diff mbox series

Fix ICE due to copy_reg_eh_region_note_forward (PR rtl-optimization/89234)

Message ID 20190207222602.GP2135@tucnak
State New
Headers show
Series Fix ICE due to copy_reg_eh_region_note_forward (PR rtl-optimization/89234) | expand

Commit Message

Jakub Jelinek Feb. 7, 2019, 10:26 p.m. UTC 
Hi!

The following testcase ICEs on ppc64le.  The problem is that
copy_reg_eh_region_note_* functions accept either some instruction, or
REG_EH_REGION note directly.  To differentiate between those it uses INSN_P
test (and returns early if the insn doesn't contain any REG_EH_REGION
notes).  If the function is called on a rtx_insn * that isn't INSN_P, like
on the testcase on a NOTE, it assumes it must be REG_EH_REGION note, uses
XEXP (note, 0) on it, which is actually PREV_INSN in this case and stores
an instruction (JUMP_INSN in this case) into REG_EH_REGION notes it creates.

I believe we should treat rtx_insn * that aren't INSN_P like instructions
that don't have any REG_EH_REGION notes.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2019-02-07  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/89234
	* except.c (copy_reg_eh_region_note_forward): Return if note_or_insn
	is a NOTE, CODE_LABEL etc. - rtx_insn * other than INSN_P.
	(copy_reg_eh_region_note_backward): Likewise.

	* g++.dg/ubsan/pr89234.C: New test.


	Jakub

Comments

Eric Botcazou Feb. 8, 2019, 10:05 a.m. UTC | #1 
> The following testcase ICEs on ppc64le.  The problem is that
> copy_reg_eh_region_note_* functions accept either some instruction, or
> REG_EH_REGION note directly.  To differentiate between those it uses INSN_P
> test (and returns early if the insn doesn't contain any REG_EH_REGION
> notes).  If the function is called on a rtx_insn * that isn't INSN_P, like
> on the testcase on a NOTE, it assumes it must be REG_EH_REGION note, uses
> XEXP (note, 0) on it, which is actually PREV_INSN in this case and stores
> an instruction (JUMP_INSN in this case) into REG_EH_REGION notes it creates.

It took me a few minutes to clear the confusion between NOTE & REG_NOTE here.

> I believe we should treat rtx_insn * that aren't INSN_P like instructions
> that don't have any REG_EH_REGION notes.
>
> 2019-02-07  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR rtl-optimization/89234
> 	* except.c (copy_reg_eh_region_note_forward): Return if note_or_insn
> 	is a NOTE, CODE_LABEL etc. - rtx_insn * other than INSN_P.
> 	(copy_reg_eh_region_note_backward): Likewise.
> 
> 	* g++.dg/ubsan/pr89234.C: New test.

OK, thanks.
diff mbox series

Patch

--- gcc/except.c.jj	2019-01-10 11:43:14.387377695 +0100
+++ gcc/except.c	2019-02-07 15:11:27.756869475 +0100
@@ -1756,6 +1756,8 @@  copy_reg_eh_region_note_forward (rtx not
       if (note == NULL)
 	return;
     }
+  else if (is_a <rtx_insn *> (note_or_insn))
+    return;
   note = XEXP (note, 0);
 
   for (insn = first; insn != last ; insn = NEXT_INSN (insn))
@@ -1778,6 +1780,8 @@  copy_reg_eh_region_note_backward (rtx no
       if (note == NULL)
 	return;
     }
+  else if (is_a <rtx_insn *> (note_or_insn))
+    return;
   note = XEXP (note, 0);
 
   for (insn = last; insn != first; insn = PREV_INSN (insn))
--- gcc/testsuite/g++.dg/ubsan/pr89234.C.jj	2019-02-07 17:12:02.081024666 +0100
+++ gcc/testsuite/g++.dg/ubsan/pr89234.C	2019-02-07 17:08:55.026097949 +0100
@@ -0,0 +1,11 @@ 
+// PR rtl-optimization/89234
+// { dg-do compile { target dfp } }
+// { dg-options "-O2 -fnon-call-exceptions -fsanitize=null" }
+
+typedef float __attribute__((mode (SD))) _Decimal32;
+
+void
+foo (_Decimal32 *b, _Decimal32 c)
+{
+  *b = c + 1.5;
+}