User account creation filtered due to spam.
The following was observed for gcc-4_7-branch r185763 as well as for trunk r186667.
The decision-tree in insn-recog.c, when generated from candidate patterns where one candidate (the first in file) pattern has a bare (label_ref ...) (without specified mode) and others (after the first in the file) has (match_operand:SI "operand_known_to_not_match_label_ref" ...) where operand_known_to_not_match_label_ref is for example, nonimmediate_operand or register_operand, inspects the mode of the operand of the insn being matched and calls the predicates despite it knowing they will fail. After the calls, it will correctly match the insn to the first pattern. This will not happen if the predicate is not known to fail or when no mode is specified for the second pattern (then there's the expected test for LABEL_REF). It should instead always first check the operand to be LABEL_REF, matching the first pattern.
I'm labelling this as "missed optimization"; I can't make a strong case that the predicate call should be considered wrong-code, though the argument would be that it could assert, as there's no point calling it for a label_ref as the first pattern will always match.
It could be that this is a special-case and that inspecting mode and calling predicates is seen as cheaper than inspecting the actual operand code and matching any mode, but at face value it doesn't make sense, in particular as the rules are that the first pattern has to match.
See <http://gcc.gnu.org/ml/gcc/2012-04/msg00771.html> for actual generated code and further discussion.