>-----Original Message----- >From: Tim Ellison [mailto:[EMAIL PROTECTED] >Sent: Wednesday, August 09, 2006 10:07 PM >To: harmony-dev@incubator.apache.org >Subject: Re: [general] platform support > >(Your post had weird quoting, I've tried to fix it up in my reply) > >Rana Dasgupta wrote: >> On 8/9/06, Tim Ellison <[EMAIL PROTECTED]> wrote: >>> Yes, it is the question you also pose elsewhere -- can we have a binary >>> that either (a) uses the lowest common denominator of the different >>> windows platforms API without incurring an undue penalty performance, or >>> (b) performs runtime checks and picks the best available APIs. >>> >> There are distinct approaches as I understand it. >> >> One option is a single binary image that contains code that supports >> multiple platforms seperately by doing a dynamic check for platform. >> Though less pernicious than a least common denominator approach, >> these runtime checks are not healthy for a binary image that targets >> performance. So if our ideal platform were XinXP, we would incur a >> penalty repeatedly when running with it to accomodate the fact that >> this binary could have also run on W2k. > >But there are degrees to which this is done too right? Somewhere along >the spectrum from a start-up check that chooses between the winxp.dll >and win2k.dll, to repeatedly choosing between any number of possible OS >function calls.
I also thought about the approach of using just separate DLLs with the same external API. At start-up the VM chooses one of them, and then forgets about platform-specific differences. This might be a rather good solution. Regards, Alexey. > >Oh, and I'm assuming that we are leaving the jitted code out of this. >Of course the jit will know what platform it is targeting and can >generate the code appropriately. So we are discussing the performance >of the interpreter and the compiler itself. > >> The second option is to use a least common denominator approach where >> we use code/functionality that is only available on the least >> platform. This is not a good idea for obvious reasons. For example it >> is not a good idea not to use the excellent vectored exception >> handling on WinXP and Win2003( which intentionally share the same >> debug and kernel codebases )If this were not, we would be writing >> code for DOS only. > >Again, there may be cases where you may well choose the least common >denominator solution because it is 'good enough' and the overhead >elsewhere (testing etc.) is not worth the gain found here. > >Is vectored exception handling a slam dunk case for making the binary >winxp only? I don't know -- what would happen if we didn't use it? >Where is the example in the current code that makes ensuring it runs on >W2K unpalatable? > >> The third is to have a single codebase with the right _WIN32_WINNT >> guards to distinguish platform specific code, and build seperate >> distributions for seperate platforms. This is the most performance >> friendly. It has a building cost, but the major overhead is not >> building, but testing. If we were to support a platform, we would >> need to test on it anyway. > >Agree, so there is a balance to be struck. But I'm guessing from you >descriptions that you favour this approach of multiple distributions for >different OS releases. > >Regards, >Tim > >-- > >Tim Ellison ([EMAIL PROTECTED]) >IBM Java technology centre, UK. > >--------------------------------------------------------------------- >Terms of use : http://incubator.apache.org/harmony/mailing.html >To unsubscribe, e-mail: [EMAIL PROTECTED] >For additional commands, e-mail: [EMAIL PROTECTED] -- Alexey A. Ivanov Intel Middleware Product Division --------------------------------------------------------------------- Terms of use : http://incubator.apache.org/harmony/mailing.html To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
