> At least for Intel, between its branch predictor and speculative > execution engine, it doesn't make a difference.
*Sigh*. We need live measurement. My testing (in your test harness!) showed a noticeable (~10%) speedup. > When I did a quick comparison of your 64-bit fast_mix2 variant, it's > much slower than either the 32-bit fast_mix2, or the original fast_mix > alrogithm. That is f***ing *bizarre*. For me, it's *significantly* faster. You *are* compiling -m64, right? Because I agree with you it'd be stupid to try to use it on 32-bit machines. Forcing max-speed CPU: # ./perftest ./ted64 fast_mix: 419 fast_mix2: 419 fast_mix4: 318 fast_mix: 386 fast_mix2: 419 fast_mix4: 112 fast_mix: 419 fast_mix2: 510 fast_mix4: 328 fast_mix: 420 fast_mix2: 510 fast_mix4: 306 fast_mix: 420 fast_mix2: 510 fast_mix4: 317 fast_mix: 419 fast_mix2: 510 fast_mix4: 318 fast_mix: 362 fast_mix2: 510 fast_mix4: 317 fast_mix: 420 fast_mix2: 510 fast_mix4: 306 fast_mix: 419 fast_mix2: 499 fast_mix4: 318 fast_mix: 420 fast_mix2: 510 fast_mix4: 328 And not: $ ./ted64 fast_mix: 328 fast_mix2: 430 fast_mix4: 272 fast_mix: 442 fast_mix2: 442 fast_mix4: 272 fast_mix: 442 fast_mix2: 430 fast_mix4: 272 fast_mix: 329 fast_mix2: 442 fast_mix4: 272 fast_mix: 329 fast_mix2: 430 fast_mix4: 272 fast_mix: 328 fast_mix2: 442 fast_mix4: 272 fast_mix: 329 fast_mix2: 431 fast_mix4: 272 fast_mix: 328 fast_mix2: 442 fast_mix4: 272 fast_mix: 328 fast_mix2: 431 fast_mix4: 272 fast_mix: 329 fast_mix2: 442 fast_mix4: 272 And on a Phenom: $ /tmp/ted64 fast_mix: 250 fast_mix2: 174 fast_mix4: 109 fast_mix: 258 fast_mix2: 170 fast_mix4: 114 fast_mix: 371 fast_mix2: 285 fast_mix4: 109 fast_mix: 516 fast_mix2: 156 fast_mix4: 90 fast_mix: 140 fast_mix2: 184 fast_mix4: 170 fast_mix: 406 fast_mix2: 146 fast_mix4: 88 fast_mix: 185 fast_mix2: 114 fast_mix4: 94 fast_mix: 161 fast_mix2: 116 fast_mix4: 98 fast_mix: 152 fast_mix2: 104 fast_mix4: 94 fast_mix: 352 fast_mix2: 140 fast_mix4: 79 > So given that 32-bit processors tend to be slower, I'm pretty sure > if we want to add a 64-bit optimization, we'll have to conditionalize > it on BITS_PER_LONG == 64 and include both the original code and the > 64-bit optimized code. Sorry I neglected to say so earlier; that has *always* been my intention. The 32-bit version is primary; the 64-bit version is a conditional optimization. If I can make it faster *and* have more avalanche (and less register pressure, too), it seems worth the hassle of having two versions. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/