[llvm-commits] [patch] call libc memcpy/memset for big arrays

[Dale Johannesen]

On Aug 24, 2007, at 6:15 PM, Chris Lattner wrote:

>
> On Aug 24, 2007, at 4:44 AM, Rafael Espindola wrote:
>
>> Currently we expand a memcpy/memset node to a call to the libc
>> implementation if
>>
>>   if ((Align & 3) != 0 ||
>>       (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold 
>> ())) {
>>
>> Shouldn't this be a ">". The libc memcpy/memset are very fast for
>> big arrays.

Do you have data to support this?  It is dependent on the OS (well  
libc really) and target CPU for sure.

> Yep, that makes sense to me! Please commit.
>
> Dale, please review the commit when you get a chance,

Intuitively it looks OK.  I'd like to see numbers, but I doubt that  
spending a lot of time on this is productive.

I've spent way too much time playing with the heuristics for memcpy  
expansion.  My firmest conclusion is that, whatever the compiler  
does, somebody will supply an example where it is better to do  
something else.:)

The situation is especially bad on x86 since different things are  
optimal for different CPUs, and compiled code will typically run on  
more than one CPU.  Here are some numbers from the last time I looked  
at it:  http://gcc.gnu.org/ml/gcc-patches/2005-08/msg01614.html
On that CPU and OS (Darwin 10.4, IIRC), it was never right to use rep/ 
mov, and the library function didn't look too good compared to  
repeated lod/sto.   ymmv.

> -Chris
>
>> If I compile
>>
>> ---------------------------------------
>> #include <string.h>
>> void f1(long *a, long *b) {
>>   memcpy(a, b, 8 * 16);
>> }
>>
>> void f2(long *a, long *b) {
>>   memcpy(a, b, 8 * 32);
>> }
>> ---------------------------------------
>>
>> with gcc 4.2 (-O2), it will call memcpy for f2 and not for f1.
>>
>> Cheers,
>> -- 
>> Rafael Avila de Espindola
>>
>> Google Ireland Ltd.
>> Gordon House
>> Barrow Street
>> Dublin 4
>> Ireland
>>
>> Registered in Dublin, Ireland
>> Registration Number: 368047
>> <memcpy.patch>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits