Author: Eugene Nalimov
Date: 21:05:12 06/16/00
Go up one level in this thread
On both Alpha and IA-64 it's better to use 64-bit index when indexing the array.
Explanation is simple -- all 64 bits in the resulting pointer would be used,
even on 32-bit NT that is supported on Alpha.
Simple examples on IA-64 (indexing of the 32-bit array, nops not included):
(1) int foo (int i, int *a)
{
return a[i];
}
Resulting assembly is
shladd r31=r32, 2, r33 ;;
ld4 r8=[r31]
br.ret b0 ;;
Run time is 2 clock cycles.
(2) int foo (__int64 i, int *a)
{
return a[i];
}
Resulting assembly is
sxt4 r32=r32 ;;
shladd r31=r32, 2, r33 ;;
ld4 r8=[r31]
br.ret b0 ;;
Run time is 3 clock cycles.
(actual output of the 64-bit MSVC compiler slightly differs; it does sign
extension in more complicated way to be more 32-bit compatible).
Eugene
On June 16, 2000 at 23:38:10, Bruce Moreland wrote:
>If you are talking about the alpha, it does 32-bits with no performance penalty,
>right? So the fact that an int is 32-bits on the Windows Alpha compiler
>shouldn't be a big deal. And they do provide good access to the 64-bit data
>type.
>
>Not that Alpha is anything these days.
>
>bruce
>
>On June 15, 2000 at 19:10:32, Eugene Nalimov wrote:
>
>>That's not true for some systems. Int is still 32 bits, but "natural" word size
>>is 64 bits.
>>
>>Eugene
>>
>>On June 15, 2000 at 17:28:26, Tom Kerrigan wrote:
>>
>>>On June 15, 2000 at 06:15:55, Christophe Theron wrote:
>>>
>>>>On June 14, 2000 at 17:29:07, Dave Gomboc wrote:
>>>>
>>>>>On June 14, 2000 at 16:17:25, Christophe Theron wrote:
>>>>>
>>>>>>The availability of 64 bits processors changes nothing. Unless some 64 bits
>>>>>>processors are so lousy that 8, 16 and 32 bits operations become slower than 64
>>>>>>bits ops...! :)
>>>>>
>>>>>I don't think it is unusual for certain operations on sizes smaller than the
>>>>>processor word size to take longer than they would if they used the processor's
>>>>>word size. Indeed, it wouldn't even be unusual for it to be possible without
>>>>>first sign-extending or zero-extending from the smaller size to the processor
>>>>>word size. AFAIK, 80x86 is a bit freaky in that it tries very hard to support
>>>>>8-bit and 16-bit operations in registers as quickly as 32-bit operations.
>>>>>
>>>>>Dave
>>>>
>>>>
>>>>That would be a very unfair way for bitboards to win the contest! :)
>>>>
>>>>
>>>> Christophe
>>>
>>>Just define everything in your program as an int. It will automatically use the
>>>processor's word size, so problem solved. :)
>>>-Tom
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.