Author: Ed Schröder
Date: 10:34:28 10/10/01
Go up one level in this thread
On October 10, 2001 at 12:39:01, Ulrich Tuerke wrote:
>On October 10, 2001 at 12:09:15, Bruce Moreland wrote:
>
>>On October 10, 2001 at 08:38:43, Travers Waker wrote:
>>
>>>Hi.
>>>
>>>I recently read that using word-sized variables for a particular machine (e.g. a
>>>32-bit int on an Intel PC CPU) yields faster code than using variables that are
>>>not the size of a word.
>>>
>>>For example, I understood the article to be saying that if you needed an index,
>>>i, to access an array, A[256], then it would yield faster code to define:
>>>
>>>int i;
>>>
>>>instead of:
>>>
>>>unsigned char i;
>>>
>>>even though the unsigend char is big enough to be able to index all the elements
>>>of the array. I find this strange, especially since it seems to me that you
>>>waste space in the memory caches when the most sinificant 3 bytes of the 32-bit
>>>int are always going to be zero's.
>>>
>>>Can someone explain to me why using word-sized variables should be faster?
>>
>>If you have a function that is diddling with a 256-element array, and the index
>>is a local variable, using an int has little to do with cache. If the function
>>is simple, the value is going to go into a register. If it's more complicated,
>>and the value can't fit in a register, there are probably more important
>>performance issues elsewhere.
>>
>>Even if it made some sense for the compiler to try to use a byte for the index,
>>at some point the compiler is going to have to stick the byte into a 32-bit
>>register. There is no instruction that says:
>>
>> mov eax, [bl]
>
>For the 386 architectures, there is the movzx ("move zero extended") which could
>do this job. But I guess, that the "movzx" is slower than the usual "mov"
>instruction ?
>
>Uli
The "movsx" / "movzx" is a very time consuming instruction in comparison
to the normal "mov" up to a factor of 3. Also if memory serves me well
"movsx" / "movzx" can't be paired which is another disadvantage.
Rule of thumb: whenever you have to "index" always use "int" instead of
"char". This will avoid the "movsx" / "movzx" use.
Ed
>>Instead, you get:
>>
>> mov eax, [ebx]
>>
>>The difference between the two is that the first one tries to use the lower 8
>>bits of ebx, and the second one uses the whole thing. The first one doesn't
>>exist.
>>
>>So in order to get to the second one, the compiler has to make sure that the top
>>bits are clear on every iteration of the loop.
>>
>>It's likely in this case that the compiler could avoid adding extra code in each
>>iteration in order to having something in the top bits, but it's not like
>>there's any up-side, so why make it worry about it?
>>
>>bruce
>>
>>>
>>>Thanks
>>>
>>>Travers
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.