https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108447
--- Comment #24 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Andrew Macleod from comment #23) > (In reply to Jakub Jelinek from comment #21) > > (In reply to Andrew Macleod from comment #19) > > > Created attachment 54313 [details] > > > better patch > > > > > > A more consistent approach.. rather than directly call > > > relation_intersect() > > > from multiple places, add the floating point fix to > > > value_relation::intersect and always go through that. > > > > > > This will cause the normal oracle and the path oracle to both make the > > > previous adjustment to intersect for floating point operations. > > > untested, > > > running tests now. > > > > I thought (unless somebody proves otherwise) that relation_intersect is > > actually just fine, the only thing that is incorrect (IMHO) for floating > > point with NaNs are the > > > I thought intersection was wrong for x <= y intersect x >= y because it > produces x == y? A floating point comparison when NANs is possible has 4 possible outcomes: (1) x < y (2) x == y (3) x > y (4) isnan (x) || isnan (y) (unlike say integral which has only the 3 options). Now, each of the C or tree comparisons can be expressed in a table for which of those 4 outcomes it returns true. Say: x < y (1) x <= y (1) || (2) x > y (3) x >= y (2) || (3) x == y (2) x != y (1) || (3) || (4) VARYING (1) || (2) || (3) || (4) UNDEFINED etc. x <= y && x >= y from the above table is ((1) || (2)) && ((2) || (3)) aka (2) aka x == y. While x <= y || x >= y is (1) || (2) || (3) which is something we don't have a VREL_* name for (it is ORDERED_EXPR in trees). Similarly x < y || x > y is (1) || (3) which we don't have either but is LTGT_EXPR on trees. For the integral cases when only (1), (2) and (3) are possible, all we need is 8 relations, LT, LE, GT, GE, EQ, NE, TRUE (VARYING) and FALSE (UNDEFINED) are needed, but for complete coverage of the 4 possibilities we need 16 (32 if it is also about whether qNaN operand(s) raise exceptions or not). E.g. AVX VCMP instruction has those 32 possibilities: Predicate imm8 Description A > B A < B A = B Unordered Signal on qNAN EQ_OQ (EQ) 0H Equal (ordered, non-signaling) F F T F N LT_OS (LT) 1H Less-than (ordered, signaling) F T F F Y LE_OS (LE) 2H Less-than-or-equal (ordered, signaling) F T T F Y UNORD_Q (UNORD) 3H Unordered (non-signaling) F F F T N NEQ_UQ (NEQ) 4H Not-equal (unordered, non-signaling) T T F T N NLT_US (NLT) 5H Not-less-than (unordered, signaling) T F T T Y NLE_US (NLE) 6H Not-less-than-or-equal (unordered, signaling) T F F T Y ORD_Q (ORD) 7H Ordered (non-signaling) T T T F N EQ_UQ 8H Equal (unordered, non-signaling) F F T T N NGE_US (NGE) 9H Not-greater-than-or-equal (unordered, signaling) F T F T Y NGT_US (NGT) AH Not-greater-than (unordered, signaling) F T T T Y FALSE_OQ (FALSE) BH False (ordered, non-signaling) F F F F N NEQ_OQ CH Not-equal (ordered, non-signaling) T T F F N GE_OS (GE) DH Greater-than-or-equal (ordered, signaling) T F T F Y GT_OS (GT) EH Greater-than (ordered, signaling) T F F F Y TRUE_UQ(TRUE) FH True (unordered, non-signaling) T T T T N EQ_OS 10H Equal (ordered, signaling) F F T F Y LT_OQ 11H Less-than (ordered, non-signaling) F T F F N LE_OQ 12H Less-than-or-equal (ordered, non-signaling) F T T F N UNORD_S 13H Unordered (signaling) F F F T Y NEQ_US 14H Not-equal (unordered, signaling) T T F T Y NLT_UQ 15H Not-less-than (unordered, non-signaling) T F T T N NLE_UQ 16H Not-less-than-or-equal (unordered, non-signaling) T F F T N ORD_S 17H Ordered (signaling) T T T F Y EQ_US 18H Equal (unordered, signaling) F F T T Y NGE_UQ 19H Not-greater-than-or-equal (unordered, non-signaling) F T F T N NGT_UQ 1AH Not-greater-than (unordered, non-signaling) F T T T N FALSE_OS 1BH False (ordered, signaling) F F F F Y NEQ_OS 1CH Not-equal (ordered, signaling) T T F F Y GE_OQ 1DH Greater-than-or-equal (ordered, non-signaling) T F T F N GT_OQ 1EH Greater-than (ordered, non-signaling) T F F F N TRUE_US 1FH True (unordered, signaling) T T T T Y I think for VREL_*, we could use just the 16, but for the time being need just some VREL_* value to represent those 8 that don't have a representation right now (i.e. stand for I don't really know what the relation is).
