Incorrect outlining for C++ OpenMP code

[ouankou]

Given the input:

#pragma omp target parallel for 
for (int i = 0; i < n; ++i)
    y[i] += a * x[i];

If it's in a C file, the kernel is OK. Variable a is rewritten as *ap__ and used correctly.

__global__ void OUT__1__7822__axpy_omp__48__kernel__(long *np__,double *ap__,double *_dev_x,double *_dev_y) {
...
for (i = _dev_lower; ((long )i) <= _dev_upper; i += 1) {
        _dev_y[i] +=  *ap__ * _dev_x[I];    // *ap__ is used. Correct. 
      }
}

However, if the same OpenMP code is in a C++ file. The generated kernel would be:

__global__ void OUT__1__7822__axpy_omp__48__kernel__(long *np__,double *ap__,double *_dev_x,double *_dev_y) {
...
for (i = _dev_lower; ((long )i) <= _dev_upper; i += 1) {
        _dev_y[i] +=  ap__ * _dev_x[i];    // ap__ is used. Incorrect.
      }
}

Please notice that ap__ is used as a pointer without dereferencing.

I've checked the lowering and outlining modules. So far, no C++-specific handler has been found.

[yanyh15]

You can use the outline tutorial to isolate the outline problem, for example:

rexompiler-build/tutorial/outliner$ ./outlineTutorial -rose:outline:enable_debug -rose:outline:enable_classic -rose:outline:line "12" -c /home/yyan7/compiler/rexompiler/tutorial/outliner/inputCode_OutlineLoop2.c