[PATCH] Ignore 0,0 entries in .debug_aranges

[PATCH] Ignore 0,0 entries in .debug_aranges

[Tom Tromey]

When running the internal AdaCore test suite against the new DWARF
indexer, I found one regression on RISC-V.  The test in question uses
--gc-sections, and winds up with an entry in the middle of a
.debug_aranges that has both address and length of 0.  In this
scenario, gdb assumes the entries are terminated and then proceeds to
reject the section because it reads a subsequent entry as if it were a
header.

It seems to me that, because each header describes the size of each
.debug_aranges CU, it's better to simply ignore 0,0 entries and simply
read to the end.  That is what this patch does.

I've patched an existing test to provide a regression test for this.
---
 gdb/dwarf2/read.c                                     | 11 +++++++++--
 .../gdb.dwarf2/locexpr-data-member-location.exp       |  3 +++
 2 files changed, 12 insertions(+), 2 deletions(-)

diff --git a/gdb/dwarf2/read.c b/gdb/dwarf2/read.c
index 698720276a9..6dcd446e5f4 100644
--- a/gdb/dwarf2/read.c
+++ b/gdb/dwarf2/read.c
@@ -2469,7 +2469,7 @@ read_addrmap_from_aranges (dwarf2_per_objfile *per_objfile,
 	 bytes.  */
       addr += (entry_end - addr) % (2 * address_size);
 
-      for (;;)
+      while (addr < entry_end)
 	{
 	  if (addr + 2 * address_size > entry_end)
 	    {
@@ -2487,7 +2487,14 @@ read_addrmap_from_aranges (dwarf2_per_objfile *per_objfile,
 						      dwarf5_byte_order);
 	  addr += address_size;
 	  if (start == 0 && length == 0)
-	    break;
+	    {
+	      /* This can happen on some targets with --gc-sections.
+		 This pair of values is also used to mark the end of
+		 the entries for a given CU, but we ignore it and
+		 instead handle termination using the check at the top
+		 of the loop.  */
+	      continue;
+	    }
 	  if (start == 0 && !per_bfd->has_section_at_zero)
 	    {
 	      /* Symbol was eliminated due to a COMDAT group.  */
diff --git a/gdb/testsuite/gdb.dwarf2/locexpr-data-member-location.exp b/gdb/testsuite/gdb.dwarf2/locexpr-data-member-location.exp
index 67e96fb1128..adb4e0a4c21 100644
--- a/gdb/testsuite/gdb.dwarf2/locexpr-data-member-location.exp
+++ b/gdb/testsuite/gdb.dwarf2/locexpr-data-member-location.exp
@@ -296,6 +296,9 @@ Dwarf::assemble ${asm_file} {
     }
 
     aranges {} cu_label {
+	# This 0,0 entry tests that the .debug_aranges reader can
+	# handle an apparent terminator before the end of the ranges.
+	arange {} 0 0
 	arange {} $foo_start $foo_end
 	arange {} $bar_start $bar_end
     }
-- 
2.34.1

Re: [PATCH] Ignore 0,0 entries in .debug_aranges

[Pedro Alves]

On 2022-04-14 16:39, Tom Tromey via Gdb-patches wrote:
> When running the internal AdaCore test suite against the new DWARF
> indexer, I found one regression on RISC-V.  The test in question uses
> --gc-sections, and winds up with an entry in the middle of a
> .debug_aranges that has both address and length of 0.  In this
> scenario, gdb assumes the entries are terminated and then proceeds to
> reject the section because it reads a subsequent entry as if it were a
> header.
> 
> It seems to me that, because each header describes the size of each
> .debug_aranges CU, it's better to simply ignore 0,0 entries and simply
> read to the end.  That is what this patch does.
> 
> I've patched an existing test to provide a regression test for this.

Makes sense to me.

Re: [PATCH] Ignore 0,0 entries in .debug_aranges

[Jim Wilson]

On Thu, Apr 14, 2022 at 8:39 AM Tom Tromey via Gdb-patches <
gdb-patches@sourceware.org> wrote:

> When running the internal AdaCore test suite against the new DWARF
> indexer, I found one regression on RISC-V.  The test in question uses
> --gc-sections, and winds up with an entry in the middle of a
> .debug_aranges that has both address and length of 0.  In this
> scenario, gdb assumes the entries are terminated and then proceeds to
> reject the section because it reads a subsequent entry as if it were a
> header.
>

FYI There is also an elfutils bug for this.
    https://sourceware.org/bugzilla/show_bug.cgi?id=27805
I don't think the elfutils problem has been fixed yet.  I did notice that I
got different results for RISC-V and x86_64.  RISC-V gave me 0,0 pairs
which is wrong, but x86_64 gave me 0,4 pairs which is OK.  I don't know why
the difference; I didn't investigate.

Jim