RE: code optomization:any way to do this better?

  • From: "Sina Bahram" <sbahram@xxxxxxxxx>
  • To: <programmingblind@xxxxxxxxxxxxx>
  • Date: Mon, 17 Jan 2011 11:05:08 -0500

The gain in this is due to the way the jump predictor does optimizations.

There are actually a lot of useless jumps in this code, and they will get taken 
unfortunately: optimization or not. There is also
code repeat.

Also, Ken, unless if tyler runs an optimizer on here, it will not get optimized 
out because the operations he is using change the
state of flags and so are considered to be not atomic.

For example, the code will actually be measurably faster, in a loop of course, 
if he removes the 4 or 5 jumps I talked about in my
last email.

There is some talk about whether intel employs macrocode optimization, but 
that's more of a cache oriented process than anything
else, from my memory on the stuff.

Also, there's a lot of byte dereferencing here, and so one possibility is to 
simply combine some of these operations into words.

Furthermore, I haven't taken a close enough look, but I think I can remove 
every single jump in this code, simply by translating the
dereferenced bytes and using ebx as a pointer into the data segment via the use 
of the xlat op code.

Take care,
Sina

-----Original Message-----
From: programmingblind-bounce@xxxxxxxxxxxxx 
[mailto:programmingblind-bounce@xxxxxxxxxxxxx] On Behalf Of Ken Perry
Sent: Sunday, January 16, 2011 10:53 PM
To: programmingblind@xxxxxxxxxxxxx
Subject: RE: code optomization:any way to do this better?

I don't think there is a really good way to speed this up.  The fact is most
assemblers will turn this into some really optimized machine code and with
chips now days they can do multiple look ahead if statements so that this
would almost happen in one pulse of a chips timing. I think Sina might be
able to talk on the speed of this better than I but I can tell you the code
is not bad.  You have to understand  Asm is different than most languages
because its normal to have some repetitive code.  You could make a block
that tests for null and returns you to a passed in location so that way
every time you wanted to test for null you jump to your test block and then
if its null jump to end or jump back to the location in the code where you
wanted to be but I don't see any gain in doing that.

Ken


-----Original Message-----
From: programmingblind-bounce@xxxxxxxxxxxxx
[mailto:programmingblind-bounce@xxxxxxxxxxxxx] On Behalf Of Littlefield,
Tyler
Sent: Sunday, January 16, 2011 9:56 PM
To: programmingblind@xxxxxxxxxxxxx
Subject: code optomization:any way to do this better?

So I've been playing with assembly a lot lately, and was curious if 
there was a better way to do this. most importantly, the whole three 
branched if check (null, not null).
section .text
global _strcmp
_strcmp:
enter 0,0
;we copy our arguments to EBX and ECX
mov EBX, [EBP+8]
mov ECX, [EBP+12]
.loop:
;we need one value in a register
mov EDX, [ECX]
;check for null termination
cmp byte [EBX], 0
je .null
jne .notnull
;we have a null termination.
;if the other string is null terminated, we jump to success. otherwise 
it fails because they obviously aren't equal.
.null:
cmp byte [ECX], 0
je .success
jne .fail
;byte wasn't null, now we check for null on the other byte.
;if one is null, it's a fail because again they aren't equal. If it is 
not null, we do another check.
.notnull:
cmp byte [ECX], 0
;not equal, we check for equalness between the two now.
jne .check
je .fail
;we check for equalness between the two bytes here.
.check:
cmp [EBX], EDX
je .next
jne .fail
;here we increase pointers and jump back up to the top of the loop.
.next:
inc EBX
inc ECX
jmp .loop
;strings compared fully
.success:
mov EAX,1
jmp .finish
;strings did not compare fully.
.fail:
mov EAX, 0
;code cleanup.
;no need for a jmp, it just falls through.
.finish:
leave
ret

-- 

Thanks,
Ty

__________
View the list's information and change your settings at 
//www.freelists.org/list/programmingblind

__________
View the list's information and change your settings at 
//www.freelists.org/list/programmingblind

__________
View the list's information and change your settings at 
//www.freelists.org/list/programmingblind

Other related posts:

  1. Home
  2. programmingblind
  3. Archive
  4. 01-2011