On 04/14/2011 12:18 PM, Patrick Rutkowski wrote: > > On Apr 14, 2011, at 5:36 AM, Andrew Haley wrote: >> >> I can't quite tell what you're asking. C is designed so that you, >> never have to know the alignment requirements: program in a >> straightforward way and you'll be fine. Is it that you don't know >> what the rules are for writing C? Or that you know the rules, but >> for some special reason you want to violate them in non-portable >> code? > > Ah, yes, a well put question. > > As to the first part, I think that I'm now starting to understand why > alignment issues exist. If I understand correctly, it's because > compilers often choose to use aligned CPU op codes (though I can't say > I know exactly why aligned op codes run faster in hardware). And I > think I know how these alignment issues are dealt with, which > would be via malloc() aligning and struct padding. Right. Some reasons for slowness are obvious: if your bus is 32 bits wide and you have an unaligned 32-bit fetch then you have to make two accesses. It's not so bad on the x86 chips most of us have in our desktop machines, which have a bunch of special hardware to make unaligned accesses fast in most cases. However, sometimes an unaligned fetch will straddle the boundary between two 64-byte cache lines, which can hurt performance, so gcc tries hard to make this not happen. Andrew.
