Re: Safety of enabling JIT

[Rik]

On 12/08/2013 10:26 PM, John W. Eaton wrote:
> On 12/09/2013 12:10 AM, Rik wrote:
>> On 12/08/2013 10:00 AM, address@hidden wrote:
>>> Message: 2
>>> Date: Sun, 8 Dec 2013 00:33:14 +0200
>>> From: Susi Lehtola <address@hidden>
>>> To: Octave Maintainers List <address@hidden>
>>> Subject: Safety of enabling JIT
>>> Message-ID: <address@hidden>
>>> Content-Type: text/plain; charset=US-ASCII
>>>
>>> Hi,
>>>
>>>
>>> configure reports
>>>    --enable-jit            (EXPERIMENTAL) enable JIT compiler
>>> but this seems to be the default when llvm is available.
>>>
>>> How safe is the option nowadays?
>>
>> It *seems* safe.  First, even when the JIT compiler is built (configure
>> time decision) the compiler is not enabled at runtime.  To do so a user
>> either needs to invoke octave with the option --jit-compiler or use
>> "enable_jit (1)" from within Octave.  Second, there are a series of JIT
>> verification tests and they all pass.  The only contra-indication has been
>> some memory leaks and weirdness when running JIT under valgrind.  Although
>> this may point to a problem, it has never manifested itself in normal
>> operation.  It might just be something about the valgrind environment.
>>>   Can we enable it for the Fedora
>>> package?
>> I think so, for the reasons above.
>
> If we are going to recommend enabling it, then I think we should make
> that the default.
>
> I still think the JIT compiler should be disabled by default, and we need
> to be sure that enabling the code but not running it doesn't cause any
> trouble.
This is the set-up I've been running for months where JIT is compiled in,
but not enabled at run-time.  It's been fine for me.  I'm not running
production code, but I do regularly execute the test suite and haven't had
any failures that could be attributed to JIT.
>
> Does linking with the LLVM libraries cause any trouble for Octave?
Not on my Linux machine.  There is a bug report about using JIT on MinGW
(https://savannah.gnu.org/bugs/?40608).  This, however, is about compiling
and using it at run-time rather than letting it sit dormant.
>
> When it is disabled at run time, we are sure it doesn't have any effect
> on Octave?
I'm only one data point.  I think Michael Godfrey runs with JIT actually
enabled at run-time and he hasn't reported problems.  Is anyone else
running with LLVM compiled in and JIT either on or off?

--Rik