> On 08/02/2016 12:07 PM, Alex Bennée wrote: > > This will work but I wonder if it is time to call it a day for 32 on 64 > > support? I mean all this can be worked around but I wonder if it is > > worth the effort if no one actually uses this combination. > > I've been meaning to bring up exactly this question during the 2.8 cycle. > > Given the changes we want to make with target atomic operations, I think it's > a complete waste of time to continue to support 64-on-32.
The only interesting 32-bit host IMO is arm, and it does support double-word atomics on ARMv7+ (only through ldrdex/strdex if < Cortex-A15). MIPS doesn't, and PPC (POWER8) has 128-bit atomics in 64-bit mode, but lacks 64-bit atomics in 32-bit mode. FWIW, OpenRISC also lacks double-word (64-bit in 32-bit processors, 128-bit in 64-bit processors) atomics. Sad. So if we want to keep 64-on-32 emulation, it might still make sense to restrict it to ARMv7+ and x86 hosts. But otherwise it might be indeed time to say goodbye... Paolo
