Re: Are 64 Int or FP registers useful?

hrubin@pop.stat.purdue.edu (Herman Rubin)
Thu, 17 Sep 1992 13:57:41 GMT

          From comp.compilers

Related articles
Re: Are 64 Int or FP registers useful? hrubin@pop.stat.purdue.edu (1992-09-14)
Re: Are 64 Int or FP registers useful? preston@helena.cs.rice.edu (1992-09-16)
Re: Are 64 Int or FP registers useful? hrubin@pop.stat.purdue.edu (1992-09-17)
Re: Are 64 Int or FP registers useful? daveg@synaptics.com (Dave Gillespie) (1992-09-17)
Re: Are 64 Int or FP registers useful? pardo@cs.washington.edu (1992-09-17)
Re: Are 64 Int or FP registers useful? lfm@pgroup.com (1992-09-19)
| List of all articles for this month |

Newsgroups: comp.arch,comp.lang.misc,comp.compilers
From: hrubin@pop.stat.purdue.edu (Herman Rubin)
Organization: Purdue University Statistics Department
Date: Thu, 17 Sep 1992 13:57:41 GMT
References: <1992Sep7.091904.2626@newsroom.bsc.no> 92-09-088
Keywords: registers, optimize, comment

preston@helena.cs.rice.edu (Preston Briggs) writes:
>Regarding 64 registers...


>I believe that 64 FP registers is probably past the point of diminishing
>returns for most application programs. On the other hand, there are
>important examples that could profitably use more than 32 registers.
>Further, I believe that as memory latency increases (in terms of FP
>speed), the number of useful registers will increase.


>For an important example, I'd cite dense matrix algebra packages. Of
>course, if your primary applications are integer based (say, you edit and
>compile all day), extra FP registers aren't going to help much.


For editing and compiling, I agree. But why should there even be separate
integer and FP registers? Part of this is because languages, and many
users, cannot see the advantage of mixing the operations. On many
machines, if a loop integer is used in floating operations, not that
uncommon, it may very well be cheaper to have both an integer and a
floating copy, which certainly goes against the idea of repeating
computations present in current language philosophy.


Also, on many machines at this time, the best way to do multiple precision
operations, which always involves integer arithmetic when it is needed
beyond what is essentially built in, the integer unit is so slow and
clumsy that the operations need to be done, albeit clumsily, in the
floating unit. Having combined registers helps, as one has to switch back
and forth. Good integer arithmetic is becoming more important, even for
such problems as cryptography.
--
Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907-1399
Phone: (317)494-6054, hrubin@pop.stat.purdue.edu (Internet, bitnet)
{purdue,pur-ee}!pop.stat!hrubin(UUCP)
[I note that the original Motorola M88K had a single combined register set,
while the second version adds FP registers. Hennesey and Patterson give
some reasons why separate register sets are often faster, e.g., simpler
register multiplexers and more opportunity for simultaneous execution. -John]
--


Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.