Re: Config results in broken build (PR#1014 and #1016)

[<Car...@hy...>]

Helena, and the list:

I believe I have tracked down the problems I found in ICU 1.8.1
for the Solaris X86 platform. I am running Solaris 8 (01/01
release), and the Forte 6 C and C++ compilers.

Here is a modified list of the problems, which now provides the
reasons for each of the errors in the release.

1) The 'configure' script generates a non-fatal error message
during the test to determine if the Solaris system is a Sparc V9
chip.

The error message is generated by the Sun C compiler, when the
test is called from configure:

  "checking host system type... i386-pc-solaris2.8
  checking if we have a 64-bit (sparcv9) compiler...
  CC: Warning: illegal option -xarch=v9 ignored
  no"

This is because this particular compiler does not accept the
-xarch=v9 switch, and it is complaining about the invalid
option on the command line. This test can be skipped when
configure finds that the processor type is i386. As you can
see from the message above, it already has the information
needed to determine if it can skip it.

Currently, the message does no hard, as the default is set
to no, and the failed test does not reset it. But it would
be nice to eliminate the message, and clean up the actual
test code in 'configure'.

2) The recent addition of the 3 lines at the end of the
$ICUROOT/source/config/mh-solaris file cause all symbols
to be removed from the libicudt18l.so symbol table. As a
result, no program (including ICU) can find any converter,
break file, or resource bundle.

The culprit here is the line:

BIR_LDFLAGS=  -Wl,-M,$(NAME).map -Wl,-B,symbolic -Wl,-B,eliminate

The last option, 'eliminate', is passed to the linker. Here is the
explanation of this option, from the 'ld' man page:

"-B eliminate
Causes any global symbols not assigned to a version definition to
be eliminated from the symbol table. This option achieves the same
symbol elimination as the auto-elimination directive available as
part of a mapfile version definition."

Since we do not appear to be using versioning within the library,
_ALL_ symbols are eliminated, including the 'icudata_dat' symbol,
which the entire ICU product depends on, in order to locate any
of the objects within the library file.

This option is only used on Solaris X86 and older Solaris Sparc
systems (i.e., non-v9-based systems). Editing the line to remove
the '-Wl,-B,eliminate' will resolve the entire problem. After
doing this, I was able to do a complete 'make ; make clean' and
obtain 100% pass results on all the internal tests.

3) The 'rpath' switch:
I was mistaken on this. The RPATH switch is not being used in
the default makefiles. I was fooled by the truss output, where
ICU was looking for the individual files, since it couldn't
find any data in the shared data library.

4) Because of item 2 above, 'make check' always fails, with the
indication that the default converter cannot be opened, no matter
what the default converter happens to be.

** NOTE: This was recently reported to the list by
zar...@ka..., but was incorrectly diagnosed as a
missing converter file. Because of item 2 above, _ALL_
converters are missing (i.e., not globally exported from the
library).

But removing the 'eliminate' switch from the mh-solaris makefile
fragment does solve the problem in a clean way. The map file
'icudt18l.map' in the build subdirectory hides all symbols except
for the 'icudata_dat' symbol, and the required housekeeping symbols
that are needed for the initialization and termination of the
shared library. The output of 'nm -C -g libicudt18l.so' follows,
showing the only symbols left in the file.

As far as I can tell, these are the only problems in the build
process, and I have attempted to show you exactly how to resolve
them. I haven't had time to prepare a set of 'diffs' for the
affected files, but it should be very easy for someone to update
the files in CVS, and/or provide a set of replacement files for
the configure and mh-solaris files. Since the changes are very
limited, even a release note with instructions on the edits
required would be a resonable response.


Regards,
--Carl

ps: This has been an amazing release. Please give the ICU team
my sincere thanks for an outstanding product. I've built the
older versions of ICU on 5 platforms now, and the builds are
amazingly clean. But this version has cleaned up so many little
things in the build environment, it makes it much simpler to
build and use. It has really matured! Bravo!


libicudt18l.so Symbols: (Solaris 8 X86, native compiler)
[Index]   Value      Size    Type  Bind  Other Shndx   Name

[2886]  |   8033624|       0|OBJT |GLOB |0    |16     |_DYNAMIC
[2882]  |   8033824|       0|OBJT |GLOB |0    |20     |_edata
[2883]  |   8033824|       0|OBJT |GLOB |0    |21     |_end
[2885]  |   7968063|       0|OBJT |GLOB |0    |14     |_etext
[2891]  |         0|       0|NOTY |WEAK |0    |UNDEF  |_ex_deregister
[2887]  |         0|       0|NOTY |WEAK |0    |UNDEF  |_ex_register
[2884]  |   8033600|       0|OBJT |GLOB |0    |15
|_GLOBAL_OFFSET_TABLE_
[2889]  |         0|       0|OBJT |GLOB |0    |ABS
|_PROCEDURE_LINKAGE_TABLE_
[2892]  |      8556|    3816|OBJT |GLOB |0    |13     |icudata_dat
[2890]  |         0|       0|NOTY |WEAK |0    |UNDEF  |void
__Cimpl::cplus_fini()

[__1cH__CimplKcplus_fini6F_v_]
[2888]  |         0|       0|NOTY |WEAK |0    |UNDEF  |void
__Cimpl::cplus_init()

[__1cH__CimplKcplus_init6F_v_]
[2881]  |         0|       0|NOTY |WEAK |0    |UNDEF  |void
__Crun::do_exit_code_in_range(void*,void*)

[__1cG__CrunVdo_exit_code_in_range6Fpv1_v_]