On 07/19/2018 05:54 AM, Stefan Markovic wrote: > From: Yongbok Kim <yongbok....@mips.com> > > Implement nanoMIPS LLWP and SCWP instruction pair. > > Signed-off-by: Yongbok Kim <yongbok....@mips.com> > Signed-off-by: Aleksandar Markovic <amarko...@wavecomp.com> > Signed-off-by: Stefan Markovic <smarko...@wavecomp.com> > --- > linux-user/mips/cpu_loop.c | 25 ++++++++--- > target/mips/cpu.h | 2 + > target/mips/helper.h | 2 + > target/mips/op_helper.c | 35 +++++++++++++++ > target/mips/translate.c | 107 > +++++++++++++++++++++++++++++++++++++++++++++ > 5 files changed, 166 insertions(+), 5 deletions(-)
Hmm. Well, it's ok as far as it goes, but I'd really really like to see target/mips to be updated to use actual atomic operations. Otherwise mips*-linux-user will never be reliable and mips*-softmmu cannot run SMP in multi-threaded mode. While converting the rest of target/mips to atomic operations is perhaps out of scope for this patch set, there's really no reason not to do these two instructions correctly from the start. It'll save the trouble of rewriting them from scratch later. Please see target/arm/translate.c, gen_load_exclusive and gen_store_exclusive, for the size == 3 case. That is arm32 doing a 64-bit "paired" atomic operation, just like you are attempting here. Note that single-copy atomic semantics apply in both cases, so LLWP must perform one 64-bit load, not two 32-bit loads. The store in SCWP must happen with a 64-bit atomic cmpxchg operation. r~
