This is the mail archive of the gcc-bugs@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[Bug tree-optimization/50599] New: -ftree-vectorize generating incorrect code


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50599

             Bug #: 50599
           Summary: -ftree-vectorize generating incorrect code
    Classification: Unclassified
           Product: gcc
           Version: 4.5.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: a5970694@nepwk.com


MinGW x86 4.5.2
A loop:
while(total_out < decompressed_size)
        {
            printf("%X\n",p);
            int chunk_size = *(int*)p;
            if(chunk_size > 0)
            {            
                int ret = params->decompressor(p+sizeof(int), chunk_size,
params->out+total_out, std::min(params->osize-total_out,params->bsize),
params->other);

                int real_out = (ret / params->ssize) * params->ssize;
                if(real_out != ret)
                    real_out += params->ssize;
                if(total_out + real_out >= params->isize)
                    total_out += ret;
                else
                    total_out += real_out;
                p += chunk_size+sizeof(int);

            }
            else
            {
                if(params->verify)
                {
                    memcpy(params->out+total_out, p+sizeof(int), -chunk_size);
                    p += -chunk_size;
                }
                total_out += -chunk_size;
                p += sizeof(int);
            }
        }
Compiled with -O3 -fno-strict-aliasing works funny - the printf is called twice
with the same address. And on some data I'm getting crashes in
params->decompressor, which is a 3rd party code.
I added other printfs and I see that when running the loop for the 1st time,
the code enters the 1st branch, goes to p += chunk_size+sizeof(int), increases,
goes out of the if and iterates the loop again with unchanged p.

Adding -fno-tree-vectorize solves the problem.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]