On Mon, Dec 28, 2009 at 11:06:20AM -0800, Richard Henderson wrote: > On 12/24/2009 07:18 AM, Aurelien Jarno wrote: >> My main concern about this patch is that I don't really understand why >> the current fp exceptions, the current rounding mode or flush_to_zero >> mode are stored in FP_STATUS. I think it would be better to have >> dedicated variable(s) in the cpu state structure, as it is done in other >> emulated architectures. > > The internal format for these features is quite different (in terms of > bit ordering) than the native Alpha format. If I were to be forcing the > bits back into the Alpha format after each operation, that would be far > more shifting and masking than only doing the conversion to Alpha format > when the FPCR register is read or written. At present I'm only saving a > value and ORing it back in > > However, if you're convinced I should keep things in Alpha format all > the time, I can rearrange the patch along those lines. >
If the format is too different, you should still use a dedicated variable in the cpu state. Thinking more, there is no guarantee that FP_STATUS is not overriden by another thread (or another CPU in system mode), if it's value is set in one TB, and read back in another TB. Also there is no need to save FP_STATUS and restore it. It should be set to 0, and the result should be ORed with the dedicated variable. -- Aurelien Jarno GPG: 1024D/F1BCDB73 aurel...@aurel32.net http://www.aurel32.net
