Verilator

Roughly yes. It tries to look for many parallel tasks (many per thread). It can't find enough, but you may still if lucky get parallelism, even 2x if you have two perfect large mtasks.

If you run with --stats we'd be interested in the output.

Well, yep looks like the parallelism factor is pretty low on this design.

Global Statistics:

  Assertions, PSL asserts                                  0
  Assertions, SystemVerilog asserts                        0
  Assertions, cover statements                             0
  Assertions, full/parallel case                           0
  MTask graph, final, critical path cost                9203
  MTask graph, final, edge count                          30
  MTask graph, final, mtask count                         21
  MTask graph, final, parallelism factor                   1

What's the instruction count total last column? Guessing this is a fairly small design? Small designs are unlikely to parallelize well, sorry.

I'm having problems disabling this warning. I've tried lint_off / lint_on as recommended in the warning message around the module, and then in a configuration file selecting a broader and broader range of files, and finally calling verilator with -Wno-UNOPTTHREADS all to no avail.

It isn't associated with an exact file so the inline pragma won't work. However we have several self tests that check -Wno-UNOPTTHREADS works, are you sure that flag is making it through to Verilator?

BTW -Wno-fatal might help in the meantime.

Stage Statistics:
  Stat                                          Link       PreOrder   Scoped     Final      Final_Fast
  --------                                      -------    -------    -------    -------    -------  

  Branch prediction,                               2195        146        178        139         61
  Branch prediction, VL_LIKELY                                  54         54         54         24
  Branch prediction, VL_UNLIKELY                                           29         59

  Instruction count, TOTAL                       172196     238398     229845     239864       6588
  Instruction count, fast critical                    0      40792      38746      43692       4738

And my problem was like many designs, our build process is complicated, and the -WNO-UNOPTTHREADS was getting passed to the wrong stage of the process.

That's small, about 5K instructions in the main loop. This suggests you might be getting around 1M cycles/second (might be quite slower). Parallelism generally needs something about 10 times or more larger, for example one design that parallelizes well is 36 million instructions giving ~1700 cycles/second unthreaded, then ~10K cycles/second threaded.