On Tue, 12 Nov 2019, Segher Boessenkool wrote: > On Tue, Nov 12, 2019 at 08:47:55AM +0100, Richard Biener wrote: > > OK, but IMHO it is not a good idea to assert UNORDERED_EXPR cannot > > appear with !HONOR_NANS, is it? > > If we allow that, we should allow UNLT/UNGT/UNLE/UNGE/UNEQ/LTGT as well, > which doesn't make much sense, especially because LTGT then is the same > as NE, but NE with HONOR_NANS means something different. > > Without HONOR_NANS all FP comparisons (and FP math in most other aspects) > behaves like integers do. Having to support ORDERED and UNORDERED for > !HONOR_NANS means we need a third codepath in many places. Without this, > the !HONOR_NANS code is very close to the integer code, so it's not all > that bad. > > Assert... Should we have an assert in some strategic places that makes > sure we never try to create NaN stuff when NaNs are disabled?
We do have some asserts in buildN but I guess verify_gimple_comparison might be a good place to catch things early. Richard.
