🚀 go-pugleaf

Reference: 10189,9851
 
Yes, the savings have to do with the load delay's from the data cache.
Another table in the paper shows the relative savings for a certain
allocation method with 52 registers and varying data cache latencies:
 
Program		Data cache speed in cycles
 
		 1	 2	 3	 4	 5
------------------------------------------------------------------------
Simulator	12 %	19 %	24 %	27 %	29 %
Verifier	10 %	15 %	19 %	21 %	23 %
 
So there is an increasing potential for savings with longer data
cache latencies, but a faster cache will help you always on those
nasty array references (combined with block refills), have a look
at the R2000/R3000 speed/Mhz numbers for these effects. And given
the availability of ridiciously fast static RAMS (I have a Cypress
Semiconductor catalogue in front of me, CY100E474 BiCMOS RAM, 4k,
3 nS...) there doesn't seem a good reason to have at least the first
level cache be single cycle, even if you go to > 100 Mhz clocks.