On Thu, Jul 14, 2011 at 11:32:59AM +0200, Richard Guenther wrote: > On Thu, Jul 14, 2011 at 11:31 AM, Richard Guenther > <richard.guent...@gmail.com> wrote: > > 2011/7/14 Michael Meissner <meiss...@linux.vnet.ibm.com>: > >> One minor note, you will need to update doc/invoke.texi to document the new > >> switch before checkin: -ftree-reassoc-width=<n> > > > > You also need approval and somebody to review the patch before checkin. > > Btw, rather than a new target hook and an option I suggest to use > a --param which default you can modify in the backend.
You need the target hook that tells how big the reassociation based on the type. Machines have different numbers of functional units, for example, maybe 3 integer units and 2 floating point units. For example, in the PowerPC, I would set the basic integer and binary floating point types to 2 to match the dispatch rules, but for decimal floating point, I would set it to one, since the machines currently only have 1 decimal unit. With the hook, I could see eliminating the switch and/or --param altogether, and doing only in the hook. -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
