home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!noc.near.net!nic.umass.edu!dime!dime.cs.umass.edu!moss
- From: moss@cs.umass.edu (Eliot Moss)
- Newsgroups: comp.arch
- Subject: Re: SuperSPARC doesn't need superscalar compiler?!
- Message-ID: <MOSS.92Jul21091941@ibis.cs.umass.edu>
- Date: 21 Jul 92 13:19:41 GMT
- References: <9207201142.AA04083@x1sun6.ccl.itri.org.tw>
- Sender: news@dime.cs.umass.edu
- Reply-To: moss@cs.umass.edu
- Organization: Dept of Comp and Info Sci, Univ of Mass (Amherst)
- Lines: 38
- In-reply-to: lycmit@X1SUN6.CCL.ITRI.ORG.TW's message of 21 Jul 92 01:42:07 GMT
-
- >>>>> On 21 Jul 92 01:42:07 GMT, lycmit@X1SUN6.CCL.ITRI.ORG.TW (Yin-Chih Lin)
- >>>>> said:
-
- Yin-Chih> ... one of the Sun Micro employees had adverted that the SuperSPARC
- Yin-Chih> (superscalar SPARC which can issue 3 intructions at max performance)
- Yin-Chih> does not require the superscalar compiler to obtain the best CPU
- Yin-Chih> performance.
-
- It could be that compilation principles followed for other models of SPARC
- will indeed yield good results for this implementation, too. For example,
- there is the load delay of two cycles, which would cause instruction
- schedulers to space loads and succeeding uses at least three instructions
- apart, so they would fall into separate windows for issue on the SuperSPARC.
- That is, a non-superscalar might require spacing things out in time,
- especially for non-fully-pipelined functional units. This spacing will tend to
- help on a superscalar with a limited number of functional units. It's a
- transformation between time and space, but there are significant similarities.
-
- Of course, you are right that in general a good schedule for a non-superscalar
- may not be the best schedule for a superscalar that does not have enough
- resources available on each cycle. For example, a non-superscalar might be
- able to issue one register to register multiply every cycle, given one fully
- pipelined multiplier. A superscalar that issues N instructions per cycle would
- need N fully pipeined units to guarantee not to stall, but most code probbaly
- does not need N fully pipelined multipliers.
-
- However, the designers of the Motorola 88110 did do some balancing analysis
- and decided to replicate some functional units. It's a dual issue machine, and
- some kinds of instructions are popular enough that it pays to be able to issue
- two in the same cycle, which requires two similar functional units.
- --
-
- J. Eliot B. Moss, Associate Professor
- Department of Computer Science
- Lederle Graduate Research Center
- University of Massachusetts
- Amherst, MA 01003
- (413) 545-4206, 545-1249 (fax); Moss@cs.umass.edu
-