Handling of THRESHOLDs

[Torbjorn Granlund]

We choose algorithm in GMP with constructs like

  if (BELOW_THRESHOLD (MUL_TOOM22_THRESHOLD))
    {
      mul_basecase (...);
    }
  else if (BELOW_THRESHOLD (MUL_TOOM33_THRESHOLD))
    {
      mul_toom22 (...);
    }
  else
    {
      mul_toom33 (...);
    }

If for some reason, mul_toom22 is never better than mul_basecase and
mul_toom33, we use a trick:

#if MUL_TOOM22_THRESHOLD >= MUL_TOOM33_THRESHOLD
#define MUL_TOOM22_THRESHOLD   MUL_TOOM33_THRESHOLD
#define MUL_TOOM33_THRESHOLD   0
#endif

When this happens, MUL_TOOM22_THRESHOLD is actually the threshold beteen
mul_basecase and mul_toom33...  A bit confusing.  (MUL_TOOM33_THRESHOLD
is made into a dummy threshold that allows the compiler to remove the
test and the call to mul_toom22.)

I am contemplating if we should rename these thresholds to make things
slightly more readable.  Now FOO_THRESHOLD-1 is the largest operands for
which algorithm foo is used.  The change would be to let FOO_THRESHOLD
be the smallest operands for which foo is used.

The code above would become

  if (BELOW_THRESHOLD (MUL_BASECASE_THRESHOLD))
    {
      mul_basecase (...);
    }
  else if (BELOW_THRESHOLD (MUL_TOOM22_THRESHOLD))
    {
      mul_toom22 (...);
    }
  else
    {
      mul_toom33 (...);
    }

When we want to switch off mul_toom22, it is MUL_TOOM22_THRESHOLD we
need to knock to 0.  I also think moving the threshold name closer to
the routine name makes things more readable.

Opinions?  Surely there is some drawback...?

-- 
Torbjörn