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]

Re: powerpc & unaligned block moves with fp registers


>>>>> Dale Johannesen writes:

Dale> Yes.  I'd recommend using 32 rather than 64 as the alignment cutoff
Dale> in DJ Delorie's patch; a 32-bit-aligned lfd/stfd does not trap.[*]

Dale> [*] at least, on the 750/7400/7450; I haven't gone back farther.

	Not even crossing a page boundary?

	I still agree that 32-bit alignment should be the cut-off.  Page
crossing exceptions should be rare.

Dale> It is true that multiple int load/stores will have the problems you say
Dale> (particularly if not perfectly aligned), but it's hard to believe those 
Dale> hits
Dale> will be larger than taking the exception.  It's true we haven't measured 
Dale> it.

	IBM has measured this when tuning its compilers for its latest
processors.

David


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