Re: [trans-mem] fix futex syscall for x86_64

[Patrick MARLIER <patrick dot marlier at unine dot ch>]

Hello,

In 32 bit version, esi (4th parameter, ie timeout) is set to 0 even if 
it is wait or wake.
So I guess you might do the same in 64bit. Otherwise, your previous 
patch seems ok if you want to avoid the useless operation in case of 
wake operation.

Patrick Marlier.

	* futex_bits.h (sys_futex0): Always set 4th parameter (r10) to 0 for 
futex system call.

Index: ../transactional-memory/libitm/config/linux/x86/futex_bits.h
===================================================================
--- ../transactional-memory/libitm/config/linux/x86/futex_bits.h 
(revision 157407)
+++ ../transactional-memory/libitm/config/linux/x86/futex_bits.h 
(working copy)
@@ -30,11 +30,13 @@
 static inline long
 sys_futex0 (int *addr, long op, long val)
 {
+  register long r10 __asm__("%r10");
   long res;

+  r10 = 0;
   __asm volatile ("syscall"
                  : "=a" (res)
-                 : "0" (SYS_futex), "D" (addr), "S" (op), "d" (val)
+                 : "0" (SYS_futex), "D" (addr), "S" (op), "d" (val), 
"r" (r10)
                  : "r11", "rcx", "memory");

   return res;

Aldy Hernandez wrote:
> On Mon, Mar 22, 2010 at 10:09:40AM -0700, Richard Henderson wrote:
> > On 03/22/2010 09:01 AM, Aldy Hernandez wrote:
> > > +/* Define sys_futex0_wait if backend's wait syscall is different than
> > > +   the wake.  See x86_64 for an example.  */
> > > +#ifndef sys_futex0_wait
> > > +#define sys_futex0_wait sys_futex0
> > > +#endif
> > Not ok.  All targets have the same arguments.
>
> I don't get it.  How am I changing the number of arguments passed to
> sys_futex0?
>
> Do you want me to set r10 to 0 both for the x86-64 futex_wait and for
> the futex_wake?  On libgomp/config/linux/x86/futex.h we set r10 only for
> the futex_wait variant.
>
> Confused.
> Aldy