gcc/8.2.0 compilation failures

[jgphpc]

Hi,

Trying to compile the .c test files with gcc/8.2.0, i am getting a few failures. I'd like to get your expert feedback about those failures, i.e can we explain those failure because gcc just does not support the directives used in the test code (then i'll be fine with it). If not, we may want to report to the compiler team or adapt the src code.

Undefined references:

• acc_copyin_async

• acc_copyout_async

• acc_create_async

• acc_delete_async

• acc_get_default_async

• acc_memcpy_from_device_async

• acc_set_default_async

• acc_update_device_async

• acc_update_self_async

• acc_free (undef acc_property_free_memory)

• acc_get_property (undef acc_property_free_memory)

• acc_malloc (undef acc_property_free_memory)

For instance:

cc -fopenacc -lm acc_copyin_async.c -o acc_copyin_async

acc_copyin_async.c: In function 'test':
acc_copyin_async.c:19:5: warning: implicit declaration of function 'acc_copyin_async'; 
did you mean 'acc_wait_async'? [-Wimplicit-function-declaration]
     acc_copyin_async(a, n * sizeof(real_t), 1);
     ^~~~~~~~~~~~~~~~
     acc_wait_async
/tmp/ccuvXq81.o:acc_copyin_async.c:function test:
 error: undefined reference to 'acc_copyin_async'

'c' appears more than once

cc -fopenacc -lm acc_wait.c -o acc_wait

acc_wait.c: In function 'test':
acc_wait.c:27:13: error: 'c' appears more than once in data clauses
     #pragma acc parallel present(c[0:n], d[0:n], c[0:n]) async(1) wait(2)

• acc_async_test
• acc_wait

expected ')'

cc -fopenacc -lm parallel_private.c -o parallel_private

parallel_private.c:21:49: error: expected ')' before '[' token
     #pragma acc parallel num_gangs(10) private(c[0:n])

• parallel_firstprivate
• parallel_private

INTERNAL-COMPILER-ERROR

cc -fopenacc -lm acc_memcpy_to_device_async.c -o acc_memcpy_to_device_async

acc_memcpy_to_device_async.c:36:5: warning: 
implicit declaration of function 'acc_memcpy_to_device_async'; 
did you mean 'acc_memcpy_to_device'? [-Wimplicit-function-declaration]
     acc_memcpy_to_device_async(devdata, a, n * sizeof(real_t), 1);
     ^~~~~~~~~~~~~~~~~~~~~~~~~~
     acc_memcpy_to_device
during GIMPLE pass: omplower
acc_memcpy_to_device_async.c:89:13: internal compiler error: Segmentation fault
     #pragma acc update host(hostdata[2*n:n]) async(3)

'tile' argument

cc -fopenacc -lm parallel_loop_tile.c -o parallel_loop_tile

parallel_loop_tile.c:45:40: error: 'tile' argument needs positive integral constant
         #pragma acc parallel loop tile(tile_arg/10, tile_arg, tile_arg*2)

expected clause before finalize

cc -fopenacc -lm exit_data_finalize.c -o exit_data_finalize

exit_data_finalize.c:63:66: error: expected '#pragma acc' clause before 'finalize'
     #pragma acc exit data delete(a[0:n], b[0:n]) copyout(c[0:n]) finalize

array section

cc -fopenacc -lm declare.c -o declare

declare.c:50:29: error: array section in '#pragma acc declare'
   #pragma acc declare copyin(b[0:n])

array section is not contiguous

cc -fopenacc -lm acc_async_test_all.c -o acc_async_test_all

acc_async_test_all.c:25:34: error: array section is not contiguous in 'map' clause
     #pragma acc enter data create(a[0:10][0:n], b[0:10][0:n], 
                                   c[0:10][0:n], d[0:10][0:n], e[0:10][0:n])

'c' undeclared

cc -fopenacc -lm atomic_capture_bitor_equals.c -o atomic_capture_bitor_equals

atomic_capture_bitor_equals.c:54:70: error: 'c' undeclared (first use in this function)
     #pragma acc data copyin(a[0:n]) copy(totals[0:n/10 + 1]) copyout(c[0:n])

[utimatu]

Ok, so I have addressed some of these by making changes to the tests. Thanks again for pointing these things out! The issues in described in the following sections have been fixed in the tests:

• 'c' appears more than once
• 'tile' argument
• array section
• 'c' undeclared

The rest, I believe I have explanations for their failures in some regard:

• Undefined References: GCC has not kept up with the runtime library. It makes sense that these might go longer without being supported since they lose the benefit of using a directive based language. These are not hard to bypass by implementing them yourself by defining each one to be a function that contains the directive form of the runtime routine.

• expected ')': GCC doesn't allow pointer array subsections as a variable in a "var-list". So when it sees the array bounds, it doesn't accept it as a proper argument to the firstprivate/private clauses. The definition for a var is found on line 3418 on page 126 in the 2.7 specifications. The variable is allowed to be a subarray specification. Section 2.7.1 describes data specification in data clauses and suggests in many ways that pointers and arrays are interchangeable in C.

• INTERNAL-COMPILER-ERROR: This shouldn't happen. This is a compiler error, as it itself recognizes. Can't help you much here.

• 'tile' argument: I bring this one up again, because once I made the fix, It still failed to complete in my next run. There might still be an issue in it, but I didn't see anything that was readily apparent.

• expected clause before finalize: This one I don't know. Last time I ran on GCC, it didn't recognize the finalize clause as syntactically valid. It seems to now accept it, but with some bugs. I don't think that my usage of it is incorrect.

• array section is not contiguous: I don't think that this is an issue with me. The data seems to fit the requirements of a multidimensional array as described in lines 1086-1091, page 37 of the 2.7 specifications.