[Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[IainS]

Hi
IIUC, the following should be the 'correct' approach.
OK for trunk? (with an appropriate changelog,  of course).
Iain

Index: libjava/classpath/configure.ac
===================================================================
--- libjava/classpath/configure.ac	(revision 163330)
+++ libjava/classpath/configure.ac	(working copy)
@@ -563,7 +563,7 @@ if test "x${COMPILE_JNI}" = xyes; then
      dnl CFLAGS that are used for all native code.  We want to compile
      dnl everything with unwinder data so that backtrace() will always
      dnl work.
-    EXTRA_CFLAGS='-fexceptions -fasynchronous-unwind-tables'
+    EXTRA_CFLAGS='-fexceptions -fnon-call-exceptions'
      AC_SUBST(EXTRA_CFLAGS)

      dnl Strict warning flags which not every module uses.

Re: [Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[Richard Guenther]

On Wed, Aug 18, 2010 at 11:30 AM, IainS
<developer@sandoe-acoustics.co.uk> wrote:
> Hi
> IIUC, the following should be the 'correct' approach.
> OK for trunk? (with an appropriate changelog,  of course).
> Iain

That doens't match what the comment says.  What is it the correct
approach for?

> Index: libjava/classpath/configure.ac
> ===================================================================
> --- libjava/classpath/configure.ac      (revision 163330)
> +++ libjava/classpath/configure.ac      (working copy)
> @@ -563,7 +563,7 @@ if test "x${COMPILE_JNI}" = xyes; then
>     dnl CFLAGS that are used for all native code.  We want to compile
>     dnl everything with unwinder data so that backtrace() will always
>     dnl work.
> -    EXTRA_CFLAGS='-fexceptions -fasynchronous-unwind-tables'
> +    EXTRA_CFLAGS='-fexceptions -fnon-call-exceptions'
>     AC_SUBST(EXTRA_CFLAGS)
>
>     dnl Strict warning flags which not every module uses.
>
>

Re: [Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[Andrew Haley]

On 08/18/2010 10:30 AM, IainS wrote:

> IIUC, the following should be the 'correct' approach.
> OK for trunk? (with an appropriate changelog,  of course).

What's the point of this change?

See
http://developer.classpath.org/pipermail/classpath-patches/2008-June/006116.html

Andrew.

Re: [Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[Jakub Jelinek]

On Wed, Aug 18, 2010 at 11:40:15AM +0200, Richard Guenther wrote:
> On Wed, Aug 18, 2010 at 11:30 AM, IainS
> <developer@sandoe-acoustics.co.uk> wrote:
> > IIUC, the following should be the 'correct' approach.
> > OK for trunk? (with an appropriate changelog,  of course).
> > Iain
> 
> That doens't match what the comment says.  What is it the correct
> approach for?

Yeah, I think -fa-u-t is right, otherwise you can't backtrace through
signal handlers.  If Darwin unwinder is lame enough that it can't grok it,
let it change just on Darwin and don't punish sane OSes.

> > --- libjava/classpath/configure.ac      (revision 163330)
> > +++ libjava/classpath/configure.ac      (working copy)
> > @@ -563,7 +563,7 @@ if test "x${COMPILE_JNI}" = xyes; then
> >     dnl CFLAGS that are used for all native code.  We want to compile
> >     dnl everything with unwinder data so that backtrace() will always
> >     dnl work.
> > -    EXTRA_CFLAGS='-fexceptions -fasynchronous-unwind-tables'
> > +    EXTRA_CFLAGS='-fexceptions -fnon-call-exceptions'
> >     AC_SUBST(EXTRA_CFLAGS)
> >
> >     dnl Strict warning flags which not every module uses.
> >
> >

	Jakub

Re: [Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[IainS]

On 18 Aug 2010, at 10:40, Richard Guenther wrote:

> On Wed, Aug 18, 2010 at 11:30 AM, IainS
> <developer@sandoe-acoustics.co.uk> wrote:
>> Hi
>> IIUC, the following should be the 'correct' approach.
>> OK for trunk? (with an appropriate changelog,  of course).
>> Iain
>
> That doens't match what the comment says.  What is it the correct
> approach for?

true, I missed adjusting the comment.

I think that the point is that the underlying requirement is to handle  
non-call-exceptions, and that the method by which that is achieved  
should be left to the target to decide.

[at present -fasynchronous-unwind-tables is set unconditionally by - 
fnon-call-exception in any case (although I'm not sure that's correct  
in all cases)]

>> Index: libjava/classpath/configure.ac
>> ===================================================================
>> --- libjava/classpath/configure.ac      (revision 163330)
>> +++ libjava/classpath/configure.ac      (working copy)
>> @@ -563,7 +563,7 @@ if test "x${COMPILE_JNI}" = xyes; then
>>     dnl CFLAGS that are used for all native code.  We want to compile
>>     dnl everything with unwinder data so that backtrace() will always
>>     dnl work.
>> -    EXTRA_CFLAGS='-fexceptions -fasynchronous-unwind-tables'
>> +    EXTRA_CFLAGS='-fexceptions -fnon-call-exceptions'
>>     AC_SUBST(EXTRA_CFLAGS)
>>
>>     dnl Strict warning flags which not every module uses.
>>
>>

Re: [Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[Richard Henderson]

On 08/18/2010 02:45 AM, Jakub Jelinek wrote:
> On Wed, Aug 18, 2010 at 11:40:15AM +0200, Richard Guenther wrote:
>> On Wed, Aug 18, 2010 at 11:30 AM, IainS
>> <developer@sandoe-acoustics.co.uk> wrote:
>>> IIUC, the following should be the 'correct' approach.
>>> OK for trunk? (with an appropriate changelog,  of course).
>>> Iain
>>
>> That doens't match what the comment says.  What is it the correct
>> approach for?
> 
> Yeah, I think -fa-u-t is right, otherwise you can't backtrace through
> signal handlers.  If Darwin unwinder is lame enough that it can't grok it,
> let it change just on Darwin and don't punish sane OSes.

Except that the signal handlers that java wants to backtrace through
are all synchronous.  They'll never occur in the epilogue of some
function, for instance.  Otherwise libjava would have failed to work
for years and years.


r~

Re: [Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[Andrew Haley]

On 08/18/2010 04:08 PM, Richard Henderson wrote:
> On 08/18/2010 02:45 AM, Jakub Jelinek wrote:
>> On Wed, Aug 18, 2010 at 11:40:15AM +0200, Richard Guenther wrote:
>>> On Wed, Aug 18, 2010 at 11:30 AM, IainS
>>> <developer@sandoe-acoustics.co.uk> wrote:
>>>> IIUC, the following should be the 'correct' approach.
>>>> OK for trunk? (with an appropriate changelog,  of course).
>>>> Iain
>>>
>>> That doens't match what the comment says.  What is it the correct
>>> approach for?
>>
>> Yeah, I think -fa-u-t is right, otherwise you can't backtrace through
>> signal handlers.  If Darwin unwinder is lame enough that it can't grok it,
>> let it change just on Darwin and don't punish sane OSes.
> 
> Except that the signal handlers that java wants to backtrace through
> are all synchronous.  They'll never occur in the epilogue of some
> function, for instance.  Otherwise libjava would have failed to work
> for years and years.

This use of -fasynchronous-unwind-tables is in order to get backtraces
for bugs such as segfaults.  I think such bugs can trigger in prologues
and epilogues.

Andrew.

Re: [Patch, libjava] use -fnon-call-exceptions rather than -fasynchronous-unwind-tables.

[Mike Stump]

On Aug 18, 2010, at 2:45 AM, Jakub Jelinek wrote:
> Yeah, I think -fa-u-t is right, otherwise you can't backtrace through
> signal handlers.  If Darwin unwinder is lame enough that it can't grok it,
> let it change just on Darwin and don't punish sane OSes.

Walking the darwin signal handers works just fine...  I can't recall exactly how well the async signals work v the sync signals.