In a recent change gcc became more strict on where trapping instructions can be placed. They are no longer valid on condexprs. This introduced some regressions. The tests have been changed to use -fno-trapping-math. We should investigate if we could optimize them even with trapping math.
The tests are: