crash with dead code from compressed oops in gcm

During promotion for b32 crashes on linux-amd64 and windows-amd64 occurred during the build when executing pack. There seems to be some cloning of loadN nodes going on during matching that should be and this result in some dead loadN nodes still connected to the graph. gcm dies in this code:

      uint use_rpo = _bbs[use->_idx]->_rpo;

because they haven't been scheduled yet so _bbs returns NULL. The use is a dead loadN.

 1123 loadN === _ o711 893 [[ 1121 ]] narrowoop: java/util/HashMap:NotNull:exact *
 1122 loadN === _ o711 893 [[ 1121 ]] narrowoop: java/util/HashMap:NotNull:exact *
o711 Proj === o709 [[o881 o720 o4488 o788 o733 o785 o783 o965 o772 o878 o756 o838 o848 o823 o980 o950 o908 o912 o4467 o4473 o4519 o4565 o4631 o4784 o4791 892 1122 1123 1121 ]] #2 Memory: @BotPTR *+bot, idx=Bot;
o4241 IfFalse === o4240 [[o4480 o4473 1121 ]] #0
 1121 incI_mem === o4241 o711 1122 1123 [[]] memory Memory: @java/util/HashMap:NotNull:exact+32 *, name=modCount, idx=24;

The incI_mem somehow wasn't matched correctly since it ended up with two different copies of the loadN instead of two references to the same node. The del_req call in incI_memNode::Expand removes one reference which causes it to go dead. So the problem seem to be that matching gives us two copies instead of one.