https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120948
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> --- A better representation than if (...) __builtin_unrechable (); for asserts would be nice. We have with __builtin_assume_aligned something very related. So in GIMPLE we could have _2 = .ASSUME (_1, !=, 0); with != encoded as integer tree code. But of course if we'd want to expose this to the user it might instead be x = __builtin_assume (x != 0); with the syntactic requirement that the argument is a compare and the left operator is a ID (IIRC OpenMP pragmas often have similar constraints). Of course there's no need to expose this to users in the first place. Having _3 = _1 != 0; _2 = .ASSUME (_1, _3); is more difficult because the separate predicate stmt will confuse size/time metrics. So _2 = 1 / x_1 would be turned into x_3 = .ASSUME (x_1, !=, 0); _2 = x_3 == 1; note that nothing prevents code motion of the above and the conditional would still be if (x_1 == 0) return 5; which means unless we rewrite all uses of x_1 to x_3 it might not achieve what we intend to.
