.debug_names entry referring to skeleton type unit

[dwblaikie]

When using .debug_names and Type Units, two entries for the type are produced ( https://godbolt.org/z/7Pd67afn7 ):

struct t1 { };
t1 v1;

     Hash: 0x59796A
      String: 0x00000085 "t1"
      Entry @ 0xa8 {
        Abbrev: 0x26c
        Tag: DW_TAG_structure_type
        DW_IDX_type_unit: 0x00
        DW_IDX_die_offset: 0x00000023
      }
      Entry @ 0xaf {
        Abbrev: 0x268
        Tag: DW_TAG_structure_type
        DW_IDX_die_offset: 0x0000002e
      }

DW_TAG_structure_type
                DW_AT_calling_convention	(DW_CC_pass_by_value)
                DW_AT_name	("t1")
                DW_AT_byte_size	(0x01)
                DW_AT_decl_file	("/app/example.cpp")
                DW_AT_decl_line	(1)

DW_TAG_structure_type
                DW_AT_declaration	(true)
                DW_AT_signature	(0xc6694e51369161f2)

Probably only want the type unit entry for the type in the index.

[llvmbot]

@llvm/issue-subscribers-debuginfo

Author: David Blaikie (dwblaikie)

When using `.debug_names` and Type Units, two entries for the type are produced ( https://godbolt.org/z/7Pd67afn7 ): ``` struct t1 { }; t1 v1; ``` ``` Hash: 0x59796A String: 0x00000085 "t1" Entry @ 0xa8 { Abbrev: 0x26c Tag: DW_TAG_structure_type DW_IDX_type_unit: 0x00 DW_IDX_die_offset: 0x00000023 } Entry @ 0xaf { Abbrev: 0x268 Tag: DW_TAG_structure_type DW_IDX_die_offset: 0x0000002e } ``` ``` DW_TAG_structure_type DW_AT_calling_convention (DW_CC_pass_by_value) DW_AT_name ("t1") DW_AT_byte_size (0x01) DW_AT_decl_file ("/app/example.cpp") DW_AT_decl_line (1) ``` ``` DW_TAG_structure_type DW_AT_declaration (true) DW_AT_signature (0xc6694e51369161f2) ``` Probably only want the type unit entry for the type in the index.

[clayborg]

We also need to verify that if a DIE has a DW_IDX_parent entry, that it doesn't add an entry to .debug_names that is a forward declaration. I am seeing this in a binary that has had BOLT run on it, so not sure if those entries are coming from clang or from BOLT.

Quoting from an email:

fpiovezan@apple.com said:
I need to find my notes from those days, but I don't think we did. As Greg points out, the standard forbids forward declarations in debug_names; entries whose parents are forward declarations should be marked as having a parent that is not indexed.
The addition of IDX_parent_entries should not cause the addition of entries in the table, if it does then it is a bug in the implementation.

In fact, in the original PR we have this bit of code: https://github.com/llvm/llvm-project/pull/77457/files#diff-587587ad06ddb6f99f9ad8d8deffbc2ea59fde9d62d1b5ff58ace1f52cc75752R405

[dwblaikie]

We also need to verify that if a DIE has a DW_IDX_parent entry, that it doesn't add an entry to .debug_names that is a forward declaration.

Probably best not to muddy other bugs like this - either file a separate one, or a new thread on discourse, etc.

but here's the verification that we don't add a DW_IDX_parent for a declaration, at least one case of not doing that: https://godbolt.org/z/Ke3so5Eb5

[dwblaikie]

Fixed in bd5c636