- removed a somwhat unlikely but possible division-by-zero
- in case the initial throttling calibration measures a slower cps rate than
the desired cps rate, check whether the measured rate is well below the
measured peak rate. If so, distrust the measured rate and instead use
half the peak rate as measured cps rate.
Otherwise (so measured cps is in the same ballpark as measured peak rate)
disable throttling
As initially reported in #815
Faster host systems today can get very fast instruction execution rates
for a short duration calibration test. These may be skewed by round
off error, so we now run the calibration for a minimum of 100ms.
delta time computation converting the difference between two timespec
structures values to milliseconds previously truncated the difference which,
depending on the value of when the starting value in the delta happened,
along with when an OS clock tick occurred, may have resulted in a small
delta and an apparent sleep time of 0. A more accurate result is produced
when the conversion from nanoseconds to milliseconds is rounded up
before the usecs/nsecs are truncated.
This affects the output of some SCP commands (including help). The
results are cosmetic, but allows the simulator to provide correct descriptive
information.
Some simulators have clocks that have dynamically programmable tick
rates. Such a clock is only a reliable candidate to be the calibrated
clock if it uses a single tick rate rather than changing the tick rate
on the fly. Generally most systems like this, under normal conditions
don't change their tick rates unless they're running something that is
examining the behavior of the clock system (like a diagnostic). Under
these conditions this clock is removed from the potential selection as
"the" calibrated clock that all others are relative to and if necessary,
an internal calibrated clock is selected.
- Fix incomplete migration to RTC structures indicated by Coverity
warnings. Some Coverity were minor warnings and not real issues.
- Add calibration recovery parameters for idle and catchup ticks
- Aggressively perform catchup ticks when in simulated idle paths
even when idling is disabled.
- All non internal clocks can have catch-up ticks triggered if they
register a tick unit.
- Catch-up ticks will be delivered to non tick acking simulators when
idling if regstered tick unit has been specified.
- Hosts with slow ticks can idle and keep sloppy OK time when
simulators have faster ticks
- Default to active calibration (ALWAYS) while idling (no skipping)
The original approach had separate parallel arrays for each relevant
state variable for each calibrated timer. That worked when there were
only a few state variables, the state info for a timer belongs in a
structure.
- Adjust calibration parameters to properly record catchup variables for
odd condition cases.
- Disable idle percentage calibration skipping until the correct set of
calibration variables are available.
- When a pre-calibrate operation has been performed, make the results
visible in the output of SHOW CLOCK
As discussed in #705 and #699
- MicroVAX I, II and 3900 don't have a DONE bit in the clock status
register, so sim_rtcn_tick_ack() wasn't being called to acknowledge
clock ticks.
- Timer catchup tick criteria didn't work unless the host had a slow
clock tick.
As discussed in #705
- Properly handle clock transitions when control flows back and forth
between instruction execution and simh commands.
- Changed Internal Timer from 10 Hz to the MAX(100Hz, HostOSClockHz)
- Changed default idle calibration percent to 50%
- Make sure that error cases (backwards and gap too big) properly advance
the real time while avoiding calibration.
- Fix selection of the calibrated clock.
- Fix logic that sets the idle percentage that controls calibration.
- sim_instr() returns to scp (during script execution) that start simulation
again in less than 1 tick of the calibrated clock now leverage the previous
calibration state when instruction execution resumes.
- Only generate catchup ticks for clocks that are still running.
- Revert windows sim_os_msec() implementation back to use the multi-media
timer which is required on Windows XP and shouldn't be affected by dynamic
OS time adjustments that do affect System Time.
- Reliably align clock measurements on tick boundary before measuring.
- Use consistent timebase for time measurements in sim_idle_ms_sleep()
Windows:
- Properly restore system clock tick size on exit.
- Use highest resolution time for ms time measurements
Backed out commit 484889ea5a since the overflow of the timespec tv_nsec
field was the real cause of the problem. No need for an extra mutex.
As reported in #595
It seems that the prior use of sim_asynch_lock RECURSIVE mutex could
cause a pthread_cond_timedwait() failure with EINVAL returned.
As discussed in #595
