home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!ferkel.ucsb.edu!taco!gatech!swrinde!zaphod.mps.ohio-state.edu!cs.utexas.edu!qt.cs.utexas.edu!yale.edu!ira.uka.de!math.fu-berlin.de!unidui!rrz.uni-koeln.de!Germany.EU.net!mcsun!sunic!dkuug!diku!thorinn
- From: thorinn@diku.dk (Lars Henrik Mathiesen)
- Newsgroups: comp.arch
- Subject: PA-RISC ``semantic loading'' (according to DEC)
- Message-ID: <1992Nov13.173228.16970@odin.diku.dk>
- Date: 13 Nov 92 17:32:28 GMT
- References: <1992Nov10.153629.27510@ryn.mro4.dec.com> <BxIM38.L9F.2@cs.cmu.edu> <15445@auspex-gw.auspex.com> <BxL3IH.KtH.2@cs.cmu.edu>
- Sender: thorinn@tyr.diku.dk
- Organization: Department of Computer Science, U of Copenhagen
- Lines: 38
-
- lindsay+@cs.cmu.edu (Donald Lindsay) writes:
- >DEC's architects also have a low opinion of HP's high semantic
- >loading. That is, HP has instructions that also do some second thing -
- >such as skip-on-condition. The Alpha team avoided that, and have
- >handwaving arguments about how this will help in future "aggressive"
- >implementations.
-
- I recently looked at the PA-RISC architecture manual, and, well ...
- it has a number of features that seem to make it easy and pleasant to
- generate compact code for, but they probably make for some pain as
- well when they have to be implemented in a deeply pipelined,
- more-than-twice superscalar chip.
-
- 0) The programmer model contains a processor status register, with
- rapidly changing contents; it is only visible on interrupts, and
- with a special instruction, but it still has to be synchronized.
-
- 1) ``Skip-on-condition'' sets a nullification bit in this PSR;
- nullification depends on the ALU results of one instruction, and
- prevents all programmer-visible effects of the next --- including
- data cache move-in, as far as I can see.
-
- 2) The instructions that support multiprecision and BCD arithmetic
- have carry bits in the status register as implicit arguments.
-
- 3) Changes to the virtual memory mapping of the instruction stream
- are guaranteed to take effect within 8 instructions.
-
- Nothing that cannot be fixed with, e.g., a file of ``post-instruction
- states'' with attending scoreboarding and bypass tricks, but perhaps
- the transistors could have been used to decrease cycle time instead.
-
- On the other hand, right now PA-RISC seems to get numbers for SPEC/MHz
- that are quite a bit better than Alpha. While some of the difference
- may depend on pipeline and cache effects, and compilers, I suspect
- that PA-RISC can get by with executing fewer instructions as well.
-
- Lars Mathiesen (U of Copenhagen CS Dep) <thorinn@diku.dk> (Humour NOT marked)
-