IBM Releases Compiler for Power4 and G5

davids-world.com writes "IBM offers its optimized XLC compiler not just for Intel CPUs, but also for its own G5 processor (article in German at Heise). Unlike gcc, it is optimized for the G5 and achieves a major boost in speed, as first results show. I guess we will have to compare the new benchmark data (once available) with the data we get with the optimized Intel compiler for Xeon. The compiler is available for download now."

Comments

What license is this under?

[burgburgburg]

This is a preview beta. What will the final product be released under?

Questions.

[emil]

1. Can this compiler be used on any open-source unix-like kernels? I know that Linux has some hooks into gcc.
2. Can this compiler be used to prepare glibc or any other major C libraries?
3. Can this compiler be used to generate native Mac OS X GUI applications (cocoa)?
4. Will the source be released?

It is good to see IBM ending the habit of charging extra for the C compiler. AIX hasn't bundled the compiler since 3.2.5.

Re:Why?

[Theaetetus]

I'd imagine one could cook up code to make whichever machine they want look "faster".

Um, that's the point. IBM cooks up code to make the G5 look as fast as possible. Intel cooks up code to make the Xeon looks as fast as possible. Now compare. Fastest code/chip against the competition's fastest code/chip. The complaint in the past was that it was always the fastest code/chip of one side against a non-optimized code on the other side. Now we can have a fair shootout.

-T

Interesting, but...

[Doktor Memory]

I wonder how the AltiVec support in XLC for OSX compares to that in GCC?

This is actually really important. One of the big reasons that the Intel C compiler spanks every other available x86 compiler is that its SSE/SIMD support is, in the words of one of my assembly-programmer friends, "awe-inspiring." Like, unrolling entire program loops and replacing them with single SIMD instructions.

As far as I know, pretty much all of the AltiVec/VMX support in GCC was contributed by Apple and Motorola, and prior to the ppc970, IBM has never produced a PPC CPU with AltiVec instructions, so prior versions of XLC have never had to support it. So I'll be really curious to hear how it stacks up against GCC's Altivec.

Re:So much for open source at IBM

[Tal Cohen]

(Disclaimer: I used to work for IBM Research. This is not inside info.)

The XL line of compilers has a long and glorious tradition. What matters most here is compatibility with existing code. IBM doesn't optimize XLC so you could get a faster kernel, it optimizes XLC so they could recompile DB/2 on AIX and get faster results; and also, obviously, for many of their other products. Another consideration is those clients that have a large codebase used with XLC.

The world of C/C++ is rather sad in that code is not only non-cross-platform in most cases, it is also commonly unlikely to compile using a different compiler on the same platform. Moving from XLC to gcc would mean throwing away an investment of decades in the compiler, not to mention the huge amount of work required to port the source code of products such as DB/2. So IBM helps gcc, and also invests in XLC. There's no conflict here. I would not be surprised if many of the various optimizations employed by XLC will eventually find their way to gcc -- by the very same IBM researchers.

BTW, talking of optimizations, people in IBM Haifa have developed an amazing post-compile optimization tool

. Now, seeing an open-source version of that would be cool. Not very likely, though, since this is the source for some of the performance edge DB/2 has over its competitors.

Re:So much for open source at IBM

[p3d0]

And is there a particular reason why IBM couldn't apply their work towards gcc? So much for the whole open-source, contribute-to-the-community philosophy.

Oh please. It's not like there's only one compiler team in IBM. IBM is contributing big time to gcc, and they also maintain their own compiler for customers who want that. IBM, like any smart company, hedges its bets, so they work on multiple overlapping projects. Do you expect them to bet the company on technology they don't control? The kind of reactionary thinking that would have them switch all their manpower to gcc is exactly what kills a lot of less prudent companies.

Plus, the team working on this compiler would probably take several man-decades---if ever---to become as fluent and comfortable with gcc as they are with their own code.

(Disclaimer: I work for IBM. In fact, I work on a compiler for IBM, though it's neither this compiler nor gcc. My opinions don't reflect those of IBM.)

I can definitely believe it

[TheTranceFan]

I personally ported Macintosh Adobe Premiere 4.0 from 68000 to PowerPC. This was right when the PowerPC-based Macs were introduced (before, actually, since Apple was providing us with prerelease product). At that time, Apple's internal C compiler (then MrC) wasn't ready for primetime, nor was ThinkC. We'd been using MPW for the development, but the only good compiler for PowerPC was IBM's compiler. So I edited/built/linked the whole thing over the network on an RS/6000(?) somewhere at Adobe. I remember when I turned the optimizer on, Premiere got twice as fast, just like that.

The IBM compiler dis some wild instruction reordering which made the optimized compiled code really hard to understand, but somehow better fitted to the processor's pipeline structure. Fortunately the only thing that broke when I turned on the optimizer was the "marching ants" used for selection, and that was the result of some way-too-fancy-casting of Pattern pointers that fooled the optimizer. I suspect the IBM compilers will continue to reign if performance is the goal.

Re:Here we go again:

[rsborg]

I'm confused. Newegg claims to sell the Opteron 240 for $300 -- oh wait, I'm out of date -- they're $255 now. Is this a scam or something? I have a hunch it's for real, and they'll actually ship the chips if I pay them.

Hmm... lets see... you're comparing a processor to a full system. Apples vs. Oranges.
Lets do some real analysis... top server vs. top server:

• A fullly specced dual G5 2.0Ghz ... $2770
  vs.
• a similarly specced dual Opteron 244 1.8Ghz.... $3013
  And this isn't even from big boys like Dell (who wouldn't be caught dead using AMD)... just a bottomfeeder from pricewatch.

Now, the G5 has all sorts of desktop/workstation goodies built in (Radeon 96/9800 Pro, awesome sound, good looks), and the with AMD server, you can cut corners with things you don't want in a server, like soundcard, highend video card, etc.

My point? im sure we're still comparing apples to oranges. People who want a mainstream kickass workstation will love to buy the G5. People who want beowulf node clusters or inexpensive uberservers will love the Opteron. I just wish I had either :-)