Re: Build issues with gettext 0.20.2 and 0.21 on Solaris 10 i386 with 64 bit

[Dagobert Michelsen]

Hi Bruno,

Am 22.11.2020 um 12:22 schrieb Dagobert Michelsen <d...@opencsw.org>:
> Am 17.11.2020 um 22:55 schrieb Bruno Haible <br...@clisp.org>:
>>> I am currently trying to compile the most recent gettext on Solaris 10 and 
>>> found an issue.
>>> The error shows on Solaris 10 i386 64 bit only, sparc and i386 32 bit are 
>>> not affected:
>>> 
>>> gmake[4]: Entering directory 
>>> '/home/dam/mgar/pkg/gettext/trunk/work/solaris10-i386/build-isa-amd64/gettext-0.20.2/libtextstyle/adhoc-tests'
>>> /opt/csw/bin/gcc-5.5 -DHAVE_CONFIG_H -I. -I..  -I. -I. -I.. -I../lib 
>>> -I./../lib -DSRCDIR=\"./\" -I/opt/csw/include  -O2 -pipe 
>>> -fno-omit-frame-pointer -m64 -march=opteron -MT hello.o -MD -MP -MF 
>>> .deps/hello.Tpo -c -o hello.o hello.c
>>> mv -f .deps/hello.Tpo .deps/hello.Po
>>> /bin/bash ../libtool  --tag=CC   --mode=link /opt/csw/bin/gcc-5.5  -O2 
>>> -pipe -fno-omit-frame-pointer -m64 -march=opteron  -m64 -march=opteron 
>>> -L/opt/csw/lib/64 -o hello hello.o ../lib/libtextstyle.la 
>>> libtool: link: /opt/csw/bin/gcc-5.5 -O2 -pipe -fno-omit-frame-pointer -m64 
>>> -march=opteron -m64 -march=opteron -o .libs/hello hello.o  
>>> -L/opt/csw/lib/64 ../lib/.libs/libtextstyle.so -lm -ltermcap 
>>> -R/opt/csw/lib/64
>>> Undefined                       first referenced
>>> symbol                             in file
>>> iconv_ostream_create                ../lib/.libs/libtextstyle.so
>> 
>> This symbol is only defined in HAVE_ICONV is set to 1 in config.h.
>> Can you search config.log for "checking for iconv" ?
> 
> That was a good hint! Indeed iconv is not detected properly on amd64 in the
> snapshot 2020-11-16, but it was on 0.20.1. For all other ISAs like sparc, 
> sparcv9
> and i386 the detection works fine for both compared versions.

I finally found the issue: gettext after 0.20.1 uses a more recent version of 
gnulib
with linker ABI check located in host-cpu-c-abi.m4. This ABI check was not 
present
up until gettext 0.20.1 hence compilation for amd64 worked.

Now regarding the error in detail: the check tries to determine which ABI a 
platform
can run as derived from the $host_cpu value. The error lies in the assumption 
that
$host_cpu == i386 always implies C ABI i386, which is wrong at least for 
Solaris x86
where $host_cpu is always i386 for all CPU types including 64 bit.
Maybe "isainfo -k“ should be used instead in conjunction with a Solaris 
platform test?

This patch works for me but as HOST_CPU_C_ABI can have i386 and x86_64-x32 as 
values
there is probably an intention to differentiate here:

diff --git a/m4/host-cpu-c-abi.m4 b/m4/host-cpu-c-abi.m4
index 6db2aa2..3e40db7 100644
--- a/m4/host-cpu-c-abi.m4
+++ b/m4/host-cpu-c-abi.m4
@@ -55,12 +55,7 @@ AC_DEFUN([gl_HOST_CPU_C_ABI],
     [case "$host_cpu" in
              
 changequote(,)dnl
-       i[34567]86 )
-changequote([,])dnl
-         gl_cv_host_cpu_c_abi=i386
-         ;;
-      
-       x86_64 )
+       i[34567]86 | x86_64 )
          # On x86_64 systems, the C compiler may be generating code in one of
          # these ABIs:
          # - 64-bit instruction set, 64-bit pointers, 64-bit 'long': x86_64.

Btw, the updated gettext is a requirement for gnulib CI.


Best regards

  — Dago


-- 
"You don't become great by trying to be great, you become great by wanting to 
do something,
and then doing it so hard that you become great in the process." - xkcd #896