One of the biggest problems with this style of "scan all memory" GC is that you have to stop all threads (flush all pointers to memory) to perform a collection. Pointers held in registers in other active threads are inaccessible to the GC thread.
Um no. First, it doesn't "scan all memory". It scans the root and fans out. You can do read/write barriers and such to avoid blocking GC. It doesn't matter what value registers have in them. What matters is what is on the stack and what is reachable from said stack. If you have control of malloc, you can be far more clever than what you are describing. Ruby had a crummy implementation of GC, but you'll notice, for example, that most JVM's do "scan all memory" without forcing all active threads to be blocked at once.
Failing to stop the world (or ensure all pointers are in memory rather than registers) can result in live objects being collected. This happened to Ruby at one point: http://timetobleed.com/the-broken-promises-of-mrireeyarv/
Edit: This boils down to breaking the compiler's model of memory as specified by the C standard; all bets are off.