On 5/29/2012 2:09 PM, Joerg Schilling wrote:
Bob Friesenhahn<[email protected]> wrote:
On Tue, 29 May 2012, Joerg Schilling wrote:
You seem to miss, that compiling Schillix-ON on Intel/AMD based HW takes 18 Ghz
hours while compiling the same on Sparc only takes 16 GHz hours. So Sparc is
still more effective per clock cycle than intel/amd.
SPARC is still a good architecture and surely better designed from the
start than x86, which has an architecture which is traceable across 8,
16, 32, and now 64-bit CPUs. Your evaluation criteria is highly
dependent on how much work the compiler needs to do to produce good
code.
I thought that optimizing for Sparc is a harder job than doing the same for
Intel. Do you believe otherwise?
Jörg
So far as my experience in the Java VM department goes (i.e the C++ VM
itself), NO, it was simpler to optimize for SPARC for most operations,
as SPARC v9 is significantly less crufty and idiosyncratic than the x86
ISA. That said, the current generation of SPARCv9 CPUs lack many of
the "advanced" features of the modern x86_64 chips, which slow them down
significantly (i.e. fewer specialized Ops to do certain specific things
in SPARC).
Fundamentally, SPARCv9 is pretty old (and, honestly, in need of an
update by now, as far as being an ISA) - the various SIMD extensions
aren't really sufficient to keep up with the rate that AMD/Intel keep
adding stuff.
The result is that SPARC v9 and the chips implementing it are
significantly "cleaner" for compiler designers, but are "feature-poor".
It's far easier to write a good SPARC compiler than a good x86_64
compiler, but a good SPARC compiler will be outperformed by a good
x86_64 compiler. GCC is a shining example here: the gcc SPARC port
produced excellent-quality optimized code back in the early-1990s, while
it didn't get decent x86 code generation until 2000, and it *really*
only started supporting well-optimized code for post-PentiumPro stuff
after 2005.
The bottom line is resources - Intel and AMD simply had probably two or
even three orders of magnitude more cash spent to produce good compilers
than Sun did on SPARC compilers, and enabled Intel/AMD to produce
significantly more complex compilers which took full advantage of all
the extra features that weren't in SPARC.
Overall, writing a well-done SPARCv9 compiler is quite easy for most
compiler folks, but a compiler wizard can write a x86/x86_64 compiler
that is significantly more efficient than a good SPARC compiler.
-Erik
-------------------------------------------
illumos-discuss
Archives: https://www.listbox.com/member/archive/182180/=now
RSS Feed: https://www.listbox.com/member/archive/rss/182180/21175430-2e6923be
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21175430&id_secret=21175430-6a77cda4
Powered by Listbox: http://www.listbox.com
