Bug: Wrong Mask initialization under SSE + Vc Master + -ffast-math

[sawenzel]

Vc-revision: 080116dec206dca965f01a9b19e6885e0f9a0cac OS: CentOS7 Compiler: g++ (GCC) 4.8.5 20150623 (Red Hat 4.8.5-4) Flags: see attachment ( combination of sse + -ffast-math ) CPU: does not matter (reproduced on different architectures)

Observation

There is a problem in mask initializations under SSE with latest master in combination with fast math compiler flag.

This problem goes away if either a) compile for AVX b) remove fast-math c) go to release tag 1.2.0

(see attached source code and compilation script for details) package.zip

Testcase

#include <Vc/Vc>
using Vec = Vc::double_v;
int main() {
  Vec a,b;
  for (size_t i=0;i<Vec::Size;++i) {
      a[i] = (i % 2) == 0 ? 1 : 4;
      b[i] = 4;
  }
  auto m = a < b;
  std::cerr << m << "\n";
}

Actual Results

m[00]

Expected Results

m[10]

[mattkretz]

Thanks Sandro, I'll see what I can find out. The likely relevant change after the 1.2 release is the change in the subscript operator return type. Instead of returing a type-punned lvalue reference to EntryType, it returns a "smart reference" proxy type to avoid the type punning. So, in theory, the master version should be correct while 1.2 relies on a compiler extension (otherwise 1.2 would invoke undefined behavior on subscript assignment) which lead to multiple GCC bug reports already (and major new issues starting with GCC 6).