Java: add flag_use_atomic_builtins

[Andrew Haley]

Some targets, particularly ARM, don't have instructions for things
such as sync_compare_and_swap.  Instead, a routine in libgcc is
used.  This patch adds a new gcj option, -fuse-atomic-builtins,
which causes gcj to use the routines in libgcc whenever atomic
operations are needed.

Andrew.

2009-08-12  Andrew Haley  <aph@...>

        * builtins.c (compareAndSwapInt_builtin): Use
        flag_use_atomic_builtins.
        (compareAndSwapLong_builtin): Likewise.
        (compareAndSwapObject_builtin): Likewise.
        * jvspec.c: Add flag_use_atomic_builtins.
        * gcj.texi: Likewise.
        * java-tree.h: Likewise.
        * lang.opt: Likewise.

Index: java/builtins.c
===================================================================
--- java/builtins.c (revision 150696)
+++ java/builtins.c (working copy)
@@ -318,7 +318,8 @@
    tree orig_call)
 {
   enum machine_mode mode = TYPE_MODE (int_type_node);
-  if (sync_compare_and_swap[mode] != CODE_FOR_nothing)
+  if (sync_compare_and_swap[mode] != CODE_FOR_nothing
+      || flag_use_atomic_builtins)
     {
       tree addr, stmt;
       UNMARSHAL5 (orig_call);
@@ -337,7 +338,12 @@
     tree orig_call)
 {
   enum machine_mode mode = TYPE_MODE (long_type_node);
-  if (sync_compare_and_swap[mode] != CODE_FOR_nothing)
+  if (sync_compare_and_swap[mode] != CODE_FOR_nothing
+      || (GET_MODE_SIZE (mode) <= GET_MODE_SIZE (word_mode)
+  && flag_use_atomic_builtins))
+    /* We don't trust flag_use_atomic_builtins for multi-word
+       compareAndSwap.  Some machines such as ARM have atomic libfuncs
+       but not the multi-word versions.  */
     {
       tree addr, stmt;
       UNMARSHAL5 (orig_call);
@@ -355,7 +361,8 @@
       tree orig_call)
 {
   enum machine_mode mode = TYPE_MODE (ptr_type_node);
-  if (sync_compare_and_swap[mode] != CODE_FOR_nothing)
+  if (sync_compare_and_swap[mode] != CODE_FOR_nothing
+      || flag_use_atomic_builtins)
   {
     tree addr, stmt;
     int builtin;
Index: java/jvspec.c
===================================================================
--- java/jvspec.c (revision 150696)
+++ java/jvspec.c (working copy)
@@ -73,6 +73,7 @@
    %<fclasspath* %<fCLASSPATH* %<fbootclasspath*\
    %<fextdirs*\
    %<fuse-divide-subroutine %<fno-use-divide-subroutine\
+   %<fuse-atomic-builtins %<fno-use-atomic-builtins\
    %<fcheck-references %<fno-check-references\
    %<ffilelist-file %<fsaw-java-file %<fsource* %<ftarget*\
    %{f*} -fdollars-in-identifiers\
Index: java/lang.opt
===================================================================
--- java/lang.opt (revision 150696)
+++ java/lang.opt (working copy)
@@ -192,6 +192,10 @@
 Java Var(flag_use_divide_subroutine) Init(1)
 Call a library routine to do integer divisions

+fuse-atomic-builtins
+Java Var(flag_use_atomic_builtins) Init(0)
+Generate code for built-in atomic operations
+
 fbootstrap-classes
 Java Var(flag_bootstrap_classes)
 Generated should be loaded by bootstrap loader
Index: java/java-tree.h
===================================================================
--- java/java-tree.h (revision 150696)
+++ java/java-tree.h (working copy)
@@ -145,6 +145,9 @@
 /* When nonzero, call a library routine to do integer divisions. */
 extern int flag_use_divide_subroutine;

+/* When nonzero, use atomic builtins. */
+extern int flag_use_atomic_builtins;
+
 /* When nonzero, generate code for the Boehm GC.  */
 extern int flag_use_boehm_gc;

Index: java/gcj.texi
===================================================================
--- java/gcj.texi (revision 150696)
+++ java/gcj.texi (working copy)
@@ -607,6 +607,13 @@
 accessing an object via a reference.  On other systems you won't need
 this because null pointer accesses are caught automatically by the
 processor.
+
+@item -fuse-atomic-builtins
+On some systems, gcc can generate code for built-in atomic operations.
+Use this option to force gcj to use these builtins when compiling Java
+code.  Where this capability is present it should be automatically
+detected, so you won't usually need to use this option.
+
 @end table

 @c man end
z

[Joseph S. Myers]

On Wed, 12 Aug 2009, Andrew Haley wrote:

> Some targets, particularly ARM, don't have instructions for things
> such as sync_compare_and_swap.  Instead, a routine in libgcc is
> used.  This patch adds a new gcj option, -fuse-atomic-builtins,
> which causes gcj to use the routines in libgcc whenever atomic
> operations are needed.

Wouldn't a target hook to allow a target to declare which operations it
provides in libgcc be better than a command-line option?

--
Joseph S. Myers
joseph@...

[Andrew Haley]

Joseph S. Myers wrote:

> On Wed, 12 Aug 2009, Andrew Haley wrote:
>
>> Some targets, particularly ARM, don't have instructions for things
>> such as sync_compare_and_swap.  Instead, a routine in libgcc is
>> used.  This patch adds a new gcj option, -fuse-atomic-builtins,
>> which causes gcj to use the routines in libgcc whenever atomic
>> operations are needed.
>
> Wouldn't a target hook to allow a target to declare which operations it
> provides in libgcc be better than a command-line option?

Maybe it would.  I'm happy to work with anyone to create such a
thing.  What would such a hook look like, and how would a front-
end use it?

Andrew.

[Paolo Bonzini-2]

On 08/12/2009 06:52 PM, Andrew Haley wrote:

> Joseph S. Myers wrote:
>> On Wed, 12 Aug 2009, Andrew Haley wrote:
>>
>>> Some targets, particularly ARM, don't have instructions for things
>>> such as sync_compare_and_swap.  Instead, a routine in libgcc is
>>> used.  This patch adds a new gcj option, -fuse-atomic-builtins,
>>> which causes gcj to use the routines in libgcc whenever atomic
>>> operations are needed.
>> Wouldn't a target hook to allow a target to declare which operations it
>> provides in libgcc be better than a command-line option?
>
> Maybe it would.  I'm happy to work with anyone to create such a
> thing.  What would such a hook look like, and how would a front-
> end use it?

Couldn't optabs be (re)used?

Paolo

[Andrew Haley]

Paolo Bonzini wrote:

> On 08/12/2009 06:52 PM, Andrew Haley wrote:
>> Joseph S. Myers wrote:
>>> On Wed, 12 Aug 2009, Andrew Haley wrote:
>>>
>>>> Some targets, particularly ARM, don't have instructions for things
>>>> such as sync_compare_and_swap.  Instead, a routine in libgcc is
>>>> used.  This patch adds a new gcj option, -fuse-atomic-builtins,
>>>> which causes gcj to use the routines in libgcc whenever atomic
>>>> operations are needed.
>>> Wouldn't a target hook to allow a target to declare which operations it
>>> provides in libgcc be better than a command-line option?
>>
>> Maybe it would.  I'm happy to work with anyone to create such a
>> thing.  What would such a hook look like, and how would a front-
>> end use it?
>
> Couldn't optabs be (re)used?

That would be nice.

Andrew.

[Andrew Haley]

Andrew Haley wrote:

> Paolo Bonzini wrote:
>> On 08/12/2009 06:52 PM, Andrew Haley wrote:
>>> Joseph S. Myers wrote:
>>>> On Wed, 12 Aug 2009, Andrew Haley wrote:
>>>>
>>>>> Some targets, particularly ARM, don't have instructions for things
>>>>> such as sync_compare_and_swap.  Instead, a routine in libgcc is
>>>>> used.  This patch adds a new gcj option, -fuse-atomic-builtins,
>>>>> which causes gcj to use the routines in libgcc whenever atomic
>>>>> operations are needed.
>>>> Wouldn't a target hook to allow a target to declare which operations it
>>>> provides in libgcc be better than a command-line option?
>>> Maybe it would.  I'm happy to work with anyone to create such a
>>> thing.  What would such a hook look like, and how would a front-
>>> end use it?
>> Couldn't optabs be (re)used?
>
> That would be nice.

I guess that's just a matter of

  set_optab_libfunc (sync_compare_and_swap_optab, SImode,, SImode, SImode "__sync_bool_compare_and_swap_4");

or somesuch?

Andrew.

[Boehm, Hans]

At the risk of asking a stupid question, shouldn't all the code inside gcc gradually migrate towards using the C++0x (and probably C1x) atomics, which seem to be generally supported by gcc 4.4?

There are known issues with __sync (no atomic loads and stores, underspecified ordering), which is why there wasn't much of an effort topush the __sync interface into C++0x.

Hans

> -----Original Message-----
> From: java-patches-owner@...
> [mailto:java-patches-owner@...] On Behalf Of Andrew Haley
> Sent: Wednesday, August 12, 2009 10:15 AM
> To: Paolo Bonzini
> Cc: Joseph S. Myers; java-patches@...; Gcc Patch List
> Subject: Re: Java: add flag_use_atomic_builtins
>
> Andrew Haley wrote:
> > Paolo Bonzini wrote:
> >> On 08/12/2009 06:52 PM, Andrew Haley wrote:
> >>> Joseph S. Myers wrote:
> >>>> On Wed, 12 Aug 2009, Andrew Haley wrote:
> >>>>
> >>>>> Some targets, particularly ARM, don't have instructions
> for things
> >>>>> such as sync_compare_and_swap.  Instead, a routine in libgcc is
> >>>>> used.  This patch adds a new gcj option, -fuse-atomic-builtins,
> >>>>> which causes gcj to use the routines in libgcc whenever atomic
> >>>>> operations are needed.
> >>>> Wouldn't a target hook to allow a target to declare which
> >>>> operations it provides in libgcc be better than a
> command-line option?
> >>> Maybe it would.  I'm happy to work with anyone to create such a
> >>> thing.  What would such a hook look like, and how would a
> front- end
> >>> use it?
> >> Couldn't optabs be (re)used?
> >
> > That would be nice.
>
> I guess that's just a matter of
>
>   set_optab_libfunc (sync_compare_and_swap_optab, SImode,,
> SImode, SImode "__sync_bool_compare_and_swap_4");
>
> or somesuch?
>
> Andrew.
>

[Paolo Bonzini-2]

Andrew Haley wrote:

> Andrew Haley wrote:
>> Paolo Bonzini wrote:
>>> On 08/12/2009 06:52 PM, Andrew Haley wrote:
>>>> Joseph S. Myers wrote:
>>>>> On Wed, 12 Aug 2009, Andrew Haley wrote:
>>>>>
>>>>>> Some targets, particularly ARM, don't have instructions for things
>>>>>> such as sync_compare_and_swap.  Instead, a routine in libgcc is
>>>>>> used.  This patch adds a new gcj option, -fuse-atomic-builtins,
>>>>>> which causes gcj to use the routines in libgcc whenever atomic
>>>>>> operations are needed.
>>>>> Wouldn't a target hook to allow a target to declare which operations it
>>>>> provides in libgcc be better than a command-line option?
>>>> Maybe it would.  I'm happy to work with anyone to create such a
>>>> thing.  What would such a hook look like, and how would a front-
>>>> end use it?
>>> Couldn't optabs be (re)used?
>> That would be nice.
>
> I guess that's just a matter of
>
>   set_optab_libfunc (sync_compare_and_swap_optab, SImode,, SImode, SImode "__sync_bool_compare_and_swap_4");
>
> or somesuch?

Yes, that's what I was too lazy to look up.  Then possibly the
middle-end needs to be taught about it rather than the front-end, I
don't know.

Paolo

[Boehm, Hans]

[Partially replying to myself]

> From:  Boehm, Hans
>
> At the risk of asking a stupid question, shouldn't all the
> code inside gcc gradually migrate towards using the C++0x
> (and probably C1x) atomics, which seem to be generally
> supported by gcc 4.4?
>
> There are known issues with __sync (no atomic loads and
> stores, underspecified ordering), which is why there wasn't
> much of an effort topush the __sync interface into C++0x.
>
> Hans
>

OK.  That was largely a stupid question, since we're talking about the compiler implementation of those primitives, which presumably are shared with the atomic<T> implementation?

Hans

[Richard Guenther-2]

On Wed, Aug 12, 2009 at 8:24 PM, Boehm, Hans<hans.boehm@...> wrote:

> [Partially replying to myself]
>> From:  Boehm, Hans
>>
>> At the risk of asking a stupid question, shouldn't all the
>> code inside gcc gradually migrate towards using the C++0x
>> (and probably C1x) atomics, which seem to be generally
>> supported by gcc 4.4?
>>
>> There are known issues with __sync (no atomic loads and
>> stores, underspecified ordering), which is why there wasn't
>> much of an effort topush the __sync interface into C++0x.
>>
>> Hans
>>
> OK.  That was largely a stupid question, since we're talking about the compiler implementation of those primitives, which presumably are shared with the atomic<T> implementation?

I'm not aware of a proper implementation of the C++1x atomics or the memory
model for gcc.

Richard.

[David Daney-4]

Boehm, Hans wrote:

> [Partially replying to myself]
>> From:  Boehm, Hans
>>
>> At the risk of asking a stupid question, shouldn't all the
>> code inside gcc gradually migrate towards using the C++0x
>> (and probably C1x) atomics, which seem to be generally
>> supported by gcc 4.4?
>>
>> There are known issues with __sync (no atomic loads and
>> stores, underspecified ordering), which is why there wasn't
>> much of an effort topush the __sync interface into C++0x.
>>
>> Hans
>>
> OK.  That was largely a stupid question, since we're talking about the compiler implementation of those primitives, which presumably are shared with the atomic<T> implementation?
>

The generic implementation uses __sync*, but some individual targets
have hand coded asm blocks.

David Daney