This patch fixes a disagreement about the form of REG_BR_PRED.
I came across it as an rtl-checking ICE when inverting a jump
after dbr_schedule.
REG_BR_PRED used to be just a single int (and is still documented
that way, although I'm going to chicken out of fixing that).
dbr_schedule created the notes at the end, so that targets could tell
whether the branch was assumed to be likely-taken or not. However,
no target now uses this information, and all other references to the
notes are for the new 2-int style. This patch therefore removes the
old code as dead.
Tested on mips64-linux-gnu and applied.
Richard
gcc/
* reorg.c (dbr_schedule): Delete REG_BR_PRED loop.
Index: gcc/reorg.c
===================================================================
--- gcc/reorg.c 2012-08-02 21:19:07.320527035 +0100
+++ gcc/reorg.c 2012-08-02 21:21:53.698526627 +0100
@@ -4054,28 +4054,6 @@ dbr_schedule (rtx first)
fprintf (dump_file, "\n");
}
- /* For all JUMP insns, fill in branch prediction notes, so that during
- assembler output a target can set branch prediction bits in the code.
- We have to do this now, as up until this point the destinations of
- JUMPS can be moved around and changed, but past right here that cannot
- happen. */
- for (insn = first; insn; insn = NEXT_INSN (insn))
- {
- int pred_flags;
-
- if (NONJUMP_INSN_P (insn))
- {
- rtx pat = PATTERN (insn);
-
- if (GET_CODE (pat) == SEQUENCE)
- insn = XVECEXP (pat, 0, 0);
- }
- if (!JUMP_P (insn))
- continue;
-
- pred_flags = get_jump_flags (insn, JUMP_LABEL (insn));
- add_reg_note (insn, REG_BR_PRED, GEN_INT (pred_flags));
- }
free_resource_info ();
free (uid_to_ruid);
#ifdef DELAY_SLOTS_FOR_EPILOGUE
