PA-8500 vs. PA-8000 - gcc support?

[Jerry Quinn]

>> "Jonathan" == Jonathan Couper <couper@g8labs.co.nz> writes:

 Jonathan> Hiya We're using a PA-8000 machine, but are expecting to port to
 Jonathan> PA-8500 later in the year.  I'm curious to know:

 Jonathan> 	- Does gcc have specific support for the 8500?

No.  We've added an optimization path for the 8000, but so far it's stuff that
should be generically applicable to any 8x00.  There's a few things.  There is
a machine model that attempts to capture retirement from the reorder buffers
and a bit of floating point latency.  There are patterns to generate
multiply/accumulate instructions.  We turn off the combined instructions such
as fmpyadd since those can tie up the reorder buffers.  And finally we turn
off the 1.1 processor-specific cost adjustments.

 Jonathan> 	- Is there any significant difference between the two?

According to public docs at HP, there are a few differences.  First, the cache
is on-chip, although I don't think that affects anything.  Second, it has
better branch prediction.  The 8000 has to choose between static and dynamic
prediction while the 8500 offers a combination of the two where dynamic
prediction can compensate for bad static predictions.  The 8500 (and 8200)
also has a larger branch history table.  The TLB is larger.  And of course the 
clock rates are higher.

I suspect that these changes probably don't mean any changes to the compiler
although someone who actually knows their stuff may counter that.

Jerry

-- 
Jerry Quinn                             Tel: (514) 761-8737
jquinn@nortelnetworks.com               Fax: (514) 761-8505
Speech Recognition Research