llvm dev - Dec 2014 - [LLVMdev] Memset/memcpy: user control of loop-idiom recognizer

On Sat, Dec 06, 2014 at 07:06:31AM -0600, Hal Finkel
wrote:
>  - Direction (should the memory be traversed forward or backward)
I don't think that this makes sense for memset and memcpy. It does
matter for memmove.

Joerg

On Sun, Dec 7, 2014 at 10:48 AM, Joerg Sonnenberger <joerg at
britannica.bec.de
> wrote:
> On Sat, Dec 06, 2014 at 07:06:31AM -0600, Hal Finkel wrote:
> >  - Direction (should the memory be traversed forward or backward)
>
> I don't think that this makes sense for memset and memcpy. It does
> matter for memmove.
>
This "memcpy" is being generated by pattern-matching an explicit loop
in
the code. Direction is certainly something that needs to be kept the same
as the original loop, if there is overlap.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20141207/d257022e/attachment.html>

On Sun, Dec 07, 2014 at 11:13:51AM +1300, Bruce Hoult
wrote:
> On Sun, Dec 7, 2014 at 10:48 AM, Joerg Sonnenberger <joerg at
britannica.bec.de
> > wrote:
> 
> > On Sat, Dec 06, 2014 at 07:06:31AM -0600, Hal Finkel wrote:
> > >  - Direction (should the memory be traversed forward or backward)
> >
> > I don't think that this makes sense for memset and memcpy. It does
> > matter for memmove.
> >
> 
> This "memcpy" is being generated by pattern-matching an explicit
loop in
> the code. Direction is certainly something that needs to be kept the same
> as the original loop, if there is overlap.
The memcpy intrinsic must not be generated in case of overlap.

Joerg

----- Original Message -----
> From: "Joerg Sonnenberger" <joerg at britannica.bec.de>
> To: llvmdev at cs.uiuc.edu
> Sent: Saturday, December 6, 2014 3:48:32 PM
> Subject: Re: [LLVMdev] Memset/memcpy: user control of loop-idiom recognizer
> 
> On Sat, Dec 06, 2014 at 07:06:31AM -0600, Hal Finkel wrote:
> >  - Direction (should the memory be traversed forward or backward)
> 
> I don't think that this makes sense for memset and memcpy. It does
> matter for memmove.
I was trying to list things that might affect performance. Even for memcpy, the
direction can matter for performance, depending on how the hardware prefetcher
and other parts of the memory subsystem work.

 -Hal
> 
> Joerg
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> 
-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory