[PATCH] gdb.rocm/watchpoint-basic: add gfx1103 to XFAILs

[PATCH] gdb.rocm/watchpoint-basic: add gfx1103 to XFAILs

[Shahab Vahedi]

Again, a confirmed KFD issue.  By confirmed, I mean that if a dummy
dispatch is done first, then everything goes OK.

  __global__ void dummy () {}

  int main (...)
  {
    ...
    dummy<<<1, 1>>> ();
    /* Break after malloc.  */
    kernel<<<1, 1>>> (global_ptr1, global_ptr2);
    ...
  }
---
 gdb/testsuite/gdb.rocm/watchpoint-basic.exp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/gdb/testsuite/gdb.rocm/watchpoint-basic.exp b/gdb/testsuite/gdb.rocm/watchpoint-basic.exp
index bc30726515e..558a1f8472c 100644
--- a/gdb/testsuite/gdb.rocm/watchpoint-basic.exp
+++ b/gdb/testsuite/gdb.rocm/watchpoint-basic.exp
@@ -96,7 +96,7 @@ proc_with_prefix test_host_watchpoint_after_runtime_load {} {
 # end and exit prematurely.
 
 gdb_caching_proc target_has_xfail {} {
-    set xfail_arches {gfx1201}
+    set xfail_arches {gfx1103 gfx1201}
 
     set targets [find_amdgpu_devices]
     if {[llength $targets] == 0} {
-- 
2.55.0

Re: [PATCH] gdb.rocm/watchpoint-basic: add gfx1103 to XFAILs

[Simon Marchi]

On 8/7/26 12:28 PM, Shahab Vahedi wrote:
> Again, a confirmed KFD issue.  By confirmed, I mean that if a dummy
> dispatch is done first, then everything goes OK.
> 
>   __global__ void dummy () {}
> 
>   int main (...)
>   {
>     ...
>     dummy<<<1, 1>>> ();
>     /* Break after malloc.  */
>     kernel<<<1, 1>>> (global_ptr1, global_ptr2);
>     ...
>   }

If it's really the case that this arch exhibits the same problem, then
it's fine to continue with the existing pattern.

Approved-By: Simon Marchi <simon.marchi@efficios.com>

As a subsequent comment, I wonder if, for these arches, instead of
xfail-ing, it would be more useful to dispatch that dummy commit and
then expect the watchpoint to work.

Simon

Re: [PATCH] gdb.rocm/watchpoint-basic: add gfx1103 to XFAILs

[Machado, Luis]

> As a subsequent comment, I wonder if, for these arches, instead of
> xfail-ing, it would be more useful to dispatch that dummy commit and
> then expect the watchpoint to work.

I think that'd makes sense. And if we want to make sure we still expose the KFD issue, we could run a quick check first (but not FAIL) and proceed to use the workaround if needed.

Re: [PATCH] gdb.rocm/watchpoint-basic: add gfx1103 to XFAILs

[Vahedi, Shahab]

On 7-Aug-2026 07:36PM, Simon Marchi wrote:

> As a subsequent comment, I wonder if, for these arches, instead of
> xfail-ing, it would be more useful to dispatch that dummy commit and
> then expect the watchpoint to work.

IMO, this would not trigger us if/when the KFD issue is resolved, so
the "list" won't go away.  Having the "list" removed has the benefit
of treating any failure from that point onward as a new one.

On 7-Aug-2026 10:58PM, Machado, Luis wrote:

> I think that'd makes sense. And if we want to make sure we still expose
> the KFD issue, we could run a quick check first (but not FAIL) and
> proceed to use the workaround if needed.

What would be this quick-check (without failures)?
How would it cover new architectures?
Or is it going to "fallback" in case of "any" failure?
Yet again, how would it trigger us when the problem is resolved?

-Shahab