Re: [boost] [gil] gcc fail with simple code

Subject: Re: [boost] [gil] gcc fail with simple code
From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2018-03-28 18:01:57

On 03/28/18 20:52, Andrey Semashev wrote:
> On 03/28/18 20:30, Christian Henning via Boost wrote:
>> Hi all,
>>
>> we are, at the boost::gil corner, investigating a runtime failure when
>> compiling with gcc 5.4. When running in release configuration the 
>> following
>> code fails:
>>
>> #include <iostream>
>> #include <boost/gil/gil_all.hpp>
>>
>> using namespace boost::gil;
>> using namespace std;
>>
>> void error_if(bool condition) {
>> Â Â Â Â  if (condition)
>> Â Â Â Â Â Â Â Â  throw std::exception();
>> }
>>
>> bits32s c32s_min = channel_traits<bits32s>::min_value();
>> bits32s c32s_max = channel_traits<bits32s>::max_value();
>>
>> // For channel values simply initialize the value directly
>> template <typename ChannelValue>
>> struct value_core {
>> Â Â Â Â  typedef ChannelValue channel_t;
>> Â Â Â Â  channel_t _min_v, _max_v;
>>
>> Â Â Â Â  value_core() : _min_v(channel_traits<ChannelValue>::min_value()),
>> _max_v(channel_traits<ChannelValue>::max_value())
>> Â Â Â Â  {}
>> };
>>
>> int main(int argc, char *argv[])
>> {
>> Â Â Â Â  value_core<bits32s> a;
>>
>> Â Â Â Â  bits32s v_min, v_max;
>>
>> Â Â Â Â  v_min = channel_convert<bits32s>(c32s_min);
>> Â Â Â Â  v_max = channel_convert<bits32s>(c32s_max);
>>
>> Â Â Â Â  //std::cout << v_min << std::endl;
>> Â Â Â Â  // std::cout << this->_min_v << std::endl;
>> Â Â Â Â  //std::cout << v_max << std::endl;
>> Â Â Â Â  // std::cout << this->_max_v << std::endl;
>>
>> Â Â Â Â  error_if(v_min != a._min_v || v_max != a._max_v);
>> }
>>
>> The interesting thing is that everything works with the cout's enabled.
>>
>> Has anyone ever dealt with a similar issue?
> 
> Most likely, you have a signed integer overflow somewhere. The code 
> crashes on gcc 7.2 with -O3 but doesn't crash with -O0 or with -O3 -fwrapv.

With -ftrapv the code crashes at this point:

#0  __GI_raise (sig=sig_at_entry=6) at ../sysdeps/unix/sysv/linux/raise.c:51
#1  0x00007ffff7490f5d in __GI_abort () at abort.c:90
#2  0x00007ffff783b2d8 in __addvsi3 () from 
/lib/x86_64-linux-gnu/libgcc_s.so.1
#3  0x00005555555548a9 in 
boost::gil::detail::channel_convert_to_unsigned<int>::operator() 
(this=<optimized out>, x=<optimized out>) at 
./boost/gil/channel_algorithm.hpp:332
#4  boost::gil::channel_converter<int, int>::operator() (this=<optimized 
out>, src=@0x555555755014: 2147483647) at 
./boost/gil/channel_algorithm.hpp:368
#5  boost::gil::channel_convert<int, int> (src=@0x555555755014: 
2147483647) at ./boost/gil/channel_algorithm.hpp:377
#6  main (argc=<optimized out>, argv=<optimized out>) at test_gil.cpp:32

channel_algorithm.hpp:332 is this line:

     type operator()(bits32s x) const { return 
static_cast<bits32>(x+(1<<31)); }

This shift is UB because it overflows (until C++17, I think?). The 
addition will also overflow unless x is 0.

I would suggest casting to unsigned first and then performing the math.