This is from definition of stg_atomicModifyMutVarzh(): (for threaded runtime) retry: x = StgMutVar_var(mv); StgThunk_payload(z,1) = x; (h) = ccall cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var, x, y); if (h != x) { goto retry; } cas() is defined in includes/stg/SMP.h like this: EXTERN_INLINE StgWord cas(StgVolatilePtr p, StgWord o, StgWord n) { return __sync_val_compare_and_swap(p, o, n); } I think this is a function we want to make sure to inline everywhere, right? It's compiled to a single instruction on my x86_64 Linux laptop. >>> disassemble cas Dump of assembler code for function cas: 0x0000000000027240 <+0>: mov %rsi,%rax 0x0000000000027243 <+3>: lock cmpxchg %rdx,(%rdi) 0x0000000000027248 <+8>: retq End of assembler dump. But it seems like it's not really inlined in Cmm functions: >>> disassemble stg_atomicModifyMutVarzh Dump of assembler code for function stg_atomicModifyMutVarzh: ... 0x0000000000046738 <+120>: callq 0x27240 <cas> ... End of assembler dump. I guess the problem is that we can't inline C code in Cmm, but I was wondering if this is important enough to try to fix maybe. Has anyone here looked at some profiling info to see how much time spent on this cas() call when threads are blocked in `atomicModifyIORef` etc?