Re: Move dwarf2 frame tables to read-only section for AIX
Hi, The following patch substantially reduces the memory requirements of GCC compiled C++ programs on AIX. Currently there are two different pieces of code to decide if a read-only or a read-write section is required for exception information: - If the target supports named sections, there is good logic which considers the dwarf encoding types, flag_pic and also has a target hook for override. - If the target doesn't support named sections, the only consideration is flag_pic. The key part of this patch is removing the second sub-optimal code and using the optimal logic on all targets. I understand David has approved the target specifics, and this just needs a review for the dwarf changes. I believe I have CCed the dwarf maintainers on this email. https://gcc.gnu.org/ml/gcc-patches/2014-09/msg01897.html Thanks, Andrew
Re: Move unwind info to read-only section on AIX
Hi David,
As discussed, I have updated the to use SYMBOL-$ syntax for
PC-relative addressing and to use a new symbol, __gcc_unwind_dbase,
rather than overloading __dso_handle.
Regards,
Andrew
2014-09-09 Andrew Dixie
Move exception tables to read-only memory on AIX.
* collect2.c (write_c_file_stat): Provide dbase on AIX.
(scan_prog_file): Don't output __dso_handle nor __gcc_unwind_dbase.
* config/rs6000/aix.h (ASM_PREFERRED_EH_DATA_FORMAT): define.
(EH_TABLES_CAN_BE_READ_ONLY): define.
(ASM_OUTPUT_DWARF_PCREL): define.
(ASM_OUTPUT_DWARF_DATAREL): define.
(EH_FRAME_IN_DATA_SECTION): undefine.
* config/rs6000/rs6000-aix.c: new file.
(rs6000_aix_asm_output_dwarf_pcrel): new function.
(rs6000_aix_asm_output_dwarf_datarel): new function.
* config/rs6000/rs6000.c (rs6000_xcoff_asm_init_sections): remove
assignment of exception_section.
* dwarf2asm.c (dw2_asm_output_encoded_addr_rtx): Add call to
ASM_OUTPUT_DWARF_DATAREL.
* dwarf2out.c (switch_to_eh_frame_section): Add call to
EMIT_COLLECT2_LABELS_FOR_EH_FRAMES.
diff -rupN orig/gcc-4.10-20140706/gcc/collect2.c
gcc-4.10-20140706/gcc/collect2.c
--- orig/gcc-4.10-20140706/gcc/collect2.c 2014-06-26
21:16:28.0 +1200
+++ gcc-4.10-20140706/gcc/collect2.c2014-09-09 19:37:49.383271225 +1200
@@ -2101,12 +2110,23 @@ write_c_file_stat (FILE *stream, const c
fprintf (stream, " struct object *next;\n");
fprintf (stream, "};\n");
+ fprintf (stream, "extern void __register_frame_info_table_bases
(void *, struct object *, void *tbase, void *dbase);\n");
fprintf (stream, "extern void __register_frame_info_table (void
*, struct object *);\n");
fprintf (stream, "extern void *__deregister_frame_info (void *);\n");
+#ifdef TARGET_AIX_VERSION
+ fprintf (stream, "extern void *__gcc_unwind_dbase;\n");
+#endif
fprintf (stream, "static void reg_frame () {\n");
fprintf (stream, "\tstatic struct object ob;\n");
+#ifdef TARGET_AIX_VERSION
+ /* As per config/rs6000/rs6000-aix.c, we use __gcc_unwind_dbase
as the dbase on AIX.
+This might not be the start of the segment, we assume signed offsets.
+ */
+ fprintf (stream, "\t__register_frame_info_table_bases
(frame_table, &ob, (void *)0, &__gcc_unwind_dbase);\n");
+#else
fprintf (stream, "\t__register_frame_info_table (frame_table, &ob);\n");
+#endif
fprintf (stream, "\t}\n");
fprintf (stream, "static void dereg_frame () {\n");
@@ -2868,7 +2888,16 @@ scan_prog_file (const char *prog_name, s
provides an explicit export list. */
if (shared_obj && !is_shared
&& which_pass == PASS_OBJ && !export_flag)
- add_to_list (&exports, name);
+ {
+ /* Do not auto-export __dso_handle
or __gcc_unwind_dbase.
+ They are required to be local
to each module.
+*/
+ if (strcmp(name, "__dso_handle") != 0
+ && strcmp(name,
"__gcc_unwind_dbase") != 0)
+{
+ add_to_list (&exports, name);
+}
+ }
#endif
continue;
}
diff -rupN orig/gcc-4.10-20140706/gcc/config/rs6000/aix61.h
gcc-4.10-20140706/gcc/config/rs6000/aix61.h
--- orig/gcc-4.10-20140706/gcc/config/rs6000/aix61.h2014-01-03
11:23:26.0 +1300
+++ gcc-4.10-20140706/gcc/config/rs6000/aix61.h 2014-09-10
17:52:02.125993022 +1200
@@ -167,7 +167,7 @@ do {
\
%{!maix64:\
%{pthread:%{pg:gcrt0_r%O%s}%{!pg:%{p:mcrt0_r%O%s}%{!p:crt0_r%O%s}}}\
%{!pthread:%{pg:gcrt0%O%s}%{!pg:%{p:mcrt0%O%s}%{!p:crt0%O%s}\
- %{shared:crtcxa_s%O%s;:crtcxa%O%s}"
+ %{shared:crtcxa_s%O%s;:crtcxa%O%s} crtdbase%O%s"
/* AIX V5 typedefs ptrdiff_t as "long" while earlier releases used "int". */
diff -rupN orig/gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c
gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c
--- orig/gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c
1970-01-01 12:00:00.0 +1200
+++ gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c2014-09-09
19:39:34.770296080 +1200
@@ -0,0 +1,48 @@
+/* Functions for AIX on PowerPC.
+ Copyright (C) 2014 Free Software Foundation, Inc.
+
+This file is part of GCC.
+
+GCC is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
Re: Move unwind info to read-only section on AIX
Hi David, On AIX, mixing objects from different GCC versions should work as long as the newest GCC is used for linking. I tested a library with some objects compiled by GCC-4.4 and some objects compiled by a patched GCC-5.0. Exceptions passed through the mixed objects without issue. Did you have a specific example or reason to think it wouldn't work? If one attempts to mix C++ objects from GCC-3.3 and GCC-4.4 on Linux or Solaris there are link time errors. If you're worried about the risk of runtime problems, would you rather have a deliberate link time error on AIX? Regards, Andrew On Wed, Sep 17, 2014 at 6:37 AM, David Edelsohn wrote: > My concern is this really is an ABI change. One cannot mix object > files compiled with the old EH format and with the new EH format. One > can mix shared libraries of different styles, but not object files. > I'm not certain how much of a problem it is to require recompiling > everything from scratch.
Move dwarf2 frame tables to read-only section for AIX
Hi,
The following patch moves dwarf2 exception frame tables to read-only
memory so they can be properly shared by multiple processes on AIX.
There are two main mechanisms to register frame tables:
- EH_FRAME_SECTION_NAME puts the frames in a specially named section,
which are then registered by crtstuff.c
- EH_FRAME_IN_DATA_SECTION assigns special labels to the frames, these
labels are then located by collect2.c and loaded by generated code.
The AIX linker doesn't meet the requirements of EH_FRAME_SECTION_NAME,
and the other option EH_FRAME_IN_DATA_SECTION defaults to the
read-write data section.
I altered the dwarf2 frame and exception table generation so the
decision on whether to use a read-only or read-write section is an
independent decision from how the frame tables are registered.
I renamed EH_FRAME_IN_DATA_SECTION to EH_FRAME_THROUGH_COLLECT2, as it
now supports read-only, has slightly changed semantics, and I think
this name better reflects what it currently does rather than what it
historically did.
The spu port may see frame tables for executables move to read-only
memory. I assume this will work and is beneficial. Alternately
#define EH_TABLES_CAN_BE_READ_ONLY 0 should make it see no semantic
changes.
I believe I have an OK for the AIX changes:
https://gcc.gnu.org/ml/gcc-patches/2014-09/msg01305.html.
Can I please have a review for the dwarf2 changes?
Thanks,
Andrew
2014-09-22 Andrew Dixie
Move exception tables to read-only memory on AIX.
* dwarf2asm.c (dw2_asm_output_encoded_addr_rtx): Add call to
ASM_OUTPUT_DWARF_DATAREL.
* dwarf2out.c (switch_to_eh_frame_section): Use a read-only section
even if EH_FRAME_SECTION_NAME is undefined. Add call to
EH_FRAME_THROUGH_COLLECT2.
* except.c (switch_to_exception_section): Use a read-only section
even if EH_FRAME_SECTION_NAME is undefined.
* collect2.c (write_c_file_stat): Provide dbase on AIX.
(scan_prog_file): Don't output __dso_handle nor __gcc_unwind_dbase.
* config/rs6000/aix.h (ASM_PREFERRED_EH_DATA_FORMAT): define.
(EH_TABLES_CAN_BE_READ_ONLY): define.
(ASM_OUTPUT_DWARF_PCREL): define.
(ASM_OUTPUT_DWARF_DATAREL): define.
(EH_FRAME_IN_DATA_SECTION): undefine.
(EH_FRAME_THROUGH_COLLECT2): define.
* config/rs6000/rs6000-aix.c: new file.
(rs6000_aix_asm_output_dwarf_pcrel): new function.
(rs6000_aix_asm_output_dwarf_datarel): new function.
* config/rs6000/rs6000.c (rs6000_xcoff_asm_init_sections): remove
assignment of exception_section.
diff -rupN orig/gcc-4.10-20140706/gcc/doc/tm.texi
gcc-4.10-20140706/gcc/doc/tm.texi
--- orig/gcc-4.10-20140706/gcc/doc/tm.texi 2014-07-03
01:03:14.0 +1200
+++ gcc-4.10-20140706/gcc/doc/tm.texi 2014-09-19 21:53:11.593386708 +1200
@@ -8780,14 +8780,15 @@ You should define this symbol if your ta
unwind information and the default definition does not work.
@end defmac
-@defmac EH_FRAME_IN_DATA_SECTION
-If defined, DWARF 2 frame unwind information will be placed in the
-data section even though the target supports named sections. This
-might be necessary, for instance, if the system linker does garbage
-collection and sections cannot be marked as not to be collected.
-
-Do not define this macro unless @code{TARGET_ASM_NAMED_SECTION} is
-also defined.
+@defmac EH_FRAME_THROUGH_COLLECT2
+If defined, DWARF 2 frame unwind information will identified by
+specially named labels. The collect2 process will locate these
+labels and generate code to register the frames.
+
+This might be necessary, for instance, if the system linker will not
+place the frames in-between the sentinals from @file{crtstuff.c},
+or if the system linker does garbage collection and sections cannot
+be marked as not to be collected.
@end defmac
@defmac EH_TABLES_CAN_BE_READ_ONLY
@@ -9400,6 +9401,11 @@ A C statement to issue assembly directiv
reference to the given @var{label}, using an integer of the given @var{size}.
@end defmac
+@defmac ASM_OUTPUT_DWARF_DATAREL (@var{stream}, @var{size}, @var{label})
+A C statement to issue assembly directives that create a reference to the
+given @var{label} relative to the dbase, using an integer of the
given @var{size}.
+@end defmac
+
@defmac ASM_OUTPUT_DWARF_TABLE_REF (@var{label})
A C statement to issue assembly directives that create a reference to
the DWARF table identifier @var{label} from the current section. This
diff -rupN orig/gcc-4.10-20140706/gcc/doc/tm.texi.in
gcc-4.10-20140706/gcc/doc/tm.texi.in
--- orig/gcc-4.10-20140706/gcc/doc/tm.texi.in 2014-06-06
13:04:22.0 +1200
+++ gcc-4.10-20140706/gcc/doc/tm.texi.in2014-09-19
21:22:37.062429084 +1200
@@ -6515,14 +6515,15 @@ You should define this symbol if your ta
unwind information and the default definition does not work.
@end defmac
-@defmac EH_FRAME_IN_DATA_SECTION
-If defined, DWARF 2 frame unwind information will be placed in the
-data se
Re: Move dwarf2 frame tables to read-only section for AIX
Hi,
On Tue, Sep 23, 2014 at 2:47 AM, Joseph S. Myers
wrote:
> If you rename a target macro, the old target macro name needs to be
> poisoned in system.h.
Fixed in patch below.
> This ChangeLog entry seems very incomplete. It doesn't mention the
> changes for other architectures, or to defaults.h, or to the
> documentation, for example.
I believe the changelog below covers everything now.
Regards,
Andrew
2014-09-23 Andrew Dixie
Move exception tables to read-only memory on AIX.
* defaults.h (EH_FRAME_SECTION_NAME): Make dependent on
EH_FRAME_THROUGH_COLLECT2.
* tm.texi: Regenerate.
* tm.texi.in (EH_FRAME_IN_DATA_SECTION): Remove.
(EH_FRAME_THROUGH_COLLECT2): Document.
(ASM_OUTPUT_DWARF_DATAREL): Document.
* dwarf2asm.c (dw2_asm_output_encoded_addr_rtx): Add call to
ASM_OUTPUT_DWARF_DATAREL.
* dwarf2out.c (switch_to_eh_frame_section): Use a read-only section
even if EH_FRAME_SECTION_NAME is undefined. Add call to
EH_FRAME_THROUGH_COLLECT2.
* except.c (switch_to_exception_section): Use a read-only section
even if EH_FRAME_SECTION_NAME is undefined.
* system.h (EH_FRAME_IN_DATA_SECTION): Poison.
* collect2.c (write_c_file_stat): Provide dbase on AIX.
(scan_prog_file): Don't output __dso_handle nor __gcc_unwind_dbase.
* config/rs6000/aix.h (ASM_PREFERRED_EH_DATA_FORMAT): define.
(EH_TABLES_CAN_BE_READ_ONLY): define.
(ASM_OUTPUT_DWARF_PCREL): define.
(ASM_OUTPUT_DWARF_DATAREL): define.
(EH_FRAME_IN_DATA_SECTION): undefine.
(EH_FRAME_THROUGH_COLLECT2): define.
* config/rs6000/rs6000-aix.c: new file.
(rs6000_aix_asm_output_dwarf_pcrel): new function.
(rs6000_aix_asm_output_dwarf_datarel): new function.
* config/rs6000/rs6000.c (rs6000_xcoff_asm_init_sections): remove
assignment of exception_section.
* config/spu/spu-elf.h: Replace EH_FRAME_IN_DATA_SECTION with
EH_FRAME_THROUGH_COLLECT2.
* config/i386/i386-interix.h: Replace EH_FRAME_IN_DATA_SECTION with
EH_FRAME_THROUGH_COLLECT2 and EH_TABLES_CAN_BE_READ_ONLY.
* config.gcc (powerpc-ibm-aix-*): Add rs6000-aix.o.
libgcc:
* config.host (powerpc-ibm-aix-*): Add crtdbase.o.
* config/rs6000/crtdbase.c: New file.
* config/rs6000/t-aix-cxa: New file.
diff -rupN orig/gcc-4.10-20140706/gcc/defaults.h
gcc-4.10-20140706/gcc/defaults.h
--- orig/gcc-4.10-20140706/gcc/defaults.h 2014-06-03
19:27:13.0 +1200
+++ gcc-4.10-20140706/gcc/defaults.h2014-09-20 03:58:41.421848000 +1200
@@ -351,7 +351,7 @@ see the files COPYING3 and COPYING.RUNTI
/* If we have named sections, and we're using crtstuff to run ctors,
use them for registering eh frame information. */
#if defined (TARGET_ASM_NAMED_SECTION) && DWARF2_UNWIND_INFO \
-&& !defined (EH_FRAME_IN_DATA_SECTION)
+&& !defined (EH_FRAME_THROUGH_COLLECT2)
#ifndef EH_FRAME_SECTION_NAME
#define EH_FRAME_SECTION_NAME ".eh_frame"
#endif
diff -rupN orig/gcc-4.10-20140706/gcc/doc/tm.texi
gcc-4.10-20140706/gcc/doc/tm.texi
--- orig/gcc-4.10-20140706/gcc/doc/tm.texi 2014-07-03
01:03:14.0 +1200
+++ gcc-4.10-20140706/gcc/doc/tm.texi 2014-09-19 21:56:14.303290995 +1200
@@ -8780,14 +8780,15 @@ You should define this symbol if your ta
unwind information and the default definition does not work.
@end defmac
-@defmac EH_FRAME_IN_DATA_SECTION
-If defined, DWARF 2 frame unwind information will be placed in the
-data section even though the target supports named sections. This
-might be necessary, for instance, if the system linker does garbage
-collection and sections cannot be marked as not to be collected.
-
-Do not define this macro unless @code{TARGET_ASM_NAMED_SECTION} is
-also defined.
+@defmac EH_FRAME_THROUGH_COLLECT2
+If defined, DWARF 2 frame unwind information will identified by
+specially named labels. The collect2 process will locate these
+labels and generate code to register the frames.
+
+This might be necessary, for instance, if the system linker will not
+place the eh_frames in-between the sentinals from @file{crtstuff.c},
+or if the system linker does garbage collection and sections cannot
+be marked as not to be collected.
@end defmac
@defmac EH_TABLES_CAN_BE_READ_ONLY
@@ -9400,6 +9401,11 @@ A C statement to issue assembly directiv
reference to the given @var{label}, using an integer of the given @var{size}.
@end defmac
+@defmac ASM_OUTPUT_DWARF_DATAREL (@var{stream}, @var{size}, @var{label})
+A C statement to issue assembly directives that create a reference to the
+given @var{label} relative to the dbase, using an integer of the
given @var{size}.
+@end defmac
+
@defmac ASM_OUTPUT_DWARF_TABLE_REF (@var{label})
A C statement to issue assembly directives that create a reference to
the DWARF table identifier @var
Move unwind info to read-only section on AIX
Hi,
The following patch moves dwarf unwind information from the data
section to the read-only/text section on AIX. This means the memory
for the unwind information can be shared across multiple processes.
The frame tables are still registered through collect2 scanning for
"_GLOBAL_F" symbol names, however these symbols now exist in the
eh_frame section.
Previously all the unwind pointers were absolute (requiring
relocations), these have been changed to PC-relative and data-relative
offsets.
The DW_EH_PE_datarel encoding doesn't appear to have previously been
used in the compiler, but it is already implemented in the runtime
unwind code.
The data relative encoding is needed as a PC relative offset cannot
point from the text section to the data section on AIX.
For the base for the data relative offset, I couldn't find any way to
get a reference to the start of the data segment, so this patch uses
the symbol __dso_handle as an arbitrary base for the data relative
offsets.
I'm considering changing the base to be the TOC.
I'm Interested in thoughts or ideas on a better base for the data
relative offsets?
Thanks,
Andrew
2014-07-25 Andrew Dixie
Move exception tables to read-only memory on AIX.
* collect2.c (write_c_file_stat): Provide dbase on AIX.
(scan_prog_file): Don't output __dso_handle.
* config/rs6000/aix.h (ASM_PREFERRED_EH_DATA_FORMAT): define.
(EH_TABLES_CAN_BE_READ_ONLY): define.
(ASM_OUTPUT_DWARF_PCREL): define.
(ASM_OUTPUT_DWARF_DATAREL): define.
(EH_FRAME_IN_DATA_SECTION): undefine.
* config/rs6000/rs6000-aix.c: new file.
(rs6000_aix_asm_output_dwarf_pcrel): new function.
(rs6000_aix_asm_output_dwarf_datarel): new function.
* config/rs6000/rs6000.c (rs6000_xcoff_asm_init_sections): remove
assignment of exception_section.
* dwarf2asm.c (dw2_asm_output_encoded_addr_rtx): Add call to
ASM_OUTPUT_DWARF_DATAREL.
* dwarf2out.c (switch_to_eh_frame_section): Add call to
EMIT_COLLECT2_LABELS_FOR_EH_FRAMES.
diff -Nrup orig/gcc-4.10-20140706/gcc/collect2.c
gcc-4.10-20140706/gcc/collect2.c
--- orig/gcc-4.10-20140706/gcc/collect2.c 2014-06-26
21:16:28.0 +1200
+++ gcc-4.10-20140706/gcc/collect2.c2014-07-28 12:41:54.852926208 +1200
@@ -2101,12 +2110,23 @@ write_c_file_stat (FILE *stream, const c
fprintf (stream, " struct object *next;\n");
fprintf (stream, "};\n");
+ fprintf (stream, "extern void __register_frame_info_table_bases
(void *, struct object *, void *tbase, void *dbase);\n");
fprintf (stream, "extern void __register_frame_info_table (void
*, struct object *);\n");
fprintf (stream, "extern void *__deregister_frame_info (void *);\n");
+#ifdef TARGET_AIX_VERSION
+ fprintf (stream, "extern void *__dso_handle;\n");
+#endif
fprintf (stream, "static void reg_frame () {\n");
fprintf (stream, "\tstatic struct object ob;\n");
+#ifdef TARGET_AIX_VERSION
+ /* As per config/rs6000/rs6000-aix.c, we use __dso_handle as
the dbase on AIX.
+This might not be the start of the segment, we assume signed offsets.
+ */
+ fprintf (stream, "\t__register_frame_info_table_bases
(frame_table, &ob, (void *)0, &__dso_handle);\n");
+#else
fprintf (stream, "\t__register_frame_info_table (frame_table, &ob);\n");
+#endif
fprintf (stream, "\t}\n");
fprintf (stream, "static void dereg_frame () {\n");
@@ -2868,7 +2888,15 @@ scan_prog_file (const char *prog_name, s
provides an explicit export list. */
if (shared_obj && !is_shared
&& which_pass == PASS_OBJ && !export_flag)
- add_to_list (&exports, name);
+ {
+ /* Do not auto-export __dso_handle.
+ It is required to be local to
each module.
+*/
+ if (strcmp(name, "__dso_handle") != 0)
+{
+ add_to_list (&exports, name);
+}
+ }
#endif
continue;
}
diff -Nrup orig/gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c
gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c
--- orig/gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c
1970-01-01 12:00:00.0 +1200
+++ gcc-4.10-20140706/gcc/config/rs6000/rs6000-aix.c2014-07-28
12:41:54.852926208 +1200
@@ -0,0 +1,54 @@
+/* Functions for AIX on PowerPC.
+ Copyright (C) 2014 Free Software Foundation, Inc.
+
+This file is part of GCC.
+
+GCC is free softwar
Re: Move unwind info to read-only section on AIX
Hi David, On Thu, Jul 31, 2014 at 2:22 PM, David Edelsohn wrote: > As the comment in the code states DWARF2 unwind info was placed in the > data section to prevent the AIX linker from garbage collecting it. > How are you avoiding that problem? That comment is very old. I'm not sure what the original problem was. The code generated by collect2 which calls register_frame_info_table has explicit references to the DWARF unwind info. The linker won't garbage collect the unwind info as the linker will see it as used through these references. > Also, the AIX linker uses "$" for the current location, so the kludge > of creating PCREL symbols should not be necessary. I will swap it over to this. > Using __dso_handle as the base address is underwhelming. Does the using the TOC sound like a reasonable solution? Or do you know how to get a reference to the start of the data section from the AIX assembler/linker? The ld manpage lists _data, but this symbol doesn't work for shared libraries. Thanks, Andrew
