JayDDee

Putting a NOP just before the call to v1 seems to have done the trick to workaround GCC's optimizer and allow both versions to coexist. Now I have to clean...

I now have a working build where both versions are present and selected appropriately. This was a lot more difficult that it should have been.

Things are starting to settle down. The final implementation includes 2 copies of essentially the same function. They differ only in how they manage non-contiguous data. One version, preferred by...

I wouldn't call them bizarre but I am familiar with competition between the compiler and the CPU. The compiler clearly misjudged the effect of the differences in the 2 functions...

Still pondering this issue. If I send a bug report to gcc I have to do a lot more work first. That won't be any time soon. I'm still monitoring...

Maybe the comment was for another weird issue where the compiler optimized beyond it's guaranteed data alignment. Specifically, vector optimzation replaced a loop with AVX2 code which requires 256 bit...

The use of Nway notation in hash functions is being changed to Nx64 or Nx32 where appropriate. This notation is already used for interleave functions. This is needed for algos...

This is typically a missing package. Review the build procedure and following it, make note of any deviations. Also provide your CPU model, OS, GCC version and any other relevant...

It was never a problem before, even when GCC-6 was new. They all seem to be inline functions that aren't static. I have no idea why this matters. In fact...

What do you mean no? That's WAY off script! Just use the default build procedure as I suggested in my first reply.