From: monnier@iro.umontreal.ca   
      
   > One must also take into consideration the relative time of each path.   
   > For example, I have Transcendental instructions in my ISA where I can   
   > perform FP64×{SIN, COS, Ln2, 2**} in {19, 19, 14, 15} cycles respectively.   
   > I make a rounding error about once every 378 calculations. Running a SW   
   > routing that gives correct rounding is 300-500 cycles.   
   >   
   > So, if one can take the Incorrect-Rounding exception in less than 100   
   > cycles (round tip) trap-and-emulate are adding only 1-cycle on average   
   > to the running of the transcendentals.   
      
   Hmm... I think that calculation isn't quite right:   
      
   - First, this assumes a random distribution.  Depending on the details   
     (e.g. if the 1/378 cases are grouped in a specific subspace that some   
     algorithm might use heavily), it can get a lot worse (or a bit better).   
      
   - During those 300-500 cycles, no other instruction is processed, IOW   
     it's not as-if this one instruction took just one more cycle, but it's   
     as-if the CPU stopped all activity during one cycle.  For an ILP of 1,   
     it's about the same, but for the ILP of 4, it's 4x more expensive.   
      
   > I CAN prove it WILL come back to Bite--but only if your architecture   
   > survives   
      
   It's part of the problem with those "lessons from" successful projects:   
   don't forget that it is sometimes indispensable to cut corners for the   
   project to survive in the first place.  Maybe you'll regret it years   
   later, yet, without it, there wouldn't even be any opportunity to regret   
   anything years later.  There are delicate trade-offs.   
      
      
           Stefan   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)