Board index » delphi » Re: Fastcode Voting - Rep Ret Rule 3

Avatar Zondertau

Delphi Developer
 

Re: Fastcode Voting - Rep Ret Rule 3


2005-10-04 11:26:27 PM
delphi166
Quote
First of all I'd like to apologize for being so bad a managing
this voting.

We need to get more votes in because it stands 3 to 3 between 2) and
6). 1) Do not allow Rep Ret on any targets
ISTM Erics vote is closer to (2) than to (6), which would solve the
problem.
If you wish i could also vote (2) a second time; that would solve it as
well :)
 
 
Eric Grange

Delphi Developer

2005-10-04 11:47:07 PM

Re: Fastcode Voting - Rep Ret Rule 3

Quote
ISTM Erics vote is closer to (2) than to (6), which would solve the
problem.
Well, I am not 100% sure we can trust AMD to "rep ret" correctly forever
on future CPUs (if only because Intel might decide to give the
instruction a meaning, and AMD would have to follow suit), and since
it's in the grand scheme of things a rather negligible speedup, consider
it a CPU-specific artefact not worth the risk bothering with.
For instance does AMD Geode support it? (Geode is based on Cyrix designs)
It's that whole "risk the gold ingots we already have in bank for a few
more dollar bills" business I am not sure with.
Eric
 
Avatar Zondertau

Delphi Developer

2005-10-05 12:31:51 AM

Re: Fastcode Voting - Rep Ret Rule 3

Quote
>ISTM Erics vote is closer to (2) than to (6), which would solve the
 >problem.
Well, I am not 100% sure we can trust AMD to "rep ret" correctly
forever on future CPUs (if only because Intel might decide to give
the instruction a meaning, and AMD would have to follow suit), and
since it is in the grand scheme of things a rather negligible speedup,
consider it a CPU-specific artefact not worth the risk bothering
with. For instance does AMD Geode support it? (Geode is based on
Cyrix designs)

It's that whole "risk the gold ingots we already have in bank for a
few more dollar bills" business I am not sure with.
Yes, i agree with that.
This shows a problem with the current way we're polling BTW. Currently
4 out of 7 this REP RET is not worth the risk. Their only difference in
opinion is if they want to completely ban it or just restrict it to AMD.
The remaining 3 on the other hand think that it may be worth taking an
incompatibility risk (however small) to add a slight speed increase on
AMD targets.
The fact that people disagree on details now makes their absolute
majority together into a first position shared with their opponents.
This situation can be avoided if we introduce a round-based voting
system like this:
1) Initially everyone can vote for all options
2) If there is an absolute majority the poll has been decided
3) The option with least votes is removed. People who chose it should
change their vote
4) Goto 2
 

Anders Isaksson

Delphi Developer

2005-10-05 01:43:09 AM

Re: Fastcode Voting - Rep Ret Rule 3

Eric Grange writes:
Quote

Well, I am not 100% sure we can trust AMD to "rep ret" correctly
forever on future CPUs (if only because Intel might decide to give the
instruction a meaning, and AMD would have to follow suit),
After reading that, it is no problem for me to vote 1)
--
Anders Isaksson, Sweden
BlockCAD: web.telia.com/~u16122508/proglego.htm
Gallery: web.telia.com/~u16122508/gallery/index.htm
 
Avatar Zondertau

Delphi Developer

2005-10-05 01:56:41 AM

Re: Fastcode Voting - Rep Ret Rule 3

Quote
>Well, I am not 100% sure we can trust AMD to "rep ret" correctly
 >forever on future CPUs (if only because Intel might decide to give
 >the instruction a meaning, and AMD would have to follow suit),
After reading that, it is no problem for me to vote 1)
Unfortunately this doesn't give us a majority.
1) Do not allow Rep Ret on any targets
2 Votes: Anders, Eric
2) Allow Rep Ret on all AMD targets.
3 Votes: Avatar, Dennis, Hallvard
3) Allow Rep Ret in all targets except RTL replacement and Blended
4) Allow Rep Ret in all targets except RTL replacement
5) Allow Rep Ret on all targets.
6) Wait for Intel and AMD reply
3 Vote: Pierre, Will, Jan Derk
 
Dennis

Delphi Developer

2005-10-05 02:01:08 AM

Re: Fastcode Voting - Rep Ret Rule 3

Hi
1) Do not allow Rep Ret on any targets
2 Votes: Eric, Anders
2) Allow Rep Ret on all AMD targets.
3 Votes: Avatar, Dennis, Hallvard
3) Allow Rep Ret in all targets except RTL replacement and Blended
4) Allow Rep Ret in all targets except RTL replacement
5) Allow Rep Ret on all targets.
6) Wait for Intel and AMD reply
3 Vote: Pierre, Will, Jan Derk
Please update the list with your votes.
Best regards
Dennis