1.2.2  2013-12-30
-----

Linux kernel 3.11, commit 14c63f17, added a feature that automatically lowers
the perf sample rate limit when it thinks that perf is using up too much CPU
time in NMI handlers.  The resulting log messages look like:

Dec 30 13:30:32 surplus2 kernel: perf samples too long (2501 > 2500), lowering kernel.perf_event_max_sample_rate to 50000

The throttling seems to be overly aggressive and with default settings it
interferes with the operation of this test suite.  To disable it, the
following command must be done by root on every boot (e.g., in rc.local),
before *any* perf recording is done:

echo 0 > /proc/sys/kernel/perf_cpu_time_max_percent

1.1
---

GCC 4.7.2 and 4.8.0 both want to do this on val5.c at -O3:

    66.54%     val5  val5               [.] cycfn.constprop.0     
               |
               --- cycfn.constprop.0
                   cycfn

To stop it, -fno-ipa-cp-clone has to be on the command line.
__attribute__((noinline,optimize("no-optimize-sibling-calls,no-ipa-cp-clone")))
does not work.

Rather than do that, the default OPTLVL has been reduced to -O2.

1.0
---

In the code used for busy-work, the multiplication M_PI*3 is done at compile
time.  The remaining run-time busy-work is sufficient, so the code was left
as-is.
