package/harfbuzz: fix static linking for test-unicode program

During static build, -lstdc++ is required to link
against libicuuc.a.

Fixes:
http://autobuild.buildroot.net/results/210/2107f9dfb39eeb6559fb4271c7af8b39aef521ca/

Signed-off-by: Romain Naour <romain.naour@openwide.fr>
---
 ...1-test-unicode-add-lstdc-for-static-build.patch | 29 ++++++++++++++++++++++
 1 file changed, 29 insertions(+)
 create mode 100644 package/harfbuzz/0001-test-unicode-add-lstdc-for-static-build.patch

Dear Romain Naour,

On Tue, 19 May 2015 00:37:40 +0200, Romain Naour wrote:
> During static build, -lstdc++ is required to link
> against libicuuc.a.
> 
> Fixes:
> http://autobuild.buildroot.net/results/210/2107f9dfb39eeb6559fb4271c7af8b39aef521ca/
> 
> Signed-off-by: Romain Naour <romain.naour@openwide.fr>
> ---
>  ...1-test-unicode-add-lstdc-for-static-build.patch | 29
> ++++++++++++++++++++++ 1 file changed, 29 insertions(+)
>  create mode 100644
> package/harfbuzz/0001-test-unicode-add-lstdc-for-static-build.patch

Maybe I'm missing something, but I still don't think this is the right
fix. harfbuzz is using pkg-config to detect icu-uc. So the problem is
really that the icu-uc .pc file does not declare its dependency on
lstdc++.

I've tried on my Ubuntu distro. Try to build the following stupid
program:

#include <unicode/utypes.h>
#include <unicode/ustring.h>

int main(void) {
 UChar         source            [42];
 u_uastrcpy(source, "This is a test.");
 return 0;
}

If you build it with:

$ gcc -static -o foo foo.c $(pkg-config --static --libs --cflags icu-uc)

it fails like our build failure. Now if you fix the icu-uc .pc file by
adding -lstdc++ -lpthread in base_libs (which gets used in
Libs.private), the build succeeds.

So shouldn't we fix icu-uc.pc instead of hardcoding the need for
-lstdc++ in harfbuzz?

Thanks,

Thomas

Hi Thomas,

Le 21/05/2015 12:37, Thomas Petazzoni a écrit :
> Dear Romain Naour,
> 
> On Tue, 19 May 2015 00:37:40 +0200, Romain Naour wrote:
>> During static build, -lstdc++ is required to link
>> against libicuuc.a.
>>
>> Fixes:
>> http://autobuild.buildroot.net/results/210/2107f9dfb39eeb6559fb4271c7af8b39aef521ca/
>>
>> Signed-off-by: Romain Naour <romain.naour@openwide.fr>
>> ---
>>  ...1-test-unicode-add-lstdc-for-static-build.patch | 29
>> ++++++++++++++++++++++ 1 file changed, 29 insertions(+)
>>  create mode 100644
>> package/harfbuzz/0001-test-unicode-add-lstdc-for-static-build.patch
> 
> Maybe I'm missing something, but I still don't think this is the right
> fix. harfbuzz is using pkg-config to detect icu-uc. So the problem is
> really that the icu-uc .pc file does not declare its dependency on
> lstdc++.
> 
> I've tried on my Ubuntu distro. Try to build the following stupid
> program:
> 
> #include <unicode/utypes.h>
> #include <unicode/ustring.h>
> 
> int main(void) {
>  UChar         source            [42];
>  u_uastrcpy(source, "This is a test.");
>  return 0;
> }
> 
> If you build it with:
> 
> $ gcc -static -o foo foo.c $(pkg-config --static --libs --cflags icu-uc)
> 
> it fails like our build failure. Now if you fix the icu-uc .pc file by
> adding -lstdc++ -lpthread in base_libs (which gets used in
> Libs.private), the build succeeds.

On Fedora 20/21 there is no icu static libraries bundled in the icu's rpm...

> 
> So shouldn't we fix icu-uc.pc instead of hardcoding the need for
> -lstdc++ in harfbuzz?

Yes you're right, but it would require invasive change in icu build system.
If we want to do things right, we needs to patch configure.ac to add -lstdc++
when --enable-static is passed to icu's configure script.

Something like:
# When building release static library, there might be some optimization flags
we can use
if test "$ENABLE_STATIC" = "YES"; then

    [snip]

    case "${host}" in
    *-linux*|i*86-*-*bsd*|i*86-pc-gnu)
        # Add -lstdc++ in Libs.private
        LIBS="${LIBS} -lstdc++"
        ;;
    *)
        ;;
    esac
fi

Unfortunately, we can't autoreconf icu since we have a patch
0003-detect-compiler-symbol-prefix.patch which modify configure script without
updating configure.ac accordingly.

./out/tmp/icudt55l_dat.S: Assembler messages:
./out/tmp/icudt55l_dat.S:1: Error: expected symbol name
./out/tmp/icudt55l_dat.S:8: Error: Missing symbol name in directive
./out/tmp/icudt55l_dat.S:8: Error: unrecognized symbol type "SYMBOL_PREFIX"
./out/tmp/icudt55l_dat.S:8: Error: junk at end of line, first unrecognized
character is `@'
./out/tmp/icudt55l_dat.S:9: Error: junk at end of line, first unrecognized
character is `@'
-- return status = 256

[snip]

/home/naourr/git/buildroot/test/harfbuzz/host/usr/bin/x86_64-linux-g++
[snip]
uconvmsg/lib@SYMBOL_PREFIX@uconvmsg.a
[snip]
uconv.cpp:(.text._ZL7initMsgPKc.part.0+0x2): undefined reference to `uconvmsg_dat'
collect2: error: ld returned 1 exit status

So at least I can patch only the configure script.

But even with icu-uc.pc fixed, test-unicode doesn't build correctly (as you said
in a previous mail)
"Unfortunately, fixing the .pc file would not fix the internal harfbuzz
test programs, since they don't use pkg-config to get the link flags."

To fix that I added $(ICU_LIBS) in test/api/Makefile.am

test_unicode_LDADD += $(top_builddir)/src/libharfbuzz-icu.la $(ICU_LIBS)

from the config.log:
ICU_LIBS='-L/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib
-licuuc -licudata -lpthread -ldl -lm  -lstdc++  '

Sadly, the build fail due to a well known uClibc issue...

  CCLD     test-unicode
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libpthread.a(lowlevellock.os):
In function `__lll_lock_wait_private':
(.text+0x0): multiple definition of `__lll_lock_wait_private'
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libc.a(libc-lowlevellock.os):(.text+0x0):
first defined here
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libpthread.a(lowlevellock.os):
In function `__lll_unlock_wake_private':
(.text+0xa0): multiple definition of `__lll_unlock_wake_private'
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libc.a(libc-lowlevellock.os):(.text+0x30):
first defined here
collect2: error: ld returned 1 exit status

I reverted this change and tried to fix libharfbuzz-icu.la instead where
-lstdc++ seems to be missing in dependency_libs

dependency_libs='
-L/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib
-licuuc -licudata -ldl
/home/naourr/buildroot-test/test/harfbuzz/build/harfbuzz-0.9.40/src/libharfbuzz.la
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libglib-2.0.la
-lintl -lpthread
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libpcre.la
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libintl.la
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libfreetype.la
-lbz2
/home/naourr/git/buildroot/test/harfbuzz/host/usr/x86_64-buildroot-linux-uclibc/sysroot/usr/lib/libpng16.la
-lz'

But I don't know how to fix this properly...

Best regards,
Romain

> 
> Thanks,
> 
> Thomas
>

>>>>> "Romain" == Romain Naour <romain.naour@openwide.fr> writes:

Hi,

>> So shouldn't we fix icu-uc.pc instead of hardcoding the need for
 >> -lstdc++ in harfbuzz?

 > Yes you're right, but it would require invasive change in icu build system.
 > If we want to do things right, we needs to patch configure.ac to add -lstdc++
 > when --enable-static is passed to icu's configure script.

Why only add it conditionally? The Libs.private stuff is only used for
static linking, so it can just be added unconditionally.