[RFC] Uninitialized section index internal error on Tru64 5.1

[RFC] Uninitialized section index internal error on Tru64 5.1

[Joel Brobecker]

This follows up on a discussion started over a year ago... Lack of time
prevented me from persuing this issue. Hopefully we can make progress
quickly. The last message on this topic was sent by me, and contains
a full description of what I believe is the source of the problem:

    http://sources.redhat.com/ml/gdb-patches/2001-11/msg00296.html

The following patch fixes many regressions in the testsuite:
    Summary 1       2
    FAIL    1644    792
    PASS    4388    6990
    XFAIL   59      149
    XPASS   1       3

I compiled on Tru64 5.1A, using GCC.

This patch is not meant to be integrated as is, this is more a first
rough fix presented so that we can discuss more precisely which
direction to take in order to complete the work.

In particular, should the "if section-is-uninitialized then discard"
bits be always there regardless of the target? Or should we put in place
some configure test in order to execute this test only on targets were
this is needed (alpha-osf is the only one that I know of).

There is also the get_section_index function that I quickly threw in. I
suppose it should be defined in a  .h file, and implemented somewhere
else than in mdebugread.c.

Cheers,
-- 
Joel

Re: [RFC] Uninitialized section index internal error on Tru64 5.1

[Joel Brobecker]

[-- Attachment #1: Type: text/plain, Size: 268 bytes --]

> The following patch fixes many regressions in the testsuite:
>     Summary 1       2
>     FAIL    1644    792
>     PASS    4388    6990
>     XFAIL   59      149
>     XPASS   1       3

Sorry, the patch is attached, this time. 

What a fool sometimes...
-- 
Joel

[-- Attachment #2: mdebugread.c.diff --]
[-- Type: text/plain, Size: 4716 bytes --]

Index: mdebugread.c
===================================================================
RCS file: /cvs/src/src/gdb/mdebugread.c,v
retrieving revision 1.24
diff -c -3 -p -r1.24 mdebugread.c
*** mdebugread.c	19 Mar 2002 19:00:03 -0000	1.24
--- mdebugread.c	12 Apr 2002 10:02:08 -0000
*************** struct symloc
*** 143,149 ****
  		   || (sc) == scPData \
  		   || (sc) == scXData)
  #define SC_IS_COMMON(sc) ((sc) == scCommon || (sc) == scSCommon)
! #define SC_IS_BSS(sc) ((sc) == scBss || (sc) == scSBss)
  #define SC_IS_UNDEF(sc) ((sc) == scUndefined || (sc) == scSUndefined)
  \f
  /* Various complaints about symbol reading that don't abort the process */
--- 143,150 ----
  		   || (sc) == scPData \
  		   || (sc) == scXData)
  #define SC_IS_COMMON(sc) ((sc) == scCommon || (sc) == scSCommon)
! #define SC_IS_BSS(sc) ((sc) == scBss)
! #define SC_IS_SBSS(sc) ((sc) == scSBss)
  #define SC_IS_UNDEF(sc) ((sc) == scUndefined || (sc) == scSUndefined)
  \f
  /* Various complaints about symbol reading that don't abort the process */
*************** static int found_ecoff_debugging_info;
*** 327,332 ****
--- 328,335 ----
  
  /* Forward declarations */
  
+ static int get_section_index (struct objfile *, char *);
+ 
  static int upgrade_type (int, struct type **, int, union aux_ext *,
  			 int, char *);
  
*************** is_pending_symbol (FDR *fh, char *sh)
*** 637,642 ****
--- 640,657 ----
    return p;
  }
  
+ /* Return the section index for the given section name. Return -1 if
+    the section was not found. */
+ static int
+ get_section_index (struct objfile *objfile, char *section_name)
+ {
+    asection *sect = bfd_get_section_by_name (objfile->obfd, section_name);
+    if (sect)
+      return sect->index;
+    else
+      return -1;
+ }
+ 
  /* Add a new symbol SH of type T */
  
  static void
*************** parse_partial_symbols (struct objfile *o
*** 2425,2450 ****
--- 2440,2511 ----
  	      ms_type = mst_bss;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_BSS (objfile));
  	    }
+           else if (SC_IS_SBSS (ext_in->asym.sc))
+             {
+               ms_type = mst_bss;
+               svalue += ANOFFSET (objfile->section_offsets, 
+                                   get_section_index (objfile, ".sbss"));
+             }
  	  else
  	    ms_type = mst_abs;
  	  break;
  	case stLabel:
  	  /* Label */
+ 
+           /* On certain platforms, some extra label symbols can be
+              generated by the linker. One possible usage for this kind
+              of symbols is to represent the address of the begining of a
+              given section. For instance, on Tru64 5.1, the address of
+              the _ftext label is the start address of the .text section.
+ 
+              The storage class of these symbols is usually directly
+              related to the section to which the symbol refers. For
+              instance, on Tru64 5.1, the storage class for the _fdata
+              label is scData, refering to the .data section.
+ 
+              It is actually possible that the section associated to the
+              storage class of the label does not exist. On True64 5.1
+              for instance, the libm.so shared library does not contain
+              any .data section, although it contains a _fpdata label
+              which storage class is scData... Since these symbols are
+              usually useless for the debugger user anyway, we just
+              discard these symbols.
+            */
+           
  	  if (SC_IS_TEXT (ext_in->asym.sc))
  	    {
+               if (objfile->sect_index_text == -1)
+                 continue;
+                 
  	      ms_type = mst_file_text;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_TEXT (objfile));
  	    }
  	  else if (SC_IS_DATA (ext_in->asym.sc))
  	    {
+               if (objfile->sect_index_data == -1)
+                 continue;
+ 
  	      ms_type = mst_file_data;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_DATA (objfile));
  	    }
  	  else if (SC_IS_BSS (ext_in->asym.sc))
  	    {
+               if (objfile->sect_index_bss == -1)
+                 continue;
+ 
  	      ms_type = mst_file_bss;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_BSS (objfile));
  	    }
+           else if (SC_IS_SBSS (ext_in->asym.sc))
+             {
+               const int sbss_sect_index = get_section_index (objfile, ".sbss");
+ 
+               if (sbss_sect_index == -1)
+                 continue;
+ 
+               ms_type = mst_file_bss;
+               svalue += ANOFFSET (objfile->section_offsets, sbss_sect_index);
+             }
  	  else
  	    ms_type = mst_abs;
  	  break;

Re: [RFC] Uninitialized section index internal error on Tru64 5.1

[Elena Zannoni]

Joel Brobecker writes:
 > This follows up on a discussion started over a year ago... Lack of time
 > prevented me from persuing this issue. Hopefully we can make progress
 > quickly. The last message on this topic was sent by me, and contains
 > a full description of what I believe is the source of the problem:
 > 
 >     http://sources.redhat.com/ml/gdb-patches/2001-11/msg00296.html
 > 
 > The following patch fixes many regressions in the testsuite:
 >     Summary 1       2
 >     FAIL    1644    792
 >     PASS    4388    6990
 >     XFAIL   59      149
 >     XPASS   1       3
 > 
 > I compiled on Tru64 5.1A, using GCC.
 > 
 > This patch is not meant to be integrated as is, this is more a first
 > rough fix presented so that we can discuss more precisely which
 > direction to take in order to complete the work.
 > 

I don't mind this patch as is actually.

 > In particular, should the "if section-is-uninitialized then discard"
 > bits be always there regardless of the target? Or should we put in place
 > some configure test in order to execute this test only on targets were
 > this is needed (alpha-osf is the only one that I know of).
 > 

Yeah, I see your point. It seems like the real problem is that a
storage class is assigned w/o a corresponding section being
there. Does the section ever exist, does it get optimized out at some
point? Your comments explain pretty well what's going on, so I am inclined
to leave the "continue's" in.

 > There is also the get_section_index function that I quickly threw in. I
 > suppose it should be defined in a  .h file, and implemented somewhere
 > else than in mdebugread.c.
 > 

I would suggest symfile.c and symfile.h.

Elena


 > Cheers,
 > -- 
 > Joel

Re: [RFC] Uninitialized section index internal error on Tru64 5.1

[Joel Brobecker]

[-- Attachment #1: Type: text/plain, Size: 964 bytes --]

Thank you Elena for reviewing my changes. May I suggest the following
patch for inclusion, then?

2002-04-16  J. Brobecker  <brobecker@gnat.com>

        * symfile.h (get_section_index): Define.
        * symfile.c (get_section_index): New function.
        * mdebugread.c (SC_IS_SBSS): New macro.
        (SC_IS_BSS): Return true for the scBss storage class only, as
        the scSBss storage class refers to the .sbss section.
        (parse_partial_symbols): Discard the symbols which associated
        section does not exist.
        Make sure to use the .sbss section index for symbols which
        storage class is scBss, rather than using the .bss section index.

One additional note: I split the SC_IS_BSS macro into 2 macros, one for
scBss and one for scSBss. I wonder if the same should not be done for
the SC_IS_TEXT and SC_IS_DATA macros as well. So far, I haven't found a
case where this causes a problem, so left them unchanged for now.

-- 
Joel

[-- Attachment #2: sect_index_uninitialized.diff --]
[-- Type: text/plain, Size: 5618 bytes --]

Index: symfile.h
===================================================================
RCS file: /cvs/src/src/gdb/symfile.h,v
retrieving revision 1.12
diff -c -3 -p -r1.12 symfile.h
*** symfile.h	1 Feb 2002 01:14:20 -0000	1.12
--- symfile.h	16 Apr 2002 13:18:36 -0000
*************** extern void find_lowest_section (bfd *, 
*** 252,257 ****
--- 252,259 ----
  
  extern bfd *symfile_bfd_open (char *);
  
+ extern int get_section_index (struct objfile *, char *);
+ 
  /* Utility functions for overlay sections: */
  extern enum overlay_debugging_state {
    ovly_off, 
Index: symfile.c
===================================================================
RCS file: /cvs/src/src/gdb/symfile.c,v
retrieving revision 1.58
diff -c -3 -p -r1.58 symfile.c
*** symfile.c	29 Mar 2002 01:09:27 -0000	1.58
--- symfile.c	16 Apr 2002 13:18:36 -0000
*************** static void cashier_psymtab (struct part
*** 120,125 ****
--- 120,127 ----
  
  bfd *symfile_bfd_open (char *);
  
+ int get_section_index (struct objfile *, char *);
+ 
  static void find_sym_fns (struct objfile *);
  
  static void decrement_reading_symtab (void *);
*************** symfile_bfd_open (char *name)
*** 1113,1118 ****
--- 1115,1132 ----
  	     bfd_errmsg (bfd_get_error ()));
      }
    return (sym_bfd);
+ }
+ 
+ /* Return the section index for the given section name. Return -1 if
+    the section was not found. */
+ int
+ get_section_index (struct objfile *objfile, char *section_name)
+ {
+   asection *sect = bfd_get_section_by_name (objfile->obfd, section_name);
+   if (sect)
+     return sect->index;
+   else
+     return -1;
  }
  
  /* Link a new symtab_fns into the global symtab_fns list.  Called on gdb
Index: mdebugread.c
===================================================================
RCS file: /cvs/src/src/gdb/mdebugread.c,v
retrieving revision 1.24
diff -c -3 -p -r1.24 mdebugread.c
*** mdebugread.c	19 Mar 2002 19:00:03 -0000	1.24
--- mdebugread.c	16 Apr 2002 13:18:36 -0000
*************** struct symloc
*** 143,149 ****
  		   || (sc) == scPData \
  		   || (sc) == scXData)
  #define SC_IS_COMMON(sc) ((sc) == scCommon || (sc) == scSCommon)
! #define SC_IS_BSS(sc) ((sc) == scBss || (sc) == scSBss)
  #define SC_IS_UNDEF(sc) ((sc) == scUndefined || (sc) == scSUndefined)
  \f
  /* Various complaints about symbol reading that don't abort the process */
--- 143,150 ----
  		   || (sc) == scPData \
  		   || (sc) == scXData)
  #define SC_IS_COMMON(sc) ((sc) == scCommon || (sc) == scSCommon)
! #define SC_IS_BSS(sc) ((sc) == scBss)
! #define SC_IS_SBSS(sc) ((sc) == scSBss)
  #define SC_IS_UNDEF(sc) ((sc) == scUndefined || (sc) == scSUndefined)
  \f
  /* Various complaints about symbol reading that don't abort the process */
*************** parse_partial_symbols (struct objfile *o
*** 2425,2450 ****
--- 2426,2497 ----
  	      ms_type = mst_bss;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_BSS (objfile));
  	    }
+           else if (SC_IS_SBSS (ext_in->asym.sc))
+             {
+               ms_type = mst_bss;
+               svalue += ANOFFSET (objfile->section_offsets, 
+                                   get_section_index (objfile, ".sbss"));
+             }
  	  else
  	    ms_type = mst_abs;
  	  break;
  	case stLabel:
  	  /* Label */
+ 
+           /* On certain platforms, some extra label symbols can be
+              generated by the linker. One possible usage for this kind
+              of symbols is to represent the address of the begining of a
+              given section. For instance, on Tru64 5.1, the address of
+              the _ftext label is the start address of the .text section.
+ 
+              The storage class of these symbols is usually directly
+              related to the section to which the symbol refers. For
+              instance, on Tru64 5.1, the storage class for the _fdata
+              label is scData, refering to the .data section.
+ 
+              It is actually possible that the section associated to the
+              storage class of the label does not exist. On True64 5.1
+              for instance, the libm.so shared library does not contain
+              any .data section, although it contains a _fpdata label
+              which storage class is scData... Since these symbols are
+              usually useless for the debugger user anyway, we just
+              discard these symbols.
+            */
+           
  	  if (SC_IS_TEXT (ext_in->asym.sc))
  	    {
+               if (objfile->sect_index_text == -1)
+                 continue;
+                 
  	      ms_type = mst_file_text;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_TEXT (objfile));
  	    }
  	  else if (SC_IS_DATA (ext_in->asym.sc))
  	    {
+               if (objfile->sect_index_data == -1)
+                 continue;
+ 
  	      ms_type = mst_file_data;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_DATA (objfile));
  	    }
  	  else if (SC_IS_BSS (ext_in->asym.sc))
  	    {
+               if (objfile->sect_index_bss == -1)
+                 continue;
+ 
  	      ms_type = mst_file_bss;
  	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_BSS (objfile));
  	    }
+           else if (SC_IS_SBSS (ext_in->asym.sc))
+             {
+               const int sbss_sect_index = get_section_index (objfile, ".sbss");
+ 
+               if (sbss_sect_index == -1)
+                 continue;
+ 
+               ms_type = mst_file_bss;
+               svalue += ANOFFSET (objfile->section_offsets, sbss_sect_index);
+             }
  	  else
  	    ms_type = mst_abs;
  	  break;

Re: [RFC] Uninitialized section index internal error on Tru64 5.1

[Elena Zannoni]

Joel Brobecker writes:
 > Thank you Elena for reviewing my changes. May I suggest the following
 > patch for inclusion, then?


Yes, OK.

Thanks
Elena


 > 
 > 2002-04-16  J. Brobecker  <brobecker@gnat.com>
 > 
 >         * symfile.h (get_section_index): Define.
 >         * symfile.c (get_section_index): New function.
 >         * mdebugread.c (SC_IS_SBSS): New macro.
 >         (SC_IS_BSS): Return true for the scBss storage class only, as
 >         the scSBss storage class refers to the .sbss section.
 >         (parse_partial_symbols): Discard the symbols which associated
 >         section does not exist.
 >         Make sure to use the .sbss section index for symbols which
 >         storage class is scBss, rather than using the .bss section index.
 > 
 > One additional note: I split the SC_IS_BSS macro into 2 macros, one for
 > scBss and one for scSBss. I wonder if the same should not be done for
 > the SC_IS_TEXT and SC_IS_DATA macros as well. So far, I haven't found a
 > case where this causes a problem, so left them unchanged for now.
 > 
 > -- 
 > Joel
 > Index: symfile.h
 > ===================================================================
 > RCS file: /cvs/src/src/gdb/symfile.h,v
 > retrieving revision 1.12
 > diff -c -3 -p -r1.12 symfile.h
 > *** symfile.h	1 Feb 2002 01:14:20 -0000	1.12
 > --- symfile.h	16 Apr 2002 13:18:36 -0000
 > *************** extern void find_lowest_section (bfd *, 
 > *** 252,257 ****
 > --- 252,259 ----
 >   
 >   extern bfd *symfile_bfd_open (char *);
 >   
 > + extern int get_section_index (struct objfile *, char *);
 > + 
 >   /* Utility functions for overlay sections: */
 >   extern enum overlay_debugging_state {
 >     ovly_off, 
 > Index: symfile.c
 > ===================================================================
 > RCS file: /cvs/src/src/gdb/symfile.c,v
 > retrieving revision 1.58
 > diff -c -3 -p -r1.58 symfile.c
 > *** symfile.c	29 Mar 2002 01:09:27 -0000	1.58
 > --- symfile.c	16 Apr 2002 13:18:36 -0000
 > *************** static void cashier_psymtab (struct part
 > *** 120,125 ****
 > --- 120,127 ----
 >   
 >   bfd *symfile_bfd_open (char *);
 >   
 > + int get_section_index (struct objfile *, char *);
 > + 
 >   static void find_sym_fns (struct objfile *);
 >   
 >   static void decrement_reading_symtab (void *);
 > *************** symfile_bfd_open (char *name)
 > *** 1113,1118 ****
 > --- 1115,1132 ----
 >   	     bfd_errmsg (bfd_get_error ()));
 >       }
 >     return (sym_bfd);
 > + }
 > + 
 > + /* Return the section index for the given section name. Return -1 if
 > +    the section was not found. */
 > + int
 > + get_section_index (struct objfile *objfile, char *section_name)
 > + {
 > +   asection *sect = bfd_get_section_by_name (objfile->obfd, section_name);
 > +   if (sect)
 > +     return sect->index;
 > +   else
 > +     return -1;
 >   }
 >   
 >   /* Link a new symtab_fns into the global symtab_fns list.  Called on gdb
 > Index: mdebugread.c
 > ===================================================================
 > RCS file: /cvs/src/src/gdb/mdebugread.c,v
 > retrieving revision 1.24
 > diff -c -3 -p -r1.24 mdebugread.c
 > *** mdebugread.c	19 Mar 2002 19:00:03 -0000	1.24
 > --- mdebugread.c	16 Apr 2002 13:18:36 -0000
 > *************** struct symloc
 > *** 143,149 ****
 >   		   || (sc) == scPData \
 >   		   || (sc) == scXData)
 >   #define SC_IS_COMMON(sc) ((sc) == scCommon || (sc) == scSCommon)
 > ! #define SC_IS_BSS(sc) ((sc) == scBss || (sc) == scSBss)
 >   #define SC_IS_UNDEF(sc) ((sc) == scUndefined || (sc) == scSUndefined)
 >   \f
 >   /* Various complaints about symbol reading that don't abort the process */
 > --- 143,150 ----
 >   		   || (sc) == scPData \
 >   		   || (sc) == scXData)
 >   #define SC_IS_COMMON(sc) ((sc) == scCommon || (sc) == scSCommon)
 > ! #define SC_IS_BSS(sc) ((sc) == scBss)
 > ! #define SC_IS_SBSS(sc) ((sc) == scSBss)
 >   #define SC_IS_UNDEF(sc) ((sc) == scUndefined || (sc) == scSUndefined)
 >   \f
 >   /* Various complaints about symbol reading that don't abort the process */
 > *************** parse_partial_symbols (struct objfile *o
 > *** 2425,2450 ****
 > --- 2426,2497 ----
 >   	      ms_type = mst_bss;
 >   	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_BSS (objfile));
 >   	    }
 > +           else if (SC_IS_SBSS (ext_in->asym.sc))
 > +             {
 > +               ms_type = mst_bss;
 > +               svalue += ANOFFSET (objfile->section_offsets, 
 > +                                   get_section_index (objfile, ".sbss"));
 > +             }
 >   	  else
 >   	    ms_type = mst_abs;
 >   	  break;
 >   	case stLabel:
 >   	  /* Label */
 > + 
 > +           /* On certain platforms, some extra label symbols can be
 > +              generated by the linker. One possible usage for this kind
 > +              of symbols is to represent the address of the begining of a
 > +              given section. For instance, on Tru64 5.1, the address of
 > +              the _ftext label is the start address of the .text section.
 > + 
 > +              The storage class of these symbols is usually directly
 > +              related to the section to which the symbol refers. For
 > +              instance, on Tru64 5.1, the storage class for the _fdata
 > +              label is scData, refering to the .data section.
 > + 
 > +              It is actually possible that the section associated to the
 > +              storage class of the label does not exist. On True64 5.1
 > +              for instance, the libm.so shared library does not contain
 > +              any .data section, although it contains a _fpdata label
 > +              which storage class is scData... Since these symbols are
 > +              usually useless for the debugger user anyway, we just
 > +              discard these symbols.
 > +            */
 > +           
 >   	  if (SC_IS_TEXT (ext_in->asym.sc))
 >   	    {
 > +               if (objfile->sect_index_text == -1)
 > +                 continue;
 > +                 
 >   	      ms_type = mst_file_text;
 >   	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_TEXT (objfile));
 >   	    }
 >   	  else if (SC_IS_DATA (ext_in->asym.sc))
 >   	    {
 > +               if (objfile->sect_index_data == -1)
 > +                 continue;
 > + 
 >   	      ms_type = mst_file_data;
 >   	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_DATA (objfile));
 >   	    }
 >   	  else if (SC_IS_BSS (ext_in->asym.sc))
 >   	    {
 > +               if (objfile->sect_index_bss == -1)
 > +                 continue;
 > + 
 >   	      ms_type = mst_file_bss;
 >   	      svalue += ANOFFSET (objfile->section_offsets, SECT_OFF_BSS (objfile));
 >   	    }
 > +           else if (SC_IS_SBSS (ext_in->asym.sc))
 > +             {
 > +               const int sbss_sect_index = get_section_index (objfile, ".sbss");
 > + 
 > +               if (sbss_sect_index == -1)
 > +                 continue;
 > + 
 > +               ms_type = mst_file_bss;
 > +               svalue += ANOFFSET (objfile->section_offsets, sbss_sect_index);
 > +             }
 >   	  else
 >   	    ms_type = mst_abs;
 >   	  break;

Re: [RFC] Uninitialized section index internal error on Tru64 5.1

[Joel Brobecker]

> Yes, OK.

Thank you, committed in the main branch.

>  > 2002-04-16  J. Brobecker  <brobecker@gnat.com>
>  > 
>  >         * symfile.h (get_section_index): Define.
>  >         * symfile.c (get_section_index): New function.
>  >         * mdebugread.c (SC_IS_SBSS): New macro.
>  >         (SC_IS_BSS): Return true for the scBss storage class only, as
>  >         the scSBss storage class refers to the .sbss section.
>  >         (parse_partial_symbols): Discard the symbols which associated
>  >         section does not exist.
>  >         Make sure to use the .sbss section index for symbols which
>  >         storage class is scBss, rather than using the .bss section index.

-- 
Joel