Re: [DynInst_API:] Function parameters inspection with GCC -O2 optimization

[Francis Deslauriers <francis.deslauriers@xxxxxxxxxxxx>]

All right.
Thank you,

Francis

On Thu, Jul 17, 2014 at 2:15 PM, Bill Williams <bill@xxxxxxxxxxx> wrote:
> After a quick check, libdwarf supports a variety of GNU extensions that
> Dyninst presently does not. Filing this one as "for 8.3" tentatively, as the
> gnu extensions will take more work than we want to tackle for 8.2.
>
>
> On 07/17/2014 01:03 PM, Francis Deslauriers wrote:
>>
>> Hi folks,
>> This email is part bug report and part question. I wanted to share
>> what I found while debugging this issue in order to helps someone that
>> would tackle this issue in the future.
>>
>> I want to retrieve the parameters of a binary that has been compile
>> with gcc -O2 optimization option.
>>
>> I made a dummy program (mutatee.c in the attached archive) that
>> contains a function named foo that has one parameter. I was wondering
>> why I was not able to get the parameter's name, type and eventually
>> value with my mutator program. So, I activated the
>> DYNINST_DEBUG_DWARF=1 env var to examine what was happening.
>> http://paste.ubuntu.com/7809741/. We can see on line 655 that there is
>> a failure parsing the third entry of the location list for the bar
>> variable.
>>
>> Here is the function foo:
>> void __attribute__((noinline, noclone))
>>   foo(int bar)
>> {
>>      printf("%d",bar);
>>      return;
>> }
>>
>> Using the following: objdump --dwarf=info mutatee
>> We can find the entry for the foo function and find the bar parameter
>> at the 0x2ec offset.
>>   <1><2cf>: Abbrev Number: 17 (DW_TAG_subprogram)
>>      <2d0>   DW_AT_external    : 1
>>      <2d0>   DW_AT_name        : foo
>>      <2d4>   DW_AT_decl_file   : 1
>>      <2d5>   DW_AT_decl_line   : 8
>>      <2d6>   DW_AT_prototyped  : 1
>>      <2d6>   DW_AT_low_pc      : 0x400650
>>      <2de>   DW_AT_high_pc     : 0x13
>>      <2e6>   DW_AT_frame_base  : 1 byte block: 9c
>> (DW_OP_call_frame_cfa)
>>      <2e8>   DW_AT_GNU_all_call_sites: 1
>>      <2e8>   DW_AT_sibling     : <0x34b>
>>   <2><2ec>: Abbrev Number: 18 (DW_TAG_formal_parameter)
>>      <2ed>   DW_AT_name        : bar
>>      <2f1>   DW_AT_decl_file   : 1
>>      <2f2>   DW_AT_decl_line   : 8
>>      <2f3>   DW_AT_type        : <0x5e>
>>      <2f7>   DW_AT_location    : 0x0     (location list)
>>
>> We see at the last line that the DW_AT_location of the foo formal
>> parameter is a location list at offset 0x0 in the loc table which we
>> can examine using: objdump --dwarf=loc mutatee
>>      00000000 0000000000400650 000000000040065c (DW_OP_reg5 (rdi))
>>      00000013 000000000040065c 0000000000400662 (DW_OP_reg1 (rdx))
>>      00000026 0000000000400662 0000000000400663 (DW_OP_GNU_entry_value:
>> (DW_OP_reg5 (rdi)); DW_OP_stack_value)
>>
>> So from my understanding, we aren't able to correctly parse the third
>> entry of the location list. Digging in the code I found that the
>> keyword
>> DW_OP_GNU_entry_value is not handle by the switch case in
>> dwarf/src/dwarfExprParser.C:126.it then falls in the default case and
>> then return false.
>> This behavior causes the mutator to return an empty vector when asked
>> for the parameters of the function.
>>
>> My setup:
>> gcc version 4.8.1 (Ubuntu/Linaro 4.8.1-10ubuntu9)
>> Dyninst branch v8.2
>>
>> Is this an issue that is within Dyninst's control, more related to
>> libdwarf or something else?
>>
>> Thanks,
>>
>> Francis
>>
>>
>>
>> _______________________________________________
>> Dyninst-api mailing list
>> Dyninst-api@xxxxxxxxxxx
>> https://lists.cs.wisc.edu/mailman/listinfo/dyninst-api
>>
>
>
> --
> --bw
>
> Bill Williams
> Paradyn Project
> bill@xxxxxxxxxxx