Board index » delphi » Re: Voting for FastCode rules on memory reading

2005-05-10 01:23:28 AM
delphi37
This rule but can be important to string-related challenges, but AFAIK
doesn't have much effect on memory managers. These are not likely to
benefit from reading extra data.
Memory managers can safely ignore this rule, for if one byte in a
memory page can be read, all can be read. The MM can not change this.
 

reason to decline this rule, just some programmer's habits:
keep a program clean (in this case while executing).
Florent
 

Avatar Zondertau writes:
good reason), but I got the threads confused after that. Please
disregard.
--
Regards,
Bruce McGee
Glooscap Software
 

Avatar Zondertau writes:
--
Regards
Torjei Kvinen
 

In favor of changing rule: Aleksandr, Avatar, Bruce, Pierre,
Thorsten, Torjei
In favor of keeping rule: Dennis, Florent
Unsure (not clear from post): Anders
Unsure (depends on speed tests): Eric
 

On 9 May 2005 12:26:14 -0700, "Avatar Zondertau" <XXXX@XXXXX.COM>
writes:
My first reaction is: "No way, you shouldn't allow this! It definitely
has a 'bad smell'."
Next: "But if we allow reading the full DWORD containing the last byte,
why not the full page?"
So,
1) How sure can we be that it will 'always' be allowed to read a full
page - at least some (very much) older computers had permission bits for
every memory cell!
2) Can the CPU actually, physically, access anything smaller than a
DWORD? Or will the memory access be a full DWORD even when I ask for a
byte?
After this short discussion with myself, I will cast my vote:
No, I don't think the rule should be changed.
(But feel free to ignore my opinion, as it is not built on any solid
fact, it is just an opinion)
--
Anders Isaksson, Sweden
BlockCAD: web.telia.com/~u16122508/proglego.htm
Gallery: web.telia.com/~u16122508/gallery/index.htm
 

Hi Anders
You and I have the same opinion and understanding and "feeling" about this
;-)
Regards
Dennis
 

can be set per page and per segment. It also specifies that pages are 4
KB, 2 MB or 4 MB. These specifications must hold for IA32 CPUs.
Segments are specified by Windows to be used for creating a flat memory
space; this means they cannot be used by Windows for protection (at
least not DS, ES and SS). To change this would break all existing
Windows programs.
In favor of changing rule: Aleksandr, Avatar, Bruce, Pierre,
Thorsten, Torjei
In favor of keeping rule: Anders, Dennis, Florent
Unsure (depends on speed tests): Eric
 

Hi Avatar
Thanks for keeping on top of this voting round. It takes some pressure of my
head ;-)
64 byte L1 in P4
128 byte L2 in P4
In favor of changing rule:
6 - Aleksandr, Avatar, Bruce, Pierre, Thorsten, Torjei
In favor of keeping rule:
3 - Anders, Dennis, Florent
Regards
Dennis
 

I don't think there is much to gain in practice, but that could mean
trouble in corner cases, so unless I am wrong and there is a whole lot of
performance to gain, I would rather keep the strict rule.
Eric
 

Hi
homepage today.
Just one question.
Do you think that Borland would like functions in the RTL that are based on
this rule?
Regards
Dennis
 

If he doesn't we could keep the old rule for the RTL replacement
category.
 

Hi Avatar
Regards
Dennis
 

use the hardware, while not scaring away Borland. Afterall since the
RTL replacement category doesn't allow MMX the rule change wouldn't
affect RTL replacement anyway.
 

In practice yes, but in theory no.
The rule allows you to read anywhere on the page no matter which
instruction
you use?
Regards
Dennis