drcachesim offline simulation does not properly combine multiple processes

[derekbruening]

It looks like we never made a test for a multi-process offline simulation. Looking at the code, the files are closed on fork, but there's no fork init event to open a new one, so I would expect a failure on writing out the child's trace.

[derekbruening]

Pasting from my notes:

*** TODO need separate subdir per process b/c of modules.log, so painful to add test

Should allow multiple -indir options. But that brings up complications:

• Do we map the binaries from multiple modules.log simultaneously?
  Going to hit scalability issues: maybe it's better to instead support
  merging two final traces?

• The droption accum separates with spaces: so what about paths with spaces
  in them?

For now I just have a test of the output from one process.

[derekbruening]

#3059 covers adding non-space separator support to droption

[derekbruening]

To summarize the current status, the generation of offline files works fine (they go to separate dirs, courtesy of the commits referenced above), but trying to merge them back together to pass to the simulator for a single multi-process simulation is not yet supported.

[derekbruening]

Note that the scheduler supports mixing multiple processes (as separate
"workloads" if different scheduling policies are desired), and we do
recommend dynamic rescheduling (see
https://dynamorio.org/sec_drcachesim_sched.html. We will be trying to make
that the default rather than the current poor static schedule default).
For analyzers, just combining the files into one directory would be a
workaround that ignores addressing differences.

Status today of multiple address space support:

• If you collect physical addresses during tracing (-use_physical, which
  requires privileges and adds overhead) and use those during simulation
  (same -use_physical) then mixing multiple processes is all handled
  nicely.

• If you don't -use_physical during tracing then:

  • If you pass -use_physical it will convert virtual to physical by
    dropping top bits but this can lead to address collisions between
    processes when there was no collision in reality.

  • You could create a "v2p" textproto and add dr$sim support akin to the
    TLB simulator support added in PR i#6662 public traces, part 9: TLB simulator with v2p #6961.

We might want to add some flag to shift addresses (but at a point where the
v2p will still differentiate them), or provide the tool to create the v2p
textprotos.