As discussed in the motivating PR44509:
...we can end up with worse code using fast-math than without. This is because the reassociate pass greedily transforms fsub into fneg/fadd and apparently (based on the regression tests seen here) expects instcombine to clean that up if it wasn't profitable. But we were missing this fold:
(X - Y) - Z --> X - (Y + Z)
There's another, more specific case that I think we should handle as shown in the "fake" fneg test (but missed with a real fneg), but that's another patch. That may be tricky to get right without conflicting with existing transforms for fneg.