User-space context switch and floating-point

Daniel Eischen eischen at vigrid.com
Wed Mar 3 06:16:31 PST 2004


On Wed, 3 Mar 2004, Peter Grehan wrote:

> Here's the comment I currently have in libpthread/.../powerpc/context.S:
> 
>   * XXX XXX
>   *  Floating-point is a big issue. Since there's no way to determine
>   * if the caller has used FP, all volatile register need to be saved.
>   * If FP hasn't been used, this results in a lazy FP exception in
>   * the kernel and from that point on FP is always switched in/out
>   * for the thread, which may be a big performance drag for the system.
>   *  An alternative is to use the the getcontext system call, which
>   * will do the right thing for floating point but will save all
>   * registers rather than the caller-saved subset, and has the overhead
>   * of a syscall.
>   *  Maybe another option would be to give a light-weight way for a
>   * thread to determine if FP is in used: perhaps a syscall that
>   * returns this info immediately in the asm traphandler, or an
>   * OSX-style read-only per-thread page with a flag to indicate FP state.
> 
>   Any opinions ? I noticed the alpha/sparc64 routines don't save FP.

You don't need to save floating point registers when saving the
context from userland.  The userland _thr_getcontext() function
is called as a result of an application calling pthread_yield(),
sleep(), nanosleep(), etc (yielding/sleeping calls), or by something
that blocks in userland like pthread_mutex_lock(), pthread_cond_wait(),
low-level libc/libpthread locks, etc.  Look at the i386 thr_getcontext();
it only needs to save the FP control word.

-- 
Dan Eischen



More information about the freebsd-threads mailing list