https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97926
--- Comment #3 from acsawdey at gcc dot gnu.org ---
So the underlying problem here is that the unordered comparisons are not
allowed with -ffinite-math-only due to this predicate:
;; Return 1 if OP is a comparison operation that is valid for a branch
;; instruction. We check the opcode against the mode of the CC value.
;; validate_condition_mode is an assertion.
(define_predicate "branch_comparison_operator"
(and (match_operand 0 "comparison_operator")
(match_test "GET_MODE_CLASS (GET_MODE (XEXP (op, 0))) == MODE_CC")
(if_then_else (match_test "GET_MODE (XEXP (op, 0)) == CCFPmode")
(if_then_else (match_test "flag_finite_math_only")
(match_code "lt,le,gt,ge,eq,ne,unordered,ordered")
(match_code "lt,gt,eq,unordered,unge,unle,ne,ordered"))
(match_code "lt,ltu,le,leu,gt,gtu,ge,geu,eq,ne"))
(match_test "validate_condition_mode (GET_CODE (op),
GET_MODE (XEXP (op, 0))),
1")))
But ubsan_instrument_float_cast() generates this:
t = fold_build2 (UNLE_EXPR, boolean_type_node, expr, min);
tt = fold_build2 (UNGE_EXPR, boolean_type_node, expr, max);
which eventually leads to the ICE. Even if this branch wasn't rewritten by
patch_dump_insn() it would not be recognized and would eventually ICE.
Segher is working on a change to that predicate for PR98092 though which may be
a workaround fix for this.
