OSNews

Take that mailing list posting with a grain of salt - not because Chris works for Apple, but mostly because there has been a big herd of madly skilled Linux and BSD hackers optimizing the **** out of gcc on Intel (and Xcode is mostly a fancy UI around gcc), where the PPC port of gcc hasn't received nearly as much devotion.

That argument is misleading, since it conflates code performance and compile performance. Lot's of people have been optimizing GCC's code generation for x86, meaning that GCC generates better code on x86 than it does on PowerPC. GCC itself hasn't gotten a lot of optimization attention until recently, and most of it has been directed at the C++ frontend and optimizer. These components, which currently dominate GCC's compile-time, are fairly unoptimized generic C code, like the rest of GCC.

Of course, what this benchmark really tells us is that the G5 is a poor performer on generic, branchy integer code with lots of random memory references. This fact shouldn't come as a surprise to anyone.

Edited 2006-01-28 00:32

Optimization was actually one of the purposes of the subtasks in the EGCS project. It's also important to differentiate between different types of optimization. But that's besides the point. At best the grandparent's argument is that an Intel-built GCC is more efficient at compiling because it is built itself with a compiler more optimized for the x86. If one were to compile GCC with XLC would that satisfy the grandparent or not? If one were to explicitly disable all optimizations for both builds of GCC, would that be sufficient? What about linker considerations? Does he have any specific complaints there? Requirements of universal binaries for both?

It's interesting to look at the SPECint GCC results for the G5 and Pentium-M. The 2.26 GHz P-M gets 1994 in the gcc subtest, the 2.2 GHz 970FX gets 1130.0. Scaling for the clockspeeds used in the second benchmark I linked to (1.83 for the P-M, 2.7 for the G5), and you get the same 16% speedup reported in the posting. This implies that GCC compiled with XLC probably isn't notably faster than GCC compiled with GCC.

The code generation issue is a factor, but its entirely hypothetical. Considerations of XLC or ICC are moot --- GCC is the compiler Apple uses. The G5's performance with any other compiler is largely irrelevent. Also, casting the blame on GCC entirely is not fair. GCC generates poor code on the G5 because the G5 is hard to generate good code for. It's two cycle integer latency and group dispatch scheme requires a lot of smarts from the compiler.

Since this is just an anecdote it's not especially compelling. If it suggested the opposite or something Rayinder doesn't expect to be true, he would criticize it as such. He might even say something about CINT2000 scores or what have you.

You dismiss the result because you don't think the PPC970 backend of GCC is as optimized as the x86 backend. Even if that were the case, what would it matter? If you have a build cluster, you don't really care if it could potentially, maybe, someday be competitive if you spent a large sum of money paying experts to improve GCC's performance. If a build cluster of Core Duo iMacs is more efficient than a build cluster of quad G5s, then you're just going to buy the latter. If that's all they're doing, and they do it cheaper while being marginally slower, that's a no-brainer. They even take up less space! Though you're probably just going to wait for the next XServe.

We can't really say why this person obtains the results he does, or if they're representative. We can't because there's not a lot of information. Maybe it's because the machine he's using doesn't have enough RAM, maybe it's simply demonstrative of his particular build, perhaps he's running into some limitation of MacOS, perhaps the 4-way G5 scales about as badly as a Xeon, or perhaps the G5 really is just that bad at integer performance compared to the Core Duo.
Until someone bothers to perform real testing, no one knows.

That should have been "former' and not "latter."

Since this is just an anecdote it's not especially compelling. If it suggested the opposite or something Rayinder doesn't expect to be true, he would criticize it as such. He might even say something about CINT2000 scores or what have you.

The anecdote is what it is: an account of a person who uses both systems can offer a performance comparison based on their own usage. It could be more scientific, but its hardly any less useful than the tests Macworld conducted.

Also, I expect certain things to be true because I have reasons to expect them to be true. I know the G5's compile performance is poor --- I have one and an Athlon64, and its obvious to me which one compiles code faster. The SPECint.gcc results for the P-M and 970FX show the same thing. I posted the anecdote because I felt it was an accurate account of the G5's performance in a "real world" situation. I'd give SPEC scores, but the PPC fanatics don't believe in SPEC...

Edited 2006-01-28 07:47