[Bug rtl-optimization/88033] [9 Regression] ICE on valid code at -O2 and -O3 on x86-64-linux-gnu: in remove_some_program_points_and_update_live_ranges, at lra-lives.c:1179

["bergner at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88033

--- Comment #3 from Peter Bergner <bergner at gcc dot gnu.org> ---
So this is a problem when we have a simple copy from a reg to itself, like so:

  (insn (set (reg:DI XXX) (reg:DI XXX)))

By definition, a register does not conflict with itself, but a copy of this
type is confusing the code that is handling non conflicting copies and we're
adding extra program points when we shouldn't.  Since a register cannot
conflict with itself, it's easier to just bail on these types of copies in
non_conflicting_reg_copy_p() rather than trying to fix up the code to handle
them.  I'm regtesting the following patch that fixes the ICE for me.


Index: gcc/ira-lives.c
===================================================================
--- gcc/ira-lives.c     (revision 266207)
+++ gcc/ira-lives.c     (working copy)
@@ -1083,11 +1083,17 @@ non_conflicting_reg_copy_p (rtx_insn *insn)
   int src_regno = REGNO (SET_SRC (set));
   machine_mode mode = GET_MODE (SET_DEST (set));

+  /* By definition, a register does not conflict with itself, therefore we
+     do not have to handle it specially.  Returning NULL_RTX now, helps
+     simplify the callers of this function.  */
+  if (dst_regno == src_regno)
+    return NULL_RTX;
+
   /* Computing conflicts for register pairs is difficult to get right, so
      for now, disallow it.  */
-  if ((dst_regno < FIRST_PSEUDO_REGISTER
+  if ((HARD_REGISTER_NUM_P (dst_regno)
        && hard_regno_nregs (dst_regno, mode) != 1)
-      || (src_regno < FIRST_PSEUDO_REGISTER
+      || (HARD_REGISTER_NUM_P (src_regno)
          && hard_regno_nregs (src_regno, mode) != 1))
     return NULL_RTX;