On Thu, 21 Nov 2013, Andrew MacLeod wrote: > If we add the hook for atomic_align_for_mode, and change the initalizers in > tree.c, any target which doesnt need/use the hook should be unaffected. So > everything remains as it is today. > > So Putting the hook in shouldn't be an issue. Then you can experiment with it > on your port and see if you get the desired effect... > > I've attached what I think the remaining bits are regarding the alignment.
Right, that's about what I expected. Nice. Of course, I'd argue that a better default for the atomic alignment is the max of the default alignment and the natural alignment, since *you can't have atomic support without extra machinery for misaligned data* - straddling a page or cache boundary has dire consequences. But then the patch would not have the nice property of "obviously" being a NOP elsewhere. Then again, where it isn't a NOP, you have breakage anyway. Bah. > All untested since I wrote it of course :-) In any case, you should just > need to provide a function to override the alignment for whatever mode(s) you > need with this... > > I can bootstrap and check this on x86 to make sure it doesnt affect anything, > and you can fool with it and see if you can get your desired results with your > port. Thanks! I'll play with it and get back. brgds, H-P
