[PATCH] Fix ICE during expand_debug_expr

[Jakub Jelinek]

Hi!

On this, slightly invalid, testcase GCC ICEs with -O2 -g.
For MEM_P op0 negative bitpos can and is handled, but for e.g. x[-1]
when x is in a register simplify_gen_subreg ICEs when seeing the negative
offset.

Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux.  Ok for
trunk?

2009-10-18  Jakub Jelinek  <jakub@redhat.com>

	* cfgexpand.c (expand_debug_expr): Fail if bitpos < 0 for non-MEM
	op0.

	* gcc.dg/debug/vta-3.c: New test.

--- gcc/cfgexpand.c.jj	2009-10-16 01:45:11.000000000 +0200
+++ gcc/cfgexpand.c	2009-10-18 14:02:25.000000000 +0200
@@ -2593,6 +2593,9 @@ expand_debug_expr (tree exp)
 	if (bitpos == 0 && mode == GET_MODE (op0))
 	  return op0;
 
+        if (bitpos < 0)
+          return NULL;
+
 	if ((bitpos % BITS_PER_UNIT) == 0
 	    && bitsize == GET_MODE_BITSIZE (mode1))
 	  {
--- gcc/testsuite/gcc.dg/debug/vta-3.c.jj	2009-10-18 14:44:12.000000000 +0200
+++ gcc/testsuite/gcc.dg/debug/vta-3.c	2009-10-18 14:44:32.000000000 +0200
@@ -0,0 +1,17 @@
+/* { dg-do compile } */
+
+int
+foo (void)
+{
+  union { char e[8]; int i; } a, b;
+  char *c, *d;
+  unsigned int i;
+  c = a.e;
+  d = &b.e[sizeof (int) - 1];
+  for (i = 0; i < sizeof (int); i++)
+    {
+      *d = *c++;
+      --d;
+    }
+  return b.i;
+}


	Jakub