... of course, I immediately found two problems by inspection. (1) Accidentally dropped the barrier for sync_test_and_set when used on post-v7. Given that we have only SImode atomic_exchange, and we do have QImode sync_test_and_set, we will probably try to use LDSTUB for the c++ atomic_bool_test_and_set, which does want to modify a single byte.
(2) We would fail to emit anything when compile-time barriers are still desired, and if the user does something silly like -mcpu=v7 -mmmodel=rmo then we'd wind up with unrecognizable insns. Full sparc64-linux testing still underway; I've pushed these two patches to the git branch previously mentioned. r~ Richard Henderson (2): fixup: emit barrier for sync_test_and_set. fixup: Handle barriers for v7, and compile barriers. gcc/config/sparc/predicates.md | 4 ++++ gcc/config/sparc/sparc.c | 14 +++++++++----- gcc/config/sparc/sync.md | 16 ++++++++++++++++ 3 files changed, 29 insertions(+), 5 deletions(-) -- 1.7.4.4
