx_end_psymtab explicitly tests for zero, so
	       don't relocate it.  */

	    if (nlist.n_value == 0
		&& gdbarch_sofun_address_maybe_missing (gdbarch))
	      {
		textlow_not_set = 1;
		valu = 0;
	      }
	    else
	      textlow_not_set = 0;

I don't know anything about stabs, so my basic idea in this area was to
make the changes as minimal as possible.

Tom
  

On 06/05/2018 10:25 AM, Tom Tromey wrote:
>>>>>> "Keith" == Keith Seitz <keiths@redhat.com> writes:
> Also the final patch in this series changes this line to:
> 
>   if (PSYMTAB_RAW_TEXTHIGH (pst) == 0 && last_function_name
>       && gdbarch_sofun_address_maybe_missing (gdbarch))
> 

Sorry, I missed that. In that case, ignore my comment.

> Keith> There's also this line in dbx_read_symtab that might qualify (from case N_SCOPE):
> 
> Keith> 	      valu = nlist.n_value + last_function_start;
> Keith> 	      if (PSYMTAB_TEXTHIGH (pst) == 0 || valu > PSYMTAB_TEXTHIGH (pst))
> Keith> 		SET_PSYMTAB_TEXTHIGH (pst, valu);
> Keith> 	      break;
> 
> This also gets changed to use the RAW_ forms.

Likewise.

> Keith> Also in this function, can textlow_not_set be replaced by
> Keith> textlow_valid? [The same with the textlow_not_set parameter to
> Keith> dbx_end_psymtab?]
> 
> I think it is a good idea, but there is one spot setting textlow_not_set
> without invoking SET_PSYMTAB_TEXTLOW.  From the N_SO case:
> 
> 	    prev_textlow_not_set = textlow_not_set;
> 
> 	    /* A zero value is probably an indication for the SunPRO 3.0
> 	       compiler.  dbx_end_psymtab explicitly tests for zero, so
> 	       don't relocate it.  */
> 
> 	    if (nlist.n_value == 0
> 		&& gdbarch_sofun_address_maybe_missing (gdbarch))
> 	      {
> 		textlow_not_set = 1;
> 		valu = 0;
> 	      }
> 	    else
> 	      textlow_not_set = 0;
> 
> I don't know anything about stabs, so my basic idea in this area was to
> make the changes as minimal as possible.

I should have looked a little more closely at the viability of my suggestion. I agree with you, erring on the side of minimal invasion in "obscure" code is always prudent.

Thank you for the explanations.

Keith