This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC] A method for forcing IFUNC selector

Please contribute to benchtests so that they are more representative of a
variety of workloads.  Those tests are how we make decisions about changes
intended to improve performance, so improving them is how we can improve
that decision-making.

There is indeed no mechanism to control the selection.  It just uses cpuid
directly to set up __cpu_features and then each selector has its own logic
to choose an implementation based on that.

I am not very sanguine either about random new environment variables, nor
about a mechanism that has users picking implementations by internal symbol

As to the basic way in to do any kind of tweak like this, that is the
long-standing subject of "tunables".  We have had (for years now, I guess)
a plan to get a plan about that subject, but nobody has put real work into
even proposing a model.  (I have previously lodged strenuous objection to
willy-nilly adding magic environment variables.  But I haven't proposed a
specific alternative mechanism.)

As to how you'd go about describing what it is you want once there were a
place to put it, that also needs some thought.  I'm more positively
disposed towards ideas like a mask for cpuid feature bits, which would be
more generic and more in terms of the well-specified aspects of the
hardware rather than feeling like an API that talks in terms of
implementation internals, which seems wrong.  But I can see how that might
not actually give you the knob you really want.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]