I was at David Sehr's presentation at the LLVM conference. There are something like a dozen Googlers working on getting PNaCL working. They are able to get pretty close to the performance of native code, which is pretty enticing. There's still plenty of work left to do though.
One of the problems that needs to be fixed is that LLVM bitcode is kind of bloated, something like 6x as compared with native code. Another issue is that bitcode is not platform independent. Things like pointer sizes and exception handling are dependent on the target you are compiling to. Thus, they have to define a bunch of specifications for PNaCL programs so that they'll run everywhere.
It will definitely be interesting to see if anyone uses it. Like you said, one killer app will do the trick. I'd like to see Google Earth reimplemented for PNaCL, or any Google project for that matter. That would give more confidence that this is a project that is going to be taken seriously within Google.
One of the problems that needs to be fixed is that LLVM bitcode is kind of bloated, something like 6x as compared with native code. Another issue is that bitcode is not platform independent. Things like pointer sizes and exception handling are dependent on the target you are compiling to. Thus, they have to define a bunch of specifications for PNaCL programs so that they'll run everywhere.
It will definitely be interesting to see if anyone uses it. Like you said, one killer app will do the trick. I'd like to see Google Earth reimplemented for PNaCL, or any Google project for that matter. That would give more confidence that this is a project that is going to be taken seriously within Google.