On Sun, Jul 17, 2011 at 9:15 PM, Walter Prins <wpr...@gmail.com> wrote: > > > On 17 July 2011 15:26, Lisi <lisi.re...@gmail.com> wrote: >> >> Sorry to be slow. Blame virtually no sleep last night ;-( But even were >> the >> power of two bit correct (and I see subsequently that it is not), how is >> 18 a >> power of two? >> > > The 18 bytes is a bit of an irrelevance. The point is that if the start of > the buffer falls on a dword (double word) alligned memory location then in > theory the access should be faster. The term is a little bit ambiguous > because strictly speaking different processors have different word sizes. > Even so, usually when people speak of double-word alignment, it's often the > case that the term word in such a context has its original meaning, e.g. 16 > bits. A dword is then 32bits or 4 bytes. A doubleword aligned memory > address is, using these assumptions, therefore an address that is divisible > by 4. Obviously if the word size is 32bits, then a double word would be > 64bits and a doubleword aligned address would need to be divisible by 8. As > an aside, this type of optimization is often taken care of by compilers > under the hood, and in any case it's generally not something that you'll > really be considering as a Python programmer. (If however you were working > on one of the Python runtimes or implementations, then you might well be > sometimes considering this type of thing, depending on exactly how > performance critical what you are working might be and what the runtime was > being implemented in.) >
It's not just about performance. Some hardware simply cannot access data that is not correctly aligned. C programs that indiscriminately cast among pointers to types of different sizes are a pain to port off lenient architectures like x86. If you're writing C code that deals with pointers, you *always* need to keep alignment in mind. -- regards, kushal _______________________________________________ Tutor maillist - Tutor@python.org To unsubscribe or change subscription options: http://mail.python.org/mailman/listinfo/tutor
