[PATCH][gdb/breakpoints] Workaround missing line-table entry

[Tom de Vries <tdevries@suse.de>]

Hi,

When running test-case gdb.opt/inline-cmds.exp, we run into this KFAIL with
gcc:
...
Breakpoint 7, main () at gdb.opt/inline-cmds.c:71^M
71        result = 0; /* set breakpoint 3 here */^M
(gdb) PASS: gdb.opt/inline-cmds.exp: continue to breakpoint: consecutive func1
next^M
73        func1 (); /* first call */^M
(gdb) PASS: gdb.opt/inline-cmds.exp: next to first func1
next^M
75        marker ();^M
(gdb) KFAIL: gdb.opt/inline-cmds.exp: next to second func1 (PRMS: gdb/25884)
...
while with clang we have instead:
...
next^M
74        func1 (); /* second call */^M
(gdb) PASS: gdb.opt/inline-cmds.exp: next to second func1
...

The relevant bit of the test source is here in inline-cmds.c:
...
    71    result = 0; /* set breakpoint 3 here */
    72
    73    func1 (); /* first call */
    74    func1 (); /* second call */
    75    marker ();
...
with func1 defined as:
...
    33  inline __attribute__((always_inline)) int func1(void)
    34  {
    35    bar ();
    36    return x * y;
    37  }
...

The corresponding insns are:
...
  40050b:       movl   $0x0,0x200b1f(%rip)        # 601034 <result>
  400515:       callq  40057b <bar>
  40051a:       callq  40057b <bar>
  40051f:       callq  400596 <marker>
...
and the line number info is:
...
Line number    Starting address    View    Stmt
         71            0x40050b               x
         35            0x400515               x
         75            0x40051f               x
...

The line number info is missing an entry for the insn at 40051a, and that is
causing the FAIL.  This is a gcc issue, filed as PR gcc/98780 -" Missing line
table entry for inlined stmt at -g -O0".

[ For contrast, with clang we have an extra entry:
...
Line number    Starting address    View    Stmt
         71            0x40050b               x
         35            0x400515               x
         35            0x40051a
         75            0x40051f               x
...
though it appears to be missing the start-of-statement marker. ]

However, there is debug info that indicates that the insn at 40051a is not
part of the line table entry for the insn at 400515:
...
<2><1c4>: Abbrev Number: 8 (DW_TAG_inlined_subroutine)
    <1c5>   DW_AT_abstract_origin: <0x2a2>
    <1c9>   DW_AT_low_pc      : 0x400515
    <1d1>   DW_AT_high_pc     : 0x5
    <1d9>   DW_AT_call_file   : 1
    <1da>   DW_AT_call_line   : 73
 <2><1db>: Abbrev Number: 8 (DW_TAG_inlined_subroutine)
    <1dc>   DW_AT_abstract_origin: <0x2a2>
    <1e0>   DW_AT_low_pc      : 0x40051a
    <1e8>   DW_AT_high_pc     : 0x5
    <1f0>   DW_AT_call_file   : 1
    <1f1>   DW_AT_call_line   : 74
...
and indeed lldb manages to "next" from line 73 to line 74.

Work around the missing line table entry, by using the inline frame info to
narrow the stepping range in prepare_one_step.

Tested on x86_64-linux.

Any comments?

Thanks,
- Tom

[gdb/breakpoints] Workaround missing line-table entry

gdb/ChangeLog:

2021-01-22  Tom de Vries  <tdevries@suse.de>

	PR breakpoints/25884
	* infcmd.c (prepare_one_step): Using inline frame info to narrow
	stepping range.

gdb/testsuite/ChangeLog:

2021-01-22  Tom de Vries  <tdevries@suse.de>

	PR breakpoints/25884
	* gdb.opt/inline-cmds.exp: Remove kfail.

---
 gdb/infcmd.c                          | 14 ++++++++++++++
 gdb/testsuite/gdb.opt/inline-cmds.exp | 17 +----------------
 2 files changed, 15 insertions(+), 16 deletions(-)

diff --git a/gdb/infcmd.c b/gdb/infcmd.c
index 6f0ed952de6..ecb3de31879 100644
--- a/gdb/infcmd.c
+++ b/gdb/infcmd.c
@@ -1009,6 +1009,20 @@ prepare_one_step (thread_info *tp, struct step_command_fsm *sm)
 				 &tp->control.step_range_start,
 				 &tp->control.step_range_end);
 
+	  /* There's a problem in gcc (PR gcc/98780) that causes missing line
+	     table entries, which results in a too large stepping range.
+	     Use inlined_subroutine info to make the range more narrow.  */
+	  if (inline_skipped_frames (tp) > 0)
+	    {
+	      symbol *sym = inline_skipped_symbol (tp);
+	      if (SYMBOL_CLASS (sym) == LOC_BLOCK)
+		{
+		  const block *block = SYMBOL_BLOCK_VALUE (sym);
+		  if (BLOCK_END (block) < tp->control.step_range_end)
+		    tp->control.step_range_end = BLOCK_END (block);
+		}
+	    }
+
 	  tp->control.may_range_step = 1;
 
 	  /* If we have no line info, switch to stepi mode.  */
diff --git a/gdb/testsuite/gdb.opt/inline-cmds.exp b/gdb/testsuite/gdb.opt/inline-cmds.exp
index 17720c46795..981dcbb4a29 100644
--- a/gdb/testsuite/gdb.opt/inline-cmds.exp
+++ b/gdb/testsuite/gdb.opt/inline-cmds.exp
@@ -222,22 +222,7 @@ gdb_breakpoint $line3
 gdb_continue_to_breakpoint "consecutive func1"
 
 gdb_test "next" ".*func1 .*first call.*" "next to first func1"
-set msg "next to second func1"
-gdb_test_multiple "next" $msg {
-    -re ".*func1 .*second call.*$gdb_prompt $" {
-	pass $msg
-    }
-    -re ".*marker .*$gdb_prompt $" {
-	# This assembles to two consecutive call instructions.
-	# Both appear to be at the same line, because they're
-	# in the body of the same inlined function.  This is
-	# reasonable for the line table.  GDB should take the
-	# containing block and/or function into account when
-	# deciding how far to step.  The single line table entry
-	# is actually two consecutive instances of the same line.
-	kfail gdb/25884 $msg
-    }
-}
+gdb_test "next" ".*func1 .*second call.*" "next to second func1"
 
 # It is easier when the two inlined functions are not on the same line.
 set line4 [gdb_get_line_number "set breakpoint 4 here"]