[patch/RFA/hppa] Make the stub identification code smarter

[patch/RFA/hppa] Make the stub identification code smarter

[Randolph Chung]

The old stub identification code only identified a stub if we are at the
first insn of the stub. If we are single stepping through the stub, it
will fail when the pc points to the middle of the stub. This is probably
one of the reasons why the watchpoint tests are failing... 

Tested on hppa-linux. ok to apply?

(Mark: for hppabsd, i don't see any stub handling code. if it uses
gcc/binutils then i think the same logic for hppa-linux-tdep applies to
hppabsd-tdep)

randolph

2004-11-19  Randolph Chung  <tausq@debian.org>

	* hppa-linux-tdep.c (insns_match_pattern_relaxed): New function.
	(hppa_linux_in_dyncall): Check that we are inside the range of 
	$$dyncall, not necessarily at the first insn.
	(hppa_linux_in_solib_call_trampoline): Identify a trampoline 
	even if the pc does not point to the first insn of the trampoline.

Index: hppa-linux-tdep.c
===================================================================
RCS file: /cvs/src/src/gdb/hppa-linux-tdep.c,v
retrieving revision 1.11
diff -u -p -r1.11 hppa-linux-tdep.c
--- hppa-linux-tdep.c	31 Oct 2004 21:09:28 -0000	1.11
+++ hppa-linux-tdep.c	18 Nov 2004 01:08:18 -0000
@@ -161,10 +162,39 @@ insns_match_pattern (CORE_ADDR pc,
   return 1;
 }
 
+/* The relaxed version of the insn matcher allows us to match from somewhere
+   inside the pattern, by looking backwards in the instruction scheme.  */
+static int
+insns_match_pattern_relaxed (CORE_ADDR pc,
+			     struct insn_pattern *pattern,
+			     unsigned int *insn)
+{
+  int pat_len = 0;
+  int offset;
+
+  while (pattern[pat_len].mask)
+    pat_len++;
+
+  for (offset = 0; offset < pat_len; offset++)
+    {
+      if (insns_match_pattern (pc - offset * 4,
+			       pattern, insn))
+	return 1;
+    }
+
+  return 0;
+}
+
 static int
 hppa_linux_in_dyncall (CORE_ADDR pc)
 {
-  return pc == hppa_symbol_address("$$dyncall");
+  struct unwind_table_entry *u;
+  u = find_unwind_entry (hppa_symbol_address ("$$dyncall"));
+
+  if (!u)
+    return 0;
+	
+  return pc >= u->region_start && pc <= u->region_end;
 }
 
 /* There are several kinds of "trampolines" that we need to deal with:
@@ -182,13 +212,20 @@ hppa_linux_in_solib_call_trampoline (COR
 {
   unsigned int insn[HPPA_MAX_INSN_PATTERN_LEN];
   int r;
+  struct unwind_table_entry *u;
+
+  /* on hppa-linux, linker stubs have no unwind information.  Since the pattern
+     matching for linker stubs can be quite slow, we try to avoid it if
+     we can.  */
+  u = find_unwind_entry (pc);
 
   r = in_plt_section (pc, name)
       || hppa_linux_in_dyncall (pc)
-      || insns_match_pattern (pc, hppa_import_stub, insn)
-      || insns_match_pattern (pc, hppa_import_pic_stub, insn)
-      || insns_match_pattern (pc, hppa_long_branch_stub, insn)
-      || insns_match_pattern (pc, hppa_long_branch_pic_stub, insn);
+      || (u == NULL
+	  && (insns_match_pattern_relaxed (pc, hppa_import_stub, insn)
+	      || insns_match_pattern_relaxed (pc, hppa_import_pic_stub, insn)
+	      || insns_match_pattern_relaxed (pc, hppa_long_branch_stub, insn)
+	      || insns_match_pattern_relaxed (pc, hppa_long_branch_pic_stub, insn)));
 
   return r;
 }
-- 
Randolph Chung
Debian GNU/Linux Developer, hppa/ia64 ports
http://www.tausq.org/

Re: [patch/RFA/hppa] Make the stub identification code smarter

[mark Kettenis]

   Date: Wed, 17 Nov 2004 17:23:27 -0800
   From: Randolph Chung <randolph@tausq.org>

   (Mark: for hppabsd, i don't see any stub handling code. if it uses
   gcc/binutils then i think the same logic for hppa-linux-tdep applies to
   hppabsd-tdep)

AFAIK the BSD's don't use a segmented memory model, and therefore
don't use stubs, but I'm not completely sure about that.

Now that I think of it, OpenBSD/hppa didn't have shared libraries
until very recently.

Mark

Re: [patch/RFA/hppa] Make the stub identification code smarter

[Randolph Chung]

> AFAIK the BSD's don't use a segmented memory model, and therefore
> don't use stubs, but I'm not completely sure about that.

neither does hppa-linux. The stubs are not for that... they are for
calling shlib functions and to handle long branches (since for pa11 the
branch target for a b,l insn is only 17 bits)

> Now that I think of it, OpenBSD/hppa didn't have shared libraries
> until very recently.

ok, anyway, just a heads up in case the code is useful for hppabsd too
:)

randolph
-- 
Randolph Chung
Debian GNU/Linux Developer, hppa/ia64 ports
http://www.tausq.org/