[rfc] Work around ARM RV compiler unwind info

[Daniel Jacobowitz <drow at false dot org>]

ARM's compilers generate, overall, very good debug info.  It's poorly
supported by GDB, for various reasons: a few places where the debug info is
in fact wrong, and a number where it is so very right that GDB can't handle
it (since we rely on less correct information typically output by GCC).
Unfortunately this is one of the former cases, but it's one that I think is
tentatively worth supporting in GDB.

The compilers have two problems.  Both have been fixed, but the fixes are
conditionalized, for maximum compatibility with other ARM tools that made
the same assumptions.

  - When generating DWARF2, with the version field in .debug_frame set to
    1, they assume that DW_CFA_def_cfa and DW_CFA_def_cfa_offset take
    factored offsets instead of unfactored.  Using --dwarf3 makes this
    problem go away; the first ARM compilers to generate DWARF3 were
    released after this bug was found.

  - When generating either DWARF2 or DWARF3 (CIE version 1 or 3), they
    assume that a reg_offset rule for the CFA is REG - OFFSET instead of
    REG + OFFSET.  RVCT 3.0 has this problem.  RVCT 3.0 SP1 does not;
    but it marks the output with an "armcc+" augmentation to indicate
    that it has the standard DWARF meaning, again for compatibility
    with other versions of ARM tools.

These aren't reverse engineered fixes, by the way - I spoke with the ARM
compiler team about the proper checks.  I only turn on either of the fixes
if the file also has .debug_info, so that we can find a producer string that
unambiguously indicates that the ARM compiler was used.

What do you think?  Useful?  Too ugly?

-- 
Daniel Jacobowitz
CodeSourcery

2006-12-18  Daniel Jacobowitz  <dan@codesourcery.com>

	* buildsym.c (start_subfile): Handle producer.
	(record_producer): New function.
	* buildsym.h (struct subfile): Include producer.
	(record_producer): New prototype.
	* dwarf2-frame.c (struct dwarf2_cie): Add version and augmentation.
	(struct dwarf2_frame_state): Add armcc_cfa_offsets_sf and
	armcc_cfa_offsets_reversed.
	(execute_cfa_program): Handle armcc_cfa_offsets_sf.
	(dwarf2_frame_find_quirks): New function.
	(dwarf2_frame_cache): Call it.  Handle armcc_cfa_offsets_reversed.
	(decode_frame_entry_1): Record the CIE version.  Record the
	augmentation.  Skip armcc augmentations.
	* dwarf2read.c (read_file_scope): Save the producer.
	* symtab.h (struct symtab): Rename unused version member to
	producer.

Index: gdb/buildsym.c
===================================================================
RCS file: /scratch/gcc/repos/src/src/gdb/buildsym.c,v
retrieving revision 1.43
diff -u -p -r1.43 buildsym.c
--- gdb/buildsym.c	25 Aug 2006 16:32:32 -0000	1.43
+++ gdb/buildsym.c	6 Nov 2006 19:46:19 -0000
@@ -596,6 +596,9 @@ start_subfile (char *name, char *dirname
      later via a call to record_debugformat. */
   subfile->debugformat = NULL;
 
+  /* Similarly for the producer.  */
+  subfile->producer = NULL;
+
   /* If the filename of this subfile ends in .C, then change the
      language of any pending subfiles from C to C++.  We also accept
      any other C++ suffixes accepted by deduce_language_from_filename.  */
@@ -1004,6 +1007,12 @@ end_symtab (CORE_ADDR end_addr, struct o
 						  &objfile->objfile_obstack);
 	    }
 
+	  /* Similarly for the producer.  */
+	  if (subfile->producer != NULL)
+	    symtab->producer = obsavestring (subfile->producer,
+					     strlen (subfile->producer),
+					     &objfile->objfile_obstack);
+
 	  /* All symtabs for the main file and the subfiles share a
 	     blockvector, so we need to clear primary for everything
 	     but the main file.  */
@@ -1026,6 +1035,8 @@ end_symtab (CORE_ADDR end_addr, struct o
 	{
 	  xfree ((void *) subfile->debugformat);
 	}
+      if (subfile->producer != NULL)
+	xfree (subfile->producer);
 
       nextsub = subfile->next;
       xfree ((void *) subfile);
@@ -1102,6 +1113,12 @@ record_debugformat (char *format)
   current_subfile->debugformat = savestring (format, strlen (format));
 }
 
+void
+record_producer (const char *producer)
+{
+  current_subfile->producer = savestring (producer, strlen (producer));
+}
+
 /* Merge the first symbol list SRCLIST into the second symbol list
    TARGETLIST by repeated calls to add_symbol_to_list().  This
    procedure "frees" each link of SRCLIST by adding it to the
Index: gdb/buildsym.h
===================================================================
RCS file: /scratch/gcc/repos/src/src/gdb/buildsym.h,v
retrieving revision 1.13
diff -u -p -r1.13 buildsym.h
--- gdb/buildsym.h	17 Dec 2005 22:33:59 -0000	1.13
+++ gdb/buildsym.h	6 Nov 2006 19:46:19 -0000
@@ -68,6 +68,7 @@ struct subfile
     struct linetable *line_vector;
     int line_vector_length;
     enum language language;
+    char *producer;
     char *debugformat;
   };
 
@@ -281,6 +282,8 @@ extern void record_pending_block (struct
 
 extern void record_debugformat (char *format);
 
+extern void record_producer (const char *producer);
+
 extern void merge_symbol_lists (struct pending **srclist,
 				struct pending **targetlist);
 
Index: gdb/dwarf2-frame.c
===================================================================
RCS file: /scratch/gcc/repos/src/src/gdb/dwarf2-frame.c,v
retrieving revision 1.63
diff -u -p -r1.63 dwarf2-frame.c
--- gdb/dwarf2-frame.c	28 May 2006 05:56:50 -0000	1.63
+++ gdb/dwarf2-frame.c	6 Nov 2006 20:24:34 -0000
@@ -64,6 +64,9 @@ struct dwarf2_cie
   gdb_byte *initial_instructions;
   gdb_byte *end;
 
+  /* Saved augmentation, in case it's needed later.  */
+  char *augmentation;
+
   /* Encoding of addresses.  */
   gdb_byte encoding;
 
@@ -73,6 +76,9 @@ struct dwarf2_cie
   /* True if an 'S' augmentation existed.  */
   unsigned char signal_frame;
 
+  /* The version recorded in the CIE.  */
+  unsigned char version;
+
   struct dwarf2_cie *next;
 };
 
@@ -138,6 +144,16 @@ struct dwarf2_frame_state
   LONGEST data_align;
   ULONGEST code_align;
   ULONGEST retaddr_column;
+
+  /* Flags for known producer quirks.  */
+
+  /* The ARM compilers, in DWARF2 mode, assume that DW_CFA_def_cfa
+     and DW_CFA_def_cfa_offset takes a factored offset.  */
+  int armcc_cfa_offsets_sf;
+
+  /* The ARM compilers, in DWARF2 or DWARF3 mode, may assume that
+     the CFA is defined as REG - OFFSET rather than REG + OFFSET.  */
+  int armcc_cfa_offsets_reversed;
 };
 
 /* Store the length the expression for the CFA in the `cfa_reg' field,
@@ -430,6 +446,10 @@ bad CFI data; mismatched DW_CFA_restore_
 	    case DW_CFA_def_cfa:
 	      insn_ptr = read_uleb128 (insn_ptr, insn_end, &fs->cfa_reg);
 	      insn_ptr = read_uleb128 (insn_ptr, insn_end, &utmp);
+
+	      if (fs->armcc_cfa_offsets_sf)
+		utmp *= fs->data_align;
+
 	      fs->cfa_offset = utmp;
 	      fs->cfa_how = CFA_REG_OFFSET;
 	      break;
@@ -444,6 +464,10 @@ bad CFI data; mismatched DW_CFA_restore_
 
 	    case DW_CFA_def_cfa_offset:
 	      insn_ptr = read_uleb128 (insn_ptr, insn_end, &utmp);
+
+	      if (fs->armcc_cfa_offsets_sf)
+		utmp *= fs->data_align;
+
 	      fs->cfa_offset = utmp;
 	      /* cfa_how deliberately not set.  */
 	      break;
@@ -715,6 +739,49 @@ dwarf2_frame_eh_frame_regnum (struct gdb
     return regnum;
   return ops->eh_frame_regnum (gdbarch, regnum);
 }
+
+static void
+dwarf2_frame_find_quirks (struct dwarf2_frame_state *fs,
+			  struct dwarf2_fde *fde)
+{
+  static const char *arm_idents[] = {
+    "ARM C Compiler, ADS",
+    "Thumb C Compiler, ADS",
+    "ARM C++ Compiler, ADS",
+    "Thumb C++ Compiler, ADS",
+    "ARM/Thumb C/C++ Compiler, RVCT"
+  };
+  int i;
+
+  struct symtab *s;
+
+  s = find_pc_symtab (fs->pc);
+  if (s == NULL || s->producer == NULL)
+    return;
+
+  for (i = 0; i < ARRAY_SIZE (arm_idents); i++)
+    if (strncmp (s->producer, arm_idents[i], strlen (arm_idents[i])) == 0)
+      {
+	if (fde->cie->version == 1)
+	  fs->armcc_cfa_offsets_sf = 1;
+
+	if (fde->cie->version == 1)
+	  fs->armcc_cfa_offsets_reversed = 1;
+
+	/* The reversed offset problem is present in some compilers
+	   using DWARF3, but it was eventually fixed.  Check the ARM
+	   defined augmentations, which are in the format "armcc" followed
+	   by a list of one-character options.  The "+" option means
+	   this problem is fixed (no quirk needed).  If the armcc
+	   augmentation is missing, the quirk is needed.  */
+	if (fde->cie->version == 3
+	    && (strncmp (fde->cie->augmentation, "armcc", 5) != 0
+		|| strchr (fde->cie->augmentation + 5, '+') == NULL))
+	  fs->armcc_cfa_offsets_reversed = 1;
+
+	return;
+      }
+}
 
 
 struct dwarf2_frame_cache
@@ -781,6 +848,9 @@ dwarf2_frame_cache (struct frame_info *n
   fs->code_align = fde->cie->code_alignment_factor;
   fs->retaddr_column = fde->cie->return_address_register;
 
+  /* Check for "quirks" - known bugs in producers.  */
+  dwarf2_frame_find_quirks (fs, fde);
+
   /* First decode all the insns in the CIE.  */
   execute_cfa_program (fde->cie->initial_instructions,
 		       fde->cie->end, next_frame, fs, fde->eh_frame_p);
@@ -798,7 +868,10 @@ dwarf2_frame_cache (struct frame_info *n
     {
     case CFA_REG_OFFSET:
       cache->cfa = read_reg (next_frame, fs->cfa_reg);
-      cache->cfa += fs->cfa_offset;
+      if (fs->armcc_cfa_offsets_reversed)
+	cache->cfa -= fs->cfa_offset;
+      else
+	cache->cfa += fs->cfa_offset;
       break;
 
     case CFA_EXP:
@@ -1584,12 +1657,18 @@ decode_frame_entry_1 (struct comp_unit *
       cie_version = read_1_byte (unit->abfd, buf);
       if (cie_version != 1 && cie_version != 3)
 	return NULL;
+      cie->version = cie_version;
       buf += 1;
 
       /* Interpret the interesting bits of the augmentation.  */
-      augmentation = (char *) buf;
+      cie->augmentation = augmentation = (char *) buf;
       buf += (strlen (augmentation) + 1);
 
+      /* Ignore armcc augmentations.  We only use them for quirks,
+	 and that doesn't happen until later.  */
+      if (strncmp (augmentation, "armcc", 5) == 0)
+	augmentation += strlen (augmentation);
+
       /* The GCC 2.x "eh" augmentation has a pointer immediately
          following the augmentation string, so it must be handled
          first.  */
Index: gdb/dwarf2read.c
===================================================================
RCS file: /scratch/gcc/repos/src/src/gdb/dwarf2read.c,v
retrieving revision 1.206
diff -u -p -r1.206 dwarf2read.c
--- gdb/dwarf2read.c	2 Nov 2006 21:34:07 -0000	1.206
+++ gdb/dwarf2read.c	6 Nov 2006 19:46:19 -0000
@@ -2803,16 +2803,9 @@ read_file_scope (struct die_info *die, s
   attr = dwarf2_attr (die, DW_AT_producer, cu);
   if (attr) 
     cu->producer = DW_STRING (attr);
-  
+
   /* We assume that we're processing GCC output. */
   processing_gcc_compilation = 2;
-#if 0
-  /* FIXME:Do something here.  */
-  if (dip->at_producer != NULL)
-    {
-      handle_producer (dip->at_producer);
-    }
-#endif
 
   /* The compilation unit may be in a different language or objfile,
      zero out all remembered fundamental types.  */
@@ -2820,6 +2813,7 @@ read_file_scope (struct die_info *die, s
 
   start_symtab (name, comp_dir, lowpc);
   record_debugformat ("DWARF 2");
+  record_producer (cu->producer);
 
   initialize_cu_func_list (cu);
 
Index: gdb/symtab.h
===================================================================
RCS file: /scratch/gcc/repos/src/src/gdb/symtab.h,v
retrieving revision 1.98
diff -u -p -r1.98 symtab.h
--- gdb/symtab.h	17 Oct 2006 20:17:45 -0000	1.98
+++ gdb/symtab.h	6 Nov 2006 19:46:19 -0000
@@ -846,9 +846,9 @@ struct symtab
 
   char *debugformat;
 
-  /* String of version information.  May be zero.  */
+  /* String of producer version information.  May be zero.  */
 
-  char *version;
+  char *producer;
 
   /* Full name of file as found by searching the source path.
      NULL if not yet known.  */