Pointer type information lost in 4.5 debuginfo

[Michael Hope]

Split from LP: #649067

Running the GDB test suite against Linaro GCC 4.5 shows more failures than against 4.4.

The following regressions remain; as mentioned in the bug report, these are due to bug PR 45088:
FAIL: gdb.cp/class2.exp: p acp->c1
FAIL: gdb.cp/class2.exp: p acp->c2
If you prefer, we can open a second bug to track these. (There is no fix available upstream yet.)

In addition, I'm now seeing yet another failure:
FAIL: gdb.cp/temargs.exp: test value of P in inner_m
Not sure yet what causes this one.

[Jakub-gcc (jakub-gcc)]

So, to add c1's type, we call gen_typedef_with_usage with A typedef variant.
19880 /* If TYPE is a typedef type variant, let's generate debug info
19881 for the parent typedef which TYPE is a type of. */
19882 if (typedef_variant_p (type))
19883 {
19884 if (TREE_ASM_WRITTEN (type))
19885 return;
19886
19887 /* Prevent broken recursion; we can't hand off to the same type. */
19888 gcc_assert (DECL_ORIGINAL_TYPE (TYPE_NAME (type)) != type);
19889
19890 /* Use the DIE of the containing namespace as the parent DIE of
19891 the type description DIE we want to generate. */
19892 if (DECL_CONTEXT (TYPE_NAME (type))
19893 && TREE_CODE (DECL_CONTEXT (TYPE_NAME (type))) == NAMESPACE_DECL)
19894 context_die = get_context_die (DECL_CONTEXT (TYPE_NAME (type)));
19895
19896 TREE_ASM_WRITTEN (type) = 1;
19897
19898 gen_decl_die (TYPE_NAME (type), NULL, context_die);
19899 return;

TREE_ASM_WRITTEN is not set on type originally, we compute context_die. Then set TREE_ASM_WRITTEN and call gen_decl_die. TYPE_NAME (type) is a redundant typedef though, so in gen_decl_die:
20544 if (is_redundant_typedef (decl))
20545 gen_type_die (TREE_TYPE (decl), context_die);
and TREE_TYPE (decl) here is the type on which we've just set TREE_ASM_WRITTEN, so gen_type_die_with_usage is recursed with the same type and as TREE_ASM_WRITTEN is now already set, it returns immediately, without creating any type. So, I think we need to special case is_redundant_typedef (TYPE_NAME (type))) in gen_type_die_with_usage in the typedef_variant_p (type) handling and ensure we actually create the DECL_ORIGINAL_TYPE in that case and equate that also to the underlying type. Or this could be a bug in the FE.

[Uweigand-gcc (uweigand-gcc)]

Any update on this bug? I'm seeing this in 4.5 as well ...

[Redi (redi)]

is this another dup of PR 43628? should be fixed if it is

[Uweigand-gcc (uweigand-gcc)]

(In reply to comment #3)
> is this another dup of PR 43628? should be fixed if it is

No, unfortunately it's not. In mainline and 4.5 (which have the 43628) fix, all other GDB C++ test suite regressions over 4.4 are fixed, *except* for the one described in this bug report.

[Michael Hope (michaelh1)]

Split from LP: #649067

Running the GDB test suite against Linaro GCC 4.5 shows more failures than against 4.4.

The following regressions remain; as mentioned in the bug report, these are due to bug PR 45088:
FAIL: gdb.cp/class2.exp: p acp->c1
FAIL: gdb.cp/class2.exp: p acp->c2
If you prefer, we can open a second bug to track these. (There is no fix available upstream yet.)

In addition, I'm now seeing yet another failure:
FAIL: gdb.cp/temargs.exp: test value of P in inner_m
Not sure yet what causes this one.

[Dodji-gcc (dodji-gcc)]

Following the (IMHO correect) analysis of Jakub, an interesting
question would be:

  Why is the type of *c1 a typedef variant of "A" instead
  of being just "A"?

It's actually the "self reference type" of A. In struct A, the
standard asks to inject the name A into the struct A itself, so that
looking up A::A or C::A succeeds. G++ creates a special typedef of A
(named self reference type in G++ speak) and injects that typedef into
A.

So from inside struct C, when considering "A *c1", the lookup of A
returns the self reference type of A. I think after the lookup
succeeds, G++ should retain A as the type of *c1; not the self
reference. Otherwise that confuses the debug info emitter as this bug
suggests.

I am about to test the patch below that hopefully does what I would
want. It's actually the second hunk that fixes this case because it
does what I am saying for the simple-type-specifier production.

The first hunk is something I noticed while looking at this issue. The
initial code in check_elaborated_type_specifier actually tries to do
what I am saying (for the elaborated-type-specifier production) but I
believe it fails in doing so, probably because the way self references
are represented has changed since the code was written.

diff --git a/gcc/cp/decl.c b/gcc/cp/decl.c
index fb5ca7f..feba130 100644
--- a/gcc/cp/decl.c
+++ b/gcc/cp/decl.c
@@ -10902,7 +10902,7 @@ check_elaborated_type_specifier (enum tag_types tag_code,
      name lookup will find the TYPE_DECL for the implicit "S::S"
      typedef. Adjust for that here. */
   if (DECL_SELF_REFERENCE_P (decl))
- decl = TYPE_NAME (TREE_TYPE (decl));
+ decl = TYPE_NAME (DECL_ORIGINAL_TYPE (decl));

   type = TREE_TYPE (decl);

diff --git a/gcc/cp/parser.c b/gcc/cp/parser.c
index 6a9e4d7..d70c621 100644
--- a/gcc/cp/parser.c
+++ b/gcc/cp/parser.c
@@ -12791,6 +12791,11 @@ cp_parser_simple_type_specifier (cp_parser* parser,
       /* Otherwise, look for a type-name. */
       else
  type = cp_parser_type_name (parser);
+
+ if (type && TREE_CODE (type) == TYPE_DECL
+ && DECL_SELF_REFERENCE_P (type))
+ type = TYPE_NAME (DECL_ORIGINAL_TYPE (type));
+
       /* Keep track of all name-lookups performed in class scopes. */
       if (type
    && !global_p

[Dodji-gcc (dodji-gcc)]

I believe this is a smaller reproducer (the type of *a1 is not present
in debug info):

struct A
{
    virtual ~A();
};

struct B : public A
{
    virtual ~B(){}
};

struct C : public B
{
    A* a1;
};

int
main()
{
    C c;
    c.a1 = 0;
    return 0;
}

[Dodji-gcc (dodji-gcc)]

Finally the patch I pasted earlier wasn't good enough. I modified it
and proposed it to
http://gcc.gnu.org/ml/gcc-patches/2010-11/msg01277.html

[Dodji-gcc (dodji-gcc)]

Author: dodji
Date: Fri Dec 17 10:39:21 2010
New Revision: 167976

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=167976
Log:
Fix for PR debug/45088

gcc/

 * dwarf2out.c (gen_type_die_with_usage): Do not try to emit debug
 info for a redundant typedef that has DECL_ORIGINAL_TYPE set. Use
 that underlying type instead.

gcc/testsuite/

 * g++.dg/debug/dwarf2/self-ref-1.C: New test.
 * g++.dg/debug/dwarf2/self-ref-2.C: Likewise.

Added:
    trunk/gcc/testsuite/g++.dg/debug/dwarf2/self-ref-1.C
    trunk/gcc/testsuite/g++.dg/debug/dwarf2/self-ref-2.C
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/dwarf2out.c
    trunk/gcc/testsuite/ChangeLog

[Ulrich Weigand (uweigand)]

Mainline fix has now been committed:
http://gcc.gnu.org/ml/gcc-patches/2010-12/msg01311.html

I'll work on a backport to Linaro GCC 4.5.

[Andrew Stubbs (ams-codesourcery)]

Related: lp:gcc-linaro/4.5,revno=99450

[Dodji-gcc (dodji-gcc)]

Fixed in 4.6

[Pluto-7 (pluto-7)]

what about 4.7 branch?

[Jason-gcc (jason-gcc)]

(In reply to comment #10)
> what about 4.7 branch?

The fix was on the trunk before 4.7 branched, so yes.