[Info-vax] Dvorak on Itanic

[Paul]

On 2009-01-30 11:36:51 -0600, JF Mezei <jfmezei.spamnot at vaxination.ca> said:

> Christopher wrote:
>> 
>> IA-64 processors are completely capable of explicit multi-threading.
> 
> But isn't this decision done by the compiler at compile time ?

Not exactly. The compile has to be much smarter than say, an X86 
compiler, but all the
real work takes place in the hardware. The compiler just has to set 
things up the right way,
so to speak.

And in Itanium, the compiler can reliably do that, which is a BIG 
difference from say, X86.
But the processor is managed by the compiled code (at runtime) just 
like most other processors.

I think you might be thinking about how the instruction level 
parallelism in particular;
the compiler can put a whole bunch of things to do inside *one* 
instruction. It just has to
pack things in there that the computer actually can do at the same 
time, and setup registers
correctly, and so forth. Without going into great detail, that is the 
$0.02 version of
what is "smart" about the compilers.

-Paul


> 
> The Alpha presentations I had mentioned that while IA64 compilers can do
> a good job of optimizing code to run fast on IA64, they cannot predict
> the flow of the code because the compilers don't know what data will be
> processed.
> 
> By having brains in the CPU at run time (as opposed to brains in the
> compiler), real world applications will run much faster.
> 
> But if you have a fixed task (like calculating PI to 50 billion
> decimals), then IA64 will fare well because the code will execute in the
> precise way that the compiler predicted.
> 
> 
> What I find interesting is that from Intel's point of view, the original
> intent of IA64 was to replace the 8086, and as such, a "general" CPU
> would have been what was needed, not a CPU whose philosophy was suited
> more for high performance calculations rather than general computing.
> 
> Note that Alpha also had some serious mistakes early on with the
> engineers having omitted fast access to unaligned bytes (such as wen
> processing strings of characters). But those were implementation
> problems, not architecture/philosophy problems.