Nope, the reason was nondeterminism - it happened on the Core2Quad running 5.9-rc4 while trying to compile todays Linux from git.This is 5.9.0-rc3-00091-ge28f0104343d on Lenovo t460s that has ran fine up to 5.8.0.
Now I reproduced the same problem with 5.9.0-rc3 on a HP desktop with Core2Quad CPU. The call trace is very similar and it's crashing gcc again while compiling 5.9-rc4.
But it seems 5.9-rc4 cures it here as well - whatever the reason might have been.