[Bug rtl-optimization/79388] [6/7 Regression] wrong code with -O -fno-tree-coalesce-vars

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79388

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
--- gcc/combine.c.jj    2017-01-30 09:31:48.000000000 +0100
+++ gcc/combine.c       2017-02-10 12:16:18.507855160 +0100
@@ -14288,6 +14288,11 @@ distribute_notes (rtx notes, rtx_insn *f
                                            NULL_RTX, NULL_RTX, NULL_RTX);
                          distribute_links (LOG_LINKS (tem_insn));

+                         unsigned int regno = REGNO (XEXP (note, 0));
+                         reg_stat_type *rsp = &reg_stat[regno];
+                         if (rsp->last_set == tem_insn)
+                           record_value_for_reg (XEXP (note, 0), NULL,
NULL_RTX);
+
                          SET_INSN_DELETED (tem_insn);
                          if (tem_insn == i2)
                            i2 = NULL;

fixes both testcases by making sure that if we delete the instruction that is
still considered to be the setter of the last value, we also invalidate what
we've remembered for it.  No idea whether this is the right spot or right way
to invalidate that though.  Segher?