On 12/05/2016 03:09 AM, Alex Bennée wrote: > > Alex Bennée <alex.ben...@linaro.org> writes: > >> While testing rth's latest TCG patches with risu I found ldaxp was >> broken. Investigating further I found it was broken by 1dd089d0 when >> the cmpxchg atomic work was merged. > > CC'ing Paolo/Richard > > Do you guys have any final TCG fixes planned for 2.8 that can take this > fix for the ldaxp regression?
I don't have any pending patchs for 2.8, no. >> As part of that change the code >> attempted to be clever by doing a single 64 bit load and then shuffle >> the data around to set the two 32 bit registers. It's not trying to be "clever", it's trying to be correct, giving an atomic 64-bit load. The fact that we didn't attempt an atomic 128-bit load here for the 64-bit pair is a bug. We have support for that in helper_atomic_ldo_le_mmu; I'm not sure why I didn't use that here. >> As I couldn't quite follow the endian magic I've simply partially >> reverted the change to the original code gen_load_exclusive code. This >> doesn't affect the cmpxchg functionality as that is all done on in >> gen_store_exclusive part which is untouched. We'll have to fix it properly eventually. But perhaps 2.9 is soon enough, since that's when mttcg will go in. r~
