diff mbox series

gdb/infrun: lazily load curr_frame_id in process_event_stop_test

Message ID 20231228193359.3888031-1-lancelot.six@amd.com
State New
Headers
Series gdb/infrun: lazily load curr_frame_id in process_event_stop_test |

Checks

Context Check Description
linaro-tcwg-bot/tcwg_gdb_build--master-aarch64 success Testing passed
linaro-tcwg-bot/tcwg_gdb_build--master-arm success Testing passed
linaro-tcwg-bot/tcwg_gdb_check--master-aarch64 success Testing passed
linaro-tcwg-bot/tcwg_gdb_check--master-arm fail Testing failed

Commit Message

Lancelot SIX Dec. 28, 2023, 7:33 p.m. UTC 
  A recent(ish) change in gdb/infrun.c made process_event_stop_test load
debug information where it would not have done so previously.  The
change is:

    commit bf2813aff8f2988ad3d53e819a0415abf295c91f
    AuthorDate: Fri Sep 1 13:47:32 2023 +0200
    CommitDate: Mon Nov 20 10:54:03 2023 +0100

        gdb/record: print frame information when exiting a recursive call

        Currently,  when GDB is reverse stepping out of a function into the same
        function due to a recursive call, it doesn't print frame information, as
        reported by PR record/29178. This happens because when the inferior
        leaves the current frame, GDB decides to refresh the step information,
        clobbering the original step_frame_id, making it impossible to figure
        out later on that the frame has been changed.

        This commit changes GDB so that, if we notice we're in this exact
        situation, we won't refresh the step information.

        Because of implementation details, this change can cause some debug
        information to be read when it normally wouldn't before, which showed up
        as a regression on gdb.dwarf2/dw2-out-of-range-end-of-seq. Since that
        isn't a problem, the test was changed to allow for the new output.

        Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29178

Although there is nothing wrong with this change in principle, it
happens to break most of the tests in gdb/testsuite/gdb.rocm/*.exp.
This is because those tests do rely on GDB not loading debug
information.  This is necessary because the debug information produced
for AMDGPU code is using DWARF extensions which are not supported by GDB
at this point.

In this patch, I propose to use a lazy loading mechanism so the frame_id
for the current frame is only computed when required instead of when
entering process_event_stop_test.  The lazy_loader class is currently
defined locally in infrun.c, but if it turns out to be useful elsewhere,
it could go somewhere under gdbsupport.

This patch should restore the behavior GDB had before
bf2813aff8f2988ad3d53e819a0415abf295c91f when it comes to load debug
info.

Another approach could have been to revert
fb84fbf8a51f5be2e78765508ebd9753af96b492 (gdb/infrun: simplify
process_event_stop_test) and adjust the implementation of
bf2813aff8f2988ad3d53e819a0415abf295c91f (gdb/record: print frame
information when exiting a recursive call).  However, I think that the
lazy loading works well with the simplification done recently, so I went
down that route.

Regression tested on x86_64-linux (Ubuntu 22.04) with AMDGPU support.

Change-Id: Ib63a162128130d1786a77c98623e9e3dcbc363b7
---
 gdb/infrun.c | 50 +++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 41 insertions(+), 9 deletions(-)


base-commit: 3bb1944a5a0527a38702084ac301d9933b0130bb

Comments

Kevin Buettner Jan. 10, 2024, 8:04 p.m. UTC | #1 
On Thu, 28 Dec 2023 19:33:59 +0000
Lancelot SIX <lancelot.six@amd.com> wrote:

> A recent(ish) change in gdb/infrun.c made process_event_stop_test load
> debug information where it would not have done so previously.  The
> change is:
> 
>     commit bf2813aff8f2988ad3d53e819a0415abf295c91f
>     AuthorDate: Fri Sep 1 13:47:32 2023 +0200
>     CommitDate: Mon Nov 20 10:54:03 2023 +0100
> 
>         gdb/record: print frame information when exiting a recursive call
> 
>         Currently,  when GDB is reverse stepping out of a function into the same
>         function due to a recursive call, it doesn't print frame information, as
>         reported by PR record/29178. This happens because when the inferior
>         leaves the current frame, GDB decides to refresh the step information,
>         clobbering the original step_frame_id, making it impossible to figure
>         out later on that the frame has been changed.
> 
>         This commit changes GDB so that, if we notice we're in this exact
>         situation, we won't refresh the step information.
> 
>         Because of implementation details, this change can cause some debug
>         information to be read when it normally wouldn't before, which showed up
>         as a regression on gdb.dwarf2/dw2-out-of-range-end-of-seq. Since that
>         isn't a problem, the test was changed to allow for the new output.
> 
>         Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29178
> 
> Although there is nothing wrong with this change in principle, it
> happens to break most of the tests in gdb/testsuite/gdb.rocm/*.exp.
> This is because those tests do rely on GDB not loading debug
> information.  This is necessary because the debug information produced
> for AMDGPU code is using DWARF extensions which are not supported by GDB
> at this point.
> 
> In this patch, I propose to use a lazy loading mechanism so the frame_id
> for the current frame is only computed when required instead of when
> entering process_event_stop_test.  The lazy_loader class is currently
> defined locally in infrun.c, but if it turns out to be useful elsewhere,
> it could go somewhere under gdbsupport.
> 
> This patch should restore the behavior GDB had before
> bf2813aff8f2988ad3d53e819a0415abf295c91f when it comes to load debug
> info.
> 
> Another approach could have been to revert
> fb84fbf8a51f5be2e78765508ebd9753af96b492 (gdb/infrun: simplify
> process_event_stop_test) and adjust the implementation of
> bf2813aff8f2988ad3d53e819a0415abf295c91f (gdb/record: print frame
> information when exiting a recursive call).  However, I think that the
> lazy loading works well with the simplification done recently, so I went
> down that route.
> 
> Regression tested on x86_64-linux (Ubuntu 22.04) with AMDGPU support.

I remember reviewing that change from Guinevere last year wondering if
there might be any unexpected fallout from reading debug info earlier
than it had been previously.  Now we know...

I'm okay with your approach, but I'd like to give other maintainers
a chance to weigh in.  So, please wait a few more days before you
push this change.

Approved-by: Kevin Buettner <kevinb@redhat.com>

Kevin
Lancelot SIX Jan. 21, 2024, 4:20 p.m. UTC | #2 
[Public]

> I remember reviewing that change from Guinevere last year wondering if
> there might be any unexpected fallout from reading debug info earlier
> than it had been previously.  Now we know...
>
> I'm okay with your approach, but I'd like to give other maintainers
> a chance to weigh in.  So, please wait a few more days before you
> push this change.
>
> Approved-by: Kevin Buettner <kevinb@redhat.com>
>
> Kevin

Thanks for the review,

I am pushing this now.

Lancelot.
diff mbox series

Patch

diff --git a/gdb/infrun.c b/gdb/infrun.c
index 1d863896c40..f7f83d540e2 100644
--- a/gdb/infrun.c
+++ b/gdb/infrun.c
@@ -7295,6 +7295,37 @@  handle_signal_stop (struct execution_control_state *ecs)
   process_event_stop_test (ecs);
 }
 
+namespace {
+
+/* Helper class for process_event_stop_test implementing lazy evaluation.  */
+template<typename T>
+class lazy_loader
+{
+  using fetcher_t = std::function<T ()>;
+
+public:
+  explicit lazy_loader (fetcher_t &&f) : m_loader (std::move (f))
+  { }
+
+  T &operator* ()
+  {
+    if (!m_value.has_value ())
+      m_value.emplace (m_loader ());
+    return m_value.value ();
+  }
+
+  T *operator-> ()
+  {
+    return &**this;
+  }
+
+private:
+  std::optional<T> m_value;
+  fetcher_t m_loader;
+};
+
+}
+
 /* Come here when we've got some debug event / signal we can explain
    (IOW, not a random signal), and test whether it should cause a
    stop, or whether we should resume the inferior (transparently).
@@ -7335,7 +7366,8 @@  process_event_stop_test (struct execution_control_state *ecs)
   /* Shorthand to make if statements smaller.  */
   struct frame_id original_frame_id
     = ecs->event_thread->control.step_frame_id;
-  struct frame_id curr_frame_id = get_frame_id (get_current_frame ());
+  lazy_loader<frame_id> curr_frame_id
+    ([] () { return get_frame_id (get_current_frame ()); });
 
   switch (what.main_action)
     {
@@ -7423,7 +7455,7 @@  process_event_stop_test (struct execution_control_state *ecs)
 
 	if (init_frame)
 	  {
-	    if (curr_frame_id == ecs->event_thread->initiating_frame)
+	    if (*curr_frame_id == ecs->event_thread->initiating_frame)
 	      {
 		/* Case 2.  Fall through.  */
 	      }
@@ -7596,7 +7628,7 @@  process_event_stop_test (struct execution_control_state *ecs)
   if (pc_in_thread_step_range (ecs->event_thread->stop_pc (),
 			       ecs->event_thread)
       && (execution_direction != EXEC_REVERSE
-	  || curr_frame_id == original_frame_id))
+	  || *curr_frame_id == original_frame_id))
     {
       infrun_debug_printf
 	("stepping inside range [%s-%s]",
@@ -8039,7 +8071,7 @@  process_event_stop_test (struct execution_control_state *ecs)
      frame machinery detected some skipped call sites, we have entered
      a new inline function.  */
 
-  if ((curr_frame_id == original_frame_id)
+  if ((*curr_frame_id == original_frame_id)
       && inline_skipped_frames (ecs->event_thread))
     {
       infrun_debug_printf ("stepped into inlined function");
@@ -8087,7 +8119,7 @@  process_event_stop_test (struct execution_control_state *ecs)
      through a more inlined call beyond its call site.  */
 
   if (get_frame_type (get_current_frame ()) == INLINE_FRAME
-      && (curr_frame_id != original_frame_id)
+      && (*curr_frame_id != original_frame_id)
       && stepped_in_from (get_current_frame (), original_frame_id))
     {
       infrun_debug_printf ("stepping through inlined function");
@@ -8118,7 +8150,7 @@  process_event_stop_test (struct execution_control_state *ecs)
 	  end_stepping_range (ecs);
 	  return;
 	}
-      else if (curr_frame_id == original_frame_id)
+      else if (*curr_frame_id == original_frame_id)
 	{
 	  /* We are not at the start of a statement, and we have not changed
 	     frame.
@@ -8144,9 +8176,9 @@  process_event_stop_test (struct execution_control_state *ecs)
 	}
     }
   else if (execution_direction == EXEC_REVERSE
-	  && curr_frame_id != original_frame_id
-	  && original_frame_id.code_addr_p && curr_frame_id.code_addr_p
-	  && original_frame_id.code_addr == curr_frame_id.code_addr)
+	  && *curr_frame_id != original_frame_id
+	  && original_frame_id.code_addr_p && curr_frame_id->code_addr_p
+	  && original_frame_id.code_addr == curr_frame_id->code_addr)
     {
       /* If we enter here, we're leaving a recursive function call.  In this
 	 situation, we shouldn't refresh the step information, because if we