avx2_shift_lib logic is wrong

[joebonrichie]

(Making this so Bryan sees it for libypkg rewrite).

Currently the avx2_shift_lib command in rc.yml dumps everything into /usr/lib64/haswell/*, this is wrong for packages which install themselves to /usr/lib64/packagename. avx2_shift_lib will install such packages to /usr/lib64/haswell/packagename, however, glibc actually tries to load haswell libs from /usr/lib64/packagename/haswell

Strace output from glibc 2.33 (which also searches the new glibc-hwcaps directories)

openat(AT_FDCWD, "/usr/lib64/gvfs/glibc-hwcaps/x86-64-v3/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/glibc-hwcaps/x86-64-v3", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/glibc-hwcaps/x86-64-v2/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/glibc-hwcaps/x86-64-v2", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/tls/haswell/x86_64/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/tls/haswell/x86_64", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/tls/haswell/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/tls/haswell", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/tls/x86_64/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/tls/x86_64", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/tls/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/tls", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/haswell/x86_64/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/haswell/x86_64", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/haswell/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/haswell", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/x86_64/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
newfstatat(AT_FDCWD, "/usr/lib64/gvfs/x86_64", 0x7fff9cfe9cb0, 0) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/usr/lib64/gvfs/libgvfscommon.so", O_RDONLY|O_CLOEXEC) = 4

Setting %libdir% back to /usr/lib64/haswell (which special care taken for the glibc package) or manually mangling the libs after the install step to /usr/lib64/packagename/haswell would fix this.

Potentially, .so.avx2files could be skipped so patching packages which dlopen their modules would be easier.

[DataDrake]

This is a non-issue at the moment, since avx2 is really only used with a few key libraries where performance has been thoroughly tested and the libs are in the libdir anyways. gvfs isn't something I'd generally consider a candidate for haswell optimizations.

[DataDrake]

I do agree that libdir is being incorrectly set for avx2 builds.

[joebonrichie]

gvfs was only used an an example as the strace was handy, but yeah it would unlock enabling avx2 in a few more packages, i believe I ran into this originally when looking into avx2 for octave.

[DataDrake]

I don't really think it's worth it in octave. I looked at this in the past and it gets most benefit from its dependencies getting AVX2 (fftw, openblas). The rest of the internal stuff is either a light wrapper for those functions or awful interpreter code that's beyond saving.