Re: [PATCH 5/5] aarch64: Allow overriding HWCAP_CPUID feature check using HWCAP_MASK

[Szabolcs Nagy <nsz at port70 dot net>]

* Siddhesh Poyarekar <siddhesh@sourceware.org> [2017-05-20 09:02:20 +0530]:
> On Saturday 20 May 2017 03:42 AM, Szabolcs Nagy wrote:
> > no oom on aarch64 does not mean there is no problem
> > 
> > LD_HWCAP_MASK changes the behaviour of _dl_init_path
> > in elf/ld-load.c which we should understand before
> > the patch is committed (because it implies the
> > env var historically had different semantics than
> > what we assumed).
> 
> Ahh, sorry, I just realized that we are talking about different things.
> I thought you were objecting to including the patch until the OOM was fixed.
> 
> One may have additional search paths for every hwcap that is set by the
> kernel, so if hwcap_cpuid is set, the linker (and ldconfig) ought to
> search for /usr/lib/hwcap_cpuid for libraries to load as well.
> 
> This has an impact only if the target system is configured with
> libraries in that hwcap path, e.g. if someone decides to have additional
> libraries in /usr/lib/<hwcap feature> and that path gets searched only
> if that hwcap feature is enabled.  This allows users to ship, say,
> separate set of libraries for avx512 on x86_64 in /usr/lib/avx512_1
> which gets searched before /usr/lib.
> 
> So masking out specific bits will disable searching of libraries in
> those paths as well.  I hope that answers your question.

ok now i understand it better, my confusion is that
the default seems to be _dl_hwcap_mask == 0, which
would disable the cpuid bit unless one explicitly
set LD_HWCAP_MASK=2048, i thought we wanted the
opposite behaviour (cpuid based dispatch by default,
which can be disabled with explicit setting).