This is the mail archive of the 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: [patch] Fix pr14796

On Thu, Jun 09, 2005 at 03:29:15PM +0100, Paul Brook wrote:
> On Thursday 09 June 2005 14:31, James A. Morrison wrote:
> >  Hi,
> >
> >   This patch adds some transformations for shifts and rotates such as
> > (a << c1) << c2 to a << (c1+c2).  This patch has been bootstrapped and
> > regtested on ia64-linux with no new regressions.  Ok for mainline?
> +      /* Transform (x >> c1) << c2...  */
> +	    /*  ...into (x >> (c1 - c2)) & (-1<<c2).  */
> +	    /*  ...into (x & (-1<<c1)) << (c2 - c1).  */
> Is this a good idea?
> There are certainly some targets (eg. Arm) where shifts tend to be cheaper 
> than masks because the mask can require a constant load, and the shift can be 
> combined with subsequent instructions.

Indeed.  Small immediate masks and single-cycle barrel shifters is a
feature common to many risc machines.  Now, it is something that should
be considered somewhere, because Intel borked the shifter on P4, and
the mask+shift is going to be faster there than two shifts.  But you
can't get at that knowledge at the fold-const level.

The only one that's likely to be a win universally is "x >> c << c"
being transformed to an AND.  The rtl expanders could convert that 
back to shifts immedately if that were the most efficient way to avoid
loading a large constant for the mask.

According to comment 2 in this pr, "x >> c1 >> c2" is already handled.
Is this statement false?  If not, why are you handing it again?


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