[PATCH] Fix ira.c indirect_jump_optimize (PR rtl-optimization/70460)
Jakub Jelinek
jakub@redhat.com
Wed Mar 30 22:42:00 GMT 2016
Hi!
As mentioned in the PR, we are miscompiling glibc on i686-linux, because
the new indirect_jump_optimize mini-pass thinks that a insn
which has REG_LABEL_OPERAND note necessarily has to set the target register
to that label, while in the glibc case it is actually that label + some
offset, where the offset is read from a table which contains other labels -
this label differences.
The following patch changes it to just look at SET_SRC of single_set and/or
REG_EQUAL note, and only consider those if one of them is a LABEL_REF.
That alone broke lots of tests, which contain non-local gotos, so I had
to add a check that we don't do anything in this mini-pass (like old ira.c
code did) if there is REG_NON_LOCAL_GOTO note.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk/5/4.9?
Bernd has preapproved the patch, but that was before the REG_NON_LOCAL_GOTO
changes.
2016-03-30 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/70460
* ira.c (indirect_jump_optimize): Don't substitute LABEL_REF
with operand from REG_LABEL_OPERAND, instead substitute
SET_SRC or REG_EQUAL note content if it is a LABEL_REF.
Don't do anything for REG_NON_LOCAL_GOTO jumps.
* gcc.c-torture/execute/pr70460.c: New test.
--- gcc/ira.c.jj 2016-03-21 10:12:32.000000000 +0100
+++ gcc/ira.c 2016-03-30 19:06:42.883892088 +0200
@@ -3870,7 +3870,8 @@ indirect_jump_optimize (void)
FOR_EACH_BB_REVERSE_FN (bb, cfun)
{
rtx_insn *insn = BB_END (bb);
- if (!JUMP_P (insn))
+ if (!JUMP_P (insn)
+ || find_reg_note (insn, REG_NON_LOCAL_GOTO, NULL_RTX))
continue;
rtx x = pc_set (insn);
@@ -3884,19 +3885,18 @@ indirect_jump_optimize (void)
if (!DF_REF_IS_ARTIFICIAL (def))
{
rtx_insn *def_insn = DF_REF_INSN (def);
- rtx note = find_reg_note (def_insn, REG_LABEL_OPERAND, NULL_RTX);
-
- if (note)
+ rtx lab = NULL_RTX;
+ rtx set = single_set (def_insn);
+ if (set && GET_CODE (SET_SRC (set)) == LABEL_REF)
+ lab = SET_SRC (set);
+ else
{
- /* Substitute a LABEL_REF to the label given by the
- note rather than using SET_SRC of DEF_INSN.
- DEF_INSN might be loading the label constant from
- a constant pool, which isn't what we want in a
- direct branch. */
- rtx lab = gen_rtx_LABEL_REF (Pmode, XEXP (note, 0));
- if (validate_replace_rtx (SET_SRC (x), lab, insn))
- rebuild_p = true;
+ rtx eqnote = find_reg_note (def_insn, REG_EQUAL, NULL_RTX);
+ if (eqnote && GET_CODE (XEXP (eqnote, 0)) == LABEL_REF)
+ lab = XEXP (eqnote, 0);
}
+ if (lab && validate_replace_rtx (SET_SRC (x), lab, insn))
+ rebuild_p = true;
}
}
}
--- gcc/testsuite/gcc.c-torture/execute/pr70460.c.jj 2016-03-30 19:08:39.295335567 +0200
+++ gcc/testsuite/gcc.c-torture/execute/pr70460.c 2016-03-30 19:08:18.000000000 +0200
@@ -0,0 +1,29 @@
+/* PR rtl-optimization/70460 */
+
+int c;
+
+__attribute__((noinline, noclone)) void
+foo (int x)
+{
+ static int b[] = { &&lab1 - &&lab0, &&lab2 - &&lab0 };
+ void *a = &&lab0 + b[x];
+ goto *a;
+lab1:
+ c += 2;
+lab2:
+ c++;
+lab0:
+ ;
+}
+
+int
+main ()
+{
+ foo (0);
+ if (c != 3)
+ __builtin_abort ();
+ foo (1);
+ if (c != 4)
+ __builtin_abort ();
+ return 0;
+}
Jakub
More information about the Gcc-patches
mailing list