Re: KDE a gcc

Re: KDE a gcc

[Jan Hubicka]

Hi,
this patch implements idea described in:
> http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment

basically in KDE/QT applications there is lots of pregenerated data that is
read only, but in PIC compilation it isn't because of dynamic linking.  This
code is intermixed with other readwrite data then.  With prelinking this is a
waste, as the pregenereated data is really readonly so it should be put
together and without it is waste too, as dynamic linker needs to COW much more
pages.

This patch adds two new sections. .data.rel is used for read/write data that
contains relocations (so it will be COWed on startup in case no prelinking is
done) and .data.rel.ro contains read/only data with relocations that are
read-only for prelinking and COWed on startup.

Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
resources with the prelinking. He estimates savings for KDE application to
about 200Kb.  I made simple test with sartup time on the c++ library from
crystal space w/o prelinking and got about 6% savings in startup time too.

Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
.data subsections on all systems.  The patch uses the named sections by
default, but allows to define new opcodes for that too, hope it is flexible
enought and will not bring considerable compatibility issues.

Honza

Re: KDE a gcc

[Jakub Jelinek]

On Mon, Nov 12, 2001 at 03:52:58PM +0100, Jan Hubicka wrote:
> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.

As the patch is missing, I cannot comment on it, but only guess.
Is .data.rel for R/W data which need relocations?
I had in mind .data.ro.rel and .data.ro.localrel where the former would be
R/O data which is expected to have dynamic relocation while the latter would
be for R/O data which is expected to have just relative relocations.
E.g. in
extern char foo[];
char * const p[] = { foo };
char * const q[] = { "foo" };
p would go into .data.ro.rel and q into .data.ro.localrel.
For prelinking, the advantage would be that data which would be never
relocated would be put together (there will be just a R_*_RELATIVE reloc
against q and prelinked libs are already relocated to the address they will
be mapped to (unless something gets in the way)), while .data.ro.rel might
be runtime relocated.

	Jakub

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.
> 
> Honza
Oops, forgot the patch :)
Bootstrapped/regtested i686

Wed Oct 31 17:44:53 CET 2001  Jan Hubicka  <jh@susue.cz>
	* output.h (data_rel_section, data_rel_ro_section): Declare.
	* varasm.c (in_section): Add in_data_rel and in_data_rel_ro.
	(default_exception_section): Use data_rel_ro_section.
	(output_constant_def_components): Likewise.
	* elfos.h (SELECT_SECTION): Use data_rel_section
	and data_rel_ro_section for relocatable insns.
	* tm.texi (DATA_REL_SECTION_OP, DATA_REL_RO_SECTION_OP):
	document.

Index: output.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/output.h,v
retrieving revision 1.87
diff -c -3 -p -r1.87 output.h
*** output.h	2001/11/04 02:47:36	1.87
--- output.h	2001/11/12 14:42:05
*************** extern void text_section		PARAMS ((void)
*** 156,161 ****
--- 156,174 ----
  /* Tell assembler to switch to data section.  */
  extern void data_section		PARAMS ((void));
  
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ extern void data_rel_section		PARAMS ((void));
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ extern void data_rel_ro_section		PARAMS ((void));
+ 
  /* Tell assembler to make sure its in the data section.  */
  extern void force_data_section		PARAMS ((void));
  
Index: varasm.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/varasm.c,v
retrieving revision 1.226
diff -c -3 -p -r1.226 varasm.c
*** varasm.c	2001/11/11 05:56:36	1.226
--- varasm.c	2001/11/12 14:42:06
*************** static void const_str_htab_del		PARAMS (
*** 193,199 ****
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
--- 193,200 ----
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named, in_data_rel,
!   in_data_rel_ro
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
*************** data_section ()
*** 271,276 ****
--- 272,334 ----
        in_section = in_data;
      }
  }
+ 
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ void
+ data_rel_section ()
+ {
+   if (in_section == in_data_rel)
+     ;
+   else if (flag_shared_data || !flag_pic)
+     data_section ();
+ #ifdef DATA_REL_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_SECTION_ASM_OP);
+       in_section = in_data_rel;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel", 0);
+       in_section = in_data_rel;
+     }
+   else
+     data_section ();
+ #endif
+ }
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ void
+ data_rel_ro_section ()
+ {
+   if (in_section == in_data_rel_ro)
+     ;
+   else if (!flag_pic)
+     readonly_data_section ();
+ #ifdef DATA_REL_RO_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_RO_SECTION_ASM_OP);
+       in_section = in_data_rel_ro;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel.ro", 0);
+       in_section = in_data_rel_ro;
+     }
+   else
+     data_rel_section ();
+ #endif
+ }
  /* Tell assembler to ALWAYS switch to data section, in case
     it's not sure where it is.  */
  
*************** default_exception_section ()
*** 614,620 ****
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_section ();
    else
      readonly_data_section ();
  }
--- 672,678 ----
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_rel_ro_section ();
    else
      readonly_data_section ();
  }
*************** output_constant_def_contents (exp, reloc
*** 3454,3462 ****
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if (((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
! 	  || (flag_pic && reloc))
  	data_section ();
        else
  	readonly_data_section ();
  #endif
--- 3512,3521 ----
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if ((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
  	data_section ();
+       else if (flag_pic && reloc)
+ 	data_rel_ro_section ();
        else
  	readonly_data_section ();
  #endif
Index: config/elfos.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/elfos.h,v
retrieving revision 1.36
diff -c -3 -p -r1.36 elfos.h
*** elfos.h	2001/11/11 05:56:38	1.36
--- elfos.h	2001/11/12 14:42:07
*************** const_section ()						\
*** 377,388 ****
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	data_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
--- 377,394 ----
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	{							\
! 	  if (flag_pic && RELOC)				\
! 	    data_rel_section ();				\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && RELOC)				\
! 	data_rel_ro_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doc/tm.texi,v
retrieving revision 1.69
diff -c -3 -p -r1.69 tm.texi
*** tm.texi	2001/11/10 21:44:52	1.69
--- tm.texi	2001/11/12 14:42:09
*************** A C expression whose value is a string, 
*** 5579,5584 ****
--- 5579,5600 ----
  assembler operation to identify the following data as writable initialized
  data.  Normally @code{"\t.data"} is right.
  
+ @findex DATA_REL_SECTION_ASM_OP
+ @item DATA_REL_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as writable initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
+ @findex DATA_REL_RO_SECTION_ASM_OP
+ @item DATA_REL_RO_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as read only initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
  @findex SHARED_SECTION_ASM_OP
  @item SHARED_SECTION_ASM_OP
  If defined, a C expression whose value is a string, including spacing,

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> As the patch is missing, I cannot comment on it, but only guess.
> Is .data.rel for R/W data which need relocations?
> I had in mind .data.ro.rel and .data.ro.localrel where the former would be
> R/O data which is expected to have dynamic relocation while the latter would
> be for R/O data which is expected to have just relative relocations.

By relative relocation you I guess you don't mean PC_REL...
> E.g. in
> extern char foo[];
> char * const p[] = { foo };
> char * const q[] = { "foo" };
> p would go into .data.ro.rel and q into .data.ro.localrel.

Hmm, so you mean reference to static data should be handled even more differently?
I guess I can easilly add it, as we note the static labelref anyway, but it will
need more changes to the SELECT_SECTION.

I guess this can be done as followup patch.
So to sumarize we want:
1) have all data with references together for non-prelinking dynlinking
2) have all read only data with local references together for dynlinking

So we need .data.rel, .data.rel.ro and .data.rel.ro.local
or whatever we want to name it. Sounds sane?

Honza

> For prelinking, the advantage would be that data which would be never
> relocated would be put together (there will be just a R_*_RELATIVE reloc
> against q and prelinked libs are already relocated to the address they will
> be mapped to (unless something gets in the way)), while .data.ro.rel might
> be runtime relocated.
> 
> 	Jakub

Re: KDE a gcc

[Jan Hubicka]

Hi,
this patch implements idea described in:
> http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment

basically in KDE/QT applications there is lots of pregenerated data that is
read only, but in PIC compilation it isn't because of dynamic linking.  This
code is intermixed with other readwrite data then.  With prelinking this is a
waste, as the pregenereated data is really readonly so it should be put
together and without it is waste too, as dynamic linker needs to COW much more
pages.

This patch adds two new sections. .data.rel is used for read/write data that
contains relocations (so it will be COWed on startup in case no prelinking is
done) and .data.rel.ro contains read/only data with relocations that are
read-only for prelinking and COWed on startup.

Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
resources with the prelinking. He estimates savings for KDE application to
about 200Kb.  I made simple test with sartup time on the c++ library from
crystal space w/o prelinking and got about 6% savings in startup time too.

Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
.data subsections on all systems.  The patch uses the named sections by
default, but allows to define new opcodes for that too, hope it is flexible
enought and will not bring considerable compatibility issues.

Honza

Re: KDE a gcc

[Jakub Jelinek]

On Mon, Nov 12, 2001 at 03:52:58PM +0100, Jan Hubicka wrote:
> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.

As the patch is missing, I cannot comment on it, but only guess.
Is .data.rel for R/W data which need relocations?
I had in mind .data.ro.rel and .data.ro.localrel where the former would be
R/O data which is expected to have dynamic relocation while the latter would
be for R/O data which is expected to have just relative relocations.
E.g. in
extern char foo[];
char * const p[] = { foo };
char * const q[] = { "foo" };
p would go into .data.ro.rel and q into .data.ro.localrel.
For prelinking, the advantage would be that data which would be never
relocated would be put together (there will be just a R_*_RELATIVE reloc
against q and prelinked libs are already relocated to the address they will
be mapped to (unless something gets in the way)), while .data.ro.rel might
be runtime relocated.

	Jakub

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> As the patch is missing, I cannot comment on it, but only guess.
> Is .data.rel for R/W data which need relocations?
> I had in mind .data.ro.rel and .data.ro.localrel where the former would be
> R/O data which is expected to have dynamic relocation while the latter would
> be for R/O data which is expected to have just relative relocations.

By relative relocation you I guess you don't mean PC_REL...
> E.g. in
> extern char foo[];
> char * const p[] = { foo };
> char * const q[] = { "foo" };
> p would go into .data.ro.rel and q into .data.ro.localrel.

Hmm, so you mean reference to static data should be handled even more differently?
I guess I can easilly add it, as we note the static labelref anyway, but it will
need more changes to the SELECT_SECTION.

I guess this can be done as followup patch.
So to sumarize we want:
1) have all data with references together for non-prelinking dynlinking
2) have all read only data with local references together for dynlinking

So we need .data.rel, .data.rel.ro and .data.rel.ro.local
or whatever we want to name it. Sounds sane?

Honza

> For prelinking, the advantage would be that data which would be never
> relocated would be put together (there will be just a R_*_RELATIVE reloc
> against q and prelinked libs are already relocated to the address they will
> be mapped to (unless something gets in the way)), while .data.ro.rel might
> be runtime relocated.
> 
> 	Jakub

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.
> 
> Honza
Oops, forgot the patch :)
Bootstrapped/regtested i686

Wed Oct 31 17:44:53 CET 2001  Jan Hubicka  <jh@susue.cz>
	* output.h (data_rel_section, data_rel_ro_section): Declare.
	* varasm.c (in_section): Add in_data_rel and in_data_rel_ro.
	(default_exception_section): Use data_rel_ro_section.
	(output_constant_def_components): Likewise.
	* elfos.h (SELECT_SECTION): Use data_rel_section
	and data_rel_ro_section for relocatable insns.
	* tm.texi (DATA_REL_SECTION_OP, DATA_REL_RO_SECTION_OP):
	document.

Index: output.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/output.h,v
retrieving revision 1.87
diff -c -3 -p -r1.87 output.h
*** output.h	2001/11/04 02:47:36	1.87
--- output.h	2001/11/12 14:42:05
*************** extern void text_section		PARAMS ((void)
*** 156,161 ****
--- 156,174 ----
  /* Tell assembler to switch to data section.  */
  extern void data_section		PARAMS ((void));
  
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ extern void data_rel_section		PARAMS ((void));
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ extern void data_rel_ro_section		PARAMS ((void));
+ 
  /* Tell assembler to make sure its in the data section.  */
  extern void force_data_section		PARAMS ((void));
  
Index: varasm.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/varasm.c,v
retrieving revision 1.226
diff -c -3 -p -r1.226 varasm.c
*** varasm.c	2001/11/11 05:56:36	1.226
--- varasm.c	2001/11/12 14:42:06
*************** static void const_str_htab_del		PARAMS (
*** 193,199 ****
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
--- 193,200 ----
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named, in_data_rel,
!   in_data_rel_ro
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
*************** data_section ()
*** 271,276 ****
--- 272,334 ----
        in_section = in_data;
      }
  }
+ 
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ void
+ data_rel_section ()
+ {
+   if (in_section == in_data_rel)
+     ;
+   else if (flag_shared_data || !flag_pic)
+     data_section ();
+ #ifdef DATA_REL_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_SECTION_ASM_OP);
+       in_section = in_data_rel;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel", 0);
+       in_section = in_data_rel;
+     }
+   else
+     data_section ();
+ #endif
+ }
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ void
+ data_rel_ro_section ()
+ {
+   if (in_section == in_data_rel_ro)
+     ;
+   else if (!flag_pic)
+     readonly_data_section ();
+ #ifdef DATA_REL_RO_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_RO_SECTION_ASM_OP);
+       in_section = in_data_rel_ro;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel.ro", 0);
+       in_section = in_data_rel_ro;
+     }
+   else
+     data_rel_section ();
+ #endif
+ }
  /* Tell assembler to ALWAYS switch to data section, in case
     it's not sure where it is.  */
  
*************** default_exception_section ()
*** 614,620 ****
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_section ();
    else
      readonly_data_section ();
  }
--- 672,678 ----
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_rel_ro_section ();
    else
      readonly_data_section ();
  }
*************** output_constant_def_contents (exp, reloc
*** 3454,3462 ****
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if (((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
! 	  || (flag_pic && reloc))
  	data_section ();
        else
  	readonly_data_section ();
  #endif
--- 3512,3521 ----
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if ((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
  	data_section ();
+       else if (flag_pic && reloc)
+ 	data_rel_ro_section ();
        else
  	readonly_data_section ();
  #endif
Index: config/elfos.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/elfos.h,v
retrieving revision 1.36
diff -c -3 -p -r1.36 elfos.h
*** elfos.h	2001/11/11 05:56:38	1.36
--- elfos.h	2001/11/12 14:42:07
*************** const_section ()						\
*** 377,388 ****
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	data_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
--- 377,394 ----
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	{							\
! 	  if (flag_pic && RELOC)				\
! 	    data_rel_section ();				\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && RELOC)				\
! 	data_rel_ro_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doc/tm.texi,v
retrieving revision 1.69
diff -c -3 -p -r1.69 tm.texi
*** tm.texi	2001/11/10 21:44:52	1.69
--- tm.texi	2001/11/12 14:42:09
*************** A C expression whose value is a string, 
*** 5579,5584 ****
--- 5579,5600 ----
  assembler operation to identify the following data as writable initialized
  data.  Normally @code{"\t.data"} is right.
  
+ @findex DATA_REL_SECTION_ASM_OP
+ @item DATA_REL_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as writable initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
+ @findex DATA_REL_RO_SECTION_ASM_OP
+ @item DATA_REL_RO_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as read only initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
  @findex SHARED_SECTION_ASM_OP
  @item SHARED_SECTION_ASM_OP
  If defined, a C expression whose value is a string, including spacing,

Re: KDE a gcc

[Jakub Jelinek]

On Mon, Nov 12, 2001 at 03:52:58PM +0100, Jan Hubicka wrote:
> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.

As the patch is missing, I cannot comment on it, but only guess.
Is .data.rel for R/W data which need relocations?
I had in mind .data.ro.rel and .data.ro.localrel where the former would be
R/O data which is expected to have dynamic relocation while the latter would
be for R/O data which is expected to have just relative relocations.
E.g. in
extern char foo[];
char * const p[] = { foo };
char * const q[] = { "foo" };
p would go into .data.ro.rel and q into .data.ro.localrel.
For prelinking, the advantage would be that data which would be never
relocated would be put together (there will be just a R_*_RELATIVE reloc
against q and prelinked libs are already relocated to the address they will
be mapped to (unless something gets in the way)), while .data.ro.rel might
be runtime relocated.

	Jakub

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> As the patch is missing, I cannot comment on it, but only guess.
> Is .data.rel for R/W data which need relocations?
> I had in mind .data.ro.rel and .data.ro.localrel where the former would be
> R/O data which is expected to have dynamic relocation while the latter would
> be for R/O data which is expected to have just relative relocations.

By relative relocation you I guess you don't mean PC_REL...
> E.g. in
> extern char foo[];
> char * const p[] = { foo };
> char * const q[] = { "foo" };
> p would go into .data.ro.rel and q into .data.ro.localrel.

Hmm, so you mean reference to static data should be handled even more differently?
I guess I can easilly add it, as we note the static labelref anyway, but it will
need more changes to the SELECT_SECTION.

I guess this can be done as followup patch.
So to sumarize we want:
1) have all data with references together for non-prelinking dynlinking
2) have all read only data with local references together for dynlinking

So we need .data.rel, .data.rel.ro and .data.rel.ro.local
or whatever we want to name it. Sounds sane?

Honza

> For prelinking, the advantage would be that data which would be never
> relocated would be put together (there will be just a R_*_RELATIVE reloc
> against q and prelinked libs are already relocated to the address they will
> be mapped to (unless something gets in the way)), while .data.ro.rel might
> be runtime relocated.
> 
> 	Jakub

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.
> 
> Honza
Oops, forgot the patch :)
Bootstrapped/regtested i686

Wed Oct 31 17:44:53 CET 2001  Jan Hubicka  <jh@susue.cz>
	* output.h (data_rel_section, data_rel_ro_section): Declare.
	* varasm.c (in_section): Add in_data_rel and in_data_rel_ro.
	(default_exception_section): Use data_rel_ro_section.
	(output_constant_def_components): Likewise.
	* elfos.h (SELECT_SECTION): Use data_rel_section
	and data_rel_ro_section for relocatable insns.
	* tm.texi (DATA_REL_SECTION_OP, DATA_REL_RO_SECTION_OP):
	document.

Index: output.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/output.h,v
retrieving revision 1.87
diff -c -3 -p -r1.87 output.h
*** output.h	2001/11/04 02:47:36	1.87
--- output.h	2001/11/12 14:42:05
*************** extern void text_section		PARAMS ((void)
*** 156,161 ****
--- 156,174 ----
  /* Tell assembler to switch to data section.  */
  extern void data_section		PARAMS ((void));
  
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ extern void data_rel_section		PARAMS ((void));
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ extern void data_rel_ro_section		PARAMS ((void));
+ 
  /* Tell assembler to make sure its in the data section.  */
  extern void force_data_section		PARAMS ((void));
  
Index: varasm.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/varasm.c,v
retrieving revision 1.226
diff -c -3 -p -r1.226 varasm.c
*** varasm.c	2001/11/11 05:56:36	1.226
--- varasm.c	2001/11/12 14:42:06
*************** static void const_str_htab_del		PARAMS (
*** 193,199 ****
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
--- 193,200 ----
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named, in_data_rel,
!   in_data_rel_ro
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
*************** data_section ()
*** 271,276 ****
--- 272,334 ----
        in_section = in_data;
      }
  }
+ 
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ void
+ data_rel_section ()
+ {
+   if (in_section == in_data_rel)
+     ;
+   else if (flag_shared_data || !flag_pic)
+     data_section ();
+ #ifdef DATA_REL_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_SECTION_ASM_OP);
+       in_section = in_data_rel;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel", 0);
+       in_section = in_data_rel;
+     }
+   else
+     data_section ();
+ #endif
+ }
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ void
+ data_rel_ro_section ()
+ {
+   if (in_section == in_data_rel_ro)
+     ;
+   else if (!flag_pic)
+     readonly_data_section ();
+ #ifdef DATA_REL_RO_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_RO_SECTION_ASM_OP);
+       in_section = in_data_rel_ro;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel.ro", 0);
+       in_section = in_data_rel_ro;
+     }
+   else
+     data_rel_section ();
+ #endif
+ }
  /* Tell assembler to ALWAYS switch to data section, in case
     it's not sure where it is.  */
  
*************** default_exception_section ()
*** 614,620 ****
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_section ();
    else
      readonly_data_section ();
  }
--- 672,678 ----
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_rel_ro_section ();
    else
      readonly_data_section ();
  }
*************** output_constant_def_contents (exp, reloc
*** 3454,3462 ****
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if (((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
! 	  || (flag_pic && reloc))
  	data_section ();
        else
  	readonly_data_section ();
  #endif
--- 3512,3521 ----
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if ((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
  	data_section ();
+       else if (flag_pic && reloc)
+ 	data_rel_ro_section ();
        else
  	readonly_data_section ();
  #endif
Index: config/elfos.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/elfos.h,v
retrieving revision 1.36
diff -c -3 -p -r1.36 elfos.h
*** elfos.h	2001/11/11 05:56:38	1.36
--- elfos.h	2001/11/12 14:42:07
*************** const_section ()						\
*** 377,388 ****
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	data_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
--- 377,394 ----
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	{							\
! 	  if (flag_pic && RELOC)				\
! 	    data_rel_section ();				\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && RELOC)				\
! 	data_rel_ro_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doc/tm.texi,v
retrieving revision 1.69
diff -c -3 -p -r1.69 tm.texi
*** tm.texi	2001/11/10 21:44:52	1.69
--- tm.texi	2001/11/12 14:42:09
*************** A C expression whose value is a string, 
*** 5579,5584 ****
--- 5579,5600 ----
  assembler operation to identify the following data as writable initialized
  data.  Normally @code{"\t.data"} is right.
  
+ @findex DATA_REL_SECTION_ASM_OP
+ @item DATA_REL_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as writable initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
+ @findex DATA_REL_RO_SECTION_ASM_OP
+ @item DATA_REL_RO_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as read only initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
  @findex SHARED_SECTION_ASM_OP
  @item SHARED_SECTION_ASM_OP
  If defined, a C expression whose value is a string, including spacing,

Re: KDE a gcc

[Jan Hubicka]

Hi,
this patch implements idea described in:
> http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment

basically in KDE/QT applications there is lots of pregenerated data that is
read only, but in PIC compilation it isn't because of dynamic linking.  This
code is intermixed with other readwrite data then.  With prelinking this is a
waste, as the pregenereated data is really readonly so it should be put
together and without it is waste too, as dynamic linker needs to COW much more
pages.

This patch adds two new sections. .data.rel is used for read/write data that
contains relocations (so it will be COWed on startup in case no prelinking is
done) and .data.rel.ro contains read/only data with relocations that are
read-only for prelinking and COWed on startup.

Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
resources with the prelinking. He estimates savings for KDE application to
about 200Kb.  I made simple test with sartup time on the c++ library from
crystal space w/o prelinking and got about 6% savings in startup time too.

Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
.data subsections on all systems.  The patch uses the named sections by
default, but allows to define new opcodes for that too, hope it is flexible
enought and will not bring considerable compatibility issues.

Honza

KDE and gcc take III

[Jan Hubicka]

> On Wed, Nov 21, 2001 at 10:45:27AM -0500, Jakub Jelinek wrote:
> > On Mon, Nov 19, 2001 at 03:57:54PM +0100, Jan Hubicka wrote:
> > > Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
> > > 	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
> > > 	containing pointers.
> > > 	(output_addressed_constants): Check for local/external relocations.
> > > 	* elfos.h (SELECT_SECTION): Classify data section.
> > > 	* tm.texi (SELECT_SECTION): Update documentation.
> > 
> > > !       else if (flag_pic && ((RELOC) & 2))			\
> > > ! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
> > > !       else if (flag_pic && (RELOC))				\
> > > ! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
> > ...
> > > !       else if (flag_pic && ((RELOC) & 2))			\
> > > ! 	named_section (NULL_TREE, ".data.ro.rel", RELOC);	\
> > > !       else if (flag_pic && (RELOC))				\
> > > ! 	named_section (NULL_TREE, ".data.ro.rel.local", RELOC);	\
> > 
> > You should keep the same order of suffixes all the time.
> > So IMHO the latter should be .data.rel.ro and .data.rel.ro.local.
True, I've missed that. Thanks!
> 
> And another question is whether it shouldn't be done for .gnu.linkonce.d
> data similarly (ie. have
> .gnu.linkonce.d.rel.*, .gnu.linkonce.d.rel.ro.*, .gnu.linkonce.d.rel.local.*
> and .gnu.linkonce.d.rel.ro.local.*).
OK, I can do that as followup patch, to have this issue settled down. OK?

Honza

Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
	containing pointers.
	(output_addressed_constants): Check for local/external relocations.
	* elfos.h (SELECT_SECTION): Classify data section.
	* tm.texi (SELECT_SECTION): Update documentation.
Index: varasm.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/varasm.c,v
retrieving revision 1.226
diff -c -3 -p -r1.226 varasm.c
*** varasm.c	2001/11/11 05:56:36	1.226
--- varasm.c	2001/11/19 14:53:17
*************** assemble_variable (decl, top_level, at_e
*** 1676,1682 ****
  
    /* Output any data that we will need to use the address of.  */
    if (DECL_INITIAL (decl) == error_mark_node)
!     reloc = contains_pointers_p (TREE_TYPE (decl));
    else if (DECL_INITIAL (decl))
      reloc = output_addressed_constants (DECL_INITIAL (decl));
  
--- 1676,1682 ----
  
    /* Output any data that we will need to use the address of.  */
    if (DECL_INITIAL (decl) == error_mark_node)
!     reloc = contains_pointers_p (TREE_TYPE (decl)) ? 3 : 0;
    else if (DECL_INITIAL (decl))
      reloc = output_addressed_constants (DECL_INITIAL (decl));
  
*************** output_addressed_constants (exp)
*** 4204,4212 ****
  	tree constant = TREE_OPERAND (exp, 0);
  
  	while (TREE_CODE (constant) == COMPONENT_REF)
! 	  {
! 	    constant = TREE_OPERAND (constant, 0);
! 	  }
  
  	if (TREE_CODE_CLASS (TREE_CODE (constant)) == 'c'
  	    || TREE_CODE (constant) == CONSTRUCTOR)
--- 4204,4217 ----
  	tree constant = TREE_OPERAND (exp, 0);
  
  	while (TREE_CODE (constant) == COMPONENT_REF)
! 	  constant = TREE_OPERAND (constant, 0);
! 
! 	if (TREE_CODE (constant) == CONSTRUCTOR)
! 	  reloc = 1;
! 	else if (TREE_PUBLIC (constant))
! 	  reloc |= 2;
! 	else
! 	  reloc |= 1;
  
  	if (TREE_CODE_CLASS (TREE_CODE (constant)) == 'c'
  	    || TREE_CODE (constant) == CONSTRUCTOR)
*************** output_addressed_constants (exp)
*** 4214,4220 ****
  	     for addresses of variables or functions.  */
  	  output_constant_def (constant, 0);
        }
-       reloc = 1;
        break;
  
      case PLUS_EXPR:
--- 4219,4224 ----
Index: config/elfos.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/elfos.h,v
retrieving revision 1.36
diff -c -3 -p -r1.36 elfos.h
*** elfos.h	2001/11/11 05:56:38	1.36
--- elfos.h	2001/11/19 14:53:17
*************** const_section ()						\
*** 363,370 ****
  /* A C statement or statements to switch to the appropriate
     section for output of DECL.  DECL is either a `VAR_DECL' node
     or a constant of some sort.  RELOC indicates whether forming
!    the initial value of DECL requires link-time relocations.  */
  
  #undef SELECT_SECTION
  #define SELECT_SECTION(DECL, RELOC, ALIGN)			\
  {								\
--- 363,386 ----
  /* A C statement or statements to switch to the appropriate
     section for output of DECL.  DECL is either a `VAR_DECL' node
     or a constant of some sort.  RELOC indicates whether forming
!    the initial value of DECL requires link-time relocations.  
!  
!    To optimize loading of shared programs, define following subsections
!    of data section by attaching:
  
+    .rel
+      Section with this string in name contains data that do have
+      relocations, so they get grouped together and dynamic linker
+      will visit fewer pages in memory.
+    .ro
+      Marks data read only otherwise.  This is usefull with prelinking
+      as most of relocations won't be dynamically linked and thus
+      stay read only.
+    .local
+      Marks data containing relocations only to local objects.  These
+      relocation will get fully resolved by prelinking.
+  */
+ 
  #undef SELECT_SECTION
  #define SELECT_SECTION(DECL, RELOC, ALIGN)			\
  {								\
*************** const_section ()						\
*** 377,388 ****
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	data_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
--- 393,414 ----
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	{							\
! 	  if (flag_pic && (RELOC) & 2)				\
! 	    named_section (NULL_TREE, ".data.rel", RELOC);	\
! 	  else if (flag_pic && (RELOC))				\
! 	    named_section (NULL_TREE, ".data.rel.local", RELOC);\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && ((RELOC) & 2))			\
! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
!       else if (flag_pic && (RELOC))				\
! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
*************** const_section ()						\
*** 397,406 ****
      }								\
    else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || ! TREE_CONSTANT (DECL))				\
! 	data_section ();					\
        else							\
  	const_section ();					\
      }								\
--- 423,442 ----
      }								\
    else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || ! TREE_CONSTANT (DECL))				\
! 	{							\
! 	  if (flag_pic && ((RELOC) & 2))			\
! 	    named_section (NULL_TREE, ".data.rel", RELOC);	\
! 	  else if (flag_pic && (RELOC))				\
! 	    named_section (NULL_TREE, ".data.rel.local", RELOC);\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && ((RELOC) & 2))			\
! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
!       else if (flag_pic && (RELOC))				\
! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
        else							\
  	const_section ();					\
      }								\
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doc/tm.texi,v
retrieving revision 1.69
diff -c -3 -p -r1.69 tm.texi
*** tm.texi	2001/11/10 21:44:52	1.69
--- tm.texi	2001/11/19 14:53:20
*************** A C statement or statements to switch to
*** 5661,5667 ****
  output of @var{exp}.  You can assume that @var{exp} is either a
  @code{VAR_DECL} node or a constant of some sort.  @var{reloc}
  indicates whether the initial value of @var{exp} requires link-time
! relocations.  Select the section by calling @code{text_section} or one
  of the alternatives for other sections.  @var{align} is the constant
  alignment in bits.
  
--- 5661,5669 ----
  output of @var{exp}.  You can assume that @var{exp} is either a
  @code{VAR_DECL} node or a constant of some sort.  @var{reloc}
  indicates whether the initial value of @var{exp} requires link-time
! relocations.  Bit 1 is set when variable contains local relocations
! only, while bit 2 is set for global relocations.
! Select the section by calling @code{text_section} or one
  of the alternatives for other sections.  @var{align} is the constant
  alignment in bits.
  

Re: KDE a gcc

[Jakub Jelinek]

On Mon, Nov 12, 2001 at 03:52:58PM +0100, Jan Hubicka wrote:
> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.

As the patch is missing, I cannot comment on it, but only guess.
Is .data.rel for R/W data which need relocations?
I had in mind .data.ro.rel and .data.ro.localrel where the former would be
R/O data which is expected to have dynamic relocation while the latter would
be for R/O data which is expected to have just relative relocations.
E.g. in
extern char foo[];
char * const p[] = { foo };
char * const q[] = { "foo" };
p would go into .data.ro.rel and q into .data.ro.localrel.
For prelinking, the advantage would be that data which would be never
relocated would be put together (there will be just a R_*_RELATIVE reloc
against q and prelinked libs are already relocated to the address they will
be mapped to (unless something gets in the way)), while .data.ro.rel might
be runtime relocated.

	Jakub

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> As the patch is missing, I cannot comment on it, but only guess.
> Is .data.rel for R/W data which need relocations?
> I had in mind .data.ro.rel and .data.ro.localrel where the former would be
> R/O data which is expected to have dynamic relocation while the latter would
> be for R/O data which is expected to have just relative relocations.

By relative relocation you I guess you don't mean PC_REL...
> E.g. in
> extern char foo[];
> char * const p[] = { foo };
> char * const q[] = { "foo" };
> p would go into .data.ro.rel and q into .data.ro.localrel.

Hmm, so you mean reference to static data should be handled even more differently?
I guess I can easilly add it, as we note the static labelref anyway, but it will
need more changes to the SELECT_SECTION.

I guess this can be done as followup patch.
So to sumarize we want:
1) have all data with references together for non-prelinking dynlinking
2) have all read only data with local references together for dynlinking

So we need .data.rel, .data.rel.ro and .data.rel.ro.local
or whatever we want to name it. Sounds sane?

Honza

> For prelinking, the advantage would be that data which would be never
> relocated would be put together (there will be just a R_*_RELATIVE reloc
> against q and prelinked libs are already relocated to the address they will
> be mapped to (unless something gets in the way)), while .data.ro.rel might
> be runtime relocated.
> 
> 	Jakub

Re: KDE a gcc

[Jan Hubicka]

Hi,
this patch implements idea described in:
> http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment

basically in KDE/QT applications there is lots of pregenerated data that is
read only, but in PIC compilation it isn't because of dynamic linking.  This
code is intermixed with other readwrite data then.  With prelinking this is a
waste, as the pregenereated data is really readonly so it should be put
together and without it is waste too, as dynamic linker needs to COW much more
pages.

This patch adds two new sections. .data.rel is used for read/write data that
contains relocations (so it will be COWed on startup in case no prelinking is
done) and .data.rel.ro contains read/only data with relocations that are
read-only for prelinking and COWed on startup.

Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
resources with the prelinking. He estimates savings for KDE application to
about 200Kb.  I made simple test with sartup time on the c++ library from
crystal space w/o prelinking and got about 6% savings in startup time too.

Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
.data subsections on all systems.  The patch uses the named sections by
default, but allows to define new opcodes for that too, hope it is flexible
enought and will not bring considerable compatibility issues.

Honza

Re: [patches] Re: KDE a gcc

[Jakub Jelinek]

On Mon, Nov 19, 2001 at 03:57:54PM +0100, Jan Hubicka wrote:
> Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
> 	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
> 	containing pointers.
> 	(output_addressed_constants): Check for local/external relocations.
> 	* elfos.h (SELECT_SECTION): Classify data section.
> 	* tm.texi (SELECT_SECTION): Update documentation.

> !       else if (flag_pic && ((RELOC) & 2))			\
> ! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
> !       else if (flag_pic && (RELOC))				\
> ! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
...
> !       else if (flag_pic && ((RELOC) & 2))			\
> ! 	named_section (NULL_TREE, ".data.ro.rel", RELOC);	\
> !       else if (flag_pic && (RELOC))				\
> ! 	named_section (NULL_TREE, ".data.ro.rel.local", RELOC);	\

You should keep the same order of suffixes all the time.
So IMHO the latter should be .data.rel.ro and .data.rel.ro.local.

	Jakub

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> Hi,
> this patch implements idea described in:
> > http://dforce.sh.cvut.cz/~seli/en/linking2/#data_segment
> 
> basically in KDE/QT applications there is lots of pregenerated data that is
> read only, but in PIC compilation it isn't because of dynamic linking.  This
> code is intermixed with other readwrite data then.  With prelinking this is a
> waste, as the pregenereated data is really readonly so it should be put
> together and without it is waste too, as dynamic linker needs to COW much more
> pages.
> 
> This patch adds two new sections. .data.rel is used for read/write data that
> contains relocations (so it will be COWed on startup in case no prelinking is
> done) and .data.rel.ro contains read/only data with relocations that are
> read-only for prelinking and COWed on startup.
> 
> Lubos measured about 60Kb saving on simple QT application requiring about 500Kb
> resources with the prelinking. He estimates savings for KDE application to
> about 200Kb.  I made simple test with sartup time on the c++ library from
> crystal space w/o prelinking and got about 6% savings in startup time too.
> 
> Note that I am not quite sure whether .data.rel / .data.rel.ro are handled as
> .data subsections on all systems.  The patch uses the named sections by
> default, but allows to define new opcodes for that too, hope it is flexible
> enought and will not bring considerable compatibility issues.
> 
> Honza
Oops, forgot the patch :)
Bootstrapped/regtested i686

Wed Oct 31 17:44:53 CET 2001  Jan Hubicka  <jh@susue.cz>
	* output.h (data_rel_section, data_rel_ro_section): Declare.
	* varasm.c (in_section): Add in_data_rel and in_data_rel_ro.
	(default_exception_section): Use data_rel_ro_section.
	(output_constant_def_components): Likewise.
	* elfos.h (SELECT_SECTION): Use data_rel_section
	and data_rel_ro_section for relocatable insns.
	* tm.texi (DATA_REL_SECTION_OP, DATA_REL_RO_SECTION_OP):
	document.

Index: output.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/output.h,v
retrieving revision 1.87
diff -c -3 -p -r1.87 output.h
*** output.h	2001/11/04 02:47:36	1.87
--- output.h	2001/11/12 14:42:05
*************** extern void text_section		PARAMS ((void)
*** 156,161 ****
--- 156,174 ----
  /* Tell assembler to switch to data section.  */
  extern void data_section		PARAMS ((void));
  
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ extern void data_rel_section		PARAMS ((void));
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ extern void data_rel_ro_section		PARAMS ((void));
+ 
  /* Tell assembler to make sure its in the data section.  */
  extern void force_data_section		PARAMS ((void));
  
Index: varasm.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/varasm.c,v
retrieving revision 1.226
diff -c -3 -p -r1.226 varasm.c
*** varasm.c	2001/11/11 05:56:36	1.226
--- varasm.c	2001/11/12 14:42:06
*************** static void const_str_htab_del		PARAMS (
*** 193,199 ****
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
--- 193,200 ----
  static void asm_emit_uninitialised	PARAMS ((tree, const char*, int, int));
  static void resolve_unique_section	PARAMS ((tree, int));
  \f
! static enum in_section { no_section, in_text, in_data, in_named, in_data_rel,
!   in_data_rel_ro
  #ifdef BSS_SECTION_ASM_OP
    , in_bss
  #endif
*************** data_section ()
*** 271,276 ****
--- 272,334 ----
        in_section = in_data;
      }
  }
+ 
+ /* Tell assembler to switch to data_rel section.
+    This section contains data that needs dynamic linking.
+    If there is no dedicated section for this purpose, data
+    section is used.  */
+ 
+ void
+ data_rel_section ()
+ {
+   if (in_section == in_data_rel)
+     ;
+   else if (flag_shared_data || !flag_pic)
+     data_section ();
+ #ifdef DATA_REL_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_SECTION_ASM_OP);
+       in_section = in_data_rel;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel", 0);
+       in_section = in_data_rel;
+     }
+   else
+     data_section ();
+ #endif
+ }
+ 
+ /* Like data_rel_section, but data_rel_ro section is used
+    for data needing dynamic linking but read only otherwise.
+    These data are read only with prelinking.  */
+ 
+ void
+ data_rel_ro_section ()
+ {
+   if (in_section == in_data_rel_ro)
+     ;
+   else if (!flag_pic)
+     readonly_data_section ();
+ #ifdef DATA_REL_RO_SECTION_ASM_OP
+   else
+     {
+       fprintf (asm_out_file, "%s\n", DATA_REL_RO_SECTION_ASM_OP);
+       in_section = in_data_rel_ro;
+     }
+ #else
+   else if (targetm.have_named_sections)
+     {
+       named_section (NULL_TREE, ".data.rel.ro", 0);
+       in_section = in_data_rel_ro;
+     }
+   else
+     data_rel_section ();
+ #endif
+ }
  /* Tell assembler to ALWAYS switch to data section, in case
     it's not sure where it is.  */
  
*************** default_exception_section ()
*** 614,620 ****
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_section ();
    else
      readonly_data_section ();
  }
--- 672,678 ----
    if (targetm.have_named_sections)
      named_section (NULL_TREE, ".gcc_except_table", 0);
    else if (flag_pic)
!     data_rel_ro_section ();
    else
      readonly_data_section ();
  }
*************** output_constant_def_contents (exp, reloc
*** 3454,3462 ****
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if (((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
! 	  || (flag_pic && reloc))
  	data_section ();
        else
  	readonly_data_section ();
  #endif
--- 3512,3521 ----
  #ifdef SELECT_SECTION
        SELECT_SECTION (exp, reloc, align);
  #else
!       if ((TREE_CODE (exp) == STRING_CST) && flag_writable_strings)
  	data_section ();
+       else if (flag_pic && reloc)
+ 	data_rel_ro_section ();
        else
  	readonly_data_section ();
  #endif
Index: config/elfos.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/elfos.h,v
retrieving revision 1.36
diff -c -3 -p -r1.36 elfos.h
*** elfos.h	2001/11/11 05:56:38	1.36
--- elfos.h	2001/11/12 14:42:07
*************** const_section ()						\
*** 377,388 ****
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	data_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
--- 377,394 ----
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	{							\
! 	  if (flag_pic && RELOC)				\
! 	    data_rel_section ();				\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && RELOC)				\
! 	data_rel_ro_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doc/tm.texi,v
retrieving revision 1.69
diff -c -3 -p -r1.69 tm.texi
*** tm.texi	2001/11/10 21:44:52	1.69
--- tm.texi	2001/11/12 14:42:09
*************** A C expression whose value is a string, 
*** 5579,5584 ****
--- 5579,5600 ----
  assembler operation to identify the following data as writable initialized
  data.  Normally @code{"\t.data"} is right.
  
+ @findex DATA_REL_SECTION_ASM_OP
+ @item DATA_REL_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as writable initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
+ @findex DATA_REL_RO_SECTION_ASM_OP
+ @item DATA_REL_RO_SECTION_ASM_OP
+ A C expression whose value is a string, including spacing, containing the
+ assembler operation to identify the following data as read only initialized
+ data that do need dynamic linking.
+ Used only in PIC compilation and if not present, data section is used
+ instead.
+ 
  @findex SHARED_SECTION_ASM_OP
  @item SHARED_SECTION_ASM_OP
  If defined, a C expression whose value is a string, including spacing,

Re: [patches] Re: KDE a gcc

[Jan Hubicka]

> On Mon, Nov 12, 2001 at 03:59:51PM +0100, Jan Hubicka wrote:
> > 	* output.h (data_rel_section, data_rel_ro_section): Declare.
> > 	* varasm.c (in_section): Add in_data_rel and in_data_rel_ro.
> > 	(default_exception_section): Use data_rel_ro_section.
> > 	(output_constant_def_components): Likewise.
> 
> I'd rather you handled this entirely in SELECT_SECTION.
Hi,
here is updated patch that does so and adds the local/global flag
requested by Jakub.  Hope it is OK now.
Bootstrapped mainline.

Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
	containing pointers.
	(output_addressed_constants): Check for local/external relocations.
	* elfos.h (SELECT_SECTION): Classify data section.
	* tm.texi (SELECT_SECTION): Update documentation.
Index: varasm.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/varasm.c,v
retrieving revision 1.226
diff -c -3 -p -r1.226 varasm.c
*** varasm.c	2001/11/11 05:56:36	1.226
--- varasm.c	2001/11/19 14:53:17
*************** assemble_variable (decl, top_level, at_e
*** 1676,1682 ****
  
    /* Output any data that we will need to use the address of.  */
    if (DECL_INITIAL (decl) == error_mark_node)
!     reloc = contains_pointers_p (TREE_TYPE (decl));
    else if (DECL_INITIAL (decl))
      reloc = output_addressed_constants (DECL_INITIAL (decl));
  
--- 1676,1682 ----
  
    /* Output any data that we will need to use the address of.  */
    if (DECL_INITIAL (decl) == error_mark_node)
!     reloc = contains_pointers_p (TREE_TYPE (decl)) ? 3 : 0;
    else if (DECL_INITIAL (decl))
      reloc = output_addressed_constants (DECL_INITIAL (decl));
  
*************** output_addressed_constants (exp)
*** 4204,4212 ****
  	tree constant = TREE_OPERAND (exp, 0);
  
  	while (TREE_CODE (constant) == COMPONENT_REF)
! 	  {
! 	    constant = TREE_OPERAND (constant, 0);
! 	  }
  
  	if (TREE_CODE_CLASS (TREE_CODE (constant)) == 'c'
  	    || TREE_CODE (constant) == CONSTRUCTOR)
--- 4204,4217 ----
  	tree constant = TREE_OPERAND (exp, 0);
  
  	while (TREE_CODE (constant) == COMPONENT_REF)
! 	  constant = TREE_OPERAND (constant, 0);
! 
! 	if (TREE_CODE (constant) == CONSTRUCTOR)
! 	  reloc = 1;
! 	else if (TREE_PUBLIC (constant))
! 	  reloc |= 2;
! 	else
! 	  reloc |= 1;
  
  	if (TREE_CODE_CLASS (TREE_CODE (constant)) == 'c'
  	    || TREE_CODE (constant) == CONSTRUCTOR)
*************** output_addressed_constants (exp)
*** 4214,4220 ****
  	     for addresses of variables or functions.  */
  	  output_constant_def (constant, 0);
        }
-       reloc = 1;
        break;
  
      case PLUS_EXPR:
--- 4219,4224 ----
Index: config/elfos.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/elfos.h,v
retrieving revision 1.36
diff -c -3 -p -r1.36 elfos.h
*** elfos.h	2001/11/11 05:56:38	1.36
--- elfos.h	2001/11/19 14:53:17
*************** const_section ()						\
*** 363,370 ****
  /* A C statement or statements to switch to the appropriate
     section for output of DECL.  DECL is either a `VAR_DECL' node
     or a constant of some sort.  RELOC indicates whether forming
!    the initial value of DECL requires link-time relocations.  */
  
  #undef SELECT_SECTION
  #define SELECT_SECTION(DECL, RELOC, ALIGN)			\
  {								\
--- 363,386 ----
  /* A C statement or statements to switch to the appropriate
     section for output of DECL.  DECL is either a `VAR_DECL' node
     or a constant of some sort.  RELOC indicates whether forming
!    the initial value of DECL requires link-time relocations.  
!  
!    To optimize loading of shared programs, define following subsections
!    of data section by attaching:
  
+    .rel
+      Section with this string in name contains data that do have
+      relocations, so they get grouped together and dynamic linker
+      will visit fewer pages in memory.
+    .ro
+      Marks data read only otherwise.  This is usefull with prelinking
+      as most of relocations won't be dynamically linked and thus
+      stay read only.
+    .local
+      Marks data containing relocations only to local objects.  These
+      relocation will get fully resolved by prelinking.
+  */
+ 
  #undef SELECT_SECTION
  #define SELECT_SECTION(DECL, RELOC, ALIGN)			\
  {								\
*************** const_section ()						\
*** 377,388 ****
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	data_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
--- 393,414 ----
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	{							\
! 	  if (flag_pic && (RELOC) & 2)				\
! 	    named_section (NULL_TREE, ".data.rel", RELOC);	\
! 	  else if (flag_pic && (RELOC))				\
! 	    named_section (NULL_TREE, ".data.rel.local", RELOC);\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && ((RELOC) & 2))			\
! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
!       else if (flag_pic && (RELOC))				\
! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
*************** const_section ()						\
*** 397,406 ****
      }								\
    else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || ! TREE_CONSTANT (DECL))				\
! 	data_section ();					\
        else							\
  	const_section ();					\
      }								\
--- 423,442 ----
      }								\
    else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || ! TREE_CONSTANT (DECL))				\
! 	{							\
! 	  if (flag_pic && ((RELOC) & 2))			\
! 	    named_section (NULL_TREE, ".data.rel", RELOC);	\
! 	  else if (flag_pic && (RELOC))				\
! 	    named_section (NULL_TREE, ".data.rel.local", RELOC);\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && ((RELOC) & 2))			\
! 	named_section (NULL_TREE, ".data.ro.rel", RELOC);	\
!       else if (flag_pic && (RELOC))				\
! 	named_section (NULL_TREE, ".data.ro.rel.local", RELOC);	\
        else							\
  	const_section ();					\
      }								\
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doc/tm.texi,v
retrieving revision 1.69
diff -c -3 -p -r1.69 tm.texi
*** tm.texi	2001/11/10 21:44:52	1.69
--- tm.texi	2001/11/19 14:53:20
*************** A C statement or statements to switch to
*** 5661,5667 ****
  output of @var{exp}.  You can assume that @var{exp} is either a
  @code{VAR_DECL} node or a constant of some sort.  @var{reloc}
  indicates whether the initial value of @var{exp} requires link-time
! relocations.  Select the section by calling @code{text_section} or one
  of the alternatives for other sections.  @var{align} is the constant
  alignment in bits.
  
--- 5661,5669 ----
  output of @var{exp}.  You can assume that @var{exp} is either a
  @code{VAR_DECL} node or a constant of some sort.  @var{reloc}
  indicates whether the initial value of @var{exp} requires link-time
! relocations.  Bit 1 is set when variable contains local relocations
! only, while bit 2 is set for global relocations.
! Select the section by calling @code{text_section} or one
  of the alternatives for other sections.  @var{align} is the constant
  alignment in bits.
  

Re: [patches] Re: KDE a gcc

[Jakub Jelinek]

On Wed, Nov 21, 2001 at 10:45:27AM -0500, Jakub Jelinek wrote:
> On Mon, Nov 19, 2001 at 03:57:54PM +0100, Jan Hubicka wrote:
> > Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
> > 	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
> > 	containing pointers.
> > 	(output_addressed_constants): Check for local/external relocations.
> > 	* elfos.h (SELECT_SECTION): Classify data section.
> > 	* tm.texi (SELECT_SECTION): Update documentation.
> 
> > !       else if (flag_pic && ((RELOC) & 2))			\
> > ! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
> > !       else if (flag_pic && (RELOC))				\
> > ! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
> ...
> > !       else if (flag_pic && ((RELOC) & 2))			\
> > ! 	named_section (NULL_TREE, ".data.ro.rel", RELOC);	\
> > !       else if (flag_pic && (RELOC))				\
> > ! 	named_section (NULL_TREE, ".data.ro.rel.local", RELOC);	\
> 
> You should keep the same order of suffixes all the time.
> So IMHO the latter should be .data.rel.ro and .data.rel.ro.local.

And another question is whether it shouldn't be done for .gnu.linkonce.d
data similarly (ie. have
.gnu.linkonce.d.rel.*, .gnu.linkonce.d.rel.ro.*, .gnu.linkonce.d.rel.local.*
and .gnu.linkonce.d.rel.ro.local.*).

	Jakub

Re: KDE and gcc take III

[Richard Henderson]

On Thu, Nov 22, 2001 at 02:06:33PM +0100, Jan Hubicka wrote:
> ! 	if (TREE_CODE (constant) == CONSTRUCTOR)
> ! 	  reloc = 1;

Overriding previous data?  Surely not.  Can't I reference global
data from a constructor?

> ! 	  if (flag_pic && (RELOC) & 2)				\

Parenthesis.

>     else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
>       {								\
> !       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
>   	  || ! TREE_CONSTANT (DECL))				\

Remove the test for TREE_READONLY for CONSTRUCTOR.


r~

Re: KDE and gcc take III

[Jan Hubicka]

> On Thu, Nov 22, 2001 at 02:06:33PM +0100, Jan Hubicka wrote:
> > ! 	if (TREE_CODE (constant) == CONSTRUCTOR)
> > ! 	  reloc = 1;
> 
> Overriding previous data?  Surely not.  Can't I reference global
Probably not needed anymore, but aren't we speaking about reference to
the constructor and not constructor itself?

Honza
> data from a constructor?

Re: KDE and gcc take III

[Richard Henderson]

On Sat, Nov 24, 2001 at 08:20:41PM +0100, Jan Hubicka wrote:
> Probably not needed anymore, but aren't we speaking about reference to
> the constructor and not constructor itself?

I don't know.  But surely a constructor itself would not
be marked TREE_PUBLIC, so why the extra test?


r~

Re: [patches] Re: KDE a gcc

[Jakub Jelinek]

On Wed, Nov 21, 2001 at 10:45:27AM -0500, Jakub Jelinek wrote:
> On Mon, Nov 19, 2001 at 03:57:54PM +0100, Jan Hubicka wrote:
> > Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
> > 	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
> > 	containing pointers.
> > 	(output_addressed_constants): Check for local/external relocations.
> > 	* elfos.h (SELECT_SECTION): Classify data section.
> > 	* tm.texi (SELECT_SECTION): Update documentation.
> 
> > !       else if (flag_pic && ((RELOC) & 2))			\
> > ! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
> > !       else if (flag_pic && (RELOC))				\
> > ! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
> ...
> > !       else if (flag_pic && ((RELOC) & 2))			\
> > ! 	named_section (NULL_TREE, ".data.ro.rel", RELOC);	\
> > !       else if (flag_pic && (RELOC))				\
> > ! 	named_section (NULL_TREE, ".data.ro.rel.local", RELOC);	\
> 
> You should keep the same order of suffixes all the time.
> So IMHO the latter should be .data.rel.ro and .data.rel.ro.local.

And another question is whether it shouldn't be done for .gnu.linkonce.d
data similarly (ie. have
.gnu.linkonce.d.rel.*, .gnu.linkonce.d.rel.ro.*, .gnu.linkonce.d.rel.local.*
and .gnu.linkonce.d.rel.ro.local.*).

	Jakub

KDE and gcc take III

[Jan Hubicka]

> On Wed, Nov 21, 2001 at 10:45:27AM -0500, Jakub Jelinek wrote:
> > On Mon, Nov 19, 2001 at 03:57:54PM +0100, Jan Hubicka wrote:
> > > Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
> > > 	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
> > > 	containing pointers.
> > > 	(output_addressed_constants): Check for local/external relocations.
> > > 	* elfos.h (SELECT_SECTION): Classify data section.
> > > 	* tm.texi (SELECT_SECTION): Update documentation.
> > 
> > > !       else if (flag_pic && ((RELOC) & 2))			\
> > > ! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
> > > !       else if (flag_pic && (RELOC))				\
> > > ! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
> > ...
> > > !       else if (flag_pic && ((RELOC) & 2))			\
> > > ! 	named_section (NULL_TREE, ".data.ro.rel", RELOC);	\
> > > !       else if (flag_pic && (RELOC))				\
> > > ! 	named_section (NULL_TREE, ".data.ro.rel.local", RELOC);	\
> > 
> > You should keep the same order of suffixes all the time.
> > So IMHO the latter should be .data.rel.ro and .data.rel.ro.local.
True, I've missed that. Thanks!
> 
> And another question is whether it shouldn't be done for .gnu.linkonce.d
> data similarly (ie. have
> .gnu.linkonce.d.rel.*, .gnu.linkonce.d.rel.ro.*, .gnu.linkonce.d.rel.local.*
> and .gnu.linkonce.d.rel.ro.local.*).
OK, I can do that as followup patch, to have this issue settled down. OK?

Honza

Mon Nov 19 15:53:29 CET 2001  Jan Hubicka  <jh@suse.cz>
	* varasm.c (assemble_variable): Set reloc to 3 for error_mark
	containing pointers.
	(output_addressed_constants): Check for local/external relocations.
	* elfos.h (SELECT_SECTION): Classify data section.
	* tm.texi (SELECT_SECTION): Update documentation.
Index: varasm.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/varasm.c,v
retrieving revision 1.226
diff -c -3 -p -r1.226 varasm.c
*** varasm.c	2001/11/11 05:56:36	1.226
--- varasm.c	2001/11/19 14:53:17
*************** assemble_variable (decl, top_level, at_e
*** 1676,1682 ****
  
    /* Output any data that we will need to use the address of.  */
    if (DECL_INITIAL (decl) == error_mark_node)
!     reloc = contains_pointers_p (TREE_TYPE (decl));
    else if (DECL_INITIAL (decl))
      reloc = output_addressed_constants (DECL_INITIAL (decl));
  
--- 1676,1682 ----
  
    /* Output any data that we will need to use the address of.  */
    if (DECL_INITIAL (decl) == error_mark_node)
!     reloc = contains_pointers_p (TREE_TYPE (decl)) ? 3 : 0;
    else if (DECL_INITIAL (decl))
      reloc = output_addressed_constants (DECL_INITIAL (decl));
  
*************** output_addressed_constants (exp)
*** 4204,4212 ****
  	tree constant = TREE_OPERAND (exp, 0);
  
  	while (TREE_CODE (constant) == COMPONENT_REF)
! 	  {
! 	    constant = TREE_OPERAND (constant, 0);
! 	  }
  
  	if (TREE_CODE_CLASS (TREE_CODE (constant)) == 'c'
  	    || TREE_CODE (constant) == CONSTRUCTOR)
--- 4204,4217 ----
  	tree constant = TREE_OPERAND (exp, 0);
  
  	while (TREE_CODE (constant) == COMPONENT_REF)
! 	  constant = TREE_OPERAND (constant, 0);
! 
! 	if (TREE_CODE (constant) == CONSTRUCTOR)
! 	  reloc = 1;
! 	else if (TREE_PUBLIC (constant))
! 	  reloc |= 2;
! 	else
! 	  reloc |= 1;
  
  	if (TREE_CODE_CLASS (TREE_CODE (constant)) == 'c'
  	    || TREE_CODE (constant) == CONSTRUCTOR)
*************** output_addressed_constants (exp)
*** 4214,4220 ****
  	     for addresses of variables or functions.  */
  	  output_constant_def (constant, 0);
        }
-       reloc = 1;
        break;
  
      case PLUS_EXPR:
--- 4219,4224 ----
Index: config/elfos.h
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/elfos.h,v
retrieving revision 1.36
diff -c -3 -p -r1.36 elfos.h
*** elfos.h	2001/11/11 05:56:38	1.36
--- elfos.h	2001/11/19 14:53:17
*************** const_section ()						\
*** 363,370 ****
  /* A C statement or statements to switch to the appropriate
     section for output of DECL.  DECL is either a `VAR_DECL' node
     or a constant of some sort.  RELOC indicates whether forming
!    the initial value of DECL requires link-time relocations.  */
  
  #undef SELECT_SECTION
  #define SELECT_SECTION(DECL, RELOC, ALIGN)			\
  {								\
--- 363,386 ----
  /* A C statement or statements to switch to the appropriate
     section for output of DECL.  DECL is either a `VAR_DECL' node
     or a constant of some sort.  RELOC indicates whether forming
!    the initial value of DECL requires link-time relocations.  
!  
!    To optimize loading of shared programs, define following subsections
!    of data section by attaching:
  
+    .rel
+      Section with this string in name contains data that do have
+      relocations, so they get grouped together and dynamic linker
+      will visit fewer pages in memory.
+    .ro
+      Marks data read only otherwise.  This is usefull with prelinking
+      as most of relocations won't be dynamically linked and thus
+      stay read only.
+    .local
+      Marks data containing relocations only to local objects.  These
+      relocation will get fully resolved by prelinking.
+  */
+ 
  #undef SELECT_SECTION
  #define SELECT_SECTION(DECL, RELOC, ALIGN)			\
  {								\
*************** const_section ()						\
*** 377,388 ****
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	data_section ();					\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
--- 393,414 ----
      }								\
    else if (TREE_CODE (DECL) == VAR_DECL)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || !DECL_INITIAL (DECL)				\
  	  || (DECL_INITIAL (DECL) != error_mark_node		\
  	      && !TREE_CONSTANT (DECL_INITIAL (DECL))))		\
! 	{							\
! 	  if (flag_pic && (RELOC) & 2)				\
! 	    named_section (NULL_TREE, ".data.rel", RELOC);	\
! 	  else if (flag_pic && (RELOC))				\
! 	    named_section (NULL_TREE, ".data.rel.local", RELOC);\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && ((RELOC) & 2))			\
! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
!       else if (flag_pic && (RELOC))				\
! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
        else if (flag_merge_constants < 2)			\
  	/* C and C++ don't allow different variables to share	\
  	   the same location.  -fmerge-all-constants allows	\
*************** const_section ()						\
*** 397,406 ****
      }								\
    else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
      {								\
!       if ((flag_pic && RELOC)					\
! 	  || !TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || ! TREE_CONSTANT (DECL))				\
! 	data_section ();					\
        else							\
  	const_section ();					\
      }								\
--- 423,442 ----
      }								\
    else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
      {								\
!       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
  	  || ! TREE_CONSTANT (DECL))				\
! 	{							\
! 	  if (flag_pic && ((RELOC) & 2))			\
! 	    named_section (NULL_TREE, ".data.rel", RELOC);	\
! 	  else if (flag_pic && (RELOC))				\
! 	    named_section (NULL_TREE, ".data.rel.local", RELOC);\
! 	  else							\
! 	    data_section ();					\
! 	}							\
!       else if (flag_pic && ((RELOC) & 2))			\
! 	named_section (NULL_TREE, ".data.rel.ro", RELOC);	\
!       else if (flag_pic && (RELOC))				\
! 	named_section (NULL_TREE, ".data.rel.ro.local", RELOC);	\
        else							\
  	const_section ();					\
      }								\
Index: doc/tm.texi
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doc/tm.texi,v
retrieving revision 1.69
diff -c -3 -p -r1.69 tm.texi
*** tm.texi	2001/11/10 21:44:52	1.69
--- tm.texi	2001/11/19 14:53:20
*************** A C statement or statements to switch to
*** 5661,5667 ****
  output of @var{exp}.  You can assume that @var{exp} is either a
  @code{VAR_DECL} node or a constant of some sort.  @var{reloc}
  indicates whether the initial value of @var{exp} requires link-time
! relocations.  Select the section by calling @code{text_section} or one
  of the alternatives for other sections.  @var{align} is the constant
  alignment in bits.
  
--- 5661,5669 ----
  output of @var{exp}.  You can assume that @var{exp} is either a
  @code{VAR_DECL} node or a constant of some sort.  @var{reloc}
  indicates whether the initial value of @var{exp} requires link-time
! relocations.  Bit 1 is set when variable contains local relocations
! only, while bit 2 is set for global relocations.
! Select the section by calling @code{text_section} or one
  of the alternatives for other sections.  @var{align} is the constant
  alignment in bits.
  

Re: KDE and gcc take III

[Richard Henderson]

On Thu, Nov 22, 2001 at 02:06:33PM +0100, Jan Hubicka wrote:
> ! 	if (TREE_CODE (constant) == CONSTRUCTOR)
> ! 	  reloc = 1;

Overriding previous data?  Surely not.  Can't I reference global
data from a constructor?

> ! 	  if (flag_pic && (RELOC) & 2)				\

Parenthesis.

>     else if (TREE_CODE (DECL) == CONSTRUCTOR)			\
>       {								\
> !       if (!TREE_READONLY (DECL) || TREE_SIDE_EFFECTS (DECL)	\
>   	  || ! TREE_CONSTANT (DECL))				\

Remove the test for TREE_READONLY for CONSTRUCTOR.


r~

Re: KDE and gcc take III

[Jan Hubicka]

> On Thu, Nov 22, 2001 at 02:06:33PM +0100, Jan Hubicka wrote:
> > ! 	if (TREE_CODE (constant) == CONSTRUCTOR)
> > ! 	  reloc = 1;
> 
> Overriding previous data?  Surely not.  Can't I reference global
Probably not needed anymore, but aren't we speaking about reference to
the constructor and not constructor itself?

Honza
> data from a constructor?

Re: KDE and gcc take III

[Richard Henderson]

On Sat, Nov 24, 2001 at 08:20:41PM +0100, Jan Hubicka wrote:
> Probably not needed anymore, but aren't we speaking about reference to
> the constructor and not constructor itself?

I don't know.  But surely a constructor itself would not
be marked TREE_PUBLIC, so why the extra test?


r~