[llvm-dev] RFC: EfficiencySanitizer

[Derek Bruening via llvm-dev]

On Wed, Apr 20, 2016 at 2:14 AM, Adam Nemet <anemet at apple.com> wrote:

> Interesting idea!  I understand how the bookkeeping in the tool is similar
> to some of the sanitizers but I am wondering whether that is really the
> best developer’s work-flow for such a tool.
>
> I could imagine that some of the opportunities discovered by the tool
> could be optimized automatically by the compiler (e.g. temporal loads, sw
> prefetching, partitioning the heap) so feeding this information back to the
> compiler could be highly useful.  I am wondering whether the PGO model is
> closer to what we want at the end.  The problem can also be thought of as a
> natural extension of PGO.  Besides instrumenting branches and indirect
> calls, it adds instrumentation for loads and stores.
>

It would be great to automatically apply the results of the tools, but we
do not think that this is straightforward for enough cases up front.  For
the cache fragmentation tool, automatically applying data structure field
reordering (or splitting or peeling) generally requires whole-program
compilation, which is not always available and currently does not scale up
to the size of applications we would like to target.  The working set tool
is not a candidate for automated action.  Acting on dead stores typically
requires programmer analysis to confirm that there is not some non-executed
path where the store is not actually dead.

We would like to start with a standalone sanitizer usage model providing
developer feedback.  How about if we start with that and in the future we
can revisit whether some subset of the results can be acted on in an
automated fashion?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160420/57e8ec56/attachment.html>