Re: [PATCH v5 2/2] Implement pahole-like 'ptype /o' option

[Sergio Durigan Junior <sergiodj@redhat.com>]

On Wednesday, December 13 2017, Pedro Alves wrote:

> On 12/13/2017 08:36 PM, Sergio Durigan Junior wrote:
>> On Wednesday, December 13 2017, I wrote:
>
>>> OK, I'll confirm on PPC64BE.
>
> Thanks.
>
>> 
>> It seems like the algorithm for calculating bitfield offsets is not
>> working correctly on BE machines.  This is not only for "ptype /o", but
>> also for regular print commands.  For example, consider this test:
>> 
>>   struct tyu
>>   {
>>     int a1 : 1;
>> 
>>     int a2 : 3;
>> 
>>     int a3 : 23;
>> 
>>     char a4 : 2;
>> 
>>     int64_t a5;
>> 
>>     int a6 : 5;
>> 
>>     int64_t a7 : 3;
>>   };
>> 
>>   int
>>   main (int argc, char *argv[])
>>   {
>>     struct tyu e;
>> 
>>     e.a1 = e.a2 = e.a3 = e.a4 = e.a6 = e.a7 = -1;
>> 
>>     return 0;
>>   }
>> 
>> After stopping GDB at the "return 0;" line, here's what we see when we
>> print "e" on x86_64:
>> 
>>   (gdb) p e
>>   $1 = {a1 = -1, a2 = -1, a3 = -1, a4 = -1 '\377', a5 = 140737488344880, a6 = -1, a7 = -1}
>> 
>> While on PPC64BE:
>> 
>>   (gdb) p e
>>   $1 = {a1 = -1, a2 = 3, a3 = 3, a4 = 3 '\003', a5 = 70367536153528, a6 = -1, a7 = -1}
>> 
>
> You didn't initialize e.a5, so even the x86_64 version looks
> wrong at first.  You're seeing stack/register garbage in
> the padding holes.

I should have initialized e.a5 to 0 in order to make the problem easier
to spot.  I did that now:

$1 = {a1 = -1, a2 = 3, a3 = 3, a4 = 3 '\003', a5 = 0, a6 = -1, a7 = -1}

> You should make that "e" a global to make sure all its
> underlying bytes are clear, including padding.  Or memset it.
> The former is easier.
>
> a2, a3 and a4 in the PPC64 version do look odd.  Though
> maybe that's something do to with the expression you used.
>
> Does it make a difference if you initialize all fields
> with separate statements, like:
>
>  e.a1 = -1;
>  e.a2 = -1;
>  etc.

After memset'ing the variable to 0, and separating all assignments, I
get:

[sergiodj@gcc1-power7 build-gdb]$ g++ -g ptype-offsets.cc 
ptype-offsets.cc: In function ‘int main(int, char**)’:
ptype-offsets.cc:170:8: warning: large integer implicitly truncated to unsigned type [-Woverflow]
   e.a4 = -1;
        ^


$1 = {a1 = -1, a2 = -1, a3 = -1, a4 = 3 '\003', a5 = 0, a6 = -1, a7 = -1}

After changing the declaration of a4 to "signed char a4 : 2;":

$1 = {a1 = -1, a2 = -1, a3 = -1, a4 = -1 '\377', a5 = 0, a6 = -1, a7 = -1}

>> As for "ptype /o", the offsets printed on x86_64 and PPC64BE are the
>> same:
>
> So it sounds like we could remove the x86-64 check in the 
> testcase and let it run on all lp64 targets?  Does it pass
> cleanly on PPC64?

I'm checking this right now, because I have to readjust the test due to
the changes in the output format.

-- 
Sergio
GPG key ID: 237A 54B1 0287 28BF 00EF  31F4 D0EB 7628 65FC 5E36
Please send encrypted e-mail if possible
http://sergiodj.net/