home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.arch:8877 comp.benchmarks:1299
- Newsgroups: comp.arch,comp.benchmarks
- Path: sparky!uunet!infonode!ingr!capalo!quintus!quintus!ludemann
- From: ludemann@quintus.com (Peter Ludemann)
- Subject: Re: MIPS ratings go up; faith in Dhrystone persists.
- Message-ID: <1992Aug12.231534.5843@quintus.com>
- Sender: news@quintus.com (USENET news account)
- Nntp-Posting-Host: ebisu
- Organization: Quintus Corporation, Palo Alto, CA
- References: <1992Aug10.165958.16152@sinix.UUCP>
- Date: Wed, 12 Aug 1992 23:15:34 GMT
- Lines: 20
-
- In article <1992Aug10.165958.16152@sinix.UUCP>, weicker@sinix.UUCP (Reinhold Weicker) writes:
- >
- > Also, I must add that Dhrystone version 2.1 is not fool-proof either:
- > If a compiler compiles both compilation units together (not intended
- > for Dhrystone, but some compilers have such a switch), and does
- > excessive constant propagation (between procedures), it can determine
- > the whole flow of computation at compile time, and pre-compute the
- > output.
-
- Indeed; I was caught by this when I tried comparing the gcc v2 compiler
- versus IBM's xlc on the RS/6000. At first, it seemed that gcc was faster,
- but then it turned out that IBM's compiler was optimizing away the dummy
- timing loops, so the correction factor for loop overhead was coming out
- wrong. I have heard of "hardened Dhrystones" to avoid such problems, but
- I suspect that each new optimizing compiler will require re-hardening.
- If we could somehow go through all the ftp sites and blow away all old
- Dhrystone sources, we might find a bit more truth in advertising ...
-
- --
- Peter Ludemann +1-415-813-3800 ludemann@quintus.com
-