This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.

Index Nav:	[Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav:	[Date Prev] [Date Next]	[Thread Prev] [Thread Next]
Other format:	[Raw text]

Re: Useless vectorization of small loops

From: Richard Henderson <rth at redhat dot com>
To: Richard Guenther <rguenth at tat dot physik dot uni-tuebingen dot de>
Cc: gcc at gcc dot gnu dot org
Date: Tue, 22 Mar 2005 13:29:48 -0800
Subject: Re: Useless vectorization of small loops
References: <Pine.LNX.4.44.0503211340190.6707-100000@alwazn.tat.physik.uni-tuebingen.de>

On Mon, Mar 21, 2005 at 01:45:19PM +0100, Richard Guenther wrote:
> I also cannot
> see why we zero the mm registers before loading and why we
> load them high/low separated:

We load hi/lo separate because movlps+movhps is faster than movups.

We zero first to break the insn dependency chain before doing two
half-register modifies.  IIRC such chain breaking is only relevant
to the p4.

r~

References:
- Useless vectorization of small loops
  - From: Richard Guenther

Index Nav:	[Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav:	[Date Prev] [Date Next]	[Thread Prev] [Thread Next]