On Sun, Jun 14, 2009 at 02:43:17PM -0700, Tim Prince wrote: > Michael Meissner wrote: > > > > > 32-bit GCC on x86 defines the alignment of long long to 4 bytes, while 64-bit > > GCC on x86 defines the alignment to be 8 bytes. Unfortunately the 32 bit x86 > > is also stuck with doubles being aligned on a 32-bit boundary, due to the fact > > that the people creating the original ABI were using a 386 which did not need > > higher alignments, and newer processors lose performance because it isn't > > properly aligned. > In x86 32-bit mode, long long continues to be implemented with pairs of > int, so the performance problems with 4-byte alignment arise only for > double and vectorizable types. Well, there is the FILD instruction that converts from a 64-bit integer and pushes the value on the 387 stack, and FIST which pops off the stack and converts it to 64-bit integer. But I would argue it was a mistake for doubles not being aligned in the ABI (and I would think long longs also should have the same alignment, even if technically the hardware doesn't require it). -- Michael Meissner, IBM 4 Technology Place Drive, MS 2203A, Westford, MA, 01886, USA meissner@xxxxxxxxxxxxxxxxxx
