PGF90-F-0000-Internal compiler error. Errors in Lowering 2

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi Mat and other developers,

I’m back. I don’t understand what the following errors mean. I have tried PGI 7.2.5 and 8.0.5. Both report the errors.

8.0.5:

FC Halos_Numbering.F90
/opt/cray/xt-asyncpe/3.0/bin/ftn: INFO: linux target is being used
Lowering Error: bad ast optype in expression [ast=1454,asttype=12,datatype=0]
Lowering Error: bad ast optype in expression [ast=1454,asttype=12,datatype=0]
PGF90-F-0000-Internal compiler error. Errors in Lowering 2 (Halos_Numbering.F90: 111)
PGF90/x86-64 Linux 8.0-5: compilation aborted
gmake[1]: *** [Halos_Numbering.o] Error 1
make: *** [lib/libdfluidity.a] Error 2

7.2.5:

FC Halos_Numbering.F90
/opt/cray/xt-asyncpe/3.0/bin/ftn: INFO: linux target is being used
Lowering Error: bad ast optype in expression [ast=1454,asttype=12,datatype=0]
Lowering Error: bad ast optype in expression [ast=1454,asttype=12,datatype=0]
PGF90-F-0000-Internal compiler error. Errors in Lowering 2 (Halos_Numbering.F90: 111)
PGF90/x86-64 Linux 7.2-5: compilation aborted
gmake[1]: *** [Halos_Numbering.o] Error 1

The code around line 111 in the .F90 is

88 #ifdef HAVE_MPI
89 integer :: communicator, i, ierr, nowned_nodes, nprocs, rank, count, node_count
90 integer, dimension(:), allocatable :: receive_types, requests,&
91 & send_types, statuses
92 logical, dimension(:), allocatable :: local_nodes
93
94 ewrite(1,*) “Creating universal numbering for a general order halo”
95
96 nprocs = halo_proc_count(halo)
97 communicator = halo_communicator(halo)
98 nowned_nodes = halo_nowned_nodes(halo)
99
100 ! Calculate the base universal node number for the owned nodes. The i th
101 ! owned node then has universal node number equal to the base + i.
102 call mpi_scan(nowned_nodes, halo%owned_nodes_unn_base, 1, getpinteger(), MPI_SUM, communicator, ierr)
103 assert(ierr == MPI_SUCCESS)
104 halo%owned_nodes_unn_base = halo%owned_nodes_unn_base - nowned_nodes
105
106 node_count = nowned_nodes + halo_all_receives_count(halo)
107
108 allocate(local_nodes(node_count))
109 local_nodes=.true.
110 assert(max_halo_receive_node(halo) <= node_count)
111 do i = 1, nprocs
112 local_nodes(halo_receives(halo, i)) = .false.
113 end do
114
115 allocate(halo%gnn_to_unn(node_count))
116 halo%gnn_to_unn=-1
117
118 count=halo%owned_nodes_unn_base

The error may refer to somewhere nearby. Is there any clue?

This is a big package. I can pass Halos_Numbering.F90 or the whole package to you.

Thanks.

2

Hi yus,

Unfortunately, this is an error with the compiler. Please send Halos_Numbering.F90 to PGI customer service (trs@pgroup.com) along with any required module source files and the command line options you use. We’ll take a look and hopefully get the problem resolved shortly.

Thanks,
Mat

3

Mat:

I’ve sent the whole package to your site. Have you got it?

4

I did and was able to put together a example reproducing the problem this morning. I have file TPR#15998 and have sent it to our engineers.

Thanks,
Mat

5

Is the fix to this bug in PGI release plan?

6

Hi Yus,

It doesn’t look like the fix will be make into this weeks 9.0-3 release. Hopefully by 9.0-4 in September.

  • Mat
7

Hi Mat,

Thanks for the news.

8

Hi yus,

The engineer who is looking at TRP#15998 sent me this update.

Isolated the errors to two calls to mpi_type_create_indexed_block()
in Halos_Numbering.F90:create_global_to_universal_numbering_order_general().
In each case, the compiler bug is with processing the 3rd argument, an array
expression that corresponds to an adjustable array formal argument:

call mpi_type_create_indexed_block(halo_send_count(halo, i), 1, &
& halo_sends(halo, i) - lbound(halo%gnn_to_unn, 1), & <<<-- BUG
& getpinteger(), send_types(i), ierr)

and

call mpi_type_create_indexed_block(halo_receive_count(halo, i), 1, &
& halo_receives(halo, i) - lbound(halo%gnn_to_unn, 1), & <<<-- BUG
& getpinteger(), receive_types(i), ierr)

A workaround is to use a temp for the lbound expression in the third
argument, e.g.,

  • Mat
9

Hi Mat,

Is this problem fixed in the new release?

Thanks.

yus

10

Hi yus,

Yes, TPR#15998 is fixed as of the 9.0-4 release.

  • Mat
11

Many thanks for the quick fix.