test/apps/test_cuda_get_symbol_address.cu build failure on Debian 12 with "relocation R_X86_64_32S against `.rodata' can not be used when making a PIE object; recompile with -fPIE"

[hexclover]

Describe the bug

On Debian 12 with GCC 12.2.0, build will fail at test/apps/test_cuda_get_symbol_address.cu, complaining about relocation R_X86_64_32S against `.rodata' can not be used when making a PIE object; recompile with -fPIE.

/usr/local/cuda-12.6/bin/nvcc -DHAVE_CONFIG_H -I. -I../..  \
-DCPU_FLAGS="" -I/home/user/ucc-dev/ucx/src -I/home/user/ucc-dev/ucx -I/home/user/ucc-dev/ucx/src -I/usr/local/cuda-12.6/include -MT test_cuda_get_symbol_address.o -MD -MF \
.deps/test_cuda_get_symbol_address.Tpo -c -o test_cuda_get_symbol_address.o \
`test -f 'test_cuda_get_symbol_address.cu' || echo './'`test_cuda_get_symbol_address.cu
mv -f .deps/test_cuda_get_symbol_address.Tpo \
.deps/test_cuda_get_symbol_address.Po
/bin/bash ../../libtool  --tag=CC   --mode=link gcc -O3 -g -Wall -Werror -funwind-tables -Wframe-larger-than=8192 -Wno-missing-field-initializers -Wno-unused-parameter -Wno-unused-label -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -Wno-pointer-sign -Werror-implicit-function-declaration -Wno-format-zero-length -Wnested-externs -Wshadow -Werror=declaration-after-statement -fPIC -L/usr/local/cuda-12.6/lib64 -L/usr/local/cuda-12.6/lib64/stubs  -o test_cuda_get_symbol_address test_cuda_get_symbol_address.o ../../src/ucp/libucp.la ../../src/ucm/libucm.la -lcudart -lrt -lrt 
libtool: link: gcc -O3 -g -Wall -Werror -funwind-tables -Wframe-larger-than=8192 -Wno-missing-field-initializers -Wno-unused-parameter -Wno-unused-label -Wno-long-long -Wno-endif-labels -Wno-sign-compare -Wno-multichar -Wno-deprecated-declarations -Winvalid-pch -Wno-pointer-sign -Werror-implicit-function-declaration -Wno-format-zero-length -Wnested-externs -Wshadow -Werror=declaration-after-statement -fPIC -o .libs/test_cuda_get_symbol_address test_cuda_get_symbol_address.o  -L/usr/local/cuda-12.6/lib64 -L/usr/local/cuda-12.6/lib64/stubs ../../src/ucp/.libs/libucp.so /home/user/ucc-dev/ucx/src/uct/.libs/libuct.so /home/user/ucc-dev/ucx/src/ucs/.libs/libucs.so -lm /home/user/ucc-dev/ucx/src/ucm/.libs/libucm.so ../../src/ucm/.libs/libucm.so -ldl -lcudart -lrt -Wl,-rpath -Wl,/home/user/ucc-dev/prefix/lib
/usr/bin/ld: test_cuda_get_symbol_address.o: relocation R_X86_64_32S against `.rodata' can not be used when making a PIE object; recompile with -fPIE

Steps to Reproduce

./autogen.sh
./configure --prefix=/home/user/ucc-dev/prefix --with-cuda=/usr/local/cuda-12.6
make -j V=1 install

Setup and versions

• OS: Debian GNU/Linux 12 + Linux Server 6.1.0-17-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.1.69-1 (2023-12-30) x86_64 GNU/Linux
• GCC: gcc (Debian 12.2.0-14) 12.2.0
• CUDA: 12.6, installed with the official .run at /usr/local/cuda-12.6
• UCX: commit e51cd2a

Additional information (depending on the issue)

It seems in this case nvcc needs a flag --compiler-options -fPIC passed to it:

diff --git a/test/apps/Makefile.am b/test/apps/Makefile.am
index 690dbc4c3..1705076c7 100644
--- a/test/apps/Makefile.am
+++ b/test/apps/Makefile.am
@@ -119,10 +119,10 @@ test_tcmalloc_LDADD    = -ldl $(TCMALLOC_LIB) \
                           $(top_builddir)/src/ucp/libucp.la
 endif
 
-.cu.o : ; $(NVCC) -c -o $@ $<
+.cu.o : ; $(NVCC) --compiler-options -fPIC -c -o $@ $<
 
 test_cuda_get_symbol_address.o: test_cuda_get_symbol_address.cu
-       $(NVCC) $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) \
+       $(NVCC) --compiler-options -fPIC $(DEFS) $(DEFAULT_INCLUDES) $(INCLUDES) \
        $(test_cuda_get_symbol_address_CPPFLAGS) -MT $@ -MD -MF \
        $(DEPDIR)/test_cuda_get_symbol_address.Tpo -c -o $@ \
        `test -f '$<' || echo '$(srcdir)/'`$<

With this change I am able to build UCX successfully.

[rakhmets]

Should be fixed by #10358.