cfgexpand: Yet another spot with debug insns references to global vars without varpool nodes [PR105630]

This is similar to the earlier patch to avoid having MEM_EXPRs
referencing global vars without varpool nodes, but this time
the difference is that during gimplification some hashing
actually created DECL_RTLs for the n VAR_DECL and the previous
change was in the if above this when DECL_RTL is NULL and we are
considering creating it.

The following patch drops on the floor references to vars where
we've optimized away the varpool node even when it has DECL_RTL.

Bootstrapped/regtested on x86_64-linux and i686-linux, plus
bootstrapped on those without the cfgexpand.cc change, reapplied
it and rebuilt stage3 cc1/cc1plus, the resulting cc1/cc1plus
binaries on both targets were identical except for the 16-byte
executable_checksum (I've done the second bootstraps in the same
directory as the first one after moving the previous one elsewhere,
so pathnames were the same, just checksum hasn't been regenerated).
So, at least on those binaries this patch doesn't affect debug info
at all.

2022-05-19  Jakub Jelinek  <jakub@redhat.com>

	PR debug/105630
	* cfgexpand.cc (expand_debug_expr): For VAR_DECL, punt for
	global vars without symtab node even when they have DECL_RTL
	set.

	* gcc.dg/pr105630.c: New test.

gcc/cfgexpand.cc

@@ -4575,6 +4575,10 @@ expand_debug_expr (tree exp)
 	      || SYMBOL_REF_DECL (XEXP (op0, 0)) != exp)
 	    return NULL;
 	}
+      else if (VAR_P (exp)
+	       && is_global_var (exp)
+	       && symtab_node::get (exp) == NULL)
+	return NULL;
       else
 	op0 = copy_rtx (op0);
 

gcc/testsuite/gcc.dg/pr105630.c

@@ -0,0 +1,22 @@
+/* PR debug/105630 */
+/* { dg-do compile { target pthread } } */
+/* { dg-options "-O1 -ftree-parallelize-loops=2 -fcompare-debug" } */
+
+int m;
+static int n;
+
+void
+foo (void)
+{
+  int *arr[] = { &n, &n, &n };
+  int unused = n;
+
+  m = 0;
+}
+
+void
+bar (int *arr, int i)
+{
+  while (i < 1)
+    arr[i++] = 1;
+}