[llvm-dev] Intel MPX support (instrumentation pass similar to gcc's Pointer Checker)

[John Criswell via llvm-dev]

On 2/3/16 6:27 AM, Dmitrii Kuvaiskii via llvm-dev wrote:
> I continue playing with Intel MPX and its support in modern compilers.
> All experiments were done on the Alienware (Dell) 15 R2, Ubuntu 15.10
> (linux 4.2.0), gcc version is 5.2.1, icc version 2016.1.150. The
> benchmark suite is PARSEC 3.0, all versions with 1 thread and default
> configs.
>
> As I described previously, PointerChecker in gcc produces very
> inefficient code. My experiments show overheads over native of up to
> 9.5X (on "raytrace"), with common overheads of 3X ("bodytrack",
> "fluidanimate", "streamcluster"). At the same time, AddressSanitizer
> performs much better -- 1.3X on "raytrace", 1.7X on "bodytrack" and so
> on.
>
> Recently I played with MPX support on Intel C/C++ Compiler (icc). This
> implementation looks *much* better, with the following example
> overheads: 1.2X on "raytrace", 1.25X on "bodytrack", 1.08X on
> "streamcluster". So the common overheads are in the range of 15%-25%!
>
> Please note that gcc-mpx and gcc-asan versions were compared against
> gcc-native, and icc-mpx version was compared against icc-native.
>
> We would like to compile a small technical report with all our
> measurements (performance and memory  overhead) and put it online.
> We'll do it in the near future, I will write an update here when it's
> done. Please tell me if anyone is interested in any specific
> benchmarks (I want to test PARSEC and some case-studies: PostgreSQL,
> Memcached, SQLite3). Any feedback is welcome.

Two comments.

First, it would be interesting to compare SoftBound to MPX to see how 
the two compare.  To the best of my understanding, MPX is intended to be 
a hardware implementation that does what SoftBound does.

Second, when you write your report, please keep in mind that what Asan 
and MPX do are not really the same thing.  MPX is designed to track 
bounds information for each pointer while Asan only tracks where memory 
objects are located.  As a result, their security guarantees are not the 
same: MPX should be able to catch pointers that jump from one memory 
object into another while Asan does not (Asan distributes memory objects 
far away from each other in the address space so that out of bounds 
accesses are likely to point into unallocated memory).  The difference 
is subtle, but I think it's important.

Regards,

John Criswell


>
> On Thu, Jan 28, 2016 at 8:29 PM, Kostya Serebryany <kcc at google.com> wrote:
>> I've recently played with the GCC implementation of pointer checker on a
>> real hardware,
>> my recent impressions are here:
>> https://github.com/google/sanitizers/wiki/AddressSanitizerIntelMemoryProtectionExtensions
>> (there is also some old pre-hardware content).
>>
>> In short, I totally agree with what David says above: MPX is a disaster.
>> (Usual disclaimer: my opinion here is too biased)
>>
>> I am glad that LLVM already has the support for MPX instructions, but I see
>> no good reason to add the MPX-based checker to LLVM.
>> Yes, it will allow us to detect intra-object overflows, something that asan
>> can not do by default, but it's not worth the extreme complexity of the
>> MPX-based checker.
>>
>> --kcc
>>
>> On Thu, Jan 28, 2016 at 9:14 AM, David Chisnall via llvm-dev
>> <llvm-dev at lists.llvm.org> wrote:
>>> On 28 Jan 2016, at 17:03, Dmitrii Kuvaiskii via llvm-dev
>>> <llvm-dev at lists.llvm.org> wrote:
>>>> Interestingly, my
>>>> preliminary experiments indicate that adding MPX bounds checking via
>>>> Pointer Checker in gcc is usually slower than using software-only
>>>> AddressSanitizer.
>>> This corresponds with other results that I have seen.  The last time I
>>> looked at the output from gcc, it also did not generate pointer updates that
>>> were safe in the presence of concurrency (they must be bracketed in
>>> transactions if you want the MPX metadata and the pointer updates to be
>>> atomic) and the overhead of this is likely to be even more.
>>>
>>> I am particularly impressed with Intel for creating a hardware
>>> implementation that is both slower than a software-only version and can not
>>> (due to its fail-open policy being embedded in the hardware) be used for
>>> security.
>>>
>>> David
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>
>


-- 
John Criswell
Assistant Professor
Department of Computer Science, University of Rochester
http://www.cs.rochester.edu/u/criswell