
Originally Posted by
llama
Wrong (about Crafty). If you'd compiled 32 bit craft on OS X, it would have been slow there. I think you're being fooled by the lack of a 32bit OS X benchmark. Crafty heavily depends on 64bit operations. Chess boards have 64 squares, and crafty uses 64 bit ints as "bit board" bitmaps. It probably has to AND, OR, and XOR 64 bit ints all over the place, so no wonder it's more than 3 times slower on a register-poor 32 bit architecture. I haven't looked up other (non PTS) crafty benchmarks, but there are lots published. I wouldn't be at all surprised if this 32 vs 64 slowdown is typical.
The other wins for OS X are interesting, and some are either better filesystem performance, or performance regressions from gcc 4.2.1 to 4.4.1 (if those were the respective compiler versions). Some probably are actually better multi-thread handling by the OS, which is of course most interesting, since that's one thing you can't solve just by trying your code with different gcc versions to find which one makes the fastest binary.
Which benchmarks are compiled from source, and which aren't? I expect most/all of the computational ones are, but what about e.g. Nexuiz? They distribute handy binaries...