Currently bisecting this issue, the range I have right now is
3db653a2c235207b347849200eeb3f89bb4ba51 ..
9dbaa10aa755f33f8050082bb443d1f9eabf6e21

That's a 3 minute commit time difference, but containing 216 commits.
Nothing immediately obvious to my eyes. Many (most?) of the intermediate
commits die well before they would start building gcc-4.9.4, making
bisection incredibly tedious.