Index
> Main > Why is the behavior of the repXX prefixes undefined forGoto page Previous 1, 2 |
| Author |
|
|
revolution 27 Dec 2009, 00:26
So what about:
Code: rep addsd xmm0,[esi+ecx*8] |
|||
27 Dec 2009, 00:26 |
|
|
revolution 27 Dec 2009, 00:33
So you impose exceptions, rep is good here, but not there. Okay, that needs silicon to decide what is good and what is not.
But what is the benefit? You still have not shown any compelling example as to where it would be useful enough to put in the extra effort to design, test, build and document it. |
|||
|
27 Dec 2009, 00:33 |
|
|
Azu 27 Dec 2009, 00:35
revolution wrote: So you impose exceptions, rep is good here, but not there. Okay, that needs silicon to decide what is good and what is not. The silicon is already there to see if an instruction is SSE. Just don't check for the rep prefix until after checking for an SSE instruction. Or whatever the way is that they currently differentiate (with SSE instructions it doesn't work like a nop, right? So the silicon to check is already there) revolution wrote: But what is the benefit? You still have not shown any compelling example as to where it would be useful enough to put in the extra effort to design, test, build and document it. Azu wrote:
|
|||
|
27 Dec 2009, 00:35 |
|
|
Borsuc 27 Dec 2009, 00:57
Azu wrote:
_________________ Previously known as The_Grey_Beast |
|||
|
27 Dec 2009, 00:57 |
|
|
Azu 27 Dec 2009, 01:02
Examples: all instructions that can take ecx as an operand or increment or decrement anything or modify the zero flag.
|
|||
|
27 Dec 2009, 01:02 |
|
|
revolution 27 Dec 2009, 02:46
Where is it useful?
|
|||
|
27 Dec 2009, 02:46 |
|
|
Azu 27 Dec 2009, 02:51
I just said.
|
|||
|
27 Dec 2009, 02:51 |
|
|
revolution 27 Dec 2009, 02:57
No, where is it useful, in what applications. Can you give a practical example of it's use in an application?
[edit to add] If you really want this to be included in the future then you have to tell Intel/AMD where it has practical uses. They won't add it just because you ask them, you have to convince them. |
|||
|
27 Dec 2009, 02:57 |
|
|
bitRAKE 27 Dec 2009, 08:46
I'm just going to babble because I don't work for Intel/AMD:
Basically, the decoder had a branch for the REP prefix which occurs comparitively early. This would make it a preferred canidate for new instructions given only a cost of increased "setup" time for REP <string> instructions. A general REP might seem like a Good Thing™, but most loops are greater than a single operation. What would be the costs of continually re-initializing the loop counter prior to each operation? As soon as two loop operations have a dependancy across them then the REP instruction paradigm fails to work. |
|||
|
27 Dec 2009, 08:46 |
|
|
Borsuc 27 Dec 2009, 20:44
Azu wrote: Examples: all instructions that can take ecx as an operand or increment or decrement anything or modify the zero flag. Borsuc wrote: [I've] Never used loop on a single instruction... example?  _________________ Previously known as The_Grey_Beast |
|||
|
27 Dec 2009, 20:44 |
|
| Goto page Previous 1, 2 < Last Thread | Next Thread > |
Forum Rules:
|