This patch changes the syscall entry path to store the
current-frame-mask (CFM) in pt_regs->cr_ifs.  This just takes one
extra instruction (a "dep" to clear the bits other than 0-37) and is
free in terms of cycles.

The advantage of doing this is that it lets ptrace() avoid having to
walk the stack to determine the end of the user-level backing-store of
a process which is in the middle of a system-call.  Since this is what
strace does all the time, this speeds up strace quite a bit (by ~50%).
More importantly, it makes the syscall vs. non-syscall case much more
symmetric, which is always something I wanted.

Note that the change to ivt.S looks big but this is just a rippling
effect of instruction-scheduling to keep syscall latency the same.
All that's really going on there is that instead of storing 0 into
cr_ifs member we store the low 38 bits of ar.pfs.
Signed-off-by: David Mosberger <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>