That's a great list of subtle issues (I can add that atomicity guarantees are different for non-aligned access sometimes) - it would be fantastic if we could amend it with actual data, say SunSpider results and/or DOM performance tests on Intel Mac and/or Windows.
If we got an improvement on SunSpider/Mac, we wouldn't have to discuss processors running in non-native endianness modes :-) - WBR, Alexey Proskuryakov On Sep 4, 2008, at 1:24 PM, Paul Pedriana wrote: > This is a hardware issue and can vary between different versions. > Some hardware of course generates an unhandled exception on > unaligned access; some hardware generates an internally handled > exception and restarts the access with a different and more > expensive pathway or with microcode; some hardware efficiently > handles it as it goes and has minimal impact. Sometimes this > behavior is different depending on whether the hardware is in big > endian vs. little endian mode. A new generation of the same > processor family or a version by a different company may act > differently than another. Finally, if that memory is accessed by a > different means than the CPU (e.g. via DMA or the video bus), then > it may require alignment where it doesn't otherwise. > > What I'm getting at here is that I think we would need to be careful > about disabling it for any hardware, including even forgiving > hardware like x86. > > > >> On Sep 4, 2008, at 12:20 PM, Paul Pedriana wrote: >> >>> I'll make a patch and attach it to >>> <https://bugs.webkit.org/show_bug.cgi?id=16925 >>> >, if that's OK. >> >> That would be great! >> >> One thing I'm not sure about is whether we want to enforce >> alignment on platforms that don't require it - performance testing >> should answer this. >> >> - WBR, Alexey Proskuryakov >> >> > _______________________________________________ webkit-dev mailing list [email protected] http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
