Tcl Source Code

This can be seen as the amendment for [b87ad7e9146832d505f9a430d779c5313c440256]

The short description:

By NRT (near real-time) development under windows, I get several issues with drifting of the [clock microseconds] etc. back, that looks like:

  clock microseconds -> 1490985783769193
  clock microseconds -> 1490985783769191

Additionally I've own clock extension with "nanoseconds", where this is more critical and very annoying:
  clock nanoseconds -> 1490985425627129800
  clock nanoseconds -> 1490985425627128400

This goes to the calibration cycle, where the new combination of fileTimeLastCall, curCounterFreq and perfCounterLastCall breaks the forward running of the native virtual clock.

I've fixed it for tcl8.5 branch (merge-able to 8.6 and trunk), solution coming soon...

Just as reminder for me (excerpt of log-message of the fix):
```
SHA-1: a746f0501229a4a76e3c9a3825c55dbe9c8c3cdf

* optimized calibration cycle (makes Tcl for windows "RTS" resp. NRT-capable):
  - the clock ticks never backwards (avoid it by negative drifts using comparison of times before and after calibration);
  - more precise, smooth/soft drifting (avoids too large drifts, already after 10 iterations the drift gets fewer as 0.1 microseconds);
  - because of more accurate drifting (aspire to the smallest difference), we can prolong calibration interval (up to 10 seconds by small tdiff-value);
```

Newest findings here - test cmdMZ-6.6 could sporadically fail with:

---- Result was:
{1 0 1 0 1 0 1 1 1} {998.039 µs/# 51 # 1002.0 #/sec 50.900 net-ms} {999.255 µs/# 51 # 1000.7 #/sec 50.962 net-ms}
---- Result should have been (glob matching):
{1 1 1 1 1 1 1 1 1} *

Probably I should rebase parts of my event-perf-branch fixing that.

Just as PoC (or for the case someone needs it for tests resp. to check for some other platforms)...

Here is an example with 30 seconds run-time:

proc test {tm} {
  set l [clock microseconds];
  timerate {
    set i [clock microseconds];
    if {$i < $l} {error "$i < $l"};
    set l $i
  } $tm
}; 
test 30000

Similar tests have been succeeded (no backward-drifts after run-time of ca. half an hour) on physical host and virtual machines of me (single- and multi-threaded).

fixed in a minimalist way in [f6637d3dd8] (merged now), the extensions are provided in event-perf-branches (thus close now).

[f6637d3dd85af2a35101e2461c5769ebd1b27cbe] closes this

As example below small excerpt with DEBUG-output of the calibration thread:

```
*****[30F4]***** 1490992535186529 ** tdiff **-1965, 14909925351865281, 14909925351867246
*****[30F4]***** before 14909925351870010, 14909925351870010
*****[30F4]***** after  14909925351870009, 14909925351870010, diff: -1 / 0
*****[30F4]***** 1490992535187944, [10] ** cfreq == 3117204, nfreq == 3117441, perfc == 2031595984296
*****[30F4]***** 1490992545326460 ** tdiff **-2006, 14909925453264590, 14909925453266596
*****[30F4]***** before 14909925469160412, 14909925469160412
*****[30F4]***** after  14909925469160386, 14909925469160412, diff: -26 / 0
*****[30F4]***** 1490992546917626, [10] ** cfreq == 3117209, nfreq == 3117441, perfc == 2031627592529
*****[30F4]***** 1490992556043572 ** tdiff **-2206, 14909925560435703, 14909925560437909
*****[30F4]***** before 14909925560441189, 14909925560441189
*****[30F4]***** after  14909925560441188, 14909925560441189, diff: -1 / 0
*****[30F4]***** 1490992556045267, [10] ** cfreq == 3117203, nfreq == 3117441, perfc == 2031660999997
*****[30F4]***** 1490992566183518 ** tdiff **-2089, 14909925661835170, 14909925661837259
*****[30F4]***** 1490992566184086, [10] ** cfreq == 3117205, nfreq == 3117441, perfc == 2031692608269
*****[30F4]***** 1490992576323403 ** tdiff **-2587, 14909925763234022, 14909925763236609
*****[30F4]***** before 14909925763238690, 14909925763238690
*****[30F4]***** after  14909925763238689, 14909925763238690, diff: -1 / 0
*****[30F4]***** 1490992576324574, [10] ** cfreq == 3117182, nfreq == 3117441, perfc == 2031724216370
*****[30F4]***** 1490992586463519 ** tdiff **-786, 14909925864635173, 14909925864635959
*****[30F4]***** 1490992586464045, [10] ** cfreq == 3117256, nfreq == 3117441, perfc == 2031755824954
```

Legend:
  tdiff  : current deviation from virtual time (vt0) to the file-time (in 100-ns);
  before : nt0, nt0 -  virtual time before calibration;
  after  : nt1, nt1c - virtual time after calibration and correction 
           (nt1c = vt0 + (nt0-nt1));
  diff   : -26 / 0 - difference without and with correction;
  [10]   : current calibration interval (10 seconds)
  cfreq  : calculated frequency;
  nfreq  : nominal frequency;
  perfc  : last performance counter;