home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.arch
- Path: sparky!uunet!utcsri!torn!cunews!nrcnet0!bnrgate!bmerh85!bcars64a!bqneh23!schow
- From: schow@bqneh23.bnr.ca (Stanley T.H. Chow)
- Subject: Re: trapping speculative ops (LONG)
- Message-ID: <1992Sep2.145533.5976@bcars64a.bnr.ca>
- Sender: news@bcars64a.bnr.ca (Usenet News)
- Organization: Bell Northern Research Ltd, Ottawa
- References: <1992Aug31.224611.5196@odin.diku.dk> <1992Sep1.152044.12253@tera.com>
- Date: Wed, 2 Sep 1992 14:55:33 GMT
- Lines: 35
-
- In article <1992Sep1.152044.12253@tera.com> bob@tera.com writes:
- >To quote William Kahan:
- >
- > Exceptions are not errors.
- > Unless they are handled badly.
- >
- >In other words, if you think you can roll over and die when you take an exception,
- >you are severely restricting hardware and software. In many cases, some fix up
- >is desired followed by a return from the exception. Continuing in this way can
- >be quite challenging if you have gone too far forward. Values that were around
- >may no longer be anywhere to be found. However, an architectural way to guarantee
- >that no traps from an operation will occur in the future is nice. DEC's Alpha
- >has the sledge-hammer-like trap barrier. Being able to selectively touch off
- >traps gives the compiler more control and may allow better pipeline usage.
-
- Isn't that the whole point of the trap-bit scheme?
-
- The compiler chooses when to trap, and the timing can be different for each
- potential trap! Instead of a sledge hammer, the compiler get to use the data-
- flow and register allocation information. As I see it, DEC Alpha forces a barrier
- per basic block since the trap handler doesn't know which instruction caused the
- trap and the multiple incoming path could all generate the same type of trap.
- In the trap-bit scheme, the compiler chooses which traps to handle and when to
- handle them. So if an operation written to R15 traps, the flow of control could
- meander through any number of basic blocks before the trap detection (as long as
- the compiler knows that R15 is not touched).
-
- In what way is fix up prevented?
-
-
- --
- Stanley Chow InterNet: schow@BNR.CA
- Bell Northern Research UUCP: ..!uunet!bnrgate!bqneh3!schow
- (613) 763-2831
- Me? Represent other people? Don't make them laugh so hard.
-