> Divide REG_LABEL notes into REG_LABEL_OPERAND and REG_LABEL_TARGET.
> * reload.c (find_reloads): Generate REG_LABEL_OPERAND, not
> REG_LABEL when replacing an operand with a LABEL_REF for a
> non-jump insn.
> (subst_reloads): When replacing a LABEL_REG with a register,
> instead of generating a REG_LABEL note, assert that there already
> is one or that the label is a known target for the insn.
> * loop.c (add_label_notes): Generate REG_LABEL_OPERAND notes.
> Adjust head comment accordingly.
> (find_and_verify_loops): Test for REG_LABEL_OPERAND notes instead
> of testing insn type and REG_LABEL presence.
> * rtlanal.c (computed_jump_p): Instead of looking for a REG_LABEL
> note, check the JUMP_LABEL field.  Remove "else" after return.
> * reorg.c (emit_delay_sequence): Replace case for REG_LABEL with
> cases for REG_LABEL_OPERAND and REG_LABEL_TARGET.
> (fill_slots_from_thread): Handle both REG_LABEL_OPERAND and
> REG_LABEL_TARGET notes, including the JUMP_TARGET field on JUMP_P
> insns.  Iterate over all notes; don't assume there's only one.
> * flow.c (propagate_block_delete_insn): Check for a
> REG_LABEL_TARGET note referring to a case-table; ignore
> REG_LABEL_OPERAND notes.
> * cse.c (recorded_label_ref): Adjust comment to refer to
> REG_LABEL_OPERAND.
> (cse_basic_block): Do LABEL_REF check for all INSN_P insns, not
> just NONJUMP_INSN_P.
> (check_for_label_ref): For JUMP_P insns, check that the LABEL_REF
> isn't a jump target.
> (rebuild_jump_labels): Adjust head comment.
> (init_label_info): Ditto.  Remove REG_LABEL_OPERAND notes only;
> don't reset REG_LABEL_TARGET notes, including the JUMP_LABEL field.
> (mark_all_labels): For JUMP_P insns without a target, check if the
> the target is noted on the previous nonjump insn.
> (mark_jump_label_1): New function, guts from mark_jump_label.
> <case IF_THEN_ELSE>: Handle first operand as a non-target when
> marking jump target labels.
> <case LABEL_REF>: Adjust for whether to generate a
> REG_LABEL_TARGET or a REG_LABEL_OPERAND note.
> For 'E' format rtl, iterate in descending element order.
> (delete_related_insns): Handle both REG_LABEL_TARGET and
> REG_LABEL_OPERAND notes.  For JUMP_P insns with labels with zero
> reference count, delete and fallthrough.  Move finding-next-
> non-deleted insn last in the function.  Look at all INSN_P insns
> for REG_LABEL_OPERAND notes.
> (redirect_jump_2): Assert that OLABEL equals the old JUMP_LABEL of
> JUMP.
> * print-rtl.c (print_rtx): For JUMP_P insns and a non-NULL
> JUMP_LABEL, output the INSN_UID of it.
> * gcse.c: Adjust comments as appropriate to say REG_LABEL_OPERAND
> and/or REG_LABEL_TARGET.
> (add_label_notes): Only add REG_LABEL_OPERAND notes.  Put in line
> with jump.c copy by only adding notes for labels actually
> referenced in the insn.
> * emit-rtl.c (try_split): Don't assume only NONJUMP_INSN_P need
> usage count increment; handle all INSN_P trial insns.
> (emit_copy_of_insn_after): Change to not copy REG_LABEL_OPERAND
> notes.
> * rtl.h (struct rtx_def) <volatil>: Adjust to mention
> REG_LABEL_TARGET and REG_LABEL_OPERAND.
> (LABEL_REF_NONLOCAL_P): Allow REG_LABEL_TARGET and
> REG_LABEL_OPERAND.
> * combine.c (distribute_notes): Adjust for REG_LABEL_TARGET on
> JUMP_P insns and REG_LABEL_OPERAND everywhere.
> * sched-rgn.c (is_cfg_nonregular): Check for REG_LABEL_OPERANDS
> on all INSN_P insns.
> * reg-notes.def (LABEL_TARGET, LABEL_OPERAND): Split from LABEL.
> * cfgrtl.c (delete_insn): Adjust to handle REG_LABEL_TARGET and
> REG_LABEL_OPERAND notes.
> * reload1.c (calculate_needs_all_insns): Adjust comments.
> (set_label_offsets): Adjust to look for REG_LABEL_OPERAND notes.
> * config/alpha/alpha.md (split for load of an address into a
> four-insn sequence on Unicos/Mk): Adjust to use
> REG_LABEL_OPERAND.
> * config/sh/sh.md (sh_reorg, final_prescan_insn): Ditto.
> * doc/rtl.texi (Insns): Specify when a label_ref makes a jump_insn.
> Similar for what label_refs can go in the JUMP_TARGET field.  Split
> REG_LABEL documentation into REG_LABEL_TARGET and REG_LABEL_OPERAND.

This patch is OK.

I think the code changes you saw were because of the way this patch  
prefers to use JUMP_TARGET rather than adding a REG_LABEL_TARGET  
note.  That's probably a good thing.

:REVIEWMAIL: