Re: Optimization o2 vs o3 issue

[Manish Baphna <manish_baphna at yahoo dot com>]

Hi
1.Its NOT multi threaded application.

2. I created a small example where this issue is visible. Let me know what you think after it.

///////////// bug.cpp ///////////
#include <iostream>
#include <string>
#include <stdio.h>

using namespace std ;

typedef unsigned char uint8_t ;

class DBus {
public:
   uint8_t data[16] ;
};

int main(){
         DBus dout ;  
         long long unsigned upper_data = 0xFFFFFFFff ; 
	 long long unsigned int mask = 0xFF ;

	 for(int count=0 ; count < 8 ; count++)
         {
	    dout.data[15 - count] = (uint8_t)((upper_data & (mask << ((8+count)*8)))>>((8+count)*8));
          //cout << "Dout data " << (uint64_t)((dout.data)[15-count]) << endl; // Line to be played with
         }

         for(int count=0 ; count < 8 ; count++)
         {
            cout << "Dout data " << (uint64_t)((dout.data)[15-count]) << endl;
         }




   return 1 ;		
};

///////////// end of bug.cpp /////////
Step1 :   Compile using o2/o3 
bash>> g++-4.1.1 -g3 -O2 -fPIC -Wall -W -Wcast-qual -Wpointer-arith 
bash>> g++-Woverloaded-virtual -Werror -m64 -Werror bug.cpp
bash>> ./a.out
Dout data 255
Dout data 255
Dout data 255
Dout data 255
Dout data 15
Dout data 0
Dout data 0
Dout data 0
bash>> g++-4.1.1 -g3 -O3 -fPIC -Wall -W -Wcast-qual -Wpointer-arith 
bash>> g++-Woverloaded-virtual -Werror -m64 -Werror bug.cpp
bash>> ./a.out
Dout data 0
Dout data 0
Dout data 0
Dout data 0
Dout data 0
Dout data 0
Dout data 0
Dout data 0

Now if you uncomment the cout line , o3 would also work !!


Let me know what you guys think of it.

Best Regards,
Manish




--- On Mon, 22/12/08, Andrew Haley <aph@redhat.com> wrote:

> From: Andrew Haley <aph@redhat.com>
> Subject: Re: Optimization o2 vs o3 issue
> To: manish_baphna@yahoo.com
> Cc: gcc-help@gcc.gnu.org, "John Fine" <johnsfine@verizon.net>
> Date: Monday, 22 December, 2008, 4:50 PM
> Manish Baphna wrote:
> 
> > A quick update, Though I couldn't really found
> real issue , I found
> > an interesting solution which worked. magic word is
> 'volatile' :)
> >
> >
> > Original two lines were
> >
> >   dout.data[15 - count]
> >     = (uint8_t)(((long long unsigned int)
> > 		 (*(soc_cpr_dbusdataout_->sig_value_upper))
> > 		 & (mask <<
> ((8+count)*8)))>>((8+count)*8));
> >
> >   cout << "Dout data " <<
> (uint64_t)((dout.data)[15-count]) << endl;
> >
> > What I did was inserting this line just above Line#1
> > volatile uint32_t shift_val = ((8+count)*8) ;
> > and used this in above Line#1 , and then I removed
> Line#2 and this time it worked.
> >  ( First I tried without using 'volatile' and
> it didn't help )
> > Another workaround that did the same thing was making
> > count as volatile.
> 
> Interesting.  Is this a multi-threaded application?  The
> other
> possibility is an array bounds overflow.
> 
> > These two lines are in for loop which goes for count =
> 0 to 7 , I
> > don't know why compiler is being lazy here to read
> it everytime
> > unless I make it volatile.
> 
> The compiler isn't supposed to read it everytime unless
> you make it
> volatile.
> 
> > Should it be logged as bug in gcc somewhere ?
> 
> Not until you can come up with a test cases that someone
> else can run.
> 
> Andrew.


      Be the first one to try the new Messenger 9 Beta! Go to http://in.messenger.yahoo.com/win/