libjava failed to build on Linux/x86

[Bryce McKinlay]

Richard Henderson wrote:

>On Thu, Jan 24, 2002 at 03:42:32PM -0800, Boehm, Hans wrote:
>
>>This happened once before.  IIRC, the conclusion then was that it might be
>>better to change "=m" to "+m" and to remove the "0"(*(addr)) argument.
>>
>
>The two are functionally identical, as "+" is expanded to "="/"0".
>With memories, matching constraints are rarely needed or useful.
>

Strangely, one works while the other does not. The following will fail 
without -O but works fine with -O:

#include <stddef.h>

static char
compare_and_swap(volatile size_t *addr, size_t *old, size_t *new_val)
{
  char result;
  __asm__ __volatile__("lock; cmpxchgl %2, %0; setz %1"
               : "=m"(*(addr)), "=q"(result)
               : "r" (new_val), "0"(*(addr)), "a"(old) : "memory");
  return (char) result;
}


The patch below seems to fix it. Does it look ok?

2002-01-25  Bryce McKinlay  <bryce@waitaki.otago.ac.nz>

    * java/lang/natObject.cc (X86 compare_and_swap): Use +m constraint.

Index: natObject.cc
===================================================================
RCS file: /cvs/gcc/gcc/libjava/java/lang/natObject.cc,v
retrieving revision 1.20
diff -u -r1.20 natObject.cc
--- natObject.cc        2001/10/31 00:48:16     1.20
+++ natObject.cc        2002/01/25 02:29:20
@@ -327,8 +327,9 @@
   {
     char result;
     __asm__ __volatile__("lock; cmpxchgl %2, %0; setz %1"
-               : "=m"(*(addr)), "=q"(result)
-               : "r" (new_val), "0"(*(addr)), "a"(old) : "memory");
+               : "+m"(*(addr)), "=q"(result)
+               : "r" (new_val), "a"(old)
+               : "memory");
     return (bool) result;
   }