On 08/11/2017 01:29 PM, Alistair Francis wrote: > On Fri, Aug 11, 2017 at 1:24 PM, Richard Henderson > <richard.hender...@linaro.org> wrote: >> On 08/11/2017 01:13 PM, Alistair Francis wrote: >>>>> + tcg_gen_ext_i64(val, val, memop); >>>> >>>> What is this addition intended to accomplish? Because of the position >>>> within >>>> the code, you know that memop contains MO_64, so that this is a no-op. >>> >>> This is when size == 2 so it's a 32bit operation so memop contains MO_32. >> >> It's a paired 32-bit operation, so we're operating on a 64-bit quantity. So >> extending from 32-bits would be actively wrong. > > From what I can see though, the 32bit memop is carried into the > tcg_gen_atomic_cmpxchg_i64() call so the value returned to tmp is > masked by the 32bit operation. > > Is passing down MO_32 into tcg_gen_atomic_cmpxchg_i64() wrong then as > it ends up as a 64-bit operation?
If we're passing MO_32 down to cmpxchg_i64 for this case, you have indeed found a bug. I'll investigate this further on Monday. r~
