Richard, this is what you suggested. Thanks. Explicit NANs in the IL can be treated as undefined for flag_finite_math_only. This causes all the right things to happen wrt threading, folding, etc. It also saves us special casing throughout.
It occurs to me that we should do something similar for infinities for -ffinite-math-only. That is, drop them to the min/max representable numbers, and adjust everything (including VARYING endpoints) accordingly. Furthermore, we should saturate to min/max representable in the setter, so (upcoming) binary operators don't have to worry about going over min/max. Sigh...floating point... the gift that keeps on giving. gcc/ChangeLog: * value-range.cc (frange::set): Set known NANs to undefined for flag_finite_math_only. --- gcc/value-range.cc | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/gcc/value-range.cc b/gcc/value-range.cc index 505eb9211a7..7e8028eced2 100644 --- a/gcc/value-range.cc +++ b/gcc/value-range.cc @@ -313,8 +313,13 @@ frange::set (tree min, tree max, value_range_kind kind) gcc_checking_assert (real_identical (TREE_REAL_CST_PTR (min), TREE_REAL_CST_PTR (max))); tree type = TREE_TYPE (min); - bool sign = real_isneg (TREE_REAL_CST_PTR (min)); - set_nan (type, sign); + if (HONOR_NANS (type)) + { + bool sign = real_isneg (TREE_REAL_CST_PTR (min)); + set_nan (type, sign); + } + else + set_undefined (); return; } -- 2.37.1