Some bugs in cilkplus-gcc

[Niklas_B_]

Hello,

I'm using the cilkplus GCC extensions and ran into a few smallish bugs. All of them are reproducible with the most recent cilkplus-gcc 'gcc version 4.9.0 20130520 (experimental) (GCC)'

$ g++ -v
Using built-in specs.
COLLECT_GCC=g++
COLLECT_LTO_WRAPPER=/usr0/home/nbaumstark/cilkplus-install/bin/../libexec/gcc/x86_64-unknown-linux-gnu/4.9.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: /home/nbaumstark/cilkplus-gcc/configure --prefix=/home/nbaumstark/cilkplus-install --enable-languages=c,c++ --disable-multilib
Thread model: posix
gcc version 4.9.0 20130520 (experimental) (GCC)

1. cilk_for pollutes local namespace in template functions. I think I saw a report of this one somewhere, but can't find it anymore and it's not been fixed:

template <typename T>
void test() {
  int test = 0;
  cilk_for (int i = 0; i < 10; ++i) test += i;
  cilk_for (int i = 0; i < 10; ++i) test += i;
}

int main() {
  test<int>();
}

Output:

$ g++ -lcilkrts -fcilkplus -mcx16 -O2 -Wall loop-var-pollution.cpp
loop-var-pollution.cpp: In instantiation of ‘void test() [with T = int]’:
loop-var-pollution.cpp:9:13:   required from here
loop-var-pollution.cpp:5:17: error: redeclaration of ‘int i’
   cilk_for (int i = 0; i < 10; ++i) test += i;
                 ^
loop-var-pollution.cpp:4:17: error: ‘int i’ previously declared here
   cilk_for (int i = 0; i < 10; ++i) test += i;
                 ^

2. unused variable warnings for variables that are used in a cilk_for header:

// compile with -Wall
void test() {
  int len = 10;
  int test = 0;
  cilk_for (int i = 0; i < len; ++i) test += i;
}

int main() {
  test();
}

Output:

$ g++ -lcilkrts -fcilkplus -mcx16 -O2 -Wall loop-var-unused.cpp
loop-var-unused.cpp: In function ‘void test()’:
loop-var-unused.cpp:3:7: warning: variable ‘len’ set but not used [-Wunused-but-set-variable]
   int len = 10;

3. Type errors with -std=c++11 and cilk_spawn with a method of the current instance:

// compile with -Wall -std=c++11
template <typename T>
struct X {
  int test;
  X() : test(0) {}
  void foo() { test++; }
  void bar() {
    cilk_spawn foo();
    foo();
    cilk_sync;
  }
};

int main() {
  X<int>().bar();
}

Output:

$ g++ -lcilkrts -fcilkplus -mcx16 -O2 -Wall -std=c++11 this-method-spawn.cpp
this-method-spawn.cpp: In instantiation of ‘void X<T>::bar() [with T = int]’:
this-method-spawn.cpp:15:16:   required from here
this-method-spawn.cpp:8:20: error: cannot convert ‘X<T>::foo<int>’ from type ‘void (X<int>::)()’ to type ‘void (X<int>::*)()’
     cilk_spawn foo();
                    ^
this-method-spawn.cpp:8:20: error: must use ‘.*’ or ‘->*’ to call pointer-to-member function in ‘_cilk_lmda_fn_var_000000 (...)’, e.g. ‘(... ->* _cilk_lmda_fn_var_000000) (...)’
cc1plus: warning: unused variable ‘sf_02’ [-Wunused-variable]

4. Warning about unused variable "sf_*". See example above. With -std=c++03, the error doesn't occur, but the warning does.

The source code is attached as a tarball.

Best regards,

Niklas Baumstark

[TimP]

Maybe you should try building the current gcc source e.g. svn co svn://gcc.gnu.org/svn/gcc/trunk

and file any problem reports on gcc bugzilla; you could mention the pr numbers here.

I'll have to confess I have little personal interest in the many lto cases which fail their standard testsuite.

I'll try your cases later, if I can figure out easily how to run.  gcc cilkplus doesn't seem well enough documented to me.

I must deal with the snow now.

[Niklas_B_]

Hi,

okay, problems 3 and 4 seem to be fixed in GCC trunk, nevermind. Actually a workaround for 3 in cilkplus-gcc is to add an explicit this:

cilk_spawn this->foo()

cilkplus-gcc might be badly documented, but GCC trunk seems to only partially implement cilkplus as of this date, if I'm not mistaken. For example, cilk_for seems to be implemented, so I can't try to reproduce problems 1 and 2. Does anyone have any knowledge of the status of GCC trunk w.r.t cilk? In particular, is the performance comparable to other cilk implementations? And is there a trick to work around the absence of the cilk_for keyword to use parallel for loops?

Best, Niklas

[Balaji_I_Intel]

Hi Niklas,

     _Cilk_for is not merged into the trunk yet. It is currently under review by the GCC reviewers. It is a different implementation and I am hoping that you problems outlined in 1 and 2 should disappear when it gets into trunk.

Thanks,

Balaji V. Iyer.

[Niklas_B_]

Okay, that sounds reasonable. Since there are workarounds for all issues, I'm going to stick with cilkplus-gcc until _Cilk_for is merged into the GCC trunk.

Thanks guys!