[Symphony] Suspicious Code

[Konvicka Filip]

Hi,

While cleaning some warnings I stumbled upon several very suspicious lines of code in Symphony 5.5.7.

Note that item 1 is a change in the headers, and unless this is fixed this causes warnings in programs that link libsymphony.
Items 2 - 4 are possible bugs in Symphony.

1.  In OsiSolverInterface.hpp and OsiClpSolverInterface.hpp, there are functions defined with the default parameter values of std::string::npos.  The type of this is size_t, but the argument for which the default is defined is unsigned. In 64bit builds, this leads to a truncation of the value.

    virtual std::string getRowName(int rowIndex,
                              unsigned maxLen = static_cast<unsigned>(std::string::npos)) const;

I think that it is better to use

    virtual std::string getRowName(int rowIndex,
                              unsigned maxLen = std::numeric_limits<unsigned>::max()) const;

If you look at how COIN_INT_MAX is defined in CoinFinite.hpp, this is seems to be an acceptable COIN coding practice.  There are 5 occurrences in total of this (in OsiSolverInterface.hpp and OsiClpSolverInterface.hpp).  The only use of this that I could find is OsiNames.cpp which uses this as a limit to std::string::substr.  Also this change should not break the ABI since this does not affect the symbols.

2.  In lp_heuristics.c, line 1410:

       if(dive_depth > dive_depth_limit/2){
          fix_incr_cnt =(int)(1.0*(frac_ip_cnt - d_fixed_cnt)/
                           (dive_depth_limit > (dive_depth ? dive_depth_limit - dive_depth : 1)));
          if(fix_incr_cnt < 1) fix_incr_cnt = 1;
       }

Look at the division.  The divider is a bool value! (dive_depth_limit > (dive_depth ? dive_depth_limit - dive_depth : 1))
I think that most likely this should have been (dive_depth_limit > dive_depth) ? dive_depth_limit-dive_depth : 1, which by the way is equivalent to std::max(dive_depth_limit - dive_depth, 1).

       if(dive_depth > dive_depth_limit/2){
          fix_incr_cnt =(int)(1.0*(frac_ip_cnt - d_fixed_cnt)/
                           ((dive_depth_limit > dive_depth) ? dive_depth_limit - dive_depth : 1));
          if(fix_incr_cnt < 1) fix_incr_cnt = 1;
       }

By the way, the warning that caught the compiler's attention was - lp_heuristics.c(1412): warning C4804: '/' : unsafe use of type 'bool' in operation

3. In master_prep_base.c:504, I can also see something very suspicious:

       if(P->ub < (P->mip->obj_sense == SYM_MAXIMIZE) ? -(P->mip->obj_offset) :
           P->mip->obj_offset){

This is actually evaluated as (P->ub < (P->mip->obj_sense == SYM_MAXIMIZE)) ? -(P->mip->obj_offset) : P->mip->obj_offset, but obviously the intention was

       if(P->ub < ((P->mip->obj_sense == SYM_MAXIMIZE) ? -(P->mip->obj_offset) :
           P->mip->obj_offset)){

Again the warning was master_prep_base.c(504): warning C4804: '<' : unsafe use of type 'bool' in operation

4. In master.c:231 and 233, there are assignments like this:

   tm_par->warm_start_node_limit = (int)SYM_INFINITY;

warm_start_node_limit is an int, while SYM_INFINITY is defined as 1e20 (i.e. a double).  This does not make much sense (the conversion results in some strange value), so I think that it's much safer to write

   tm_par->warm_start_node_limit = INT_MAX;

Note that 1e20 mod (UINT_MAX+1) is the number that gets assigned, which is 1661992960 for 4byte ints but zero for 2byte ints.  Lucky we don't have small ints anymore :-)

Cheers,
Filip

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/symphony/attachments/20140916/dba489ee/attachment.html>