Comment Re:Benchmarking Across Platforms (Score 1) 1595
One note up front: I'm guessing based on knowledge of the PowerPC architecture...
> G5 Specific instruction usage sounds suspicious
This one would probably be the instructions that exist in a 64 bit PPC chip but not in a 32 bit version. Things like 64 bit loads/stores and 64 bit shifts/rotates.
> I really like the relaxed IEEE math operations.
The PPC floating point unit has a few instructions that don't follow the IEEE spec. In particular, there are combined multiply/add and multiply/subtract instructions that are much faster than doing the two operations separately. However, they don't round the intermediate results, so they don't always result in the exact same output. Because of this, they normally aren't allowed to be used in any program expecting exact IEEE compliant results.
I'll agree that the malloc thing does look somewhat suspicious. There's was a big flamewar over the performance of the default malloc on OSX on one of Apple's mailing lists not too long ago.
> G5 Specific instruction usage sounds suspicious
This one would probably be the instructions that exist in a 64 bit PPC chip but not in a 32 bit version. Things like 64 bit loads/stores and 64 bit shifts/rotates.
> I really like the relaxed IEEE math operations.
The PPC floating point unit has a few instructions that don't follow the IEEE spec. In particular, there are combined multiply/add and multiply/subtract instructions that are much faster than doing the two operations separately. However, they don't round the intermediate results, so they don't always result in the exact same output. Because of this, they normally aren't allowed to be used in any program expecting exact IEEE compliant results.
I'll agree that the malloc thing does look somewhat suspicious. There's was a big flamewar over the performance of the default malloc on OSX on one of Apple's mailing lists not too long ago.
