I appreciate the information. I did try with O2 and the results were more or less the same, unfortunately. Still about 5x slower than 4.8.2. On Tue, Apr 24, 2018, 7:51 AM Mason <slash.tmp@xxxxxxx> wrote: > On 18/04/2018 23:55, Jack Stalnaker wrote: > > > Yeah, I realize it's a difficult question with no example. It would be > > quite difficult to chop out a chunk of the code showing the 5x slowdown, > > though, I think. I was probing for something big or obvious I may have > > missed, or some known issue with the later versions, since this is so > > dramatic, and since the only thing that has changed is the compiler. The > > compilation options are minimal. I'm using -g -Og, and that's it. I have > > tried a release build with some finer tuning, but that doesn't show much > > improvement. > > A 5x slowdown sounds like SIMD-vs-integer code, for embarrassingly-parallel > source (auto-vectorized vs scalar). > > Although, I think GCC is just now getting some auto-vectorization caps, > and it is almost certain it did not in 4.8 o_O > > As Marc points out, you should test with -O2 instead of -Og > > Regards. >
