[PATCH] asan: optimization experiments

[Dmitry Vyukov]

In http://reviews.llvm.org/D8198#140697, @kcc wrote:

> > > Experiment ID is a must, otherwise you will need to ask users to send you a full reproducer (something they won't be able to do), and then you will have to stare at the code trying to figure out what llvm byte code it produced and how asan pass handled it. You won't be able to just compile it, because it uses a custom build system and the bug is not reproducible without a non-trivial input.
>
> > 
>
>
> I don't understand that. If we have just one experiment at any given time -- ID is completely useless. 
>  If we have multiple experiments in flight it may be a bit useful, but we will still need reproducer to be able to perform complete analysis. 
>  The downside of Exp parameter is added complexity, code size and performance loss (probably not huge though)


Even if you are doing a single change, you still may want several experiments within it: e.g. several levels of aggressiveness and/or application to stack, globals, generic pointers. Doing it in several multi-month consecutive experiments is unbearable.
For example, in the optimization I've slightly changed handling of globals as well. After the experiment I want to get info for stack/globals separately.


http://reviews.llvm.org/D8198

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/