PPC64 segfault using ifunc with -static and -pie

[adalava]

While investigating a problem with ifunc on FreeBSD/PPC64 kernel I came with a example code[1] that crashes with segmentation fault on FreeBSD. The key appears to be the combination of -static and -pie flags. The binary works if linked with GNU BFD (-fuse-ld=bfd)

On Linux PPC64 the binary works fine, but clang isn't actually producing a PIE binary, and it warns that -pie argument is not being used.

Results:

Build command	Result
CC=clang CFLAGS="-fuse-ld=lld" make && ./ifunc2_static	OK
CC=clang CFLAGS="-pie -fuse-ld=lld" make && ./ifunc2_static	SEGFAULT
CC=clang CFLAGS="-pie -fuse-ld=bfd" make && ./ifunc2_static	OK

I'm always availabe on IRC as adalava and I'll be happy to do any test or provide a FreeBSD VM. Thanks!

[1] https://people.freebsd.org/~alfredo/tmp/ifunc-crash.tar.gz

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

[MaskRay]

This issue is invalid.

Using -Bdynamic with -static is incorrect: it links against libc.so while the used crt file does not expect non-IRELATIVE dynamic relocations.

In driver options, -static overrides -pie so you get a -Wunused-command-line-argument warning.
I believe early static PIE systems (OpenBSD, musl) actually made -static -pie mean static PIE, but upstream GCC made a decision to use -static-pie instead
and keep the -static -pie => -static status quo. Clang emulates its behavior while giving a useful warning.

ld.lld has an extension that --export-dynamic adds PT_DYNAMIC to a position-dependent executable. A PDE almost never has PT_DYNAMIC and therefore some system libc/crt may not support it. The libc/crt should be fixed to no longer assume PT_DYNAMIC implies PIE/shared object.

On Linux, this command produces a working executable: bmake CC='clang --target=powerpc64le-linux-gnu' LDFLAGS=-fuse-ld=lld _CFLAGS='-g -Xlinker --discard-locals'