AMD after X86

[OneOfTheseDays]

Originally posted by: dmens
Given the option of ripping out legacy support to regain that dinky little bit of space and running the risk of losing all backwards compatibility, the choice is obvious.

As for register management, making the physical rf non-visible is not a waste. In fact, the hardware will do a much better job than any compiler, excluding the use of predicates and static hints (which is not out of the question for x86). For basic assignments, there is nothing better than dynamic renaming, which by the way can take into account any optimizations done in the front end, such as macrofusion, idioms, etc.

In regards to decoder "nativity", there is no such thing as native mode, all code is treated the same. The only caveat is how much complexity is added to the frontend decoders to support new instructions. But given the importance of x86-64 performance in the near future, there is no reason to back off from an aggressive implementation.

any chance you can dumb this down for the rest of us?

 

[dmens]

That is not a valid comparison, since I assume vista will sound a lot better, hopefully.

Would moving to software emulation gain anything? No, because hardware support is cheap, well understood, and much easier and less work than software emulation. Plus there's barely any performance to be gained, so what is the point. I personally would like to redo x87, but that's just because it is annoying. FP performance on x86 is limited by resource priorities, not by definitions. In fact, nothing is specification limited, imo. Otherwise, it would've been changed already.

In regards to the blurb on architectural registers, the point is that having a small number of registers (like x86) doesn't really matter that much because everything gets renamed to available physical resources. Plus, a large architectural register set will be more expensive to rename.

 

[dexvx]

Originally posted by: dmens
That is not a valid comparison, since I assume vista will sound a lot better, hopefully.

So basically Vista vastly improves the sound management scheme by screwing over existing products (this is what I'm basically inferring based on the CL posts). Not much different than a new ISA coming in vastly improving performance, while screwing over existing software and requiring a rewrite.

Originally posted by: dmens
Would moving to software emulation gain anything? No, because hardware support is cheap, well understood, and much easier and less work than software emulation. Plus there's barely any performance to be gained, so what is the point. I personally would like to redo x87, but that's just because it is annoying. FP performance on x86 is limited by resource priorities, not by definitions. In fact, nothing is specification limited, imo. Otherwise, it would've been changed already.

Looks like x86-64 already redoes the FPU. I was wrong in the fact that the 8FP Registers are not available. They are in 64bit mode, but the coders are heavily encouraged in using the SSE units instead of the FP ones. So now that you have compilers using SSE registers, what happens to older machines using the FPR's? Their performance drops considerably. There is absolutely no way you use a compiler that takes full advantage of 16 128bit SSE/2 registers and "downgrade" that to 8 32bit FPR's (on old hardware) and not lose performance.

Remember the venerable AthlonXP? It cannot do hardware SSE2. Even old Willamettes are now beginning to smack them around in SSE2 optimized code, something unheard of when they were released.

Originally posted by: dmens
In regards to the blurb on architectural registers, the point is that having a small number of registers (like x86) doesn't really matter that much because everything gets renamed to available physical resources. Plus, a large architectural register set will be more expensive to rename.

I've always thought of having a small register set is hinding performance at the hardware level. You're right in the aspect that the legacy decoders don't take as much % of die space as say a Pentium, but come now, there is a reason for expanding the number of registers physically available to the compiler, and it usually stems from performance at the ISA level.

 

[dmens]

The vista/audio comparison is still invalid, because of the sheer amount of software that would be rendered useless if x86 cpu's stop supporting legacy code.

So you're saying new extensions make old hardware become obsolete faster? Yeah, that's pretty cool, more money to be made. But how is the point relevant to the current discussion, which is focusing on new hardware supporting old code? Either way, defintions cannot limit performance, especially a loose one like x86.

ISA and performance for x86 has been a dead issue for ages, given today's hybrid designs. ucode-driven machines using a loose ISA like x86 are infinitely expandable, imo. better keep the exposed parts simple, which in turn allows more flexibility to implement aggressive designs inside the machine.

 

[kenji4life]

Thanks for being a smartass. I wasn't posting the news because it was 'new' I am well aware of it's age. I posted the article because I thought it might spark more conversation on this topic, it being relevant. But again, thanks for being a smartass.