[Bug gdb/26851] New: Make next not skip over GOMP_parallel

[Bug gdb/26851] New: Make next not skip over GOMP_parallel

[vries at gcc dot gnu.org]

https://sourceware.org/bugzilla/show_bug.cgi?id=26851

            Bug ID: 26851
           Summary: Make next not skip over GOMP_parallel
           Product: gdb
           Version: HEAD
            Status: NEW
          Severity: normal
          Priority: P2
         Component: gdb
          Assignee: unassigned at sourceware dot org
          Reporter: vries at gcc dot gnu.org
  Target Milestone: ---

Consider test-case gdb/testsuite/gdb.fortran/omp-step.f90 as found in fedora
patch gdb-bz533176-fortran-omp-step.patch (
https://src.fedoraproject.org/rpms/gdb/blob/HEAD/f/gdb-bz533176-fortran-omp-step.patch
) :
...
    16        use omp_lib
    17        integer nthreads, i, a(1000)
    18        nthreads = omp_get_num_threads()
    19        if (nthreads .gt. 1000) call abort
    20
    21        do i = 1, nthreads
    22            a(i) = 0
    23        end do
    24        print *, "start-here"
    25  !$omp parallel
    26        a(omp_get_thread_num() + 1) = 1
    27  !$omp end parallel
    28        do i = 1, nthreads
    29            if (a(i) .ne. 1) call abort
    30        end do
    31        print *, "success"
    32        end
...

Ideally, a next at line 25 would get us to line 26, the parallel body.

However, that's not the case, we end up at line 28, after the parallel body:
...
$ gdb -q outputs/gdb.fortran/omp-step/omp-step    
Reading symbols from omp-step...
(gdb) b 24
Breakpoint 1 at 0x400910: file omp-step.f90, line 24.
(gdb) r
Starting program: omp-step 
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".

Breakpoint 1, MAIN__ ()
    at omp-step.f90:24
24            print *, "start-here"
(gdb) n
 start-here
25      !$omp parallel
(gdb) n
[New Thread 0x7ffff684a700 (LWP 3404)]
[New Thread 0x7ffff6049700 (LWP 3405)]
[New Thread 0x7ffff5848700 (LWP 3406)]
28            do i = 1, nthreads
(gdb) 
...

The fedora patch contains code to fix this, but it doesn't seem to work.

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug gdb/26851] Make next not skip over GOMP_parallel

[vries at gcc dot gnu.org]

https://sourceware.org/bugzilla/show_bug.cgi?id=26851

Tom de Vries <vries at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |keiths at redhat dot com,
                   |                            |kevinb at sourceware dot org,
                   |                            |palves at sourceware dot org

--- Comment #1 from Tom de Vries <vries at gcc dot gnu.org> ---
This works for me, but has hardcoded x86_64 abi knowledge (reading first arg to
GOMP_parallel from rdi reg at index 5):
...
diff --git a/gdb/infrun.c b/gdb/infrun.c
index 4c2e1f56d2..782c6726c4 100644
--- a/gdb/infrun.c
+++ b/gdb/infrun.c
@@ -6788,6 +6788,22 @@ process_event_stop_test (struct execution_control_state
*ecs)

       if (ecs->event_thread->control.step_over_calls == STEP_OVER_ALL)
        {
+         const char *gomp_parallel = "GOMP_parallel";
+         bool gomp_parallel_p
+           = (ecs->stop_func_name != NULL
+              && strncmp (ecs->stop_func_name, gomp_parallel,
+                          strlen (gomp_parallel)) == 0);
+         if (gomp_parallel_p)
+           {
+             struct regcache *regcache
+               = get_thread_regcache (ecs->event_thread);
+             ULONGEST val;
+             regcache_cooked_read_unsigned (regcache, 5, &val);
+
+             ecs->stop_func_start = val;
+             goto step_into;
+           }
+
          /* We're doing a "next".

             Normal (forward) execution: set a breakpoint at the
@@ -6846,6 +6862,7 @@ process_event_stop_test (struct execution_control_state
*ecs)
          return;
        }

+ step_into:
       /* If we have line number information for the function we are
         thinking of stepping into and the function isn't on the skip
         list, step into it.
...

And with this patch, the test-case from the fedora patch does:
...
25      !$omp parallel^M
(gdb) PASS: gdb.fortran/omp-step.exp: step closer
next^M
[New Thread 0x7ffff684a700 (LWP 2877)]^M
[New Thread 0x7ffff6049700 (LWP 2878)]^M
[New Thread 0x7ffff5848700 (LWP 2879)]^M
MAIN__._omp_fn.0 () at omp-step.f90:26^M
26            a(omp_get_thread_num() + 1) = 1^M
(gdb) PASS: gdb.fortran/omp-step.exp: step into omp
...

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug gdb/26851] Make next not skip over GOMP_parallel

[vries at gcc dot gnu.org]

https://sourceware.org/bugzilla/show_bug.cgi?id=26851

--- Comment #2 from Tom de Vries <vries at gcc dot gnu.org> ---
See also https://bugzilla.redhat.com/show_bug.cgi?id=533176

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug gdb/26851] Make next not skip over GOMP_parallel

[vries at gcc dot gnu.org]

https://sourceware.org/bugzilla/show_bug.cgi?id=26851

--- Comment #3 from Tom de Vries <vries at gcc dot gnu.org> ---
(In reply to Tom de Vries from comment #1)
> This works for me, but has hardcoded x86_64 abi knowledge (reading first arg
> to GOMP_parallel from rdi reg at index 5):

Alternatively, we can use libgomp debuginfo to get the arg.

-- 
You are receiving this mail because:
You are on the CC list for the bug.