combine: Fix for PR82024

With the testcase in the PR, with all the command line options mentioned
there, a (comparison) instruction becomes dead in fwprop1 but is not
deleted until all the way in rtl_dce.

Before combine this insn look like:

20: flags:CC=cmp(r106:DI,0xffffffffffffffff)
      REG_DEAD r106:DI
      REG_UNUSED flags:CC
      REG_EQUAL cmp(0,0xffffffffffffffff)

(note the only output is unused).

Combining some earlier insns gives

13: r106:DI=0
14: r105:DI=r101:DI+r103:DI

14+13+20 then gives

(parallel [
        (set (reg:CC 17 flags)
            (compare:CC (const_int 0 [0])
                (const_int -1 [0xffffffffffffffff])))
        (set (reg:DI 105)
            (plus:DI (reg/v:DI 101 [ e ])
                (reg:DI 103)))
    ])

which doesn't match; but the set of flags is dead, so combine makes the
set of r105 the whole new instruction, which it then places at i3.  But
that is wrong, because r105 is used after i2 but before i3!  We forget
to check for that in this case.

This patch fixes it.

	PR rtl-optimization/82024
	* combine.c (try_combine): If the combination result is a PARALLEL,
	and we only need to retain the SET in there that would be placed
	at I2, check that we can place that at I3 instead, before doing so.

From-SVN: r251607

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 4cc3ac68d87baff9f8207e295d232f062fac50ce..3f632ca31c2e42fcb68303b524973a8b308dc8e6 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2017-09-01  Segher Boessenkool  <segher@kernel.crashing.org>
+
+       PR rtl-optimization/82024
+       * combine.c (try_combine): If the combination result is a PARALLEL,
+       and we only need to retain the SET in there that would be placed
+       at I2, check that we can place that at I3 instead, before doing so.
+
 2017-09-01  Jakub Jelinek  <jakub@redhat.com>
 
        PR target/81766

diff --git a/gcc/combine.c b/gcc/combine.c
index 86055385d78f51bfe35f8631651ad9d862dc5997..c748c92d2eda7f8a35de905a6d02fec43fe41986 100644 (file)
--- a/gcc/combine.c
+++ b/gcc/combine.c
@@ -3488,7 +3488,10 @@ try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, rtx_insn *i0,
      i3, and one from i2.  Combining then splitting the parallel results
      in the original i2 again plus an invalid insn (which we delete).
      The net effect is only to move instructions around, which makes
-     debug info less accurate.  */
+     debug info less accurate.
+
+     If the remaining SET came from I2 its destination should not be used
+     between I2 and I3.  See PR82024.  */
 
   if (!(added_sets_2 && i1 == 0)
       && is_parallel_of_n_reg_sets (newpat, 2)
@@ -3517,11 +3520,17 @@ try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, rtx_insn *i0,
               && insn_nothrow_p (i3)
               && !side_effects_p (SET_SRC (set0)))
        {
-         newpat = set1;
-         insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
+         rtx dest = SET_DEST (set1);
+         if (GET_CODE (dest) == SUBREG)
+           dest = SUBREG_REG (dest);
+         if (!reg_used_between_p (dest, i2, i3))
+           {
+             newpat = set1;
+             insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes);
 
-         if (insn_code_number >= 0)
-           changed_i3_dest = 1;
+             if (insn_code_number >= 0)
+               changed_i3_dest = 1;
+           }
        }
 
       if (insn_code_number < 0)