zig build-exe should not strip indirectly needed libraries

[martinetd]

Zig Version

0.13.0
also tested latest 0.15.0-dev.37+423907c27

Steps to Reproduce and Observed Behavior

Building the following main.c against libgsl produces a binary that cannot run

• main.c:

#include <gsl/gsl_cdf.h>
#include <stdio.h>

int main(){
    double bottom_tail = gsl_cdf_gaussian_P(-1.96, 1);
    printf("Area betweeen [-1.96, 1.96]: %g\n", 1 - 2*bottom_tail);
}

• build (and comparison with gcc/clang)

$ pkg-config --cflags --libs gsl
 -lgsl -lgslcblas -lm
$ zig build-exe main.c -femit-bin=test-zig -lgsl -lgslcblas -lm
$ gcc main.c -o test-gcc -lgsl -lgslcblas -lm
$ clang main.c -o test-clang -lgsl -lgslcblas -lm
$ for t in test-*; do echo $t; ldd $t | grep gsl; done
test-clang
	libgsl.so.27 => /lib64/libgsl.so.27 (0x00007f7c2fe00000)
	libgslcblas.so.0 => /lib64/libgslcblas.so.0 (0x00007f7c301dc000)
test-gcc
	libgsl.so.27 => /lib64/libgsl.so.27 (0x00007f7213600000)
	libgslcblas.so.0 => /lib64/libgslcblas.so.0 (0x00007f7213914000)
test-zig
	libgsl.so.27 => /lib64/libgsl.so.27 (0x00007fdc05400000)
$ ./test-zig 
./test-zig: symbol lookup error: /lib64/libgsl.so.27: undefined symbol: cblas_ctrmv
$ LD_PRELOAD=/lib64/libgslcblas.so.0 ./test-zig
Area betweeen [-1.96, 1.96]: 0.950004

Expected Behavior

The immediate problem is that main.c itself doesn't depend on libgslcblas, so the library is stripped from the elf's NEEDED libs, but libgsl does need it and it's not explicit with libgsl on purpose to allow using other cblas libraries (according to https://www.gnu.org/software/gsl/doc/html/usage.html )

gcc/clang properly keep the indirectly needed libraries (even with -Wl,--as-needed), so zig probably should detect it is needed as well somehow.

Thank you!

Credits

Thanks to @ju1m for the reproducer, coming from https://discourse.nixos.org/t/gsl-dependency-cblas-in-nix-flake/38133/3

[martinetd]

(wow, there just was something in the pipes for that?) (looks unrelated)

workaround until then, just using any symbol from gslcblas works so this main.c properly runs:

#include <gsl/gsl_cdf.h>
#include <stdio.h>

extern void *cblas_ctrmv;

int main(){
    void *foo = cblas_ctrmv;
    (void) foo;
    double bottom_tail = gsl_cdf_gaussian_P(-1.96, 1);
    printf("Area betweeen [-1.96, 1.96]: %g\n", 1 - 2*bottom_tail);
}

[xdBronch]

I'm not sure how exactly this detection works but I find it odd it works differently on clang since we use LLD, you can use -needed-l instead of -l to force it

[martinetd]

Yeah, that's the reason I tried clang...
OTOH I can confirm I reproduce this with LLD, so I guess I can report this up to LLD next :|

$ ld.lld --as-needed main.o -L /lib64 -l c -l gsl -l gslcblas -l m /lib64/crt1.o 
$ ldd a.out
	linux-vdso.so.1 (0x00007f66309cb000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f66307a1000)
	libgsl.so.27 => /lib64/libgsl.so.27 (0x00007f6630400000)
	/lib64/ld-linux-x86-64.so.2 (0x00007f66309cd000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f663031a000)

$ ld.bfd --as-needed main.o -L /lib64 -l c -l gsl -l gslcblas -l m /lib64/crt1.o 
$ ldd a.out
	linux-vdso.so.1 (0x00007f6c67641000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f6c67417000)
	libgsl.so.27 => /lib64/libgsl.so.27 (0x00007f6c67000000)
	libgslcblas.so.0 => /lib64/libgslcblas.so.0 (0x00007f6c673d1000)
	/lib/ld64.so.1 => /lib64/ld-linux-x86-64.so.2 (0x00007f6c67643000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f6c672eb000)

(oh, mold does the same too, fun...)

Well, up to you what to do with this then; when I find time I'll prod lld folks about it.