[rfc][ARM] Support for safepoint check based on memory protect rather than polling

[Andrew Haley]

On 05/15/2012 10:12 AM, Andrew Dinn wrote:
>> I'm not really convinced that all this magic word stuff is necessary.
>> > I guess if someone other than a safepoint check does fault we'll see
>> > it in the backtrace, so this is OK.  But is it really worth all the
>> > complexity?

> If we remove this check and some bug happens to invalidly read or write
> the safepoint page then we will blithely skip the PC forwards 6 or 8
> bytes and return from the signal handler i.e. trying to continue
> execution from wherever that adjustment takes us. We might be lucky and
> see another SEGV or an  illegal instruction causing a fatal exit.
> However, we might also be unlucky and continue Java execution with
> corrupt Java or VM data. I would not want to debug the latter -- the
> error might manifest many instructions later.

Well, yes.  But we don't check that some rogue code elsewhere hasn't,
say, corrupted the stack either.  What's different about this?

Andrew.