[PATCH] Fix SUBREG_BYTE bug in dbxout.c

[Jakub Jelinek]

Hi!

The SUBREG_BYTE patch introduced a bug in dbxout - regno should only be set
for register SUBREGs, not memory SUBREGs.
Previously, the code did:
      if (GET_CODE (value) == REG)
        {
          regno = REGNO (value);
          if (regno >= FIRST_PSEUDO_REGISTER)
            return 0;
          regno += offset;
        }
      alter_subreg (home);
so regno was not set for memory SUBREGs.
I can reproduce it only with gcc-2.96-RH since gcc 3.1 does not pass such
SUBREG to dbxout_symbol_location in this testcase, but it might happen on
some other code.
Ok to commit?

2001-08-03  Jakub Jelinek  <jakub@redhat.com>

	* dbxout.c (dbxout_symbol_location): Only set regno for register
	SUBREG.

	* g++.old-deja/g++.other/debug10.C: New testcase.

--- gcc/dbxout.c.jj	Tue Jul 24 20:41:21 2001
+++ gcc/dbxout.c	Fri Aug  3 20:26:17 2001
@@ -2024,8 +2024,10 @@ dbxout_symbol_location (decl, type, suff
 	{
 	  if (REGNO (value) >= FIRST_PSEUDO_REGISTER)
 	    return 0;
+	  regno = REGNO (alter_subreg (home));
 	}
-      regno = REGNO (alter_subreg (home));
+      else
+	alter_subreg (home);
     }
 
   /* The kind-of-variable letter depends on where
--- gcc/testsuite/g++.old-deja/g++.other/debug10.C.jj	Fri Aug  3 20:31:05 2001
+++ gcc/testsuite/g++.old-deja/g++.other/debug10.C	Fri Aug  3 20:31:12 2001
@@ -0,0 +1,24 @@
+// Build don't link:
+// Special g++ Options: -g -O
+
+struct A
+{
+  int e();
+  A &f(double &);
+};
+
+int A::e()
+{
+  return 0;
+}
+
+A &A::f(double &a)
+{
+  union {
+    double x;
+    char y[8];
+  };
+  y[0] = e();
+  a = x;
+  return *this;
+}

	Jakub