DWARF information parsing with TI C6000

Hello everyone,

We have detected an interesting case when reading debug information for our tools for files 
that have been built with C6000 compiler. 
Besides the specific usage of DIE elements with name CIE_RETA that have been explained in
TI documentation we have found that name attribute in DIEs containing CIE_RETA have one
extra byte in length.
It was possible to detect this issue by debugging inside the parser and comparing the parsing 
steps with the output produced by ofd6x.exe

1. Output from ofd6x.exe:

00016a64 27 DW_TAG_TI_assign_register
00016a65 DW_AT_location DW_OP_regx 127
00016a68 DW_AT_name DETR <- the size is OK (6d - 68=5 bytes=4 chars + term)
00016a6d 27 DW_TAG_TI_assign_register
00016a6e DW_AT_location DW_OP_regx 228

00016a72 DW_AT_name CIE_RETA <- the size is bigger (7c-72=10 bytes>8 chars + term)!
00016a7c 28 DW_TAG_pointer_type
00016a7d DW_AT_address_class DW_ADDR_TI_PTR32

The fact that this string attribute (it is possible to confirm it is a string because the attribute type 
is DW_FORM_string or 0x08) has different size than it should have makes our parser read 
incorrect abbrev. identifier (because of the wrong offset) and terminate the parsing of DIEs 
for current compilation unit. 
Could you please help me with a link to a document that describes this behavior 
or with additional information what would be the preferred way of handling such cases.

Thank you very much in advance for your help,

  • I am using v7.3.1 of the C6000 compiler tools.  And I cannot reproduce what you describe.  So, one possible solution is to upgrade your compiler to v7.3.x.  If that isn't possible for some reason, please submit a test case.  The last part of the forum usage guidelines gives the details on doing that.

    Thanks and regards,


  • In reply to George Mock:

    Dear George,

    Thank you very much for your reply. As this is a project of our customer I am not able to upload it directly without a permission. However I will try to obtain one and provide you with an example.

    As far as I'm aware the compiler version that have been used for building this out file is built with compiler version

        File Name:               broken.out                         
    Format: TI-COFF Version 2
    File Type: executable file
    Time Stamp: Fri Feb 24 18:14:30 2012
    Machine: TI C6x
    Machine Endian: little endian
    Entry Point: 0xc01277c0
    Vendor: Texas Instruments, Inc.
    Vendor Version: 6001
    Number of Sections: 29
    File Length: 36663922
    TI-COFF f_flags: 0x00000193
    TI-COFF f_flag: F_RELFLG (relocation info stripped)
    CPU Generation: 0x9
    Control Data Endian: little endian
    Start of .text Section: 0xc0000020
    Start of .data Section: 0xc0000010
    Size of .text Section: 1268224
    Size of .data Section: 0
    Size of .bss Section: 6626

    For a number of reason this customer wants to stick this compiler version, so we're trying to find out why exactly this situation happens. 

    What I can provide as an additional information is that this special case happened when building C++ application which also has stlport library.

    Since we have tested with other projects it was possible to observe similar size behavior of CIE_RETA, but in our other projects the CIE_RETA tag

    was always put at the end of the compilation unit, and in this case having one extra byte and reading zero abbrev id is actually correct and does not have

    side effects on our parsing:

        00013569  56    DW_TAG_TI_assign_register
    0001356a DW_AT_location DW_OP_regx 127
    0001356d DW_AT_name DETR
    00013572 56 DW_TAG_TI_assign_register
    00013573 DW_AT_location DW_OP_regx 228
    00013577 DW_AT_name CIE_RETA

    DIE Table at Offset 00013581

    Length: 55403
    Version: 2
    .debug_abbrev Offset: 00000823
    Address Size: 4
    Thank you very much once again for your information,
  • In reply to Stanimir Kabaivanov:

    I have not exactly reproduced your issue.  But I did experiment and find that, for a simple test case, the v7.3.4 compiler always encodes the CIE_RETA register name in 9 bytes, and the v6.1.20 compiler appears to encode it in 10 bytes.  So, I filed the entry SDSCM00043576 in the SDOWP system.  In it, I describe this difference in detail and ask for an explanation.  Perhaps it is a bug that has been fixed.  I don't know.  But I think this is the best way to find out.  You can track this issue with SDOWP link below in my signature.

    BTW, I used v6.1.20 and not v6.1.19 because I wanted to use the latest version along that release stream.  Release v6.1.20 differs from v6.1.19 only in bug fixes.

    Thanks and regards,