[Info-vax] Poulson info from Dave Cantor

[Michael Kraemer]

glen herrmannsfeldt schrieb:

> Well, for RISC, and especially for VLIW, the problem isn't 
> allowing it to run, but to be able to extend the architecture
> and allow older programs to run reasonably fast.
> 
> Expecting the compiler to optimize for a specific processor tends
> to cause it to be much less optimal for other similar processors.

This seems to be much more of a problem for that EPIC contraption
than for traditional RISC. Newer versions of the same RISC
chip will execute old code considerably faster,
without any modification, that's what the vendors
claim and that's my experience. Of course one can "tune"
for a specific CPU version, but then one may loose compatibility
with future chip versions. It's up to the developer to decide
whether the gain in speed is worth it.

> For VLIW, a new implementation might have completely different
> timing requirements, but is still forced to allow for the old ones.

one reason one might consider VLIW to be a dead end.

> There is one solution, though rarely used, and that is to have
> the compilers output intermediate code, before the final bundling.
> Then, either at program load time or install time, run the bundling
> program for the specific processor available.  It will then generate
> the optimal code for that specific version.

Isn't that what RISC chips do automagically via instruction
scheduling? If the newer CPU has more functional units
it could throw more instructions at them.