Bill Todd (billtodd@metrocast.net) on 6/20/03 wrote:
---------------------------
>David Wang (dwang@realworldtech.com) on 6/20/03 wrote:

>>The "headroom" that may exist in platform performance that could be extracted out
>>by compiler would presumably be larger on the Itanium/EPIC side as compared to the
>>hammer/64 bit side.
>
>Not in the short term that we're talking about. Multiple well-funded and highly-qualified
>compiler teams have been working on EPIC compilers for half a decade or more now,
>so while it's not unreasonable to speculate that there may still be a good deal
>more compiler potential to mine all the low-hanging fruit was picked quite some
>time ago and that potential is likely to be realized only gradually (if at all)
>over time. By contrast, my impression is that the AMD64 compiler efforts started
>considerably more recently and have been far less well funded: AFAIK we have only
>a single 64-bit SPECint result to look at so far (from the gcc compiler; nothing
>from either the Portland or Microsoft compilers), not a couple of generations' worth
>on multiple compilers as Itanic has.

People have been working on various incarnations of EPIC/VLIW compilers for a while, and there are still some serious "issues" outstanding, but there's one thing that can't be beat that's different from the past year alone as opposed to all the years before, and that's the availability of actual hardware to run the compiler on. You can work on all the compiler issue you want, and you can check the assumptions against a simulator, perhaps even one that's cycle accurate, but accurate simulators are at least a couple of orders of magnitude slower than actual hardware.

The result is that if you wanted to check the net effect of a few compiler switches on the SPEC suite. You can simulate the entire run, and that will take perhaps several compute-months. (or you just take a short run and pretend that it's just as good as the entire run)

If you had actual hardware, your assumptions could be verified or rejected on actual hardware in a matter of few compute-hours. That has to be significant along the way somewhere.