Re: jit test failure octave 3.8 on windows

[Stefan Mahr]

Philip Nienhuis wrote:
> tmacchant wrote
>> Hello
>>
>> On the thread concerning GNU Octave in Japan, there was a report for jit
>> feature of octave-3.8.0 downloaded from http://mxeoctave.osuv.de/ 
>>
>> octave-3.8.0-5, octave-3.8.1-rc3
>> Test
>>
>>> a = 1; b = 1; tic; for i=1:1000; for j=1:1000; a = a + b; end; end; toc 
>> Elapsed time is 1.5 seconds. 
>>> jit_enable(1) 
>>> a = 1; b = 1; tic; for i=1:1000; for j=1:1000; a = a + b; end; end; toc
>> (No response)
>>
>> I have tested the same test on Ubuntu 12.04 LTS (32bit) on octave-3.8.0
>> built by myself on my old notebook .
>>
>> octave:1> a = 1; b = 1; tic; for i=1:1000; for j=1:1000; a = a + b; end;
>> end; toc 
>> Elapsed time is 4.21553 seconds. 
>> octave:2> jit_enable(1) 
>> octave:3> a = 1; b = 1; tic; for i=1:1000; for j=1:1000; a = a + b; end;
>> end; toc 
>> Elapsed time is 0.891052 seconds. 
>>
>> In this case jit feature worked correctly.
>>
>> Have anyone met the same phenomena on mxe-build mingw octave ?
> 
> Hi Tatsuro,
> 
> Yes I think I've reported a similar issue already (IIRC) some months ago for
> the 3.7.7+ series, see here:
> https://mailman.cae.wisc.edu/pipermail/octave-maintainers/2013-November/037260.html
> 
> I hit it when running __run_test_suite__.m. 
> Symptoms:
> - it seems that jit.tst hangs, but:
> - is doesn't, after 20 minutes or so it segfaults.
> IIRC it happened already at the 3rd (still very simple) test in jit.tst.
> More details in bug #40608
> 

Even if it not segfaults, sometimes the results are wrong. Especially if
complex numbers are involved.


> Moving jit.tst out of the way (from share/octave/3.8.0/etc/tests/fixed to
> e.g., share/octave/3.8.0) makes __run_test_suite__.m finish.
> 
> At the time one of the replies mentioned that JIT still has many problems,
> at least on Windows. The advice was to ignore it for the time being (see
> here:
> http://lists.gnu.org/archive/html/octave-bug-tracker/2013-11/msg00519.html)
> 
> On the patch tracker there are improvements for llvm/JIT, made in the course
> of GSoC last year, that haven't been merged yet. Perhaps these patches
> contain fixes that make llvm work better on Windows, who knows.

You mean http://savannah.gnu.org/patch/?8271 ? This is already merged.
The support for integer types is still unmerged.

Also, there's a workaround for compiling functions like sin, exp and
norm. (http://savannah.gnu.org/bugs/?41312)

Unfortunately no bugfixes for MinGW or other platforms than linux.

I stopped to look deeper into this bug, since I don't use Windows.
Furthermore David Malcolm is working on jit support using his libgccjit,
so LLVM is obsolete :)

Stefan