Self-modifying code (really Bli

[mcdonald at uxe.cso.uiuc.edu]

>This is indeed part of the answer. The other part is that
>instruction-fetch behavior is much simpler and more predictable
>than operand fetch, and that I-caches don't need to be
>written-through, allowing an isolated I-cache to be simpler
>and more effective than a combined operand/instruction cache.

This statement, unfortunately, connects up with the recurrent self-
modifying-code theme. SEPARATE I and D caches can result in incorrect
program execution if entries in the I cache are not voided if the
D cache gets written at the same location. Fixing this requires either
hardware to do this, or a "flush-cache" instruction. I strongly
prefer that hardware do this sort of housekeeping. It is OK if
there is a "dead period" of a reasonable number of cycles or
instructions during which the I cache is wrong. Some machines,
however, seem to be able to get this down to zero.

Doug McDonald