Slashdot Mirror
Despite Aging Design, x86 Still in Charge
An anonymous reader writes "The x86 chip architecture is still kicking, almost 30 years after it was first introduced. A News.com article looks into the reasons why we're not likely to see it phased out any time soon, and the history of a well-known instruction set architecture. 'Every time [there is a dramatic new requirement or change in the marketplace], whether it's the invention of the browser or low-cost network computers that were supposed to make PCs go away, the engineers behind x86 find a way to make it adapt to the situation. Is that a problem? Critics say x86 is saddled with the burden of supporting outdated features and software, and that improvements in energy efficiency and software development have been sacrificed to its legacy. And a comedian would say it all depends on what you think about disco.'"
It should be replaced with Esperanto when we all upgrade to Vista.
technical writing / development
I'm going to go with:
Did I miss anything?
If you disagree, post your argument. (-1, Overrated) isn't your personal censorship tool for views you don't like.
The x86 instruction set will be retired in the same year as the QWERTY keyboard layout.
Just like the four stroke engine. It's not the best one, it can be largely enhanced and made better, but it's still here.
And just like the four stroke engine, modern engines just burn gasoline and push car forward. This is where the similarity with the original engines end.
Maybe Computers will never be as intelligent as Humans.
For sure they won't ever become so stupid. [VR-1988]
At this point, does it matter as much? As we move on the future is clearly x86-64 which is MASSIVELY cleaned up compared to x86 and is really rather clean compared to that. Sure at this point we still boot into 8086 mode and have to switch up to x86-64 but that's not that important, it only lasts a short while.
As we move off of x86 onto -64, are things really still that bad? Memory isn't segmented, you have like 32 different registers, you don't have operands tied to registers (all add instructions must use AX or something like that) as some 16/32 bit instructions were.
Of course, we should have used a nice clean architecture like 68k from the start, but that wasn't what was in the first IBM.... and we all know how things went from there.
Comment forecast: Bits of genius surrounded by a sea of mediocrity.
Yes, the instruction set is old, but, it does still work. As a consumer, why should I have to re-invest in software that I purchased and does the job, just becuase my hardware failed, or faster hardware becomes available and I upgrade. Apple bit that one some time ago. Last year, I had an investment of $4000.00 in software when Intel came out with a significantly faster part that was dropping in price. Just by upgrading my hardware (cost $800) my invenstment improved significantly. $4800.00 did not justify the upgrade but the low cost of hardware only, did. Also, there was not learning curve involved.
You don't buy a new car just becuase the tires need replaceing (well some people do, but that is rarely the fiscally responsible thing).
If it ain't broke, it doesn't need fixing.
Athiesm is a religion like not collecting stamps is a hobby.
It has been said that people will not change unless something is preceived to be 10 times better. The problem is nothing has been perceived to be that much better, so people stay with what they know.
Paul
Things would be a lot easier if the darned thing wasn't so bloody complex to emulate. I mean if we were "stuck" with (say) an ARM or even a 68K we'd be able to use virtual machines to dig ourselves out of a similar architectural hole (though with an ARM we'd be unlikely to want to).
:-)
The x86 has so many modes of operation (SMM, real/protected, lots of choices for vectorizing instructions, 16/32/64 bit modes) and special cases that it's a pretty big project to get emulation working correctly (much less fast). You're pretty much stuck with a 10x reduction clock-for-clock on a host. Making an emulated environment secure is hard, too; you don't necessarily need specialized hardware here (e.g., specialized MMU mapping modes), but it helps.
And now, with transistor speeds bottoming-out, they want to go multicore and make *more* of the things, which is exactly the opposite direction that I want to go in...
Any sufficiently advanced technology is insufficiently documented.
If a chipmaker declared its chip could run only software written past some date such as 1990 or 1995, you would see a dramatic decrease in cost and power consumption, Crosby said. The problem is that deep inside Windows is code taken from the MS-DOS operating system of the early 1980s, and that code looks for certain instructions when it boots.
Even new software might (and often does) use the so-called old instructions. If you want to completely redesign the hardware you would also have to completely rewrite the software from scratch as you would not be able to rely on previously written code and libraries. This is simply not feasible on a global scale...
09 f9 11 02 9d 74 e3 5b d8 41 56 c5 63
Who is this guy and what is he smoking? Over half of a modern processor is cache. The instruction decoding and address decoding are a small fraction of the remainder. Where does he get the 60% from?
Sometimes I doubt your committment to SparkleMotion!
I know we all bitch about old designs, legacy support for outdated features, but, one of the things that keep people from moving from one OS to another is "existing base of installed software" and "knowledge of exisiting software". Like it or not, the major player is Microsoft. No matter how much a geek says, MS UI's suck, people are comfy with them. If alternative OS's had the same software offerings with the same UI, people would be able to move to them. The same holds true for processors.
No matter how well a processor performs, if there is no application base for it, no one is going to buy a machine with that processor. In this case, perception is reality. You walk into a software store, you see 16 rows of Windows applications, half a row of Linux, and 5 rows of Apple.
What processor family runs each of these? Guess who has moved to the dominant processor?
The only way to build a software base is to build in legacy support. Then start weening users away from the legacy features, get programmers to stop using those features (mainly those building the compilers that developers use), and move towards the more advanced features.
x86 rules for a reason. Microsoft rules for a reason. The customer is comfortable with them, and their perception is reinforced everytime they go to the store.
Politics is the art of looking for trouble, finding it everywhere, diagnosing it incorrectly and applying the wrong fix.
4. Price / performance. A segment the x86 have done well in.
:D
5. Security. Will my x86 progs be supported in 20 years? The answer: yes.
6. Availability. Hmm... Intel, I'd like to 1 000 000 CPUs. Intel: Sure thing.
7. Good will. What should we buy, Intel or PPC. PPC? What's that? Go Intel! Yes boss. (Just look how far Itanium got on Intel's name, alone.)
Already been done, didn't catch on (see Itanium).
Because there is such a massive amount of installed x86 software base that you'd be throwing away silicon. To be sure that software ran on the most systems possible, software would still be written for x86 and not the 'desired' architecture.
That being said, OSS tends to have good inroads in that you get all the source so can recompile to whatever architecture you want. However, since x86 is still the huge marketshare, other architectures get less attention. Also, all of the JIT languages (Java, C#, etc.) make transitioning easier IF you can get the frameworks ported to a stable environment on the 'desired' architecture.
The main problem is that there is *so* much legacy code in binary (EXE) format only (the source code for many of those has been literally lost) that can be directly tracked to money. There are systems that companies continue to use and have so much momentum that changing platforms would require extreme amounts of money to reverse engineer the current system - complete with quirks and oddities, rewrite, and (here is a big part that many people fail to add in) retest and revalidate, that many companies don't want to spend that kind of money to replace something that 'works'.
There's so much work/time/effort invested in x86 now that it's hard to jump off that train. AMD's x86-64 is a good approach in that you can run all the old stuff and develop on the new at the same time with few performance penalties. However, I don't know if we'll ever be able to shrug off the burden of x86.... at least not for a long time to come. It'd take something truly disruptive to divert from it (and what people are currently invisioning as quantum computing is not that disruption).
Boot loaders tend to be 16bit segment model code 8086, at least they contain enough code to get into 32bit mode. The BIOS will be 16bit legacy code, at least some anyway as a x86 PC chip still boots in Real Mode (there is a 386 embedded variant that doesn't). Windows 9x series is _RIDDLED_ with 16 bit code esp the display drivers, although many of these switch to 32bit mode ASAP the entry points are 16 bit code. Any attempt at killing off 16bit code would stop any 9X system running.
For WinNT and variants (2K, XP) I don't know how much 16bit code is in there. I've written drivers for 2K/XP and could not find a single 16bit style instruction however even NT series for x86 uses segments. FS is used for process & thread info. IIRC even AMD64 long mode implements FS & GS to make OS porting easier.
Lastly. 16bit code (instruction operating on 16bits of a 32bit register) are trivial in 32bit mode - all you have to do is preceed an instruction with 0x66 and/or 0x67 to switch a 32bit instruction to a 16bit instruction.
The problem transcends MSDOS and goes to the BIOS and boot sequence itself. Intel tried to address the with EFI but that seems to be slow gaining traction - probably because of backwards compatibility.
Time flies like an arrow. Fruit flies like a banana.
Computer manufacturers have tried making non-compatible machines. Commodore 64, VIC 20, Coleco Adam, Atari ST. They all had their place in time and their niche in the market before fading out.
Something they all had in common, though, is that they sold better than IBM's mostly-compatible PCjr. I attribute that difference to software and compatibility problems. Because of BIOS differences, a number of programs written for the PC couldn't run on the PCjr. That led to a fragmentation of shelf space at software retailers and confusion among retail customers, and led to customers avoiding the platform in favor of easier-to-understand options.
I would expect something similar to happen if Intel, AMD, or anyone else started making mostly-compatible x86 processors. It wouldn't sell unless all of the software people are used to running still worked. Sure, someone could take Transmeta's approach and emulate little-used functionality in firmware rather than continuing to implement everything in silicon, but it all pretty much needs to keep working, so why bother?
Seriously, why would anyone undertake the effort and expense needed to slim-down x86 processors when the potential gains are small and the market risk is pretty huge? No chip manufacturer wants to replace the math-challenged Pentium as the most recent mass-market processor to demonstrably not work right.
Pundits and nerds can talk all they want about why the x86 architecture should be put out to pasture, but it won't happen until a successor is available that can run Windows, OSX, and virtually all current software titles at acceptable speeds. At that seems pretty unlikely to happen on anything other than yet another generation of x86 chips.
You know, there is a difference between trolling and pointing out the flaws in your reasoning. Just saying.
Basically x86 isn't a perfect instruction set for today's landscape, but then again UNIX isn't a perfect operating system for today's landscape; that doesn't mean it's not still very good and we shouldn't praise those who have made it so good.
Some say plan9 has a better design than Linux, some say that PPC has a better design than x86, but apparently design isn't everything.
Lots of things could be better if we could get everyone to migrate from what they currently use, but would it be worth it in this case? I don't think so, at least not until we reach the limits that better design & hardware can do.
// MD_Update(&m,buf,j);
If free software ever goes truly mainstream, and the stacks people use are free from top to bottom, lock in goes away in general. Even hardware lock in.
A couple of years ago, I was shifting some stuff around and I needed to clean off my main desktop machine, an x86 box. I installed the same linux distro on a G4 mac and just copied my home directory over. Everything was exactly the same -- my browser bookmarks and stored passwords, my email, my office docs, etc.
A lot of people take Apple's jump from PowerPC to x86 as a sign that x86 is unstoppable. But I'd argue that the comparative ease with which the migration took place shows how weak processor lock in is becoming. The shift from PPC to x86 was nothing compared to the jump from MacOS Classic to OS X.
The real reason x86 won't go away any time soon is that MS has decided that's the only thing it's going to support, and MS powers most of the computers in the world. Windows is closed, so MS's decision on this is final, and impossible to appeal.
I would add to this that ISA mattered a lot more when I wrote code in assembly language. For a clean (and simple) instruction set architecture, I fondly remember the PDP-11. Later on, the 680x0 offered more powerful addressing modes for less simplicity (and consistency). Compared to both, the x86 was infuriating to work with.
ISA's still mattered, but less, in my early "C" days when source-level debugging was less robust, or even to understand what the compiler was turning my code into so I could figure out where to optimize.
Today, it hardly matters at all. Looking at generated code tells me little about how the processor with multiple execution units is going to process it; it is necessary to trust the compiler and its optimization strategy. It matters even less with interpreted or JIT'd languages, where the work eventually performed by the processor is far removed from my code. Knowing what's happening at runtime involves much more important factors than the ISA.
Actually the encoding is VERY efficient where it matters most, cache density and limiting the number of calls to main memory. Having complex instructions helps in the areas where real world performance is most hurt and that is why we have a CISC frontend to an efficient RISC backend. This balance was reached even in the "RISC" camp, look at the PPC970 with the more complex instructions that get broken down in uops and dispatched to execution units, very similar in many ways to how modern x86 processors work. The translation layer is less than one percent of die space and probably a much lower percent of power usage on modern x86 chips.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
We lose the X86 when another processor comes along that is cheaper, 10x more powerful, and runs all X86 software at the speed that the users consider to be the same as a PC. Until then we keep the X86. Simple as that. Next tech issue, please.
The irony of being a grammar Nazi by pointing out that the statement "People like you make Nazi's look good." is incorrect. It should be "People like you make Nazis look good." - unless there is some unseen object that the Nazis posses, the apostrophe before the "s" is unnecessary and incorrect. Simply adding a "s" to the end is sufficient to make it plural.
"But this one goes to 11!"
The bottom line is: has any other architecture enabled apps run significantly faster over multiple CPU generations at comparable costs? Nope. As other architecture fads have come and gone, but the X86 just absorbs the best ideas from each and keeps marching along.
The Instruction Set of a processor architecture with so many resources available to it doesn't really matter, so long as it isn't utterly and completely braindead. X86 isn't braindead enough to qualify... if you had an intercal instruction set or an One Instruction Set Computer it might.
You really want to do several things to get performance out of an instruction stream -- register renaming, instruction manipulation (breaking them apart or joining them together or changing them into other instructions), elimination of some bad instruction choices, and a host of other things. You would want to do these things even on a "clean" ISA like Alpha or PPC or MIPS. And if you are doing them, the x86 instruction set suddenly becomes much less of a problem. There are even advantages: the code size on x86 tends to be better than a 32-bits-per-instruction architecture.
Instruction sets are languages with exact meanings. Which means that you can precisely translate from one instruction set to another. And, as it turns out, you can do it fairly easily and efficiently. Which is why Transmeta did pretty well. Which is why Apple's rosetta and Java JIT compilers work (and Alpha FX32 before that). Which is why AMD and Intel are right there at the top of the performance curve with x86-style instruction sets, because it JUST DOESN'T MATTER THAT MUCH.
Why didn't Transmeta kick more butt? Because they didn't have the economies of scale that AMD and Intel have. Because they didn't have the design resources that AMD and intel have. Because AMD and Intel had better-tuned processes faster than TSMC or whoever was fabbing Transmeta's chips. THOSE are the most important things, not the instruction set that you have on disk.
Now a good ISA can help in many ways: SIMD instructions really help to point out data level parallelism. More registers helps a wee bit to prevent unnecessary work done around the stack for correctness. You can get rid of a bit of logic if you can execute without translation. But these things can either be added to x86 (SSE/x86-64) or aren't expensive enough to be worth it on a 100 sq mm, >50W processor. Maybe in an embedded, low-power processor.
-- Erich
Slashdot reader since 1997
...rather than intelligent design.
"How to Do Nothing," kids activities, back in print!
> Now, this is the important part: He's used to XP. He's used to an OS, that while sucky, worked well enough for him, was relatively speedy, so why can't he just have that? Why does he have to have something replaced that worked just to put up with this shit?
If instead of giving up after a day, he had tried it for a week or a month, he would have found out how great everything is. Then in a few months he would be used to it and if you try to make him downgrade to XP he will cry.
There are many great features in Vista, but you have to try it for yourself.
I'll probably be modded down for this...
Any processor has to do the exact same work, whether the user-visible encoding is done this way or as an "SP indexed" addressing mode. At the micro-op level, it all gets renamed, reordered, etc. so that the same things are happening. Moreover, that particular sequence is so common, in all probability most X86 CPUs have special logic just to optimally execute that entire sequence faster that the naive RISC equivalent.
What's with all this dissing of the X86?
Like you, I'm an old fart; I wrote assembler code for the PDP-8, PDP/LSI-11 and the 68k. They were ok: easy to learn and use, but I always preferred the X86.
Sure, it was harder to learn and I never got past having the blue book on my desk when I was coding but, in the end, it produced smaller, faster code. There were a number of apps I wrote for multiple platforms, so I got to compare. Also, (the same reason I love perl) you could do astounding things with side-effects.
Commercially, X86 has staying power because it was architected to scale. Variable-length instructions with lots of space in the operator range lets Intel adapt the design to any new demands. Most, if not all, of the complaints about X86 (e.g. too few registers) are just version features—yesterday's news if there's a market demand for an improvement.
Bottom line—it ain't neat, but that doesn't matter; it's programmed once and used millions of times. Programmer convenience is irrelevant.
I'm a Programmer. That's one level above Software Engineer and one level below Engineer.
I tried looking into my heart but it asked me to "allow" or "deny". When I hit "allow" I got a BSOD. I'll have to get back to you on that one.
The x86 instruction set is a surprisingly good way to build a computer. The reasons aren't obvious.
First, the original x86 was a huge pain, with that stupid segmented memory arrangement. But IA-32 was better and cleaner; at last there was a flat 32-bit address space. (Yes, there's a segmented 48-bit mode, and Linux even supports it, but at least apps see a flat address space.) AMD-64 is even more regular; the segmented memory stuff is completely gone in 64 bit mode. So there is progress.
RISC architectures could yield simple machines that could execute one simple fixed-width instruction per clock cycle. The early DEC Alphas, the MIPS machines, and early IBM Power chips are examples of straightforward RISC machines. This looked like a big win. The ALU was simple, design teams were small (one midrange MIPS CPU was designed by about six people), and debugging wasn't hard. RISC looked like the future around 1990.
What really changed everything was advanced superscalar architecture. The Pentium Pro, which could execute significantly more than one instruction per clock, changed everything. The complexity was appallingly high, far beyond that of supercomputers. The design teams required were huge; Intel peaked somewhere around 3000 people on that project. But it worked. All the clever stuff, like the "retirement unit" actually worked. Even the horrible cases, like code that stored into instructions just ahead of execution, worked. It was possible to beat the RISC machines without changing the software.
The Pentium Pro was a bit ahead of the available fab technology. It required a multi-chip module, and was expensive to make. But soon fab caught up with architecture, and the result was the Pentium II and III, which delivered this technology to the masses. Then AMD figured out how to do superscalar x86, too, using different approaches than Intel had taken.
The RISC CPUs went superscalar too. But they lost simplicity when they did. One of the big RISC ideas was to have many, many programmer-visible registers and do as much as possible register-to-register. But superscalar technology used register renaming, where the CPU has more internal registers than the programmer sees. The effect is that references to locations near the top of the stack are as efficient as register references. Once the CPU has that capability, all those programmer-visible registers don't help performance.
Making all the instructions the same size, as in most RISC machines, leads to code bloat. Look at RISC code in hex, and you'll see that the middle third of most instructions is zero. Not only does this eat up RAM, it eats up memory and cache bandwidth, which is today's scarce resource. Fixed size instructions simplify instruction decode, but that doesn't really affect performance all that much. So x86, which is a rather compact code representation, actually turns out to be useful.
I drive a '96 cavalier; Its not stylish, its not particulalry fast, no power windows or locks and due to some dings, its not even orthogonal anymore. But it was cheap, relatively fuel-efficient, reliable and it gets me from A to B as fast as I'm otherwise allowed. We geeks tend to pine over these sleek ISAs like MIPS or Power in much the same way that car enthusiasts wax romantic about the latest sports car. For most of us however, practicality forces us to drive more modest vehicles. Its not practical to drive a vehicle that requires some exotic fuel in the same way that its not practical to run a CPU that digests some exotic instruction set, and for the same reasons: Limited use and availability leads to higher cost-of-ownership overall. Economies of scale and past investment lead to comparatively rock-bottom prices. The PC is also bogged by something far more sinister than the x86 instruction set, namely, the PC BIOS. This is only just beginning to go away with Apple having adopted Intel's EFI firmware (OpenBIOS on their PPC systems before that) and the growing list of LinuxBIOS supported motherboards (still not ready for personal use, but getting there). Widespread EFI adoption might take place if Microsoft releases a home OS with the capability of using EFI without the BIOS compatability layer. Another point to watch for in the future is the proliferation of platforms such as the CLR (.NET) and to a lesser extent, the JVM. These sort of platforms serve as an abstraction layer between the instruction set the software is written in, and the instruction set of the hardware on which it runs. With a performance difference of 10% or so now, and that difference shrinking as the technology matures, we'll begin to see that the underlying architecture will loose its hold on being the defining element of the platform. We're already beginning to see x86 technolgy moving towards extensions to make virtualization (such as Xen) more efficient, and I suspect it will not be long before it moves to include features to make the .net platform and similar technologies run more efficently as well. If these sort of technologies eventually become the defacto target for software, we may see a future in which the CPU's sole purpose becomes to efficiently support a higher-level platform that is defined by software.
In the Embedded world, x86 does not reign - in fact, x86 is a very small portion of the embedded market. PowerPC rules, followed by ARM and 68k, this doesn't even mention smaller processing tasks run by Microcontrollers like the 8051 or PIC devices. x86 has all but been ousted where engineers are freed from the concerns of backwards compatibility and high performance is not required.