[patch] PR30513

[patch] PR30513

[Andreas Tobler]

[-- Attachment #1: Type: text/plain, Size: 597 bytes --]

Hello all,

for details have a look at the PR.

Here comes my proposal for a patch. It is the easiest one in terms of 
testing and modifying files.

Thanks goes to Tom and PinskiA.

Ok for trunk?

I'd like a sparc expert to look at the read/write barriers. Are they 
correct ? Who is one ??


Andreas

2007-01-24  Andreas Tobler  <a.tobler@schweiz.org>

	PR libgcj/30513
	* configure.host: Add forgottten sysdep_dir to sparc. Add a flag to
	libgcj_flags to undefine 'sun' at compile time.
	* sysdep/sparc/locks.h (read_barrier): New functions for 32 and 64 bit
	Sparc.
	(write_barrier): Likewise.

[-- Attachment #2: pr30513-2.diff --]
[-- Type: text/plain, Size: 2081 bytes --]

Index: configure.host
===================================================================
--- configure.host	(revision 121008)
+++ configure.host	(working copy)
@@ -159,6 +159,10 @@
 	enable_hash_synchronization_default=yes
 	;;
   sparc*-*)
+	sysdeps_dir=sparc
+	# On Solaris we have defined 'sun' which later conflicts with 
+	# namespace usage. So to work this away we use the below undefine.
+	libgcj_flags="${libgcj_flags} -Usun"
 	libgcj_interpreter=yes
         ;;
   ia64-*)
Index: sysdep/sparc/locks.h
===================================================================
--- sysdep/sparc/locks.h	(revision 121008)
+++ sysdep/sparc/locks.h	(working copy)
@@ -1,6 +1,6 @@
 // locks.h - Thread synchronization primitives. Sparc implementation.
 
-/* Copyright (C) 2002  Free Software Foundation
+/* Copyright (C) 2002, 2007  Free Software Foundation
 
    This file is part of libgcj.
 
@@ -38,12 +38,23 @@
 }
 
 inline static bool
-compare_and_swap_release(volatile obj_addr_t *addr,
-		  				     obj_addr_t old,
-						     obj_addr_t new_val)
+compare_and_swap_release(volatile obj_addr_t *addr, obj_addr_t old,
+			 obj_addr_t new_val)
 {
   return compare_and_swap(addr, old, new_val);
 }
+
+inline static void
+read_barrier()
+{
+  __asm__ __volatile__("membar #LoadLoad | #LoadStore" : : : "memory");
+}
+
+inline static void
+write_barrier()
+{
+  __asm__ __volatile__("membar #StoreLoad | #StoreStore" : : : "memory");
+}
 #else
 /* Sparc32 implementation, use a spinlock.  */
 static unsigned char __cas_lock = 0;
@@ -109,12 +120,23 @@
 }
 
 inline static bool
-compare_and_swap_release(volatile obj_addr_t *addr,
-		  				     obj_addr_t old,
-						     obj_addr_t new_val)
+compare_and_swap_release(volatile obj_addr_t *addr, obj_addr_t old,
+			 obj_addr_t new_val)
 {
   return compare_and_swap(addr, old, new_val);
 }
+
+inline static void
+read_barrier()
+{
+  __asm__ __volatile__ ("" : : : "memory");
+}
+
+inline static void
+write_barrier()
+{
+  __asm__ __volatile__ ("" : : : "memory");
+}
 #endif /* __arch64__ */
 
 #endif /* ! __SYSDEP_LOCKS_H__ */

Re: [patch] PR30513

[Andrew Haley]

Andreas Tobler writes:
 > Hello all,
 > 
 > for details have a look at the PR.
 > 
 > Here comes my proposal for a patch. It is the easiest one in terms of 
 > testing and modifying files.
 > 
 > Thanks goes to Tom and PinskiA.
 > 
 > Ok for trunk?

OK.

 > I'd like a sparc expert to look at the read/write barriers. Are they 
 > correct ? Who is one ??

The SPARC maintainers, surely.

Andrew.

RE: [patch] PR30513

[Boehm, Hans]

Is there a spec of what these should do?

The patch, taken together with support on other architectures, is quite
inconsistent.

I would expect that the 32 and 64 bit variants would need the same kind
of fences.

It looks like x86-64 and i386 read_barrier is just broken, in that it
doesn't prevent compiler reordering.

I would have expected at least the PowerPC write barrier implementation
to use lwsync.

I'm probably biased, but it seems to me that the atomic_ops package
(used in gc7) is more coherent here.

Hans

> -----Original Message-----
> From: java-patches-owner@gcc.gnu.org 
> [mailto:java-patches-owner@gcc.gnu.org] On Behalf Of Andreas Tobler
> Sent: Wednesday, January 24, 2007 4:33 AM
> To: Java Patches
> Subject: [patch] PR30513
> 
> Hello all,
> 
> for details have a look at the PR.
> 
> Here comes my proposal for a patch. It is the easiest one in 
> terms of testing and modifying files.
> 
> Thanks goes to Tom and PinskiA.
> 
> Ok for trunk?
> 
> I'd like a sparc expert to look at the read/write barriers. 
> Are they correct ? Who is one ??
> 
> 
> Andreas
> 
> 2007-01-24  Andreas Tobler  <a.tobler@schweiz.org>
> 
> 	PR libgcj/30513
> 	* configure.host: Add forgottten sysdep_dir to sparc. 
> Add a flag to
> 	libgcj_flags to undefine 'sun' at compile time.
> 	* sysdep/sparc/locks.h (read_barrier): New functions 
> for 32 and 64 bit
> 	Sparc.
> 	(write_barrier): Likewise.
> 

Re: [patch] PR30513

[Andreas Tobler]

Boehm, Hans wrote:
> Is there a spec of what these should do?

I don't know. My take was to get gcc back in bootstrap land for 
sparc-solaris. I found on this way that these two functions, 
read_barrier and write_barrier were not implemented for. So I tried to 
implement them for sparc.

> The patch, taken together with support on other architectures, is quite
> inconsistent.
> 
> I would expect that the 32 and 64 bit variants would need the same kind
> of fences.

As said, I only tried to get it build again. I see that there are 
differences but I didn't investigate further. Also, due to whatever 
reason, this code was not used until this patch comes in. The 
configure.host bit where it tells sparc* to have sysdep_dir=sparc was 
missing until now. So it always used the generic part for sparc.

> It looks like x86-64 and i386 read_barrier is just broken, in that it
> doesn't prevent compiler reordering.
> 
> I would have expected at least the PowerPC write barrier implementation
> to use lwsync.

Hm.

> I'm probably biased, but it seems to me that the atomic_ops package
> (used in gc7) is more coherent here.

:)

Yeah, but there we lack some functions in the sparc-solaris port.

BTW, you got my ppc64 additions for the atomic_ops package ?

Thanks,
Andreas