Segmentation violation/assertion violation when using stateful functors in interpreted mode

[markusscherer]

Hi!

Problem: When trying to use stateful external functors with the interpreted mode, souffle crashes.

How to Reproduce:

# build souffle according to documentation
# add test.dl and functors.cpp to souffle's root folder
# compile functors.cpp
 g++ -I src/include/ functors.cpp -std=c++17 -c -fPIC -o functors.o
 g++ -shared -o libfunctors.so functors.o
 # execute souffle
./build/src/souffle test.dl

where test.dl contains

.functor mycat(a:symbol, b:symbol):symbol stateful

.decl A(S1: symbol)
.decl B(S1: symbol)

.output A
.output B

A("AA").
A("BB").
A("CC").

B(@mycat(X,Y)) :- A(X), A(Y).

and functors.cpp contains

#include "souffle/SymbolTable.h"
#include "souffle/RecordTable.h"

extern "C" {
 souffle::RamDomain mycat(souffle::SymbolTable* symbolTable, souffle::RecordTable* recordTable,
        souffle::RamDomain arg1, souffle::RamDomain arg2) {
    std::cout << "hallo 1" << std::endl;
    assert(symbolTable && "NULL symbol table");
    std::cout << "hallo 2" << std::endl;
    assert(recordTable && "NULL record table");
//    std::cout << "HHH " << symbolTable->encode("Hall") << std::endl;
    std::cout << "hallo 3" << std::endl;
    const std::string& sarg1 = symbolTable->decode(arg1);
    std::cout << "hallo 4" << std::endl;
    const std::string& sarg2 = symbolTable->decode(arg2);
    std::cout << "hallo 5" << std::endl;
    std::string result = sarg1 + sarg2;
    std::cout << "hallo 6" << std::endl;
    return symbolTable->encode(result);
 }
}

output:

hallo 1
hallo 2
hallo 3
souffle: src/include/souffle/datastructure/ConcurrentFlyweight.h:290: const Key& souffle::ConcurrentFlyweight<LanesPolicy, Key, Hash, KeyEqual, KeyFactory>::fetch(souffle::ConcurrentFlyweight<LanesPolicy, Key, Hash, KeyEqual, KeyFactory>::lane_id, souffle::ConcurrentFlyweight<LanesPolicy, Key, Hash, KeyEqual, KeyFactory>::index_type) const [with LanesPolicy = souffle::SeqConcurrentLanes; Key = std::__cxx11::basic_string<char>; Hash = std::hash<std::__cxx11::basic_string<char> >; KeyEqual = std::equal_to<std::__cxx11::basic_string<char> >; KeyFactory = souffle::details::Factory<std::__cxx11::basic_string<char> >; souffle::ConcurrentFlyweight<LanesPolicy, Key, Hash, KeyEqual, KeyFactory>::lane_id = long unsigned int; souffle::ConcurrentFlyweight<LanesPolicy, Key, Hash, KeyEqual, KeyFactory>::index_type = long unsigned int]: Assertion `Idx < SlotCount.load(std::memory_order_relaxed)' failed.
Aborted (core dumped)

The outputs indicates, that the libfunctors.so is correctly loaded, but that is not possible to reenter the souffle code again.

I instrumented souffle and die pointers to the symbol and record table are passed correctly (i.e. have the same value on the souffle side as on the library side). It makes a difference if encode or decode is used. When when enabling the call to encode in functor.cpp, the output is:

hallo 1
hallo 2
Segmentation violation signal in rule:
B(@mycat(X,Y)) :-
   A(X),
   A(Y).
in file /home/markus/souffle/test.dl

The error also appears with

./build/src/souffle test.dl -c

What works is:

./build/src/souffle test.dl -g test.cpp
g++ -Isrc/include test.cpp -L. -lfunctors -std=c++17 -o test
LD_LIBRARY_PATH=. ./test

but for our use case (for now: executing many smallish test cases) this is unfortunately not too practical. I thought that maybe this has to do with some multithreading problems, but the error also get thrown with -j1.

I first encountered the error on tag 2.2 (and don't no if it ever worked like that) but just reproduced it on a fresh Ubuntu VM with current master (89aae85).

Am I using stateful functors wrong or is this a bug?

[quentin]

Hi,
Can you make sure the functor library is built with openmp support if the souffle interpreter was too ?
Try adding -fopenmp to the g++ command.

[markusscherer]

Indeed, that did the trick! Since that is currently not mentioned in the documentation's chapter on stateful functors: shall I add a sentence or two about compiling them correctly?

[b-scholz]

Please, could you update the documentation? Many thanks in advance.

[quentin]

I believe we could have some kind of souffle-compile-functors similar to souffle-compile.

Or alternatively we could isolate the SymbolTable interface from the implementation, so that functors only have to know about the interface. I think it's already what was done to isolate the RecordTable.

[b-scholz]

That would be great!

[b-scholz]

I have added a sentence to the documentation about functors.

I thought more about the script. I feel this is tricky because usually, users will have external dependencies. Therefore, I would prefer that build scripts are user-provided.

[b-scholz]

I like the idea of separating the interface. Perhaps we can raise a new issue for this.